blob: 086cf1d1d71d820e13baca26091d6104abea9eb9 [file] [log] [blame]
Don Zickus1d489222011-09-30 15:06:19 -04001/*
2 * Copyright (C) 1991, 1992 Linus Torvalds
3 * Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
Don Zickus9c48f1c2011-09-30 15:06:21 -04004 * Copyright (C) 2011 Don Zickus Red Hat, Inc.
Don Zickus1d489222011-09-30 15:06:19 -04005 *
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>
Ingo Molnarb17b0152017-02-08 18:51:35 +010016#include <linux/sched/debug.h>
Don Zickus1d489222011-09-30 15:06:19 -040017#include <linux/nmi.h>
Dave Hansen2ab004562013-06-21 08:51:35 -070018#include <linux/debugfs.h>
Don Zickusc9126b22011-09-30 15:06:20 -040019#include <linux/delay.h>
20#include <linux/hardirq.h>
Arnd Bergmannc361db52016-06-06 16:09:52 +020021#include <linux/ratelimit.h>
Don Zickusc9126b22011-09-30 15:06:20 -040022#include <linux/slab.h>
Paul Gortmaker69c60c82011-05-26 12:22:53 -040023#include <linux/export.h>
Ingo Molnare6017572017-02-01 16:36:40 +010024#include <linux/sched/clock.h>
Don Zickus1d489222011-09-30 15:06:19 -040025
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>
Don Zickusc9126b22011-09-30 15:06:20 -040033#include <asm/nmi.h>
Mathias Nyman6fd36ba2011-11-10 13:45:24 +000034#include <asm/x86_init.h>
Hidehiro Kawaib279d672015-12-14 11:19:13 +010035#include <asm/reboot.h>
Kostenzer Felix8e2a7f52016-03-06 23:20:06 +010036#include <asm/cache.h>
Thomas Gleixner04dcbdb2019-02-18 23:42:51 +010037#include <asm/nospec-branch.h>
Don Zickusc9126b22011-09-30 15:06:20 -040038
Dave Hansen0c4df022013-06-21 08:51:38 -070039#define CREATE_TRACE_POINTS
40#include <trace/events/nmi.h>
41
Don Zickusc9126b22011-09-30 15:06:20 -040042struct nmi_desc {
Scott Woodc455fd92017-07-24 16:32:42 -050043 raw_spinlock_t lock;
Don Zickusc9126b22011-09-30 15:06:20 -040044 struct list_head head;
45};
46
47static struct nmi_desc nmi_desc[NMI_MAX] =
48{
49 {
Scott Woodc455fd92017-07-24 16:32:42 -050050 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
Don Zickusc9126b22011-09-30 15:06:20 -040051 .head = LIST_HEAD_INIT(nmi_desc[0].head),
52 },
53 {
Scott Woodc455fd92017-07-24 16:32:42 -050054 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
Don Zickusc9126b22011-09-30 15:06:20 -040055 .head = LIST_HEAD_INIT(nmi_desc[1].head),
56 },
Don Zickus553222f2012-03-29 16:11:16 -040057 {
Scott Woodc455fd92017-07-24 16:32:42 -050058 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
Don Zickus553222f2012-03-29 16:11:16 -040059 .head = LIST_HEAD_INIT(nmi_desc[2].head),
60 },
61 {
Scott Woodc455fd92017-07-24 16:32:42 -050062 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
Don Zickus553222f2012-03-29 16:11:16 -040063 .head = LIST_HEAD_INIT(nmi_desc[3].head),
64 },
Don Zickusc9126b22011-09-30 15:06:20 -040065
66};
Don Zickus1d489222011-09-30 15:06:19 -040067
Don Zickusefc3aac2011-09-30 15:06:23 -040068struct nmi_stats {
69 unsigned int normal;
70 unsigned int unknown;
71 unsigned int external;
72 unsigned int swallow;
73};
74
75static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
76
Kostenzer Felix8e2a7f52016-03-06 23:20:06 +010077static int ignore_nmis __read_mostly;
Don Zickus1d489222011-09-30 15:06:19 -040078
79int unknown_nmi_panic;
80/*
81 * Prevent NMI reason port (0x61) being accessed simultaneously, can
82 * only be used in NMI handler.
