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 | 15 | #include <linux/cpu.h> |
58687acb | 16 | #include <linux/init.h> |
e9a9292e BH |
17 | #include <linux/irq.h> |
18 | #include <linux/irqdesc.h> | |
d7037381 | 19 | #include <linux/kernel_stat.h> |
e9a9292e | 20 | #include <linux/kvm_para.h> |
d7037381 | 21 | #include <linux/math64.h> |
e9a9292e | 22 | #include <linux/mm.h> |
58687acb | 23 | #include <linux/module.h> |
e9a9292e BH |
24 | #include <linux/nmi.h> |
25 | #include <linux/stop_machine.h> | |
58687acb | 26 | #include <linux/sysctl.h> |
fe4ba3c3 | 27 | #include <linux/tick.h> |
e9a9292e | 28 | |
e6017571 | 29 | #include <linux/sched/clock.h> |
b17b0153 | 30 | #include <linux/sched/debug.h> |
78634061 | 31 | #include <linux/sched/isolation.h> |
58687acb DZ |
32 | |
33 | #include <asm/irq_regs.h> | |
58687acb | 34 | |
946d1977 | 35 | static DEFINE_MUTEX(watchdog_mutex); |
ab992dc3 | 36 | |
47f4cb43 | 37 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HARDLOCKUP_DETECTOR_SPARC64) |
df95d308 | 38 | # define WATCHDOG_HARDLOCKUP_DEFAULT 1 |
84d56e66 | 39 | #else |
df95d308 | 40 | # define WATCHDOG_HARDLOCKUP_DEFAULT 0 |
84d56e66 | 41 | #endif |
05a4a952 | 42 | |
d7037381 BH |
43 | #define NUM_SAMPLE_PERIODS 5 |
44 | ||
09154985 TG |
45 | unsigned long __read_mostly watchdog_enabled; |
46 | int __read_mostly watchdog_user_enabled = 1; | |
df95d308 DA |
47 | static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT; |
48 | static int __read_mostly watchdog_softlockup_user_enabled = 1; | |
7feeb9cd | 49 | int __read_mostly watchdog_thresh = 10; |
df95d308 | 50 | static int __read_mostly watchdog_hardlockup_available; |
7feeb9cd | 51 | |
7feeb9cd TG |
52 | struct cpumask watchdog_cpumask __read_mostly; |
53 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
54 | ||
05a4a952 | 55 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
f117955a GP |
56 | |
57 | # ifdef CONFIG_SMP | |
58 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
59 | # endif /* CONFIG_SMP */ | |
60 | ||
05a4a952 NP |
61 | /* |
62 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
63 | */ | |
64 | unsigned int __read_mostly hardlockup_panic = | |
67fca000 | 65 | IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC); |
05a4a952 NP |
66 | /* |
67 | * We may not want to enable hard lockup detection by default in all cases, | |
68 | * for example when running the kernel as a guest on a hypervisor. In these | |
69 | * cases this function can be called to disable hard lockup detection. This | |
70 | * function should only be executed once by the boot processor before the | |
71 | * kernel command line parameters are parsed, because otherwise it is not | |
72 | * possible to override this in hardlockup_panic_setup(). | |
73 | */ | |
7a355820 | 74 | void __init hardlockup_detector_disable(void) |
05a4a952 | 75 | { |
df95d308 | 76 | watchdog_hardlockup_user_enabled = 0; |
05a4a952 NP |
77 | } |
78 | ||
79 | static int __init hardlockup_panic_setup(char *str) | |
80 | { | |
602ba773 | 81 | next: |
05a4a952 NP |
82 | if (!strncmp(str, "panic", 5)) |
83 | hardlockup_panic = 1; | |
84 | else if (!strncmp(str, "nopanic", 7)) | |
85 | hardlockup_panic = 0; | |
86 | else if (!strncmp(str, "0", 1)) | |
df95d308 | 87 | watchdog_hardlockup_user_enabled = 0; |
05a4a952 | 88 | else if (!strncmp(str, "1", 1)) |
df95d308 | 89 | watchdog_hardlockup_user_enabled = 1; |
393fb313 SL |
90 | else if (!strncmp(str, "r", 1)) |
91 | hardlockup_config_perf_event(str + 1); | |
602ba773 SL |
92 | while (*(str++)) { |
93 | if (*str == ',') { | |
94 | str++; | |
95 | goto next; | |
96 | } | |
97 | } | |
05a4a952 NP |
98 | return 1; |
99 | } | |
100 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
101 | ||
368a7e2c | 102 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
05a4a952 | 103 | |
1f423c90 | 104 | #if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER) |
81972551 | 105 | |
77c12fc9 DA |
106 | static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts); |
107 | static DEFINE_PER_CPU(int, hrtimer_interrupts_saved); | |
1610611a | 108 | static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned); |
ed92e1ef | 109 | static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched); |
6dcde5d5 | 110 | static unsigned long hard_lockup_nmi_warn; |
81972551 | 111 | |
ed92e1ef DA |
112 | notrace void arch_touch_nmi_watchdog(void) |
113 | { | |
114 | /* | |
115 | * Using __raw here because some code paths have | |
116 | * preemption enabled. If preemption is enabled | |
117 | * then interrupts should be enabled too, in which | |
118 | * case we shouldn't have to worry about the watchdog | |
119 | * going off. | |
120 | */ | |
121 | raw_cpu_write(watchdog_hardlockup_touched, true); | |
122 | } | |
123 | EXPORT_SYMBOL(arch_touch_nmi_watchdog); | |
124 | ||
1f423c90 DA |
125 | void watchdog_hardlockup_touch_cpu(unsigned int cpu) |
126 | { | |
127 | per_cpu(watchdog_hardlockup_touched, cpu) = true; | |
1f423c90 DA |
128 | } |
129 | ||
77c12fc9 | 130 | static bool is_hardlockup(unsigned int cpu) |
81972551 | 131 | { |
77c12fc9 | 132 | int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu)); |
81972551 | 133 | |
77c12fc9 | 134 | if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint) |
81972551 DA |
135 | return true; |
136 | ||
77c12fc9 DA |
137 | /* |
138 | * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE | |
139 | * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is | |
140 | * written/read by a single CPU. | |
141 | */ | |
142 | per_cpu(hrtimer_interrupts_saved, cpu) = hrint; | |
1610611a | 143 | |
81972551 DA |
144 | return false; |
145 | } | |
146 | ||
d3b62ace | 147 | static void watchdog_hardlockup_kick(void) |
81972551 | 148 | { |
d3b62ace DA |
149 | int new_interrupts; |
150 | ||
151 | new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts)); | |
152 | watchdog_buddy_check_hardlockup(new_interrupts); | |
81972551 DA |
153 | } |
154 | ||
77c12fc9 | 155 | void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs) |
81972551 | 156 | { |
ed92e1ef DA |
157 | if (per_cpu(watchdog_hardlockup_touched, cpu)) { |
158 | per_cpu(watchdog_hardlockup_touched, cpu) = false; | |
159 | return; | |
160 | } | |
161 | ||
1610611a DA |
