2 * Machine check handler.
4 * K8 parts Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 * Rest from unknown author(s).
6 * 2004 Andi Kleen. Rewrote most of it.
7 * Copyright 2008 Intel Corporation
11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13 #include <linux/thread_info.h>
14 #include <linux/capability.h>
15 #include <linux/miscdevice.h>
16 #include <linux/ratelimit.h>
17 #include <linux/kallsyms.h>
18 #include <linux/rcupdate.h>
19 #include <linux/kobject.h>
20 #include <linux/uaccess.h>
21 #include <linux/kdebug.h>
22 #include <linux/kernel.h>
23 #include <linux/percpu.h>
24 #include <linux/string.h>
25 #include <linux/device.h>
26 #include <linux/syscore_ops.h>
27 #include <linux/delay.h>
28 #include <linux/ctype.h>
29 #include <linux/sched.h>
30 #include <linux/sysfs.h>
31 #include <linux/types.h>
32 #include <linux/slab.h>
33 #include <linux/init.h>
34 #include <linux/kmod.h>
35 #include <linux/poll.h>
36 #include <linux/nmi.h>
37 #include <linux/cpu.h>
38 #include <linux/ras.h>
39 #include <linux/smp.h>
42 #include <linux/debugfs.h>
43 #include <linux/irq_work.h>
44 #include <linux/export.h>
45 #include <linux/jump_label.h>
47 #include <asm/intel-family.h>
48 #include <asm/processor.h>
49 #include <asm/traps.h>
50 #include <asm/tlbflush.h>
53 #include <asm/reboot.h>
55 #include "mce-internal.h"
57 static DEFINE_MUTEX(mce_log_mutex
);
59 #define CREATE_TRACE_POINTS
60 #include <trace/events/mce.h>
62 #define SPINUNIT 100 /* 100ns */
64 DEFINE_PER_CPU(unsigned, mce_exception_count
);
66 struct mce_bank
*mce_banks __read_mostly
;
67 struct mce_vendor_flags mce_flags __read_mostly
;
69 struct mca_config mca_cfg __read_mostly
= {
73 * 0: always panic on uncorrected errors, log corrected errors
74 * 1: panic or SIGBUS on uncorrected errors, log corrected errors
75 * 2: SIGBUS or log uncorrected errors (if possible), log corr. errors
76 * 3: never panic or SIGBUS, log all errors (for testing only)
82 static DEFINE_PER_CPU(struct mce
, mces_seen
);
83 static unsigned long mce_need_notify
;
84 static int cpu_missing
;
87 * MCA banks polled by the period polling timer for corrected events.
88 * With Intel CMCI, this only has MCA banks which do not support CMCI (if any).
90 DEFINE_PER_CPU(mce_banks_t
, mce_poll_banks
) = {
91 [0 ... BITS_TO_LONGS(MAX_NR_BANKS
)-1] = ~0UL
95 * MCA banks controlled through firmware first for corrected errors.
96 * This is a global list of banks for which we won't enable CMCI and we
97 * won't poll. Firmware controls these banks and is responsible for
98 * reporting corrected errors through GHES. Uncorrected/recoverable
99 * errors are still notified through a machine check.
101 mce_banks_t mce_banks_ce_disabled
;
103 static struct work_struct mce_work
;
104 static struct irq_work mce_irq_work
;
106 static void (*quirk_no_way_out
)(int bank
, struct mce
*m
, struct pt_regs
*regs
);
109 * CPU/chipset specific EDAC code can register a notifier call here to print
110 * MCE errors in a human-readable form.
112 BLOCKING_NOTIFIER_HEAD(x86_mce_decoder_chain
);
114 /* Do initial initialization of a struct mce */
115 void mce_setup(struct mce
*m
)
117 memset(m
, 0, sizeof(struct mce
));
118 m
->cpu
= m
->extcpu
= smp_processor_id();
119 /* We hope get_seconds stays lockless */
120 m
->time
= get_seconds();
121 m
->cpuvendor
= boot_cpu_data
.x86_vendor
;
122 m
->cpuid
= cpuid_eax(1);
123 m
->socketid
= cpu_data(m
->extcpu
).phys_proc_id
;
124 m
->apicid
= cpu_data(m
->extcpu
).initial_apicid
;
125 rdmsrl(MSR_IA32_MCG_CAP
, m
->mcgcap
);
127 if (this_cpu_has(X86_FEATURE_INTEL_PPIN
))
128 rdmsrl(MSR_PPIN
, m
->ppin
);
131 DEFINE_PER_CPU(struct mce
, injectm
);
132 EXPORT_PER_CPU_SYMBOL_GPL(injectm
);
134 void mce_log(struct mce
*m
)
136 if (!mce_gen_pool_add(m
))
137 irq_work_queue(&mce_irq_work
);
140 void mce_inject_log(struct mce
*m
)
142 mutex_lock(&mce_log_mutex
);
144 mutex_unlock(&mce_log_mutex
);
146 EXPORT_SYMBOL_GPL(mce_inject_log
);
148 static struct notifier_block mce_srao_nb
;
151 * We run the default notifier if we have only the SRAO, the first and the
152 * default notifier registered. I.e., the mandatory NUM_DEFAULT_NOTIFIERS
153 * notifiers registered on the chain.
155 #define NUM_DEFAULT_NOTIFIERS 3
156 static atomic_t num_notifiers
;
158 void mce_register_decode_chain(struct notifier_block
*nb
)
160 if (WARN_ON(nb
->priority
> MCE_PRIO_MCELOG
&& nb
->priority
< MCE_PRIO_EDAC
))
163 atomic_inc(&num_notifiers
);
165 blocking_notifier_chain_register(&x86_mce_decoder_chain
, nb
);
167 EXPORT_SYMBOL_GPL(mce_register_decode_chain
);
169 void mce_unregister_decode_chain(struct notifier_block
*nb
)
171 atomic_dec(&num_notifiers
);
173 blocking_notifier_chain_unregister(&x86_mce_decoder_chain
, nb
);
175 EXPORT_SYMBOL_GPL(mce_unregister_decode_chain
);
177 static inline u32
ctl_reg(int bank
)
179 return MSR_IA32_MCx_CTL(bank
);
182 static inline u32
status_reg(int bank
)
184 return MSR_IA32_MCx_STATUS(bank
);
187 static inline u32
addr_reg(int bank
)
189 return MSR_IA32_MCx_ADDR(bank
);
192 static inline u32
misc_reg(int bank
)
194 return MSR_IA32_MCx_MISC(bank
);
197 static inline u32
smca_ctl_reg(int bank
)
199 return MSR_AMD64_SMCA_MCx_CTL(bank
);
202 static inline u32
smca_status_reg(int bank
)
204 return MSR_AMD64_SMCA_MCx_STATUS(bank
);
207 static inline u32
smca_addr_reg(int bank
)
209 return MSR_AMD64_SMCA_MCx_ADDR(bank
);
212 static inline u32
smca_misc_reg(int bank
)
214 return MSR_AMD64_SMCA_MCx_MISC(bank
);
217 struct mca_msr_regs msr_ops
= {
219 .status
= status_reg
,
224 static void __print_mce(struct mce
*m
)
226 pr_emerg(HW_ERR
"CPU %d: Machine Check%s: %Lx Bank %d: %016Lx\n",
228 (m
->mcgstatus
& MCG_STATUS_MCIP
? " Exception" : ""),
229 m
->mcgstatus
, m
->bank
, m
->status
);
232 pr_emerg(HW_ERR
"RIP%s %02x:<%016Lx> ",
233 !(m
->mcgstatus
& MCG_STATUS_EIPV
) ? " !INEXACT!" : "",
236 if (m
->cs
== __KERNEL_CS
)
237 print_symbol("{%s}", m
->ip
);
241 pr_emerg(HW_ERR
"TSC %llx ", m
->tsc
);
243 pr_cont("ADDR %llx ", m
->addr
);
245 pr_cont("MISC %llx ", m
->misc
);
247 if (mce_flags
.smca
) {
249 pr_cont("SYND %llx ", m
->synd
);
251 pr_cont("IPID %llx ", m
->ipid
);
256 * Note this output is parsed by external tools and old fields
257 * should not be changed.
