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 bool mce_is_memory_error(struct mce *m)
504 if (m->cpuvendor == X86_VENDOR_AMD) {
505 /* ErrCodeExt[20:16] */
506 u8 xec = (m->status >> 16) & 0x1f;
508 return (xec == 0x0 || xec == 0x8);
509 } else if (m->cpuvendor == X86_VENDOR_INTEL) {
511 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
513 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
514 * indicating a memory error. Bit 8 is used for indicating a
515 * cache hierarchy error. The combination of bit 2 and bit 3
516 * is used for indicating a `generic' cache hierarchy error
517 * But we can't just blindly check the above bits, because if
518 * bit 11 is set, then it is a bus/interconnect error - and
519 * either way the above bits just gives more detail on what
520 * bus/interconnect error happened. Note that bit 12 can be
521 * ignored, as it's the "filter" bit.
523 return (m->status & 0xef80) == BIT(7) ||
524 (m->status & 0xef00) == BIT(8) ||
525 (m->status & 0xeffc) == 0xc;
530 EXPORT_SYMBOL_GPL(mce_is_memory_error);
532 static bool cec_add_mce(struct mce *m)
537 /* We eat only correctable DRAM errors with usable addresses. */
538 if (mce_is_memory_error(m) &&
539 !(m->status & MCI_STATUS_UC) &&
540 mce_usable_address(m))
541 if (!cec_add_elem(m->addr >> PAGE_SHIFT))
547 static int mce_first_notifier(struct notifier_block *nb, unsigned long val,
550 struct mce *m = (struct mce *)data;
558 /* Emit the trace record: */
561 set_bit(0, &mce_need_notify);
568 static struct notifier_block first_nb = {
569 .notifier_call = mce_first_notifier,
570 .priority = MCE_PRIO_FIRST,
573 static int srao_decode_notifier(struct notifier_block *nb, unsigned long val,
576 struct mce *mce = (struct mce *)data;
582 if (mce_usable_address(mce) && (mce->severity == MCE_AO_SEVERITY)) {
583 pfn = mce->addr >> PAGE_SHIFT;
584 memory_failure(pfn, MCE_VECTOR, 0);
589 static struct notifier_block mce_srao_nb = {
590 .notifier_call = srao_decode_notifier,
591 .priority = MCE_PRIO_SRAO,
594 static int mce_default_notifier(struct notifier_block *nb, unsigned long val,
597 struct mce *m = (struct mce *)data;
602 if (atomic_read(&num_notifiers) > NUM_DEFAULT_NOTIFIERS)
610 static struct notifier_block mce_default_nb = {
611 .notifier_call = mce_default_notifier,
612 /* lowest prio, we want it to run last. */
613 .priority = MCE_PRIO_LOWEST,
617 * Read ADDR and MISC registers.
619 static void mce_read_aux(struct mce *m, int i)
621 if (m->status & MCI_STATUS_MISCV)
622 m->misc = mce_rdmsrl(msr_ops.misc(i));
624 if (m->status & MCI_STATUS_ADDRV) {
625 m->addr = mce_rdmsrl(msr_ops.addr(i));
628 * Mask the reported address by the reported granularity.
630 if (mca_cfg.ser && (m->status & MCI_STATUS_MISCV)) {
631 u8 shift = MCI_MISC_ADDR_LSB(m->misc);
637 * Extract [55:<lsb>] where lsb is the least significant
638 * *valid* bit of the address bits.
640 if (mce_flags.smca) {
641 u8 lsb = (m->addr >> 56) & 0x3f;
643 m->addr &= GENMASK_ULL(55, lsb);
647 if (mce_flags.smca) {
648 m->ipid = mce_rdmsrl(MSR_AMD64_SMCA_MCx_IPID(i));
650 if (m->status & MCI_STATUS_SYNDV)
651 m->synd = mce_rdmsrl(MSR_AMD64_SMCA_MCx_SYND(i));
655 DEFINE_PER_CPU(unsigned, mce_poll_count);
658 * Poll for corrected events or events that happened before reset.
659 * Those are just logged through /dev/mcelog.
661 * This is executed in standard interrupt context.
663 * Note: spec recommends to panic for fatal unsignalled
664 * errors here. However this would be quite problematic --
665 * we would need to reimplement the Monarch handling and
666 * it would mess up the exclusion between exception handler
667 * and poll hander -- * so we skip this for now.
668 * These cases should not happen anyways, or only when the CPU
669 * is already totally * confused. In this case it's likely it will
670 * not fully execute the machine check handler either.
672 bool machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
674 bool error_seen = false;
678 this_cpu_inc(mce_poll_count);
680 mce_gather_info(&m, NULL);
682 if (flags & MCP_TIMESTAMP)
685 for (i = 0; i < mca_cfg.banks; i++) {
686 if (!mce_banks[i].ctl || !test_bit(i, *b))
694 m.status = mce_rdmsrl(msr_ops.status(i));
695 if (!(m.status & MCI_STATUS_VAL))
699 * Uncorrected or signalled events are handled by the exception
700 * handler when it is enabled, so don't process those here.
702 * TBD do the same check for MCI_STATUS_EN here?
704 if (!(flags & MCP_UC) &&
705 (m.status & (mca_cfg.ser ? MCI_STATUS_S : MCI_STATUS_UC)))
712 m.severity = mce_severity(&m, mca_cfg.tolerant, NULL, false);
715 * Don't get the IP here because it's unlikely to
716 * have anything to do with the actual error location.
718 if (!(flags & MCP_DONTLOG) && !mca_cfg.dont_log_ce)
720 else if (mce_usable_address(&m)) {
722 * Although we skipped logging this, we still want
723 * to take action. Add to the pool so the registered
724 * notifiers will see it.
726 if (!mce_gen_pool_add(&m))
731 * Clear state for this bank.
733 mce_wrmsrl(msr_ops.status(i), 0);
737 * Don't clear MCG_STATUS here because it's only defined for
745 EXPORT_SYMBOL_GPL(machine_check_poll);
748 * Do a quick check if any of the events requires a panic.
