x86, NMI: fix watchdog failure message
[linux-2.6] / arch / x86 / kernel / nmi.c
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 <asm/apic.h>
15
16 #include <linux/nmi.h>
17 #include <linux/mm.h>
18 #include <linux/delay.h>
19 #include <linux/interrupt.h>
20 #include <linux/module.h>
21 #include <linux/sysdev.h>
22 #include <linux/sysctl.h>
23 #include <linux/percpu.h>
24 #include <linux/kprobes.h>
25 #include <linux/cpumask.h>
26 #include <linux/kernel_stat.h>
27 #include <linux/kdebug.h>
28 #include <linux/smp.h>
29
30 #include <asm/i8259.h>
31 #include <asm/io_apic.h>
32 #include <asm/smp.h>
33 #include <asm/nmi.h>
34 #include <asm/proto.h>
35 #include <asm/timer.h>
36
37 #include <asm/mce.h>
38
39 #include <mach_traps.h>
40
41 int unknown_nmi_panic;
42 int nmi_watchdog_enabled;
43
44 static cpumask_t backtrace_mask = CPU_MASK_NONE;
45
46 /* nmi_active:
47  * >0: the lapic NMI watchdog is active, but can be disabled
48  * <0: the lapic NMI watchdog has not been set up, and cannot
49  *     be enabled
50  *  0: the lapic NMI watchdog is disabled, but can be enabled
51  */
52 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
53 EXPORT_SYMBOL(nmi_active);
54
55 unsigned int nmi_watchdog = NMI_NONE;
56 EXPORT_SYMBOL(nmi_watchdog);
57
58 static int panic_on_timeout;
59
60 static unsigned int nmi_hz = HZ;
61 static DEFINE_PER_CPU(short, wd_enabled);
62 static int endflag __initdata;
63
64 static inline unsigned int get_nmi_count(int cpu)
65 {
66 #ifdef CONFIG_X86_64
67         return cpu_pda(cpu)->__nmi_count;
68 #else
69         return nmi_count(cpu);
70 #endif
71 }
72
73 static inline int mce_in_progress(void)
74 {
75 #if defined(CONFIG_X86_64) && defined(CONFIG_X86_MCE)
76         return atomic_read(&mce_entry) > 0;
77 #endif
78         return 0;
79 }
80
81 /*
82  * Take the local apic timer and PIT/HPET into account. We don't
83  * know which one is active, when we have highres/dyntick on
84  */
85 static inline unsigned int get_timer_irqs(int cpu)
86 {
87 #ifdef CONFIG_X86_64
88         return read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
89 #else
90         return per_cpu(irq_stat, cpu).apic_timer_irqs +
91                 per_cpu(irq_stat, cpu).irq0_irqs;
92 #endif
93 }
94
95 #ifdef CONFIG_SMP
96 /*
97  * The performance counters used by NMI_LOCAL_APIC don't trigger when
98  * the CPU is idle. To make sure the NMI watchdog really ticks on all
99  * CPUs during the test make them busy.
100  */
101 static __init void nmi_cpu_busy(void *data)
102 {
103         local_irq_enable_in_hardirq();
104         /*
105          * Intentionally don't use cpu_relax here. This is
106          * to make sure that the performance counter really ticks,
107          * even if there is a simulator or similar that catches the
108          * pause instruction. On a real HT machine this is fine because
109          * all other CPUs are busy with "useless" delay loops and don't
110          * care if they get somewhat less cycles.
