aboutsummaryrefslogtreecommitdiffstats
path: root/arch/x86/kernel/nmi_64.c
diff options
context:
space:
mode:
authorThomas Gleixner <tglx@linutronix.de>2007-10-11 05:17:24 -0400
committerThomas Gleixner <tglx@linutronix.de>2007-10-11 05:17:24 -0400
commit250c22777fe1ccd7ac588579a6c16db4c0161cc5 (patch)
tree55c317efb7d792ec6fdae1d1937c67a502c48dec /arch/x86/kernel/nmi_64.c
parent2db55d344e529492545cb3b755c7e9ba8e4fa94e (diff)
x86_64: move kernel
Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'arch/x86/kernel/nmi_64.c')
-rw-r--r--arch/x86/kernel/nmi_64.c483
1 files changed, 483 insertions, 0 deletions
diff --git a/arch/x86/kernel/nmi_64.c b/arch/x86/kernel/nmi_64.c
new file mode 100644
index 000000000000..0ec6d2ddb931
--- /dev/null
+++ b/arch/x86/kernel/nmi_64.c
@@ -0,0 +1,483 @@
1/*
2 * linux/arch/x86_64/nmi.c
3 *
4 * NMI watchdog support on APIC systems
5 *
6 * Started by Ingo Molnar <mingo@redhat.com>
7 *
8 * Fixes:
9 * Mikael Pettersson : AMD K7 support for local APIC NMI watchdog.
10 * Mikael Pettersson : Power Management for local APIC NMI watchdog.
11 * Pavel Machek and
12 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
13 */
14
15#include <linux/nmi.h>
16#include <linux/mm.h>
17#include <linux/delay.h>
18#include <linux/interrupt.h>
19#include <linux/module.h>
20#include <linux/sysdev.h>
21#include <linux/sysctl.h>
22#include <linux/kprobes.h>
23#include <linux/cpumask.h>
24#include <linux/kdebug.h>
25
26#include <asm/smp.h>
27#include <asm/nmi.h>
28#include <asm/proto.h>
29#include <asm/mce.h>
30
31int unknown_nmi_panic;
32int nmi_watchdog_enabled;
33int panic_on_unrecovered_nmi;
34
35static 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 */
43atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
44int panic_on_timeout;
45
46unsigned int nmi_watchdog = NMI_DEFAULT;
47static unsigned int nmi_hz = HZ;
48
49static DEFINE_PER_CPU(short, wd_enabled);
50
51/* local prototypes */
52static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
53
54/* Run after command line and cpu_init init, but before all other checks */
55void nmi_watchdog_default(void)
56{
57 if (nmi_watchdog != NMI_DEFAULT)
58 return;
59 nmi_watchdog = NMI_NONE;
60}
61
62static int endflag __initdata = 0;
63
64#ifdef CONFIG_SMP
65/* The performance counters used by NMI_LOCAL_APIC don't trigger when
66 * the CPU is idle. To make sure the NMI watchdog really ticks on all
67 * CPUs during the test make them busy.
68 */
69static __init void nmi_cpu_busy(void *data)
70{
71 local_irq_enable_in_hardirq();
72 /* Intentionally don't use cpu_relax here. This is
73 to make sure that the performance counter really ticks,
74 even if there is a simulator or similar that catches the
75 pause instruction. On a real HT machine this is fine because
76 all other CPUs are busy with "useless" delay loops and don't
77 care if they get somewhat less cycles. */
78 while (endflag == 0)
79 mb();
80}
81#endif
82
83int __init check_nmi_watchdog (void)
84{
85 int *counts;
86 int cpu;
87
88 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
89 return 0;
90
91 if (!atomic_read(&nmi_active))
92 return 0;
93
94 counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
95 if (!counts)
96 return -1;
97
98 printk(KERN_INFO "testing NMI watchdog ... ");
99
100#ifdef CONFIG_SMP
101 if (nmi_watchdog == NMI_LOCAL_APIC)
102 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
103#endif
104
105 for (cpu = 0; cpu < NR_CPUS; cpu++)
106 counts[cpu] = cpu_pda(cpu)->__nmi_count;
107 local_irq_enable();
108 mdelay((20*1000)/nmi_hz); // wait 20 ticks
109
110 for_each_online_cpu(cpu) {
111 if (!per_cpu(wd_enabled, cpu))
112 continue;
113 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
114 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
115 cpu,
116 counts[cpu],
117 cpu_pda(cpu)->__nmi_count);
118 per_cpu(wd_enabled, cpu) = 0;
119 atomic_dec(&nmi_active);
120 }
121 }
122 if (!atomic_read(&nmi_active)) {
123 kfree(counts);
124 atomic_set(&nmi_active, -1);
125 endflag = 1;
126 return -1;
127 }
128 endflag = 1;
129 printk("OK.\n");
130
131 /* now that we know it works we can reduce NMI frequency to
132 something more reasonable; makes a difference in some configs */
133 if (nmi_watchdog == NMI_LOCAL_APIC)
