1 // SPDX-License-Identifier: GPL-2.0-only
3 * Machine check handler.
5 * K8 parts Copyright 2002,2003 Andi Kleen, SuSE Labs.
6 * Rest from unknown author(s).
7 * 2004 Andi Kleen. Rewrote most of it.
8 * Copyright 2008 Intel Corporation
12 #include <linux/thread_info.h>
13 #include <linux/capability.h>
14 #include <linux/miscdevice.h>
15 #include <linux/ratelimit.h>
16 #include <linux/rcupdate.h>
17 #include <linux/kobject.h>
18 #include <linux/uaccess.h>
19 #include <linux/kdebug.h>
20 #include <linux/kernel.h>
21 #include <linux/percpu.h>
22 #include <linux/string.h>
23 #include <linux/device.h>
24 #include <linux/syscore_ops.h>
25 #include <linux/delay.h>
26 #include <linux/ctype.h>
27 #include <linux/sched.h>
28 #include <linux/sysfs.h>
29 #include <linux/types.h>
30 #include <linux/slab.h>
31 #include <linux/init.h>
32 #include <linux/kmod.h>
33 #include <linux/poll.h>
34 #include <linux/nmi.h>
35 #include <linux/cpu.h>
36 #include <linux/ras.h>
37 #include <linux/smp.h>
40 #include <linux/debugfs.h>
41 #include <linux/irq_work.h>
42 #include <linux/export.h>
43 #include <linux/jump_label.h>
44 #include <linux/set_memory.h>
46 #include <asm/intel-family.h>
47 #include <asm/processor.h>
48 #include <asm/traps.h>
49 #include <asm/tlbflush.h>
52 #include <asm/reboot.h>
56 static DEFINE_MUTEX(mce_log_mutex);
58 /* sysfs synchronization */
59 static DEFINE_MUTEX(mce_sysfs_mutex);
61 #define CREATE_TRACE_POINTS
62 #include <trace/events/mce.h>
64 #define SPINUNIT 100 /* 100ns */
66 DEFINE_PER_CPU(unsigned, mce_exception_count);
69 u64 ctl; /* subevents to enable */
70 bool init; /* initialise bank? */
72 static DEFINE_PER_CPU_READ_MOSTLY(struct mce_bank[MAX_NR_BANKS], mce_banks_array);
75 /* One object for each MCE bank, shared by all CPUs */
77 struct device_attribute attr; /* device attribute */
78 char attrname[ATTR_LEN]; /* attribute name */
79 u8 bank; /* bank number */
81 static struct mce_bank_dev mce_bank_devs[MAX_NR_BANKS];
83 struct mce_vendor_flags mce_flags __read_mostly;
85 struct mca_config mca_cfg __read_mostly = {
89 * 0: always panic on uncorrected errors, log corrected errors
90 * 1: panic or SIGBUS on uncorrected errors, log corrected errors
91 * 2: SIGBUS or log uncorrected errors (if possible), log corr. errors
92 * 3: never panic or SIGBUS, log all errors (for testing only)
98 static DEFINE_PER_CPU(struct mce, mces_seen);
99 static unsigned long mce_need_notify;
100 static int cpu_missing;
103 * MCA banks polled by the period polling timer for corrected events.
104 * With Intel CMCI, this only has MCA banks which do not support CMCI (if any).
106 DEFINE_PER_CPU(mce_banks_t, mce_poll_banks) = {
107 [0 ... BITS_TO_LONGS(MAX_NR_BANKS)-1] = ~0UL
111 * MCA banks controlled through firmware first for corrected errors.
112 * This is a global list of banks for which we won't enable CMCI and we
113 * won't poll. Firmware controls these banks and is responsible for
114 * reporting corrected errors through GHES. Uncorrected/recoverable
115 * errors are still notified through a machine check.
117 mce_banks_t mce_banks_ce_disabled;
119 static struct work_struct mce_work;
120 static struct irq_work mce_irq_work;
122 static void (*quirk_no_way_out)(int bank, struct mce *m, struct pt_regs *regs);
125 * CPU/chipset specific EDAC code can register a notifier call here to print
126 * MCE errors in a human-readable form.
128 BLOCKING_NOTIFIER_HEAD(x86_mce_decoder_chain);
130 /* Do initial initialization of a struct mce */
131 void mce_setup(struct mce *m)
133 memset(m, 0, sizeof(struct mce));
134 m->cpu = m->extcpu = smp_processor_id();
135 /* need the internal __ version to avoid deadlocks */
136 m->time = __ktime_get_real_seconds();
137 m->cpuvendor = boot_cpu_data.x86_vendor;
138 m->cpuid = cpuid_eax(1);
139 m->socketid = cpu_data(m->extcpu).phys_proc_id;
140 m->apicid = cpu_data(m->extcpu).initial_apicid;
141 rdmsrl(MSR_IA32_MCG_CAP, m->mcgcap);
143 if (this_cpu_has(X86_FEATURE_INTEL_PPIN))
144 rdmsrl(MSR_PPIN, m->ppin);
146 m->microcode = boot_cpu_data.microcode;
149 DEFINE_PER_CPU(struct mce, injectm);
150 EXPORT_PER_CPU_SYMBOL_GPL(injectm);
152 void mce_log(struct mce *m)
154 if (!mce_gen_pool_add(m))
155 irq_work_queue(&mce_irq_work);
158 void mce_inject_log(struct mce *m)
160 mutex_lock(&mce_log_mutex);
162 mutex_unlock(&mce_log_mutex);
164 EXPORT_SYMBOL_GPL(mce_inject_log);
166 static struct notifier_block mce_srao_nb;
169 * We run the default notifier if we have only the SRAO, the first and the
170 * default notifier registered. I.e., the mandatory NUM_DEFAULT_NOTIFIERS
171 * notifiers registered on the chain.
173 #define NUM_DEFAULT_NOTIFIERS 3
174 static atomic_t num_notifiers;
176 void mce_register_decode_chain(struct notifier_block *nb)
178 if (WARN_ON(nb->priority > MCE_PRIO_MCELOG && nb->priority < MCE_PRIO_EDAC))
181 atomic_inc(&num_notifiers);
183 blocking_notifier_chain_register(&x86_mce_decoder_chain, nb);
185 EXPORT_SYMBOL_GPL(mce_register_decode_chain);
187 void mce_unregister_decode_chain(struct notifier_block *nb)
189 atomic_dec(&num_notifiers);
191 blocking_notifier_chain_unregister(&x86_mce_decoder_chain, nb);
193 EXPORT_SYMBOL_GPL(mce_unregister_decode_chain);
195 static inline u32 ctl_reg(int bank)
197 return MSR_IA32_MCx_CTL(bank);
200 static inline u32 status_reg(int bank)
202 return MSR_IA32_MCx_STATUS(bank);
205 static inline u32 addr_reg(int bank)
207 return MSR_IA32_MCx_ADDR(bank);
210 static inline u32 misc_reg(int bank)
212 return MSR_IA32_MCx_MISC(bank);
215 static inline u32 smca_ctl_reg(int bank)
217 return MSR_AMD64_SMCA_MCx_CTL(bank);
220 static inline u32 smca_status_reg(int bank)
222 return MSR_AMD64_SMCA_MCx_STATUS(bank);
225 static inline u32 smca_addr_reg(int bank)
227 return MSR_AMD64_SMCA_MCx_ADDR(bank);
230 static inline u32 smca_misc_reg(int bank)
232 return MSR_AMD64_SMCA_MCx_MISC(bank);
235 struct mca_msr_regs msr_ops = {
237 .status = status_reg,
242 static void __print_mce(struct mce *m)
244 pr_emerg(HW_ERR "CPU %d: Machine Check%s: %Lx Bank %d: %016Lx\n",
246 (m->mcgstatus & MCG_STATUS_MCIP ? " Exception" : ""),
247 m->mcgstatus, m->bank, m->status);
250 pr_emerg(HW_ERR "RIP%s %02x:<%016Lx> ",
251 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
254 if (m->cs == __KERNEL_CS)
255 pr_cont("{%pS}", (void *)(unsigned long)m->ip);
259 pr_emerg(HW_ERR "TSC %llx ", m->tsc);
261 pr_cont("ADDR %llx ", m->addr);
263 pr_cont("MISC %llx ", m->misc);
265 if (mce_flags.smca) {
267 pr_cont("SYND %llx ", m->synd);
269 pr_cont("IPID %llx ", m->ipid);
274 * Note this output is parsed by external tools and old fields
275 * should not be changed.
