Linux Audio

Check our new training course

Loading...
v4.17
 
  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/sched/clock.h>
 25
 26#if defined(CONFIG_EDAC)
 27#include <linux/edac.h>
 28#endif
 29
 30#include <linux/atomic.h>
 31#include <asm/traps.h>
 32#include <asm/mach_traps.h>
 33#include <asm/nmi.h>
 34#include <asm/x86_init.h>
 35#include <asm/reboot.h>
 36#include <asm/cache.h>
 
 
 
 37
 38#define CREATE_TRACE_POINTS
 39#include <trace/events/nmi.h>
 40
 41struct nmi_desc {
 42	raw_spinlock_t lock;
 43	struct list_head head;
 44};
 45
 46static struct nmi_desc nmi_desc[NMI_MAX] = 
 47{
 48	{
 49		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
 50		.head = LIST_HEAD_INIT(nmi_desc[0].head),
 51	},
 52	{
 53		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
 54		.head = LIST_HEAD_INIT(nmi_desc[1].head),
 55	},
 56	{
 57		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
 58		.head = LIST_HEAD_INIT(nmi_desc[2].head),
 59	},
 60	{
 61		.lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
 62		.head = LIST_HEAD_INIT(nmi_desc[3].head),
 63	},
 64
 65};
 66
 67struct nmi_stats {
 68	unsigned int normal;
 69	unsigned int unknown;
 70	unsigned int external;
 71	unsigned int swallow;
 
 
 
 
 
 
 
 
 
 72};
 73
 74static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
 75
 76static int ignore_nmis __read_mostly;
 77
 78int unknown_nmi_panic;
 79/*
 80 * Prevent NMI reason port (0x61) being accessed simultaneously, can
 81 * only be used in NMI handler.
 82 */
 83static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
 84
 85static int __init setup_unknown_nmi_panic(char *str)
 86{
 87	unknown_nmi_panic = 1;
 88	return 1;
 89}
 90__setup("unknown_nmi_panic", setup_unknown_nmi_panic);
 91
 92#define nmi_to_desc(type) (&nmi_desc[type])
 93
 94static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
 95
 96static int __init nmi_warning_debugfs(void)
 97{
 98	debugfs_create_u64("nmi_longest_ns", 0644,
 99			arch_debugfs_dir, &nmi_longest_ns);
100	return 0;
101}
102fs_initcall(nmi_warning_debugfs);
103
104static void nmi_max_handler(struct irq_work *w)
105{
106	struct nmiaction *a = container_of(w, struct nmiaction, irq_work);
107	int remainder_ns, decimal_msecs;
108	u64 whole_msecs = READ_ONCE(a->max_duration);
109
110	remainder_ns = do_div(whole_msecs, (1000 * 1000));
 
 
 
 
 
111	decimal_msecs = remainder_ns / 1000;
112
113	printk_ratelimited(KERN_INFO
114		"INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
115		a->handler, whole_msecs, decimal_msecs);
116}
117
118static int nmi_handle(unsigned int type, struct pt_regs *regs)
119{
120	struct nmi_desc *desc = nmi_to_desc(type);
121	struct nmiaction *a;
122	int handled=0;
123
124	rcu_read_lock();
125
126	/*
127	 * NMIs are edge-triggered, which means if you have enough
128	 * of them concurrently, you can lose some because only one
129	 * can be latched at any given time.  Walk the whole list
130	 * to handle those situations.
131	 */
132	list_for_each_entry_rcu(a, &desc->head, list) {
133		int thishandled;
134		u64 delta;
135
136		delta = sched_clock();
137		thishandled = a->handler(type, regs);
138		handled += thishandled;
139		delta = sched_clock() - delta;
140		trace_nmi_handler(a->handler, (int)delta, thishandled);
141
142		if (delta < nmi_longest_ns || delta < a->max_duration)
143			continue;
144
145		a->max_duration = delta;
146		irq_work_queue(&a->irq_work);
147	}
148
149	rcu_read_unlock();
150
151	/* return total number of NMI events handled */
152	return handled;
153}
154NOKPROBE_SYMBOL(nmi_handle);
155
156int __register_nmi_handler(unsigned int type, struct nmiaction *action)
157{
158	struct nmi_desc *desc = nmi_to_desc(type);
159	unsigned long flags;
160
161	if (!action->handler)
162		return -EINVAL;
163
164	init_irq_work(&action->irq_work, nmi_max_handler);
165
166	raw_spin_lock_irqsave(&desc->lock, flags);
167
168	/*
169	 * Indicate if there are multiple registrations on the
170	 * internal NMI handler call chains (SERR and IO_CHECK).
171	 */
172	WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
173	WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
174
175	/*
176	 * some handlers need to be executed first otherwise a fake
177	 * event confuses some handlers (kdump uses this flag)
178	 */
179	if (action->flags & NMI_FLAG_FIRST)
180		list_add_rcu(&action->list, &desc->head);
181	else
182		list_add_tail_rcu(&action->list, &desc->head);
183	
184	raw_spin_unlock_irqrestore(&desc->lock, flags);
185	return 0;
186}
187EXPORT_SYMBOL(__register_nmi_handler);
188
189void unregister_nmi_handler(unsigned int type, const char *name)
190{
191	struct nmi_desc *desc = nmi_to_desc(type);
192	struct nmiaction *n;
193	unsigned long flags;
194
195	raw_spin_lock_irqsave(&desc->lock, flags);
196
197	list_for_each_entry_rcu(n, &desc->head, list) {
198		/*
199		 * the name passed in to describe the nmi handler
200		 * is used as the lookup key
201		 */
202		if (!strcmp(n->name, name)) {
203			WARN(in_nmi(),
204				"Trying to free NMI (%s) from NMI context!\n", n->name);
205			list_del_rcu(&n->list);
 
