xref: /openbmc/linux/arch/x86/kernel/cpu/mce/amd.c (revision db970bd2)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  (c) 2005-2016 Advanced Micro Devices, Inc.
4  *
5  *  Written by Jacob Shin - AMD, Inc.
6  *  Maintained by: Borislav Petkov <bp@alien8.de>
7  *
8  *  All MC4_MISCi registers are shared between cores on a node.
9  */
10 #include <linux/interrupt.h>
11 #include <linux/notifier.h>
12 #include <linux/kobject.h>
13 #include <linux/percpu.h>
14 #include <linux/errno.h>
15 #include <linux/sched.h>
16 #include <linux/sysfs.h>
17 #include <linux/slab.h>
18 #include <linux/init.h>
19 #include <linux/cpu.h>
20 #include <linux/smp.h>
21 #include <linux/string.h>
22 
23 #include <asm/amd_nb.h>
24 #include <asm/traps.h>
25 #include <asm/apic.h>
26 #include <asm/mce.h>
27 #include <asm/msr.h>
28 #include <asm/trace/irq_vectors.h>
29 
30 #include "internal.h"
31 
32 #define NR_BLOCKS         5
33 #define THRESHOLD_MAX     0xFFF
34 #define INT_TYPE_APIC     0x00020000
35 #define MASK_VALID_HI     0x80000000
36 #define MASK_CNTP_HI      0x40000000
37 #define MASK_LOCKED_HI    0x20000000
38 #define MASK_LVTOFF_HI    0x00F00000
39 #define MASK_COUNT_EN_HI  0x00080000
40 #define MASK_INT_TYPE_HI  0x00060000
41 #define MASK_OVERFLOW_HI  0x00010000
42 #define MASK_ERR_COUNT_HI 0x00000FFF
43 #define MASK_BLKPTR_LO    0xFF000000
44 #define MCG_XBLK_ADDR     0xC0000400
45 
46 /* Deferred error settings */
47 #define MSR_CU_DEF_ERR		0xC0000410
48 #define MASK_DEF_LVTOFF		0x000000F0
49 #define MASK_DEF_INT_TYPE	0x00000006
50 #define DEF_LVT_OFF		0x2
51 #define DEF_INT_TYPE_APIC	0x2
52 
53 /* Scalable MCA: */
54 
55 /* Threshold LVT offset is at MSR0xC0000410[15:12] */
56 #define SMCA_THR_LVT_OFF	0xF000
57 
58 static bool thresholding_irq_en;
59 
60 static const char * const th_names[] = {
61 	"load_store",
62 	"insn_fetch",
63 	"combined_unit",
64 	"decode_unit",
65 	"northbridge",
66 	"execution_unit",
67 };
68 
69 static const char * const smca_umc_block_names[] = {
70 	"dram_ecc",
71 	"misc_umc"
72 };
73 
74 struct smca_bank_name {
75 	const char *name;	/* Short name for sysfs */
76 	const char *long_name;	/* Long name for pretty-printing */
77 };
78 
79 static struct smca_bank_name smca_names[] = {
80 	[SMCA_LS]	= { "load_store",	"Load Store Unit" },
81 	[SMCA_LS_V2]	= { "load_store",	"Load Store Unit" },
82 	[SMCA_IF]	= { "insn_fetch",	"Instruction Fetch Unit" },
83 	[SMCA_L2_CACHE]	= { "l2_cache",		"L2 Cache" },
84 	[SMCA_DE]	= { "decode_unit",	"Decode Unit" },
85 	[SMCA_RESERVED]	= { "reserved",		"Reserved" },
86 	[SMCA_EX]	= { "execution_unit",	"Execution Unit" },
87 	[SMCA_FP]	= { "floating_point",	"Floating Point Unit" },
88 	[SMCA_L3_CACHE]	= { "l3_cache",		"L3 Cache" },
89 	[SMCA_CS]	= { "coherent_slave",	"Coherent Slave" },
90 	[SMCA_CS_V2]	= { "coherent_slave",	"Coherent Slave" },
91 	[SMCA_PIE]	= { "pie",		"Power, Interrupts, etc." },
92 	[SMCA_UMC]	= { "umc",		"Unified Memory Controller" },
93 	[SMCA_PB]	= { "param_block",	"Parameter Block" },
94 	[SMCA_PSP]	= { "psp",		"Platform Security Processor" },
95 	[SMCA_PSP_V2]	= { "psp",		"Platform Security Processor" },
96 	[SMCA_SMU]	= { "smu",		"System Management Unit" },
97 	[SMCA_SMU_V2]	= { "smu",		"System Management Unit" },
98 	[SMCA_MP5]	= { "mp5",		"Microprocessor 5 Unit" },
99 	[SMCA_NBIO]	= { "nbio",		"Northbridge IO Unit" },
100 	[SMCA_PCIE]	= { "pcie",		"PCI Express Unit" },
101 };
102 
103 static const char *smca_get_name(enum smca_bank_types t)
104 {
105 	if (t >= N_SMCA_BANK_TYPES)
106 		return NULL;
107 
108 	return smca_names[t].name;
109 }
110 
111 const char *smca_get_long_name(enum smca_bank_types t)
112 {
113 	if (t >= N_SMCA_BANK_TYPES)
114 		return NULL;
115 
116 	return smca_names[t].long_name;
117 }
118 EXPORT_SYMBOL_GPL(smca_get_long_name);
119 
120 static enum smca_bank_types smca_get_bank_type(unsigned int bank)
121 {
122 	struct smca_bank *b;
123 
124 	if (bank >= MAX_NR_BANKS)
125 		return N_SMCA_BANK_TYPES;
126 
127 	b = &smca_banks[bank];
128 	if (!b->hwid)
129 		return N_SMCA_BANK_TYPES;
130 
131 	return b->hwid->bank_type;
132 }
133 
134 static struct smca_hwid smca_hwid_mcatypes[] = {
135 	/* { bank_type, hwid_mcatype } */
136 
137 	/* Reserved type */
138 	{ SMCA_RESERVED, HWID_MCATYPE(0x00, 0x0)	},
139 
140 	/* ZN Core (HWID=0xB0) MCA types */
141 	{ SMCA_LS,	 HWID_MCATYPE(0xB0, 0x0)	},
142 	{ SMCA_LS_V2,	 HWID_MCATYPE(0xB0, 0x10)	},
143 	{ SMCA_IF,	 HWID_MCATYPE(0xB0, 0x1)	},
144 	{ SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2)	},
145 	{ SMCA_DE,	 HWID_MCATYPE(0xB0, 0x3)	},
146 	/* HWID 0xB0 MCATYPE 0x4 is Reserved */
147 	{ SMCA_EX,	 HWID_MCATYPE(0xB0, 0x5)	},
148 	{ SMCA_FP,	 HWID_MCATYPE(0xB0, 0x6)	},
149 	{ SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7)	},
150 
151 	/* Data Fabric MCA types */
152 	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0)	},
153 	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1)	},
154 	{ SMCA_CS_V2,	 HWID_MCATYPE(0x2E, 0x2)	},
155 
156 	/* Unified Memory Controller MCA type */
