Linux Audio

Check our new training course

Loading...
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 *  Copyright (C) 1991, 1992  Linus Torvalds
  4 *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
  5 *  Copyright (C) 2011	Don Zickus Red Hat, Inc.
  6 *
  7 *  Pentium III FXSR, SSE support
  8 *	Gareth Hughes <gareth@valinux.com>, May 2000
  9 */
 10
 11/*
 12 * Handle hardware traps and faults.
 13 */
 14#include <linux/spinlock.h>
 15#include <linux/kprobes.h>
 16#include <linux/kdebug.h>
 17#include <linux/sched/debug.h>
 18#include <linux/nmi.h>
 19#include <linux/debugfs.h>
 20#include <linux/delay.h>
 21#include <linux/hardirq.h>
 22#include <linux/ratelimit.h>
 23#include <linux/slab.h>
 24#include <linux/export.h>
 25#include <linux/atomic.h>
 26#include <linux/sched/clock.h>
 27
 28#if defined(CONFIG_EDAC)
 29#include <linux/edac.h>
 30#endif
 31
 32#include <asm/cpu_entry_area.h>
 33#include <asm/traps.h>
 34#include <asm/mach_traps.h>
 35#include <asm/nmi.h>
 36#include <asm/x86_init.h>
 37#include <asm/reboot.h>
 38#include <asm/cache.h>
 39#include <asm/nospec-branch.h>
 40
 41#define CREATE_TRACE_POINTS
 42#include <trace/events/nmi.h>
 43
 44struct nmi_desc {
 45	raw_spinlock_t lock;
 46	struct list_head head;
 47};
 48
 49static struct nmi_desc nmi_desc[NMI_MAX] = 
 50{
 51	{
 52		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
 53		.head = LIST_HEAD_INIT(nmi_desc[0].head),
 54	},
 55	{
 56		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
 57		.head = LIST_HEAD_INIT(nmi_desc[1].head),
 58	},
 59	{
 60		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
 61		.head = LIST_HEAD_INIT(nmi_desc[2].head),
 62	},
 63	{
 64		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
 65		.head = LIST_HEAD_INIT(nmi_desc[3].head),
 66	},
 67
 68};
 69
 70struct nmi_stats {
 71	unsigned int normal;
 72	unsigned int unknown;
 73	unsigned int external;
 74	unsigned int swallow;
 75};
 76
 77static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
 78
 79static int ignore_nmis __read_mostly;
 80
 81int unknown_nmi_panic;
 82/*
 83 * Prevent NMI reason port (0x61) being accessed simultaneously, can
 84 * only be used in NMI handler.
 85 */
 86static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
 87
 88static int __init setup_unknown_nmi_panic(char *str)
 89{
 90	unknown_nmi_panic = 1;
 91	return 1;
 92}
 93__setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 94
 95#define nmi_to_desc(type) (&nmi_desc[type])
 96
 97static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
 98
 99static int __init nmi_warning_debugfs(void)
100{
101	debugfs_create_u64("nmi_longest_ns", 0644,
102			arch_debugfs_dir, &nmi_longest_ns);
103	return 0;
104}
105fs_initcall(nmi_warning_debugfs);
106
107static void nmi_max_handler(struct irq_work *w)
108{
109	struct nmiaction *a = container_of(w, struct nmiaction, irq_work);
110	int remainder_ns, decimal_msecs;
111	u64 whole_msecs = READ_ONCE(a->max_duration);
 
 
 
