xref: /linux/arch/x86/kernel/cpu/mce/amd.c (revision 720909a7abd351535bfb485a0ecce03c2e4467e2)
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, xec_bitmap } */
136 
137 	/* Reserved type */
138 	{ SMCA_RESERVED, HWID_MCATYPE(0x00, 0x0), 0x0 },
139 
140 	/* ZN Core (HWID=0xB0) MCA types */
141 	{ SMCA_LS,	 HWID_MCATYPE(0xB0, 0x0), 0x1FFFFF },
142 	{ SMCA_LS_V2,	 HWID_MCATYPE(0xB0, 0x10), 0xFFFFFF },
143 	{ SMCA_IF,	 HWID_MCATYPE(0xB0, 0x1), 0x3FFF },
144 	{ SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2), 0xF },
145 	{ SMCA_DE,	 HWID_MCATYPE(0xB0, 0x3), 0x1FF },
146 	/* HWID 0xB0 MCATYPE 0x4 is Reserved */
147 	{ SMCA_EX,	 HWID_MCATYPE(0xB0, 0x5), 0xFFF },
148 	{ SMCA_FP,	 HWID_MCATYPE(0xB0, 0x6), 0x7F },
149 	{ SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7), 0xFF },
150 
151 	/* Data Fabric MCA types */
152 	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0), 0x1FF },
153 	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1), 0x1F },
154 	{ SMCA_CS_V2,	 HWID_MCATYPE(0x2E, 0x2), 0x3FFF },
155 
156 	/* Unified Memory Controller MCA type */
157 	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0), 0xFF },
158 
159 	/* Parameter Block MCA type */
160 	{ SMCA_PB,	 HWID_MCATYPE(0x05, 0x0), 0x1 },
161 
162 	/* Platform Security Processor MCA type */
163 	{ SMCA_PSP,	 HWID_MCATYPE(0xFF, 0x0), 0x1 },
164 	{ SMCA_PSP_V2,	 HWID_MCATYPE(0xFF, 0x1), 0x3FFFF },
165 
166 	/* System Management Unit MCA type */
167 	{ SMCA_SMU,	 HWID_MCATYPE(0x01, 0x0), 0x1 },
168 	{ SMCA_SMU_V2,	 HWID_MCATYPE(0x01, 0x1), 0x7FF },
169 
170 	/* Microprocessor 5 Unit MCA type */
171 	{ SMCA_MP5,	 HWID_MCATYPE(0x01, 0x2), 0x3FF },
172 
173 	/* Northbridge IO Unit MCA type */
174 	{ SMCA_NBIO,	 HWID_MCATYPE(0x18, 0x0), 0x1F },
175 
176 	/* PCI Express Unit MCA type */
177 	{ SMCA_PCIE,	 HWID_MCATYPE(0x46, 0x0), 0x1F },
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 static DEFINE_PER_CPU(unsigned int, bank_map);	/* see which banks are on */
196 
197 /* Map of banks that have more than MCA_MISC0 available. */
198 static DEFINE_PER_CPU(u32, smca_misc_banks_map);
199 
200 static void amd_threshold_interrupt(void);
201 static void amd_deferred_error_interrupt(void);
202 
203 static void default_deferred_error_interrupt(void)
204 {
205 	pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
206 }
207 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
208 
209 static void smca_set_misc_banks_map(unsigned int bank, unsigned int cpu)
210 {
211 	u32 low, high;
212 
213 	/*
214 	 * For SMCA enabled processors, BLKPTR field of the first MISC register
215 	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
216 	 */
217 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
218 		return;
219 
220 	if (!(low & MCI_CONFIG_MCAX))
221 		return;
222 
223 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high))
224 		return;
225 
226 	if (low & MASK_BLKPTR_LO)
227 		per_cpu(smca_misc_banks_map, cpu) |= BIT(bank);
228 
229 }
230 
231 static void smca_configure(unsigned int bank, unsigned int cpu)
232 {
233 	unsigned int i, hwid_mcatype;
234 	struct smca_hwid *s_hwid;
235 	u32 high, low;
236 	u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
237 
238 	/* Set appropriate bits in MCA_CONFIG */
239 	if (!rdmsr_safe(smca_config, &low, &high)) {
240 		/*
241 		 * OS is required to set the MCAX bit to acknowledge that it is
242 		 * now using the new MSR ranges and new registers under each
243 		 * bank. It also means that the OS will configure deferred
244 		 * errors in the new MCx_CONFIG register. If the bit is not set,
245 		 * uncorrectable errors will cause a system panic.
246 		 *
247 		 * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
248 		 */
249 		high |= BIT(0);
250 
251 		/*
252 		 * SMCA sets the Deferred Error Interrupt type per bank.
253 		 *
254 		 * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
255 		 * if the DeferredIntType bit field is available.
256 		 *
257 		 * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
258 		 * high portion of the MSR). OS should set this to 0x1 to enable
259 		 * APIC based interrupt. First, check that no interrupt has been
260 		 * set.
261 		 */
262 		if ((low & BIT(5)) && !((high >> 5) & 0x3))
263 			high |= BIT(5);
264 
265 		wrmsr(smca_config, low, high);
266 	}
267 
268 	smca_set_misc_banks_map(bank, cpu);
269 
270 	/* Return early if this bank was already initialized. */
271 	if (smca_banks[bank].hwid && smca_banks[bank].hwid->hwid_mcatype != 0)
272 		return;
273 
274 	if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
275 		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
276 		return;
277 	}
278 
279 	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
280 				    (high & MCI_IPID_MCATYPE) >> 16);
281 
282 	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
283 		s_hwid = &smca_hwid_mcatypes[i];
284 		if (hwid_mcatype == s_hwid->hwid_mcatype) {
285 			smca_banks[bank].hwid = s_hwid;
286 			smca_banks[bank].id = low;
287 			smca_banks[bank].sysfs_id = s_hwid->count++;
288 			break;
289 		}
290 	}
291 }
292 
293 struct thresh_restart {
294 	struct threshold_block	*b;
295 	int			reset;
296 	int			set_lvt_off;
297 	int			lvt_off;
298 	u16			old_limit;
299 };
300 
301 static inline bool is_shared_bank(int bank)
302 {
303 	/*
304 	 * Scalable MCA provides for only one core to have access to the MSRs of
305 	 * a shared bank.
