2 * linux/arch/x86_64/nmi.c
4 * NMI watchdog support on APIC systems
6 * Started by Ingo Molnar <mingo@redhat.com>
9 * Mikael Pettersson : AMD K7 support for local APIC NMI watchdog.
10 * Mikael Pettersson : Power Management for local APIC NMI watchdog.
12 * Mikael Pettersson : PM converted to driver model. Disable/enable API.
16 #include <linux/delay.h>
17 #include <linux/interrupt.h>
18 #include <linux/module.h>
19 #include <linux/sysdev.h>
20 #include <linux/nmi.h>
21 #include <linux/sysctl.h>
22 #include <linux/kprobes.h>
26 #include <asm/proto.h>
27 #include <asm/kdebug.h>
31 * lapic_nmi_owner tracks the ownership of the lapic NMI hardware:
32 * - it may be reserved by some other driver, or not
33 * - when not reserved by some other driver, it may be used for
34 * the NMI watchdog, or not
36 * This is maintained separately from nmi_active because the NMI
37 * watchdog may also be driven from the I/O APIC timer.
39 static DEFINE_SPINLOCK(lapic_nmi_owner_lock);
40 static unsigned int lapic_nmi_owner;
41 #define LAPIC_NMI_WATCHDOG (1<<0)
42 #define LAPIC_NMI_RESERVED (1<<1)
45 * +1: the lapic NMI watchdog is active, but can be disabled
46 * 0: the lapic NMI watchdog has not been set up, and cannot
48 * -1: the lapic NMI watchdog is disabled, but can be enabled
50 int nmi_active; /* oprofile uses this */
53 unsigned int nmi_watchdog = NMI_DEFAULT;
54 static unsigned int nmi_hz = HZ;
55 static unsigned int nmi_perfctr_msr; /* the MSR to reset in NMI handler */
56 static unsigned int nmi_p4_cccr_val;
58 /* Note that these events don't tick when the CPU idles. This means
59 the frequency varies with CPU load. */
61 #define K7_EVNTSEL_ENABLE (1 << 22)
62 #define K7_EVNTSEL_INT (1 << 20)
63 #define K7_EVNTSEL_OS (1 << 17)
64 #define K7_EVNTSEL_USR (1 << 16)
65 #define K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING 0x76
66 #define K7_NMI_EVENT K7_EVENT_CYCLES_PROCESSOR_IS_RUNNING
68 #define MSR_P4_MISC_ENABLE 0x1A0
69 #define MSR_P4_MISC_ENABLE_PERF_AVAIL (1<<7)
70 #define MSR_P4_MISC_ENABLE_PEBS_UNAVAIL (1<<12)
71 #define MSR_P4_PERFCTR0 0x300
72 #define MSR_P4_CCCR0 0x360
73 #define P4_ESCR_EVENT_SELECT(N) ((N)<<25)
74 #define P4_ESCR_OS (1<<3)
75 #define P4_ESCR_USR (1<<2)
76 #define P4_CCCR_OVF_PMI0 (1<<26)
77 #define P4_CCCR_OVF_PMI1 (1<<27)
78 #define P4_CCCR_THRESHOLD(N) ((N)<<20)
79 #define P4_CCCR_COMPLEMENT (1<<19)
80 #define P4_CCCR_COMPARE (1<<18)
81 #define P4_CCCR_REQUIRED (3<<16)
82 #define P4_CCCR_ESCR_SELECT(N) ((N)<<13)
83 #define P4_CCCR_ENABLE (1<<12)
84 /* Set up IQ_COUNTER0 to behave like a clock, by having IQ_CCCR0 filter
85 CRU_ESCR0 (with any non-null event selector) through a complemented
86 max threshold. [IA32-Vol3, Section 14.9.9] */
87 #define MSR_P4_IQ_COUNTER0 0x30C
88 #define P4_NMI_CRU_ESCR0 (P4_ESCR_EVENT_SELECT(0x3F)|P4_ESCR_OS|P4_ESCR_USR)
89 #define P4_NMI_IQ_CCCR0 \
90 (P4_CCCR_OVF_PMI0|P4_CCCR_THRESHOLD(15)|P4_CCCR_COMPLEMENT| \
91 P4_CCCR_COMPARE|P4_CCCR_REQUIRED|P4_CCCR_ESCR_SELECT(4)|P4_CCCR_ENABLE)
93 static __cpuinit inline int nmi_known_cpu(void)
95 switch (boot_cpu_data.x86_vendor) {
97 return boot_cpu_data.x86 == 15;
98 case X86_VENDOR_INTEL:
99 return boot_cpu_data.x86 == 15;
104 /* Run after command line and cpu_init init, but before all other checks */
105 void __cpuinit nmi_watchdog_default(void)
107 if (nmi_watchdog != NMI_DEFAULT)
110 nmi_watchdog = NMI_LOCAL_APIC;
112 nmi_watchdog = NMI_IO_APIC;
116 /* The performance counters used by NMI_LOCAL_APIC don't trigger when
117 * the CPU is idle. To make sure the NMI watchdog really ticks on all
118 * CPUs during the test make them busy.
