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/cpu.h>
19 #include <linux/percpu.h>
20 #include <asm/processor.h>
23 #include <asm/kdebug.h>
24 #include <asm/uaccess.h>
26 #define MISC_MCELOG_MINOR 227
29 static int mce_dont_init;
31 /* 0: always panic, 1: panic if deadlock possible, 2: try to avoid panic,
32 3: never panic or exit (for testing only) */
33 static int tolerant = 1;
35 static unsigned long bank[NR_BANKS] = { [0 ... NR_BANKS-1] = ~0UL };
36 static unsigned long console_logged;
37 static int notify_user;
39 static int mce_bootlog;
42 * Lockless MCE logging infrastructure.
43 * This avoids deadlocks on printk locks without having to break locks. Also
44 * separate MCEs from kernel messages to avoid bogus bug reports.
47 struct mce_log mcelog = {
52 void mce_log(struct mce *mce)
58 entry = rcu_dereference(mcelog.next);
59 /* When the buffer fills up discard new entries. Assume
60 that the earlier errors are the more interesting. */
61 if (entry >= MCE_LOG_LEN) {
62 set_bit(MCE_OVERFLOW, &mcelog.flags);
65 /* Old left over entry. Skip. */
66 if (mcelog.entry[entry].finished)
70 if (cmpxchg(&mcelog.next, entry, next) == entry)
73 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
75 mcelog.entry[entry].finished = 1;
78 if (!test_and_set_bit(0, &console_logged))
82 static void print_mce(struct mce *m)
84 printk(KERN_EMERG "\n"
86 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
87 m->cpu, m->mcgstatus, m->bank, m->status);
90 "RIP%s %02x:<%016Lx> ",
91 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
93 if (m->cs == __KERNEL_CS)
94 print_symbol("{%s}", m->rip);
97 printk(KERN_EMERG "TSC %Lx ", m->tsc);
99 printk("ADDR %Lx ", m->addr);
101 printk("MISC %Lx ", m->misc);
105 static void mce_panic(char *msg, struct mce *backup, unsigned long start)
109 for (i = 0; i < MCE_LOG_LEN; i++) {
110 unsigned long tsc = mcelog.entry[i].tsc;
111 if (time_before(tsc, start))
113 print_mce(&mcelog.entry[i]);
114 if (backup && mcelog.entry[i].tsc == backup->tsc)
120 printk("Fake panic: %s\n", msg);
125 static int mce_available(struct cpuinfo_x86 *c)
127 return test_bit(X86_FEATURE_MCE, &c->x86_capability) &&
128 test_bit(X86_FEATURE_MCA, &c->x86_capability);
131 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
133 if (regs && (m->mcgstatus & MCG_STATUS_RIPV)) {
141 /* Assume the RIP in the MSR is exact. Is this true? */
142 m->mcgstatus |= MCG_STATUS_EIPV;
143 rdmsrl(rip_msr, m->rip);
149 * The actual machine check handler
152 void do_machine_check(struct pt_regs * regs, long error_code)
154 struct mce m, panicm;
155 int nowayout = (tolerant < 1);
159 int panicm_found = 0;
162 notify_die(DIE_NMI, "machine check", regs, error_code, 255, SIGKILL);
166 memset(&m, 0, sizeof(struct mce));
167 m.cpu = hard_smp_processor_id();
168 rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
169 if (!(m.mcgstatus & MCG_STATUS_RIPV))
175 for (i = 0; i < banks; i++) {
184 rdmsrl(MSR_IA32_MC0_STATUS + i*4, m.status);
185 if ((m.status & MCI_STATUS_VAL) == 0)
188 if (m.status & MCI_STATUS_EN) {
189 /* In theory _OVER could be a nowayout too, but
190 assume any overflowed errors were no fatal. */
191 nowayout |= !!(m.status & MCI_STATUS_PCC);
192 kill_it |= !!(m.status & MCI_STATUS_UC);
195 if (m.status & MCI_STATUS_MISCV)
196 rdmsrl(MSR_IA32_MC0_MISC + i*4, m.misc);
197 if (m.status & MCI_STATUS_ADDRV)
198 rdmsrl(MSR_IA32_MC0_ADDR + i*4, m.addr);
200 mce_get_rip(&m, regs);
203 wrmsrl(MSR_IA32_MC0_STATUS + i*4, 0);
204 if (error_code != -2)
207 /* Did this bank cause the exception? */
208 /* Assume that the bank with uncorrectable errors did it,
209 and that there is only a single one. */
210 if ((m.status & MCI_STATUS_UC) && (m.status & MCI_STATUS_EN)) {
215 tainted |= TAINT_MACHINE_CHECK;
218 /* Never do anything final in the polling timer */
222 /* If we didn't find an uncorrectable error, pick
223 the last one (shouldn't happen, just being safe). */
227 mce_panic("Machine check", &panicm, mcestart);
231 if (m.mcgstatus & MCG_STATUS_RIPV)
232 user_space = panicm.rip && (panicm.cs & 3);
234 /* When the machine was in user space and the CPU didn't get
235 confused it's normally not necessary to panic, unless you
236 are paranoid (tolerant == 0)
238 RED-PEN could be more tolerant for MCEs in idle,
239 but most likely they occur at boot anyways, where
240 it is best to just halt the machine. */
241 if ((!user_space && (panic_on_oops || tolerant < 2)) ||
242 (unsigned)current->pid <= 1)
243 mce_panic("Uncorrected machine check", &panicm, mcestart);
245 /* do_exit takes an awful lot of locks and has as
246 slight risk of deadlocking. If you don't want that
247 don't set tolerant >= 2 */
253 /* Last thing done in the machine check exception to clear state. */
254 wrmsrl(MSR_IA32_MCG_STATUS, 0);
258 * Periodic polling timer for "silent" machine check errors.
