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
10 #include <linux/thread_info.h>
11 #include <linux/capability.h>
12 #include <linux/miscdevice.h>
13 #include <linux/ratelimit.h>
14 #include <linux/kallsyms.h>
15 #include <linux/rcupdate.h>
16 #include <linux/kobject.h>
17 #include <linux/kdebug.h>
18 #include <linux/kernel.h>
19 #include <linux/percpu.h>
20 #include <linux/string.h>
21 #include <linux/sysdev.h>
22 #include <linux/ctype.h>
23 #include <linux/sched.h>
24 #include <linux/sysfs.h>
25 #include <linux/types.h>
26 #include <linux/init.h>
27 #include <linux/kmod.h>
28 #include <linux/poll.h>
29 #include <linux/cpu.h>
32 #include <asm/processor.h>
33 #include <asm/uaccess.h>
41 /* Handle unconfigured int18 (should never happen) */
42 static void unexpected_machine_check(struct pt_regs *regs, long error_code)
44 printk(KERN_ERR "CPU#%d: Unexpected int18 (Machine Check).\n",
48 /* Call the installed machine check handler for this CPU setup. */
49 void (*machine_check_vector)(struct pt_regs *, long error_code) =
50 unexpected_machine_check;
54 #ifdef CONFIG_X86_NEW_MCE
56 #define MISC_MCELOG_MINOR 227
62 * 0: always panic on uncorrected errors, log corrected errors
63 * 1: panic or SIGBUS on uncorrected errors, log corrected errors
64 * 2: SIGBUS or log uncorrected errors (if possible), log corrected errors
65 * 3: never panic or SIGBUS, log all errors (for testing only)
67 static int tolerant = 1;
70 static unsigned long notify_user;
72 static int mce_bootlog = -1;
74 static char trigger[128];
75 static char *trigger_argv[2] = { trigger, NULL };
77 static unsigned long dont_init_banks;
79 static DECLARE_WAIT_QUEUE_HEAD(mce_wait);
81 /* MCA banks polled by the period polling timer for corrected events */
82 DEFINE_PER_CPU(mce_banks_t, mce_poll_banks) = {
83 [0 ... BITS_TO_LONGS(MAX_NR_BANKS)-1] = ~0UL
86 static inline int skip_bank_init(int i)
88 return i < BITS_PER_LONG && test_bit(i, &dont_init_banks);
91 /* Do initial initialization of a struct mce */
92 void mce_setup(struct mce *m)
94 memset(m, 0, sizeof(struct mce));
95 m->cpu = smp_processor_id();
99 DEFINE_PER_CPU(struct mce, injectm);
100 EXPORT_PER_CPU_SYMBOL_GPL(injectm);
103 * Lockless MCE logging infrastructure.
104 * This avoids deadlocks on printk locks without having to break locks. Also
105 * separate MCEs from kernel messages to avoid bogus bug reports.
108 static struct mce_log mcelog = {
113 void mce_log(struct mce *mce)
115 unsigned next, entry;
120 entry = rcu_dereference(mcelog.next);
123 * When the buffer fills up discard new entries.
