2 * linux/arch/i386/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.
11 * Mikael Pettersson : Pentium 4 support for local APIC NMI watchdog.
13 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
16 #include <linux/delay.h>
17 #include <linux/interrupt.h>
18 #include <linux/module.h>
19 #include <linux/nmi.h>
20 #include <linux/sysdev.h>
21 #include <linux/sysctl.h>
22 #include <linux/percpu.h>
23 #include <linux/kprobes.h>
24 #include <linux/cpumask.h>
25 #include <linux/kernel_stat.h>
26 #include <linux/kdebug.h>
31 #include "mach_traps.h"
33 int unknown_nmi_panic;
34 int nmi_watchdog_enabled;
36 static cpumask_t backtrace_mask = CPU_MASK_NONE;
39 * >0: the lapic NMI watchdog is active, but can be disabled
40 * <0: the lapic NMI watchdog has not been set up, and cannot
42 * 0: the lapic NMI watchdog is disabled, but can be enabled
44 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 static int endflag __initdata = 0;
57 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
58 * the CPU is idle. To make sure the NMI watchdog really ticks on all
59 * CPUs during the test make them busy.
61 static __init void nmi_cpu_busy(void *data)
63 local_irq_enable_in_hardirq();
64 /* Intentionally don't use cpu_relax here. This is
65 to make sure that the performance counter really ticks,
66 even if there is a simulator or similar that catches the
67 pause instruction. On a real HT machine this is fine because
68 all other CPUs are busy with "useless" delay loops and don't
69 care if they get somewhat less cycles. */
75 static int __init check_nmi_watchdog(void)
77 unsigned int *prev_nmi_count;
80 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
83 if (!atomic_read(&nmi_active))
86 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
90 printk(KERN_INFO "Testing NMI watchdog ... ");
92 if (nmi_watchdog == NMI_LOCAL_APIC)
93 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
95 for_each_possible_cpu(cpu)
96 prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
98 mdelay((20*1000)/nmi_hz); // wait 20 ticks
100 for_each_possible_cpu(cpu) {
102 /* Check cpu_callin_map here because that is set
103 after the timer is started. */
104 if (!cpu_isset(cpu, cpu_callin_map))
107 if (!per_cpu(wd_enabled, cpu))
109 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
110 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
114 per_cpu(wd_enabled, cpu) = 0;
115 atomic_dec(&nmi_active);
119 if (!atomic_read(&nmi_active)) {
120 kfree(prev_nmi_count);
121 atomic_set(&nmi_active, -1);
126 /* now that we know it works we can reduce NMI frequency to
127 something more reasonable; makes a difference in some configs */
128 if (nmi_watchdog == NMI_LOCAL_APIC)
129 nmi_hz = lapic_adjust_nmi_hz(1);
131 kfree(prev_nmi_count);
134 /* This needs to happen later in boot so counters are working */
135 late_initcall(check_nmi_watchdog);
137 static int __init setup_nmi_watchdog(char *str)
141 get_option(&str, &nmi);
143 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
150 __setup("nmi_watchdog=", setup_nmi_watchdog);
153 /* Suspend/resume support */
157 static int nmi_pm_active; /* nmi_active before suspend */
159 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
161 /* only CPU0 goes here, other CPUs should be offline */
162 nmi_pm_active = atomic_read(&nmi_active);
163 stop_apic_nmi_watchdog(NULL);
164 BUG_ON(atomic_read(&nmi_active) != 0);
168 static int lapic_nmi_resume(struct sys_device *dev)
170 /* only CPU0 goes here, other CPUs should be offline */
171 if (nmi_pm_active > 0) {
172 setup_apic_nmi_watchdog(NULL);
173 touch_nmi_watchdog();
179 static struct sysdev_class nmi_sysclass = {
180 set_kset_name("lapic_nmi"),
181 .resume = lapic_nmi_resume,
182 .suspend = lapic_nmi_suspend,
185 static struct sys_device device_lapic_nmi = {
187 .cls = &nmi_sysclass,
190 static int __init init_lapic_nmi_sysfs(void)
194 /* should really be a BUG_ON but b/c this is an
195 * init call, it just doesn't work. -dcz
197 if (nmi_watchdog != NMI_LOCAL_APIC)
200 if (atomic_read(&nmi_active) < 0)
203 error = sysdev_class_register(&nmi_sysclass);
205 error = sysdev_register(&device_lapic_nmi);
208 /* must come after the local APIC's device_initcall() */
209 late_initcall(init_lapic_nmi_sysfs);
211 #endif /* CONFIG_PM */
213 static void __acpi_nmi_enable(void *__unused)
215 apic_write_around(APIC_LVT0, APIC_DM_NMI);
219 * Enable timer based NMIs on all CPUs:
221 void acpi_nmi_enable(void)
223 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
224 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
227 static void __acpi_nmi_disable(void *__unused)
229 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
233 * Disable timer based NMIs on all CPUs:
235 void acpi_nmi_disable(void)
237 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
238 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
241 void setup_apic_nmi_watchdog (void *unused)
243 if (__get_cpu_var(wd_enabled))
246 /* cheap hack to support suspend/resume */
247 /* if cpu0 is not active neither should the other cpus */
248 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
251 switch (nmi_watchdog) {
253 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
254 if (lapic_watchdog_init(nmi_hz) < 0) {
255 __get_cpu_var(wd_enabled) = 0;
260 __get_cpu_var(wd_enabled) = 1;
261 atomic_inc(&nmi_active);
265 void stop_apic_nmi_watchdog(void *unused)
267 /* only support LOCAL and IO APICs for now */
268 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
269 (nmi_watchdog != NMI_IO_APIC))
271 if (__get_cpu_var(wd_enabled) == 0)
273 if (nmi_watchdog == NMI_LOCAL_APIC)
274 lapic_watchdog_stop();
275 __get_cpu_var(wd_enabled) = 0;
276 atomic_dec(&nmi_active);
280 * the best way to detect whether a CPU has a 'hard lockup' problem
281 * is to check it's local APIC timer IRQ counts. If they are not
282 * changing then that CPU has some problem.
