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