f1c3769bbd6433344968bf5f78ad223170ef6878
[cascardo/linux.git] / arch / x86 / kernel / cpu / mcheck / mce_amd.c
1 /*
2  *  (c) 2005-2012 Advanced Micro Devices, Inc.
3  *  Your use of this code is subject to the terms and conditions of the
4  *  GNU general public license version 2. See "COPYING" or
5  *  http://www.gnu.org/licenses/gpl.html
6  *
7  *  Written by Jacob Shin - AMD, Inc.
8  *
9  *  Maintained by: Borislav Petkov <bp@alien8.de>
10  *
11  *  April 2006
12  *     - added support for AMD Family 0x10 processors
13  *  May 2012
14  *     - major scrubbing
15  *
16  *  All MC4_MISCi registers are shared between multi-cores
17  */
18 #include <linux/interrupt.h>
19 #include <linux/notifier.h>
20 #include <linux/kobject.h>
21 #include <linux/percpu.h>
22 #include <linux/errno.h>
23 #include <linux/sched.h>
24 #include <linux/sysfs.h>
25 #include <linux/slab.h>
26 #include <linux/init.h>
27 #include <linux/cpu.h>
28 #include <linux/smp.h>
29
30 #include <asm/amd_nb.h>
31 #include <asm/apic.h>
32 #include <asm/idle.h>
33 #include <asm/mce.h>
34 #include <asm/msr.h>
35
36 #define NR_BLOCKS         9
37 #define THRESHOLD_MAX     0xFFF
38 #define INT_TYPE_APIC     0x00020000
39 #define MASK_VALID_HI     0x80000000
40 #define MASK_CNTP_HI      0x40000000
41 #define MASK_LOCKED_HI    0x20000000
42 #define MASK_LVTOFF_HI    0x00F00000
43 #define MASK_COUNT_EN_HI  0x00080000
44 #define MASK_INT_TYPE_HI  0x00060000
45 #define MASK_OVERFLOW_HI  0x00010000
46 #define MASK_ERR_COUNT_HI 0x00000FFF
47 #define MASK_BLKPTR_LO    0xFF000000
48 #define MCG_XBLK_ADDR     0xC0000400
49
50 static const char * const th_names[] = {
51         "load_store",
52         "insn_fetch",
53         "combined_unit",
54         "",
55         "northbridge",
56         "execution_unit",
57 };
58
59 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
60 static DEFINE_PER_CPU(unsigned char, bank_map); /* see which banks are on */
61
62 static void amd_threshold_interrupt(void);
63
64 /*
65  * CPU Initialization
66  */
67
68 struct thresh_restart {
69         struct threshold_block  *b;
70         int                     reset;
71         int                     set_lvt_off;
72         int                     lvt_off;
73         u16                     old_limit;
74 };
75
76 static inline bool is_shared_bank(int bank)
77 {
78         /* Bank 4 is for northbridge reporting and is thus shared */
79         return (bank == 4);
80 }
81
82 static const char * const bank4_names(struct threshold_block *b)
83 {
84         switch (b->address) {
85         /* MSR4_MISC0 */
86         case 0x00000413:
87                 return "dram";
88
89         case 0xc0000408:
90                 return "ht_links";
91
92         case 0xc0000409:
93                 return "l3_cache";
94
95         default:
96                 WARN(1, "Funny MSR: 0x%08x\n", b->address);
97                 return "";
98         }
99 };
100
101
102 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
103 {
104         /*
105          * bank 4 supports APIC LVT interrupts implicitly since forever.
106          */
107         if (bank == 4)
108                 return true;
109
110         /*
111          * IntP: interrupt present; if this bit is set, the thresholding
112          * bank can generate APIC LVT interrupts
113          */
114         return msr_high_bits & BIT(28);
115 }
116
117 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
118 {
119         int msr = (hi & MASK_LVTOFF_HI) >> 20;
120
121         if (apic < 0) {
122                 pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
123                        "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
124                        b->bank, b->block, b->address, hi, lo);
125                 return 0;
126         }
127
128         if (apic != msr) {
129                 pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
130                        "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
131                        b->cpu, apic, b->bank, b->block, b->address, hi, lo);
132                 return 0;
133         }
134
135         return 1;
136 };
137
138 /*
139  * Called via smp_call_function_single(), must be called with correct
140  * cpu affinity.