206			break;
207		}
208	}
209
210	raw_spin_unlock_irqrestore(&desc->lock, flags);
211	synchronize_rcu();
 
 
 
212}
213EXPORT_SYMBOL_GPL(unregister_nmi_handler);
214
215static void
216pci_serr_error(unsigned char reason, struct pt_regs *regs)
217{
218	/* check to see if anyone registered against these types of errors */
219	if (nmi_handle(NMI_SERR, regs))
220		return;
221
222	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
223		 reason, smp_processor_id());
224
225	if (panic_on_unrecovered_nmi)
226		nmi_panic(regs, "NMI: Not continuing");
227
228	pr_emerg("Dazed and confused, but trying to continue\n");
229
230	/* Clear and disable the PCI SERR error line. */
231	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
232	outb(reason, NMI_REASON_PORT);
233}
234NOKPROBE_SYMBOL(pci_serr_error);
235
236static void
237io_check_error(unsigned char reason, struct pt_regs *regs)
238{
239	unsigned long i;
240
241	/* check to see if anyone registered against these types of errors */
242	if (nmi_handle(NMI_IO_CHECK, regs))
243		return;
244
245	pr_emerg(
246	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
247		 reason, smp_processor_id());
248	show_regs(regs);
249
250	if (panic_on_io_nmi) {
251		nmi_panic(regs, "NMI IOCK error: Not continuing");
252
253		/*
254		 * If we end up here, it means we have received an NMI while
255		 * processing panic(). Simply return without delaying and
256		 * re-enabling NMIs.
257		 */
258		return;
259	}
260
261	/* Re-enable the IOCK line, wait for a few seconds */
262	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
263	outb(reason, NMI_REASON_PORT);
264
265	i = 20000;
266	while (--i) {
267		touch_nmi_watchdog();
268		udelay(100);
269	}
270
271	reason &= ~NMI_REASON_CLEAR_IOCHK;
272	outb(reason, NMI_REASON_PORT);
273}
274NOKPROBE_SYMBOL(io_check_error);
275
276static void
277unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
278{
279	int handled;
280
281	/*
282	 * Use 'false' as back-to-back NMIs are dealt with one level up.
283	 * Of course this makes having multiple 'unknown' handlers useless
284	 * as only the first one is ever run (unless it can actually determine
285	 * if it caused the NMI)
286	 */
287	handled = nmi_handle(NMI_UNKNOWN, regs);
288	if (handled) {
289		__this_cpu_add(nmi_stats.unknown, handled);
290		return;
291	}
292
293	__this_cpu_add(nmi_stats.unknown, 1);
294
295	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
296		 reason, smp_processor_id());
297
298	pr_emerg("Do you have a strange power saving mode enabled?\n");
299	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
300		nmi_panic(regs, "NMI: Not continuing");
301
302	pr_emerg("Dazed and confused, but trying to continue\n");
303}
304NOKPROBE_SYMBOL(unknown_nmi_error);
305
306static DEFINE_PER_CPU(bool, swallow_nmi);
307static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
308
309static void default_do_nmi(struct pt_regs *regs)
310{
311	unsigned char reason = 0;
312	int handled;
313	bool b2b = false;
314
315	/*
316	 * CPU-specific NMI must be processed before non-CPU-specific
317	 * NMI, otherwise we may lose it, because the CPU-specific
318	 * NMI can not be detected/processed on other CPUs.
319	 */
320
321	/*
322	 * Back-to-back NMIs are interesting because they can either
323	 * be two NMI or more than two NMIs (any thing over two is dropped
324	 * due to NMI being edge-triggered).  If this is the second half
325	 * of the back-to-back NMI, assume we dropped things and process
326	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
327	 */
328	if (regs->ip == __this_cpu_read(last_nmi_rip))
329		b2b = true;
330	else
331		__this_cpu_write(swallow_nmi, false);
332
333	__this_cpu_write(last_nmi_rip, regs->ip);
334
 
 
 
 
 