749 * This decides if we keep the events around or clear them.
751 static int mce_no_way_out(struct mce *m, char **msg, unsigned long *validp,
752 struct pt_regs *regs)
757 for (i = 0; i < mca_cfg.banks; i++) {
758 m->status = mce_rdmsrl(msr_ops.status(i));
759 if (m->status & MCI_STATUS_VAL) {
760 __set_bit(i, validp);
761 if (quirk_no_way_out)
762 quirk_no_way_out(i, m, regs);
765 if (mce_severity(m, mca_cfg.tolerant, &tmp, true) >= MCE_PANIC_SEVERITY) {
774 * Variable to establish order between CPUs while scanning.
775 * Each CPU spins initially until executing is equal its number.
777 static atomic_t mce_executing;
780 * Defines order of CPUs on entry. First CPU becomes Monarch.
782 static atomic_t mce_callin;
785 * Check if a timeout waiting for other CPUs happened.
787 static int mce_timed_out(u64 *t, const char *msg)
790 * The others already did panic for some reason.
791 * Bail out like in a timeout.
792 * rmb() to tell the compiler that system_state
793 * might have been modified by someone else.
796 if (atomic_read(&mce_panicked))
798 if (!mca_cfg.monarch_timeout)
800 if ((s64)*t < SPINUNIT) {
801 if (mca_cfg.tolerant <= 1)
802 mce_panic(msg, NULL, NULL);
808 touch_nmi_watchdog();
813 * The Monarch's reign. The Monarch is the CPU who entered
814 * the machine check handler first. It waits for the others to
815 * raise the exception too and then grades them. When any
816 * error is fatal panic. Only then let the others continue.
818 * The other CPUs entering the MCE handler will be controlled by the
819 * Monarch. They are called Subjects.
821 * This way we prevent any potential data corruption in a unrecoverable case
822 * and also makes sure always all CPU's errors are examined.
824 * Also this detects the case of a machine check event coming from outer
825 * space (not detected by any CPUs) In this case some external agent wants
826 * us to shut down, so panic too.
828 * The other CPUs might still decide to panic if the handler happens
829 * in a unrecoverable place, but in this case the system is in a semi-stable
830 * state and won't corrupt anything by itself. It's ok to let the others
831 * continue for a bit first.
833 * All the spin loops have timeouts; when a timeout happens a CPU
834 * typically elects itself to be Monarch.
836 static void mce_reign(void)
839 struct mce *m = NULL;
840 int global_worst = 0;
845 * This CPU is the Monarch and the other CPUs have run
846 * through their handlers.
847 * Grade the severity of the errors of all the CPUs.
849 for_each_possible_cpu(cpu) {
850 int severity = mce_severity(&per_cpu(mces_seen, cpu),
853 if (severity > global_worst) {
855 global_worst = severity;
856 m = &per_cpu(mces_seen, cpu);
861 * Cannot recover? Panic here then.
862 * This dumps all the mces in the log buffer and stops the
865 if (m && global_worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3)
866 mce_panic("Fatal machine check", m, msg);
869 * For UC somewhere we let the CPU who detects it handle it.
870 * Also must let continue the others, otherwise the handling
871 * CPU could deadlock on a lock.
875 * No machine check event found. Must be some external
876 * source or one CPU is hung. Panic.
878 if (global_worst <= MCE_KEEP_SEVERITY && mca_cfg.tolerant < 3)
879 mce_panic("Fatal machine check from unknown source", NULL, NULL);
882 * Now clear all the mces_seen so that they don't reappear on
885 for_each_possible_cpu(cpu)
886 memset(&per_cpu(mces_seen, cpu), 0, sizeof(struct mce));
889 static atomic_t global_nwo;
892 * Start of Monarch synchronization. This waits until all CPUs have
893 * entered the exception handler and then determines if any of them
894 * saw a fatal event that requires panic. Then it executes them
895 * in the entry order.
896 * TBD double check parallel CPU hotunplug
898 static int mce_start(int *no_way_out)
901 int cpus = num_online_cpus();
902 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
907 atomic_add(*no_way_out, &global_nwo);
909 * Rely on the implied barrier below, such that global_nwo
910 * is updated before mce_callin.
912 order = atomic_inc_return(&mce_callin);
917 while (atomic_read(&mce_callin) != cpus) {
918 if (mce_timed_out(&timeout,
919 "Timeout: Not all CPUs entered broadcast exception handler")) {
920 atomic_set(&global_nwo, 0);
927 * mce_callin should be read before global_nwo
933 * Monarch: Starts executing now, the others wait.
935 atomic_set(&mce_executing, 1);
938 * Subject: Now start the scanning loop one by one in
939 * the original callin order.
940 * This way when there are any shared banks it will be
941 * only seen by one CPU before cleared, avoiding duplicates.
943 while (atomic_read(&mce_executing) < order) {
944 if (mce_timed_out(&timeout,
945 "Timeout: Subject CPUs unable to finish machine check processing")) {
946 atomic_set(&global_nwo, 0);
954 * Cache the global no_way_out state.
956 *no_way_out = atomic_read(&global_nwo);
962 * Synchronize between CPUs after main scanning loop.
963 * This invokes the bulk of the Monarch processing.
965 static int mce_end(int order)
968 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
976 * Allow others to run.
978 atomic_inc(&mce_executing);
981 /* CHECKME: Can this race with a parallel hotplug? */
982 int cpus = num_online_cpus();
985 * Monarch: Wait for everyone to go through their scanning
988 while (atomic_read(&mce_executing) <= cpus) {
989 if (mce_timed_out(&timeout,
990 "Timeout: Monarch CPU unable to finish machine check processing"))
1000 * Subject: Wait for Monarch to finish.
1002 while (atomic_read(&mce_executing) != 0) {
1003 if (mce_timed_out(&timeout,
1004 "Timeout: Monarch CPU did not finish machine check processing"))
1010 * Don't reset anything. That's done by the Monarch.