111          */
112         while (endflag == 0)
113                 mb();
114 }
115 #endif
116
117 int __init check_nmi_watchdog(void)
118 {
119         unsigned int *prev_nmi_count;
120         int cpu;
121
122         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
123                 return 0;
124
125         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
126         if (!prev_nmi_count)
127                 goto error;
128
129         printk(KERN_INFO "Testing NMI watchdog ... ");
130
131 #ifdef CONFIG_SMP
132         if (nmi_watchdog == NMI_LOCAL_APIC)
133                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
134 #endif
135
136         for_each_possible_cpu(cpu)
137                 prev_nmi_count[cpu] = get_nmi_count(cpu);
138         local_irq_enable();
139         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
140
141         for_each_online_cpu(cpu) {
142                 if (!per_cpu(wd_enabled, cpu))
143                         continue;
144                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
145                         printk("\n");
146                         printk(KERN_WARNING "WARNING: CPU#%d: NMI "
147                                 "appears to be stuck (%d->%d)!\n",
148                                 cpu,
149                                 prev_nmi_count[cpu],
150                                 get_nmi_count(cpu));
151                         printk(KERN_WARNING "Please report this to "
152                                "linux-kernel@vger.kernel.org and attach "
153                                "the output of 'dmesg' command.\n");
154                         per_cpu(wd_enabled, cpu) = 0;
155                         atomic_dec(&nmi_active);
156                 }
157         }
158         endflag = 1;
159         if (!atomic_read(&nmi_active)) {
160                 kfree(prev_nmi_count);
161                 atomic_set(&nmi_active, -1);
162                 goto error;
163         }
164         printk("OK.\n");
165
166         /*
167          * now that we know it works we can reduce NMI frequency to
168          * something more reasonable; makes a difference in some configs
169          */
170         if (nmi_watchdog == NMI_LOCAL_APIC)
171                 nmi_hz = lapic_adjust_nmi_hz(1);
172
173         kfree(prev_nmi_count);
174         return 0;
175 error:
176         if (nmi_watchdog == NMI_IO_APIC && !timer_through_8259)
177                 disable_8259A_irq(0);
178 #ifdef CONFIG_X86_32
179         timer_ack = 0;
180 #endif
181         return -1;
182 }
183
184 static int __init setup_nmi_watchdog(char *str)
185 {
186         unsigned int nmi;
187
188         if (!strncmp(str, "panic", 5)) {
189                 panic_on_timeout = 1;
190                 str = strchr(str, ',');
191                 if (!str)
192                         return 1;
193                 ++str;
194         }
195
196         get_option(&str, &nmi);
197
198         if (nmi >= NMI_INVALID)
199                 return 0;
200
201         nmi_watchdog = nmi;
202         return 1;
203 }
204 __setup("nmi_watchdog=", setup_nmi_watchdog);
205
206 /*
207  * Suspend/resume support
208  */
209 #ifdef CONFIG_PM
210
211 static int nmi_pm_active; /* nmi_active before suspend */
212
213 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
214 {
215         /* only CPU0 goes here, other CPUs should be offline */
216         nmi_pm_active = atomic_read(&nmi_active);
217         stop_apic_nmi_watchdog(NULL);
218         BUG_ON(atomic_read(&nmi_active) != 0);
219         return 0;
220 }
221
222 static int lapic_nmi_resume(struct sys_device *dev)
223 {
224         /* only CPU0 goes here, other CPUs should be offline */
225         if (nmi_pm_active > 0) {
226                 setup_apic_nmi_watchdog(NULL);
227                 touch_nmi_watchdog();
228         }
229         return 0;
230 }
231
232 static struct sysdev_class nmi_sysclass = {
233         .name           = "lapic_nmi",
234         .resume         = lapic_nmi_resume,
235         .suspend        = lapic_nmi_suspend,
236 };
237
238 static struct sys_device device_lapic_nmi = {
239         .id     = 0,
240         .cls    = &nmi_sysclass,
241 };
242
243 static int __init init_lapic_nmi_sysfs(void)
244 {
245         int error;
246
247         /*
248          * should really be a BUG_ON but b/c this is an
249          * init call, it just doesn't work.  -dcz