335	handled = nmi_handle(NMI_LOCAL, regs);
336	__this_cpu_add(nmi_stats.normal, handled);
337	if (handled) {
338		/*
339		 * There are cases when a NMI handler handles multiple
340		 * events in the current NMI.  One of these events may
341		 * be queued for in the next NMI.  Because the event is
342		 * already handled, the next NMI will result in an unknown
343		 * NMI.  Instead lets flag this for a potential NMI to
344		 * swallow.
345		 */
346		if (handled > 1)
347			__this_cpu_write(swallow_nmi, true);
348		return;
349	}
350
351	/*
352	 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
353	 *
354	 * Another CPU may be processing panic routines while holding
355	 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
356	 * and if so, call its callback directly.  If there is no CPU preparing
357	 * crash dump, we simply loop here.
358	 */
359	while (!raw_spin_trylock(&nmi_reason_lock)) {
360		run_crash_ipi_callback(regs);
361		cpu_relax();
362	}
363
364	reason = x86_platform.get_nmi_reason();
365
366	if (reason & NMI_REASON_MASK) {
367		if (reason & NMI_REASON_SERR)
368			pci_serr_error(reason, regs);
369		else if (reason & NMI_REASON_IOCHK)
370			io_check_error(reason, regs);
371#ifdef CONFIG_X86_32
372		/*
373		 * Reassert NMI in case it became active
374		 * meanwhile as it's edge-triggered:
375		 */
376		reassert_nmi();
377#endif
378		__this_cpu_add(nmi_stats.external, 1);
379		raw_spin_unlock(&nmi_reason_lock);
380		return;
381	}
382	raw_spin_unlock(&nmi_reason_lock);
383
384	/*
385	 * Only one NMI can be latched at a time.  To handle
386	 * this we may process multiple nmi handlers at once to
387	 * cover the case where an NMI is dropped.  The downside
388	 * to this approach is we may process an NMI prematurely,
389	 * while its real NMI is sitting latched.  This will cause
390	 * an unknown NMI on the next run of the NMI processing.
391	 *
392	 * We tried to flag that condition above, by setting the
393	 * swallow_nmi flag when we process more than one event.
394	 * This condition is also only present on the second half
395	 * of a back-to-back NMI, so we flag that condition too.
396	 *
397	 * If both are true, we assume we already processed this
398	 * NMI previously and we swallow it.  Otherwise we reset
399	 * the logic.
400	 *
401	 * There are scenarios where we may accidentally swallow
402	 * a 'real' unknown NMI.  For example, while processing
403	 * a perf NMI another perf NMI comes in along with a
404	 * 'real' unknown NMI.  These two NMIs get combined into
405	 * one (as descibed above).  When the next NMI gets
406	 * processed, it will be flagged by perf as handled, but
407	 * noone will know that there was a 'real' unknown NMI sent
408	 * also.  As a result it gets swallowed.  Or if the first
409	 * perf NMI returns two events handled then the second
410	 * NMI will get eaten by the logic below, again losing a
411	 * 'real' unknown NMI.  But this is the best we can do
412	 * for now.
413	 */
414	if (b2b && __this_cpu_read(swallow_nmi))
415		__this_cpu_add(nmi_stats.swallow, 1);
416	else
417		unknown_nmi_error(reason, regs);
 
 
 
418}
419NOKPROBE_SYMBOL(default_do_nmi);
420
421/*
422 * NMIs can page fault or hit breakpoints which will cause it to lose
423 * its NMI context with the CPU when the breakpoint or page fault does an IRET.
424 *
425 * As a result, NMIs can nest if NMIs get unmasked due an IRET during
426 * NMI processing.  On x86_64, the asm glue protects us from nested NMIs
427 * if the outer NMI came from kernel mode, but we can still nest if the
428 * outer NMI came from user mode.
429 *
430 * To handle these nested NMIs, we have three states:
431 *
432 *  1) not running
433 *  2) executing
434 *  3) latched
435 *
436 * When no NMI is in progress, it is in the "not running" state.
437 * When an NMI comes in, it goes into the "executing" state.
438 * Normally, if another NMI is triggered, it does not interrupt
439 * the running NMI and the HW will simply latch it so that when
440 * the first NMI finishes, it will restart the second NMI.
441 * (Note, the latch is binary, thus multiple NMIs triggering,
442 *  when one is running, are ignored. Only one NMI is restarted.)
443 *
444 * If an NMI executes an iret, another NMI can preempt it. We do not
445 * want to allow this new NMI to run, but we want to execute it when the
446 * first one finishes.  We set the state to "latched", and the exit of
447 * the first NMI will perform a dec_return, if the result is zero
448 * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the
449 * dec_return would have set the state to NMI_EXECUTING (what we want it
450 * to be when we are running). In this case, we simply jump back to
451 * rerun the NMI handler again, and restart the 'latched' NMI.
452 *
453 * No trap (breakpoint or page fault) should be hit before nmi_restart,
454 * thus there is no race between the first check of state for NOT_RUNNING
455 * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs
456 * at this point.
457 *
458 * In case the NMI takes a page fault, we need to save off the CR2
459 * because the NMI could have preempted another page fault and corrupt
460 * the CR2 that is about to be read. As nested NMIs must be restarted
461 * and they can not take breakpoints or page faults, the update of the
462 * CR2 must be done before converting the nmi state back to NOT_RUNNING.
463 * Otherwise, there would be a race of another nested NMI coming in
464 * after setting state to NOT_RUNNING but before updating the nmi_cr2.
465 */
466enum nmi_states {
467	NMI_NOT_RUNNING = 0,
468	NMI_EXECUTING,
469	NMI_LATCHED,
470};
471static DEFINE_PER_CPU(enum nmi_states, nmi_state);
472static DEFINE_PER_CPU(unsigned long, nmi_cr2);
 