162 | /* |
163 | * Check for a hardlockup by making sure the CPU's timer | |
164 | * interrupt is incrementing. The timer interrupt should have | |
165 | * fired multiple times before we overflow'd. If it hasn't | |
81972551 DA |
166 | * then this is a good indication the cpu is stuck |
167 | */ | |
77c12fc9 | 168 | if (is_hardlockup(cpu)) { |
1610611a | 169 | unsigned int this_cpu = smp_processor_id(); |
ee6bdb3f | 170 | unsigned long flags; |
81972551 | 171 | |
1610611a | 172 | /* Only print hardlockups once. */ |
77c12fc9 | 173 | if (per_cpu(watchdog_hardlockup_warned, cpu)) |
81972551 DA |
174 | return; |
175 | ||
6dcde5d5 DA |
176 | /* |
177 | * Prevent multiple hard-lockup reports if one cpu is already | |
178 | * engaged in dumping all cpu back traces. | |
179 | */ | |
180 | if (sysctl_hardlockup_all_cpu_backtrace) { | |
181 | if (test_and_set_bit_lock(0, &hard_lockup_nmi_warn)) | |
182 | return; | |
183 | } | |
184 | ||
ee6bdb3f DA |
185 | /* |
186 | * NOTE: we call printk_cpu_sync_get_irqsave() after printing | |
187 | * the lockup message. While it would be nice to serialize | |
188 | * that printout, we really want to make sure that if some | |
189 | * other CPU somehow locked up while holding the lock associated | |
190 | * with printk_cpu_sync_get_irqsave() that we can still at least | |
191 | * get the message about the lockup out. | |
192 | */ | |
77c12fc9 | 193 | pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu); |
ee6bdb3f DA |
194 | printk_cpu_sync_get_irqsave(flags); |
195 | ||
81972551 DA |
196 | print_modules(); |
197 | print_irqtrace_events(current); | |
77c12fc9 DA |
198 | if (cpu == this_cpu) { |
199 | if (regs) | |
200 | show_regs(regs); | |
201 | else | |
202 | dump_stack(); | |
ee6bdb3f | 203 | printk_cpu_sync_put_irqrestore(flags); |
77c12fc9 | 204 | } else { |
ee6bdb3f | 205 | printk_cpu_sync_put_irqrestore(flags); |
1f38c86b | 206 | trigger_single_cpu_backtrace(cpu); |
77c12fc9 | 207 | } |
81972551 | 208 | |
6dcde5d5 | 209 | if (sysctl_hardlockup_all_cpu_backtrace) { |
1f38c86b | 210 | trigger_allbutcpu_cpu_backtrace(cpu); |
55efe4ab DA |
211 | if (!hardlockup_panic) |
212 | clear_bit_unlock(0, &hard_lockup_nmi_warn); | |
6dcde5d5 | 213 | } |
81972551 DA |
214 | |
215 | if (hardlockup_panic) | |
216 | nmi_panic(regs, "Hard LOCKUP"); | |
217 | ||
77c12fc9 | 218 | per_cpu(watchdog_hardlockup_warned, cpu) = true; |
1610611a | 219 | } else { |
77c12fc9 | 220 | per_cpu(watchdog_hardlockup_warned, cpu) = false; |
81972551 | 221 | } |
81972551 DA |
222 | } |
223 | ||
1f423c90 | 224 | #else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */ |
81972551 | 225 | |
d3b62ace | 226 | static inline void watchdog_hardlockup_kick(void) { } |
81972551 | 227 | |
1f423c90 | 228 | #endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */ |
81972551 | 229 | |
05a4a952 | 230 | /* |
d9b3629a | 231 | * These functions can be overridden based on the configured hardlockdup detector. |
a10a842f | 232 | * |
df95d308 | 233 | * watchdog_hardlockup_enable/disable can be implemented to start and stop when |
d9b3629a | 234 | * softlockup watchdog start and stop. The detector must select the |
a10a842f | 235 | * SOFTLOCKUP_DETECTOR Kconfig. |
05a4a952 | 236 | */ |
d9b3629a | 237 | void __weak watchdog_hardlockup_enable(unsigned int cpu) { } |
941154bd | 238 | |
d9b3629a | 239 | void __weak watchdog_hardlockup_disable(unsigned int cpu) { } |
05a4a952 | 240 | |
930d8f8d LC |
241 | /* |
242 | * Watchdog-detector specific API. | |
243 | * | |
244 | * Return 0 when hardlockup watchdog is available, negative value otherwise. | |
245 | * Note that the negative value means that a delayed probe might | |
246 | * succeed later. | |
247 | */ | |
df95d308 | 248 | int __weak __init watchdog_hardlockup_probe(void) |
a994a314 | 249 | { |
d9b3629a | 250 | return -ENODEV; |
a994a314 TG |
251 | } |
252 | ||
6592ad2f | 253 | /** |
df95d308 | 254 | * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration |
6592ad2f | 255 | * |
6b9dc480 | 256 | * The reconfiguration steps are: |
df95d308 | 257 | * watchdog_hardlockup_stop(); |
6592ad2f | 258 | * update_variables(); |
df95d308 | 259 | * watchdog_hardlockup_start(); |
6b9dc480 | 260 | */ |
df95d308 | 261 | void __weak watchdog_hardlockup_stop(void) { } |
6b9dc480 TG |
262 | |
263 | /** | |
df95d308 | 264 | * watchdog_hardlockup_start - Start the watchdog after reconfiguration |
6592ad2f | 265 | * |
df95d308 | 266 | * Counterpart to watchdog_hardlockup_stop(). |
6b9dc480 TG |
267 | * |
268 | * The following variables have been updated in update_variables() and | |
269 | * contain the currently valid configuration: | |
7feeb9cd | 270 | * - watchdog_enabled |
a10a842f NP |
271 | * - watchdog_thresh |
272 | * - watchdog_cpumask | |
a10a842f | 273 | */ |
df95d308 | 274 | void __weak watchdog_hardlockup_start(void) { } |
a10a842f | 275 | |
09154985 TG |
276 | /** |
277 | * lockup_detector_update_enable - Update the sysctl enable bit | |
278 | * | |
df95d308 DA |
279 | * Caller needs to make sure that the hard watchdogs are off, so this |
280 | * can't race with watchdog_hardlockup_disable(). | |
09154985 TG |
281 | */ |
282 | static void lockup_detector_update_enable(void) | |
283 | { | |
284 | watchdog_enabled = 0; | |
285 | if (!watchdog_user_enabled) | |
286 | return; | |
df95d308 DA |
287 | if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled) |
288 | watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED; | |
289 | if (watchdog_softlockup_user_enabled) | |
290 | watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED; | |
09154985 TG |
291 | } |
292 | ||
05a4a952 NP |
293 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
294 | ||
fef06efc PM |
295 | /* |
296 | * Delay the soflockup report when running a known slow code. | |
297 | * It does _not_ affect the timestamp of the last successdul reschedule. | |
298 | */ | |
299 | #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX | |
11e31f60 | 300 | |
f117955a GP |
301 | #ifdef CONFIG_SMP |
302 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
303 | #endif | |
304 | ||
e7e04615 SS |
305 | static struct cpumask watchdog_allowed_mask __read_mostly; |
306 | ||
2b9d7f23 TG |
307 | /* Global variables, exported for sysctl */ |