250          */
251         if (nmi_watchdog != NMI_LOCAL_APIC)
252                 return 0;
253
254         if (atomic_read(&nmi_active) < 0)
255                 return 0;
256
257         error = sysdev_class_register(&nmi_sysclass);
258         if (!error)
259                 error = sysdev_register(&device_lapic_nmi);
260         return error;
261 }
262
263 /* must come after the local APIC's device_initcall() */
264 late_initcall(init_lapic_nmi_sysfs);
265
266 #endif  /* CONFIG_PM */
267
268 static void __acpi_nmi_enable(void *__unused)
269 {
270         apic_write(APIC_LVT0, APIC_DM_NMI);
271 }
272
273 /*
274  * Enable timer based NMIs on all CPUs:
275  */
276 void acpi_nmi_enable(void)
277 {
278         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
279                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
280 }
281
282 static void __acpi_nmi_disable(void *__unused)
283 {
284         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
285 }
286
287 /*
288  * Disable timer based NMIs on all CPUs:
289  */
290 void acpi_nmi_disable(void)
291 {
292         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
293                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
294 }
295
296 void setup_apic_nmi_watchdog(void *unused)
297 {
298         if (__get_cpu_var(wd_enabled))
299                 return;
300
301         /* cheap hack to support suspend/resume */
302         /* if cpu0 is not active neither should the other cpus */
303         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
304                 return;
305
306         switch (nmi_watchdog) {
307         case NMI_LOCAL_APIC:
308                  /* enable it before to avoid race with handler */
309                 __get_cpu_var(wd_enabled) = 1;
310                 if (lapic_watchdog_init(nmi_hz) < 0) {
311                         __get_cpu_var(wd_enabled) = 0;
312                         return;
313                 }
314                 /* FALL THROUGH */
315         case NMI_IO_APIC:
316                 __get_cpu_var(wd_enabled) = 1;
317                 atomic_inc(&nmi_active);
318         }
319 }
320
321 void stop_apic_nmi_watchdog(void *unused)
322 {
323         /* only support LOCAL and IO APICs for now */
324         if (!nmi_watchdog_active())
325                 return;
326         if (__get_cpu_var(wd_enabled) == 0)
327                 return;
328         if (nmi_watchdog == NMI_LOCAL_APIC)
329                 lapic_watchdog_stop();
330         __get_cpu_var(wd_enabled) = 0;
331         atomic_dec(&nmi_active);
332 }
333
334 /*
335  * the best way to detect whether a CPU has a 'hard lockup' problem
336  * is to check it's local APIC timer IRQ counts. If they are not
337  * changing then that CPU has some problem.
338  *
339  * as these watchdog NMI IRQs are generated on every CPU, we only
340  * have to check the current processor.
341  *
342  * since NMIs don't listen to _any_ locks, we have to be extremely
343  * careful not to rely on unsafe variables. The printk might lock
344  * up though, so we have to break up any console locks first ...
345  * [when there will be more tty-related locks, break them up here too!]
346  */
347
348 static DEFINE_PER_CPU(unsigned, last_irq_sum);
349 static DEFINE_PER_CPU(local_t, alert_counter);
350 static DEFINE_PER_CPU(int, nmi_touch);
351
352 void touch_nmi_watchdog(void)
353 {
354         if (nmi_watchdog_active()) {
355                 unsigned cpu;
356
357                 /*
358                  * Tell other CPUs to reset their alert counters. We cannot
359                  * do it ourselves because the alert count increase is not
360                  * atomic.
361                  */
362                 for_each_present_cpu(cpu) {
363                         if (per_cpu(nmi_touch, cpu) != 1)
364                                 per_cpu(nmi_touch, cpu) = 1;
365                 }
366         }
367
368         /*
369          * Tickle the softlockup detector too:
370          */
371         touch_softlockup_watchdog();
372 }
373 EXPORT_SYMBOL(touch_nmi_watchdog);
374
375 notrace __kprobes int
376 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
377 {
378         /*
379          * Since current_thread_info()-> is always on the stack, and we
380          * always switch the stack NMI-atomically, it's safe to use
381          * smp_processor_id().