83 */
84static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
85
86static int __init setup_unknown_nmi_panic(char *str)
87{
88 unknown_nmi_panic = 1;
89 return 1;
90}
91__setup("unknown_nmi_panic", setup_unknown_nmi_panic);
92
Don Zickusc9126b22011-09-30 15:06:20 -040093#define nmi_to_desc(type) (&nmi_desc[type])
94
Dave Hansen2ab004562013-06-21 08:51:35 -070095static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
Peter Zijlstrae90c7852014-02-03 18:02:09 +010096
Dave Hansen2ab004562013-06-21 08:51:35 -070097static int __init nmi_warning_debugfs(void)
98{
99 debugfs_create_u64("nmi_longest_ns", 0644,
100 arch_debugfs_dir, &nmi_longest_ns);
101 return 0;
102}
103fs_initcall(nmi_warning_debugfs);
104
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100105static void nmi_max_handler(struct irq_work *w)
106{
107 struct nmiaction *a = container_of(w, struct nmiaction, irq_work);
108 int remainder_ns, decimal_msecs;
Mark Rutland6aa7de02017-10-23 14:07:29 -0700109 u64 whole_msecs = READ_ONCE(a->max_duration);
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100110
111 remainder_ns = do_div(whole_msecs, (1000 * 1000));
112 decimal_msecs = remainder_ns / 1000;
113
114 printk_ratelimited(KERN_INFO
115 "INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
116 a->handler, whole_msecs, decimal_msecs);
117}
118
Andy Lutomirskibf9f2ee2015-07-20 11:52:23 -0700119static int nmi_handle(unsigned int type, struct pt_regs *regs)
Don Zickusc9126b22011-09-30 15:06:20 -0400120{
121 struct nmi_desc *desc = nmi_to_desc(type);
122 struct nmiaction *a;
123 int handled=0;
124
125 rcu_read_lock();
126
127 /*
128 * NMIs are edge-triggered, which means if you have enough
129 * of them concurrently, you can lose some because only one
130 * can be latched at any given time. Walk the whole list
131 * to handle those situations.
132 */
Dave Hansen2ab004562013-06-21 08:51:35 -0700133 list_for_each_entry_rcu(a, &desc->head, list) {
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100134 int thishandled;
135 u64 delta;
Dave Hansen2ab004562013-06-21 08:51:35 -0700136
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100137 delta = sched_clock();
Dave Hansen0c4df022013-06-21 08:51:38 -0700138 thishandled = a->handler(type, regs);
139 handled += thishandled;
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100140 delta = sched_clock() - delta;
Dave Hansen0c4df022013-06-21 08:51:38 -0700141 trace_nmi_handler(a->handler, (int)delta, thishandled);
Dave Hansen2ab004562013-06-21 08:51:35 -0700142
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100143 if (delta < nmi_longest_ns || delta < a->max_duration)
Dave Hansen2ab004562013-06-21 08:51:35 -0700144 continue;
145
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100146 a->max_duration = delta;
147 irq_work_queue(&a->irq_work);
Dave Hansen2ab004562013-06-21 08:51:35 -0700148 }
Don Zickusc9126b22011-09-30 15:06:20 -0400149
Don Zickusc9126b22011-09-30 15:06:20 -0400150 rcu_read_unlock();
151
152 /* return total number of NMI events handled */
153 return handled;
154}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900155NOKPROBE_SYMBOL(nmi_handle);
Don Zickusc9126b22011-09-30 15:06:20 -0400156
Li Zhong72b3fb22012-03-29 16:11:17 -0400157int __register_nmi_handler(unsigned int type, struct nmiaction *action)
Don Zickusc9126b22011-09-30 15:06:20 -0400158{
159 struct nmi_desc *desc = nmi_to_desc(type);
160 unsigned long flags;
161
Li Zhong72b3fb22012-03-29 16:11:17 -0400162 if (!action->handler)
163 return -EINVAL;
164
Peter Zijlstrae90c7852014-02-03 18:02:09 +0100165 init_irq_work(&action->irq_work, nmi_max_handler);
166
Scott Woodc455fd92017-07-24 16:32:42 -0500167 raw_spin_lock_irqsave(&desc->lock, flags);
Don Zickusc9126b22011-09-30 15:06:20 -0400168
169 /*
Mike Travis0d443b72017-03-07 15:08:42 -0600170 * Indicate if there are multiple registrations on the
171 * internal NMI handler call chains (SERR and IO_CHECK).