 
112
113	remainder_ns = do_div(whole_msecs, (1000 * 1000));
114	decimal_msecs = remainder_ns / 1000;
115
116	printk_ratelimited(KERN_INFO
117		"INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
118		a->handler, whole_msecs, decimal_msecs);
119}
120
121static int nmi_handle(unsigned int type, struct pt_regs *regs)
122{
123	struct nmi_desc *desc = nmi_to_desc(type);
124	struct nmiaction *a;
125	int handled=0;
126
127	rcu_read_lock();
128
129	/*
130	 * NMIs are edge-triggered, which means if you have enough
131	 * of them concurrently, you can lose some because only one
132	 * can be latched at any given time.  Walk the whole list
133	 * to handle those situations.
134	 */
135	list_for_each_entry_rcu(a, &desc->head, list) {
136		int thishandled;
137		u64 delta;
138
139		delta = sched_clock();
140		thishandled = a->handler(type, regs);
141		handled += thishandled;
142		delta = sched_clock() - delta;
143		trace_nmi_handler(a->handler, (int)delta, thishandled);
144
145		if (delta < nmi_longest_ns || delta < a->max_duration)
146			continue;
147
148		a->max_duration = delta;
149		irq_work_queue(&a->irq_work);
150	}
151
152	rcu_read_unlock();
153
154	/* return total number of NMI events handled */
155	return handled;
156}
157NOKPROBE_SYMBOL(nmi_handle);
158
159int __register_nmi_handler(unsigned int type, struct nmiaction *action)
160{
161	struct nmi_desc *desc = nmi_to_desc(type);
162	unsigned long flags;
163
164	if (!action->handler)
165		return -EINVAL;
166
167	init_irq_work(&action->irq_work, nmi_max_handler);
168
169	raw_spin_lock_irqsave(&desc->lock, flags);
170
171	/*
172	 * Indicate if there are multiple registrations on the
173	 * internal NMI handler call chains (SERR and IO_CHECK).
174	 */
175	WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
176	WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
177
178	/*
179	 * some handlers need to be executed first otherwise a fake
180	 * event confuses some handlers (kdump uses this flag)
181	 */
182	if (action->flags & NMI_FLAG_FIRST)
183		list_add_rcu(&action->list, &desc->head);
184	else
185		list_add_tail_rcu(&action->list, &desc->head);
186	
187	raw_spin_unlock_irqrestore(&desc->lock, flags);
188	return 0;
189}
190EXPORT_SYMBOL(__register_nmi_handler);
191
192void unregister_nmi_handler(unsigned int type, const char *name)
193{
194	struct nmi_desc *desc = nmi_to_desc(type);
195	struct nmiaction *n;
196	unsigned long flags;
197
198	raw_spin_lock_irqsave(&desc->lock, flags);
199
200	list_for_each_entry_rcu(n, &desc->head, list) {
201		/*
202		 * the name passed in to describe the nmi handler
203		 * is used as the lookup key
204		 */
205		if (!strcmp(n->name, name)) {
206			WARN(in_nmi(),
207				"Trying to free NMI (%s) from NMI context!\n", n->name);
208			list_del_rcu(&n->list);
209			break;
210		}
211	}
212
213	raw_spin_unlock_irqrestore(&desc->lock, flags);
214	synchronize_rcu();
215}
216EXPORT_SYMBOL_GPL(unregister_nmi_handler);
217
218static void
219pci_serr_error(unsigned char reason, struct pt_regs *regs)
220{
221	/* check to see if anyone registered against these types of errors */
222	if (nmi_handle(NMI_SERR, regs))
223		return;
224
225	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
226		 reason, smp_processor_id());
227
228	if (panic_on_unrecovered_nmi)
229		nmi_panic(regs, "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}
237NOKPROBE_SYMBOL(pci_serr_error);
238
239static void
240io_check_error(unsigned char reason, struct pt_regs *regs)
241{
242	unsigned long i;
243
244	/* check to see if anyone registered against these types of errors */
245	if (nmi_handle(NMI_IO_CHECK, regs))
246		return;
247
248	pr_emerg(
249	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
250		 reason, smp_processor_id());
251	show_regs(regs);
252
253	if (panic_on_io_nmi) {
254		nmi_panic(regs, "NMI IOCK error: Not continuing");
255
256		/*
257		 * If we end up here, it means we have received an NMI while
258		 * processing panic(). Simply return without delaying and
259		 * re-enabling NMIs.
260		 */
261		return;
262	}
263
264	/* Re-enable the IOCK line, wait for a few seconds */
265	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
266	outb(reason, NMI_REASON_PORT);
267
268	i = 20000;
269	while (--i) {
270		touch_nmi_watchdog();
271		udelay(100);
272	}
273
274	reason &= ~NMI_REASON_CLEAR_IOCHK;
275	outb(reason, NMI_REASON_PORT);
276}
277NOKPROBE_SYMBOL(io_check_error);
278
279static void
280unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
281{
282	int handled;
283
284	/*
285	 * Use 'false' as back-to-back NMIs are dealt with one level up.
286	 * Of course this makes having multiple 'unknown' handlers useless
287	 * as only the first one is ever run (unless it can actually determine
288	 * if it caused the NMI)
289	 */
290	handled = nmi_handle(NMI_UNKNOWN, regs);
291	if (handled) {
292		__this_cpu_add(nmi_stats.unknown, handled);
293		return;
294	}
295
296	__this_cpu_add(nmi_stats.unknown, 1);
297
298	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
299		 reason, smp_processor_id());
300
301	pr_emerg("Do you have a strange power saving mode enabled?\n");
302	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
303		nmi_panic(regs, "NMI: Not continuing");
304
305	pr_emerg("Dazed and confused, but trying to continue\n");
306}
307NOKPROBE_SYMBOL(unknown_nmi_error);
308
309static DEFINE_PER_CPU(bool, swallow_nmi);
310static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
311
312static void default_do_nmi(struct pt_regs *regs)
313{
314	unsigned char reason = 0;
315	int handled;
316	bool b2b = false;
317
318	/*
319	 * CPU-specific NMI must be processed before non-CPU-specific
320	 * NMI, otherwise we may lose it, because the CPU-specific
321	 * NMI can not be detected/processed on other CPUs.
322	 */
323
324	/*
325	 * Back-to-back NMIs are interesting because they can either
326	 * be two NMI or more than two NMIs (any thing over two is dropped
327	 * due to NMI being edge-triggered).  If this is the second half
328	 * of the back-to-back NMI, assume we dropped things and process
329	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
330	 */
331	if (regs->ip == __this_cpu_read(last_nmi_rip))
332		b2b = true;
333	else
334		__this_cpu_write(swallow_nmi, false);
335
336	__this_cpu_write(last_nmi_rip, regs->ip);
337
 
