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.
10 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
13 #include <linux/nmi.h>
15 #include <linux/delay.h>
16 #include <linux/interrupt.h>
17 #include <linux/module.h>
18 #include <linux/sysdev.h>
19 #include <linux/sysctl.h>
20 #include <linux/kprobes.h>
21 #include <linux/cpumask.h>
22 #include <linux/kdebug.h>
26 #include <asm/proto.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 */
43 static int panic_on_timeout;
45 unsigned int nmi_watchdog = NMI_DEFAULT;
46 static unsigned int nmi_hz = HZ;
48 static DEFINE_PER_CPU(short, wd_enabled);
50 /* Run after command line and cpu_init init, but before all other checks */
51 void nmi_watchdog_default(void)
53 if (nmi_watchdog != NMI_DEFAULT)
55 nmi_watchdog = NMI_NONE;
58 static int endflag __initdata = 0;
61 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
62 * the CPU is idle. To make sure the NMI watchdog really ticks on all
63 * CPUs during the test make them busy.
65 static __init void nmi_cpu_busy(void *data)
67 local_irq_enable_in_hardirq();
68 /* Intentionally don't use cpu_relax here. This is
69 to make sure that the performance counter really ticks,
70 even if there is a simulator or similar that catches the
71 pause instruction. On a real HT machine this is fine because
72 all other CPUs are busy with "useless" delay loops and don't
73 care if they get somewhat less cycles. */
79 int __init check_nmi_watchdog(void)
81 unsigned int *prev_nmi_count;
84 if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
87 if (!atomic_read(&nmi_active))
90 prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
94 printk(KERN_INFO "Testing NMI watchdog ... ");
97 if (nmi_watchdog == NMI_LOCAL_APIC)
98 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
101 for (cpu = 0; cpu < NR_CPUS; cpu++)
102 prev_nmi_count[cpu] = cpu_pda(cpu)->__nmi_count;
104 mdelay((20*1000)/nmi_hz); // wait 20 ticks
106 for_each_online_cpu(cpu) {
107 if (!per_cpu(wd_enabled, cpu))
109 if (cpu_pda(cpu)->__nmi_count - prev_nmi_count[cpu] <= 5) {
110 printk(KERN_WARNING "WARNING: CPU#%d: NMI "
111 "appears to be stuck (%d->%d)!\n",
114 cpu_pda(cpu)->__nmi_count);
115 per_cpu(wd_enabled, cpu) = 0;
116 atomic_dec(&nmi_active);
120 if (!atomic_read(&nmi_active)) {
121 kfree(prev_nmi_count);
122 atomic_set(&nmi_active, -1);
127 /* now that we know it works we can reduce NMI frequency to
128 something more reasonable; makes a difference in some configs */
129 if (nmi_watchdog == NMI_LOCAL_APIC)
130 nmi_hz = lapic_adjust_nmi_hz(1);
132 kfree(prev_nmi_count);
136 static int __init setup_nmi_watchdog(char *str)
140 if (!strncmp(str,"panic",5)) {
141 panic_on_timeout = 1;
142 str = strchr(str, ',');
148 get_option(&str, &nmi);
150 if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
157 __setup("nmi_watchdog=", setup_nmi_watchdog);
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();
182 static struct sysdev_class nmi_sysclass = {
184 .resume = lapic_nmi_resume,
185 .suspend = lapic_nmi_suspend,
188 static struct sys_device device_lapic_nmi = {
190 .cls = &nmi_sysclass,
193 static int __init init_lapic_nmi_sysfs(void)
197 /* should really be a BUG_ON but b/c this is an
198 * init call, it just doesn't work. -dcz
200 if (nmi_watchdog != NMI_LOCAL_APIC)
203 if (atomic_read(&nmi_active) < 0)
206 error = sysdev_class_register(&nmi_sysclass);
208 error = sysdev_register(&device_lapic_nmi);
211 /* must come after the local APIC's device_initcall() */
212 late_initcall(init_lapic_nmi_sysfs);
214 #endif /* CONFIG_PM */
216 static void __acpi_nmi_enable(void *__unused)
218 apic_write(APIC_LVT0, APIC_DM_NMI);
222 * Enable timer based NMIs on all CPUs:
224 void acpi_nmi_enable(void)
226 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
227 on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
230 static void __acpi_nmi_disable(void *__unused)
232 apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
236 * Disable timer based NMIs on all CPUs:
238 void acpi_nmi_disable(void)
240 if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
241 on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
244 void setup_apic_nmi_watchdog(void *unused)
246 if (__get_cpu_var(wd_enabled))
249 /* cheap hack to support suspend/resume */
250 /* if cpu0 is not active neither should the other cpus */
251 if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
254 switch (nmi_watchdog) {
256 __get_cpu_var(wd_enabled) = 1;
257 if (lapic_watchdog_init(nmi_hz) < 0) {
258 __get_cpu_var(wd_enabled) = 0;
263 __get_cpu_var(wd_enabled) = 1;
264 atomic_inc(&nmi_active);
268 void stop_apic_nmi_watchdog(void *unused)
270 /* only support LOCAL and IO APICs for now */
271 if ((nmi_watchdog != NMI_LOCAL_APIC) &&
272 (nmi_watchdog != NMI_IO_APIC))
274 if (__get_cpu_var(wd_enabled) == 0)
276 if (nmi_watchdog == NMI_LOCAL_APIC)
277 lapic_watchdog_stop();
278 __get_cpu_var(wd_enabled) = 0;
279 atomic_dec(&nmi_active);
283 * the best way to detect whether a CPU has a 'hard lockup' problem
284 * is to check it's local APIC timer IRQ counts. If they are not
285 * changing then that CPU has some problem.