308 | unsigned int __read_mostly softlockup_panic = | |
67fca000 | 309 | IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC); |
2eb2527f | 310 | |
9cf57731 | 311 | static bool softlockup_initialized __read_mostly; |
0f34c400 | 312 | static u64 __read_mostly sample_period; |
58687acb | 313 | |
fef06efc | 314 | /* Timestamp taken after the last successful reschedule. */ |
58687acb | 315 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); |
fef06efc PM |
316 | /* Timestamp of the last softlockup report. */ |
317 | static DEFINE_PER_CPU(unsigned long, watchdog_report_ts); | |
58687acb DZ |
318 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); |
319 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
ed235875 | 320 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 321 | |
8b793bcd KJ |
322 | static int __init softlockup_panic_setup(char *str) |
323 | { | |
324 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
325 | return 1; | |
326 | } | |
327 | __setup("softlockup_panic=", softlockup_panic_setup); | |
328 | ||
58687acb DZ |
329 | static int __init nowatchdog_setup(char *str) |
330 | { | |
09154985 | 331 | watchdog_user_enabled = 0; |
58687acb DZ |
332 | return 1; |
333 | } | |
334 | __setup("nowatchdog", nowatchdog_setup); | |
335 | ||
58687acb DZ |
336 | static int __init nosoftlockup_setup(char *str) |
337 | { | |
df95d308 | 338 | watchdog_softlockup_user_enabled = 0; |
58687acb DZ |
339 | return 1; |
340 | } | |
341 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 342 | |
11295055 LO |
343 | static int __init watchdog_thresh_setup(char *str) |
344 | { | |
345 | get_option(&str, &watchdog_thresh); | |
346 | return 1; | |
347 | } | |
348 | __setup("watchdog_thresh=", watchdog_thresh_setup); | |
349 | ||
941154bd TG |
350 | static void __lockup_detector_cleanup(void); |
351 | ||
d7037381 BH |
352 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR_INTR_STORM |
353 | enum stats_per_group { | |
354 | STATS_SYSTEM, | |
355 | STATS_SOFTIRQ, | |
356 | STATS_HARDIRQ, | |
357 | STATS_IDLE, | |
358 | NUM_STATS_PER_GROUP, | |
359 | }; | |
360 | ||
361 | static const enum cpu_usage_stat tracked_stats[NUM_STATS_PER_GROUP] = { | |
362 | CPUTIME_SYSTEM, | |
363 | CPUTIME_SOFTIRQ, | |
364 | CPUTIME_IRQ, | |
365 | CPUTIME_IDLE, | |
366 | }; | |
367 | ||
368 | static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]); | |
369 | static DEFINE_PER_CPU(u8, cpustat_util[NUM_SAMPLE_PERIODS][NUM_STATS_PER_GROUP]); | |
370 | static DEFINE_PER_CPU(u8, cpustat_tail); | |
371 | ||
372 | /* | |
373 | * We don't need nanosecond resolution. A granularity of 16ms is | |
374 | * sufficient for our precision, allowing us to use u16 to store | |
375 | * cpustats, which will roll over roughly every ~1000 seconds. | |
376 | * 2^24 ~= 16 * 10^6 | |
377 | */ | |
378 | static u16 get_16bit_precision(u64 data_ns) | |
379 | { | |
380 | return data_ns >> 24LL; /* 2^24ns ~= 16.8ms */ | |
381 | } | |
382 | ||
383 | static void update_cpustat(void) | |
384 | { | |
385 | int i; | |
386 | u8 util; | |
387 | u16 old_stat, new_stat; | |
388 | struct kernel_cpustat kcpustat; | |
389 | u64 *cpustat = kcpustat.cpustat; | |
390 | u8 tail = __this_cpu_read(cpustat_tail); | |
391 | u16 sample_period_16 = get_16bit_precision(sample_period); | |
392 | ||
393 | kcpustat_cpu_fetch(&kcpustat, smp_processor_id()); | |
394 | ||
395 | for (i = 0; i < NUM_STATS_PER_GROUP; i++) { | |
396 | old_stat = __this_cpu_read(cpustat_old[i]); | |
397 | new_stat = get_16bit_precision(cpustat[tracked_stats[i]]); | |
398 | util = DIV_ROUND_UP(100 * (new_stat - old_stat), sample_period_16); | |
399 | __this_cpu_write(cpustat_util[tail][i], util); | |
400 | __this_cpu_write(cpustat_old[i], new_stat); | |
401 | } | |
402 | ||
403 | __this_cpu_write(cpustat_tail, (tail + 1) % NUM_SAMPLE_PERIODS); | |
404 | } | |
405 | ||
406 | static void print_cpustat(void) | |
407 | { | |
408 | int i, group; | |
409 | u8 tail = __this_cpu_read(cpustat_tail); | |
410 | u64 sample_period_second = sample_period; | |
411 | ||
412 | do_div(sample_period_second, NSEC_PER_SEC); | |
413 | ||
414 | /* | |
415 | * Outputting the "watchdog" prefix on every line is redundant and not | |
416 | * concise, and the original alarm information is sufficient for | |
417 | * positioning in logs, hence here printk() is used instead of pr_crit(). | |
418 | */ | |
419 | printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n", | |
420 | smp_processor_id(), sample_period_second); | |
421 | ||
422 | for (i = 0; i < NUM_SAMPLE_PERIODS; i++) { | |
423 | group = (tail + i) % NUM_SAMPLE_PERIODS; | |
424 | printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t" | |
425 | "%3u%% hardirq,\t%3u%% idle\n", i + 1, | |
426 | __this_cpu_read(cpustat_util[group][STATS_SYSTEM]), | |
427 | __this_cpu_read(cpustat_util[group][STATS_SOFTIRQ]), | |
428 | __this_cpu_read(cpustat_util[group][STATS_HARDIRQ]), | |
429 | __this_cpu_read(cpustat_util[group][STATS_IDLE])); | |
430 | } | |
431 | } | |
432 | ||
e9a9292e BH |
433 | #define HARDIRQ_PERCENT_THRESH 50 |
434 | #define NUM_HARDIRQ_REPORT 5 | |
435 | struct irq_counts { | |
436 | int irq; | |
437 | u32 counts; | |
438 | }; | |
439 | ||
440 | static DEFINE_PER_CPU(bool, snapshot_taken); | |
441 | ||
442 | /* Tabulate the most frequent interrupts. */ | |
443 | static void tabulate_irq_count(struct irq_counts *irq_counts, int irq, u32 counts, int rank) | |
444 | { | |
445 | int i; | |
446 | struct irq_counts new_count = {irq, counts}; | |
447 | ||
448 | for (i = 0; i < rank; i++) { | |
449 | if (counts > irq_counts[i].counts) | |
450 | swap(new_count, irq_counts[i]); | |
451 | } | |
452 | } | |
453 | ||
454 | /* | |
455 | * If the hardirq time exceeds HARDIRQ_PERCENT_THRESH% of the sample_period, | |
456 | * then the cause of softlockup might be interrupt storm. In this case, it | |
457 | * would be useful to start interrupt counting. | |
458 | */ | |
459 | static bool need_counting_irqs(void) | |
460 | { | |
461 | u8 util; | |
462 | int tail = __this_cpu_read(cpustat_tail); | |
463 | ||
464 | tail = (tail + NUM_HARDIRQ_REPORT - 1) % NUM_HARDIRQ_REPORT; | |
465 | util = __this_cpu_read(cpustat_util[tail][STATS_HARDIRQ]); | |
466 | return util > HARDIRQ_PERCENT_THRESH; | |