124 * Assume that the earlier errors are the more
127 if (entry >= MCE_LOG_LEN) {
128 set_bit(MCE_OVERFLOW, (unsigned long *)&mcelog.flags);
131 /* Old left over entry. Skip: */
132 if (mcelog.entry[entry].finished) {
140 if (cmpxchg(&mcelog.next, entry, next) == entry)
143 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
145 mcelog.entry[entry].finished = 1;
148 set_bit(0, ¬ify_user);
151 static void print_mce(struct mce *m)
153 printk(KERN_EMERG "\n"
154 KERN_EMERG "HARDWARE ERROR\n"
156 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
157 m->cpu, m->mcgstatus, m->bank, m->status);
159 printk(KERN_EMERG "RIP%s %02x:<%016Lx> ",
160 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
162 if (m->cs == __KERNEL_CS)
163 print_symbol("{%s}", m->ip);
166 printk(KERN_EMERG "TSC %llx ", m->tsc);
168 printk("ADDR %llx ", m->addr);
170 printk("MISC %llx ", m->misc);
172 printk(KERN_EMERG "This is not a software problem!\n");
173 printk(KERN_EMERG "Run through mcelog --ascii to decode "
174 "and contact your hardware vendor\n");
177 static void mce_panic(char *msg, struct mce *backup, u64 start)
183 for (i = 0; i < MCE_LOG_LEN; i++) {
184 u64 tsc = mcelog.entry[i].tsc;
186 if ((s64)(tsc - start) < 0)
188 print_mce(&mcelog.entry[i]);
189 if (backup && mcelog.entry[i].tsc == backup->tsc)
197 /* Support code for software error injection */
199 static int msr_to_offset(u32 msr)
201 unsigned bank = __get_cpu_var(injectm.bank);
203 return offsetof(struct mce, ip);
204 if (msr == MSR_IA32_MC0_STATUS + bank*4)
205 return offsetof(struct mce, status);
206 if (msr == MSR_IA32_MC0_ADDR + bank*4)
207 return offsetof(struct mce, addr);
208 if (msr == MSR_IA32_MC0_MISC + bank*4)
209 return offsetof(struct mce, misc);
210 if (msr == MSR_IA32_MCG_STATUS)
211 return offsetof(struct mce, mcgstatus);
215 /* MSR access wrappers used for error injection */
216 static u64 mce_rdmsrl(u32 msr)
219 if (__get_cpu_var(injectm).finished) {
220 int offset = msr_to_offset(msr);
223 return *(u64 *)((char *)&__get_cpu_var(injectm) + offset);
229 static void mce_wrmsrl(u32 msr, u64 v)
231 if (__get_cpu_var(injectm).finished) {
232 int offset = msr_to_offset(msr);
234 *(u64 *)((char *)&__get_cpu_var(injectm) + offset) = v;
240 int mce_available(struct cpuinfo_x86 *c)
244 return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
247 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
249 if (regs && (m->mcgstatus & MCG_STATUS_RIPV)) {
257 /* Assume the RIP in the MSR is exact. Is this true? */
258 m->mcgstatus |= MCG_STATUS_EIPV;
259 m->ip = mce_rdmsrl(rip_msr);
265 * Poll for corrected events or events that happened before reset.
266 * Those are just logged through /dev/mcelog.
268 * This is executed in standard interrupt context.
270 void machine_check_poll(enum mcp_flags flags, mce_banks_t *b)
277 m.mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
278 for (i = 0; i < banks; i++) {
279 if (!bank[i] || !test_bit(i, *b))
288 m.status = mce_rdmsrl(MSR_IA32_MC0_STATUS + i*4);
289 if (!(m.status & MCI_STATUS_VAL))
293 * Uncorrected events are handled by the exception handler
294 * when it is enabled. But when the exception is disabled log
297 * TBD do the same check for MCI_STATUS_EN here?
299 if ((m.status & MCI_STATUS_UC) && !(flags & MCP_UC))
302 if (m.status & MCI_STATUS_MISCV)
303 m.misc = mce_rdmsrl(MSR_IA32_MC0_MISC + i*4);
304 if (m.status & MCI_STATUS_ADDRV)
305 m.addr = mce_rdmsrl(MSR_IA32_MC0_ADDR + i*4);
307 if (!(flags & MCP_TIMESTAMP))
310 * Don't get the IP here because it's unlikely to
311 * have anything to do with the actual error location.
313 if (!(flags & MCP_DONTLOG)) {
315 add_taint(TAINT_MACHINE_CHECK);
319 * Clear state for this bank.
321 mce_wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
325 * Don't clear MCG_STATUS here because it's only defined for
331 EXPORT_SYMBOL_GPL(machine_check_poll);
334 * The actual machine check handler. This only handles real
335 * exceptions when something got corrupted coming in through int 18.
337 * This is executed in NMI context not subject to normal locking rules. This
338 * implies that most kernel services cannot be safely used. Don't even
339 * think about putting a printk in there!
