xref: /linux/arch/x86/kernel/nmi.c (revision 424f0750edd5af866f80f5e65998e0610503cb5c)
1 /*
2  *  Copyright (C) 1991, 1992  Linus Torvalds
3  *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
4  *  Copyright (C) 2011	Don Zickus Red Hat, Inc.
5  *
6  *  Pentium III FXSR, SSE support
7  *	Gareth Hughes <gareth@valinux.com>, May 2000
8  */
9 
10 /*
11  * Handle hardware traps and faults.
12  */
13 #include <linux/spinlock.h>
14 #include <linux/kprobes.h>
15 #include <linux/kdebug.h>
16 #include <linux/nmi.h>
17 #include <linux/delay.h>
18 #include <linux/hardirq.h>
19 #include <linux/slab.h>
20 #include <linux/export.h>
21 
22 #include <linux/mca.h>
23 
24 #if defined(CONFIG_EDAC)
25 #include <linux/edac.h>
26 #endif
27 
28 #include <linux/atomic.h>
29 #include <asm/traps.h>
30 #include <asm/mach_traps.h>
31 #include <asm/nmi.h>
32 
33 #define NMI_MAX_NAMELEN	16
34 struct nmiaction {
35 	struct list_head list;
36 	nmi_handler_t handler;
37 	unsigned int flags;
38 	char *name;
39 };
40 
41 struct nmi_desc {
42 	spinlock_t lock;
43 	struct list_head head;
44 };
45 
46 static struct nmi_desc nmi_desc[NMI_MAX] =
47 {
48 	{
49 		.lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
50 		.head = LIST_HEAD_INIT(nmi_desc[0].head),
51 	},
52 	{
53 		.lock = __SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
54 		.head = LIST_HEAD_INIT(nmi_desc[1].head),
55 	},
56 
57 };
58 
59 struct nmi_stats {
60 	unsigned int normal;
61 	unsigned int unknown;
62 	unsigned int external;
63 	unsigned int swallow;
64 };
65 
66 static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
67 
68 static int ignore_nmis;
69 
70 int unknown_nmi_panic;
71 /*
72  * Prevent NMI reason port (0x61) being accessed simultaneously, can
73  * only be used in NMI handler.
74  */
75 static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
76 
77 static int __init setup_unknown_nmi_panic(char *str)
78 {
79 	unknown_nmi_panic = 1;
80 	return 1;
81 }
82 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
83 
84 #define nmi_to_desc(type) (&nmi_desc[type])
85 
86 static int notrace __kprobes nmi_handle(unsigned int type, struct pt_regs *regs, bool b2b)
87 {
88 	struct nmi_desc *desc = nmi_to_desc(type);
89 	struct nmiaction *a;
90 	int handled=0;
91 
92 	rcu_read_lock();
93 
94 	/*
95 	 * NMIs are edge-triggered, which means if you have enough
96 	 * of them concurrently, you can lose some because only one
97 	 * can be latched at any given time.  Walk the whole list
98 	 * to handle those situations.
99 	 */
100 	list_for_each_entry_rcu(a, &desc->head, list)
101 		handled += a->handler(type, regs);
102 
103 	rcu_read_unlock();
104 
105 	/* return total number of NMI events handled */
106 	return handled;
107 }
108 
109 static int __setup_nmi(unsigned int type, struct nmiaction *action)
110 {
111 	struct nmi_desc *desc = nmi_to_desc(type);
112 	unsigned long flags;
113 
114 	spin_lock_irqsave(&desc->lock, flags);
115 
116 	/*
117 	 * most handlers of type NMI_UNKNOWN never return because
118 	 * they just assume the NMI is theirs.  Just a sanity check
119 	 * to manage expectations
120 	 */
121 	WARN_ON_ONCE(type == NMI_UNKNOWN && !list_empty(&desc->head));
122 
123 	/*
124 	 * some handlers need to be executed first otherwise a fake
125 	 * event confuses some handlers (kdump uses this flag)
126 	 */
127 	if (action->flags & NMI_FLAG_FIRST)
128 		list_add_rcu(&action->list, &desc->head);
129 	else
130 		list_add_tail_rcu(&action->list, &desc->head);
131 
132 	spin_unlock_irqrestore(&desc->lock, flags);
133 	return 0;
134 }
135 
136 static struct nmiaction *__free_nmi(unsigned int type, const char *name)
137 {
138 	struct nmi_desc *desc = nmi_to_desc(type);
139 	struct nmiaction *n;
140 	unsigned long flags;
141 
142 	spin_lock_irqsave(&desc->lock, flags);
143 
144 	list_for_each_entry_rcu(n, &desc->head, list) {
145 		/*
146 		 * the name passed in to describe the nmi handler
147 		 * is used as the lookup key
148 		 */
149 		if (!strcmp(n->name, name)) {
150 			WARN(in_nmi(),
151 				"Trying to free NMI (%s) from NMI context!\n", n->name);
152 			list_del_rcu(&n->list);
153 			break;
154 		}
155 	}
156 
157 	spin_unlock_irqrestore(&desc->lock, flags);