157 	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0)	},
158 
159 	/* Parameter Block MCA type */
160 	{ SMCA_PB,	 HWID_MCATYPE(0x05, 0x0)	},
161 
162 	/* Platform Security Processor MCA type */
163 	{ SMCA_PSP,	 HWID_MCATYPE(0xFF, 0x0)	},
164 	{ SMCA_PSP_V2,	 HWID_MCATYPE(0xFF, 0x1)	},
165 
166 	/* System Management Unit MCA type */
167 	{ SMCA_SMU,	 HWID_MCATYPE(0x01, 0x0)	},
168 	{ SMCA_SMU_V2,	 HWID_MCATYPE(0x01, 0x1)	},
169 
170 	/* Microprocessor 5 Unit MCA type */
171 	{ SMCA_MP5,	 HWID_MCATYPE(0x01, 0x2)	},
172 
173 	/* Northbridge IO Unit MCA type */
174 	{ SMCA_NBIO,	 HWID_MCATYPE(0x18, 0x0)	},
175 
176 	/* PCI Express Unit MCA type */
177 	{ SMCA_PCIE,	 HWID_MCATYPE(0x46, 0x0)	},
178 };
179 
180 struct smca_bank smca_banks[MAX_NR_BANKS];
181 EXPORT_SYMBOL_GPL(smca_banks);
182 
183 /*
184  * In SMCA enabled processors, we can have multiple banks for a given IP type.
185  * So to define a unique name for each bank, we use a temp c-string to append
186  * the MCA_IPID[InstanceId] to type's name in get_name().
187  *
188  * InstanceId is 32 bits which is 8 characters. Make sure MAX_MCATYPE_NAME_LEN
189  * is greater than 8 plus 1 (for underscore) plus length of longest type name.
190  */
191 #define MAX_MCATYPE_NAME_LEN	30
192 static char buf_mcatype[MAX_MCATYPE_NAME_LEN];
193 
194 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
195 
196 /*
197  * A list of the banks enabled on each logical CPU. Controls which respective
198  * descriptors to initialize later in mce_threshold_create_device().
199  */
200 static DEFINE_PER_CPU(unsigned int, bank_map);
201 
202 /* Map of banks that have more than MCA_MISC0 available. */
203 static DEFINE_PER_CPU(u32, smca_misc_banks_map);
204 
205 static void amd_threshold_interrupt(void);
206 static void amd_deferred_error_interrupt(void);
207 
208 static void default_deferred_error_interrupt(void)
209 {
210 	pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
211 }
212 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
213 
214 static void smca_set_misc_banks_map(unsigned int bank, unsigned int cpu)
215 {
216 	u32 low, high;
217 
218 	/*
219 	 * For SMCA enabled processors, BLKPTR field of the first MISC register
220 	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
221 	 */
222 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
223 		return;
224 
225 	if (!(low & MCI_CONFIG_MCAX))
226 		return;
227 
228 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high))
229 		return;
230 
231 	if (low & MASK_BLKPTR_LO)
232 		per_cpu(smca_misc_banks_map, cpu) |= BIT(bank);
233 
234 }
235 
236 static void smca_configure(unsigned int bank, unsigned int cpu)
237 {
238 	unsigned int i, hwid_mcatype;
239 	struct smca_hwid *s_hwid;
240 	u32 high, low;
241 	u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
242 
243 	/* Set appropriate bits in MCA_CONFIG */
244 	if (!rdmsr_safe(smca_config, &low, &high)) {
245 		/*
246 		 * OS is required to set the MCAX bit to acknowledge that it is
247 		 * now using the new MSR ranges and new registers under each
248 		 * bank. It also means that the OS will configure deferred
249 		 * errors in the new MCx_CONFIG register. If the bit is not set,
250 		 * uncorrectable errors will cause a system panic.
251 		 *
252 		 * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
253 		 */
254 		high |= BIT(0);
255 
256 		/*
257 		 * SMCA sets the Deferred Error Interrupt type per bank.
258 		 *
259 		 * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
260 		 * if the DeferredIntType bit field is available.
261 		 *
262 		 * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
263 		 * high portion of the MSR). OS should set this to 0x1 to enable
264 		 * APIC based interrupt. First, check that no interrupt has been
265 		 * set.
266 		 */
267 		if ((low & BIT(5)) && !((high >> 5) & 0x3))
268 			high |= BIT(5);
269 
270 		wrmsr(smca_config, low, high);
271 	}
272 
273 	smca_set_misc_banks_map(bank, cpu);
274 
275 	/* Return early if this bank was already initialized. */
276 	if (smca_banks[bank].hwid && smca_banks[bank].hwid->hwid_mcatype != 0)
277 		return;
278 
279 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
280 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
281 		return;
282 	}
283 
284 	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
285 				    (high & MCI_IPID_MCATYPE) >> 16);
286 
287 	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
288 		s_hwid = &smca_hwid_mcatypes[i];
289 		if (hwid_mcatype == s_hwid->hwid_mcatype) {
290 			smca_banks[bank].hwid = s_hwid;
291 			smca_banks[bank].id = low;
292 			smca_banks[bank].sysfs_id = s_hwid->count++;
293 			break;
294 		}
295 	}
296 }
297 
298 struct thresh_restart {
299 	struct threshold_block	*b;
300 	int			reset;
301 	int			set_lvt_off;
302 	int			lvt_off;
303 	u16			old_limit;
304 };
305 
306 static inline bool is_shared_bank(int bank)
307 {
308 	/*
309 	 * Scalable MCA provides for only one core to have access to the MSRs of
310 	 * a shared bank.
