drivers/edac/edac_mce_amd.c

   1 #include <linux/module.h>
   2 #include "edac_mce_amd.h"
   3
   4 static bool report_gart_errors;
   5 static void (*nb_bus_decoder)(int node_id, struct err_regs *regs);
   6
   7 void amd_report_gart_errors(bool v)
   8 {
   9         report_gart_errors = v;
  10 }
  11 EXPORT_SYMBOL_GPL(amd_report_gart_errors);
  12
  13 void amd_register_ecc_decoder(void (*f)(int, struct err_regs *))
  14 {
  15         nb_bus_decoder = f;
  16 }
  17 EXPORT_SYMBOL_GPL(amd_register_ecc_decoder);
  18
  19 void amd_unregister_ecc_decoder(void (*f)(int, struct err_regs *))
  20 {
  21         if (nb_bus_decoder) {
  22                 WARN_ON(nb_bus_decoder != f);
  23
  24                 nb_bus_decoder = NULL;
  25         }
  26 }
  27 EXPORT_SYMBOL_GPL(amd_unregister_ecc_decoder);
  28
  29 /*
  30  * string representation for the different MCA reported error types, see F3x48
  31  * or MSR0000_0411.
  32  */
  33 const char *tt_msgs[] = {        /* transaction type */
  34         "instruction",
  35         "data",
  36         "generic",
  37         "reserved"
  38 };
  39 EXPORT_SYMBOL_GPL(tt_msgs);
  40
  41 const char *ll_msgs[] = {       /* cache level */
  42         "L0",
  43         "L1",
  44         "L2",
  45         "L3/generic"
  46 };
  47 EXPORT_SYMBOL_GPL(ll_msgs);
  48
  49 const char *rrrr_msgs[] = {
  50         "generic",
  51         "generic read",
  52         "generic write",
  53         "data read",
  54         "data write",
  55         "inst fetch",
  56         "prefetch",
  57         "evict",
  58         "snoop",
  59         "reserved RRRR= 9",
  60         "reserved RRRR= 10",
  61         "reserved RRRR= 11",
  62         "reserved RRRR= 12",
  63         "reserved RRRR= 13",
  64         "reserved RRRR= 14",
  65         "reserved RRRR= 15"
  66 };
  67 EXPORT_SYMBOL_GPL(rrrr_msgs);
  68
  69 const char *pp_msgs[] = {       /* participating processor */
  70         "local node originated (SRC)",
  71         "local node responded to request (RES)",
  72         "local node observed as 3rd party (OBS)",
  73         "generic"
  74 };
  75 EXPORT_SYMBOL_GPL(pp_msgs);
  76
  77 const char *to_msgs[] = {
  78         "no timeout",
  79         "timed out"
  80 };
  81 EXPORT_SYMBOL_GPL(to_msgs);
  82
  83 const char *ii_msgs[] = {       /* memory or i/o */
  84         "mem access",
  85         "reserved",
  86         "i/o access",
  87         "generic"
  88 };
  89 EXPORT_SYMBOL_GPL(ii_msgs);
  90
  91 /*
  92  * Map the 4 or 5 (family-specific) bits of Extended Error code to the
  93  * string table.
  94  */
  95 const char *ext_msgs[] = {
  96         "K8 ECC error",                                 /* 0_0000b */
  97         "CRC error on link",                            /* 0_0001b */
  98         "Sync error packets on link",                   /* 0_0010b */
  99         "Master Abort during link operation",           /* 0_0011b */
 100         "Target Abort during link operation",           /* 0_0100b */
 101         "Invalid GART PTE entry during table walk",     /* 0_0101b */
 102         "Unsupported atomic RMW command received",      /* 0_0110b */
 103         "WDT error: NB transaction timeout",            /* 0_0111b */
 104         "ECC/ChipKill ECC error",                       /* 0_1000b */
 105         "SVM DEV Error",                                /* 0_1001b */
 106         "Link Data error",                              /* 0_1010b */
 107         "Link/L3/Probe Filter Protocol error",          /* 0_1011b */
 108         "NB Internal Arrays Parity error",              /* 0_1100b */
 109         "DRAM Address/Control Parity error",            /* 0_1101b */
 110         "Link Transmission error",                      /* 0_1110b */
 111         "GART/DEV Table Walk Data error"                /* 0_1111b */
 112         "Res 0x100 error",                              /* 1_0000b */
 113         "Res 0x101 error",                              /* 1_0001b */