 
338	handled = nmi_handle(NMI_LOCAL, regs);
339	__this_cpu_add(nmi_stats.normal, handled);
340	if (handled) {
341		/*
342		 * There are cases when a NMI handler handles multiple
343		 * events in the current NMI.  One of these events may
344		 * be queued for in the next NMI.  Because the event is
345		 * already handled, the next NMI will result in an unknown
346		 * NMI.  Instead lets flag this for a potential NMI to
347		 * swallow.
348		 */
349		if (handled > 1)
350			__this_cpu_write(swallow_nmi, true);
351		return;
352	}
353
354	/*
355	 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
356	 *
357	 * Another CPU may be processing panic routines while holding
358	 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
359	 * and if so, call its callback directly.  If there is no CPU preparing
360	 * crash dump, we simply loop here.
361	 */
362	while (!raw_spin_trylock(&nmi_reason_lock)) {
363		run_crash_ipi_callback(regs);
364		cpu_relax();
365	}
366
367	reason = x86_platform.get_nmi_reason();
368
369	if (reason & NMI_REASON_MASK) {
370		if (reason & NMI_REASON_SERR)
371			pci_serr_error(reason, regs);
372		else if (reason & NMI_REASON_IOCHK)
373			io_check_error(reason, regs);
374#ifdef CONFIG_X86_32
375		/*
376		 * Reassert NMI in case it became active
377		 * meanwhile as it's edge-triggered:
378		 */
379		reassert_nmi();
380#endif
381		__this_cpu_add(nmi_stats.external, 1);
382		raw_spin_unlock(&nmi_reason_lock);
383		return;
384	}
385	raw_spin_unlock(&nmi_reason_lock);
386
387	/*
388	 * Only one NMI can be latched at a time.  To handle
389	 * this we may process multiple nmi handlers at once to
390	 * cover the case where an NMI is dropped.  The downside
391	 * to this approach is we may process an NMI prematurely,
392	 * while its real NMI is sitting latched.  This will cause
393	 * an unknown NMI on the next run of the NMI processing.
394	 *
395	 * We tried to flag that condition above, by setting the
396	 * swallow_nmi flag when we process more than one event.
397	 * This condition is also only present on the second half
398	 * of a back-to-back NMI, so we flag that condition too.
399	 *
400	 * If both are true, we assume we already processed this
401	 * NMI previously and we swallow it.  Otherwise we reset
402	 * the logic.
403	 *
404	 * There are scenarios where we may accidentally swallow
405	 * a 'real' unknown NMI.  For example, while processing
406	 * a perf NMI another perf NMI comes in along with a
407	 * 'real' unknown NMI.  These two NMIs get combined into
408	 * one (as descibed above).  When the next NMI gets
409	 * processed, it will be flagged by perf as handled, but
410	 * noone will know that there was a 'real' unknown NMI sent
411	 * also.  As a result it gets swallowed.  Or if the first
412	 * perf NMI returns two events handled then the second
413	 * NMI will get eaten by the logic below, again losing a
414	 * 'real' unknown NMI.  But this is the best we can do
415	 * for now.
416	 */
417	if (b2b && __this_cpu_read(swallow_nmi))
418		__this_cpu_add(nmi_stats.swallow, 1);
419	else
420		unknown_nmi_error(reason, regs);
 