287 * as these watchdog NMI IRQs are generated on every CPU, we only
288 * have to check the current processor.
291 static DEFINE_PER_CPU(unsigned, last_irq_sum);
292 static DEFINE_PER_CPU(local_t, alert_counter);
293 static DEFINE_PER_CPU(int, nmi_touch);
295 void touch_nmi_watchdog(void)
297 if (nmi_watchdog > 0) {
301 * Tell other CPUs to reset their alert counters. We cannot
302 * do it ourselves because the alert count increase is not
305 for_each_present_cpu(cpu) {
306 if (per_cpu(nmi_touch, cpu) != 1)
307 per_cpu(nmi_touch, cpu) = 1;
311 touch_softlockup_watchdog();
313 EXPORT_SYMBOL(touch_nmi_watchdog);
315 notrace __kprobes int
316 nmi_watchdog_tick(struct pt_regs *regs, unsigned reason)
320 int cpu = smp_processor_id();
323 /* check for other users first */
324 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
330 sum = read_pda(apic_timer_irqs) + read_pda(irq0_irqs);
331 if (__get_cpu_var(nmi_touch)) {
332 __get_cpu_var(nmi_touch) = 0;
336 if (cpu_isset(cpu, backtrace_mask)) {
337 static DEFINE_SPINLOCK(lock); /* Serialise the printks */
340 printk("NMI backtrace for cpu %d\n", cpu);
343 cpu_clear(cpu, backtrace_mask);
346 #ifdef CONFIG_X86_MCE
347 /* Could check oops_in_progress here too, but it's safer
349 if (atomic_read(&mce_entry) > 0)
352 /* if the apic timer isn't firing, this cpu isn't doing much */
353 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
355 * Ayiee, looks like this CPU is stuck ...
356 * wait a few IRQs (5 seconds) before doing the oops ...
358 local_inc(&__get_cpu_var(alert_counter));
359 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz)
360 die_nmi("NMI Watchdog detected LOCKUP",
361 regs, panic_on_timeout);
363 __get_cpu_var(last_irq_sum) = sum;
364 local_set(&__get_cpu_var(alert_counter), 0);
367 /* see if the nmi watchdog went off */
368 if (!__get_cpu_var(wd_enabled))
370 switch (nmi_watchdog) {
372 rc |= lapic_wd_event(nmi_hz);
375 /* don't know how to accurately check for this.
376 * just assume it was a watchdog timer interrupt
377 * This matches the old behaviour.
387 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
389 unsigned char reason = get_nmi_reason();
392 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
393 die_nmi(buf, regs, 1); /* Always panic here */
398 * proc handler for /proc/sys/kernel/nmi
400 int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
401 void __user *buffer, size_t *length, loff_t *ppos)
405 nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
406 old_state = nmi_watchdog_enabled;
407 proc_dointvec(table, write, file, buffer, length, ppos);
408 if (!!old_state == !!nmi_watchdog_enabled)
411 if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
412 printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
416 /* if nmi_watchdog is not set yet, then set it */
417 nmi_watchdog_default();
419 if (nmi_watchdog == NMI_LOCAL_APIC) {
420 if (nmi_watchdog_enabled)
421 enable_lapic_nmi_watchdog();
423 disable_lapic_nmi_watchdog();
426 "NMI watchdog doesn't know what hardware to touch\n");
434 int do_nmi_callback(struct pt_regs *regs, int cpu)
437 if (unknown_nmi_panic)
438 return unknown_nmi_panic_callback(regs, cpu);
443 void __trigger_all_cpu_backtrace(void)
447 backtrace_mask = cpu_online_map;
448 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
449 for (i = 0; i < 10 * 1000; i++) {
450 if (cpus_empty(backtrace_mask))
456 EXPORT_SYMBOL(nmi_active);
457 EXPORT_SYMBOL(nmi_watchdog);