1 // SPDX-License-Identifier: GPL-2.0-only
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.
7 * Pentium III FXSR, SSE support
12 * Handle hardware traps and faults.
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>
28 #include <asm/cpu_entry_area.h>
29 #include <asm/traps.h>
30 #include <asm/mach_traps.h>
32 #include <asm/x86_init.h>
33 #include <asm/reboot.h>
34 #include <asm/cache.h>
35 #include <asm/nospec-branch.h>
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/nmi.h>
43 struct list_head head;
46 static struct nmi_desc nmi_desc[NMI_MAX] =
49 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[0].lock),
50 .head = LIST_HEAD_INIT(nmi_desc[0].head),
53 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[1].lock),
54 .head = LIST_HEAD_INIT(nmi_desc[1].head),
57 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[2].lock),
58 .head = LIST_HEAD_INIT(nmi_desc[2].head),
61 .lock = __RAW_SPIN_LOCK_UNLOCKED(&nmi_desc[3].lock),
62 .head = LIST_HEAD_INIT(nmi_desc[3].head),
70 unsigned int external;
74 static DEFINE_PER_CPU(struct nmi_stats, nmi_stats);
76 static int ignore_nmis __read_mostly;
78 int unknown_nmi_panic;
80 * Prevent NMI reason port (0x61) being accessed simultaneously, can
81 * only be used in NMI handler.
83 static DEFINE_RAW_SPINLOCK(nmi_reason_lock);
85 static int __init setup_unknown_nmi_panic(char *str)
87 unknown_nmi_panic = 1;
90 __setup("unknown_nmi_panic", setup_unknown_nmi_panic);
92 #define nmi_to_desc(type) (&nmi_desc[type])
94 static u64 nmi_longest_ns = 1 * NSEC_PER_MSEC;
96 static int __init nmi_warning_debugfs(void)
98 debugfs_create_u64("nmi_longest_ns", 0644,
99 arch_debugfs_dir, &nmi_longest_ns);
102 fs_initcall(nmi_warning_debugfs);
104 static void nmi_check_duration(struct nmiaction *action, u64 duration)
106 int remainder_ns, decimal_msecs;
108 if (duration < nmi_longest_ns || duration < action->max_duration)
111 action->max_duration = duration;
113 remainder_ns = do_div(duration, (1000 * 1000));
114 decimal_msecs = remainder_ns / 1000;
116 printk_ratelimited(KERN_INFO
117 "INFO: NMI handler (%ps) took too long to run: %lld.%03d msecs\n",
118 action->handler, duration, decimal_msecs);
121 static int nmi_handle(unsigned int type, struct pt_regs *regs)
123 struct nmi_desc *desc = nmi_to_desc(type);
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.
135 list_for_each_entry_rcu(a, &desc->head, list) {
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);
145 nmi_check_duration(a, delta);
150 /* return total number of NMI events handled */
153 NOKPROBE_SYMBOL(nmi_handle);
155 int __register_nmi_handler(unsigned int type, struct nmiaction *action)
157 struct nmi_desc *desc = nmi_to_desc(type);
160 if (WARN_ON_ONCE(!action->handler || !list_empty(&action->list)))
163 raw_spin_lock_irqsave(&desc->lock, flags);
166 * Indicate if there are multiple registrations on the
167 * internal NMI handler call chains (SERR and IO_CHECK).
