• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Pseudo NMI support on sparc64 systems.
2  *
3  * Copyright (C) 2009 David S. Miller <davem@davemloft.net>
4  *
5  * The NMI watchdog support and infrastructure is based almost
6  * entirely upon the x86 NMI support code.
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 /* We don't have a real NMI on sparc64, but we can fake one
29  * up using profiling counter overflow interrupts and interrupt
30  * levels.
31  *
32  * The profile overflow interrupts at level 15, so we use
33  * level 14 as our IRQ off level.
34  */
35 
36 static int panic_on_timeout;
37 
38 /* nmi_active:
39  * >0: the NMI watchdog is active, but can be disabled
40  * <0: the NMI watchdog has not been set up, and cannot be enabled
41  *  0: the NMI watchdog is disabled, but can be enabled
42  */
43 atomic_t nmi_active = ATOMIC_INIT(0);		/* oprofile uses this */
44 EXPORT_SYMBOL(nmi_active);
45 
46 static unsigned int nmi_hz = HZ;
47 static DEFINE_PER_CPU(short, wd_enabled);
48 static int endflag __initdata;
49 
50 static DEFINE_PER_CPU(unsigned int, last_irq_sum);
51 static DEFINE_PER_CPU(long, alert_counter);
52 static DEFINE_PER_CPU(int, nmi_touch);
53 
touch_nmi_watchdog(void)54 void 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 
65 	touch_softlockup_watchdog();
66 }
67 EXPORT_SYMBOL(touch_nmi_watchdog);
68 
die_nmi(const char * str,struct pt_regs * regs,int do_panic)69 static void die_nmi(const char *str, struct pt_regs *regs, int do_panic)
70 {
71 	int this_cpu = smp_processor_id();
72 
73 	if (notify_die(DIE_NMIWATCHDOG, str, regs, 0,
74 		       pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
75 		return;
76 
77 	if (do_panic || panic_on_oops)
78 		panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
79 	else
80 		WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
81 }
82 
perfctr_irq(int irq,struct pt_regs * regs)83 notrace __kprobes void perfctr_irq(int irq, struct pt_regs *regs)
84 {
85 	unsigned int sum, touched = 0;
86 	void *orig_sp;
87 
88 	clear_softint(1 << irq);
89 
90 	local_cpu_data().__nmi_count++;
91 
92 	nmi_enter();
93 
94 	orig_sp = set_hardirq_stack();
95 
96 	if (notify_die(DIE_NMI, "nmi", regs, 0,
97 		       pt_regs_trap_type(regs), SIGINT) == NOTIFY_STOP)
98 		touched = 1;
99 	else
100 		pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
101 
102 	sum = local_cpu_data().irq0_irqs;
103 	if (__this_cpu_read(nmi_touch)) {
104 		__this_cpu_write(nmi_touch, 0);
105 		touched = 1;
106 	}
107 	if (!touched && __this_cpu_read(last_irq_sum) == sum) {
108 		__this_cpu_inc(alert_counter);
109 		if (__this_cpu_read(alert_counter) == 30 * nmi_hz)
110 			die_nmi("BUG: NMI Watchdog detected LOCKUP",
111 				regs, panic_on_timeout);
112 	} else {
113 		__this_cpu_write(last_irq_sum, sum);
114 		__this_cpu_write(alert_counter, 0);
115 	}
116 	if (__this_cpu_read(wd_enabled)) {
117 		pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
118 		pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
119 	}
120 
121 	restore_hardirq_stack(orig_sp);
122 
123 	nmi_exit();
124 }
125 
get_nmi_count(int cpu)126 static inline unsigned int get_nmi_count(int cpu)
127 {
128 	return cpu_data(cpu).__nmi_count;
129 }
130 
nmi_cpu_busy(void * data)131 static __init void nmi_cpu_busy(void *data)
132 {
133 	while (endflag == 0)
134 		mb();
135 }
136 
report_broken_nmi(int cpu,int * prev_nmi_count)137 static void report_broken_nmi(int cpu, int *prev_nmi_count)