158 	synchronize_rcu();
159 	return (n);
160 }
161 
162 int register_nmi_handler(unsigned int type, nmi_handler_t handler,
163 			unsigned long nmiflags, const char *devname)
164 {
165 	struct nmiaction *action;
166 	int retval = -ENOMEM;
167 
168 	if (!handler)
169 		return -EINVAL;
170 
171 	action = kzalloc(sizeof(struct nmiaction), GFP_KERNEL);
172 	if (!action)
173 		goto fail_action;
174 
175 	action->handler = handler;
176 	action->flags = nmiflags;
177 	action->name = kstrndup(devname, NMI_MAX_NAMELEN, GFP_KERNEL);
178 	if (!action->name)
179 		goto fail_action_name;
180 
181 	retval = __setup_nmi(type, action);
182 
183 	if (retval)
184 		goto fail_setup_nmi;
185 
186 	return retval;
187 
188 fail_setup_nmi:
189 	kfree(action->name);
190 fail_action_name:
191 	kfree(action);
192 fail_action:
193 
194 	return retval;
195 }
196 EXPORT_SYMBOL_GPL(register_nmi_handler);
197 
198 void unregister_nmi_handler(unsigned int type, const char *name)
199 {
200 	struct nmiaction *a;
201 
202 	a = __free_nmi(type, name);
203 	if (a) {
204 		kfree(a->name);
205 		kfree(a);
206 	}
207 }
208 
209 EXPORT_SYMBOL_GPL(unregister_nmi_handler);
210 
211 static notrace __kprobes void
212 pci_serr_error(unsigned char reason, struct pt_regs *regs)
213 {
214 	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
215 		 reason, smp_processor_id());
216 
217 	/*
218 	 * On some machines, PCI SERR line is used to report memory
219 	 * errors. EDAC makes use of it.
220 	 */
221 #if defined(CONFIG_EDAC)
222 	if (edac_handler_set()) {
223 		edac_atomic_assert_error();
224 		return;
225 	}
226 #endif
227 
228 	if (panic_on_unrecovered_nmi)
229 		panic("NMI: Not continuing");
230 
231 	pr_emerg("Dazed and confused, but trying to continue\n");
232 
233 	/* Clear and disable the PCI SERR error line. */
234 	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
235 	outb(reason, NMI_REASON_PORT);
236 }
237 
238 static notrace __kprobes void
239 io_check_error(unsigned char reason, struct pt_regs *regs)
240 {
241 	unsigned long i;
242 
243 	pr_emerg(
244 	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
245 		 reason, smp_processor_id());
246 	show_registers(regs);
247 
248 	if (panic_on_io_nmi)
249 		panic("NMI IOCK error: Not continuing");
250 
251 	/* Re-enable the IOCK line, wait for a few seconds */
252 	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
253 	outb(reason, NMI_REASON_PORT);
254 
255 	i = 20000;
256 	while (--i) {
257 		touch_nmi_watchdog();
258 		udelay(100);
259 	}
260 
261 	reason &= ~NMI_REASON_CLEAR_IOCHK;
262 	outb(reason, NMI_REASON_PORT);
263 }
264 
265 static notrace __kprobes void
266 unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
267 {
268 	int handled;
269 
270 	/*
271 	 * Use 'false' as back-to-back NMIs are dealt with one level up.
272 	 * Of course this makes having multiple 'unknown' handlers useless
273 	 * as only the first one is ever run (unless it can actually determine
274 	 * if it caused the NMI)
275 	 */
276 	handled = nmi_handle(NMI_UNKNOWN, regs, false);
277 	if (handled) {
278 		__this_cpu_add(nmi_stats.unknown, handled);
279 		return;
280 	}
281 
282 	__this_cpu_add(nmi_stats.unknown, 1);
283 
284 #ifdef CONFIG_MCA
285 	/*
286 	 * Might actually be able to figure out what the guilty party
287 	 * is:
288 	 */
289 	if (MCA_bus) {
290 		mca_handle_nmi();
291 		return;
292 	}
293 #endif
294 	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
295 		 reason, smp_processor_id());
296 
297 	pr_emerg("Do you have a strange power saving mode enabled?\n");
298 	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
299 		panic("NMI: Not continuing");
300 
301 	pr_emerg("Dazed and confused, but trying to continue\n");
302 }
303 
304 static DEFINE_PER_CPU(bool, swallow_nmi);
305 static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
306 
307 static notrace __kprobes void default_do_nmi(struct pt_regs *regs)
308 {
309 	unsigned char reason = 0;
310 	int handled;
311 	bool b2b = false;
312 
313 	/*
314 	 * CPU-specific NMI must be processed before non-CPU-specific
315 	 * NMI, otherwise we may lose it, because the CPU-specific
316 	 * NMI can not be detected/processed on other CPUs.
