2 * Machine check handler.
3 * K8 parts Copyright 2002,2003 Andi Kleen, SuSE Labs.
4 * Rest from unknown author(s).
5 * 2004 Andi Kleen. Rewrote most of it.
8 #include <linux/init.h>
9 #include <linux/types.h>
10 #include <linux/kernel.h>
11 #include <linux/sched.h>
12 #include <linux/string.h>
13 #include <linux/rcupdate.h>
14 #include <linux/kallsyms.h>
15 #include <linux/sysdev.h>
16 #include <linux/miscdevice.h>
18 #include <linux/capability.h>
19 #include <linux/cpu.h>
20 #include <linux/percpu.h>
21 #include <linux/poll.h>
22 #include <linux/thread_info.h>
23 #include <linux/ctype.h>
24 #include <linux/kmod.h>
25 #include <linux/kdebug.h>
26 #include <asm/processor.h>
29 #include <asm/uaccess.h>
33 #define MISC_MCELOG_MINOR 227
38 static int mce_dont_init;
42 * 0: always panic on uncorrected errors, log corrected errors
43 * 1: panic or SIGBUS on uncorrected errors, log corrected errors
44 * 2: SIGBUS or log uncorrected errors (if possible), log corrected errors
45 * 3: never panic or SIGBUS, log all errors (for testing only)
47 static int tolerant = 1;
49 static unsigned long bank[NR_BANKS] = { [0 ... NR_BANKS-1] = ~0UL };
50 static unsigned long notify_user;
52 static int mce_bootlog = -1;
53 static atomic_t mce_events;
55 static char trigger[128];
56 static char *trigger_argv[2] = { trigger, NULL };
58 static DECLARE_WAIT_QUEUE_HEAD(mce_wait);
61 * Lockless MCE logging infrastructure.
62 * This avoids deadlocks on printk locks without having to break locks. Also
63 * separate MCEs from kernel messages to avoid bogus bug reports.
66 static struct mce_log mcelog = {
71 void mce_log(struct mce *mce)
74 atomic_inc(&mce_events);
78 entry = rcu_dereference(mcelog.next);
80 /* When the buffer fills up discard new entries. Assume
81 that the earlier errors are the more interesting. */
82 if (entry >= MCE_LOG_LEN) {
83 set_bit(MCE_OVERFLOW, (unsigned long *)&mcelog.flags);
86 /* Old left over entry. Skip. */
87 if (mcelog.entry[entry].finished) {
95 if (cmpxchg(&mcelog.next, entry, next) == entry)
98 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
100 mcelog.entry[entry].finished = 1;
103 set_bit(0, ¬ify_user);
106 static void print_mce(struct mce *m)
108 printk(KERN_EMERG "\n"
109 KERN_EMERG "HARDWARE ERROR\n"
111 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
112 m->cpu, m->mcgstatus, m->bank, m->status);
114 printk(KERN_EMERG "RIP%s %02x:<%016Lx> ",
115 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
117 if (m->cs == __KERNEL_CS)
118 print_symbol("{%s}", m->ip);
121 printk(KERN_EMERG "TSC %Lx ", m->tsc);
123 printk("ADDR %Lx ", m->addr);
125 printk("MISC %Lx ", m->misc);
127 printk(KERN_EMERG "This is not a software problem!\n");
128 printk(KERN_EMERG "Run through mcelog --ascii to decode "
129 "and contact your hardware vendor\n");
132 static void mce_panic(char *msg, struct mce *backup, unsigned long start)
137 for (i = 0; i < MCE_LOG_LEN; i++) {
138 unsigned long tsc = mcelog.entry[i].tsc;
140 if (time_before(tsc, start))
142 print_mce(&mcelog.entry[i]);
143 if (backup && mcelog.entry[i].tsc == backup->tsc)
151 static int mce_available(struct cpuinfo_x86 *c)
153 return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
156 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
158 if (regs && (m->mcgstatus & MCG_STATUS_RIPV)) {
166 /* Assume the RIP in the MSR is exact. Is this true? */
167 m->mcgstatus |= MCG_STATUS_EIPV;
168 rdmsrl(rip_msr, m->ip);
174 * The actual machine check handler
176 void do_machine_check(struct pt_regs * regs, long error_code)
178 struct mce m, panicm;
181 int panicm_found = 0;
183 * If no_way_out gets set, there is no safe way to recover from this
184 * MCE. If tolerant is cranked up, we'll try anyway.
