cb66a22d98ad72a3ad9eae5fc5b606beb15bcb94
[linux-2.6.git] / arch / x86 / kernel / apic / 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/proto.h>
33 #include <asm/timer.h>
34
35 #include <asm/mce.h>
36
37 #include <asm/mach_traps.h>
38
39 int unknown_nmi_panic;
40 int nmi_watchdog_enabled;
41
42 static cpumask_t backtrace_mask __read_mostly;
43
44 /* nmi_active:
45  * >0: the lapic NMI watchdog is active, but can be disabled
46  * <0: the lapic NMI watchdog has not been set up, and cannot
47  *     be enabled
48  *  0: the lapic NMI watchdog is disabled, but can be enabled
49  */
50 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
51 EXPORT_SYMBOL(nmi_active);
52
53 unsigned int nmi_watchdog = NMI_NONE;
54 EXPORT_SYMBOL(nmi_watchdog);
55
56 static int panic_on_timeout;
57
58 static unsigned int nmi_hz = HZ;
59 static DEFINE_PER_CPU(short, wd_enabled);
60 static int endflag __initdata;
61
62 static inline unsigned int get_nmi_count(int cpu)
63 {
64         return per_cpu(irq_stat, cpu).__nmi_count;
65 }
66
67 static inline int mce_in_progress(void)
68 {
69 #if defined(CONFIG_X86_MCE)
70         return atomic_read(&mce_entry) > 0;
71 #endif
72         return 0;
73 }
74
75 /*
76  * Take the local apic timer and PIT/HPET into account. We don't
77  * know which one is active, when we have highres/dyntick on
78  */
79 static inline unsigned int get_timer_irqs(int cpu)
80 {
81         return per_cpu(irq_stat, cpu).apic_timer_irqs +
82                 per_cpu(irq_stat, cpu).irq0_irqs;
83 }
84
85 #ifdef CONFIG_SMP
86 /*
87  * The performance counters used by NMI_LOCAL_APIC don't trigger when
88  * the CPU is idle. To make sure the NMI watchdog really ticks on all
89  * CPUs during the test make them busy.
90  */
91 static __init void nmi_cpu_busy(void *data)
92 {
93         local_irq_enable_in_hardirq();
94         /*
95          * Intentionally don't use cpu_relax here. This is
96          * to make sure that the performance counter really ticks,
97          * even if there is a simulator or similar that catches the
98          * pause instruction. On a real HT machine this is fine because
99          * all other CPUs are busy with "useless" delay loops and don't
100          * care if they get somewhat less cycles.
101          */
102         while (endflag == 0)
103                 mb();
104 }
105 #endif
106
107 static void report_broken_nmi(int cpu, unsigned int *prev_nmi_count)
108 {
109         printk(KERN_CONT "\n");
110
111         printk(KERN_WARNING
112                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
113                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
114
115         printk(KERN_WARNING
116                 "Please report this to bugzilla.kernel.org,\n");
117         printk(KERN_WARNING
118                 "and attach the output of the 'dmesg' command.\n");
119
120         per_cpu(wd_enabled, cpu) = 0;
121         atomic_dec(&nmi_active);
122 }
123
124 static void __acpi_nmi_disable(void *__unused)
125 {
126         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
127 }
128
129 int __init check_nmi_watchdog(void)
130 {
131         unsigned int *prev_nmi_count;
132         int cpu;
133
134         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
135                 return 0;
136
137         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
138         if (!prev_nmi_count)
139                 goto error;
140
141         printk(KERN_INFO "Testing NMI watchdog ... ");
142
143 #ifdef CONFIG_SMP
144         if (nmi_watchdog == NMI_LOCAL_APIC)
145                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
146 #endif
147
148         for_each_possible_cpu(cpu)
149                 prev_nmi_count[cpu] = get_nmi_count(cpu);
150         local_irq_enable();
151         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
152
153         for_each_online_cpu(cpu) {
154                 if (!per_cpu(wd_enabled, cpu))
155                         continue;
156                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
157                         report_broken_nmi(cpu, prev_nmi_count);
158         }
159         endflag = 1;
160         if (!atomic_read(&nmi_active)) {
161                 kfree(prev_nmi_count);
162                 atomic_set(&nmi_active, -1);
163                 goto error;
164         }
165         printk("OK.\n");
