Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
58687acb DZ |
2 | /* |
3 | * Detect hard and soft lockups on a system | |
4 | * | |
5 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. | |
6 | * | |
86f5e6a7 FLVC |
7 | * Note: Most of this code is borrowed heavily from the original softlockup |
8 | * detector, so thanks to Ingo for the initial implementation. | |
9 | * Some chunks also taken from the old x86-specific nmi watchdog code, thanks | |
58687acb DZ |
10 | * to those contributors as well. |
11 | */ | |
12 | ||
5f92a7b0 | 13 | #define pr_fmt(fmt) "watchdog: " fmt |
4501980a | 14 | |
58687acb DZ |
15 | #include <linux/mm.h> |
16 | #include <linux/cpu.h> | |
17 | #include <linux/nmi.h> | |
18 | #include <linux/init.h> | |
58687acb DZ |
19 | #include <linux/module.h> |
20 | #include <linux/sysctl.h> | |
bcd951cf | 21 | #include <linux/smpboot.h> |
8bd75c77 | 22 | #include <linux/sched/rt.h> |
ae7e81c0 | 23 | #include <uapi/linux/sched/types.h> |
fe4ba3c3 | 24 | #include <linux/tick.h> |
82607adc | 25 | #include <linux/workqueue.h> |
e6017571 | 26 | #include <linux/sched/clock.h> |
b17b0153 | 27 | #include <linux/sched/debug.h> |
58687acb DZ |
28 | |
29 | #include <asm/irq_regs.h> | |
5d1c0f4a | 30 | #include <linux/kvm_para.h> |
81a4beef | 31 | #include <linux/kthread.h> |
58687acb | 32 | |
946d1977 | 33 | static DEFINE_MUTEX(watchdog_mutex); |
ab992dc3 | 34 | |
6ff42c2c HK |
35 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) \ |
36 | || defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU) | |
09154985 TG |
37 | # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED) |
38 | # define NMI_WATCHDOG_DEFAULT 1 | |
84d56e66 | 39 | #else |
09154985 TG |
40 | # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED) |
41 | # define NMI_WATCHDOG_DEFAULT 0 | |
84d56e66 | 42 | #endif |
05a4a952 | 43 | |
09154985 TG |
44 | unsigned long __read_mostly watchdog_enabled; |
45 | int __read_mostly watchdog_user_enabled = 1; | |
46 | int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT; | |
47 | int __read_mostly soft_watchdog_user_enabled = 1; | |
7feeb9cd | 48 | int __read_mostly watchdog_thresh = 10; |
a994a314 | 49 | int __read_mostly nmi_watchdog_available; |
6ff42c2c HK |
50 | #if defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU) |
51 | int __read_mostly watchdog_other_cpu_available = WATCHDOG_DEFAULT; | |
52 | #endif | |
7feeb9cd TG |
53 | |
54 | struct cpumask watchdog_allowed_mask __read_mostly; | |
7feeb9cd TG |
55 | |
56 | struct cpumask watchdog_cpumask __read_mostly; | |
57 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
58 | ||
6ff42c2c | 59 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU) |
05a4a952 NP |
60 | /* |
61 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
62 | */ | |
63 | unsigned int __read_mostly hardlockup_panic = | |
64 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
65 | /* | |
66 | * We may not want to enable hard lockup detection by default in all cases, | |
67 | * for example when running the kernel as a guest on a hypervisor. In these | |
68 | * cases this function can be called to disable hard lockup detection. This | |
69 | * function should only be executed once by the boot processor before the | |
70 | * kernel command line parameters are parsed, because otherwise it is not | |
71 | * possible to override this in hardlockup_panic_setup(). | |
72 | */ | |
7a355820 | 73 | void __init hardlockup_detector_disable(void) |
05a4a952 | 74 | { |
09154985 | 75 | nmi_watchdog_user_enabled = 0; |
05a4a952 NP |
76 | } |
77 | ||
78 | static int __init hardlockup_panic_setup(char *str) | |
79 | { | |
80 | if (!strncmp(str, "panic", 5)) | |
81 | hardlockup_panic = 1; | |
82 | else if (!strncmp(str, "nopanic", 7)) | |
83 | hardlockup_panic = 0; | |
84 | else if (!strncmp(str, "0", 1)) | |
09154985 | 85 | nmi_watchdog_user_enabled = 0; |
05a4a952 | 86 | else if (!strncmp(str, "1", 1)) |
09154985 | 87 | nmi_watchdog_user_enabled = 1; |
05a4a952 NP |
88 | return 1; |
89 | } | |
90 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
91 | ||
368a7e2c TG |
92 | # ifdef CONFIG_SMP |
93 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 94 | |
368a7e2c TG |
95 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
96 | { | |
97 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
98 | return 1; | |
99 | } | |
100 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
101 | # endif /* CONFIG_SMP */ | |
102 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 103 | |
05a4a952 NP |
104 | /* |
105 | * These functions can be overridden if an architecture implements its | |
106 | * own hardlockup detector. | |
a10a842f NP |
