1/*
2 *  linux/arch/i386/nmi.c
3 *
4 *  NMI watchdog support on APIC systems
5 *
6 *  Started by Ingo Molnar <mingo@redhat.com>
7 *
8 *  Fixes:
9 *  Mikael Pettersson	: AMD K7 support for local APIC NMI watchdog.
10 *  Mikael Pettersson	: Power Management for local APIC NMI watchdog.
11 *  Mikael Pettersson	: Pentium 4 support for local APIC NMI watchdog.
12 *  Pavel Machek and
13 *  Mikael Pettersson	: PM converted to driver model. Disable/enable API.
14 */
15
16#include <linux/delay.h>
17#include <linux/interrupt.h>
18#include <linux/module.h>
19#include <linux/nmi.h>
20#include <linux/sysdev.h>
21#include <linux/sysctl.h>
22#include <linux/percpu.h>
23#include <linux/kprobes.h>
24#include <linux/cpumask.h>
25#include <linux/kernel_stat.h>
26#include <linux/kdebug.h>
27
28#include <asm/smp.h>
29#include <asm/nmi.h>
30
31#include "mach_traps.h"
32
33int unknown_nmi_panic;
34int nmi_watchdog_enabled;
35
36static cpumask_t backtrace_mask = CPU_MASK_NONE;
37
38/* nmi_active:
39 * >0: the lapic NMI watchdog is active, but can be disabled
40 * <0: the lapic NMI watchdog has not been set up, and cannot
41 *     be enabled
42 *  0: the lapic NMI watchdog is disabled, but can be enabled
43 */
44atomic_t nmi_active = ATOMIC_INIT(0);		/* oprofile uses this */
45
46unsigned int nmi_watchdog = NMI_DEFAULT;
47static unsigned int nmi_hz = HZ;
48
49static DEFINE_PER_CPU(short, wd_enabled);
50
51/* local prototypes */
52static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu);
53
54static int endflag __initdata = 0;
55
56#ifdef CONFIG_SMP
57/* The performance counters used by NMI_LOCAL_APIC don't trigger when
58 * the CPU is idle. To make sure the NMI watchdog really ticks on all
59 * CPUs during the test make them busy.
60 */
61static __init void nmi_cpu_busy(void *data)
62{
63	local_irq_enable_in_hardirq();
64	/* Intentionally don't use cpu_relax here. This is
65	   to make sure that the performance counter really ticks,
66	   even if there is a simulator or similar that catches the
67	   pause instruction. On a real HT machine this is fine because
68	   all other CPUs are busy with "useless" delay loops and don't
69	   care if they get somewhat less cycles. */
70	while (endflag == 0)
71		mb();
72}
73#endif
74
75static int __init check_nmi_watchdog(void)
76{
77	unsigned int *prev_nmi_count;
78	int cpu;
79
80	if ((nmi_watchdog == NMI_NONE) || (nmi_watchdog == NMI_DEFAULT))
81		return 0;
82
83	if (!atomic_read(&nmi_active))
84		return 0;
85
86	prev_nmi_count = kmalloc(NR_CPUS * sizeof(int), GFP_KERNEL);
87	if (!prev_nmi_count)
88		return -1;
89
90	printk(KERN_INFO "Testing NMI watchdog ... ");
91
92	if (nmi_watchdog == NMI_LOCAL_APIC)
93		smp_call_function(nmi_cpu_busy, (void *)&endflag, 0, 0);
94
95	for_each_possible_cpu(cpu)
96		prev_nmi_count[cpu] = per_cpu(irq_stat, cpu).__nmi_count;
97	local_irq_enable();
98	mdelay((20*1000)/nmi_hz); // wait 20 ticks
99
100	for_each_possible_cpu(cpu) {
101#ifdef CONFIG_SMP
102		/* Check cpu_callin_map here because that is set
103		   after the timer is started. */
104		if (!cpu_isset(cpu, cpu_callin_map))
105			continue;
106#endif
107		if (!per_cpu(wd_enabled, cpu))
108			continue;
109		if (nmi_count(cpu) - prev_nmi_count[cpu] <= 5) {
110			printk("CPU#%d: NMI appears to be stuck (%d->%d)!\n",
