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>
28 #include <asm/timer.h>
30 #include "mach_traps.h"
32 int unknown_nmi_panic;
33 int nmi_watchdog_enabled;
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 */
45 unsigned int nmi_watchdog = NMI_DEFAULT;
46 static unsigned int nmi_hz = HZ;
48 static DEFINE_PER_CPU(short, wd_enabled);
50 /* local prototypes */
51 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
53 static int endflag __initdata = 0;
55 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
56 * the CPU is idle. To make sure the NMI watchdog really ticks on all
57 * CPUs during the test make them busy.
59 static __init void nmi_cpu_busy(void *data)
62 local_irq_enable_in_hardirq();
63 /* Intentionally don't use cpu_relax here. This is
64 to make sure that the performance counter really ticks,
65 even if there is a simulator or similar that catches the
66 pause instruction. On a real HT machine this is fine because
67 all other CPUs are busy with "useless" delay loops and don't
68 care if they get somewhat less cycles. */
74 static int __init check_nmi_watchdog(void)
76 unsigned int *prev_nmi_count;
79 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
82 if (!atomic_read(&nmi_active))
85 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
89 printk(KERN_INFO "Testing NMI watchdog ... ");
91 if (nmi_watchdog == NMI_LOCAL_APIC)
92 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
94 for_each_possible_cpu(cpu)
95 prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
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;
138 /* This needs to happen later in boot so counters are working */
139 late_initcall(check_nmi_watchdog);
141 static int __init setup_nmi_watchdog(char *str)
145 get_option(&str, &nmi);
147 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
154 __setup("nmi_watchdog=", setup_nmi_watchdog);
157 /* Suspend/resume support */
161 static int nmi_pm_active; /* nmi_active before suspend */
163 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
165 /* only CPU0 goes here, other CPUs should be offline */
166 nmi_pm_active = atomic_read(&nmi_active);
167 stop_apic_nmi_watchdog(NULL);
168 BUG_ON(atomic_read(&nmi_active) != 0);
172 static int lapic_nmi_resume(struct sys_device *dev)
174 /* only CPU0 goes here, other CPUs should be offline */
175 if (nmi_pm_active > 0) {
176 setup_apic_nmi_watchdog(NULL);
177 touch_nmi_watchdog();
183 static struct sysdev_class nmi_sysclass = {
184 set_kset_name("lapic_nmi"),
185 .resume = lapic_nmi_resume,
186 .suspend = lapic_nmi_suspend,
189 static struct sys_device device_lapic_nmi = {
191 .cls = &nmi_sysclass,
194 static int __init init_lapic_nmi_sysfs(void)
198 /* should really be a BUG_ON but b/c this is an
199 * init call, it just doesn't work. -dcz
201 if (nmi_watchdog != NMI_LOCAL_APIC)
204 if (atomic_read(&nmi_active) < 0)
207 error = sysdev_class_register(&nmi_sysclass);
209 error = sysdev_register(&device_lapic_nmi);
212 /* must come after the local APIC's device_initcall() */
213 late_initcall(init_lapic_nmi_sysfs);
215 #endif /* CONFIG_PM */
217 static void __acpi_nmi_enable(void *__unused)
219 apic_write_around(APIC_LVT0, APIC_DM_NMI);
223 * Enable timer based NMIs on all CPUs:
225 void acpi_nmi_enable(void)
227 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
228 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
231 static void __acpi_nmi_disable(void *__unused)
233 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
237 * Disable timer based NMIs on all CPUs:
239 void acpi_nmi_disable(void)
241 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
242 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
245 void setup_apic_nmi_watchdog (void *unused)
247 if (__get_cpu_var(wd_enabled))
250 /* cheap hack to support suspend/resume */
251 /* if cpu0 is not active neither should the other cpus */
252 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
255 switch (nmi_watchdog) {
257 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
258 if (lapic_watchdog_init(nmi_hz) < 0) {
259 __get_cpu_var(wd_enabled) = 0;
264 __get_cpu_var(wd_enabled) = 1;
265 atomic_inc(&nmi_active);
269 void stop_apic_nmi_watchdog(void *unused)
271 /* only support LOCAL and IO APICs for now */
272 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
273 (nmi_watchdog != NMI_IO_APIC))
275 if (__get_cpu_var(wd_enabled) == 0)
277 if (nmi_watchdog == NMI_LOCAL_APIC)
278 lapic_watchdog_stop();
279 __get_cpu_var(wd_enabled) = 0;
280 atomic_dec(&nmi_active);
284 * the best way to detect whether a CPU has a 'hard lockup' problem
285 * is to check it's local APIC timer IRQ counts. If they are not
286 * changing then that CPU has some problem.