306 	 */
307 	if (mce_flags.smca)
308 		return false;
309 
310 	/* Bank 4 is for northbridge reporting and is thus shared */
311 	return (bank == 4);
312 }
313 
314 static const char *bank4_names(const struct threshold_block *b)
315 {
316 	switch (b->address) {
317 	/* MSR4_MISC0 */
318 	case 0x00000413:
319 		return "dram";
320 
321 	case 0xc0000408:
322 		return "ht_links";
323 
324 	case 0xc0000409:
325 		return "l3_cache";
326 
327 	default:
328 		WARN(1, "Funny MSR: 0x%08x\n", b->address);
329 		return "";
330 	}
331 };
332 
333 
334 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
335 {
336 	/*
337 	 * bank 4 supports APIC LVT interrupts implicitly since forever.
338 	 */
339 	if (bank == 4)
340 		return true;
341 
342 	/*
343 	 * IntP: interrupt present; if this bit is set, the thresholding
344 	 * bank can generate APIC LVT interrupts
345 	 */
346 	return msr_high_bits & BIT(28);
347 }
348 
349 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
350 {
351 	int msr = (hi & MASK_LVTOFF_HI) >> 20;
352 
353 	if (apic < 0) {
354 		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
355 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
356 		       b->bank, b->block, b->address, hi, lo);
357 		return 0;
358 	}
359 
360 	if (apic != msr) {
361 		/*
362 		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
363 		 * the BIOS provides the value. The original field where LVT offset
364 		 * was set is reserved. Return early here:
365 		 */
366 		if (mce_flags.smca)
367 			return 0;
368 
369 		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
370 		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
371 		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
372 		return 0;
373 	}
374 
375 	return 1;
376 };
377 
378 /* Reprogram MCx_MISC MSR behind this threshold bank. */
379 static void threshold_restart_bank(void *_tr)
380 {
381 	struct thresh_restart *tr = _tr;
382 	u32 hi, lo;
383 
384 	rdmsr(tr->b->address, lo, hi);
385 
386 	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
387 		tr->reset = 1;	/* limit cannot be lower than err count */
388 
389 	if (tr->reset) {		/* reset err count and overflow bit */
390 		hi =
391 		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
392 		    (THRESHOLD_MAX - tr->b->threshold_limit);
393 	} else if (tr->old_limit) {	/* change limit w/o reset */
394 		int new_count = (hi & THRESHOLD_MAX) +
395 		    (tr->old_limit - tr->b->threshold_limit);
396 
397 		hi = (hi & ~MASK_ERR_COUNT_HI) |
398 		    (new_count & THRESHOLD_MAX);
399 	}
400 
401 	/* clear IntType */
402 	hi &= ~MASK_INT_TYPE_HI;
403 
404 	if (!tr->b->interrupt_capable)
405 		goto done;
406 
407 	if (tr->set_lvt_off) {
408 		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
409 			/* set new lvt offset */
410 			hi &= ~MASK_LVTOFF_HI;
411 			hi |= tr->lvt_off << 20;
412 		}
413 	}
414 
415 	if (tr->b->interrupt_enable)
416 		hi |= INT_TYPE_APIC;
417 
418  done:
419 
420 	hi |= MASK_COUNT_EN_HI;
421 	wrmsr(tr->b->address, lo, hi);
422 }
423 
424 static void mce_threshold_block_init(struct threshold_block *b, int offset)
425 {
426 	struct thresh_restart tr = {
427 		.b			= b,
428 		.set_lvt_off		= 1,
429 		.lvt_off		= offset,
430 	};
431 
432 	b->threshold_limit		= THRESHOLD_MAX;
433 	threshold_restart_bank(&tr);
434 };
435 
436 static int setup_APIC_mce_threshold(int reserved, int new)
437 {
438 	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
439 					      APIC_EILVT_MSG_FIX, 0))
440 		return new;
441 
442 	return reserved;
443 }
444 
445 static int setup_APIC_deferred_error(int reserved, int new)
446 {
447 	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
448 					      APIC_EILVT_MSG_FIX, 0))
449 		return new;
450 
451 	return reserved;
452 }
453 
454 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
455 {
456 	u32 low = 0, high = 0;
457 	int def_offset = -1, def_new;
458 
459 	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
460 		return;
461 
462 	def_new = (low & MASK_DEF_LVTOFF) >> 4;
463 	if (!(low & MASK_DEF_LVTOFF)) {
464 		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
465 		def_new = DEF_LVT_OFF;
466 		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
467 	}
468 
469 	def_offset = setup_APIC_deferred_error(def_offset, def_new);
470 	if ((def_offset == def_new) &&
471 	    (deferred_error_int_vector != amd_deferred_error_interrupt))
472 		deferred_error_int_vector = amd_deferred_error_interrupt;
473 
474 	if (!mce_flags.smca)
475 		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
476 
477 	wrmsr(MSR_CU_DEF_ERR, low, high);
478 }
479 
480 static u32 smca_get_block_address(unsigned int bank, unsigned int block,
481 				  unsigned int cpu)
482 {
483 	if (!block)
484 		return MSR_AMD64_SMCA_MCx_MISC(bank);
