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/smp.h>
41 #include <linux/debugfs.h>
42 #include <linux/irq_work.h>
43 #include <linux/export.h>
45 #include <asm/processor.h>
46 #include <asm/traps.h>
47 #include <asm/tlbflush.h>
51 #include "mce-internal.h"
53 static DEFINE_MUTEX(mce_chrdev_read_mutex);
55 #define mce_log_get_idx_check(p) \
57 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
58 !lockdep_is_held(&mce_chrdev_read_mutex), \
59 "suspicious mce_log_get_idx_check() usage"); \
60 smp_load_acquire(&(p)); \
63 #define CREATE_TRACE_POINTS
64 #include <trace/events/mce.h>
66 #define SPINUNIT 100 /* 100ns */
68 DEFINE_PER_CPU(unsigned, mce_exception_count);
70 struct mce_bank *mce_banks __read_mostly;
71 struct mce_vendor_flags mce_flags __read_mostly;
73 struct mca_config mca_cfg __read_mostly = {
77 * 0: always panic on uncorrected errors, log corrected errors
78 * 1: panic or SIGBUS on uncorrected errors, log corrected errors
79 * 2: SIGBUS or log uncorrected errors (if possible), log corr. errors
80 * 3: never panic or SIGBUS, log all errors (for testing only)
86 /* User mode helper program triggered by machine check event */
87 static unsigned long mce_need_notify;
88 static char mce_helper[128];
89 static char *mce_helper_argv[2] = { mce_helper, NULL };
91 static DECLARE_WAIT_QUEUE_HEAD(mce_chrdev_wait);
93 static DEFINE_PER_CPU(struct mce, mces_seen);
94 static int cpu_missing;
97 * MCA banks polled by the period polling timer for corrected events.
98 * With Intel CMCI, this only has MCA banks which do not support CMCI (if any).
100 DEFINE_PER_CPU(mce_banks_t, mce_poll_banks) = {
101 [0 ... BITS_TO_LONGS(MAX_NR_BANKS)-1] = ~0UL
105 * MCA banks controlled through firmware first for corrected errors.
106 * This is a global list of banks for which we won't enable CMCI and we
107 * won't poll. Firmware controls these banks and is responsible for
108 * reporting corrected errors through GHES. Uncorrected/recoverable
109 * errors are still notified through a machine check.
111 mce_banks_t mce_banks_ce_disabled;
113 static struct work_struct mce_work;
114 static struct irq_work mce_irq_work;
116 static void (*quirk_no_way_out)(int bank, struct mce *m, struct pt_regs *regs);
119 * CPU/chipset specific EDAC code can register a notifier call here to print
120 * MCE errors in a human-readable form.
122 ATOMIC_NOTIFIER_HEAD(x86_mce_decoder_chain);
124 /* Do initial initialization of a struct mce */
125 void mce_setup(struct mce *m)
127 memset(m, 0, sizeof(struct mce));
128 m->cpu = m->extcpu = smp_processor_id();
130 /* We hope get_seconds stays lockless */
131 m->time = get_seconds();
132 m->cpuvendor = boot_cpu_data.x86_vendor;
133 m->cpuid = cpuid_eax(1);
134 m->socketid = cpu_data(m->extcpu).phys_proc_id;
135 m->apicid = cpu_data(m->extcpu).initial_apicid;
136 rdmsrl(MSR_IA32_MCG_CAP, m->mcgcap);
139 DEFINE_PER_CPU(struct mce, injectm);
140 EXPORT_PER_CPU_SYMBOL_GPL(injectm);
143 * Lockless MCE logging infrastructure.
144 * This avoids deadlocks on printk locks without having to break locks. Also
145 * separate MCEs from kernel messages to avoid bogus bug reports.
148 static struct mce_log mcelog = {
149 .signature = MCE_LOG_SIGNATURE,
151 .recordlen = sizeof(struct mce),
154 void mce_log(struct mce *mce)
156 unsigned next, entry;
158 /* Emit the trace record: */
159 trace_mce_record(mce);
161 if (!mce_gen_pool_add(mce))
162 irq_work_queue(&mce_irq_work);
167 entry = mce_log_get_idx_check(mcelog.next);
171 * When the buffer fills up discard new entries.
172 * Assume that the earlier errors are the more
175 if (entry >= MCE_LOG_LEN) {
176 set_bit(MCE_OVERFLOW,
177 (unsigned long *)&mcelog.flags);
180 /* Old left over entry. Skip: */
181 if (mcelog.entry[entry].finished) {
189 if (cmpxchg(&mcelog.next, entry, next) == entry)
192 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
194 mcelog.entry[entry].finished = 1;
198 set_bit(0, &mce_need_notify);
201 void mce_inject_log(struct mce *m)
203 mutex_lock(&mce_chrdev_read_mutex);
205 mutex_unlock(&mce_chrdev_read_mutex);
207 EXPORT_SYMBOL_GPL(mce_inject_log);
209 static struct notifier_block mce_srao_nb;
211 void mce_register_decode_chain(struct notifier_block *nb)
213 /* Ensure SRAO notifier has the highest priority in the decode chain. */
214 if (nb != &mce_srao_nb && nb->priority == INT_MAX)
217 atomic_notifier_chain_register(&x86_mce_decoder_chain, nb);
219 EXPORT_SYMBOL_GPL(mce_register_decode_chain);
221 void mce_unregister_decode_chain(struct notifier_block *nb)
223 atomic_notifier_chain_unregister(&x86_mce_decoder_chain, nb);
225 EXPORT_SYMBOL_GPL(mce_unregister_decode_chain);
227 static inline u32 ctl_reg(int bank)
229 return MSR_IA32_MCx_CTL(bank);
232 static inline u32 status_reg(int bank)
234 return MSR_IA32_MCx_STATUS(bank);
237 static inline u32 addr_reg(int bank)
239 return MSR_IA32_MCx_ADDR(bank);
242 static inline u32 misc_reg(int bank)
244 return MSR_IA32_MCx_MISC(bank);
247 static inline u32 smca_ctl_reg(int bank)
249 return MSR_AMD64_SMCA_MCx_CTL(bank);
252 static inline u32 smca_status_reg(int bank)
254 return MSR_AMD64_SMCA_MCx_STATUS(bank);
257 static inline u32 smca_addr_reg(int bank)
259 return MSR_AMD64_SMCA_MCx_ADDR(bank);
262 static inline u32 smca_misc_reg(int bank)
264 return MSR_AMD64_SMCA_MCx_MISC(bank);
267 struct mca_msr_regs msr_ops = {
269 .status = status_reg,
274 static void print_mce(struct mce *m)
278 pr_emerg(HW_ERR "CPU %d: Machine Check Exception: %Lx Bank %d: %016Lx\n",
279 m->extcpu, m->mcgstatus, m->bank, m->status);
282 pr_emerg(HW_ERR "RIP%s %02x:<%016Lx> ",
283 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
286 if (m->cs == __KERNEL_CS)
287 print_symbol("{%s}", m->ip);
291 pr_emerg(HW_ERR "TSC %llx ", m->tsc);
293 pr_cont("ADDR %llx ", m->addr);
295 pr_cont("MISC %llx ", m->misc);
299 * Note this output is parsed by external tools and old fields
300 * should not be changed.