 
 
421}
422NOKPROBE_SYMBOL(default_do_nmi);
423
424/*
425 * NMIs can page fault or hit breakpoints which will cause it to lose
426 * its NMI context with the CPU when the breakpoint or page fault does an IRET.
427 *
428 * As a result, NMIs can nest if NMIs get unmasked due an IRET during
429 * NMI processing.  On x86_64, the asm glue protects us from nested NMIs
430 * if the outer NMI came from kernel mode, but we can still nest if the
431 * outer NMI came from user mode.
432 *
433 * To handle these nested NMIs, we have three states:
434 *
435 *  1) not running
436 *  2) executing
437 *  3) latched
438 *
439 * When no NMI is in progress, it is in the "not running" state.
440 * When an NMI comes in, it goes into the "executing" state.
441 * Normally, if another NMI is triggered, it does not interrupt
442 * the running NMI and the HW will simply latch it so that when
443 * the first NMI finishes, it will restart the second NMI.
444 * (Note, the latch is binary, thus multiple NMIs triggering,
445 *  when one is running, are ignored. Only one NMI is restarted.)
446 *
447 * If an NMI executes an iret, another NMI can preempt it. We do not
448 * want to allow this new NMI to run, but we want to execute it when the
449 * first one finishes.  We set the state to "latched", and the exit of
450 * the first NMI will perform a dec_return, if the result is zero
451 * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the
452 * dec_return would have set the state to NMI_EXECUTING (what we want it
453 * to be when we are running). In this case, we simply jump back to
454 * rerun the NMI handler again, and restart the 'latched' NMI.
455 *
456 * No trap (breakpoint or page fault) should be hit before nmi_restart,
457 * thus there is no race between the first check of state for NOT_RUNNING
458 * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs
459 * at this point.
460 *
461 * In case the NMI takes a page fault, we need to save off the CR2
462 * because the NMI could have preempted another page fault and corrupt
463 * the CR2 that is about to be read. As nested NMIs must be restarted
464 * and they can not take breakpoints or page faults, the update of the
465 * CR2 must be done before converting the nmi state back to NOT_RUNNING.
466 * Otherwise, there would be a race of another nested NMI coming in
467 * after setting state to NOT_RUNNING but before updating the nmi_cr2.
468 */
469enum nmi_states {
470	NMI_NOT_RUNNING = 0,
471	NMI_EXECUTING,
472	NMI_LATCHED,
473};
474static DEFINE_PER_CPU(enum nmi_states, nmi_state);
475static DEFINE_PER_CPU(unsigned long, nmi_cr2);
 