277 pr_emerg(HW_ERR "PROCESSOR %u:%x TIME %llu SOCKET %u APIC %x microcode %x\n",
278 m->cpuvendor, m->cpuid, m->time, m->socketid, m->apicid,
282 static void print_mce(struct mce *m)
286 if (m->cpuvendor != X86_VENDOR_AMD && m->cpuvendor != X86_VENDOR_HYGON)
287 pr_emerg_ratelimited(HW_ERR "Run the above through 'mcelog --ascii'\n");
290 #define PANIC_TIMEOUT 5 /* 5 seconds */
292 static atomic_t mce_panicked;
294 static int fake_panic;
295 static atomic_t mce_fake_panicked;
297 /* Panic in progress. Enable interrupts and wait for final IPI */
298 static void wait_for_panic(void)
300 long timeout = PANIC_TIMEOUT*USEC_PER_SEC;
304 while (timeout-- > 0)
306 if (panic_timeout == 0)
307 panic_timeout = mca_cfg.panic_timeout;
308 panic("Panicing machine check CPU died");
311 static void mce_panic(const char *msg, struct mce *final, char *exp)
314 struct llist_node *pending;
315 struct mce_evt_llist *l;
319 * Make sure only one CPU runs in machine check panic
321 if (atomic_inc_return(&mce_panicked) > 1)
328 /* Don't log too much for fake panic */
329 if (atomic_inc_return(&mce_fake_panicked) > 1)
332 pending = mce_gen_pool_prepare_records();
333 /* First print corrected ones that are still unlogged */
334 llist_for_each_entry(l, pending, llnode) {
335 struct mce *m = &l->mce;
336 if (!(m->status & MCI_STATUS_UC)) {
339 apei_err = apei_write_mce(m);
342 /* Now print uncorrected but with the final one last */
343 llist_for_each_entry(l, pending, llnode) {
344 struct mce *m = &l->mce;
345 if (!(m->status & MCI_STATUS_UC))
347 if (!final || mce_cmp(m, final)) {
350 apei_err = apei_write_mce(m);
356 apei_err = apei_write_mce(final);
359 pr_emerg(HW_ERR "Some CPUs didn't answer in synchronization\n");
361 pr_emerg(HW_ERR "Machine check: %s\n", exp);
363 if (panic_timeout == 0)
364 panic_timeout = mca_cfg.panic_timeout;
367 pr_emerg(HW_ERR "Fake kernel panic: %s\n", msg);
370 /* Support code for software error injection */
372 static int msr_to_offset(u32 msr)
374 unsigned bank = __this_cpu_read(injectm.bank);
376 if (msr == mca_cfg.rip_msr)
377 return offsetof(struct mce, ip);
378 if (msr == msr_ops.status(bank))
379 return offsetof(struct mce, status);
380 if (msr == msr_ops.addr(bank))
381 return offsetof(struct mce, addr);
382 if (msr == msr_ops.misc(bank))
383 return offsetof(struct mce, misc);
384 if (msr == MSR_IA32_MCG_STATUS)
385 return offsetof(struct mce, mcgstatus);
389 /* MSR access wrappers used for error injection */
390 static u64 mce_rdmsrl(u32 msr)
394 if (__this_cpu_read(injectm.finished)) {
395 int offset = msr_to_offset(msr);
399 return *(u64 *)((char *)this_cpu_ptr(&injectm) + offset);
402 if (rdmsrl_safe(msr, &v)) {
403 WARN_ONCE(1, "mce: Unable to read MSR 0x%x!\n", msr);
405 * Return zero in case the access faulted. This should
406 * not happen normally but can happen if the CPU does
407 * something weird, or if the code is buggy.
415 static void mce_wrmsrl(u32 msr, u64 v)
417 if (__this_cpu_read(injectm.finished)) {
418 int offset = msr_to_offset(msr);
421 *(u64 *)((char *)this_cpu_ptr(&injectm) + offset) = v;
428 * Collect all global (w.r.t. this processor) status about this machine
429 * check into our "mce" struct so that we can use it later to assess
430 * the severity of the problem as we read per-bank specific details.
432 static inline void mce_gather_info(struct mce *m, struct pt_regs *regs)
436 m->mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
439 * Get the address of the instruction at the time of
440 * the machine check error.
442 if (m->mcgstatus & (MCG_STATUS_RIPV|MCG_STATUS_EIPV)) {
447 * When in VM86 mode make the cs look like ring 3
448 * always. This is a lie, but it's better than passing
449 * the additional vm86 bit around everywhere.
451 if (v8086_mode(regs))
454 /* Use accurate RIP reporting if available. */
456 m->ip = mce_rdmsrl(mca_cfg.rip_msr);
460 int mce_available(struct cpuinfo_x86 *c)
462 if (mca_cfg.disabled)
464 return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
467 static void mce_schedule_work(void)
469 if (!mce_gen_pool_empty())
470 schedule_work(&mce_work);
473 static void mce_irq_work_cb(struct irq_work *entry)
479 * Check if the address reported by the CPU is in a format we can parse.
480 * It would be possible to add code for most other cases, but all would
481 * be somewhat complicated (e.g. segment offset would require an instruction
482 * parser). So only support physical addresses up to page granuality for now.
484 int mce_usable_address(struct mce *m)
486 if (!(m->status & MCI_STATUS_ADDRV))
489 /* Checks after this one are Intel-specific: */
490 if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
493 if (!(m->status & MCI_STATUS_MISCV))
496 if (MCI_MISC_ADDR_LSB(m->misc) > PAGE_SHIFT)
499 if (MCI_MISC_ADDR_MODE(m->misc) != MCI_MISC_ADDR_PHYS)
504 EXPORT_SYMBOL_GPL(mce_usable_address);
506 bool mce_is_memory_error(struct mce *m)
508 if (m->cpuvendor == X86_VENDOR_AMD ||
509 m->cpuvendor == X86_VENDOR_HYGON) {
510 return amd_mce_is_memory_error(m);
511 } else if (m->cpuvendor == X86_VENDOR_INTEL) {
513 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
515 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
516 * indicating a memory error. Bit 8 is used for indicating a
517 * cache hierarchy error. The combination of bit 2 and bit 3
518 * is used for indicating a `generic' cache hierarchy error
519 * But we can't just blindly check the above bits, because if
520 * bit 11 is set, then it is a bus/interconnect error - and
521 * either way the above bits just gives more detail on what
522 * bus/interconnect error happened. Note that bit 12 can be
523 * ignored, as it's the "filter" bit.
525 return (m->status & 0xef80) == BIT(7) ||
526 (m->status & 0xef00) == BIT(8) ||
527 (m->status & 0xeffc) == 0xc;
532 EXPORT_SYMBOL_GPL(mce_is_memory_error);
534 bool mce_is_correctable(struct mce *m)
536 if (m->cpuvendor == X86_VENDOR_AMD && m->status & MCI_STATUS_DEFERRED)
539 if (m->cpuvendor == X86_VENDOR_HYGON && m->status & MCI_STATUS_DEFERRED)
542 if (m->status & MCI_STATUS_UC)
547 EXPORT_SYMBOL_GPL(mce_is_correctable);
549 static bool cec_add_mce(struct mce *m)
554 /* We eat only correctable DRAM errors with usable addresses. */
555 if (mce_is_memory_error(m) &&
556 mce_is_correctable(m) &&
557 mce_usable_address(m))
558 if (!cec_add_elem(m->addr >> PAGE_SHIFT))
564 static int mce_first_notifier(struct notifier_block *nb, unsigned long val,
567 struct mce *m = (struct mce *)data;
575 /* Emit the trace record: */
578 set_bit(0, &mce_need_notify);
585 static struct notifier_block first_nb = {
586 .notifier_call = mce_first_notifier,
587 .priority = MCE_PRIO_FIRST,
590 static int srao_decode_notifier(struct notifier_block *nb, unsigned long val,
593 struct mce *mce = (struct mce *)data;
599 if (mce_usable_address(mce) && (mce->severity == MCE_AO_SEVERITY)) {
600 pfn = mce->addr >> PAGE_SHIFT;
601 if (!memory_failure(pfn, 0))
607 static struct notifier_block mce_srao_nb = {
608 .notifier_call = srao_decode_notifier,
609 .priority = MCE_PRIO_SRAO,
612 static int mce_default_notifier(struct notifier_block *nb, unsigned long val,
615 struct mce *m = (struct mce *)data;
620 if (atomic_read(&num_notifiers) > NUM_DEFAULT_NOTIFIERS)
628 static struct notifier_block mce_default_nb = {
629 .notifier_call = mce_default_notifier,
630 /* lowest prio, we want it to run last. */
631 .priority = MCE_PRIO_LOWEST,
635 * Read ADDR and MISC registers.