473
474#ifdef CONFIG_X86_64
475/*
476 * In x86_64, we need to handle breakpoint -> NMI -> breakpoint.  Without
477 * some care, the inner breakpoint will clobber the outer breakpoint's
478 * stack.
479 *
480 * If a breakpoint is being processed, and the debug stack is being
481 * used, if an NMI comes in and also hits a breakpoint, the stack
482 * pointer will be set to the same fixed address as the breakpoint that
483 * was interrupted, causing that stack to be corrupted. To handle this
484 * case, check if the stack that was interrupted is the debug stack, and
485 * if so, change the IDT so that new breakpoints will use the current
486 * stack and not switch to the fixed address. On return of the NMI,
487 * switch back to the original IDT.
488 */
489static DEFINE_PER_CPU(int, update_debug_stack);
490#endif
491
492dotraplinkage notrace void
493do_nmi(struct pt_regs *regs, long error_code)
494{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
495	if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
496		this_cpu_write(nmi_state, NMI_LATCHED);
497		return;
498	}
499	this_cpu_write(nmi_state, NMI_EXECUTING);
500	this_cpu_write(nmi_cr2, read_cr2());
 
501nmi_restart:
 
 
 
 
 
502
503#ifdef CONFIG_X86_64
504	/*
505	 * If we interrupted a breakpoint, it is possible that
506	 * the nmi handler will have breakpoints too. We need to
507	 * change the IDT such that breakpoints that happen here
508	 * continue to use the NMI stack.
509	 */
510	if (unlikely(is_debug_stack(regs->sp))) {
511		debug_stack_set_zero();
512		this_cpu_write(update_debug_stack, 1);
513	}
514#endif
515
516	nmi_enter();
 
 
517
518	inc_irq_stat(__nmi_count);
519
520	if (!ignore_nmis)
 
 
 
 
 
 
521		default_do_nmi(regs);
 
 
 
 
 
522
523	nmi_exit();
524
525#ifdef CONFIG_X86_64
526	if (unlikely(this_cpu_read(update_debug_stack))) {
527		debug_stack_reset();
528		this_cpu_write(update_debug_stack, 0);
529	}
530#endif
531
532	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
533		write_cr2(this_cpu_read(nmi_cr2));
 
 
 
 
 
534	if (this_cpu_dec_return(nmi_state))
535		goto nmi_restart;
536}
537NOKPROBE_SYMBOL(do_nmi);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
538
539void stop_nmi(void)
540{
541	ignore_nmis++;
542}
543
544void restart_nmi(void)
545{
546	ignore_nmis--;
547}
548
549/* reset the back-to-back NMI logic */
550void local_touch_nmi(void)
551{
552	__this_cpu_write(last_nmi_rip, 0);
553}
554EXPORT_SYMBOL_GPL(local_touch_nmi);
v6.8
  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#include <asm/microcode.h>
 37#include <asm/sev.h>
 38
 39#define CREATE_TRACE_POINTS
 40#include <trace/events/nmi.h>
 41
 42struct nmi_desc {
 43	raw_spinlock_t lock;
 44	struct list_head head;
 45};
 46
 47static 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
 68struct nmi_stats {
 69	unsigned int normal;
 70	unsigned int unknown;
 71	unsigned int external;
 72	unsigned int swallow;
 73	unsigned long recv_jiffies;
 74	unsigned long idt_seq;
 75	unsigned long idt_nmi_seq;
 76	unsigned long idt_ignored;
 77	atomic_long_t idt_calls;
 78	unsigned long idt_seq_snap;
 79	unsigned long idt_nmi_seq_snap;
 80	unsigned long idt_ignored_snap;
 81	long idt_calls_snap;
 82};
 83
 84static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
 85
 86static int ignore_nmis __read_mostly;
 87
 88int unknown_nmi_panic;
 89/*
 90 * Prevent NMI reason port (0x61) being accessed simultaneously, can
 91 * only be used in NMI handler.
 92 */
 93static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
 94
 95static int __init setup_unknown_nmi_panic(char *str)
 96{
 97	unknown_nmi_panic = 1;
 98	return 1;
 99}
100__setup("unknown_nmi_panic", setup_unknown_nmi_panic);
101
102#define nmi_to_desc(type) (&nmi_desc[type])
103
104static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
105
106static int __init nmi_warning_debugfs(void)
107{
108	debugfs_create_u64("nmi_longest_ns", 0644,
109			arch_debugfs_dir, &nmi_longest_ns);
110	return 0;
111}
112fs_initcall(nmi_warning_debugfs);
113
114static void nmi_check_duration(struct nmiaction *action, u64 duration)
115{
 
116	int remainder_ns, decimal_msecs;
 