134 nmi_hz = lapic_adjust_nmi_hz(1);
135
136 kfree(counts);
137 return 0;
138}
139
140int __init setup_nmi_watchdog(char *str)
141{
142 int nmi;
143
144 if (!strncmp(str,"panic",5)) {
145 panic_on_timeout = 1;
146 str = strchr(str, ',');
147 if (!str)
148 return 1;
149 ++str;
150 }
151
152 get_option(&str, &nmi);
153
154 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
155 return 0;
156
157 nmi_watchdog = nmi;
158 return 1;
159}
160
161__setup("nmi_watchdog=", setup_nmi_watchdog);
162
163
164static void __acpi_nmi_disable(void *__unused)
165{
166 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
167}
168
169/*
170 * Disable timer based NMIs on all CPUs:
171 */
172void acpi_nmi_disable(void)
173{
174 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
175 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
176}
177
178static void __acpi_nmi_enable(void *__unused)
179{
180 apic_write(APIC_LVT0, APIC_DM_NMI);
181}
182
183/*
184 * Enable timer based NMIs on all CPUs:
185 */
186void acpi_nmi_enable(void)
187{
188 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
189 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
190}
191#ifdef CONFIG_PM
192
193static int nmi_pm_active; /* nmi_active before suspend */
194
195static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
196{
197 /* only CPU0 goes here, other CPUs should be offline */
198 nmi_pm_active = atomic_read(&nmi_active);
199 stop_apic_nmi_watchdog(NULL);
200 BUG_ON(atomic_read(&nmi_active) != 0);
201 return 0;
202}
203
204static int lapic_nmi_resume(struct sys_device *dev)
205{
206 /* only CPU0 goes here, other CPUs should be offline */
207 if (nmi_pm_active > 0) {
208 setup_apic_nmi_watchdog(NULL);
209 touch_nmi_watchdog();
210 }
211 return 0;
212}
213
214static struct sysdev_class nmi_sysclass = {
215 set_kset_name("lapic_nmi"),
216 .resume = lapic_nmi_resume,
217 .suspend = lapic_nmi_suspend,
218};
219
220static struct sys_device device_lapic_nmi = {
221 .id = 0,
222 .cls = &nmi_sysclass,
223};
224
225static int __init init_lapic_nmi_sysfs(void)
226{
227 int error;
228
229 /* should really be a BUG_ON but b/c this is an
230 * init call, it just doesn't work. -dcz
231 */
232 if (nmi_watchdog != NMI_LOCAL_APIC)
233 return 0;
234
235 if ( atomic_read(&nmi_active) < 0 )
236 return 0;
237
238 error = sysdev_class_register(&nmi_sysclass);
239 if (!error)
240 error = sysdev_register(&device_lapic_nmi);
241 return error;
242}
243/* must come after the local APIC's device_initcall() */
244late_initcall(init_lapic_nmi_sysfs);
245
246#endif /* CONFIG_PM */
247
248void setup_apic_nmi_watchdog(void *unused)
249{
250 if (__get_cpu_var(wd_enabled) == 1)
251 return;
252
253 /* cheap hack to support suspend/resume */
254 /* if cpu0 is not active neither should the other cpus */
255 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
256 return;
257
258 switch (nmi_watchdog) {
259 case NMI_LOCAL_APIC:
260 __get_cpu_var(wd_enabled) = 1;
261 if (lapic_watchdog_init(nmi_hz) < 0) {
262 __get_cpu_var(wd_enabled) = 0;
263 return;
264 }
265 /* FALL THROUGH */
266 case NMI_IO_APIC:
267 __get_cpu_var(wd_enabled) = 1;
268 atomic_inc(&nmi_active);
269 }
270}
271
272void stop_apic_nmi_watchdog(void *unused)
273{
274 /* only support LOCAL and IO APICs for now */
275 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
276 (nmi_watchdog != NMI_IO_APIC))
277 return;
278 if (__get_cpu_var(wd_enabled) == 0)
279 return;
280 if (nmi_watchdog == NMI_LOCAL_APIC)
281 lapic_watchdog_stop();
282 __get_cpu_var(wd_enabled) = 0;
283 atomic_dec(&nmi_active);
284}
285
286/*
287 * the best way to detect whether a CPU has a 'hard lockup' problem
288 * is to check it's local APIC timer IRQ counts. If they are not
289 * changing then that CPU has some problem.
290 *
291 * as these watchdog NMI IRQs are generated on every CPU, we only
292 * have to check the current processor.
293 */
294
295static DEFINE_PER_CPU(unsigned, last_irq_sum);
296static DEFINE_PER_CPU(local_t, alert_counter);
297static DEFINE_PER_CPU(int, nmi_touch);
298
299void touch_nmi_watchdog(void)
300{
301 if (nmi_watchdog > 0) {
302 unsigned cpu;
303
304 /*
305 * Tell other CPUs to reset their alert counters. We cannot
306 * do it ourselves because the alert count increase is not
307 * atomic.