259 pr_emerg(HW_ERR
"PROCESSOR %u:%x TIME %llu SOCKET %u APIC %x microcode %x\n",
260 m
->cpuvendor
, m
->cpuid
, m
->time
, m
->socketid
, m
->apicid
,
261 cpu_data(m
->extcpu
).microcode
);
264 static void print_mce(struct mce
*m
)
267 pr_emerg_ratelimited(HW_ERR
"Run the above through 'mcelog --ascii'\n");
270 #define PANIC_TIMEOUT 5 /* 5 seconds */
272 static atomic_t mce_panicked
;
274 static int fake_panic
;
275 static atomic_t mce_fake_panicked
;
277 /* Panic in progress. Enable interrupts and wait for final IPI */
278 static void wait_for_panic(void)
280 long timeout
= PANIC_TIMEOUT
*USEC_PER_SEC
;
284 while (timeout
-- > 0)
286 if (panic_timeout
== 0)
287 panic_timeout
= mca_cfg
.panic_timeout
;
288 panic("Panicing machine check CPU died");
291 static void mce_panic(const char *msg
, struct mce
*final
, char *exp
)
294 struct llist_node
*pending
;
295 struct mce_evt_llist
*l
;
299 * Make sure only one CPU runs in machine check panic
301 if (atomic_inc_return(&mce_panicked
) > 1)
308 /* Don't log too much for fake panic */
309 if (atomic_inc_return(&mce_fake_panicked
) > 1)
312 pending
= mce_gen_pool_prepare_records();
313 /* First print corrected ones that are still unlogged */
314 llist_for_each_entry(l
, pending
, llnode
) {
315 struct mce
*m
= &l
->mce
;
316 if (!(m
->status
& MCI_STATUS_UC
)) {
319 apei_err
= apei_write_mce(m
);
322 /* Now print uncorrected but with the final one last */
323 llist_for_each_entry(l
, pending
, llnode
) {
324 struct mce
*m
= &l
->mce
;
325 if (!(m
->status
& MCI_STATUS_UC
))
327 if (!final
|| mce_cmp(m
, final
)) {
330 apei_err
= apei_write_mce(m
);
336 apei_err
= apei_write_mce(final
);
339 pr_emerg(HW_ERR
"Some CPUs didn't answer in synchronization\n");
341 pr_emerg(HW_ERR
"Machine check: %s\n", exp
);
343 if (panic_timeout
== 0)
344 panic_timeout
= mca_cfg
.panic_timeout
;
347 pr_emerg(HW_ERR
"Fake kernel panic: %s\n", msg
);
350 /* Support code for software error injection */
352 static int msr_to_offset(u32 msr
)
354 unsigned bank
= __this_cpu_read(injectm
.bank
);
356 if (msr
== mca_cfg
.rip_msr
)
357 return offsetof(struct mce
, ip
);
358 if (msr
== msr_ops
.status(bank
))
359 return offsetof(struct mce
, status
);
360 if (msr
== msr_ops
.addr(bank
))
361 return offsetof(struct mce
, addr
);
362 if (msr
== msr_ops
.misc(bank
))
363 return offsetof(struct mce
, misc
);
364 if (msr
== MSR_IA32_MCG_STATUS
)
365 return offsetof(struct mce
, mcgstatus
);
369 /* MSR access wrappers used for error injection */
370 static u64
mce_rdmsrl(u32 msr
)
374 if (__this_cpu_read(injectm
.finished
)) {
375 int offset
= msr_to_offset(msr
);
379 return *(u64
*)((char *)this_cpu_ptr(&injectm
) + offset
);
382 if (rdmsrl_safe(msr
, &v
)) {
383 WARN_ONCE(1, "mce: Unable to read MSR 0x%x!\n", msr
);
385 * Return zero in case the access faulted. This should
386 * not happen normally but can happen if the CPU does
387 * something weird, or if the code is buggy.
395 static void mce_wrmsrl(u32 msr
, u64 v
)
397 if (__this_cpu_read(injectm
.finished
)) {
398 int offset
= msr_to_offset(msr
);
401 *(u64
*)((char *)this_cpu_ptr(&injectm
) + offset
) = v
;
408 * Collect all global (w.r.t. this processor) status about this machine
409 * check into our "mce" struct so that we can use it later to assess
410 * the severity of the problem as we read per-bank specific details.
412 static inline void mce_gather_info(struct mce
*m
, struct pt_regs
*regs
)
416 m
->mcgstatus
= mce_rdmsrl(MSR_IA32_MCG_STATUS
);
419 * Get the address of the instruction at the time of
420 * the machine check error.
422 if (m
->mcgstatus
& (MCG_STATUS_RIPV
|MCG_STATUS_EIPV
)) {
427 * When in VM86 mode make the cs look like ring 3
428 * always. This is a lie, but it's better than passing
429 * the additional vm86 bit around everywhere.
431 if (v8086_mode(regs
))
434 /* Use accurate RIP reporting if available. */
436 m
->ip
= mce_rdmsrl(mca_cfg
.rip_msr
);
440 int mce_available(struct cpuinfo_x86
*c
)
442 if (mca_cfg
.disabled
)
444 return cpu_has(c
, X86_FEATURE_MCE
) && cpu_has(c
, X86_FEATURE_MCA
);
447 static void mce_schedule_work(void)
449 if (!mce_gen_pool_empty())
450 schedule_work(&mce_work
);
453 static void mce_irq_work_cb(struct irq_work
*entry
)
458 static void mce_report_event(struct pt_regs
*regs
)
460 if (regs
->flags
& (X86_VM_MASK
|X86_EFLAGS_IF
)) {
463 * Triggering the work queue here is just an insurance
464 * policy in case the syscall exit notify handler
465 * doesn't run soon enough or ends up running on the
466 * wrong CPU (can happen when audit sleeps)
472 irq_work_queue(&mce_irq_work
);
476 * Check if the address reported by the CPU is in a format we can parse.
477 * It would be possible to add code for most other cases, but all would
478 * be somewhat complicated (e.g. segment offset would require an instruction
479 * parser). So only support physical addresses up to page granuality for now.
481 static int mce_usable_address(struct mce
*m
)
483 if (!(m
->status
& MCI_STATUS_ADDRV
))
486 /* Checks after this one are Intel-specific: */
487 if (boot_cpu_data
.x86_vendor
!= X86_VENDOR_INTEL
)
490 if (!(m
->status
& MCI_STATUS_MISCV
))
493 if (MCI_MISC_ADDR_LSB(m
->misc
) > PAGE_SHIFT
)
496 if (MCI_MISC_ADDR_MODE(m
->misc
) != MCI_MISC_ADDR_PHYS
)
502 static bool memory_error(struct mce
*m
)
504 struct cpuinfo_x86
*c
= &boot_cpu_data
;
506 if (c
->x86_vendor
== X86_VENDOR_AMD
) {
507 /* ErrCodeExt[20:16] */
508 u8 xec
= (m
->status
>> 16) & 0x1f;
510 return (xec
== 0x0 || xec
== 0x8);
511 } else if (c
->x86_vendor
== X86_VENDOR_INTEL
) {
513 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
515 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
516 * indicating a memory error. Bit 8 is used for indicating a
517 * cache hierarchy error. The combination of bit 2 and bit 3
518 * is used for indicating a `generic' cache hierarchy error
519 * But we can't just blindly check the above bits, because if
520 * bit 11 is set, then it is a bus/interconnect error - and
521 * either way the above bits just gives more detail on what
522 * bus/interconnect error happened. Note that bit 12 can be
523 * ignored, as it's the "filter" bit.
