arch/x86/kernel/cpu/mcheck/mce_amd.c

   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/traps.h>
  27 #include <asm/apic.h>
  28 #include <asm/mce.h>
  29 #include <asm/msr.h>
  30 #include <asm/trace/irq_vectors.h>
  31
  32 #include "mce-internal.h"
  33
  34 #define NR_BLOCKS         5
  35 #define THRESHOLD_MAX     0xFFF
  36 #define INT_TYPE_APIC     0x00020000
  37 #define MASK_VALID_HI     0x80000000
  38 #define MASK_CNTP_HI      0x40000000
  39 #define MASK_LOCKED_HI    0x20000000
  40 #define MASK_LVTOFF_HI    0x00F00000
  41 #define MASK_COUNT_EN_HI  0x00080000
  42 #define MASK_INT_TYPE_HI  0x00060000
  43 #define MASK_OVERFLOW_HI  0x00010000
  44 #define MASK_ERR_COUNT_HI 0x00000FFF
  45 #define MASK_BLKPTR_LO    0xFF000000
  46 #define MCG_XBLK_ADDR     0xC0000400
  47
  48 /* Deferred error settings */
  49 #define MSR_CU_DEF_ERR          0xC0000410
  50 #define MASK_DEF_LVTOFF         0x000000F0
  51 #define MASK_DEF_INT_TYPE       0x00000006
  52 #define DEF_LVT_OFF             0x2
  53 #define DEF_INT_TYPE_APIC       0x2
  54
  55 /* Scalable MCA: */
  56
  57 /* Threshold LVT offset is at MSR0xC0000410[15:12] */
  58 #define SMCA_THR_LVT_OFF        0xF000
  59
  60 static bool thresholding_irq_en;
  61
  62 static const char * const th_names[] = {
  63         "load_store",
  64         "insn_fetch",
  65         "combined_unit",
  66         "decode_unit",
  67         "northbridge",
  68         "execution_unit",
  69 };
  70
  71 static const char * const smca_umc_block_names[] = {
  72         "dram_ecc",
  73         "misc_umc"
  74 };
  75
  76 struct smca_bank_name {
  77         const char *name;       /* Short name for sysfs */
  78         const char *long_name;  /* Long name for pretty-printing */
  79 };
  80
  81 static struct smca_bank_name smca_names[] = {
  82         [SMCA_LS]       = { "load_store",       "Load Store Unit" },
  83         [SMCA_IF]       = { "insn_fetch",       "Instruction Fetch Unit" },
  84         [SMCA_L2_CACHE] = { "l2_cache",         "L2 Cache" },
  85         [SMCA_DE]       = { "decode_unit",      "Decode Unit" },
  86         [SMCA_RESERVED] = { "reserved",         "Reserved" },
  87         [SMCA_EX]       = { "execution_unit",   "Execution Unit" },
  88         [SMCA_FP]       = { "floating_point",   "Floating Point Unit" },
  89         [SMCA_L3_CACHE] = { "l3_cache",         "L3 Cache" },
  90         [SMCA_CS]       = { "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_SMU]      = { "smu",              "System Management Unit" },
  96 };
  97
  98 static u32 smca_bank_addrs[MAX_NR_BANKS][NR_BLOCKS] __ro_after_init =
  99 {
 100         [0 ... MAX_NR_BANKS - 1] = { [0 ... NR_BLOCKS - 1] = -1 }
 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), 0x1FFFEF },
 142         { SMCA_IF,       HWID_MCATYPE(0xB0, 0x1), 0x3FFF },
 143         { SMCA_L2_CACHE, HWID_MCATYPE(0xB0, 0x2), 0xF },
 144         { SMCA_DE,       HWID_MCATYPE(0xB0, 0x3), 0x1FF },
 145         /* HWID 0xB0 MCATYPE 0x4 is Reserved */
 146         { SMCA_EX,       HWID_MCATYPE(0xB0, 0x5), 0x7FF },
 147         { SMCA_FP,       HWID_MCATYPE(0xB0, 0x6), 0x7F },
 148         { SMCA_L3_CACHE, HWID_MCATYPE(0xB0, 0x7), 0xFF },
 149
 150         /* Data Fabric MCA types */
 151         { SMCA_CS,       HWID_MCATYPE(0x2E, 0x0), 0x1FF },
 152         { SMCA_PIE,      HWID_MCATYPE(0x2E, 0x1), 0xF },
 153
 154         /* Unified Memory Controller MCA type */
 155         { SMCA_UMC,      HWID_MCATYPE(0x96, 0x0), 0x3F },
 156
 157         /* Parameter Block MCA type */
 158         { SMCA_PB,       HWID_MCATYPE(0x05, 0x0), 0x1 },
 159
 160         /* Platform Security Processor MCA type */
 161         { SMCA_PSP,      HWID_MCATYPE(0xFF, 0x0), 0x1 },
 162
 163         /* System Management Unit MCA type */
 164         { SMCA_SMU,      HWID_MCATYPE(0x01, 0x0), 0x1 },
 165 };
 166
 167 struct smca_bank smca_banks[MAX_NR_BANKS];
 168 EXPORT_SYMBOL_GPL(smca_banks);
 169
 170 /*
 171  * In SMCA enabled processors, we can have multiple banks for a given IP type.
 172  * So to define a unique name for each bank, we use a temp c-string to append
 173  * the MCA_IPID[InstanceId] to type's name in get_name().
 174  *
 175  * InstanceId is 32 bits which is 8 characters. Make sure MAX_MCATYPE_NAME_LEN
 176  * is greater than 8 plus 1 (for underscore) plus length of longest type name.
 177  */
 178 #define MAX_MCATYPE_NAME_LEN    30
 179 static char buf_mcatype[MAX_MCATYPE_NAME_LEN];
 180
 181 static DEFINE_PER_CPU(struct threshold_bank **, threshold_banks);
 182 static DEFINE_PER_CPU(unsigned int, bank_map);  /* see which banks are on */
 183
 184 static void amd_threshold_interrupt(void);
 185 static void amd_deferred_error_interrupt(void);
 186
 187 static void default_deferred_error_interrupt(void)
 188 {
 189         pr_err("Unexpected deferred interrupt at vector %x\n", DEFERRED_ERROR_VECTOR);
 190 }
 191 void (*deferred_error_int_vector)(void) = default_deferred_error_interrupt;
 192
 193 static void smca_configure(unsigned int bank, unsigned int cpu)
 194 {
 195         unsigned int i, hwid_mcatype;
 196         struct smca_hwid *s_hwid;
 197         u32 high, low;
 198         u32 smca_config = MSR_AMD64_SMCA_MCx_CONFIG(bank);
 199
 200         /* Set appropriate bits in MCA_CONFIG */
 201         if (!rdmsr_safe(smca_config, &low, &high)) {
 202                 /*
 203                  * OS is required to set the MCAX bit to acknowledge that it is
 204                  * now using the new MSR ranges and new registers under each
 205                  * bank. It also means that the OS will configure deferred
 206                  * errors in the new MCx_CONFIG register. If the bit is not set,
 207                  * uncorrectable errors will cause a system panic.