637 static void mce_read_aux(struct mce *m, int i)
639 if (m->status & MCI_STATUS_MISCV)
640 m->misc = mce_rdmsrl(msr_ops.misc(i));
642 if (m->status & MCI_STATUS_ADDRV) {
643 m->addr = mce_rdmsrl(msr_ops.addr(i));
646 * Mask the reported address by the reported granularity.
648 if (mca_cfg.ser && (m->status & MCI_STATUS_MISCV)) {
649 u8 shift = MCI_MISC_ADDR_LSB(m->misc);
655 * Extract [55:<lsb>] where lsb is the least significant
656 * *valid* bit of the address bits.
658 if (mce_flags.smca) {
659 u8 lsb = (m->addr >> 56) & 0x3f;
661 m->addr &= GENMASK_ULL(55, lsb);
665 if (mce_flags.smca) {
666 m->ipid = mce_rdmsrl(MSR_AMD64_SMCA_MCx_IPID(i));
668 if (m->status & MCI_STATUS_SYNDV)
669 m->synd = mce_rdmsrl(MSR_AMD64_SMCA_MCx_SYND(i));
673 DEFINE_PER_CPU(unsigned, mce_poll_count);
676 * Poll for corrected events or events that happened before reset.
677 * Those are just logged through /dev/mcelog.
679 * This is executed in standard interrupt context.
681 * Note: spec recommends to panic for fatal unsignalled
682 * errors here. However this would be quite problematic --
683 * we would need to reimplement the Monarch handling and
684 * it would mess up the exclusion between exception handler
685 * and poll handler -- * so we skip this for now.
686 * These cases should not happen anyways, or only when the CPU
687 * is already totally * confused. In this case it's likely it will
688 * not fully execute the machine check handler either.
690 bool machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
692 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
693 bool error_seen = false;
697 this_cpu_inc(mce_poll_count);
699 mce_gather_info(&m, NULL);
701 if (flags & MCP_TIMESTAMP)
704 for (i = 0; i < mca_cfg.banks; i++) {
705 if (!mce_banks[i].ctl || !test_bit(i, *b))
713 m.status = mce_rdmsrl(msr_ops.status(i));
715 /* If this entry is not valid, ignore it */
716 if (!(m.status & MCI_STATUS_VAL))
720 * If we are logging everything (at CPU online) or this
721 * is a corrected error, then we must log it.
723 if ((flags & MCP_UC) || !(m.status & MCI_STATUS_UC))
727 * Newer Intel systems that support software error
728 * recovery need to make additional checks. Other
729 * CPUs should skip over uncorrected errors, but log
733 if (m.status & MCI_STATUS_UC)
738 /* Log "not enabled" (speculative) errors */
739 if (!(m.status & MCI_STATUS_EN))
743 * Log UCNA (SDM: 15.6.3 "UCR Error Classification")
744 * UC == 1 && PCC == 0 && S == 0
746 if (!(m.status & MCI_STATUS_PCC) && !(m.status & MCI_STATUS_S))
750 * Skip anything else. Presumption is that our read of this
751 * bank is racing with a machine check. Leave the log alone
752 * for do_machine_check() to deal with it.
761 m.severity = mce_severity(&m, mca_cfg.tolerant, NULL, false);
764 * Don't get the IP here because it's unlikely to
765 * have anything to do with the actual error location.
767 if (!(flags & MCP_DONTLOG) && !mca_cfg.dont_log_ce)
769 else if (mce_usable_address(&m)) {
771 * Although we skipped logging this, we still want
772 * to take action. Add to the pool so the registered
773 * notifiers will see it.
775 if (!mce_gen_pool_add(&m))
780 * Clear state for this bank.
782 mce_wrmsrl(msr_ops.status(i), 0);
786 * Don't clear MCG_STATUS here because it's only defined for
794 EXPORT_SYMBOL_GPL(machine_check_poll);
797 * Do a quick check if any of the events requires a panic.
798 * This decides if we keep the events around or clear them.
800 static int mce_no_way_out(struct mce *m, char **msg, unsigned long *validp,
801 struct pt_regs *regs)
806 for (i = 0; i < mca_cfg.banks; i++) {
807 m->status = mce_rdmsrl(msr_ops.status(i));
808 if (!(m->status & MCI_STATUS_VAL))
811 __set_bit(i, validp);
812 if (quirk_no_way_out)
813 quirk_no_way_out(i, m, regs);
815 if (mce_severity(m, mca_cfg.tolerant, &tmp, true) >= MCE_PANIC_SEVERITY) {
826 * Variable to establish order between CPUs while scanning.
827 * Each CPU spins initially until executing is equal its number.
829 static atomic_t mce_executing;
832 * Defines order of CPUs on entry. First CPU becomes Monarch.
834 static atomic_t mce_callin;
837 * Check if a timeout waiting for other CPUs happened.
839 static int mce_timed_out(u64 *t, const char *msg)
842 * The others already did panic for some reason.
843 * Bail out like in a timeout.
844 * rmb() to tell the compiler that system_state
845 * might have been modified by someone else.
848 if (atomic_read(&mce_panicked))
850 if (!mca_cfg.monarch_timeout)
852 if ((s64)*t < SPINUNIT) {
853 if (mca_cfg.tolerant <= 1)
854 mce_panic(msg, NULL, NULL);
860 touch_nmi_watchdog();
865 * The Monarch's reign. The Monarch is the CPU who entered
866 * the machine check handler first. It waits for the others to
867 * raise the exception too and then grades them. When any
868 * error is fatal panic. Only then let the others continue.
870 * The other CPUs entering the MCE handler will be controlled by the
871 * Monarch. They are called Subjects.
873 * This way we prevent any potential data corruption in a unrecoverable case
874 * and also makes sure always all CPU's errors are examined.
876 * Also this detects the case of a machine check event coming from outer
877 * space (not detected by any CPUs) In this case some external agent wants
878 * us to shut down, so panic too.
880 * The other CPUs might still decide to panic if the handler happens
881 * in a unrecoverable place, but in this case the system is in a semi-stable
882 * state and won't corrupt anything by itself. It's ok to let the others
883 * continue for a bit first.
885 * All the spin loops have timeouts; when a timeout happens a CPU
886 * typically elects itself to be Monarch.
888 static void mce_reign(void)
891 struct mce *m = NULL;
892 int global_worst = 0;
897 * This CPU is the Monarch and the other CPUs have run
898 * through their handlers.
899 * Grade the severity of the errors of all the CPUs.
901 for_each_possible_cpu(cpu) {
902 int severity = mce_severity(&per_cpu(mces_seen, cpu),
905 if (severity > global_worst) {
907 global_worst = severity;
908 m = &per_cpu(mces_seen, cpu);
913 * Cannot recover? Panic here then.
914 * This dumps all the mces in the log buffer and stops the
917 if (m && global_worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3)
918 mce_panic("Fatal machine check", m, msg);
921 * For UC somewhere we let the CPU who detects it handle it.
922 * Also must let continue the others, otherwise the handling
923 * CPU could deadlock on a lock.