525 return (m
->status
& 0xef80) == BIT(7) ||
526 (m
->status
& 0xef00) == BIT(8) ||
527 (m
->status
& 0xeffc) == 0xc;
533 static bool cec_add_mce(struct mce
*m
)
538 /* We eat only correctable DRAM errors with usable addresses. */
539 if (memory_error(m
) &&
540 !(m
->status
& MCI_STATUS_UC
) &&
541 mce_usable_address(m
))
542 if (!cec_add_elem(m
->addr
>> PAGE_SHIFT
))
548 static int mce_first_notifier(struct notifier_block
*nb
, unsigned long val
,
551 struct mce
*m
= (struct mce
*)data
;
559 /* Emit the trace record: */
562 set_bit(0, &mce_need_notify
);
569 static struct notifier_block first_nb
= {
570 .notifier_call
= mce_first_notifier
,
571 .priority
= MCE_PRIO_FIRST
,
574 static int srao_decode_notifier(struct notifier_block
*nb
, unsigned long val
,
577 struct mce
*mce
= (struct mce
*)data
;
583 if (mce_usable_address(mce
) && (mce
->severity
== MCE_AO_SEVERITY
)) {
584 pfn
= mce
->addr
>> PAGE_SHIFT
;
585 memory_failure(pfn
, MCE_VECTOR
, 0);
590 static struct notifier_block mce_srao_nb
= {
591 .notifier_call
= srao_decode_notifier
,
592 .priority
= MCE_PRIO_SRAO
,
595 static int mce_default_notifier(struct notifier_block
*nb
, unsigned long val
,
598 struct mce
*m
= (struct mce
*)data
;
603 if (atomic_read(&num_notifiers
) > NUM_DEFAULT_NOTIFIERS
)
611 static struct notifier_block mce_default_nb
= {
612 .notifier_call
= mce_default_notifier
,
613 /* lowest prio, we want it to run last. */
614 .priority
= MCE_PRIO_LOWEST
,
618 * Read ADDR and MISC registers.
620 static void mce_read_aux(struct mce
*m
, int i
)
622 if (m
->status
& MCI_STATUS_MISCV
)
623 m
->misc
= mce_rdmsrl(msr_ops
.misc(i
));
625 if (m
->status
& MCI_STATUS_ADDRV
) {
626 m
->addr
= mce_rdmsrl(msr_ops
.addr(i
));
629 * Mask the reported address by the reported granularity.
631 if (mca_cfg
.ser
&& (m
->status
& MCI_STATUS_MISCV
)) {
632 u8 shift
= MCI_MISC_ADDR_LSB(m
->misc
);
638 * Extract [55:<lsb>] where lsb is the least significant
639 * *valid* bit of the address bits.
641 if (mce_flags
.smca
) {
642 u8 lsb
= (m
->addr
>> 56) & 0x3f;
644 m
->addr
&= GENMASK_ULL(55, lsb
);
648 if (mce_flags
.smca
) {
649 m
->ipid
= mce_rdmsrl(MSR_AMD64_SMCA_MCx_IPID(i
));
651 if (m
->status
& MCI_STATUS_SYNDV
)
652 m
->synd
= mce_rdmsrl(MSR_AMD64_SMCA_MCx_SYND(i
));
656 DEFINE_PER_CPU(unsigned, mce_poll_count
);
659 * Poll for corrected events or events that happened before reset.
660 * Those are just logged through /dev/mcelog.
662 * This is executed in standard interrupt context.
664 * Note: spec recommends to panic for fatal unsignalled
665 * errors here. However this would be quite problematic --
666 * we would need to reimplement the Monarch handling and
667 * it would mess up the exclusion between exception handler
668 * and poll hander -- * so we skip this for now.
669 * These cases should not happen anyways, or only when the CPU
670 * is already totally * confused. In this case it's likely it will
671 * not fully execute the machine check handler either.
673 bool machine_check_poll(enum mcp_flags flags
, mce_banks_t
*b
)
675 bool error_seen
= false;
680 this_cpu_inc(mce_poll_count
);
682 mce_gather_info(&m
, NULL
);
684 if (flags
& MCP_TIMESTAMP
)
687 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
688 if (!mce_banks
[i
].ctl
|| !test_bit(i
, *b
))
696 m
.status
= mce_rdmsrl(msr_ops
.status(i
));
697 if (!(m
.status
& MCI_STATUS_VAL
))
701 * Uncorrected or signalled events are handled by the exception
702 * handler when it is enabled, so don't process those here.
704 * TBD do the same check for MCI_STATUS_EN here?
706 if (!(flags
& MCP_UC
) &&
707 (m
.status
& (mca_cfg
.ser
? MCI_STATUS_S
: MCI_STATUS_UC
)))
714 severity
= mce_severity(&m
, mca_cfg
.tolerant
, NULL
, false);
716 if (severity
== MCE_DEFERRED_SEVERITY
&& memory_error(&m
))
717 if (m
.status
& MCI_STATUS_ADDRV
)
718 m
.severity
= severity
;
721 * Don't get the IP here because it's unlikely to
722 * have anything to do with the actual error location.
724 if (!(flags
& MCP_DONTLOG
) && !mca_cfg
.dont_log_ce
)
726 else if (mce_usable_address(&m
)) {
728 * Although we skipped logging this, we still want
729 * to take action. Add to the pool so the registered
730 * notifiers will see it.
732 if (!mce_gen_pool_add(&m
))
737 * Clear state for this bank.
739 mce_wrmsrl(msr_ops
.status(i
), 0);
743 * Don't clear MCG_STATUS here because it's only defined for
751 EXPORT_SYMBOL_GPL(machine_check_poll
);
754 * Do a quick check if any of the events requires a panic.
755 * This decides if we keep the events around or clear them.
757 static int mce_no_way_out(struct mce
*m
, char **msg
, unsigned long *validp
,
758 struct pt_regs
*regs
)
763 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
764 m
->status
= mce_rdmsrl(msr_ops
.status(i
));
765 if (m
->status
& MCI_STATUS_VAL
) {
766 __set_bit(i
, validp
);
767 if (quirk_no_way_out
)
768 quirk_no_way_out(i
, m
, regs
);
771 if (mce_severity(m
, mca_cfg
.tolerant
, &tmp
, true) >= MCE_PANIC_SEVERITY
) {
780 * Variable to establish order between CPUs while scanning.
781 * Each CPU spins initially until executing is equal its number.
783 static atomic_t mce_executing
;
786 * Defines order of CPUs on entry. First CPU becomes Monarch.
788 static atomic_t mce_callin
;
791 * Check if a timeout waiting for other CPUs happened.
793 static int mce_timed_out(u64
*t
, const char *msg
)
796 * The others already did panic for some reason.
797 * Bail out like in a timeout.
798 * rmb() to tell the compiler that system_state
799 * might have been modified by someone else.
802 if (atomic_read(&mce_panicked
))
804 if (!mca_cfg
.monarch_timeout
)
806 if ((s64
)*t
< SPINUNIT
) {
807 if (mca_cfg
.tolerant
<= 1)
808 mce_panic(msg
, NULL
, NULL
);
814 touch_nmi_watchdog();
819 * The Monarch's reign. The Monarch is the CPU who entered
820 * the machine check handler first. It waits for the others to
821 * raise the exception too and then grades them. When any
822 * error is fatal panic. Only then let the others continue.
824 * The other CPUs entering the MCE handler will be controlled by the
825 * Monarch. They are called Subjects.
827 * This way we prevent any potential data corruption in a unrecoverable case
828 * and also makes sure always all CPU's errors are examined.
830 * Also this detects the case of a machine check event coming from outer
831 * space (not detected by any CPUs) In this case some external agent wants
832 * us to shut down, so panic too.
834 * The other CPUs might still decide to panic if the handler happens
835 * in a unrecoverable place, but in this case the system is in a semi-stable
836 * state and won't corrupt anything by itself. It's ok to let the others
837 * continue for a bit first.
839 * All the spin loops have timeouts; when a timeout happens a CPU
840 * typically elects itself to be Monarch.
842 static void mce_reign(void)
845 struct mce
*m
= NULL
;
846 int global_worst
= 0;
851 * This CPU is the Monarch and the other CPUs have run
852 * through their handlers.
853 * Grade the severity of the errors of all the CPUs.
855 for_each_possible_cpu(cpu
) {
856 int severity
= mce_severity(&per_cpu(mces_seen
, cpu
),
859 if (severity
> global_worst
) {
861 global_worst
= severity
;
862 m
= &per_cpu(mces_seen
, cpu
);
867 * Cannot recover? Panic here then.
868 * This dumps all the mces in the log buffer and stops the
871 if (m
&& global_worst
>= MCE_PANIC_SEVERITY
&& mca_cfg
.tolerant
< 3)
872 mce_panic("Fatal machine check", m
, msg
);
875 * For UC somewhere we let the CPU who detects it handle it.
876 * Also must let continue the others, otherwise the handling
877 * CPU could deadlock on a lock.