117
118	if (duration < nmi_longest_ns || duration < action->max_duration)
119		return;
120
121	action->max_duration = duration;
122
123	remainder_ns = do_div(duration, (1000 * 1000));
124	decimal_msecs = remainder_ns / 1000;
125
126	printk_ratelimited(KERN_INFO
127		"INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
128		action->handler, duration, decimal_msecs);
129}
130
131static int nmi_handle(unsigned int type, struct pt_regs *regs)
132{
133	struct nmi_desc *desc = nmi_to_desc(type);
134	struct nmiaction *a;
135	int handled=0;
136
137	rcu_read_lock();
138
139	/*
140	 * NMIs are edge-triggered, which means if you have enough
141	 * of them concurrently, you can lose some because only one
142	 * can be latched at any given time.  Walk the whole list
143	 * to handle those situations.
144	 */
145	list_for_each_entry_rcu(a, &desc->head, list) {
146		int thishandled;
147		u64 delta;
148
149		delta = sched_clock();
150		thishandled = a->handler(type, regs);
151		handled += thishandled;
152		delta = sched_clock() - delta;
153		trace_nmi_handler(a->handler, (int)delta, thishandled);
154
155		nmi_check_duration(a, delta);
 
 
 
 
156	}
157
158	rcu_read_unlock();
159
160	/* return total number of NMI events handled */
161	return handled;
162}
163NOKPROBE_SYMBOL(nmi_handle);
164
165int __register_nmi_handler(unsigned int type, struct nmiaction *action)
166{
167	struct nmi_desc *desc = nmi_to_desc(type);
168	unsigned long flags;
169
170	if (WARN_ON_ONCE(!action->handler || !list_empty(&action->list)))
171		return -EINVAL;
172
 
 
173	raw_spin_lock_irqsave(&desc->lock, flags);
174
175	/*
176	 * Indicate if there are multiple registrations on the
177	 * internal NMI handler call chains (SERR and IO_CHECK).
178	 */
179	WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
180	WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
181
182	/*
183	 * some handlers need to be executed first otherwise a fake
184	 * event confuses some handlers (kdump uses this flag)
185	 */
186	if (action->flags & NMI_FLAG_FIRST)
187		list_add_rcu(&action->list, &desc->head);
188	else
189		list_add_tail_rcu(&action->list, &desc->head);
190
191	raw_spin_unlock_irqrestore(&desc->lock, flags);
192	return 0;
193}
194EXPORT_SYMBOL(__register_nmi_handler);
195
196void unregister_nmi_handler(unsigned int type, const char *name)
197{
198	struct nmi_desc *desc = nmi_to_desc(type);
199	struct nmiaction *n, *found = NULL;
200	unsigned long flags;
201
202	raw_spin_lock_irqsave(&desc->lock, flags);
203
204	list_for_each_entry_rcu(n, &desc->head, list) {
205		/*
206		 * the name passed in to describe the nmi handler
207		 * is used as the lookup key
208		 */
209		if (!strcmp(n->name, name)) {
210			WARN(in_nmi(),
211				"Trying to free NMI (%s) from NMI context!\n", n->name);
212			list_del_rcu(&n->list);
213			found = n;
214			break;
215		}
216	}
217
218	raw_spin_unlock_irqrestore(&desc->lock, flags);
219	if (found) {
220		synchronize_rcu();
221		INIT_LIST_HEAD(&found->list);
222	}
223}
224EXPORT_SYMBOL_GPL(unregister_nmi_handler);
225
226static void
227pci_serr_error(unsigned char reason, struct pt_regs *regs)
228{
229	/* check to see if anyone registered against these types of errors */
230	if (nmi_handle(NMI_SERR, regs))
231		return;
232
233	pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
234		 reason, smp_processor_id());
235
236	if (panic_on_unrecovered_nmi)
237		nmi_panic(regs, "NMI: Not continuing");
238
239	pr_emerg("Dazed and confused, but trying to continue\n");
240
241	/* Clear and disable the PCI SERR error line. */
242	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
243	outb(reason, NMI_REASON_PORT);
244}
245NOKPROBE_SYMBOL(pci_serr_error);
246
247static void
248io_check_error(unsigned char reason, struct pt_regs *regs)
249{
250	unsigned long i;
251
252	/* check to see if anyone registered against these types of errors */
253	if (nmi_handle(NMI_IO_CHECK, regs))
254		return;
255
256	pr_emerg(
257	"NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
258		 reason, smp_processor_id());
259	show_regs(regs);
260
261	if (panic_on_io_nmi) {
262		nmi_panic(regs, "NMI IOCK error: Not continuing");
263
264		/*
265		 * If we end up here, it means we have received an NMI while
266		 * processing panic(). Simply return without delaying and
267		 * re-enabling NMIs.
268		 */
269		return;
270	}
271
272	/* Re-enable the IOCK line, wait for a few seconds */
273	reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
274	outb(reason, NMI_REASON_PORT);
275
276	i = 20000;
277	while (--i) {
278		touch_nmi_watchdog();
279		udelay(100);
280	}
281
282	reason &= ~NMI_REASON_CLEAR_IOCHK;
283	outb(reason, NMI_REASON_PORT);
284}
285NOKPROBE_SYMBOL(io_check_error);
286
287static void
288unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
289{
290	int handled;
291
292	/*
293	 * Use 'false' as back-to-back NMIs are dealt with one level up.
294	 * Of course this makes having multiple 'unknown' handlers useless
295	 * as only the first one is ever run (unless it can actually determine
296	 * if it caused the NMI)
297	 */
298	handled = nmi_handle(NMI_UNKNOWN, regs);
299	if (handled) {
300		__this_cpu_add(nmi_stats.unknown, handled);
301		return;
302	}
303
304	__this_cpu_add(nmi_stats.unknown, 1);
305
306	pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
307		 reason, smp_processor_id());
308
 