485 
486 	if (!(per_cpu(smca_misc_banks_map, cpu) & BIT(bank)))
487 		return 0;
488 
489 	return MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
490 }
491 
492 static u32 get_block_address(u32 current_addr, u32 low, u32 high,
493 			     unsigned int bank, unsigned int block,
494 			     unsigned int cpu)
495 {
496 	u32 addr = 0, offset = 0;
497 
498 	if ((bank >= per_cpu(mce_num_banks, cpu)) || (block >= NR_BLOCKS))
499 		return addr;
500 
501 	if (mce_flags.smca)
502 		return smca_get_block_address(bank, block, cpu);
503 
504 	/* Fall back to method we used for older processors: */
505 	switch (block) {
506 	case 0:
507 		addr = msr_ops.misc(bank);
508 		break;
509 	case 1:
510 		offset = ((low & MASK_BLKPTR_LO) >> 21);
511 		if (offset)
512 			addr = MCG_XBLK_ADDR + offset;
513 		break;
514 	default:
515 		addr = ++current_addr;
516 	}
517 	return addr;
518 }
519 
520 static int
521 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
522 			int offset, u32 misc_high)
523 {
524 	unsigned int cpu = smp_processor_id();
525 	u32 smca_low, smca_high;
526 	struct threshold_block b;
527 	int new;
528 
529 	if (!block)
530 		per_cpu(bank_map, cpu) |= (1 << bank);
531 
532 	memset(&b, 0, sizeof(b));
533 	b.cpu			= cpu;
534 	b.bank			= bank;
535 	b.block			= block;
536 	b.address		= addr;
537 	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
538 
539 	if (!b.interrupt_capable)
540 		goto done;
541 
542 	b.interrupt_enable = 1;
543 
544 	if (!mce_flags.smca) {
545 		new = (misc_high & MASK_LVTOFF_HI) >> 20;
546 		goto set_offset;
547 	}
548 
549 	/* Gather LVT offset for thresholding: */
550 	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
551 		goto out;
552 
553 	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
554 
555 set_offset:
556 	offset = setup_APIC_mce_threshold(offset, new);
557 	if (offset == new)
558 		thresholding_irq_en = true;
559 
560 done:
561 	mce_threshold_block_init(&b, offset);
562 
563 out:
564 	return offset;
565 }
566 
567 bool amd_filter_mce(struct mce *m)
568 {
569 	enum smca_bank_types bank_type = smca_get_bank_type(m->bank);
570 	struct cpuinfo_x86 *c = &boot_cpu_data;
571 	u8 xec = (m->status >> 16) & 0x3F;
572 
573 	/* See Family 17h Models 10h-2Fh Erratum #1114. */
574 	if (c->x86 == 0x17 &&
575 	    c->x86_model >= 0x10 && c->x86_model <= 0x2F &&
576 	    bank_type == SMCA_IF && xec == 10)
577 		return true;
578 
579 	return false;
580 }
581 
582 /*
583  * Turn off thresholding banks for the following conditions:
584  * - MC4_MISC thresholding is not supported on Family 0x15.
585  * - Prevent possible spurious interrupts from the IF bank on Family 0x17
586  *   Models 0x10-0x2F due to Erratum #1114.
587  */
588 static void disable_err_thresholding(struct cpuinfo_x86 *c, unsigned int bank)
589 {
590 	int i, num_msrs;
591 	u64 hwcr;
592 	bool need_toggle;
593 	u32 msrs[NR_BLOCKS];
594 
595 	if (c->x86 == 0x15 && bank == 4) {
596 		msrs[0] = 0x00000413; /* MC4_MISC0 */
597 		msrs[1] = 0xc0000408; /* MC4_MISC1 */
598 		num_msrs = 2;
599 	} else if (c->x86 == 0x17 &&
600 		   (c->x86_model >= 0x10 && c->x86_model <= 0x2F)) {
601 
602 		if (smca_get_bank_type(bank) != SMCA_IF)
603 			return;
604 
605 		msrs[0] = MSR_AMD64_SMCA_MCx_MISC(bank);
606 		num_msrs = 1;
607 	} else {
608 		return;
609 	}
610 
611 	rdmsrl(MSR_K7_HWCR, hwcr);
612 
613 	/* McStatusWrEn has to be set */
614 	need_toggle = !(hwcr & BIT(18));
615 	if (need_toggle)
616 		wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
617 
618 	/* Clear CntP bit safely */
619 	for (i = 0; i < num_msrs; i++)
620 		msr_clear_bit(msrs[i], 62);
621 
622 	/* restore old settings */
623 	if (need_toggle)
624 		wrmsrl(MSR_K7_HWCR, hwcr);
625 }
626 
627 /* cpu init entry point, called from mce.c with preempt off */
628 void mce_amd_feature_init(struct cpuinfo_x86 *c)
629 {
630 	unsigned int bank, block, cpu = smp_processor_id();
631 	u32 low = 0, high = 0, address = 0;
632 	int offset = -1;
633 
634 
635 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
636 		if (mce_flags.smca)
637 			smca_configure(bank, cpu);
638 
639 		disable_err_thresholding(c, bank);
640 
641 		for (block = 0; block < NR_BLOCKS; ++block) {
642 			address = get_block_address(address, low, high, bank, block, cpu);
643 			if (!address)
644 				break;
645 
646 			if (rdmsr_safe(address, &low, &high))
647 				break;
648 
649 			if (!(high & MASK_VALID_HI))
650 				continue;
651 
652 			if (!(high & MASK_CNTP_HI)  ||
653 			     (high & MASK_LOCKED_HI))
654 				continue;
655 
656 			offset = prepare_threshold_block(bank, block, address, offset, high);
657 		}
658 	}
659 
660 	if (mce_flags.succor)
661 		deferred_error_interrupt_enable(c);
662 }
663 