1016 * Reset all global state.
1019 atomic_set(&global_nwo, 0);
1020 atomic_set(&mce_callin, 0);
1024 * Let others run again.
1026 atomic_set(&mce_executing, 0);
1030 static void mce_clear_state(unsigned long *toclear)
1034 for (i = 0; i < mca_cfg.banks; i++) {
1035 if (test_bit(i, toclear))
1036 mce_wrmsrl(msr_ops.status(i), 0);
1040 static int do_memory_failure(struct mce *m)
1042 int flags = MF_ACTION_REQUIRED;
1045 pr_err("Uncorrected hardware memory error in user-access at %llx", m->addr);
1046 if (!(m->mcgstatus & MCG_STATUS_RIPV))
1047 flags |= MF_MUST_KILL;
1048 ret = memory_failure(m->addr >> PAGE_SHIFT, MCE_VECTOR, flags);
1050 pr_err("Memory error not recovered");
1055 * The actual machine check handler. This only handles real
1056 * exceptions when something got corrupted coming in through int 18.
1058 * This is executed in NMI context not subject to normal locking rules. This
1059 * implies that most kernel services cannot be safely used. Don't even
1060 * think about putting a printk in there!
1062 * On Intel systems this is entered on all CPUs in parallel through
1063 * MCE broadcast. However some CPUs might be broken beyond repair,
1064 * so be always careful when synchronizing with others.
1066 void do_machine_check(struct pt_regs *regs, long error_code)
1068 struct mca_config *cfg = &mca_cfg;
1069 struct mce m, *final;
1075 * Establish sequential order between the CPUs entering the machine
1080 * If no_way_out gets set, there is no safe way to recover from this
1081 * MCE. If mca_cfg.tolerant is cranked up, we'll try anyway.
1085 * If kill_it gets set, there might be a way to recover from this
1089 DECLARE_BITMAP(toclear, MAX_NR_BANKS);
1090 DECLARE_BITMAP(valid_banks, MAX_NR_BANKS);
1091 char *msg = "Unknown";
1094 * MCEs are always local on AMD. Same is determined by MCG_STATUS_LMCES
1098 int cpu = smp_processor_id();
1101 * Cases where we avoid rendezvous handler timeout:
1102 * 1) If this CPU is offline.
1104 * 2) If crashing_cpu was set, e.g. we're entering kdump and we need to
1105 * skip those CPUs which remain looping in the 1st kernel - see
1106 * crash_nmi_callback().
1108 * Note: there still is a small window between kexec-ing and the new,
1109 * kdump kernel establishing a new #MC handler where a broadcasted MCE
1110 * might not get handled properly.
1112 if (cpu_is_offline(cpu) ||
1113 (crashing_cpu != -1 && crashing_cpu != cpu)) {
1116 mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
1117 if (mcgstatus & MCG_STATUS_RIPV) {
1118 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1125 this_cpu_inc(mce_exception_count);
1130 mce_gather_info(&m, regs);
1133 final = this_cpu_ptr(&mces_seen);
1136 memset(valid_banks, 0, sizeof(valid_banks));
1137 no_way_out = mce_no_way_out(&m, &msg, valid_banks, regs);
1142 * When no restart IP might need to kill or panic.
1143 * Assume the worst for now, but if we find the
1144 * severity is MCE_AR_SEVERITY we have other options.
1146 if (!(m.mcgstatus & MCG_STATUS_RIPV))
1150 * Check if this MCE is signaled to only this logical processor,
1153 if (m.cpuvendor == X86_VENDOR_INTEL)
1154 lmce = m.mcgstatus & MCG_STATUS_LMCES;
1157 * Go through all banks in exclusion of the other CPUs. This way we
1158 * don't report duplicated events on shared banks because the first one
1159 * to see it will clear it. If this is a Local MCE, then no need to
1160 * perform rendezvous.
1163 order = mce_start(&no_way_out);
1165 for (i = 0; i < cfg->banks; i++) {
1166 __clear_bit(i, toclear);
1167 if (!test_bit(i, valid_banks))
1169 if (!mce_banks[i].ctl)
1176 m.status = mce_rdmsrl(msr_ops.status(i));
1177 if ((m.status & MCI_STATUS_VAL) == 0)
1181 * Non uncorrected or non signaled errors are handled by
1182 * machine_check_poll. Leave them alone, unless this panics.
1184 if (!(m.status & (cfg->ser ? MCI_STATUS_S : MCI_STATUS_UC)) &&
1189 * Set taint even when machine check was not enabled.
1191 add_taint(TAINT_MACHINE_CHECK, LOCKDEP_NOW_UNRELIABLE);
1193 severity = mce_severity(&m, cfg->tolerant, NULL, true);
1196 * When machine check was for corrected/deferred handler don't
1197 * touch, unless we're panicing.
1199 if ((severity == MCE_KEEP_SEVERITY ||
1200 severity == MCE_UCNA_SEVERITY) && !no_way_out)
1202 __set_bit(i, toclear);
1203 if (severity == MCE_NO_SEVERITY) {
1205 * Machine check event was not enabled. Clear, but
1211 mce_read_aux(&m, i);
1213 /* assuming valid severity level != 0 */
1214 m.severity = severity;
1218 if (severity > worst) {
1224 /* mce_clear_state will clear *final, save locally for use later */
1228 mce_clear_state(toclear);
1231 * Do most of the synchronization with other CPUs.
1232 * When there's any problem use only local no_way_out state.
1235 if (mce_end(order) < 0)
1236 no_way_out = worst >= MCE_PANIC_SEVERITY;
1239 * Local MCE skipped calling mce_reign()
1240 * If we found a fatal error, we need to panic here.
1242 if (worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3)
1243 mce_panic("Machine check from unknown source",
1248 * If tolerant is at an insane level we drop requests to kill
1249 * processes and continue even when there is no way out.