341 void do_machine_check(struct pt_regs *regs, long error_code)
343 struct mce m, panicm;
344 int panicm_found = 0;
348 * If no_way_out gets set, there is no safe way to recover from this
349 * MCE. If tolerant is cranked up, we'll try anyway.
353 * If kill_it gets set, there might be a way to recover from this
357 DECLARE_BITMAP(toclear, MAX_NR_BANKS);
359 atomic_inc(&mce_entry);
361 if (notify_die(DIE_NMI, "machine check", regs, error_code,
362 18, SIGKILL) == NOTIFY_STOP)
369 m.mcgstatus = mce_rdmsrl(MSR_IA32_MCG_STATUS);
371 /* if the restart IP is not valid, we're done for */
372 if (!(m.mcgstatus & MCG_STATUS_RIPV))
378 for (i = 0; i < banks; i++) {
379 __clear_bit(i, toclear);
387 m.status = mce_rdmsrl(MSR_IA32_MC0_STATUS + i*4);
388 if ((m.status & MCI_STATUS_VAL) == 0)
392 * Non uncorrected errors are handled by machine_check_poll
395 if ((m.status & MCI_STATUS_UC) == 0)
399 * Set taint even when machine check was not enabled.
401 add_taint(TAINT_MACHINE_CHECK);
403 __set_bit(i, toclear);
405 if (m.status & MCI_STATUS_EN) {
406 /* if PCC was set, there's no way out */
407 no_way_out |= !!(m.status & MCI_STATUS_PCC);
409 * If this error was uncorrectable and there was
410 * an overflow, we're in trouble. If no overflow,
411 * we might get away with just killing a task.
413 if (m.status & MCI_STATUS_UC) {
414 if (tolerant < 1 || m.status & MCI_STATUS_OVER)
420 * Machine check event was not enabled. Clear, but
426 if (m.status & MCI_STATUS_MISCV)
427 m.misc = mce_rdmsrl(MSR_IA32_MC0_MISC + i*4);
428 if (m.status & MCI_STATUS_ADDRV)
429 m.addr = mce_rdmsrl(MSR_IA32_MC0_ADDR + i*4);
431 mce_get_rip(&m, regs);
435 * Did this bank cause the exception?
437 * Assume that the bank with uncorrectable errors did it,
438 * and that there is only a single one:
440 if ((m.status & MCI_STATUS_UC) &&
441 (m.status & MCI_STATUS_EN)) {
448 * If we didn't find an uncorrectable error, pick
449 * the last one (shouldn't happen, just being safe).
455 * If we have decided that we just CAN'T continue, and the user
456 * has not set tolerant to an insane level, give up and die.
458 if (no_way_out && tolerant < 3)
459 mce_panic("Machine check", &panicm, mcestart);
462 * If the error seems to be unrecoverable, something should be
463 * done. Try to kill as little as possible. If we can kill just
464 * one task, do that. If the user has set the tolerance very
465 * high, don't try to do anything at all.
467 if (kill_it && tolerant < 3) {
471 * If the EIPV bit is set, it means the saved IP is the
472 * instruction which caused the MCE.
474 if (m.mcgstatus & MCG_STATUS_EIPV)
475 user_space = panicm.ip && (panicm.cs & 3);
478 * If we know that the error was in user space, send a
479 * SIGBUS. Otherwise, panic if tolerance is low.
481 * force_sig() takes an awful lot of locks and has a slight
482 * risk of deadlocking.
485 force_sig(SIGBUS, current);
486 } else if (panic_on_oops || tolerant < 2) {
487 mce_panic("Uncorrected machine check",
492 /* notify userspace ASAP */
493 set_thread_flag(TIF_MCE_NOTIFY);
495 /* the last thing we do is clear state */
496 for (i = 0; i < banks; i++) {
497 if (test_bit(i, toclear))
498 mce_wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
500 mce_wrmsrl(MSR_IA32_MCG_STATUS, 0);
502 atomic_dec(&mce_entry);
505 EXPORT_SYMBOL_GPL(do_machine_check);
507 #ifdef CONFIG_X86_MCE_INTEL
509 * mce_log_therm_throt_event - Logs the thermal throttling event to mcelog
510 * @cpu: The CPU on which the event occurred.