120 static __init void nmi_cpu_busy(void *data)
122 volatile int *endflag = data;
123 local_irq_enable_in_hardirq();
124 /* Intentionally don't use cpu_relax here. This is
125 to make sure that the performance counter really ticks,
126 even if there is a simulator or similar that catches the
127 pause instruction. On a real HT machine this is fine because
128 all other CPUs are busy with "useless" delay loops and don't
129 care if they get somewhat less cycles. */
130 while (*endflag == 0)
135 int __init check_nmi_watchdog (void)
137 volatile int endflag = 0;
141 counts = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
145 printk(KERN_INFO "testing NMI watchdog ... ");
148 if (nmi_watchdog == NMI_LOCAL_APIC)
149 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
152 for (cpu = 0; cpu < NR_CPUS; cpu++)
153 counts[cpu] = cpu_pda(cpu)->__nmi_count;
155 mdelay((10*1000)/nmi_hz); // wait 10 ticks
157 for_each_online_cpu(cpu) {
158 if (cpu_pda(cpu)->__nmi_count - counts[cpu] <= 5) {
160 printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
163 cpu_pda(cpu)->__nmi_count);
165 lapic_nmi_owner &= ~LAPIC_NMI_WATCHDOG;
174 /* now that we know it works we can reduce NMI frequency to
175 something more reasonable; makes a difference in some configs */
176 if (nmi_watchdog == NMI_LOCAL_APIC)
183 int __init setup_nmi_watchdog(char *str)
187 if (!strncmp(str,"panic",5)) {
188 panic_on_timeout = 1;
189 str = strchr(str, ',');
195 get_option(&str, &nmi);
197 if (nmi >= NMI_INVALID)
203 __setup("nmi_watchdog=", setup_nmi_watchdog);
205 static void disable_lapic_nmi_watchdog(void)
209 switch (boot_cpu_data.x86_vendor) {
211 wrmsr(MSR_K7_EVNTSEL0, 0, 0);
213 case X86_VENDOR_INTEL:
214 if (boot_cpu_data.x86 == 15) {
215 wrmsr(MSR_P4_IQ_CCCR0, 0, 0);
216 wrmsr(MSR_P4_CRU_ESCR0, 0, 0);
221 /* tell do_nmi() and others that we're not active any more */
225 static void enable_lapic_nmi_watchdog(void)
227 if (nmi_active < 0) {
228 nmi_watchdog = NMI_LOCAL_APIC;
229 touch_nmi_watchdog();
230 setup_apic_nmi_watchdog();
234 int reserve_lapic_nmi(void)
236 unsigned int old_owner;
238 spin_lock(&lapic_nmi_owner_lock);
239 old_owner = lapic_nmi_owner;
240 lapic_nmi_owner |= LAPIC_NMI_RESERVED;
241 spin_unlock(&lapic_nmi_owner_lock);
242 if (old_owner & LAPIC_NMI_RESERVED)
244 if (old_owner & LAPIC_NMI_WATCHDOG)
245 disable_lapic_nmi_watchdog();
249 void release_lapic_nmi(void)
251 unsigned int new_owner;
253 spin_lock(&lapic_nmi_owner_lock);
254 new_owner = lapic_nmi_owner & ~LAPIC_NMI_RESERVED;
255 lapic_nmi_owner = new_owner;
256 spin_unlock(&lapic_nmi_owner_lock);
257 if (new_owner & LAPIC_NMI_WATCHDOG)
258 enable_lapic_nmi_watchdog();
261 void disable_timer_nmi_watchdog(void)
263 if ((nmi_watchdog != NMI_IO_APIC) || (nmi_active <= 0))
267 unset_nmi_callback();
269 nmi_watchdog = NMI_NONE;
272 void enable_timer_nmi_watchdog(void)
274 if (nmi_active < 0) {
275 nmi_watchdog = NMI_IO_APIC;
276 touch_nmi_watchdog();
284 static int nmi_pm_active; /* nmi_active before suspend */
286 static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
288 nmi_pm_active = nmi_active;
289 disable_lapic_nmi_watchdog();
293 static int lapic_nmi_resume(struct sys_device *dev)
295 if (nmi_pm_active > 0)
296 enable_lapic_nmi_watchdog();
300 static struct sysdev_class nmi_sysclass = {
301 set_kset_name("lapic_nmi"),
302 .resume = lapic_nmi_resume,
303 .suspend = lapic_nmi_suspend,
306 static struct sys_device device_lapic_nmi = {
308 .cls = &nmi_sysclass,
311 static int __init init_lapic_nmi_sysfs(void)
315 if (nmi_active == 0 || nmi_watchdog != NMI_LOCAL_APIC)
318 error = sysdev_class_register(&nmi_sysclass);
320 error = sysdev_register(&device_lapic_nmi);
323 /* must come after the local APIC's device_initcall() */
324 late_initcall(init_lapic_nmi_sysfs);
326 #endif /* CONFIG_PM */
329 * Activate the NMI watchdog via the local APIC.