 208                  *
 209                  * MCA_CONFIG[MCAX] is bit 32 (0 in the high portion of the MSR.)
 210                  */
 211                 high |= BIT(0);
 212
 213                 /*
 214                  * SMCA sets the Deferred Error Interrupt type per bank.
 215                  *
 216                  * MCA_CONFIG[DeferredIntTypeSupported] is bit 5, and tells us
 217                  * if the DeferredIntType bit field is available.
 218                  *
 219                  * MCA_CONFIG[DeferredIntType] is bits [38:37] ([6:5] in the
 220                  * high portion of the MSR). OS should set this to 0x1 to enable
 221                  * APIC based interrupt. First, check that no interrupt has been
 222                  * set.
 223                  */
 224                 if ((low & BIT(5)) && !((high >> 5) & 0x3))
 225                         high |= BIT(5);
 226
 227                 wrmsr(smca_config, low, high);
 228         }
 229
 230         /* Return early if this bank was already initialized. */
 231         if (smca_banks[bank].hwid && smca_banks[bank].hwid->hwid_mcatype != 0)
 232                 return;
 233
 234         if (rdmsr_safe(MSR_AMD64_SMCA_MCx_IPID(bank), &low, &high)) {
 235                 pr_warn("Failed to read MCA_IPID for bank %d\n", bank);
 236                 return;
 237         }
 238
 239         hwid_mcatype = HWID_MCATYPE(high & MCI_IPID_HWID,
 240                                     (high & MCI_IPID_MCATYPE) >> 16);
 241
 242         for (i = 0; i < ARRAY_SIZE(smca_hwid_mcatypes); i++) {
 243                 s_hwid = &smca_hwid_mcatypes[i];
 244                 if (hwid_mcatype == s_hwid->hwid_mcatype) {
 245                         smca_banks[bank].hwid = s_hwid;
 246                         smca_banks[bank].id = low;
 247                         smca_banks[bank].sysfs_id = s_hwid->count++;
 248                         break;
 249                 }
 250         }
 251 }
 252
 253 struct thresh_restart {
 254         struct threshold_block  *b;
 255         int                     reset;
 256         int                     set_lvt_off;
 257         int                     lvt_off;
 258         u16                     old_limit;
 259 };
 260
 261 static inline bool is_shared_bank(int bank)
 262 {
 263         /*
 264          * Scalable MCA provides for only one core to have access to the MSRs of
 265          * a shared bank.
 266          */
 267         if (mce_flags.smca)
 268                 return false;
 269
 270         /* Bank 4 is for northbridge reporting and is thus shared */
 271         return (bank == 4);
 272 }
 273
 274 static const char *bank4_names(const struct threshold_block *b)
 275 {
 276         switch (b->address) {
 277         /* MSR4_MISC0 */
 278         case 0x00000413:
 279                 return "dram";
 280
 281         case 0xc0000408:
 282                 return "ht_links";
 283
 284         case 0xc0000409:
 285                 return "l3_cache";
 286
 287         default:
 288                 WARN(1, "Funny MSR: 0x%08x\n", b->address);
 289                 return "";
 290         }
 291 };
 292
 293
 294 static bool lvt_interrupt_supported(unsigned int bank, u32 msr_high_bits)
 295 {
 296         /*
 297          * bank 4 supports APIC LVT interrupts implicitly since forever.
 298          */
 299         if (bank == 4)
 300                 return true;
 301
 302         /*
 303          * IntP: interrupt present; if this bit is set, the thresholding
 304          * bank can generate APIC LVT interrupts
 305          */
 306         return msr_high_bits & BIT(28);
 307 }
 308
 309 static int lvt_off_valid(struct threshold_block *b, int apic, u32 lo, u32 hi)
 310 {
 311         int msr = (hi & MASK_LVTOFF_HI) >> 20;
 312
 313         if (apic < 0) {
 314                 pr_err(FW_BUG "cpu %d, failed to setup threshold interrupt "
 315                        "for bank %d, block %d (MSR%08X=0x%x%08x)\n", b->cpu,
 316                        b->bank, b->block, b->address, hi, lo);
 317                 return 0;
 318         }
 319
 320         if (apic != msr) {
 321                 /*
 322                  * On SMCA CPUs, LVT offset is programmed at a different MSR, and
 323                  * the BIOS provides the value. The original field where LVT offset
 324                  * was set is reserved. Return early here:
 325                  */
 326                 if (mce_flags.smca)
 327                         return 0;
 328
 329                 pr_err(FW_BUG "cpu %d, invalid threshold interrupt offset %d "
 330                        "for bank %d, block %d (MSR%08X=0x%x%08x)\n",
 331                        b->cpu, apic, b->bank, b->block, b->address, hi, lo);
 332                 return 0;
 333         }
 334
 335         return 1;
 336 };
 337
 338 /* Reprogram MCx_MISC MSR behind this threshold bank. */
 339 static void threshold_restart_bank(void *_tr)
 340 {
 341         struct thresh_restart *tr = _tr;
 342         u32 hi, lo;
 343
 344         rdmsr(tr->b->address, lo, hi);
 345
 346         if (tr->b->threshold_limit < (hi & THRESHOLD_MAX))
 347                 tr->reset = 1;  /* limit cannot be lower than err count */
 348
 349         if (tr->reset) {                /* reset err count and overflow bit */
 350                 hi =
 351                     (hi & ~(MASK_ERR_COUNT_HI | MASK_OVERFLOW_HI)) |
 352                     (THRESHOLD_MAX - tr->b->threshold_limit);
 353         } else if (tr->old_limit) {     /* change limit w/o reset */
 354                 int new_count = (hi & THRESHOLD_MAX) +
 355                     (tr->old_limit - tr->b->threshold_limit);
 356
 357                 hi = (hi & ~MASK_ERR_COUNT_HI) |
 358                     (new_count & THRESHOLD_MAX);
 359         }
 360
 361         /* clear IntType */
 362         hi &= ~MASK_INT_TYPE_HI;
 363
 364         if (!tr->b->interrupt_capable)
 365                 goto done;
 366
 367         if (tr->set_lvt_off) {
 368                 if (lvt_off_valid(tr->b, tr->lvt_off, lo, hi)) {
 369                         /* set new lvt offset */
 370                         hi &= ~MASK_LVTOFF_HI;