927 * No machine check event found. Must be some external
928 * source or one CPU is hung. Panic.
930 if (global_worst <= MCE_KEEP_SEVERITY && mca_cfg.tolerant < 3)
931 mce_panic("Fatal machine check from unknown source", NULL, NULL);
934 * Now clear all the mces_seen so that they don't reappear on
937 for_each_possible_cpu(cpu)
938 memset(&per_cpu(mces_seen, cpu), 0, sizeof(struct mce));
941 static atomic_t global_nwo;
944 * Start of Monarch synchronization. This waits until all CPUs have
945 * entered the exception handler and then determines if any of them
946 * saw a fatal event that requires panic. Then it executes them
947 * in the entry order.
948 * TBD double check parallel CPU hotunplug
950 static int mce_start(int *no_way_out)
953 int cpus = num_online_cpus();
954 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
959 atomic_add(*no_way_out, &global_nwo);
961 * Rely on the implied barrier below, such that global_nwo
962 * is updated before mce_callin.
964 order = atomic_inc_return(&mce_callin);
969 while (atomic_read(&mce_callin) != cpus) {
970 if (mce_timed_out(&timeout,
971 "Timeout: Not all CPUs entered broadcast exception handler")) {
972 atomic_set(&global_nwo, 0);
979 * mce_callin should be read before global_nwo
985 * Monarch: Starts executing now, the others wait.
987 atomic_set(&mce_executing, 1);
990 * Subject: Now start the scanning loop one by one in
991 * the original callin order.
992 * This way when there are any shared banks it will be
993 * only seen by one CPU before cleared, avoiding duplicates.
995 while (atomic_read(&mce_executing) < order) {
996 if (mce_timed_out(&timeout,
997 "Timeout: Subject CPUs unable to finish machine check processing")) {
998 atomic_set(&global_nwo, 0);
1006 * Cache the global no_way_out state.
1008 *no_way_out = atomic_read(&global_nwo);
1014 * Synchronize between CPUs after main scanning loop.
1015 * This invokes the bulk of the Monarch processing.
1017 static int mce_end(int order)
1020 u64 timeout = (u64)mca_cfg.monarch_timeout * NSEC_PER_USEC;
1028 * Allow others to run.
1030 atomic_inc(&mce_executing);
1033 /* CHECKME: Can this race with a parallel hotplug? */
1034 int cpus = num_online_cpus();
1037 * Monarch: Wait for everyone to go through their scanning
1040 while (atomic_read(&mce_executing) <= cpus) {
1041 if (mce_timed_out(&timeout,
1042 "Timeout: Monarch CPU unable to finish machine check processing"))
1052 * Subject: Wait for Monarch to finish.
1054 while (atomic_read(&mce_executing) != 0) {
1055 if (mce_timed_out(&timeout,
1056 "Timeout: Monarch CPU did not finish machine check processing"))
1062 * Don't reset anything. That's done by the Monarch.
1068 * Reset all global state.
1071 atomic_set(&global_nwo, 0);
1072 atomic_set(&mce_callin, 0);
1076 * Let others run again.
1078 atomic_set(&mce_executing, 0);
1082 static void mce_clear_state(unsigned long *toclear)
1086 for (i = 0; i < mca_cfg.banks; i++) {
1087 if (test_bit(i, toclear))
1088 mce_wrmsrl(msr_ops.status(i), 0);
1092 static int do_memory_failure(struct mce *m)
1094 int flags = MF_ACTION_REQUIRED;
1097 pr_err("Uncorrected hardware memory error in user-access at %llx", m->addr);
1098 if (!(m->mcgstatus & MCG_STATUS_RIPV))
1099 flags |= MF_MUST_KILL;
1100 ret = memory_failure(m->addr >> PAGE_SHIFT, flags);
1102 pr_err("Memory error not recovered");
1104 set_mce_nospec(m->addr >> PAGE_SHIFT);
1110 * Cases where we avoid rendezvous handler timeout:
1111 * 1) If this CPU is offline.
1113 * 2) If crashing_cpu was set, e.g. we're entering kdump and we need to
1114 * skip those CPUs which remain looping in the 1st kernel - see
1115 * crash_nmi_callback().
1117 * Note: there still is a small window between kexec-ing and the new,
1118 * kdump kernel establishing a new #MC handler where a broadcasted MCE
1119 * might not get handled properly.
1121 static bool __mc_check_crashing_cpu(int cpu)
1123 if (cpu_is_offline(cpu) ||
1124 (crashing_cpu != -1 && crashing_cpu != cpu)) {
1127 mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
1128 if (mcgstatus & MCG_STATUS_RIPV) {
1129 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1136 static void __mc_scan_banks(struct mce *m, struct mce *final,
1137 unsigned long *toclear, unsigned long *valid_banks,
1138 int no_way_out, int *worst)
1140 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
1141 struct mca_config *cfg = &mca_cfg;
1144 for (i = 0; i < cfg->banks; i++) {
1145 __clear_bit(i, toclear);
1146 if (!test_bit(i, valid_banks))
1149 if (!mce_banks[i].ctl)
1156 m->status = mce_rdmsrl(msr_ops.status(i));
1157 if (!(m->status & MCI_STATUS_VAL))
1161 * Corrected or non-signaled errors are handled by
1162 * machine_check_poll(). Leave them alone, unless this panics.
1164 if (!(m->status & (cfg->ser ? MCI_STATUS_S : MCI_STATUS_UC)) &&
1168 /* Set taint even when machine check was not enabled. */
1169 add_taint(TAINT_MACHINE_CHECK, LOCKDEP_NOW_UNRELIABLE);
1171 severity = mce_severity(m, cfg->tolerant, NULL, true);
1174 * When machine check was for corrected/deferred handler don't
1175 * touch, unless we're panicking.
1177 if ((severity == MCE_KEEP_SEVERITY ||
1178 severity == MCE_UCNA_SEVERITY) && !no_way_out)
1181 __set_bit(i, toclear);
1183 /* Machine check event was not enabled. Clear, but ignore. */
1184 if (severity == MCE_NO_SEVERITY)
1189 /* assuming valid severity level != 0 */
1190 m->severity = severity;
1194 if (severity > *worst) {
1200 /* mce_clear_state will clear *final, save locally for use later */
1205 * The actual machine check handler. This only handles real
1206 * exceptions when something got corrupted coming in through int 18.
1208 * This is executed in NMI context not subject to normal locking rules. This
1209 * implies that most kernel services cannot be safely used. Don't even
1210 * think about putting a printk in there!
1212 * On Intel systems this is entered on all CPUs in parallel through
1213 * MCE broadcast. However some CPUs might be broken beyond repair,
1214 * so be always careful when synchronizing with others.
1216 void do_machine_check(struct pt_regs *regs, long error_code)
1218 DECLARE_BITMAP(valid_banks, MAX_NR_BANKS);
1219 DECLARE_BITMAP(toclear, MAX_NR_BANKS);
1220 struct mca_config *cfg = &mca_cfg;
1221 int cpu = smp_processor_id();
1222 char *msg = "Unknown";
1223 struct mce m, *final;
1227 * Establish sequential order between the CPUs entering the machine
1233 * If no_way_out gets set, there is no safe way to recover from this
1234 * MCE. If mca_cfg.tolerant is cranked up, we'll try anyway.
1239 * If kill_it gets set, there might be a way to recover from this
1245 * MCEs are always local on AMD. Same is determined by MCG_STATUS_LMCES
1250 if (__mc_check_crashing_cpu(cpu))
1255 this_cpu_inc(mce_exception_count);
1257 mce_gather_info(&m, regs);
1260 final = this_cpu_ptr(&mces_seen);
1263 memset(valid_banks, 0, sizeof(valid_banks));
1264 no_way_out = mce_no_way_out(&m, &msg, valid_banks, regs);
1269 * When no restart IP might need to kill or panic.
1270 * Assume the worst for now, but if we find the
1271 * severity is MCE_AR_SEVERITY we have other options.