664 int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
665 {
666 	u64 dram_base_addr, dram_limit_addr, dram_hole_base;
667 	/* We start from the normalized address */
668 	u64 ret_addr = norm_addr;
669 
670 	u32 tmp;
671 
672 	u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
673 	u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
674 	u8 intlv_addr_sel, intlv_addr_bit;
675 	u8 num_intlv_bits, hashed_bit;
676 	u8 lgcy_mmio_hole_en, base = 0;
677 	u8 cs_mask, cs_id = 0;
678 	bool hash_enabled = false;
679 
680 	/* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
681 	if (amd_df_indirect_read(nid, 0, 0x1B4, umc, &tmp))
682 		goto out_err;
683 
684 	/* Remove HiAddrOffset from normalized address, if enabled: */
685 	if (tmp & BIT(0)) {
686 		u64 hi_addr_offset = (tmp & GENMASK_ULL(31, 20)) << 8;
687 
688 		if (norm_addr >= hi_addr_offset) {
689 			ret_addr -= hi_addr_offset;
690 			base = 1;
691 		}
692 	}
693 
694 	/* Read D18F0x110 (DramBaseAddress). */
695 	if (amd_df_indirect_read(nid, 0, 0x110 + (8 * base), umc, &tmp))
696 		goto out_err;
697 
698 	/* Check if address range is valid. */
699 	if (!(tmp & BIT(0))) {
700 		pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
701 			__func__, tmp);
702 		goto out_err;
703 	}
704 
705 	lgcy_mmio_hole_en = tmp & BIT(1);
706 	intlv_num_chan	  = (tmp >> 4) & 0xF;
707 	intlv_addr_sel	  = (tmp >> 8) & 0x7;
708 	dram_base_addr	  = (tmp & GENMASK_ULL(31, 12)) << 16;
709 
710 	/* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
711 	if (intlv_addr_sel > 3) {
712 		pr_err("%s: Invalid interleave address select %d.\n",
713 			__func__, intlv_addr_sel);
714 		goto out_err;
715 	}
716 
717 	/* Read D18F0x114 (DramLimitAddress). */
718 	if (amd_df_indirect_read(nid, 0, 0x114 + (8 * base), umc, &tmp))
719 		goto out_err;
720 
721 	intlv_num_sockets = (tmp >> 8) & 0x1;
722 	intlv_num_dies	  = (tmp >> 10) & 0x3;
723 	dram_limit_addr	  = ((tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
724 
725 	intlv_addr_bit = intlv_addr_sel + 8;
726 
727 	/* Re-use intlv_num_chan by setting it equal to log2(#channels) */
728 	switch (intlv_num_chan) {
729 	case 0:	intlv_num_chan = 0; break;
730 	case 1: intlv_num_chan = 1; break;
731 	case 3: intlv_num_chan = 2; break;
732 	case 5:	intlv_num_chan = 3; break;
733 	case 7:	intlv_num_chan = 4; break;
734 
735 	case 8: intlv_num_chan = 1;
736 		hash_enabled = true;
737 		break;
738 	default:
739 		pr_err("%s: Invalid number of interleaved channels %d.\n",
740 			__func__, intlv_num_chan);
741 		goto out_err;
742 	}
743 
744 	num_intlv_bits = intlv_num_chan;
745 
746 	if (intlv_num_dies > 2) {
747 		pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
748 			__func__, intlv_num_dies);
749 		goto out_err;
750 	}
751 
752 	num_intlv_bits += intlv_num_dies;
753 
754 	/* Add a bit if sockets are interleaved. */
755 	num_intlv_bits += intlv_num_sockets;
756 
757 	/* Assert num_intlv_bits <= 4 */
758 	if (num_intlv_bits > 4) {
759 		pr_err("%s: Invalid interleave bits %d.\n",
760 			__func__, num_intlv_bits);
761 		goto out_err;
762 	}
763 
764 	if (num_intlv_bits > 0) {
765 		u64 temp_addr_x, temp_addr_i, temp_addr_y;
766 		u8 die_id_bit, sock_id_bit, cs_fabric_id;
767 
768 		/*
769 		 * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
770 		 * This is the fabric id for this coherent slave. Use
771 		 * umc/channel# as instance id of the coherent slave
772 		 * for FICAA.
773 		 */
774 		if (amd_df_indirect_read(nid, 0, 0x50, umc, &tmp))
775 			goto out_err;
776 
777 		cs_fabric_id = (tmp >> 8) & 0xFF;
778 		die_id_bit   = 0;
779 
780 		/* If interleaved over more than 1 channel: */
781 		if (intlv_num_chan) {
782 			die_id_bit = intlv_num_chan;
783 			cs_mask	   = (1 << die_id_bit) - 1;
784 			cs_id	   = cs_fabric_id & cs_mask;
785 		}
786 
787 		sock_id_bit = die_id_bit;
788 
789 		/* Read D18F1x208 (SystemFabricIdMask). */
790 		if (intlv_num_dies || intlv_num_sockets)
791 			if (amd_df_indirect_read(nid, 1, 0x208, umc, &tmp))
792 				goto out_err;
793 
794 		/* If interleaved over more than 1 die. */
795 		if (intlv_num_dies) {
796 			sock_id_bit  = die_id_bit + intlv_num_dies;
797 			die_id_shift = (tmp >> 24) & 0xF;
798 			die_id_mask  = (tmp >> 8) & 0xFF;
799 
800 			cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
801 		}
802 
803 		/* If interleaved over more than 1 socket. */
804 		if (intlv_num_sockets) {
805 			socket_id_shift	= (tmp >> 28) & 0xF;
806 			socket_id_mask	= (tmp >> 16) & 0xFF;
807 
808 			cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
809 		}
810 
811 		/*
812 		 * The pre-interleaved address consists of XXXXXXIIIYYYYY
813 		 * where III is the ID for this CS, and XXXXXXYYYYY are the
814 		 * address bits from the post-interleaved address.