261 static int check_interval = 5 * 60; /* 5 minutes */
262 static void mcheck_timer(void *data);
263 static DECLARE_WORK(mcheck_work, mcheck_timer, NULL);
265 static void mcheck_check_cpu(void *info)
267 if (mce_available(¤t_cpu_data))
268 do_machine_check(NULL, 0);
271 static void mcheck_timer(void *data)
273 on_each_cpu(mcheck_check_cpu, NULL, 1, 1);
274 schedule_delayed_work(&mcheck_work, check_interval * HZ);
277 * It's ok to read stale data here for notify_user and
278 * console_logged as we'll simply get the updated versions
279 * on the next mcheck_timer execution and atomic operations
280 * on console_logged act as synchronization for notify_user
283 if (notify_user && console_logged) {
285 clear_bit(0, &console_logged);
286 printk(KERN_INFO "Machine check events logged\n");
291 static __init int periodic_mcheck_init(void)
294 schedule_delayed_work(&mcheck_work, check_interval*HZ);
297 __initcall(periodic_mcheck_init);
301 * Initialize Machine Checks for a CPU.
303 static void mce_init(void *dummy)
308 rdmsrl(MSR_IA32_MCG_CAP, cap);
310 if (banks > NR_BANKS) {
311 printk(KERN_INFO "MCE: warning: using only %d banks\n", banks);
314 /* Use accurate RIP reporting if available. */
315 if ((cap & (1<<9)) && ((cap >> 16) & 0xff) >= 9)
316 rip_msr = MSR_IA32_MCG_EIP;
318 /* Log the machine checks left over from the previous reset.
319 This also clears all registers */
320 do_machine_check(NULL, mce_bootlog ? -1 : -2);
322 set_in_cr4(X86_CR4_MCE);
325 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
327 for (i = 0; i < banks; i++) {
328 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
329 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
333 /* Add per CPU specific workarounds here */
334 static void __cpuinit mce_cpu_quirks(struct cpuinfo_x86 *c)
336 /* This should be disabled by the BIOS, but isn't always */
337 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 == 15) {
338 /* disable GART TBL walk error reporting, which trips off
339 incorrectly with the IOMMU & 3ware & Cerberus. */
340 clear_bit(10, &bank[4]);
344 static void __cpuinit mce_cpu_features(struct cpuinfo_x86 *c)
346 switch (c->x86_vendor) {
347 case X86_VENDOR_INTEL:
348 mce_intel_feature_init(c);
356 * Called for each booted CPU to set up machine checks.
357 * Must be called with preempt off.
359 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
361 static cpumask_t mce_cpus __initdata = CPU_MASK_NONE;
366 cpu_test_and_set(smp_processor_id(), mce_cpus) ||
375 * Character device to read and clear the MCE log.
378 static void collect_tscs(void *data)
380 unsigned long *cpu_tsc = (unsigned long *)data;
381 rdtscll(cpu_tsc[smp_processor_id()]);
384 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize, loff_t *off)
386 unsigned long *cpu_tsc;
387 static DECLARE_MUTEX(mce_read_sem);
389 char __user *buf = ubuf;
392 cpu_tsc = kmalloc(NR_CPUS * sizeof(long), GFP_KERNEL);
397 next = rcu_dereference(mcelog.next);
399 /* Only supports full reads right now */
400 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
407 for (i = 0; i < next; i++) {
408 if (!mcelog.entry[i].finished)
411 err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
412 buf += sizeof(struct mce);
415 memset(mcelog.entry, 0, next * sizeof(struct mce));
420 /* Collect entries that were still getting written before the synchronize. */
422 on_each_cpu(collect_tscs, cpu_tsc, 1, 1);
423 for (i = next; i < MCE_LOG_LEN; i++) {
424 if (mcelog.entry[i].finished &&
425 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
426 err |= copy_to_user(buf, mcelog.entry+i, sizeof(struct mce));
428 buf += sizeof(struct mce);
429 memset(&mcelog.entry[i], 0, sizeof(struct mce));
434 return err ? -EFAULT : buf - ubuf;
437 static int mce_ioctl(struct inode *i, struct file *f,unsigned int cmd, unsigned long arg)
439 int __user *p = (int __user *)arg;
440 if (!capable(CAP_SYS_ADMIN))
443 case MCE_GET_RECORD_LEN:
444 return put_user(sizeof(struct mce), p);
445 case MCE_GET_LOG_LEN:
446 return put_user(MCE_LOG_LEN, p);
447 case MCE_GETCLEAR_FLAGS: {
450 flags = mcelog.flags;
451 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
452 return put_user(flags, p);
459 static struct file_operations mce_chrdev_ops = {
464 static struct miscdevice mce_log_device = {
471 * Old style boot options parsing. Only for compatibility.
