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.
6 * Pentium III FXSR, SSE support
7 * Gareth Hughes <gareth@valinux.com>, May 2000
11 * Handle hardware traps and faults.
13 #include <linux/spinlock.h>
14 #include <linux/kprobes.h>
15 #include <linux/kdebug.h>
16 #include <linux/nmi.h>
17 #include <linux/delay.h>
18 #include <linux/hardirq.h>
19 #include <linux/slab.h>
20 #include <linux/export.h>
22 #include <linux/mca.h>
24 #if defined(CONFIG_EDAC)
25 #include <linux/edac.h>
28 #include <linux/atomic.h>
29 #include <asm/traps.h>
30 #include <asm/mach_traps.h>
32 #include <asm/x86_init.h>
36 struct list_head head
;
39 static struct nmi_desc nmi_desc
[NMI_MAX
] =
42 .lock
= __SPIN_LOCK_UNLOCKED(&nmi_desc
[0].lock
),
43 .head
= LIST_HEAD_INIT(nmi_desc
[0].head
),
46 .lock
= __SPIN_LOCK_UNLOCKED(&nmi_desc
[1].lock
),
47 .head
= LIST_HEAD_INIT(nmi_desc
[1].head
),
50 .lock
= __SPIN_LOCK_UNLOCKED(&nmi_desc
[2].lock
),
51 .head
= LIST_HEAD_INIT(nmi_desc
[2].head
),
54 .lock
= __SPIN_LOCK_UNLOCKED(&nmi_desc
[3].lock
),
55 .head
= LIST_HEAD_INIT(nmi_desc
[3].head
),
63 unsigned int external
;
67 static DEFINE_PER_CPU(struct nmi_stats
, nmi_stats
);
69 static int ignore_nmis
;
71 int unknown_nmi_panic
;
73 * Prevent NMI reason port (0x61) being accessed simultaneously, can
74 * only be used in NMI handler.
76 static DEFINE_RAW_SPINLOCK(nmi_reason_lock
);
78 static int __init
setup_unknown_nmi_panic(char *str
)
80 unknown_nmi_panic
= 1;
83 __setup("unknown_nmi_panic", setup_unknown_nmi_panic
);
85 #define nmi_to_desc(type) (&nmi_desc[type])
87 static int __kprobes
nmi_handle(unsigned int type
, struct pt_regs
*regs
, bool b2b
)
89 struct nmi_desc
*desc
= nmi_to_desc(type
);
96 * NMIs are edge-triggered, which means if you have enough
97 * of them concurrently, you can lose some because only one
98 * can be latched at any given time. Walk the whole list
99 * to handle those situations.
101 list_for_each_entry_rcu(a
, &desc
->head
, list
)
102 handled
+= a
->handler(type
, regs
);
106 /* return total number of NMI events handled */
110 int __register_nmi_handler(unsigned int type
, struct nmiaction
*action
)
112 struct nmi_desc
*desc
= nmi_to_desc(type
);
115 if (!action
->handler
)
118 spin_lock_irqsave(&desc
->lock
, flags
);
121 * most handlers of type NMI_UNKNOWN never return because
122 * they just assume the NMI is theirs. Just a sanity check
123 * to manage expectations
125 WARN_ON_ONCE(type
== NMI_UNKNOWN
&& !list_empty(&desc
->head
));
126 WARN_ON_ONCE(type
== NMI_SERR
&& !list_empty(&desc
->head
));
127 WARN_ON_ONCE(type
== NMI_IO_CHECK
&& !list_empty(&desc
->head
));
130 * some handlers need to be executed first otherwise a fake
131 * event confuses some handlers (kdump uses this flag)
133 if (action
->flags
& NMI_FLAG_FIRST
)
134 list_add_rcu(&action
->list
, &desc
->head
);
136 list_add_tail_rcu(&action
->list
, &desc
->head
);
138 spin_unlock_irqrestore(&desc
->lock
, flags
);
141 EXPORT_SYMBOL(__register_nmi_handler
);
143 void unregister_nmi_handler(unsigned int type
, const char *name
)
145 struct nmi_desc
*desc
= nmi_to_desc(type
);
149 spin_lock_irqsave(&desc
->lock
, flags
);
151 list_for_each_entry_rcu(n
, &desc
->head
, list
) {
153 * the name passed in to describe the nmi handler
154 * is used as the lookup key
156 if (!strcmp(n
->name
, name
)) {
158 "Trying to free NMI (%s) from NMI context!\n", n
->name
);
159 list_del_rcu(&n
->list
);
164 spin_unlock_irqrestore(&desc
->lock
, flags
);
167 EXPORT_SYMBOL_GPL(unregister_nmi_handler
);
169 static __kprobes
void
170 pci_serr_error(unsigned char reason
, struct pt_regs
*regs
)
172 /* check to see if anyone registered against these types of errors */
173 if (nmi_handle(NMI_SERR
, regs
, false))
176 pr_emerg("NMI: PCI system error (SERR) for reason %02x on CPU %d.\n",
177 reason
, smp_processor_id());
180 * On some machines, PCI SERR line is used to report memory
181 * errors. EDAC makes use of it.