309	if (unknown_nmi_panic || panic_on_unrecovered_nmi)
310		nmi_panic(regs, "NMI: Not continuing");
311
312	pr_emerg("Dazed and confused, but trying to continue\n");
313}
314NOKPROBE_SYMBOL(unknown_nmi_error);
315
316static DEFINE_PER_CPU(bool, swallow_nmi);
317static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
318
319static noinstr void default_do_nmi(struct pt_regs *regs)
320{
321	unsigned char reason = 0;
322	int handled;
323	bool b2b = false;
324
325	/*
326	 * CPU-specific NMI must be processed before non-CPU-specific
327	 * NMI, otherwise we may lose it, because the CPU-specific
328	 * NMI can not be detected/processed on other CPUs.
329	 */
330
331	/*
332	 * Back-to-back NMIs are interesting because they can either
333	 * be two NMI or more than two NMIs (any thing over two is dropped
334	 * due to NMI being edge-triggered).  If this is the second half
335	 * of the back-to-back NMI, assume we dropped things and process
336	 * more handlers.  Otherwise reset the 'swallow' NMI behaviour
337	 */
338	if (regs->ip == __this_cpu_read(last_nmi_rip))
339		b2b = true;
340	else
341		__this_cpu_write(swallow_nmi, false);
342
343	__this_cpu_write(last_nmi_rip, regs->ip);
344
345	instrumentation_begin();
346
347	if (microcode_nmi_handler_enabled() && microcode_nmi_handler())
348		goto out;
349
350	handled = nmi_handle(NMI_LOCAL, regs);
351	__this_cpu_add(nmi_stats.normal, handled);
352	if (handled) {
353		/*
354		 * There are cases when a NMI handler handles multiple
355		 * events in the current NMI.  One of these events may
356		 * be queued for in the next NMI.  Because the event is
357		 * already handled, the next NMI will result in an unknown
358		 * NMI.  Instead lets flag this for a potential NMI to
359		 * swallow.
360		 */
361		if (handled > 1)
362			__this_cpu_write(swallow_nmi, true);
363		goto out;
364	}
365
366	/*
367	 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
368	 *
369	 * Another CPU may be processing panic routines while holding
370	 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
371	 * and if so, call its callback directly.  If there is no CPU preparing
372	 * crash dump, we simply loop here.
373	 */
374	while (!raw_spin_trylock(&nmi_reason_lock)) {
375		run_crash_ipi_callback(regs);
376		cpu_relax();
377	}
378
379	reason = x86_platform.get_nmi_reason();
380
381	if (reason & NMI_REASON_MASK) {
382		if (reason & NMI_REASON_SERR)
383			pci_serr_error(reason, regs);
384		else if (reason & NMI_REASON_IOCHK)
385			io_check_error(reason, regs);
386#ifdef CONFIG_X86_32
387		/*
388		 * Reassert NMI in case it became active
389		 * meanwhile as it's edge-triggered:
390		 */
391		reassert_nmi();
392#endif
393		__this_cpu_add(nmi_stats.external, 1);
394		raw_spin_unlock(&nmi_reason_lock);
395		goto out;
396	}
397	raw_spin_unlock(&nmi_reason_lock);
398
399	/*
400	 * Only one NMI can be latched at a time.  To handle
401	 * this we may process multiple nmi handlers at once to
402	 * cover the case where an NMI is dropped.  The downside
403	 * to this approach is we may process an NMI prematurely,
404	 * while its real NMI is sitting latched.  This will cause
405	 * an unknown NMI on the next run of the NMI processing.
406	 *
407	 * We tried to flag that condition above, by setting the
408	 * swallow_nmi flag when we process more than one event.
409	 * This condition is also only present on the second half
410	 * of a back-to-back NMI, so we flag that condition too.
411	 *
412	 * If both are true, we assume we already processed this
413	 * NMI previously and we swallow it.  Otherwise we reset
414	 * the logic.
415	 *
416	 * There are scenarios where we may accidentally swallow
417	 * a 'real' unknown NMI.  For example, while processing
418	 * a perf NMI another perf NMI comes in along with a
419	 * 'real' unknown NMI.  These two NMIs get combined into
420	 * one (as described above).  When the next NMI gets
421	 * processed, it will be flagged by perf as handled, but
422	 * no one will know that there was a 'real' unknown NMI sent
423	 * also.  As a result it gets swallowed.  Or if the first
424	 * perf NMI returns two events handled then the second
425	 * NMI will get eaten by the logic below, again losing a
426	 * 'real' unknown NMI.  But this is the best we can do
427	 * for now.
428	 */
429	if (b2b && __this_cpu_read(swallow_nmi))
430		__this_cpu_add(nmi_stats.swallow, 1);
431	else
432		unknown_nmi_error(reason, regs);
433
434out:
435	instrumentation_end();
436}
 