107 | * |
108 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
109 | * softlockup watchdog threads start and stop. The arch must select the | |
110 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 | 111 | */ |
6ff42c2c | 112 | |
d718da59 | 113 | #ifndef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU |
05a4a952 NP |
114 | int __weak watchdog_nmi_enable(unsigned int cpu) |
115 | { | |
146c9d0e | 116 | hardlockup_detector_perf_enable(); |
05a4a952 NP |
117 | return 0; |
118 | } | |
941154bd | 119 | |
05a4a952 NP |
120 | void __weak watchdog_nmi_disable(unsigned int cpu) |
121 | { | |
941154bd | 122 | hardlockup_detector_perf_disable(); |
05a4a952 | 123 | } |
6ff42c2c | 124 | #endif |
05a4a952 | 125 | |
a994a314 TG |
126 | /* Return 0, if a NMI watchdog is available. Error code otherwise */ |
127 | int __weak __init watchdog_nmi_probe(void) | |
128 | { | |
129 | return hardlockup_detector_perf_init(); | |
130 | } | |
131 | ||
6592ad2f | 132 | /** |
6b9dc480 | 133 | * watchdog_nmi_stop - Stop the watchdog for reconfiguration |
6592ad2f | 134 | * |
6b9dc480 TG |
135 | * The reconfiguration steps are: |
136 | * watchdog_nmi_stop(); | |
6592ad2f | 137 | * update_variables(); |
6b9dc480 TG |
138 | * watchdog_nmi_start(); |
139 | */ | |
140 | void __weak watchdog_nmi_stop(void) { } | |
141 | ||
142 | /** | |
143 | * watchdog_nmi_start - Start the watchdog after reconfiguration | |
6592ad2f | 144 | * |
6b9dc480 TG |
145 | * Counterpart to watchdog_nmi_stop(). |
146 | * | |
147 | * The following variables have been updated in update_variables() and | |
148 | * contain the currently valid configuration: | |
7feeb9cd | 149 | * - watchdog_enabled |
a10a842f NP |
150 | * - watchdog_thresh |
151 | * - watchdog_cpumask | |
a10a842f | 152 | */ |
6b9dc480 | 153 | void __weak watchdog_nmi_start(void) { } |
a10a842f | 154 | |
09154985 TG |
155 | /** |
156 | * lockup_detector_update_enable - Update the sysctl enable bit | |
157 | * | |
158 | * Caller needs to make sure that the NMI/perf watchdogs are off, so this | |
159 | * can't race with watchdog_nmi_disable(). | |
160 | */ | |
161 | static void lockup_detector_update_enable(void) | |
162 | { | |
163 | watchdog_enabled = 0; | |
164 | if (!watchdog_user_enabled) | |
165 | return; | |
6ff42c2c HK |
166 | #if defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU) |
167 | if (watchdog_other_cpu_available && nmi_watchdog_user_enabled) | |
168 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
169 | #endif | |
a994a314 | 170 | if (nmi_watchdog_available && nmi_watchdog_user_enabled) |
09154985 TG |
171 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; |
172 | if (soft_watchdog_user_enabled) | |
173 | watchdog_enabled |= SOFT_WATCHDOG_ENABLED; | |
174 | } | |
175 | ||
05a4a952 NP |
176 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
177 | ||
2b9d7f23 TG |
178 | /* Global variables, exported for sysctl */ |
179 | unsigned int __read_mostly softlockup_panic = | |
180 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
2eb2527f | 181 | |
0b62bf86 | 182 | static bool softlockup_threads_initialized __read_mostly; |
0f34c400 | 183 | static u64 __read_mostly sample_period; |
6ff42c2c | 184 | static unsigned long __read_mostly hardlockup_thresh; |
58687acb DZ |
185 | |
186 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
6ff42c2c | 187 | static DEFINE_PER_CPU(unsigned long, hardlockup_touch_ts); |
58687acb DZ |
188 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
189 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
190 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 191 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
192 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
193 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 194 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 195 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 196 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 197 | |
58687acb DZ |
198 | static int __init softlockup_panic_setup(char *str) |
199 | { | |
200 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
201 | return 1; |
202 | } | |
203 | __setup("softlockup_panic=", softlockup_panic_setup); | |
204 | ||
205 | static int __init nowatchdog_setup(char *str) | |
206 | { | |
09154985 | 207 | watchdog_user_enabled = 0; |
58687acb DZ |
208 | return 1; |
209 | } | |
210 | __setup("nowatchdog", nowatchdog_setup); | |
211 | ||
58687acb DZ |
212 | static int __init nosoftlockup_setup(char *str) |
213 | { | |
09154985 | 214 | soft_watchdog_user_enabled = 0; |
58687acb DZ |
215 | return 1; |
216 | } | |
217 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 218 | |
