Merge branch 'master' into percpu
[sfrench/cifs-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 /* For reliability, we're prepared to waste bits here. */
43 static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
44
45 /* nmi_active:
46  * >0: the lapic NMI watchdog is active, but can be disabled
47  * <0: the lapic NMI watchdog has not been set up, and cannot
48  *     be enabled
49  *  0: the lapic NMI watchdog is disabled, but can be enabled
50  */
51 atomic_t nmi_active = ATOMIC_INIT(0);           /* oprofile uses this */
52 EXPORT_SYMBOL(nmi_active);
53
54 unsigned int nmi_watchdog = NMI_NONE;
55 EXPORT_SYMBOL(nmi_watchdog);
56
57 static int panic_on_timeout;
58
59 static unsigned int nmi_hz = HZ;
60 static DEFINE_PER_CPU(short, wd_enabled);
61 static int endflag __initdata;
62
63 static inline unsigned int get_nmi_count(int cpu)
64 {
65         return per_cpu(irq_stat, cpu).__nmi_count;
66 }
67
68 static inline int mce_in_progress(void)
69 {
70 #if defined(CONFIG_X86_MCE)
71         return atomic_read(&mce_entry) > 0;
72 #endif
73         return 0;
74 }
75
76 /*
77  * Take the local apic timer and PIT/HPET into account. We don't
78  * know which one is active, when we have highres/dyntick on
79  */
80 static inline unsigned int get_timer_irqs(int cpu)
81 {
82         return per_cpu(irq_stat, cpu).apic_timer_irqs +
83                 per_cpu(irq_stat, cpu).irq0_irqs;
84 }
85
86 #ifdef CONFIG_SMP
87 /*
88  * The performance counters used by NMI_LOCAL_APIC don't trigger when
89  * the CPU is idle. To make sure the NMI watchdog really ticks on all
90  * CPUs during the test make them busy.
91  */
92 static __init void nmi_cpu_busy(void *data)
93 {
94         local_irq_enable_in_hardirq();
95         /*
96          * Intentionally don't use cpu_relax here. This is
97          * to make sure that the performance counter really ticks,
98          * even if there is a simulator or similar that catches the
99          * pause instruction. On a real HT machine this is fine because
100          * all other CPUs are busy with "useless" delay loops and don't
101          * care if they get somewhat less cycles.
102          */
103         while (endflag == 0)
104                 mb();
105 }
106 #endif
107
108 static void report_broken_nmi(int cpu, unsigned int *prev_nmi_count)
109 {
110         printk(KERN_CONT "\n");
111
112         printk(KERN_WARNING
113                 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
114                         cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
115
116         printk(KERN_WARNING
117                 "Please report this to bugzilla.kernel.org,\n");
118         printk(KERN_WARNING
119                 "and attach the output of the 'dmesg' command.\n");
120
121         per_cpu(wd_enabled, cpu) = 0;
122         atomic_dec(&nmi_active);
123 }
124
125 static void __acpi_nmi_disable(void *__unused)
126 {
127         apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
128 }
129
130 int __init check_nmi_watchdog(void)
131 {
132         unsigned int *prev_nmi_count;
133         int cpu;
134
135         if (!nmi_watchdog_active() || !atomic_read(&nmi_active))
136                 return 0;
137
138         prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
139         if (!prev_nmi_count)
140                 goto error;
141
142         printk(KERN_INFO "Testing NMI watchdog ... ");
143
144 #ifdef CONFIG_SMP
145         if (nmi_watchdog == NMI_LOCAL_APIC)
146                 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
147 #endif
148
149         for_each_possible_cpu(cpu)
150                 prev_nmi_count[cpu] = get_nmi_count(cpu);
151         local_irq_enable();
152         mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
153
154         for_each_online_cpu(cpu) {
155                 if (!per_cpu(wd_enabled, cpu))
156                         continue;
157                 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
158                         report_broken_nmi(cpu, prev_nmi_count);
159         }
160         endflag = 1;
161         if (!atomic_read(&nmi_active)) {
162                 kfree(prev_nmi_count);
163                 atomic_set(&nmi_active, -1);
164                 goto error;
165         }
166         printk("OK.\n");
167
168         /*
169          * now that we know it works we can reduce NMI frequency to
170          * something more reasonable; makes a difference in some configs
171          */
172         if (nmi_watchdog == NMI_LOCAL_APIC)
173                 nmi_hz = lapic_adjust_nmi_hz(1);
174
175         kfree(prev_nmi_count);
176         return 0;
177 error:
178         if (nmi_watchdog == NMI_IO_APIC) {
179                 if (!timer_through_8259)
180                         disable_8259A_irq(0);
181                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
182         }
183
184 #ifdef CONFIG_X86_32
185         timer_ack = 0;
186 #endif
187         return -1;
