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>
29 #include "mach_traps.h"
31 int unknown_nmi_panic;
32 int nmi_watchdog_enabled;
34 static cpumask_t backtrace_mask = CPU_MASK_NONE;
37 * >0: the lapic NMI watchdog is active, but can be disabled
38 * <0: the lapic NMI watchdog has not been set up, and cannot
40 * 0: the lapic NMI watchdog is disabled, but can be enabled
42 atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
44 unsigned int nmi_watchdog = NMI_DEFAULT;
45 static unsigned int nmi_hz = HZ;
47 static DEFINE_PER_CPU(short, wd_enabled);
49 /* local prototypes */
50 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
52 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)
61 local_irq_enable_in_hardirq();
62 /* Intentionally don't use cpu_relax here. This is
63 to make sure that the performance counter really ticks,
64 even if there is a simulator or similar that catches the
65 pause instruction. On a real HT machine this is fine because
66 all other CPUs are busy with "useless" delay loops and don't
67 care if they get somewhat less cycles. */
73 static int __init check_nmi_watchdog(void)
75 unsigned int *prev_nmi_count;
78 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
81 if (!atomic_read(&nmi_active))
84 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
88 printk(KERN_INFO "Testing NMI watchdog ... ");
90 if (nmi_watchdog == NMI_LOCAL_APIC)
91 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
93 for_each_possible_cpu(cpu)
94 prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
96 mdelay((20*1000)/nmi_hz); // wait 20 ticks
98 for_each_possible_cpu(cpu) {
100 /* Check cpu_callin_map here because that is set
101 after the timer is started. */
102 if (!cpu_isset(cpu, cpu_callin_map))
105 if (!per_cpu(wd_enabled, cpu))
107 if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
108 printk(KERN_WARNING "WARNING: CPU#%d: NMI "
109 "appears to be stuck (%d->%d)!\n",
113 per_cpu(wd_enabled, cpu) = 0;
114 atomic_dec(&nmi_active);
118 if (!atomic_read(&nmi_active)) {
119 kfree(prev_nmi_count);
120 atomic_set(&nmi_active, -1);
125 /* now that we know it works we can reduce NMI frequency to
126 something more reasonable; makes a difference in some configs */
127 if (nmi_watchdog == NMI_LOCAL_APIC)
128 nmi_hz = lapic_adjust_nmi_hz(1);
130 kfree(prev_nmi_count);
133 /* This needs to happen later in boot so counters are working */
134 late_initcall(check_nmi_watchdog);
136 static int __init setup_nmi_watchdog(char *str)
140 get_option(&str, &nmi);
142 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
149 __setup("nmi_watchdog=", setup_nmi_watchdog);
152 /* Suspend/resume support */
156 static int nmi_pm_active; /* nmi_active before suspend */
158 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
160 /* only CPU0 goes here, other CPUs should be offline */
161 nmi_pm_active = atomic_read(&nmi_active);
162 stop_apic_nmi_watchdog(NULL);
163 BUG_ON(atomic_read(&nmi_active) != 0);
167 static int lapic_nmi_resume(struct sys_device *dev)
169 /* only CPU0 goes here, other CPUs should be offline */
170 if (nmi_pm_active > 0) {
171 setup_apic_nmi_watchdog(NULL);
172 touch_nmi_watchdog();
178 static struct sysdev_class nmi_sysclass = {
179 set_kset_name("lapic_nmi"),
180 .resume = lapic_nmi_resume,
181 .suspend = lapic_nmi_suspend,
184 static struct sys_device device_lapic_nmi = {
186 .cls = &nmi_sysclass,
189 static int __init init_lapic_nmi_sysfs(void)
193 /* should really be a BUG_ON but b/c this is an
194 * init call, it just doesn't work. -dcz
196 if (nmi_watchdog != NMI_LOCAL_APIC)
199 if (atomic_read(&nmi_active) < 0)
202 error = sysdev_class_register(&nmi_sysclass);
204 error = sysdev_register(&device_lapic_nmi);
207 /* must come after the local APIC's device_initcall() */
208 late_initcall(init_lapic_nmi_sysfs);
210 #endif /* CONFIG_PM */
212 static void __acpi_nmi_enable(void *__unused)
214 apic_write_around(APIC_LVT0, APIC_DM_NMI);
218 * Enable timer based NMIs on all CPUs:
220 void acpi_nmi_enable(void)
222 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
223 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
226 static void __acpi_nmi_disable(void *__unused)
228 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
232 * Disable timer based NMIs on all CPUs:
234 void acpi_nmi_disable(void)
236 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
237 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
240 void setup_apic_nmi_watchdog (void *unused)
242 if (__get_cpu_var(wd_enabled))
245 /* cheap hack to support suspend/resume */
246 /* if cpu0 is not active neither should the other cpus */
247 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
250 switch (nmi_watchdog) {
252 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
253 if (lapic_watchdog_init(nmi_hz) < 0) {
254 __get_cpu_var(wd_enabled) = 0;
259 __get_cpu_var(wd_enabled) = 1;
260 atomic_inc(&nmi_active);
264 void stop_apic_nmi_watchdog(void *unused)
266 /* only support LOCAL and IO APICs for now */
267 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
268 (nmi_watchdog != NMI_IO_APIC))
270 if (__get_cpu_var(wd_enabled) == 0)
272 if (nmi_watchdog == NMI_LOCAL_APIC)
273 lapic_watchdog_stop();
274 __get_cpu_var(wd_enabled) = 0;
275 atomic_dec(&nmi_active);
279 * the best way to detect whether a CPU has a 'hard lockup' problem
280 * is to check it's local APIC timer IRQ counts. If they are not
281 * changing then that CPU has some problem.
