Merge commit 'v2.6.28-rc9' into x86/apic
[sfrench/cifs-2.6.git] / 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 static void report_broken_nmi(int cpu, int *prev_nmi_count)
118 {
119         printk(KERN_CONT "\n");
120
121         printk(KERN_WARNING
122                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
123                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
124
125         printk(KERN_WARNING
126                 "Please report this to bugzilla.kernel.org,\n");
127         printk(KERN_WARNING
128                 "and attach the output of the 'dmesg' command.\n");
129
130         per_cpu(wd_enabled, cpu) = 0;
131         atomic_dec(&nmi_active);
132 }
133
134 static void __acpi_nmi_disable(void *__unused)
135 {
136         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
137 }
138
139 int __init check_nmi_watchdog(void)
140 {
141         unsigned int *prev_nmi_count;
142         int cpu;
143
144         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
145                 return 0;
146
147         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
148         if (!prev_nmi_count)
149                 goto error;
150
151         printk(KERN_INFO "Testing NMI watchdog ... ");
152
153 #ifdef CONFIG_SMP
154         if (nmi_watchdog == NMI_LOCAL_APIC)
155                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
156 #endif
157
158         for_each_possible_cpu(cpu)
159                 prev_nmi_count[cpu] = get_nmi_count(cpu);
160         local_irq_enable();
161         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
162
163         for_each_online_cpu(cpu) {
164                 if (!per_cpu(wd_enabled, cpu))
165                         continue;
166                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
167                         report_broken_nmi(cpu, prev_nmi_count);
168         }
169         endflag = 1;
170         if (!atomic_read(&nmi_active)) {
171                 kfree(prev_nmi_count);
172                 atomic_set(&nmi_active, -1);
173                 goto error;
174         }
175         printk("OK.\n");
176
177         /*
178          * now that we know it works we can reduce NMI frequency to
179          * something more reasonable; makes a difference in some configs
180          */
181         if (nmi_watchdog == NMI_LOCAL_APIC)
182                 nmi_hz = lapic_adjust_nmi_hz(1);
183
184         kfree(prev_nmi_count);
185         return 0;
186 error:
187         if (nmi_watchdog == NMI_IO_APIC) {
188                 if (!timer_through_8259)
189                         disable_8259A_irq(0);
190                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
191         }
192
193 #ifdef CONFIG_X86_32
194         timer_ack = 0;
195 #endif
196         return -1;
197 }
198
199 static int __init setup_nmi_watchdog(char *str)
200 {
201         unsigned int nmi;
202
203         if (!strncmp(str, "panic", 5)) {
204                 panic_on_timeout = 1;
205                 str = strchr(str, ',');
206                 if (!str)
207                         return 1;
208                 ++str;
209         }
210
211         get_option(&str, &nmi);
212
213         if (nmi >= NMI_INVALID)
214                 return 0;
215
216         nmi_watchdog = nmi;
217         return 1;
218 }
219 __setup("nmi_watchdog=", setup_nmi_watchdog);
220
221 /*
222  * Suspend/resume support
223  */
224 #ifdef CONFIG_PM
225
226 static int nmi_pm_active; /* nmi_active before suspend */
227
228 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
229 {
230         /* only CPU0 goes here, other CPUs should be offline */
231         nmi_pm_active = atomic_read(&nmi_active);
232         stop_apic_nmi_watchdog(NULL);
233         BUG_ON(atomic_read(&nmi_active) != 0);
234         return 0;
235 }
236
237 static int lapic_nmi_resume(struct sys_device *dev)
238 {
239         /* only CPU0 goes here, other CPUs should be offline */
240         if (nmi_pm_active > 0) {
241                 setup_apic_nmi_watchdog(NULL);
242                 touch_nmi_watchdog();
243         }
244         return 0;
245 }
246
247 static struct sysdev_class nmi_sysclass = {
248         .name           = "lapic_nmi",
249         .resume         = lapic_nmi_resume,
250         .suspend        = lapic_nmi_suspend,
251 };
252
253 static struct sys_device device_lapic_nmi = {
254         .id     = 0,
255         .cls    = &nmi_sysclass,
256 };
257
258 static int __init init_lapic_nmi_sysfs(void)
259 {
260         int error;
261
262         /*
263          * should really be a BUG_ON but b/c this is an
264          * init call, it just doesn't work.  -dcz
265          */
266         if (nmi_watchdog != NMI_LOCAL_APIC)
267                 return 0;
268
269         if (atomic_read(&nmi_active) < 0)
270                 return 0;
271
272         error = sysdev_class_register(&nmi_sysclass);
273         if (!error)
274                 error = sysdev_register(&device_lapic_nmi);
275         return error;
276 }
277
278 /* must come after the local APIC's device_initcall() */
279 late_initcall(init_lapic_nmi_sysfs);
280
281 #endif  /* CONFIG_PM */
282
283 static void __acpi_nmi_enable(void *__unused)
284 {