511 * @status: Event status information
513 * This function should be called by the thermal interrupt after the
514 * event has been processed and the decision was made to log the event
517 * The status parameter will be saved to the 'status' field of 'struct mce'
518 * and historically has been the register value of the
519 * MSR_IA32_THERMAL_STATUS (Intel) msr.
521 void mce_log_therm_throt_event(__u64 status)
526 m.bank = MCE_THERMAL_BANK;
530 #endif /* CONFIG_X86_MCE_INTEL */
533 * Periodic polling timer for "silent" machine check errors. If the
534 * poller finds an MCE, poll 2x faster. When the poller finds no more
535 * errors, poll 2x slower (up to check_interval seconds).
537 static int check_interval = 5 * 60; /* 5 minutes */
539 static DEFINE_PER_CPU(int, next_interval); /* in jiffies */
540 static DEFINE_PER_CPU(struct timer_list, mce_timer);
542 static void mcheck_timer(unsigned long data)
544 struct timer_list *t = &per_cpu(mce_timer, data);
547 WARN_ON(smp_processor_id() != data);
549 if (mce_available(¤t_cpu_data)) {
550 machine_check_poll(MCP_TIMESTAMP,
551 &__get_cpu_var(mce_poll_banks));
555 * Alert userspace if needed. If we logged an MCE, reduce the
556 * polling interval, otherwise increase the polling interval.
558 n = &__get_cpu_var(next_interval);
559 if (mce_notify_user()) {
560 *n = max(*n/2, HZ/100);
562 *n = min(*n*2, (int)round_jiffies_relative(check_interval*HZ));
565 t->expires = jiffies + *n;
569 static void mce_do_trigger(struct work_struct *work)
571 call_usermodehelper(trigger, trigger_argv, NULL, UMH_NO_WAIT);
574 static DECLARE_WORK(mce_trigger_work, mce_do_trigger);
577 * Notify the user(s) about new machine check events.
578 * Can be called from interrupt context, but not from machine check/NMI
581 int mce_notify_user(void)
583 /* Not more than two messages every minute */
584 static DEFINE_RATELIMIT_STATE(ratelimit, 60*HZ, 2);
586 clear_thread_flag(TIF_MCE_NOTIFY);
588 if (test_and_clear_bit(0, ¬ify_user)) {
589 wake_up_interruptible(&mce_wait);
592 * There is no risk of missing notifications because
593 * work_pending is always cleared before the function is
596 if (trigger[0] && !work_pending(&mce_trigger_work))
597 schedule_work(&mce_trigger_work);
599 if (__ratelimit(&ratelimit))
600 printk(KERN_INFO "Machine check events logged\n");
606 EXPORT_SYMBOL_GPL(mce_notify_user);
609 * Initialize Machine Checks for a CPU.
611 static int mce_cap_init(void)
616 rdmsrl(MSR_IA32_MCG_CAP, cap);
618 b = cap & MCG_BANKCNT_MASK;
619 printk(KERN_INFO "mce: CPU supports %d MCE banks\n", b);
621 if (b > MAX_NR_BANKS) {
623 "MCE: Using only %u machine check banks out of %u\n",
628 /* Don't support asymmetric configurations today */
629 WARN_ON(banks != 0 && b != banks);
632 bank = kmalloc(banks * sizeof(u64), GFP_KERNEL);
635 memset(bank, 0xff, banks * sizeof(u64));
638 /* Use accurate RIP reporting if available. */
639 if ((cap & MCG_EXT_P) && MCG_EXT_CNT(cap) >= 9)
640 rip_msr = MSR_IA32_MCG_EIP;
645 static void mce_init(void)
647 mce_banks_t all_banks;
652 * Log the machine checks left over from the previous reset.