815 		 * "num_intlv_bits" has been calculated to tell us how many "I"
816 		 * bits there are. "intlv_addr_bit" tells us how many "Y" bits
817 		 * there are (where "I" starts).
818 		 */
819 		temp_addr_y = ret_addr & GENMASK_ULL(intlv_addr_bit-1, 0);
820 		temp_addr_i = (cs_id << intlv_addr_bit);
821 		temp_addr_x = (ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
822 		ret_addr    = temp_addr_x | temp_addr_i | temp_addr_y;
823 	}
824 
825 	/* Add dram base address */
826 	ret_addr += dram_base_addr;
827 
828 	/* If legacy MMIO hole enabled */
829 	if (lgcy_mmio_hole_en) {
830 		if (amd_df_indirect_read(nid, 0, 0x104, umc, &tmp))
831 			goto out_err;
832 
833 		dram_hole_base = tmp & GENMASK(31, 24);
834 		if (ret_addr >= dram_hole_base)
835 			ret_addr += (BIT_ULL(32) - dram_hole_base);
836 	}
837 
838 	if (hash_enabled) {
839 		/* Save some parentheses and grab ls-bit at the end. */
840 		hashed_bit =	(ret_addr >> 12) ^
841 				(ret_addr >> 18) ^
842 				(ret_addr >> 21) ^
843 				(ret_addr >> 30) ^
844 				cs_id;
845 
846 		hashed_bit &= BIT(0);
847 
848 		if (hashed_bit != ((ret_addr >> intlv_addr_bit) & BIT(0)))
849 			ret_addr ^= BIT(intlv_addr_bit);
850 	}
851 
852 	/* Is calculated system address is above DRAM limit address? */
853 	if (ret_addr > dram_limit_addr)
854 		goto out_err;
855 
856 	*sys_addr = ret_addr;
857 	return 0;
858 
859 out_err:
860 	return -EINVAL;
861 }
862 EXPORT_SYMBOL_GPL(umc_normaddr_to_sysaddr);
863 
864 bool amd_mce_is_memory_error(struct mce *m)
865 {
866 	/* ErrCodeExt[20:16] */
867 	u8 xec = (m->status >> 16) & 0x1f;
868 
869 	if (mce_flags.smca)
870 		return smca_get_bank_type(m->bank) == SMCA_UMC && xec == 0x0;
871 
872 	return m->bank == 4 && xec == 0x8;
873 }
874 
875 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
876 {
877 	struct mce m;
878 
879 	mce_setup(&m);
880 
881 	m.status = status;
882 	m.misc   = misc;
883 	m.bank   = bank;
884 	m.tsc	 = rdtsc();
885 
886 	if (m.status & MCI_STATUS_ADDRV) {
887 		m.addr = addr;
888 
889 		/*
890 		 * Extract [55:<lsb>] where lsb is the least significant
891 		 * *valid* bit of the address bits.
892 		 */
893 		if (mce_flags.smca) {
894 			u8 lsb = (m.addr >> 56) & 0x3f;
895 
896 			m.addr &= GENMASK_ULL(55, lsb);
897 		}
898 	}
899 
900 	if (mce_flags.smca) {
901 		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
902 
903 		if (m.status & MCI_STATUS_SYNDV)
904 			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
905 	}
906 
907 	mce_log(&m);
908 }
909 
910 DEFINE_IDTENTRY_SYSVEC(sysvec_deferred_error)
911 {
912 	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
913 	inc_irq_stat(irq_deferred_error_count);
914 	deferred_error_int_vector();
915 	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
916 	ack_APIC_irq();
917 }
918 
919 /*
920  * Returns true if the logged error is deferred. False, otherwise.
921  */
922 static inline bool
923 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
924 {
925 	u64 status, addr = 0;
926 
927 	rdmsrl(msr_stat, status);
928 	if (!(status & MCI_STATUS_VAL))
929 		return false;
930 
931 	if (status & MCI_STATUS_ADDRV)
932 		rdmsrl(msr_addr, addr);
933 
934 	__log_error(bank, status, addr, misc);
935 
936 	wrmsrl(msr_stat, 0);
937 
938 	return status & MCI_STATUS_DEFERRED;
939 }
940 
941 /*
942  * We have three scenarios for checking for Deferred errors:
943  *
944  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
945  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
946  *    clear MCA_DESTAT.
947  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
948  *    log it.
949  */
950 static void log_error_deferred(unsigned int bank)
951 {
952 	bool defrd;
953 
954 	defrd = _log_error_bank(bank, msr_ops.status(bank),
955 					msr_ops.addr(bank), 0);
956 
957 	if (!mce_flags.smca)
958 		return;
959 
960 	/* Clear MCA_DESTAT if we logged the deferred error from MCA_STATUS. */
961 	if (defrd) {
962 		wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
963 		return;
964 	}
965 
966 	/*
967 	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
968 	 * for a valid error.