467 | } | |
468 | ||
469 | static void start_counting_irqs(void) | |
470 | { | |
471 | if (!__this_cpu_read(snapshot_taken)) { | |
472 | kstat_snapshot_irqs(); | |
473 | __this_cpu_write(snapshot_taken, true); | |
474 | } | |
475 | } | |
476 | ||
477 | static void stop_counting_irqs(void) | |
478 | { | |
479 | __this_cpu_write(snapshot_taken, false); | |
480 | } | |
481 | ||
482 | static void print_irq_counts(void) | |
483 | { | |
484 | unsigned int i, count; | |
485 | struct irq_counts irq_counts_sorted[NUM_HARDIRQ_REPORT] = { | |
486 | {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0} | |
487 | }; | |
488 | ||
489 | if (__this_cpu_read(snapshot_taken)) { | |
490 | for_each_active_irq(i) { | |
491 | count = kstat_get_irq_since_snapshot(i); | |
492 | tabulate_irq_count(irq_counts_sorted, i, count, NUM_HARDIRQ_REPORT); | |
493 | } | |
494 | ||
495 | /* | |
496 | * Outputting the "watchdog" prefix on every line is redundant and not | |
497 | * concise, and the original alarm information is sufficient for | |
498 | * positioning in logs, hence here printk() is used instead of pr_crit(). | |
499 | */ | |
500 | printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n", | |
501 | smp_processor_id(), HARDIRQ_PERCENT_THRESH); | |
502 | ||
503 | for (i = 0; i < NUM_HARDIRQ_REPORT; i++) { | |
504 | if (irq_counts_sorted[i].irq == -1) | |
505 | break; | |
506 | ||
507 | printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n", | |
508 | i + 1, irq_counts_sorted[i].counts, | |
509 | irq_counts_sorted[i].irq); | |
510 | } | |
511 | ||
512 | /* | |
513 | * If the hardirq time is less than HARDIRQ_PERCENT_THRESH% in the last | |
514 | * sample_period, then we suspect the interrupt storm might be subsiding. | |
515 | */ | |
516 | if (!need_counting_irqs()) | |
517 | stop_counting_irqs(); | |
518 | } | |
519 | } | |
520 | ||
d7037381 BH |
521 | static void report_cpu_status(void) |
522 | { | |
523 | print_cpustat(); | |
e9a9292e | 524 | print_irq_counts(); |
d7037381 BH |
525 | } |
526 | #else | |
527 | static inline void update_cpustat(void) { } | |
528 | static inline void report_cpu_status(void) { } | |
e9a9292e BH |
529 | static inline bool need_counting_irqs(void) { return false; } |
530 | static inline void start_counting_irqs(void) { } | |
531 | static inline void stop_counting_irqs(void) { } | |
d7037381 BH |
532 | #endif |
533 | ||
4eec42f3 MSB |
534 | /* |
535 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
536 | * lockups can have false positives under extreme conditions. So we generally | |
537 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
538 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
539 | * time the hard threshold is. | |
540 | */ | |
6e9101ae | 541 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
542 | { |
543 | return watchdog_thresh * 2; | |
544 | } | |
58687acb DZ |
545 | |
546 | /* | |
547 | * Returns seconds, approximately. We don't need nanosecond | |
548 | * resolution, and we don't need to waste time with a big divide when | |
549 | * 2^30ns == 1.074s. | |
550 | */ | |
c06b4f19 | 551 | static unsigned long get_timestamp(void) |
58687acb | 552 | { |
545a2bf7 | 553 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
554 | } |
555 | ||
0f34c400 | 556 | static void set_sample_period(void) |
58687acb DZ |
557 | { |
558 | /* | |
586692a5 | 559 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
560 | * the divide by 5 is to give hrtimer several chances (two |
561 | * or three with the current relation between the soft | |
562 | * and hard thresholds) to increment before the | |
563 | * hardlockup detector generates a warning | |
58687acb | 564 | */ |
d7037381 | 565 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / NUM_SAMPLE_PERIODS); |
7edaeb68 | 566 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
567 | } |
568 | ||
fef06efc PM |
569 | static void update_report_ts(void) |
570 | { | |
571 | __this_cpu_write(watchdog_report_ts, get_timestamp()); | |
572 | } | |
573 | ||
58687acb | 574 | /* Commands for resetting the watchdog */ |
7c0012f5 | 575 | static void update_touch_ts(void) |
58687acb | 576 | { |
c06b4f19 | 577 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
fef06efc | 578 | update_report_ts(); |
58687acb DZ |
579 | } |
580 | ||
03e0d461 TH |
581 | /** |
582 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
583 | * | |
584 | * Call when the scheduler may have stalled for legitimate reasons | |
585 | * preventing the watchdog task from executing - e.g. the scheduler | |
586 | * entering idle state. This should only be used for scheduler events. | |
587 | * Use touch_softlockup_watchdog() for everything else. | |
588 | */ | |
cb9d7fd5 | 589 | notrace void touch_softlockup_watchdog_sched(void) |
58687acb | 590 | { |
7861144b | 591 | /* |
fef06efc PM |
592 | * Preemption can be enabled. It doesn't matter which CPU's watchdog |
593 | * report period gets restarted here, so use the raw_ operation. | |
7861144b | 594 | */ |
fef06efc | 595 | raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); |
58687acb | 596 | } |
03e0d461 | 597 | |
cb9d7fd5 | 598 | notrace void touch_softlockup_watchdog(void) |
03e0d461 TH |
599 | { |
600 | touch_softlockup_watchdog_sched(); | |
82607adc | 601 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 602 | } |
0167c781 | 603 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 604 | |
332fbdbc | 605 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
606 | { |
607 | int cpu; | |
608 | ||
609 | /* | |
d57108d4 TG |
610 | * watchdog_mutex cannpt be taken here, as this might be called |
611 | * from (soft)interrupt context, so the access to | |
612 | * watchdog_allowed_cpumask might race with a concurrent update. | |
613 | * | |
614 | * The watchdog time stamp can race against a concurrent real | |
615 | * update as well, the only side effect might be a cycle delay for | |
616 | * the softlockup check. | |
58687acb | 617 | */ |
89e28ce6 | 618 | for_each_cpu(cpu, &watchdog_allowed_mask) { |
fef06efc | 619 | per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT; |
89e28ce6 WQ |
620 | wq_watchdog_touch(cpu); |
621 | } | |
58687acb DZ |
622 | } |
623 | ||
58687acb DZ |
624 | void touch_softlockup_watchdog_sync(void) |