188 * If kill_it gets set, there might be a way to recover from this
193 atomic_inc(&mce_entry);
196 && notify_die(DIE_NMI, "machine check", regs, error_code,
197 18, SIGKILL) == NOTIFY_STOP)
201 memset(&m, 0, sizeof(struct mce));
202 m.cpu = smp_processor_id();
203 rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
204 /* if the restart IP is not valid, we're done for */
205 if (!(m.mcgstatus & MCG_STATUS_RIPV))
211 for (i = 0; i < banks; i++) {
220 rdmsrl(MSR_IA32_MC0_STATUS + i*4, m.status);
221 if ((m.status & MCI_STATUS_VAL) == 0)
224 if (m.status & MCI_STATUS_EN) {
225 /* if PCC was set, there's no way out */
226 no_way_out |= !!(m.status & MCI_STATUS_PCC);
228 * If this error was uncorrectable and there was
229 * an overflow, we're in trouble. If no overflow,
230 * we might get away with just killing a task.
232 if (m.status & MCI_STATUS_UC) {
233 if (tolerant < 1 || m.status & MCI_STATUS_OVER)
239 if (m.status & MCI_STATUS_MISCV)
240 rdmsrl(MSR_IA32_MC0_MISC + i*4, m.misc);
241 if (m.status & MCI_STATUS_ADDRV)
242 rdmsrl(MSR_IA32_MC0_ADDR + i*4, m.addr);
244 mce_get_rip(&m, regs);
247 if (error_code != -2)
250 /* Did this bank cause the exception? */
251 /* Assume that the bank with uncorrectable errors did it,
252 and that there is only a single one. */
253 if ((m.status & MCI_STATUS_UC) && (m.status & MCI_STATUS_EN)) {
258 add_taint(TAINT_MACHINE_CHECK);
261 /* Never do anything final in the polling timer */
265 /* If we didn't find an uncorrectable error, pick
266 the last one (shouldn't happen, just being safe). */
271 * If we have decided that we just CAN'T continue, and the user
272 * has not set tolerant to an insane level, give up and die.
274 if (no_way_out && tolerant < 3)
275 mce_panic("Machine check", &panicm, mcestart);
278 * If the error seems to be unrecoverable, something should be
279 * done. Try to kill as little as possible. If we can kill just
280 * one task, do that. If the user has set the tolerance very
281 * high, don't try to do anything at all.
283 if (kill_it && tolerant < 3) {
287 * If the EIPV bit is set, it means the saved IP is the
288 * instruction which caused the MCE.
290 if (m.mcgstatus & MCG_STATUS_EIPV)
291 user_space = panicm.ip && (panicm.cs & 3);
294 * If we know that the error was in user space, send a
295 * SIGBUS. Otherwise, panic if tolerance is low.
297 * do_exit() takes an awful lot of locks and has a slight
298 * risk of deadlocking.
302 } else if (panic_on_oops || tolerant < 2) {
303 mce_panic("Uncorrected machine check",
308 /* notify userspace ASAP */
309 set_thread_flag(TIF_MCE_NOTIFY);
312 /* the last thing we do is clear state */
313 for (i = 0; i < banks; i++)
314 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
315 wrmsrl(MSR_IA32_MCG_STATUS, 0);
317 atomic_dec(&mce_entry);
320 #ifdef CONFIG_X86_MCE_INTEL
322 * mce_log_therm_throt_event - Logs the thermal throttling event to mcelog
323 * @cpu: The CPU on which the event occurred.
324 * @status: Event status information
326 * This function should be called by the thermal interrupt after the
327 * event has been processed and the decision was made to log the event
330 * The status parameter will be saved to the 'status' field of 'struct mce'
331 * and historically has been the register value of the
332 * MSR_IA32_THERMAL_STATUS (Intel) msr.
334 void mce_log_therm_throt_event(unsigned int cpu, __u64 status)
338 memset(&m, 0, sizeof(m));
340 m.bank = MCE_THERMAL_BANK;
345 #endif /* CONFIG_X86_MCE_INTEL */
348 * Periodic polling timer for "silent" machine check errors. If the
349 * poller finds an MCE, poll 2x faster. When the poller finds no more
350 * errors, poll 2x slower (up to check_interval seconds).
