Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
Note: File does not exist in v5.4.
   1/*
   2 *  (c) 2005-2016 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 *  Maintained by: Borislav Petkov <bp@alien8.de>
   9 *
  10 *  All MC4_MISCi registers are shared between cores on a node.
  11 */
  12#include <linux/interrupt.h>
  13#include <linux/notifier.h>
  14#include <linux/kobject.h>
  15#include <linux/percpu.h>
  16#include <linux/errno.h>
  17#include <linux/sched.h>
  18#include <linux/sysfs.h>
  19#include <linux/slab.h>
  20#include <linux/init.h>
  21#include <linux/cpu.h>
  22#include <linux/smp.h>
  23#include <linux/string.h>
  24
  25#include <asm/amd_nb.h>
  26#include <asm/apic.h>
  27#include <asm/mce.h>
  28#include <asm/msr.h>
  29#include <asm/trace/irq_vectors.h>
  30
  31#include "mce-internal.h"
  32
  33#define NR_BLOCKS         5
  34#define THRESHOLD_MAX     0xFFF
  35#define INT_TYPE_APIC     0x00020000
  36#define MASK_VALID_HI     0x80000000
  37#define MASK_CNTP_HI      0x40000000
  38#define MASK_LOCKED_HI    0x20000000
  39#define MASK_LVTOFF_HI    0x00F00000
  40#define MASK_COUNT_EN_HI  0x00080000
  41#define MASK_INT_TYPE_HI  0x00060000
  42#define MASK_OVERFLOW_HI  0x00010000
  43#define MASK_ERR_COUNT_HI 0x00000FFF
  44#define MASK_BLKPTR_LO    0xFF000000
  45#define MCG_XBLK_ADDR     0xC0000400
  46
  47/* Deferred error settings */
  48#define MSR_CU_DEF_ERR		0xC0000410
  49#define MASK_DEF_LVTOFF		0x000000F0
  50#define MASK_DEF_INT_TYPE	0x00000006
  51#define DEF_LVT_OFF		0x2
  52#define DEF_INT_TYPE_APIC	0x2
  53
  54/* Scalable MCA: */
  55
  56/* Threshold LVT offset is at MSR0xC0000410[15:12] */
  57#define SMCA_THR_LVT_OFF	0xF000
  58
  59static bool thresholding_en;
  60
  61static const char * const th_names[] = {
  62	"load_store",
  63	"insn_fetch",
  64	"combined_unit",
  65	"decode_unit",
  66	"northbridge",
  67	"execution_unit",
  68};
  69
  70static const char * const smca_umc_block_names[] = {
  71	"dram_ecc",
  72	"misc_umc"
  73};
  74
  75struct smca_bank_name {
  76	const char *name;	/* Short name for sysfs */
  77	const char *long_name;	/* Long name for pretty-printing */
  78};
  79
  80static struct smca_bank_name smca_names[] = {
  81	[SMCA_LS]	= { "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_PIE]	= { "pie",		"Power, Interrupts, etc." },
  91	[SMCA_UMC]	= { "umc",		"Unified Memory Controller" },
  92	[SMCA_PB]	= { "param_block",	"Parameter Block" },
  93	[SMCA_PSP]	= { "psp",		"Platform Security Processor" },
  94	[SMCA_SMU]	= { "smu",		"System Management Unit" },
  95};
  96
  97static u32 smca_bank_addrs[MAX_NR_BANKS][NR_BLOCKS] __ro_after_init =
  98{
  99	[0 ... MAX_NR_BANKS - 1] = { [0 ... NR_BLOCKS - 1] = -1 }
 100};
 101
 102const char *smca_get_name(enum smca_bank_types t)
 103{
 104	if (t >= N_SMCA_BANK_TYPES)
 105		return NULL;
 106
 107	return smca_names[t].name;
 108}
 109
 110const char *smca_get_long_name(enum smca_bank_types t)
 111{
 112	if (t >= N_SMCA_BANK_TYPES)
 113		return NULL;
 114
 115	return smca_names[t].long_name;
 116}
 117EXPORT_SYMBOL_GPL(smca_get_long_name);
 118
 119static enum smca_bank_types smca_get_bank_type(unsigned int bank)
 120{
 121	struct smca_bank *b;
 122
 123	if (bank >= MAX_NR_BANKS)
 124		return N_SMCA_BANK_TYPES;
 125
 126	b = &smca_banks[bank];
 127	if (!b->hwid)
 128		return N_SMCA_BANK_TYPES;
 129
 130	return b->hwid->bank_type;
 131}
 132
 133static struct smca_hwid smca_hwid_mcatypes[] = {
 134	/* { bank_type, hwid_mcatype, xec_bitmap } */
 135
 136	/* Reserved type */
 137	{ SMCA_RESERVED, HWID_MCATYPE(0x00, 0x0), 0x0 },
 138
 139	/* ZN Core (HWID=0xB0) MCA types */
 140	{ SMCA_LS,	 HWID_MCATYPE(0xB0, 0x0), 0x1FFFEF },
 141	{ SMCA_IF,	 HWID_MCATYPE(0xB0, 0x1), 0x3FFF },
 142	{ SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2), 0xF },
 143	{ SMCA_DE,	 HWID_MCATYPE(0xB0, 0x3), 0x1FF },
 144	/* HWID 0xB0 MCATYPE 0x4 is Reserved */
 145	{ SMCA_EX,	 HWID_MCATYPE(0xB0, 0x5), 0x7FF },
 146	{ SMCA_FP,	 HWID_MCATYPE(0xB0, 0x6), 0x7F },
 147	{ SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7), 0xFF },
 148
 149	/* Data Fabric MCA types */
 150	{ SMCA_CS,	 HWID_MCATYPE(0x2E, 0x0), 0x1FF },
 151	{ SMCA_PIE,	 HWID_MCATYPE(0x2E, 0x1), 0xF },
 152
 153	/* Unified Memory Controller MCA type */
 154	{ SMCA_UMC,	 HWID_MCATYPE(0x96, 0x0), 0x3F },
 155
 156	/* Parameter Block MCA type */
 157	{ SMCA_PB,	 HWID_MCATYPE(0x05, 0x0), 0x1 },
 158
 159	/* Platform Security Processor MCA type */
 160	{ SMCA_PSP,	 HWID_MCATYPE(0xFF, 0x0), 0x1 },
 161
 162	/* System Management Unit MCA type */
 163	{ SMCA_SMU,	 HWID_MCATYPE(0x01, 0x0), 0x1 },
 164};
 165
 166struct smca_bank smca_banks[MAX_NR_BANKS];
 167EXPORT_SYMBOL_GPL(smca_banks);
 168
 169/*
 170 * In SMCA enabled processors, we can have multiple banks for a given IP type.
