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.
16 #include <linux/nmi.h>
18 #include <linux/delay.h>
19 #include <linux/interrupt.h>
20 #include <linux/module.h>
21 #include <linux/sysdev.h>
22 #include <linux/sysctl.h>
23 #include <linux/percpu.h>
24 #include <linux/kprobes.h>
25 #include <linux/cpumask.h>
26 #include <linux/kernel_stat.h>
27 #include <linux/kdebug.h>
28 #include <linux/smp.h>
30 #include <asm/i8259.h>
31 #include <asm/io_apic.h>
34 #include <asm/proto.h>
35 #include <asm/timer.h>
39 #include <mach_traps.h>
41 int unknown_nmi_panic
;
42 int nmi_watchdog_enabled
;
44 static cpumask_t backtrace_mask
= CPU_MASK_NONE
;
47 * >0: the lapic NMI watchdog is active, but can be disabled
48 * <0: the lapic NMI watchdog has not been set up, and cannot
50 * 0: the lapic NMI watchdog is disabled, but can be enabled
52 atomic_t nmi_active
= ATOMIC_INIT(0); /* oprofile uses this */
53 EXPORT_SYMBOL(nmi_active
);
55 unsigned int nmi_watchdog
= NMI_NONE
;
56 EXPORT_SYMBOL(nmi_watchdog
);
58 static int panic_on_timeout
;
60 static unsigned int nmi_hz
= HZ
;
61 static DEFINE_PER_CPU(short, wd_enabled
);
62 static int endflag __initdata
;
64 static inline unsigned int get_nmi_count(int cpu
)
67 return cpu_pda(cpu
)->__nmi_count
;
69 return nmi_count(cpu
);
73 static inline int mce_in_progress(void)
75 #if defined(CONFIG_X86_64) && defined(CONFIG_X86_MCE)
76 return atomic_read(&mce_entry
) > 0;
82 * Take the local apic timer and PIT/HPET into account. We don't
83 * know which one is active, when we have highres/dyntick on
85 static inline unsigned int get_timer_irqs(int cpu
)
88 return read_pda(apic_timer_irqs
) + read_pda(irq0_irqs
);
90 return per_cpu(irq_stat
, cpu
).apic_timer_irqs
+
91 per_cpu(irq_stat
, cpu
).irq0_irqs
;
97 * The performance counters used by NMI_LOCAL_APIC don't trigger when
98 * the CPU is idle. To make sure the NMI watchdog really ticks on all
99 * CPUs during the test make them busy.
101 static __init
void nmi_cpu_busy(void *data
)
103 local_irq_enable_in_hardirq();
105 * Intentionally don't use cpu_relax here. This is
106 * to make sure that the performance counter really ticks,
107 * even if there is a simulator or similar that catches the
108 * pause instruction. On a real HT machine this is fine because
109 * all other CPUs are busy with "useless" delay loops and don't
110 * care if they get somewhat less cycles.
117 int __init
check_nmi_watchdog(void)
119 unsigned int *prev_nmi_count
;
122 if (!nmi_watchdog_active() || !atomic_read(&nmi_active
))
125 prev_nmi_count
= kmalloc(nr_cpu_ids
* sizeof(int), GFP_KERNEL
);
129 printk(KERN_INFO
"Testing NMI watchdog ... ");
132 if (nmi_watchdog
== NMI_LOCAL_APIC
)
133 smp_call_function(nmi_cpu_busy
, (void *)&endflag
, 0, 0);
136 for_each_possible_cpu(cpu
)
137 prev_nmi_count
[cpu
] = get_nmi_count(cpu
);
139 mdelay((20 * 1000) / nmi_hz
); /* wait 20 ticks */
141 for_each_online_cpu(cpu
) {
142 if (!per_cpu(wd_enabled
, cpu
))
144 if (get_nmi_count(cpu
) - prev_nmi_count
[cpu
] <= 5) {
145 printk(KERN_WARNING
"WARNING: CPU#%d: NMI "
146 "appears to be stuck (%d->%d)!\n",
150 per_cpu(wd_enabled
, cpu
) = 0;
151 atomic_dec(&nmi_active
);
155 if (!atomic_read(&nmi_active
)) {
156 kfree(prev_nmi_count
);
157 atomic_set(&nmi_active
, -1);
163 * now that we know it works we can reduce NMI frequency to
164 * something more reasonable; makes a difference in some configs
166 if (nmi_watchdog
== NMI_LOCAL_APIC
)
167 nmi_hz
= lapic_adjust_nmi_hz(1);
169 kfree(prev_nmi_count
);
172 if (nmi_watchdog
== NMI_IO_APIC
