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 <asm/processor.h>
21 #include <asm/kdebug.h>
22 #include <asm/uaccess.h>
24 #define MISC_MCELOG_MINOR 227
27 static int mce_dont_init;
29 /* 0: always panic, 1: panic if deadlock possible, 2: try to avoid panic,
30 3: never panic or exit (for testing only) */
31 static int tolerant = 1;
33 static unsigned long bank[NR_BANKS] = { [0 ... NR_BANKS-1] = ~0UL };
34 static unsigned long console_logged;
35 static int notify_user;
39 * Lockless MCE logging infrastructure.
40 * This avoids deadlocks on printk locks without having to break locks. Also
41 * separate MCEs from kernel messages to avoid bogus bug reports.
44 struct mce_log mcelog = {
49 void mce_log(struct mce *mce)
55 entry = rcu_dereference(mcelog.next);
56 /* When the buffer fills up discard new entries. Assume
57 that the earlier errors are the more interesting. */
58 if (entry >= MCE_LOG_LEN) {
59 set_bit(MCE_OVERFLOW, &mcelog.flags);
62 /* Old left over entry. Skip. */
63 if (mcelog.entry[entry].finished)
67 if (cmpxchg(&mcelog.next, entry, next) == entry)
70 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
72 mcelog.entry[entry].finished = 1;
75 if (!test_and_set_bit(0, &console_logged))
79 static void print_mce(struct mce *m)
81 printk(KERN_EMERG "\n"
83 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
84 m->cpu, m->mcgstatus, m->bank, m->status);
87 "RIP%s %02x:<%016Lx> ",
88 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
90 if (m->cs == __KERNEL_CS)
91 print_symbol("{%s}", m->rip);
94 printk(KERN_EMERG "TSC %Lx ", m->tsc);
96 printk("ADDR %Lx ", m->addr);
98 printk("MISC %Lx ", m->misc);
102 static void mce_panic(char *msg, struct mce *backup, unsigned long start)
106 for (i = 0; i < MCE_LOG_LEN; i++) {
107 unsigned long tsc = mcelog.entry[i].tsc;
108 if (time_before(tsc, start))
110 print_mce(&mcelog.entry[i]);
111 if (backup && mcelog.entry[i].tsc == backup->tsc)
117 printk("Fake panic: %s\n", msg);
122 static int mce_available(struct cpuinfo_x86 *c)
124 return test_bit(X86_FEATURE_MCE, &c->x86_capability) &&
125 test_bit(X86_FEATURE_MCA, &c->x86_capability);
128 static inline void mce_get_rip(struct mce *m, struct pt_regs *regs)
130 if (regs && (m->mcgstatus & MCG_STATUS_RIPV)) {
138 /* Assume the RIP in the MSR is exact. Is this true? */
139 m->mcgstatus |= MCG_STATUS_EIPV;
140 rdmsrl(rip_msr, m->rip);
146 * The actual machine check handler
149 void do_machine_check(struct pt_regs * regs, long error_code)
151 struct mce m, panicm;
152 int nowayout = (tolerant < 1);
156 int panicm_found = 0;
159 notify_die(DIE_NMI, "machine check", regs, error_code, 255, SIGKILL);
163 memset(&m, 0, sizeof(struct mce));
164 m.cpu = hard_smp_processor_id();
165 rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
166 if (!(m.mcgstatus & MCG_STATUS_RIPV))
172 for (i = 0; i < banks; i++) {
181 rdmsrl(MSR_IA32_MC0_STATUS + i*4, m.status);
182 if ((m.status & MCI_STATUS_VAL) == 0)
185 if (m.status & MCI_STATUS_EN) {
186 /* In theory _OVER could be a nowayout too, but
187 assume any overflowed errors were no fatal. */
188 nowayout |= !!(m.status & MCI_STATUS_PCC);
189 kill_it |= !!(m.status & MCI_STATUS_UC);
192 if (m.status & MCI_STATUS_MISCV)
193 rdmsrl(MSR_IA32_MC0_MISC + i*4, m.misc);
194 if (m.status & MCI_STATUS_ADDRV)
195 rdmsrl(MSR_IA32_MC0_ADDR + i*4, m.addr);
197 mce_get_rip(&m, regs);
198 if (error_code != -1)
200 wrmsrl(MSR_IA32_MC0_STATUS + i*4, 0);
203 /* Did this bank cause the exception? */
204 /* Assume that the bank with uncorrectable errors did it,
205 and that there is only a single one. */
206 if ((m.status & MCI_STATUS_UC) && (m.status & MCI_STATUS_EN)) {
211 tainted |= TAINT_MACHINE_CHECK;
214 /* Never do anything final in the polling timer */
218 /* If we didn't find an uncorrectable error, pick
219 the last one (shouldn't happen, just being safe). */
223 mce_panic("Machine check", &panicm, mcestart);
227 if (m.mcgstatus & MCG_STATUS_RIPV)
228 user_space = panicm.rip && (panicm.cs & 3);
230 /* When the machine was in user space and the CPU didn't get
231 confused it's normally not necessary to panic, unless you
232 are paranoid (tolerant == 0)
234 RED-PEN could be more tolerant for MCEs in idle,
235 but most likely they occur at boot anyways, where
236 it is best to just halt the machine. */
237 if ((!user_space && (panic_on_oops || tolerant < 2)) ||
238 (unsigned)current->pid <= 1)
239 mce_panic("Uncorrected machine check", &panicm, mcestart);
241 /* do_exit takes an awful lot of locks and has as
242 slight risk of deadlocking. If you don't want that
243 don't set tolerant >= 2 */
249 /* Last thing done in the machine check exception to clear state. */
250 wrmsrl(MSR_IA32_MCG_STATUS, 0);
254 * Periodic polling timer for "silent" machine check errors.