 114         "Res 0x102 error",                              /* 1_0010b */
 115         "Res 0x103 error",                              /* 1_0011b */
 116         "Res 0x104 error",                              /* 1_0100b */
 117         "Res 0x105 error",                              /* 1_0101b */
 118         "Res 0x106 error",                              /* 1_0110b */
 119         "Res 0x107 error",                              /* 1_0111b */
 120         "Res 0x108 error",                              /* 1_1000b */
 121         "Res 0x109 error",                              /* 1_1001b */
 122         "Res 0x10A error",                              /* 1_1010b */
 123         "Res 0x10B error",                              /* 1_1011b */
 124         "ECC error in L3 Cache Data",                   /* 1_1100b */
 125         "L3 Cache Tag error",                           /* 1_1101b */
 126         "L3 Cache LRU Parity error",                    /* 1_1110b */
 127         "Probe Filter error"                            /* 1_1111b */
 128 };
 129 EXPORT_SYMBOL_GPL(ext_msgs);
 130
 131 static void amd_decode_dc_mce(u64 mc0_status)
 132 {
 133         u32 ec  = mc0_status & 0xffff;
 134         u32 xec = (mc0_status >> 16) & 0xf;
 135
 136         pr_emerg(" Data Cache Error");
 137
 138         if (xec == 1 && TLB_ERROR(ec))
 139                 pr_cont(": %s TLB multimatch.\n", LL_MSG(ec));
 140         else if (xec == 0) {
 141                 if (mc0_status & (1ULL << 40))
 142                         pr_cont(" during Data Scrub.\n");
 143                 else if (TLB_ERROR(ec))
 144                         pr_cont(": %s TLB parity error.\n", LL_MSG(ec));
 145                 else if (MEM_ERROR(ec)) {
 146                         u8 ll   = ec & 0x3;
 147                         u8 tt   = (ec >> 2) & 0x3;
 148                         u8 rrrr = (ec >> 4) & 0xf;
 149
 150                         /* see F10h BKDG (31116), Table 92. */
 151                         if (ll == 0x1) {
 152                                 if (tt != 0x1)
 153                                         goto wrong_dc_mce;
 154
 155                                 pr_cont(": Data/Tag %s error.\n", RRRR_MSG(ec));
 156
 157                         } else if (ll == 0x2 && rrrr == 0x3)
 158                                 pr_cont(" during L1 linefill from L2.\n");
 159                         else
 160                                 goto wrong_dc_mce;
 161                 } else if (BUS_ERROR(ec) && boot_cpu_data.x86 == 0xf)
 162                         pr_cont(" during system linefill.\n");
 163                 else
 164                         goto wrong_dc_mce;
 165         } else
 166                 goto wrong_dc_mce;
 167
 168         return;
 169
 170 wrong_dc_mce:
 171         pr_warning("Corrupted DC MCE info?\n");
 172 }
 173
 174 static void amd_decode_ic_mce(u64 mc1_status)
 175 {
 176         u32 ec  = mc1_status & 0xffff;
 177         u32 xec = (mc1_status >> 16) & 0xf;
 178
 179         pr_emerg(" Instruction Cache Error");
 180
 181         if (xec == 1 && TLB_ERROR(ec))
 182                 pr_cont(": %s TLB multimatch.\n", LL_MSG(ec));
 183         else if (xec == 0) {
 184                 if (TLB_ERROR(ec))
 185                         pr_cont(": %s TLB Parity error.\n", LL_MSG(ec));
 186                 else if (BUS_ERROR(ec)) {
 187                         if (boot_cpu_data.x86 == 0xf &&
 188                             (mc1_status & (1ULL << 58)))
 189                                 pr_cont(" during system linefill.\n");
 190                         else
 191                                 pr_cont(" during attempted NB data read.\n");
 192                 } else if (MEM_ERROR(ec)) {
 193                         u8 ll   = ec & 0x3;
 194                         u8 rrrr = (ec >> 4) & 0xf;
 195
 196                         if (ll == 0x2)
 197                                 pr_cont(" during a linefill from L2.\n");
