1 /* Pseudo NMI support on sparc64 systems.
3 * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
5 * The NMI watchdog support and infrastructure is based almost
6 * entirely upon the x86 NMI support code.
8 #include <linux/kernel.h>
9 #include <linux/param.h>
10 #include <linux/init.h>
11 #include <linux/percpu.h>
12 #include <linux/nmi.h>
13 #include <linux/module.h>
14 #include <linux/kprobes.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/reboot.h>
17 #include <linux/slab.h>
18 #include <linux/kdebug.h>
19 #include <linux/delay.h>
20 #include <linux/smp.h>
22 #include <asm/perf_event.h>
23 #include <asm/ptrace.h>
26 /* We don't have a real NMI on sparc64, but we can fake one
27 * up using profiling counter overflow interrupts and interrupt
30 * The profile overflow interrupts at level 15, so we use
31 * level 14 as our IRQ off level.
34 static int panic_on_timeout;
37 * >0: the NMI watchdog is active, but can be disabled
38 * <0: the NMI watchdog has not been set up, and cannot be enabled
39 * 0: the NMI watchdog is disabled, but can be enabled
41 atomic_t nmi_active = ATOMIC_INIT(0); /* oprofile uses this */
42 EXPORT_SYMBOL(nmi_active);
44 static unsigned int nmi_hz = HZ;
45 static DEFINE_PER_CPU(short, wd_enabled);
46 static int endflag __initdata;
48 static DEFINE_PER_CPU(unsigned int, last_irq_sum);
49 static DEFINE_PER_CPU(long, alert_counter);
50 static DEFINE_PER_CPU(int, nmi_touch);
52 void touch_nmi_watchdog(void)
54 if (atomic_read(&nmi_active)) {
57 for_each_present_cpu(cpu) {
58 if (per_cpu(nmi_touch, cpu) != 1)
59 per_cpu(nmi_touch, cpu) = 1;
63 touch_softlockup_watchdog();
65 EXPORT_SYMBOL(touch_nmi_watchdog);
67 static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
69 if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
70 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
76 printk(KERN_EMERG "%s", str);
77 printk(" on CPU%d, ip %08lx, registers:\n",
78 smp_processor_id(), regs->tpc);
84 if (do_panic || panic_on_oops)
85 panic("Non maskable interrupt");
92 notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
94 unsigned int sum, touched = 0;
96 clear_softint(1 << irq);
98 local_cpu_data().__nmi_count++;
102 if (notify_die(DIE_NMI, "nmi", regs, 0,
103 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
106 pcr_ops->write(PCR_PIC_PRIV);
108 sum = local_cpu_data().irq0_irqs;
109 if (__get_cpu_var(nmi_touch)) {
110 __get_cpu_var(nmi_touch) = 0;
113 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
114 __this_cpu_inc(alert_counter);
115 if (__this_cpu_read(alert_counter) == 30 * nmi_hz)
116 die_nmi("BUG: NMI Watchdog detected LOCKUP",
117 regs, panic_on_timeout);
119 __get_cpu_var(last_irq_sum) = sum;
120 __this_cpu_write(alert_counter, 0);
122 if (__get_cpu_var(wd_enabled)) {
123 write_pic(picl_value(nmi_hz));
124 pcr_ops->write(pcr_enable);
130 static inline unsigned int get_nmi_count(int cpu)
132 return cpu_data(cpu).__nmi_count;
135 static __init void nmi_cpu_busy(void *data)
137 local_irq_enable_in_hardirq();
142 static void report_broken_nmi(int cpu, int *prev_nmi_count)
144 printk(KERN_CONT "\n");
147 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
148 cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
151 "Please report this to bugzilla.kernel.org,\n");
153 "and attach the output of the 'dmesg' command.\n");
155 per_cpu(wd_enabled, cpu) = 0;
156 atomic_dec(&nmi_active);
159 void stop_nmi_watchdog(void *unused)
161 pcr_ops->write(PCR_PIC_PRIV);
162 __get_cpu_var(wd_enabled) = 0;
163 atomic_dec(&nmi_active);
166 static int __init check_nmi_watchdog(void)
168 unsigned int *prev_nmi_count;
171 if (!atomic_read(&nmi_active))
174 prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(unsigned int), GFP_KERNEL);
175 if (!prev_nmi_count) {
180 printk(KERN_INFO "Testing NMI watchdog ... ");
182 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
184 for_each_possible_cpu(cpu)
185 prev_nmi_count[cpu] = get_nmi_count(cpu);
187 mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
189 for_each_online_cpu(cpu) {
190 if (!per_cpu(wd_enabled, cpu))
192 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
193 report_broken_nmi(cpu, prev_nmi_count);
196 if (!atomic_read(&nmi_active)) {
197 kfree(prev_nmi_count);
198 atomic_set(&nmi_active, -1);
206 kfree(prev_nmi_count);
209 on_each_cpu(stop_nmi_watchdog, NULL, 1);
213 void start_nmi_watchdog(void *unused)
215 __get_cpu_var(wd_enabled) = 1;
216 atomic_inc(&nmi_active);
218 pcr_ops->write(PCR_PIC_PRIV);
219 write_pic(picl_value(nmi_hz));
221 pcr_ops->write(pcr_enable);
224 static void nmi_adjust_hz_one(void *unused)
226 if (!__get_cpu_var(wd_enabled))
229 pcr_ops->write(PCR_PIC_PRIV);
230 write_pic(picl_value(nmi_hz));
232 pcr_ops->write(pcr_enable);
235 void nmi_adjust_hz(unsigned int new_hz)
238 on_each_cpu(nmi_adjust_hz_one, NULL, 1);
240 EXPORT_SYMBOL_GPL(nmi_adjust_hz);
242 static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
244 on_each_cpu(stop_nmi_watchdog, NULL, 1);
248 static struct notifier_block nmi_reboot_notifier = {
249 .notifier_call = nmi_shutdown,
252 int __init nmi_init(void)
256 on_each_cpu(start_nmi_watchdog, NULL, 1);
258 err = check_nmi_watchdog();
260 err = register_reboot_notifier(&nmi_reboot_notifier);
262 on_each_cpu(stop_nmi_watchdog, NULL, 1);
263 atomic_set(&nmi_active, -1);
267 init_hw_perf_events();
272 static int __init setup_nmi_watchdog(char *str)
274 if (!strncmp(str, "panic", 5))
275 panic_on_timeout = 1;
279 __setup("nmi_watchdog=", setup_nmi_watchdog);