141  */
142 static void threshold_restart_bank(void *_tr)
143 {
144         struct thresh_restart *tr = _tr;
145         u32 hi, lo;
146
147         rdmsr(tr->b->address, lo, hi);
148
149         if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
150                 tr->reset = 1;  /* limit cannot be lower than err count */
151
152         if (tr->reset) {                /* reset err count and overflow bit */
153                 hi =
154                     (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
155                     (THRESHOLD_MAX - tr->b->threshold_limit);
156         } else if (tr->old_limit) {     /* change limit w/o reset */
157                 int new_count = (hi & THRESHOLD_MAX) +
158                     (tr->old_limit - tr->b->threshold_limit);
159
160                 hi = (hi & ~MASK_ERR_COUNT_HI) |
161                     (new_count & THRESHOLD_MAX);
162         }
163
164         /* clear IntType */
165         hi &= ~MASK_INT_TYPE_HI;
166
167         if (!tr->b->interrupt_capable)
168                 goto done;
169
170         if (tr->set_lvt_off) {
171                 if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
172                         /* set new lvt offset */
173                         hi &= ~MASK_LVTOFF_HI;
174                         hi |= tr->lvt_off << 20;
175                 }
176         }
177
178         if (tr->b->interrupt_enable)
179                 hi |= INT_TYPE_APIC;
180
181  done:
182
183         hi |= MASK_COUNT_EN_HI;
184         wrmsr(tr->b->address, lo, hi);
185 }
186
187 static void mce_threshold_block_init(struct threshold_block *b, int offset)
188 {
189         struct thresh_restart tr = {
190                 .b                      = b,
191                 .set_lvt_off            = 1,
192                 .lvt_off                = offset,
193         };
194
195         b->threshold_limit              = THRESHOLD_MAX;
196         threshold_restart_bank(&tr);
197 };
198
199 static int setup_APIC_mce(int reserved, int new)
200 {
201         if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
202                                               APIC_EILVT_MSG_FIX, 0))
203                 return new;
204
205         return reserved;
206 }
207
208 /* cpu init entry point, called from mce.c with preempt off */
209 void mce_amd_feature_init(struct cpuinfo_x86 *c)
210 {
211         struct threshold_block b;
212         unsigned int cpu = smp_processor_id();
213         u32 low = 0, high = 0, address = 0;
214         unsigned int bank, block;
215         int offset = -1, new;
216
217         for (bank = 0; bank < mca_cfg.banks; ++bank) {
218                 for (block = 0; block < NR_BLOCKS; ++block) {
219                         if (block == 0)
220                                 address = MSR_IA32_MCx_MISC(bank);
221                         else if (block == 1) {
222                                 address = (low & MASK_BLKPTR_LO) >> 21;
223                                 if (!address)
224                                         break;
225
226                                 address += MCG_XBLK_ADDR;
227                         } else
228                                 ++address;
229
230                         if (rdmsr_safe(address, &low, &high))
231                                 break;
232
233                         if (!(high & MASK_VALID_HI))
234                                 continue;
235
236                         if (!(high & MASK_CNTP_HI)  ||
237                              (high & MASK_LOCKED_HI))
238                                 continue;
239
240                         if (!block)
241                                 per_cpu(bank_map, cpu) |= (1 << bank);
242
243                         memset(&b, 0, sizeof(b));
244                         b.cpu                   = cpu;
245                         b.bank                  = bank;
246                         b.block                 = block;
247                         b.address               = address;
248                         b.interrupt_capable     = lvt_interrupt_supported(bank, high);
249
250                         if (!b.interrupt_capable)
251                                 goto init;
252
253                         new     = (high & MASK_LVTOFF_HI) >> 20;
254                         offset  = setup_APIC_mce(offset, new);
255
256                         if ((offset == new) &&
257                             (mce_threshold_vector != amd_threshold_interrupt))
258                                 mce_threshold_vector = amd_threshold_interrupt;
259
260 init:
261                         mce_threshold_block_init(&b, offset);
262                 }
263         }
264 }
265
266 /*
267  * APIC Interrupt Handler
268  */
269
270 /*
271  * threshold interrupt handler will service THRESHOLD_APIC_VECTOR.