437
438/*
439 * NMIs can page fault or hit breakpoints which will cause it to lose
440 * its NMI context with the CPU when the breakpoint or page fault does an IRET.
441 *
442 * As a result, NMIs can nest if NMIs get unmasked due an IRET during
443 * NMI processing.  On x86_64, the asm glue protects us from nested NMIs
444 * if the outer NMI came from kernel mode, but we can still nest if the
445 * outer NMI came from user mode.
446 *
447 * To handle these nested NMIs, we have three states:
448 *
449 *  1) not running
450 *  2) executing
451 *  3) latched
452 *
453 * When no NMI is in progress, it is in the "not running" state.
454 * When an NMI comes in, it goes into the "executing" state.
455 * Normally, if another NMI is triggered, it does not interrupt
456 * the running NMI and the HW will simply latch it so that when
457 * the first NMI finishes, it will restart the second NMI.
458 * (Note, the latch is binary, thus multiple NMIs triggering,
459 *  when one is running, are ignored. Only one NMI is restarted.)
460 *
461 * If an NMI executes an iret, another NMI can preempt it. We do not
462 * want to allow this new NMI to run, but we want to execute it when the
463 * first one finishes.  We set the state to "latched", and the exit of
464 * the first NMI will perform a dec_return, if the result is zero
465 * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the
466 * dec_return would have set the state to NMI_EXECUTING (what we want it
467 * to be when we are running). In this case, we simply jump back to
468 * rerun the NMI handler again, and restart the 'latched' NMI.
469 *
470 * No trap (breakpoint or page fault) should be hit before nmi_restart,
471 * thus there is no race between the first check of state for NOT_RUNNING
472 * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs
473 * at this point.
474 *
475 * In case the NMI takes a page fault, we need to save off the CR2
476 * because the NMI could have preempted another page fault and corrupt
477 * the CR2 that is about to be read. As nested NMIs must be restarted
478 * and they can not take breakpoints or page faults, the update of the
479 * CR2 must be done before converting the nmi state back to NOT_RUNNING.
480 * Otherwise, there would be a race of another nested NMI coming in
481 * after setting state to NOT_RUNNING but before updating the nmi_cr2.
482 */
483enum nmi_states {
484	NMI_NOT_RUNNING = 0,
485	NMI_EXECUTING,
486	NMI_LATCHED,
487};
488static DEFINE_PER_CPU(enum nmi_states, nmi_state);
489static DEFINE_PER_CPU(unsigned long, nmi_cr2);
490static DEFINE_PER_CPU(unsigned long, nmi_dr7);
491
492DEFINE_IDTENTRY_RAW(exc_nmi)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
493{
494	irqentry_state_t irq_state;
495	struct nmi_stats *nsp = this_cpu_ptr(&nmi_stats);
496
497	/*
498	 * Re-enable NMIs right here when running as an SEV-ES guest. This might
499	 * cause nested NMIs, but those can be handled safely.
500	 */
501	sev_es_nmi_complete();
502	if (IS_ENABLED(CONFIG_NMI_CHECK_CPU))
503		raw_atomic_long_inc(&nsp->idt_calls);
504
505	if (IS_ENABLED(CONFIG_SMP) && arch_cpu_is_offline(smp_processor_id())) {
506		if (microcode_nmi_handler_enabled())
507			microcode_offline_nmi_handler();
508		return;
509	}
510
511	if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
512		this_cpu_write(nmi_state, NMI_LATCHED);
513		return;
514	}
515	this_cpu_write(nmi_state, NMI_EXECUTING);
516	this_cpu_write(nmi_cr2, read_cr2());
517
518nmi_restart:
519	if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
520		WRITE_ONCE(nsp->idt_seq, nsp->idt_seq + 1);
521		WARN_ON_ONCE(!(nsp->idt_seq & 0x1));
522		WRITE_ONCE(nsp->recv_jiffies, jiffies);
523	}
524
 
525	/*
526	 * Needs to happen before DR7 is accessed, because the hypervisor can
527	 * intercept DR7 reads/writes, turning those into #VC exceptions.
 
 
528	 */
529	sev_es_ist_enter(regs);
 
 
 
 
530
531	this_cpu_write(nmi_dr7, local_db_save());
532
533	irq_state = irqentry_nmi_enter(regs);
534
535	inc_irq_stat(__nmi_count);
536
537	if (IS_ENABLED(CONFIG_NMI_CHECK_CPU) && ignore_nmis) {
538		WRITE_ONCE(nsp->idt_ignored, nsp->idt_ignored + 1);
539	} else if (!ignore_nmis) {
540		if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
541			WRITE_ONCE(nsp->idt_nmi_seq, nsp->idt_nmi_seq + 1);
542			WARN_ON_ONCE(!(nsp->idt_nmi_seq & 0x1));
543		}
544		default_do_nmi(regs);
545		if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
546			WRITE_ONCE(nsp->idt_nmi_seq, nsp->idt_nmi_seq + 1);
547			WARN_ON_ONCE(nsp->idt_nmi_seq & 0x1);
548		}
549	}
550
551	irqentry_nmi_exit(regs, irq_state);
552
553	local_db_restore(this_cpu_read(nmi_dr7));
554
555	sev_es_ist_exit();
 
 
 