288 * as these watchdog NMI IRQs are generated on every CPU, we only
289 * have to check the current processor.
291 * since NMIs don't listen to _any_ locks, we have to be extremely
292 * careful not to rely on unsafe variables. The printk might lock
293 * up though, so we have to break up any console locks first ...
294 * [when there will be more tty-related locks, break them up
299 last_irq_sums [NR_CPUS],
300 alert_counter [NR_CPUS];
302 void touch_nmi_watchdog(void)
304 if (nmi_watchdog > 0) {
308 * Just reset the alert counters, (other CPUs might be
309 * spinning on locks we hold):
311 for_each_present_cpu(cpu) {
312 if (alert_counter[cpu])
313 alert_counter[cpu] = 0;
318 * Tickle the softlockup detector too:
320 touch_softlockup_watchdog();
322 EXPORT_SYMBOL(touch_nmi_watchdog);
324 extern void die_nmi(struct pt_regs *, const char *msg);
326 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
330 * Since current_thread_info()-> is always on the stack, and we
331 * always switch the stack NMI-atomically, it's safe to use
332 * smp_processor_id().
336 int cpu = smp_processor_id();
339 /* check for other users first */
340 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
346 if (cpu_isset(cpu, backtrace_mask)) {
347 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
350 printk("NMI backtrace for cpu %d\n", cpu);
353 cpu_clear(cpu, backtrace_mask);
357 * Take the local apic timer and PIT/HPET into account. We don't
358 * know which one is active, when we have highres/dyntick on
360 sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
361 per_cpu(irq_stat, cpu).irq0_irqs;
363 /* if the none of the timers isn't firing, this cpu isn't doing much */
364 if (!touched && last_irq_sums[cpu] == sum) {
366 * Ayiee, looks like this CPU is stuck ...
367 * wait a few IRQs (5 seconds) before doing the oops ...
369 alert_counter[cpu]++;
370 if (alert_counter[cpu] == 5*nmi_hz)
372 * die_nmi will return ONLY if NOTIFY_STOP happens..
374 die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
376 last_irq_sums[cpu] = sum;
377 alert_counter[cpu] = 0;
379 /* see if the nmi watchdog went off */
380 if (!__get_cpu_var(wd_enabled))
382 switch (nmi_watchdog) {
384 rc |= lapic_wd_event(nmi_hz);
387 /* don't know how to accurately check for this.
388 * just assume it was a watchdog timer interrupt
389 * This matches the old behaviour.
397 int do_nmi_callback(struct pt_regs * regs, int cpu)
400 if (unknown_nmi_panic)
401 return unknown_nmi_panic_callback(regs, cpu);
408 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
410 unsigned char reason = get_nmi_reason();
413 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
419 * proc handler for /proc/sys/kernel/nmi
421 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
422 void __user *buffer, size_t *length, loff_t *ppos)
426 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
427 old_state = nmi_watchdog_enabled;
428 proc_dointvec(table, write, file, buffer, length, ppos);
429 if (!!old_state == !!nmi_watchdog_enabled)
432 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
433 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
437 if (nmi_watchdog == NMI_DEFAULT) {
438 if (lapic_watchdog_ok())
439 nmi_watchdog = NMI_LOCAL_APIC;
441 nmi_watchdog = NMI_IO_APIC;
444 if (nmi_watchdog == NMI_LOCAL_APIC) {
445 if (nmi_watchdog_enabled)
446 enable_lapic_nmi_watchdog();
448 disable_lapic_nmi_watchdog();
451 "NMI watchdog doesn't know what hardware to touch\n");
459 void __trigger_all_cpu_backtrace(void)
463 backtrace_mask = cpu_online_map;
464 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
465 for (i = 0; i < 10 * 1000; i++) {
466 if (cpus_empty(backtrace_mask))
472 EXPORT_SYMBOL(nmi_active);
473 EXPORT_SYMBOL(nmi_watchdog);