311 	 */
312 	if (mce_flags.smca)
313 		return false;
314 
315 	/* Bank 4 is for northbridge reporting and is thus shared */
316 	return (bank == 4);
317 }
318 
319 static const char *bank4_names(const struct threshold_block *b)
320 {
321 	switch (b->address) {
322 	/* MSR4_MISC0 */
323 	case 0x00000413:
324 		return "dram";
325 
326 	case 0xc0000408:
327 		return "ht_links";
328 
329 	case 0xc0000409:
330 		return "l3_cache";
331 
332 	default:
333 		WARN(1, "Funny MSR: 0x%08x\n", b->address);
334 		return "";
335 	}
336 };
337 
338 
339 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
340 {
341 	/*
342 	 * bank 4 supports APIC LVT interrupts implicitly since forever.
343 	 */
344 	if (bank == 4)
345 		return true;
346 
347 	/*
348 	 * IntP: interrupt present; if this bit is set, the thresholding
349 	 * bank can generate APIC LVT interrupts
350 	 */
351 	return msr_high_bits & BIT(28);
352 }
353 
354 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
355 {
356 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
357 
358 	if (apic < 0) {
359 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
360 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
361 		       b->bank, b->block, b->address, hi, lo);
362 		return 0;
363 	}
364 
365 	if (apic != msr) {
366 		/*
367 		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
368 		 * the BIOS provides the value. The original field where LVT offset
369 		 * was set is reserved. Return early here:
370 		 */
371 		if (mce_flags.smca)
372 			return 0;
373 
374 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
375 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
376 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
377 		return 0;
378 	}
379 
380 	return 1;
381 };
382 
383 /* Reprogram MCx_MISC MSR behind this threshold bank. */
384 static void threshold_restart_bank(void *_tr)
385 {
386 	struct thresh_restart *tr = _tr;
387 	u32 hi, lo;
388 
389 	/* sysfs write might race against an offline operation */
390 	if (this_cpu_read(threshold_banks))
391 		return;
392 
393 	rdmsr(tr->b->address, lo, hi);
394 
395 	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
396 		tr->reset = 1;	/* limit cannot be lower than err count */
397 
398 	if (tr->reset) {		/* reset err count and overflow bit */
399 		hi =
400 		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
401 		    (THRESHOLD_MAX - tr->b->threshold_limit);
402 	} else if (tr->old_limit) {	/* change limit w/o reset */
403 		int new_count = (hi & THRESHOLD_MAX) +
404 		    (tr->old_limit - tr->b->threshold_limit);
405 
406 		hi = (hi & ~MASK_ERR_COUNT_HI) |
407 		    (new_count & THRESHOLD_MAX);
408 	}
409 
410 	/* clear IntType */
411 	hi &= ~MASK_INT_TYPE_HI;
412 
413 	if (!tr->b->interrupt_capable)
414 		goto done;
415 
416 	if (tr->set_lvt_off) {
417 		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
418 			/* set new lvt offset */
419 			hi &= ~MASK_LVTOFF_HI;
420 			hi |= tr->lvt_off << 20;
421 		}
422 	}
423 
424 	if (tr->b->interrupt_enable)
425 		hi |= INT_TYPE_APIC;
426 
427  done:
428 
429 	hi |= MASK_COUNT_EN_HI;
430 	wrmsr(tr->b->address, lo, hi);
431 }
432 
433 static void mce_threshold_block_init(struct threshold_block *b, int offset)
434 {
435 	struct thresh_restart tr = {
436 		.b			= b,
437 		.set_lvt_off		= 1,
438 		.lvt_off		= offset,
439 	};
440 
441 	b->threshold_limit		= THRESHOLD_MAX;
442 	threshold_restart_bank(&tr);
443 };
444 
445 static int setup_APIC_mce_threshold(int reserved, int new)
446 {
447 	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
448 					      APIC_EILVT_MSG_FIX, 0))
449 		return new;
450 
451 	return reserved;
452 }
453 
454 static int setup_APIC_deferred_error(int reserved, int new)
455 {
456 	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
457 					      APIC_EILVT_MSG_FIX, 0))
458 		return new;
459 
460 	return reserved;
461 }
462 
463 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
464 {
465 	u32 low = 0, high = 0;
466 	int def_offset = -1, def_new;
467 
468 	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
469 		return;
470 
471 	def_new = (low & MASK_DEF_LVTOFF) >> 4;
472 	if (!(low & MASK_DEF_LVTOFF)) {
473 		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
474 		def_new = DEF_LVT_OFF;
475 		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
476 	}
477 
478 	def_offset = setup_APIC_deferred_error(def_offset, def_new);
479 	if ((def_offset == def_new) &&
480 	    (deferred_error_int_vector != amd_deferred_error_interrupt))
481 		deferred_error_int_vector = amd_deferred_error_interrupt;
482 
483 	if (!mce_flags.smca)
484 		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
485 
486 	wrmsr(MSR_CU_DEF_ERR, low, high);
487 }
488 
489 static u32 smca_get_block_address(unsigned int bank, unsigned int block,
490 				  unsigned int cpu)
491 {
492 	if (!block)
493 		return MSR_AMD64_SMCA_MCx_MISC(bank);
494 
495 	if (!(per_cpu(smca_misc_banks_map, cpu) & BIT(bank)))
496 		return 0;
497 
498 	return MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
499 }
500 
501 static u32 get_block_address(u32 current_addr, u32 low, u32 high,
502 			     unsigned int bank, unsigned int block,
503 			     unsigned int cpu)
504 {
505 	u32 addr = 0, offset = 0;
506 
507 	if ((bank >= per_cpu(mce_num_banks, cpu)) || (block >= NR_BLOCKS))
508 		return addr;
509 
510 	if (mce_flags.smca)
511 		return smca_get_block_address(bank, block, cpu);
512 
513 	/* Fall back to method we used for older processors: */
514 	switch (block) {
515 	case 0:
516 		addr = msr_ops.misc(bank);
517 		break;
518 	case 1:
519 		offset = ((low & MASK_BLKPTR_LO) >> 21);
520 		if (offset)
521 			addr = MCG_XBLK_ADDR + offset;
522 		break;
523 	default:
524 		addr = ++current_addr;
525 	}
526 	return addr;
527 }
528 
529 static int
530 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
531 			int offset, u32 misc_high)
532 {
533 	unsigned int cpu = smp_processor_id();
534 	u32 smca_low, smca_high;
535 	struct threshold_block b;
536 	int new;
537 
538 	if (!block)
539 		per_cpu(bank_map, cpu) |= (1 << bank);
540 
541 	memset(&b, 0, sizeof(b));
542 	b.cpu			= cpu;
543 	b.bank			= bank;
544 	b.block			= block;
545 	b.address		= addr;
546 	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
547 
548 	if (!b.interrupt_capable)
549 		goto done;
550 
551 	b.interrupt_enable = 1;
552 
553 	if (!mce_flags.smca) {
554 		new = (misc_high & MASK_LVTOFF_HI) >> 20;
555 		goto set_offset;
556 	}
557 
558 	/* Gather LVT offset for thresholding: */
559 	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
560 		goto out;
561 
562 	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
563 
564 set_offset:
565 	offset = setup_APIC_mce_threshold(offset, new);
566 	if (offset == new)
567 		thresholding_irq_en = true;
568 
569 done:
570 	mce_threshold_block_init(&b, offset);
571 
572 out:
573 	return offset;
574 }
575 
576 bool amd_filter_mce(struct mce *m)
577 {
578 	enum smca_bank_types bank_type = smca_get_bank_type(m->bank);
579 	struct cpuinfo_x86 *c = &boot_cpu_data;
580 
581 	/* See Family 17h Models 10h-2Fh Erratum #1114. */
582 	if (c->x86 == 0x17 &&
583 	    c->x86_model >= 0x10 && c->x86_model <= 0x2F &&
584 	    bank_type == SMCA_IF && XEC(m->status, 0x3f) == 10)
585 		return true;
586 
587 	/* NB GART TLB error reporting is disabled by default. */
588 	if (c->x86 < 0x17) {
589 		if (m->bank == 4 && XEC(m->status, 0x1f) == 0x5)
590 			return true;
591 	}
592 
593 	return false;
594 }
595 
596 /*
597  * Turn off thresholding banks for the following conditions:
598  * - MC4_MISC thresholding is not supported on Family 0x15.