308 */
309 for_each_present_cpu(cpu) {
310 if (per_cpu(nmi_touch, cpu) != 1)
311 per_cpu(nmi_touch, cpu) = 1;
312 }
313 }
314
315 touch_softlockup_watchdog();
316}
317
318int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
319{
320 int sum;
321 int touched = 0;
322 int cpu = smp_processor_id();
323 int rc = 0;
324
325 /* check for other users first */
326 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
327 == NOTIFY_STOP) {
328 rc = 1;
329 touched = 1;
330 }
331
332 sum = read_pda(apic_timer_irqs);
333 if (__get_cpu_var(nmi_touch)) {
334 __get_cpu_var(nmi_touch) = 0;
335 touched = 1;
336 }
337
338 if (cpu_isset(cpu, backtrace_mask)) {
339 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
340
341 spin_lock(&lock);
342 printk("NMI backtrace for cpu %d\n", cpu);
343 dump_stack();
344 spin_unlock(&lock);
345 cpu_clear(cpu, backtrace_mask);
346 }
347
348#ifdef CONFIG_X86_MCE
349 /* Could check oops_in_progress here too, but it's safer
350 not too */
351 if (atomic_read(&mce_entry) > 0)
352 touched = 1;
353#endif
354 /* if the apic timer isn't firing, this cpu isn't doing much */
355 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
356 /*
357 * Ayiee, looks like this CPU is stuck ...
358 * wait a few IRQs (5 seconds) before doing the oops ...
359 */
360 local_inc(&__get_cpu_var(alert_counter));
361 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
362 die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs,
363 panic_on_timeout);
364 } else {
365 __get_cpu_var(last_irq_sum) = sum;
366 local_set(&__get_cpu_var(alert_counter), 0);
367 }
368
369 /* see if the nmi watchdog went off */
370 if (!__get_cpu_var(wd_enabled))
371 return rc;
372 switch (nmi_watchdog) {
373 case NMI_LOCAL_APIC:
374 rc |= lapic_wd_event(nmi_hz);
375 break;
376 case NMI_IO_APIC:
377 /* don't know how to accurately check for this.
378 * just assume it was a watchdog timer interrupt
379 * This matches the old behaviour.
380 */
381 rc = 1;
382 break;
383 }
384 return rc;
385}
386
387static unsigned ignore_nmis;
388
389asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
390{
391 nmi_enter();
392 add_pda(__nmi_count,1);
393 if (!ignore_nmis)
394 default_do_nmi(regs);
395 nmi_exit();
396}
397
398int do_nmi_callback(struct pt_regs * regs, int cpu)
399{
400#ifdef CONFIG_SYSCTL
401 if (unknown_nmi_panic)
402 return unknown_nmi_panic_callback(regs, cpu);
403#endif
404 return 0;
405}
406
407void stop_nmi(void)
408{
409 acpi_nmi_disable();
410 ignore_nmis++;
411}
412
413void restart_nmi(void)
414{
415 ignore_nmis--;
416 acpi_nmi_enable();
417}
418
419#ifdef CONFIG_SYSCTL
420
421static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
422{
423 unsigned char reason = get_nmi_reason();
424 char buf[64];
425
426 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
427 die_nmi(buf, regs, 1); /* Always panic here */
428 return 0;
429}
430
431/*
432 * proc handler for /proc/sys/kernel/nmi
433 */
434int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
435 void __user *buffer, size_t *length, loff_t *ppos)
436{
437 int old_state;
438
439 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
440 old_state = nmi_watchdog_enabled;
441 proc_dointvec(table, write, file, buffer, length, ppos);
442 if (!!old_state == !!nmi_watchdog_enabled)
443 return 0;
444
445 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
446 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
447 return -EIO;
448 }
449
450 /* if nmi_watchdog is not set yet, then set it */
451 nmi_watchdog_default();
452
453 if (nmi_watchdog == NMI_LOCAL_APIC) {
454 if (nmi_watchdog_enabled)
455 enable_lapic_nmi_watchdog();
456 else
457 disable_lapic_nmi_watchdog();
458 } else {
459 printk( KERN_WARNING
460 "NMI watchdog doesn't know what hardware to touch\n");
461 return -EIO;
462 }
463 return 0;
464}
465
466#endif
467
468void __trigger_all_cpu_backtrace(void)
469{
470 int i;
471
472 backtrace_mask = cpu_online_map;
473 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
474 for (i = 0; i < 10 * 1000; i++) {
475 if (cpus_empty(backtrace_mask))
476 break;
477 mdelay(1);
478 }
479}
480
481EXPORT_SYMBOL(nmi_active);
482EXPORT_SYMBOL(nmi_watchdog);
483EXPORT_SYMBOL(touch_nmi_watchdog);