 371                         hi |= tr->lvt_off << 20;
 372                 }
 373         }
 374
 375         if (tr->b->interrupt_enable)
 376                 hi |= INT_TYPE_APIC;
 377
 378  done:
 379
 380         hi |= MASK_COUNT_EN_HI;
 381         wrmsr(tr->b->address, lo, hi);
 382 }
 383
 384 static void mce_threshold_block_init(struct threshold_block *b, int offset)
 385 {
 386         struct thresh_restart tr = {
 387                 .b                      = b,
 388                 .set_lvt_off            = 1,
 389                 .lvt_off                = offset,
 390         };
 391
 392         b->threshold_limit              = THRESHOLD_MAX;
 393         threshold_restart_bank(&tr);
 394 };
 395
 396 static int setup_APIC_mce_threshold(int reserved, int new)
 397 {
 398         if (reserved < 0 && !setup_APIC_eilvt(new, THRESHOLD_APIC_VECTOR,
 399                                               APIC_EILVT_MSG_FIX, 0))
 400                 return new;
 401
 402         return reserved;
 403 }
 404
 405 static int setup_APIC_deferred_error(int reserved, int new)
 406 {
 407         if (reserved < 0 && !setup_APIC_eilvt(new, DEFERRED_ERROR_VECTOR,
 408                                               APIC_EILVT_MSG_FIX, 0))
 409                 return new;
 410
 411         return reserved;
 412 }
 413
 414 static void deferred_error_interrupt_enable(struct cpuinfo_x86 *c)
 415 {
 416         u32 low = 0, high = 0;
 417         int def_offset = -1, def_new;
 418
 419         if (rdmsr_safe(MSR_CU_DEF_ERR, &low, &high))
 420                 return;
 421
 422         def_new = (low & MASK_DEF_LVTOFF) >> 4;
 423         if (!(low & MASK_DEF_LVTOFF)) {
 424                 pr_err(FW_BUG "Your BIOS is not setting up LVT offset 0x2 for deferred error IRQs correctly.\n");
 425                 def_new = DEF_LVT_OFF;
 426                 low = (low & ~MASK_DEF_LVTOFF) | (DEF_LVT_OFF << 4);
 427         }
 428
 429         def_offset = setup_APIC_deferred_error(def_offset, def_new);
 430         if ((def_offset == def_new) &&
 431             (deferred_error_int_vector != amd_deferred_error_interrupt))
 432                 deferred_error_int_vector = amd_deferred_error_interrupt;
 433
 434         low = (low & ~MASK_DEF_INT_TYPE) | DEF_INT_TYPE_APIC;
 435         wrmsr(MSR_CU_DEF_ERR, low, high);
 436 }
 437
 438 static u32 smca_get_block_address(unsigned int cpu, unsigned int bank,
 439                                   unsigned int block)
 440 {
 441         u32 low, high;
 442         u32 addr = 0;
 443
 444         if (smca_get_bank_type(bank) == SMCA_RESERVED)
 445                 return addr;
 446
 447         if (!block)
 448                 return MSR_AMD64_SMCA_MCx_MISC(bank);
 449
 450         /* Check our cache first: */
 451         if (smca_bank_addrs[bank][block] != -1)
 452                 return smca_bank_addrs[bank][block];
 453
 454         /*
 455          * For SMCA enabled processors, BLKPTR field of the first MISC register
 456          * (MCx_MISC0) indicates presence of additional MISC regs set (MISC1-4).
 457          */
 458         if (rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_CONFIG(bank), &low, &high))
 459                 goto out;
 460
 461         if (!(low & MCI_CONFIG_MCAX))
 462                 goto out;
 463
 464         if (!rdmsr_safe_on_cpu(cpu, MSR_AMD64_SMCA_MCx_MISC(bank), &low, &high) &&
 465             (low & MASK_BLKPTR_LO))
 466                 addr = MSR_AMD64_SMCA_MCx_MISCy(bank, block - 1);
 467
 468 out:
 469         smca_bank_addrs[bank][block] = addr;
 470         return addr;
 471 }
 472
 473 static u32 get_block_address(unsigned int cpu, u32 current_addr, u32 low, u32 high,
 474                              unsigned int bank, unsigned int block)
 475 {
 476         u32 addr = 0, offset = 0;
 477
 478         if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
 479                 return addr;
 480
 481         if (mce_flags.smca)
 482                 return smca_get_block_address(cpu, bank, block);
 483
 484         /* Fall back to method we used for older processors: */
 485         switch (block) {
 486         case 0:
 487                 addr = msr_ops.misc(bank);
 488                 break;
 489         case 1:
 490                 offset = ((low & MASK_BLKPTR_LO) >> 21);
 491                 if (offset)
 492                         addr = MCG_XBLK_ADDR + offset;
 493                 break;
 494         default:
 495                 addr = ++current_addr;
 496         }
 497         return addr;
 498 }
 499
 500 static int
 501 prepare_threshold_block(unsigned int bank, unsigned int block, u32 addr,
 502                         int offset, u32 misc_high)
 503 {
 504         unsigned int cpu = smp_processor_id();
 505         u32 smca_low, smca_high;
 506         struct threshold_block b;
 507         int new;
 508
 509         if (!block)
 510                 per_cpu(bank_map, cpu) |= (1 << bank);
 511
 512         memset(&b, 0, sizeof(b));
 513         b.cpu                   = cpu;
 514         b.bank                  = bank;
 515         b.block                 = block;
 516         b.address               = addr;
 517         b.interrupt_capable     = lvt_interrupt_supported(bank, misc_high);
 518
 519         if (!b.interrupt_capable)
 520                 goto done;
 521
 522         b.interrupt_enable = 1;
 523
 524         if (!mce_flags.smca) {
 525                 new = (misc_high & MASK_LVTOFF_HI) >> 20;
 526                 goto set_offset;
 527         }
 528
 529         /* Gather LVT offset for thresholding: */
 530         if (rdmsr_safe(MSR_CU_DEF_ERR, &smca_low, &smca_high))
 531                 goto out;
 532
 533         new = (smca_low & SMCA_THR_LVT_OFF) >> 12;
 534
 535 set_offset:
 536         offset = setup_APIC_mce_threshold(offset, new);
 537         if (offset == new)
 538                 thresholding_irq_en = true;
 539
 540 done:
 541         mce_threshold_block_init(&b, offset);
 542
 543 out:
 544         return offset;
 545 }
 546
 547 /*
 548  * Turn off MC4_MISC thresholding banks on all family 0x15 models since
 549  * they're not supported there.