 171 * So to define a unique name for each bank, we use a temp c-string to append
 172 * the MCA_IPID[InstanceId] to type's name in get_name().
 173 *
 174 * InstanceId is 32 bits which is 8 characters. Make sure MAX_MCATYPE_NAME_LEN
 175 * is greater than 8 plus 1 (for underscore) plus length of longest type name.
 176 */
 177#define MAX_MCATYPE_NAME_LEN	30
 178static char buf_mcatype[MAX_MCATYPE_NAME_LEN];
 179
 180static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
 181static DEFINE_PER_CPU(unsigned int, bank_map);	/* see which banks are on */
 182
 183static void amd_threshold_interrupt(void);
 184static void amd_deferred_error_interrupt(void);
 185
 186static void default_deferred_error_interrupt(void)
 187{
 188	pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
 189}
 190void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
 191
 192static void smca_configure(unsigned int bank, unsigned int cpu)
 193{
 194	unsigned int i, hwid_mcatype;
 195	struct smca_hwid *s_hwid;
 196	u32 high, low;
 197	u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
 198
 199	/* Set appropriate bits in MCA_CONFIG */
 200	if (!rdmsr_safe(smca_config, &low, &high)) {
 201		/*
 202		 * OS is required to set the MCAX bit to acknowledge that it is
 203		 * now using the new MSR ranges and new registers under each
 204		 * bank. It also means that the OS will configure deferred
 205		 * errors in the new MCx_CONFIG register. If the bit is not set,
 206		 * uncorrectable errors will cause a system panic.
 207		 *
 208		 * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
 209		 */
 210		high |= BIT(0);
 211
 212		/*
 213		 * SMCA sets the Deferred Error Interrupt type per bank.
 214		 *
 215		 * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
 216		 * if the DeferredIntType bit field is available.
 217		 *
 218		 * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
 219		 * high portion of the MSR). OS should set this to 0x1 to enable
 220		 * APIC based interrupt. First, check that no interrupt has been
 221		 * set.
 222		 */
 223		if ((low & BIT(5)) && !((high >> 5) & 0x3))
 224			high |= BIT(5);
 225
 226		wrmsr(smca_config, low, high);
 227	}
 228
 229	/* Return early if this bank was already initialized. */
 230	if (smca_banks[bank].hwid)
 231		return;
 232
 233	if (rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
 234		pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
 235		return;
 236	}
 237
 238	hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
 239				    (high & MCI_IPID_MCATYPE) >> 16);
 240
 241	for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
 242		s_hwid = &smca_hwid_mcatypes[i];
 243		if (hwid_mcatype == s_hwid->hwid_mcatype) {
 244			smca_banks[bank].hwid = s_hwid;
 245			smca_banks[bank].id = low;
 246			smca_banks[bank].sysfs_id = s_hwid->count++;
 247			break;
 248		}
 249	}
 250}
 251
 252struct thresh_restart {
 253	struct threshold_block	*b;
 254	int			reset;
 255	int			set_lvt_off;
 256	int			lvt_off;
 257	u16			old_limit;
 258};
 259
 260static inline bool is_shared_bank(int bank)
 261{
 262	/*
 263	 * Scalable MCA provides for only one core to have access to the MSRs of
 264	 * a shared bank.
 265	 */
 266	if (mce_flags.smca)
 267		return false;
 268
 269	/* Bank 4 is for northbridge reporting and is thus shared */
 270	return (bank == 4);
 271}
 272
 273static const char *bank4_names(const struct threshold_block *b)
 274{
 275	switch (b->address) {
 276	/* MSR4_MISC0 */
 277	case 0x00000413:
 278		return "dram";
 279
 280	case 0xc0000408:
 281		return "ht_links";
 282
 283	case 0xc0000409:
 284		return "l3_cache";
 285
 286	default:
 287		WARN(1, "Funny MSR: 0x%08x\n", b->address);
 288		return "";
 289	}
 290};
 291
 292
 293static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
 294{
 295	/*
 296	 * bank 4 supports APIC LVT interrupts implicitly since forever.
