]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * NMI watchdog support on APIC systems | |
3 | * | |
4 | * Started by Ingo Molnar <mingo@redhat.com> | |
5 | * | |
6 | * Fixes: | |
7 | * Mikael Pettersson : AMD K7 support for local APIC NMI watchdog. | |
8 | * Mikael Pettersson : Power Management for local APIC NMI watchdog. | |
9 | * Mikael Pettersson : Pentium 4 support for local APIC NMI watchdog. | |
10 | * Pavel Machek and | |
11 | * Mikael Pettersson : PM converted to driver model. Disable/enable API. | |
12 | */ | |
13 | ||
14 | #include <linux/delay.h> | |
15 | #include <linux/interrupt.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/nmi.h> | |
18 | #include <linux/sysdev.h> | |
19 | #include <linux/sysctl.h> | |
20 | #include <linux/percpu.h> | |
21 | #include <linux/kprobes.h> | |
22 | #include <linux/cpumask.h> | |
23 | #include <linux/kernel_stat.h> | |
24 | #include <linux/kdebug.h> | |
25 | #include <linux/slab.h> | |
26 | ||
27 | #include <asm/smp.h> | |
28 | #include <asm/nmi.h> | |
29 | ||
30 | #include "mach_traps.h" | |
31 | ||
32 | int unknown_nmi_panic; | |
33 | int nmi_watchdog_enabled; | |
34 | ||
35 | static cpumask_t backtrace_mask = CPU_MASK_NONE; | |
36 | ||
37 | /* nmi_active: | |
38 | * >0: the lapic NMI watchdog is active, but can be disabled | |
39 | * <0: the lapic NMI watchdog has not been set up, and cannot | |
40 | * be enabled | |
41 | * 0: the lapic NMI watchdog is disabled, but can be enabled | |
42 | */ | |
43 | atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */ | |
44 | ||
45 | unsigned int nmi_watchdog = NMI_DEFAULT; | |
46 | static unsigned int nmi_hz = HZ; | |
47 | ||
48 | static DEFINE_PER_CPU(short, wd_enabled); | |
49 | ||
50 | static int endflag __initdata = 0; | |
51 | ||
52 | #ifdef CONFIG_SMP | |
53 | /* The performance counters used by NMI_LOCAL_APIC don't trigger when | |
54 | * the CPU is idle. To make sure the NMI watchdog really ticks on all | |
55 | * CPUs during the test make them busy. | |
56 | */ | |
57 | static __init void nmi_cpu_busy(void *data) | |
58 | { | |
59 | local_irq_enable_in_hardirq(); | |
60 | /* Intentionally don't use cpu_relax here. This is | |
61 | to make sure that the performance counter really ticks, | |
62 | even if there is a simulator or similar that catches the | |
63 | pause instruction. On a real HT machine this is fine because | |
64 | all other CPUs are busy with "useless" delay loops and don't | |
65 | care if they get somewhat less cycles. */ | |
66 | while (endflag == 0) | |
67 | mb(); | |
68 | } | |
69 | #endif | |
70 | ||
71 | int __init check_nmi_watchdog(void) | |
72 | { | |
73 | unsigned int *prev_nmi_count; | |
74 | int cpu; | |
75 | ||
76 | if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED)) | |
77 | return 0; | |
78 | ||
79 | if (!atomic_read(&nmi_active)) | |
80 | return 0; | |
81 | ||
82 | prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL); | |
83 | if (!prev_nmi_count) | |
84 | return -1; | |
85 | ||
86 | printk(KERN_INFO "Testing NMI watchdog ... "); | |
87 | ||
88 | #ifdef CONFIG_SMP | |
89 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
90 | smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0); | |
91 | #endif | |
92 | ||
93 | for_each_possible_cpu(cpu) | |
94 | prev_nmi_count[cpu] = nmi_count(cpu); | |
95 | local_irq_enable(); | |
96 | mdelay((20*1000)/nmi_hz); // wait 20 ticks | |
97 | ||
98 | for_each_possible_cpu(cpu) { | |
99 | #ifdef CONFIG_SMP | |
100 | /* Check cpu_callin_map here because that is set | |
101 | after the timer is started. */ | |
102 | if (!cpu_isset(cpu, cpu_callin_map)) | |
103 | continue; | |
104 | #endif | |
105 | if (!per_cpu(wd_enabled, cpu)) | |
106 | continue; | |
107 | if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) { | |