 550  */
 551 void disable_err_thresholding(struct cpuinfo_x86 *c)
 552 {
 553         int i;
 554         u64 hwcr;
 555         bool need_toggle;
 556         u32 msrs[] = {
 557                 0x00000413, /* MC4_MISC0 */
 558                 0xc0000408, /* MC4_MISC1 */
 559         };
 560
 561         if (c->x86 != 0x15)
 562                 return;
 563
 564         rdmsrl(MSR_K7_HWCR, hwcr);
 565
 566         /* McStatusWrEn has to be set */
 567         need_toggle = !(hwcr & BIT(18));
 568
 569         if (need_toggle)
 570                 wrmsrl(MSR_K7_HWCR, hwcr | BIT(18));
 571
 572         /* Clear CntP bit safely */
 573         for (i = 0; i < ARRAY_SIZE(msrs); i++)
 574                 msr_clear_bit(msrs[i], 62);
 575
 576         /* restore old settings */
 577         if (need_toggle)
 578                 wrmsrl(MSR_K7_HWCR, hwcr);
 579 }
 580
 581 /* cpu init entry point, called from mce.c with preempt off */
 582 void mce_amd_feature_init(struct cpuinfo_x86 *c)
 583 {
 584         u32 low = 0, high = 0, address = 0;
 585         unsigned int bank, block, cpu = smp_processor_id();
 586         int offset = -1;
 587
 588         disable_err_thresholding(c);
 589
 590         for (bank = 0; bank < mca_cfg.banks; ++bank) {
 591                 if (mce_flags.smca)
 592                         smca_configure(bank, cpu);
 593
 594                 for (block = 0; block < NR_BLOCKS; ++block) {
 595                         address = get_block_address(cpu, address, low, high, bank, block);
 596                         if (!address)
 597                                 break;
 598
 599                         if (rdmsr_safe(address, &low, &high))
 600                                 break;
 601
 602                         if (!(high & MASK_VALID_HI))
 603                                 continue;
 604
 605                         if (!(high & MASK_CNTP_HI)  ||
 606                              (high & MASK_LOCKED_HI))
 607                                 continue;
 608
 609                         offset = prepare_threshold_block(bank, block, address, offset, high);
 610                 }
 611         }
 612
 613         if (mce_flags.succor)
 614                 deferred_error_interrupt_enable(c);
 615 }
 616
 617 int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
 618 {
 619         u64 dram_base_addr, dram_limit_addr, dram_hole_base;
 620         /* We start from the normalized address */
 621         u64 ret_addr = norm_addr;
 622
 623         u32 tmp;
 624
 625         u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
 626         u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
 627         u8 intlv_addr_sel, intlv_addr_bit;
 628         u8 num_intlv_bits, hashed_bit;
 629         u8 lgcy_mmio_hole_en, base = 0;
 630         u8 cs_mask, cs_id = 0;
 631         bool hash_enabled = false;
 632
 633         /* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
 634         if (amd_df_indirect_read(nid, 0, 0x1B4, umc, &tmp))
 635                 goto out_err;
 636
 637         /* Remove HiAddrOffset from normalized address, if enabled: */
 638         if (tmp & BIT(0)) {
 639                 u64 hi_addr_offset = (tmp & GENMASK_ULL(31, 20)) << 8;
 640
 641                 if (norm_addr >= hi_addr_offset) {
 642                         ret_addr -= hi_addr_offset;
 643                         base = 1;
 644                 }
 645         }
 646
 647         /* Read D18F0x110 (DramBaseAddress). */
 648         if (amd_df_indirect_read(nid, 0, 0x110 + (8 * base), umc, &tmp))
 649                 goto out_err;
 650
 651         /* Check if address range is valid. */
 652         if (!(tmp & BIT(0))) {
 653                 pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
 654                         __func__, tmp);
 655                 goto out_err;
 656         }
 657
 658         lgcy_mmio_hole_en = tmp & BIT(1);
 659         intlv_num_chan    = (tmp >> 4) & 0xF;
 660         intlv_addr_sel    = (tmp >> 8) & 0x7;
 661         dram_base_addr    = (tmp & GENMASK_ULL(31, 12)) << 16;
 662
 663         /* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
 664         if (intlv_addr_sel > 3) {
 665                 pr_err("%s: Invalid interleave address select %d.\n",
 666                         __func__, intlv_addr_sel);
 667                 goto out_err;
 668         }
 669
 670         /* Read D18F0x114 (DramLimitAddress). */
 671         if (amd_df_indirect_read(nid, 0, 0x114 + (8 * base), umc, &tmp))
 672                 goto out_err;
 673
 674         intlv_num_sockets = (tmp >> 8) & 0x1;
 675         intlv_num_dies    = (tmp >> 10) & 0x3;
 676         dram_limit_addr   = ((tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
 677
 678         intlv_addr_bit = intlv_addr_sel + 8;
 679
 680         /* Re-use intlv_num_chan by setting it equal to log2(#channels) */
 681         switch (intlv_num_chan) {
 682         case 0: intlv_num_chan = 0; break;
 683         case 1: intlv_num_chan = 1; break;
 684         case 3: intlv_num_chan = 2; break;
 685         case 5: intlv_num_chan = 3; break;
 686         case 7: intlv_num_chan = 4; break;
 687
 688         case 8: intlv_num_chan = 1;
 689                 hash_enabled = true;
 690                 break;
 691         default:
 692                 pr_err("%s: Invalid number of interleaved channels %d.\n",
 693                         __func__, intlv_num_chan);
 694                 goto out_err;
 695         }
 696
 697         num_intlv_bits = intlv_num_chan;
 698
 699         if (intlv_num_dies > 2) {
 700                 pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
 701                         __func__, intlv_num_dies);
 702                 goto out_err;
 703         }
 704
 705         num_intlv_bits += intlv_num_dies;
 706
 707         /* Add a bit if sockets are interleaved. */
 708         num_intlv_bits += intlv_num_sockets;
 709
 710         /* Assert num_intlv_bits <= 4 */
 711         if (num_intlv_bits > 4) {
 712                 pr_err("%s: Invalid interleave bits %d.\n",
 713                         __func__, num_intlv_bits);
 714                 goto out_err;
 715         }
 716
 717         if (num_intlv_bits > 0) {
 718                 u64 temp_addr_x, temp_addr_i, temp_addr_y;
 719                 u8 die_id_bit, sock_id_bit, cs_fabric_id;
 720
 721                 /*
 722                  * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
 723                  * This is the fabric id for this coherent slave. Use
 724                  * umc/channel# as instance id of the coherent slave
 725                  * for FICAA.