ed235875 | 219 | #ifdef CONFIG_SMP |
368a7e2c TG |
220 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
221 | ||
ed235875 AT |
222 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
223 | { | |
368a7e2c | 224 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
225 | return 1; |
226 | } | |
227 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 228 | #endif |
58687acb | 229 | |
941154bd TG |
230 | static void __lockup_detector_cleanup(void); |
231 | ||
4eec42f3 MSB |
232 | /* |
233 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
234 | * lockups can have false positives under extreme conditions. So we generally | |
235 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
236 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
237 | * time the hard threshold is. | |
238 | */ | |
6e9101ae | 239 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
240 | { |
241 | return watchdog_thresh * 2; | |
242 | } | |
58687acb DZ |
243 | |
244 | /* | |
245 | * Returns seconds, approximately. We don't need nanosecond | |
246 | * resolution, and we don't need to waste time with a big divide when | |
247 | * 2^30ns == 1.074s. | |
248 | */ | |
c06b4f19 | 249 | static unsigned long get_timestamp(void) |
58687acb | 250 | { |
545a2bf7 | 251 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
252 | } |
253 | ||
0f34c400 | 254 | static void set_sample_period(void) |
58687acb DZ |
255 | { |
256 | /* | |
586692a5 | 257 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
258 | * the divide by 5 is to give hrtimer several chances (two |
259 | * or three with the current relation between the soft | |
260 | * and hard thresholds) to increment before the | |
261 | * hardlockup detector generates a warning | |
58687acb | 262 | */ |
0f34c400 | 263 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 264 | watchdog_update_hrtimer_threshold(sample_period); |
6ff42c2c | 265 | hardlockup_thresh = sample_period * 3 / NSEC_PER_SEC; |
58687acb DZ |
266 | } |
267 | ||
268 | /* Commands for resetting the watchdog */ | |
269 | static void __touch_watchdog(void) | |
270 | { | |
c06b4f19 | 271 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
6ff42c2c | 272 | __this_cpu_write(hardlockup_touch_ts, get_timestamp()); |
58687acb DZ |
273 | } |
274 | ||
03e0d461 TH |
275 | /** |
276 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
277 | * | |
278 | * Call when the scheduler may have stalled for legitimate reasons | |
279 | * preventing the watchdog task from executing - e.g. the scheduler | |
280 | * entering idle state. This should only be used for scheduler events. | |
281 | * Use touch_softlockup_watchdog() for everything else. | |
282 | */ | |
63a0f9de | 283 | notrace void touch_softlockup_watchdog_sched(void) |
58687acb | 284 | { |
7861144b AM |
285 | /* |
286 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
287 | * gets zeroed here, so use the raw_ operation. | |
288 | */ | |
289 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 290 | } |
03e0d461 | 291 | |
63a0f9de | 292 | notrace void touch_softlockup_watchdog(void) |
03e0d461 TH |
293 | { |
294 | touch_softlockup_watchdog_sched(); | |
82607adc | 295 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 296 | } |
0167c781 | 297 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 298 | |
332fbdbc | 299 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
300 | { |
301 | int cpu; | |
302 | ||
303 | /* | |
d57108d4 TG |
304 | * watchdog_mutex cannpt be taken here, as this might be called |
305 | * from (soft)interrupt context, so the access to | |
306 | * watchdog_allowed_cpumask might race with a concurrent update. | |
307 | * | |
308 | * The watchdog time stamp can race against a concurrent real | |
309 | * update as well, the only side effect might be a cycle delay for | |
310 | * the softlockup check. | |
58687acb | 311 | */ |
d57108d4 | 312 | for_each_cpu(cpu, &watchdog_allowed_mask) |
58687acb | 313 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 314 | wq_watchdog_touch(-1); |
58687acb DZ |
315 | } |
316 | ||
58687acb DZ |
317 | void touch_softlockup_watchdog_sync(void) |
318 | { | |
f7f66b05 CL |
319 | __this_cpu_write(softlockup_touch_sync, true); |
320 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
321 | } |
322 | ||
6ff42c2c HK |
323 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU |
324 | static void watchdog_check_hardlockup_other_cpu(void); | |
325 | #else | |
326 | static inline void watchdog_check_hardlockup_other_cpu(void) { return; } | |
327 | #endif | |
328 | ||
26e09c6e | 329 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 330 | { |
c06b4f19 | 331 | unsigned long now = get_timestamp(); |