1273 if (!(m.mcgstatus & MCG_STATUS_RIPV))
1277 * Check if this MCE is signaled to only this logical processor,
1280 if (m.cpuvendor == X86_VENDOR_INTEL)
1281 lmce = m.mcgstatus & MCG_STATUS_LMCES;
1284 * Local machine check may already know that we have to panic.
1285 * Broadcast machine check begins rendezvous in mce_start()
1286 * Go through all banks in exclusion of the other CPUs. This way we
1287 * don't report duplicated events on shared banks because the first one
1288 * to see it will clear it.
1292 mce_panic("Fatal local machine check", &m, msg);
1294 order = mce_start(&no_way_out);
1297 __mc_scan_banks(&m, final, toclear, valid_banks, no_way_out, &worst);
1300 mce_clear_state(toclear);
1303 * Do most of the synchronization with other CPUs.
1304 * When there's any problem use only local no_way_out state.
1307 if (mce_end(order) < 0)
1308 no_way_out = worst >= MCE_PANIC_SEVERITY;
1311 * If there was a fatal machine check we should have
1312 * already called mce_panic earlier in this function.
1313 * Since we re-read the banks, we might have found
1314 * something new. Check again to see if we found a
1315 * fatal error. We call "mce_severity()" again to
1316 * make sure we have the right "msg".
1318 if (worst >= MCE_PANIC_SEVERITY && mca_cfg.tolerant < 3) {
1319 mce_severity(&m, cfg->tolerant, &msg, true);
1320 mce_panic("Local fatal machine check!", &m, msg);
1325 * If tolerant is at an insane level we drop requests to kill
1326 * processes and continue even when there is no way out.
1328 if (cfg->tolerant == 3)
1330 else if (no_way_out)
1331 mce_panic("Fatal machine check on current CPU", &m, msg);
1334 irq_work_queue(&mce_irq_work);
1336 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
1340 if (worst != MCE_AR_SEVERITY && !kill_it)
1343 /* Fault was in user mode and we need to take some action */
1344 if ((m.cs & 3) == 3) {
1345 ist_begin_non_atomic(regs);
1348 if (kill_it || do_memory_failure(&m))
1349 force_sig(SIGBUS, current);
1350 local_irq_disable();
1351 ist_end_non_atomic();
1353 if (!fixup_exception(regs, X86_TRAP_MC, error_code, 0))
1354 mce_panic("Failed kernel mode recovery", &m, NULL);
1360 EXPORT_SYMBOL_GPL(do_machine_check);
1362 #ifndef CONFIG_MEMORY_FAILURE
1363 int memory_failure(unsigned long pfn, int flags)
1365 /* mce_severity() should not hand us an ACTION_REQUIRED error */
1366 BUG_ON(flags & MF_ACTION_REQUIRED);
1367 pr_err("Uncorrected memory error in page 0x%lx ignored\n"
1368 "Rebuild kernel with CONFIG_MEMORY_FAILURE=y for smarter handling\n",
1376 * Periodic polling timer for "silent" machine check errors. If the
1377 * poller finds an MCE, poll 2x faster. When the poller finds no more
1378 * errors, poll 2x slower (up to check_interval seconds).
1380 static unsigned long check_interval = INITIAL_CHECK_INTERVAL;
1382 static DEFINE_PER_CPU(unsigned long, mce_next_interval); /* in jiffies */
1383 static DEFINE_PER_CPU(struct timer_list, mce_timer);
1385 static unsigned long mce_adjust_timer_default(unsigned long interval)
1390 static unsigned long (*mce_adjust_timer)(unsigned long interval) = mce_adjust_timer_default;
1392 static void __start_timer(struct timer_list *t, unsigned long interval)
1394 unsigned long when = jiffies + interval;
1395 unsigned long flags;
1397 local_irq_save(flags);
1399 if (!timer_pending(t) || time_before(when, t->expires))
1400 mod_timer(t, round_jiffies(when));
1402 local_irq_restore(flags);
1405 static void mce_timer_fn(struct timer_list *t)
1407 struct timer_list *cpu_t = this_cpu_ptr(&mce_timer);
1410 WARN_ON(cpu_t != t);
1412 iv = __this_cpu_read(mce_next_interval);
1414 if (mce_available(this_cpu_ptr(&cpu_info))) {
1415 machine_check_poll(0, this_cpu_ptr(&mce_poll_banks));
1417 if (mce_intel_cmci_poll()) {
1418 iv = mce_adjust_timer(iv);
1424 * Alert userspace if needed. If we logged an MCE, reduce the polling
1425 * interval, otherwise increase the polling interval.
1427 if (mce_notify_irq())
1428 iv = max(iv / 2, (unsigned long) HZ/100);
1430 iv = min(iv * 2, round_jiffies_relative(check_interval * HZ));
1433 __this_cpu_write(mce_next_interval, iv);
1434 __start_timer(t, iv);
1438 * Ensure that the timer is firing in @interval from now.
1440 void mce_timer_kick(unsigned long interval)
1442 struct timer_list *t = this_cpu_ptr(&mce_timer);
1443 unsigned long iv = __this_cpu_read(mce_next_interval);
1445 __start_timer(t, interval);
1448 __this_cpu_write(mce_next_interval, interval);
1451 /* Must not be called in IRQ context where del_timer_sync() can deadlock */
1452 static void mce_timer_delete_all(void)
1456 for_each_online_cpu(cpu)
1457 del_timer_sync(&per_cpu(mce_timer, cpu));
1461 * Notify the user(s) about new machine check events.
1462 * Can be called from interrupt context, but not from machine check/NMI
1465 int mce_notify_irq(void)
1467 /* Not more than two messages every minute */
1468 static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2);
1470 if (test_and_clear_bit(0, &mce_need_notify)) {
1473 if (__ratelimit(&ratelimit))
1474 pr_info(HW_ERR "Machine check events logged\n");
1480 EXPORT_SYMBOL_GPL(mce_notify_irq);
1482 static void __mcheck_cpu_mce_banks_init(void)
1484 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
1487 for (i = 0; i < MAX_NR_BANKS; i++) {
1488 struct mce_bank *b = &mce_banks[i];
1496 * Initialize Machine Checks for a CPU.
1498 static void __mcheck_cpu_cap_init(void)
1503 rdmsrl(MSR_IA32_MCG_CAP, cap);
1505 b = cap & MCG_BANKCNT_MASK;
1506 if (WARN_ON_ONCE(b > MAX_NR_BANKS))
1509 mca_cfg.banks = max(mca_cfg.banks, b);
1511 __mcheck_cpu_mce_banks_init();
1513 /* Use accurate RIP reporting if available. */
1514 if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
1515 mca_cfg.rip_msr = MSR_IA32_MCG_EIP;
1517 if (cap & MCG_SER_P)
1521 static void __mcheck_cpu_init_generic(void)
1523 enum mcp_flags m_fl = 0;
1524 mce_banks_t all_banks;
1527 if (!mca_cfg.bootlog)
1531 * Log the machine checks left over from the previous reset.
1533 bitmap_fill(all_banks, MAX_NR_BANKS);
1534 machine_check_poll(MCP_UC | m_fl, &all_banks);
1536 cr4_set_bits(X86_CR4_MCE);
1538 rdmsrl(MSR_IA32_MCG_CAP, cap);
1539 if (cap & MCG_CTL_P)
1540 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
1543 static void __mcheck_cpu_init_clear_banks(void)
1545 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
1548 for (i = 0; i < mca_cfg.banks; i++) {
1549 struct mce_bank *b = &mce_banks[i];
1553 wrmsrl(msr_ops.ctl(i), b->ctl);
1554 wrmsrl(msr_ops.status(i), 0);
1559 * During IFU recovery Sandy Bridge -EP4S processors set the RIPV and
1560 * EIPV bits in MCG_STATUS to zero on the affected logical processor (SDM
1561 * Vol 3B Table 15-20). But this confuses both the code that determines
1562 * whether the machine check occurred in kernel or user mode, and also
1563 * the severity assessment code. Pretend that EIPV was set, and take the
1564 * ip/cs values from the pt_regs that mce_gather_info() ignored earlier.
