nmi_32.c 10.9 KB
Newer Older
L
Linus Torvalds 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/*
 *  NMI watchdog support on APIC systems
 *
 *  Started by Ingo Molnar <mingo@redhat.com>
 *
 *  Fixes:
 *  Mikael Pettersson	: AMD K7 support for local APIC NMI watchdog.
 *  Mikael Pettersson	: Power Management for local APIC NMI watchdog.
 *  Mikael Pettersson	: Pentium 4 support for local APIC NMI watchdog.
 *  Pavel Machek and
 *  Mikael Pettersson	: PM converted to driver model. Disable/enable API.
 */

#include <linux/delay.h>
#include <linux/interrupt.h>
#include <linux/module.h>
#include <linux/nmi.h>
#include <linux/sysdev.h>
#include <linux/sysctl.h>
20
#include <linux/percpu.h>
21
#include <linux/kprobes.h>
A
Andrew Morton 已提交
22
#include <linux/cpumask.h>
23
#include <linux/kernel_stat.h>
24
#include <linux/kdebug.h>
L
Linus Torvalds 已提交
25 26 27 28 29 30

#include <asm/smp.h>
#include <asm/nmi.h>

#include "mach_traps.h"

31 32 33
int unknown_nmi_panic;
int nmi_watchdog_enabled;

34
static cpumask_t backtrace_mask = CPU_MASK_NONE;
35

L
Linus Torvalds 已提交
36
/* nmi_active:
37 38
 * >0: the lapic NMI watchdog is active, but can be disabled
 * <0: the lapic NMI watchdog has not been set up, and cannot
L
Linus Torvalds 已提交
39
 *     be enabled
40
 *  0: the lapic NMI watchdog is disabled, but can be enabled
L
Linus Torvalds 已提交
41
 */
42
atomic_t nmi_active = ATOMIC_INIT(0);		/* oprofile uses this */
L
Linus Torvalds 已提交
43

44 45
unsigned int nmi_watchdog = NMI_DEFAULT;
static unsigned int nmi_hz = HZ;
L
Linus Torvalds 已提交
46

47
static DEFINE_PER_CPU(short, wd_enabled);
L
Linus Torvalds 已提交
48

49 50
static int endflag __initdata = 0;

H
Hiroshi Shimamoto 已提交
51
#ifdef CONFIG_SMP
52 53 54 55 56 57
/* The performance counters used by NMI_LOCAL_APIC don't trigger when
 * the CPU is idle. To make sure the NMI watchdog really ticks on all
 * CPUs during the test make them busy.
 */
static __init void nmi_cpu_busy(void *data)
{
58
	local_irq_enable_in_hardirq();
59 60 61 62 63 64
	/* Intentionally don't use cpu_relax here. This is
	   to make sure that the performance counter really ticks,
	   even if there is a simulator or similar that catches the
	   pause instruction. On a real HT machine this is fine because
	   all other CPUs are busy with "useless" delay loops and don't
	   care if they get somewhat less cycles. */
65 66
	while (endflag == 0)
		mb();
67
}
H
Hiroshi Shimamoto 已提交
68
#endif
69

70
static int __init check_nmi_watchdog(void)
L
Linus Torvalds 已提交
71
{
72
	unsigned int *prev_nmi_count;
L
Linus Torvalds 已提交
73 74
	int cpu;

75
	if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DISABLED))
76 77 78
		return 0;

	if (!atomic_read(&nmi_active))
79 80
		return 0;

81 82
	prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
	if (!prev_nmi_count)
83
		return -1;
84

85
	printk(KERN_INFO "Testing NMI watchdog ... ");
L
Linus Torvalds 已提交
86

H
Hiroshi Shimamoto 已提交
87
#ifdef CONFIG_SMP
88 89
	if (nmi_watchdog == NMI_LOCAL_APIC)
		smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
H
Hiroshi Shimamoto 已提交
90
#endif
91

92
	for_each_possible_cpu(cpu)
H
Hiroshi Shimamoto 已提交
93
		prev_nmi_count[cpu] = nmi_count(cpu);
L
Linus Torvalds 已提交
94
	local_irq_enable();
95
	mdelay((20*1000)/nmi_hz); // wait 20 ticks
L
Linus Torvalds 已提交
96

