blob: 9a42b8bc0cc6010947e6c0491761953a543b0bb0
1 | /* |
2 | * Detect hard lockups on a system |
3 | * |
4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. |
5 | * |
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 |
9 | * to those contributors as well. |
10 | */ |
11 | |
12 | #define pr_fmt(fmt) "NMI watchdog: " fmt |
13 | |
14 | #include <linux/nmi.h> |
15 | #include <linux/module.h> |
16 | #include <asm/irq_regs.h> |
17 | #include <linux/perf_event.h> |
18 | |
19 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
20 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); |
21 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU |
22 | static cpumask_t __read_mostly watchdog_cpus; |
23 | #endif |
24 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI |
25 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); |
26 | #endif |
27 | DECLARE_PER_CPU(unsigned long, hrtimer_interrupts); |
28 | DECLARE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
29 | |
30 | /* boot commands */ |
31 | /* |
32 | * Should we panic when a soft-lockup or hard-lockup occurs: |
33 | */ |
34 | unsigned int __read_mostly hardlockup_panic = |
35 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; |
36 | static unsigned long __maybe_unused hardlockup_allcpu_dumped; |
37 | /* |
38 | * We may not want to enable hard lockup detection by default in all cases, |
39 | * for example when running the kernel as a guest on a hypervisor. In these |
40 | * cases this function can be called to disable hard lockup detection. This |
41 | * function should only be executed once by the boot processor before the |
42 | * kernel command line parameters are parsed, because otherwise it is not |
43 | * possible to override this in hardlockup_panic_setup(). |
44 | */ |
45 | void hardlockup_detector_disable(void) |
46 | { |
47 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
48 | } |
49 | |
50 | static int __init hardlockup_panic_setup(char *str) |
51 | { |
52 | if (!strncmp(str, "panic", 5)) |
53 | hardlockup_panic = 1; |
54 | else if (!strncmp(str, "nopanic", 7)) |
55 | hardlockup_panic = 0; |
56 | else if (!strncmp(str, "0", 1)) |
57 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
58 | else if (!strncmp(str, "1", 1)) |
59 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; |
60 | return 1; |
61 | } |
62 | __setup("nmi_watchdog=", hardlockup_panic_setup); |
63 | |
64 | void touch_nmi_watchdog(void) |
65 | { |
66 | /* |
67 | * Using __raw here because some code paths have |
68 | * preemption enabled. If preemption is enabled |
69 | * then interrupts should be enabled too, in which |
70 | * case we shouldn't have to worry about the watchdog |
71 | * going off. |
72 | */ |
73 | raw_cpu_write(watchdog_nmi_touch, true); |
74 | touch_softlockup_watchdog(); |
75 | } |
76 | EXPORT_SYMBOL(touch_nmi_watchdog); |
77 | |
78 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI |
79 | static struct perf_event_attr wd_hw_attr = { |
80 | .type = PERF_TYPE_HARDWARE, |
81 | .config = PERF_COUNT_HW_CPU_CYCLES, |
82 | .size = sizeof(struct perf_event_attr), |
83 | .pinned = 1, |
84 | .disabled = 1, |
85 | }; |
86 | |
87 | /* Callback function for perf event subsystem */ |
88 | static void watchdog_overflow_callback(struct perf_event *event, |
89 | struct perf_sample_data *data, |
90 | struct pt_regs *regs) |
91 | { |
92 | /* Ensure the watchdog never gets throttled */ |
93 | event->hw.interrupts = 0; |
94 | |
95 | if (atomic_read(&watchdog_park_in_progress) != 0) |
96 | return; |
97 | |
98 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
99 | __this_cpu_write(watchdog_nmi_touch, false); |
100 | return; |
101 | } |
102 | |
103 | /* check for a hardlockup |
104 | * This is done by making sure our timer interrupt |
105 | * is incrementing. The timer interrupt should have |
106 | * fired multiple times before we overflow'd. If it hasn't |
107 | * then this is a good indication the cpu is stuck |
108 | */ |
109 | if (is_hardlockup()) { |
110 | int this_cpu = smp_processor_id(); |
111 | |
112 | /* only print hardlockups once */ |
113 | if (__this_cpu_read(hard_watchdog_warn) == true) |