1566 static void quirk_sandybridge_ifu(int bank, struct mce *m, struct pt_regs *regs)
1570 if ((m->mcgstatus & (MCG_STATUS_EIPV|MCG_STATUS_RIPV)) != 0)
1572 if ((m->status & (MCI_STATUS_OVER|MCI_STATUS_UC|
1573 MCI_STATUS_EN|MCI_STATUS_MISCV|MCI_STATUS_ADDRV|
1574 MCI_STATUS_PCC|MCI_STATUS_S|MCI_STATUS_AR|
1576 (MCI_STATUS_UC|MCI_STATUS_EN|
1577 MCI_STATUS_MISCV|MCI_STATUS_ADDRV|MCI_STATUS_S|
1578 MCI_STATUS_AR|MCACOD_INSTR))
1581 m->mcgstatus |= MCG_STATUS_EIPV;
1586 /* Add per CPU specific workarounds here */
1587 static int __mcheck_cpu_apply_quirks(struct cpuinfo_x86 *c)
1589 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
1590 struct mca_config *cfg = &mca_cfg;
1592 if (c->x86_vendor == X86_VENDOR_UNKNOWN) {
1593 pr_info("unknown CPU type - not enabling MCE support\n");
1597 /* This should be disabled by the BIOS, but isn't always */
1598 if (c->x86_vendor == X86_VENDOR_AMD) {
1599 if (c->x86 == 15 && cfg->banks > 4) {
1601 * disable GART TBL walk error reporting, which
1602 * trips off incorrectly with the IOMMU & 3ware
1605 clear_bit(10, (unsigned long *)&mce_banks[4].ctl);
1607 if (c->x86 < 0x11 && cfg->bootlog < 0) {
1609 * Lots of broken BIOS around that don't clear them
1610 * by default and leave crap in there. Don't log:
1615 * Various K7s with broken bank 0 around. Always disable
1618 if (c->x86 == 6 && cfg->banks > 0)
1619 mce_banks[0].ctl = 0;
1622 * overflow_recov is supported for F15h Models 00h-0fh
1623 * even though we don't have a CPUID bit for it.
1625 if (c->x86 == 0x15 && c->x86_model <= 0xf)
1626 mce_flags.overflow_recov = 1;
1630 if (c->x86_vendor == X86_VENDOR_INTEL) {
1632 * SDM documents that on family 6 bank 0 should not be written
1633 * because it aliases to another special BIOS controlled
1635 * But it's not aliased anymore on model 0x1a+
1636 * Don't ignore bank 0 completely because there could be a
1637 * valid event later, merely don't write CTL0.
1640 if (c->x86 == 6 && c->x86_model < 0x1A && cfg->banks > 0)
1641 mce_banks[0].init = 0;
1644 * All newer Intel systems support MCE broadcasting. Enable
1645 * synchronization with a one second timeout.
1647 if ((c->x86 > 6 || (c->x86 == 6 && c->x86_model >= 0xe)) &&
1648 cfg->monarch_timeout < 0)
1649 cfg->monarch_timeout = USEC_PER_SEC;
1652 * There are also broken BIOSes on some Pentium M and
1655 if (c->x86 == 6 && c->x86_model <= 13 && cfg->bootlog < 0)
1658 if (c->x86 == 6 && c->x86_model == 45)
1659 quirk_no_way_out = quirk_sandybridge_ifu;
1661 if (cfg->monarch_timeout < 0)
1662 cfg->monarch_timeout = 0;
1663 if (cfg->bootlog != 0)
1664 cfg->panic_timeout = 30;
1669 static int __mcheck_cpu_ancient_init(struct cpuinfo_x86 *c)
1674 switch (c->x86_vendor) {
1675 case X86_VENDOR_INTEL:
1676 intel_p5_mcheck_init(c);
1679 case X86_VENDOR_CENTAUR:
1680 winchip_mcheck_init(c);
1691 * Init basic CPU features needed for early decoding of MCEs.
1693 static void __mcheck_cpu_init_early(struct cpuinfo_x86 *c)
1695 if (c->x86_vendor == X86_VENDOR_AMD || c->x86_vendor == X86_VENDOR_HYGON) {
1696 mce_flags.overflow_recov = !!cpu_has(c, X86_FEATURE_OVERFLOW_RECOV);
1697 mce_flags.succor = !!cpu_has(c, X86_FEATURE_SUCCOR);
1698 mce_flags.smca = !!cpu_has(c, X86_FEATURE_SMCA);
1700 if (mce_flags.smca) {
1701 msr_ops.ctl = smca_ctl_reg;
1702 msr_ops.status = smca_status_reg;
1703 msr_ops.addr = smca_addr_reg;
1704 msr_ops.misc = smca_misc_reg;
1709 static void mce_centaur_feature_init(struct cpuinfo_x86 *c)
1711 struct mca_config *cfg = &mca_cfg;
1714 * All newer Centaur CPUs support MCE broadcasting. Enable
1715 * synchronization with a one second timeout.
1717 if ((c->x86 == 6 && c->x86_model == 0xf && c->x86_stepping >= 0xe) ||
1719 if (cfg->monarch_timeout < 0)
1720 cfg->monarch_timeout = USEC_PER_SEC;
1724 static void __mcheck_cpu_init_vendor(struct cpuinfo_x86 *c)
1726 switch (c->x86_vendor) {
1727 case X86_VENDOR_INTEL:
1728 mce_intel_feature_init(c);
1729 mce_adjust_timer = cmci_intel_adjust_timer;
1732 case X86_VENDOR_AMD: {
1733 mce_amd_feature_init(c);
1737 case X86_VENDOR_HYGON:
1738 mce_hygon_feature_init(c);
1741 case X86_VENDOR_CENTAUR:
1742 mce_centaur_feature_init(c);
1750 static void __mcheck_cpu_clear_vendor(struct cpuinfo_x86 *c)
1752 switch (c->x86_vendor) {
1753 case X86_VENDOR_INTEL:
1754 mce_intel_feature_clear(c);
1761 static void mce_start_timer(struct timer_list *t)
1763 unsigned long iv = check_interval * HZ;
1765 if (mca_cfg.ignore_ce || !iv)
1768 this_cpu_write(mce_next_interval, iv);
1769 __start_timer(t, iv);
1772 static void __mcheck_cpu_setup_timer(void)
1774 struct timer_list *t = this_cpu_ptr(&mce_timer);
1776 timer_setup(t, mce_timer_fn, TIMER_PINNED);
1779 static void __mcheck_cpu_init_timer(void)
1781 struct timer_list *t = this_cpu_ptr(&mce_timer);
1783 timer_setup(t, mce_timer_fn, TIMER_PINNED);
1787 bool filter_mce(struct mce *m)
1789 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1790 return amd_filter_mce(m);
1795 /* Handle unconfigured int18 (should never happen) */
1796 static void unexpected_machine_check(struct pt_regs *regs, long error_code)
1798 pr_err("CPU#%d: Unexpected int18 (Machine Check)\n",
1799 smp_processor_id());
1802 /* Call the installed machine check handler for this CPU setup. */
1803 void (*machine_check_vector)(struct pt_regs *, long error_code) =
1804 unexpected_machine_check;
1806 dotraplinkage void do_mce(struct pt_regs *regs, long error_code)
1808 machine_check_vector(regs, error_code);
1812 * Called for each booted CPU to set up machine checks.