599  * - Prevent possible spurious interrupts from the IF bank on Family 0x17
600  *   Models 0x10-0x2F due to Erratum #1114.
601  */
602 static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
603 {
604 	int i, num_msrs;
605 	u64 hwcr;
606 	bool need_toggle;
607 	u32 msrs[NR_BLOCKS];
608 
609 	if (c->x86 == 0x15 && bank == 4) {
610 		msrs[0] = 0x00000413; /* MC4_MISC0 */
611 		msrs[1] = 0xc0000408; /* MC4_MISC1 */
612 		num_msrs = 2;
613 	} else if (c->x86 == 0x17 &&
614 		   (c->x86_model >= 0x10 && c->x86_model <= 0x2F)) {
615 
616 		if (smca_get_bank_type(bank) != SMCA_IF)
617 			return;
618 
619 		msrs[0] = MSR_AMD64_SMCA_MCx_MISC(bank);
620 		num_msrs = 1;
621 	} else {
622 		return;
623 	}
624 
625 	rdmsrl(MSR_K7_HWCR, hwcr);
626 
627 	/* McStatusWrEn has to be set */
628 	need_toggle = !(hwcr & BIT(18));
629 	if (need_toggle)
630 		wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
631 
632 	/* Clear CntP bit safely */
633 	for (i = 0; i < num_msrs; i++)
634 		msr_clear_bit(msrs[i], 62);
635 
636 	/* restore old settings */
637 	if (need_toggle)
638 		wrmsrl(MSR_K7_HWCR, hwcr);
639 }
640 
641 /* cpu init entry point, called from mce.c with preempt off */
642 void mce_amd_feature_init(struct cpuinfo_x86 *c)
643 {
644 	unsigned int bank, block, cpu = smp_processor_id();
645 	u32 low = 0, high = 0, address = 0;
646 	int offset = -1;
647 
648 
649 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
650 		if (mce_flags.smca)
651 			smca_configure(bank, cpu);
652 
653 		disable_err_thresholding(c, bank);
654 
655 		for (block = 0; block < NR_BLOCKS; ++block) {
656 			address = get_block_address(address, low, high, bank, block, cpu);
657 			if (!address)
658 				break;
659 
660 			if (rdmsr_safe(address, &low, &high))
661 				break;
662 
663 			if (!(high & MASK_VALID_HI))
664 				continue;
665 
666 			if (!(high & MASK_CNTP_HI)  ||
667 			     (high & MASK_LOCKED_HI))
668 				continue;
669 
670 			offset = prepare_threshold_block(bank, block, address, offset, high);
671 		}
672 	}
673 
674 	if (mce_flags.succor)
675 		deferred_error_interrupt_enable(c);
676 }
677 
678 int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
679 {
680 	u64 dram_base_addr, dram_limit_addr, dram_hole_base;
681 	/* We start from the normalized address */
682 	u64 ret_addr = norm_addr;
683 
684 	u32 tmp;
685 
686 	u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
687 	u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
688 	u8 intlv_addr_sel, intlv_addr_bit;
689 	u8 num_intlv_bits, hashed_bit;
690 	u8 lgcy_mmio_hole_en, base = 0;
691 	u8 cs_mask, cs_id = 0;
692 	bool hash_enabled = false;
693 
694 	/* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
695 	if (amd_df_indirect_read(nid, 0, 0x1B4, umc, &tmp))
696 		goto out_err;
697 
698 	/* Remove HiAddrOffset from normalized address, if enabled: */
699 	if (tmp & BIT(0)) {
700 		u64 hi_addr_offset = (tmp & GENMASK_ULL(31, 20)) << 8;
701 
702 		if (norm_addr >= hi_addr_offset) {
703 			ret_addr -= hi_addr_offset;
704 			base = 1;
705 		}
706 	}
707 
708 	/* Read D18F0x110 (DramBaseAddress). */
709 	if (amd_df_indirect_read(nid, 0, 0x110 + (8 * base), umc, &tmp))
710 		goto out_err;
711 
712 	/* Check if address range is valid. */
713 	if (!(tmp & BIT(0))) {
714 		pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
715 			__func__, tmp);
716 		goto out_err;
717 	}
718 
719 	lgcy_mmio_hole_en = tmp & BIT(1);
720 	intlv_num_chan	  = (tmp >> 4) & 0xF;
721 	intlv_addr_sel	  = (tmp >> 8) & 0x7;
722 	dram_base_addr	  = (tmp & GENMASK_ULL(31, 12)) << 16;
723 
724 	/* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
725 	if (intlv_addr_sel > 3) {
726 		pr_err("%s: Invalid interleave address select %d.\n",
727 			__func__, intlv_addr_sel);
728 		goto out_err;
729 	}
730 
731 	/* Read D18F0x114 (DramLimitAddress). */
732 	if (amd_df_indirect_read(nid, 0, 0x114 + (8 * base), umc, &tmp))
733 		goto out_err;
734 
735 	intlv_num_sockets = (tmp >> 8) & 0x1;
736 	intlv_num_dies	  = (tmp >> 10) & 0x3;
737 	dram_limit_addr	  = ((tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
738 
739 	intlv_addr_bit = intlv_addr_sel + 8;
740 
741 	/* Re-use intlv_num_chan by setting it equal to log2(#channels) */
742 	switch (intlv_num_chan) {
743 	case 0:	intlv_num_chan = 0; break;
744 	case 1: intlv_num_chan = 1; break;
745 	case 3: intlv_num_chan = 2; break;
746 	case 5:	intlv_num_chan = 3; break;
747 	case 7:	intlv_num_chan = 4; break;
748 
749 	case 8: intlv_num_chan = 1;
750 		hash_enabled = true;
