2 * linux/arch/x86_64/nmi.c
4 * NMI watchdog support on APIC systems
6 * Started by Ingo Molnar <mingo@redhat.com>
9 * Mikael Pettersson : AMD K7 support for local APIC NMI watchdog.
10 * Mikael Pettersson : Power Management for local APIC NMI watchdog.
12 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
15 #include <linux/nmi.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>
28 #include <asm/proto.h>
31 int unknown_nmi_panic;
32 int nmi_watchdog_enabled;
33 int panic_on_unrecovered_nmi;
35 static cpumask_t backtrace_mask = CPU_MASK_NONE;
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
41 * 0: the lapic NMI watchdog is disabled, but can be enabled
43 atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
46 unsigned int nmi_watchdog = NMI_DEFAULT;
47 static unsigned int nmi_hz = HZ;
49 static DEFINE_PER_CPU(short, wd_enabled);
51 /* local prototypes */
52 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
54 /* Run after command line and cpu_init init, but before all other checks */
55 void nmi_watchdog_default(void)
57 if (nmi_watchdog != NMI_DEFAULT)
59 nmi_watchdog = NMI_NONE;
62 static int endflag __initdata = 0;
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.
69 static __init void nmi_cpu_busy(void *data)
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. */
83 int __init check_nmi_watchdog (void)
88 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DEFAULT))
91 if (!atomic_read(&nmi_active))
94 counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
98 printk(KERN_INFO "testing NMI watchdog ... ");
101 if (nmi_watchdog == NMI_LOCAL_APIC)
102 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
105 for (cpu = 0; cpu < NR_CPUS; cpu++)
106 counts[cpu] = cpu_pda(cpu)->__nmi_count;
108 mdelay((20*1000)/nmi_hz); // wait 20 ticks
110 for_each_online_cpu(cpu) {
111 if (!per_cpu(wd_enabled, cpu))
113 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
114 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
117 cpu_pda(cpu)->__nmi_count);
118 per_cpu(wd_enabled, cpu) = 0;
119 atomic_dec(&nmi_active);
122 if (!atomic_read(&nmi_active)) {
124 atomic_set(&nmi_active, -1);
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);
140 int __init setup_nmi_watchdog(char *str)
144 if (!strncmp(str,"panic",5)) {
145 panic_on_timeout = 1;
146 str = strchr(str, ',');
152 get_option(&str, &nmi);
154 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
161 __setup("nmi_watchdog=", setup_nmi_watchdog);
164 static void __acpi_nmi_disable(void *__unused)
166 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
170 * Disable timer based NMIs on all CPUs:
172 void acpi_nmi_disable(void)
174 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
175 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
178 static void __acpi_nmi_enable(void *__unused)
180 apic_write(APIC_LVT0, APIC_DM_NMI);
184 * Enable timer based NMIs on all CPUs:
186 void acpi_nmi_enable(void)
188 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
189 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
193 static int nmi_pm_active; /* nmi_active before suspend */
195 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
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);
204 static int lapic_nmi_resume(struct sys_device *dev)
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();
214 static struct sysdev_class nmi_sysclass = {
215 set_kset_name("lapic_nmi"),
216 .resume = lapic_nmi_resume,
217 .suspend = lapic_nmi_suspend,
220 static struct sys_device device_lapic_nmi = {
222 .cls = &nmi_sysclass,
225 static int __init init_lapic_nmi_sysfs(void)
229 /* should really be a BUG_ON but b/c this is an
230 * init call, it just doesn't work. -dcz
232 if (nmi_watchdog != NMI_LOCAL_APIC)
235 if ( atomic_read(&nmi_active) < 0 )
238 error = sysdev_class_register(&nmi_sysclass);
240 error = sysdev_register(&device_lapic_nmi);
243 /* must come after the local APIC's device_initcall() */
244 late_initcall(init_lapic_nmi_sysfs);
246 #endif /* CONFIG_PM */
248 void setup_apic_nmi_watchdog(void *unused)
250 if (__get_cpu_var(wd_enabled) == 1)
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))
258 switch (nmi_watchdog) {
260 __get_cpu_var(wd_enabled) = 1;
261 if (lapic_watchdog_init(nmi_hz) < 0) {
262 __get_cpu_var(wd_enabled) = 0;
267 __get_cpu_var(wd_enabled) = 1;
268 atomic_inc(&nmi_active);
272 void stop_apic_nmi_watchdog(void *unused)
274 /* only support LOCAL and IO APICs for now */
275 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
276 (nmi_watchdog != NMI_IO_APIC))
278 if (__get_cpu_var(wd_enabled) == 0)
280 if (nmi_watchdog == NMI_LOCAL_APIC)
281 lapic_watchdog_stop();
282 __get_cpu_var(wd_enabled) = 0;
283 atomic_dec(&nmi_active);
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.
