1
2
3
4
5
6
7
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>
21
22#include <asm/perf_event.h>
23#include <asm/ptrace.h>
24#include <asm/local.h>
25#include <asm/pcr.h>
26
27
28
29
30
31
32
33
34
35static int panic_on_timeout;
36
37
38
39
40
41
42atomic_t nmi_active = ATOMIC_INIT(0);
43EXPORT_SYMBOL(nmi_active);
44
45static unsigned int nmi_hz = HZ;
46static DEFINE_PER_CPU(short, wd_enabled);
47static int endflag __initdata;
48
49static DEFINE_PER_CPU(unsigned int, last_irq_sum);
50static DEFINE_PER_CPU(local_t, alert_counter);
51static DEFINE_PER_CPU(int, nmi_touch);
52
53void touch_nmi_watchdog(void)
54{
55 if (atomic_read(&nmi_active)) {
56 int cpu;
57
58 for_each_present_cpu(cpu) {
59 if (per_cpu(nmi_touch, cpu) != 1)
60 per_cpu(nmi_touch, cpu) = 1;
61 }
62 }
63
64 touch_softlockup_watchdog();
65}
66EXPORT_SYMBOL(touch_nmi_watchdog);
67
68static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
69{
70 if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
71 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
72 return;
73
74 console_verbose();
75 bust_spinlocks(1);
76
77 printk(KERN_EMERG "%s", str);
78 printk(" on CPU%d, ip %08lx, registers:\n",
79 smp_processor_id(), regs->tpc);
80 show_regs(regs);
81 dump_stack();
82
83 bust_spinlocks(0);
84
85 if (do_panic || panic_on_oops)
86 panic("Non maskable interrupt");
87
88 nmi_exit();
89 local_irq_enable();
90 do_exit(SIGBUS);
91}
92
93notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
94{
95 unsigned int sum, touched = 0;
96 int cpu = smp_processor_id();
97
98 clear_softint(1 << irq);
99 pcr_ops->write(PCR_PIC_PRIV);
100
101 local_cpu_data().__nmi_count++;
102
103 nmi_enter();
104
105 if (notify_die(DIE_NMI, "nmi", regs, 0,
106 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
107 touched = 1;
108
109 sum = kstat_irqs_cpu(0, cpu);
110 if (__get_cpu_var(nmi_touch)) {
111 __get_cpu_var(nmi_touch) = 0;
112 touched = 1;
113 }
114 if (!touched && __get_cpu_var(last_irq_sum) == sum) {
115 local_inc(&__get_cpu_var(alert_counter));
116 if (local_read(&__get_cpu_var(alert_counter)) == 30 * nmi_hz)
117 die_nmi("BUG: NMI Watchdog detected LOCKUP",
118 regs, panic_on_timeout);
119 } else {
120 __get_cpu_var(last_irq_sum) = sum;
121 local_set(&__get_cpu_var(alert_counter), 0);
122 }
123 if (__get_cpu_var(wd_enabled)) {
124 write_pic(picl_value(nmi_hz));
125 pcr_ops->write(pcr_enable);
126 }
127
128 nmi_exit();
129}
130
131static inline unsigned int get_nmi_count(int cpu)
132{
133 return cpu_data(cpu).__nmi_count;
134}
135
136static __init void nmi_cpu_busy(void *data)
137{
138 local_irq_enable_in_hardirq();
139 while (endflag == 0)
140 mb();
141}
142
143static void report_broken_nmi(int cpu, int *prev_nmi_count)
144{
145 printk(KERN_CONT "\n");
146
147 printk(KERN_WARNING
148 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
149 cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
150
151 printk(KERN_WARNING
152 "Please report this to bugzilla.kernel.org,\n");
153 printk(KERN_WARNING
154 "and attach the output of the 'dmesg' command.\n");
155
156 per_cpu(wd_enabled, cpu) = 0;
157 atomic_dec(&nmi_active);
158}
159
160void stop_nmi_watchdog(void *unused)
161{
162 pcr_ops->write(PCR_PIC_PRIV);
163 __get_cpu_var(wd_enabled) = 0;
164 atomic_dec(&nmi_active);
165}
166
167static int __init check_nmi_watchdog(void)
168{
169 unsigned int *prev_nmi_count;
170 int cpu, err;
171
172 if (!atomic_read(&nmi_active))
173 return 0;
174
175 prev_nmi_count = kmalloc(nr_cpu_ids * sizeof(unsigned int), GFP_KERNEL);
176 if (!prev_nmi_count) {
177 err = -ENOMEM;
178 goto error;
179 }
180
181 printk(KERN_INFO "Testing NMI watchdog ... ");
182
183 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
184
185 for_each_possible_cpu(cpu)
186 prev_nmi_count[cpu] = get_nmi_count(cpu);
187 local_irq_enable();
188 mdelay((20 * 1000) / nmi_hz);
189
190 for_each_online_cpu(cpu) {
191 if (!per_cpu(wd_enabled, cpu))
192 continue;
193 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
194 report_broken_nmi(cpu, prev_nmi_count);
195 }
196 endflag = 1;
197 if (!atomic_read(&nmi_active)) {
198 kfree(prev_nmi_count);
199 atomic_set(&nmi_active, -1);
200 err = -ENODEV;
201 goto error;
202 }
203 printk("OK.\n");
204
205 nmi_hz = 1;
206
207 kfree(prev_nmi_count);
208 return 0;
209error:
210 on_each_cpu(stop_nmi_watchdog, NULL, 1);
211 return err;
212}
213
214void start_nmi_watchdog(void *unused)
215{
216 __get_cpu_var(wd_enabled) = 1;
217 atomic_inc(&nmi_active);
218
219 pcr_ops->write(PCR_PIC_PRIV);
220 write_pic(picl_value(nmi_hz));
221
222 pcr_ops->write(pcr_enable);
223}
224
225static void nmi_adjust_hz_one(void *unused)
226{
227 if (!__get_cpu_var(wd_enabled))
228 return;
229
230 pcr_ops->write(PCR_PIC_PRIV);
231 write_pic(picl_value(nmi_hz));
232
233 pcr_ops->write(pcr_enable);
234}
235
236void nmi_adjust_hz(unsigned int new_hz)
237{
238 nmi_hz = new_hz;
239 on_each_cpu(nmi_adjust_hz_one, NULL, 1);
240}
241EXPORT_SYMBOL_GPL(nmi_adjust_hz);
242
243static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
244{
245 on_each_cpu(stop_nmi_watchdog, NULL, 1);
246 return 0;
247}
248
249static struct notifier_block nmi_reboot_notifier = {
250 .notifier_call = nmi_shutdown,
251};
252
253int __init nmi_init(void)
254{
255 int err;
256
257 on_each_cpu(start_nmi_watchdog, NULL, 1);
258
259 err = check_nmi_watchdog();
260 if (!err) {
261 err = register_reboot_notifier(&nmi_reboot_notifier);
262 if (err) {
263 on_each_cpu(stop_nmi_watchdog, NULL, 1);
264 atomic_set(&nmi_active, -1);
265 }
266 }
267 if (!err)
268 init_hw_perf_events();
269
270 return err;
271}
272
273static int __init setup_nmi_watchdog(char *str)
274{
275 if (!strncmp(str, "panic", 5))
276 panic_on_timeout = 1;
277
278 return 0;
279}
280__setup("nmi_watchdog=", setup_nmi_watchdog);
281