302 pr_emerg(HW_ERR "PROCESSOR %u:%x TIME %llu SOCKET %u APIC %x microcode %x\n",
303 m->cpuvendor, m->cpuid, m->time, m->socketid, m->apicid,
304 cpu_data(m->extcpu).microcode);
307 * Print out human-readable details about the MCE error,
308 * (if the CPU has an implementation for that)
310 ret = atomic_notifier_call_chain(&x86_mce_decoder_chain, 0, m);
311 if (ret == NOTIFY_STOP)
314 pr_emerg_ratelimited(HW_ERR "Run the above through 'mcelog --ascii'\n");
317 #define PANIC_TIMEOUT 5 /* 5 seconds */
319 static atomic_t mce_panicked;
321 static int fake_panic;
322 static atomic_t mce_fake_panicked;
324 /* Panic in progress. Enable interrupts and wait for final IPI */
325 static void wait_for_panic(void)
327 long timeout = PANIC_TIMEOUT*USEC_PER_SEC;
331 while (timeout-- > 0)
333 if (panic_timeout == 0)
334 panic_timeout = mca_cfg.panic_timeout;
335 panic("Panicing machine check CPU died");
338 static void mce_panic(const char *msg, struct mce *final, char *exp)
344 * Make sure only one CPU runs in machine check panic
346 if (atomic_inc_return(&mce_panicked) > 1)
353 /* Don't log too much for fake panic */
354 if (atomic_inc_return(&mce_fake_panicked) > 1)
357 /* First print corrected ones that are still unlogged */
358 for (i = 0; i < MCE_LOG_LEN; i++) {
359 struct mce *m = &mcelog.entry[i];
360 if (!(m->status & MCI_STATUS_VAL))
362 if (!(m->status & MCI_STATUS_UC)) {
365 apei_err = apei_write_mce(m);
368 /* Now print uncorrected but with the final one last */
369 for (i = 0; i < MCE_LOG_LEN; i++) {
370 struct mce *m = &mcelog.entry[i];
371 if (!(m->status & MCI_STATUS_VAL))
373 if (!(m->status & MCI_STATUS_UC))
375 if (!final || memcmp(m, final, sizeof(struct mce))) {
378 apei_err = apei_write_mce(m);
384 apei_err = apei_write_mce(final);
387 pr_emerg(HW_ERR "Some CPUs didn't answer in synchronization\n");
389 pr_emerg(HW_ERR "Machine check: %s\n", exp);
391 if (panic_timeout == 0)
392 panic_timeout = mca_cfg.panic_timeout;
395 pr_emerg(HW_ERR "Fake kernel panic: %s\n", msg);
398 /* Support code for software error injection */
400 static int msr_to_offset(u32 msr)
402 unsigned bank = __this_cpu_read(injectm.bank);
404 if (msr == mca_cfg.rip_msr)
405 return offsetof(struct mce, ip);
406 if (msr == msr_ops.status(bank))
407 return offsetof(struct mce, status);
408 if (msr == msr_ops.addr(bank))
409 return offsetof(struct mce, addr);
410 if (msr == msr_ops.misc(bank))
411 return offsetof(struct mce, misc);
412 if (msr == MSR_IA32_MCG_STATUS)
413 return offsetof(struct mce, mcgstatus);
417 /* MSR access wrappers used for error injection */
418 static u64 mce_rdmsrl(u32 msr)
422 if (__this_cpu_read(injectm.finished)) {
423 int offset = msr_to_offset(msr);
427 return *(u64 *)((char *)this_cpu_ptr(&injectm) + offset);
430 if (rdmsrl_safe(msr, &v)) {
431 WARN_ONCE(1, "mce: Unable to read msr %d!\n", msr);
433 * Return zero in case the access faulted. This should
434 * not happen normally but can happen if the CPU does
435 * something weird, or if the code is buggy.
443 static void mce_wrmsrl(u32 msr, u64 v)
445 if (__this_cpu_read(injectm.finished)) {
446 int offset = msr_to_offset(msr);
449 *(u64 *)((char *)this_cpu_ptr(&injectm) + offset) = v;
456 * Collect all global (w.r.t. this processor) status about this machine
457 * check into our "mce" struct so that we can use it later to assess
458 * the severity of the problem as we read per-bank specific details.
460 static inline void mce_gather_info(struct mce *m, struct pt_regs *regs)
464 m->mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
467 * Get the address of the instruction at the time of
468 * the machine check error.
470 if (m->mcgstatus & (MCG_STATUS_RIPV|MCG_STATUS_EIPV)) {
475 * When in VM86 mode make the cs look like ring 3
476 * always. This is a lie, but it's better than passing
477 * the additional vm86 bit around everywhere.
479 if (v8086_mode(regs))
482 /* Use accurate RIP reporting if available. */
484 m->ip = mce_rdmsrl(mca_cfg.rip_msr);
488 int mce_available(struct cpuinfo_x86 *c)
490 if (mca_cfg.disabled)
492 return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
495 static void mce_schedule_work(void)
497 if (!mce_gen_pool_empty() && keventd_up())
498 schedule_work(&mce_work);
501 static void mce_irq_work_cb(struct irq_work *entry)
507 static void mce_report_event(struct pt_regs *regs)
509 if (regs->flags & (X86_VM_MASK|X86_EFLAGS_IF)) {
512 * Triggering the work queue here is just an insurance
513 * policy in case the syscall exit notify handler
514 * doesn't run soon enough or ends up running on the
515 * wrong CPU (can happen when audit sleeps)
521 irq_work_queue(&mce_irq_work);
525 * Check if the address reported by the CPU is in a format we can parse.
526 * It would be possible to add code for most other cases, but all would
527 * be somewhat complicated (e.g. segment offset would require an instruction
528 * parser). So only support physical addresses up to page granuality for now.
530 static int mce_usable_address(struct mce *m)
532 if (!(m->status & MCI_STATUS_MISCV) || !(m->status & MCI_STATUS_ADDRV))
535 /* Checks after this one are Intel-specific: */
536 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
539 if (MCI_MISC_ADDR_LSB(m->misc) > PAGE_SHIFT)
541 if (MCI_MISC_ADDR_MODE(m->misc) != MCI_MISC_ADDR_PHYS)
546 static int srao_decode_notifier(struct notifier_block *nb, unsigned long val,
549 struct mce *mce = (struct mce *)data;
555 if (mce_usable_address(mce) && (mce->severity == MCE_AO_SEVERITY)) {
556 pfn = mce->addr >> PAGE_SHIFT;
557 memory_failure(pfn, MCE_VECTOR, 0);
562 static struct notifier_block mce_srao_nb = {
563 .notifier_call = srao_decode_notifier,
568 * Read ADDR and MISC registers.
570 static void mce_read_aux(struct mce *m, int i)
572 if (m->status & MCI_STATUS_MISCV)
573 m->misc = mce_rdmsrl(msr_ops.misc(i));
574 if (m->status & MCI_STATUS_ADDRV) {
575 m->addr = mce_rdmsrl(msr_ops.addr(i));
578 * Mask the reported address by the reported granularity.
580 if (mca_cfg.ser && (m->status & MCI_STATUS_MISCV)) {
581 u8 shift = MCI_MISC_ADDR_LSB(m->misc);
588 static bool memory_error(struct mce *m)
590 struct cpuinfo_x86 *c = &boot_cpu_data;
592 if (c->x86_vendor == X86_VENDOR_AMD) {
593 /* ErrCodeExt[20:16] */
594 u8 xec = (m->status >> 16) & 0x1f;
596 return (xec == 0x0 || xec == 0x8);
597 } else if (c->x86_vendor == X86_VENDOR_INTEL) {
599 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
601 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
602 * indicating a memory error. Bit 8 is used for indicating a
603 * cache hierarchy error. The combination of bit 2 and bit 3
604 * is used for indicating a `generic' cache hierarchy error
605 * But we can't just blindly check the above bits, because if
606 * bit 11 is set, then it is a bus/interconnect error - and
607 * either way the above bits just gives more detail on what
608 * bus/interconnect error happened. Note that bit 12 can be
609 * ignored, as it's the "filter" bit.
611 return (m->status & 0xef80) == BIT(7) ||
612 (m->status & 0xef00) == BIT(8) ||
613 (m->status & 0xeffc) == 0xc;
619 DEFINE_PER_CPU(unsigned, mce_poll_count);
622 * Poll for corrected events or events that happened before reset.
623 * Those are just logged through /dev/mcelog.
625 * This is executed in standard interrupt context.
627 * Note: spec recommends to panic for fatal unsignalled
628 * errors here. However this would be quite problematic --
629 * we would need to reimplement the Monarch handling and
630 * it would mess up the exclusion between exception handler
631 * and poll hander -- * so we skip this for now.
632 * These cases should not happen anyways, or only when the CPU
633 * is already totally * confused. In this case it's likely it will
634 * not fully execute the machine check handler either.
636 bool machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
638 bool error_seen = false;
643 this_cpu_inc(mce_poll_count);
645 mce_gather_info(&m, NULL);
647 for (i = 0; i < mca_cfg.banks; i++) {
648 if (!mce_banks[i].ctl || !test_bit(i, *b))
657 m.status = mce_rdmsrl(msr_ops.status(i));
658 if (!(m.status & MCI_STATUS_VAL))
663 * Uncorrected or signalled events are handled by the exception
664 * handler when it is enabled, so don't process those here.
666 * TBD do the same check for MCI_STATUS_EN here?