625 | { | |
f7f66b05 | 626 | __this_cpu_write(softlockup_touch_sync, true); |
fef06efc | 627 | __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT); |
58687acb DZ |
628 | } |
629 | ||
0f90b88d PM |
630 | static int is_softlockup(unsigned long touch_ts, |
631 | unsigned long period_ts, | |
632 | unsigned long now) | |
58687acb | 633 | { |
df95d308 | 634 | if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) { |
e9a9292e BH |
635 | /* |
636 | * If period_ts has not been updated during a sample_period, then | |
637 | * in the subsequent few sample_periods, period_ts might also not | |
638 | * be updated, which could indicate a potential softlockup. In | |
639 | * this case, if we suspect the cause of the potential softlockup | |
640 | * might be interrupt storm, then we need to count the interrupts | |
641 | * to find which interrupt is storming. | |
642 | */ | |
643 | if (time_after_eq(now, period_ts + get_softlockup_thresh() / NUM_SAMPLE_PERIODS) && | |
644 | need_counting_irqs()) | |
645 | start_counting_irqs(); | |
646 | ||
195daf66 | 647 | /* Warn about unreasonable delays. */ |
fef06efc | 648 | if (time_after(now, period_ts + get_softlockup_thresh())) |
195daf66 UO |
649 | return now - touch_ts; |
650 | } | |
58687acb DZ |
651 | return 0; |
652 | } | |
653 | ||
05a4a952 | 654 | /* watchdog detector functions */ |
be45bf53 PZ |
655 | static DEFINE_PER_CPU(struct completion, softlockup_completion); |
656 | static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work); | |
657 | ||
9cf57731 | 658 | /* |
b124ac45 | 659 | * The watchdog feed function - touches the timestamp. |
9cf57731 PZ |
660 | * |
661 | * It only runs once every sample_period seconds (4 seconds by | |
662 | * default) to reset the softlockup timestamp. If this gets delayed | |
663 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
664 | * triggers in watchdog_timer_fn(). | |
665 | */ | |
666 | static int softlockup_fn(void *data) | |
667 | { | |
7c0012f5 | 668 | update_touch_ts(); |
e9a9292e | 669 | stop_counting_irqs(); |
be45bf53 | 670 | complete(this_cpu_ptr(&softlockup_completion)); |
9cf57731 PZ |
671 | |
672 | return 0; | |
673 | } | |
674 | ||
58687acb DZ |
675 | /* watchdog kicker functions */ |
676 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
677 | { | |
0f90b88d | 678 | unsigned long touch_ts, period_ts, now; |
58687acb DZ |
679 | struct pt_regs *regs = get_irq_regs(); |
680 | int duration; | |
ed235875 | 681 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
896260a6 | 682 | unsigned long flags; |
58687acb | 683 | |
01f0a027 | 684 | if (!watchdog_enabled) |
b94f5118 DZ |
685 | return HRTIMER_NORESTART; |
686 | ||
d3b62ace | 687 | watchdog_hardlockup_kick(); |
58687acb DZ |
688 | |
689 | /* kick the softlockup detector */ | |
be45bf53 PZ |
690 | if (completion_done(this_cpu_ptr(&softlockup_completion))) { |
691 | reinit_completion(this_cpu_ptr(&softlockup_completion)); | |
692 | stop_one_cpu_nowait(smp_processor_id(), | |
693 | softlockup_fn, NULL, | |
694 | this_cpu_ptr(&softlockup_stop_work)); | |
695 | } | |
58687acb DZ |
696 | |
697 | /* .. and repeat */ | |
0f34c400 | 698 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb | 699 | |
0f90b88d PM |
700 | /* |
701 | * Read the current timestamp first. It might become invalid anytime | |
702 | * when a virtual machine is stopped by the host or when the watchog | |
703 | * is touched from NMI. | |
704 | */ | |
705 | now = get_timestamp(); | |
9bf3bc94 PM |
706 | /* |
707 | * If a virtual machine is stopped by the host it can look to | |
0f90b88d | 708 | * the watchdog like a soft lockup. This function touches the watchdog. |
9bf3bc94 PM |
709 | */ |
710 | kvm_check_and_clear_guest_paused(); | |
0f90b88d PM |
711 | /* |
712 | * The stored timestamp is comparable with @now only when not touched. | |
713 | * It might get touched anytime from NMI. Make sure that is_softlockup() | |
714 | * uses the same (valid) value. | |
715 | */ | |
716 | period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts)); | |
9bf3bc94 | 717 | |
d7037381 BH |
718 | update_cpustat(); |
719 | ||
9bf3bc94 | 720 | /* Reset the interval when touched by known problematic code. */ |
fef06efc | 721 | if (period_ts == SOFTLOCKUP_DELAY_REPORT) { |
909ea964 | 722 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
723 | /* |
724 | * If the time stamp was touched atomically | |
725 | * make sure the scheduler tick is up to date. | |
726 | */ | |
909ea964 | 727 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
728 | sched_clock_tick(); |
729 | } | |
5d1c0f4a | 730 | |
fef06efc | 731 | update_report_ts(); |
58687acb DZ |
732 | return HRTIMER_RESTART; |
733 | } | |
734 | ||
0f90b88d PM |
735 | /* Check for a softlockup. */ |
736 | touch_ts = __this_cpu_read(watchdog_touch_ts); | |
737 | duration = is_softlockup(touch_ts, period_ts, now); | |
58687acb | 738 | if (unlikely(duration)) { |
9f113bf7 PM |
739 | /* |
740 | * Prevent multiple soft-lockup reports if one cpu is already | |
741 | * engaged in dumping all cpu back traces. | |
742 | */ | |
ed235875 | 743 | if (softlockup_all_cpu_backtrace) { |
9f113bf7 | 744 | if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn)) |
ed235875 | 745 | return HRTIMER_RESTART; |
ed235875 AT |
746 | } |
747 | ||
c9ad17c9 | 748 | /* Start period for the next softlockup warning. */ |
fef06efc | 749 | update_report_ts(); |
c9ad17c9 | 750 | |
896260a6 | 751 | printk_cpu_sync_get_irqsave(flags); |
656c3b79 | 752 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 753 | smp_processor_id(), duration, |
58687acb | 754 | current->comm, task_pid_nr(current)); |
d7037381 | 755 | report_cpu_status(); |
58687acb DZ |
756 | print_modules(); |
757 | print_irqtrace_events(current); | |
758 | if (regs) | |
759 | show_regs(regs); | |
760 | else | |
761 | dump_stack(); | |
896260a6 | 762 | printk_cpu_sync_put_irqrestore(flags); |
58687acb | 763 | |
ed235875 | 764 | if (softlockup_all_cpu_backtrace) { |
8d539b84 | 765 | trigger_allbutcpu_cpu_backtrace(smp_processor_id()); |
55efe4ab DA |
766 | if (!softlockup_panic) |
767 | clear_bit_unlock(0, &soft_lockup_nmi_warn); | |
ed235875 AT |
768 | } |
769 | ||
69361eef | 770 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