257 static int check_interval = 5 * 60; /* 5 minutes */
258 static void mcheck_timer(void *data);
259 static DECLARE_WORK(mcheck_work, mcheck_timer, NULL);
261 static void mcheck_check_cpu(void *info)
263 if (mce_available(¤t_cpu_data))
264 do_machine_check(NULL, 0);
267 static void mcheck_timer(void *data)
269 on_each_cpu(mcheck_check_cpu, NULL, 1, 1);
270 schedule_delayed_work(&mcheck_work, check_interval * HZ);
273 * It's ok to read stale data here for notify_user and
274 * console_logged as we'll simply get the updated versions
275 * on the next mcheck_timer execution and atomic operations
276 * on console_logged act as synchronization for notify_user
279 if (notify_user && console_logged) {
281 clear_bit(0, &console_logged);
282 printk(KERN_INFO "Machine check events logged\n");
287 static __init int periodic_mcheck_init(void)
290 schedule_delayed_work(&mcheck_work, check_interval*HZ);
293 __initcall(periodic_mcheck_init);
297 * Initialize Machine Checks for a CPU.
299 static void mce_init(void *dummy)
304 rdmsrl(MSR_IA32_MCG_CAP, cap);
306 if (banks > NR_BANKS) {
307 printk(KERN_INFO "MCE: warning: using only %d banks\n", banks);
310 /* Use accurate RIP reporting if available. */
311 if ((cap & (1<<9)) && ((cap >> 16) & 0xff) >= 9)
312 rip_msr = MSR_IA32_MCG_EIP;
314 /* Log the machine checks left over from the previous reset.
315 This also clears all registers */
316 do_machine_check(NULL, -1);
318 set_in_cr4(X86_CR4_MCE);
321 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
323 for (i = 0; i < banks; i++) {
324 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
325 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
329 /* Add per CPU specific workarounds here */
330 static void __cpuinit mce_cpu_quirks(struct cpuinfo_x86 *c)
332 /* This should be disabled by the BIOS, but isn't always */
333 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 == 15) {
334 /* disable GART TBL walk error reporting, which trips off
335 incorrectly with the IOMMU & 3ware & Cerberus. */
336 clear_bit(10, &bank[4]);
340 static void __cpuinit mce_cpu_features(struct cpuinfo_x86 *c)
342 switch (c->x86_vendor) {
343 case X86_VENDOR_INTEL:
344 mce_intel_feature_init(c);
352 * Called for each booted CPU to set up machine checks.
353 * Must be called with preempt off.
355 void __cpuinit mcheck_init(struct cpuinfo_x86 *c)
357 static cpumask_t mce_cpus __initdata = CPU_MASK_NONE;
362 cpu_test_and_set(smp_processor_id(), mce_cpus) ||
371 * Character device to read and clear the MCE log.
