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_disabled __initdata;
28 /* 0: always panic, 1: panic if deadlock possible, 2: try to avoid panic,
29 3: never panic or exit (for testing only) */
30 static int tolerant = 1;
32 static unsigned long bank[NR_BANKS] = { [0 ... NR_BANKS-1] = ~0UL };
35 * Lockless MCE logging infrastructure.
36 * This avoids deadlocks on printk locks without having to break locks. Also
37 * separate MCEs from kernel messages to avoid bogus bug reports.
40 struct mce_log mcelog = {
45 static void mce_log(struct mce *mce)
52 read_barrier_depends();
53 /* When the buffer fills up discard new entries. Assume
54 that the earlier errors are the more interesting. */
55 if (entry >= MCE_LOG_LEN) {
56 set_bit(MCE_OVERFLOW, &mcelog.flags);
59 /* Old left over entry. Skip. */
60 if (mcelog.entry[entry].finished)
64 if (cmpxchg(&mcelog.next, entry, next) == entry)
67 memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
69 mcelog.entry[entry].finished = 1;
73 static void print_mce(struct mce *m)
76 "CPU %d: Machine Check Exception: %16Lx Bank %d: %016Lx\n",
77 m->cpu, m->mcgstatus, m->bank, m->status);
80 "RIP%s %02x:<%016Lx> ",
81 !(m->mcgstatus & MCG_STATUS_EIPV) ? " !INEXACT!" : "",
83 if (m->cs == __KERNEL_CS)
84 print_symbol("{%s}", m->rip);
87 printk(KERN_EMERG "TSC %Lx ", m->tsc);
89 printk("ADDR %Lx ", m->addr);
91 printk("MISC %Lx ", m->misc);
95 static void mce_panic(char *msg, struct mce *backup, unsigned long start)
99 for (i = 0; i < MCE_LOG_LEN; i++) {
100 unsigned long tsc = mcelog.entry[i].tsc;
101 if (time_before(tsc, start))
103 print_mce(&mcelog.entry[i]);
104 if (mcelog.entry[i].tsc == backup->tsc)
110 printk("Fake panic: %s\n", msg);
115 static int mce_available(struct cpuinfo_x86 *c)
117 return !mce_disabled &&
118 test_bit(X86_FEATURE_MCE, &c->x86_capability) &&
119 test_bit(X86_FEATURE_MCA, &c->x86_capability);
123 * The actual machine check handler
126 void do_machine_check(struct pt_regs * regs, long error_code)
128 struct mce m, panicm;
129 int nowayout = (tolerant < 1);
135 notify_die(DIE_NMI, "machine check", regs, error_code, 255, SIGKILL);
139 memset(&m, 0, sizeof(struct mce));
140 m.cpu = hard_smp_processor_id();
141 rdmsrl(MSR_IA32_MCG_STATUS, m.mcgstatus);
142 if (!regs && (m.mcgstatus & MCG_STATUS_MCIP))
144 if (!(m.mcgstatus & MCG_STATUS_RIPV))
154 for (i = 0; i < banks; i++) {
161 rdmsrl(MSR_IA32_MC0_STATUS + i*4, m.status);
162 if ((m.status & MCI_STATUS_VAL) == 0)
164 /* Should be implied by the banks check above, but
166 if ((m.status & MCI_STATUS_EN) == 0)
169 /* Did this bank cause the exception? */
170 /* Assume that the bank with uncorrectable errors did it,
171 and that there is only a single one. */
172 if (m.status & MCI_STATUS_UC) {
179 /* In theory _OVER could be a nowayout too, but
180 assume any overflowed errors were no fatal. */
181 nowayout |= !!(m.status & MCI_STATUS_PCC);
182 kill_it |= !!(m.status & MCI_STATUS_UC);
185 if (m.status & MCI_STATUS_MISCV)
186 rdmsrl(MSR_IA32_MC0_MISC + i*4, m.misc);
187 if (m.status & MCI_STATUS_ADDRV)
188 rdmsrl(MSR_IA32_MC0_ADDR + i*4, m.addr);
191 wrmsrl(MSR_IA32_MC0_STATUS + i*4, 0);
194 wrmsrl(MSR_IA32_MCG_STATUS, 0);
196 /* Never do anything final in the polling timer */
200 mce_panic("Machine check", &m, mcestart);
204 if (m.mcgstatus & MCG_STATUS_RIPV)
205 user_space = m.rip && (m.cs & 3);
207 /* When the machine was in user space and the CPU didn't get
208 confused it's normally not necessary to panic, unless you
209 are paranoid (tolerant == 0)
211 RED-PEN could be more tolerant for MCEs in idle,
212 but most likely they occur at boot anyways, where
213 it is best to just halt the machine. */
214 if ((!user_space && (panic_on_oops || tolerant < 2)) ||
215 (unsigned)current->pid <= 1)
216 mce_panic("Uncorrected machine check", &panicm, mcestart);
218 /* do_exit takes an awful lot of locks and has as slight risk
219 of deadlocking. If you don't want that don't set tolerant >= 2 */
225 static void mce_clear_all(void)
228 for (i = 0; i < banks; i++)
229 wrmsrl(MSR_IA32_MC0_STATUS + i*4, 0);
230 wrmsrl(MSR_IA32_MCG_STATUS, 0);
234 * Periodic polling timer for "silent" machine check errors.