188 }
189
190 static int __init setup_nmi_watchdog(char *str)
191 {
192         unsigned int nmi;
193
194         if (!strncmp(str, "panic", 5)) {
195                 panic_on_timeout = 1;
196                 str = strchr(str, ',');
197                 if (!str)
198                         return 1;
199                 ++str;
200         }
201
202         if (!strncmp(str, "lapic", 5))
203                 nmi_watchdog = NMI_LOCAL_APIC;
204         else if (!strncmp(str, "ioapic", 6))
205                 nmi_watchdog = NMI_IO_APIC;
206         else {
207                 get_option(&str, &nmi);
208                 if (nmi >= NMI_INVALID)
209                         return 0;
210                 nmi_watchdog = nmi;
211         }
212
213         return 1;
214 }
215 __setup("nmi_watchdog=", setup_nmi_watchdog);
216
217 /*
218  * Suspend/resume support
219  */
220 #ifdef CONFIG_PM
221
222 static int nmi_pm_active; /* nmi_active before suspend */
223
224 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
225 {
226         /* only CPU0 goes here, other CPUs should be offline */
227         nmi_pm_active = atomic_read(&nmi_active);
228         stop_apic_nmi_watchdog(NULL);
229         BUG_ON(atomic_read(&nmi_active) != 0);
230         return 0;
231 }
232
233 static int lapic_nmi_resume(struct sys_device *dev)
234 {
235         /* only CPU0 goes here, other CPUs should be offline */
236         if (nmi_pm_active > 0) {
237                 setup_apic_nmi_watchdog(NULL);
238                 touch_nmi_watchdog();
239         }
240         return 0;
241 }
242
243 static struct sysdev_class nmi_sysclass = {
244         .name           = "lapic_nmi",
245         .resume         = lapic_nmi_resume,
246         .suspend        = lapic_nmi_suspend,
247 };
248
249 static struct sys_device device_lapic_nmi = {
250         .id     = 0,
251         .cls    = &nmi_sysclass,
252 };
253
254 static int __init init_lapic_nmi_sysfs(void)
255 {
256         int error;
257
258         /*
259          * should really be a BUG_ON but b/c this is an
260          * init call, it just doesn't work.  -dcz
261          */
262         if (nmi_watchdog != NMI_LOCAL_APIC)
263                 return 0;
264
265         if (atomic_read(&nmi_active) < 0)
266                 return 0;
267
268         error = sysdev_class_register(&nmi_sysclass);
269         if (!error)
270                 error = sysdev_register(&device_lapic_nmi);
271         return error;
272 }
273
274 /* must come after the local APIC's device_initcall() */
275 late_initcall(init_lapic_nmi_sysfs);
276
277 #endif  /* CONFIG_PM */
278
279 static void __acpi_nmi_enable(void *__unused)
280 {
281         apic_write(APIC_LVT0, APIC_DM_NMI);
282 }
283
284 /*
285  * Enable timer based NMIs on all CPUs:
286  */
287 void acpi_nmi_enable(void)
288 {
289         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
290                 on_each_cpu(__acpi_nmi_enable, NULL, 1);
291 }
292
293 /*
294  * Disable timer based NMIs on all CPUs:
295  */
296 void acpi_nmi_disable(void)
297 {
298         if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
299                 on_each_cpu(__acpi_nmi_disable, NULL, 1);
300 }
301
302 /*
303  * This function is called as soon the LAPIC NMI watchdog driver has everything
304  * in place and it's ready to check if the NMIs belong to the NMI watchdog
305  */
306 void cpu_nmi_set_wd_enabled(void)
307 {
308         __get_cpu_var(wd_enabled) = 1;
309 }
310
311 void setup_apic_nmi_watchdog(void *unused)
312 {
313         if (__get_cpu_var(wd_enabled))
314                 return;
315
316         /* cheap hack to support suspend/resume */
317         /* if cpu0 is not active neither should the other cpus */
318         if (smp_processor_id() != 0 && atomic_read(&nmi_active) <= 0)
319                 return;
320
321         switch (nmi_watchdog) {
322         case NMI_LOCAL_APIC:
323                 if (lapic_watchdog_init(nmi_hz) < 0) {
324                         __get_cpu_var(wd_enabled) = 0;
325                         return;
326                 }
327                 /* FALL THROUGH */
328         case NMI_IO_APIC:
329                 __get_cpu_var(wd_enabled) = 1;
330                 atomic_inc(&nmi_active);
331         }
332 }
333
334 void stop_apic_nmi_watchdog(void *unused)
335 {
336         /* only support LOCAL and IO APICs for now */
337         if (!nmi_watchdog_active())
338                 return;
339         if (__get_cpu_var(wd_enabled) == 0)
340                 return;
341         if (nmi_watchdog == NMI_LOCAL_APIC)
342                 lapic_watchdog_stop();
343         else
344                 __acpi_nmi_disable(NULL);
345         __get_cpu_var(wd_enabled) = 0;
346         atomic_dec(&nmi_active);
347 }
348
349 /*
350  * the best way to detect whether a CPU has a 'hard lockup' problem
351  * is to check it's local APIC timer IRQ counts. If they are not
352  * changing then that CPU has some problem.