285         apic_write(APIC_LVT0, APIC_DM_NMI);
286 }
287
288 /*
289  * Enable timer based NMIs on all CPUs:
290  */
291 void acpi_nmi_enable(void)
292 {
293         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
294                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
295 }
296
297 /*
298  * Disable timer based NMIs on all CPUs:
299  */
300 void acpi_nmi_disable(void)
301 {
302         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
303                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
304 }
305
306 /*
307  * This function is called as soon the LAPIC NMI watchdog driver has everything
308  * in place and it's ready to check if the NMIs belong to the NMI watchdog
309  */
310 void cpu_nmi_set_wd_enabled(void)
311 {
312         __get_cpu_var(wd_enabled) = 1;
313 }
314
315 void setup_apic_nmi_watchdog(void *unused)
316 {
317         if (__get_cpu_var(wd_enabled))
318                 return;
319
320         /* cheap hack to support suspend/resume */
321         /* if cpu0 is not active neither should the other cpus */
322         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
323                 return;
324
325         switch (nmi_watchdog) {
326         case NMI_LOCAL_APIC:
327                 if (lapic_watchdog_init(nmi_hz) < 0) {
328                         __get_cpu_var(wd_enabled) = 0;
329                         return;
330                 }
331                 /* FALL THROUGH */
332         case NMI_IO_APIC:
333                 __get_cpu_var(wd_enabled) = 1;
334                 atomic_inc(&nmi_active);
335         }
336 }
337
338 void stop_apic_nmi_watchdog(void *unused)
339 {
340         /* only support LOCAL and IO APICs for now */
341         if (!nmi_watchdog_active())
342                 return;
343         if (__get_cpu_var(wd_enabled) == 0)
344                 return;
345         if (nmi_watchdog == NMI_LOCAL_APIC)
346                 lapic_watchdog_stop();
347         else
348                 __acpi_nmi_disable(NULL);
349         __get_cpu_var(wd_enabled) = 0;
350         atomic_dec(&nmi_active);
351 }
352
353 /*
354  * the best way to detect whether a CPU has a 'hard lockup' problem
355  * is to check it's local APIC timer IRQ counts. If they are not
356  * changing then that CPU has some problem.
357  *
358  * as these watchdog NMI IRQs are generated on every CPU, we only
359  * have to check the current processor.
360  *
361  * since NMIs don't listen to _any_ locks, we have to be extremely
362  * careful not to rely on unsafe variables. The printk might lock
363  * up though, so we have to break up any console locks first ...
364  * [when there will be more tty-related locks, break them up here too!]
365  */
366
367 static DEFINE_PER_CPU(unsigned, last_irq_sum);
368 static DEFINE_PER_CPU(local_t, alert_counter);
369 static DEFINE_PER_CPU(int, nmi_touch);
370
371 void touch_nmi_watchdog(void)
372 {
373         if (nmi_watchdog_active()) {
374                 unsigned cpu;
375
376                 /*
377                  * Tell other CPUs to reset their alert counters. We cannot
378                  * do it ourselves because the alert count increase is not
379                  * atomic.
380                  */
381                 for_each_present_cpu(cpu) {
382                         if (per_cpu(nmi_touch, cpu) != 1)
383                                 per_cpu(nmi_touch, cpu) = 1;
384                 }
385         }
386
387         /*
388          * Tickle the softlockup detector too:
389          */
390         touch_softlockup_watchdog();
391 }
392 EXPORT_SYMBOL(touch_nmi_watchdog);
393
394 notrace __kprobes int
395 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
396 {
397         /*
398          * Since current_thread_info()-> is always on the stack, and we
399          * always switch the stack NMI-atomically, it's safe to use
400          * smp_processor_id().
401          */
402         unsigned int sum;
403         int touched = 0;
404         int cpu = smp_processor_id();
405         int rc = 0;
406
407         /* check for other users first */
408         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
409                         == NOTIFY_STOP) {
410                 rc = 1;
411                 touched = 1;
412         }
413
414         sum = get_timer_irqs(cpu);
415
416         if (__get_cpu_var(nmi_touch)) {
417                 __get_cpu_var(nmi_touch) = 0;
418                 touched = 1;
419         }
420
421         if (cpu_isset(cpu, backtrace_mask)) {
422                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
423
424                 spin_lock(&lock);
425                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
426                 dump_stack();
427                 spin_unlock(&lock);
428                 cpu_clear(cpu, backtrace_mask);
429         }
430
431         /* Could check oops_in_progress here too, but it's safer not to */
432         if (mce_in_progress())
433                 touched = 1;
434
435         /* if the none of the timers isn't firing, this cpu isn't doing much */
436         if (!touched && __get_cpu_var(last_irq_sum) == sum) {
437                 /*
438                  * Ayiee, looks like this CPU is stuck ...