 297	 */
 298	if (bank == 4)
 299		return true;
 300
 301	/*
 302	 * IntP: interrupt present; if this bit is set, the thresholding
 303	 * bank can generate APIC LVT interrupts
 304	 */
 305	return msr_high_bits & BIT(28);
 306}
 307
 308static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
 309{
 310	int msr = (hi & MASK_LVTOFF_HI) >> 20;
 311
 312	if (apic < 0) {
 313		pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
 314		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
 315		       b->bank, b->block, b->address, hi, lo);
 316		return 0;
 317	}
 318
 319	if (apic != msr) {
 320		/*
 321		 * On SMCA CPUs, LVT offset is programmed at a different MSR, and
 322		 * the BIOS provides the value. The original field where LVT offset
 323		 * was set is reserved. Return early here:
 324		 */
 325		if (mce_flags.smca)
 326			return 0;
 327
 328		pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
 329		       "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
 330		       b->cpu, apic, b->bank, b->block, b->address, hi, lo);
 331		return 0;
 332	}
 333
 334	return 1;
 335};
 336
 337/* Reprogram MCx_MISC MSR behind this threshold bank. */
 338static void threshold_restart_bank(void *_tr)
 339{
 340	struct thresh_restart *tr = _tr;
 341	u32 hi, lo;
 342
 343	rdmsr(tr->b->address, lo, hi);
 344
 345	if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
 346		tr->reset = 1;	/* limit cannot be lower than err count */
 347
 348	if (tr->reset) {		/* reset err count and overflow bit */
 349		hi =
 350		    (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
 351		    (THRESHOLD_MAX - tr->b->threshold_limit);
 352	} else if (tr->old_limit) {	/* change limit w/o reset */
 353		int new_count = (hi & THRESHOLD_MAX) +
 354		    (tr->old_limit - tr->b->threshold_limit);
 355
 356		hi = (hi & ~MASK_ERR_COUNT_HI) |
 357		    (new_count & THRESHOLD_MAX);
 358	}
 359
 360	/* clear IntType */
 361	hi &= ~MASK_INT_TYPE_HI;
 362
 363	if (!tr->b->interrupt_capable)
 364		goto done;
 365
 366	if (tr->set_lvt_off) {
 367		if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
 368			/* set new lvt offset */
 369			hi &= ~MASK_LVTOFF_HI;
 370			hi |= tr->lvt_off << 20;
 371		}
 372	}
 373
 374	if (tr->b->interrupt_enable)
 375		hi |= INT_TYPE_APIC;
 376
 377 done:
 378
 379	hi |= MASK_COUNT_EN_HI;
 380	wrmsr(tr->b->address, lo, hi);
 381}
 382
 383static void mce_threshold_block_init(struct threshold_block *b, int offset)
 384{
 385	struct thresh_restart tr = {
 386		.b			= b,
 387		.set_lvt_off		= 1,
 388		.lvt_off		= offset,
 389	};
 390
 391	b->threshold_limit		= THRESHOLD_MAX;
 392	threshold_restart_bank(&tr);
 393};
 394
 395static int setup_APIC_mce_threshold(int reserved, int new)
 396{
 397	if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
 398					      APIC_EILVT_MSG_FIX, 0))
 399		return new;
 400
 401	return reserved;
 402}
 403
 404static int setup_APIC_deferred_error(int reserved, int new)
 405{
 406	if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
 407					      APIC_EILVT_MSG_FIX, 0))
 408		return new;
 409
 410	return reserved;
 411}
 412
 413static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
 414{
 415	u32 low = 0, high = 0;
 416	int def_offset = -1, def_new;
 417
 418	if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
 419		return;
 420
 421	def_new = (low & MASK_DEF_LVTOFF) >> 4;
 422	if (!(low & MASK_DEF_LVTOFF)) {
 423		pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
 424		def_new = DEF_LVT_OFF;
 425		low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
 426	}
 427
 428	def_offset = setup_APIC_deferred_error(def_offset, def_new);
 429	if ((def_offset == def_new) &&
 430	    (deferred_error_int_vector != amd_deferred_error_interrupt))
 431		deferred_error_int_vector = amd_deferred_error_interrupt;
 432
 433	if (!mce_flags.smca)
 434		low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
 435
 436	wrmsr(MSR_CU_DEF_ERR, low, high);
 437}
 438
 439static u32 smca_get_block_address(unsigned int cpu, unsigned int bank,
 440				  unsigned int block)
 441{
 442	u32 low, high;
 443	u32 addr = 0;
 444
 445	if (smca_get_bank_type(bank) == SMCA_RESERVED)
 446		return addr;
 447
 448	if (!block)
 449		return MSR_AMD64_SMCA_MCx_MISC(bank);
 450
 451	/* Check our cache first: */
 452	if (smca_bank_addrs[bank][block] != -1)
 453		return smca_bank_addrs[bank][block];
 454
 455	/*
 456	 * For SMCA enabled processors, BLKPTR field of the first MISC register
 457	 * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
 458	 */
 459	if (rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
 460		goto out;
 461
 462	if (!(low & MCI_CONFIG_MCAX))
 463		goto out;
 464
 465	if (!rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high) &&
 466	    (low & MASK_BLKPTR_LO))
 467		addr = MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
 468
 469out:
 470	smca_bank_addrs[bank][block] = addr;
 471	return addr;
 472}
 473
 474static u32 get_block_address(unsigned int cpu, u32 current_addr, u32 low, u32 high,
 475			     unsigned int bank, unsigned int block)
 476{
 477	u32 addr = 0, offset = 0;
 478
 479	if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
 480		return addr;
 481
 482	if (mce_flags.smca)
 483		return smca_get_block_address(cpu, bank, block);
 484
 485	/* Fall back to method we used for older processors: */
 486	switch (block) {
 487	case 0:
 488		addr = msr_ops.misc(bank);
 489		break;
 490	case 1:
 491		offset = ((low & MASK_BLKPTR_LO) >> 21);
 492		if (offset)
 493			addr = MCG_XBLK_ADDR + offset;
 494		break;
 495	default:
 496		addr = ++current_addr;