108 | printk(KERN_WARNING "WARNING: CPU#%d: NMI " | |
109 | "appears to be stuck (%d->%d)!\n", | |
110 | cpu, | |
111 | prev_nmi_count[cpu], | |
112 | nmi_count(cpu)); | |
113 | per_cpu(wd_enabled, cpu) = 0; | |
114 | atomic_dec(&nmi_active); | |
115 | } | |
116 | } | |
117 | endflag = 1; | |
118 | if (!atomic_read(&nmi_active)) { | |
119 | kfree(prev_nmi_count); | |
120 | atomic_set(&nmi_active, -1); | |
121 | return -1; | |
122 | } | |
123 | printk("OK.\n"); | |
124 | ||
125 | /* now that we know it works we can reduce NMI frequency to | |
126 | something more reasonable; makes a difference in some configs */ | |
127 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
128 | nmi_hz = lapic_adjust_nmi_hz(1); | |
129 | ||
130 | kfree(prev_nmi_count); | |
131 | return 0; | |
132 | } | |
133 | ||
134 | static int __init setup_nmi_watchdog(char *str) | |
135 | { | |
136 | int nmi; | |
137 | ||
138 | get_option(&str, &nmi); | |
139 | ||
140 | if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE)) | |
141 | return 0; | |
142 | ||
143 | nmi_watchdog = nmi; | |
144 | return 1; | |
145 | } | |
146 | ||
147 | __setup("nmi_watchdog=", setup_nmi_watchdog); | |
148 | ||
149 | ||
150 | /* Suspend/resume support */ | |
151 | ||
152 | #ifdef CONFIG_PM | |
153 | ||
154 | static int nmi_pm_active; /* nmi_active before suspend */ | |
155 | ||
156 | static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state) | |
157 | { | |
158 | /* only CPU0 goes here, other CPUs should be offline */ | |
159 | nmi_pm_active = atomic_read(&nmi_active); | |
160 | stop_apic_nmi_watchdog(NULL); | |
161 | BUG_ON(atomic_read(&nmi_active) != 0); | |
162 | return 0; | |
163 | } | |
164 | ||
165 | static int lapic_nmi_resume(struct sys_device *dev) | |
166 | { | |
167 | /* only CPU0 goes here, other CPUs should be offline */ | |
168 | if (nmi_pm_active > 0) { | |
169 | setup_apic_nmi_watchdog(NULL); | |
170 | touch_nmi_watchdog(); | |
171 | } | |
172 | return 0; | |
173 | } | |
174 | ||
175 | ||
176 | static struct sysdev_class nmi_sysclass = { | |
177 | .name = "lapic_nmi", | |
178 | .resume = lapic_nmi_resume, | |
179 | .suspend = lapic_nmi_suspend, | |
180 | }; | |
181 | ||
182 | static struct sys_device device_lapic_nmi = { | |
183 | .id = 0, | |
184 | .cls = &nmi_sysclass, | |
185 | }; | |
186 | ||
187 | static int __init init_lapic_nmi_sysfs(void) | |
188 | { | |
189 | int error; | |
190 | ||
191 | /* should really be a BUG_ON but b/c this is an | |
192 | * init call, it just doesn't work. -dcz | |
193 | */ | |
194 | if (nmi_watchdog != NMI_LOCAL_APIC) | |
195 | return 0; | |
196 | ||
197 | if (atomic_read(&nmi_active) < 0) | |
198 | return 0; | |
199 | ||
200 | error = sysdev_class_register(&nmi_sysclass); | |
201 | if (!error) | |
202 | error = sysdev_register(&device_lapic_nmi); | |
203 | return error; | |
204 | } | |
205 | /* must come after the local APIC's device_initcall() */ | |
206 | late_initcall(init_lapic_nmi_sysfs); | |
207 | ||
208 | #endif /* CONFIG_PM */ | |
209 | ||
210 | static void __acpi_nmi_enable(void *__unused) | |
211 | { | |
212 | apic_write_around(APIC_LVT0, APIC_DM_NMI); | |
213 | } | |
214 | ||
215 | /* | |
216 | * Enable timer based NMIs on all CPUs: | |
217 | */ | |
218 | void acpi_nmi_enable(void) | |
219 | { | |
220 | if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC) | |
221 | on_each_cpu(__acpi_nmi_enable, NULL, 0, 1); | |
222 | } | |
223 | ||
224 | static void __acpi_nmi_disable(void *__unused) | |
225 | { | |
226 | apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED); | |
227 | } | |
228 | ||
229 | /* | |
230 | * Disable timer based NMIs on all CPUs: | |
231 | */ | |
232 | void acpi_nmi_disable(void) | |
233 | { | |
234 | if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC) | |