237 static int check_interval = 5 * 60; /* 5 minutes */
238 static void mcheck_timer(void *data);
239 static DECLARE_WORK(mcheck_work, mcheck_timer, NULL);
241 static void mcheck_check_cpu(void *info)
243 if (mce_available(¤t_cpu_data))
244 do_machine_check(NULL, 0);
247 static void mcheck_timer(void *data)
249 on_each_cpu(mcheck_check_cpu, NULL, 1, 1);
250 schedule_delayed_work(&mcheck_work, check_interval * HZ);
254 static __init int periodic_mcheck_init(void)
257 schedule_delayed_work(&mcheck_work, check_interval*HZ);
260 __initcall(periodic_mcheck_init);
264 * Initialize Machine Checks for a CPU.
266 static void mce_init(void *dummy)
271 rdmsrl(MSR_IA32_MCG_CAP, cap);
273 wrmsr(MSR_IA32_MCG_CTL, 0xffffffff, 0xffffffff);
276 if (banks > NR_BANKS) {
277 printk(KERN_INFO "MCE: warning: using only %d banks\n", banks);
282 for (i = 0; i < banks; i++) {
283 wrmsrl(MSR_IA32_MC0_CTL+4*i, bank[i]);
284 wrmsrl(MSR_IA32_MC0_STATUS+4*i, 0);
287 set_in_cr4(X86_CR4_MCE);
290 /* Add per CPU specific workarounds here */
291 static void __init mce_cpu_quirks(struct cpuinfo_x86 *c)
293 /* This should be disabled by the BIOS, but isn't always */
294 if (c->x86_vendor == X86_VENDOR_AMD && c->x86 == 15) {
295 /* disable GART TBL walk error reporting, which trips off
296 incorrectly with the IOMMU & 3ware & Cerberus. */
297 clear_bit(10, &bank[4]);
302 * Called for each booted CPU to set up machine checks.
303 * Must be called with preempt off.
305 void __init mcheck_init(struct cpuinfo_x86 *c)
307 static unsigned long mce_cpus __initdata = 0;
311 if (test_and_set_bit(smp_processor_id(), &mce_cpus) || !mce_available(c))
318 * Character device to read and clear the MCE log.