353 static int check_interval = 5 * 60; /* 5 minutes */
354 static int next_interval; /* in jiffies */
355 static void mcheck_timer(struct work_struct *work);
356 static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer);
358 static void mcheck_check_cpu(void *info)
360 if (mce_available(¤t_cpu_data))
361 do_machine_check(NULL, 0);
364 static void mcheck_timer(struct work_struct *work)
366 on_each_cpu(mcheck_check_cpu, NULL, 1, 1);
369 * Alert userspace if needed. If we logged an MCE, reduce the
370 * polling interval, otherwise increase the polling interval.
372 if (mce_notify_user()) {
373 next_interval = max(next_interval/2, HZ/100);
375 next_interval = min(next_interval * 2,
376 (int)round_jiffies_relative(check_interval*HZ));
379 schedule_delayed_work(&mcheck_work, next_interval);
383 * This is only called from process context. This is where we do
384 * anything we need to alert userspace about new MCEs. This is called
385 * directly from the poller and also from entry.S and idle, thanks to
388 int mce_notify_user(void)
390 clear_thread_flag(TIF_MCE_NOTIFY);
391 if (test_and_clear_bit(0, ¬ify_user)) {
392 static unsigned long last_print;
393 unsigned long now = jiffies;
395 wake_up_interruptible(&mce_wait);
397 call_usermodehelper(trigger, trigger_argv, NULL,
400 if (time_after_eq(now, last_print + (check_interval*HZ))) {
402 printk(KERN_INFO "Machine check events logged\n");
410 /* see if the idle task needs to notify userspace */
412 mce_idle_callback(struct notifier_block *nfb, unsigned long action, void *junk)
414 /* IDLE_END should be safe - interrupts are back on */
415 if (action == IDLE_END && test_thread_flag(TIF_MCE_NOTIFY))
421 static struct notifier_block mce_idle_notifier = {
422 .notifier_call = mce_idle_callback,
425 static __init int periodic_mcheck_init(void)
427 next_interval = check_interval * HZ;
429 schedule_delayed_work(&mcheck_work,
430 round_jiffies_relative(next_interval));
431 idle_notifier_register(&mce_idle_notifier);
434 __initcall(periodic_mcheck_init);
438 * Initialize Machine Checks for a CPU.
440 static void mce_init(void *dummy)
445 rdmsrl(MSR_IA32_MCG_CAP, cap);
447 if (banks > NR_BANKS) {
448 printk(KERN_INFO "MCE: warning: using only %d banks\n", banks);
451 /* Use accurate RIP reporting if available. */
452 if ((cap & (1<<9)) && ((cap >> 16) & 0xff) >= 9)
453 rip_msr = MSR_IA32_MCG_EIP;
455 /* Log the machine checks left over from the previous reset.
456 This also clears all registers */
457 do_machine_check(NULL, mce_bootlog ? -1 : -2);
459 set_in_cr4(X86_CR4_MCE);
462 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
464 for (i = 0; i < banks; i++) {
465 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
466 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
470 /* Add per CPU specific workarounds here */
471 static void __cpuinit mce_cpu_quirks(struct cpuinfo_x86 *c)
473 /* This should be disabled by the BIOS, but isn't always */
474 if (c->x86_vendor == X86_VENDOR_AMD) {
476 /* disable GART TBL walk error reporting, which trips off
477 incorrectly with the IOMMU & 3ware & Cerberus. */
478 clear_bit(10, &bank[4]);
479 if(c->x86 <= 17 && mce_bootlog < 0)
480 /* Lots of broken BIOS around that don't clear them
481 by default and leave crap in there. Don't log. */
487 static void __cpuinit mce_cpu_features(struct cpuinfo_x86 *c)
489 switch (c->x86_vendor) {
490 case X86_VENDOR_INTEL:
491 mce_intel_feature_init(c);
494 mce_amd_feature_init(c);
502 * Called for each booted CPU to set up machine checks.
503 * Must be called with preempt off.
505 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
507 static cpumask_t mce_cpus = CPU_MASK_NONE;
512 cpu_test_and_set(smp_processor_id(), mce_cpus) ||
521 * Character device to read and clear the MCE log.