166
167         /*
168          * now that we know it works we can reduce NMI frequency to
169          * something more reasonable; makes a difference in some configs
170          */
171         if (nmi_watchdog == NMI_LOCAL_APIC)
172                 nmi_hz = lapic_adjust_nmi_hz(1);
173
174         kfree(prev_nmi_count);
175         return 0;
176 error:
177         if (nmi_watchdog == NMI_IO_APIC) {
178                 if (!timer_through_8259)
179                         disable_8259A_irq(0);
180                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
181         }
182
183 #ifdef CONFIG_X86_32
184         timer_ack = 0;
185 #endif
186         return -1;
187 }
188
189 static int __init setup_nmi_watchdog(char *str)
190 {
191         unsigned int nmi;
192
193         if (!strncmp(str, "panic", 5)) {
194                 panic_on_timeout = 1;
195                 str = strchr(str, ',');
196                 if (!str)
197                         return 1;
198                 ++str;
199         }
200
201         if (!strncmp(str, "lapic", 5))
202                 nmi_watchdog = NMI_LOCAL_APIC;
203         else if (!strncmp(str, "ioapic", 6))
204                 nmi_watchdog = NMI_IO_APIC;
205         else {
206                 get_option(&str, &nmi);
207                 if (nmi >= NMI_INVALID)
208                         return 0;
209                 nmi_watchdog = nmi;
210         }
211
212         return 1;
213 }
214 __setup("nmi_watchdog=", setup_nmi_watchdog);
215
216 /*
217  * Suspend/resume support
218  */
219 #ifdef CONFIG_PM
220
221 static int nmi_pm_active; /* nmi_active before suspend */
222
223 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
224 {
225         /* only CPU0 goes here, other CPUs should be offline */
226         nmi_pm_active = atomic_read(&nmi_active);
227         stop_apic_nmi_watchdog(NULL);
228         BUG_ON(atomic_read(&nmi_active) != 0);
229         return 0;
230 }
231
232 static int lapic_nmi_resume(struct sys_device *dev)
233 {
234         /* only CPU0 goes here, other CPUs should be offline */
235         if (nmi_pm_active > 0) {
236                 setup_apic_nmi_watchdog(NULL);
237                 touch_nmi_watchdog();
238         }
239         return 0;
240 }
241
242 static struct sysdev_class nmi_sysclass = {
243         .name           = "lapic_nmi",
244         .resume         = lapic_nmi_resume,
245         .suspend        = lapic_nmi_suspend,
246 };
247
248 static struct sys_device device_lapic_nmi = {
249         .id     = 0,
250         .cls    = &nmi_sysclass,
251 };
252
253 static int __init init_lapic_nmi_sysfs(void)
254 {
255         int error;
256
257         /*
258          * should really be a BUG_ON but b/c this is an
259          * init call, it just doesn't work.  -dcz
260          */
261         if (nmi_watchdog != NMI_LOCAL_APIC)
262                 return 0;
263
264         if (atomic_read(&nmi_active) < 0)
265                 return 0;
266
267         error = sysdev_class_register(&nmi_sysclass);
268         if (!error)
269                 error = sysdev_register(&device_lapic_nmi);
270         return error;
271 }
272
273 /* must come after the local APIC's device_initcall() */
274 late_initcall(init_lapic_nmi_sysfs);
275
276 #endif  /* CONFIG_PM */
277
278 static void __acpi_nmi_enable(void *__unused)
279 {
280         apic_write(APIC_LVT0, APIC_DM_NMI);
281 }
282
283 /*
284  * Enable timer based NMIs on all CPUs:
285  */
286 void acpi_nmi_enable(void)
287 {
288         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
289                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
290 }
291
292 /*
293  * Disable timer based NMIs on all CPUs:
294  */
295 void acpi_nmi_disable(void)
296 {
297         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
298                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
299 }
300
301 /*
302  * This function is called as soon the LAPIC NMI watchdog driver has everything
303  * in place and it's ready to check if the NMIs belong to the NMI watchdog
304  */
305 void cpu_nmi_set_wd_enabled(void)
306 {
307         __get_cpu_var(wd_enabled) = 1;
308 }
309
310 void setup_apic_nmi_watchdog(void *unused)
311 {
312         if (__get_cpu_var(wd_enabled))
313                 return;
314
315         /* cheap hack to support suspend/resume */