58687acb | 332 | |
39d2da21 | 333 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
334 | /* Warn about unreasonable delays. */ |
335 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
336 | return now - touch_ts; | |
337 | } | |
58687acb DZ |
338 | return 0; |
339 | } | |
340 | ||
05a4a952 NP |
341 | /* watchdog detector functions */ |
342 | bool is_hardlockup(void) | |
58687acb | 343 | { |
05a4a952 | 344 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 345 | |
05a4a952 NP |
346 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
347 | return true; | |
348 | ||
349 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
350 | return false; | |
73ce0511 | 351 | } |
05a4a952 NP |
352 | |
353 | static void watchdog_interrupt_count(void) | |
73ce0511 | 354 | { |
05a4a952 | 355 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 356 | } |
58687acb DZ |
357 | |
358 | /* watchdog kicker functions */ | |
359 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
360 | { | |
909ea964 | 361 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
362 | struct pt_regs *regs = get_irq_regs(); |
363 | int duration; | |
ed235875 | 364 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 365 | |
6ff42c2c HK |
366 | /* try to enable log_kevent of exynos-snapshot if log_kevent was off because of rcu stall */ |
367 | dbg_snapshot_try_enable("log_kevent", NSEC_PER_SEC * 15); | |
01f0a027 | 368 | if (!watchdog_enabled) |
b94f5118 DZ |
369 | return HRTIMER_NORESTART; |
370 | ||
58687acb DZ |
371 | /* kick the hardlockup detector */ |
372 | watchdog_interrupt_count(); | |
373 | ||
6ff42c2c HK |
374 | /* test for hardlockups on the next cpu */ |
375 | watchdog_check_hardlockup_other_cpu(); | |
376 | ||
58687acb | 377 | /* kick the softlockup detector */ |
909ea964 | 378 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
379 | |
380 | /* .. and repeat */ | |
0f34c400 | 381 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
382 | |
383 | if (touch_ts == 0) { | |
909ea964 | 384 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
385 | /* |
386 | * If the time stamp was touched atomically | |
387 | * make sure the scheduler tick is up to date. | |
388 | */ | |
909ea964 | 389 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
390 | sched_clock_tick(); |
391 | } | |
5d1c0f4a EM |
392 | |
393 | /* Clear the guest paused flag on watchdog reset */ | |
394 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
395 | __touch_watchdog(); |
396 | return HRTIMER_RESTART; | |
397 | } | |
398 | ||
399 | /* check for a softlockup | |
400 | * This is done by making sure a high priority task is | |
401 | * being scheduled. The task touches the watchdog to | |
402 | * indicate it is getting cpu time. If it hasn't then | |
403 | * this is a good indication some task is hogging the cpu | |
404 | */ | |
26e09c6e | 405 | duration = is_softlockup(touch_ts); |
58687acb | 406 | if (unlikely(duration)) { |
5d1c0f4a EM |
407 | /* |
408 | * If a virtual machine is stopped by the host it can look to | |
409 | * the watchdog like a soft lockup, check to see if the host | |
410 | * stopped the vm before we issue the warning | |
411 | */ | |
412 | if (kvm_check_and_clear_guest_paused()) | |
413 | return HRTIMER_RESTART; | |
414 | ||
58687acb | 415 | /* only warn once */ |
b1a8de1f | 416 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
417 | /* | |
418 | * When multiple processes are causing softlockups the | |
419 | * softlockup detector only warns on the first one | |
420 | * because the code relies on a full quiet cycle to | |
421 | * re-arm. The second process prevents the quiet cycle | |
422 | * and never gets reported. Use task pointers to detect | |
423 | * this. | |
424 | */ | |
425 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
426 | current) { | |
427 | __this_cpu_write(soft_watchdog_warn, false); | |
428 | __touch_watchdog(); | |
429 | } | |
58687acb | 430 | return HRTIMER_RESTART; |
b1a8de1f | 431 | } |
58687acb | 432 | |
ed235875 AT |
433 | if (softlockup_all_cpu_backtrace) { |
434 | /* Prevent multiple soft-lockup reports if one cpu is already | |
435 | * engaged in dumping cpu back traces | |
436 | */ | |
437 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
438 | /* Someone else will report us. Let's give up */ | |
439 | __this_cpu_write(soft_watchdog_warn, true); | |
440 | return HRTIMER_RESTART; | |
441 | } | |
442 | } | |
443 | ||
656c3b79 | 444 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 445 | smp_processor_id(), duration, |
58687acb | 446 | current->comm, task_pid_nr(current)); |
b1a8de1f | 447 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
448 | print_modules(); |