1813 * Must be called with preempt off:
1815 void mcheck_cpu_init(struct cpuinfo_x86 *c)
1817 if (mca_cfg.disabled)
1820 if (__mcheck_cpu_ancient_init(c))
1823 if (!mce_available(c))
1826 __mcheck_cpu_cap_init();
1828 if (__mcheck_cpu_apply_quirks(c) < 0) {
1829 mca_cfg.disabled = 1;
1833 if (mce_gen_pool_init()) {
1834 mca_cfg.disabled = 1;
1835 pr_emerg("Couldn't allocate MCE records pool!\n");
1839 machine_check_vector = do_machine_check;
1841 __mcheck_cpu_init_early(c);
1842 __mcheck_cpu_init_generic();
1843 __mcheck_cpu_init_vendor(c);
1844 __mcheck_cpu_init_clear_banks();
1845 __mcheck_cpu_setup_timer();
1849 * Called for each booted CPU to clear some machine checks opt-ins
1851 void mcheck_cpu_clear(struct cpuinfo_x86 *c)
1853 if (mca_cfg.disabled)
1856 if (!mce_available(c))
1860 * Possibly to clear general settings generic to x86
1861 * __mcheck_cpu_clear_generic(c);
1863 __mcheck_cpu_clear_vendor(c);
1867 static void __mce_disable_bank(void *arg)
1869 int bank = *((int *)arg);
1870 __clear_bit(bank, this_cpu_ptr(mce_poll_banks));
1871 cmci_disable_bank(bank);
1874 void mce_disable_bank(int bank)
1876 if (bank >= mca_cfg.banks) {
1878 "Ignoring request to disable invalid MCA bank %d.\n",
1882 set_bit(bank, mce_banks_ce_disabled);
1883 on_each_cpu(__mce_disable_bank, &bank, 1);
1887 * mce=off Disables machine check
1888 * mce=no_cmci Disables CMCI
1889 * mce=no_lmce Disables LMCE
1890 * mce=dont_log_ce Clears corrected events silently, no log created for CEs.
1891 * mce=ignore_ce Disables polling and CMCI, corrected events are not cleared.
1892 * mce=TOLERANCELEVEL[,monarchtimeout] (number, see above)
1893 * monarchtimeout is how long to wait for other CPUs on machine
1894 * check, or 0 to not wait
1895 * mce=bootlog Log MCEs from before booting. Disabled by default on AMD Fam10h
1897 * mce=nobootlog Don't log MCEs from before booting.
1898 * mce=bios_cmci_threshold Don't program the CMCI threshold
1899 * mce=recovery force enable memcpy_mcsafe()
1901 static int __init mcheck_enable(char *str)
1903 struct mca_config *cfg = &mca_cfg;
1911 if (!strcmp(str, "off"))
1913 else if (!strcmp(str, "no_cmci"))
1914 cfg->cmci_disabled = true;
1915 else if (!strcmp(str, "no_lmce"))
1916 cfg->lmce_disabled = 1;
1917 else if (!strcmp(str, "dont_log_ce"))
1918 cfg->dont_log_ce = true;
1919 else if (!strcmp(str, "ignore_ce"))
1920 cfg->ignore_ce = true;
1921 else if (!strcmp(str, "bootlog") || !strcmp(str, "nobootlog"))
1922 cfg->bootlog = (str[0] == 'b');
1923 else if (!strcmp(str, "bios_cmci_threshold"))
1924 cfg->bios_cmci_threshold = 1;
1925 else if (!strcmp(str, "recovery"))
1927 else if (isdigit(str[0])) {
1928 if (get_option(&str, &cfg->tolerant) == 2)
1929 get_option(&str, &(cfg->monarch_timeout));
1931 pr_info("mce argument %s ignored. Please use /sys\n", str);
1936 __setup("mce", mcheck_enable);
1938 int __init mcheck_init(void)
1940 mcheck_intel_therm_init();
1941 mce_register_decode_chain(&first_nb);
1942 mce_register_decode_chain(&mce_srao_nb);
1943 mce_register_decode_chain(&mce_default_nb);
1944 mcheck_vendor_init_severity();
1946 INIT_WORK(&mce_work, mce_gen_pool_process);
1947 init_irq_work(&mce_irq_work, mce_irq_work_cb);
1953 * mce_syscore: PM support
1957 * Disable machine checks on suspend and shutdown. We can't really handle
1960 static void mce_disable_error_reporting(void)
1962 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
1965 for (i = 0; i < mca_cfg.banks; i++) {
1966 struct mce_bank *b = &mce_banks[i];
1969 wrmsrl(msr_ops.ctl(i), 0);
1974 static void vendor_disable_error_reporting(void)
1977 * Don't clear on Intel or AMD or Hygon CPUs. Some of these MSRs
1979 * Disabling them for just a single offlined CPU is bad, since it will
1980 * inhibit reporting for all shared resources on the socket like the
1981 * last level cache (LLC), the integrated memory controller (iMC), etc.
1983 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL ||
1984 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ||
1985 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1988 mce_disable_error_reporting();
1991 static int mce_syscore_suspend(void)
1993 vendor_disable_error_reporting();
1997 static void mce_syscore_shutdown(void)
1999 vendor_disable_error_reporting();
2003 * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
2004 * Only one CPU is active at this time, the others get re-added later using
2007 static void mce_syscore_resume(void)
2009 __mcheck_cpu_init_generic();
2010 __mcheck_cpu_init_vendor(raw_cpu_ptr(&cpu_info));
2011 __mcheck_cpu_init_clear_banks();
2014 static struct syscore_ops mce_syscore_ops = {
2015 .suspend = mce_syscore_suspend,
2016 .shutdown = mce_syscore_shutdown,
2017 .resume = mce_syscore_resume,
2021 * mce_device: Sysfs support
2024 static void mce_cpu_restart(void *data)
2026 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2028 __mcheck_cpu_init_generic();
2029 __mcheck_cpu_init_clear_banks();
2030 __mcheck_cpu_init_timer();
2033 /* Reinit MCEs after user configuration changes */
2034 static void mce_restart(void)
2036 mce_timer_delete_all();
2037 on_each_cpu(mce_cpu_restart, NULL, 1);
2040 /* Toggle features for corrected errors */
2041 static void mce_disable_cmci(void *data)
2043 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2048 static void mce_enable_ce(void *all)
2050 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2055 __mcheck_cpu_init_timer();
2058 static struct bus_type mce_subsys = {
2059 .name = "machinecheck",
2060 .dev_name = "machinecheck",
2063 DEFINE_PER_CPU(struct device *, mce_device);
2065 static inline struct mce_bank_dev *attr_to_bank(struct device_attribute *attr)
2067 return container_of(attr, struct mce_bank_dev, attr);
2070 static ssize_t show_bank(struct device *s, struct device_attribute *attr,
2073 u8 bank = attr_to_bank(attr)->bank;
2076 if (bank >= mca_cfg.banks)
2079 b = &per_cpu(mce_banks_array, s->id)[bank];
2081 return sprintf(buf, "%llx\n", b->ctl);
2084 static ssize_t set_bank(struct device *s, struct device_attribute *attr,
2085 const char *buf, size_t size)
2087 u8 bank = attr_to_bank(attr)->bank;
2091 if (kstrtou64(buf, 0, &new) < 0)
2094 if (bank >= mca_cfg.banks)
2097 b = &per_cpu(mce_banks_array, s->id)[bank];
2105 static ssize_t set_ignore_ce(struct device *s,
2106 struct device_attribute *attr,
2107 const char *buf, size_t size)
2111 if (kstrtou64(buf, 0, &new) < 0)
2114 mutex_lock(&mce_sysfs_mutex);
2115 if (mca_cfg.ignore_ce ^ !!new) {
2117 /* disable ce features */
2118 mce_timer_delete_all();
2119 on_each_cpu(mce_disable_cmci, NULL, 1);
2120 mca_cfg.ignore_ce = true;
2122 /* enable ce features */
2123 mca_cfg.ignore_ce = false;
2124 on_each_cpu(mce_enable_ce, (void *)1, 1);
2127 mutex_unlock(&mce_sysfs_mutex);
2132 static ssize_t set_cmci_disabled(struct device *s,