Don Zickusb227e232011-09-30 15:06:22 -0400172 */
Don Zickus553222f2012-03-29 16:11:16 -0400173 WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
174 WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
Don Zickusb227e232011-09-30 15:06:22 -0400175
176 /*
Don Zickusc9126b22011-09-30 15:06:20 -0400177 * some handlers need to be executed first otherwise a fake
178 * event confuses some handlers (kdump uses this flag)
179 */
180 if (action->flags & NMI_FLAG_FIRST)
181 list_add_rcu(&action->list, &desc->head);
182 else
183 list_add_tail_rcu(&action->list, &desc->head);
184
Scott Woodc455fd92017-07-24 16:32:42 -0500185 raw_spin_unlock_irqrestore(&desc->lock, flags);
Don Zickusc9126b22011-09-30 15:06:20 -0400186 return 0;
187}
Li Zhong72b3fb22012-03-29 16:11:17 -0400188EXPORT_SYMBOL(__register_nmi_handler);
Don Zickusc9126b22011-09-30 15:06:20 -0400189
Li Zhong72b3fb22012-03-29 16:11:17 -0400190void unregister_nmi_handler(unsigned int type, const char *name)
Don Zickusc9126b22011-09-30 15:06:20 -0400191{
192 struct nmi_desc *desc = nmi_to_desc(type);
193 struct nmiaction *n;
194 unsigned long flags;
195
Scott Woodc455fd92017-07-24 16:32:42 -0500196 raw_spin_lock_irqsave(&desc->lock, flags);
Don Zickusc9126b22011-09-30 15:06:20 -0400197
198 list_for_each_entry_rcu(n, &desc->head, list) {
199 /*
200 * the name passed in to describe the nmi handler
201 * is used as the lookup key
202 */
203 if (!strcmp(n->name, name)) {
204 WARN(in_nmi(),
205 "Trying to free NMI (%s) from NMI context!\n", n->name);
206 list_del_rcu(&n->list);
207 break;
208 }
209 }
210
Scott Woodc455fd92017-07-24 16:32:42 -0500211 raw_spin_unlock_irqrestore(&desc->lock, flags);
Don Zickusc9126b22011-09-30 15:06:20 -0400212 synchronize_rcu();
Don Zickusc9126b22011-09-30 15:06:20 -0400213}
Don Zickusc9126b22011-09-30 15:06:20 -0400214EXPORT_SYMBOL_GPL(unregister_nmi_handler);
215
Masami Hiramatsu93266382014-04-17 17:18:14 +0900216static void
Don Zickus1d489222011-09-30 15:06:19 -0400217pci_serr_error(unsigned char reason, struct pt_regs *regs)
218{
Don Zickus553222f2012-03-29 16:11:16 -0400219 /* check to see if anyone registered against these types of errors */
Andy Lutomirskibf9f2ee2015-07-20 11:52:23 -0700220 if (nmi_handle(NMI_SERR, regs))
Don Zickus553222f2012-03-29 16:11:16 -0400221 return;
222
Don Zickus1d489222011-09-30 15:06:19 -0400223 pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
224 reason, smp_processor_id());
225
Don Zickus1d489222011-09-30 15:06:19 -0400226 if (panic_on_unrecovered_nmi)