449 | print_irqtrace_events(current); | |
450 | if (regs) | |
451 | show_regs(regs); | |
452 | else | |
453 | dump_stack(); | |
454 | ||
ed235875 AT |
455 | if (softlockup_all_cpu_backtrace) { |
456 | /* Avoid generating two back traces for current | |
457 | * given that one is already made above | |
458 | */ | |
459 | trigger_allbutself_cpu_backtrace(); | |
460 | ||
461 | clear_bit(0, &soft_lockup_nmi_warn); | |
462 | /* Barrier to sync with other cpus */ | |
463 | smp_mb__after_atomic(); | |
464 | } | |
465 | ||
69361eef | 466 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
467 | if (softlockup_panic) |
468 | panic("softlockup: hung tasks"); | |
909ea964 | 469 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 470 | } else |
909ea964 | 471 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
472 | |
473 | return HRTIMER_RESTART; | |
474 | } | |
475 | ||
bcd951cf TG |
476 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
477 | { | |
478 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 479 | |
bcd951cf TG |
480 | sched_setscheduler(current, policy, ¶m); |
481 | } | |
482 | ||
483 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 484 | { |
01f0a027 | 485 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 486 | |
01f0a027 TG |
487 | /* |
488 | * Start the timer first to prevent the NMI watchdog triggering | |
489 | * before the timer has a chance to fire. | |
490 | */ | |
3935e895 BM |
491 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
492 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
493 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
494 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 495 | |
01f0a027 TG |
496 | /* Initialize timestamp */ |
497 | __touch_watchdog(); | |
bcd951cf | 498 | /* Enable the perf event */ |
146c9d0e TG |
499 | if (watchdog_enabled & NMI_WATCHDOG_ENABLED) |
500 | watchdog_nmi_enable(cpu); | |
58687acb | 501 | |
bcd951cf | 502 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 503 | } |
58687acb | 504 | |
bcd951cf TG |
505 | static void watchdog_disable(unsigned int cpu) |
506 | { | |
01f0a027 | 507 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 508 | |
bcd951cf | 509 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
510 | /* |
511 | * Disable the perf event first. That prevents that a large delay | |
512 | * between disabling the timer and disabling the perf event causes | |
513 | * the perf NMI to detect a false positive. | |
514 | */ | |
bcd951cf | 515 | watchdog_nmi_disable(cpu); |
01f0a027 | 516 | hrtimer_cancel(hrtimer); |
58687acb DZ |
517 | } |
518 | ||
b8900bc0 FW |
519 | static void watchdog_cleanup(unsigned int cpu, bool online) |
520 | { | |
521 | watchdog_disable(cpu); | |
522 | } | |
523 | ||
bcd951cf TG |
524 | static int watchdog_should_run(unsigned int cpu) |
525 | { | |
526 | return __this_cpu_read(hrtimer_interrupts) != | |
527 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
528 | } | |
529 | ||
530 | /* | |
531 | * The watchdog thread function - touches the timestamp. | |
532 | * | |
0f34c400 | 533 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
534 | * default) to reset the softlockup timestamp. If this gets delayed |
535 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
536 | * triggers in watchdog_timer_fn(). | |
537 | */ | |
538 | static void watchdog(unsigned int cpu) | |
539 | { | |
540 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
541 | __this_cpu_read(hrtimer_interrupts)); | |
542 | __touch_watchdog(); | |
543 | } | |
58687acb | 544 | |
b8900bc0 FW |
545 | static struct smp_hotplug_thread watchdog_threads = { |
546 | .store = &softlockup_watchdog, | |
547 | .thread_should_run = watchdog_should_run, | |
548 | .thread_fn = watchdog, | |
549 | .thread_comm = "watchdog/%u", | |
550 | .setup = watchdog_enable, | |
551 | .cleanup = watchdog_cleanup, | |
552 | .park = watchdog_disable, | |
553 | .unpark = watchdog_enable, | |
554 | }; | |
555 | ||
2eb2527f TG |
556 | static void softlockup_update_smpboot_threads(void) |
557 | { | |
558 | lockdep_assert_held(&watchdog_mutex); | |
559 | ||
560 | if (!softlockup_threads_initialized) | |
561 | return; | |
562 | ||
563 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
564 | &watchdog_allowed_mask); | |
2eb2527f TG |
565 | } |
566 | ||
567 | /* Temporarily park all watchdog threads */ | |
568 | static void softlockup_park_all_threads(void) | |
569 | { | |
570 | cpumask_clear(&watchdog_allowed_mask); | |
571 | softlockup_update_smpboot_threads(); | |
572 | } | |
573 | ||
e8b62b2d TG |
574 | /* Unpark enabled threads */ |
575 | static void softlockup_unpark_threads(void) | |
2eb2527f TG |
576 | { |