1251 if (cfg->tolerant == 3)
1253 else if (no_way_out)
1254 mce_panic("Fatal machine check on current CPU", &m, msg);
1257 mce_report_event(regs);
1258 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1262 if (worst != MCE_AR_SEVERITY && !kill_it)
1265 /* Fault was in user mode and we need to take some action */
1266 if ((m.cs & 3) == 3) {
1267 ist_begin_non_atomic(regs);
1270 if (kill_it || do_memory_failure(&m))
1271 force_sig(SIGBUS, current);
1272 local_irq_disable();
1273 ist_end_non_atomic();
1275 if (!fixup_exception(regs, X86_TRAP_MC))
1276 mce_panic("Failed kernel mode recovery", &m, NULL);
1282 EXPORT_SYMBOL_GPL(do_machine_check);
1284 #ifndef CONFIG_MEMORY_FAILURE
1285 int memory_failure(unsigned long pfn, int vector, int flags)
1287 /* mce_severity() should not hand us an ACTION_REQUIRED error */
1288 BUG_ON(flags & MF_ACTION_REQUIRED);
1289 pr_err("Uncorrected memory error in page 0x%lx ignored\n"
1290 "Rebuild kernel with CONFIG_MEMORY_FAILURE=y for smarter handling\n",
1298 * Periodic polling timer for "silent" machine check errors. If the
1299 * poller finds an MCE, poll 2x faster. When the poller finds no more
1300 * errors, poll 2x slower (up to check_interval seconds).
1302 static unsigned long check_interval = INITIAL_CHECK_INTERVAL;
1304 static DEFINE_PER_CPU(unsigned long, mce_next_interval); /* in jiffies */
1305 static DEFINE_PER_CPU(struct timer_list, mce_timer);
1307 static unsigned long mce_adjust_timer_default(unsigned long interval)
1312 static unsigned long (*mce_adjust_timer)(unsigned long interval) = mce_adjust_timer_default;
1314 static void __start_timer(struct timer_list *t, unsigned long interval)
1316 unsigned long when = jiffies + interval;
1317 unsigned long flags;
1319 local_irq_save(flags);
1321 if (!timer_pending(t) || time_before(when, t->expires))
1322 mod_timer(t, round_jiffies(when));
1324 local_irq_restore(flags);
1327 static void mce_timer_fn(unsigned long data)
1329 struct timer_list *t = this_cpu_ptr(&mce_timer);
1330 int cpu = smp_processor_id();
1333 WARN_ON(cpu != data);
1335 iv = __this_cpu_read(mce_next_interval);
1337 if (mce_available(this_cpu_ptr(&cpu_info))) {
1338 machine_check_poll(0, this_cpu_ptr(&mce_poll_banks));
1340 if (mce_intel_cmci_poll()) {
1341 iv = mce_adjust_timer(iv);
1347 * Alert userspace if needed. If we logged an MCE, reduce the polling
1348 * interval, otherwise increase the polling interval.
1350 if (mce_notify_irq())
1351 iv = max(iv / 2, (unsigned long) HZ/100);
1353 iv = min(iv * 2, round_jiffies_relative(check_interval * HZ));
1356 __this_cpu_write(mce_next_interval, iv);
1357 __start_timer(t, iv);
1361 * Ensure that the timer is firing in @interval from now.
1363 void mce_timer_kick(unsigned long interval)
1365 struct timer_list *t = this_cpu_ptr(&mce_timer);
1366 unsigned long iv = __this_cpu_read(mce_next_interval);
1368 __start_timer(t, interval);
1371 __this_cpu_write(mce_next_interval, interval);
1374 /* Must not be called in IRQ context where del_timer_sync() can deadlock */
1375 static void mce_timer_delete_all(void)
1379 for_each_online_cpu(cpu)
1380 del_timer_sync(&per_cpu(mce_timer, cpu));
1384 * Notify the user(s) about new machine check events.
1385 * Can be called from interrupt context, but not from machine check/NMI
1388 int mce_notify_irq(void)
1390 /* Not more than two messages every minute */
1391 static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2);
1393 if (test_and_clear_bit(0, &mce_need_notify)) {
1396 if (__ratelimit(&ratelimit))
1397 pr_info(HW_ERR "Machine check events logged\n");
1403 EXPORT_SYMBOL_GPL(mce_notify_irq);
1405 static int __mcheck_cpu_mce_banks_init(void)
1408 u8 num_banks = mca_cfg.banks;
1410 mce_banks = kzalloc(num_banks * sizeof(struct mce_bank), GFP_KERNEL);
1414 for (i = 0; i < num_banks; i++) {
1415 struct mce_bank *b = &mce_banks[i];
1424 * Initialize Machine Checks for a CPU.
1426 static int __mcheck_cpu_cap_init(void)
1431 rdmsrl(MSR_IA32_MCG_CAP, cap);
1433 b = cap & MCG_BANKCNT_MASK;
1435 pr_info("CPU supports %d MCE banks\n", b);
1437 if (b > MAX_NR_BANKS) {
1438 pr_warn("Using only %u machine check banks out of %u\n",
1443 /* Don't support asymmetric configurations today */
1444 WARN_ON(mca_cfg.banks != 0 && b != mca_cfg.banks);
1448 int err = __mcheck_cpu_mce_banks_init();
1454 /* Use accurate RIP reporting if available. */
1455 if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
1456 mca_cfg.rip_msr = MSR_IA32_MCG_EIP;
1458 if (cap & MCG_SER_P)
1464 static void __mcheck_cpu_init_generic(void)
1466 enum mcp_flags m_fl = 0;
1467 mce_banks_t all_banks;
1470 if (!mca_cfg.bootlog)
1474 * Log the machine checks left over from the previous reset.
