1
2
3
4
5
6
7
8
9#include <linux/kernel.h>
10#include <linux/param.h>
11#include <linux/init.h>
12#include <linux/percpu.h>
13#include <linux/nmi.h>
14#include <linux/export.h>
15#include <linux/kprobes.h>
16#include <linux/kernel_stat.h>
17#include <linux/reboot.h>
18#include <linux/slab.h>
19#include <linux/kdebug.h>
20#include <linux/delay.h>
21#include <linux/smp.h>
22
23#include <asm/perf_event.h>
24#include <asm/ptrace.h>
25#include <asm/pcr.h>
26
27#include "kstack.h"
28
29
30
31
32
33
34
35
36
37static int panic_on_timeout;
38
39
40
41
42
43
44atomic_t nmi_active = ATOMIC_INIT(0);
45EXPORT_SYMBOL(nmi_active);
46static int nmi_init_done;
47static unsigned int nmi_hz = HZ;
48static DEFINE_PER_CPU(short, wd_enabled);
49static int endflag __initdata;
50
51static DEFINE_PER_CPU(unsigned int, last_irq_sum);
52static DEFINE_PER_CPU(long, alert_counter);
53static DEFINE_PER_CPU(int, nmi_touch);
54
55void arch_touch_nmi_watchdog(void)
56{
57 if (atomic_read(&nmi_active)) {
58 int cpu;
59
60 for_each_present_cpu(cpu) {
61 if (per_cpu(nmi_touch, cpu) != 1)
62 per_cpu(nmi_touch, cpu) = 1;
63 }
64 }
65}
66EXPORT_SYMBOL(arch_touch_nmi_watchdog);
67
68static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
69{
70 int this_cpu = smp_processor_id();
71
72 if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
73 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
74 return;
75
76 if (do_panic || panic_on_oops)
77 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
78 else
79 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
80}
81
82notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
83{
84 unsigned int sum, touched = 0;
85 void *orig_sp;
86
87 clear_softint(1 << irq);
88
89 local_cpu_data().__nmi_count++;
90
91 nmi_enter();
92
93 orig_sp = set_hardirq_stack();
94
95 if (notify_die(DIE_NMI, "nmi", regs, 0,
96 pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
97 touched = 1;
98 else
99 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
100
101 sum = local_cpu_data().irq0_irqs;
102 if (__this_cpu_read(nmi_touch)) {
103 __this_cpu_write(nmi_touch, 0);
104 touched = 1;
105 }
106 if (!touched && __this_cpu_read(last_irq_sum) == sum) {
107 __this_cpu_inc(alert_counter);
108 if (__this_cpu_read(alert_counter) == 30 * nmi_hz)
109 die_nmi("BUG: NMI Watchdog detected LOCKUP",
110 regs, panic_on_timeout);
111 } else {
112 __this_cpu_write(last_irq_sum, sum);
113 __this_cpu_write(alert_counter, 0);
114 }
115 if (__this_cpu_read(wd_enabled)) {
116 pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
117 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
118 }
119
120 restore_hardirq_stack(orig_sp);
121
122 nmi_exit();
123}
124
125static inline unsigned int get_nmi_count(int cpu)
126{
127 return cpu_data(cpu).__nmi_count;
128}
129
130static __init void nmi_cpu_busy(void *data)
131{
132 while (endflag == 0)
133 mb();
134}
135
136static void report_broken_nmi(int cpu, int *prev_nmi_count)
137{
138 printk(KERN_CONT "\n");
139
140 printk(KERN_WARNING
141 "WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
142 cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
143
144 printk(KERN_WARNING
145 "Please report this to bugzilla.kernel.org,\n");
146 printk(KERN_WARNING
147 "and attach the output of the 'dmesg' command.\n");
148
149 per_cpu(wd_enabled, cpu) = 0;
150 atomic_dec(&nmi_active);
151}
152
153void stop_nmi_watchdog(void *unused)
154{
155 if (!__this_cpu_read(wd_enabled))
