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/ctype.h>
22 #include <linux/kmod.h>
23 #include <asm/processor.h>
26 #include <asm/kdebug.h>
27 #include <asm/uaccess.h>
30 #define MISC_MCELOG_MINOR 227
35 static int mce_dont_init;
37 /* 0: always panic, 1: panic if deadlock possible, 2: try to avoid panic,
38 3: never panic or exit (for testing only) */
39 static int tolerant = 1;
41 static unsigned long bank[NR_BANKS] = { [0 ... NR_BANKS-1] = ~0UL };
42 static unsigned long console_logged;
43 static int notify_user;
45 static int mce_bootlog = 1;
46 static atomic_t mce_events;
48 static char trigger[128];
49 static char *trigger_argv[2] = { trigger, NULL };
52 * Lockless MCE logging infrastructure.
53 * This avoids deadlocks on printk locks without having to break locks. Also
54 * separate MCEs from kernel messages to avoid bogus bug reports.
57 struct mce_log mcelog = {
62 void mce_log(struct mce *mce)
65 atomic_inc(&mce_events);
69 entry = rcu_dereference(mcelog.next);
70 /* The rmb forces the compiler to reload next in each
74 /* When the buffer fills up discard new entries. Assume
75 that the earlier errors are the more interesting. */
76 if (entry >= MCE_LOG_LEN) {
77 set_bit(MCE_OVERFLOW, &mcelog.flags);
80 /* Old left over entry. Skip. */
81 if (mcelog.entry[entry].finished) {
89 if (cmpxchg(&mcelog.next, entry, next) == entry)
92 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
94 mcelog.entry[entry].finished = 1;
97 if (!test_and_set_bit(0, &console_logged))
101 static void print_mce(struct mce *m)
103 printk(KERN_EMERG "\n"
104 KERN_EMERG "HARDWARE ERROR\n"
106 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
107 m->cpu, m->mcgstatus, m->bank, m->status);
110 "RIP%s %02x:<%016Lx> ",
111 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
113 if (m->cs == __KERNEL_CS)
114 print_symbol("{%s}", m->rip);
117 printk(KERN_EMERG "TSC %Lx ", m->tsc);
119 printk("ADDR %Lx ", m->addr);
121 printk("MISC %Lx ", m->misc);
123 printk(KERN_EMERG "This is not a software problem!\n");
125 "Run through mcelog --ascii to decode and contact your hardware vendor\n");
128 static void mce_panic(char *msg, struct mce *backup, unsigned long start)
132 for (i = 0; i < MCE_LOG_LEN; i++) {
133 unsigned long tsc = mcelog.entry[i].tsc;
134 if (time_before(tsc, start))
136 print_mce(&mcelog.entry[i]);
137 if (backup && mcelog.entry[i].tsc == backup->tsc)
143 printk("Fake panic: %s\n", msg);
148 static int mce_available(struct cpuinfo_x86 *c)
150 return cpu_has(c, X86_FEATURE_MCE) && cpu_has(c, X86_FEATURE_MCA);
153 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
155 if (regs && (m->mcgstatus & MCG_STATUS_RIPV)) {
163 /* Assume the RIP in the MSR is exact. Is this true? */
164 m->mcgstatus |= MCG_STATUS_EIPV;
165 rdmsrl(rip_msr, m->rip);
170 static void do_mce_trigger(void)
172 static atomic_t mce_logged;
173 int events = atomic_read(&mce_events);
174 if (events != atomic_read(&mce_logged) && trigger[0]) {
175 /* Small race window, but should be harmless. */
176 atomic_set(&mce_logged, events);
177 call_usermodehelper(trigger, trigger_argv, NULL, -1);
182 * The actual machine check handler
185 void do_machine_check(struct pt_regs * regs, long error_code)
187 struct mce m, panicm;
188 int nowayout = (tolerant < 1);
192 int panicm_found = 0;
194 atomic_inc(&mce_entry);
197 notify_die(DIE_NMI, "machine check", regs, error_code, 18, SIGKILL);
201 memset(&m, 0, sizeof(struct mce));
202 m.cpu = smp_processor_id();
203 rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
204 if (!(m.mcgstatus & MCG_STATUS_RIPV))
210 for (i = 0; i < banks; i++) {
219 rdmsrl(MSR_IA32_MC0_STATUS + i*4, m.status);
220 if ((m.status & MCI_STATUS_VAL) == 0)
223 if (m.status & MCI_STATUS_EN) {
224 /* In theory _OVER could be a nowayout too, but
225 assume any overflowed errors were no fatal. */
226 nowayout |= !!(m.status & MCI_STATUS_PCC);
227 kill_it |= !!(m.status & MCI_STATUS_UC);