577 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
578 | softlockup_update_smpboot_threads(); | |
579 | } | |
580 | ||
5587185d | 581 | static void lockup_detector_reconfigure(void) |
2eb2527f | 582 | { |
e31d6883 | 583 | cpus_read_lock(); |
6b9dc480 | 584 | watchdog_nmi_stop(); |
2eb2527f TG |
585 | softlockup_park_all_threads(); |
586 | set_sample_period(); | |
09154985 TG |
587 | lockup_detector_update_enable(); |
588 | if (watchdog_enabled && watchdog_thresh) | |
e8b62b2d | 589 | softlockup_unpark_threads(); |
6b9dc480 | 590 | watchdog_nmi_start(); |
e31d6883 TG |
591 | cpus_read_unlock(); |
592 | /* | |
593 | * Must be called outside the cpus locked section to prevent | |
594 | * recursive locking in the perf code. | |
595 | */ | |
596 | __lockup_detector_cleanup(); | |
2eb2527f TG |
597 | } |
598 | ||
599 | /* | |
5587185d | 600 | * Create the watchdog thread infrastructure and configure the detector(s). |
2eb2527f TG |
601 | * |
602 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
603 | * the threads are sucessfully initialized, take the proper locks and | |
604 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
605 | */ | |
5587185d | 606 | static __init void lockup_detector_setup(void) |
2eb2527f TG |
607 | { |
608 | int ret; | |
609 | ||
610 | /* | |
611 | * If sysctl is off and watchdog got disabled on the command line, | |
612 | * nothing to do here. | |
613 | */ | |
09154985 TG |
614 | lockup_detector_update_enable(); |
615 | ||
2eb2527f TG |
616 | if (!IS_ENABLED(CONFIG_SYSCTL) && |
617 | !(watchdog_enabled && watchdog_thresh)) | |
618 | return; | |
619 | ||
620 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
621 | &watchdog_allowed_mask); | |
622 | if (ret) { | |
623 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
624 | return; | |
625 | } | |
626 | ||
627 | mutex_lock(&watchdog_mutex); | |
628 | softlockup_threads_initialized = true; | |
5587185d | 629 | lockup_detector_reconfigure(); |
2eb2527f TG |
630 | mutex_unlock(&watchdog_mutex); |
631 | } | |
632 | ||
2b9d7f23 TG |
633 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
634 | static inline int watchdog_park_threads(void) { return 0; } | |
635 | static inline void watchdog_unpark_threads(void) { } | |
636 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
637 | static inline void watchdog_disable_all_cpus(void) { } | |
5587185d | 638 | static void lockup_detector_reconfigure(void) |
6592ad2f | 639 | { |
e31d6883 | 640 | cpus_read_lock(); |
6b9dc480 | 641 | watchdog_nmi_stop(); |
09154985 | 642 | lockup_detector_update_enable(); |
6b9dc480 | 643 | watchdog_nmi_start(); |
e31d6883 | 644 | cpus_read_unlock(); |
6592ad2f | 645 | } |
5587185d | 646 | static inline void lockup_detector_setup(void) |
34ddaa3e | 647 | { |
5587185d | 648 | lockup_detector_reconfigure(); |
34ddaa3e | 649 | } |
2b9d7f23 | 650 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 651 | |
941154bd TG |
652 | static void __lockup_detector_cleanup(void) |
653 | { | |
654 | lockdep_assert_held(&watchdog_mutex); | |
655 | hardlockup_detector_perf_cleanup(); | |
656 | } | |
657 | ||
658 | /** | |
659 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
660 | * | |
661 | * Caller must not hold the cpu hotplug rwsem. | |
662 | */ | |
663 | void lockup_detector_cleanup(void) | |
664 | { | |
665 | mutex_lock(&watchdog_mutex); | |
666 | __lockup_detector_cleanup(); | |
667 | mutex_unlock(&watchdog_mutex); | |
668 | } | |
669 | ||
6554fd8c TG |
670 | /** |
671 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
672 | * | |
673 | * Special interface for parisc. It prevents lockup detector warnings from | |
674 | * the default pm_poweroff() function which busy loops forever. | |
675 | */ | |
676 | void lockup_detector_soft_poweroff(void) | |
677 | { | |
678 | watchdog_enabled = 0; | |
679 | } | |
680 | ||
58cf690a UO |
681 | #ifdef CONFIG_SYSCTL |
682 | ||
e8b62b2d | 683 | /* Propagate any changes to the watchdog threads */ |
d57108d4 | 684 | static void proc_watchdog_update(void) |
a0c9cbb9 | 685 | { |
e8b62b2d TG |
686 | /* Remove impossible cpus to keep sysctl output clean. */ |
687 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
5587185d | 688 | lockup_detector_reconfigure(); |
a0c9cbb9 UO |
689 | } |
690 | ||
ef246a21 UO |
691 | /* |
692 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
693 | * | |
7feeb9cd TG |
694 | * caller | table->data points to | 'which' |
695 | * -------------------|----------------------------|-------------------------- | |
696 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED | | |
697 | * | | SOFT_WATCHDOG_ENABLED | |
698 | * -------------------|----------------------------|-------------------------- | |