138 {
139 	printk(KERN_CONT "\n");
140 
141 	printk(KERN_WARNING
142 		"WARNING: CPU#%d: NMI appears to be stuck (%d->%d)!\n",
143 			cpu, prev_nmi_count[cpu], get_nmi_count(cpu));
144 
145 	printk(KERN_WARNING
146 		"Please report this to bugzilla.kernel.org,\n");
147 	printk(KERN_WARNING
148 		"and attach the output of the 'dmesg' command.\n");
149 
150 	per_cpu(wd_enabled, cpu) = 0;
151 	atomic_dec(&nmi_active);
152 }
153 
stop_nmi_watchdog(void * unused)154 void stop_nmi_watchdog(void *unused)
155 {
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 
check_nmi_watchdog(void)161 static 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(nr_cpu_ids * sizeof(unsigned int), GFP_KERNEL);
170 	if (!prev_nmi_count) {
171 		err = -ENOMEM;
172 		goto error;
173 	}
174 
175 	printk(KERN_INFO "Testing NMI watchdog ... ");
176 
177 	smp_call_function(nmi_cpu_busy, (void *)&endflag, 0);
178 
179 	for_each_possible_cpu(cpu)
180 		prev_nmi_count[cpu] = get_nmi_count(cpu);
181 	local_irq_enable();
182 	mdelay((20 * 1000) / nmi_hz); /* wait 20 ticks */
183 
184 	for_each_online_cpu(cpu) {
185 		if (!per_cpu(wd_enabled, cpu))
186 			continue;
187 		if (get_nmi_count(cpu) - prev_nmi_count[cpu] <= 5)
188 			report_broken_nmi(cpu, prev_nmi_count);
189 	}
190 	endflag = 1;
191 	if (!atomic_read(&nmi_active)) {
192 		kfree(prev_nmi_count);
193 		atomic_set(&nmi_active, -1);
194 		err = -ENODEV;
195 		goto error;
196 	}
197 	printk("OK.\n");
198 
199 	nmi_hz = 1;
200 
201 	kfree(prev_nmi_count);
202 	return 0;
203 error:
204 	on_each_cpu(stop_nmi_watchdog, NULL, 1);
205 	return err;
206 }
207 
start_nmi_watchdog(void * unused)208 void start_nmi_watchdog(void *unused)
209 {
210 	__this_cpu_write(wd_enabled, 1);
211 	atomic_inc(&nmi_active);
212 
213 	pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
214 	pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
215 
216 	pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
217 }
218 
nmi_adjust_hz_one(void * unused)219 static void nmi_adjust_hz_one(void *unused)
220 {
221 	if (!__this_cpu_read(wd_enabled))
222 		return;
223 
224 	pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_disable);
225 	pcr_ops->write_pic(0, pcr_ops->nmi_picl_value(nmi_hz));
226 
227 	pcr_ops->write_pcr(0, pcr_ops->pcr_nmi_enable);
228 }
229 
nmi_adjust_hz(unsigned int new_hz)230 void nmi_adjust_hz(unsigned int new_hz)
231 {
232 	nmi_hz = new_hz;
233 	on_each_cpu(nmi_adjust_hz_one, NULL, 1);
234 }
235 EXPORT_SYMBOL_GPL(nmi_adjust_hz);
236 
nmi_shutdown(struct notifier_block * nb,unsigned long cmd,void * p)237 static int nmi_shutdown(struct notifier_block *nb, unsigned long cmd, void *p)
238 {
239 	on_each_cpu(stop_nmi_watchdog, NULL, 1);
240 	return 0;
241 }
242 
243 static struct notifier_block nmi_reboot_notifier = {
244 	.notifier_call = nmi_shutdown,
245 };
246 
nmi_init(void)247 int __init nmi_init(void)
248 {
249 	int err;
250 
251 	on_each_cpu(start_nmi_watchdog, NULL, 1);
252 
253 	err = check_nmi_watchdog();
254 	if (!err) {
255 		err = register_reboot_notifier(&nmi_reboot_notifier);
256 		if (err) {
257 			on_each_cpu(stop_nmi_watchdog, NULL, 1);
258 			atomic_set(&nmi_active, -1);
259 		}
260 	}
261 
262 	return err;
263 }
264 
setup_nmi_watchdog(char * str)265 static int __init setup_nmi_watchdog(char *str)
266 {
267 	if (!strncmp(str, "panic", 5))
268 		panic_on_timeout = 1;
269 
270 	return 0;
271 }
272 __setup("nmi_watchdog=", setup_nmi_watchdog);
273