230 if (m.status & MCI_STATUS_MISCV)
231 rdmsrl(MSR_IA32_MC0_MISC + i*4, m.misc);
232 if (m.status & MCI_STATUS_ADDRV)
233 rdmsrl(MSR_IA32_MC0_ADDR + i*4, m.addr);
235 mce_get_rip(&m, regs);
238 wrmsrl(MSR_IA32_MC0_STATUS + i*4, 0);
239 if (error_code != -2)
242 /* Did this bank cause the exception? */
243 /* Assume that the bank with uncorrectable errors did it,
244 and that there is only a single one. */
245 if ((m.status & MCI_STATUS_UC) && (m.status & MCI_STATUS_EN)) {
250 add_taint(TAINT_MACHINE_CHECK);
253 /* Never do anything final in the polling timer */
255 /* Normal interrupt context here. Call trigger for any new
261 /* If we didn't find an uncorrectable error, pick
262 the last one (shouldn't happen, just being safe). */
266 mce_panic("Machine check", &panicm, mcestart);
270 if (m.mcgstatus & MCG_STATUS_RIPV)
271 user_space = panicm.rip && (panicm.cs & 3);
273 /* When the machine was in user space and the CPU didn't get
274 confused it's normally not necessary to panic, unless you
275 are paranoid (tolerant == 0)
277 RED-PEN could be more tolerant for MCEs in idle,
278 but most likely they occur at boot anyways, where
279 it is best to just halt the machine. */
280 if ((!user_space && (panic_on_oops || tolerant < 2)) ||
281 (unsigned)current->pid <= 1)
282 mce_panic("Uncorrected machine check", &panicm, mcestart);
284 /* do_exit takes an awful lot of locks and has as
285 slight risk of deadlocking. If you don't want that
286 don't set tolerant >= 2 */
292 /* Last thing done in the machine check exception to clear state. */
293 wrmsrl(MSR_IA32_MCG_STATUS, 0);
295 atomic_dec(&mce_entry);
298 #ifdef CONFIG_X86_MCE_INTEL
300 * mce_log_therm_throt_event - Logs the thermal throttling event to mcelog
301 * @cpu: The CPU on which the event occured.
302 * @status: Event status information
304 * This function should be called by the thermal interrupt after the
305 * event has been processed and the decision was made to log the event
308 * The status parameter will be saved to the 'status' field of 'struct mce'
309 * and historically has been the register value of the
310 * MSR_IA32_THERMAL_STATUS (Intel) msr.
312 void mce_log_therm_throt_event(unsigned int cpu, __u64 status)
316 memset(&m, 0, sizeof(m));
318 m.bank = MCE_THERMAL_BANK;
323 #endif /* CONFIG_X86_MCE_INTEL */
326 * Periodic polling timer for "silent" machine check errors. If the
327 * poller finds an MCE, poll 2x faster. When the poller finds no more
328 * errors, poll 2x slower (up to check_interval seconds).
331 static int check_interval = 5 * 60; /* 5 minutes */
332 static int next_interval; /* in jiffies */
333 static void mcheck_timer(struct work_struct *work);
334 static DECLARE_DELAYED_WORK(mcheck_work, mcheck_timer);
336 static void mcheck_check_cpu(void *info)
338 if (mce_available(¤t_cpu_data))
339 do_machine_check(NULL, 0);
342 static void mcheck_timer(struct work_struct *work)
344 on_each_cpu(mcheck_check_cpu, NULL, 1, 1);
347 * It's ok to read stale data here for notify_user and
348 * console_logged as we'll simply get the updated versions
349 * on the next mcheck_timer execution and atomic operations
350 * on console_logged act as synchronization for notify_user
353 if (notify_user && console_logged) {
354 static unsigned long last_print;
355 unsigned long now = jiffies;
357 /* if we logged an MCE, reduce the polling interval */
358 next_interval = max(next_interval/2, HZ/100);
360 clear_bit(0, &console_logged);
361 if (time_after_eq(now, last_print + (check_interval*HZ))) {
363 printk(KERN_INFO "Machine check events logged\n");
366 next_interval = min(next_interval*2, check_interval*HZ);
369 schedule_delayed_work(&mcheck_work, next_interval);
373 static __init int periodic_mcheck_init(void)
375 next_interval = check_interval * HZ;
377 schedule_delayed_work(&mcheck_work, next_interval);
380 __initcall(periodic_mcheck_init);
384 * Initialize Machine Checks for a CPU.