1476 bitmap_fill(all_banks, MAX_NR_BANKS);
1477 machine_check_poll(MCP_UC | m_fl, &all_banks);
1479 cr4_set_bits(X86_CR4_MCE);
1481 rdmsrl(MSR_IA32_MCG_CAP, cap);
1482 if (cap & MCG_CTL_P)
1483 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
1486 static void __mcheck_cpu_init_clear_banks(void)
1490 for (i = 0; i < mca_cfg.banks; i++) {
1491 struct mce_bank *b = &mce_banks[i];
1495 wrmsrl(msr_ops.ctl(i), b->ctl);
1496 wrmsrl(msr_ops.status(i), 0);
1501 * During IFU recovery Sandy Bridge -EP4S processors set the RIPV and
1502 * EIPV bits in MCG_STATUS to zero on the affected logical processor (SDM
1503 * Vol 3B Table 15-20). But this confuses both the code that determines
1504 * whether the machine check occurred in kernel or user mode, and also
1505 * the severity assessment code. Pretend that EIPV was set, and take the
1506 * ip/cs values from the pt_regs that mce_gather_info() ignored earlier.
1508 static void quirk_sandybridge_ifu(int bank, struct mce *m, struct pt_regs *regs)
1512 if ((m->mcgstatus & (MCG_STATUS_EIPV|MCG_STATUS_RIPV)) != 0)
1514 if ((m->status & (MCI_STATUS_OVER|MCI_STATUS_UC|
1515 MCI_STATUS_EN|MCI_STATUS_MISCV|MCI_STATUS_ADDRV|
1516 MCI_STATUS_PCC|MCI_STATUS_S|MCI_STATUS_AR|
1518 (MCI_STATUS_UC|MCI_STATUS_EN|
1519 MCI_STATUS_MISCV|MCI_STATUS_ADDRV|MCI_STATUS_S|
1520 MCI_STATUS_AR|MCACOD_INSTR))
1523 m->mcgstatus |= MCG_STATUS_EIPV;
1528 /* Add per CPU specific workarounds here */
1529 static int __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
1531 struct mca_config *cfg = &mca_cfg;
1533 if (c->x86_vendor == X86_VENDOR_UNKNOWN) {
1534 pr_info("unknown CPU type - not enabling MCE support\n");
1538 /* This should be disabled by the BIOS, but isn't always */
1539 if (c->x86_vendor == X86_VENDOR_AMD) {
1540 if (c->x86 == 15 && cfg->banks > 4) {
1542 * disable GART TBL walk error reporting, which
1543 * trips off incorrectly with the IOMMU & 3ware
1546 clear_bit(10, (unsigned long *)&mce_banks[4].ctl);
1548 if (c->x86 < 0x11 && cfg->bootlog < 0) {
1550 * Lots of broken BIOS around that don't clear them
1551 * by default and leave crap in there. Don't log:
1556 * Various K7s with broken bank 0 around. Always disable
1559 if (c->x86 == 6 && cfg->banks > 0)
1560 mce_banks[0].ctl = 0;
1563 * overflow_recov is supported for F15h Models 00h-0fh
1564 * even though we don't have a CPUID bit for it.
1566 if (c->x86 == 0x15 && c->x86_model <= 0xf)
1567 mce_flags.overflow_recov = 1;
1570 * Turn off MC4_MISC thresholding banks on those models since
1571 * they're not supported there.
1573 if (c->x86 == 0x15 &&
1574 (c->x86_model >= 0x10 && c->x86_model <= 0x1f)) {
1579 0x00000413, /* MC4_MISC0 */
1580 0xc0000408, /* MC4_MISC1 */
1583 rdmsrl(MSR_K7_HWCR, hwcr);
1585 /* McStatusWrEn has to be set */
1586 need_toggle = !(hwcr & BIT(18));
1589 wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
1591 /* Clear CntP bit safely */
1592 for (i = 0; i < ARRAY_SIZE(msrs); i++)
1593 msr_clear_bit(msrs[i], 62);
1595 /* restore old settings */
1597 wrmsrl(MSR_K7_HWCR, hwcr);
1601 if (c->x86_vendor == X86_VENDOR_INTEL) {
1603 * SDM documents that on family 6 bank 0 should not be written
1604 * because it aliases to another special BIOS controlled
1606 * But it's not aliased anymore on model 0x1a+
1607 * Don't ignore bank 0 completely because there could be a
1608 * valid event later, merely don't write CTL0.
1611 if (c->x86 == 6 && c->x86_model < 0x1A && cfg->banks > 0)
1612 mce_banks[0].init = 0;
1615 * All newer Intel systems support MCE broadcasting. Enable
1616 * synchronization with a one second timeout.
1618 if ((c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xe)) &&
1619 cfg->monarch_timeout < 0)
1620 cfg->monarch_timeout = USEC_PER_SEC;
1623 * There are also broken BIOSes on some Pentium M and
1626 if (c->x86 == 6 && c->x86_model <= 13 && cfg->bootlog < 0)
1629 if (c->x86 == 6 && c->x86_model == 45)
1630 quirk_no_way_out = quirk_sandybridge_ifu;
1632 if (cfg->monarch_timeout < 0)
1633 cfg->monarch_timeout = 0;
1634 if (cfg->bootlog != 0)
1635 cfg->panic_timeout = 30;
1640 static int __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c)
1645 switch (c->x86_vendor) {
1646 case X86_VENDOR_INTEL:
1647 intel_p5_mcheck_init(c);
1650 case X86_VENDOR_CENTAUR:
1651 winchip_mcheck_init(c);
1662 * Init basic CPU features needed for early decoding of MCEs.