2133 struct device_attribute *attr,
2134 const char *buf, size_t size)
2138 if (kstrtou64(buf, 0, &new) < 0)
2141 mutex_lock(&mce_sysfs_mutex);
2142 if (mca_cfg.cmci_disabled ^ !!new) {
2145 on_each_cpu(mce_disable_cmci, NULL, 1);
2146 mca_cfg.cmci_disabled = true;
2149 mca_cfg.cmci_disabled = false;
2150 on_each_cpu(mce_enable_ce, NULL, 1);
2153 mutex_unlock(&mce_sysfs_mutex);
2158 static ssize_t store_int_with_restart(struct device *s,
2159 struct device_attribute *attr,
2160 const char *buf, size_t size)
2162 unsigned long old_check_interval = check_interval;
2163 ssize_t ret = device_store_ulong(s, attr, buf, size);
2165 if (check_interval == old_check_interval)
2168 mutex_lock(&mce_sysfs_mutex);
2170 mutex_unlock(&mce_sysfs_mutex);
2175 static DEVICE_INT_ATTR(tolerant, 0644, mca_cfg.tolerant);
2176 static DEVICE_INT_ATTR(monarch_timeout, 0644, mca_cfg.monarch_timeout);
2177 static DEVICE_BOOL_ATTR(dont_log_ce, 0644, mca_cfg.dont_log_ce);
2179 static struct dev_ext_attribute dev_attr_check_interval = {
2180 __ATTR(check_interval, 0644, device_show_int, store_int_with_restart),
2184 static struct dev_ext_attribute dev_attr_ignore_ce = {
2185 __ATTR(ignore_ce, 0644, device_show_bool, set_ignore_ce),
2189 static struct dev_ext_attribute dev_attr_cmci_disabled = {
2190 __ATTR(cmci_disabled, 0644, device_show_bool, set_cmci_disabled),
2191 &mca_cfg.cmci_disabled
2194 static struct device_attribute *mce_device_attrs[] = {
2195 &dev_attr_tolerant.attr,
2196 &dev_attr_check_interval.attr,
2197 #ifdef CONFIG_X86_MCELOG_LEGACY
2200 &dev_attr_monarch_timeout.attr,
2201 &dev_attr_dont_log_ce.attr,
2202 &dev_attr_ignore_ce.attr,
2203 &dev_attr_cmci_disabled.attr,
2207 static cpumask_var_t mce_device_initialized;
2209 static void mce_device_release(struct device *dev)
2214 /* Per CPU device init. All of the CPUs still share the same bank device: */
2215 static int mce_device_create(unsigned int cpu)
2221 if (!mce_available(&boot_cpu_data))
2224 dev = per_cpu(mce_device, cpu);
2228 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2232 dev->bus = &mce_subsys;
2233 dev->release = &mce_device_release;
2235 err = device_register(dev);
2241 for (i = 0; mce_device_attrs[i]; i++) {
2242 err = device_create_file(dev, mce_device_attrs[i]);
2246 for (j = 0; j < mca_cfg.banks; j++) {
2247 err = device_create_file(dev, &mce_bank_devs[j].attr);
2251 cpumask_set_cpu(cpu, mce_device_initialized);
2252 per_cpu(mce_device, cpu) = dev;
2257 device_remove_file(dev, &mce_bank_devs[j].attr);
2260 device_remove_file(dev, mce_device_attrs[i]);
2262 device_unregister(dev);
2267 static void mce_device_remove(unsigned int cpu)
2269 struct device *dev = per_cpu(mce_device, cpu);
2272 if (!cpumask_test_cpu(cpu, mce_device_initialized))
2275 for (i = 0; mce_device_attrs[i]; i++)
2276 device_remove_file(dev, mce_device_attrs[i]);
2278 for (i = 0; i < mca_cfg.banks; i++)
2279 device_remove_file(dev, &mce_bank_devs[i].attr);
2281 device_unregister(dev);
2282 cpumask_clear_cpu(cpu, mce_device_initialized);
2283 per_cpu(mce_device, cpu) = NULL;
2286 /* Make sure there are no machine checks on offlined CPUs. */
2287 static void mce_disable_cpu(void)
2289 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2292 if (!cpuhp_tasks_frozen)
2295 vendor_disable_error_reporting();
2298 static void mce_reenable_cpu(void)
2300 struct mce_bank *mce_banks = this_cpu_ptr(mce_banks_array);
2303 if (!mce_available(raw_cpu_ptr(&cpu_info)))
2306 if (!cpuhp_tasks_frozen)
2308 for (i = 0; i < mca_cfg.banks; i++) {
2309 struct mce_bank *b = &mce_banks[i];
2312 wrmsrl(msr_ops.ctl(i), b->ctl);
2316 static int mce_cpu_dead(unsigned int cpu)
2318 mce_intel_hcpu_update(cpu);
2320 /* intentionally ignoring frozen here */
2321 if (!cpuhp_tasks_frozen)
2326 static int mce_cpu_online(unsigned int cpu)
2328 struct timer_list *t = this_cpu_ptr(&mce_timer);
2331 mce_device_create(cpu);
2333 ret = mce_threshold_create_device(cpu);
2335 mce_device_remove(cpu);
2343 static int mce_cpu_pre_down(unsigned int cpu)
2345 struct timer_list *t = this_cpu_ptr(&mce_timer);
2349 mce_threshold_remove_device(cpu);
2350 mce_device_remove(cpu);
2354 static __init void mce_init_banks(void)
2358 for (i = 0; i < MAX_NR_BANKS; i++) {
2359 struct mce_bank_dev *b = &mce_bank_devs[i];
2360 struct device_attribute *a = &b->attr;
2364 sysfs_attr_init(&a->attr);
2365 a->attr.name = b->attrname;
2366 snprintf(b->attrname, ATTR_LEN, "bank%d", i);
2368 a->attr.mode = 0644;
2369 a->show = show_bank;
2370 a->store = set_bank;
2374 static __init int mcheck_init_device(void)
2379 * Check if we have a spare virtual bit. This will only become
2380 * a problem if/when we move beyond 5-level page tables.
2382 MAYBE_BUILD_BUG_ON(__VIRTUAL_MASK_SHIFT >= 63);
2384 if (!mce_available(&boot_cpu_data)) {
2389 if (!zalloc_cpumask_var(&mce_device_initialized, GFP_KERNEL)) {
2396 err = subsys_system_register(&mce_subsys, NULL);
2400 err = cpuhp_setup_state(CPUHP_X86_MCE_DEAD, "x86/mce:dead", NULL,
2405 err = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "x86/mce:online",
2406 mce_cpu_online, mce_cpu_pre_down);
2408 goto err_out_online;
2410 register_syscore_ops(&mce_syscore_ops);
2415 cpuhp_remove_state(CPUHP_X86_MCE_DEAD);
2418 free_cpumask_var(mce_device_initialized);
2421 pr_err("Unable to init MCE device (rc: %d)\n", err);
2425 device_initcall_sync(mcheck_init_device);
2428 * Old style boot options parsing. Only for compatibility.
2430 static int __init mcheck_disable(char *str)
2432 mca_cfg.disabled = 1;
2435 __setup("nomce", mcheck_disable);
2437 #ifdef CONFIG_DEBUG_FS
2438 struct dentry *mce_get_debugfs_dir(void)
2440 static struct dentry *dmce;
2443 dmce = debugfs_create_dir("mce", NULL);
2448 static void mce_reset(void)
2451 atomic_set(&mce_fake_panicked, 0);
2452 atomic_set(&mce_executing, 0);
2453 atomic_set(&mce_callin, 0);
2454 atomic_set(&global_nwo, 0);
2457 static int fake_panic_get(void *data, u64 *val)
2463 static int fake_panic_set(void *data, u64 val)
2470 DEFINE_DEBUGFS_ATTRIBUTE(fake_panic_fops, fake_panic_get, fake_panic_set,
2473 static int __init mcheck_debugfs_init(void)
2475 struct dentry *dmce, *ffake_panic;
2477 dmce = mce_get_debugfs_dir();
2480 ffake_panic = debugfs_create_file_unsafe("fake_panic", 0444, dmce,
2481 NULL, &fake_panic_fops);
2488 static int __init mcheck_debugfs_init(void) { return -EINVAL; }
2491 DEFINE_STATIC_KEY_FALSE(mcsafe_key);
2492 EXPORT_SYMBOL_GPL(mcsafe_key);
2494 static int __init mcheck_late_init(void)
2496 pr_info("Using %d MCE banks\n", mca_cfg.banks);
2498 if (mca_cfg.recovery)
2499 static_branch_inc(&mcsafe_key);
2501 mcheck_debugfs_init();
2505 * Flush out everything that has been logged during early boot, now that
2506 * everything has been initialized (workqueues, decoders, ...).
2508 mce_schedule_work();
2512 late_initcall(mcheck_late_init);