xref: /linux/arch/x86/kernel/nmi.c (revision 73b0140bf0fe9df90fb267c00673c4b9bf285430)
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/sched/debug.h>
17 #include <linux/nmi.h>
18 #include <linux/debugfs.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/ratelimit.h>
22 #include <linux/slab.h>
23 #include <linux/export.h>
24 #include <linux/atomic.h>
25 #include <linux/sched/clock.h>
26 
27 #if defined(CONFIG_EDAC)
28 #include <linux/edac.h>
29 #endif
30 
31 #include <asm/cpu_entry_area.h>
32 #include <asm/traps.h>
33 #include <asm/mach_traps.h>
34 #include <asm/nmi.h>
35 #include <asm/x86_init.h>
36 #include <asm/reboot.h>
37 #include <asm/cache.h>
38 
39 #define CREATE_TRACE_POINTS
40 #include <trace/events/nmi.h>
41 
42 struct nmi_desc {
43 	raw_spinlock_t lock;
44 	struct list_head head;
45 };
46 
47 static struct nmi_desc nmi_desc[NMI_MAX] =
48 {
49 	{
50 		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
51 		.head = LIST_HEAD_INIT(nmi_desc[0].head),
52 	},
53 	{
54 		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
55 		.head = LIST_HEAD_INIT(nmi_desc[1].head),
56 	},
57 	{
58 		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
59 		.head = LIST_HEAD_INIT(nmi_desc[2].head),
60 	},
61 	{
62 		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
63 		.head = LIST_HEAD_INIT(nmi_desc[3].head),
64 	},
65 
66 };
67 
68 struct nmi_stats {
69 	unsigned int normal;
70 	unsigned int unknown;
71 	unsigned int external;
72 	unsigned int swallow;
73 };
74 
75 static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
76 
77 static int ignore_nmis __read_mostly;
78 
79 int unknown_nmi_panic;
80 /*
81  * Prevent NMI reason port (0x61) being accessed simultaneously, can
82  * only be used in NMI handler.
83  */
84 static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
85 
86 static int __init setup_unknown_nmi_panic(char *str)
87 {
88 	unknown_nmi_panic = 1;
89 	return 1;
90 }
91 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
92 
93 #define nmi_to_desc(type) (&nmi_desc[type])
94 
95 static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
96 
97 static int __init nmi_warning_debugfs(void)
98 {
99 	debugfs_create_u64("nmi_longest_ns", 0644,
100 			arch_debugfs_dir, &nmi_longest_ns);
101 	return 0;
102 }
103 fs_initcall(nmi_warning_debugfs);
104 
105 static void nmi_max_handler(struct irq_work *w)
106 {
107 	struct nmiaction *a = container_of(w, struct nmiaction, irq_work);
108 	int remainder_ns, decimal_msecs;
109 	u64 whole_msecs = READ_ONCE(a->max_duration);
110 
111 	remainder_ns = do_div(whole_msecs, (1000 * 1000));
112 	decimal_msecs = remainder_ns / 1000;
113 
114 	printk_ratelimited(KERN_INFO
115 		"INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
116 		a->handler, whole_msecs, decimal_msecs);
117 }
118 
119 static int nmi_handle(unsigned int type, struct pt_regs *regs)
120 {
121 	struct nmi_desc *desc = nmi_to_desc(type);
122 	struct nmiaction *a;
123 	int handled=0;
124 
125 	rcu_read_lock();
126 
127 	/*
128 	 * NMIs are edge-triggered, which means if you have enough
129 	 * of them concurrently, you can lose some because only one
130 	 * can be latched at any given time.  Walk the whole list
131 	 * to handle those situations.