386 static void mce_init(void *dummy)
391 rdmsrl(MSR_IA32_MCG_CAP, cap);
393 if (banks > NR_BANKS) {
394 printk(KERN_INFO "MCE: warning: using only %d banks\n", banks);
397 /* Use accurate RIP reporting if available. */
398 if ((cap & (1<<9)) && ((cap >> 16) & 0xff) >= 9)
399 rip_msr = MSR_IA32_MCG_EIP;
401 /* Log the machine checks left over from the previous reset.
402 This also clears all registers */
403 do_machine_check(NULL, mce_bootlog ? -1 : -2);
405 set_in_cr4(X86_CR4_MCE);
408 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
410 for (i = 0; i < banks; i++) {
411 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
412 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
416 /* Add per CPU specific workarounds here */
417 static void __cpuinit mce_cpu_quirks(struct cpuinfo_x86 *c)
419 /* This should be disabled by the BIOS, but isn't always */
420 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 == 15) {
421 /* disable GART TBL walk error reporting, which trips off
422 incorrectly with the IOMMU & 3ware & Cerberus. */
423 clear_bit(10, &bank[4]);
424 /* Lots of broken BIOS around that don't clear them
425 by default and leave crap in there. Don't log. */
431 static void __cpuinit mce_cpu_features(struct cpuinfo_x86 *c)
433 switch (c->x86_vendor) {
434 case X86_VENDOR_INTEL:
435 mce_intel_feature_init(c);
438 mce_amd_feature_init(c);
446 * Called for each booted CPU to set up machine checks.
447 * Must be called with preempt off.
449 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
451 static cpumask_t mce_cpus = CPU_MASK_NONE;
456 cpu_test_and_set(smp_processor_id(), mce_cpus) ||
465 * Character device to read and clear the MCE log.
468 static void collect_tscs(void *data)
470 unsigned long *cpu_tsc = (unsigned long *)data;
471 rdtscll(cpu_tsc[smp_processor_id()]);
474 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize, loff_t *off)
476 unsigned long *cpu_tsc;
477 static DECLARE_MUTEX(mce_read_sem);
479 char __user *buf = ubuf;
482 cpu_tsc = kmalloc(NR_CPUS * sizeof(long), GFP_KERNEL);
487 next = rcu_dereference(mcelog.next);
489 /* Only supports full reads right now */
490 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
497 for (i = 0; i < next; i++) {
498 unsigned long start = jiffies;
499 while (!mcelog.entry[i].finished) {
500 if (!time_before(jiffies, start + 2)) {
501 memset(mcelog.entry + i,0, sizeof(struct mce));
507 err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
508 buf += sizeof(struct mce);
511 memset(mcelog.entry, 0, next * sizeof(struct mce));
516 /* Collect entries that were still getting written before the synchronize. */
518 on_each_cpu(collect_tscs, cpu_tsc, 1, 1);
519 for (i = next; i < MCE_LOG_LEN; i++) {
520 if (mcelog.entry[i].finished &&
521 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
522 err |= copy_to_user(buf, mcelog.entry+i, sizeof(struct mce));
524 buf += sizeof(struct mce);
525 memset(&mcelog.entry[i], 0, sizeof(struct mce));
530 return err ? -EFAULT : buf - ubuf;
533 static int mce_ioctl(struct inode *i, struct file *f,unsigned int cmd, unsigned long arg)
535 int __user *p = (int __user *)arg;
536 if (!capable(CAP_SYS_ADMIN))
539 case MCE_GET_RECORD_LEN:
540 return put_user(sizeof(struct mce), p);
541 case MCE_GET_LOG_LEN:
542 return put_user(MCE_LOG_LEN, p);
543 case MCE_GETCLEAR_FLAGS: {
546 flags = mcelog.flags;
547 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
548 return put_user(flags, p);
555 static const struct file_operations mce_chrdev_ops = {
560 static struct miscdevice mce_log_device = {
567 * Old style boot options parsing. Only for compatibility.
570 static int __init mcheck_disable(char *str)
576 /* mce=off disables machine check. Note you can reenable it later
578 mce=TOLERANCELEVEL (number, see above)
579 mce=bootlog Log MCEs from before booting. Disabled by default on AMD.