&& !timer_through_8259
)
173 disable_8259A_irq(0);
177 static int __init
setup_nmi_watchdog(char *str
)
181 if (!strncmp(str
, "panic", 5)) {
182 panic_on_timeout
= 1;
183 str
= strchr(str
, ',');
189 get_option(&str
, &nmi
);
191 if (nmi
>= NMI_INVALID
)
197 __setup("nmi_watchdog=", setup_nmi_watchdog
);
200 * Suspend/resume support
204 static int nmi_pm_active
; /* nmi_active before suspend */
206 static int lapic_nmi_suspend(struct sys_device
*dev
, pm_message_t state
)
208 /* only CPU0 goes here, other CPUs should be offline */
209 nmi_pm_active
= atomic_read(&nmi_active
);
210 stop_apic_nmi_watchdog(NULL
);
211 BUG_ON(atomic_read(&nmi_active
) != 0);
215 static int lapic_nmi_resume(struct sys_device
*dev
)
217 /* only CPU0 goes here, other CPUs should be offline */
218 if (nmi_pm_active
> 0) {
219 setup_apic_nmi_watchdog(NULL
);
220 touch_nmi_watchdog();
225 static struct sysdev_class nmi_sysclass
= {
227 .resume
= lapic_nmi_resume
,
228 .suspend
= lapic_nmi_suspend
,
231 static struct sys_device device_lapic_nmi
= {
233 .cls
= &nmi_sysclass
,
236 static int __init
init_lapic_nmi_sysfs(void)
241 * should really be a BUG_ON but b/c this is an
242 * init call, it just doesn't work. -dcz
244 if (nmi_watchdog
!= NMI_LOCAL_APIC
)
247 if (atomic_read(&nmi_active
) < 0)
250 error
= sysdev_class_register(&nmi_sysclass
);
252 error
= sysdev_register(&device_lapic_nmi
);
256 /* must come after the local APIC's device_initcall() */
257 late_initcall(init_lapic_nmi_sysfs
);
259 #endif /* CONFIG_PM */
261 static void __acpi_nmi_enable(void *__unused
)
263 apic_write_around(APIC_LVT0
, APIC_DM_NMI
);
267 * Enable timer based NMIs on all CPUs:
269 void acpi_nmi_enable(void)
271 if (atomic_read(&nmi_active
) && nmi_watchdog
== NMI_IO_APIC
)
272 on_each_cpu(__acpi_nmi_enable
, NULL
, 0, 1);
275 static void __acpi_nmi_disable(void *__unused
)
277 apic_write_around(APIC_LVT0
, APIC_DM_NMI
| APIC_LVT_MASKED
);
281 * Disable timer based NMIs on all CPUs:
283 void acpi_nmi_disable(void)
285 if (atomic_read(&nmi_active
) && nmi_watchdog
== NMI_IO_APIC
)
286 on_each_cpu(__acpi_nmi_disable
, NULL
, 0, 1);
289 void setup_apic_nmi_watchdog(void *unused
)
291 if (__get_cpu_var(wd_enabled
))
294 /* cheap hack to support suspend/resume */
295 /* if cpu0 is not active neither should the other cpus */
296 if (smp_processor_id() != 0 && atomic_read(&nmi_active
) <= 0)
299 switch (nmi_watchdog
) {
301 /* enable it before to avoid race with handler */
302 __get_cpu_var(wd_enabled
) = 1;
303 if (lapic_watchdog_init(nmi_hz
) < 0) {
304 __get_cpu_var(wd_enabled
) = 0;
309 __get_cpu_var(wd_enabled
) = 1;
310 atomic_inc(&nmi_active
);
314 void stop_apic_nmi_watchdog(void *unused
)
316 /* only support LOCAL and IO APICs for now */
317 if (!nmi_watchdog_active())
319 if (__get_cpu_var(wd_enabled
) == 0)
321 if (nmi_watchdog
== NMI_LOCAL_APIC
)
322 lapic_watchdog_stop();
323 __get_cpu_var(wd_enabled
) = 0;
324 atomic_dec(&nmi_active
);
328 * the best way to detect whether a CPU has a 'hard lockup' problem
329 * is to check it's local APIC timer IRQ counts. If they are not
330 * changing then that CPU has some problem.
332 * as these watchdog NMI IRQs are generated on every CPU, we only
333 * have to check the current processor.
335 * since NMIs don't listen to _any_ locks, we have to be extremely
336 * careful not to rely on unsafe variables. The printk might lock
337 * up though, so we have to break up any console locks first ...
338 * [when there will be more tty-related locks, break them up here too!]