169 WARN_ON_ONCE(type == NMI_SERR && !list_empty(&desc->head));
170 WARN_ON_ONCE(type == NMI_IO_CHECK && !list_empty(&desc->head));
173 * some handlers need to be executed first otherwise a fake
174 * event confuses some handlers (kdump uses this flag)
176 if (action->flags & NMI_FLAG_FIRST)
177 list_add_rcu(&action->list, &desc->head);
179 list_add_tail_rcu(&action->list, &desc->head);
181 raw_spin_unlock_irqrestore(&desc->lock, flags);
184 EXPORT_SYMBOL(__register_nmi_handler);
186 void unregister_nmi_handler(unsigned int type, const char *name)
188 struct nmi_desc *desc = nmi_to_desc(type);
189 struct nmiaction *n, *found = NULL;
192 raw_spin_lock_irqsave(&desc->lock, flags);
194 list_for_each_entry_rcu(n, &desc->head, list) {
196 * the name passed in to describe the nmi handler
197 * is used as the lookup key
199 if (!strcmp(n->name, name)) {
201 "Trying to free NMI (%s) from NMI context!\n", n->name);
202 list_del_rcu(&n->list);
208 raw_spin_unlock_irqrestore(&desc->lock, flags);
211 INIT_LIST_HEAD(&found->list);
214 EXPORT_SYMBOL_GPL(unregister_nmi_handler);
217 pci_serr_error(unsigned char reason, struct pt_regs *regs)
219 /* check to see if anyone registered against these types of errors */
220 if (nmi_handle(NMI_SERR, regs))
223 pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
224 reason, smp_processor_id());
226 if (panic_on_unrecovered_nmi)
227 nmi_panic(regs, "NMI: Not continuing");
229 pr_emerg("Dazed and confused, but trying to continue\n");
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);
235 NOKPROBE_SYMBOL(pci_serr_error);
238 io_check_error(unsigned char reason, struct pt_regs *regs)
242 /* check to see if anyone registered against these types of errors */
243 if (nmi_handle(NMI_IO_CHECK, regs))
247 "NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
248 reason, smp_processor_id());
251 if (panic_on_io_nmi) {
252 nmi_panic(regs, "NMI IOCK error: Not continuing");
255 * If we end up here, it means we have received an NMI while
256 * processing panic(). Simply return without delaying and
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);
268 touch_nmi_watchdog();
272 reason &= ~NMI_REASON_CLEAR_IOCHK;
273 outb(reason, NMI_REASON_PORT);
275 NOKPROBE_SYMBOL(io_check_error);
278 unknown_nmi_error(unsigned char reason, struct pt_regs *regs)
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)
288 handled = nmi_handle(NMI_UNKNOWN, regs);
290 __this_cpu_add(nmi_stats.unknown, handled);
294 __this_cpu_add(nmi_stats.unknown, 1);
296 pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
297 reason, smp_processor_id());
299 if (unknown_nmi_panic || panic_on_unrecovered_nmi)
300 nmi_panic(regs, "NMI: Not continuing");
302 pr_emerg("Dazed and confused, but trying to continue\n");
304 NOKPROBE_SYMBOL(unknown_nmi_error);
306 static DEFINE_PER_CPU(bool, swallow_nmi);
307 static DEFINE_PER_CPU(unsigned long, last_nmi_rip);
309 static noinstr void default_do_nmi(struct pt_regs *regs)
311 unsigned char reason = 0;
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.
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
328 if (regs->ip == __this_cpu_read(last_nmi_rip))
331 __this_cpu_write(swallow_nmi, false);
333 __this_cpu_write(last_nmi_rip, regs->ip);
335 instrumentation_begin();
337 handled = nmi_handle(NMI_LOCAL, regs);
338 __this_cpu_add(nmi_stats.normal, handled);
341 * There are cases when a NMI handler handles multiple
342 * events in the current NMI. One of these events may
343 * be queued for in the next NMI. Because the event is
344 * already handled, the next NMI will result in an unknown
345 * NMI. Instead lets flag this for a potential NMI to
349 __this_cpu_write(swallow_nmi, true);
354 * Non-CPU-specific NMI: NMI sources can be processed on any CPU.
356 * Another CPU may be processing panic routines while holding
357 * nmi_reason_lock. Check if the CPU issued the IPI for crash dumping,
358 * and if so, call its callback directly. If there is no CPU preparing
359 * crash dump, we simply loop here.
361 while (!raw_spin_trylock(&nmi_reason_lock)) {
362 run_crash_ipi_callback(regs);
366 reason = x86_platform.get_nmi_reason();
368 if (reason & NMI_REASON_MASK) {
369 if (reason & NMI_REASON_SERR)
370 pci_serr_error(reason, regs);
371 else if (reason & NMI_REASON_IOCHK)
372 io_check_error(reason, regs);
375 * Reassert NMI in case it became active
376 * meanwhile as it's edge-triggered:
380 __this_cpu_add(nmi_stats.external, 1);
381 raw_spin_unlock(&nmi_reason_lock);
384 raw_spin_unlock(&nmi_reason_lock);
387 * Only one NMI can be latched at a time. To handle
388 * this we may process multiple nmi handlers at once to
389 * cover the case where an NMI is dropped. The downside
390 * to this approach is we may process an NMI prematurely,
391 * while its real NMI is sitting latched. This will cause
392 * an unknown NMI on the next run of the NMI processing.
394 * We tried to flag that condition above, by setting the
395 * swallow_nmi flag when we process more than one event.
396 * This condition is also only present on the second half
397 * of a back-to-back NMI, so we flag that condition too.
399 * If both are true, we assume we already processed this
400 * NMI previously and we swallow it. Otherwise we reset
403 * There are scenarios where we may accidentally swallow
404 * a 'real' unknown NMI. For example, while processing
405 * a perf NMI another perf NMI comes in along with a
406 * 'real' unknown NMI. These two NMIs get combined into
407 * one (as described above). When the next NMI gets
408 * processed, it will be flagged by perf as handled, but
409 * no one will know that there was a 'real' unknown NMI sent
410 * also. As a result it gets swallowed. Or if the first
411 * perf NMI returns two events handled then the second
412 * NMI will get eaten by the logic below, again losing a
413 * 'real' unknown NMI. But this is the best we can do
416 if (b2b && __this_cpu_read(swallow_nmi))
417 __this_cpu_add(nmi_stats.swallow, 1);
419 unknown_nmi_error(reason, regs);
422 instrumentation_end();
426 * NMIs can page fault or hit breakpoints which will cause it to lose
427 * its NMI context with the CPU when the breakpoint or page fault does an IRET.