969 	 */
970 	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
971 			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
972 }
973 
974 /* APIC interrupt handler for deferred errors */
975 static void amd_deferred_error_interrupt(void)
976 {
977 	unsigned int bank;
978 
979 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank)
980 		log_error_deferred(bank);
981 }
982 
983 static void log_error_thresholding(unsigned int bank, u64 misc)
984 {
985 	_log_error_bank(bank, msr_ops.status(bank), msr_ops.addr(bank), misc);
986 }
987 
988 static void log_and_reset_block(struct threshold_block *block)
989 {
990 	struct thresh_restart tr;
991 	u32 low = 0, high = 0;
992 
993 	if (!block)
994 		return;
995 
996 	if (rdmsr_safe(block->address, &low, &high))
997 		return;
998 
999 	if (!(high & MASK_OVERFLOW_HI))
1000 		return;
1001 
1002 	/* Log the MCE which caused the threshold event. */
1003 	log_error_thresholding(block->bank, ((u64)high << 32) | low);
1004 
1005 	/* Reset threshold block after logging error. */
1006 	memset(&tr, 0, sizeof(tr));
1007 	tr.b = block;
1008 	threshold_restart_bank(&tr);
1009 }
1010 
1011 /*
1012  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
1013  * goes off when error_count reaches threshold_limit.
1014  */
1015 static void amd_threshold_interrupt(void)
1016 {
1017 	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
1018 	unsigned int bank, cpu = smp_processor_id();
1019 
1020 	for (bank = 0; bank < this_cpu_read(mce_num_banks); ++bank) {
1021 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1022 			continue;
1023 
1024 		first_block = per_cpu(threshold_banks, cpu)[bank]->blocks;
1025 		if (!first_block)
1026 			continue;
1027 
1028 		/*
1029 		 * The first block is also the head of the list. Check it first
1030 		 * before iterating over the rest.
1031 		 */
1032 		log_and_reset_block(first_block);
1033 		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
1034 			log_and_reset_block(block);
1035 	}
1036 }
1037 
1038 /*
1039  * Sysfs Interface
1040  */
1041 
1042 struct threshold_attr {
1043 	struct attribute attr;
1044 	ssize_t (*show) (struct threshold_block *, char *);
1045 	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
1046 };
1047 
1048 #define SHOW_FIELDS(name)						\
1049 static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
1050 {									\
1051 	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
1052 }
1053 SHOW_FIELDS(interrupt_enable)
1054 SHOW_FIELDS(threshold_limit)
1055 
1056 static ssize_t
1057 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
1058 {
1059 	struct thresh_restart tr;
1060 	unsigned long new;
1061 
1062 	if (!b->interrupt_capable)
1063 		return -EINVAL;
1064 
1065 	if (kstrtoul(buf, 0, &new) < 0)
1066 		return -EINVAL;
1067 
1068 	b->interrupt_enable = !!new;
1069 
1070 	memset(&tr, 0, sizeof(tr));
1071 	tr.b		= b;
1072 
1073 	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1074 
1075 	return size;
1076 }
1077 
1078 static ssize_t
1079 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
1080 {
1081 	struct thresh_restart tr;
1082 	unsigned long new;
1083 
1084 	if (kstrtoul(buf, 0, &new) < 0)
1085 		return -EINVAL;
1086 
1087 	if (new > THRESHOLD_MAX)
1088 		new = THRESHOLD_MAX;
1089 	if (new < 1)
1090 		new = 1;
1091 
1092 	memset(&tr, 0, sizeof(tr));
1093 	tr.old_limit = b->threshold_limit;
1094 	b->threshold_limit = new;
1095 	tr.b = b;
1096 
1097 	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1098 
1099 	return size;
1100 }
1101 
1102 static ssize_t show_error_count(struct threshold_block *b, char *buf)
1103 {
1104 	u32 lo, hi;
1105 
1106 	rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
1107 
1108 	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
1109 				     (THRESHOLD_MAX - b->threshold_limit)));
1110 }
1111 
1112 static struct threshold_attr error_count = {
1113 	.attr = {.name = __stringify(error_count), .mode = 0444 },
1114 	.show = show_error_count,
1115 };
1116 
1117 #define RW_ATTR(val)							\
1118 static struct threshold_attr val = {					\
1119 	.attr	= {.name = __stringify(val), .mode = 0644 },		\
1120 	.show	= show_## val,						\
1121 	.store	= store_## val,						\
1122 };
1123 
1124 RW_ATTR(interrupt_enable);
1125 RW_ATTR(threshold_limit);
1126 
1127 static struct attribute *default_attrs[] = {
1128 	&threshold_limit.attr,
1129 	&error_count.attr,
1130 	NULL,	/* possibly interrupt_enable if supported, see below */
1131 	NULL,
1132 };
1133 
1134 #define to_block(k)	container_of(k, struct threshold_block, kobj)
1135 #define to_attr(a)	container_of(a, struct threshold_attr, attr)
1136 
1137 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1138 {
1139 	struct threshold_block *b = to_block(kobj);
1140 	struct threshold_attr *a = to_attr(attr);
1141 	ssize_t ret;
1142 
1143 	ret = a->show ? a->show(b, buf) : -EIO;
1144 
1145 	return ret;
1146 }
1147 
1148 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1149 		     const char *buf, size_t count)