156 return;
157 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
158 __this_cpu_write(wd_enabled, 0);
159 atomic_dec(&nmi_active);
160}
161
162static int __init check_nmi_watchdog(void)
163{
164 unsigned int *prev_nmi_count;
165 int cpu, err;
166
167 if (!atomic_read(&nmi_active))
168 return 0;
169
170 prev_nmi_count = kmalloc_array(nr_cpu_ids, sizeof(unsigned int),
171 GFP_KERNEL);
172 if (!prev_nmi_count) {
173 err = -ENOMEM;
174 goto error;
175 }
176
177 printk(KERN_INFO "Testing NMI watchdog ... ");
178
179 smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
180
181 for_each_possible_cpu(cpu)
182 prev_nmi_count[cpu] = get_nmi_count(cpu);
183 local_irq_enable();
184 mdelay((20 * 1000) / nmi_hz);
185
186 for_each_online_cpu(cpu) {
187 if (!per_cpu(wd_enabled, cpu))
188 continue;
189 if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
190 report_broken_nmi(cpu, prev_nmi_count);
191 }
192 endflag = 1;
193 if (!atomic_read(&nmi_active)) {
194 kfree(prev_nmi_count);
195 atomic_set(&nmi_active, -1);
196 err = -ENODEV;
197 goto error;
198 }
199 printk("OK.\n");
200
201 nmi_hz = 1;
202
203 kfree(prev_nmi_count);
204 return 0;
205error:
206 on_each_cpu(stop_nmi_watchdog, NULL, 1);
207 return err;
208}
209
210void start_nmi_watchdog(void *unused)
211{
212 if (__this_cpu_read(wd_enabled))
213 return;
214
215 __this_cpu_write(wd_enabled, 1);
216 atomic_inc(&nmi_active);
217
218 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
219 pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
220
221 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
222}
223
224static void nmi_adjust_hz_one(void *unused)
225{
226 if (!__this_cpu_read(wd_enabled))
227 return;
228
229 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
230 pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
231
232 pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
233}
234
235void nmi_adjust_hz(unsigned int new_hz)
236{
237 nmi_hz = new_hz;
238 on_each_cpu(nmi_adjust_hz_one, NULL, 1);
239}
240EXPORT_SYMBOL_GPL(nmi_adjust_hz);
241
242static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
243{
244 on_each_cpu(stop_nmi_watchdog, NULL, 1);
245 return 0;
246}
247
248static struct notifier_block nmi_reboot_notifier = {
249 .notifier_call = nmi_shutdown,
250};
251
252int __init nmi_init(void)
253{
254 int err;
255
256 on_each_cpu(start_nmi_watchdog, NULL, 1);
257
258 err = check_nmi_watchdog();
259 if (!err) {
260 err = register_reboot_notifier(&nmi_reboot_notifier);
261 if (err) {
262 on_each_cpu(stop_nmi_watchdog, NULL, 1);
263 atomic_set(&nmi_active, -1);
264 }
265 }
266
267 nmi_init_done = 1;
268
269 return err;
270}
271
272static int __init setup_nmi_watchdog(char *str)
273{
274 if (!strncmp(str, "panic", 5))
275 panic_on_timeout = 1;
276
277 return 0;
278}
279__setup("nmi_watchdog=", setup_nmi_watchdog);
280
281
282
283
284
285int watchdog_nmi_enable(unsigned int cpu)
286{
287 if (atomic_read(&nmi_active) == -1) {
288 pr_warn("NMI watchdog cannot be enabled or disabled\n");
289 return -1;
290 }
291
292
293
294
295
296
297 if (!nmi_init_done)
298 return 0;
299
300 smp_call_function_single(cpu, start_nmi_watchdog, NULL, 1);
301
302 return 0;
303}
304
305
306
307
308void watchdog_nmi_disable(unsigned int cpu)
309{
310 if (atomic_read(&nmi_active) == -1)
311 pr_warn_once("NMI watchdog cannot be enabled or disabled\n");
312 else
313 smp_call_function_single(cpu, stop_nmi_watchdog, NULL, 1);
314}
315