2 * NMI watchdog support on APIC systems
4 * Started by Ingo Molnar <mingo@redhat.com>
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.
11 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
14 #include <linux/delay.h>
15 #include <linux/interrupt.h>
16 #include <linux/module.h>
17 #include <linux/nmi.h>
18 #include <linux/sysdev.h>
19 #include <linux/sysctl.h>
20 #include <linux/percpu.h>
21 #include <linux/kprobes.h>
22 #include <linux/cpumask.h>
23 #include <linux/kernel_stat.h>
24 #include <linux/kdebug.h>
25 #include <linux/slab.h>
29 #include <asm/timer.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 static int endflag __initdata = 0;
54 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
55 * the CPU is idle. To make sure the NMI watchdog really ticks on all
56 * CPUs during the test make them busy.
58 static __init void nmi_cpu_busy(void *data)
60 local_irq_enable_in_hardirq();
61 /* Intentionally don't use cpu_relax here. This is
62 to make sure that the performance counter really ticks,
63 even if there is a simulator or similar that catches the
64 pause instruction. On a real HT machine this is fine because
65 all other CPUs are busy with "useless" delay loops and don't
66 care if they get somewhat less cycles. */
72 int __init check_nmi_watchdog(void)
74 unsigned int *prev_nmi_count;
77 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
80 if (!atomic_read(&nmi_active))
83 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
87 printk(KERN_INFO "Testing NMI watchdog ... ");
90 if (nmi_watchdog == NMI_LOCAL_APIC)
91 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
94 for_each_possible_cpu(cpu)
95 prev_nmi_count[cpu] = nmi_count(cpu);
97 mdelay((20*1000)/nmi_hz); // wait 20 ticks
99 for_each_possible_cpu(cpu) {
101 /* Check cpu_callin_map here because that is set
102 after the timer is started. */
103 if (!cpu_isset(cpu, cpu_callin_map))
106 if (!per_cpu(wd_enabled, cpu))
108 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
109 printk(KERN_WARNING "WARNING: CPU#%d: NMI "
110 "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 timer_ack = !cpu_has_tsc;
139 static int __init setup_nmi_watchdog(char *str)
143 get_option(&str, &nmi);
145 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
152 __setup("nmi_watchdog=", setup_nmi_watchdog);
155 /* Suspend/resume support */
159 static int nmi_pm_active; /* nmi_active before suspend */
161 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
163 /* only CPU0 goes here, other CPUs should be offline */
164 nmi_pm_active = atomic_read(&nmi_active);
165 stop_apic_nmi_watchdog(NULL);
166 BUG_ON(atomic_read(&nmi_active) != 0);
170 static int lapic_nmi_resume(struct sys_device *dev)
172 /* only CPU0 goes here, other CPUs should be offline */
173 if (nmi_pm_active > 0) {
174 setup_apic_nmi_watchdog(NULL);
175 touch_nmi_watchdog();
181 static struct sysdev_class nmi_sysclass = {
183 .resume = lapic_nmi_resume,
184 .suspend = lapic_nmi_suspend,
187 static struct sys_device device_lapic_nmi = {
189 .cls = &nmi_sysclass,
192 static int __init init_lapic_nmi_sysfs(void)
196 /* should really be a BUG_ON but b/c this is an
197 * init call, it just doesn't work. -dcz
199 if (nmi_watchdog != NMI_LOCAL_APIC)
202 if (atomic_read(&nmi_active) < 0)
205 error = sysdev_class_register(&nmi_sysclass);
207 error = sysdev_register(&device_lapic_nmi);
210 /* must come after the local APIC's device_initcall() */
211 late_initcall(init_lapic_nmi_sysfs);
213 #endif /* CONFIG_PM */
215 static void __acpi_nmi_enable(void *__unused)
217 apic_write_around(APIC_LVT0, APIC_DM_NMI);
221 * Enable timer based NMIs on all CPUs:
223 void acpi_nmi_enable(void)
225 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
226 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
229 static void __acpi_nmi_disable(void *__unused)
231 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
235 * Disable timer based NMIs on all CPUs:
237 void acpi_nmi_disable(void)
239 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
240 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
243 void setup_apic_nmi_watchdog(void *unused)
245 if (__get_cpu_var(wd_enabled))
248 /* cheap hack to support suspend/resume */
249 /* if cpu0 is not active neither should the other cpus */
250 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
253 switch (nmi_watchdog) {
255 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
256 if (lapic_watchdog_init(nmi_hz) < 0) {
257 __get_cpu_var(wd_enabled) = 0;
262 __get_cpu_var(wd_enabled) = 1;
263 atomic_inc(&nmi_active);
267 void stop_apic_nmi_watchdog(void *unused)
269 /* only support LOCAL and IO APICs for now */
270 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
271 (nmi_watchdog != NMI_IO_APIC))
273 if (__get_cpu_var(wd_enabled) == 0)
275 if (nmi_watchdog == NMI_LOCAL_APIC)
276 lapic_watchdog_stop();
277 __get_cpu_var(wd_enabled) = 0;
278 atomic_dec(&nmi_active);
282 * the best way to detect whether a CPU has a 'hard lockup' problem
283 * is to check it's local APIC timer IRQ counts. If they are not
284 * changing then that CPU has some problem.