Hidehiro Kawai58c56612015-12-14 11:19:10 +0100227 nmi_panic(regs, "NMI: Not continuing");
Don Zickus1d489222011-09-30 15:06:19 -0400228
229 pr_emerg("Dazed and confused, but trying to continue\n");
230
231 /* Clear and disable the PCI SERR error line. */
232 reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_SERR;
233 outb(reason, NMI_REASON_PORT);
234}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900235NOKPROBE_SYMBOL(pci_serr_error);
Don Zickus1d489222011-09-30 15:06:19 -0400236
Masami Hiramatsu93266382014-04-17 17:18:14 +0900237static void
Don Zickus1d489222011-09-30 15:06:19 -0400238io_check_error(unsigned char reason, struct pt_regs *regs)
239{
240 unsigned long i;
241
Don Zickus553222f2012-03-29 16:11:16 -0400242 /* check to see if anyone registered against these types of errors */
Andy Lutomirskibf9f2ee2015-07-20 11:52:23 -0700243 if (nmi_handle(NMI_IO_CHECK, regs))
Don Zickus553222f2012-03-29 16:11:16 -0400244 return;
245
Don Zickus1d489222011-09-30 15:06:19 -0400246 pr_emerg(
247 "NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
248 reason, smp_processor_id());
Jan Beulich57da8b92012-05-09 08:47:37 +0100249 show_regs(regs);
Don Zickus1d489222011-09-30 15:06:19 -0400250
Hidehiro Kawai1717f202015-12-14 11:19:09 +0100251 if (panic_on_io_nmi) {
Hidehiro Kawai58c56612015-12-14 11:19:10 +0100252 nmi_panic(regs, "NMI IOCK error: Not continuing");
Hidehiro Kawai1717f202015-12-14 11:19:09 +0100253
254 /*
255 * If we end up here, it means we have received an NMI while
256 * processing panic(). Simply return without delaying and
257 * re-enabling NMIs.
258 */
259 return;
260 }
Don Zickus1d489222011-09-30 15:06:19 -0400261
262 /* Re-enable the IOCK line, wait for a few seconds */
263 reason = (reason & NMI_REASON_CLEAR_MASK) | NMI_REASON_CLEAR_IOCHK;
264 outb(reason, NMI_REASON_PORT);
265
266 i = 20000;
267 while (--i) {
268 touch_nmi_watchdog();
269 udelay(100);
270 }
271
272 reason &= ~NMI_REASON_CLEAR_IOCHK;
273 outb(reason, NMI_REASON_PORT);
274}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900275NOKPROBE_SYMBOL(io_check_error);
Don Zickus1d489222011-09-30 15:06:19 -0400276
Masami Hiramatsu93266382014-04-17 17:18:14 +0900277static void
Don Zickus1d489222011-09-30 15:06:19 -0400278unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
279{
Don Zickus9c48f1c2011-09-30 15:06:21 -0400280 int handled;
281
Don Zickusb227e232011-09-30 15:06:22 -0400282 /*
283 * Use 'false' as back-to-back NMIs are dealt with one level up.