1664 static void __mcheck_cpu_init_early(struct cpuinfo_x86 *c)
1666 if (c->x86_vendor == X86_VENDOR_AMD) {
1667 mce_flags.overflow_recov = !!cpu_has(c, X86_FEATURE_OVERFLOW_RECOV);
1668 mce_flags.succor = !!cpu_has(c, X86_FEATURE_SUCCOR);
1669 mce_flags.smca = !!cpu_has(c, X86_FEATURE_SMCA);
1671 if (mce_flags.smca) {
1672 msr_ops.ctl = smca_ctl_reg;
1673 msr_ops.status = smca_status_reg;
1674 msr_ops.addr = smca_addr_reg;
1675 msr_ops.misc = smca_misc_reg;
1680 static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c)
1682 switch (c->x86_vendor) {
1683 case X86_VENDOR_INTEL:
1684 mce_intel_feature_init(c);
1685 mce_adjust_timer = cmci_intel_adjust_timer;
1688 case X86_VENDOR_AMD: {
1689 mce_amd_feature_init(c);
1698 static void __mcheck_cpu_clear_vendor(struct cpuinfo_x86 *c)
1700 switch (c->x86_vendor) {
1701 case X86_VENDOR_INTEL:
1702 mce_intel_feature_clear(c);
1709 static void mce_start_timer(struct timer_list *t)
1711 unsigned long iv = check_interval * HZ;
1713 if (mca_cfg.ignore_ce || !iv)
1716 this_cpu_write(mce_next_interval, iv);
1717 __start_timer(t, iv);
1720 static void __mcheck_cpu_setup_timer(void)
1722 struct timer_list *t = this_cpu_ptr(&mce_timer);
1723 unsigned int cpu = smp_processor_id();
1725 setup_pinned_timer(t, mce_timer_fn, cpu);
1728 static void __mcheck_cpu_init_timer(void)
1730 struct timer_list *t = this_cpu_ptr(&mce_timer);
1731 unsigned int cpu = smp_processor_id();
1733 setup_pinned_timer(t, mce_timer_fn, cpu);
1737 /* Handle unconfigured int18 (should never happen) */
1738 static void unexpected_machine_check(struct pt_regs *regs, long error_code)
1740 pr_err("CPU#%d: Unexpected int18 (Machine Check)\n",
1741 smp_processor_id());
1744 /* Call the installed machine check handler for this CPU setup. */
1745 void (*machine_check_vector)(struct pt_regs *, long error_code) =
1746 unexpected_machine_check;
1749 * Called for each booted CPU to set up machine checks.
1750 * Must be called with preempt off:
1752 void mcheck_cpu_init(struct cpuinfo_x86 *c)
1754 if (mca_cfg.disabled)
1757 if (__mcheck_cpu_ancient_init(c))
1760 if (!mce_available(c))
1763 if (__mcheck_cpu_cap_init() < 0 || __mcheck_cpu_apply_quirks(c) < 0) {
1764 mca_cfg.disabled = true;
1768 if (mce_gen_pool_init()) {
1769 mca_cfg.disabled = true;
1770 pr_emerg("Couldn't allocate MCE records pool!\n");
1774 machine_check_vector = do_machine_check;
1776 __mcheck_cpu_init_early(c);
1777 __mcheck_cpu_init_generic();
1778 __mcheck_cpu_init_vendor(c);
1779 __mcheck_cpu_init_clear_banks();
1780 __mcheck_cpu_setup_timer();
1784 * Called for each booted CPU to clear some machine checks opt-ins
1786 void mcheck_cpu_clear(struct cpuinfo_x86 *c)
1788 if (mca_cfg.disabled)
1791 if (!mce_available(c))
1795 * Possibly to clear general settings generic to x86
1796 * __mcheck_cpu_clear_generic(c);
1798 __mcheck_cpu_clear_vendor(c);
1802 static void __mce_disable_bank(void *arg)
1804 int bank = *((int *)arg);
1805 __clear_bit(bank, this_cpu_ptr(mce_poll_banks));
1806 cmci_disable_bank(bank);
1809 void mce_disable_bank(int bank)
1811 if (bank >= mca_cfg.banks) {
1813 "Ignoring request to disable invalid MCA bank %d.\n",
1817 set_bit(bank, mce_banks_ce_disabled);
1818 on_each_cpu(__mce_disable_bank, &bank, 1);
1822 * mce=off Disables machine check
1823 * mce=no_cmci Disables CMCI
1824 * mce=no_lmce Disables LMCE
1825 * mce=dont_log_ce Clears corrected events silently, no log created for CEs.
1826 * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared.
1827 * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above)
1828 * monarchtimeout is how long to wait for other CPUs on machine
1829 * check, or 0 to not wait
1830 * mce=bootlog Log MCEs from before booting. Disabled by default on AMD Fam10h
1832 * mce=nobootlog Don't log MCEs from before booting.
1833 * mce=bios_cmci_threshold Don't program the CMCI threshold
1834 * mce=recovery force enable memcpy_mcsafe()
1836 static int __init mcheck_enable(char *str)
1838 struct mca_config *cfg = &mca_cfg;
1846 if (!strcmp(str, "off"))
1847 cfg->disabled = true;
1848 else if (!strcmp(str, "no_cmci"))
1849 cfg->cmci_disabled = true;
1850 else if (!strcmp(str, "no_lmce"))
1851 cfg->lmce_disabled = true;
1852 else if (!strcmp(str, "dont_log_ce"))
1853 cfg->dont_log_ce = true;
1854 else if (!strcmp(str, "ignore_ce"))
1855 cfg->ignore_ce = true;
1856 else if (!strcmp(str, "bootlog") || !strcmp(str, "nobootlog"))
1857 cfg->bootlog = (str[0] == 'b');
1858 else if (!strcmp(str, "bios_cmci_threshold"))
1859 cfg->bios_cmci_threshold = true;
1860 else if (!strcmp(str, "recovery"))
1861 cfg->recovery = true;
1862 else if (isdigit(str[0])) {
1863 if (get_option(&str, &cfg->tolerant) == 2)
1864 get_option(&str, &(cfg->monarch_timeout));
1866 pr_info("mce argument %s ignored. Please use /sys\n", str);
1871 __setup("mce", mcheck_enable);
1873 int __init mcheck_init(void)
1875 mcheck_intel_therm_init();
1876 mce_register_decode_chain(&first_nb);
1877 mce_register_decode_chain(&mce_srao_nb);
1878 mce_register_decode_chain(&mce_default_nb);
1879 mcheck_vendor_init_severity();
1881 INIT_WORK(&mce_work, mce_gen_pool_process);
1882 init_irq_work(&mce_irq_work, mce_irq_work_cb);
1888 * mce_syscore: PM support
1892 * Disable machine checks on suspend and shutdown. We can't really handle
1895 static void mce_disable_error_reporting(void)
1899 for (i = 0; i < mca_cfg.banks; i++) {
1900 struct mce_bank *b = &mce_banks[i];
1903 wrmsrl(msr_ops.ctl(i), 0);
1908 static void vendor_disable_error_reporting(void)
1911 * Don't clear on Intel or AMD CPUs. Some of these MSRs are socket-wide.