286 * as these watchdog NMI IRQs are generated on every CPU, we only
287 * have to check the current processor.
289 * since NMIs don't listen to _any_ locks, we have to be extremely
290 * careful not to rely on unsafe variables. The printk might lock
291 * up though, so we have to break up any console locks first ...
292 * [when there will be more tty-related locks, break them up
297 last_irq_sums [NR_CPUS],
298 alert_counter [NR_CPUS];
300 void touch_nmi_watchdog(void)
302 if (nmi_watchdog > 0) {
306 * Just reset the alert counters, (other CPUs might be
307 * spinning on locks we hold):
309 for_each_present_cpu(cpu) {
310 if (alert_counter[cpu])
311 alert_counter[cpu] = 0;
316 * Tickle the softlockup detector too:
318 touch_softlockup_watchdog();
320 EXPORT_SYMBOL(touch_nmi_watchdog);
322 extern void die_nmi(struct pt_regs *, const char *msg);
324 notrace __kprobes int
325 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
329 * Since current_thread_info()-> is always on the stack, and we
330 * always switch the stack NMI-atomically, it's safe to use
331 * smp_processor_id().
335 int cpu = smp_processor_id();
338 /* check for other users first */
339 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
345 if (cpu_isset(cpu, backtrace_mask)) {
346 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
349 printk("NMI backtrace for cpu %d\n", cpu);
352 cpu_clear(cpu, backtrace_mask);
356 * Take the local apic timer and PIT/HPET into account. We don't
357 * know which one is active, when we have highres/dyntick on
359 sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
360 per_cpu(irq_stat, cpu).irq0_irqs;
362 /* if the none of the timers isn't firing, this cpu isn't doing much */
363 if (!touched && last_irq_sums[cpu] == sum) {
365 * Ayiee, looks like this CPU is stuck ...
366 * wait a few IRQs (5 seconds) before doing the oops ...
368 alert_counter[cpu]++;
369 if (alert_counter[cpu] == 5*nmi_hz)
371 * die_nmi will return ONLY if NOTIFY_STOP happens..
373 die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
375 last_irq_sums[cpu] = sum;
376 alert_counter[cpu] = 0;
378 /* see if the nmi watchdog went off */
379 if (!__get_cpu_var(wd_enabled))
381 switch (nmi_watchdog) {
383 rc |= lapic_wd_event(nmi_hz);
386 /* don't know how to accurately check for this.
387 * just assume it was a watchdog timer interrupt
388 * This matches the old behaviour.
398 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
400 unsigned char reason = get_nmi_reason();
403 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
409 * proc handler for /proc/sys/kernel/nmi
411 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
412 void __user *buffer, size_t *length, loff_t *ppos)
416 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
417 old_state = nmi_watchdog_enabled;
418 proc_dointvec(table, write, file, buffer, length, ppos);
419 if (!!old_state == !!nmi_watchdog_enabled)
422 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
423 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
427 if (nmi_watchdog == NMI_DEFAULT) {
428 if (lapic_watchdog_ok())
429 nmi_watchdog = NMI_LOCAL_APIC;
431 nmi_watchdog = NMI_IO_APIC;
434 if (nmi_watchdog == NMI_LOCAL_APIC) {
435 if (nmi_watchdog_enabled)
436 enable_lapic_nmi_watchdog();
438 disable_lapic_nmi_watchdog();
441 "NMI watchdog doesn't know what hardware to touch\n");
449 int do_nmi_callback(struct pt_regs *regs, int cpu)
452 if (unknown_nmi_panic)
453 return unknown_nmi_panic_callback(regs, cpu);
458 void __trigger_all_cpu_backtrace(void)
462 backtrace_mask = cpu_online_map;
463 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
464 for (i = 0; i < 10 * 1000; i++) {
465 if (cpus_empty(backtrace_mask))
471 EXPORT_SYMBOL(nmi_active);
472 EXPORT_SYMBOL(nmi_watchdog);