699 | * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED | |
700 | * -------------------|----------------------------|-------------------------- | |
701 | * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED | |
ef246a21 UO |
702 | */ |
703 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
704 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
705 | { | |
09154985 | 706 | int err, old, *param = table->data; |
ef246a21 | 707 | |
946d1977 | 708 | mutex_lock(&watchdog_mutex); |
ef246a21 | 709 | |
ef246a21 | 710 | if (!write) { |
09154985 TG |
711 | /* |
712 | * On read synchronize the userspace interface. This is a | |
713 | * racy snapshot. | |
714 | */ | |
715 | *param = (watchdog_enabled & which) != 0; | |
ef246a21 UO |
716 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
717 | } else { | |
09154985 | 718 | old = READ_ONCE(*param); |
ef246a21 | 719 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
09154985 | 720 | if (!err && old != READ_ONCE(*param)) |
d57108d4 | 721 | proc_watchdog_update(); |
ef246a21 | 722 | } |
946d1977 | 723 | mutex_unlock(&watchdog_mutex); |
ef246a21 UO |
724 | return err; |
725 | } | |
726 | ||
83a80a39 UO |
727 | /* |
728 | * /proc/sys/kernel/watchdog | |
729 | */ | |
730 | int proc_watchdog(struct ctl_table *table, int write, | |
731 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
732 | { | |
733 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
734 | table, write, buffer, lenp, ppos); | |
735 | } | |
736 | ||
737 | /* | |
738 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 739 | */ |
83a80a39 UO |
740 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
741 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
742 | { | |
a994a314 TG |
743 | if (!nmi_watchdog_available && write) |
744 | return -ENOTSUPP; | |
83a80a39 UO |
745 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, |
746 | table, write, buffer, lenp, ppos); | |
747 | } | |
748 | ||
749 | /* | |
750 | * /proc/sys/kernel/soft_watchdog | |
751 | */ | |
752 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
753 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
754 | { | |
755 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
756 | table, write, buffer, lenp, ppos); | |
757 | } | |
58687acb | 758 | |
83a80a39 UO |
759 | /* |
760 | * /proc/sys/kernel/watchdog_thresh | |
761 | */ | |
762 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
763 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 764 | { |
d57108d4 | 765 | int err, old; |
58687acb | 766 | |
946d1977 | 767 | mutex_lock(&watchdog_mutex); |
bcd951cf | 768 | |
d57108d4 | 769 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 770 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 771 | |
d57108d4 TG |
772 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
773 | proc_watchdog_update(); | |
e04ab2bc | 774 | |
946d1977 | 775 | mutex_unlock(&watchdog_mutex); |
b8900bc0 | 776 | return err; |
58687acb | 777 | } |
fe4ba3c3 CM |
778 | |
779 | /* | |
780 | * The cpumask is the mask of possible cpus that the watchdog can run | |
781 | * on, not the mask of cpus it is actually running on. This allows the | |
782 | * user to specify a mask that will include cpus that have not yet | |
783 | * been brought online, if desired. | |
784 | */ | |
785 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
786 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
787 | { | |
788 | int err; | |
789 | ||
946d1977 | 790 | mutex_lock(&watchdog_mutex); |
8c073d27 | 791 | |
fe4ba3c3 | 792 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 793 | if (!err && write) |
e8b62b2d | 794 | proc_watchdog_update(); |
5490125d | 795 | |
946d1977 | 796 | mutex_unlock(&watchdog_mutex); |
fe4ba3c3 CM |
797 | return err; |
798 | } | |
58687acb DZ |
799 | #endif /* CONFIG_SYSCTL */ |
800 | ||
004417a6 | 801 | void __init lockup_detector_init(void) |
58687acb | 802 | { |
fe4ba3c3 CM |
803 | #ifdef CONFIG_NO_HZ_FULL |
804 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
805 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
806 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
807 | } else |
808 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
809 | #else | |
810 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
811 | #endif | |
812 | ||
a994a314 TG |
813 | if (!watchdog_nmi_probe()) |
814 | nmi_watchdog_available = true; | |
5587185d | 815 | lockup_detector_setup(); |
58687acb | 816 | } |
6ff42c2c HK |
817 | |
818 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU | |
819 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); | |