316         /* if cpu0 is not active neither should the other cpus */
317         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
318                 return;
319
320         switch (nmi_watchdog) {
321         case NMI_LOCAL_APIC:
322                 if (lapic_watchdog_init(nmi_hz) < 0) {
323                         __get_cpu_var(wd_enabled) = 0;
324                         return;
325                 }
326                 /* FALL THROUGH */
327         case NMI_IO_APIC:
328                 __get_cpu_var(wd_enabled) = 1;
329                 atomic_inc(&nmi_active);
330         }
331 }
332
333 void stop_apic_nmi_watchdog(void *unused)
334 {
335         /* only support LOCAL and IO APICs for now */
336         if (!nmi_watchdog_active())
337                 return;
338         if (__get_cpu_var(wd_enabled) == 0)
339                 return;
340         if (nmi_watchdog == NMI_LOCAL_APIC)
341                 lapic_watchdog_stop();
342         else
343                 __acpi_nmi_disable(NULL);
344         __get_cpu_var(wd_enabled) = 0;
345         atomic_dec(&nmi_active);
346 }
347
348 /*
349  * the best way to detect whether a CPU has a 'hard lockup' problem
350  * is to check it's local APIC timer IRQ counts. If they are not
351  * changing then that CPU has some problem.
352  *
353  * as these watchdog NMI IRQs are generated on every CPU, we only
354  * have to check the current processor.
355  *
356  * since NMIs don't listen to _any_ locks, we have to be extremely
357  * careful not to rely on unsafe variables. The printk might lock
358  * up though, so we have to break up any console locks first ...
359  * [when there will be more tty-related locks, break them up here too!]
360  */
361
362 static DEFINE_PER_CPU(unsigned, last_irq_sum);
363 static DEFINE_PER_CPU(local_t, alert_counter);
364 static DEFINE_PER_CPU(int, nmi_touch);
365
366 void touch_nmi_watchdog(void)
367 {
368         if (nmi_watchdog_active()) {
369                 unsigned cpu;
370
371                 /*
372                  * Tell other CPUs to reset their alert counters. We cannot
373                  * do it ourselves because the alert count increase is not
374                  * atomic.
375                  */
376                 for_each_present_cpu(cpu) {
377                         if (per_cpu(nmi_touch, cpu) != 1)
378                                 per_cpu(nmi_touch, cpu) = 1;
379                 }
380         }
381
382         /*
383          * Tickle the softlockup detector too:
384          */
385         touch_softlockup_watchdog();
386 }
387 EXPORT_SYMBOL(touch_nmi_watchdog);
388
389 notrace __kprobes int
390 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
391 {
392         /*
393          * Since current_thread_info()-> is always on the stack, and we
394          * always switch the stack NMI-atomically, it's safe to use
395          * smp_processor_id().
396          */
397         unsigned int sum;
398         int touched = 0;
399         int cpu = smp_processor_id();
400         int rc = 0;
401
402         /* check for other users first */
403         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
404                         == NOTIFY_STOP) {
405                 rc = 1;
406                 touched = 1;
407         }
408
409         sum = get_timer_irqs(cpu);
410
411         if (__get_cpu_var(nmi_touch)) {
412                 __get_cpu_var(nmi_touch) = 0;
413                 touched = 1;
414         }
415
416         /* We can be called before check_nmi_watchdog, hence NULL check. */
417         if (cpumask_test_cpu(cpu, &backtrace_mask)) {
418                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
419
420                 spin_lock(&lock);
421                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
422                 show_regs(regs);
423                 dump_stack();
424                 spin_unlock(&lock);
425                 cpumask_clear_cpu(cpu, &backtrace_mask);
426
427                 rc = 1;
428         }
429
430         /* Could check oops_in_progress here too, but it's safer not to */
431         if (mce_in_progress())
432                 touched = 1;
433
434         /* if the none of the timers isn't firing, this cpu isn't doing much */
435         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
436                 /*
437                  * Ayiee, looks like this CPU is stuck ...