97
	for_each_possible_cpu(cpu) {
L
Linus Torvalds 已提交
98 99 100 101 102 103
#ifdef CONFIG_SMP
		/* Check cpu_callin_map here because that is set
		   after the timer is started. */
		if (!cpu_isset(cpu, cpu_callin_map))
			continue;
#endif
104
		if (!per_cpu(wd_enabled, cpu))
105
			continue;
L
Linus Torvalds 已提交
106
		if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
107 108
			printk(KERN_WARNING "WARNING: CPU#%d: NMI "
				"appears to be stuck (%d->%d)!\n",
109 110 111
				cpu,
				prev_nmi_count[cpu],
				nmi_count(cpu));
112
			per_cpu(wd_enabled, cpu) = 0;
113
			atomic_dec(&nmi_active);
L
Linus Torvalds 已提交
114 115
		}
	}
116
	endflag = 1;
117 118 119
	if (!atomic_read(&nmi_active)) {
		kfree(prev_nmi_count);
		atomic_set(&nmi_active, -1);
120
		return -1;
121
	}
L
Linus Torvalds 已提交
122 123 124 125
	printk("OK.\n");

	/* now that we know it works we can reduce NMI frequency to
	   something more reasonable; makes a difference in some configs */
126 127
	if (nmi_watchdog == NMI_LOCAL_APIC)
		nmi_hz = lapic_adjust_nmi_hz(1);
L
Linus Torvalds 已提交
128

129
	kfree(prev_nmi_count);
L
Linus Torvalds 已提交
130 131
	return 0;
}
132 133
/* This needs to happen later in boot so counters are working */
late_initcall(check_nmi_watchdog);
L
Linus Torvalds 已提交
134 135 136 137 138 139 140

static int __init setup_nmi_watchdog(char *str)
{
	int nmi;

	get_option(&str, &nmi);

141
	if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
L
Linus Torvalds 已提交
142
		return 0;
143

144
	nmi_watchdog = nmi;
L
Linus Torvalds 已提交
145 146 147 148 149
	return 1;
}

__setup("nmi_watchdog=", setup_nmi_watchdog);

150

151
/* Suspend/resume support */
152

L
Linus Torvalds 已提交
153 154 155 156
#ifdef CONFIG_PM

static int nmi_pm_active; /* nmi_active before suspend */

157
static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
L
Linus Torvalds 已提交
158
{
159
	/* only CPU0 goes here, other CPUs should be offline */
160
	nmi_pm_active = atomic_read(&nmi_active);
161 162
	stop_apic_nmi_watchdog(NULL);
	BUG_ON(atomic_read(&nmi_active) != 0);
L
Linus Torvalds 已提交
163 164 165 166 167
	return 0;
}

static int lapic_nmi_resume(struct sys_device *dev)
{
168 169 170 171 172
	/* only CPU0 goes here, other CPUs should be offline */
	if (nmi_pm_active > 0) {
		setup_apic_nmi_watchdog(NULL);
		touch_nmi_watchdog();
	}
L
Linus Torvalds 已提交
173 174 175 176 177
	return 0;
}


static struct sysdev_class nmi_sysclass = {
178
	.name		= "lapic_nmi",
L
Linus Torvalds 已提交
179 180 181 182 183 184 185 186 187 188 189 190 191
	.resume		= lapic_nmi_resume,
	.suspend	= lapic_nmi_suspend,
};

static struct sys_device device_lapic_nmi = {
	.id	= 0,
	.cls	= &nmi_sysclass,
};

static int __init init_lapic_nmi_sysfs(void)
{
	int error;

192 193 194 195 196 197
	/* should really be a BUG_ON but b/c this is an
	 * init call, it just doesn't work.  -dcz
	 */
	if (nmi_watchdog != NMI_LOCAL_APIC)
		return 0;

198
	if (atomic_read(&nmi_active) < 0)
L
Linus Torvalds 已提交
199 200 201 202 203 204 205 206 207 208 209 210
		return 0;

	error = sysdev_class_register(&nmi_sysclass);
	if (!error)
		error = sysdev_register(&device_lapic_nmi);
	return error;
}
/* must come after the local APIC's device_initcall() */
late_initcall(init_lapic_nmi_sysfs);

#endif	/* CONFIG_PM */

211
static void __acpi_nmi_enable(void *__unused)
L
Linus Torvalds 已提交
212
{
213
	apic_write_around(APIC_LVT0, APIC_DM_NMI);
L
Linus Torvalds 已提交
214 215
}

216 217 218 219
/*
 * Enable timer based NMIs on all CPUs:
 */
void acpi_nmi_enable(void)
L
Linus Torvalds 已提交
220
{
221 222
	if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
		on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
223 224
}