272  * the interrupt goes off when error_count reaches threshold_limit.
273  * the handler will simply log mcelog w/ software defined bank number.
274  */
275 static void amd_threshold_interrupt(void)
276 {
277         u32 low = 0, high = 0, address = 0;
278         int cpu = smp_processor_id();
279         unsigned int bank, block;
280         struct mce m;
281
282         /* assume first bank caused it */
283         for (bank = 0; bank < mca_cfg.banks; ++bank) {
284                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
285                         continue;
286                 for (block = 0; block < NR_BLOCKS; ++block) {
287                         if (block == 0) {
288                                 address = MSR_IA32_MCx_MISC(bank);
289                         } else if (block == 1) {
290                                 address = (low & MASK_BLKPTR_LO) >> 21;
291                                 if (!address)
292                                         break;
293                                 address += MCG_XBLK_ADDR;
294                         } else {
295                                 ++address;
296                         }
297
298                         if (rdmsr_safe(address, &low, &high))
299                                 break;
300
301                         if (!(high & MASK_VALID_HI)) {
302                                 if (block)
303                                         continue;
304                                 else
305                                         break;
306                         }
307
308                         if (!(high & MASK_CNTP_HI)  ||
309                              (high & MASK_LOCKED_HI))
310                                 continue;
311
312                         /*
313                          * Log the machine check that caused the threshold
314                          * event.
315                          */
316                         if (high & MASK_OVERFLOW_HI)
317                                 goto log;
318                 }
319         }
320         return;
321
322 log:
323         mce_setup(&m);
324         rdmsrl(MSR_IA32_MCx_STATUS(bank), m.status);
325         m.misc = ((u64)high << 32) | low;
326         m.bank = bank;
327         mce_log(&m);
328
329         wrmsrl(MSR_IA32_MCx_STATUS(bank), 0);
330 }
331
332 /*
333  * Sysfs Interface
334  */
335
336 struct threshold_attr {
337         struct attribute attr;
338         ssize_t (*show) (struct threshold_block *, char *);
339         ssize_t (*store) (struct threshold_block *, const char *, size_t count);
340 };
341
342 #define SHOW_FIELDS(name)                                               \
343 static ssize_t show_ ## name(struct threshold_block *b, char *buf)      \
344 {                                                                       \
345         return sprintf(buf, "%lu\n", (unsigned long) b->name);          \
346 }
347 SHOW_FIELDS(interrupt_enable)
348 SHOW_FIELDS(threshold_limit)
349
350 static ssize_t
351 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
352 {
353         struct thresh_restart tr;
354         unsigned long new;
355
356         if (!b->interrupt_capable)
357                 return -EINVAL;
358
359         if (kstrtoul(buf, 0, &new) < 0)
360                 return -EINVAL;
361
362         b->interrupt_enable = !!new;
363
364         memset(&tr, 0, sizeof(tr));
365         tr.b            = b;
366
367         smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
368
369         return size;
370 }
371
372 static ssize_t
373 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
374 {
375         struct thresh_restart tr;
376         unsigned long new;
377
378         if (kstrtoul(buf, 0, &new) < 0)
379                 return -EINVAL;
380
381         if (new > THRESHOLD_MAX)
382                 new = THRESHOLD_MAX;
383         if (new < 1)
384                 new = 1;
385
386         memset(&tr, 0, sizeof(tr));
387         tr.old_limit = b->threshold_limit;
388         b->threshold_limit = new;
389         tr.b = b;
390
391         smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
392
393         return size;
394 }
395
396 static ssize_t show_error_count(struct threshold_block *b, char *buf)
397 {
398         u32 lo, hi;
399
400         rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