 198                         else if (ll == 0x1) {
 199
 200                                 switch (rrrr) {
 201                                 case 0x5:
 202                                         pr_cont(": Parity error during "
 203                                                "data load.\n");
 204                                         break;
 205
 206                                 case 0x7:
 207                                         pr_cont(": Copyback Parity/Victim"
 208                                                 " error.\n");
 209                                         break;
 210
 211                                 case 0x8:
 212                                         pr_cont(": Tag Snoop error.\n");
 213                                         break;
 214
 215                                 default:
 216                                         goto wrong_ic_mce;
 217                                         break;
 218                                 }
 219                         }
 220                 } else
 221                         goto wrong_ic_mce;
 222         } else
 223                 goto wrong_ic_mce;
 224
 225         return;
 226
 227 wrong_ic_mce:
 228         pr_warning("Corrupted IC MCE info?\n");
 229 }
 230
 231 static void amd_decode_bu_mce(u64 mc2_status)
 232 {
 233         u32 ec = mc2_status & 0xffff;
 234         u32 xec = (mc2_status >> 16) & 0xf;
 235
 236         pr_emerg(" Bus Unit Error");
 237
 238         if (xec == 0x1)
 239                 pr_cont(" in the write data buffers.\n");
 240         else if (xec == 0x3)
 241                 pr_cont(" in the victim data buffers.\n");
 242         else if (xec == 0x2 && MEM_ERROR(ec))
 243                 pr_cont(": %s error in the L2 cache tags.\n", RRRR_MSG(ec));
 244         else if (xec == 0x0) {
 245                 if (TLB_ERROR(ec))
 246                         pr_cont(": %s error in a Page Descriptor Cache or "
 247                                 "Guest TLB.\n", TT_MSG(ec));
 248                 else if (BUS_ERROR(ec))
 249                         pr_cont(": %s/ECC error in data read from NB: %s.\n",
 250                                 RRRR_MSG(ec), PP_MSG(ec));
 251                 else if (MEM_ERROR(ec)) {
 252                         u8 rrrr = (ec >> 4) & 0xf;
 253
 254                         if (rrrr >= 0x7)
 255                                 pr_cont(": %s error during data copyback.\n",
 256                                         RRRR_MSG(ec));
 257                         else if (rrrr <= 0x1)
 258                                 pr_cont(": %s parity/ECC error during data "
 259                                         "access from L2.\n", RRRR_MSG(ec));
 260                         else
 261                                 goto wrong_bu_mce;
 262                 } else
 263                         goto wrong_bu_mce;
 264         } else
 265                 goto wrong_bu_mce;
 266
 267         return;
 268
 269 wrong_bu_mce:
 270         pr_warning("Corrupted BU MCE info?\n");
 271 }
 272
 273 static void amd_decode_ls_mce(u64 mc3_status)
 274 {
 275         u32 ec  = mc3_status & 0xffff;
 276         u32 xec = (mc3_status >> 16) & 0xf;
 277
 278         pr_emerg(" Load Store Error");
 279
 280         if (xec == 0x0) {
 281                 u8 rrrr = (ec >> 4) & 0xf;
 282
 283                 if (!BUS_ERROR(ec) || (rrrr != 0x3 && rrrr != 0x4))
 284                         goto wrong_ls_mce;
 285
 286                 pr_cont(" during %s.\n", RRRR_MSG(ec));
 287         }
 288         return;
 289
 290 wrong_ls_mce:
 291         pr_warning("Corrupted LS MCE info?\n");
 292 }
 293
 294 void amd_decode_nb_mce(int node_id, struct err_regs *regs, int handle_errors)
 295 {
 296         u32 ec  = ERROR_CODE(regs->nbsl);
 297
 298         if (!handle_errors)
 299                 return;
 300
 301         /*
 302          * GART TLB error reporting is disabled by default. Bail out early.