751 		break;
752 	default:
753 		pr_err("%s: Invalid number of interleaved channels %d.\n",
754 			__func__, intlv_num_chan);
755 		goto out_err;
756 	}
757 
758 	num_intlv_bits = intlv_num_chan;
759 
760 	if (intlv_num_dies > 2) {
761 		pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
762 			__func__, intlv_num_dies);
763 		goto out_err;
764 	}
765 
766 	num_intlv_bits += intlv_num_dies;
767 
768 	/* Add a bit if sockets are interleaved. */
769 	num_intlv_bits += intlv_num_sockets;
770 
771 	/* Assert num_intlv_bits <= 4 */
772 	if (num_intlv_bits > 4) {
773 		pr_err("%s: Invalid interleave bits %d.\n",
774 			__func__, num_intlv_bits);
775 		goto out_err;
776 	}
777 
778 	if (num_intlv_bits > 0) {
779 		u64 temp_addr_x, temp_addr_i, temp_addr_y;
780 		u8 die_id_bit, sock_id_bit, cs_fabric_id;
781 
782 		/*
783 		 * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
784 		 * This is the fabric id for this coherent slave. Use
785 		 * umc/channel# as instance id of the coherent slave
786 		 * for FICAA.
787 		 */
788 		if (amd_df_indirect_read(nid, 0, 0x50, umc, &tmp))
789 			goto out_err;
790 
791 		cs_fabric_id = (tmp >> 8) & 0xFF;
792 		die_id_bit   = 0;
793 
794 		/* If interleaved over more than 1 channel: */
795 		if (intlv_num_chan) {
796 			die_id_bit = intlv_num_chan;
797 			cs_mask	   = (1 << die_id_bit) - 1;
798 			cs_id	   = cs_fabric_id & cs_mask;
799 		}
800 
801 		sock_id_bit = die_id_bit;
802 
803 		/* Read D18F1x208 (SystemFabricIdMask). */
804 		if (intlv_num_dies || intlv_num_sockets)
805 			if (amd_df_indirect_read(nid, 1, 0x208, umc, &tmp))
806 				goto out_err;
807 
808 		/* If interleaved over more than 1 die. */
809 		if (intlv_num_dies) {
810 			sock_id_bit  = die_id_bit + intlv_num_dies;
811 			die_id_shift = (tmp >> 24) & 0xF;
812 			die_id_mask  = (tmp >> 8) & 0xFF;
813 
814 			cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
815 		}
816 
817 		/* If interleaved over more than 1 socket. */
818 		if (intlv_num_sockets) {
819 			socket_id_shift	= (tmp >> 28) & 0xF;
820 			socket_id_mask	= (tmp >> 16) & 0xFF;
821 
822 			cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
823 		}
824 
825 		/*
826 		 * The pre-interleaved address consists of XXXXXXIIIYYYYY
827 		 * where III is the ID for this CS, and XXXXXXYYYYY are the
828 		 * address bits from the post-interleaved address.
829 		 * "num_intlv_bits" has been calculated to tell us how many "I"
830 		 * bits there are. "intlv_addr_bit" tells us how many "Y" bits
831 		 * there are (where "I" starts).
832 		 */
833 		temp_addr_y = ret_addr & GENMASK_ULL(intlv_addr_bit-1, 0);
834 		temp_addr_i = (cs_id << intlv_addr_bit);
835 		temp_addr_x = (ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
836 		ret_addr    = temp_addr_x | temp_addr_i | temp_addr_y;
837 	}
838 
839 	/* Add dram base address */
840 	ret_addr += dram_base_addr;
841 
842 	/* If legacy MMIO hole enabled */
843 	if (lgcy_mmio_hole_en) {
844 		if (amd_df_indirect_read(nid, 0, 0x104, umc, &tmp))
845 			goto out_err;
846 
847 		dram_hole_base = tmp & GENMASK(31, 24);
848 		if (ret_addr >= dram_hole_base)
849 			ret_addr += (BIT_ULL(32) - dram_hole_base);
850 	}
851 
852 	if (hash_enabled) {
853 		/* Save some parentheses and grab ls-bit at the end. */
854 		hashed_bit =	(ret_addr >> 12) ^
855 				(ret_addr >> 18) ^
856 				(ret_addr >> 21) ^
857 				(ret_addr >> 30) ^
858 				cs_id;
859 
860 		hashed_bit &= BIT(0);
861 
862 		if (hashed_bit != ((ret_addr >> intlv_addr_bit) & BIT(0)))
863 			ret_addr ^= BIT(intlv_addr_bit);
864 	}
865 
866 	/* Is calculated system address is above DRAM limit address? */
867 	if (ret_addr > dram_limit_addr)
868 		goto out_err;
869 
870 	*sys_addr = ret_addr;
871 	return 0;
872 
873 out_err:
874 	return -EINVAL;
875 }
876 EXPORT_SYMBOL_GPL(umc_normaddr_to_sysaddr);
877 
878 bool amd_mce_is_memory_error(struct mce *m)
879 {
880 	/* ErrCodeExt[20:16] */
881 	u8 xec = (m->status >> 16) & 0x1f;
882 
883 	if (mce_flags.smca)
884 		return smca_get_bank_type(m->bank) == SMCA_UMC && xec == 0x0;
885 
886 	return m->bank == 4 && xec == 0x8;
887 }
888 
889 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
890 {
891 	struct mce m;
892 
893 	mce_setup(&m);
894 
895 	m.status = status;
896 	m.misc   = misc;
897 	m.bank   = bank;
898 	m.tsc	 = rdtsc();
899 
900 	if (m.status & MCI_STATUS_ADDRV) {
901 		m.addr = addr;
902 
903 		/*
904 		 * Extract [55:<lsb>] where lsb is the least significant
905 		 * *valid* bit of the address bits.