524 static DEFINE_SPINLOCK(mce_state_lock);
525 static int open_count; /* #times opened */
526 static int open_exclu; /* already open exclusive? */
528 static int mce_open(struct inode *inode, struct file *file)
530 spin_lock(&mce_state_lock);
532 if (open_exclu || (open_count && (file->f_flags & O_EXCL))) {
533 spin_unlock(&mce_state_lock);
537 if (file->f_flags & O_EXCL)
541 spin_unlock(&mce_state_lock);
543 return nonseekable_open(inode, file);
546 static int mce_release(struct inode *inode, struct file *file)
548 spin_lock(&mce_state_lock);
553 spin_unlock(&mce_state_lock);
558 static void collect_tscs(void *data)
560 unsigned long *cpu_tsc = (unsigned long *)data;
562 rdtscll(cpu_tsc[smp_processor_id()]);
565 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize,
568 unsigned long *cpu_tsc;
569 static DEFINE_MUTEX(mce_read_mutex);
571 char __user *buf = ubuf;
574 cpu_tsc = kmalloc(NR_CPUS * sizeof(long), GFP_KERNEL);
578 mutex_lock(&mce_read_mutex);
579 next = rcu_dereference(mcelog.next);
581 /* Only supports full reads right now */
582 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
583 mutex_unlock(&mce_read_mutex);
589 for (i = 0; i < next; i++) {
590 unsigned long start = jiffies;
592 while (!mcelog.entry[i].finished) {
593 if (time_after_eq(jiffies, start + 2)) {
594 memset(mcelog.entry + i,0, sizeof(struct mce));
600 err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
601 buf += sizeof(struct mce);
606 memset(mcelog.entry, 0, next * sizeof(struct mce));
612 * Collect entries that were still getting written before the
615 on_each_cpu(collect_tscs, cpu_tsc, 1, 1);
616 for (i = next; i < MCE_LOG_LEN; i++) {
617 if (mcelog.entry[i].finished &&
618 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
619 err |= copy_to_user(buf, mcelog.entry+i,
622 buf += sizeof(struct mce);
623 memset(&mcelog.entry[i], 0, sizeof(struct mce));
626 mutex_unlock(&mce_read_mutex);
628 return err ? -EFAULT : buf - ubuf;
631 static unsigned int mce_poll(struct file *file, poll_table *wait)
633 poll_wait(file, &mce_wait, wait);
634 if (rcu_dereference(mcelog.next))
635 return POLLIN | POLLRDNORM;
639 static long mce_ioctl(struct file *f, unsigned int cmd, unsigned long arg)
641 int __user *p = (int __user *)arg;
643 if (!capable(CAP_SYS_ADMIN))
646 case MCE_GET_RECORD_LEN:
647 return put_user(sizeof(struct mce), p);
648 case MCE_GET_LOG_LEN:
649 return put_user(MCE_LOG_LEN, p);
650 case MCE_GETCLEAR_FLAGS: {
654 flags = mcelog.flags;
655 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
656 return put_user(flags, p);
663 static const struct file_operations mce_chrdev_ops = {
665 .release = mce_release,
668 .unlocked_ioctl = mce_ioctl,
671 static struct miscdevice mce_log_device = {
677 static unsigned long old_cr4 __initdata;
679 void __init stop_mce(void)
681 old_cr4 = read_cr4();
682 clear_in_cr4(X86_CR4_MCE);
685 void __init restart_mce(void)
687 if (old_cr4 & X86_CR4_MCE)
688 set_in_cr4(X86_CR4_MCE);
692 * Old style boot options parsing. Only for compatibility.
694 static int __init mcheck_disable(char *str)
700 /* mce=off disables machine check. Note you can re-enable it later
702 mce=TOLERANCELEVEL (number, see above)
703 mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
704 mce=nobootlog Don't log MCEs from before booting. */
705 static int __init mcheck_enable(char *str)
707 if (!strcmp(str, "off"))
709 else if (!strcmp(str, "bootlog") || !strcmp(str,"nobootlog"))
710 mce_bootlog = str[0] == 'b';
711 else if (isdigit(str[0]))
712 get_option(&str, &tolerant);
714 printk("mce= argument %s ignored. Please use /sys", str);
718 __setup("nomce", mcheck_disable);
719 __setup("mce=", mcheck_enable);
725 /* On resume clear all MCE state. Don't want to see leftovers from the BIOS.
