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 */
45 static int panic_on_timeout;
47 unsigned int nmi_watchdog = NMI_DEFAULT;
48 static unsigned int nmi_hz = HZ;
50 static DEFINE_PER_CPU(short, wd_enabled);
52 static int endflag __initdata = 0;
54 static inline unsigned int get_nmi_count(int cpu)
56 return nmi_count(cpu);
59 static inline int mce_in_progress(void)
65 * Take the local apic timer and PIT/HPET into account. We don't
66 * know which one is active, when we have highres/dyntick on
68 static inline unsigned int get_timer_irqs(int cpu)
70 return per_cpu(irq_stat, cpu).apic_timer_irqs +
71 per_cpu(irq_stat, cpu).irq0_irqs;
74 /* Run after command line and cpu_init init, but before all other checks */
75 void nmi_watchdog_default(void)
77 if (nmi_watchdog != NMI_DEFAULT)
79 if (lapic_watchdog_ok())
80 nmi_watchdog = NMI_LOCAL_APIC;
82 nmi_watchdog = NMI_IO_APIC;
86 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
87 * the CPU is idle. To make sure the NMI watchdog really ticks on all
88 * CPUs during the test make them busy.
90 static __init void nmi_cpu_busy(void *data)
92 local_irq_enable_in_hardirq();
93 /* Intentionally don't use cpu_relax here. This is
94 to make sure that the performance counter really ticks,
95 even if there is a simulator or similar that catches the
96 pause instruction. On a real HT machine this is fine because
97 all other CPUs are busy with "useless" delay loops and don't
98 care if they get somewhat less cycles. */
104 int __init check_nmi_watchdog(void)
106 unsigned int *prev_nmi_count;
109 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
112 if (!atomic_read(&nmi_active))
115 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
119 printk(KERN_INFO "Testing NMI watchdog ... ");
122 if (nmi_watchdog == NMI_LOCAL_APIC)
123 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
126 for_each_possible_cpu(cpu)
127 prev_nmi_count[cpu] = get_nmi_count(cpu);
129 mdelay((20*1000)/nmi_hz); // wait 20 ticks
131 for_each_online_cpu(cpu) {
132 if (!per_cpu(wd_enabled, cpu))
134 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
135 printk(KERN_WARNING "WARNING: CPU#%d: NMI "
136 "appears to be stuck (%d->%d)!\n",
140 per_cpu(wd_enabled, cpu) = 0;
141 atomic_dec(&nmi_active);
145 if (!atomic_read(&nmi_active)) {
146 kfree(prev_nmi_count);
147 atomic_set(&nmi_active, -1);
152 /* now that we know it works we can reduce NMI frequency to
153 something more reasonable; makes a difference in some configs */
154 if (nmi_watchdog == NMI_LOCAL_APIC)
155 nmi_hz = lapic_adjust_nmi_hz(1);
157 kfree(prev_nmi_count);
160 timer_ack = !cpu_has_tsc;
165 static int __init setup_nmi_watchdog(char *str)
169 if (!strncmp(str, "panic", 5)) {
170 panic_on_timeout = 1;
171 str = strchr(str, ',');
177 get_option(&str, &nmi);
179 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
186 __setup("nmi_watchdog=", setup_nmi_watchdog);
189 /* Suspend/resume support */
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();
215 static struct sysdev_class nmi_sysclass = {
217 .resume = lapic_nmi_resume,
218 .suspend = lapic_nmi_suspend,
221 static struct sys_device device_lapic_nmi = {
223 .cls = &nmi_sysclass,
226 static int __init init_lapic_nmi_sysfs(void)
230 /* should really be a BUG_ON but b/c this is an
231 * init call, it just doesn't work. -dcz
233 if (nmi_watchdog != NMI_LOCAL_APIC)
236 if (atomic_read(&nmi_active) < 0)
239 error = sysdev_class_register(&nmi_sysclass);
241 error = sysdev_register(&device_lapic_nmi);
244 /* must come after the local APIC's device_initcall() */
245 late_initcall(init_lapic_nmi_sysfs);
247 #endif /* CONFIG_PM */
249 static void __acpi_nmi_enable(void *__unused)
251 apic_write_around(APIC_LVT0, APIC_DM_NMI);
255 * Enable timer based NMIs on all CPUs:
257 void acpi_nmi_enable(void)
259 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
260 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
263 static void __acpi_nmi_disable(void *__unused)
265 apic_write_around(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
269 * Disable timer based NMIs on all CPUs:
271 void acpi_nmi_disable(void)
273 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
274 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
277 void setup_apic_nmi_watchdog(void *unused)
279 if (__get_cpu_var(wd_enabled))
282 /* cheap hack to support suspend/resume */
283 /* if cpu0 is not active neither should the other cpus */
284 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
287 switch (nmi_watchdog) {
289 __get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
290 if (lapic_watchdog_init(nmi_hz) < 0) {
291 __get_cpu_var(wd_enabled) = 0;
296 __get_cpu_var(wd_enabled) = 1;
297 atomic_inc(&nmi_active);
301 void stop_apic_nmi_watchdog(void *unused)
303 /* only support LOCAL and IO APICs for now */
304 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
305 (nmi_watchdog != NMI_IO_APIC))
307 if (__get_cpu_var(wd_enabled) == 0)
309 if (nmi_watchdog == NMI_LOCAL_APIC)
310 lapic_watchdog_stop();
311 __get_cpu_var(wd_enabled) = 0;
312 atomic_dec(&nmi_active);
316 * the best way to detect whether a CPU has a 'hard lockup' problem
317 * is to check it's local APIC timer IRQ counts. If they are not
318 * changing then that CPU has some problem.