654 bitmap_fill(all_banks, MAX_NR_BANKS);
655 machine_check_poll(MCP_UC|(!mce_bootlog ? MCP_DONTLOG : 0), &all_banks);
657 set_in_cr4(X86_CR4_MCE);
659 rdmsrl(MSR_IA32_MCG_CAP, cap);
661 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
663 for (i = 0; i < banks; i++) {
664 if (skip_bank_init(i))
666 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
667 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
671 /* Add per CPU specific workarounds here */
672 static void mce_cpu_quirks(struct cpuinfo_x86 *c)
674 /* This should be disabled by the BIOS, but isn't always */
675 if (c->x86_vendor == X86_VENDOR_AMD) {
676 if (c->x86 == 15 && banks > 4) {
678 * disable GART TBL walk error reporting, which
679 * trips off incorrectly with the IOMMU & 3ware
682 clear_bit(10, (unsigned long *)&bank[4]);
684 if (c->x86 <= 17 && mce_bootlog < 0) {
686 * Lots of broken BIOS around that don't clear them
687 * by default and leave crap in there. Don't log:
692 * Various K7s with broken bank 0 around. Always disable
699 if (c->x86_vendor == X86_VENDOR_INTEL) {
701 * SDM documents that on family 6 bank 0 should not be written
702 * because it aliases to another special BIOS controlled
704 * But it's not aliased anymore on model 0x1a+
705 * Don't ignore bank 0 completely because there could be a
706 * valid event later, merely don't write CTL0.
709 if (c->x86 == 6 && c->x86_model < 0x1A)
710 __set_bit(0, &dont_init_banks);
714 static void __cpuinit mce_ancient_init(struct cpuinfo_x86 *c)
718 switch (c->x86_vendor) {
719 case X86_VENDOR_INTEL:
720 if (mce_p5_enabled())
721 intel_p5_mcheck_init(c);
723 case X86_VENDOR_CENTAUR:
724 winchip_mcheck_init(c);
729 static void mce_cpu_features(struct cpuinfo_x86 *c)
731 switch (c->x86_vendor) {
732 case X86_VENDOR_INTEL:
733 mce_intel_feature_init(c);
736 mce_amd_feature_init(c);
743 static void mce_init_timer(void)
745 struct timer_list *t = &__get_cpu_var(mce_timer);
746 int *n = &__get_cpu_var(next_interval);
748 *n = check_interval * HZ;
751 setup_timer(t, mcheck_timer, smp_processor_id());
752 t->expires = round_jiffies(jiffies + *n);
757 * Called for each booted CPU to set up machine checks.
758 * Must be called with preempt off:
760 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
767 if (!mce_available(c))
770 if (mce_cap_init() < 0) {
776 machine_check_vector = do_machine_check;
784 * Character device to read and clear the MCE log.
787 static DEFINE_SPINLOCK(mce_state_lock);
788 static int open_count; /* #times opened */
789 static int open_exclu; /* already open exclusive? */
791 static int mce_open(struct inode *inode, struct file *file)
793 spin_lock(&mce_state_lock);
795 if (open_exclu || (open_count && (file->f_flags & O_EXCL))) {
796 spin_unlock(&mce_state_lock);
801 if (file->f_flags & O_EXCL)
805 spin_unlock(&mce_state_lock);
807 return nonseekable_open(inode, file);
810 static int mce_release(struct inode *inode, struct file *file)
812 spin_lock(&mce_state_lock);
817 spin_unlock(&mce_state_lock);
822 static void collect_tscs(void *data)
824 unsigned long *cpu_tsc = (unsigned long *)data;
826 rdtscll(cpu_tsc[smp_processor_id()]);