726 Only one CPU is active at this time, the others get readded later using
728 static int mce_resume(struct sys_device *dev)
734 /* Reinit MCEs after user configuration changes */
735 static void mce_restart(void)
738 cancel_delayed_work(&mcheck_work);
739 /* Timer race is harmless here */
740 on_each_cpu(mce_init, NULL, 1, 1);
741 next_interval = check_interval * HZ;
743 schedule_delayed_work(&mcheck_work,
744 round_jiffies_relative(next_interval));
747 static struct sysdev_class mce_sysclass = {
748 .resume = mce_resume,
749 .name = "machinecheck",
752 DEFINE_PER_CPU(struct sys_device, device_mce);
754 /* Why are there no generic functions for this? */
755 #define ACCESSOR(name, var, start) \
756 static ssize_t show_ ## name(struct sys_device *s, char *buf) { \
757 return sprintf(buf, "%lx\n", (unsigned long)var); \
759 static ssize_t set_ ## name(struct sys_device *s,const char *buf,size_t siz) { \
761 unsigned long new = simple_strtoul(buf, &end, 0); \
762 if (end == buf) return -EINVAL; \
767 static SYSDEV_ATTR(name, 0644, show_ ## name, set_ ## name);
769 /* TBD should generate these dynamically based on number of available banks */
770 ACCESSOR(bank0ctl,bank[0],mce_restart())
771 ACCESSOR(bank1ctl,bank[1],mce_restart())
772 ACCESSOR(bank2ctl,bank[2],mce_restart())
773 ACCESSOR(bank3ctl,bank[3],mce_restart())
774 ACCESSOR(bank4ctl,bank[4],mce_restart())
775 ACCESSOR(bank5ctl,bank[5],mce_restart())
777 static ssize_t show_trigger(struct sys_device *s, char *buf)
779 strcpy(buf, trigger);
781 return strlen(trigger) + 1;
784 static ssize_t set_trigger(struct sys_device *s,const char *buf,size_t siz)
788 strncpy(trigger, buf, sizeof(trigger));
789 trigger[sizeof(trigger)-1] = 0;
790 len = strlen(trigger);
791 p = strchr(trigger, '\n');
796 static SYSDEV_ATTR(trigger, 0644, show_trigger, set_trigger);
797 ACCESSOR(tolerant,tolerant,)
798 ACCESSOR(check_interval,check_interval,mce_restart())
799 static struct sysdev_attribute *mce_attributes[] = {
800 &attr_bank0ctl, &attr_bank1ctl, &attr_bank2ctl,
801 &attr_bank3ctl, &attr_bank4ctl, &attr_bank5ctl,
802 &attr_tolerant, &attr_check_interval, &attr_trigger,
806 static cpumask_t mce_device_initialized = CPU_MASK_NONE;
808 /* Per cpu sysdev init. All of the cpus still share the same ctl bank */
809 static __cpuinit int mce_create_device(unsigned int cpu)
814 if (!mce_available(&boot_cpu_data))
817 memset(&per_cpu(device_mce, cpu).kobj, 0, sizeof(struct kobject));
818 per_cpu(device_mce,cpu).id = cpu;
819 per_cpu(device_mce,cpu).cls = &mce_sysclass;
821 err = sysdev_register(&per_cpu(device_mce,cpu));
825 for (i = 0; mce_attributes[i]; i++) {
826 err = sysdev_create_file(&per_cpu(device_mce,cpu),
831 cpu_set(cpu, mce_device_initialized);
836 sysdev_remove_file(&per_cpu(device_mce,cpu),
839 sysdev_unregister(&per_cpu(device_mce,cpu));
844 static void mce_remove_device(unsigned int cpu)
848 if (!cpu_isset(cpu, mce_device_initialized))
851 for (i = 0; mce_attributes[i]; i++)
852 sysdev_remove_file(&per_cpu(device_mce,cpu),
854 sysdev_unregister(&per_cpu(device_mce,cpu));
855 cpu_clear(cpu, mce_device_initialized);
858 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
859 static int __cpuinit mce_cpu_callback(struct notifier_block *nfb,
860 unsigned long action, void *hcpu)
862 unsigned int cpu = (unsigned long)hcpu;
866 case CPU_ONLINE_FROZEN:
867 mce_create_device(cpu);
870 case CPU_DEAD_FROZEN:
871 mce_remove_device(cpu);
877 static struct notifier_block mce_cpu_notifier __cpuinitdata = {
878 .notifier_call = mce_cpu_callback,
881 static __init int mce_init_device(void)
886 if (!mce_available(&boot_cpu_data))
888 err = sysdev_class_register(&mce_sysclass);
892 for_each_online_cpu(i) {
893 err = mce_create_device(i);
898 register_hotcpu_notifier(&mce_cpu_notifier);
899 misc_register(&mce_log_device);
903 device_initcall(mce_init_device);