881 * No machine check event found. Must be some external
882 * source or one CPU is hung. Panic.
884 if (global_worst
<= MCE_KEEP_SEVERITY
&& mca_cfg
.tolerant
< 3)
885 mce_panic("Fatal machine check from unknown source", NULL
, NULL
);
888 * Now clear all the mces_seen so that they don't reappear on
891 for_each_possible_cpu(cpu
)
892 memset(&per_cpu(mces_seen
, cpu
), 0, sizeof(struct mce
));
895 static atomic_t global_nwo
;
898 * Start of Monarch synchronization. This waits until all CPUs have
899 * entered the exception handler and then determines if any of them
900 * saw a fatal event that requires panic. Then it executes them
901 * in the entry order.
902 * TBD double check parallel CPU hotunplug
904 static int mce_start(int *no_way_out
)
907 int cpus
= num_online_cpus();
908 u64 timeout
= (u64
)mca_cfg
.monarch_timeout
* NSEC_PER_USEC
;
913 atomic_add(*no_way_out
, &global_nwo
);
915 * Rely on the implied barrier below, such that global_nwo
916 * is updated before mce_callin.
918 order
= atomic_inc_return(&mce_callin
);
923 while (atomic_read(&mce_callin
) != cpus
) {
924 if (mce_timed_out(&timeout
,
925 "Timeout: Not all CPUs entered broadcast exception handler")) {
926 atomic_set(&global_nwo
, 0);
933 * mce_callin should be read before global_nwo
939 * Monarch: Starts executing now, the others wait.
941 atomic_set(&mce_executing
, 1);
944 * Subject: Now start the scanning loop one by one in
945 * the original callin order.
946 * This way when there are any shared banks it will be
947 * only seen by one CPU before cleared, avoiding duplicates.
949 while (atomic_read(&mce_executing
) < order
) {
950 if (mce_timed_out(&timeout
,
951 "Timeout: Subject CPUs unable to finish machine check processing")) {
952 atomic_set(&global_nwo
, 0);
960 * Cache the global no_way_out state.
962 *no_way_out
= atomic_read(&global_nwo
);
968 * Synchronize between CPUs after main scanning loop.
969 * This invokes the bulk of the Monarch processing.
971 static int mce_end(int order
)
974 u64 timeout
= (u64
)mca_cfg
.monarch_timeout
* NSEC_PER_USEC
;
982 * Allow others to run.
984 atomic_inc(&mce_executing
);
987 /* CHECKME: Can this race with a parallel hotplug? */
988 int cpus
= num_online_cpus();
991 * Monarch: Wait for everyone to go through their scanning
994 while (atomic_read(&mce_executing
) <= cpus
) {
995 if (mce_timed_out(&timeout
,
996 "Timeout: Monarch CPU unable to finish machine check processing"))
1006 * Subject: Wait for Monarch to finish.
1008 while (atomic_read(&mce_executing
) != 0) {
1009 if (mce_timed_out(&timeout
,
1010 "Timeout: Monarch CPU did not finish machine check processing"))
1016 * Don't reset anything. That's done by the Monarch.
1022 * Reset all global state.
1025 atomic_set(&global_nwo
, 0);
1026 atomic_set(&mce_callin
, 0);
1030 * Let others run again.
1032 atomic_set(&mce_executing
, 0);
1036 static void mce_clear_state(unsigned long *toclear
)
1040 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
1041 if (test_bit(i
, toclear
))
1042 mce_wrmsrl(msr_ops
.status(i
), 0);
1046 static int do_memory_failure(struct mce
*m
)
1048 int flags
= MF_ACTION_REQUIRED
;
1051 pr_err("Uncorrected hardware memory error in user-access at %llx", m
->addr
);
1052 if (!(m
->mcgstatus
& MCG_STATUS_RIPV
))
1053 flags
|= MF_MUST_KILL
;
1054 ret
= memory_failure(m
->addr
>> PAGE_SHIFT
, MCE_VECTOR
, flags
);
1056 pr_err("Memory error not recovered");
1061 * The actual machine check handler. This only handles real
1062 * exceptions when something got corrupted coming in through int 18.
1064 * This is executed in NMI context not subject to normal locking rules. This
1065 * implies that most kernel services cannot be safely used. Don't even
1066 * think about putting a printk in there!
1068 * On Intel systems this is entered on all CPUs in parallel through
1069 * MCE broadcast. However some CPUs might be broken beyond repair,
1070 * so be always careful when synchronizing with others.
1072 void do_machine_check(struct pt_regs
*regs
, long error_code
)
1074 struct mca_config
*cfg
= &mca_cfg
;
1075 struct mce m
, *final
;
1081 * Establish sequential order between the CPUs entering the machine
1086 * If no_way_out gets set, there is no safe way to recover from this
1087 * MCE. If mca_cfg.tolerant is cranked up, we'll try anyway.
1091 * If kill_it gets set, there might be a way to recover from this
1095 DECLARE_BITMAP(toclear
, MAX_NR_BANKS
);
1096 DECLARE_BITMAP(valid_banks
, MAX_NR_BANKS
);
1097 char *msg
= "Unknown";
1100 * MCEs are always local on AMD. Same is determined by MCG_STATUS_LMCES
1104 int cpu
= smp_processor_id();
1107 * Cases where we avoid rendezvous handler timeout:
1108 * 1) If this CPU is offline.
1110 * 2) If crashing_cpu was set, e.g. we're entering kdump and we need to
1111 * skip those CPUs which remain looping in the 1st kernel - see
1112 * crash_nmi_callback().
1114 * Note: there still is a small window between kexec-ing and the new,
1115 * kdump kernel establishing a new #MC handler where a broadcasted MCE
1116 * might not get handled properly.
1118 if (cpu_is_offline(cpu
) ||
1119 (crashing_cpu
!= -1 && crashing_cpu
!= cpu
)) {
1122 mcgstatus
= mce_rdmsrl(MSR_IA32_MCG_STATUS
);
1123 if (mcgstatus
& MCG_STATUS_RIPV
) {
1124 mce_wrmsrl(MSR_IA32_MCG_STATUS
, 0);
1131 this_cpu_inc(mce_exception_count
);
1136 mce_gather_info(&m
, regs
);
1139 final
= this_cpu_ptr(&mces_seen
);
1142 memset(valid_banks
, 0, sizeof(valid_banks
));
1143 no_way_out
= mce_no_way_out(&m
, &msg
, valid_banks
, regs
);
1148 * When no restart IP might need to kill or panic.
1149 * Assume the worst for now, but if we find the
1150 * severity is MCE_AR_SEVERITY we have other options.
1152 if (!(m
.mcgstatus
& MCG_STATUS_RIPV
))
1156 * Check if this MCE is signaled to only this logical processor,
1159 if (m
.cpuvendor
== X86_VENDOR_INTEL
)
1160 lmce
= m
.mcgstatus
& MCG_STATUS_LMCES
;
1163 * Go through all banks in exclusion of the other CPUs. This way we
1164 * don't report duplicated events on shared banks because the first one
1165 * to see it will clear it. If this is a Local MCE, then no need to
1166 * perform rendezvous.
1169 order
= mce_start(&no_way_out
);
1171 for (i
= 0; i
< cfg
->banks
; i
++) {
1172 __clear_bit(i
, toclear
);
1173 if (!test_bit(i
, valid_banks
))
1175 if (!mce_banks
[i
].ctl
)
1182 m
.status
= mce_rdmsrl(msr_ops
.status(i
));
1183 if ((m
.status
& MCI_STATUS_VAL
) == 0)
1187 * Non uncorrected or non signaled errors are handled by
1188 * machine_check_poll. Leave them alone, unless this panics.
1190 if (!(m
.status
& (cfg
->ser
? MCI_STATUS_S
: MCI_STATUS_UC
)) &&
1195 * Set taint even when machine check was not enabled.
1197 add_taint(TAINT_MACHINE_CHECK
, LOCKDEP_NOW_UNRELIABLE
);
1199 severity
= mce_severity(&m
, cfg
->tolerant
, NULL
, true);
1202 * When machine check was for corrected/deferred handler don't
1203 * touch, unless we're panicing.