 497	}
 498	return addr;
 499}
 500
 501static int
 502prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 503			int offset, u32 misc_high)
 504{
 505	unsigned int cpu = smp_processor_id();
 506	u32 smca_low, smca_high;
 507	struct threshold_block b;
 508	int new;
 509
 510	if (!block)
 511		per_cpu(bank_map, cpu) |= (1 << bank);
 512
 513	memset(&b, 0, sizeof(b));
 514	b.cpu			= cpu;
 515	b.bank			= bank;
 516	b.block			= block;
 517	b.address		= addr;
 518	b.interrupt_capable	= lvt_interrupt_supported(bank, misc_high);
 519
 520	if (!b.interrupt_capable)
 521		goto done;
 522
 523	b.interrupt_enable = 1;
 524
 525	if (!mce_flags.smca) {
 526		new = (misc_high & MASK_LVTOFF_HI) >> 20;
 527		goto set_offset;
 528	}
 529
 530	/* Gather LVT offset for thresholding: */
 531	if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
 532		goto out;
 533
 534	new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
 535
 536set_offset:
 537	offset = setup_APIC_mce_threshold(offset, new);
 538
 539	if ((offset == new) && (mce_threshold_vector != amd_threshold_interrupt))
 540		mce_threshold_vector = amd_threshold_interrupt;
 541
 542done:
 543	mce_threshold_block_init(&b, offset);
 544
 545out:
 546	return offset;
 547}
 548
 549/* cpu init entry point, called from mce.c with preempt off */
 550void mce_amd_feature_init(struct cpuinfo_x86 *c)
 551{
 552	u32 low = 0, high = 0, address = 0;
 553	unsigned int bank, block, cpu = smp_processor_id();
 554	int offset = -1;
 555
 556	for (bank = 0; bank < mca_cfg.banks; ++bank) {
 557		if (mce_flags.smca)
 558			smca_configure(bank, cpu);
 559
 560		for (block = 0; block < NR_BLOCKS; ++block) {
 561			address = get_block_address(cpu, address, low, high, bank, block);
 562			if (!address)
 563				break;
 564
 565			if (rdmsr_safe(address, &low, &high))
 566				break;
 567
 568			if (!(high & MASK_VALID_HI))
 569				continue;
 570
 571			if (!(high & MASK_CNTP_HI)  ||
 572			     (high & MASK_LOCKED_HI))
 573				continue;
 574
 575			offset = prepare_threshold_block(bank, block, address, offset, high);
 576		}
 577	}
 578
 579	if (mce_flags.succor)
 580		deferred_error_interrupt_enable(c);
 581}
 582
 583int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
 584{
 585	u64 dram_base_addr, dram_limit_addr, dram_hole_base;
 586	/* We start from the normalized address */
 587	u64 ret_addr = norm_addr;
 588
 589	u32 tmp;
 590
 591	u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
 592	u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
 593	u8 intlv_addr_sel, intlv_addr_bit;
 594	u8 num_intlv_bits, hashed_bit;
 595	u8 lgcy_mmio_hole_en, base = 0;
 596	u8 cs_mask, cs_id = 0;
 597	bool hash_enabled = false;
 598
 599	/* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
 600	if (amd_df_indirect_read(nid, 0, 0x1B4, umc, &tmp))
 601		goto out_err;
 602
 603	/* Remove HiAddrOffset from normalized address, if enabled: */
 604	if (tmp & BIT(0)) {
 605		u64 hi_addr_offset = (tmp & GENMASK_ULL(31, 20)) << 8;
 606
 607		if (norm_addr >= hi_addr_offset) {
 608			ret_addr -= hi_addr_offset;
 609			base = 1;
 610		}
 611	}
 612
 613	/* Read D18F0x110 (DramBaseAddress). */
 614	if (amd_df_indirect_read(nid, 0, 0x110 + (8 * base), umc, &tmp))
 615		goto out_err;
 616
 617	/* Check if address range is valid. */
 618	if (!(tmp & BIT(0))) {
 619		pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
 620			__func__, tmp);
 621		goto out_err;
 622	}
 623
 624	lgcy_mmio_hole_en = tmp & BIT(1);
 625	intlv_num_chan	  = (tmp >> 4) & 0xF;
 626	intlv_addr_sel	  = (tmp >> 8) & 0x7;
 627	dram_base_addr	  = (tmp & GENMASK_ULL(31, 12)) << 16;
 628
 629	/* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
 630	if (intlv_addr_sel > 3) {
 631		pr_err("%s: Invalid interleave address select %d.\n",
 632			__func__, intlv_addr_sel);
 633		goto out_err;
 634	}
 635
 636	/* Read D18F0x114 (DramLimitAddress). */
 637	if (amd_df_indirect_read(nid, 0, 0x114 + (8 * base), umc, &tmp))
 638		goto out_err;
 639
 640	intlv_num_sockets = (tmp >> 8) & 0x1;
 641	intlv_num_dies	  = (tmp >> 10) & 0x3;
 642	dram_limit_addr	  = ((tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
 643
 644	intlv_addr_bit = intlv_addr_sel + 8;
 645
 646	/* Re-use intlv_num_chan by setting it equal to log2(#channels) */
 647	switch (intlv_num_chan) {
 648	case 0:	intlv_num_chan = 0; break;
 649	case 1: intlv_num_chan = 1; break;
 650	case 3: intlv_num_chan = 2; break;
 651	case 5:	intlv_num_chan = 3; break;
 652	case 7:	intlv_num_chan = 4; break;
 653
 654	case 8: intlv_num_chan = 1;
 655		hash_enabled = true;
 656		break;
 657	default:
 658		pr_err("%s: Invalid number of interleaved channels %d.\n",
 659			__func__, intlv_num_chan);
 660		goto out_err;
 661	}
 662
 663	num_intlv_bits = intlv_num_chan;
 664
 665	if (intlv_num_dies > 2) {
 666		pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
 667			__func__, intlv_num_dies);
 668		goto out_err;
 669	}
 670
 671	num_intlv_bits += intlv_num_dies;
 672
 673	/* Add a bit if sockets are interleaved. */
 674	num_intlv_bits += intlv_num_sockets;
 675
 676	/* Assert num_intlv_bits <= 4 */
 677	if (num_intlv_bits > 4) {
 678		pr_err("%s: Invalid interleave bits %d.\n",
 679			__func__, num_intlv_bits);
 680		goto out_err;
 681	}
 682
 683	if (num_intlv_bits > 0) {
 684		u64 temp_addr_x, temp_addr_i, temp_addr_y;
 685		u8 die_id_bit, sock_id_bit, cs_fabric_id;
 686
 687		/*
 688		 * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
 689		 * This is the fabric id for this coherent slave. Use
 690		 * umc/channel# as instance id of the coherent slave
 691		 * for FICAA.