132 	 */
133 	list_for_each_entry_rcu(a, &desc->head, list) {
134 		int thishandled;
135 		u64 delta;
136 
137 		delta = sched_clock();
138 		thishandled = a->handler(type, regs);
139 		handled += thishandled;
140 		delta = sched_clock() - delta;
141 		trace_nmi_handler(a->handler, (int)delta, thishandled);
142 
143 		if (delta < nmi_longest_ns || delta < a->max_duration)
144 			continue;
145 
146 		a->max_duration = delta;
147 		irq_work_queue(&a->irq_work);
148 	}
149 
150 	rcu_read_unlock();
151 
152 	/* return total number of NMI events handled */
153 	return handled;
154 }
155 NOKPROBE_SYMBOL(nmi_handle);
156 
157 int __register_nmi_handler(unsigned int type, struct nmiaction *action)
158 {
159 	struct nmi_desc *desc = nmi_to_desc(type);
160 	unsigned long flags;
161 
162 	if (!action->handler)
163 		return -EINVAL;
164 
165 	init_irq_work(&action->irq_work, nmi_max_handler);
166 
167 	raw_spin_lock_irqsave(&desc->lock, flags);
168 
169 	/*
170 	 * Indicate if there are multiple registrations on the
171 	 * internal NMI handler call chains (SERR and IO_CHECK).
172 	 */
173 	WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
174 	WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
175 
176 	/*
177 	 * some handlers need to be executed first otherwise a fake
178 	 * event confuses some handlers (kdump uses this flag)
179 	 */
180 	if (action->flags & NMI_FLAG_FIRST)
181 		list_add_rcu(&action->list, &desc->head);
182 	else
183 		list_add_tail_rcu(&action->list, &desc->head);
184 
185 	raw_spin_unlock_irqrestore(&desc->lock, flags);
186 	return 0;
187 }
188 EXPORT_SYMBOL(__register_nmi_handler);
189 
190 void unregister_nmi_handler(unsigned int type, const char *name)
191 {
192 	struct nmi_desc *desc = nmi_to_desc(type);
193 	struct nmiaction *n;
194 	unsigned long flags;
195 
196 	raw_spin_lock_irqsave(&desc->lock, flags);
197 
198 	list_for_each_entry_rcu(n, &desc->head, list) {
199 		/*
200 		 * the name passed in to describe the nmi handler
201 		 * is used as the lookup key
202 		 */
203 		if (!strcmp(n->name, name)) {
204 			WARN(in_nmi(),
205 				"Trying to free NMI (%s) from NMI context!\n", n->name);
206 			list_del_rcu(&n->list);
207 			break;
208 		}
209 	}
210 
211 	raw_spin_unlock_irqrestore(&desc->lock, flags);
212 	synchronize_rcu();
213 }
214 EXPORT_SYMBOL_GPL(unregister_nmi_handler);
215 
216 static void
217 pci_serr_error(unsigned char reason, struct pt_regs *regs)
218 {
219 	/* check to see if anyone registered against these types of errors */
220 	if (nmi_handle(NMI_SERR, regs))
221 		return;
222 
223 	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
224 		 reason, smp_processor_id());
225 
226 	if (panic_on_unrecovered_nmi)
227 		nmi_panic(regs, "NMI: Not continuing");
228 
229 	pr_emerg("Dazed and confused, but trying to continue\n");
230 
231 	/* Clear and disable the PCI SERR error line. */
232 	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
233 	outb(reason, NMI_REASON_PORT);
234 }
235 NOKPROBE_SYMBOL(pci_serr_error);
236 
237 static void
238 io_check_error(unsigned char reason, struct pt_regs *regs)
239 {
240 	unsigned long i;
241 
242 	/* check to see if anyone registered against these types of errors */
243 	if (nmi_handle(NMI_IO_CHECK, regs))
244 		return;
245 
246 	pr_emerg(
247 	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
248 		 reason, smp_processor_id());
249 	show_regs(regs);
250 
251 	if (panic_on_io_nmi) {
252 		nmi_panic(regs, "NMI IOCK error: Not continuing");
253 
254 		/*
255 		 * If we end up here, it means we have received an NMI while
256 		 * processing panic(). Simply return without delaying and
257 		 * re-enabling NMIs.