668 if (!(flags & MCP_UC) &&
669 (m.status & (mca_cfg.ser ? MCI_STATUS_S : MCI_STATUS_UC)))
676 if (!(flags & MCP_TIMESTAMP))
679 severity = mce_severity(&m, mca_cfg.tolerant, NULL, false);
681 if (severity == MCE_DEFERRED_SEVERITY && memory_error(&m))
682 if (m.status & MCI_STATUS_ADDRV)
683 m.severity = severity;
686 * Don't get the IP here because it's unlikely to
687 * have anything to do with the actual error location.
689 if (!(flags & MCP_DONTLOG) && !mca_cfg.dont_log_ce)
691 else if (mce_usable_address(&m)) {
693 * Although we skipped logging this, we still want
694 * to take action. Add to the pool so the registered
695 * notifiers will see it.
697 if (!mce_gen_pool_add(&m))
702 * Clear state for this bank.
704 mce_wrmsrl(msr_ops.status(i), 0);
708 * Don't clear MCG_STATUS here because it's only defined for
716 EXPORT_SYMBOL_GPL(machine_check_poll);
719 * Do a quick check if any of the events requires a panic.
720 * This decides if we keep the events around or clear them.
722 static int mce_no_way_out(struct mce *m, char **msg, unsigned long *validp,
723 struct pt_regs *regs)
728 for (i = 0; i < mca_cfg.banks; i++) {
729 m->status = mce_rdmsrl(msr_ops.status(i));
730 if (m->status & MCI_STATUS_VAL) {
731 __set_bit(i, validp);
732 if (quirk_no_way_out)
733 quirk_no_way_out(i, m, regs);
736 if (mce_severity(m, mca_cfg.tolerant, &tmp, true) >= MCE_PANIC_SEVERITY) {
745 * Variable to establish order between CPUs while scanning.
746 * Each CPU spins initially until executing is equal its number.
748 static atomic_t mce_executing;
751 * Defines order of CPUs on entry. First CPU becomes Monarch.
753 static atomic_t mce_callin;
756 * Check if a timeout waiting for other CPUs happened.
758 static int mce_timed_out(u64 *t, const char *msg)
761 * The others already did panic for some reason.
762 * Bail out like in a timeout.
763 * rmb() to tell the compiler that system_state
764 * might have been modified by someone else.
767 if (atomic_read(&mce_panicked))
769 if (!mca_cfg.monarch_timeout)
771 if ((s64)*t < SPINUNIT) {
772 if (mca_cfg.tolerant <= 1)
773 mce_panic(msg, NULL, NULL);
779 touch_nmi_watchdog();
784 * The Monarch's reign. The Monarch is the CPU who entered
785 * the machine check handler first. It waits for the others to
786 * raise the exception too and then grades them. When any
787 * error is fatal panic. Only then let the others continue.
789 * The other CPUs entering the MCE handler will be controlled by the
790 * Monarch. They are called Subjects.
792 * This way we prevent any potential data corruption in a unrecoverable case
793 * and also makes sure always all CPU's errors are examined.
795 * Also this detects the case of a machine check event coming from outer
796 * space (not detected by any CPUs) In this case some external agent wants
797 * us to shut down, so panic too.
799 * The other CPUs might still decide to panic if the handler happens
800 * in a unrecoverable place, but in this case the system is in a semi-stable
801 * state and won't corrupt anything by itself. It's ok to let the others
802 * continue for a bit first.
804 * All the spin loops have timeouts; when a timeout happens a CPU
805 * typically elects itself to be Monarch.
807 static void mce_reign(void)
810 struct mce *m = NULL;
811 int global_worst = 0;
816 * This CPU is the Monarch and the other CPUs have run
817 * through their handlers.
818 * Grade the severity of the errors of all the CPUs.
820 for_each_possible_cpu(cpu) {
821 int severity = mce_severity(&per_cpu(mces_seen, cpu),
824 if (severity > global_worst) {
826 global_worst = severity;
827 m = &per_cpu(mces_seen, cpu);
832 * Cannot recover? Panic here then.
833 * This dumps all the mces in the log buffer and stops the
836 if (m && global_worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3)
837 mce_panic("Fatal machine check", m, msg);
840 * For UC somewhere we let the CPU who detects it handle it.
841 * Also must let continue the others, otherwise the handling
842 * CPU could deadlock on a lock.
846 * No machine check event found. Must be some external
847 * source or one CPU is hung. Panic.
849 if (global_worst <= MCE_KEEP_SEVERITY && mca_cfg.tolerant < 3)
850 mce_panic("Fatal machine check from unknown source", NULL, NULL);
853 * Now clear all the mces_seen so that they don't reappear on
856 for_each_possible_cpu(cpu)
857 memset(&per_cpu(mces_seen, cpu), 0, sizeof(struct mce));
860 static atomic_t global_nwo;
863 * Start of Monarch synchronization. This waits until all CPUs have
864 * entered the exception handler and then determines if any of them
865 * saw a fatal event that requires panic. Then it executes them
866 * in the entry order.
867 * TBD double check parallel CPU hotunplug
869 static int mce_start(int *no_way_out)
872 int cpus = num_online_cpus();
873 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
878 atomic_add(*no_way_out, &global_nwo);
880 * Rely on the implied barrier below, such that global_nwo
881 * is updated before mce_callin.
883 order = atomic_inc_return(&mce_callin);
888 while (atomic_read(&mce_callin) != cpus) {
889 if (mce_timed_out(&timeout,
890 "Timeout: Not all CPUs entered broadcast exception handler")) {
891 atomic_set(&global_nwo, 0);
898 * mce_callin should be read before global_nwo
904 * Monarch: Starts executing now, the others wait.
906 atomic_set(&mce_executing, 1);
909 * Subject: Now start the scanning loop one by one in
910 * the original callin order.
911 * This way when there are any shared banks it will be
912 * only seen by one CPU before cleared, avoiding duplicates.
914 while (atomic_read(&mce_executing) < order) {
915 if (mce_timed_out(&timeout,
916 "Timeout: Subject CPUs unable to finish machine check processing")) {
917 atomic_set(&global_nwo, 0);
925 * Cache the global no_way_out state.
927 *no_way_out = atomic_read(&global_nwo);
933 * Synchronize between CPUs after main scanning loop.
934 * This invokes the bulk of the Monarch processing.
936 static int mce_end(int order)
939 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
947 * Allow others to run.
949 atomic_inc(&mce_executing);
952 /* CHECKME: Can this race with a parallel hotplug? */
953 int cpus = num_online_cpus();
956 * Monarch: Wait for everyone to go through their scanning
959 while (atomic_read(&mce_executing) <= cpus) {
960 if (mce_timed_out(&timeout,
961 "Timeout: Monarch CPU unable to finish machine check processing"))
971 * Subject: Wait for Monarch to finish.
973 while (atomic_read(&mce_executing) != 0) {
974 if (mce_timed_out(&timeout,
975 "Timeout: Monarch CPU did not finish machine check processing"))
981 * Don't reset anything. That's done by the Monarch.
987 * Reset all global state.
990 atomic_set(&global_nwo, 0);
991 atomic_set(&mce_callin, 0);
995 * Let others run again.
997 atomic_set(&mce_executing, 0);
1001 static void mce_clear_state(unsigned long *toclear)
1005 for (i = 0; i < mca_cfg.banks; i++) {
1006 if (test_bit(i, toclear))
1007 mce_wrmsrl(msr_ops.status(i), 0);
1011 static int do_memory_failure(struct mce *m)
1013 int flags = MF_ACTION_REQUIRED;
1016 pr_err("Uncorrected hardware memory error in user-access at %llx", m->addr);
1017 if (!(m->mcgstatus & MCG_STATUS_RIPV))
1018 flags |= MF_MUST_KILL;
1019 ret = memory_failure(m->addr >> PAGE_SHIFT, MCE_VECTOR, flags);
1021 pr_err("Memory error not recovered");
1026 * The actual machine check handler. This only handles real
1027 * exceptions when something got corrupted coming in through int 18.
1029 * This is executed in NMI context not subject to normal locking rules. This
1030 * implies that most kernel services cannot be safely used. Don't even
1031 * think about putting a printk in there!
1033 * On Intel systems this is entered on all CPUs in parallel through
1034 * MCE broadcast. However some CPUs might be broken beyond repair,
1035 * so be always careful when synchronizing with others.
