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