114 | return; |
115 | |
116 | pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu); |
117 | print_modules(); |
118 | print_irqtrace_events(current); |
119 | if (regs) |
120 | show_regs(regs); |
121 | else |
122 | dump_stack(); |
123 | |
124 | /* |
125 | * Perform all-CPU dump only once to avoid multiple hardlockups |
126 | * generating interleaving traces |
127 | */ |
128 | if (sysctl_hardlockup_all_cpu_backtrace && |
129 | !test_and_set_bit(0, &hardlockup_allcpu_dumped)) |
130 | trigger_allbutself_cpu_backtrace(); |
131 | |
132 | #ifdef CONFIG_AMLOGIC_DEBUG_LOCKUP |
133 | pr_lockup_info(next_cpu); |
134 | #endif |
135 | if (hardlockup_panic) |
136 | nmi_panic(regs, "Hard LOCKUP"); |
137 | |
138 | __this_cpu_write(hard_watchdog_warn, true); |
139 | return; |
140 | } |
141 | |
142 | __this_cpu_write(hard_watchdog_warn, false); |
143 | return; |
144 | } |
145 | #endif |
146 | |
147 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU |
148 | static unsigned int watchdog_next_cpu(unsigned int cpu) |
149 | { |
150 | cpumask_t cpus = watchdog_cpus; |
151 | unsigned int next_cpu; |
152 | |
153 | next_cpu = cpumask_next(cpu, &cpus); |
154 | if (next_cpu >= nr_cpu_ids) |
155 | next_cpu = cpumask_first(&cpus); |
156 | |
157 | if (next_cpu == cpu) |
158 | return nr_cpu_ids; |
159 | |
160 | return next_cpu; |
161 | } |
162 | |
163 | static int is_hardlockup_other_cpu(unsigned int cpu) |
164 | { |
165 | unsigned long hrint = per_cpu(hrtimer_interrupts, cpu); |
166 | |
167 | if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint) |
168 | return 1; |
169 | |
170 | per_cpu(hrtimer_interrupts_saved, cpu) = hrint; |
171 | return 0; |
172 | } |
173 | |
174 | void watchdog_check_hardlockup_other_cpu(void) |
175 | { |
176 | unsigned int next_cpu; |
177 | |
178 | /* |
179 | * Test for hardlockups every 3 samples. The sample period is |
180 | * watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over |
181 | * watchdog_thresh (over by 20%). |
182 | */ |
183 | if (__this_cpu_read(hrtimer_interrupts) % 3 != 0) |
184 | return; |
185 | |
186 | /* check for a hardlockup on the next cpu */ |
187 | next_cpu = watchdog_next_cpu(smp_processor_id()); |
188 | if (next_cpu >= nr_cpu_ids) |
189 | return; |
190 | |
191 | /*mem barrier*/ |
192 | smp_rmb(); |
193 | |
194 | if (per_cpu(watchdog_nmi_touch, next_cpu) == true) { |
195 | per_cpu(watchdog_nmi_touch, next_cpu) = false; |
196 | return; |
197 | } |
198 | |
199 | if (is_hardlockup_other_cpu(next_cpu)) { |
200 | /* only warn once */ |
201 | if (per_cpu(hard_watchdog_warn, next_cpu) == true) |
202 | return; |
203 | |
204 | #ifdef CONFIG_AMLOGIC_DEBUG_LOCKUP |
205 | pr_lockup_info(next_cpu); |
206 | #endif |
207 | if (hardlockup_panic) |
208 | panic("Watchdog detected hard LOCKUP on cpu %u", |
209 | next_cpu); |
210 | else |
211 | WARN(1, "Watchdog detected hard LOCKUP on cpu %u", |
212 | next_cpu); |
213 | |
214 | per_cpu(hard_watchdog_warn, next_cpu) = true; |
215 | } else { |
216 | per_cpu(hard_watchdog_warn, next_cpu) = false; |
217 | } |
218 | } |
219 | #else |
220 | static inline void watchdog_check_hardlockup_other_cpu(void) { return; } |
221 | #endif |
222 | |
223 | |
224 | |
225 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI |
226 | /* |
227 | * People like the simple clean cpu node info on boot. |
228 | * Reduce the watchdog noise by only printing messages |
229 | * that are different from what cpu0 displayed. |
230 | */ |
231 | static unsigned long cpu0_err; |
232 | |
233 | int watchdog_nmi_enable(unsigned int cpu) |
234 | { |
235 | struct perf_event_attr *wd_attr; |
236 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
237 | |
238 | /* nothing to do if the hard lockup detector is disabled */ |
239 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) |
240 | goto out; |
241 | |
242 | /* is it already setup and enabled? */ |
243 | if (event && event->state > PERF_EVENT_STATE_OFF) |
244 | goto out; |
245 | |
246 | /* it is setup but not enabled */ |
247 | if (event != NULL) |
248 | goto out_enable; |
249 | |
250 | wd_attr = &wd_hw_attr; |
251 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); |