1037 void do_machine_check(struct pt_regs *regs, long error_code)
1039 struct mca_config *cfg = &mca_cfg;
1040 struct mce m, *final;
1045 * Establish sequential order between the CPUs entering the machine
1050 * If no_way_out gets set, there is no safe way to recover from this
1051 * MCE. If mca_cfg.tolerant is cranked up, we'll try anyway.
1055 * If kill_it gets set, there might be a way to recover from this
1059 DECLARE_BITMAP(toclear, MAX_NR_BANKS);
1060 DECLARE_BITMAP(valid_banks, MAX_NR_BANKS);
1061 char *msg = "Unknown";
1064 /* If this CPU is offline, just bail out. */
1065 if (cpu_is_offline(smp_processor_id())) {
1068 mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
1069 if (mcgstatus & MCG_STATUS_RIPV) {
1070 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1077 this_cpu_inc(mce_exception_count);
1082 mce_gather_info(&m, regs);
1084 final = this_cpu_ptr(&mces_seen);
1087 memset(valid_banks, 0, sizeof(valid_banks));
1088 no_way_out = mce_no_way_out(&m, &msg, valid_banks, regs);
1093 * When no restart IP might need to kill or panic.
1094 * Assume the worst for now, but if we find the
1095 * severity is MCE_AR_SEVERITY we have other options.
1097 if (!(m.mcgstatus & MCG_STATUS_RIPV))
1101 * Check if this MCE is signaled to only this logical processor
1103 if (m.mcgstatus & MCG_STATUS_LMCES)
1107 * Go through all the banks in exclusion of the other CPUs.
1108 * This way we don't report duplicated events on shared banks
1109 * because the first one to see it will clear it.
1110 * If this is a Local MCE, then no need to perform rendezvous.
1112 order = mce_start(&no_way_out);
1115 for (i = 0; i < cfg->banks; i++) {
1116 __clear_bit(i, toclear);
1117 if (!test_bit(i, valid_banks))
1119 if (!mce_banks[i].ctl)
1126 m.status = mce_rdmsrl(msr_ops.status(i));
1127 if ((m.status & MCI_STATUS_VAL) == 0)
1131 * Non uncorrected or non signaled errors are handled by
1132 * machine_check_poll. Leave them alone, unless this panics.
1134 if (!(m.status & (cfg->ser ? MCI_STATUS_S : MCI_STATUS_UC)) &&
1139 * Set taint even when machine check was not enabled.
1141 add_taint(TAINT_MACHINE_CHECK, LOCKDEP_NOW_UNRELIABLE);
1143 severity = mce_severity(&m, cfg->tolerant, NULL, true);
1146 * When machine check was for corrected/deferred handler don't
1147 * touch, unless we're panicing.
1149 if ((severity == MCE_KEEP_SEVERITY ||
1150 severity == MCE_UCNA_SEVERITY) && !no_way_out)
1152 __set_bit(i, toclear);
1153 if (severity == MCE_NO_SEVERITY) {
1155 * Machine check event was not enabled. Clear, but
1161 mce_read_aux(&m, i);
1163 /* assuming valid severity level != 0 */
1164 m.severity = severity;
1168 if (severity > worst) {
1174 /* mce_clear_state will clear *final, save locally for use later */
1178 mce_clear_state(toclear);
1181 * Do most of the synchronization with other CPUs.
1182 * When there's any problem use only local no_way_out state.
1185 if (mce_end(order) < 0)
1186 no_way_out = worst >= MCE_PANIC_SEVERITY;
1189 * Local MCE skipped calling mce_reign()
1190 * If we found a fatal error, we need to panic here.
1192 if (worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3)
1193 mce_panic("Machine check from unknown source",
1198 * If tolerant is at an insane level we drop requests to kill
1199 * processes and continue even when there is no way out.
1201 if (cfg->tolerant == 3)
1203 else if (no_way_out)
1204 mce_panic("Fatal machine check on current CPU", &m, msg);
1207 mce_report_event(regs);
1208 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1212 if (worst != MCE_AR_SEVERITY && !kill_it)
1215 /* Fault was in user mode and we need to take some action */
1216 if ((m.cs & 3) == 3) {
1217 ist_begin_non_atomic(regs);
1220 if (kill_it || do_memory_failure(&m))
1221 force_sig(SIGBUS, current);
1222 local_irq_disable();
1223 ist_end_non_atomic();
1225 if (!fixup_exception(regs, X86_TRAP_MC))
1226 mce_panic("Failed kernel mode recovery", &m, NULL);
1232 EXPORT_SYMBOL_GPL(do_machine_check);
1234 #ifndef CONFIG_MEMORY_FAILURE
1235 int memory_failure(unsigned long pfn, int vector, int flags)
1237 /* mce_severity() should not hand us an ACTION_REQUIRED error */
1238 BUG_ON(flags & MF_ACTION_REQUIRED);
1239 pr_err("Uncorrected memory error in page 0x%lx ignored\n"
1240 "Rebuild kernel with CONFIG_MEMORY_FAILURE=y for smarter handling\n",
1248 * Action optional processing happens here (picking up
1249 * from the list of faulting pages that do_machine_check()
1250 * placed into the genpool).
1252 static void mce_process_work(struct work_struct *dummy)
1254 mce_gen_pool_process();
1257 #ifdef CONFIG_X86_MCE_INTEL
1259 * mce_log_therm_throt_event - Logs the thermal throttling event to mcelog
1260 * @cpu: The CPU on which the event occurred.
1261 * @status: Event status information
1263 * This function should be called by the thermal interrupt after the
1264 * event has been processed and the decision was made to log the event
1267 * The status parameter will be saved to the 'status' field of 'struct mce'
1268 * and historically has been the register value of the
1269 * MSR_IA32_THERMAL_STATUS (Intel) msr.
1271 void mce_log_therm_throt_event(__u64 status)
1276 m.bank = MCE_THERMAL_BANK;
1280 #endif /* CONFIG_X86_MCE_INTEL */
1283 * Periodic polling timer for "silent" machine check errors. If the
1284 * poller finds an MCE, poll 2x faster. When the poller finds no more
1285 * errors, poll 2x slower (up to check_interval seconds).
1287 static unsigned long check_interval = INITIAL_CHECK_INTERVAL;
1289 static DEFINE_PER_CPU(unsigned long, mce_next_interval); /* in jiffies */
1290 static DEFINE_PER_CPU(struct timer_list, mce_timer);
1292 static unsigned long mce_adjust_timer_default(unsigned long interval)
1297 static unsigned long (*mce_adjust_timer)(unsigned long interval) = mce_adjust_timer_default;
1299 static void __restart_timer(struct timer_list *t, unsigned long interval)
1301 unsigned long when = jiffies + interval;
1302 unsigned long flags;
1304 local_irq_save(flags);
1306 if (timer_pending(t)) {
1307 if (time_before(when, t->expires))
1308 mod_timer_pinned(t, when);
1310 t->expires = round_jiffies(when);
1311 add_timer_on(t, smp_processor_id());
1314 local_irq_restore(flags);
1317 static void mce_timer_fn(unsigned long data)
1319 struct timer_list *t = this_cpu_ptr(&mce_timer);
1320 int cpu = smp_processor_id();
1323 WARN_ON(cpu != data);
1325 iv = __this_cpu_read(mce_next_interval);
1327 if (mce_available(this_cpu_ptr(&cpu_info))) {
1328 machine_check_poll(MCP_TIMESTAMP, this_cpu_ptr(&mce_poll_banks));
1330 if (mce_intel_cmci_poll()) {
1331 iv = mce_adjust_timer(iv);
1337 * Alert userspace if needed. If we logged an MCE, reduce the polling
1338 * interval, otherwise increase the polling interval.
1340 if (mce_notify_irq())
1341 iv = max(iv / 2, (unsigned long) HZ/100);
1343 iv = min(iv * 2, round_jiffies_relative(check_interval * HZ));
1346 __this_cpu_write(mce_next_interval, iv);
1347 __restart_timer(t, iv);
1351 * Ensure that the timer is firing in @interval from now.
1353 void mce_timer_kick(unsigned long interval)
1355 struct timer_list *t = this_cpu_ptr(&mce_timer);
1356 unsigned long iv = __this_cpu_read(mce_next_interval);
1358 __restart_timer(t, interval);
1361 __this_cpu_write(mce_next_interval, interval);
1364 /* Must not be called in IRQ context where del_timer_sync() can deadlock */
1365 static void mce_timer_delete_all(void)
1369 for_each_online_cpu(cpu)
1370 del_timer_sync(&per_cpu(mce_timer, cpu));
1373 static void mce_do_trigger(struct work_struct *work)
1375 call_usermodehelper(mce_helper, mce_helper_argv, NULL, UMH_NO_WAIT);
1378 static DECLARE_WORK(mce_trigger_work, mce_do_trigger);
1381 * Notify the user(s) about new machine check events.