284 * Of course this makes having multiple 'unknown' handlers useless
285 * as only the first one is ever run (unless it can actually determine
286 * if it caused the NMI)
287 */
Andy Lutomirskibf9f2ee2015-07-20 11:52:23 -0700288 handled = nmi_handle(NMI_UNKNOWN, regs);
Don Zickusefc3aac2011-09-30 15:06:23 -0400289 if (handled) {
290 __this_cpu_add(nmi_stats.unknown, handled);
Don Zickus1d489222011-09-30 15:06:19 -0400291 return;
Don Zickusefc3aac2011-09-30 15:06:23 -0400292 }
293
294 __this_cpu_add(nmi_stats.unknown, 1);
295
Don Zickus1d489222011-09-30 15:06:19 -0400296 pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
297 reason, smp_processor_id());
298
299 pr_emerg("Do you have a strange power saving mode enabled?\n");
300 if (unknown_nmi_panic || panic_on_unrecovered_nmi)
Hidehiro Kawai58c56612015-12-14 11:19:10 +0100301 nmi_panic(regs, "NMI: Not continuing");
Don Zickus1d489222011-09-30 15:06:19 -0400302
303 pr_emerg("Dazed and confused, but trying to continue\n");
304}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900305NOKPROBE_SYMBOL(unknown_nmi_error);
Don Zickus1d489222011-09-30 15:06:19 -0400306
Don Zickusb227e232011-09-30 15:06:22 -0400307static DEFINE_PER_CPU(bool, swallow_nmi);
308static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
309
Masami Hiramatsu93266382014-04-17 17:18:14 +0900310static void default_do_nmi(struct pt_regs *regs)
Don Zickus1d489222011-09-30 15:06:19 -0400311{
312 unsigned char reason = 0;
Don Zickus9c48f1c2011-09-30 15:06:21 -0400313 int handled;
Don Zickusb227e232011-09-30 15:06:22 -0400314 bool b2b = false;
Don Zickus1d489222011-09-30 15:06:19 -0400315
316 /*
317 * CPU-specific NMI must be processed before non-CPU-specific
318 * NMI, otherwise we may lose it, because the CPU-specific
319 * NMI can not be detected/processed on other CPUs.
320 */
Don Zickusb227e232011-09-30 15:06:22 -0400321
322 /*
323 * Back-to-back NMIs are interesting because they can either
324 * be two NMI or more than two NMIs (any thing over two is dropped
325 * due to NMI being edge-triggered). If this is the second half
326 * of the back-to-back NMI, assume we dropped things and process
327 * more handlers. Otherwise reset the 'swallow' NMI behaviour
328 */
329 if (regs->ip == __this_cpu_read(last_nmi_rip))
330 b2b = true;
331 else
332 __this_cpu_write(swallow_nmi, false);
333
334 __this_cpu_write(last_nmi_rip, regs->ip);
335
Andy Lutomirskibf9f2ee2015-07-20 11:52:23 -0700336 handled = nmi_handle(NMI_LOCAL, regs);
Don Zickusefc3aac2011-09-30 15:06:23 -0400337 __this_cpu_add(nmi_stats.normal, handled);
Don Zickusb227e232011-09-30 15:06:22 -0400338 if (handled) {
339 /*
340 * There are cases when a NMI handler handles multiple
341 * events in the current NMI. One of these events may
342 * be queued for in the next NMI. Because the event is
343 * already handled, the next NMI will result in an unknown
344 * NMI. Instead lets flag this for a potential NMI to
345 * swallow.
346 */
347 if (handled > 1)
348 __this_cpu_write(swallow_nmi, true);
Don Zickus1d489222011-09-30 15:06:19 -0400349 return;
Don Zickusb227e232011-09-30 15:06:22 -0400350 }
Don Zickus1d489222011-09-30 15:06:19 -0400351
Hidehiro Kawaib279d672015-12-14 11:19:13 +0100352 /*
353 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
354 *
355 * Another CPU may be processing panic routines while holding
356 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
357 * and if so, call its callback directly. If there is no CPU preparing
358 * crash dump, we simply loop here.
359 */
360 while (!raw_spin_trylock(&nmi_reason_lock)) {
361 run_crash_ipi_callback(regs);
362 cpu_relax();
363 }
364
Jacob Pan064a59b2011-11-10 13:43:05 +0000365 reason = x86_platform.get_nmi_reason();
Don Zickus1d489222011-09-30 15:06:19 -0400366
367 if (reason & NMI_REASON_MASK) {
368 if (reason & NMI_REASON_SERR)
369 pci_serr_error(reason, regs);
370 else if (reason & NMI_REASON_IOCHK)
371 io_check_error(reason, regs);
372#ifdef CONFIG_X86_32
373 /*
374 * Reassert NMI in case it became active
375 * meanwhile as it's edge-triggered:
376 */
377 reassert_nmi();
378#endif
Don Zickusefc3aac2011-09-30 15:06:23 -0400379 __this_cpu_add(nmi_stats.external, 1);
Don Zickus1d489222011-09-30 15:06:19 -0400380 raw_spin_unlock(&nmi_reason_lock);
381 return;
382 }
383 raw_spin_unlock(&nmi_reason_lock);
384
Don Zickusb227e232011-09-30 15:06:22 -0400385 /*
386 * Only one NMI can be latched at a time. To handle
387 * this we may process multiple nmi handlers at once to
388 * cover the case where an NMI is dropped. The downside
389 * to this approach is we may process an NMI prematurely,
390 * while its real NMI is sitting latched. This will cause
391 * an unknown NMI on the next run of the NMI processing.