258 		 */
259 		return;
260 	}
261 
262 	/* Re-enable the IOCK line, wait for a few seconds */
263 	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
264 	outb(reason, NMI_REASON_PORT);
265 
266 	i = 20000;
267 	while (--i) {
268 		touch_nmi_watchdog();
269 		udelay(100);
270 	}
271 
272 	reason &= ~NMI_REASON_CLEAR_IOCHK;
273 	outb(reason, NMI_REASON_PORT);
274 }
275 NOKPROBE_SYMBOL(io_check_error);
276 
277 static void
278 unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
279 {
280 	int handled;
281 
282 	/*
283 	 * Use 'false' as back-to-back NMIs are dealt with one level up.
284 	 * Of course this makes having multiple 'unknown' handlers useless
285 	 * as only the first one is ever run (unless it can actually determine
286 	 * if it caused the NMI)
287 	 */
288 	handled = nmi_handle(NMI_UNKNOWN, regs);
289 	if (handled) {
290 		__this_cpu_add(nmi_stats.unknown, handled);
291 		return;
292 	}
293 
294 	__this_cpu_add(nmi_stats.unknown, 1);
295 
296 	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
297 		 reason, smp_processor_id());
298 
299 	pr_emerg("Do you have a strange power saving mode enabled?\n");
300 	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
301 		nmi_panic(regs, "NMI: Not continuing");
302 
303 	pr_emerg("Dazed and confused, but trying to continue\n");
304 }
305 NOKPROBE_SYMBOL(unknown_nmi_error);
306 
307 static DEFINE_PER_CPU(bool, swallow_nmi);
308 static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
309 
310 static void default_do_nmi(struct pt_regs *regs)
311 {
312 	unsigned char reason = 0;
313 	int handled;
314 	bool b2b = false;
315 
316 	/*
317 	 * CPU-specific NMI must be processed before non-CPU-specific
318 	 * NMI, otherwise we may lose it, because the CPU-specific
319 	 * NMI can not be detected/processed on other CPUs.
320 	 */
321 
322 	/*
323 	 * Back-to-back NMIs are interesting because they can either
324 	 * be two NMI or more than two NMIs (any thing over two is dropped
325 	 * due to NMI being edge-triggered).  If this is the second half
326 	 * of the back-to-back NMI, assume we dropped things and process
327 	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
328 	 */
329 	if (regs->ip == __this_cpu_read(last_nmi_rip))
330 		b2b = true;
331 	else
332 		__this_cpu_write(swallow_nmi, false);
333 
334 	__this_cpu_write(last_nmi_rip, regs->ip);
335 
336 	handled = nmi_handle(NMI_LOCAL, regs);
337 	__this_cpu_add(nmi_stats.normal, handled);
338 	if (handled) {
339 		/*
340 		 * There are cases when a NMI handler handles multiple
341 		 * events in the current NMI.  One of these events may
342 		 * be queued for in the next NMI.  Because the event is
343 		 * already handled, the next NMI will result in an unknown
344 		 * NMI.  Instead lets flag this for a potential NMI to
345 		 * swallow.
346 		 */
347 		if (handled > 1)
348 			__this_cpu_write(swallow_nmi, true);
349 		return;
350 	}
351 
352 	/*
353 	 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
354 	 *
355 	 * Another CPU may be processing panic routines while holding
356 	 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
357 	 * and if so, call its callback directly.  If there is no CPU preparing
358 	 * crash dump, we simply loop here.
359 	 */
360 	while (!raw_spin_trylock(&nmi_reason_lock)) {
361 		run_crash_ipi_callback(regs);
362 		cpu_relax();
363 	}
364 
365 	reason = x86_platform.get_nmi_reason();
366 
367 	if (reason & NMI_REASON_MASK) {
368 		if (reason & NMI_REASON_SERR)
369 			pci_serr_error(reason, regs);
370 		else if (reason & NMI_REASON_IOCHK)
371 			io_check_error(reason, regs);
372 #ifdef CONFIG_X86_32
373 		/*
374 		 * Reassert NMI in case it became active
375 		 * meanwhile as it's edge-triggered:
376 		 */
377 		reassert_nmi();
378 #endif
379 		__this_cpu_add(nmi_stats.external, 1);
380 		raw_spin_unlock(&nmi_reason_lock);
381 		return;
382 	}
383 	raw_spin_unlock(&nmi_reason_lock);
384 
385 	/*
386 	 * Only one NMI can be latched at a time.  To handle
387 	 * this we may process multiple nmi handlers at once to
388 	 * cover the case where an NMI is dropped.  The downside
389 	 * to this approach is we may process an NMI prematurely,
390 	 * while its real NMI is sitting latched.  This will cause
391 	 * an unknown NMI on the next run of the NMI processing.