1382 * Can be called from interrupt context, but not from machine check/NMI
1385 int mce_notify_irq(void)
1387 /* Not more than two messages every minute */
1388 static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2);
1390 if (test_and_clear_bit(0, &mce_need_notify)) {
1391 /* wake processes polling /dev/mcelog */
1392 wake_up_interruptible(&mce_chrdev_wait);
1395 schedule_work(&mce_trigger_work);
1397 if (__ratelimit(&ratelimit))
1398 pr_info(HW_ERR "Machine check events logged\n");
1404 EXPORT_SYMBOL_GPL(mce_notify_irq);
1406 static int __mcheck_cpu_mce_banks_init(void)
1409 u8 num_banks = mca_cfg.banks;
1411 mce_banks = kzalloc(num_banks * sizeof(struct mce_bank), GFP_KERNEL);
1415 for (i = 0; i < num_banks; i++) {
1416 struct mce_bank *b = &mce_banks[i];
1425 * Initialize Machine Checks for a CPU.
1427 static int __mcheck_cpu_cap_init(void)
1432 rdmsrl(MSR_IA32_MCG_CAP, cap);
1434 b = cap & MCG_BANKCNT_MASK;
1436 pr_info("CPU supports %d MCE banks\n", b);
1438 if (b > MAX_NR_BANKS) {
1439 pr_warn("Using only %u machine check banks out of %u\n",
1444 /* Don't support asymmetric configurations today */
1445 WARN_ON(mca_cfg.banks != 0 && b != mca_cfg.banks);
1449 int err = __mcheck_cpu_mce_banks_init();
1455 /* Use accurate RIP reporting if available. */
1456 if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
1457 mca_cfg.rip_msr = MSR_IA32_MCG_EIP;
1459 if (cap & MCG_SER_P)
1465 static void __mcheck_cpu_init_generic(void)
1467 enum mcp_flags m_fl = 0;
1468 mce_banks_t all_banks;
1471 if (!mca_cfg.bootlog)
1475 * Log the machine checks left over from the previous reset.
1477 bitmap_fill(all_banks, MAX_NR_BANKS);
1478 machine_check_poll(MCP_UC | m_fl, &all_banks);
1480 cr4_set_bits(X86_CR4_MCE);
1482 rdmsrl(MSR_IA32_MCG_CAP, cap);
1483 if (cap & MCG_CTL_P)
1484 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
1487 static void __mcheck_cpu_init_clear_banks(void)
1491 for (i = 0; i < mca_cfg.banks; i++) {
1492 struct mce_bank *b = &mce_banks[i];
1496 wrmsrl(msr_ops.ctl(i), b->ctl);
1497 wrmsrl(msr_ops.status(i), 0);
1502 * During IFU recovery Sandy Bridge -EP4S processors set the RIPV and
1503 * EIPV bits in MCG_STATUS to zero on the affected logical processor (SDM
1504 * Vol 3B Table 15-20). But this confuses both the code that determines
1505 * whether the machine check occurred in kernel or user mode, and also
1506 * the severity assessment code. Pretend that EIPV was set, and take the
1507 * ip/cs values from the pt_regs that mce_gather_info() ignored earlier.
1509 static void quirk_sandybridge_ifu(int bank, struct mce *m, struct pt_regs *regs)
1513 if ((m->mcgstatus & (MCG_STATUS_EIPV|MCG_STATUS_RIPV)) != 0)
1515 if ((m->status & (MCI_STATUS_OVER|MCI_STATUS_UC|
1516 MCI_STATUS_EN|MCI_STATUS_MISCV|MCI_STATUS_ADDRV|
1517 MCI_STATUS_PCC|MCI_STATUS_S|MCI_STATUS_AR|
1519 (MCI_STATUS_UC|MCI_STATUS_EN|
1520 MCI_STATUS_MISCV|MCI_STATUS_ADDRV|MCI_STATUS_S|
1521 MCI_STATUS_AR|MCACOD_INSTR))
1524 m->mcgstatus |= MCG_STATUS_EIPV;
1529 /* Add per CPU specific workarounds here */
1530 static int __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
1532 struct mca_config *cfg = &mca_cfg;
1534 if (c->x86_vendor == X86_VENDOR_UNKNOWN) {
1535 pr_info("unknown CPU type - not enabling MCE support\n");
1539 /* This should be disabled by the BIOS, but isn't always */
1540 if (c->x86_vendor == X86_VENDOR_AMD) {
1541 if (c->x86 == 15 && cfg->banks > 4) {
1543 * disable GART TBL walk error reporting, which
1544 * trips off incorrectly with the IOMMU & 3ware
1547 clear_bit(10, (unsigned long *)&mce_banks[4].ctl);
1549 if (c->x86 < 17 && cfg->bootlog < 0) {
1551 * Lots of broken BIOS around that don't clear them
1552 * by default and leave crap in there. Don't log:
1557 * Various K7s with broken bank 0 around. Always disable
1560 if (c->x86 == 6 && cfg->banks > 0)
1561 mce_banks[0].ctl = 0;
1564 * overflow_recov is supported for F15h Models 00h-0fh
1565 * even though we don't have a CPUID bit for it.
1567 if (c->x86 == 0x15 && c->x86_model <= 0xf)
1568 mce_flags.overflow_recov = 1;
1571 * Turn off MC4_MISC thresholding banks on those models since
1572 * they're not supported there.
1574 if (c->x86 == 0x15 &&
1575 (c->x86_model >= 0x10 && c->x86_model <= 0x1f)) {
1580 0x00000413, /* MC4_MISC0 */
1581 0xc0000408, /* MC4_MISC1 */
1584 rdmsrl(MSR_K7_HWCR, hwcr);
1586 /* McStatusWrEn has to be set */
1587 need_toggle = !(hwcr & BIT(18));
1590 wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
1592 /* Clear CntP bit safely */
1593 for (i = 0; i < ARRAY_SIZE(msrs); i++)
1594 msr_clear_bit(msrs[i], 62);
1596 /* restore old settings */
1598 wrmsrl(MSR_K7_HWCR, hwcr);
1602 if (c->x86_vendor == X86_VENDOR_INTEL) {
1604 * SDM documents that on family 6 bank 0 should not be written
1605 * because it aliases to another special BIOS controlled
1607 * But it's not aliased anymore on model 0x1a+
1608 * Don't ignore bank 0 completely because there could be a
1609 * valid event later, merely don't write CTL0.
1612 if (c->x86 == 6 && c->x86_model < 0x1A && cfg->banks > 0)
1613 mce_banks[0].init = 0;
1616 * All newer Intel systems support MCE broadcasting. Enable
1617 * synchronization with a one second timeout.
1619 if ((c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xe)) &&
1620 cfg->monarch_timeout < 0)
1621 cfg->monarch_timeout = USEC_PER_SEC;
1624 * There are also broken BIOSes on some Pentium M and
1627 if (c->x86 == 6 && c->x86_model <= 13 && cfg->bootlog < 0)
1630 if (c->x86 == 6 && c->x86_model == 45)
1631 quirk_no_way_out = quirk_sandybridge_ifu;
1633 * MCG_CAP.MCG_SER_P is necessary but not sufficient to know
1634 * whether this processor will actually generate recoverable
1635 * machine checks. Check to see if this is an E7 model Xeon.
1636 * We can't do a model number check because E5 and E7 use the
1637 * same model number. E5 doesn't support recovery, E7 does.