321 static void collect_tscs(void *data)
323 unsigned long *cpu_tsc = (unsigned long *)data;
324 rdtscll(cpu_tsc[smp_processor_id()]);
327 static ssize_t mce_read(struct file *filp, char __user *ubuf, size_t usize, loff_t *off)
329 unsigned long cpu_tsc[NR_CPUS];
330 static DECLARE_MUTEX(mce_read_sem);
332 char __user *buf = ubuf;
337 read_barrier_depends();
339 /* Only supports full reads right now */
340 if (*off != 0 || usize < MCE_LOG_LEN*sizeof(struct mce)) {
346 for (i = 0; i < next; i++) {
347 if (!mcelog.entry[i].finished)
350 err |= copy_to_user(buf, mcelog.entry + i, sizeof(struct mce));
351 buf += sizeof(struct mce);
354 memset(mcelog.entry, 0, next * sizeof(struct mce));
358 synchronize_kernel();
360 /* Collect entries that were still getting written before the synchronize. */
362 on_each_cpu(collect_tscs, cpu_tsc, 1, 1);
363 for (i = next; i < MCE_LOG_LEN; i++) {
364 if (mcelog.entry[i].finished &&
365 mcelog.entry[i].tsc < cpu_tsc[mcelog.entry[i].cpu]) {
366 err |= copy_to_user(buf, mcelog.entry+i, sizeof(struct mce));
368 buf += sizeof(struct mce);
369 memset(&mcelog.entry[i], 0, sizeof(struct mce));
373 return err ? -EFAULT : buf - ubuf;
376 static int mce_ioctl(struct inode *i, struct file *f,unsigned int cmd, unsigned long arg)
378 int __user *p = (int __user *)arg;
379 if (!capable(CAP_SYS_ADMIN))
382 case MCE_GET_RECORD_LEN:
383 return put_user(sizeof(struct mce), p);
384 case MCE_GET_LOG_LEN:
385 return put_user(MCE_LOG_LEN, p);
386 case MCE_GETCLEAR_FLAGS: {
389 flags = mcelog.flags;
390 } while (cmpxchg(&mcelog.flags, flags, 0) != flags);
391 return put_user(flags, p);
398 static struct file_operations mce_chrdev_ops = {
403 static struct miscdevice mce_log_device = {
410 * Old style boot options parsing. Only for compatibility.
413 static int __init mcheck_disable(char *str)
419 /* mce=off disable machine check */
420 static int __init mcheck_enable(char *str)
422 if (!strcmp(str, "off"))
425 printk("mce= argument %s ignored. Please use /sys", str);
429 __setup("nomce", mcheck_disable);
430 __setup("mce", mcheck_enable);
436 /* On resume clear all MCE state. Don't want to see leftovers from the BIOS. */
437 static int mce_resume(struct sys_device *dev)
440 on_each_cpu(mce_init, NULL, 1, 1);
444 /* Reinit MCEs after user configuration changes */
445 static void mce_restart(void)
448 cancel_delayed_work(&mcheck_work);
449 /* Timer race is harmless here */
450 on_each_cpu(mce_init, NULL, 1, 1);
452 schedule_delayed_work(&mcheck_work, check_interval*HZ);
455 static struct sysdev_class mce_sysclass = {
456 .resume = mce_resume,
457 set_kset_name("machinecheck"),
460 static struct sys_device device_mce = {
462 .cls = &mce_sysclass,
465 /* Why are there no generic functions for this? */
466 #define ACCESSOR(name, var, start) \
467 static ssize_t show_ ## name(struct sys_device *s, char *buf) { \
468 return sprintf(buf, "%lu\n", (unsigned long)var); \
470 static ssize_t set_ ## name(struct sys_device *s,const char *buf,size_t siz) { \
472 unsigned long new = simple_strtoul(buf, &end, 0); \
473 if (end == buf) return -EINVAL; \
478 static SYSDEV_ATTR(name, 0644, show_ ## name, set_ ## name);
480 ACCESSOR(bank0ctl,bank[0],mce_restart())
481 ACCESSOR(bank1ctl,bank[1],mce_restart())
482 ACCESSOR(bank2ctl,bank[2],mce_restart())
483 ACCESSOR(bank3ctl,bank[3],mce_restart())
484 ACCESSOR(bank4ctl,bank[4],mce_restart())
485 ACCESSOR(tolerant,tolerant,)
486 ACCESSOR(check_interval,check_interval,mce_restart())
488 static __init int mce_init_device(void)
491 if (!mce_available(&boot_cpu_data))
493 err = sysdev_class_register(&mce_sysclass);
495 err = sysdev_register(&device_mce);
497 /* could create per CPU objects, but is not worth it. */
498 sysdev_create_file(&device_mce, &attr_bank0ctl);
499 sysdev_create_file(&device_mce, &attr_bank1ctl);
500 sysdev_create_file(&device_mce, &attr_bank2ctl);
501 sysdev_create_file(&device_mce, &attr_bank3ctl);
502 sysdev_create_file(&device_mce, &attr_bank4ctl);
503 sysdev_create_file(&device_mce, &attr_tolerant);
504 sysdev_create_file(&device_mce, &attr_check_interval);
507 misc_register(&mce_log_device);
511 device_initcall(mce_init_device);