1205 if ((severity
== MCE_KEEP_SEVERITY
||
1206 severity
== MCE_UCNA_SEVERITY
) && !no_way_out
)
1208 __set_bit(i
, toclear
);
1209 if (severity
== MCE_NO_SEVERITY
) {
1211 * Machine check event was not enabled. Clear, but
1217 mce_read_aux(&m
, i
);
1219 /* assuming valid severity level != 0 */
1220 m
.severity
= severity
;
1224 if (severity
> worst
) {
1230 /* mce_clear_state will clear *final, save locally for use later */
1234 mce_clear_state(toclear
);
1237 * Do most of the synchronization with other CPUs.
1238 * When there's any problem use only local no_way_out state.
1241 if (mce_end(order
) < 0)
1242 no_way_out
= worst
>= MCE_PANIC_SEVERITY
;
1245 * Local MCE skipped calling mce_reign()
1246 * If we found a fatal error, we need to panic here.
1248 if (worst
>= MCE_PANIC_SEVERITY
&& mca_cfg
.tolerant
< 3)
1249 mce_panic("Machine check from unknown source",
1254 * If tolerant is at an insane level we drop requests to kill
1255 * processes and continue even when there is no way out.
1257 if (cfg
->tolerant
== 3)
1259 else if (no_way_out
)
1260 mce_panic("Fatal machine check on current CPU", &m
, msg
);
1263 mce_report_event(regs
);
1264 mce_wrmsrl(MSR_IA32_MCG_STATUS
, 0);
1268 if (worst
!= MCE_AR_SEVERITY
&& !kill_it
)
1271 /* Fault was in user mode and we need to take some action */
1272 if ((m
.cs
& 3) == 3) {
1273 ist_begin_non_atomic(regs
);
1276 if (kill_it
|| do_memory_failure(&m
))
1277 force_sig(SIGBUS
, current
);
1278 local_irq_disable();
1279 ist_end_non_atomic();
1281 if (!fixup_exception(regs
, X86_TRAP_MC
))
1282 mce_panic("Failed kernel mode recovery", &m
, NULL
);
1288 EXPORT_SYMBOL_GPL(do_machine_check
);
1290 #ifndef CONFIG_MEMORY_FAILURE
1291 int memory_failure(unsigned long pfn
, int vector
, int flags
)
1293 /* mce_severity() should not hand us an ACTION_REQUIRED error */
1294 BUG_ON(flags
& MF_ACTION_REQUIRED
);
1295 pr_err("Uncorrected memory error in page 0x%lx ignored\n"
1296 "Rebuild kernel with CONFIG_MEMORY_FAILURE=y for smarter handling\n",
1304 * Periodic polling timer for "silent" machine check errors. If the
1305 * poller finds an MCE, poll 2x faster. When the poller finds no more
1306 * errors, poll 2x slower (up to check_interval seconds).
1308 static unsigned long check_interval
= INITIAL_CHECK_INTERVAL
;
1310 static DEFINE_PER_CPU(unsigned long, mce_next_interval
); /* in jiffies */
1311 static DEFINE_PER_CPU(struct timer_list
, mce_timer
);
1313 static unsigned long mce_adjust_timer_default(unsigned long interval
)
1318 static unsigned long (*mce_adjust_timer
)(unsigned long interval
) = mce_adjust_timer_default
;
1320 static void __start_timer(struct timer_list
*t
, unsigned long interval
)
1322 unsigned long when
= jiffies
+ interval
;
1323 unsigned long flags
;
1325 local_irq_save(flags
);
1327 if (!timer_pending(t
) || time_before(when
, t
->expires
))
1328 mod_timer(t
, round_jiffies(when
));
1330 local_irq_restore(flags
);
1333 static void mce_timer_fn(unsigned long data
)
1335 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
1336 int cpu
= smp_processor_id();
1339 WARN_ON(cpu
!= data
);
1341 iv
= __this_cpu_read(mce_next_interval
);
1343 if (mce_available(this_cpu_ptr(&cpu_info
))) {
1344 machine_check_poll(0, this_cpu_ptr(&mce_poll_banks
));
1346 if (mce_intel_cmci_poll()) {
1347 iv
= mce_adjust_timer(iv
);
1353 * Alert userspace if needed. If we logged an MCE, reduce the polling
1354 * interval, otherwise increase the polling interval.
1356 if (mce_notify_irq())
1357 iv
= max(iv
/ 2, (unsigned long) HZ
/100);
1359 iv
= min(iv
* 2, round_jiffies_relative(check_interval
* HZ
));
1362 __this_cpu_write(mce_next_interval
, iv
);
1363 __start_timer(t
, iv
);
1367 * Ensure that the timer is firing in @interval from now.
1369 void mce_timer_kick(unsigned long interval
)
1371 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
1372 unsigned long iv
= __this_cpu_read(mce_next_interval
);
1374 __start_timer(t
, interval
);
1377 __this_cpu_write(mce_next_interval
, interval
);
1380 /* Must not be called in IRQ context where del_timer_sync() can deadlock */
1381 static void mce_timer_delete_all(void)
1385 for_each_online_cpu(cpu
)
1386 del_timer_sync(&per_cpu(mce_timer
, cpu
));
1390 * Notify the user(s) about new machine check events.
1391 * Can be called from interrupt context, but not from machine check/NMI
1394 int mce_notify_irq(void)
1396 /* Not more than two messages every minute */
1397 static DEFINE_RATELIMIT_STATE(ratelimit
, 60*HZ
, 2);
1399 if (test_and_clear_bit(0, &mce_need_notify
)) {
1402 if (__ratelimit(&ratelimit
))
1403 pr_info(HW_ERR
"Machine check events logged\n");
1409 EXPORT_SYMBOL_GPL(mce_notify_irq
);
1411 static int __mcheck_cpu_mce_banks_init(void)
1414 u8 num_banks
= mca_cfg
.banks
;
1416 mce_banks
= kzalloc(num_banks
* sizeof(struct mce_bank
), GFP_KERNEL
);
1420 for (i
= 0; i
< num_banks
; i
++) {
1421 struct mce_bank
*b
= &mce_banks
[i
];
1430 * Initialize Machine Checks for a CPU.
1432 static int __mcheck_cpu_cap_init(void)
1437 rdmsrl(MSR_IA32_MCG_CAP
, cap
);
1439 b
= cap
& MCG_BANKCNT_MASK
;
1441 pr_info("CPU supports %d MCE banks\n", b
);
1443 if (b
> MAX_NR_BANKS
) {
1444 pr_warn("Using only %u machine check banks out of %u\n",
1449 /* Don't support asymmetric configurations today */
1450 WARN_ON(mca_cfg
.banks
!= 0 && b
!= mca_cfg
.banks
);
1454 int err
= __mcheck_cpu_mce_banks_init();
1460 /* Use accurate RIP reporting if available. */
1461 if ((cap
& MCG_EXT_P
) && MCG_EXT_CNT(cap
) >= 9)
1462 mca_cfg
.rip_msr
= MSR_IA32_MCG_EIP
;
1464 if (cap
& MCG_SER_P
)
1470 static void __mcheck_cpu_init_generic(void)
1472 enum mcp_flags m_fl
= 0;
1473 mce_banks_t all_banks
;
1476 if (!mca_cfg
.bootlog
)
1480 * Log the machine checks left over from the previous reset.
1482 bitmap_fill(all_banks
, MAX_NR_BANKS
);
1483 machine_check_poll(MCP_UC
| m_fl
, &all_banks
);
1485 cr4_set_bits(X86_CR4_MCE
);
1487 rdmsrl(MSR_IA32_MCG_CAP
, cap
);
1488 if (cap
& MCG_CTL_P
)
1489 wrmsr(MSR_IA32_MCG_CTL
, 0xffffffff, 0xffffffff);
1492 static void __mcheck_cpu_init_clear_banks(void)
1496 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
1497 struct mce_bank
*b
= &mce_banks
[i
];
1501 wrmsrl(msr_ops
.ctl(i
), b
->ctl
);
1502 wrmsrl(msr_ops
.status(i
), 0);
1507 * During IFU recovery Sandy Bridge -EP4S processors set the RIPV and
1508 * EIPV bits in MCG_STATUS to zero on the affected logical processor (SDM
1509 * Vol 3B Table 15-20). But this confuses both the code that determines
1510 * whether the machine check occurred in kernel or user mode, and also
1511 * the severity assessment code. Pretend that EIPV was set, and take the
1512 * ip/cs values from the pt_regs that mce_gather_info() ignored earlier.