 726                  */
 727                 if (amd_df_indirect_read(nid, 0, 0x50, umc, &tmp))
 728                         goto out_err;
 729
 730                 cs_fabric_id = (tmp >> 8) & 0xFF;
 731                 die_id_bit   = 0;
 732
 733                 /* If interleaved over more than 1 channel: */
 734                 if (intlv_num_chan) {
 735                         die_id_bit = intlv_num_chan;
 736                         cs_mask    = (1 << die_id_bit) - 1;
 737                         cs_id      = cs_fabric_id & cs_mask;
 738                 }
 739
 740                 sock_id_bit = die_id_bit;
 741
 742                 /* Read D18F1x208 (SystemFabricIdMask). */
 743                 if (intlv_num_dies || intlv_num_sockets)
 744                         if (amd_df_indirect_read(nid, 1, 0x208, umc, &tmp))
 745                                 goto out_err;
 746
 747                 /* If interleaved over more than 1 die. */
 748                 if (intlv_num_dies) {
 749                         sock_id_bit  = die_id_bit + intlv_num_dies;
 750                         die_id_shift = (tmp >> 24) & 0xF;
 751                         die_id_mask  = (tmp >> 8) & 0xFF;
 752
 753                         cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
 754                 }
 755
 756                 /* If interleaved over more than 1 socket. */
 757                 if (intlv_num_sockets) {
 758                         socket_id_shift = (tmp >> 28) & 0xF;
 759                         socket_id_mask  = (tmp >> 16) & 0xFF;
 760
 761                         cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
 762                 }
 763
 764                 /*
 765                  * The pre-interleaved address consists of XXXXXXIIIYYYYY
 766                  * where III is the ID for this CS, and XXXXXXYYYYY are the
 767                  * address bits from the post-interleaved address.
 768                  * "num_intlv_bits" has been calculated to tell us how many "I"
 769                  * bits there are. "intlv_addr_bit" tells us how many "Y" bits
 770                  * there are (where "I" starts).
 771                  */
 772                 temp_addr_y = ret_addr & GENMASK_ULL(intlv_addr_bit-1, 0);
 773                 temp_addr_i = (cs_id << intlv_addr_bit);
 774                 temp_addr_x = (ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
 775                 ret_addr    = temp_addr_x | temp_addr_i | temp_addr_y;
 776         }
 777
 778         /* Add dram base address */
 779         ret_addr += dram_base_addr;
 780
 781         /* If legacy MMIO hole enabled */
 782         if (lgcy_mmio_hole_en) {
 783                 if (amd_df_indirect_read(nid, 0, 0x104, umc, &tmp))
 784                         goto out_err;
 785
 786                 dram_hole_base = tmp & GENMASK(31, 24);
 787                 if (ret_addr >= dram_hole_base)
 788                         ret_addr += (BIT_ULL(32) - dram_hole_base);
 789         }
 790
 791         if (hash_enabled) {
 792                 /* Save some parentheses and grab ls-bit at the end. */
 793                 hashed_bit =    (ret_addr >> 12) ^
 794                                 (ret_addr >> 18) ^
 795                                 (ret_addr >> 21) ^
 796                                 (ret_addr >> 30) ^
 797                                 cs_id;
 798
 799                 hashed_bit &= BIT(0);
 800
 801                 if (hashed_bit != ((ret_addr >> intlv_addr_bit) & BIT(0)))
 802                         ret_addr ^= BIT(intlv_addr_bit);
 803         }
 804
 805         /* Is calculated system address is above DRAM limit address? */
 806         if (ret_addr > dram_limit_addr)
 807                 goto out_err;
 808
 809         *sys_addr = ret_addr;
 810         return 0;
 811
 812 out_err:
 813         return -EINVAL;
 814 }
 815 EXPORT_SYMBOL_GPL(umc_normaddr_to_sysaddr);
 816
 817 bool amd_mce_is_memory_error(struct mce *m)
 818 {
 819         /* ErrCodeExt[20:16] */
 820         u8 xec = (m->status >> 16) & 0x1f;
 821
 822         if (mce_flags.smca)
 823                 return smca_get_bank_type(m->bank) == SMCA_UMC && xec == 0x0;
 824
 825         return m->bank == 4 && xec == 0x8;
 826 }
 827
 828 static void __log_error(unsigned int bank, u64 status, u64 addr, u64 misc)
 829 {
 830         struct mce m;
 831
 832         mce_setup(&m);
 833
 834         m.status = status;
 835         m.misc   = misc;
 836         m.bank   = bank;
 837         m.tsc    = rdtsc();
 838
 839         if (m.status & MCI_STATUS_ADDRV) {
 840                 m.addr = addr;
 841
 842                 /*
 843                  * Extract [55:<lsb>] where lsb is the least significant
 844                  * *valid* bit of the address bits.
 845                  */
 846                 if (mce_flags.smca) {
 847                         u8 lsb = (m.addr >> 56) & 0x3f;
 848
 849                         m.addr &= GENMASK_ULL(55, lsb);
 850                 }
 851         }
 852
 853         if (mce_flags.smca) {
 854                 rdmsrl(MSR_AMD64_SMCA_MCx_IPID(bank), m.ipid);
 855
 856                 if (m.status & MCI_STATUS_SYNDV)
 857                         rdmsrl(MSR_AMD64_SMCA_MCx_SYND(bank), m.synd);
 858         }
 859
 860         mce_log(&m);
 861 }
 862
 863 asmlinkage __visible void __irq_entry smp_deferred_error_interrupt(struct pt_regs *regs)
 864 {
 865         entering_irq();
 866         trace_deferred_error_apic_entry(DEFERRED_ERROR_VECTOR);
 867         inc_irq_stat(irq_deferred_error_count);
 868         deferred_error_int_vector();
 869         trace_deferred_error_apic_exit(DEFERRED_ERROR_VECTOR);
 870         exiting_ack_irq();
 871 }
 872
 873 /*
 874  * Returns true if the logged error is deferred. False, otherwise.
 875  */
 876 static inline bool
 877 _log_error_bank(unsigned int bank, u32 msr_stat, u32 msr_addr, u64 misc)
 878 {
 879         u64 status, addr = 0;
 880
 881         rdmsrl(msr_stat, status);
 882         if (!(status & MCI_STATUS_VAL))
 883                 return false;
 884
 885         if (status & MCI_STATUS_ADDRV)
 886                 rdmsrl(msr_addr, addr);
 887
 888         __log_error(bank, status, addr, misc);
 889
 890         wrmsrl(msr_stat, 0);
 891
 892         return status & MCI_STATUS_DEFERRED;
 893 }
 894
 895 /*
 896  * We have three scenarios for checking for Deferred errors:
 897  *
 898  * 1) Non-SMCA systems check MCA_STATUS and log error if found.