1639 if (mca_cfg.recovery || (mca_cfg.ser &&
1640 !strncmp(c->x86_model_id,
1641 "Intel(R) Xeon(R) CPU E7-", 24)))
1642 set_cpu_cap(c, X86_FEATURE_MCE_RECOVERY);
1644 if (cfg->monarch_timeout < 0)
1645 cfg->monarch_timeout = 0;
1646 if (cfg->bootlog != 0)
1647 cfg->panic_timeout = 30;
1652 static int __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c)
1657 switch (c->x86_vendor) {
1658 case X86_VENDOR_INTEL:
1659 intel_p5_mcheck_init(c);
1662 case X86_VENDOR_CENTAUR:
1663 winchip_mcheck_init(c);
1673 static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c)
1675 switch (c->x86_vendor) {
1676 case X86_VENDOR_INTEL:
1677 mce_intel_feature_init(c);
1678 mce_adjust_timer = cmci_intel_adjust_timer;
1681 case X86_VENDOR_AMD: {
1682 u32 ebx = cpuid_ebx(0x80000007);
1684 mce_flags.overflow_recov = !!(ebx & BIT(0));
1685 mce_flags.succor = !!(ebx & BIT(1));
1686 mce_flags.smca = !!(ebx & BIT(3));
1689 * Install proper ops for Scalable MCA enabled processors
1691 if (mce_flags.smca) {
1692 msr_ops.ctl = smca_ctl_reg;
1693 msr_ops.status = smca_status_reg;
1694 msr_ops.addr = smca_addr_reg;
1695 msr_ops.misc = smca_misc_reg;
1697 mce_amd_feature_init(c);
1707 static void __mcheck_cpu_clear_vendor(struct cpuinfo_x86 *c)
1709 switch (c->x86_vendor) {
1710 case X86_VENDOR_INTEL:
1711 mce_intel_feature_clear(c);
1718 static void mce_start_timer(unsigned int cpu, struct timer_list *t)
1720 unsigned long iv = check_interval * HZ;
1722 if (mca_cfg.ignore_ce || !iv)
1725 per_cpu(mce_next_interval, cpu) = iv;
1727 t->expires = round_jiffies(jiffies + iv);
1728 add_timer_on(t, cpu);
1731 static void __mcheck_cpu_init_timer(void)
1733 struct timer_list *t = this_cpu_ptr(&mce_timer);
1734 unsigned int cpu = smp_processor_id();
1736 setup_timer(t, mce_timer_fn, cpu);
1737 mce_start_timer(cpu, t);
1740 /* Handle unconfigured int18 (should never happen) */
1741 static void unexpected_machine_check(struct pt_regs *regs, long error_code)
1743 pr_err("CPU#%d: Unexpected int18 (Machine Check)\n",
1744 smp_processor_id());
1747 /* Call the installed machine check handler for this CPU setup. */
1748 void (*machine_check_vector)(struct pt_regs *, long error_code) =
1749 unexpected_machine_check;
1752 * Called for each booted CPU to set up machine checks.
1753 * Must be called with preempt off:
1755 void mcheck_cpu_init(struct cpuinfo_x86 *c)
1757 if (mca_cfg.disabled)
1760 if (__mcheck_cpu_ancient_init(c))
1763 if (!mce_available(c))
1766 if (__mcheck_cpu_cap_init() < 0 || __mcheck_cpu_apply_quirks(c) < 0) {
1767 mca_cfg.disabled = true;
1771 if (mce_gen_pool_init()) {
1772 mca_cfg.disabled = true;
1773 pr_emerg("Couldn't allocate MCE records pool!\n");
1777 machine_check_vector = do_machine_check;
1779 __mcheck_cpu_init_generic();
1780 __mcheck_cpu_init_vendor(c);
1781 __mcheck_cpu_init_clear_banks();
1782 __mcheck_cpu_init_timer();
1786 * Called for each booted CPU to clear some machine checks opt-ins
1788 void mcheck_cpu_clear(struct cpuinfo_x86 *c)
1790 if (mca_cfg.disabled)
1793 if (!mce_available(c))
1797 * Possibly to clear general settings generic to x86
1798 * __mcheck_cpu_clear_generic(c);
1800 __mcheck_cpu_clear_vendor(c);
1805 * mce_chrdev: Character device /dev/mcelog to read and clear the MCE log.
1808 static DEFINE_SPINLOCK(mce_chrdev_state_lock);
1809 static int mce_chrdev_open_count; /* #times opened */
1810 static int mce_chrdev_open_exclu; /* already open exclusive? */
1812 static int mce_chrdev_open(struct inode *inode, struct file *file)
1814 spin_lock(&mce_chrdev_state_lock);
1816 if (mce_chrdev_open_exclu ||
1817 (mce_chrdev_open_count && (file->f_flags & O_EXCL))) {
1818 spin_unlock(&mce_chrdev_state_lock);
1823 if (file->f_flags & O_EXCL)
1824 mce_chrdev_open_exclu = 1;
1825 mce_chrdev_open_count++;
1827 spin_unlock(&mce_chrdev_state_lock);
1829 return nonseekable_open(inode, file);
1832 static int mce_chrdev_release(struct inode *inode, struct file *file)
1834 spin_lock(&mce_chrdev_state_lock);
1836 mce_chrdev_open_count--;
1837 mce_chrdev_open_exclu = 0;
1839 spin_unlock(&mce_chrdev_state_lock);
1844 static void collect_tscs(void *data)
1846 unsigned long *cpu_tsc = (unsigned long *)data;
1848 cpu_tsc[smp_processor_id()] = rdtsc();
1851 static int mce_apei_read_done;
1853 /* Collect MCE record of previous boot in persistent storage via APEI ERST. */
1854 static int __mce_read_apei(char __user **ubuf, size_t usize)
1860 if (usize < sizeof(struct mce))
1863 rc = apei_read_mce(&m, &record_id);
1864 /* Error or no more MCE record */
1866 mce_apei_read_done = 1;
1868 * When ERST is disabled, mce_chrdev_read() should return
1869 * "no record" instead of "no device."
1876 if (copy_to_user(*ubuf, &m, sizeof(struct mce)))
1879 * In fact, we should have cleared the record after that has
1880 * been flushed to the disk or sent to network in
1881 * /sbin/mcelog, but we have no interface to support that now,
1882 * so just clear it to avoid duplication.
1884 rc = apei_clear_mce(record_id);
1886 mce_apei_read_done = 1;
1889 *ubuf += sizeof(struct mce);
1894 static ssize_t mce_chrdev_read(struct file *filp, char __user *ubuf,
1895 size_t usize, loff_t *off)
1897 char __user *buf = ubuf;
1898 unsigned long *cpu_tsc;
1899 unsigned prev, next;
1902 cpu_tsc = kmalloc(nr_cpu_ids * sizeof(long), GFP_KERNEL);
1906 mutex_lock(&mce_chrdev_read_mutex);
1908 if (!mce_apei_read_done) {
1909 err = __mce_read_apei(&buf, usize);
1910 if (err || buf != ubuf)
1914 next = mce_log_get_idx_check(mcelog.next);
1916 /* Only supports full reads right now */
1918 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce))
1924 for (i = prev; i < next; i++) {
1925 unsigned long start = jiffies;
1926 struct mce *m = &mcelog.entry[i];
1928 while (!m->finished) {
1929 if (time_after_eq(jiffies, start + 2)) {
1930 memset(m, 0, sizeof(*m));
1936 err |= copy_to_user(buf, m, sizeof(*m));
1942 memset(mcelog.entry + prev, 0,
1943 (next - prev) * sizeof(struct mce));
1945 next = cmpxchg(&mcelog.next, prev, 0);
1946 } while (next != prev);
1948 synchronize_sched();
1951 * Collect entries that were still getting written before the
1954 on_each_cpu(collect_tscs, cpu_tsc, 1);
1956 for (i = next; i < MCE_LOG_LEN; i++) {
1957 struct mce *m = &mcelog.entry[i];
1959 if (m->finished && m->tsc < cpu_tsc[m->cpu]) {
1960 err |= copy_to_user(buf, m, sizeof(*m));
1963 memset(m, 0, sizeof(*m));
1971 mutex_unlock(&mce_chrdev_read_mutex);
1974 return err ? err : buf - ubuf;
1977 static unsigned int mce_chrdev_poll(struct file *file, poll_table *wait)
1979 poll_wait(file, &mce_chrdev_wait, wait);
1980 if (READ_ONCE(mcelog.next))
1981 return POLLIN | POLLRDNORM;
1982 if (!mce_apei_read_done && apei_check_mce())
1983 return POLLIN | POLLRDNORM;
1987 static long mce_chrdev_ioctl(struct file *f, unsigned int cmd,
1990 int __user *p = (int __user *)arg;
1992 if (!capable(CAP_SYS_ADMIN))
1996 case MCE_GET_RECORD_LEN:
1997 return put_user(sizeof(struct mce), p);
1998 case MCE_GET_LOG_LEN:
1999 return put_user(MCE_LOG_LEN, p);
2000 case MCE_GETCLEAR_FLAGS: {
2004 flags = mcelog.flags;
2005 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
2007 return put_user(flags, p);
2014 static ssize_t (*mce_write)(struct file *filp, const char __user *ubuf,
2015 size_t usize, loff_t *off);
2017 void register_mce_write_callback(ssize_t (*fn)(struct file *filp,
2018 const char __user *ubuf,
2019 size_t usize, loff_t *off))
2023 EXPORT_SYMBOL_GPL(register_mce_write_callback);
2025 static ssize_t mce_chrdev_write(struct file *filp, const char __user *ubuf,
2026 size_t usize, loff_t *off)
2029 return mce_write(filp, ubuf, usize, off);
2034 static const struct file_operations mce_chrdev_ops = {
2035 .open = mce_chrdev_open,
2036 .release = mce_chrdev_release,
2037 .read = mce_chrdev_read,
2038 .write = mce_chrdev_write,
2039 .poll = mce_chrdev_poll,
2040 .unlocked_ioctl = mce_chrdev_ioctl,
2041 .llseek = no_llseek,
2044 static struct miscdevice mce_chrdev_device = {
2050 static void __mce_disable_bank(void *arg)
2052 int bank = *((int *)arg);
2053 __clear_bit(bank, this_cpu_ptr(mce_poll_banks));
2054 cmci_disable_bank(bank);
2057 void mce_disable_bank(int bank)
2059 if (bank >= mca_cfg.banks) {
2061 "Ignoring request to disable invalid MCA bank %d.\n",
2065 set_bit(bank, mce_banks_ce_disabled);
2066 on_each_cpu(__mce_disable_bank, &bank, 1);
2070 * mce=off Disables machine check
2071 * mce=no_cmci Disables CMCI
2072 * mce=no_lmce Disables LMCE
2073 * mce=dont_log_ce Clears corrected events silently, no log created for CEs.