771 | if (softlockup_panic) |
772 | panic("softlockup: hung tasks"); | |
1bc503cb | 773 | } |
58687acb DZ |
774 | |
775 | return HRTIMER_RESTART; | |
776 | } | |
777 | ||
bcd951cf | 778 | static void watchdog_enable(unsigned int cpu) |
58687acb | 779 | { |
01f0a027 | 780 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
be45bf53 | 781 | struct completion *done = this_cpu_ptr(&softlockup_completion); |
58687acb | 782 | |
9cf57731 PZ |
783 | WARN_ON_ONCE(cpu != smp_processor_id()); |
784 | ||
be45bf53 PZ |
785 | init_completion(done); |
786 | complete(done); | |
787 | ||
01f0a027 | 788 | /* |
df95d308 | 789 | * Start the timer first to prevent the hardlockup watchdog triggering |
01f0a027 TG |
790 | * before the timer has a chance to fire. |
791 | */ | |
d2ab4cf4 | 792 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD); |
3935e895 | 793 | hrtimer->function = watchdog_timer_fn; |
01f0a027 | 794 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
d2ab4cf4 | 795 | HRTIMER_MODE_REL_PINNED_HARD); |
3935e895 | 796 | |
01f0a027 | 797 | /* Initialize timestamp */ |
7c0012f5 | 798 | update_touch_ts(); |
df95d308 DA |
799 | /* Enable the hardlockup detector */ |
800 | if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED) | |
801 | watchdog_hardlockup_enable(cpu); | |
bcd951cf | 802 | } |
58687acb | 803 | |
bcd951cf TG |
804 | static void watchdog_disable(unsigned int cpu) |
805 | { | |
01f0a027 | 806 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 807 | |
9cf57731 PZ |
808 | WARN_ON_ONCE(cpu != smp_processor_id()); |
809 | ||
01f0a027 | 810 | /* |
df95d308 DA |
811 | * Disable the hardlockup detector first. That prevents that a large |
812 | * delay between disabling the timer and disabling the hardlockup | |
813 | * detector causes a false positive. | |
01f0a027 | 814 | */ |
df95d308 | 815 | watchdog_hardlockup_disable(cpu); |
01f0a027 | 816 | hrtimer_cancel(hrtimer); |
be45bf53 | 817 | wait_for_completion(this_cpu_ptr(&softlockup_completion)); |
58687acb DZ |
818 | } |
819 | ||
9cf57731 | 820 | static int softlockup_stop_fn(void *data) |
b8900bc0 | 821 | { |
9cf57731 PZ |
822 | watchdog_disable(smp_processor_id()); |
823 | return 0; | |
b8900bc0 FW |
824 | } |
825 | ||
9cf57731 | 826 | static void softlockup_stop_all(void) |
bcd951cf | 827 | { |
9cf57731 PZ |
828 | int cpu; |
829 | ||
830 | if (!softlockup_initialized) | |
831 | return; | |
832 | ||
833 | for_each_cpu(cpu, &watchdog_allowed_mask) | |
834 | smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false); | |
835 | ||
836 | cpumask_clear(&watchdog_allowed_mask); | |
bcd951cf TG |
837 | } |
838 | ||
9cf57731 | 839 | static int softlockup_start_fn(void *data) |
bcd951cf | 840 | { |
9cf57731 PZ |
841 | watchdog_enable(smp_processor_id()); |
842 | return 0; | |
bcd951cf | 843 | } |
58687acb | 844 | |
9cf57731 | 845 | static void softlockup_start_all(void) |
2eb2527f | 846 | { |
9cf57731 | 847 | int cpu; |
2eb2527f | 848 | |
9cf57731 PZ |
849 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); |
850 | for_each_cpu(cpu, &watchdog_allowed_mask) | |
851 | smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false); | |
2eb2527f TG |
852 | } |
853 | ||
9cf57731 | 854 | int lockup_detector_online_cpu(unsigned int cpu) |
2eb2527f | 855 | { |
7dd47617 TG |
856 | if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) |
857 | watchdog_enable(cpu); | |
9cf57731 | 858 | return 0; |
2eb2527f TG |
859 | } |
860 | ||
9cf57731 | 861 | int lockup_detector_offline_cpu(unsigned int cpu) |
2eb2527f | 862 | { |
7dd47617 TG |
863 | if (cpumask_test_cpu(cpu, &watchdog_allowed_mask)) |
864 | watchdog_disable(cpu); | |
9cf57731 | 865 | return 0; |
2eb2527f TG |
866 | } |
867 | ||
7c56a873 | 868 | static void __lockup_detector_reconfigure(void) |
2eb2527f | 869 | { |
e31d6883 | 870 | cpus_read_lock(); |
df95d308 | 871 | watchdog_hardlockup_stop(); |
9cf57731 PZ |
872 | |
873 | softlockup_stop_all(); | |
2eb2527f | 874 | set_sample_period(); |
09154985 TG |
875 | lockup_detector_update_enable(); |
876 | if (watchdog_enabled && watchdog_thresh) | |
9cf57731 PZ |
877 | softlockup_start_all(); |
878 | ||
df95d308 | 879 | watchdog_hardlockup_start(); |
e31d6883 TG |
880 | cpus_read_unlock(); |
881 | /* | |
882 | * Must be called outside the cpus locked section to prevent | |
883 | * recursive locking in the perf code. | |
884 | */ | |
885 | __lockup_detector_cleanup(); | |
2eb2527f TG |
886 | } |
887 | ||
7c56a873 LD |
888 | void lockup_detector_reconfigure(void) |
889 | { | |
890 | mutex_lock(&watchdog_mutex); | |
891 | __lockup_detector_reconfigure(); | |
892 | mutex_unlock(&watchdog_mutex); | |
893 | } | |
894 | ||
2eb2527f | 895 | /* |
b124ac45 | 896 | * Create the watchdog infrastructure and configure the detector(s). |
2eb2527f | 897 | */ |
5587185d | 898 | static __init void lockup_detector_setup(void) |
2eb2527f | 899 | { |
2eb2527f TG |
900 | /* |
901 | * If sysctl is off and watchdog got disabled on the command line, | |
902 | * nothing to do here. | |
903 | */ | |
09154985 TG |
904 | lockup_detector_update_enable(); |
905 | ||
2eb2527f TG |
906 | if (!IS_ENABLED(CONFIG_SYSCTL) && |
907 | !(watchdog_enabled && watchdog_thresh)) | |
908 | return; | |
909 | ||
2eb2527f | 910 | mutex_lock(&watchdog_mutex); |
7c56a873 | 911 | __lockup_detector_reconfigure(); |
9cf57731 | 912 | softlockup_initialized = true; |
2eb2527f TG |
913 | mutex_unlock(&watchdog_mutex); |
914 | } | |
915 | ||
2b9d7f23 | 916 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
7c56a873 | 917 | static void __lockup_detector_reconfigure(void) |
6592ad2f | 918 | { |
e31d6883 | 919 | cpus_read_lock(); |
df95d308 | 920 | watchdog_hardlockup_stop(); |
09154985 | 921 | lockup_detector_update_enable(); |
df95d308 | 922 | watchdog_hardlockup_start(); |
e31d6883 | 923 | cpus_read_unlock(); |
6592ad2f | 924 | } |
7c56a873 LD |
925 | void lockup_detector_reconfigure(void) |
926 | { | |
927 | __lockup_detector_reconfigure(); | |
928 | } | |
5587185d | 929 | static inline void lockup_detector_setup(void) |
34ddaa3e | 930 | { |
7c56a873 | 931 | __lockup_detector_reconfigure(); |
34ddaa3e | 932 | } |
2b9d7f23 | 933 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 934 | |
941154bd TG |