580 mce=nobootlog Don't log MCEs from before booting. */
581 static int __init mcheck_enable(char *str)
585 if (!strcmp(str, "off"))
587 else if (!strcmp(str, "bootlog") || !strcmp(str,"nobootlog"))
588 mce_bootlog = str[0] == 'b';
589 else if (isdigit(str[0]))
590 get_option(&str, &tolerant);
592 printk("mce= argument %s ignored. Please use /sys", str);
596 __setup("nomce", mcheck_disable);
597 __setup("mce", mcheck_enable);
603 /* On resume clear all MCE state. Don't want to see leftovers from the BIOS.
604 Only one CPU is active at this time, the others get readded later using
606 static int mce_resume(struct sys_device *dev)
612 /* Reinit MCEs after user configuration changes */
613 static void mce_restart(void)
616 cancel_delayed_work(&mcheck_work);
617 /* Timer race is harmless here */
618 on_each_cpu(mce_init, NULL, 1, 1);
619 next_interval = check_interval * HZ;
621 schedule_delayed_work(&mcheck_work, next_interval);
624 static struct sysdev_class mce_sysclass = {
625 .resume = mce_resume,
626 set_kset_name("machinecheck"),
629 DEFINE_PER_CPU(struct sys_device, device_mce);
631 /* Why are there no generic functions for this? */
632 #define ACCESSOR(name, var, start) \
633 static ssize_t show_ ## name(struct sys_device *s, char *buf) { \
634 return sprintf(buf, "%lx\n", (unsigned long)var); \
636 static ssize_t set_ ## name(struct sys_device *s,const char *buf,size_t siz) { \
638 unsigned long new = simple_strtoul(buf, &end, 0); \
639 if (end == buf) return -EINVAL; \
644 static SYSDEV_ATTR(name, 0644, show_ ## name, set_ ## name);
646 /* TBD should generate these dynamically based on number of available banks */
647 ACCESSOR(bank0ctl,bank[0],mce_restart())
648 ACCESSOR(bank1ctl,bank[1],mce_restart())
649 ACCESSOR(bank2ctl,bank[2],mce_restart())
650 ACCESSOR(bank3ctl,bank[3],mce_restart())
651 ACCESSOR(bank4ctl,bank[4],mce_restart())
652 ACCESSOR(bank5ctl,bank[5],mce_restart())
654 static ssize_t show_trigger(struct sys_device *s, char *buf)
656 strcpy(buf, trigger);
658 return strlen(trigger) + 1;
661 static ssize_t set_trigger(struct sys_device *s,const char *buf,size_t siz)
665 strncpy(trigger, buf, sizeof(trigger));
666 trigger[sizeof(trigger)-1] = 0;
667 len = strlen(trigger);
668 p = strchr(trigger, '\n');
673 static SYSDEV_ATTR(trigger, 0644, show_trigger, set_trigger);
674 ACCESSOR(tolerant,tolerant,)
675 ACCESSOR(check_interval,check_interval,mce_restart())
676 static struct sysdev_attribute *mce_attributes[] = {
677 &attr_bank0ctl, &attr_bank1ctl, &attr_bank2ctl,
678 &attr_bank3ctl, &attr_bank4ctl, &attr_bank5ctl,
679 &attr_tolerant, &attr_check_interval, &attr_trigger,
683 /* Per cpu sysdev init. All of the cpus still share the same ctl bank */
684 static __cpuinit int mce_create_device(unsigned int cpu)
688 if (!mce_available(&cpu_data[cpu]))
691 per_cpu(device_mce,cpu).id = cpu;
692 per_cpu(device_mce,cpu).cls = &mce_sysclass;
694 err = sysdev_register(&per_cpu(device_mce,cpu));
697 for (i = 0; mce_attributes[i]; i++)
698 sysdev_create_file(&per_cpu(device_mce,cpu),
704 static void mce_remove_device(unsigned int cpu)
708 for (i = 0; mce_attributes[i]; i++)
709 sysdev_remove_file(&per_cpu(device_mce,cpu),
711 sysdev_unregister(&per_cpu(device_mce,cpu));
712 memset(&per_cpu(device_mce, cpu).kobj, 0, sizeof(struct kobject));
715 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
717 mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
719 unsigned int cpu = (unsigned long)hcpu;
723 mce_create_device(cpu);
726 mce_remove_device(cpu);
732 static struct notifier_block mce_cpu_notifier = {
733 .notifier_call = mce_cpu_callback,
736 static __init int mce_init_device(void)
741 if (!mce_available(&boot_cpu_data))
743 err = sysdev_class_register(&mce_sysclass);
745 for_each_online_cpu(i) {
746 mce_create_device(i);
749 register_hotcpu_notifier(&mce_cpu_notifier);
750 misc_register(&mce_log_device);
754 device_initcall(mce_init_device);