392 	 *
393 	 * We tried to flag that condition above, by setting the
394 	 * swallow_nmi flag when we process more than one event.
395 	 * This condition is also only present on the second half
396 	 * of a back-to-back NMI, so we flag that condition too.
397 	 *
398 	 * If both are true, we assume we already processed this
399 	 * NMI previously and we swallow it.  Otherwise we reset
400 	 * the logic.
401 	 *
402 	 * There are scenarios where we may accidentally swallow
403 	 * a 'real' unknown NMI.  For example, while processing
404 	 * a perf NMI another perf NMI comes in along with a
405 	 * 'real' unknown NMI.  These two NMIs get combined into
406 	 * one (as descibed above).  When the next NMI gets
407 	 * processed, it will be flagged by perf as handled, but
408 	 * noone will know that there was a 'real' unknown NMI sent
409 	 * also.  As a result it gets swallowed.  Or if the first
410 	 * perf NMI returns two events handled then the second
411 	 * NMI will get eaten by the logic below, again losing a
412 	 * 'real' unknown NMI.  But this is the best we can do
413 	 * for now.
414 	 */
415 	if (b2b && __this_cpu_read(swallow_nmi))
416 		__this_cpu_add(nmi_stats.swallow, 1);
417 	else
418 		unknown_nmi_error(reason, regs);
419 }
420 NOKPROBE_SYMBOL(default_do_nmi);
421 
422 /*
423  * NMIs can page fault or hit breakpoints which will cause it to lose
424  * its NMI context with the CPU when the breakpoint or page fault does an IRET.
425  *
426  * As a result, NMIs can nest if NMIs get unmasked due an IRET during
427  * NMI processing.  On x86_64, the asm glue protects us from nested NMIs
428  * if the outer NMI came from kernel mode, but we can still nest if the
429  * outer NMI came from user mode.
430  *
431  * To handle these nested NMIs, we have three states:
432  *
433  *  1) not running
434  *  2) executing
435  *  3) latched
436  *
437  * When no NMI is in progress, it is in the "not running" state.
438  * When an NMI comes in, it goes into the "executing" state.
439  * Normally, if another NMI is triggered, it does not interrupt
440  * the running NMI and the HW will simply latch it so that when
441  * the first NMI finishes, it will restart the second NMI.
442  * (Note, the latch is binary, thus multiple NMIs triggering,
443  *  when one is running, are ignored. Only one NMI is restarted.)
444  *
445  * If an NMI executes an iret, another NMI can preempt it. We do not
446  * want to allow this new NMI to run, but we want to execute it when the
447  * first one finishes.  We set the state to "latched", and the exit of
448  * the first NMI will perform a dec_return, if the result is zero
449  * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the
450  * dec_return would have set the state to NMI_EXECUTING (what we want it
451  * to be when we are running). In this case, we simply jump back to
452  * rerun the NMI handler again, and restart the 'latched' NMI.
453  *
454  * No trap (breakpoint or page fault) should be hit before nmi_restart,
455  * thus there is no race between the first check of state for NOT_RUNNING
456  * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs
457  * at this point.
458  *
459  * In case the NMI takes a page fault, we need to save off the CR2
460  * because the NMI could have preempted another page fault and corrupt
461  * the CR2 that is about to be read. As nested NMIs must be restarted
462  * and they can not take breakpoints or page faults, the update of the
463  * CR2 must be done before converting the nmi state back to NOT_RUNNING.