2074 * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared.
2075 * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above)
2076 * monarchtimeout is how long to wait for other CPUs on machine
2077 * check, or 0 to not wait
2078 * mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
2079 * mce=nobootlog Don't log MCEs from before booting.
2080 * mce=bios_cmci_threshold Don't program the CMCI threshold
2082 static int __init mcheck_enable(char *str)
2084 struct mca_config *cfg = &mca_cfg;
2092 if (!strcmp(str, "off"))
2093 cfg->disabled = true;
2094 else if (!strcmp(str, "no_cmci"))
2095 cfg->cmci_disabled = true;
2096 else if (!strcmp(str, "no_lmce"))
2097 cfg->lmce_disabled = true;
2098 else if (!strcmp(str, "dont_log_ce"))
2099 cfg->dont_log_ce = true;
2100 else if (!strcmp(str, "ignore_ce"))
2101 cfg->ignore_ce = true;
2102 else if (!strcmp(str, "bootlog") || !strcmp(str, "nobootlog"))
2103 cfg->bootlog = (str[0] == 'b');
2104 else if (!strcmp(str, "bios_cmci_threshold"))
2105 cfg->bios_cmci_threshold = true;
2106 else if (!strcmp(str, "recovery"))
2107 cfg->recovery = true;
2108 else if (isdigit(str[0])) {
2109 if (get_option(&str, &cfg->tolerant) == 2)
2110 get_option(&str, &(cfg->monarch_timeout));
2112 pr_info("mce argument %s ignored. Please use /sys\n", str);
2117 __setup("mce", mcheck_enable);
2119 int __init mcheck_init(void)
2121 mcheck_intel_therm_init();
2122 mce_register_decode_chain(&mce_srao_nb);
2123 mcheck_vendor_init_severity();
2125 INIT_WORK(&mce_work, mce_process_work);
2126 init_irq_work(&mce_irq_work, mce_irq_work_cb);
2132 * mce_syscore: PM support
2136 * Disable machine checks on suspend and shutdown. We can't really handle
2139 static void mce_disable_error_reporting(void)
2143 for (i = 0; i < mca_cfg.banks; i++) {
2144 struct mce_bank *b = &mce_banks[i];
2147 wrmsrl(msr_ops.ctl(i), 0);
2152 static void vendor_disable_error_reporting(void)
2155 * Don't clear on Intel CPUs. Some of these MSRs are socket-wide.
2156 * Disabling them for just a single offlined CPU is bad, since it will
2157 * inhibit reporting for all shared resources on the socket like the
2158 * last level cache (LLC), the integrated memory controller (iMC), etc.
2160 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
2163 mce_disable_error_reporting();
2166 static int mce_syscore_suspend(void)
2168 vendor_disable_error_reporting();
2172 static void mce_syscore_shutdown(void)
2174 vendor_disable_error_reporting();
2178 * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
2179 * Only one CPU is active at this time, the others get re-added later using
2182 static void mce_syscore_resume(void)
2184 __mcheck_cpu_init_generic();
2185 __mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info));
2186 __mcheck_cpu_init_clear_banks();
2189 static struct syscore_ops mce_syscore_ops = {
2190 .suspend = mce_syscore_suspend,
2191 .shutdown = mce_syscore_shutdown,
2192 .resume = mce_syscore_resume,
2196 * mce_device: Sysfs support
2199 static void mce_cpu_restart(void *data)
2201 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2203 __mcheck_cpu_init_generic();
2204 __mcheck_cpu_init_clear_banks();
2205 __mcheck_cpu_init_timer();
2208 /* Reinit MCEs after user configuration changes */
2209 static void mce_restart(void)
2211 mce_timer_delete_all();
2212 on_each_cpu(mce_cpu_restart, NULL, 1);
2215 /* Toggle features for corrected errors */
2216 static void mce_disable_cmci(void *data)
2218 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2223 static void mce_enable_ce(void *all)
2225 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2230 __mcheck_cpu_init_timer();
2233 static struct bus_type mce_subsys = {
2234 .name = "machinecheck",
2235 .dev_name = "machinecheck",
2238 DEFINE_PER_CPU(struct device *, mce_device);
2240 void (*threshold_cpu_callback)(unsigned long action, unsigned int cpu);
2242 static inline struct mce_bank *attr_to_bank(struct device_attribute *attr)
2244 return container_of(attr, struct mce_bank, attr);
2247 static ssize_t show_bank(struct device *s, struct device_attribute *attr,
2250 return sprintf(buf, "%llx\n", attr_to_bank(attr)->ctl);
2253 static ssize_t set_bank(struct device *s, struct device_attribute *attr,
2254 const char *buf, size_t size)
2258 if (kstrtou64(buf, 0, &new) < 0)
2261 attr_to_bank(attr)->ctl = new;
2268 show_trigger(struct device *s, struct device_attribute *attr, char *buf)
2270 strcpy(buf, mce_helper);
2272 return strlen(mce_helper) + 1;
2275 static ssize_t set_trigger(struct device *s, struct device_attribute *attr,
2276 const char *buf, size_t siz)
2280 strncpy(mce_helper, buf, sizeof(mce_helper));
2281 mce_helper[sizeof(mce_helper)-1] = 0;
2282 p = strchr(mce_helper, '\n');
2287 return strlen(mce_helper) + !!p;
2290 static ssize_t set_ignore_ce(struct device *s,
2291 struct device_attribute *attr,
2292 const char *buf, size_t size)
2296 if (kstrtou64(buf, 0, &new) < 0)
2299 if (mca_cfg.ignore_ce ^ !!new) {
2301 /* disable ce features */
2302 mce_timer_delete_all();
2303 on_each_cpu(mce_disable_cmci, NULL, 1);
2304 mca_cfg.ignore_ce = true;
2306 /* enable ce features */
2307 mca_cfg.ignore_ce = false;
2308 on_each_cpu(mce_enable_ce, (void *)1, 1);
2314 static ssize_t set_cmci_disabled(struct device *s,
2315 struct device_attribute *attr,
2316 const char *buf, size_t size)
2320 if (kstrtou64(buf, 0, &new) < 0)
2323 if (mca_cfg.cmci_disabled ^ !!new) {
2326 on_each_cpu(mce_disable_cmci, NULL, 1);
2327 mca_cfg.cmci_disabled = true;
2330 mca_cfg.cmci_disabled = false;
2331 on_each_cpu(mce_enable_ce, NULL, 1);
2337 static ssize_t store_int_with_restart(struct device *s,
2338 struct device_attribute *attr,
2339 const char *buf, size_t size)
2341 ssize_t ret = device_store_int(s, attr, buf, size);
2346 static DEVICE_ATTR(trigger, 0644, show_trigger, set_trigger);
2347 static DEVICE_INT_ATTR(tolerant, 0644, mca_cfg.tolerant);
2348 static DEVICE_INT_ATTR(monarch_timeout, 0644, mca_cfg.monarch_timeout);
2349 static DEVICE_BOOL_ATTR(dont_log_ce, 0644, mca_cfg.dont_log_ce);
2351 static struct dev_ext_attribute dev_attr_check_interval = {
2352 __ATTR(check_interval, 0644, device_show_int, store_int_with_restart),
2356 static struct dev_ext_attribute dev_attr_ignore_ce = {