1514 static void quirk_sandybridge_ifu(int bank
, struct mce
*m
, struct pt_regs
*regs
)
1518 if ((m
->mcgstatus
& (MCG_STATUS_EIPV
|MCG_STATUS_RIPV
)) != 0)
1520 if ((m
->status
& (MCI_STATUS_OVER
|MCI_STATUS_UC
|
1521 MCI_STATUS_EN
|MCI_STATUS_MISCV
|MCI_STATUS_ADDRV
|
1522 MCI_STATUS_PCC
|MCI_STATUS_S
|MCI_STATUS_AR
|
1524 (MCI_STATUS_UC
|MCI_STATUS_EN
|
1525 MCI_STATUS_MISCV
|MCI_STATUS_ADDRV
|MCI_STATUS_S
|
1526 MCI_STATUS_AR
|MCACOD_INSTR
))
1529 m
->mcgstatus
|= MCG_STATUS_EIPV
;
1534 /* Add per CPU specific workarounds here */
1535 static int __mcheck_cpu_apply_quirks(struct cpuinfo_x86
*c
)
1537 struct mca_config
*cfg
= &mca_cfg
;
1539 if (c
->x86_vendor
== X86_VENDOR_UNKNOWN
) {
1540 pr_info("unknown CPU type - not enabling MCE support\n");
1544 /* This should be disabled by the BIOS, but isn't always */
1545 if (c
->x86_vendor
== X86_VENDOR_AMD
) {
1546 if (c
->x86
== 15 && cfg
->banks
> 4) {
1548 * disable GART TBL walk error reporting, which
1549 * trips off incorrectly with the IOMMU & 3ware
1552 clear_bit(10, (unsigned long *)&mce_banks
[4].ctl
);
1554 if (c
->x86
< 17 && cfg
->bootlog
< 0) {
1556 * Lots of broken BIOS around that don't clear them
1557 * by default and leave crap in there. Don't log:
1562 * Various K7s with broken bank 0 around. Always disable
1565 if (c
->x86
== 6 && cfg
->banks
> 0)
1566 mce_banks
[0].ctl
= 0;
1569 * overflow_recov is supported for F15h Models 00h-0fh
1570 * even though we don't have a CPUID bit for it.
1572 if (c
->x86
== 0x15 && c
->x86_model
<= 0xf)
1573 mce_flags
.overflow_recov
= 1;
1576 * Turn off MC4_MISC thresholding banks on those models since
1577 * they're not supported there.
1579 if (c
->x86
== 0x15 &&
1580 (c
->x86_model
>= 0x10 && c
->x86_model
<= 0x1f)) {
1585 0x00000413, /* MC4_MISC0 */
1586 0xc0000408, /* MC4_MISC1 */
1589 rdmsrl(MSR_K7_HWCR
, hwcr
);
1591 /* McStatusWrEn has to be set */
1592 need_toggle
= !(hwcr
& BIT(18));
1595 wrmsrl(MSR_K7_HWCR
, hwcr
| BIT(18));
1597 /* Clear CntP bit safely */
1598 for (i
= 0; i
< ARRAY_SIZE(msrs
); i
++)
1599 msr_clear_bit(msrs
[i
], 62);
1601 /* restore old settings */
1603 wrmsrl(MSR_K7_HWCR
, hwcr
);
1607 if (c
->x86_vendor
== X86_VENDOR_INTEL
) {
1609 * SDM documents that on family 6 bank 0 should not be written
1610 * because it aliases to another special BIOS controlled
1612 * But it's not aliased anymore on model 0x1a+
1613 * Don't ignore bank 0 completely because there could be a
1614 * valid event later, merely don't write CTL0.
1617 if (c
->x86
== 6 && c
->x86_model
< 0x1A && cfg
->banks
> 0)
1618 mce_banks
[0].init
= 0;
1621 * All newer Intel systems support MCE broadcasting. Enable
1622 * synchronization with a one second timeout.
1624 if ((c
->x86
> 6 || (c
->x86
== 6 && c
->x86_model
>= 0xe)) &&
1625 cfg
->monarch_timeout
< 0)
1626 cfg
->monarch_timeout
= USEC_PER_SEC
;
1629 * There are also broken BIOSes on some Pentium M and
1632 if (c
->x86
== 6 && c
->x86_model
<= 13 && cfg
->bootlog
< 0)
1635 if (c
->x86
== 6 && c
->x86_model
== 45)
1636 quirk_no_way_out
= quirk_sandybridge_ifu
;
1638 if (cfg
->monarch_timeout
< 0)
1639 cfg
->monarch_timeout
= 0;
1640 if (cfg
->bootlog
!= 0)
1641 cfg
->panic_timeout
= 30;
1646 static int __mcheck_cpu_ancient_init(struct cpuinfo_x86
*c
)
1651 switch (c
->x86_vendor
) {
1652 case X86_VENDOR_INTEL
:
1653 intel_p5_mcheck_init(c
);
1656 case X86_VENDOR_CENTAUR
:
1657 winchip_mcheck_init(c
);
1668 * Init basic CPU features needed for early decoding of MCEs.
1670 static void __mcheck_cpu_init_early(struct cpuinfo_x86
*c
)
1672 if (c
->x86_vendor
== X86_VENDOR_AMD
) {
1673 mce_flags
.overflow_recov
= !!cpu_has(c
, X86_FEATURE_OVERFLOW_RECOV
);
1674 mce_flags
.succor
= !!cpu_has(c
, X86_FEATURE_SUCCOR
);
1675 mce_flags
.smca
= !!cpu_has(c
, X86_FEATURE_SMCA
);
1677 if (mce_flags
.smca
) {
1678 msr_ops
.ctl
= smca_ctl_reg
;
1679 msr_ops
.status
= smca_status_reg
;
1680 msr_ops
.addr
= smca_addr_reg
;
1681 msr_ops
.misc
= smca_misc_reg
;
1686 static void __mcheck_cpu_init_vendor(struct cpuinfo_x86
*c
)
1688 switch (c
->x86_vendor
) {
1689 case X86_VENDOR_INTEL
:
1690 mce_intel_feature_init(c
);
1691 mce_adjust_timer
= cmci_intel_adjust_timer
;
1694 case X86_VENDOR_AMD
: {
1695 mce_amd_feature_init(c
);
1704 static void __mcheck_cpu_clear_vendor(struct cpuinfo_x86
*c
)
1706 switch (c
->x86_vendor
) {
1707 case X86_VENDOR_INTEL
:
1708 mce_intel_feature_clear(c
);
1715 static void mce_start_timer(struct timer_list
*t
)
1717 unsigned long iv
= check_interval
* HZ
;
1719 if (mca_cfg
.ignore_ce
|| !iv
)
1722 this_cpu_write(mce_next_interval
, iv
);
1723 __start_timer(t
, iv
);
1726 static void __mcheck_cpu_setup_timer(void)
1728 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
1729 unsigned int cpu
= smp_processor_id();
1731 setup_pinned_timer(t
, mce_timer_fn
, cpu
);
1734 static void __mcheck_cpu_init_timer(void)
1736 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
1737 unsigned int cpu
= smp_processor_id();
1739 setup_pinned_timer(t
, mce_timer_fn
, cpu
);
1743 /* Handle unconfigured int18 (should never happen) */
1744 static void unexpected_machine_check(struct pt_regs
*regs
, long error_code
)
1746 pr_err("CPU#%d: Unexpected int18 (Machine Check)\n",
1747 smp_processor_id());
1750 /* Call the installed machine check handler for this CPU setup. */
1751 void (*machine_check_vector
)(struct pt_regs
*, long error_code
) =
1752 unexpected_machine_check
;
1755 * Called for each booted CPU to set up machine checks.