476
477#ifdef CONFIG_X86_64
478/*
479 * In x86_64, we need to handle breakpoint -> NMI -> breakpoint.  Without
480 * some care, the inner breakpoint will clobber the outer breakpoint's
481 * stack.
482 *
483 * If a breakpoint is being processed, and the debug stack is being
484 * used, if an NMI comes in and also hits a breakpoint, the stack
485 * pointer will be set to the same fixed address as the breakpoint that
486 * was interrupted, causing that stack to be corrupted. To handle this
487 * case, check if the stack that was interrupted is the debug stack, and
488 * if so, change the IDT so that new breakpoints will use the current
489 * stack and not switch to the fixed address. On return of the NMI,
490 * switch back to the original IDT.
491 */
492static DEFINE_PER_CPU(int, update_debug_stack);
493
494static bool notrace is_debug_stack(unsigned long addr)
495{
496	struct cea_exception_stacks *cs = __this_cpu_read(cea_exception_stacks);
497	unsigned long top = CEA_ESTACK_TOP(cs, DB);
498	unsigned long bot = CEA_ESTACK_BOT(cs, DB1);
499
500	if (__this_cpu_read(debug_stack_usage))
501		return true;
502	/*
503	 * Note, this covers the guard page between DB and DB1 as well to
504	 * avoid two checks. But by all means @addr can never point into
505	 * the guard page.
506	 */
507	return addr >= bot && addr < top;
508}
509NOKPROBE_SYMBOL(is_debug_stack);
510#endif
511
512dotraplinkage notrace void
513do_nmi(struct pt_regs *regs, long error_code)
514{
515	if (IS_ENABLED(CONFIG_SMP) && cpu_is_offline(smp_processor_id()))
516		return;
517
518	if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
519		this_cpu_write(nmi_state, NMI_LATCHED);
520		return;
521	}
522	this_cpu_write(nmi_state, NMI_EXECUTING);
523	this_cpu_write(nmi_cr2, read_cr2());
524nmi_restart:
525
526#ifdef CONFIG_X86_64
527	/*
528	 * If we interrupted a breakpoint, it is possible that
529	 * the nmi handler will have breakpoints too. We need to
530	 * change the IDT such that breakpoints that happen here
531	 * continue to use the NMI stack.
532	 */
533	if (unlikely(is_debug_stack(regs->sp))) {
534		debug_stack_set_zero();
535		this_cpu_write(update_debug_stack, 1);
536	}
537#endif
538
539	nmi_enter();
540
541	inc_irq_stat(__nmi_count);
542
543	if (!ignore_nmis)
544		default_do_nmi(regs);
545
546	nmi_exit();
547
548#ifdef CONFIG_X86_64
549	if (unlikely(this_cpu_read(update_debug_stack))) {
550		debug_stack_reset();
551		this_cpu_write(update_debug_stack, 0);
552	}
553#endif
554
555	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
556		write_cr2(this_cpu_read(nmi_cr2));
557	if (this_cpu_dec_return(nmi_state))
558		goto nmi_restart;
559
560	if (user_mode(regs))
561		mds_user_clear_cpu_buffers();
562}
563NOKPROBE_SYMBOL(do_nmi);
564
565void stop_nmi(void)
566{
567	ignore_nmis++;
568}
569
570void restart_nmi(void)
571{
572	ignore_nmis--;
573}
574
575/* reset the back-to-back NMI logic */
576void local_touch_nmi(void)
577{
578	__this_cpu_write(last_nmi_rip, 0);
579}
580EXPORT_SYMBOL_GPL(local_touch_nmi);
v5.9
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 *  Copyright (C) 1991, 1992  Linus Torvalds
  4 *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
  5 *  Copyright (C) 2011	Don Zickus Red Hat, Inc.
  6 *
  7 *  Pentium III FXSR, SSE support
  8 *	Gareth Hughes <gareth@valinux.com>, May 2000
  9 */
 10
 11/*
 12 * Handle hardware traps and faults.
 13 */
 14#include <linux/spinlock.h>
 15#include <linux/kprobes.h>
 16#include <linux/kdebug.h>
 17#include <linux/sched/debug.h>
 18#include <linux/nmi.h>
 19#include <linux/debugfs.h>
 20#include <linux/delay.h>
 21#include <linux/hardirq.h>
 22#include <linux/ratelimit.h>
 23#include <linux/slab.h>
 24#include <linux/export.h>
 25#include <linux/atomic.h>
 26#include <linux/sched/clock.h>
 27
 
 
 