392 *
393 * We tried to flag that condition above, by setting the
394 * swallow_nmi flag when we process more than one event.
395 * This condition is also only present on the second half
396 * of a back-to-back NMI, so we flag that condition too.
397 *
398 * If both are true, we assume we already processed this
399 * NMI previously and we swallow it. Otherwise we reset
400 * the logic.
401 *
402 * There are scenarios where we may accidentally swallow
403 * a 'real' unknown NMI. For example, while processing
404 * a perf NMI another perf NMI comes in along with a
405 * 'real' unknown NMI. These two NMIs get combined into
406 * one (as descibed above). When the next NMI gets
407 * processed, it will be flagged by perf as handled, but
408 * noone will know that there was a 'real' unknown NMI sent
409 * also. As a result it gets swallowed. Or if the first
410 * perf NMI returns two events handled then the second
411 * NMI will get eaten by the logic below, again losing a
412 * 'real' unknown NMI. But this is the best we can do
413 * for now.
414 */
415 if (b2b && __this_cpu_read(swallow_nmi))
Don Zickusefc3aac2011-09-30 15:06:23 -0400416 __this_cpu_add(nmi_stats.swallow, 1);
Don Zickusb227e232011-09-30 15:06:22 -0400417 else
418 unknown_nmi_error(reason, regs);
Don Zickus1d489222011-09-30 15:06:19 -0400419}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900420NOKPROBE_SYMBOL(default_do_nmi);
Don Zickus1d489222011-09-30 15:06:19 -0400421
Steven Rostedtccd49c22011-12-13 16:44:16 -0500422/*
Andy Lutomirski0b229302015-07-15 10:29:36 -0700423 * 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.
Andy Lutomirski9d050412015-07-15 10:29:33 -0700425 *
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:
Steven Rostedtccd49c22011-12-13 16:44:16 -0500432 *
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 *
Andy Lutomirski9d050412015-07-15 10:29:33 -0700445 * 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.
Steven Rostedtc7d65a72012-06-07 11:03:00 -0400453 *
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.
Steven Rostedt70fb74a2012-06-07 11:54:37 -0400458 *
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.
Steven Rostedtccd49c22011-12-13 16:44:16 -0500466 */
467enum nmi_states {
Steven Rostedtc7d65a72012-06-07 11:03:00 -0400468 NMI_NOT_RUNNING = 0,
Steven Rostedtccd49c22011-12-13 16:44:16 -0500469 NMI_EXECUTING,
470 NMI_LATCHED,
471};
472static DEFINE_PER_CPU(enum nmi_states, nmi_state);
Steven Rostedt70fb74a2012-06-07 11:54:37 -0400473static DEFINE_PER_CPU(unsigned long, nmi_cr2);
Steven Rostedt228bdaa2011-12-09 03:02:19 -0500474
Andy Lutomirski9d050412015-07-15 10:29:33 -0700475#ifdef CONFIG_X86_64
Steven Rostedtccd49c22011-12-13 16:44:16 -0500476/*
Andy Lutomirski9d050412015-07-15 10:29:33 -0700477 * In x86_64, we need to handle breakpoint -> NMI -> breakpoint. Without
478 * some care, the inner breakpoint will clobber the outer breakpoint's
479 * stack.