235 | on_each_cpu(__acpi_nmi_disable, NULL, 0, 1); | |
236 | } | |
237 | ||
238 | void setup_apic_nmi_watchdog(void *unused) | |
239 | { | |
240 | if (__get_cpu_var(wd_enabled)) | |
241 | return; | |
242 | ||
243 | /* cheap hack to support suspend/resume */ | |
244 | /* if cpu0 is not active neither should the other cpus */ | |
245 | if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0)) | |
246 | return; | |
247 | ||
248 | switch (nmi_watchdog) { | |
249 | case NMI_LOCAL_APIC: | |
250 | __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */ | |
251 | if (lapic_watchdog_init(nmi_hz) < 0) { | |
252 | __get_cpu_var(wd_enabled) = 0; | |
253 | return; | |
254 | } | |
255 | /* FALL THROUGH */ | |
256 | case NMI_IO_APIC: | |
257 | __get_cpu_var(wd_enabled) = 1; | |
258 | atomic_inc(&nmi_active); | |
259 | } | |
260 | } | |
261 | ||
262 | void stop_apic_nmi_watchdog(void *unused) | |
263 | { | |
264 | /* only support LOCAL and IO APICs for now */ | |
265 | if ((nmi_watchdog != NMI_LOCAL_APIC) && | |
266 | (nmi_watchdog != NMI_IO_APIC)) | |
267 | return; | |
268 | if (__get_cpu_var(wd_enabled) == 0) | |
269 | return; | |
270 | if (nmi_watchdog == NMI_LOCAL_APIC) | |
271 | lapic_watchdog_stop(); | |
272 | __get_cpu_var(wd_enabled) = 0; | |
273 | atomic_dec(&nmi_active); | |
274 | } | |
275 | ||
276 | /* | |
277 | * the best way to detect whether a CPU has a 'hard lockup' problem | |
278 | * is to check it's local APIC timer IRQ counts. If they are not | |
279 | * changing then that CPU has some problem. | |
280 | * | |
281 | * as these watchdog NMI IRQs are generated on every CPU, we only | |
282 | * have to check the current processor. | |
283 | * | |
284 | * since NMIs don't listen to _any_ locks, we have to be extremely | |
285 | * careful not to rely on unsafe variables. The printk might lock | |
286 | * up though, so we have to break up any console locks first ... | |
287 | * [when there will be more tty-related locks, break them up | |
288 | * here too!] | |
289 | */ | |
290 | ||
291 | static unsigned int | |
292 | last_irq_sums [NR_CPUS], | |
293 | alert_counter [NR_CPUS]; | |
294 | ||
295 | void touch_nmi_watchdog(void) | |
296 | { | |
297 | if (nmi_watchdog > 0) { | |
298 | unsigned cpu; | |
299 | ||
300 | /* | |
301 | * Just reset the alert counters, (other CPUs might be | |
302 | * spinning on locks we hold): | |
303 | */ | |
304 | for_each_present_cpu(cpu) { | |
305 | if (alert_counter[cpu]) | |
306 | alert_counter[cpu] = 0; | |
307 | } | |
308 | } | |
309 | ||
310 | /* | |
311 | * Tickle the softlockup detector too: | |
312 | */ | |
313 | touch_softlockup_watchdog(); | |
314 | } | |
315 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
316 | ||
317 | extern void die_nmi(struct pt_regs *, const char *msg); | |
318 | ||
319 | notrace __kprobes int | |
320 | nmi_watchdog_tick(struct pt_regs *regs, unsigned reason) | |
321 | { | |
322 | ||
323 | /* | |
324 | * Since current_thread_info()-> is always on the stack, and we | |
325 | * always switch the stack NMI-atomically, it's safe to use | |
326 | * smp_processor_id(). | |
327 | */ | |
328 | unsigned int sum; | |
329 | int touched = 0; | |
330 | int cpu = smp_processor_id(); | |
331 | int rc = 0; | |
332 | ||
333 | /* check for other users first */ | |
334 | if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT) | |
335 | == NOTIFY_STOP) { | |
336 | rc = 1; | |
337 | touched = 1; | |
338 | } | |
339 | ||
340 | if (cpu_isset(cpu, backtrace_mask)) { | |
341 | static DEFINE_SPINLOCK(lock); /* Serialise the printks */ | |
342 | ||
343 | spin_lock(&lock); | |
344 | printk("NMI backtrace for cpu %d\n", cpu); | |
345 | dump_stack(); | |
346 | spin_unlock(&lock); | |
347 | cpu_clear(cpu, backtrace_mask); | |
348 | } | |