 
 28#include <asm/cpu_entry_area.h>
 29#include <asm/traps.h>
 30#include <asm/mach_traps.h>
 31#include <asm/nmi.h>
 32#include <asm/x86_init.h>
 33#include <asm/reboot.h>
 34#include <asm/cache.h>
 35#include <asm/nospec-branch.h>
 36
 37#define CREATE_TRACE_POINTS
 38#include <trace/events/nmi.h>
 39
 40struct nmi_desc {
 41	raw_spinlock_t lock;
 42	struct list_head head;
 43};
 44
 45static struct nmi_desc nmi_desc[NMI_MAX] = 
 46{
 47	{
 48		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
 49		.head = LIST_HEAD_INIT(nmi_desc[0].head),
 50	},
 51	{
 52		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
 53		.head = LIST_HEAD_INIT(nmi_desc[1].head),
 54	},
 55	{
 56		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
 57		.head = LIST_HEAD_INIT(nmi_desc[2].head),
 58	},
 59	{
 60		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
 61		.head = LIST_HEAD_INIT(nmi_desc[3].head),
 62	},
 63
 64};
 65
 66struct nmi_stats {
 67	unsigned int normal;
 68	unsigned int unknown;
 69	unsigned int external;
 70	unsigned int swallow;
 71};
 72
 73static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
 74
 75static int ignore_nmis __read_mostly;
 76
 77int unknown_nmi_panic;
 78/*
 79 * Prevent NMI reason port (0x61) being accessed simultaneously, can
 80 * only be used in NMI handler.
 81 */
 82static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
 83
 84static int __init setup_unknown_nmi_panic(char *str)
 85{
 86	unknown_nmi_panic = 1;
 87	return 1;
 88}
 89__setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 90
 91#define nmi_to_desc(type) (&nmi_desc[type])
 92
 93static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
 94
 95static int __init nmi_warning_debugfs(void)
 96{
 97	debugfs_create_u64("nmi_longest_ns", 0644,
 98			arch_debugfs_dir, &nmi_longest_ns);
 99	return 0;
100}
101fs_initcall(nmi_warning_debugfs);
102
103static void nmi_check_duration(struct nmiaction *action, u64 duration)
104{
105	u64 whole_msecs = READ_ONCE(action->max_duration);
106	int remainder_ns, decimal_msecs;
107
108	if (duration < nmi_longest_ns || duration < action->max_duration)
109		return;
110
111	action->max_duration = duration;
112
113	remainder_ns = do_div(whole_msecs, (1000 * 1000));
114	decimal_msecs = remainder_ns / 1000;
115
116	printk_ratelimited(KERN_INFO
117		"INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
118		action->handler, whole_msecs, decimal_msecs);
119}
120
121static int nmi_handle(unsigned int type, struct pt_regs *regs)
122{
123	struct nmi_desc *desc = nmi_to_desc(type);
124	struct nmiaction *a;
125	int handled=0;
126
127	rcu_read_lock();
128
129	/*
130	 * NMIs are edge-triggered, which means if you have enough
131	 * of them concurrently, you can lose some because only one
132	 * can be latched at any given time.  Walk the whole list
133	 * to handle those situations.
134	 */
135	list_for_each_entry_rcu(a, &desc->head, list) {
136		int thishandled;
137		u64 delta;
138
139		delta = sched_clock();
140		thishandled = a->handler(type, regs);
141		handled += thishandled;
142		delta = sched_clock() - delta;
143		trace_nmi_handler(a->handler, (int)delta, thishandled);
144
145		nmi_check_duration(a, delta);
 
 
 
 
146	}
147
148	rcu_read_unlock();
149
150	/* return total number of NMI events handled */
151	return handled;
152}
153NOKPROBE_SYMBOL(nmi_handle);
154
155int __register_nmi_handler(unsigned int type, struct nmiaction *action)
156{
157	struct nmi_desc *desc = nmi_to_desc(type);
158	unsigned long flags;
159
160	if (!action->handler)
161		return -EINVAL;
162
 