 692		 */
 693		if (amd_df_indirect_read(nid, 0, 0x50, umc, &tmp))
 694			goto out_err;
 695
 696		cs_fabric_id = (tmp >> 8) & 0xFF;
 697		die_id_bit   = 0;
 698
 699		/* If interleaved over more than 1 channel: */
 700		if (intlv_num_chan) {
 701			die_id_bit = intlv_num_chan;
 702			cs_mask	   = (1 << die_id_bit) - 1;
 703			cs_id	   = cs_fabric_id & cs_mask;
 704		}
 705
 706		sock_id_bit = die_id_bit;
 707
 708		/* Read D18F1x208 (SystemFabricIdMask). */
 709		if (intlv_num_dies || intlv_num_sockets)
 710			if (amd_df_indirect_read(nid, 1, 0x208, umc, &tmp))
 711				goto out_err;
 712
 713		/* If interleaved over more than 1 die. */
 714		if (intlv_num_dies) {
 715			sock_id_bit  = die_id_bit + intlv_num_dies;
 716			die_id_shift = (tmp >> 24) & 0xF;
 717			die_id_mask  = (tmp >> 8) & 0xFF;
 718
 719			cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
 720		}
 721
 722		/* If interleaved over more than 1 socket. */
 723		if (intlv_num_sockets) {
 724			socket_id_shift	= (tmp >> 28) & 0xF;
 725			socket_id_mask	= (tmp >> 16) & 0xFF;
 726
 727			cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
 728		}
 729
 730		/*
 731		 * The pre-interleaved address consists of XXXXXXIIIYYYYY
 732		 * where III is the ID for this CS, and XXXXXXYYYYY are the
 733		 * address bits from the post-interleaved address.
 734		 * "num_intlv_bits" has been calculated to tell us how many "I"
 735		 * bits there are. "intlv_addr_bit" tells us how many "Y" bits
 736		 * there are (where "I" starts).
 737		 */
 738		temp_addr_y = ret_addr & GENMASK_ULL(intlv_addr_bit-1, 0);
 739		temp_addr_i = (cs_id << intlv_addr_bit);
 740		temp_addr_x = (ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
 741		ret_addr    = temp_addr_x | temp_addr_i | temp_addr_y;
 742	}
 743
 744	/* Add dram base address */
 745	ret_addr += dram_base_addr;
 746
 747	/* If legacy MMIO hole enabled */
 748	if (lgcy_mmio_hole_en) {
 749		if (amd_df_indirect_read(nid, 0, 0x104, umc, &tmp))
 750			goto out_err;
 751
 752		dram_hole_base = tmp & GENMASK(31, 24);
 753		if (ret_addr >= dram_hole_base)
 754			ret_addr += (BIT_ULL(32) - dram_hole_base);
 755	}
 756
 757	if (hash_enabled) {
 758		/* Save some parentheses and grab ls-bit at the end. */
 759		hashed_bit =	(ret_addr >> 12) ^
 760				(ret_addr >> 18) ^
 761				(ret_addr >> 21) ^
 762				(ret_addr >> 30) ^
 763				cs_id;
 764
 765		hashed_bit &= BIT(0);
 766
 767		if (hashed_bit != ((ret_addr >> intlv_addr_bit) & BIT(0)))
 768			ret_addr ^= BIT(intlv_addr_bit);
 769	}
 770
 771	/* Is calculated system address is above DRAM limit address? */
 772	if (ret_addr > dram_limit_addr)
 773		goto out_err;
 774
 775	*sys_addr = ret_addr;
 776	return 0;
 777
 778out_err:
 779	return -EINVAL;
 780}
 781EXPORT_SYMBOL_GPL(umc_normaddr_to_sysaddr);
 782
 783bool amd_mce_is_memory_error(struct mce *m)
 784{
 785	/* ErrCodeExt[20:16] */
 786	u8 xec = (m->status >> 16) & 0x1f;
 787
 788	if (mce_flags.smca)
 789		return smca_get_bank_type(m->bank) == SMCA_UMC && xec == 0x0;
 790
 791	return m->bank == 4 && xec == 0x8;
 792}
 793
 794static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
 795{
 796	struct mce m;
 797
 798	mce_setup(&m);
 799
 800	m.status = status;
 801	m.misc   = misc;
 802	m.bank   = bank;
 803	m.tsc	 = rdtsc();
 804
 805	if (m.status & MCI_STATUS_ADDRV) {
 806		m.addr = addr;
 807
 808		/*
 809		 * Extract [55:<lsb>] where lsb is the least significant
 810		 * *valid* bit of the address bits.
 811		 */
 812		if (mce_flags.smca) {
 813			u8 lsb = (m.addr >> 56) & 0x3f;
 814
 815			m.addr &= GENMASK_ULL(55, lsb);
 816		}
 817	}
 818
 819	if (mce_flags.smca) {
 820		rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
 821
 822		if (m.status & MCI_STATUS_SYNDV)
 823			rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
 824	}
 825
 826	mce_log(&m);
 827}
 828
 829asmlinkage __visible void __irq_entry smp_deferred_error_interrupt(void)
 830{
 831	entering_irq();
 832	trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
 833	inc_irq_stat(irq_deferred_error_count);
 834	deferred_error_int_vector();
 835	trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
 836	exiting_ack_irq();
 837}
 838
 839/*
 840 * Returns true if the logged error is deferred. False, otherwise.
 841 */
 842static inline bool
 843_log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
 844{
 845	u64 status, addr = 0;
 846
 847	rdmsrl(msr_stat, status);
 848	if (!(status & MCI_STATUS_VAL))
 849		return false;
 850
 851	if (status & MCI_STATUS_ADDRV)
 852		rdmsrl(msr_addr, addr);
 853
 854	__log_error(bank, status, addr, misc);
 855
 856	wrmsrl(msr_stat, 0);
 857
 858	return status & MCI_STATUS_DEFERRED;
 859}
 860
 861/*
 862 * We have three scenarios for checking for Deferred errors:
 863 *
 864 * 1) Non-SMCA systems check MCA_STATUS and log error if found.