183 #if defined(CONFIG_EDAC)
184 if (edac_handler_set()) {
185 edac_atomic_assert_error();
190 if (panic_on_unrecovered_nmi
)
191 panic("NMI: Not continuing");
193 pr_emerg("Dazed and confused, but trying to continue\n");
195 /* Clear and disable the PCI SERR error line. */
196 reason
= (reason
& NMI_REASON_CLEAR_MASK
) | NMI_REASON_CLEAR_SERR
;
197 outb(reason
, NMI_REASON_PORT
);
200 static __kprobes
void
201 io_check_error(unsigned char reason
, struct pt_regs
*regs
)
205 /* check to see if anyone registered against these types of errors */
206 if (nmi_handle(NMI_IO_CHECK
, regs
, false))
210 "NMI: IOCK error (debug interrupt?) for reason %02x on CPU %d.\n",
211 reason
, smp_processor_id());
212 show_registers(regs
);
215 panic("NMI IOCK error: Not continuing");
217 /* Re-enable the IOCK line, wait for a few seconds */
218 reason
= (reason
& NMI_REASON_CLEAR_MASK
) | NMI_REASON_CLEAR_IOCHK
;
219 outb(reason
, NMI_REASON_PORT
);
223 touch_nmi_watchdog();
227 reason
&= ~NMI_REASON_CLEAR_IOCHK
;
228 outb(reason
, NMI_REASON_PORT
);
231 static __kprobes
void
232 unknown_nmi_error(unsigned char reason
, struct pt_regs
*regs
)
237 * Use 'false' as back-to-back NMIs are dealt with one level up.
238 * Of course this makes having multiple 'unknown' handlers useless
239 * as only the first one is ever run (unless it can actually determine
240 * if it caused the NMI)
242 handled
= nmi_handle(NMI_UNKNOWN
, regs
, false);
244 __this_cpu_add(nmi_stats
.unknown
, handled
);
248 __this_cpu_add(nmi_stats
.unknown
, 1);
252 * Might actually be able to figure out what the guilty party
260 pr_emerg("Uhhuh. NMI received for unknown reason %02x on CPU %d.\n",
261 reason
, smp_processor_id());
263 pr_emerg("Do you have a strange power saving mode enabled?\n");
264 if (unknown_nmi_panic
|| panic_on_unrecovered_nmi
)
265 panic("NMI: Not continuing");
267 pr_emerg("Dazed and confused, but trying to continue\n");
270 static DEFINE_PER_CPU(bool, swallow_nmi
);
271 static DEFINE_PER_CPU(unsigned long, last_nmi_rip
);
273 static __kprobes
void default_do_nmi(struct pt_regs
*regs
)
275 unsigned char reason
= 0;
280 * CPU-specific NMI must be processed before non-CPU-specific
281 * NMI, otherwise we may lose it, because the CPU-specific
282 * NMI can not be detected/processed on other CPUs.
286 * Back-to-back NMIs are interesting because they can either
287 * be two NMI or more than two NMIs (any thing over two is dropped
288 * due to NMI being edge-triggered). If this is the second half
289 * of the back-to-back NMI, assume we dropped things and process
290 * more handlers. Otherwise reset the 'swallow' NMI behaviour
292 if (regs
->ip
== __this_cpu_read(last_nmi_rip
))
295 __this_cpu_write(swallow_nmi
, false);
297 __this_cpu_write(last_nmi_rip
, regs
->ip
);
299 handled
= nmi_handle(NMI_LOCAL
, regs
, b2b
);
300 __this_cpu_add(nmi_stats
.normal
, handled
);
303 * There are cases when a NMI handler handles multiple
304 * events in the current NMI. One of these events may
305 * be queued for in the next NMI. Because the event is
306 * already handled, the next NMI will result in an unknown
307 * NMI. Instead lets flag this for a potential NMI to
311 __this_cpu_write(swallow_nmi
, true);
315 /* Non-CPU-specific NMI: NMI sources can be processed on any CPU */
316 raw_spin_lock(&nmi_reason_lock
);
317 reason
= x86_platform
.get_nmi_reason();
319 if (reason
& NMI_REASON_MASK
) {
320 if (reason
& NMI_REASON_SERR
)
321 pci_serr_error(reason
, regs
);
322 else if (reason
& NMI_REASON_IOCHK
)
323 io_check_error(reason
, regs
);
326 * Reassert NMI in case it became active
327 * meanwhile as it's edge-triggered:
331 __this_cpu_add(nmi_stats
.external
, 1);
332 raw_spin_unlock(&nmi_reason_lock
);
335 raw_spin_unlock(&nmi_reason_lock
);
338 * Only one NMI can be latched at a time. To handle
339 * this we may process multiple nmi handlers at once to
340 * cover the case where an NMI is dropped. The downside
341 * to this approach is we may process an NMI prematurely,
342 * while its real NMI is sitting latched. This will cause
343 * an unknown NMI on the next run of the NMI processing.