 
163	raw_spin_lock_irqsave(&desc->lock, flags);
164
165	/*
166	 * Indicate if there are multiple registrations on the
167	 * internal NMI handler call chains (SERR and IO_CHECK).
168	 */
169	WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
170	WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
171
172	/*
173	 * some handlers need to be executed first otherwise a fake
174	 * event confuses some handlers (kdump uses this flag)
175	 */
176	if (action->flags & NMI_FLAG_FIRST)
177		list_add_rcu(&action->list, &desc->head);
178	else
179		list_add_tail_rcu(&action->list, &desc->head);
180	
181	raw_spin_unlock_irqrestore(&desc->lock, flags);
182	return 0;
183}
184EXPORT_SYMBOL(__register_nmi_handler);
185
186void unregister_nmi_handler(unsigned int type, const char *name)
187{
188	struct nmi_desc *desc = nmi_to_desc(type);
189	struct nmiaction *n;
190	unsigned long flags;
191
192	raw_spin_lock_irqsave(&desc->lock, flags);
193
194	list_for_each_entry_rcu(n, &desc->head, list) {
195		/*
196		 * the name passed in to describe the nmi handler
197		 * is used as the lookup key
198		 */
199		if (!strcmp(n->name, name)) {
200			WARN(in_nmi(),
201				"Trying to free NMI (%s) from NMI context!\n", n->name);
202			list_del_rcu(&n->list);
203			break;
204		}
205	}
206
207	raw_spin_unlock_irqrestore(&desc->lock, flags);
208	synchronize_rcu();
209}
210EXPORT_SYMBOL_GPL(unregister_nmi_handler);
211
212static void
213pci_serr_error(unsigned char reason, struct pt_regs *regs)
214{
215	/* check to see if anyone registered against these types of errors */
216	if (nmi_handle(NMI_SERR, regs))
217		return;
218
219	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
220		 reason, smp_processor_id());
221
222	if (panic_on_unrecovered_nmi)
223		nmi_panic(regs, "NMI: Not continuing");
224
225	pr_emerg("Dazed and confused, but trying to continue\n");
226
227	/* Clear and disable the PCI SERR error line. */
228	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
229	outb(reason, NMI_REASON_PORT);
230}
231NOKPROBE_SYMBOL(pci_serr_error);
232
233static void
234io_check_error(unsigned char reason, struct pt_regs *regs)
235{
236	unsigned long i;
237
238	/* check to see if anyone registered against these types of errors */
239	if (nmi_handle(NMI_IO_CHECK, regs))
240		return;
241
242	pr_emerg(
243	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
244		 reason, smp_processor_id());
245	show_regs(regs);
246
247	if (panic_on_io_nmi) {
248		nmi_panic(regs, "NMI IOCK error: Not continuing");
249
250		/*
251		 * If we end up here, it means we have received an NMI while
252		 * processing panic(). Simply return without delaying and
253		 * re-enabling NMIs.
254		 */
255		return;
256	}
257
258	/* Re-enable the IOCK line, wait for a few seconds */
259	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
260	outb(reason, NMI_REASON_PORT);
261
262	i = 20000;
263	while (--i) {
264		touch_nmi_watchdog();
265		udelay(100);
266	}
267
268	reason &= ~NMI_REASON_CLEAR_IOCHK;
269	outb(reason, NMI_REASON_PORT);
270}
271NOKPROBE_SYMBOL(io_check_error);
272
273static void
274unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
275{
276	int handled;
277
278	/*
279	 * Use 'false' as back-to-back NMIs are dealt with one level up.
280	 * Of course this makes having multiple 'unknown' handlers useless
281	 * as only the first one is ever run (unless it can actually determine
282	 * if it caused the NMI)
283	 */
284	handled = nmi_handle(NMI_UNKNOWN, regs);
285	if (handled) {
286		__this_cpu_add(nmi_stats.unknown, handled);
287		return;
288	}
289
290	__this_cpu_add(nmi_stats.unknown, 1);
291
292	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
293		 reason, smp_processor_id());
294
295	pr_emerg("Do you have a strange power saving mode enabled?\n");
296	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
297		nmi_panic(regs, "NMI: Not continuing");
298
299	pr_emerg("Dazed and confused, but trying to continue\n");
300}
301NOKPROBE_SYMBOL(unknown_nmi_error);
302
303static DEFINE_PER_CPU(bool, swallow_nmi);
304static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
305
306static noinstr void default_do_nmi(struct pt_regs *regs)
307{
308	unsigned char reason = 0;
309	int handled;
310	bool b2b = false;
311
312	/*
313	 * CPU-specific NMI must be processed before non-CPU-specific
314	 * NMI, otherwise we may lose it, because the CPU-specific
315	 * NMI can not be detected/processed on other CPUs.
316	 */
317
318	/*
319	 * Back-to-back NMIs are interesting because they can either
320	 * be two NMI or more than two NMIs (any thing over two is dropped
321	 * due to NMI being edge-triggered).  If this is the second half
322	 * of the back-to-back NMI, assume we dropped things and process
323	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
324	 */
325	if (regs->ip == __this_cpu_read(last_nmi_rip))
326		b2b = true;
327	else
328		__this_cpu_write(swallow_nmi, false);
329
330	__this_cpu_write(last_nmi_rip, regs->ip);
331
332	instrumentation_begin();
333
334	handled = nmi_handle(NMI_LOCAL, regs);
335	__this_cpu_add(nmi_stats.normal, handled);
336	if (handled) {
337		/*
338		 * There are cases when a NMI handler handles multiple
339		 * events in the current NMI.  One of these events may
340		 * be queued for in the next NMI.  Because the event is
341		 * already handled, the next NMI will result in an unknown
342		 * NMI.  Instead lets flag this for a potential NMI to
343		 * swallow.
344		 */
345		if (handled > 1)
346			__this_cpu_write(swallow_nmi, true);
347		goto out;
348	}
349
350	/*
351	 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
352	 *
353	 * Another CPU may be processing panic routines while holding
354	 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
355	 * and if so, call its callback directly.  If there is no CPU preparing
356	 * crash dump, we simply loop here.
357	 */
358	while (!raw_spin_trylock(&nmi_reason_lock)) {
359		run_crash_ipi_callback(regs);
360		cpu_relax();
361	}
362
363	reason = x86_platform.get_nmi_reason();
364
365	if (reason & NMI_REASON_MASK) {
366		if (reason & NMI_REASON_SERR)
367			pci_serr_error(reason, regs);
368		else if (reason & NMI_REASON_IOCHK)
369			io_check_error(reason, regs);
370#ifdef CONFIG_X86_32
371		/*
372		 * Reassert NMI in case it became active
373		 * meanwhile as it's edge-triggered:
374		 */
375		reassert_nmi();
376#endif
377		__this_cpu_add(nmi_stats.external, 1);
378		raw_spin_unlock(&nmi_reason_lock);
379		goto out;
380	}
381	raw_spin_unlock(&nmi_reason_lock);
382
383	/*
384	 * Only one NMI can be latched at a time.  To handle
385	 * this we may process multiple nmi handlers at once to
386	 * cover the case where an NMI is dropped.  The downside
387	 * to this approach is we may process an NMI prematurely,
388	 * while its real NMI is sitting latched.  This will cause
389	 * an unknown NMI on the next run of the NMI processing.
390	 *
391	 * We tried to flag that condition above, by setting the
392	 * swallow_nmi flag when we process more than one event.
393	 * This condition is also only present on the second half
394	 * of a back-to-back NMI, so we flag that condition too.
395	 *
396	 * If both are true, we assume we already processed this
397	 * NMI previously and we swallow it.  Otherwise we reset
398	 * the logic.
399	 *
400	 * There are scenarios where we may accidentally swallow
401	 * a 'real' unknown NMI.  For example, while processing
402	 * a perf NMI another perf NMI comes in along with a
403	 * 'real' unknown NMI.  These two NMIs get combined into
404	 * one (as described above).  When the next NMI gets
405	 * processed, it will be flagged by perf as handled, but
406	 * no one will know that there was a 'real' unknown NMI sent
407	 * also.  As a result it gets swallowed.  Or if the first
408	 * perf NMI returns two events handled then the second
409	 * NMI will get eaten by the logic below, again losing a
410	 * 'real' unknown NMI.  But this is the best we can do
411	 * for now.
412	 */
413	if (b2b && __this_cpu_read(swallow_nmi))
414		__this_cpu_add(nmi_stats.swallow, 1);
415	else
416		unknown_nmi_error(reason, regs);
417
418out:
419	instrumentation_end();
420}
 