283 * as these watchdog NMI IRQs are generated on every CPU, we only
284 * have to check the current processor.
286 * since NMIs don't listen to _any_ locks, we have to be extremely
287 * careful not to rely on unsafe variables. The printk might lock
288 * up though, so we have to break up any console locks first ...
289 * [when there will be more tty-related locks, break them up
294 last_irq_sums [NR_CPUS],
295 alert_counter [NR_CPUS];
297 void touch_nmi_watchdog(void)
299 if (nmi_watchdog > 0) {
303 * Just reset the alert counters, (other CPUs might be
304 * spinning on locks we hold):
306 for_each_present_cpu(cpu) {
307 if (alert_counter[cpu])
308 alert_counter[cpu] = 0;
313 * Tickle the softlockup detector too:
315 touch_softlockup_watchdog();
317 EXPORT_SYMBOL(touch_nmi_watchdog);
319 extern void die_nmi(struct pt_regs *, const char *msg);
321 __kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
325 * Since current_thread_info()-> is always on the stack, and we
326 * always switch the stack NMI-atomically, it's safe to use
327 * smp_processor_id().
331 int cpu = smp_processor_id();
334 /* check for other users first */
335 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
341 if (cpu_isset(cpu, backtrace_mask)) {
342 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
345 printk("NMI backtrace for cpu %d\n", cpu);
348 cpu_clear(cpu, backtrace_mask);
352 * Take the local apic timer and PIT/HPET into account. We don't
353 * know which one is active, when we have highres/dyntick on
355 sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
356 per_cpu(irq_stat, cpu).irq0_irqs;
358 /* if the none of the timers isn't firing, this cpu isn't doing much */
359 if (!touched && last_irq_sums[cpu] == sum) {
361 * Ayiee, looks like this CPU is stuck ...
362 * wait a few IRQs (5 seconds) before doing the oops ...
364 alert_counter[cpu]++;
365 if (alert_counter[cpu] == 5*nmi_hz)
367 * die_nmi will return ONLY if NOTIFY_STOP happens..
369 die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
371 last_irq_sums[cpu] = sum;
372 alert_counter[cpu] = 0;
374 /* see if the nmi watchdog went off */
375 if (!__get_cpu_var(wd_enabled))
377 switch (nmi_watchdog) {
379 rc |= lapic_wd_event(nmi_hz);
382 /* don't know how to accurately check for this.
383 * just assume it was a watchdog timer interrupt
384 * This matches the old behaviour.
392 int do_nmi_callback(struct pt_regs * regs, int cpu)
395 if (unknown_nmi_panic)
396 return unknown_nmi_panic_callback(regs, cpu);
403 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
405 unsigned char reason = get_nmi_reason();
408 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
414 * proc handler for /proc/sys/kernel/nmi
416 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
417 void __user *buffer, size_t *length, loff_t *ppos)
421 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
422 old_state = nmi_watchdog_enabled;
423 proc_dointvec(table, write, file, buffer, length, ppos);
424 if (!!old_state == !!nmi_watchdog_enabled)
427 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
428 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
432 if (nmi_watchdog == NMI_DEFAULT) {
433 if (lapic_watchdog_ok())
434 nmi_watchdog = NMI_LOCAL_APIC;
436 nmi_watchdog = NMI_IO_APIC;
439 if (nmi_watchdog == NMI_LOCAL_APIC) {
440 if (nmi_watchdog_enabled)
441 enable_lapic_nmi_watchdog();
443 disable_lapic_nmi_watchdog();
446 "NMI watchdog doesn't know what hardware to touch\n");
454 void __trigger_all_cpu_backtrace(void)
458 backtrace_mask = cpu_online_map;
459 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
460 for (i = 0; i < 10 * 1000; i++) {
461 if (cpus_empty(backtrace_mask))
467 EXPORT_SYMBOL(nmi_active);
468 EXPORT_SYMBOL(nmi_watchdog);