 899  * 2) SMCA systems check MCA_STATUS. If error is found then log it and also
 900  *    clear MCA_DESTAT.
 901  * 3) SMCA systems check MCA_DESTAT, if error was not found in MCA_STATUS, and
 902  *    log it.
 903  */
 904 static void log_error_deferred(unsigned int bank)
 905 {
 906         bool defrd;
 907
 908         defrd = _log_error_bank(bank, msr_ops.status(bank),
 909                                         msr_ops.addr(bank), 0);
 910
 911         if (!mce_flags.smca)
 912                 return;
 913
 914         /* Clear MCA_DESTAT if we logged the deferred error from MCA_STATUS. */
 915         if (defrd) {
 916                 wrmsrl(MSR_AMD64_SMCA_MCx_DESTAT(bank), 0);
 917                 return;
 918         }
 919
 920         /*
 921          * Only deferred errors are logged in MCA_DE{STAT,ADDR} so just check
 922          * for a valid error.
 923          */
 924         _log_error_bank(bank, MSR_AMD64_SMCA_MCx_DESTAT(bank),
 925                               MSR_AMD64_SMCA_MCx_DEADDR(bank), 0);
 926 }
 927
 928 /* APIC interrupt handler for deferred errors */
 929 static void amd_deferred_error_interrupt(void)
 930 {
 931         unsigned int bank;
 932
 933         for (bank = 0; bank < mca_cfg.banks; ++bank)
 934                 log_error_deferred(bank);
 935 }
 936
 937 static void log_error_thresholding(unsigned int bank, u64 misc)
 938 {
 939         _log_error_bank(bank, msr_ops.status(bank), msr_ops.addr(bank), misc);
 940 }
 941
 942 static void log_and_reset_block(struct threshold_block *block)
 943 {
 944         struct thresh_restart tr;
 945         u32 low = 0, high = 0;
 946
 947         if (!block)
 948                 return;
 949
 950         if (rdmsr_safe(block->address, &low, &high))
 951                 return;
 952
 953         if (!(high & MASK_OVERFLOW_HI))
 954                 return;
 955
 956         /* Log the MCE which caused the threshold event. */
 957         log_error_thresholding(block->bank, ((u64)high << 32) | low);
 958
 959         /* Reset threshold block after logging error. */
 960         memset(&tr, 0, sizeof(tr));
 961         tr.b = block;
 962         threshold_restart_bank(&tr);
 963 }
 964
 965 /*
 966  * Threshold interrupt handler will service THRESHOLD_APIC_VECTOR. The interrupt
 967  * goes off when error_count reaches threshold_limit.
 968  */
 969 static void amd_threshold_interrupt(void)
 970 {
 971         struct threshold_block *first_block = NULL, *block = NULL, *tmp = NULL;
 972         unsigned int bank, cpu = smp_processor_id();
 973
 974         for (bank = 0; bank < mca_cfg.banks; ++bank) {
 975                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
 976                         continue;
 977
 978                 first_block = per_cpu(threshold_banks, cpu)[bank]->blocks;
 979                 if (!first_block)
 980                         continue;
 981
 982                 /*
 983                  * The first block is also the head of the list. Check it first
 984                  * before iterating over the rest.
 985                  */
 986                 log_and_reset_block(first_block);
 987                 list_for_each_entry_safe(block, tmp, &first_block->miscj, miscj)
 988                         log_and_reset_block(block);
 989         }
 990 }
 991
 992 /*
 993  * Sysfs Interface
 994  */
 995
 996 struct threshold_attr {
 997         struct attribute attr;
 998         ssize_t (*show) (struct threshold_block *, char *);
 999         ssize_t (*store) (struct threshold_block *, const char *, size_t count);
1000 };
1001
1002 #define SHOW_FIELDS(name)                                               \
1003 static ssize_t show_ ## name(struct threshold_block *b, char *buf)      \
1004 {                                                                       \
1005         return sprintf(buf, "%lu\n", (unsigned long) b->name);          \
1006 }
1007 SHOW_FIELDS(interrupt_enable)
1008 SHOW_FIELDS(threshold_limit)
1009
1010 static ssize_t
1011 store_interrupt_enable(struct threshold_block *b, const char *buf, size_t size)
1012 {
1013         struct thresh_restart tr;
1014         unsigned long new;
1015
1016         if (!b->interrupt_capable)
1017                 return -EINVAL;
1018
1019         if (kstrtoul(buf, 0, &new) < 0)
1020                 return -EINVAL;
1021
1022         b->interrupt_enable = !!new;
1023
1024         memset(&tr, 0, sizeof(tr));
1025         tr.b            = b;
1026
1027         smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1028
1029         return size;
1030 }
1031
1032 static ssize_t
1033 store_threshold_limit(struct threshold_block *b, const char *buf, size_t size)
1034 {
1035         struct thresh_restart tr;
1036         unsigned long new;
1037
1038         if (kstrtoul(buf, 0, &new) < 0)
1039                 return -EINVAL;
1040
1041         if (new > THRESHOLD_MAX)
1042                 new = THRESHOLD_MAX;
1043         if (new < 1)
1044                 new = 1;
1045
1046         memset(&tr, 0, sizeof(tr));
1047         tr.old_limit = b->threshold_limit;
1048         b->threshold_limit = new;
1049         tr.b = b;
1050
1051         smp_call_function_single(b->cpu, threshold_restart_bank, &tr, 1);
1052
1053         return size;
1054 }
1055
1056 static ssize_t show_error_count(struct threshold_block *b, char *buf)
1057 {
1058         u32 lo, hi;
1059
1060         rdmsr_on_cpu(b->cpu, b->address, &lo, &hi);
1061
1062         return sprintf(buf, "%u\n", ((hi & THRESHOLD_MAX) -
1063                                      (THRESHOLD_MAX - b->threshold_limit)));
1064 }
1065
1066 static struct threshold_attr error_count = {
1067         .attr = {.name = __stringify(error_count), .mode = 0444 },
1068         .show = show_error_count,
1069 };
1070
1071 #define RW_ATTR(val)                                                    \
1072 static struct threshold_attr val = {                                    \
1073         .attr   = {.name = __stringify(val), .mode = 0644 },            \
1074         .show   = show_## val,                                          \
1075         .store  = store_## val,                                         \
1076 };
1077
1078 RW_ATTR(interrupt_enable);
1079 RW_ATTR(threshold_limit);
1080
1081 static struct attribute *default_attrs[] = {
1082         &threshold_limit.attr,
1083         &error_count.attr,
1084         NULL,   /* possibly interrupt_enable if supported, see below */
1085         NULL,
1086 };
1087
1088 #define to_block(k)     container_of(k, struct threshold_block, kobj)