320 * as these watchdog NMI IRQs are generated on every CPU, we only
321 * have to check the current processor.
323 * since NMIs don't listen to _any_ locks, we have to be extremely
324 * careful not to rely on unsafe variables. The printk might lock
325 * up though, so we have to break up any console locks first ...
326 * [when there will be more tty-related locks, break them up
330 static DEFINE_PER_CPU(unsigned, last_irq_sum);
331 static DEFINE_PER_CPU(local_t, alert_counter);
332 static DEFINE_PER_CPU(int, nmi_touch);
334 void touch_nmi_watchdog(void)
336 if (nmi_watchdog > 0) {
340 * Tell other CPUs to reset their alert counters. We cannot
341 * do it ourselves because the alert count increase is not
344 for_each_present_cpu(cpu) {
345 if (per_cpu(nmi_touch, cpu) != 1)
346 per_cpu(nmi_touch, cpu) = 1;
351 * Tickle the softlockup detector too:
353 touch_softlockup_watchdog();
355 EXPORT_SYMBOL(touch_nmi_watchdog);
357 notrace __kprobes int
358 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
362 * Since current_thread_info()-> is always on the stack, and we
363 * always switch the stack NMI-atomically, it's safe to use
364 * smp_processor_id().
368 int cpu = smp_processor_id();
371 /* check for other users first */
372 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
378 sum = get_timer_irqs(cpu);
380 if (__get_cpu_var(nmi_touch)) {
381 __get_cpu_var(nmi_touch) = 0;
385 if (cpu_isset(cpu, backtrace_mask)) {
386 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
389 printk("NMI backtrace for cpu %d\n", cpu);
392 cpu_clear(cpu, backtrace_mask);
395 /* Could check oops_in_progress here too, but it's safer not to */
396 if (mce_in_progress())
399 /* if the none of the timers isn't firing, this cpu isn't doing much */
400 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
402 * Ayiee, looks like this CPU is stuck ...
403 * wait a few IRQs (5 seconds) before doing the oops ...
405 local_inc(&__get_cpu_var(alert_counter));
406 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
408 * die_nmi will return ONLY if NOTIFY_STOP happens..
410 die_nmi("BUG: NMI Watchdog detected LOCKUP",
411 regs, panic_on_timeout);
413 __get_cpu_var(last_irq_sum) = sum;
414 local_set(&__get_cpu_var(alert_counter), 0);
416 /* see if the nmi watchdog went off */
417 if (!__get_cpu_var(wd_enabled))
419 switch (nmi_watchdog) {
421 rc |= lapic_wd_event(nmi_hz);
424 /* don't know how to accurately check for this.
425 * just assume it was a watchdog timer interrupt
426 * This matches the old behaviour.
436 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
438 unsigned char reason = get_nmi_reason();
441 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
442 die_nmi(buf, regs, 1); /* Always panic here */
447 * proc handler for /proc/sys/kernel/nmi
449 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
450 void __user *buffer, size_t *length, loff_t *ppos)
454 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
455 old_state = nmi_watchdog_enabled;
456 proc_dointvec(table, write, file, buffer, length, ppos);
457 if (!!old_state == !!nmi_watchdog_enabled)
460 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
461 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
465 /* if nmi_watchdog is not set yet, then set it */
466 nmi_watchdog_default();
468 if (nmi_watchdog == NMI_LOCAL_APIC) {
469 if (nmi_watchdog_enabled)
470 enable_lapic_nmi_watchdog();
472 disable_lapic_nmi_watchdog();
475 "NMI watchdog doesn't know what hardware to touch\n");
483 int do_nmi_callback(struct pt_regs *regs, int cpu)
486 if (unknown_nmi_panic)
487 return unknown_nmi_panic_callback(regs, cpu);
492 void __trigger_all_cpu_backtrace(void)
496 backtrace_mask = cpu_online_map;
497 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
498 for (i = 0; i < 10 * 1000; i++) {
499 if (cpus_empty(backtrace_mask))
505 EXPORT_SYMBOL(nmi_active);
506 EXPORT_SYMBOL(nmi_watchdog);