439                  * wait a few IRQs (5 seconds) before doing the oops ...
440                  */
441                 local_inc(&__get_cpu_var(alert_counter));
442                 if (local_read(&__get_cpu_var(alert_counter)) == 5 * nmi_hz)
443                         /*
444                          * die_nmi will return ONLY if NOTIFY_STOP happens..
445                          */
446                         die_nmi("BUG: NMI Watchdog detected LOCKUP",
447                                 regs, panic_on_timeout);
448         } else {
449                 __get_cpu_var(last_irq_sum) = sum;
450                 local_set(&__get_cpu_var(alert_counter), 0);
451         }
452
453         /* see if the nmi watchdog went off */
454         if (!__get_cpu_var(wd_enabled))
455                 return rc;
456         switch (nmi_watchdog) {
457         case NMI_LOCAL_APIC:
458                 rc |= lapic_wd_event(nmi_hz);
459                 break;
460         case NMI_IO_APIC:
461                 /*
462                  * don't know how to accurately check for this.
463                  * just assume it was a watchdog timer interrupt
464                  * This matches the old behaviour.
465                  */
466                 rc = 1;
467                 break;
468         }
469         return rc;
470 }
471
472 #ifdef CONFIG_SYSCTL
473
474 static void enable_ioapic_nmi_watchdog_single(void *unused)
475 {
476         __get_cpu_var(wd_enabled) = 1;
477         atomic_inc(&nmi_active);
478         __acpi_nmi_enable(NULL);
479 }
480
481 static void enable_ioapic_nmi_watchdog(void)
482 {
483         on_each_cpu(enable_ioapic_nmi_watchdog_single, NULL, 1);
484         touch_nmi_watchdog();
485 }
486
487 static void disable_ioapic_nmi_watchdog(void)
488 {
489         on_each_cpu(stop_apic_nmi_watchdog, NULL, 1);
490 }
491
492 static int __init setup_unknown_nmi_panic(char *str)
493 {
494         unknown_nmi_panic = 1;
495         return 1;
496 }
497 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
498
499 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
500 {
501         unsigned char reason = get_nmi_reason();
502         char buf[64];
503
504         sprintf(buf, "NMI received for unknown reason %02x\n", reason);
505         die_nmi(buf, regs, 1); /* Always panic here */
506         return 0;
507 }
508
509 /*
510  * proc handler for /proc/sys/kernel/nmi
511  */
512 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
513                         void __user *buffer, size_t *length, loff_t *ppos)
514 {
515         int old_state;
516
517         nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
518         old_state = nmi_watchdog_enabled;
519         proc_dointvec(table, write, file, buffer, length, ppos);
520         if (!!old_state == !!nmi_watchdog_enabled)
521                 return 0;
522
523         if (atomic_read(&nmi_active) < 0 || !nmi_watchdog_active()) {
524                 printk(KERN_WARNING
525                         "NMI watchdog is permanently disabled\n");
526                 return -EIO;
527         }
528
529         if (nmi_watchdog == NMI_LOCAL_APIC) {
530                 if (nmi_watchdog_enabled)
531                         enable_lapic_nmi_watchdog();
532                 else
533                         disable_lapic_nmi_watchdog();
534         } else if (nmi_watchdog == NMI_IO_APIC) {
535                 if (nmi_watchdog_enabled)
536                         enable_ioapic_nmi_watchdog();
537                 else
538                         disable_ioapic_nmi_watchdog();
539         } else {
540                 printk(KERN_WARNING
541                         "NMI watchdog doesn't know what hardware to touch\n");
542                 return -EIO;
543         }
544         return 0;
545 }
546
547 #endif /* CONFIG_SYSCTL */
548
549 int do_nmi_callback(struct pt_regs *regs, int cpu)
550 {
551 #ifdef CONFIG_SYSCTL
552         if (unknown_nmi_panic)
553                 return unknown_nmi_panic_callback(regs, cpu);
554 #endif
555         return 0;
556 }
557
558 void __trigger_all_cpu_backtrace(void)
559 {
560         int i;
561
562         backtrace_mask = cpu_online_map;
563         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
564         for (i = 0; i < 10 * 1000; i++) {
565                 if (cpus_empty(backtrace_mask))
566                         break;
567                 mdelay(1);
568         }
569 }