935 | static void __lockup_detector_cleanup(void) |
936 | { | |
937 | lockdep_assert_held(&watchdog_mutex); | |
938 | hardlockup_detector_perf_cleanup(); | |
939 | } | |
940 | ||
941 | /** | |
942 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
943 | * | |
944 | * Caller must not hold the cpu hotplug rwsem. | |
945 | */ | |
946 | void lockup_detector_cleanup(void) | |
947 | { | |
948 | mutex_lock(&watchdog_mutex); | |
949 | __lockup_detector_cleanup(); | |
950 | mutex_unlock(&watchdog_mutex); | |
951 | } | |
952 | ||
6554fd8c TG |
953 | /** |
954 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
955 | * | |
956 | * Special interface for parisc. It prevents lockup detector warnings from | |
957 | * the default pm_poweroff() function which busy loops forever. | |
958 | */ | |
959 | void lockup_detector_soft_poweroff(void) | |
960 | { | |
961 | watchdog_enabled = 0; | |
962 | } | |
963 | ||
58cf690a UO |
964 | #ifdef CONFIG_SYSCTL |
965 | ||
b124ac45 | 966 | /* Propagate any changes to the watchdog infrastructure */ |
d57108d4 | 967 | static void proc_watchdog_update(void) |
a0c9cbb9 | 968 | { |
e8b62b2d TG |
969 | /* Remove impossible cpus to keep sysctl output clean. */ |
970 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
7c56a873 | 971 | __lockup_detector_reconfigure(); |
a0c9cbb9 UO |
972 | } |
973 | ||
ef246a21 UO |
974 | /* |
975 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
976 | * | |
df95d308 DA |
977 | * caller | table->data points to | 'which' |
978 | * -------------------|----------------------------------|------------------------------- | |
979 | * proc_watchdog | watchdog_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED | | |
980 | * | | WATCHDOG_SOFTOCKUP_ENABLED | |
981 | * -------------------|----------------------------------|------------------------------- | |
982 | * proc_nmi_watchdog | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED | |
983 | * -------------------|----------------------------------|------------------------------- | |
984 | * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED | |
ef246a21 UO |
985 | */ |
986 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
32927393 | 987 | void *buffer, size_t *lenp, loff_t *ppos) |
ef246a21 | 988 | { |
09154985 | 989 | int err, old, *param = table->data; |
ef246a21 | 990 | |
946d1977 | 991 | mutex_lock(&watchdog_mutex); |
ef246a21 | 992 | |
ef246a21 | 993 | if (!write) { |
09154985 TG |
994 | /* |
995 | * On read synchronize the userspace interface. This is a | |
996 | * racy snapshot. | |
997 | */ | |
998 | *param = (watchdog_enabled & which) != 0; | |
ef246a21 UO |
999 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
1000 | } else { | |
09154985 | 1001 | old = READ_ONCE(*param); |
ef246a21 | 1002 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
09154985 | 1003 | if (!err && old != READ_ONCE(*param)) |
d57108d4 | 1004 | proc_watchdog_update(); |
ef246a21 | 1005 | } |
946d1977 | 1006 | mutex_unlock(&watchdog_mutex); |
ef246a21 UO |
1007 | return err; |
1008 | } | |
1009 | ||
83a80a39 UO |
1010 | /* |
1011 | * /proc/sys/kernel/watchdog | |
1012 | */ | |
75060b6e TW |
1013 | static int proc_watchdog(struct ctl_table *table, int write, |
1014 | void *buffer, size_t *lenp, loff_t *ppos) | |
83a80a39 | 1015 | { |
df95d308 DA |
1016 | return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED | |
1017 | WATCHDOG_SOFTOCKUP_ENABLED, | |
83a80a39 UO |
1018 | table, write, buffer, lenp, ppos); |
1019 | } | |
1020 | ||
1021 | /* | |
1022 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 1023 | */ |
75060b6e TW |
1024 | static int proc_nmi_watchdog(struct ctl_table *table, int write, |
1025 | void *buffer, size_t *lenp, loff_t *ppos) | |
83a80a39 | 1026 | { |
df95d308 | 1027 | if (!watchdog_hardlockup_available && write) |
a994a314 | 1028 | return -ENOTSUPP; |
df95d308 | 1029 | return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED, |
83a80a39 UO |
1030 | table, write, buffer, lenp, ppos); |
1031 | } | |
1032 | ||
75060b6e | 1033 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
83a80a39 UO |
1034 | /* |
1035 | * /proc/sys/kernel/soft_watchdog | |
1036 | */ | |
75060b6e TW |
1037 | static int proc_soft_watchdog(struct ctl_table *table, int write, |
1038 | void *buffer, size_t *lenp, loff_t *ppos) | |
83a80a39 | 1039 | { |
df95d308 | 1040 | return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED, |
83a80a39 UO |
1041 | table, write, buffer, lenp, ppos); |
1042 | } | |
75060b6e | 1043 | #endif |
58687acb | 1044 | |
83a80a39 UO |
1045 | /* |
1046 | * /proc/sys/kernel/watchdog_thresh | |
1047 | */ | |
75060b6e TW |
1048 | static int proc_watchdog_thresh(struct ctl_table *table, int write, |
1049 | void *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 1050 | { |
d57108d4 | 1051 | int err, old; |
58687acb | 1052 | |
946d1977 | 1053 | mutex_lock(&watchdog_mutex); |
bcd951cf | 1054 | |
d57108d4 | 1055 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 1056 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 1057 | |
d57108d4 TG |
1058 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
1059 | proc_watchdog_update(); | |
e04ab2bc | 1060 | |
946d1977 | 1061 | mutex_unlock(&watchdog_mutex); |
b8900bc0 | 1062 | return err; |
58687acb | 1063 | } |
fe4ba3c3 CM |
1064 | |
1065 | /* | |
1066 | * The cpumask is the mask of possible cpus that the watchdog can run | |
1067 | * on, not the mask of cpus it is actually running on. This allows the | |
1068 | * user to specify a mask that will include cpus that have not yet | |
1069 | * been brought online, if desired. | |
1070 | */ | |
75060b6e TW |
1071 | static int proc_watchdog_cpumask(struct ctl_table *table, int write, |
1072 | void *buffer, size_t *lenp, loff_t *ppos) | |
fe4ba3c3 CM |
1073 | { |
1074 | int err; | |
1075 | ||
946d1977 | 1076 | mutex_lock(&watchdog_mutex); |
8c073d27 | 1077 | |
fe4ba3c3 | 1078 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 1079 | if (!err && write) |
e8b62b2d | 1080 | proc_watchdog_update(); |
5490125d | 1081 | |
946d1977 | 1082 | mutex_unlock(&watchdog_mutex); |
fe4ba3c3 CM |
1083 | return err; |
1084 | } | |
dd0693fd XN |
1085 | |
1086 | static const int sixty = 60; | |
1087 | ||