906 		 */
907 		if (mce_flags.smca) {
908 			u8 lsb = (m.addr >> 56) & 0x3f;
909 
910 			m.addr &= GENMASK_ULL(55, lsb);
911 		}
912 	}
913 
914 	if (mce_flags.smca) {
915 		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
916 
917 		if (m.status & MCI_STATUS_SYNDV)
918 			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
919 	}
920 
921 	mce_log(&m);
922 }
923 
924 DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
925 {
926 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
927 	inc_irq_stat(irq_deferred_error_count);
928 	deferred_error_int_vector();
929 	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
930 	ack_APIC_irq();
931 }
932 
933 /*
934  * Returns true if the logged error is deferred. False, otherwise.
935  */
936 static inline bool
937 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
938 {
939 	u64 status, addr = 0;
940 
941 	rdmsrl(msr_stat, status);
942 	if (!(status & MCI_STATUS_VAL))
943 		return false;
944 
945 	if (status & MCI_STATUS_ADDRV)
946 		rdmsrl(msr_addr, addr);
947 
948 	__log_error(bank, status, addr, misc);
949 
950 	wrmsrl(msr_stat, 0);
951 
952 	return status & MCI_STATUS_DEFERRED;
953 }
954 
955 /*
956  * We have three scenarios for checking for Deferred errors:
957  *
958  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
959  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
960  *    clear MCA_DESTAT.
961  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
962  *    log it.
963  */
964 static void log_error_deferred(unsigned int bank)
965 {
966 	bool defrd;
967 
968 	defrd = _log_error_bank(bank, msr_ops.status(bank),
969 					msr_ops.addr(bank), 0);
970 
971 	if (!mce_flags.smca)
972 		return;
973 
974 	/* Clear MCA_DESTAT if we logged the deferred error from MCA_STATUS. */
975 	if (defrd) {
976 		wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
977 		return;
978 	}
979 
980 	/*
981 	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
982 	 * for a valid error.
983 	 */
984 	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
985 			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
986 }
987 
988 /* APIC interrupt handler for deferred errors */
989 static void amd_deferred_error_interrupt(void)
990 {
991 	unsigned int bank;
992 
993 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank)
994 		log_error_deferred(bank);
995 }
996 
997 static void log_error_thresholding(unsigned int bank, u64 misc)
998 {
999 	_log_error_bank(bank, msr_ops.status(bank), msr_ops.addr(bank), misc);
1000 }
1001 
1002 static void log_and_reset_block(struct threshold_block *block)
1003 {
1004 	struct thresh_restart tr;
1005 	u32 low = 0, high = 0;
1006 
1007 	if (!block)
1008 		return;
1009 
1010 	if (rdmsr_safe(block->address, &low, &high))
1011 		return;
1012 
1013 	if (!(high & MASK_OVERFLOW_HI))
1014 		return;
1015 
1016 	/* Log the MCE which caused the threshold event. */
1017 	log_error_thresholding(block->bank, ((u64)high << 32) | low);
1018 
1019 	/* Reset threshold block after logging error. */
1020 	memset(&tr, 0, sizeof(tr));
1021 	tr.b = block;
1022 	threshold_restart_bank(&tr);
1023 }
1024 
1025 /*
1026  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
1027  * goes off when error_count reaches threshold_limit.
1028  */
1029 static void amd_threshold_interrupt(void)
1030 {
1031 	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
1032 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
1033 	unsigned int bank, cpu = smp_processor_id();
1034 
1035 	/*
1036 	 * Validate that the threshold bank has been initialized already. The
1037 	 * handler is installed at boot time, but on a hotplug event the
1038 	 * interrupt might fire before the data has been initialized.
1039 	 */
1040 	if (!bp)
1041 		return;
1042 
1043 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
1044 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1045 			continue;
1046 
1047 		first_block = bp[bank]->blocks;
1048 		if (!first_block)
1049 			continue;
1050 
1051 		/*
1052 		 * The first block is also the head of the list. Check it first
1053 		 * before iterating over the rest.
1054 		 */
1055 		log_and_reset_block(first_block);
1056 		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
1057 			log_and_reset_block(block);
1058 	}
1059 }
1060 
1061 /*
1062  * Sysfs Interface
1063  */
1064 
1065 struct threshold_attr {
1066 	struct attribute attr;
1067 	ssize_t (*show) (struct threshold_block *, char *);
1068 	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
1069 };
1070 
1071 #define SHOW_FIELDS(name)						\
1072 static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
1073 {									\
1074 	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
1075 }
1076 SHOW_FIELDS(interrupt_enable)
1077 SHOW_FIELDS(threshold_limit)
1078 
1079 static ssize_t
1080 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
1081 {
1082 	struct thresh_restart tr;
1083 	unsigned long new;
1084 
1085 	if (!b->interrupt_capable)
1086 		return -EINVAL;
1087 
1088 	if (kstrtoul(buf, 0, &new) < 0)
1089 		return -EINVAL;
1090 
1091 	b->interrupt_enable = !!new;
1092 
1093 	memset(&tr, 0, sizeof(tr));
1094 	tr.b		= b;
1095 
1096 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
1097 		return -ENODEV;
1098 
1099 	return size;
1100 }
1101 
1102 static ssize_t
1103 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
1104 {
1105 	struct thresh_restart tr;
1106 	unsigned long new;
1107 
1108 	if (kstrtoul(buf, 0, &new) < 0)
1109 		return -EINVAL;
1110 
1111 	if (new > THRESHOLD_MAX)