 303          */
 304         if (TLB_ERROR(ec) && !report_gart_errors)
 305                 return;
 306
 307         pr_emerg(" Northbridge Error, node %d", node_id);
 308
 309         /*
 310          * F10h, revD can disable ErrCpu[3:0] so check that first and also the
 311          * value encoding has changed so interpret those differently
 312          */
 313         if ((boot_cpu_data.x86 == 0x10) &&
 314             (boot_cpu_data.x86_model > 7)) {
 315                 if (regs->nbsh & K8_NBSH_ERR_CPU_VAL)
 316                         pr_cont(", core: %u\n", (u8)(regs->nbsh & 0xf));
 317         } else {
 318                 u8 assoc_cpus = regs->nbsh & 0xf;
 319
 320                 if (assoc_cpus > 0)
 321                         pr_cont(", core: %d", fls(assoc_cpus) - 1);
 322
 323                 pr_cont("\n");
 324         }
 325
 326         pr_emerg("%s.\n", EXT_ERR_MSG(regs->nbsl));
 327
 328         if (BUS_ERROR(ec) && nb_bus_decoder)
 329                 nb_bus_decoder(node_id, regs);
 330 }
 331 EXPORT_SYMBOL_GPL(amd_decode_nb_mce);
 332
 333 static void amd_decode_fr_mce(u64 mc5_status)
 334 {
 335         /* we have only one error signature so match all fields at once. */
 336         if ((mc5_status & 0xffff) == 0x0f0f)
 337                 pr_emerg(" FR Error: CPU Watchdog timer expire.\n");
 338         else
 339                 pr_warning("Corrupted FR MCE info?\n");
 340 }
 341
 342 static inline void amd_decode_err_code(unsigned int ec)
 343 {
 344         if (TLB_ERROR(ec)) {
 345                 pr_emerg(" Transaction: %s, Cache Level %s\n",
 346                          TT_MSG(ec), LL_MSG(ec));
 347         } else if (MEM_ERROR(ec)) {
 348                 pr_emerg(" Transaction: %s, Type: %s, Cache Level: %s",
 349                          RRRR_MSG(ec), TT_MSG(ec), LL_MSG(ec));
 350         } else if (BUS_ERROR(ec)) {
 351                 pr_emerg(" Transaction type: %s(%s), %s, Cache Level: %s, "
 352                          "Participating Processor: %s\n",
 353                           RRRR_MSG(ec), II_MSG(ec), TO_MSG(ec), LL_MSG(ec),
 354                           PP_MSG(ec));
 355         } else
 356                 pr_warning("Huh? Unknown MCE error 0x%x\n", ec);
 357 }
 358
 359 static int amd_decode_mce(struct notifier_block *nb, unsigned long val,
 360                            void *data)
 361 {
 362         struct mce *m = (struct mce *)data;
 363         struct err_regs regs;
 364         int node, ecc;
 365
 366         pr_emerg("MC%d_STATUS: ", m->bank);
 367
 368         pr_cont("%sorrected error, report: %s, MiscV: %svalid, "
 369                  "CPU context corrupt: %s",
 370                  ((m->status & MCI_STATUS_UC) ? "Unc"  : "C"),
 371                  ((m->status & MCI_STATUS_EN) ? "yes"  : "no"),
 372                  ((m->status & MCI_STATUS_MISCV) ? ""  : "in"),
 373                  ((m->status & MCI_STATUS_PCC) ? "yes" : "no"));
 374
 375         /* do the two bits[14:13] together */
 376         ecc = (m->status >> 45) & 0x3;
 377         if (ecc)
 378                 pr_cont(", %sECC Error", ((ecc == 2) ? "C" : "U"));
 379
 380         pr_cont("\n");
 381
 382         switch (m->bank) {
 383         case 0:
 384                 amd_decode_dc_mce(m->status);
 385                 break;
 386
 387         case 1:
 388                 amd_decode_ic_mce(m->status);
 389                 break;
 390
 391         case 2:
 392                 amd_decode_bu_mce(m->status);
 393                 break;
 394
 395         case 3:
 396                 amd_decode_ls_mce(m->status);
 397                 break;
 398
 399         case 4:
 400                 regs.nbsl  = (u32) m->status;
 401                 regs.nbsh  = (u32)(m->status >> 32);
 402                 regs.nbeal = (u32) m->addr;
 403                 regs.nbeah = (u32)(m->addr >> 32);
 404                 node       = amd_get_nb_id(m->extcpu);
 405
 406                 amd_decode_nb_mce(node, &regs, 1);
 407                 break;
 408
 409         case 5:
 410                 amd_decode_fr_mce(m->status);
 411                 break;
 412
 413         default:
 414                 break;
 415         }
 416
 417         amd_decode_err_code(m->status & 0xffff);
 418
 419         return NOTIFY_STOP;
 420 }
 421
 422 static struct notifier_block amd_mce_dec_nb = {
 423         .notifier_call  = amd_decode_mce,
 424 };
 425
 426 static int __init mce_amd_init(void)
 427 {
 428         /*
 429          * We can decode MCEs for Opteron and later CPUs:
 430          */
 431         if ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) &&
 432             (boot_cpu_data.x86 >= 0xf))
 433                 atomic_notifier_chain_register(&x86_mce_decoder_chain, &amd_mce_dec_nb);
 434
 435         return 0;
 436 }
 437 early_initcall(mce_amd_init);
 438
 439 #ifdef MODULE
 440 static void __exit mce_amd_exit(void)
 441 {
 442         atomic_notifier_chain_unregister(&x86_mce_decoder_chain, &amd_mce_dec_nb);
 443 }
 444
 445 MODULE_DESCRIPTION("AMD MCE decoder");
 446 MODULE_ALIAS("edac-mce-amd");
 447 MODULE_LICENSE("GPL");
 448 module_exit(mce_amd_exit);
 449 #endif