1088 | static struct ctl_table watchdog_sysctls[] = { | |
1089 | { | |
1090 | .procname = "watchdog", | |
1091 | .data = &watchdog_user_enabled, | |
1092 | .maxlen = sizeof(int), | |
1093 | .mode = 0644, | |
1094 | .proc_handler = proc_watchdog, | |
1095 | .extra1 = SYSCTL_ZERO, | |
1096 | .extra2 = SYSCTL_ONE, | |
1097 | }, | |
1098 | { | |
1099 | .procname = "watchdog_thresh", | |
1100 | .data = &watchdog_thresh, | |
1101 | .maxlen = sizeof(int), | |
1102 | .mode = 0644, | |
1103 | .proc_handler = proc_watchdog_thresh, | |
1104 | .extra1 = SYSCTL_ZERO, | |
1105 | .extra2 = (void *)&sixty, | |
1106 | }, | |
dd0693fd XN |
1107 | { |
1108 | .procname = "watchdog_cpumask", | |
1109 | .data = &watchdog_cpumask_bits, | |
1110 | .maxlen = NR_CPUS, | |
1111 | .mode = 0644, | |
1112 | .proc_handler = proc_watchdog_cpumask, | |
1113 | }, | |
1114 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR | |
1115 | { | |
1116 | .procname = "soft_watchdog", | |
df95d308 | 1117 | .data = &watchdog_softlockup_user_enabled, |
dd0693fd XN |
1118 | .maxlen = sizeof(int), |
1119 | .mode = 0644, | |
1120 | .proc_handler = proc_soft_watchdog, | |
1121 | .extra1 = SYSCTL_ZERO, | |
1122 | .extra2 = SYSCTL_ONE, | |
1123 | }, | |
1124 | { | |
1125 | .procname = "softlockup_panic", | |
1126 | .data = &softlockup_panic, | |
1127 | .maxlen = sizeof(int), | |
1128 | .mode = 0644, | |
1129 | .proc_handler = proc_dointvec_minmax, | |
1130 | .extra1 = SYSCTL_ZERO, | |
1131 | .extra2 = SYSCTL_ONE, | |
1132 | }, | |
1133 | #ifdef CONFIG_SMP | |
1134 | { | |
1135 | .procname = "softlockup_all_cpu_backtrace", | |
1136 | .data = &sysctl_softlockup_all_cpu_backtrace, | |
1137 | .maxlen = sizeof(int), | |
1138 | .mode = 0644, | |
1139 | .proc_handler = proc_dointvec_minmax, | |
1140 | .extra1 = SYSCTL_ZERO, | |
1141 | .extra2 = SYSCTL_ONE, | |
1142 | }, | |
1143 | #endif /* CONFIG_SMP */ | |
1144 | #endif | |
1145 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
1146 | { | |
1147 | .procname = "hardlockup_panic", | |
1148 | .data = &hardlockup_panic, | |
1149 | .maxlen = sizeof(int), | |
1150 | .mode = 0644, | |
1151 | .proc_handler = proc_dointvec_minmax, | |
1152 | .extra1 = SYSCTL_ZERO, | |
1153 | .extra2 = SYSCTL_ONE, | |
1154 | }, | |
1155 | #ifdef CONFIG_SMP | |
1156 | { | |
1157 | .procname = "hardlockup_all_cpu_backtrace", | |
1158 | .data = &sysctl_hardlockup_all_cpu_backtrace, | |
1159 | .maxlen = sizeof(int), | |
1160 | .mode = 0644, | |
1161 | .proc_handler = proc_dointvec_minmax, | |
1162 | .extra1 = SYSCTL_ZERO, | |
1163 | .extra2 = SYSCTL_ONE, | |
1164 | }, | |
1165 | #endif /* CONFIG_SMP */ | |
1166 | #endif | |
dd0693fd XN |
1167 | }; |
1168 | ||
9ec272c5 DA |
1169 | static struct ctl_table watchdog_hardlockup_sysctl[] = { |
1170 | { | |
1171 | .procname = "nmi_watchdog", | |
1172 | .data = &watchdog_hardlockup_user_enabled, | |
1173 | .maxlen = sizeof(int), | |
1174 | .mode = 0444, | |
1175 | .proc_handler = proc_nmi_watchdog, | |
1176 | .extra1 = SYSCTL_ZERO, | |
1177 | .extra2 = SYSCTL_ONE, | |
1178 | }, | |
9ec272c5 DA |
1179 | }; |
1180 | ||
dd0693fd XN |
1181 | static void __init watchdog_sysctl_init(void) |
1182 | { | |
1183 | register_sysctl_init("kernel", watchdog_sysctls); | |
9ec272c5 DA |
1184 | |
1185 | if (watchdog_hardlockup_available) | |
1186 | watchdog_hardlockup_sysctl[0].mode = 0644; | |
1187 | register_sysctl_init("kernel", watchdog_hardlockup_sysctl); | |
dd0693fd | 1188 | } |
9ec272c5 | 1189 | |
dd0693fd XN |
1190 | #else |
1191 | #define watchdog_sysctl_init() do { } while (0) | |
58687acb DZ |
1192 | #endif /* CONFIG_SYSCTL */ |
1193 | ||
930d8f8d LC |
1194 | static void __init lockup_detector_delay_init(struct work_struct *work); |
1195 | static bool allow_lockup_detector_init_retry __initdata; | |
1196 | ||
1197 | static struct work_struct detector_work __initdata = | |
1198 | __WORK_INITIALIZER(detector_work, lockup_detector_delay_init); | |
1199 | ||
1200 | static void __init lockup_detector_delay_init(struct work_struct *work) | |
1201 | { | |
1202 | int ret; | |
1203 | ||
1204 | ret = watchdog_hardlockup_probe(); | |
1205 | if (ret) { | |
1206 | pr_info("Delayed init of the lockup detector failed: %d\n", ret); | |
1207 | pr_info("Hard watchdog permanently disabled\n"); | |
1208 | return; | |
1209 | } | |
1210 | ||
1211 | allow_lockup_detector_init_retry = false; | |
1212 | ||
1213 | watchdog_hardlockup_available = true; | |
1214 | lockup_detector_setup(); | |
1215 | } | |
1216 | ||
1217 | /* | |
1218 | * lockup_detector_retry_init - retry init lockup detector if possible. | |
1219 | * | |
1220 | * Retry hardlockup detector init. It is useful when it requires some | |
1221 | * functionality that has to be initialized later on a particular | |
1222 | * platform. | |
1223 | */ | |
1224 | void __init lockup_detector_retry_init(void) | |
1225 | { | |
1226 | /* Must be called before late init calls */ | |
1227 | if (!allow_lockup_detector_init_retry) | |
1228 | return; | |
1229 | ||
1230 | schedule_work(&detector_work); | |
1231 | } | |
1232 | ||
1233 | /* | |
1234 | * Ensure that optional delayed hardlockup init is proceed before | |
1235 | * the init code and memory is freed. | |
1236 | */ | |
1237 | static int __init lockup_detector_check(void) | |
1238 | { | |
1239 | /* Prevent any later retry. */ | |
1240 | allow_lockup_detector_init_retry = false; | |
1241 | ||
1242 | /* Make sure no work is pending. */ | |
1243 | flush_work(&detector_work); | |
1244 | ||
9ec272c5 DA |
1245 | watchdog_sysctl_init(); |
1246 | ||
930d8f8d LC |
1247 | return 0; |
1248 | ||
1249 | } | |
1250 | late_initcall_sync(lockup_detector_check); | |
1251 | ||
004417a6 | 1252 | void __init lockup_detector_init(void) |
58687acb | 1253 | { |
13316b31 | 1254 | if (tick_nohz_full_enabled()) |
314b08ff | 1255 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
13316b31 | 1256 | |
de201559 | 1257 | cpumask_copy(&watchdog_cpumask, |
04d4e665 | 1258 | housekeeping_cpumask(HK_TYPE_TIMER)); |
fe4ba3c3 | 1259 | |
df95d308 DA |
1260 | if (!watchdog_hardlockup_probe()) |
1261 | watchdog_hardlockup_available = true; | |
930d8f8d LC |
1262 | else |
1263 | allow_lockup_detector_init_retry = true; | |
1264 | ||
5587185d | 1265 | lockup_detector_setup(); |
58687acb | 1266 | } |