291 * as these watchdog NMI IRQs are generated on every CPU, we only
292 * have to check the current processor.
295 static DEFINE_PER_CPU(unsigned, last_irq_sum);
296 static DEFINE_PER_CPU(local_t, alert_counter);
297 static DEFINE_PER_CPU(int, nmi_touch);
299 void touch_nmi_watchdog(void)
301 if (nmi_watchdog > 0) {
305 * Tell other CPUs to reset their alert counters. We cannot
306 * do it ourselves because the alert count increase is not
309 for_each_present_cpu(cpu) {
310 if (per_cpu(nmi_touch, cpu) != 1)
311 per_cpu(nmi_touch, cpu) = 1;
315 touch_softlockup_watchdog();
318 int __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
322 int cpu = smp_processor_id();
325 /* check for other users first */
326 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
332 sum = read_pda(apic_timer_irqs);
333 if (__get_cpu_var(nmi_touch)) {
334 __get_cpu_var(nmi_touch) = 0;
338 if (cpu_isset(cpu, backtrace_mask)) {
339 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
342 printk("NMI backtrace for cpu %d\n", cpu);
345 cpu_clear(cpu, backtrace_mask);
348 #ifdef CONFIG_X86_MCE
349 /* Could check oops_in_progress here too, but it's safer
351 if (atomic_read(&mce_entry) > 0)
354 /* if the apic timer isn't firing, this cpu isn't doing much */
355 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
357 * Ayiee, looks like this CPU is stuck ...
358 * wait a few IRQs (5 seconds) before doing the oops ...
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,
365 __get_cpu_var(last_irq_sum) = sum;
366 local_set(&__get_cpu_var(alert_counter), 0);
369 /* see if the nmi watchdog went off */
370 if (!__get_cpu_var(wd_enabled))
372 switch (nmi_watchdog) {
374 rc |= lapic_wd_event(nmi_hz);
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.
387 static unsigned ignore_nmis;
389 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
392 add_pda(__nmi_count,1);
394 default_do_nmi(regs);
398 int do_nmi_callback(struct pt_regs * regs, int cpu)
401 if (unknown_nmi_panic)
402 return unknown_nmi_panic_callback(regs, cpu);
413 void restart_nmi(void)
421 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
423 unsigned char reason = get_nmi_reason();
426 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
427 die_nmi(buf, regs, 1); /* Always panic here */
432 * proc handler for /proc/sys/kernel/nmi
434 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
435 void __user *buffer, size_t *length, loff_t *ppos)
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)
445 if (atomic_read(&nmi_active) < 0) {
446 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
450 /* if nmi_watchdog is not set yet, then set it */
451 nmi_watchdog_default();
453 if (nmi_watchdog == NMI_LOCAL_APIC) {
454 if (nmi_watchdog_enabled)
455 enable_lapic_nmi_watchdog();
457 disable_lapic_nmi_watchdog();
460 "NMI watchdog doesn't know what hardware to touch\n");
468 void __trigger_all_cpu_backtrace(void)
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))
481 EXPORT_SYMBOL(nmi_active);
482 EXPORT_SYMBOL(nmi_watchdog);
483 EXPORT_SYMBOL(touch_nmi_watchdog);