2357 __ATTR(ignore_ce, 0644, device_show_bool, set_ignore_ce),
2361 static struct dev_ext_attribute dev_attr_cmci_disabled = {
2362 __ATTR(cmci_disabled, 0644, device_show_bool, set_cmci_disabled),
2363 &mca_cfg.cmci_disabled
2366 static struct device_attribute *mce_device_attrs[] = {
2367 &dev_attr_tolerant.attr,
2368 &dev_attr_check_interval.attr,
2370 &dev_attr_monarch_timeout.attr,
2371 &dev_attr_dont_log_ce.attr,
2372 &dev_attr_ignore_ce.attr,
2373 &dev_attr_cmci_disabled.attr,
2377 static cpumask_var_t mce_device_initialized;
2379 static void mce_device_release(struct device *dev)
2384 /* Per cpu device init. All of the cpus still share the same ctrl bank: */
2385 static int mce_device_create(unsigned int cpu)
2391 if (!mce_available(&boot_cpu_data))
2394 dev = kzalloc(sizeof *dev, GFP_KERNEL);
2398 dev->bus = &mce_subsys;
2399 dev->release = &mce_device_release;
2401 err = device_register(dev);
2407 for (i = 0; mce_device_attrs[i]; i++) {
2408 err = device_create_file(dev, mce_device_attrs[i]);
2412 for (j = 0; j < mca_cfg.banks; j++) {
2413 err = device_create_file(dev, &mce_banks[j].attr);
2417 cpumask_set_cpu(cpu, mce_device_initialized);
2418 per_cpu(mce_device, cpu) = dev;
2423 device_remove_file(dev, &mce_banks[j].attr);
2426 device_remove_file(dev, mce_device_attrs[i]);
2428 device_unregister(dev);
2433 static void mce_device_remove(unsigned int cpu)
2435 struct device *dev = per_cpu(mce_device, cpu);
2438 if (!cpumask_test_cpu(cpu, mce_device_initialized))
2441 for (i = 0; mce_device_attrs[i]; i++)
2442 device_remove_file(dev, mce_device_attrs[i]);
2444 for (i = 0; i < mca_cfg.banks; i++)
2445 device_remove_file(dev, &mce_banks[i].attr);
2447 device_unregister(dev);
2448 cpumask_clear_cpu(cpu, mce_device_initialized);
2449 per_cpu(mce_device, cpu) = NULL;
2452 /* Make sure there are no machine checks on offlined CPUs. */
2453 static void mce_disable_cpu(void *h)
2455 unsigned long action = *(unsigned long *)h;
2457 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2460 if (!(action & CPU_TASKS_FROZEN))
2463 vendor_disable_error_reporting();
2466 static void mce_reenable_cpu(void *h)
2468 unsigned long action = *(unsigned long *)h;
2471 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2474 if (!(action & CPU_TASKS_FROZEN))
2476 for (i = 0; i < mca_cfg.banks; i++) {
2477 struct mce_bank *b = &mce_banks[i];
2480 wrmsrl(msr_ops.ctl(i), b->ctl);
2484 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
2486 mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
2488 unsigned int cpu = (unsigned long)hcpu;
2489 struct timer_list *t = &per_cpu(mce_timer, cpu);
2491 switch (action & ~CPU_TASKS_FROZEN) {
2493 mce_device_create(cpu);
2494 if (threshold_cpu_callback)
2495 threshold_cpu_callback(action, cpu);
2498 if (threshold_cpu_callback)
2499 threshold_cpu_callback(action, cpu);
2500 mce_device_remove(cpu);
2501 mce_intel_hcpu_update(cpu);
2503 /* intentionally ignoring frozen here */
2504 if (!(action & CPU_TASKS_FROZEN))
2507 case CPU_DOWN_PREPARE:
2508 smp_call_function_single(cpu, mce_disable_cpu, &action, 1);
2511 case CPU_DOWN_FAILED:
2512 smp_call_function_single(cpu, mce_reenable_cpu, &action, 1);
2513 mce_start_timer(cpu, t);
2520 static struct notifier_block mce_cpu_notifier = {
2521 .notifier_call = mce_cpu_callback,
2524 static __init void mce_init_banks(void)
2528 for (i = 0; i < mca_cfg.banks; i++) {
2529 struct mce_bank *b = &mce_banks[i];
2530 struct device_attribute *a = &b->attr;
2532 sysfs_attr_init(&a->attr);
2533 a->attr.name = b->attrname;
2534 snprintf(b->attrname, ATTR_LEN, "bank%d", i);
2536 a->attr.mode = 0644;
2537 a->show = show_bank;
2538 a->store = set_bank;
2542 static __init int mcheck_init_device(void)
2547 if (!mce_available(&boot_cpu_data)) {
2552 if (!zalloc_cpumask_var(&mce_device_initialized, GFP_KERNEL)) {
2559 err = subsys_system_register(&mce_subsys, NULL);
2563 cpu_notifier_register_begin();
2564 for_each_online_cpu(i) {
2565 err = mce_device_create(i);
2568 * Register notifier anyway (and do not unreg it) so
2569 * that we don't leave undeleted timers, see notifier
2572 __register_hotcpu_notifier(&mce_cpu_notifier);
2573 cpu_notifier_register_done();
2574 goto err_device_create;
2578 __register_hotcpu_notifier(&mce_cpu_notifier);
2579 cpu_notifier_register_done();
2581 register_syscore_ops(&mce_syscore_ops);
2583 /* register character device /dev/mcelog */
2584 err = misc_register(&mce_chrdev_device);
2591 unregister_syscore_ops(&mce_syscore_ops);
2595 * We didn't keep track of which devices were created above, but
2596 * even if we had, the set of online cpus might have changed.
2597 * Play safe and remove for every possible cpu, since
2598 * mce_device_remove() will do the right thing.
2600 for_each_possible_cpu(i)
2601 mce_device_remove(i);
2604 free_cpumask_var(mce_device_initialized);
2607 pr_err("Unable to init device /dev/mcelog (rc: %d)\n", err);
2611 device_initcall_sync(mcheck_init_device);
2614 * Old style boot options parsing. Only for compatibility.
2616 static int __init mcheck_disable(char *str)
2618 mca_cfg.disabled = true;
2621 __setup("nomce", mcheck_disable);
2623 #ifdef CONFIG_DEBUG_FS
2624 struct dentry *mce_get_debugfs_dir(void)
2626 static struct dentry *dmce;
2629 dmce = debugfs_create_dir("mce", NULL);
2634 static void mce_reset(void)
2637 atomic_set(&mce_fake_panicked, 0);
2638 atomic_set(&mce_executing, 0);
2639 atomic_set(&mce_callin, 0);
2640 atomic_set(&global_nwo, 0);
2643 static int fake_panic_get(void *data, u64 *val)
2649 static int fake_panic_set(void *data, u64 val)
2656 DEFINE_SIMPLE_ATTRIBUTE(fake_panic_fops, fake_panic_get,
2657 fake_panic_set, "%llu\n");
2659 static int __init mcheck_debugfs_init(void)
2661 struct dentry *dmce, *ffake_panic;
2663 dmce = mce_get_debugfs_dir();
2666 ffake_panic = debugfs_create_file("fake_panic", 0444, dmce, NULL,
2674 static int __init mcheck_debugfs_init(void) { return -EINVAL; }
2677 static int __init mcheck_late_init(void)
2679 mcheck_debugfs_init();
2682 * Flush out everything that has been logged during early boot, now that
2683 * everything has been initialized (workqueues, decoders, ...).
2685 mce_schedule_work();
2689 late_initcall(mcheck_late_init);