349 | ||
350 | /* | |
351 | * Take the local apic timer and PIT/HPET into account. We don't | |
352 | * know which one is active, when we have highres/dyntick on | |
353 | */ | |
354 | sum = per_cpu(irq_stat, cpu).apic_timer_irqs + | |
355 | per_cpu(irq_stat, cpu).irq0_irqs; | |
356 | ||
357 | /* if the none of the timers isn't firing, this cpu isn't doing much */ | |
358 | if (!touched && last_irq_sums[cpu] == sum) { | |
359 | /* | |
360 | * Ayiee, looks like this CPU is stuck ... | |
361 | * wait a few IRQs (5 seconds) before doing the oops ... | |
362 | */ | |
363 | alert_counter[cpu]++; | |
364 | if (alert_counter[cpu] == 5*nmi_hz) | |
365 | /* | |
366 | * die_nmi will return ONLY if NOTIFY_STOP happens.. | |
367 | */ | |
368 | die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP"); | |
369 | } else { | |
370 | last_irq_sums[cpu] = sum; | |
371 | alert_counter[cpu] = 0; | |
372 | } | |
373 | /* see if the nmi watchdog went off */ | |
374 | if (!__get_cpu_var(wd_enabled)) | |
375 | return rc; | |
376 | switch (nmi_watchdog) { | |
377 | case NMI_LOCAL_APIC: | |
378 | rc |= lapic_wd_event(nmi_hz); | |
379 | break; | |
380 | case NMI_IO_APIC: | |
381 | /* don't know how to accurately check for this. | |
382 | * just assume it was a watchdog timer interrupt | |
383 | * This matches the old behaviour. | |
384 | */ | |
385 | rc = 1; | |
386 | break; | |
387 | } | |
388 | return rc; | |
389 | } | |
390 | ||
391 | #ifdef CONFIG_SYSCTL | |
392 | ||
393 | static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu) | |
394 | { | |
395 | unsigned char reason = get_nmi_reason(); | |
396 | char buf[64]; | |
397 | ||
398 | sprintf(buf, "NMI received for unknown reason %02x\n", reason); | |
399 | die_nmi(regs, buf); | |
400 | return 0; | |
401 | } | |
402 | ||
403 | /* | |
404 | * proc handler for /proc/sys/kernel/nmi | |
405 | */ | |
406 | int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file, | |
407 | void __user *buffer, size_t *length, loff_t *ppos) | |
408 | { | |
409 | int old_state; | |
410 | ||
411 | nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0; | |
412 | old_state = nmi_watchdog_enabled; | |
413 | proc_dointvec(table, write, file, buffer, length, ppos); | |
414 | if (!!old_state == !!nmi_watchdog_enabled) | |
415 | return 0; | |
416 | ||
417 | if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) { | |
418 | printk( KERN_WARNING "NMI watchdog is permanently disabled\n"); | |
419 | return -EIO; | |
420 | } | |
421 | ||
422 | if (nmi_watchdog == NMI_DEFAULT) { | |
423 | if (lapic_watchdog_ok()) | |
424 | nmi_watchdog = NMI_LOCAL_APIC; | |
425 | else | |
426 | nmi_watchdog = NMI_IO_APIC; | |
427 | } | |
428 | ||
429 | if (nmi_watchdog == NMI_LOCAL_APIC) { | |
430 | if (nmi_watchdog_enabled) | |
431 | enable_lapic_nmi_watchdog(); | |
432 | else | |
433 | disable_lapic_nmi_watchdog(); | |
434 | } else { | |
435 | printk( KERN_WARNING | |
436 | "NMI watchdog doesn't know what hardware to touch\n"); | |
437 | return -EIO; | |
438 | } | |
439 | return 0; | |
440 | } | |
441 | ||
442 | #endif | |
443 | ||
444 | int do_nmi_callback(struct pt_regs *regs, int cpu) | |
445 | { | |
446 | #ifdef CONFIG_SYSCTL | |
447 | if (unknown_nmi_panic) | |
448 | return unknown_nmi_panic_callback(regs, cpu); | |
449 | #endif | |
450 | return 0; | |
451 | } | |
452 | ||
453 | void __trigger_all_cpu_backtrace(void) | |
454 | { | |
455 | int i; | |
456 | ||
457 | backtrace_mask = cpu_online_map; | |
458 | /* Wait for up to 10 seconds for all CPUs to do the backtrace */ | |
459 | for (i = 0; i < 10 * 1000; i++) { | |
460 | if (cpus_empty(backtrace_mask)) | |
461 | break; | |
462 | mdelay(1); | |
463 | } | |
464 | } | |
465 | ||
466 | EXPORT_SYMBOL(nmi_active); | |
467 | EXPORT_SYMBOL(nmi_watchdog); |