829 static DEFINE_MUTEX(mce_read_mutex);
831 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize,
834 char __user *buf = ubuf;
835 unsigned long *cpu_tsc;
839 cpu_tsc = kmalloc(nr_cpu_ids * sizeof(long), GFP_KERNEL);
843 mutex_lock(&mce_read_mutex);
844 next = rcu_dereference(mcelog.next);
846 /* Only supports full reads right now */
847 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
848 mutex_unlock(&mce_read_mutex);
857 for (i = prev; i < next; i++) {
858 unsigned long start = jiffies;
860 while (!mcelog.entry[i].finished) {
861 if (time_after_eq(jiffies, start + 2)) {
862 memset(mcelog.entry + i, 0,
869 err |= copy_to_user(buf, mcelog.entry + i,
871 buf += sizeof(struct mce);
876 memset(mcelog.entry + prev, 0,
877 (next - prev) * sizeof(struct mce));
879 next = cmpxchg(&mcelog.next, prev, 0);
880 } while (next != prev);
885 * Collect entries that were still getting written before the
888 on_each_cpu(collect_tscs, cpu_tsc, 1);
890 for (i = next; i < MCE_LOG_LEN; i++) {
891 if (mcelog.entry[i].finished &&
892 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
893 err |= copy_to_user(buf, mcelog.entry+i,
896 buf += sizeof(struct mce);
897 memset(&mcelog.entry[i], 0, sizeof(struct mce));
900 mutex_unlock(&mce_read_mutex);
903 return err ? -EFAULT : buf - ubuf;
906 static unsigned int mce_poll(struct file *file, poll_table *wait)
908 poll_wait(file, &mce_wait, wait);
909 if (rcu_dereference(mcelog.next))
910 return POLLIN | POLLRDNORM;
914 static long mce_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
916 int __user *p = (int __user *)arg;
918 if (!capable(CAP_SYS_ADMIN))
922 case MCE_GET_RECORD_LEN:
923 return put_user(sizeof(struct mce), p);
924 case MCE_GET_LOG_LEN:
925 return put_user(MCE_LOG_LEN, p);
926 case MCE_GETCLEAR_FLAGS: {
930 flags = mcelog.flags;
931 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
933 return put_user(flags, p);
940 /* Modified in mce-inject.c, so not static or const */
941 struct file_operations mce_chrdev_ops = {
943 .release = mce_release,
946 .unlocked_ioctl = mce_ioctl,
948 EXPORT_SYMBOL_GPL(mce_chrdev_ops);
950 static struct miscdevice mce_log_device = {
957 * mce=off disables machine check
958 * mce=TOLERANCELEVEL (number, see above)
959 * mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
960 * mce=nobootlog Don't log MCEs from before booting.
962 static int __init mcheck_enable(char *str)
968 if (!strcmp(str, "off"))
970 else if (!strcmp(str, "bootlog") || !strcmp(str, "nobootlog"))
971 mce_bootlog = (str[0] == 'b');
972 else if (isdigit(str[0]))
973 get_option(&str, &tolerant);
975 printk(KERN_INFO "mce argument %s ignored. Please use /sys\n",
981 __setup("mce", mcheck_enable);
988 * Disable machine checks on suspend and shutdown. We can't really handle
991 static int mce_disable(void)
995 for (i = 0; i < banks; i++) {
996 if (!skip_bank_init(i))
997 wrmsrl(MSR_IA32_MC0_CTL + i*4, 0);
1002 static int mce_suspend(struct sys_device *dev, pm_message_t state)
1004 return mce_disable();
1007 static int mce_shutdown(struct sys_device *dev)
1009 return mce_disable();
1013 * On resume clear all MCE state. Don't want to see leftovers from the BIOS.