374 static void collect_tscs(void *data)
376 unsigned long *cpu_tsc = (unsigned long *)data;
377 rdtscll(cpu_tsc[smp_processor_id()]);
380 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize, loff_t *off)
382 unsigned long *cpu_tsc;
383 static DECLARE_MUTEX(mce_read_sem);
385 char __user *buf = ubuf;
388 cpu_tsc = kmalloc(NR_CPUS * sizeof(long), GFP_KERNEL);
393 next = rcu_dereference(mcelog.next);
395 /* Only supports full reads right now */
396 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
403 for (i = 0; i < next; i++) {
404 if (!mcelog.entry[i].finished)
407 err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
408 buf += sizeof(struct mce);
411 memset(mcelog.entry, 0, next * sizeof(struct mce));
416 /* Collect entries that were still getting written before the synchronize. */
418 on_each_cpu(collect_tscs, cpu_tsc, 1, 1);
419 for (i = next; i < MCE_LOG_LEN; i++) {
420 if (mcelog.entry[i].finished &&
421 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
422 err |= copy_to_user(buf, mcelog.entry+i, sizeof(struct mce));
424 buf += sizeof(struct mce);
425 memset(&mcelog.entry[i], 0, sizeof(struct mce));
430 return err ? -EFAULT : buf - ubuf;
433 static int mce_ioctl(struct inode *i, struct file *f,unsigned int cmd, unsigned long arg)
435 int __user *p = (int __user *)arg;
436 if (!capable(CAP_SYS_ADMIN))
439 case MCE_GET_RECORD_LEN:
440 return put_user(sizeof(struct mce), p);
441 case MCE_GET_LOG_LEN:
442 return put_user(MCE_LOG_LEN, p);
443 case MCE_GETCLEAR_FLAGS: {
446 flags = mcelog.flags;
447 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
448 return put_user(flags, p);
455 static struct file_operations mce_chrdev_ops = {
460 static struct miscdevice mce_log_device = {
467 * Old style boot options parsing. Only for compatibility.
470 static int __init mcheck_disable(char *str)
476 /* mce=off disables machine check. Note you can reenable it later
478 static int __init mcheck_enable(char *str)
480 if (!strcmp(str, "off"))
483 printk("mce= argument %s ignored. Please use /sys", str);
487 __setup("nomce", mcheck_disable);
488 __setup("mce", mcheck_enable);
494 /* On resume clear all MCE state. Don't want to see leftovers from the BIOS. */
495 static int mce_resume(struct sys_device *dev)
497 on_each_cpu(mce_init, NULL, 1, 1);
501 /* Reinit MCEs after user configuration changes */
502 static void mce_restart(void)
505 cancel_delayed_work(&mcheck_work);
506 /* Timer race is harmless here */
507 on_each_cpu(mce_init, NULL, 1, 1);
509 schedule_delayed_work(&mcheck_work, check_interval*HZ);
512 static struct sysdev_class mce_sysclass = {
513 .resume = mce_resume,
514 set_kset_name("machinecheck"),
517 static struct sys_device device_mce = {
519 .cls = &mce_sysclass,
522 /* Why are there no generic functions for this? */
523 #define ACCESSOR(name, var, start) \
524 static ssize_t show_ ## name(struct sys_device *s, char *buf) { \
525 return sprintf(buf, "%lx\n", (unsigned long)var); \
527 static ssize_t set_ ## name(struct sys_device *s,const char *buf,size_t siz) { \
529 unsigned long new = simple_strtoul(buf, &end, 0); \
530 if (end == buf) return -EINVAL; \
535 static SYSDEV_ATTR(name, 0644, show_ ## name, set_ ## name);
537 ACCESSOR(bank0ctl,bank[0],mce_restart())
538 ACCESSOR(bank1ctl,bank[1],mce_restart())
539 ACCESSOR(bank2ctl,bank[2],mce_restart())
540 ACCESSOR(bank3ctl,bank[3],mce_restart())
541 ACCESSOR(bank4ctl,bank[4],mce_restart())
542 ACCESSOR(tolerant,tolerant,)
543 ACCESSOR(check_interval,check_interval,mce_restart())
545 static __cpuinit int mce_init_device(void)
548 if (!mce_available(&boot_cpu_data))
550 err = sysdev_class_register(&mce_sysclass);
552 err = sysdev_register(&device_mce);
554 /* could create per CPU objects, but it is not worth it. */
555 sysdev_create_file(&device_mce, &attr_bank0ctl);
556 sysdev_create_file(&device_mce, &attr_bank1ctl);
557 sysdev_create_file(&device_mce, &attr_bank2ctl);
558 sysdev_create_file(&device_mce, &attr_bank3ctl);
559 sysdev_create_file(&device_mce, &attr_bank4ctl);
560 sysdev_create_file(&device_mce, &attr_tolerant);
561 sysdev_create_file(&device_mce, &attr_check_interval);
564 misc_register(&mce_log_device);
568 device_initcall(mce_init_device);