382          */
383         unsigned int sum;
384         int touched = 0;
385         int cpu = smp_processor_id();
386         int rc = 0;
387
388         /* check for other users first */
389         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
390                         == NOTIFY_STOP) {
391                 rc = 1;
392                 touched = 1;
393         }
394
395         sum = get_timer_irqs(cpu);
396
397         if (__get_cpu_var(nmi_touch)) {
398                 __get_cpu_var(nmi_touch) = 0;
399                 touched = 1;
400         }
401
402         if (cpu_isset(cpu, backtrace_mask)) {
403                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
404
405                 spin_lock(&lock);
406                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
407                 dump_stack();
408                 spin_unlock(&lock);
409                 cpu_clear(cpu, backtrace_mask);
410         }
411
412         /* Could check oops_in_progress here too, but it's safer not to */
413         if (mce_in_progress())
414                 touched = 1;
415
416         /* if the none of the timers isn't firing, this cpu isn't doing much */
417         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
418                 /*
419                  * Ayiee, looks like this CPU is stuck ...
420                  * wait a few IRQs (5 seconds) before doing the oops ...
421                  */
422                 local_inc(&__get_cpu_var(alert_counter));
423                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
424                         /*
425                          * die_nmi will return ONLY if NOTIFY_STOP happens..
426                          */
427                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
428                                 regs, panic_on_timeout);
429         } else {
430                 __get_cpu_var(last_irq_sum) = sum;
431                 local_set(&__get_cpu_var(alert_counter), 0);
432         }
433
434         /* see if the nmi watchdog went off */
435         if (!__get_cpu_var(wd_enabled))
436                 return rc;
437         switch (nmi_watchdog) {
438         case NMI_LOCAL_APIC:
439                 rc |= lapic_wd_event(nmi_hz);
440                 break;
441         case NMI_IO_APIC:
442                 /*
443                  * don't know how to accurately check for this.
444                  * just assume it was a watchdog timer interrupt
445                  * This matches the old behaviour.
446                  */
447                 rc = 1;
448                 break;
449         }
450         return rc;
451 }
452
453 #ifdef CONFIG_SYSCTL
454
455 static int __init setup_unknown_nmi_panic(char *str)
456 {
457         unknown_nmi_panic = 1;
458         return 1;
459 }
460 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
461
462 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
463 {
464         unsigned char reason = get_nmi_reason();
465         char buf[64];
466
467         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
468         die_nmi(buf, regs, 1); /* Always panic here */
469         return 0;
470 }
471
472 /*
473  * proc handler for /proc/sys/kernel/nmi
474  */
475 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
476                         void __user *buffer, size_t *length, loff_t *ppos)
477 {
478         int old_state;
479
480         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
481         old_state = nmi_watchdog_enabled;
482         proc_dointvec(table, write, file, buffer, length, ppos);
483         if (!!old_state == !!nmi_watchdog_enabled)
484                 return 0;
485
486         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
487                 printk(KERN_WARNING
488                         "NMI watchdog is permanently disabled\n");
489                 return -EIO;
490         }
491
492         if (nmi_watchdog == NMI_LOCAL_APIC) {
493                 if (nmi_watchdog_enabled)
494                         enable_lapic_nmi_watchdog();
495                 else
496                         disable_lapic_nmi_watchdog();
497         } else {
498                 printk(KERN_WARNING
499                         "NMI watchdog doesn't know what hardware to touch\n");
500                 return -EIO;
501         }
502         return 0;
503 }
504
505 #endif /* CONFIG_SYSCTL */
506
507 int do_nmi_callback(struct pt_regs *regs, int cpu)
508 {
509 #ifdef CONFIG_SYSCTL
510         if (unknown_nmi_panic)
511                 return unknown_nmi_panic_callback(regs, cpu);
512 #endif
513         return 0;
514 }
515
516 void __trigger_all_cpu_backtrace(void)
517 {
518         int i;
519
520         backtrace_mask = cpu_online_map;
521         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
522         for (i = 0; i < 10 * 1000; i++) {
523                 if (cpus_empty(backtrace_mask))
524                         break;
525                 mdelay(1);
526         }
527 }