1014 * Only one CPU is active at this time, the others get re-added later using
1017 static int mce_resume(struct sys_device *dev)
1020 mce_cpu_features(¤t_cpu_data);
1025 static void mce_cpu_restart(void *data)
1027 del_timer_sync(&__get_cpu_var(mce_timer));
1028 if (mce_available(¤t_cpu_data))
1033 /* Reinit MCEs after user configuration changes */
1034 static void mce_restart(void)
1036 on_each_cpu(mce_cpu_restart, NULL, 1);
1039 static struct sysdev_class mce_sysclass = {
1040 .suspend = mce_suspend,
1041 .shutdown = mce_shutdown,
1042 .resume = mce_resume,
1043 .name = "machinecheck",
1046 DEFINE_PER_CPU(struct sys_device, mce_dev);
1049 void (*threshold_cpu_callback)(unsigned long action, unsigned int cpu);
1051 static struct sysdev_attribute *bank_attrs;
1053 static ssize_t show_bank(struct sys_device *s, struct sysdev_attribute *attr,
1056 u64 b = bank[attr - bank_attrs];
1058 return sprintf(buf, "%llx\n", b);
1061 static ssize_t set_bank(struct sys_device *s, struct sysdev_attribute *attr,
1062 const char *buf, size_t siz)
1065 u64 new = simple_strtoull(buf, &end, 0);
1070 bank[attr - bank_attrs] = new;
1077 show_trigger(struct sys_device *s, struct sysdev_attribute *attr, char *buf)
1079 strcpy(buf, trigger);
1081 return strlen(trigger) + 1;
1084 static ssize_t set_trigger(struct sys_device *s, struct sysdev_attribute *attr,
1085 const char *buf, size_t siz)
1090 strncpy(trigger, buf, sizeof(trigger));
1091 trigger[sizeof(trigger)-1] = 0;
1092 len = strlen(trigger);
1093 p = strchr(trigger, '\n');
1101 static ssize_t store_int_with_restart(struct sys_device *s,
1102 struct sysdev_attribute *attr,
1103 const char *buf, size_t size)
1105 ssize_t ret = sysdev_store_int(s, attr, buf, size);
1110 static SYSDEV_ATTR(trigger, 0644, show_trigger, set_trigger);
1111 static SYSDEV_INT_ATTR(tolerant, 0644, tolerant);
1113 static struct sysdev_ext_attribute attr_check_interval = {
1114 _SYSDEV_ATTR(check_interval, 0644, sysdev_show_int,
1115 store_int_with_restart),
1119 static struct sysdev_attribute *mce_attrs[] = {
1120 &attr_tolerant.attr, &attr_check_interval.attr, &attr_trigger,
1124 static cpumask_var_t mce_dev_initialized;
1126 /* Per cpu sysdev init. All of the cpus still share the same ctrl bank: */
1127 static __cpuinit int mce_create_device(unsigned int cpu)
1132 if (!mce_available(&boot_cpu_data))
1135 memset(&per_cpu(mce_dev, cpu).kobj, 0, sizeof(struct kobject));
1136 per_cpu(mce_dev, cpu).id = cpu;
1137 per_cpu(mce_dev, cpu).cls = &mce_sysclass;
1139 err = sysdev_register(&per_cpu(mce_dev, cpu));
1143 for (i = 0; mce_attrs[i]; i++) {
1144 err = sysdev_create_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1148 for (i = 0; i < banks; i++) {
1149 err = sysdev_create_file(&per_cpu(mce_dev, cpu),
1154 cpumask_set_cpu(cpu, mce_dev_initialized);
1159 sysdev_remove_file(&per_cpu(mce_dev, cpu), &bank_attrs[i]);
1162 sysdev_remove_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1164 sysdev_unregister(&per_cpu(mce_dev, cpu));
1169 static __cpuinit void mce_remove_device(unsigned int cpu)
1173 if (!cpumask_test_cpu(cpu, mce_dev_initialized))
1176 for (i = 0; mce_attrs[i]; i++)
1177 sysdev_remove_file(&per_cpu(mce_dev, cpu), mce_attrs[i]);
1179 for (i = 0; i < banks; i++)
1180 sysdev_remove_file(&per_cpu(mce_dev, cpu), &bank_attrs[i]);