111				cpu,
112				prev_nmi_count[cpu],
113				nmi_count(cpu));
114			per_cpu(wd_enabled, cpu) = 0;
115			atomic_dec(&nmi_active);
116		}
117	}
118	if (!atomic_read(&nmi_active)) {
119		kfree(prev_nmi_count);
120		atomic_set(&nmi_active, -1);
121		return -1;
122	}
123	endflag = 1;
124	printk("OK.\n");
125
126	/* now that we know it works we can reduce NMI frequency to
127	   something more reasonable; makes a difference in some configs */
128	if (nmi_watchdog == NMI_LOCAL_APIC)
129		nmi_hz = lapic_adjust_nmi_hz(1);
130
131	kfree(prev_nmi_count);
132	return 0;
133}
134/* This needs to happen later in boot so counters are working */
135late_initcall(check_nmi_watchdog);
136
137static int __init setup_nmi_watchdog(char *str)
138{
139	int nmi;
140
141	get_option(&str, &nmi);
142
143	if ((nmi >= NMI_INVALID) || (nmi < NMI_NONE))
144		return 0;
145
146	nmi_watchdog = nmi;
147	return 1;
148}
149
150__setup("nmi_watchdog=", setup_nmi_watchdog);
151
152
153/* Suspend/resume support */
154
155#ifdef CONFIG_PM
156
157static int nmi_pm_active; /* nmi_active before suspend */
158
159static int lapic_nmi_suspend(struct sys_device *dev, pm_message_t state)
160{
161	/* only CPU0 goes here, other CPUs should be offline */
162	nmi_pm_active = atomic_read(&nmi_active);
163	stop_apic_nmi_watchdog(NULL);
164	BUG_ON(atomic_read(&nmi_active) != 0);
165	return 0;
166}
167
168static int lapic_nmi_resume(struct sys_device *dev)
169{
170	/* only CPU0 goes here, other CPUs should be offline */
171	if (nmi_pm_active > 0) {
172		setup_apic_nmi_watchdog(NULL);
173		touch_nmi_watchdog();
174	}
175	return 0;
176}
177
178
179static struct sysdev_class nmi_sysclass = {
180	set_kset_name("lapic_nmi"),
181	.resume		= lapic_nmi_resume,
182	.suspend	= lapic_nmi_suspend,
183};
184
185static struct sys_device device_lapic_nmi = {
186	.id	= 0,
187	.cls	= &nmi_sysclass,
188};
189
190static int __init init_lapic_nmi_sysfs(void)
191{
192	int error;
193
194	/* should really be a BUG_ON but b/c this is an
195	 * init call, it just doesn't work.  -dcz
196	 */
197	if (nmi_watchdog != NMI_LOCAL_APIC)
198		return 0;
199
200	if (atomic_read(&nmi_active) < 0)
201		return 0;
202
203	error = sysdev_class_register(&nmi_sysclass);
204	if (!error)
205		error = sysdev_register(&device_lapic_nmi);
206	return error;
207}
208/* must come after the local APIC's device_initcall() */
209late_initcall(init_lapic_nmi_sysfs);
210
211#endif	/* CONFIG_PM */
212
213static void __acpi_nmi_enable(void *__unused)
214{
215	apic_write_around(APIC_LVT0, APIC_DM_NMI);
216}
217
218/*
219 * Enable timer based NMIs on all CPUs:
220 */
221void acpi_nmi_enable(void)
222{
223	if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
224		on_each_cpu(__acpi_nmi_enable, NULL, 0, 1);
225}
226
227static void __acpi_nmi_disable(void *__unused)
228{
229	apic_write(APIC_LVT0, APIC_DM_NMI | APIC_LVT_MASKED);
230}
231
232/*
233 * Disable timer based NMIs on all CPUs:
234 */
235void acpi_nmi_disable(void)
236{
237	if (atomic_read(&nmi_active) && nmi_watchdog == NMI_IO_APIC)
238		on_each_cpu(__acpi_nmi_disable, NULL, 0, 1);
239}
240
241void setup_apic_nmi_watchdog (void *unused)
242{
243	if (__get_cpu_var(wd_enabled))
244 		return;
245
246	/* cheap hack to support suspend/resume */