1089 #define to_attr(a)      container_of(a, struct threshold_attr, attr)
1090
1091 static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf)
1092 {
1093         struct threshold_block *b = to_block(kobj);
1094         struct threshold_attr *a = to_attr(attr);
1095         ssize_t ret;
1096
1097         ret = a->show ? a->show(b, buf) : -EIO;
1098
1099         return ret;
1100 }
1101
1102 static ssize_t store(struct kobject *kobj, struct attribute *attr,
1103                      const char *buf, size_t count)
1104 {
1105         struct threshold_block *b = to_block(kobj);
1106         struct threshold_attr *a = to_attr(attr);
1107         ssize_t ret;
1108
1109         ret = a->store ? a->store(b, buf, count) : -EIO;
1110
1111         return ret;
1112 }
1113
1114 static const struct sysfs_ops threshold_ops = {
1115         .show                   = show,
1116         .store                  = store,
1117 };
1118
1119 static void threshold_block_release(struct kobject *kobj);
1120
1121 static struct kobj_type threshold_ktype = {
1122         .sysfs_ops              = &threshold_ops,
1123         .default_attrs          = default_attrs,
1124         .release                = threshold_block_release,
1125 };
1126
1127 static const char *get_name(unsigned int bank, struct threshold_block *b)
1128 {
1129         enum smca_bank_types bank_type;
1130
1131         if (!mce_flags.smca) {
1132                 if (b && bank == 4)
1133                         return bank4_names(b);
1134
1135                 return th_names[bank];
1136         }
1137
1138         bank_type = smca_get_bank_type(bank);
1139         if (bank_type >= N_SMCA_BANK_TYPES)
1140                 return NULL;
1141
1142         if (b && bank_type == SMCA_UMC) {
1143                 if (b->block < ARRAY_SIZE(smca_umc_block_names))
1144                         return smca_umc_block_names[b->block];
1145                 return NULL;
1146         }
1147
1148         if (smca_banks[bank].hwid->count == 1)
1149                 return smca_get_name(bank_type);
1150
1151         snprintf(buf_mcatype, MAX_MCATYPE_NAME_LEN,
1152                  "%s_%x", smca_get_name(bank_type),
1153                           smca_banks[bank].sysfs_id);
1154         return buf_mcatype;
1155 }
1156
1157 static int allocate_threshold_blocks(unsigned int cpu, struct threshold_bank *tb,
1158                                      unsigned int bank, unsigned int block,
1159                                      u32 address)
1160 {
1161         struct threshold_block *b = NULL;
1162         u32 low, high;
1163         int err;
1164
1165         if ((bank >= mca_cfg.banks) || (block >= NR_BLOCKS))
1166                 return 0;
1167
1168         if (rdmsr_safe_on_cpu(cpu, address, &low, &high))
1169                 return 0;
1170
1171         if (!(high & MASK_VALID_HI)) {
1172                 if (block)
1173                         goto recurse;
1174                 else
1175                         return 0;
1176         }
1177
1178         if (!(high & MASK_CNTP_HI)  ||
1179              (high & MASK_LOCKED_HI))
1180                 goto recurse;
1181
1182         b = kzalloc(sizeof(struct threshold_block), GFP_KERNEL);
1183         if (!b)
1184                 return -ENOMEM;
1185
1186         b->block                = block;
1187         b->bank                 = bank;
1188         b->cpu                  = cpu;
1189         b->address              = address;
1190         b->interrupt_enable     = 0;
1191         b->interrupt_capable    = lvt_interrupt_supported(bank, high);
1192         b->threshold_limit      = THRESHOLD_MAX;
1193
1194         if (b->interrupt_capable) {
1195                 threshold_ktype.default_attrs[2] = &interrupt_enable.attr;
1196                 b->interrupt_enable = 1;
1197         } else {
1198                 threshold_ktype.default_attrs[2] = NULL;
1199         }
1200
1201         INIT_LIST_HEAD(&b->miscj);
1202
1203         if (tb->blocks)
1204                 list_add(&b->miscj, &tb->blocks->miscj);
1205         else
1206                 tb->blocks = b;
1207
1208         err = kobject_init_and_add(&b->kobj, &threshold_ktype, tb->kobj, get_name(bank, b));
1209         if (err)
1210                 goto out_free;
1211 recurse:
1212         address = get_block_address(cpu, address, low, high, bank, ++block);
1213         if (!address)
1214                 return 0;
1215
1216         err = allocate_threshold_blocks(cpu, tb, bank, block, address);
1217         if (err)
1218                 goto out_free;
1219
1220         if (b)
1221                 kobject_uevent(&b->kobj, KOBJ_ADD);
1222
1223         return err;
1224
1225 out_free:
1226         if (b) {
1227                 kobject_put(&b->kobj);
1228                 list_del(&b->miscj);
1229                 kfree(b);
1230         }
1231         return err;
1232 }
1233
1234 static int __threshold_add_blocks(struct threshold_bank *b)
1235 {
1236         struct list_head *head = &b->blocks->miscj;
1237         struct threshold_block *pos = NULL;
1238         struct threshold_block *tmp = NULL;
1239         int err = 0;
1240
1241         err = kobject_add(&b->blocks->kobj, b->kobj, b->blocks->kobj.name);
1242         if (err)
1243                 return err;
1244
1245         list_for_each_entry_safe(pos, tmp, head, miscj) {
1246
1247                 err = kobject_add(&pos->kobj, b->kobj, pos->kobj.name);
1248                 if (err) {
1249                         list_for_each_entry_safe_reverse(pos, tmp, head, miscj)
1250                                 kobject_del(&pos->kobj);
1251
1252                         return err;
1253                 }
1254         }
1255         return err;
1256 }
1257
1258 static int threshold_create_bank(unsigned int cpu, unsigned int bank)
1259 {
1260         struct device *dev = per_cpu(mce_device, cpu);
1261         struct amd_northbridge *nb = NULL;
1262         struct threshold_bank *b = NULL;
1263         const char *name = get_name(bank, NULL);
1264         int err = 0;
1265
1266         if (!dev)
1267                 return -ENODEV;
1268
1269         if (is_shared_bank(bank)) {
1270                 nb = node_to_amd_nb(amd_get_nb_id(cpu));
1271
1272                 /* threshold descriptor already initialized on this node? */
1273                 if (nb && nb->bank4) {
1274                         /* yes, use it */
1275                         b = nb->bank4;
1276                         err = kobject_add(b->kobj, &dev->kobj, name);