1912 * Disabling them for just a single offlined CPU is bad, since it will
1913 * inhibit reporting for all shared resources on the socket like the
1914 * last level cache (LLC), the integrated memory controller (iMC), etc.
1916 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL ||
1917 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1920 mce_disable_error_reporting();
1923 static int mce_syscore_suspend(void)
1925 vendor_disable_error_reporting();
1929 static void mce_syscore_shutdown(void)
1931 vendor_disable_error_reporting();
1935 * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
1936 * Only one CPU is active at this time, the others get re-added later using
1939 static void mce_syscore_resume(void)
1941 __mcheck_cpu_init_generic();
1942 __mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info));
1943 __mcheck_cpu_init_clear_banks();
1946 static struct syscore_ops mce_syscore_ops = {
1947 .suspend = mce_syscore_suspend,
1948 .shutdown = mce_syscore_shutdown,
1949 .resume = mce_syscore_resume,
1953 * mce_device: Sysfs support
1956 static void mce_cpu_restart(void *data)
1958 if (!mce_available(raw_cpu_ptr(&cpu_info)))
1960 __mcheck_cpu_init_generic();
1961 __mcheck_cpu_init_clear_banks();
1962 __mcheck_cpu_init_timer();
1965 /* Reinit MCEs after user configuration changes */
1966 static void mce_restart(void)
1968 mce_timer_delete_all();
1969 on_each_cpu(mce_cpu_restart, NULL, 1);
1972 /* Toggle features for corrected errors */
1973 static void mce_disable_cmci(void *data)
1975 if (!mce_available(raw_cpu_ptr(&cpu_info)))
1980 static void mce_enable_ce(void *all)
1982 if (!mce_available(raw_cpu_ptr(&cpu_info)))
1987 __mcheck_cpu_init_timer();
1990 static struct bus_type mce_subsys = {
1991 .name = "machinecheck",
1992 .dev_name = "machinecheck",
1995 DEFINE_PER_CPU(struct device *, mce_device);
1997 static inline struct mce_bank *attr_to_bank(struct device_attribute *attr)
1999 return container_of(attr, struct mce_bank, attr);
2002 static ssize_t show_bank(struct device *s, struct device_attribute *attr,
2005 return sprintf(buf, "%llx\n", attr_to_bank(attr)->ctl);
2008 static ssize_t set_bank(struct device *s, struct device_attribute *attr,
2009 const char *buf, size_t size)
2013 if (kstrtou64(buf, 0, &new) < 0)
2016 attr_to_bank(attr)->ctl = new;
2022 static ssize_t set_ignore_ce(struct device *s,
2023 struct device_attribute *attr,
2024 const char *buf, size_t size)
2028 if (kstrtou64(buf, 0, &new) < 0)
2031 if (mca_cfg.ignore_ce ^ !!new) {
2033 /* disable ce features */
2034 mce_timer_delete_all();
2035 on_each_cpu(mce_disable_cmci, NULL, 1);
2036 mca_cfg.ignore_ce = true;
2038 /* enable ce features */
2039 mca_cfg.ignore_ce = false;
2040 on_each_cpu(mce_enable_ce, (void *)1, 1);
2046 static ssize_t set_cmci_disabled(struct device *s,
2047 struct device_attribute *attr,
2048 const char *buf, size_t size)
2052 if (kstrtou64(buf, 0, &new) < 0)
2055 if (mca_cfg.cmci_disabled ^ !!new) {
2058 on_each_cpu(mce_disable_cmci, NULL, 1);
2059 mca_cfg.cmci_disabled = true;
2062 mca_cfg.cmci_disabled = false;
2063 on_each_cpu(mce_enable_ce, NULL, 1);
2069 static ssize_t store_int_with_restart(struct device *s,
2070 struct device_attribute *attr,
2071 const char *buf, size_t size)
2073 ssize_t ret = device_store_int(s, attr, buf, size);
2078 static DEVICE_INT_ATTR(tolerant, 0644, mca_cfg.tolerant);
2079 static DEVICE_INT_ATTR(monarch_timeout, 0644, mca_cfg.monarch_timeout);
2080 static DEVICE_BOOL_ATTR(dont_log_ce, 0644, mca_cfg.dont_log_ce);
2082 static struct dev_ext_attribute dev_attr_check_interval = {
2083 __ATTR(check_interval, 0644, device_show_int, store_int_with_restart),
2087 static struct dev_ext_attribute dev_attr_ignore_ce = {
2088 __ATTR(ignore_ce, 0644, device_show_bool, set_ignore_ce),
2092 static struct dev_ext_attribute dev_attr_cmci_disabled = {
2093 __ATTR(cmci_disabled, 0644, device_show_bool, set_cmci_disabled),
2094 &mca_cfg.cmci_disabled
2097 static struct device_attribute *mce_device_attrs[] = {
2098 &dev_attr_tolerant.attr,
2099 &dev_attr_check_interval.attr,
2100 #ifdef CONFIG_X86_MCELOG_LEGACY
2103 &dev_attr_monarch_timeout.attr,
2104 &dev_attr_dont_log_ce.attr,
2105 &dev_attr_ignore_ce.attr,
2106 &dev_attr_cmci_disabled.attr,
2110 static cpumask_var_t mce_device_initialized;
2112 static void mce_device_release(struct device *dev)
2117 /* Per cpu device init. All of the cpus still share the same ctrl bank: */
2118 static int mce_device_create(unsigned int cpu)
2124 if (!mce_available(&boot_cpu_data))