 865 * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
 866 *    clear MCA_DESTAT.
 867 * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
 868 *    log it.
 869 */
 870static void log_error_deferred(unsigned int bank)
 871{
 872	bool defrd;
 873
 874	defrd = _log_error_bank(bank, msr_ops.status(bank),
 875					msr_ops.addr(bank), 0);
 876
 877	if (!mce_flags.smca)
 878		return;
 879
 880	/* Clear MCA_DESTAT if we logged the deferred error from MCA_STATUS. */
 881	if (defrd) {
 882		wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
 883		return;
 884	}
 885
 886	/*
 887	 * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
 888	 * for a valid error.
 889	 */
 890	_log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
 891			      MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
 892}
 893
 894/* APIC interrupt handler for deferred errors */
 895static void amd_deferred_error_interrupt(void)
 896{
 897	unsigned int bank;
 898
 899	for (bank = 0; bank < mca_cfg.banks; ++bank)
 900		log_error_deferred(bank);
 901}
 902
 903static void log_error_thresholding(unsigned int bank, u64 misc)
 904{
 905	_log_error_bank(bank, msr_ops.status(bank), msr_ops.addr(bank), misc);
 906}
 907
 908static void log_and_reset_block(struct threshold_block *block)
 909{
 910	struct thresh_restart tr;
 911	u32 low = 0, high = 0;
 912
 913	if (!block)
 914		return;
 915
 916	if (rdmsr_safe(block->address, &low, &high))
 917		return;
 918
 919	if (!(high & MASK_OVERFLOW_HI))
 920		return;
 921
 922	/* Log the MCE which caused the threshold event. */
 923	log_error_thresholding(block->bank, ((u64)high << 32) | low);
 924
 925	/* Reset threshold block after logging error. */
 926	memset(&tr, 0, sizeof(tr));
 927	tr.b = block;
 928	threshold_restart_bank(&tr);
 929}
 930
 931/*
 932 * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
 933 * goes off when error_count reaches threshold_limit.
 934 */
 935static void amd_threshold_interrupt(void)
 936{
 937	struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
 938	unsigned int bank, cpu = smp_processor_id();
 939
 940	for (bank = 0; bank < mca_cfg.banks; ++bank) {
 941		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
 942			continue;
 943
 944		first_block = per_cpu(threshold_banks, cpu)[bank]->blocks;
 945		if (!first_block)
 946			continue;
 947
 948		/*
 949		 * The first block is also the head of the list. Check it first
 950		 * before iterating over the rest.
 951		 */
 952		log_and_reset_block(first_block);
 953		list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
 954			log_and_reset_block(block);
 955	}
 956}
 957
 958/*
 959 * Sysfs Interface
 960 */
 961
 962struct threshold_attr {
 963	struct attribute attr;
 964	ssize_t (*show) (struct threshold_block *, char *);
 965	ssize_t (*store) (struct threshold_block *, const char *, size_t count);
 966};
 967
 968#define SHOW_FIELDS(name)						\
 969static ssize_t show_ ## name(struct threshold_block *b, char *buf)	\
 970{									\
 971	return sprintf(buf, "%lu\n", (unsigned long) b->name);		\
 972}
 973SHOW_FIELDS(interrupt_enable)
 974SHOW_FIELDS(threshold_limit)
 975
 976static ssize_t
 977store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
 978{
 979	struct thresh_restart tr;
 980	unsigned long new;
 981
 982	if (!b->interrupt_capable)
 983		return -EINVAL;
 984
 985	if (kstrtoul(buf, 0, &new) < 0)
 986		return -EINVAL;
 987
 988	b->interrupt_enable = !!new;
 989
 990	memset(&tr, 0, sizeof(tr));
 991	tr.b		= b;
 992
 993	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
 994
 995	return size;
 996}
 997
 998static ssize_t
 999store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
1000{
1001	struct thresh_restart tr;
1002	unsigned long new;
1003
1004	if (kstrtoul(buf, 0, &new) < 0)
1005		return -EINVAL;
1006
1007	if (new > THRESHOLD_MAX)
1008		new = THRESHOLD_MAX;
1009	if (new < 1)
1010		new = 1;
1011
1012	memset(&tr, 0, sizeof(tr));
1013	tr.old_limit = b->threshold_limit;
1014	b->threshold_limit = new;
1015	tr.b = b;
1016
1017	smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1018
1019	return size;
1020}
1021
1022static ssize_t show_error_count(struct threshold_block *b, char *buf)
1023{
1024	u32 lo, hi;
1025
1026	rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
1027
1028	return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
1029				     (THRESHOLD_MAX - b->threshold_limit)));
1030}
1031
1032static struct threshold_attr error_count = {
1033	.attr = {.name = __stringify(error_count), .mode = 0444 },
1034	.show = show_error_count,
1035};
1036
1037#define RW_ATTR(val)							\
1038static struct threshold_attr val = {					\
1039	.attr	= {.name = __stringify(val), .mode = 0644 },		\
1040	.show	= show_## val,						\
1041	.store	= store_## val,						\
1042};
1043
1044RW_ATTR(interrupt_enable);
1045RW_ATTR(threshold_limit);
1046
1047static struct attribute *default_attrs[] = {
1048	&threshold_limit.attr,
1049	&error_count.attr,
1050	NULL,	/* possibly interrupt_enable if supported, see below */
1051	NULL,
1052};