1112 		new = THRESHOLD_MAX;
1113 	if (new < 1)
1114 		new = 1;
1115 
1116 	memset(&tr, 0, sizeof(tr));
1117 	tr.old_limit = b->threshold_limit;
1118 	b->threshold_limit = new;
1119 	tr.b = b;
1120 
1121 	if (smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1))
1122 		return -ENODEV;
1123 
1124 	return size;
1125 }
1126 
1127 static ssize_t show_error_count(struct threshold_block *b, char *buf)
1128 {
1129 	u32 lo, hi;
1130 
1131 	/* CPU might be offline by now */
1132 	if (rdmsr_on_cpu(b->cpu, b->address, &lo, &hi))
1133 		return -ENODEV;
1134 
1135 	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
1136 				     (THRESHOLD_MAX - b->threshold_limit)));
1137 }
1138 
1139 static struct threshold_attr error_count = {
1140 	.attr = {.name = __stringify(error_count), .mode = 0444 },
1141 	.show = show_error_count,
1142 };
1143 
1144 #define RW_ATTR(val)							\
1145 static struct threshold_attr val = {					\
1146 	.attr	= {.name = __stringify(val), .mode = 0644 },		\
1147 	.show	= show_## val,						\
1148 	.store	= store_## val,						\
1149 };
1150 
1151 RW_ATTR(interrupt_enable);
1152 RW_ATTR(threshold_limit);
1153 
1154 static struct attribute *default_attrs[] = {
1155 	&threshold_limit.attr,
1156 	&error_count.attr,
1157 	NULL,	/* possibly interrupt_enable if supported, see below */
1158 	NULL,
1159 };
1160 
1161 #define to_block(k)	container_of(k, struct threshold_block, kobj)
1162 #define to_attr(a)	container_of(a, struct threshold_attr, attr)
1163 
1164 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1165 {
1166 	struct threshold_block *b = to_block(kobj);
1167 	struct threshold_attr *a = to_attr(attr);
1168 	ssize_t ret;
1169 
1170 	ret = a->show ? a->show(b, buf) : -EIO;
1171 
1172 	return ret;
1173 }
1174 
1175 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1176 		     const char *buf, size_t count)
1177 {
1178 	struct threshold_block *b = to_block(kobj);
1179 	struct threshold_attr *a = to_attr(attr);
1180 	ssize_t ret;
1181 
1182 	ret = a->store ? a->store(b, buf, count) : -EIO;
1183 
1184 	return ret;
1185 }
1186 
1187 static const struct sysfs_ops threshold_ops = {
1188 	.show			= show,
1189 	.store			= store,
1190 };
1191 
1192 static void threshold_block_release(struct kobject *kobj);
1193 
1194 static struct kobj_type threshold_ktype = {
1195 	.sysfs_ops		= &threshold_ops,
1196 	.default_attrs		= default_attrs,
1197 	.release		= threshold_block_release,
1198 };
1199 
1200 static const char *get_name(unsigned int bank, struct threshold_block *b)
1201 {
1202 	enum smca_bank_types bank_type;
1203 
1204 	if (!mce_flags.smca) {
1205 		if (b && bank == 4)
1206 			return bank4_names(b);
1207 
1208 		return th_names[bank];
1209 	}
1210 
1211 	bank_type = smca_get_bank_type(bank);
1212 	if (bank_type >= N_SMCA_BANK_TYPES)
1213 		return NULL;
1214 
1215 	if (b && bank_type == SMCA_UMC) {
1216 		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1217 			return smca_umc_block_names[b->block];
1218 		return NULL;
1219 	}
1220 
1221 	if (smca_banks[bank].hwid->count == 1)
1222 		return smca_get_name(bank_type);
1223 
1224 	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1225 		 "%s_%x", smca_get_name(bank_type),
1226 			  smca_banks[bank].sysfs_id);
1227 	return buf_mcatype;
1228 }
1229 
1230 static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb,
1231 				     unsigned int bank, unsigned int block,
1232 				     u32 address)
1233 {
1234 	struct threshold_block *b = NULL;
1235 	u32 low, high;
1236 	int err;
1237 
1238 	if ((bank >= this_cpu_read(mce_num_banks)) || (block >= NR_BLOCKS))
1239 		return 0;
1240 
1241 	if (rdmsr_safe(address, &low, &high))
1242 		return 0;
1243 
1244 	if (!(high & MASK_VALID_HI)) {
1245 		if (block)
1246 			goto recurse;
1247 		else
1248 			return 0;
1249 	}
1250 
1251 	if (!(high & MASK_CNTP_HI)  ||
1252 	     (high & MASK_LOCKED_HI))
1253 		goto recurse;
1254 
1255 	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1256 	if (!b)
1257 		return -ENOMEM;
1258 
1259 	b->block		= block;
1260 	b->bank			= bank;
1261 	b->cpu			= cpu;
1262 	b->address		= address;
1263 	b->interrupt_enable	= 0;
1264 	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1265 	b->threshold_limit	= THRESHOLD_MAX;
1266 
1267 	if (b->interrupt_capable) {
1268 		threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
1269 		b->interrupt_enable = 1;
1270 	} else {
1271 		threshold_ktype.default_attrs[2] = NULL;
1272 	}
1273 
1274 	INIT_LIST_HEAD(&b->miscj);
1275 
1276 	/* This is safe as @tb is not visible yet */
1277 	if (tb->blocks)
1278 		list_add(&b->miscj, &tb->blocks->miscj);
1279 	else
1280 		tb->blocks = b;
1281 
1282 	err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(bank, b));
1283 	if (err)
1284 		goto out_free;
1285 recurse:
1286 	address = get_block_address(address, low, high, bank, ++block, cpu);
1287 	if (!address)
1288 		return 0;
1289 
1290 	err = allocate_threshold_blocks(cpu, tb, bank, block, address);
1291 	if (err)
1292 		goto out_free;
1293 
1294 	if (b)
1295 		kobject_uevent(&b->kobj, KOBJ_ADD);
1296 
1297 	return 0;
1298 
1299 out_free:
1300 	if (b) {
1301 		list_del(&b->miscj);
1302 		kobject_put(&b->kobj);
1303 	}
1304 	return err;
1305 }
1306 
1307 static int __threshold_add_blocks(struct threshold_bank *b)
1308 {
1309 	struct list_head *head = &b->blocks->miscj;
1310 	struct threshold_block *pos = NULL;
1311 	struct threshold_block *tmp = NULL;
1312 	int err = 0;
1313 
1314 	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1315 	if (err)