1756 * Must be called with preempt off:
1758 void mcheck_cpu_init(struct cpuinfo_x86
*c
)
1760 if (mca_cfg
.disabled
)
1763 if (__mcheck_cpu_ancient_init(c
))
1766 if (!mce_available(c
))
1769 if (__mcheck_cpu_cap_init() < 0 || __mcheck_cpu_apply_quirks(c
) < 0) {
1770 mca_cfg
.disabled
= true;
1774 if (mce_gen_pool_init()) {
1775 mca_cfg
.disabled
= true;
1776 pr_emerg("Couldn't allocate MCE records pool!\n");
1780 machine_check_vector
= do_machine_check
;
1782 __mcheck_cpu_init_early(c
);
1783 __mcheck_cpu_init_generic();
1784 __mcheck_cpu_init_vendor(c
);
1785 __mcheck_cpu_init_clear_banks();
1786 __mcheck_cpu_setup_timer();
1790 * Called for each booted CPU to clear some machine checks opt-ins
1792 void mcheck_cpu_clear(struct cpuinfo_x86
*c
)
1794 if (mca_cfg
.disabled
)
1797 if (!mce_available(c
))
1801 * Possibly to clear general settings generic to x86
1802 * __mcheck_cpu_clear_generic(c);
1804 __mcheck_cpu_clear_vendor(c
);
1808 static void __mce_disable_bank(void *arg
)
1810 int bank
= *((int *)arg
);
1811 __clear_bit(bank
, this_cpu_ptr(mce_poll_banks
));
1812 cmci_disable_bank(bank
);
1815 void mce_disable_bank(int bank
)
1817 if (bank
>= mca_cfg
.banks
) {
1819 "Ignoring request to disable invalid MCA bank %d.\n",
1823 set_bit(bank
, mce_banks_ce_disabled
);
1824 on_each_cpu(__mce_disable_bank
, &bank
, 1);
1828 * mce=off Disables machine check
1829 * mce=no_cmci Disables CMCI
1830 * mce=no_lmce Disables LMCE
1831 * mce=dont_log_ce Clears corrected events silently, no log created for CEs.
1832 * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared.
1833 * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above)
1834 * monarchtimeout is how long to wait for other CPUs on machine
1835 * check, or 0 to not wait
1836 * mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
1837 * mce=nobootlog Don't log MCEs from before booting.
1838 * mce=bios_cmci_threshold Don't program the CMCI threshold
1839 * mce=recovery force enable memcpy_mcsafe()
1841 static int __init
mcheck_enable(char *str
)
1843 struct mca_config
*cfg
= &mca_cfg
;
1851 if (!strcmp(str
, "off"))
1852 cfg
->disabled
= true;
1853 else if (!strcmp(str
, "no_cmci"))
1854 cfg
->cmci_disabled
= true;
1855 else if (!strcmp(str
, "no_lmce"))
1856 cfg
->lmce_disabled
= true;
1857 else if (!strcmp(str
, "dont_log_ce"))
1858 cfg
->dont_log_ce
= true;
1859 else if (!strcmp(str
, "ignore_ce"))
1860 cfg
->ignore_ce
= true;
1861 else if (!strcmp(str
, "bootlog") || !strcmp(str
, "nobootlog"))
1862 cfg
->bootlog
= (str
[0] == 'b');
1863 else if (!strcmp(str
, "bios_cmci_threshold"))
1864 cfg
->bios_cmci_threshold
= true;
1865 else if (!strcmp(str
, "recovery"))
1866 cfg
->recovery
= true;
1867 else if (isdigit(str
[0])) {
1868 if (get_option(&str
, &cfg
->tolerant
) == 2)
1869 get_option(&str
, &(cfg
->monarch_timeout
));
1871 pr_info("mce argument %s ignored. Please use /sys\n", str
);
1876 __setup("mce", mcheck_enable
);
1878 int __init
mcheck_init(void)
1880 mcheck_intel_therm_init();
1881 mce_register_decode_chain(&first_nb
);
1882 mce_register_decode_chain(&mce_srao_nb
);
1883 mce_register_decode_chain(&mce_default_nb
);
1884 mcheck_vendor_init_severity();
1886 INIT_WORK(&mce_work
, mce_gen_pool_process
);
1887 init_irq_work(&mce_irq_work
, mce_irq_work_cb
);
1893 * mce_syscore: PM support
1897 * Disable machine checks on suspend and shutdown. We can't really handle
1900 static void mce_disable_error_reporting(void)
1904 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
1905 struct mce_bank
*b
= &mce_banks
[i
];
1908 wrmsrl(msr_ops
.ctl(i
), 0);
1913 static void vendor_disable_error_reporting(void)
1916 * Don't clear on Intel CPUs. Some of these MSRs are socket-wide.
1917 * Disabling them for just a single offlined CPU is bad, since it will
1918 * inhibit reporting for all shared resources on the socket like the
1919 * last level cache (LLC), the integrated memory controller (iMC), etc.
1921 if (boot_cpu_data
.x86_vendor
== X86_VENDOR_INTEL
)
1924 mce_disable_error_reporting();
1927 static int mce_syscore_suspend(void)
1929 vendor_disable_error_reporting();
1933 static void mce_syscore_shutdown(void)
1935 vendor_disable_error_reporting();
1939 * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
1940 * Only one CPU is active at this time, the others get re-added later using
1943 static void mce_syscore_resume(void)
1945 __mcheck_cpu_init_generic();
1946 __mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info
));
1947 __mcheck_cpu_init_clear_banks();
1950 static struct syscore_ops mce_syscore_ops
= {
1951 .suspend
= mce_syscore_suspend
,
1952 .shutdown
= mce_syscore_shutdown
,
1953 .resume
= mce_syscore_resume
,
1957 * mce_device: Sysfs support
1960 static void mce_cpu_restart(void *data
)
1962 if (!mce_available(raw_cpu_ptr(&cpu_info
)))
1964 __mcheck_cpu_init_generic();
1965 __mcheck_cpu_init_clear_banks();
1966 __mcheck_cpu_init_timer();
1969 /* Reinit MCEs after user configuration changes */
1970 static void mce_restart(void)
1972 mce_timer_delete_all();
1973 on_each_cpu(mce_cpu_restart
, NULL
, 1);
1976 /* Toggle features for corrected errors */
1977 static void mce_disable_cmci(void *data
)
1979 if (!mce_available(raw_cpu_ptr(&cpu_info
)))
1984 static void mce_enable_ce(void *all
)
1986 if (!mce_available(raw_cpu_ptr(&cpu_info
)))
1991 __mcheck_cpu_init_timer();
1994 static struct bus_type mce_subsys
= {
1995 .name
= "machinecheck",
1996 .dev_name
= "machinecheck",
1999 DEFINE_PER_CPU(struct device
*, mce_device
);
2001 static inline struct mce_bank
*attr_to_bank(struct device_attribute
*attr
)
2003 return container_of(attr
, struct mce_bank
, attr
);
2006 static ssize_t
show_bank(struct device
*s
, struct device_attribute
*attr
,
2009 return sprintf(buf
, "%llx\n", attr_to_bank(attr
)->ctl
);
2012 static ssize_t
set_bank(struct device
*s
, struct device_attribute
*attr
,
2013 const char *buf
, size_t size
)
2017 if (kstrtou64(buf
, 0, &new) < 0)
2020 attr_to_bank(attr
)->ctl
= new;
2026 static ssize_t
set_ignore_ce(struct device
*s
,
2027 struct device_attribute
*attr
,
2028 const char *buf
, size_t size
)
2032 if (kstrtou64(buf
, 0, &new) < 0)
2035 if (mca_cfg
.ignore_ce
^ !!new) {
2037 /* disable ce features */
2038 mce_timer_delete_all();
2039 on_each_cpu(mce_disable_cmci
, NULL
, 1);
2040 mca_cfg
.ignore_ce
= true;
2042 /* enable ce features */
2043 mca_cfg
.ignore_ce
= false;
2044 on_each_cpu(mce_enable_ce
, (void *)1, 1);
2050 static ssize_t