401
402         return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
403                                      (THRESHOLD_MAX - b->threshold_limit)));
404 }
405
406 static struct threshold_attr error_count = {
407         .attr = {.name = __stringify(error_count), .mode = 0444 },
408         .show = show_error_count,
409 };
410
411 #define RW_ATTR(val)                                                    \
412 static struct threshold_attr val = {                                    \
413         .attr   = {.name = __stringify(val), .mode = 0644 },            \
414         .show   = show_## val,                                          \
415         .store  = store_## val,                                         \
416 };
417
418 RW_ATTR(interrupt_enable);
419 RW_ATTR(threshold_limit);
420
421 static struct attribute *default_attrs[] = {
422         &threshold_limit.attr,
423         &error_count.attr,
424         NULL,   /* possibly interrupt_enable if supported, see below */
425         NULL,
426 };
427
428 #define to_block(k)     container_of(k, struct threshold_block, kobj)
429 #define to_attr(a)      container_of(a, struct threshold_attr, attr)
430
431 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
432 {
433         struct threshold_block *b = to_block(kobj);
434         struct threshold_attr *a = to_attr(attr);
435         ssize_t ret;
436
437         ret = a->show ? a->show(b, buf) : -EIO;
438
439         return ret;
440 }
441
442 static ssize_t store(struct kobject *kobj, struct attribute *attr,
443                      const char *buf, size_t count)
444 {
445         struct threshold_block *b = to_block(kobj);
446         struct threshold_attr *a = to_attr(attr);
447         ssize_t ret;
448
449         ret = a->store ? a->store(b, buf, count) : -EIO;
450
451         return ret;
452 }
453
454 static const struct sysfs_ops threshold_ops = {
455         .show                   = show,
456         .store                  = store,
457 };
458
459 static struct kobj_type threshold_ktype = {
460         .sysfs_ops              = &threshold_ops,
461         .default_attrs          = default_attrs,
462 };
463
464 static int allocate_threshold_blocks(unsigned int cpu, unsigned int bank,
465                                      unsigned int block, u32 address)
466 {
467         struct threshold_block *b = NULL;
468         u32 low, high;
469         int err;
470
471         if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
472                 return 0;
473
474         if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
475                 return 0;
476
477         if (!(high & MASK_VALID_HI)) {
478                 if (block)
479                         goto recurse;
480                 else
481                         return 0;
482         }
483
484         if (!(high & MASK_CNTP_HI)  ||
485              (high & MASK_LOCKED_HI))
486                 goto recurse;
487
488         b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
489         if (!b)
490                 return -ENOMEM;
491
492         b->block                = block;
493         b->bank                 = bank;
494         b->cpu                  = cpu;
495         b->address              = address;
496         b->interrupt_enable     = 0;
497         b->interrupt_capable    = lvt_interrupt_supported(bank, high);
498         b->threshold_limit      = THRESHOLD_MAX;
499
500         if (b->interrupt_capable)
501                 threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
502         else
503                 threshold_ktype.default_attrs[2] = NULL;
504
505         INIT_LIST_HEAD(&b->miscj);
506
507         if (per_cpu(threshold_banks, cpu)[bank]->blocks) {
508                 list_add(&b->miscj,
509                          &per_cpu(threshold_banks, cpu)[bank]->blocks->miscj);
510         } else {
511                 per_cpu(threshold_banks, cpu)[bank]->blocks = b;
512         }
513
514         err = kobject_init_and_add(&b->kobj, &threshold_ktype,
515                                    per_cpu(threshold_banks, cpu)[bank]->kobj,
516                                    (bank == 4 ? bank4_names(b) : th_names[bank]));
517         if (err)
518                 goto out_free;
519 recurse:
520         if (!block) {
521                 address = (low & MASK_BLKPTR_LO) >> 21;