556
557	if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
558		write_cr2(this_cpu_read(nmi_cr2));
559	if (IS_ENABLED(CONFIG_NMI_CHECK_CPU)) {
560		WRITE_ONCE(nsp->idt_seq, nsp->idt_seq + 1);
561		WARN_ON_ONCE(nsp->idt_seq & 0x1);
562		WRITE_ONCE(nsp->recv_jiffies, jiffies);
563	}
564	if (this_cpu_dec_return(nmi_state))
565		goto nmi_restart;
566}
567
568#if IS_ENABLED(CONFIG_KVM_INTEL)
569DEFINE_IDTENTRY_RAW(exc_nmi_kvm_vmx)
570{
571	exc_nmi(regs);
572}
573#if IS_MODULE(CONFIG_KVM_INTEL)
574EXPORT_SYMBOL_GPL(asm_exc_nmi_kvm_vmx);
575#endif
576#endif
577
578#ifdef CONFIG_NMI_CHECK_CPU
579
580static char *nmi_check_stall_msg[] = {
581/*									*/
582/* +--------- nsp->idt_seq_snap & 0x1: CPU is in NMI handler.		*/
583/* | +------ cpu_is_offline(cpu)					*/
584/* | | +--- nsp->idt_calls_snap != atomic_long_read(&nsp->idt_calls):	*/
585/* | | |	NMI handler has been invoked.				*/
586/* | | |								*/
587/* V V V								*/
588/* 0 0 0 */ "NMIs are not reaching exc_nmi() handler",
589/* 0 0 1 */ "exc_nmi() handler is ignoring NMIs",
590/* 0 1 0 */ "CPU is offline and NMIs are not reaching exc_nmi() handler",
591/* 0 1 1 */ "CPU is offline and exc_nmi() handler is legitimately ignoring NMIs",
592/* 1 0 0 */ "CPU is in exc_nmi() handler and no further NMIs are reaching handler",
593/* 1 0 1 */ "CPU is in exc_nmi() handler which is legitimately ignoring NMIs",
594/* 1 1 0 */ "CPU is offline in exc_nmi() handler and no more NMIs are reaching exc_nmi() handler",
595/* 1 1 1 */ "CPU is offline in exc_nmi() handler which is legitimately ignoring NMIs",
596};
597
598void nmi_backtrace_stall_snap(const struct cpumask *btp)
599{
600	int cpu;
601	struct nmi_stats *nsp;
602
603	for_each_cpu(cpu, btp) {
604		nsp = per_cpu_ptr(&nmi_stats, cpu);
605		nsp->idt_seq_snap = READ_ONCE(nsp->idt_seq);
606		nsp->idt_nmi_seq_snap = READ_ONCE(nsp->idt_nmi_seq);
607		nsp->idt_ignored_snap = READ_ONCE(nsp->idt_ignored);
608		nsp->idt_calls_snap = atomic_long_read(&nsp->idt_calls);
609	}
610}
611
612void nmi_backtrace_stall_check(const struct cpumask *btp)
613{
614	int cpu;
615	int idx;
616	unsigned long nmi_seq;
617	unsigned long j = jiffies;
618	char *modp;
619	char *msgp;
620	char *msghp;
621	struct nmi_stats *nsp;
622
623	for_each_cpu(cpu, btp) {
624		nsp = per_cpu_ptr(&nmi_stats, cpu);
625		modp = "";
626		msghp = "";
627		nmi_seq = READ_ONCE(nsp->idt_nmi_seq);
628		if (nsp->idt_nmi_seq_snap + 1 == nmi_seq && (nmi_seq & 0x1)) {
629			msgp = "CPU entered NMI handler function, but has not exited";
630		} else if ((nsp->idt_nmi_seq_snap & 0x1) != (nmi_seq & 0x1)) {
631			msgp = "CPU is handling NMIs";
632		} else {
633			idx = ((nsp->idt_seq_snap & 0x1) << 2) |
634			      (cpu_is_offline(cpu) << 1) |
635			      (nsp->idt_calls_snap != atomic_long_read(&nsp->idt_calls));
636			msgp = nmi_check_stall_msg[idx];
637			if (nsp->idt_ignored_snap != READ_ONCE(nsp->idt_ignored) && (idx & 0x1))
638				modp = ", but OK because ignore_nmis was set";
639			if (nmi_seq & ~0x1)
640				msghp = " (CPU currently in NMI handler function)";
641			else if (nsp->idt_nmi_seq_snap + 1 == nmi_seq)
642				msghp = " (CPU exited one NMI handler function)";
643		}
644		pr_alert("%s: CPU %d: %s%s%s, last activity: %lu jiffies ago.\n",
645			 __func__, cpu, msgp, modp, msghp, j - READ_ONCE(nsp->recv_jiffies));
646	}
647}
648
649#endif
650
651void stop_nmi(void)
652{
653	ignore_nmis++;
654}
655
656void restart_nmi(void)
657{
658	ignore_nmis--;
659}
660
661/* reset the back-to-back NMI logic */
662void local_touch_nmi(void)
663{
664	__this_cpu_write(last_nmi_rip, 0);
665}
666EXPORT_SYMBOL_GPL(local_touch_nmi);