464  * Otherwise, there would be a race of another nested NMI coming in
465  * after setting state to NOT_RUNNING but before updating the nmi_cr2.
466  */
467 enum nmi_states {
468 	NMI_NOT_RUNNING = 0,
469 	NMI_EXECUTING,
470 	NMI_LATCHED,
471 };
472 static DEFINE_PER_CPU(enum nmi_states, nmi_state);
473 static DEFINE_PER_CPU(unsigned long, nmi_cr2);
474 
475 #ifdef CONFIG_X86_64
476 /*
477  * In x86_64, we need to handle breakpoint -> NMI -> breakpoint.  Without
478  * some care, the inner breakpoint will clobber the outer breakpoint's
479  * stack.
480  *
481  * If a breakpoint is being processed, and the debug stack is being
482  * used, if an NMI comes in and also hits a breakpoint, the stack
483  * pointer will be set to the same fixed address as the breakpoint that
484  * was interrupted, causing that stack to be corrupted. To handle this
485  * case, check if the stack that was interrupted is the debug stack, and
486  * if so, change the IDT so that new breakpoints will use the current
487  * stack and not switch to the fixed address. On return of the NMI,
488  * switch back to the original IDT.
489  */
490 static DEFINE_PER_CPU(int, update_debug_stack);
491 
492 static bool notrace is_debug_stack(unsigned long addr)
493 {
494 	struct cea_exception_stacks *cs = __this_cpu_read(cea_exception_stacks);
495 	unsigned long top = CEA_ESTACK_TOP(cs, DB);
496 	unsigned long bot = CEA_ESTACK_BOT(cs, DB1);
497 
498 	if (__this_cpu_read(debug_stack_usage))
499 		return true;
500 	/*
501 	 * Note, this covers the guard page between DB and DB1 as well to
502 	 * avoid two checks. But by all means @addr can never point into
503 	 * the guard page.
504 	 */
505 	return addr >= bot && addr < top;
506 }
507 NOKPROBE_SYMBOL(is_debug_stack);
508 #endif
509 
510 dotraplinkage notrace void
511 do_nmi(struct pt_regs *regs, long error_code)
512 {
513 	if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
514 		this_cpu_write(nmi_state, NMI_LATCHED);
515 		return;
516 	}
517 	this_cpu_write(nmi_state, NMI_EXECUTING);
518 	this_cpu_write(nmi_cr2, read_cr2());
519 nmi_restart:
520 
521 #ifdef CONFIG_X86_64
522 	/*
523 	 * If we interrupted a breakpoint, it is possible that
524 	 * the nmi handler will have breakpoints too. We need to
525 	 * change the IDT such that breakpoints that happen here
526 	 * continue to use the NMI stack.
527 	 */
528 	if (unlikely(is_debug_stack(regs->sp))) {
529 		debug_stack_set_zero();
530 		this_cpu_write(update_debug_stack, 1);
531 	}
532 #endif
533 
534 	nmi_enter();
535 
536 	inc_irq_stat(__nmi_count);
537 
538 	if (!ignore_nmis)
539 		default_do_nmi(regs);
540 
541 	nmi_exit();
542 
543 #ifdef CONFIG_X86_64
544 	if (unlikely(this_cpu_read(update_debug_stack))) {
545 		debug_stack_reset();
546 		this_cpu_write(update_debug_stack, 0);
547 	}
548 #endif
549 
550 	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
551 		write_cr2(this_cpu_read(nmi_cr2));
552 	if (this_cpu_dec_return(nmi_state))
553 		goto nmi_restart;
554 }
555 NOKPROBE_SYMBOL(do_nmi);
556 
557 void stop_nmi(void)
558 {
559 	ignore_nmis++;
560 }
561 
562 void restart_nmi(void)
563 {
564 	ignore_nmis--;
565 }
566 
567 /* reset the back-to-back NMI logic */
568 void local_touch_nmi(void)
569 {
570 	__this_cpu_write(last_nmi_rip, 0);
571 }
572 EXPORT_SYMBOL_GPL(local_touch_nmi);
573