522                 if (!address)
523                         return 0;
524                 address += MCG_XBLK_ADDR;
525         } else {
526                 ++address;
527         }
528
529         err = allocate_threshold_blocks(cpu, bank, ++block, address);
530         if (err)
531                 goto out_free;
532
533         if (b)
534                 kobject_uevent(&b->kobj, KOBJ_ADD);
535
536         return err;
537
538 out_free:
539         if (b) {
540                 kobject_put(&b->kobj);
541                 list_del(&b->miscj);
542                 kfree(b);
543         }
544         return err;
545 }
546
547 static int __threshold_add_blocks(struct threshold_bank *b)
548 {
549         struct list_head *head = &b->blocks->miscj;
550         struct threshold_block *pos = NULL;
551         struct threshold_block *tmp = NULL;
552         int err = 0;
553
554         err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
555         if (err)
556                 return err;
557
558         list_for_each_entry_safe(pos, tmp, head, miscj) {
559
560                 err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
561                 if (err) {
562                         list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
563                                 kobject_del(&pos->kobj);
564
565                         return err;
566                 }
567         }
568         return err;
569 }
570
571 static int threshold_create_bank(unsigned int cpu, unsigned int bank)
572 {
573         struct device *dev = per_cpu(mce_device, cpu);
574         struct amd_northbridge *nb = NULL;
575         struct threshold_bank *b = NULL;
576         const char *name = th_names[bank];
577         int err = 0;
578
579         if (is_shared_bank(bank)) {
580                 nb = node_to_amd_nb(amd_get_nb_id(cpu));
581
582                 /* threshold descriptor already initialized on this node? */
583                 if (nb && nb->bank4) {
584                         /* yes, use it */
585                         b = nb->bank4;
586                         err = kobject_add(b->kobj, &dev->kobj, name);
587                         if (err)
588                                 goto out;
589
590                         per_cpu(threshold_banks, cpu)[bank] = b;
591                         atomic_inc(&b->cpus);
592
593                         err = __threshold_add_blocks(b);
594
595                         goto out;
596                 }
597         }
598
599         b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
600         if (!b) {
601                 err = -ENOMEM;
602                 goto out;
603         }
604
605         b->kobj = kobject_create_and_add(name, &dev->kobj);
606         if (!b->kobj) {
607                 err = -EINVAL;
608                 goto out_free;
609         }
610
611         per_cpu(threshold_banks, cpu)[bank] = b;
612
613         if (is_shared_bank(bank)) {
614                 atomic_set(&b->cpus, 1);
615
616                 /* nb is already initialized, see above */
617                 if (nb) {
618                         WARN_ON(nb->bank4);
619                         nb->bank4 = b;
620                 }
621         }
622
623         err = allocate_threshold_blocks(cpu, bank, 0, MSR_IA32_MCx_MISC(bank));
624         if (!err)
625                 goto out;
626
627  out_free:
628         kfree(b);
629
630  out:
631         return err;
632 }
633
634 /* create dir/files for all valid threshold banks */
635 static int threshold_create_device(unsigned int cpu)
636 {
637         unsigned int bank;
638         struct threshold_bank **bp;
639         int err = 0;
640
641         bp = kzalloc(sizeof(struct threshold_bank *) * mca_cfg.banks,
642                      GFP_KERNEL);
643         if (!bp)
644                 return -ENOMEM;
645
646         per_cpu(threshold_banks, cpu) = bp;
647
648         for (bank = 0; bank < mca_cfg.banks; ++bank) {
649                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
650                         continue;
651                 err = threshold_create_bank(cpu, bank);
652                 if (err)
653                         return err;
654         }
655
656         return err;
657 }
658
659 static void deallocate_threshold_block(unsigned int cpu,
660                                                  unsigned int bank)