1053
1054#define to_block(k)	container_of(k, struct threshold_block, kobj)
1055#define to_attr(a)	container_of(a, struct threshold_attr, attr)
1056
1057static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1058{
1059	struct threshold_block *b = to_block(kobj);
1060	struct threshold_attr *a = to_attr(attr);
1061	ssize_t ret;
1062
1063	ret = a->show ? a->show(b, buf) : -EIO;
1064
1065	return ret;
1066}
1067
1068static ssize_t store(struct kobject *kobj, struct attribute *attr,
1069		     const char *buf, size_t count)
1070{
1071	struct threshold_block *b = to_block(kobj);
1072	struct threshold_attr *a = to_attr(attr);
1073	ssize_t ret;
1074
1075	ret = a->store ? a->store(b, buf, count) : -EIO;
1076
1077	return ret;
1078}
1079
1080static const struct sysfs_ops threshold_ops = {
1081	.show			= show,
1082	.store			= store,
1083};
1084
1085static struct kobj_type threshold_ktype = {
1086	.sysfs_ops		= &threshold_ops,
1087	.default_attrs		= default_attrs,
1088};
1089
1090static const char *get_name(unsigned int bank, struct threshold_block *b)
1091{
1092	enum smca_bank_types bank_type;
1093
1094	if (!mce_flags.smca) {
1095		if (b && bank == 4)
1096			return bank4_names(b);
1097
1098		return th_names[bank];
1099	}
1100
1101	bank_type = smca_get_bank_type(bank);
1102	if (bank_type >= N_SMCA_BANK_TYPES)
1103		return NULL;
1104
1105	if (b && bank_type == SMCA_UMC) {
1106		if (b->block < ARRAY_SIZE(smca_umc_block_names))
1107			return smca_umc_block_names[b->block];
1108		return NULL;
1109	}
1110
1111	if (smca_banks[bank].hwid->count == 1)
1112		return smca_get_name(bank_type);
1113
1114	snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1115		 "%s_%x", smca_get_name(bank_type),
1116			  smca_banks[bank].sysfs_id);
1117	return buf_mcatype;
1118}
1119
1120static int allocate_threshold_blocks(unsigned int cpu, unsigned int bank,
1121				     unsigned int block, u32 address)
1122{
1123	struct threshold_block *b = NULL;
1124	u32 low, high;
1125	int err;
1126
1127	if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
1128		return 0;
1129
1130	if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
1131		return 0;
1132
1133	if (!(high & MASK_VALID_HI)) {
1134		if (block)
1135			goto recurse;
1136		else
1137			return 0;
1138	}
1139
1140	if (!(high & MASK_CNTP_HI)  ||
1141	     (high & MASK_LOCKED_HI))
1142		goto recurse;
1143
1144	b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1145	if (!b)
1146		return -ENOMEM;
1147
1148	b->block		= block;
1149	b->bank			= bank;
1150	b->cpu			= cpu;
1151	b->address		= address;
1152	b->interrupt_enable	= 0;
1153	b->interrupt_capable	= lvt_interrupt_supported(bank, high);
1154	b->threshold_limit	= THRESHOLD_MAX;
1155
1156	if (b->interrupt_capable) {
1157		threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
1158		b->interrupt_enable = 1;
1159	} else {
1160		threshold_ktype.default_attrs[2] = NULL;
1161	}
1162
1163	INIT_LIST_HEAD(&b->miscj);
1164
1165	if (per_cpu(threshold_banks, cpu)[bank]->blocks) {
1166		list_add(&b->miscj,
1167			 &per_cpu(threshold_banks, cpu)[bank]->blocks->miscj);
1168	} else {
1169		per_cpu(threshold_banks, cpu)[bank]->blocks = b;
1170	}
1171
1172	err = kobject_init_and_add(&b->kobj, &threshold_ktype,
1173				   per_cpu(threshold_banks, cpu)[bank]->kobj,
1174				   get_name(bank, b));
1175	if (err)
1176		goto out_free;
1177recurse:
1178	address = get_block_address(cpu, address, low, high, bank, ++block);
1179	if (!address)
1180		return 0;
1181
1182	err = allocate_threshold_blocks(cpu, bank, block, address);
1183	if (err)
1184		goto out_free;
1185
1186	if (b)
1187		kobject_uevent(&b->kobj, KOBJ_ADD);
1188
1189	return err;
1190
1191out_free:
1192	if (b) {
1193		kobject_put(&b->kobj);
1194		list_del(&b->miscj);
1195		kfree(b);
1196	}
1197	return err;
1198}
1199
1200static int __threshold_add_blocks(struct threshold_bank *b)
1201{
1202	struct list_head *head = &b->blocks->miscj;
1203	struct threshold_block *pos = NULL;
1204	struct threshold_block *tmp = NULL;
1205	int err = 0;
1206
1207	err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1208	if (err)
1209		return err;
1210
1211	list_for_each_entry_safe(pos, tmp, head, miscj) {
1212
1213		err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1214		if (err) {
1215			list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1216				kobject_del(&pos->kobj);
1217
1218			return err;
1219		}
1220	}
1221	return err;
1222}
1223
1224static int threshold_create_bank(unsigned int cpu, unsigned int bank)
1225{
1226	struct device *dev = per_cpu(mce_device, cpu);
1227	struct amd_northbridge *nb = NULL;
1228	struct threshold_bank *b = NULL;
1229	const char *name = get_name(bank, NULL);
1230	int err = 0;
1231
1232	if (!dev)
1233		return -ENODEV;
1234
1235	if (is_shared_bank(bank)) {
1236		nb = node_to_amd_nb(amd_get_nb_id(cpu));
1237
1238		/* threshold descriptor already initialized on this node? */
1239		if (nb && nb->bank4) {
1240			/* yes, use it */
1241			b = nb->bank4;
1242			err = kobject_add(b->kobj, &dev->kobj, name);
1243			if (err)
1244				goto out;