247	/* if cpu0 is not active neither should the other cpus */
248	if ((smp_processor_id() != 0) && (atomic_read(&nmi_active) <= 0))
249		return;
250
251	switch (nmi_watchdog) {
252	case NMI_LOCAL_APIC:
253		__get_cpu_var(wd_enabled) = 1; /* enable it before to avoid race with handler */
254		if (lapic_watchdog_init(nmi_hz) < 0) {
255			__get_cpu_var(wd_enabled) = 0;
256			return;
257		}
258		/* FALL THROUGH */
259	case NMI_IO_APIC:
260		__get_cpu_var(wd_enabled) = 1;
261		atomic_inc(&nmi_active);
262	}
263}
264
265void stop_apic_nmi_watchdog(void *unused)
266{
267	/* only support LOCAL and IO APICs for now */
268	if ((nmi_watchdog != NMI_LOCAL_APIC) &&
269	    (nmi_watchdog != NMI_IO_APIC))
270	    	return;
271	if (__get_cpu_var(wd_enabled) == 0)
272		return;
273	if (nmi_watchdog == NMI_LOCAL_APIC)
274		lapic_watchdog_stop();
275	__get_cpu_var(wd_enabled) = 0;
276	atomic_dec(&nmi_active);
277}
278
279/*
280 * the best way to detect whether a CPU has a 'hard lockup' problem
281 * is to check it's local APIC timer IRQ counts. If they are not
282 * changing then that CPU has some problem.
283 *
284 * as these watchdog NMI IRQs are generated on every CPU, we only
285 * have to check the current processor.
286 *
287 * since NMIs don't listen to _any_ locks, we have to be extremely
288 * careful not to rely on unsafe variables. The printk might lock
289 * up though, so we have to break up any console locks first ...
290 * [when there will be more tty-related locks, break them up
291 *  here too!]
292 */
293
294static unsigned int
295	last_irq_sums [NR_CPUS],
296	alert_counter [NR_CPUS];
297
298void touch_nmi_watchdog (void)
299{
300	if (nmi_watchdog > 0) {
301		unsigned cpu;
302
303		/*
304		 * Just reset the alert counters, (other CPUs might be
305		 * spinning on locks we hold):
306		 */
307		for_each_present_cpu (cpu)
308			alert_counter[cpu] = 0;
309	}
310
311	/*
312	 * Tickle the softlockup detector too:
313	 */
314	touch_softlockup_watchdog();
315}
316EXPORT_SYMBOL(touch_nmi_watchdog);
317
318extern void die_nmi(struct pt_regs *, const char *msg);
319
320__kprobes int nmi_watchdog_tick(struct pt_regs * regs, unsigned reason)
321{
322
323	/*
324	 * Since current_thread_info()-> is always on the stack, and we
325	 * always switch the stack NMI-atomically, it's safe to use
326	 * smp_processor_id().
327	 */
328	unsigned int sum;
329	int touched = 0;
330	int cpu = smp_processor_id();
331	int rc=0;
332
333	/* check for other users first */
334	if (notify_die(DIE_NMI, "nmi", regs, reason, 2, SIGINT)
335			== NOTIFY_STOP) {
336		rc = 1;
337		touched = 1;
338	}
339
340	if (cpu_isset(cpu, backtrace_mask)) {
341		static DEFINE_SPINLOCK(lock);	/* Serialise the printks */
342
343		spin_lock(&lock);
344		printk("NMI backtrace for cpu %d\n", cpu);
345		dump_stack();
346		spin_unlock(&lock);
347		cpu_clear(cpu, backtrace_mask);
348	}
349
350	/*
351	 * Take the local apic timer and PIT/HPET into account. We don't
352	 * know which one is active, when we have highres/dyntick on
353	 */
354	sum = per_cpu(irq_stat, cpu).apic_timer_irqs + kstat_irqs(0);
355
356	/* if the none of the timers isn't firing, this cpu isn't doing much */
357	if (!touched && last_irq_sums[cpu] == sum) {
358		/*
359		 * Ayiee, looks like this CPU is stuck ...