1150 {
1151 	struct threshold_block *b = to_block(kobj);
1152 	struct threshold_attr *a = to_attr(attr);
1153 	ssize_t ret;
1154 
1155 	ret = a->store ? a->store(b, buf, count) : -EIO;
1156 
1157 	return ret;
1158 }
1159 
1160 static const struct sysfs_ops threshold_ops = {
1161 	.show			= show,
1162 	.store			= store,
1163 };
1164 
1165 static void threshold_block_release(struct kobject *kobj);
1166 
1167 static struct kobj_type threshold_ktype = {
1168 	.sysfs_ops		= &threshold_ops,
1169 	.default_attrs		= default_attrs,
1170 	.release		= threshold_block_release,
1171 };
1172 
1173 static const char *get_name(unsigned int bank, struct threshold_block *b)
1174 {
1175 	enum smca_bank_types bank_type;
1176 
1177 	if (!mce_flags.smca) {
1178 		if (b && bank == 4)
1179 			return bank4_names(b);
1180 
1181 		return th_names[bank];
1182 	}
1183 
1184 	bank_type = smca_get_bank_type(bank);
1185 	if (bank_type >= N_SMCA_BANK_TYPES)
1186 		return NULL;
1187 
1188 	if (b && bank_type == SMCA_UMC) {
1189 		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1190 			return smca_umc_block_names[b->block];
1191 		return NULL;
1192 	}
1193 
1194 	if (smca_banks[bank].hwid->count == 1)
1195 		return smca_get_name(bank_type);
1196 
1197 	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1198 		 "%s_%x", smca_get_name(bank_type),
1199 			  smca_banks[bank].sysfs_id);
1200 	return buf_mcatype;
1201 }
1202 
1203 static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb,
1204 				     unsigned int bank, unsigned int block,
1205 				     u32 address)
1206 {
1207 	struct threshold_block *b = NULL;
1208 	u32 low, high;
1209 	int err;
1210 
1211 	if ((bank >= per_cpu(mce_num_banks, cpu)) || (block >= NR_BLOCKS))
1212 		return 0;
1213 
1214 	if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
1215 		return 0;
1216 
1217 	if (!(high & MASK_VALID_HI)) {
1218 		if (block)
1219 			goto recurse;
1220 		else
1221 			return 0;
1222 	}
1223 
1224 	if (!(high & MASK_CNTP_HI)  ||
1225 	     (high & MASK_LOCKED_HI))
1226 		goto recurse;
1227 
1228 	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1229 	if (!b)
1230 		return -ENOMEM;
1231 
1232 	b->block		= block;
1233 	b->bank			= bank;
1234 	b->cpu			= cpu;
1235 	b->address		= address;
1236 	b->interrupt_enable	= 0;
1237 	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1238 	b->threshold_limit	= THRESHOLD_MAX;
1239 
1240 	if (b->interrupt_capable) {
1241 		threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
1242 		b->interrupt_enable = 1;
1243 	} else {
1244 		threshold_ktype.default_attrs[2] = NULL;
1245 	}
1246 
1247 	INIT_LIST_HEAD(&b->miscj);
1248 
1249 	if (tb->blocks)
1250 		list_add(&b->miscj, &tb->blocks->miscj);
1251 	else
1252 		tb->blocks = b;
1253 
1254 	err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(bank, b));
1255 	if (err)
1256 		goto out_free;
1257 recurse:
1258 	address = get_block_address(address, low, high, bank, ++block, cpu);
1259 	if (!address)
1260 		return 0;
1261 
1262 	err = allocate_threshold_blocks(cpu, tb, bank, block, address);
1263 	if (err)
1264 		goto out_free;
1265 
1266 	if (b)
1267 		kobject_uevent(&b->kobj, KOBJ_ADD);
1268 
1269 	return err;
1270 
1271 out_free:
1272 	if (b) {
1273 		kobject_put(&b->kobj);
1274 		list_del(&b->miscj);
1275 		kfree(b);
1276 	}
1277 	return err;
1278 }
1279 
1280 static int __threshold_add_blocks(struct threshold_bank *b)
1281 {
1282 	struct list_head *head = &b->blocks->miscj;
1283 	struct threshold_block *pos = NULL;
1284 	struct threshold_block *tmp = NULL;
1285 	int err = 0;
1286 
1287 	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1288 	if (err)
1289 		return err;
1290 
1291 	list_for_each_entry_safe(pos, tmp, head, miscj) {
1292 
1293 		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1294 		if (err) {
1295 			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1296 				kobject_del(&pos->kobj);
1297 
1298 			return err;
1299 		}
1300 	}
1301 	return err;
1302 }
1303 
1304 static int threshold_create_bank(unsigned int cpu, unsigned int bank)
1305 {
1306 	struct device *dev = per_cpu(mce_device, cpu);
1307 	struct amd_northbridge *nb = NULL;
1308 	struct threshold_bank *b = NULL;
1309 	const char *name = get_name(bank, NULL);
1310 	int err = 0;
1311 
1312 	if (!dev)
1313 		return -ENODEV;
1314 
1315 	if (is_shared_bank(bank)) {
1316 		nb = node_to_amd_nb(amd_get_nb_id(cpu));
1317 
1318 		/* threshold descriptor already initialized on this node? */
1319 		if (nb && nb->bank4) {
1320 			/* yes, use it */
1321 			b = nb->bank4;
1322 			err = kobject_add(b->kobj, &dev->kobj, name);
1323 			if (err)
1324 				goto out;
1325 
1326 			per_cpu(threshold_banks, cpu)[bank] = b;
1327 			refcount_inc(&b->cpus);
1328 
1329 			err = __threshold_add_blocks(b);
1330 
1331 			goto out;
1332 		}
1333 	}
1334 
1335 	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1336 	if (!b) {
1337 		err = -ENOMEM;