1245
1246			per_cpu(threshold_banks, cpu)[bank] = b;
1247			refcount_inc(&b->cpus);
1248
1249			err = __threshold_add_blocks(b);
1250
1251			goto out;
1252		}
1253	}
1254
1255	b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1256	if (!b) {
1257		err = -ENOMEM;
1258		goto out;
1259	}
1260
1261	b->kobj = kobject_create_and_add(name, &dev->kobj);
1262	if (!b->kobj) {
1263		err = -EINVAL;
1264		goto out_free;
1265	}
1266
1267	per_cpu(threshold_banks, cpu)[bank] = b;
1268
1269	if (is_shared_bank(bank)) {
1270		refcount_set(&b->cpus, 1);
1271
1272		/* nb is already initialized, see above */
1273		if (nb) {
1274			WARN_ON(nb->bank4);
1275			nb->bank4 = b;
1276		}
1277	}
1278
1279	err = allocate_threshold_blocks(cpu, bank, 0, msr_ops.misc(bank));
1280	if (!err)
1281		goto out;
1282
1283 out_free:
1284	kfree(b);
1285
1286 out:
1287	return err;
1288}
1289
1290static void deallocate_threshold_block(unsigned int cpu,
1291						 unsigned int bank)
1292{
1293	struct threshold_block *pos = NULL;
1294	struct threshold_block *tmp = NULL;
1295	struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
1296
1297	if (!head)
1298		return;
1299
1300	list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
1301		kobject_put(&pos->kobj);
1302		list_del(&pos->miscj);
1303		kfree(pos);
1304	}
1305
1306	kfree(per_cpu(threshold_banks, cpu)[bank]->blocks);
1307	per_cpu(threshold_banks, cpu)[bank]->blocks = NULL;
1308}
1309
1310static void __threshold_remove_blocks(struct threshold_bank *b)
1311{
1312	struct threshold_block *pos = NULL;
1313	struct threshold_block *tmp = NULL;
1314
1315	kobject_del(b->kobj);
1316
1317	list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1318		kobject_del(&pos->kobj);
1319}
1320
1321static void threshold_remove_bank(unsigned int cpu, int bank)
1322{
1323	struct amd_northbridge *nb;
1324	struct threshold_bank *b;
1325
1326	b = per_cpu(threshold_banks, cpu)[bank];
1327	if (!b)
1328		return;
1329
1330	if (!b->blocks)
1331		goto free_out;
1332
1333	if (is_shared_bank(bank)) {
1334		if (!refcount_dec_and_test(&b->cpus)) {
1335			__threshold_remove_blocks(b);
1336			per_cpu(threshold_banks, cpu)[bank] = NULL;
1337			return;
1338		} else {
1339			/*
1340			 * the last CPU on this node using the shared bank is
1341			 * going away, remove that bank now.
1342			 */
1343			nb = node_to_amd_nb(amd_get_nb_id(cpu));
1344			nb->bank4 = NULL;
1345		}
1346	}
1347
1348	deallocate_threshold_block(cpu, bank);
1349
1350free_out:
1351	kobject_del(b->kobj);
1352	kobject_put(b->kobj);
1353	kfree(b);
1354	per_cpu(threshold_banks, cpu)[bank] = NULL;
1355}
1356
1357int mce_threshold_remove_device(unsigned int cpu)
1358{
1359	unsigned int bank;
1360
1361	if (!thresholding_en)
1362		return 0;
1363
1364	for (bank = 0; bank < mca_cfg.banks; ++bank) {
1365		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1366			continue;
1367		threshold_remove_bank(cpu, bank);
1368	}
1369	kfree(per_cpu(threshold_banks, cpu));
1370	per_cpu(threshold_banks, cpu) = NULL;
1371	return 0;
1372}
1373
1374/* create dir/files for all valid threshold banks */
1375int mce_threshold_create_device(unsigned int cpu)
1376{
1377	unsigned int bank;
1378	struct threshold_bank **bp;
1379	int err = 0;
1380
1381	if (!thresholding_en)
1382		return 0;
1383
1384	bp = per_cpu(threshold_banks, cpu);
1385	if (bp)
1386		return 0;
1387
1388	bp = kzalloc(sizeof(struct threshold_bank *) * mca_cfg.banks,
1389		     GFP_KERNEL);
1390	if (!bp)
1391		return -ENOMEM;
1392
1393	per_cpu(threshold_banks, cpu) = bp;
1394
1395	for (bank = 0; bank < mca_cfg.banks; ++bank) {
1396		if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1397			continue;
1398		err = threshold_create_bank(cpu, bank);
1399		if (err)
1400			goto err;
1401	}
1402	return err;
1403err:
1404	mce_threshold_remove_device(cpu);
1405	return err;
1406}
1407
1408static __init int threshold_init_device(void)
1409{
1410	unsigned lcpu = 0;
1411
1412	if (mce_threshold_vector == amd_threshold_interrupt)
1413		thresholding_en = true;
1414
1415	/* to hit CPUs online before the notifier is up */
1416	for_each_online_cpu(lcpu) {
1417		int err = mce_threshold_create_device(lcpu);
1418
1419		if (err)
1420			return err;
1421	}
1422
1423	return 0;
1424}
1425/*
1426 * there are 3 funcs which need to be _initcalled in a logic sequence:
1427 * 1. xen_late_init_mcelog
1428 * 2. mcheck_init_device
1429 * 3. threshold_init_device
1430 *
1431 * xen_late_init_mcelog must register xen_mce_chrdev_device before
1432 * native mce_chrdev_device registration if running under xen platform;
1433 *
1434 * mcheck_init_device should be inited before threshold_init_device to
1435 * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
1436 *
1437 * so we use following _initcalls
1438 * 1. device_initcall(xen_late_init_mcelog);
1439 * 2. device_initcall_sync(mcheck_init_device);
1440 * 3. late_initcall(threshold_init_device);
1441 *
1442 * when running under xen, the initcall order is 1,2,3;
1443 * on baremetal, we skip 1 and we do only 2 and 3.
1444 */
1445late_initcall(threshold_init_device);