225
static void __acpi_nmi_disable(void *__unused)
226
{
227
	apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
228 229
}

230 231 232 233
/*
 * Disable timer based NMIs on all CPUs:
 */
void acpi_nmi_disable(void)
234
{
235 236
	if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
		on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
237 238
}

H
Hiroshi Shimamoto 已提交
239
void setup_apic_nmi_watchdog(void *unused)
240
{
241
	if (__get_cpu_var(wd_enabled))
H
Hiroshi Shimamoto 已提交
242
		return;
243 244 245 246 247 248

	/* cheap hack to support suspend/resume */
	/* if cpu0 is not active neither should the other cpus */
	if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
		return;

249 250 251 252 253
	switch (nmi_watchdog) {
	case NMI_LOCAL_APIC:
		__get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
		if (lapic_watchdog_init(nmi_hz) < 0) {
			__get_cpu_var(wd_enabled) = 0;
254 255
			return;
		}
256 257 258 259
		/* FALL THROUGH */
	case NMI_IO_APIC:
		__get_cpu_var(wd_enabled) = 1;
		atomic_inc(&nmi_active);
260 261 262
	}
}

263
void stop_apic_nmi_watchdog(void *unused)
264 265 266 267 268
{
	/* only support LOCAL and IO APICs for now */
	if ((nmi_watchdog != NMI_LOCAL_APIC) &&
	    (nmi_watchdog != NMI_IO_APIC))
	    	return;
269
	if (__get_cpu_var(wd_enabled) == 0)
270
		return;
271 272 273
	if (nmi_watchdog == NMI_LOCAL_APIC)
		lapic_watchdog_stop();
	__get_cpu_var(wd_enabled) = 0;
274
	atomic_dec(&nmi_active);
L
Linus Torvalds 已提交
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295
}

/*
 * the best way to detect whether a CPU has a 'hard lockup' problem
 * is to check it's local APIC timer IRQ counts. If they are not
 * changing then that CPU has some problem.
 *
 * as these watchdog NMI IRQs are generated on every CPU, we only
 * have to check the current processor.
 *
 * since NMIs don't listen to _any_ locks, we have to be extremely
 * careful not to rely on unsafe variables. The printk might lock
 * up though, so we have to break up any console locks first ...
 * [when there will be more tty-related locks, break them up
 *  here too!]
 */

static unsigned int
	last_irq_sums [NR_CPUS],
	alert_counter [NR_CPUS];

A
Andrew Morton 已提交
296
void touch_nmi_watchdog(void)
L
Linus Torvalds 已提交
297
{
298 299
	if (nmi_watchdog > 0) {
		unsigned cpu;
L
Linus Torvalds 已提交
300

301 302 303 304
		/*
		 * Just reset the alert counters, (other CPUs might be
		 * spinning on locks we hold):
		 */
A
Andrew Morton 已提交
305 306 307 308
		for_each_present_cpu(cpu) {
			if (alert_counter[cpu])
				alert_counter[cpu] = 0;
		}
309
	}
I
Ingo Molnar 已提交
310 311 312 313 314

	/*
	 * Tickle the softlockup detector too:
	 */
	touch_softlockup_watchdog();
L
Linus Torvalds 已提交
315
}
316
EXPORT_SYMBOL(touch_nmi_watchdog);
L
Linus Torvalds 已提交
317 318 319

extern void die_nmi(struct pt_regs *, const char *msg);

320
__kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
L
Linus Torvalds 已提交
321 322 323 324 325 326 327
{

	/*
	 * Since current_thread_info()-> is always on the stack, and we
	 * always switch the stack NMI-atomically, it's safe to use
	 * smp_processor_id().
	 */
328
	unsigned int sum;
329
	int touched = 0;
330
	int cpu = smp_processor_id();
H
Hiroshi Shimamoto 已提交
331
	int rc = 0;
332 333 334 335

	/* check for other users first */
	if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
			== NOTIFY_STOP) {
336
		rc = 1;
337 338
		touched = 1;
	}
L
Linus Torvalds 已提交
339

A
Andrew Morton 已提交
340 341 342 343 344 345 346 347 348 349
	if (cpu_isset(cpu, backtrace_mask)) {
		static DEFINE_SPINLOCK(lock);	/* Serialise the printks */

		spin_lock(&lock);
		printk("NMI backtrace for cpu %d\n", cpu);
		dump_stack();
		spin_unlock(&lock);
		cpu_clear(cpu, backtrace_mask);
	}