330 * Original code written by Keith Owens.
333 static void clear_msr_range(unsigned int base, unsigned int n)
337 for(i = 0; i < n; ++i)
341 static void setup_k7_watchdog(void)
344 unsigned int evntsel;
346 nmi_perfctr_msr = MSR_K7_PERFCTR0;
348 for(i = 0; i < 4; ++i) {
349 /* Simulator may not support it */
350 if (checking_wrmsrl(MSR_K7_EVNTSEL0+i, 0UL)) {
354 wrmsrl(MSR_K7_PERFCTR0+i, 0UL);
357 evntsel = K7_EVNTSEL_INT
362 wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
363 wrmsrl(MSR_K7_PERFCTR0, -((u64)cpu_khz * 1000 / nmi_hz));
364 apic_write(APIC_LVTPC, APIC_DM_NMI);
365 evntsel |= K7_EVNTSEL_ENABLE;
366 wrmsr(MSR_K7_EVNTSEL0, evntsel, 0);
370 static int setup_p4_watchdog(void)
372 unsigned int misc_enable, dummy;
374 rdmsr(MSR_P4_MISC_ENABLE, misc_enable, dummy);
375 if (!(misc_enable & MSR_P4_MISC_ENABLE_PERF_AVAIL))
378 nmi_perfctr_msr = MSR_P4_IQ_COUNTER0;
379 nmi_p4_cccr_val = P4_NMI_IQ_CCCR0;
381 if (smp_num_siblings == 2)
382 nmi_p4_cccr_val |= P4_CCCR_OVF_PMI1;
385 if (!(misc_enable & MSR_P4_MISC_ENABLE_PEBS_UNAVAIL))
386 clear_msr_range(0x3F1, 2);
387 /* MSR 0x3F0 seems to have a default value of 0xFC00, but current
388 docs doesn't fully define it, so leave it alone for now. */
389 if (boot_cpu_data.x86_model >= 0x3) {
390 /* MSR_P4_IQ_ESCR0/1 (0x3ba/0x3bb) removed */
391 clear_msr_range(0x3A0, 26);
392 clear_msr_range(0x3BC, 3);
394 clear_msr_range(0x3A0, 31);
396 clear_msr_range(0x3C0, 6);
397 clear_msr_range(0x3C8, 6);
398 clear_msr_range(0x3E0, 2);
399 clear_msr_range(MSR_P4_CCCR0, 18);
400 clear_msr_range(MSR_P4_PERFCTR0, 18);
402 wrmsr(MSR_P4_CRU_ESCR0, P4_NMI_CRU_ESCR0, 0);
403 wrmsr(MSR_P4_IQ_CCCR0, P4_NMI_IQ_CCCR0 & ~P4_CCCR_ENABLE, 0);
404 Dprintk("setting P4_IQ_COUNTER0 to 0x%08lx\n", -(cpu_khz * 1000UL / nmi_hz));
405 wrmsrl(MSR_P4_IQ_COUNTER0, -((u64)cpu_khz * 1000 / nmi_hz));
406 apic_write(APIC_LVTPC, APIC_DM_NMI);
407 wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
411 void setup_apic_nmi_watchdog(void)
413 switch (boot_cpu_data.x86_vendor) {
415 if (boot_cpu_data.x86 != 15)
417 if (strstr(boot_cpu_data.x86_model_id, "Screwdriver"))
421 case X86_VENDOR_INTEL:
422 if (boot_cpu_data.x86 != 15)
424 if (!setup_p4_watchdog())
431 lapic_nmi_owner = LAPIC_NMI_WATCHDOG;
436 * the best way to detect whether a CPU has a 'hard lockup' problem
437 * is to check it's local APIC timer IRQ counts. If they are not
438 * changing then that CPU has some problem.