1277                         if (err)
1278                                 goto out;
1279
1280                         per_cpu(threshold_banks, cpu)[bank] = b;
1281                         refcount_inc(&b->cpus);
1282
1283                         err = __threshold_add_blocks(b);
1284
1285                         goto out;
1286                 }
1287         }
1288
1289         b = kzalloc(sizeof(struct threshold_bank), GFP_KERNEL);
1290         if (!b) {
1291                 err = -ENOMEM;
1292                 goto out;
1293         }
1294
1295         b->kobj = kobject_create_and_add(name, &dev->kobj);
1296         if (!b->kobj) {
1297                 err = -EINVAL;
1298                 goto out_free;
1299         }
1300
1301         if (is_shared_bank(bank)) {
1302                 refcount_set(&b->cpus, 1);
1303
1304                 /* nb is already initialized, see above */
1305                 if (nb) {
1306                         WARN_ON(nb->bank4);
1307                         nb->bank4 = b;
1308                 }
1309         }
1310
1311         err = allocate_threshold_blocks(cpu, b, bank, 0, msr_ops.misc(bank));
1312         if (err)
1313                 goto out_free;
1314
1315         per_cpu(threshold_banks, cpu)[bank] = b;
1316
1317         return 0;
1318
1319  out_free:
1320         kfree(b);
1321
1322  out:
1323         return err;
1324 }
1325
1326 static void threshold_block_release(struct kobject *kobj)
1327 {
1328         kfree(to_block(kobj));
1329 }
1330
1331 static void deallocate_threshold_block(unsigned int cpu, unsigned int bank)
1332 {
1333         struct threshold_block *pos = NULL;
1334         struct threshold_block *tmp = NULL;
1335         struct threshold_bank *head = per_cpu(threshold_banks, cpu)[bank];
1336
1337         if (!head)
1338                 return;
1339
1340         list_for_each_entry_safe(pos, tmp, &head->blocks->miscj, miscj) {
1341                 list_del(&pos->miscj);
1342                 kobject_put(&pos->kobj);
1343         }
1344
1345         kobject_put(&head->blocks->kobj);
1346 }
1347
1348 static void __threshold_remove_blocks(struct threshold_bank *b)
1349 {
1350         struct threshold_block *pos = NULL;
1351         struct threshold_block *tmp = NULL;
1352
1353         kobject_del(b->kobj);
1354
1355         list_for_each_entry_safe(pos, tmp, &b->blocks->miscj, miscj)
1356                 kobject_del(&pos->kobj);
1357 }
1358
1359 static void threshold_remove_bank(unsigned int cpu, int bank)
1360 {
1361         struct amd_northbridge *nb;
1362         struct threshold_bank *b;
1363
1364         b = per_cpu(threshold_banks, cpu)[bank];
1365         if (!b)
1366                 return;
1367
1368         if (!b->blocks)
1369                 goto free_out;
1370
1371         if (is_shared_bank(bank)) {
1372                 if (!refcount_dec_and_test(&b->cpus)) {
1373                         __threshold_remove_blocks(b);
1374                         per_cpu(threshold_banks, cpu)[bank] = NULL;
1375                         return;
1376                 } else {
1377                         /*
1378                          * the last CPU on this node using the shared bank is
1379                          * going away, remove that bank now.
1380                          */
1381                         nb = node_to_amd_nb(amd_get_nb_id(cpu));
1382                         nb->bank4 = NULL;
1383                 }
1384         }
1385
1386         deallocate_threshold_block(cpu, bank);
1387
1388 free_out:
1389         kobject_del(b->kobj);
1390         kobject_put(b->kobj);
1391         kfree(b);
1392         per_cpu(threshold_banks, cpu)[bank] = NULL;
1393 }
1394
1395 int mce_threshold_remove_device(unsigned int cpu)
1396 {
1397         unsigned int bank;
1398
1399         for (bank = 0; bank < mca_cfg.banks; ++bank) {
1400                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1401                         continue;
1402                 threshold_remove_bank(cpu, bank);
1403         }
1404         kfree(per_cpu(threshold_banks, cpu));
1405         per_cpu(threshold_banks, cpu) = NULL;
1406         return 0;
1407 }
1408
1409 /* create dir/files for all valid threshold banks */
1410 int mce_threshold_create_device(unsigned int cpu)
1411 {
1412         unsigned int bank;
1413         struct threshold_bank **bp;
1414         int err = 0;
1415
1416         bp = per_cpu(threshold_banks, cpu);
1417         if (bp)
1418                 return 0;
1419
1420         bp = kzalloc(sizeof(struct threshold_bank *) * mca_cfg.banks,
1421                      GFP_KERNEL);
1422         if (!bp)
1423                 return -ENOMEM;
1424
1425         per_cpu(threshold_banks, cpu) = bp;
1426
1427         for (bank = 0; bank < mca_cfg.banks; ++bank) {
1428                 if (!(per_cpu(bank_map, cpu) & (1 << bank)))
1429                         continue;
1430                 err = threshold_create_bank(cpu, bank);
1431                 if (err)
1432                         goto err;
1433         }
1434         return err;
1435 err:
1436         mce_threshold_remove_device(cpu);
1437         return err;
1438 }
1439
1440 static __init int threshold_init_device(void)
1441 {
1442         unsigned lcpu = 0;
1443
1444         /* to hit CPUs online before the notifier is up */
1445         for_each_online_cpu(lcpu) {
1446                 int err = mce_threshold_create_device(lcpu);
1447
1448                 if (err)
1449                         return err;
1450         }
1451
1452         if (thresholding_irq_en)
1453                 mce_threshold_vector = amd_threshold_interrupt;
1454
1455         return 0;
1456 }
1457 /*
1458  * there are 3 funcs which need to be _initcalled in a logic sequence:
1459  * 1. xen_late_init_mcelog
1460  * 2. mcheck_init_device
1461  * 3. threshold_init_device
1462  *
1463  * xen_late_init_mcelog must register xen_mce_chrdev_device before
1464  * native mce_chrdev_device registration if running under xen platform;
1465  *
1466  * mcheck_init_device should be inited before threshold_init_device to
1467  * initialize mce_device, otherwise a NULL ptr dereference will cause panic.
1468  *
1469  * so we use following _initcalls
1470  * 1. device_initcall(xen_late_init_mcelog);
1471  * 2. device_initcall_sync(mcheck_init_device);
1472  * 3. late_initcall(threshold_init_device);
1473  *
1474  * when running under xen, the initcall order is 1,2,3;
1475  * on baremetal, we skip 1 and we do only 2 and 3.
1476  */
1477 late_initcall(threshold_init_device);