474 static int __init mcheck_disable(char *str)
480 /* mce=off disables machine check. Note you can reenable it later
482 mce=bootlog Log MCEs from before booting. Disabled by default to work
483 around buggy BIOS that leave bogus MCEs. */
484 static int __init mcheck_enable(char *str)
488 if (!strcmp(str, "off"))
490 else if (!strcmp(str, "bootlog"))
493 printk("mce= argument %s ignored. Please use /sys", str);
497 __setup("nomce", mcheck_disable);
498 __setup("mce", mcheck_enable);
504 /* On resume clear all MCE state. Don't want to see leftovers from the BIOS. */
505 static int mce_resume(struct sys_device *dev)
507 on_each_cpu(mce_init, NULL, 1, 1);
511 /* Reinit MCEs after user configuration changes */
512 static void mce_restart(void)
515 cancel_delayed_work(&mcheck_work);
516 /* Timer race is harmless here */
517 on_each_cpu(mce_init, NULL, 1, 1);
519 schedule_delayed_work(&mcheck_work, check_interval*HZ);
522 static struct sysdev_class mce_sysclass = {
523 .resume = mce_resume,
524 set_kset_name("machinecheck"),
527 static DEFINE_PER_CPU(struct sys_device, device_mce);
529 /* Why are there no generic functions for this? */
530 #define ACCESSOR(name, var, start) \
531 static ssize_t show_ ## name(struct sys_device *s, char *buf) { \
532 return sprintf(buf, "%lx\n", (unsigned long)var); \
534 static ssize_t set_ ## name(struct sys_device *s,const char *buf,size_t siz) { \
536 unsigned long new = simple_strtoul(buf, &end, 0); \
537 if (end == buf) return -EINVAL; \
542 static SYSDEV_ATTR(name, 0644, show_ ## name, set_ ## name);
544 ACCESSOR(bank0ctl,bank[0],mce_restart())
545 ACCESSOR(bank1ctl,bank[1],mce_restart())
546 ACCESSOR(bank2ctl,bank[2],mce_restart())
547 ACCESSOR(bank3ctl,bank[3],mce_restart())
548 ACCESSOR(bank4ctl,bank[4],mce_restart())
549 ACCESSOR(tolerant,tolerant,)
550 ACCESSOR(check_interval,check_interval,mce_restart())
552 /* Per cpu sysdev init. All of the cpus still share the same ctl bank */
553 static __cpuinit int mce_create_device(unsigned int cpu)
556 if (!mce_available(&cpu_data[cpu]))
559 per_cpu(device_mce,cpu).id = cpu;
560 per_cpu(device_mce,cpu).cls = &mce_sysclass;
562 err = sysdev_register(&per_cpu(device_mce,cpu));
565 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_bank0ctl);
566 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_bank1ctl);
567 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_bank2ctl);
568 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_bank3ctl);
569 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_bank4ctl);
570 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_tolerant);
571 sysdev_create_file(&per_cpu(device_mce,cpu), &attr_check_interval);
576 #ifdef CONFIG_HOTPLUG_CPU
577 static __cpuinit void mce_remove_device(unsigned int cpu)
579 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_bank0ctl);
580 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_bank1ctl);
581 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_bank2ctl);
582 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_bank3ctl);
583 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_bank4ctl);
584 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_tolerant);
585 sysdev_remove_file(&per_cpu(device_mce,cpu), &attr_check_interval);
586 sysdev_unregister(&per_cpu(device_mce,cpu));
590 /* Get notified when a cpu comes on/off. Be hotplug friendly. */
592 mce_cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
594 unsigned int cpu = (unsigned long)hcpu;
598 mce_create_device(cpu);
600 #ifdef CONFIG_HOTPLUG_CPU
602 mce_remove_device(cpu);
609 static struct notifier_block mce_cpu_notifier = {
610 .notifier_call = mce_cpu_callback,
613 static __init int mce_init_device(void)
618 if (!mce_available(&boot_cpu_data))
620 err = sysdev_class_register(&mce_sysclass);
622 for_each_online_cpu(i) {
623 mce_create_device(i);
626 register_cpu_notifier(&mce_cpu_notifier);
627 misc_register(&mce_log_device);
631 device_initcall(mce_init_device);