820 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); | |
821 | static cpumask_t __read_mostly watchdog_cpus; | |
822 | ATOMIC_NOTIFIER_HEAD(hardlockup_notifier_list); | |
823 | EXPORT_SYMBOL(hardlockup_notifier_list); | |
824 | ||
825 | static unsigned int watchdog_next_cpu(unsigned int cpu) | |
826 | { | |
827 | cpumask_t cpus = watchdog_cpus; | |
828 | unsigned int next_cpu; | |
829 | ||
830 | next_cpu = cpumask_next(cpu, &cpus); | |
831 | if (next_cpu >= nr_cpu_ids) | |
832 | next_cpu = cpumask_first(&cpus); | |
833 | ||
834 | if (next_cpu == cpu) | |
835 | return nr_cpu_ids; | |
836 | ||
837 | return next_cpu; | |
838 | } | |
839 | ||
840 | static int is_hardlockup_other_cpu(unsigned int cpu) | |
841 | { | |
842 | unsigned long hrint = per_cpu(hrtimer_interrupts, cpu); | |
843 | ||
844 | if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint) { | |
845 | unsigned long now = get_timestamp(); | |
846 | unsigned long touch_ts = per_cpu(hardlockup_touch_ts, cpu); | |
847 | ||
848 | if (time_after(now, touch_ts) && | |
849 | (now - touch_ts >= hardlockup_thresh)) | |
850 | return 1; | |
851 | } | |
852 | ||
853 | per_cpu(hrtimer_interrupts_saved, cpu) = hrint; | |
854 | return 0; | |
855 | } | |
856 | ||
857 | static void watchdog_check_hardlockup_other_cpu(void) | |
858 | { | |
859 | unsigned int next_cpu; | |
860 | ||
861 | /* | |
862 | * Test for hardlockups every 3 samples. The sample period is | |
863 | * watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over | |
864 | * watchdog_thresh (over by 20%). | |
865 | */ | |
866 | if (__this_cpu_read(hrtimer_interrupts) % 3 != 0) | |
867 | return; | |
868 | ||
869 | /* check for a hardlockup on the next cpu */ | |
870 | next_cpu = watchdog_next_cpu(smp_processor_id()); | |
871 | if (next_cpu >= nr_cpu_ids) | |
872 | return; | |
873 | ||
874 | smp_rmb(); | |
875 | ||
876 | if (per_cpu(watchdog_nmi_touch, next_cpu) == true) { | |
877 | per_cpu(watchdog_nmi_touch, next_cpu) = false; | |
878 | return; | |
879 | } | |
880 | ||
881 | if (is_hardlockup_other_cpu(next_cpu)) { | |
882 | /* only warn once */ | |
883 | if (per_cpu(hard_watchdog_warn, next_cpu) == true) | |
884 | return; | |
885 | ||
886 | if (hardlockup_panic) { | |
887 | dbg_snapshot_set_hardlockup(hardlockup_panic); | |
888 | atomic_notifier_call_chain(&hardlockup_notifier_list, 0, (void *)&next_cpu); | |
889 | panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu); | |
890 | } else { | |
891 | WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu); | |
892 | } | |
893 | ||
894 | per_cpu(hard_watchdog_warn, next_cpu) = true; | |
895 | } else { | |
896 | per_cpu(hard_watchdog_warn, next_cpu) = false; | |
897 | } | |
898 | } | |
899 | ||
900 | void touch_nmi_watchdog(void) | |
901 | { | |
902 | /* | |
903 | * Using __raw here because some code paths have | |
904 | * preemption enabled. If preemption is enabled | |
905 | * then interrupts should be enabled too, in which | |
906 | * case we shouldn't have to worry about the watchdog | |
907 | * going off. | |
908 | */ | |
909 | raw_cpu_write(watchdog_nmi_touch, true); | |
910 | arch_touch_nmi_watchdog(); | |
911 | touch_softlockup_watchdog(); | |
912 | } | |
913 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
914 | ||
d718da59 | 915 | int watchdog_nmi_enable(unsigned int cpu) |
6ff42c2c HK |
916 | { |
917 | /* | |
918 | * The new cpu will be marked online before the first hrtimer interrupt | |
919 | * runs on it. If another cpu tests for a hardlockup on the new cpu | |
920 | * before it has run its first hrtimer, it will get a false positive. | |
921 | * Touch the watchdog on the new cpu to delay the first check for at | |
922 | * least 3 sampling periods to guarantee one hrtimer has run on the new | |
923 | * cpu. | |
924 | */ | |
925 | per_cpu(watchdog_nmi_touch, cpu) = true; | |
926 | smp_wmb(); | |
927 | cpumask_set_cpu(cpu, &watchdog_cpus); | |
928 | return 0; | |
929 | } | |
930 | ||
d718da59 | 931 | void watchdog_nmi_disable(unsigned int cpu) |
6ff42c2c HK |
932 | { |
933 | unsigned int next_cpu = watchdog_next_cpu(cpu); | |
934 | ||
935 | /* | |
936 | * Offlining this cpu will cause the cpu before this one to start | |
937 | * checking the one after this one. If this cpu just finished checking | |
938 | * the next cpu and updating hrtimer_interrupts_saved, and then the | |
939 | * previous cpu checks it within one sample period, it will trigger a | |
940 | * false positive. Touch the watchdog on the next cpu to prevent it. | |
941 | */ | |
942 | if (next_cpu < nr_cpu_ids) | |
943 | per_cpu(watchdog_nmi_touch, next_cpu) = true; | |
944 | smp_wmb(); | |
945 | cpumask_clear_cpu(cpu, &watchdog_cpus); | |
946 | } | |
947 | #endif |