2127 dev = per_cpu(mce_device, cpu);
2131 dev = kzalloc(sizeof *dev, GFP_KERNEL);
2135 dev->bus = &mce_subsys;
2136 dev->release = &mce_device_release;
2138 err = device_register(dev);
2144 for (i = 0; mce_device_attrs[i]; i++) {
2145 err = device_create_file(dev, mce_device_attrs[i]);
2149 for (j = 0; j < mca_cfg.banks; j++) {
2150 err = device_create_file(dev, &mce_banks[j].attr);
2154 cpumask_set_cpu(cpu, mce_device_initialized);
2155 per_cpu(mce_device, cpu) = dev;
2160 device_remove_file(dev, &mce_banks[j].attr);
2163 device_remove_file(dev, mce_device_attrs[i]);
2165 device_unregister(dev);
2170 static void mce_device_remove(unsigned int cpu)
2172 struct device *dev = per_cpu(mce_device, cpu);
2175 if (!cpumask_test_cpu(cpu, mce_device_initialized))
2178 for (i = 0; mce_device_attrs[i]; i++)
2179 device_remove_file(dev, mce_device_attrs[i]);
2181 for (i = 0; i < mca_cfg.banks; i++)
2182 device_remove_file(dev, &mce_banks[i].attr);
2184 device_unregister(dev);
2185 cpumask_clear_cpu(cpu, mce_device_initialized);
2186 per_cpu(mce_device, cpu) = NULL;
2189 /* Make sure there are no machine checks on offlined CPUs. */
2190 static void mce_disable_cpu(void)
2192 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2195 if (!cpuhp_tasks_frozen)
2198 vendor_disable_error_reporting();
2201 static void mce_reenable_cpu(void)
2205 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2208 if (!cpuhp_tasks_frozen)
2210 for (i = 0; i < mca_cfg.banks; i++) {
2211 struct mce_bank *b = &mce_banks[i];
2214 wrmsrl(msr_ops.ctl(i), b->ctl);
2218 static int mce_cpu_dead(unsigned int cpu)
2220 mce_intel_hcpu_update(cpu);
2222 /* intentionally ignoring frozen here */
2223 if (!cpuhp_tasks_frozen)
2228 static int mce_cpu_online(unsigned int cpu)
2230 struct timer_list *t = this_cpu_ptr(&mce_timer);
2233 mce_device_create(cpu);
2235 ret = mce_threshold_create_device(cpu);
2237 mce_device_remove(cpu);
2245 static int mce_cpu_pre_down(unsigned int cpu)
2247 struct timer_list *t = this_cpu_ptr(&mce_timer);
2251 mce_threshold_remove_device(cpu);
2252 mce_device_remove(cpu);
2256 static __init void mce_init_banks(void)
2260 for (i = 0; i < mca_cfg.banks; i++) {
2261 struct mce_bank *b = &mce_banks[i];
2262 struct device_attribute *a = &b->attr;
2264 sysfs_attr_init(&a->attr);
2265 a->attr.name = b->attrname;
2266 snprintf(b->attrname, ATTR_LEN, "bank%d", i);
2268 a->attr.mode = 0644;
2269 a->show = show_bank;
2270 a->store = set_bank;
2274 static __init int mcheck_init_device(void)
2278 if (!mce_available(&boot_cpu_data)) {
2283 if (!zalloc_cpumask_var(&mce_device_initialized, GFP_KERNEL)) {
2290 err = subsys_system_register(&mce_subsys, NULL);
2294 err = cpuhp_setup_state(CPUHP_X86_MCE_DEAD, "x86/mce:dead", NULL,
2299 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "x86/mce:online",
2300 mce_cpu_online, mce_cpu_pre_down);
2302 goto err_out_online;
2304 register_syscore_ops(&mce_syscore_ops);
2309 cpuhp_remove_state(CPUHP_X86_MCE_DEAD);
2312 free_cpumask_var(mce_device_initialized);
2315 pr_err("Unable to init MCE device (rc: %d)\n", err);
2319 device_initcall_sync(mcheck_init_device);
2322 * Old style boot options parsing. Only for compatibility.
2324 static int __init mcheck_disable(char *str)
2326 mca_cfg.disabled = true;
2329 __setup("nomce", mcheck_disable);
2331 #ifdef CONFIG_DEBUG_FS
2332 struct dentry *mce_get_debugfs_dir(void)
2334 static struct dentry *dmce;
2337 dmce = debugfs_create_dir("mce", NULL);
2342 static void mce_reset(void)
2345 atomic_set(&mce_fake_panicked, 0);
2346 atomic_set(&mce_executing, 0);
2347 atomic_set(&mce_callin, 0);
2348 atomic_set(&global_nwo, 0);
2351 static int fake_panic_get(void *data, u64 *val)
2357 static int fake_panic_set(void *data, u64 val)
2364 DEFINE_SIMPLE_ATTRIBUTE(fake_panic_fops, fake_panic_get,
2365 fake_panic_set, "%llu\n");
2367 static int __init mcheck_debugfs_init(void)
2369 struct dentry *dmce, *ffake_panic;
2371 dmce = mce_get_debugfs_dir();
2374 ffake_panic = debugfs_create_file("fake_panic", 0444, dmce, NULL,
2382 static int __init mcheck_debugfs_init(void) { return -EINVAL; }
2385 DEFINE_STATIC_KEY_FALSE(mcsafe_key);
2386 EXPORT_SYMBOL_GPL(mcsafe_key);
2388 static int __init mcheck_late_init(void)
2390 if (mca_cfg.recovery)
2391 static_branch_inc(&mcsafe_key);
2393 mcheck_debugfs_init();
2397 * Flush out everything that has been logged during early boot, now that
2398 * everything has been initialized (workqueues, decoders, ...).
2400 mce_schedule_work();
2404 late_initcall(mcheck_late_init);