317 	 */
318 
319 	/*
320 	 * Back-to-back NMIs are interesting because they can either
321 	 * be two NMI or more than two NMIs (any thing over two is dropped
322 	 * due to NMI being edge-triggered).  If this is the second half
323 	 * of the back-to-back NMI, assume we dropped things and process
324 	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
325 	 */
326 	if (regs->ip == __this_cpu_read(last_nmi_rip))
327 		b2b = true;
328 	else
329 		__this_cpu_write(swallow_nmi, false);
330 
331 	__this_cpu_write(last_nmi_rip, regs->ip);
332 
333 	handled = nmi_handle(NMI_LOCAL, regs, b2b);
334 	__this_cpu_add(nmi_stats.normal, handled);
335 	if (handled) {
336 		/*
337 		 * There are cases when a NMI handler handles multiple
338 		 * events in the current NMI.  One of these events may
339 		 * be queued for in the next NMI.  Because the event is
340 		 * already handled, the next NMI will result in an unknown
341 		 * NMI.  Instead lets flag this for a potential NMI to
342 		 * swallow.
343 		 */
344 		if (handled > 1)
345 			__this_cpu_write(swallow_nmi, true);
346 		return;
347 	}
348 
349 	/* Non-CPU-specific NMI: NMI sources can be processed on any CPU */
350 	raw_spin_lock(&nmi_reason_lock);
351 	reason = get_nmi_reason();
352 
353 	if (reason & NMI_REASON_MASK) {
354 		if (reason & NMI_REASON_SERR)
355 			pci_serr_error(reason, regs);
356 		else if (reason & NMI_REASON_IOCHK)
357 			io_check_error(reason, regs);
358 #ifdef CONFIG_X86_32
359 		/*
360 		 * Reassert NMI in case it became active
361 		 * meanwhile as it's edge-triggered:
362 		 */
363 		reassert_nmi();
364 #endif
365 		__this_cpu_add(nmi_stats.external, 1);
366 		raw_spin_unlock(&nmi_reason_lock);
367 		return;
368 	}
369 	raw_spin_unlock(&nmi_reason_lock);
370 
371 	/*
372 	 * Only one NMI can be latched at a time.  To handle
373 	 * this we may process multiple nmi handlers at once to
374 	 * cover the case where an NMI is dropped.  The downside
375 	 * to this approach is we may process an NMI prematurely,
376 	 * while its real NMI is sitting latched.  This will cause
377 	 * an unknown NMI on the next run of the NMI processing.
378 	 *
379 	 * We tried to flag that condition above, by setting the
380 	 * swallow_nmi flag when we process more than one event.
381 	 * This condition is also only present on the second half
382 	 * of a back-to-back NMI, so we flag that condition too.
383 	 *
384 	 * If both are true, we assume we already processed this
385 	 * NMI previously and we swallow it.  Otherwise we reset
386 	 * the logic.
387 	 *
388 	 * There are scenarios where we may accidentally swallow
389 	 * a 'real' unknown NMI.  For example, while processing
390 	 * a perf NMI another perf NMI comes in along with a
391 	 * 'real' unknown NMI.  These two NMIs get combined into
392 	 * one (as descibed above).  When the next NMI gets
393 	 * processed, it will be flagged by perf as handled, but
394 	 * noone will know that there was a 'real' unknown NMI sent
395 	 * also.  As a result it gets swallowed.  Or if the first
396 	 * perf NMI returns two events handled then the second
397 	 * NMI will get eaten by the logic below, again losing a
398 	 * 'real' unknown NMI.  But this is the best we can do
399 	 * for now.
400 	 */
401 	if (b2b && __this_cpu_read(swallow_nmi))
402 		__this_cpu_add(nmi_stats.swallow, 1);
403 	else
404 		unknown_nmi_error(reason, regs);
405 }
406 
407 dotraplinkage notrace __kprobes void
408 do_nmi(struct pt_regs *regs, long error_code)
409 {
410 	nmi_enter();
411 
412 	inc_irq_stat(__nmi_count);
413 
414 	if (!ignore_nmis)
415 		default_do_nmi(regs);
416 
417 	nmi_exit();
418 }
419 
420 void stop_nmi(void)
421 {
422 	ignore_nmis++;
423 }
424 
425 void restart_nmi(void)
426 {
427 	ignore_nmis--;
428 }
429 
430 /* reset the back-to-back NMI logic */
431 void local_touch_nmi(void)
432 {
433 	__this_cpu_write(last_nmi_rip, 0);
434 }
435