345 * We tried to flag that condition above, by setting the
346 * swallow_nmi flag when we process more than one event.
347 * This condition is also only present on the second half
348 * of a back-to-back NMI, so we flag that condition too.
350 * If both are true, we assume we already processed this
351 * NMI previously and we swallow it. Otherwise we reset
354 * There are scenarios where we may accidentally swallow
355 * a 'real' unknown NMI. For example, while processing
356 * a perf NMI another perf NMI comes in along with a
357 * 'real' unknown NMI. These two NMIs get combined into
358 * one (as descibed above). When the next NMI gets
359 * processed, it will be flagged by perf as handled, but
360 * noone will know that there was a 'real' unknown NMI sent
361 * also. As a result it gets swallowed. Or if the first
362 * perf NMI returns two events handled then the second
363 * NMI will get eaten by the logic below, again losing a
364 * 'real' unknown NMI. But this is the best we can do
367 if (b2b
&& __this_cpu_read(swallow_nmi
))
368 __this_cpu_add(nmi_stats
.swallow
, 1);
370 unknown_nmi_error(reason
, regs
);
374 * NMIs can hit breakpoints which will cause it to lose its
375 * NMI context with the CPU when the breakpoint does an iret.
379 * For i386, NMIs use the same stack as the kernel, and we can
380 * add a workaround to the iret problem in C. Simply have 3 states
387 * When no NMI is in progress, it is in the "not running" state.
388 * When an NMI comes in, it goes into the "executing" state.
389 * Normally, if another NMI is triggered, it does not interrupt
390 * the running NMI and the HW will simply latch it so that when
391 * the first NMI finishes, it will restart the second NMI.
392 * (Note, the latch is binary, thus multiple NMIs triggering,
393 * when one is running, are ignored. Only one NMI is restarted.)
395 * If an NMI hits a breakpoint that executes an iret, another
396 * NMI can preempt it. We do not want to allow this new NMI
397 * to run, but we want to execute it when the first one finishes.
398 * We set the state to "latched", and the first NMI will perform
399 * an cmpxchg on the state, and if it doesn't successfully
400 * reset the state to "not running" it will restart the next
408 static DEFINE_PER_CPU(enum nmi_states
, nmi_state
);
410 #define nmi_nesting_preprocess(regs) \
412 if (__get_cpu_var(nmi_state) != NMI_NOT_RUNNING) { \
413 __get_cpu_var(nmi_state) = NMI_LATCHED; \
417 __get_cpu_var(nmi_state) = NMI_EXECUTING; \
420 #define nmi_nesting_postprocess() \
422 if (cmpxchg(&__get_cpu_var(nmi_state), \
423 NMI_EXECUTING, NMI_NOT_RUNNING) != NMI_EXECUTING) \
428 * In x86_64 things are a bit more difficult. This has the same problem
429 * where an NMI hitting a breakpoint that calls iret will remove the
430 * NMI context, allowing a nested NMI to enter. What makes this more
431 * difficult is that both NMIs and breakpoints have their own stack.
432 * When a new NMI or breakpoint is executed, the stack is set to a fixed
433 * point. If an NMI is nested, it will have its stack set at that same
434 * fixed address that the first NMI had, and will start corrupting the
435 * stack. This is handled in entry_64.S, but the same problem exists with
436 * the breakpoint stack.
438 * If a breakpoint is being processed, and the debug stack is being used,
439 * if an NMI comes in and also hits a breakpoint, the stack pointer
440 * will be set to the same fixed address as the breakpoint that was
441 * interrupted, causing that stack to be corrupted. To handle this case,
442 * check if the stack that was interrupted is the debug stack, and if
443 * so, change the IDT so that new breakpoints will use the current stack
444 * and not switch to the fixed address. On return of the NMI, switch back
445 * to the original IDT.
447 static DEFINE_PER_CPU(int, update_debug_stack
);
449 static inline void nmi_nesting_preprocess(struct pt_regs
*regs
)
452 * If we interrupted a breakpoint, it is possible that
453 * the nmi handler will have breakpoints too. We need to
454 * change the IDT such that breakpoints that happen here
455 * continue to use the NMI stack.
457 if (unlikely(is_debug_stack(regs
->sp
))) {
458 debug_stack_set_zero();
459 __get_cpu_var(update_debug_stack
) = 1;
463 static inline void nmi_nesting_postprocess(void)
465 if (unlikely(__get_cpu_var(update_debug_stack
)))
470 dotraplinkage notrace __kprobes
void
471 do_nmi(struct pt_regs
*regs
, long error_code
)
473 nmi_nesting_preprocess(regs
);
477 inc_irq_stat(__nmi_count
);
480 default_do_nmi(regs
);
484 /* On i386, may loop back to preprocess */
485 nmi_nesting_postprocess();
493 void restart_nmi(void)
498 /* reset the back-to-back NMI logic */
499 void local_touch_nmi(void)
501 __this_cpu_write(last_nmi_rip
, 0);