350 351 352 353
	/*
	 * Take the local apic timer and PIT/HPET into account. We don't
	 * know which one is active, when we have highres/dyntick on
	 */
354 355
	sum = per_cpu(irq_stat, cpu).apic_timer_irqs +
		per_cpu(irq_stat, cpu).irq0_irqs;
L
Linus Torvalds 已提交
356

357
	/* if the none of the timers isn't firing, this cpu isn't doing much */
358
	if (!touched && last_irq_sums[cpu] == sum) {
L
Linus Torvalds 已提交
359 360 361 362 363 364
		/*
		 * Ayiee, looks like this CPU is stuck ...
		 * wait a few IRQs (5 seconds) before doing the oops ...
		 */
		alert_counter[cpu]++;
		if (alert_counter[cpu] == 5*nmi_hz)
365 366 367
			/*
			 * die_nmi will return ONLY if NOTIFY_STOP happens..
			 */
368
			die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
369
	} else {
L
Linus Torvalds 已提交
370 371 372
		last_irq_sums[cpu] = sum;
		alert_counter[cpu] = 0;
	}
373
	/* see if the nmi watchdog went off */
374 375 376 377 378 379 380 381 382 383 384 385 386
	if (!__get_cpu_var(wd_enabled))
		return rc;
	switch (nmi_watchdog) {
	case NMI_LOCAL_APIC:
		rc |= lapic_wd_event(nmi_hz);
		break;
	case NMI_IO_APIC:
		/* don't know how to accurately check for this.
		 * just assume it was a watchdog timer interrupt
		 * This matches the old behaviour.
		 */
		rc = 1;
		break;
L
Linus Torvalds 已提交
387
	}
388
	return rc;
L
Linus Torvalds 已提交
389 390 391 392 393 394 395 396 397
}

#ifdef CONFIG_SYSCTL

static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
{
	unsigned char reason = get_nmi_reason();
	char buf[64];

398 399
	sprintf(buf, "NMI received for unknown reason %02x\n", reason);
	die_nmi(regs, buf);
L
Linus Torvalds 已提交
400 401 402
	return 0;
}

403
/*
404
 * proc handler for /proc/sys/kernel/nmi
405 406 407 408 409 410 411 412 413 414 415 416
 */
int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
			void __user *buffer, size_t *length, loff_t *ppos)
{
	int old_state;

	nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
	old_state = nmi_watchdog_enabled;
	proc_dointvec(table, write, file, buffer, length, ppos);
	if (!!old_state == !!nmi_watchdog_enabled)
		return 0;

417
	if (atomic_read(&nmi_active) < 0 || nmi_watchdog == NMI_DISABLED) {
418 419
		printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
		return -EIO;
420 421 422
	}

	if (nmi_watchdog == NMI_DEFAULT) {
423
		if (lapic_watchdog_ok())
424 425 426 427 428
			nmi_watchdog = NMI_LOCAL_APIC;
		else
			nmi_watchdog = NMI_IO_APIC;
	}

429
	if (nmi_watchdog == NMI_LOCAL_APIC) {
430 431 432 433 434 435 436 437 438 439 440 441
		if (nmi_watchdog_enabled)
			enable_lapic_nmi_watchdog();
		else
			disable_lapic_nmi_watchdog();
	} else {
		printk( KERN_WARNING
			"NMI watchdog doesn't know what hardware to touch\n");
		return -EIO;
	}
	return 0;
}

L
Linus Torvalds 已提交
442 443
#endif

444 445 446 447 448 449 450 451 452
int do_nmi_callback(struct pt_regs *regs, int cpu)
{
#ifdef CONFIG_SYSCTL
	if (unknown_nmi_panic)
		return unknown_nmi_panic_callback(regs, cpu);
#endif
	return 0;
}

A
Andrew Morton 已提交
453 454 455 456 457 458 459 460 461 462 463 464 465
void __trigger_all_cpu_backtrace(void)
{
	int i;

	backtrace_mask = cpu_online_map;
	/* Wait for up to 10 seconds for all CPUs to do the backtrace */
	for (i = 0; i < 10 * 1000; i++) {
		if (cpus_empty(backtrace_mask))
			break;
		mdelay(1);
	}
}

L
Linus Torvalds 已提交
466 467
EXPORT_SYMBOL(nmi_active);
EXPORT_SYMBOL(nmi_watchdog);