284 * as these watchdog NMI IRQs are generated on every CPU, we only
285 * have to check the current processor.
287 * since NMIs don't listen to _any_ locks, we have to be extremely
288 * careful not to rely on unsafe variables. The printk might lock
289 * up though, so we have to break up any console locks first ...
290 * [when there will be more tty-related locks, break them up
295 last_irq_sums [NR_CPUS],
296 alert_counter [NR_CPUS];
298 void touch_nmi_watchdog(void)
300 if (nmi_watchdog > 0) {
304 * Just reset the alert counters, (other CPUs might be
305 * spinning on locks we hold):
307 for_each_present_cpu(cpu) {
308 if (alert_counter[cpu])
309 alert_counter[cpu] = 0;
314 * Tickle the softlockup detector too:
316 touch_softlockup_watchdog();
318 EXPORT_SYMBOL(touch_nmi_watchdog);
320 extern void die_nmi(struct pt_regs *, const char *msg);
322 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
326 * Since current_thread_info()-> is always on the stack, and we
327 * always switch the stack NMI-atomically, it's safe to use
328 * smp_processor_id().
332 int cpu = smp_processor_id();
335 /* check for other users first */
336 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
342 if (cpu_isset(cpu, backtrace_mask)) {
343 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
346 printk("NMI backtrace for cpu %d\n", cpu);
349 cpu_clear(cpu, backtrace_mask);
353 * Take the local apic timer and PIT/HPET into account. We don't
354 * know which one is active, when we have highres/dyntick on
356 sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
357 per_cpu(irq_stat, cpu).irq0_irqs;
359 /* if the none of the timers isn't firing, this cpu isn't doing much */
360 if (!touched && last_irq_sums[cpu] == sum) {
362 * Ayiee, looks like this CPU is stuck ...
363 * wait a few IRQs (5 seconds) before doing the oops ...
365 alert_counter[cpu]++;
366 if (alert_counter[cpu] == 5*nmi_hz)
368 * die_nmi will return ONLY if NOTIFY_STOP happens..
370 die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
372 last_irq_sums[cpu] = sum;
373 alert_counter[cpu] = 0;
375 /* see if the nmi watchdog went off */
376 if (!__get_cpu_var(wd_enabled))
378 switch (nmi_watchdog) {
380 rc |= lapic_wd_event(nmi_hz);
383 /* don't know how to accurately check for this.
384 * just assume it was a watchdog timer interrupt
385 * This matches the old behaviour.
393 int do_nmi_callback(struct pt_regs * regs, int cpu)
396 if (unknown_nmi_panic)
397 return unknown_nmi_panic_callback(regs, cpu);
404 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
406 unsigned char reason = get_nmi_reason();
409 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
415 * proc handler for /proc/sys/kernel/nmi
417 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
418 void __user *buffer, size_t *length, loff_t *ppos)
422 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
423 old_state = nmi_watchdog_enabled;
424 proc_dointvec(table, write, file, buffer, length, ppos);
425 if (!!old_state == !!nmi_watchdog_enabled)
428 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
429 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
433 if (nmi_watchdog == NMI_DEFAULT) {
434 if (lapic_watchdog_ok())
435 nmi_watchdog = NMI_LOCAL_APIC;
437 nmi_watchdog = NMI_IO_APIC;
440 if (nmi_watchdog == NMI_LOCAL_APIC) {
441 if (nmi_watchdog_enabled)
442 enable_lapic_nmi_watchdog();
444 disable_lapic_nmi_watchdog();
447 "NMI watchdog doesn't know what hardware to touch\n");
455 void __trigger_all_cpu_backtrace(void)
459 backtrace_mask = cpu_online_map;
460 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
461 for (i = 0; i < 10 * 1000; i++) {
462 if (cpus_empty(backtrace_mask))
468 EXPORT_SYMBOL(nmi_active);
469 EXPORT_SYMBOL(nmi_watchdog);