440 * as these watchdog NMI IRQs are generated on every CPU, we only
441 * have to check the current processor.
444 static DEFINE_PER_CPU(unsigned, last_irq_sum);
445 static DEFINE_PER_CPU(local_t, alert_counter);
446 static DEFINE_PER_CPU(int, nmi_touch);
448 void touch_nmi_watchdog (void)
450 if (nmi_watchdog > 0) {
454 * Tell other CPUs to reset their alert counters. We cannot
455 * do it ourselves because the alert count increase is not
458 for_each_present_cpu (cpu)
459 per_cpu(nmi_touch, cpu) = 1;
462 touch_softlockup_watchdog();
465 void __kprobes nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
470 sum = read_pda(apic_timer_irqs);
471 if (__get_cpu_var(nmi_touch)) {
472 __get_cpu_var(nmi_touch) = 0;
475 #ifdef CONFIG_X86_MCE
476 /* Could check oops_in_progress here too, but it's safer
478 if (atomic_read(&mce_entry) > 0)
481 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
483 * Ayiee, looks like this CPU is stuck ...
484 * wait a few IRQs (5 seconds) before doing the oops ...
486 local_inc(&__get_cpu_var(alert_counter));
487 if (local_read(&__get_cpu_var(alert_counter)) == 5*nmi_hz) {
488 if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
490 local_set(&__get_cpu_var(alert_counter), 0);
493 die_nmi("NMI Watchdog detected LOCKUP on CPU %d\n", regs);
496 __get_cpu_var(last_irq_sum) = sum;
497 local_set(&__get_cpu_var(alert_counter), 0);
499 if (nmi_perfctr_msr) {
500 if (nmi_perfctr_msr == MSR_P4_IQ_COUNTER0) {
503 * - An overflown perfctr will assert its interrupt
504 * until the OVF flag in its CCCR is cleared.
505 * - LVTPC is masked on interrupt and must be
506 * unmasked by the LVTPC handler.
508 wrmsr(MSR_P4_IQ_CCCR0, nmi_p4_cccr_val, 0);
509 apic_write(APIC_LVTPC, APIC_DM_NMI);
511 wrmsrl(nmi_perfctr_msr, -((u64)cpu_khz * 1000 / nmi_hz));
515 static __kprobes int dummy_nmi_callback(struct pt_regs * regs, int cpu)
520 static nmi_callback_t nmi_callback = dummy_nmi_callback;
522 asmlinkage __kprobes void do_nmi(struct pt_regs * regs, long error_code)
524 int cpu = safe_smp_processor_id();
527 add_pda(__nmi_count,1);
528 if (!rcu_dereference(nmi_callback)(regs, cpu))
529 default_do_nmi(regs);
533 void set_nmi_callback(nmi_callback_t callback)
536 rcu_assign_pointer(nmi_callback, callback);
538 EXPORT_SYMBOL_GPL(set_nmi_callback);
540 void unset_nmi_callback(void)
542 nmi_callback = dummy_nmi_callback;
544 EXPORT_SYMBOL_GPL(unset_nmi_callback);
548 static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
550 unsigned char reason = get_nmi_reason();
553 if (!(reason & 0xc0)) {
554 sprintf(buf, "NMI received for unknown reason %02x\n", reason);
561 * proc handler for /proc/sys/kernel/unknown_nmi_panic
563 int proc_unknown_nmi_panic(struct ctl_table *table, int write, struct file *file,
564 void __user *buffer, size_t *length, loff_t *ppos)
568 old_state = unknown_nmi_panic;
569 proc_dointvec(table, write, file, buffer, length, ppos);
570 if (!!old_state == !!unknown_nmi_panic)
573 if (unknown_nmi_panic) {
574 if (reserve_lapic_nmi() < 0) {
575 unknown_nmi_panic = 0;
578 set_nmi_callback(unknown_nmi_panic_callback);
582 unset_nmi_callback();
589 EXPORT_SYMBOL(nmi_active);
590 EXPORT_SYMBOL(nmi_watchdog);
591 EXPORT_SYMBOL(reserve_lapic_nmi);
592 EXPORT_SYMBOL(release_lapic_nmi);
593 EXPORT_SYMBOL(disable_timer_nmi_watchdog);
594 EXPORT_SYMBOL(enable_timer_nmi_watchdog);
595 EXPORT_SYMBOL(touch_nmi_watchdog);