353  *
354  * as these watchdog NMI IRQs are generated on every CPU, we only
355  * have to check the current processor.
356  *
357  * since NMIs don't listen to _any_ locks, we have to be extremely
358  * careful not to rely on unsafe variables. The printk might lock
359  * up though, so we have to break up any console locks first ...
360  * [when there will be more tty-related locks, break them up here too!]
361  */
362
363 static DEFINE_PER_CPU(unsigned, last_irq_sum);
364 static DEFINE_PER_CPU(long, alert_counter);
365 static DEFINE_PER_CPU(int, nmi_touch);
366
367 void touch_nmi_watchdog(void)
368 {
369         if (nmi_watchdog_active()) {
370                 unsigned cpu;
371
372                 /*
373                  * Tell other CPUs to reset their alert counters. We cannot
374                  * do it ourselves because the alert count increase is not
375                  * atomic.
376                  */
377                 for_each_present_cpu(cpu) {
378                         if (per_cpu(nmi_touch, cpu) != 1)
379                                 per_cpu(nmi_touch, cpu) = 1;
380                 }
381         }
382
383         /*
384          * Tickle the softlockup detector too:
385          */
386         touch_softlockup_watchdog();
387 }
388 EXPORT_SYMBOL(touch_nmi_watchdog);
389
390 notrace __kprobes int
391 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
392 {
393         /*
394          * Since current_thread_info()-> is always on the stack, and we
395          * always switch the stack NMI-atomically, it's safe to use
396          * smp_processor_id().
397          */
398         unsigned int sum;
399         int touched = 0;
400         int cpu = smp_processor_id();
401         int rc = 0;
402
403         /* check for other users first */
404         if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
405                         == NOTIFY_STOP) {
406                 rc = 1;
407                 touched = 1;
408         }
409
410         sum = get_timer_irqs(cpu);
411
412         if (__get_cpu_var(nmi_touch)) {
413                 __get_cpu_var(nmi_touch) = 0;
414                 touched = 1;
415         }
416
417         /* We can be called before check_nmi_watchdog, hence NULL check. */
418         if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
419                 static DEFINE_SPINLOCK(lock);   /* Serialise the printks */
420
421                 spin_lock(&lock);
422                 printk(KERN_WARNING "NMI backtrace for cpu %d\n", cpu);
423                 show_regs(regs);
424                 dump_stack();
425                 spin_unlock(&lock);
426                 cpumask_clear_cpu(cpu, to_cpumask(backtrace_mask));
427
428                 rc = 1;
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                 __this_cpu_inc(alert_counter);
442                 if (__this_cpu_read(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                 __this_cpu_write(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,
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, 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 arch_trigger_all_cpu_backtrace(void)
559 {
560         int i;
561
562         cpumask_copy(to_cpumask(backtrace_mask), cpu_online_mask);
563
564         printk(KERN_INFO "sending NMI to all CPUs:\n");
565         apic->send_IPI_all(NMI_VECTOR);
566
567         /* Wait for up to 10 seconds for all CPUs to do the backtrace */
568         for (i = 0; i < 10 * 1000; i++) {
569                 if (cpumask_empty(to_cpumask(backtrace_mask)))
570                         break;
571                 mdelay(1);
572         }
573 }