360		 * wait a few IRQs (5 seconds) before doing the oops ...
361		 */
362		alert_counter[cpu]++;
363		if (alert_counter[cpu] == 5*nmi_hz)
364			/*
365			 * die_nmi will return ONLY if NOTIFY_STOP happens..
366			 */
367			die_nmi(regs, "BUG: NMI Watchdog detected LOCKUP");
368	} else {
369		last_irq_sums[cpu] = sum;
370		alert_counter[cpu] = 0;
371	}
372	/* see if the nmi watchdog went off */
373	if (!__get_cpu_var(wd_enabled))
374		return rc;
375	switch (nmi_watchdog) {
376	case NMI_LOCAL_APIC:
377		rc |= lapic_wd_event(nmi_hz);
378		break;
379	case NMI_IO_APIC:
380		/* don't know how to accurately check for this.
381		 * just assume it was a watchdog timer interrupt
382		 * This matches the old behaviour.
383		 */
384		rc = 1;
385		break;
386	}
387	return rc;
388}
389
390int do_nmi_callback(struct pt_regs * regs, int cpu)
391{
392#ifdef CONFIG_SYSCTL
393	if (unknown_nmi_panic)
394		return unknown_nmi_panic_callback(regs, cpu);
395#endif
396	return 0;
397}
398
399#ifdef CONFIG_SYSCTL
400
401static int unknown_nmi_panic_callback(struct pt_regs *regs, int cpu)
402{
403	unsigned char reason = get_nmi_reason();
404	char buf[64];
405
406	sprintf(buf, "NMI received for unknown reason %02x\n", reason);
407	die_nmi(regs, buf);
408	return 0;
409}
410
411/*
412 * proc handler for /proc/sys/kernel/nmi
413 */
414int proc_nmi_enabled(struct ctl_table *table, int write, struct file *file,
415			void __user *buffer, size_t *length, loff_t *ppos)
416{
417	int old_state;
418
419	nmi_watchdog_enabled = (atomic_read(&nmi_active) > 0) ? 1 : 0;
420	old_state = nmi_watchdog_enabled;
421	proc_dointvec(table, write, file, buffer, length, ppos);
422	if (!!old_state == !!nmi_watchdog_enabled)
423		return 0;
424
425	if (atomic_read(&nmi_active) < 0) {
426		printk( KERN_WARNING "NMI watchdog is permanently disabled\n");
427		return -EIO;
428	}
429
430	if (nmi_watchdog == NMI_DEFAULT) {
431		if (lapic_watchdog_ok())
432			nmi_watchdog = NMI_LOCAL_APIC;
433		else
434			nmi_watchdog = NMI_IO_APIC;
435	}
436
437	if (nmi_watchdog == NMI_LOCAL_APIC) {
438		if (nmi_watchdog_enabled)
439			enable_lapic_nmi_watchdog();
440		else
441			disable_lapic_nmi_watchdog();
442	} else {
443		printk( KERN_WARNING
444			"NMI watchdog doesn't know what hardware to touch\n");
445		return -EIO;
446	}
447	return 0;
448}
449
450#endif
451
452void __trigger_all_cpu_backtrace(void)
453{
454	int i;
455
456	backtrace_mask = cpu_online_map;
457	/* Wait for up to 10 seconds for all CPUs to do the backtrace */
458	for (i = 0; i < 10 * 1000; i++) {
459		if (cpus_empty(backtrace_mask))
460			break;
461		mdelay(1);
462	}
463}
464
465EXPORT_SYMBOL(nmi_active);
466EXPORT_SYMBOL(nmi_watchdog);
467