Steven Rostedtccd49c22011-12-13 16:44:16 -0500480 *
Andy Lutomirski9d050412015-07-15 10:29:33 -0700481 * If a breakpoint is being processed, and the debug stack is being
482 * used, if an NMI comes in and also hits a breakpoint, the stack
483 * pointer will be set to the same fixed address as the breakpoint that
484 * was interrupted, causing that stack to be corrupted. To handle this
485 * case, check if the stack that was interrupted is the debug stack, and
486 * if so, change the IDT so that new breakpoints will use the current
487 * stack and not switch to the fixed address. On return of the NMI,
488 * switch back to the original IDT.
Steven Rostedtccd49c22011-12-13 16:44:16 -0500489 */
490static DEFINE_PER_CPU(int, update_debug_stack);
Andy Lutomirski9d050412015-07-15 10:29:33 -0700491#endif
Steven Rostedtccd49c22011-12-13 16:44:16 -0500492
Andy Lutomirski9d050412015-07-15 10:29:33 -0700493dotraplinkage notrace void
494do_nmi(struct pt_regs *regs, long error_code)
Steven Rostedtccd49c22011-12-13 16:44:16 -0500495{
Andy Lutomirski9d050412015-07-15 10:29:33 -0700496 if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
497 this_cpu_write(nmi_state, NMI_LATCHED);
498 return;
499 }
500 this_cpu_write(nmi_state, NMI_EXECUTING);
501 this_cpu_write(nmi_cr2, read_cr2());
502nmi_restart:
503
504#ifdef CONFIG_X86_64
Steven Rostedt228bdaa2011-12-09 03:02:19 -0500505 /*
506 * If we interrupted a breakpoint, it is possible that
507 * the nmi handler will have breakpoints too. We need to
508 * change the IDT such that breakpoints that happen here
509 * continue to use the NMI stack.
510 */
511 if (unlikely(is_debug_stack(regs->sp))) {
512 debug_stack_set_zero();
Steven Rostedtc0525a62012-05-30 11:43:19 -0400513 this_cpu_write(update_debug_stack, 1);
Steven Rostedt228bdaa2011-12-09 03:02:19 -0500514 }
Steven Rostedtccd49c22011-12-13 16:44:16 -0500515#endif
516
Don Zickus1d489222011-09-30 15:06:19 -0400517 nmi_enter();
518
519 inc_irq_stat(__nmi_count);
520
521 if (!ignore_nmis)
522 default_do_nmi(regs);
523
524 nmi_exit();
Steven Rostedt228bdaa2011-12-09 03:02:19 -0500525
Andy Lutomirski9d050412015-07-15 10:29:33 -0700526#ifdef CONFIG_X86_64
527 if (unlikely(this_cpu_read(update_debug_stack))) {
528 debug_stack_reset();
529 this_cpu_write(update_debug_stack, 0);
530 }
531#endif
532
533 if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
534 write_cr2(this_cpu_read(nmi_cr2));
535 if (this_cpu_dec_return(nmi_state))
536 goto nmi_restart;
Thomas Gleixner04dcbdb2019-02-18 23:42:51 +0100537
538 if (user_mode(regs))
539 mds_user_clear_cpu_buffers();
Don Zickus1d489222011-09-30 15:06:19 -0400540}
Masami Hiramatsu93266382014-04-17 17:18:14 +0900541NOKPROBE_SYMBOL(do_nmi);
Don Zickus1d489222011-09-30 15:06:19 -0400542
543void stop_nmi(void)
544{
545 ignore_nmis++;
546}
547
548void restart_nmi(void)
549{
550 ignore_nmis--;
551}
Don Zickusb227e232011-09-30 15:06:22 -0400552
553/* reset the back-to-back NMI logic */
554void local_touch_nmi(void)
555{
556 __this_cpu_write(last_nmi_rip, 0);
557}
Jacob Pan29c6fb72013-01-04 11:12:44 +0000558EXPORT_SYMBOL_GPL(local_touch_nmi);