438                  * wait a few IRQs (5 seconds) before doing the oops ...
439                  */
440                 local_inc(&__get_cpu_var(alert_counter));
441                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
442                         /*
443                          * die_nmi will return ONLY if NOTIFY_STOP happens..
444                          */
445                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
446                                 regs, panic_on_timeout);
447         } else {
448                 __get_cpu_var(last_irq_sum) = sum;
449                 local_set(&__get_cpu_var(alert_counter), 0);
450         }
451
452         /* see if the nmi watchdog went off */
453         if (!__get_cpu_var(wd_enabled))
454                 return rc;
455         switch (nmi_watchdog) {
456         case NMI_LOCAL_APIC:
457                 rc |= lapic_wd_event(nmi_hz);
458                 break;
459         case NMI_IO_APIC:
460                 /*
461                  * don't know how to accurately check for this.
462                  * just assume it was a watchdog timer interrupt
463                  * This matches the old behaviour.
464                  */
465                 rc = 1;
466                 break;
467         }
468         return rc;
469 }
470
471 #ifdef CONFIG_SYSCTL
472
473 static void enable_ioapic_nmi_watchdog_single(void *unused)
474 {
475         __get_cpu_var(wd_enabled) = 1;
476         atomic_inc(&nmi_active);
477         __acpi_nmi_enable(NULL);
478 }
479
480 static void enable_ioapic_nmi_watchdog(void)
481 {
482         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
483         touch_nmi_watchdog();
484 }
485
486 static void disable_ioapic_nmi_watchdog(void)
487 {
488         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
489 }
490
491 static int __init setup_unknown_nmi_panic(char *str)
492 {
493         unknown_nmi_panic = 1;
494         return 1;
495 }
496 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
497
498 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
499 {
500         unsigned char reason = get_nmi_reason();
501         char buf[64];
502
503         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
504         die_nmi(buf, regs, 1); /* Always panic here */
505         return 0;
506 }
507
508 /*
509  * proc handler for /proc/sys/kernel/nmi
510  */
511 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
512                         void __user *buffer, size_t *length, loff_t *ppos)
513 {
514         int old_state;
515
516         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
517         old_state = nmi_watchdog_enabled;
518         proc_dointvec(table, write, file, buffer, length, ppos);
519         if (!!old_state == !!nmi_watchdog_enabled)
520                 return 0;
521
522         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
523                 printk(KERN_WARNING
524                         "NMI watchdog is permanently disabled\n");
525                 return -EIO;
526         }
527
528         if (nmi_watchdog == NMI_LOCAL_APIC) {
529                 if (nmi_watchdog_enabled)
530                         enable_lapic_nmi_watchdog();
531                 else
532                         disable_lapic_nmi_watchdog();
533         } else if (nmi_watchdog == NMI_IO_APIC) {
534                 if (nmi_watchdog_enabled)
535                         enable_ioapic_nmi_watchdog();
536                 else
537                         disable_ioapic_nmi_watchdog();
538         } else {
539                 printk(KERN_WARNING
540                         "NMI watchdog doesn't know what hardware to touch\n");
541                 return -EIO;
542         }
543         return 0;
544 }
545
546 #endif /* CONFIG_SYSCTL */
547
548 int do_nmi_callback(struct pt_regs *regs, int cpu)
549 {
550 #ifdef CONFIG_SYSCTL
551         if (unknown_nmi_panic)
552                 return unknown_nmi_panic_callback(regs, cpu);
553 #endif
554         return 0;
555 }
556
557 void arch_trigger_all_cpu_backtrace(void)
558 {
559         int i;
560
561         cpumask_copy(&backtrace_mask, cpu_online_mask);
562
563         printk(KERN_INFO "sending NMI to all CPUs:\n");
564         apic->send_IPI_all(NMI_VECTOR);
565
566         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
567         for (i = 0; i < 10 * 1000; i++) {
568                 if (cpumask_empty(&backtrace_mask))
569                         break;
570                 mdelay(1);
571         }
572 }