set_cmci_disabled(struct device
*s
,
2051 struct device_attribute
*attr
,
2052 const char *buf
, size_t size
)
2056 if (kstrtou64(buf
, 0, &new) < 0)
2059 if (mca_cfg
.cmci_disabled
^ !!new) {
2062 on_each_cpu(mce_disable_cmci
, NULL
, 1);
2063 mca_cfg
.cmci_disabled
= true;
2066 mca_cfg
.cmci_disabled
= false;
2067 on_each_cpu(mce_enable_ce
, NULL
, 1);
2073 static ssize_t
store_int_with_restart(struct device
*s
,
2074 struct device_attribute
*attr
,
2075 const char *buf
, size_t size
)
2077 ssize_t ret
= device_store_int(s
, attr
, buf
, size
);
2082 static DEVICE_INT_ATTR(tolerant
, 0644, mca_cfg
.tolerant
);
2083 static DEVICE_INT_ATTR(monarch_timeout
, 0644, mca_cfg
.monarch_timeout
);
2084 static DEVICE_BOOL_ATTR(dont_log_ce
, 0644, mca_cfg
.dont_log_ce
);
2086 static struct dev_ext_attribute dev_attr_check_interval
= {
2087 __ATTR(check_interval
, 0644, device_show_int
, store_int_with_restart
),
2091 static struct dev_ext_attribute dev_attr_ignore_ce
= {
2092 __ATTR(ignore_ce
, 0644, device_show_bool
, set_ignore_ce
),
2096 static struct dev_ext_attribute dev_attr_cmci_disabled
= {
2097 __ATTR(cmci_disabled
, 0644, device_show_bool
, set_cmci_disabled
),
2098 &mca_cfg
.cmci_disabled
2101 static struct device_attribute
*mce_device_attrs
[] = {
2102 &dev_attr_tolerant
.attr
,
2103 &dev_attr_check_interval
.attr
,
2104 #ifdef CONFIG_X86_MCELOG_LEGACY
2107 &dev_attr_monarch_timeout
.attr
,
2108 &dev_attr_dont_log_ce
.attr
,
2109 &dev_attr_ignore_ce
.attr
,
2110 &dev_attr_cmci_disabled
.attr
,
2114 static cpumask_var_t mce_device_initialized
;
2116 static void mce_device_release(struct device
*dev
)
2121 /* Per cpu device init. All of the cpus still share the same ctrl bank: */
2122 static int mce_device_create(unsigned int cpu
)
2128 if (!mce_available(&boot_cpu_data
))
2131 dev
= per_cpu(mce_device
, cpu
);
2135 dev
= kzalloc(sizeof *dev
, GFP_KERNEL
);
2139 dev
->bus
= &mce_subsys
;
2140 dev
->release
= &mce_device_release
;
2142 err
= device_register(dev
);
2148 for (i
= 0; mce_device_attrs
[i
]; i
++) {
2149 err
= device_create_file(dev
, mce_device_attrs
[i
]);
2153 for (j
= 0; j
< mca_cfg
.banks
; j
++) {
2154 err
= device_create_file(dev
, &mce_banks
[j
].attr
);
2158 cpumask_set_cpu(cpu
, mce_device_initialized
);
2159 per_cpu(mce_device
, cpu
) = dev
;
2164 device_remove_file(dev
, &mce_banks
[j
].attr
);
2167 device_remove_file(dev
, mce_device_attrs
[i
]);
2169 device_unregister(dev
);
2174 static void mce_device_remove(unsigned int cpu
)
2176 struct device
*dev
= per_cpu(mce_device
, cpu
);
2179 if (!cpumask_test_cpu(cpu
, mce_device_initialized
))
2182 for (i
= 0; mce_device_attrs
[i
]; i
++)
2183 device_remove_file(dev
, mce_device_attrs
[i
]);
2185 for (i
= 0; i
< mca_cfg
.banks
; i
++)
2186 device_remove_file(dev
, &mce_banks
[i
].attr
);
2188 device_unregister(dev
);
2189 cpumask_clear_cpu(cpu
, mce_device_initialized
);
2190 per_cpu(mce_device
, cpu
) = NULL
;
2193 /* Make sure there are no machine checks on offlined CPUs. */
2194 static void mce_disable_cpu(void)
2196 if (!mce_available(raw_cpu_ptr(&cpu_info
)))
2199 if (!cpuhp_tasks_frozen
)
2202 vendor_disable_error_reporting();
2205 static void mce_reenable_cpu(void)
2209 if (!mce_available(raw_cpu_ptr(&cpu_info
)))
2212 if (!cpuhp_tasks_frozen
)
2214 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
2215 struct mce_bank
*b
= &mce_banks
[i
];
2218 wrmsrl(msr_ops
.ctl(i
), b
->ctl
);
2222 static int mce_cpu_dead(unsigned int cpu
)
2224 mce_intel_hcpu_update(cpu
);
2226 /* intentionally ignoring frozen here */
2227 if (!cpuhp_tasks_frozen
)
2232 static int mce_cpu_online(unsigned int cpu
)
2234 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
2237 mce_device_create(cpu
);
2239 ret
= mce_threshold_create_device(cpu
);
2241 mce_device_remove(cpu
);
2249 static int mce_cpu_pre_down(unsigned int cpu
)
2251 struct timer_list
*t
= this_cpu_ptr(&mce_timer
);
2255 mce_threshold_remove_device(cpu
);
2256 mce_device_remove(cpu
);
2260 static __init
void mce_init_banks(void)
2264 for (i
= 0; i
< mca_cfg
.banks
; i
++) {
2265 struct mce_bank
*b
= &mce_banks
[i
];
2266 struct device_attribute
*a
= &b
->attr
;
2268 sysfs_attr_init(&a
->attr
);
2269 a
->attr
.name
= b
->attrname
;
2270 snprintf(b
->attrname
, ATTR_LEN
, "bank%d", i
);
2272 a
->attr
.mode
= 0644;
2273 a
->show
= show_bank
;
2274 a
->store
= set_bank
;
2278 static __init
int mcheck_init_device(void)
2282 if (!mce_available(&boot_cpu_data
)) {
2287 if (!zalloc_cpumask_var(&mce_device_initialized
, GFP_KERNEL
)) {
2294 err
= subsys_system_register(&mce_subsys
, NULL
);
2298 err
= cpuhp_setup_state(CPUHP_X86_MCE_DEAD
, "x86/mce:dead", NULL
,
2303 err
= cpuhp_setup_state(CPUHP_AP_ONLINE_DYN
, "x86/mce:online",
2304 mce_cpu_online
, mce_cpu_pre_down
);
2306 goto err_out_online
;
2308 register_syscore_ops(&mce_syscore_ops
);
2313 cpuhp_remove_state(CPUHP_X86_MCE_DEAD
);
2316 free_cpumask_var(mce_device_initialized
);
2319 pr_err("Unable to init MCE device (rc: %d)\n", err
);
2323 device_initcall_sync(mcheck_init_device
);
2326 * Old style boot options parsing. Only for compatibility.
2328 static int __init
mcheck_disable(char *str
)
2330 mca_cfg
.disabled
= true;
2333 __setup("nomce", mcheck_disable
);
2335 #ifdef CONFIG_DEBUG_FS
2336 struct dentry
*mce_get_debugfs_dir(void)
2338 static struct dentry
*dmce
;
2341 dmce
= debugfs_create_dir("mce", NULL
);
2346 static void mce_reset(void)
2349 atomic_set(&mce_fake_panicked
, 0);
2350 atomic_set(&mce_executing
, 0);
2351 atomic_set(&mce_callin
, 0);
2352 atomic_set(&global_nwo
, 0);
2355 static int fake_panic_get(void *data
, u64
*val
)
2361 static int fake_panic_set(void *data
, u64 val
)
2368 DEFINE_SIMPLE_ATTRIBUTE(fake_panic_fops
, fake_panic_get
,
2369 fake_panic_set
, "%llu\n");
2371 static int __init
mcheck_debugfs_init(void)
2373 struct dentry
*dmce
, *ffake_panic
;
2375 dmce
= mce_get_debugfs_dir();
2378 ffake_panic
= debugfs_create_file("fake_panic", 0444, dmce
, NULL
,
2386 static int __init
mcheck_debugfs_init(void) { return -EINVAL
; }
2389 DEFINE_STATIC_KEY_FALSE(mcsafe_key
);
2390 EXPORT_SYMBOL_GPL(mcsafe_key
);
2392 static int __init
mcheck_late_init(void)
2394 if (mca_cfg
.recovery
)
2395 static_branch_inc(&mcsafe_key
);
2397 mcheck_debugfs_init();
2401 * Flush out everything that has been logged during early boot, now that
2402 * everything has been initialized (workqueues, decoders, ...).
2404 mce_schedule_work();
2408 late_initcall(mcheck_late_init
);