1338 		goto out;
1339 	}
1340 
1341 	b->kobj = kobject_create_and_add(name, &dev->kobj);
1342 	if (!b->kobj) {
1343 		err = -EINVAL;
1344 		goto out_free;
1345 	}
1346 
1347 	if (is_shared_bank(bank)) {
1348 		refcount_set(&b->cpus, 1);
1349 
1350 		/* nb is already initialized, see above */
1351 		if (nb) {
1352 			WARN_ON(nb->bank4);
1353 			nb->bank4 = b;
1354 		}
1355 	}
1356 
1357 	err = allocate_threshold_blocks(cpu, b, bank, 0, msr_ops.misc(bank));
1358 	if (err)
1359 		goto out_free;
1360 
1361 	per_cpu(threshold_banks, cpu)[bank] = b;
1362 
1363 	return 0;
1364 
1365  out_free:
1366 	kfree(b);
1367 
1368  out:
1369 	return err;
1370 }
1371 
1372 static void threshold_block_release(struct kobject *kobj)
1373 {
1374 	kfree(to_block(kobj));
1375 }
1376 
1377 static void deallocate_threshold_block(unsigned int cpu, unsigned int bank)
1378 {
1379 	struct threshold_block *pos = NULL;
1380 	struct threshold_block *tmp = NULL;
1381 	struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
1382 
1383 	if (!head)
1384 		return;
1385 
1386 	list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
1387 		list_del(&pos->miscj);
1388 		kobject_put(&pos->kobj);
1389 	}
1390 
1391 	kobject_put(&head->blocks->kobj);
1392 }
1393 
1394 static void __threshold_remove_blocks(struct threshold_bank *b)
1395 {
1396 	struct threshold_block *pos = NULL;
1397 	struct threshold_block *tmp = NULL;
1398 
1399 	kobject_del(b->kobj);
1400 
1401 	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1402 		kobject_del(&pos->kobj);
1403 }
1404 
1405 static void threshold_remove_bank(unsigned int cpu, int bank)
1406 {
1407 	struct amd_northbridge *nb;
1408 	struct threshold_bank *b;
1409 
1410 	b = per_cpu(threshold_banks, cpu)[bank];
1411 	if (!b)
1412 		return;
1413 
1414 	if (!b->blocks)
1415 		goto free_out;
1416 
1417 	if (is_shared_bank(bank)) {
1418 		if (!refcount_dec_and_test(&b->cpus)) {
1419 			__threshold_remove_blocks(b);
1420 			per_cpu(threshold_banks, cpu)[bank] = NULL;
1421 			return;
1422 		} else {
1423 			/*
1424 			 * the last CPU on this node using the shared bank is
1425 			 * going away, remove that bank now.
1426 			 */
1427 			nb = node_to_amd_nb(amd_get_nb_id(cpu));
1428 			nb->bank4 = NULL;
1429 		}
1430 	}
1431 
1432 	deallocate_threshold_block(cpu, bank);
1433 
1434 free_out:
1435 	kobject_del(b->kobj);
1436 	kobject_put(b->kobj);
1437 	kfree(b);
1438 	per_cpu(threshold_banks, cpu)[bank] = NULL;
1439 }
1440 
1441 int mce_threshold_remove_device(unsigned int cpu)
1442 {
1443 	unsigned int bank;
1444 
1445 	for (bank = 0; bank < per_cpu(mce_num_banks, cpu); ++bank) {
1446 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1447 			continue;
1448 		threshold_remove_bank(cpu, bank);
1449 	}
1450 	kfree(per_cpu(threshold_banks, cpu));
1451 	per_cpu(threshold_banks, cpu) = NULL;
1452 	return 0;
1453 }
1454 
1455 /* create dir/files for all valid threshold banks */
1456 int mce_threshold_create_device(unsigned int cpu)
1457 {
1458 	unsigned int bank;
1459 	struct threshold_bank **bp;
1460 	int err = 0;
1461 
1462 	bp = per_cpu(threshold_banks, cpu);
1463 	if (bp)
1464 		return 0;
1465 
1466 	bp = kcalloc(per_cpu(mce_num_banks, cpu), sizeof(struct threshold_bank *),
1467 		     GFP_KERNEL);
1468 	if (!bp)
1469 		return -ENOMEM;
1470 
1471 	per_cpu(threshold_banks, cpu) = bp;
1472 
1473 	for (bank = 0; bank < per_cpu(mce_num_banks, cpu); ++bank) {
1474 		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1475 			continue;
1476 		err = threshold_create_bank(cpu, bank);
1477 		if (err)
1478 			goto err;
1479 	}
1480 	return err;
1481 err:
1482 	mce_threshold_remove_device(cpu);
1483 	return err;
1484 }
1485 
1486 static __init int threshold_init_device(void)
1487 {
1488 	unsigned lcpu = 0;
1489 
1490 	/* to hit CPUs online before the notifier is up */
1491 	for_each_online_cpu(lcpu) {
1492 		int err = mce_threshold_create_device(lcpu);
1493 
1494 		if (err)
1495 			return err;
1496 	}
1497 
1498 	if (thresholding_irq_en)
1499 		mce_threshold_vector = amd_threshold_interrupt;
1500 
1501 	return 0;
1502 }
1503 /*
1504  * there are 3 funcs which need to be _initcalled in a logic sequence:
1505  * 1. xen_late_init_mcelog
1506  * 2. mcheck_init_device
1507  * 3. threshold_init_device
1508  *
1509  * xen_late_init_mcelog must register xen_mce_chrdev_device before
1510  * native mce_chrdev_device registration if running under xen platform;
1511  *
1512  * mcheck_init_device should be inited before threshold_init_device to
1513  * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
1514  *
1515  * so we use following _initcalls
1516  * 1. device_initcall(xen_late_init_mcelog);
1517  * 2. device_initcall_sync(mcheck_init_device);
1518  * 3. late_initcall(threshold_init_device);
1519  *
1520  * when running under xen, the initcall order is 1,2,3;
1521  * on baremetal, we skip 1 and we do only 2 and 3.
1522  */
1523 late_initcall(threshold_init_device);
1524