1316 		return err;
1317 
1318 	list_for_each_entry_safe(pos, tmp, head, miscj) {
1319 
1320 		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1321 		if (err) {
1322 			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1323 				kobject_del(&pos->kobj);
1324 
1325 			return err;
1326 		}
1327 	}
1328 	return err;
1329 }
1330 
1331 static int threshold_create_bank(struct threshold_bank **bp, unsigned int cpu,
1332 				 unsigned int bank)
1333 {
1334 	struct device *dev = this_cpu_read(mce_device);
1335 	struct amd_northbridge *nb = NULL;
1336 	struct threshold_bank *b = NULL;
1337 	const char *name = get_name(bank, NULL);
1338 	int err = 0;
1339 
1340 	if (!dev)
1341 		return -ENODEV;
1342 
1343 	if (is_shared_bank(bank)) {
1344 		nb = node_to_amd_nb(topology_die_id(cpu));
1345 
1346 		/* threshold descriptor already initialized on this node? */
1347 		if (nb && nb->bank4) {
1348 			/* yes, use it */
1349 			b = nb->bank4;
1350 			err = kobject_add(b->kobj, &dev->kobj, name);
1351 			if (err)
1352 				goto out;
1353 
1354 			bp[bank] = b;
1355 			refcount_inc(&b->cpus);
1356 
1357 			err = __threshold_add_blocks(b);
1358 
1359 			goto out;
1360 		}
1361 	}
1362 
1363 	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1364 	if (!b) {
1365 		err = -ENOMEM;
1366 		goto out;
1367 	}
1368 
1369 	/* Associate the bank with the per-CPU MCE device */
1370 	b->kobj = kobject_create_and_add(name, &dev->kobj);
1371 	if (!b->kobj) {
1372 		err = -EINVAL;
1373 		goto out_free;
1374 	}
1375 
1376 	if (is_shared_bank(bank)) {
1377 		b->shared = 1;
1378 		refcount_set(&b->cpus, 1);
1379 
1380 		/* nb is already initialized, see above */
1381 		if (nb) {
1382 			WARN_ON(nb->bank4);
1383 			nb->bank4 = b;
1384 		}
1385 	}
1386 
1387 	err = allocate_threshold_blocks(cpu, b, bank, 0, msr_ops.misc(bank));
1388 	if (err)
1389 		goto out_kobj;
1390 
1391 	bp[bank] = b;
1392 	return 0;
1393 
1394 out_kobj:
1395 	kobject_put(b->kobj);
1396 out_free:
1397 	kfree(b);
1398 out:
1399 	return err;
1400 }
1401 
1402 static void threshold_block_release(struct kobject *kobj)
1403 {
1404 	kfree(to_block(kobj));
1405 }
1406 
1407 static void deallocate_threshold_blocks(struct threshold_bank *bank)
1408 {
1409 	struct threshold_block *pos, *tmp;
1410 
1411 	list_for_each_entry_safe(pos, tmp, &bank->blocks->miscj, miscj) {
1412 		list_del(&pos->miscj);
1413 		kobject_put(&pos->kobj);
1414 	}
1415 
1416 	kobject_put(&bank->blocks->kobj);
1417 }
1418 
1419 static void __threshold_remove_blocks(struct threshold_bank *b)
1420 {
1421 	struct threshold_block *pos = NULL;
1422 	struct threshold_block *tmp = NULL;
1423 
1424 	kobject_del(b->kobj);
1425 
1426 	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1427 		kobject_del(&pos->kobj);
1428 }
1429 
1430 static void threshold_remove_bank(struct threshold_bank *bank)
1431 {
1432 	struct amd_northbridge *nb;
1433 
1434 	if (!bank->blocks)
1435 		goto out_free;
1436 
1437 	if (!bank->shared)
1438 		goto out_dealloc;
1439 
1440 	if (!refcount_dec_and_test(&bank->cpus)) {
1441 		__threshold_remove_blocks(bank);
1442 		return;
1443 	} else {
1444 		/*
1445 		 * The last CPU on this node using the shared bank is going
1446 		 * away, remove that bank now.
1447 		 */
1448 		nb = node_to_amd_nb(topology_die_id(smp_processor_id()));
1449 		nb->bank4 = NULL;
1450 	}
1451 
1452 out_dealloc:
1453 	deallocate_threshold_blocks(bank);
1454 
1455 out_free:
1456 	kobject_put(bank->kobj);
1457 	kfree(bank);
1458 }
1459 
1460 int mce_threshold_remove_device(unsigned int cpu)
1461 {
1462 	struct threshold_bank **bp = this_cpu_read(threshold_banks);
1463 	unsigned int bank, numbanks = this_cpu_read(mce_num_banks);
1464 
1465 	if (!bp)
1466 		return 0;
1467 
1468 	/*
1469 	 * Clear the pointer before cleaning up, so that the interrupt won't
1470 	 * touch anything of this.
1471 	 */
1472 	this_cpu_write(threshold_banks, NULL);
1473 
1474 	for (bank = 0; bank < numbanks; bank++) {
1475 		if (bp[bank]) {
1476 			threshold_remove_bank(bp[bank]);
1477 			bp[bank] = NULL;
1478 		}
1479 	}
1480 	kfree(bp);
1481 	return 0;
1482 }
1483 
1484 /**
1485  * mce_threshold_create_device - Create the per-CPU MCE threshold device
1486  * @cpu:	The plugged in CPU
1487  *
1488  * Create directories and files for all valid threshold banks.
1489  *
1490  * This is invoked from the CPU hotplug callback which was installed in
1491  * mcheck_init_device(). The invocation happens in context of the hotplug
1492  * thread running on @cpu.  The callback is invoked on all CPUs which are
1493  * online when the callback is installed or during a real hotplug event.
1494  */
1495 int mce_threshold_create_device(unsigned int cpu)
1496 {
1497 	unsigned int numbanks, bank;
1498 	struct threshold_bank **bp;
1499 	int err;
1500 
1501 	if (!mce_flags.amd_threshold)
1502 		return 0;
1503 
1504 	bp = this_cpu_read(threshold_banks);
1505 	if (bp)
1506 		return 0;
1507 
1508 	numbanks = this_cpu_read(mce_num_banks);
1509 	bp = kcalloc(numbanks, sizeof(*bp), GFP_KERNEL);
1510 	if (!bp)
1511 		return -ENOMEM;
1512 
1513 	for (bank = 0; bank < numbanks; ++bank) {
1514 		if (!(this_cpu_read(bank_map) & (1 << bank)))
1515 			continue;
1516 		err = threshold_create_bank(bp, cpu, bank);
1517 		if (err)
1518 			goto out_err;
1519 	}
1520 	this_cpu_write(threshold_banks, bp);
1521 
1522 	if (thresholding_irq_en)
1523 		mce_threshold_vector = amd_threshold_interrupt;
1524 	return 0;
1525 out_err:
1526 	mce_threshold_remove_device(cpu);
1527 	return err;
1528 }
1529