429 * As a result, NMIs can nest if NMIs get unmasked due an IRET during
430 * NMI processing. On x86_64, the asm glue protects us from nested NMIs
431 * if the outer NMI came from kernel mode, but we can still nest if the
432 * outer NMI came from user mode.
434 * To handle these nested NMIs, we have three states:
440 * When no NMI is in progress, it is in the "not running" state.
441 * When an NMI comes in, it goes into the "executing" state.
442 * Normally, if another NMI is triggered, it does not interrupt
443 * the running NMI and the HW will simply latch it so that when
444 * the first NMI finishes, it will restart the second NMI.
445 * (Note, the latch is binary, thus multiple NMIs triggering,
446 * when one is running, are ignored. Only one NMI is restarted.)
448 * If an NMI executes an iret, another NMI can preempt it. We do not
449 * want to allow this new NMI to run, but we want to execute it when the
450 * first one finishes. We set the state to "latched", and the exit of
451 * the first NMI will perform a dec_return, if the result is zero
452 * (NOT_RUNNING), then it will simply exit the NMI handler. If not, the
453 * dec_return would have set the state to NMI_EXECUTING (what we want it
454 * to be when we are running). In this case, we simply jump back to
455 * rerun the NMI handler again, and restart the 'latched' NMI.
457 * No trap (breakpoint or page fault) should be hit before nmi_restart,
458 * thus there is no race between the first check of state for NOT_RUNNING
459 * and setting it to NMI_EXECUTING. The HW will prevent nested NMIs
462 * In case the NMI takes a page fault, we need to save off the CR2
463 * because the NMI could have preempted another page fault and corrupt
464 * the CR2 that is about to be read. As nested NMIs must be restarted
465 * and they can not take breakpoints or page faults, the update of the
466 * CR2 must be done before converting the nmi state back to NOT_RUNNING.
467 * Otherwise, there would be a race of another nested NMI coming in
468 * after setting state to NOT_RUNNING but before updating the nmi_cr2.
475 static DEFINE_PER_CPU(enum nmi_states, nmi_state);
476 static DEFINE_PER_CPU(unsigned long, nmi_cr2);
477 static DEFINE_PER_CPU(unsigned long, nmi_dr7);
479 DEFINE_IDTENTRY_RAW(exc_nmi)
481 irqentry_state_t irq_state;
484 * Re-enable NMIs right here when running as an SEV-ES guest. This might
485 * cause nested NMIs, but those can be handled safely.
487 sev_es_nmi_complete();
489 if (IS_ENABLED(CONFIG_SMP) && arch_cpu_is_offline(smp_processor_id()))
492 if (this_cpu_read(nmi_state) != NMI_NOT_RUNNING) {
493 this_cpu_write(nmi_state, NMI_LATCHED);
496 this_cpu_write(nmi_state, NMI_EXECUTING);
497 this_cpu_write(nmi_cr2, read_cr2());
501 * Needs to happen before DR7 is accessed, because the hypervisor can
502 * intercept DR7 reads/writes, turning those into #VC exceptions.
504 sev_es_ist_enter(regs);
506 this_cpu_write(nmi_dr7, local_db_save());
508 irq_state = irqentry_nmi_enter(regs);
510 inc_irq_stat(__nmi_count);
513 default_do_nmi(regs);
515 irqentry_nmi_exit(regs, irq_state);
517 local_db_restore(this_cpu_read(nmi_dr7));
521 if (unlikely(this_cpu_read(nmi_cr2) != read_cr2()))
522 write_cr2(this_cpu_read(nmi_cr2));
523 if (this_cpu_dec_return(nmi_state))
527 mds_user_clear_cpu_buffers();
530 #if defined(CONFIG_X86_64) && IS_ENABLED(CONFIG_KVM_INTEL)
531 DEFINE_IDTENTRY_RAW(exc_nmi_noist)
536 #if IS_MODULE(CONFIG_KVM_INTEL)
537 EXPORT_SYMBOL_GPL(asm_exc_nmi_noist);
545 void restart_nmi(void)
550 /* reset the back-to-back NMI logic */
551 void local_touch_nmi(void)
553 __this_cpu_write(last_nmi_rip, 0);
555 EXPORT_SYMBOL_GPL(local_touch_nmi);