252 | |
253 | /* Try to register using hardware perf events */ |
254 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); |
255 | |
256 | /* save cpu0 error for future comparision */ |
257 | if (cpu == 0 && IS_ERR(event)) |
258 | cpu0_err = PTR_ERR(event); |
259 | |
260 | if (!IS_ERR(event)) { |
261 | /* only print for cpu0 or different than cpu0 */ |
262 | if (cpu == 0 || cpu0_err) |
263 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); |
264 | goto out_save; |
265 | } |
266 | |
267 | /* |
268 | * Disable the hard lockup detector if _any_ CPU fails to set up |
269 | * set up the hardware perf event. The watchdog() function checks |
270 | * the NMI_WATCHDOG_ENABLED bit periodically. |
271 | * |
272 | * The barriers are for syncing up watchdog_enabled across all the |
273 | * cpus, as clear_bit() does not use barriers. |
274 | */ |
275 | smp_mb__before_atomic(); |
276 | clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled); |
277 | smp_mb__after_atomic(); |
278 | |
279 | /* skip displaying the same error again */ |
280 | if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) |
281 | return PTR_ERR(event); |
282 | |
283 | /* vary the KERN level based on the returned errno */ |
284 | if (PTR_ERR(event) == -EOPNOTSUPP) |
285 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); |
286 | else if (PTR_ERR(event) == -ENOENT) |
287 | pr_warn("disabled (cpu%i): hardware events not enabled\n", |
288 | cpu); |
289 | else |
290 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", |
291 | cpu, PTR_ERR(event)); |
292 | |
293 | pr_info("Shutting down hard lockup detector on all cpus\n"); |
294 | |
295 | return PTR_ERR(event); |
296 | |
297 | /* success path */ |
298 | out_save: |
299 | per_cpu(watchdog_ev, cpu) = event; |
300 | out_enable: |
301 | perf_event_enable(per_cpu(watchdog_ev, cpu)); |
302 | out: |
303 | return 0; |
304 | } |
305 | |
306 | void watchdog_nmi_disable(unsigned int cpu) |
307 | { |
308 | struct perf_event *event = per_cpu(watchdog_ev, cpu); |
309 | |
310 | if (event) { |
311 | perf_event_disable(event); |
312 | per_cpu(watchdog_ev, cpu) = NULL; |
313 | |
314 | /* should be in cleanup, but blocks oprofile */ |
315 | perf_event_release_kernel(event); |
316 | } |
317 | if (cpu == 0) { |
318 | /* watchdog_nmi_enable() expects this to be zero initially. */ |
319 | cpu0_err = 0; |
320 | } |
321 | } |
322 | #else |
323 | #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU |
324 | int watchdog_nmi_enable(unsigned int cpu) |
325 | { |
326 | /* |
327 | * The new cpu will be marked online before the first hrtimer interrupt |
328 | * runs on it. If another cpu tests for a hardlockup on the new cpu |
329 | * before it has run its first hrtimer, it will get a false positive. |
330 | * Touch the watchdog on the new cpu to delay the first check for at |
331 | * least 3 sampling periods to guarantee one hrtimer has run on the new |
332 | * cpu. |
333 | */ |
334 | per_cpu(watchdog_nmi_touch, cpu) = true; |
335 | /*mem barrier*/ |
336 | smp_wmb(); |
337 | cpumask_set_cpu(cpu, &watchdog_cpus); |
338 | return 0; |
339 | } |
340 | |
341 | void watchdog_nmi_disable(unsigned int cpu) |
342 | { |
343 | unsigned int next_cpu = watchdog_next_cpu(cpu); |
344 | |
345 | /* |
346 | * Offlining this cpu will cause the cpu before this one to start |
347 | * checking the one after this one. If this cpu just finished checking |
348 | * the next cpu and updating hrtimer_interrupts_saved, and then the |
349 | * previous cpu checks it within one sample period, it will trigger a |
350 | * false positive. Touch the watchdog on the next cpu to prevent it. |
351 | */ |
352 | if (next_cpu < nr_cpu_ids) |
353 | per_cpu(watchdog_nmi_touch, next_cpu) = true; |
354 | /*mem barrier*/ |
355 | smp_wmb(); |
356 | cpumask_clear_cpu(cpu, &watchdog_cpus); |
357 | } |
358 | #else |
359 | int watchdog_nmi_enable(unsigned int cpu) { return 0; } |
360 | void watchdog_nmi_disable(unsigned int cpu) { return; } |
361 | #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */ |
362 | #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */ |
363 |