661 {
662         struct threshold_block *pos = NULL;
663         struct threshold_block *tmp = NULL;
664         struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
665
666         if (!head)
667                 return;
668
669         list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
670                 kobject_put(&pos->kobj);
671                 list_del(&pos->miscj);
672                 kfree(pos);
673         }
674
675         kfree(per_cpu(threshold_banks, cpu)[bank]->blocks);
676         per_cpu(threshold_banks, cpu)[bank]->blocks = NULL;
677 }
678
679 static void __threshold_remove_blocks(struct threshold_bank *b)
680 {
681         struct threshold_block *pos = NULL;
682         struct threshold_block *tmp = NULL;
683
684         kobject_del(b->kobj);
685
686         list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
687                 kobject_del(&pos->kobj);
688 }
689
690 static void threshold_remove_bank(unsigned int cpu, int bank)
691 {
692         struct amd_northbridge *nb;
693         struct threshold_bank *b;
694
695         b = per_cpu(threshold_banks, cpu)[bank];
696         if (!b)
697                 return;
698
699         if (!b->blocks)
700                 goto free_out;
701
702         if (is_shared_bank(bank)) {
703                 if (!atomic_dec_and_test(&b->cpus)) {
704                         __threshold_remove_blocks(b);
705                         per_cpu(threshold_banks, cpu)[bank] = NULL;
706                         return;
707                 } else {
708                         /*
709                          * the last CPU on this node using the shared bank is
710                          * going away, remove that bank now.
711                          */
712                         nb = node_to_amd_nb(amd_get_nb_id(cpu));
713                         nb->bank4 = NULL;
714                 }
715         }
716
717         deallocate_threshold_block(cpu, bank);
718
719 free_out:
720         kobject_del(b->kobj);
721         kobject_put(b->kobj);
722         kfree(b);
723         per_cpu(threshold_banks, cpu)[bank] = NULL;
724 }
725
726 static void threshold_remove_device(unsigned int cpu)
727 {
728         unsigned int bank;
729
730         for (bank = 0; bank < mca_cfg.banks; ++bank) {
731                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
732                         continue;
733                 threshold_remove_bank(cpu, bank);
734         }
735         kfree(per_cpu(threshold_banks, cpu));
736 }
737
738 /* get notified when a cpu comes on/off */
739 static void
740 amd_64_threshold_cpu_callback(unsigned long action, unsigned int cpu)
741 {
742         switch (action) {
743         case CPU_ONLINE:
744         case CPU_ONLINE_FROZEN:
745                 threshold_create_device(cpu);
746                 break;
747         case CPU_DEAD:
748         case CPU_DEAD_FROZEN:
749                 threshold_remove_device(cpu);
750                 break;
751         default:
752                 break;
753         }
754 }
755
756 static __init int threshold_init_device(void)
757 {
758         unsigned lcpu = 0;
759
760         /* to hit CPUs online before the notifier is up */
761         for_each_online_cpu(lcpu) {
762                 int err = threshold_create_device(lcpu);
763
764                 if (err)
765                         return err;
766         }
767         threshold_cpu_callback = amd_64_threshold_cpu_callback;
768
769         return 0;
770 }
771 /*
772  * there are 3 funcs which need to be _initcalled in a logic sequence:
773  * 1. xen_late_init_mcelog
774  * 2. mcheck_init_device
775  * 3. threshold_init_device
776  *
777  * xen_late_init_mcelog must register xen_mce_chrdev_device before
778  * native mce_chrdev_device registration if running under xen platform;
779  *
780  * mcheck_init_device should be inited before threshold_init_device to
781  * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
782  *
783  * so we use following _initcalls
784  * 1. device_initcall(xen_late_init_mcelog);
785  * 2. device_initcall_sync(mcheck_init_device);
786  * 3. late_initcall(threshold_init_device);
787  *
788  * when running under xen, the initcall order is 1,2,3;
789  * on baremetal, we skip 1 and we do only 2 and 3.
790  */
791 late_initcall(threshold_init_device);