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 */
467enum nmi_states {
468	NMI_NOT_RUNNING = 0,
469	NMI_EXECUTING,
470	NMI_LATCHED,
471};
472static DEFINE_PER_CPU(enum nmi_states, nmi_state);
473static DEFINE_PER_CPU(unsigned long, nmi_cr2);
474static DEFINE_PER_CPU(unsigned long, nmi_dr7);
475
476DEFINE_IDTENTRY_RAW(exc_nmi)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
477{
478	bool irq_state;
 
 
479
480	if (IS_ENABLED(CONFIG_SMP) && arch_cpu_is_offline(smp_processor_id()))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
481		return;
482
483	if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
484		this_cpu_write(nmi_state, NMI_LATCHED);
485		return;
486	}
487	this_cpu_write(nmi_state, NMI_EXECUTING);
488	this_cpu_write(nmi_cr2, read_cr2());
489nmi_restart:
490
491	this_cpu_write(nmi_dr7, local_db_save());
 
 
 
 
 
 
 
 
 
 
 
492
493	irq_state = idtentry_enter_nmi(regs);
494
495	inc_irq_stat(__nmi_count);
496
497	if (!ignore_nmis)
498		default_do_nmi(regs);
499
500	idtentry_exit_nmi(regs, irq_state);
501
502	local_db_restore(this_cpu_read(nmi_dr7));
 
 
 
 
 
503
504	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
505		write_cr2(this_cpu_read(nmi_cr2));
506	if (this_cpu_dec_return(nmi_state))
507		goto nmi_restart;
508
509	if (user_mode(regs))
510		mds_user_clear_cpu_buffers();
511}
 
512
513void stop_nmi(void)
514{
515	ignore_nmis++;
516}
517
518void restart_nmi(void)
519{
520	ignore_nmis--;
521}
522
523/* reset the back-to-back NMI logic */
524void local_touch_nmi(void)
525{
526	__this_cpu_write(last_nmi_rip, 0);
527}
528EXPORT_SYMBOL_GPL(local_touch_nmi);