1182 sysdev_unregister(&per_cpu(mce_dev, cpu));
1183 cpumask_clear_cpu(cpu, mce_dev_initialized);
1186 /* Make sure there are no machine checks on offlined CPUs. */
1187 static void mce_disable_cpu(void *h)
1189 unsigned long action = *(unsigned long *)h;
1192 if (!mce_available(¤t_cpu_data))
1194 if (!(action & CPU_TASKS_FROZEN))
1196 for (i = 0; i < banks; i++) {
1197 if (!skip_bank_init(i))
1198 wrmsrl(MSR_IA32_MC0_CTL + i*4, 0);
1202 static void mce_reenable_cpu(void *h)
1204 unsigned long action = *(unsigned long *)h;
1207 if (!mce_available(¤t_cpu_data))
1210 if (!(action & CPU_TASKS_FROZEN))
1212 for (i = 0; i < banks; i++) {
1213 if (!skip_bank_init(i))
1214 wrmsrl(MSR_IA32_MC0_CTL + i*4, bank[i]);
1218 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
1219 static int __cpuinit
1220 mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
1222 unsigned int cpu = (unsigned long)hcpu;
1223 struct timer_list *t = &per_cpu(mce_timer, cpu);
1227 case CPU_ONLINE_FROZEN:
1228 mce_create_device(cpu);
1229 if (threshold_cpu_callback)
1230 threshold_cpu_callback(action, cpu);
1233 case CPU_DEAD_FROZEN:
1234 if (threshold_cpu_callback)
1235 threshold_cpu_callback(action, cpu);
1236 mce_remove_device(cpu);
1238 case CPU_DOWN_PREPARE:
1239 case CPU_DOWN_PREPARE_FROZEN:
1241 smp_call_function_single(cpu, mce_disable_cpu, &action, 1);
1243 case CPU_DOWN_FAILED:
1244 case CPU_DOWN_FAILED_FROZEN:
1245 t->expires = round_jiffies(jiffies +
1246 __get_cpu_var(next_interval));
1247 add_timer_on(t, cpu);
1248 smp_call_function_single(cpu, mce_reenable_cpu, &action, 1);
1251 /* intentionally ignoring frozen here */
1252 cmci_rediscover(cpu);
1258 static struct notifier_block mce_cpu_notifier __cpuinitdata = {
1259 .notifier_call = mce_cpu_callback,
1262 static __init int mce_init_banks(void)
1266 bank_attrs = kzalloc(sizeof(struct sysdev_attribute) * banks,
1271 for (i = 0; i < banks; i++) {
1272 struct sysdev_attribute *a = &bank_attrs[i];
1274 a->attr.name = kasprintf(GFP_KERNEL, "bank%d", i);
1278 a->attr.mode = 0644;
1279 a->show = show_bank;
1280 a->store = set_bank;
1286 kfree(bank_attrs[i].attr.name);
1293 static __init int mce_init_device(void)
1298 if (!mce_available(&boot_cpu_data))
1301 alloc_cpumask_var(&mce_dev_initialized, GFP_KERNEL);
1303 err = mce_init_banks();
1307 err = sysdev_class_register(&mce_sysclass);
1311 for_each_online_cpu(i) {
1312 err = mce_create_device(i);
1317 register_hotcpu_notifier(&mce_cpu_notifier);
1318 misc_register(&mce_log_device);
1323 device_initcall(mce_init_device);
1325 #else /* CONFIG_X86_OLD_MCE: */
1328 EXPORT_SYMBOL_GPL(nr_mce_banks); /* non-fatal.o */
1330 /* This has to be run for each processor */
1331 void mcheck_init(struct cpuinfo_x86 *c)
1333 if (mce_disabled == 1)
1336 switch (c->x86_vendor) {
1337 case X86_VENDOR_AMD:
1341 case X86_VENDOR_INTEL:
1343 intel_p5_mcheck_init(c);
1345 intel_p6_mcheck_init(c);
1347 intel_p4_mcheck_init(c);
1350 case X86_VENDOR_CENTAUR:
1352 winchip_mcheck_init(c);
1358 printk(KERN_INFO "mce: CPU supports %d MCE banks\n", nr_mce_banks);
1361 static int __init mcheck_enable(char *str)
1367 __setup("mce", mcheck_enable);
1369 #endif /* CONFIG_X86_OLD_MCE */
1372 * Old style boot options parsing. Only for compatibility.
1374 static int __init mcheck_disable(char *str)
1379 __setup("nomce", mcheck_disable);