341 static DEFINE_PER_CPU(unsigned, last_irq_sum
);
342 static DEFINE_PER_CPU(local_t
, alert_counter
);
343 static DEFINE_PER_CPU(int, nmi_touch
);
345 void touch_nmi_watchdog(void)
347 if (nmi_watchdog_active()) {
351 * Tell other CPUs to reset their alert counters. We cannot
352 * do it ourselves because the alert count increase is not
355 for_each_present_cpu(cpu
) {
356 if (per_cpu(nmi_touch
, cpu
) != 1)
357 per_cpu(nmi_touch
, cpu
) = 1;
362 * Tickle the softlockup detector too:
364 touch_softlockup_watchdog();
366 EXPORT_SYMBOL(touch_nmi_watchdog
);
368 notrace __kprobes
int
369 nmi_watchdog_tick(struct pt_regs
*regs
, unsigned reason
)
372 * Since current_thread_info()-> is always on the stack, and we
373 * always switch the stack NMI-atomically, it's safe to use
374 * smp_processor_id().
378 int cpu
= smp_processor_id();
381 /* check for other users first */
382 if (notify_die(DIE_NMI
, "nmi", regs
, reason
, 2, SIGINT
)
388 sum
= get_timer_irqs(cpu
);
390 if (__get_cpu_var(nmi_touch
)) {
391 __get_cpu_var(nmi_touch
) = 0;
395 if (cpu_isset(cpu
, backtrace_mask
)) {
396 static DEFINE_SPINLOCK(lock
); /* Serialise the printks */
399 printk(KERN_WARNING
"NMI backtrace for cpu %d\n", cpu
);
402 cpu_clear(cpu
, backtrace_mask
);
405 /* Could check oops_in_progress here too, but it's safer not to */
406 if (mce_in_progress())
409 /* if the none of the timers isn't firing, this cpu isn't doing much */
410 if (!touched
&& __get_cpu_var(last_irq_sum
) == sum
) {
412 * Ayiee, looks like this CPU is stuck ...
413 * wait a few IRQs (5 seconds) before doing the oops ...
415 local_inc(&__get_cpu_var(alert_counter
));
416 if (local_read(&__get_cpu_var(alert_counter
)) == 5 * nmi_hz
)
418 * die_nmi will return ONLY if NOTIFY_STOP happens..
420 die_nmi("BUG: NMI Watchdog detected LOCKUP",
421 regs
, panic_on_timeout
);
423 __get_cpu_var(last_irq_sum
) = sum
;
424 local_set(&__get_cpu_var(alert_counter
), 0);
427 /* see if the nmi watchdog went off */
428 if (!__get_cpu_var(wd_enabled
))
430 switch (nmi_watchdog
) {
432 rc
|= lapic_wd_event(nmi_hz
);
436 * don't know how to accurately check for this.
437 * just assume it was a watchdog timer interrupt
438 * This matches the old behaviour.
448 static int unknown_nmi_panic_callback(struct pt_regs
*regs
, int cpu
)
450 unsigned char reason
= get_nmi_reason();
453 sprintf(buf
, "NMI received for unknown reason %02x\n", reason
);
454 die_nmi(buf
, regs
, 1); /* Always panic here */
459 * proc handler for /proc/sys/kernel/nmi
461 int proc_nmi_enabled(struct ctl_table
*table
, int write
, struct file
*file
,
462 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
466 nmi_watchdog_enabled
= (atomic_read(&nmi_active
) > 0) ? 1 : 0;
467 old_state
= nmi_watchdog_enabled
;
468 proc_dointvec(table
, write
, file
, buffer
, length
, ppos
);
469 if (!!old_state
== !!nmi_watchdog_enabled
)
472 if (atomic_read(&nmi_active
) < 0 || !nmi_watchdog_active()) {
474 "NMI watchdog is permanently disabled\n");
478 if (nmi_watchdog
== NMI_LOCAL_APIC
) {
479 if (nmi_watchdog_enabled
)
480 enable_lapic_nmi_watchdog();
482 disable_lapic_nmi_watchdog();
485 "NMI watchdog doesn't know what hardware to touch\n");
491 #endif /* CONFIG_SYSCTL */
493 int do_nmi_callback(struct pt_regs
*regs
, int cpu
)
496 if (unknown_nmi_panic
)
497 return unknown_nmi_panic_callback(regs
, cpu
);
502 void __trigger_all_cpu_backtrace(void)
506 backtrace_mask
= cpu_online_map
;
507 /* Wait for up to 10 seconds for all CPUs to do the backtrace */
508 for (i
= 0; i
< 10 * 1000; i
++) {
509 if (cpus_empty(backtrace_mask
))