arch/x86/kernel/cpu/perf_event_intel.c

   1 /*
   2  * Per core/cpu state
   3  *
   4  * Used to coordinate shared registers between HT threads or
   5  * among events on a single PMU.
   6  */
   7
   8 #include <linux/stddef.h>
   9 #include <linux/types.h>
  10 #include <linux/init.h>
  11 #include <linux/slab.h>
  12 #include <linux/export.h>
  13
  14 #include <asm/hardirq.h>
  15 #include <asm/apic.h>
  16
  17 #include "perf_event.h"
  18
  19 /*
  20  * Intel PerfMon, used on Core and later.
  21  */
  22 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly =
  23 {
  24   [PERF_COUNT_HW_CPU_CYCLES]            = 0x003c,
  25   [PERF_COUNT_HW_INSTRUCTIONS]          = 0x00c0,
  26   [PERF_COUNT_HW_CACHE_REFERENCES]      = 0x4f2e,
  27   [PERF_COUNT_HW_CACHE_MISSES]          = 0x412e,
  28   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]   = 0x00c4,
  29   [PERF_COUNT_HW_BRANCH_MISSES]         = 0x00c5,
  30   [PERF_COUNT_HW_BUS_CYCLES]            = 0x013c,
  31 };
  32
  33 static struct event_constraint intel_core_event_constraints[] __read_mostly =
  34 {
  35         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
  36         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
  37         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
  38         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
  39         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
  40         INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */
  41         EVENT_CONSTRAINT_END
  42 };
  43
  44 static struct event_constraint intel_core2_event_constraints[] __read_mostly =
  45 {
  46         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  47         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  48         /*
  49          * Core2 has Fixed Counter 2 listed as CPU_CLK_UNHALTED.REF and event
  50          * 0x013c as CPU_CLK_UNHALTED.BUS and specifies there is a fixed
  51          * ratio between these counters.
  52          */
  53         /* FIXED_EVENT_CONSTRAINT(0x013c, 2),  CPU_CLK_UNHALTED.REF */
  54         INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */
  55         INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */
  56         INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */
  57         INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */
  58         INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */
  59         INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */
  60         INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */
  61         INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */
  62         INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */
  63         INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */
  64         EVENT_CONSTRAINT_END
  65 };
  66
  67 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly =
  68 {
  69         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  70         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  71         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
  72         INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */
  73         INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */
  74         INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */
  75         INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */
  76         INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */
  77         INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */
  78         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
  79         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
  80         EVENT_CONSTRAINT_END
  81 };
  82
  83 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly =
  84 {
  85         INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
  86         EVENT_EXTRA_END
  87 };
  88
  89 static struct event_constraint intel_westmere_event_constraints[] __read_mostly =
  90 {
  91         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
  92         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
  93         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
  94         INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */
  95         INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */
  96         INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */
  97         INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */
  98         EVENT_CONSTRAINT_END
  99 };
 100
 101 static struct event_constraint intel_snb_event_constraints[] __read_mostly =
 102 {
 103         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 104         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 105         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
 106         INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */
 107         INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */
 108         INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */
 109         EVENT_CONSTRAINT_END
 110 };
 111
 112 static struct extra_reg intel_westmere_extra_regs[] __read_mostly =
 113 {
 114         INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0),
 115         INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1),
 116         EVENT_EXTRA_END
 117 };
 118
 119 static struct event_constraint intel_v1_event_constraints[] __read_mostly =
 120 {
 121         EVENT_CONSTRAINT_END
 122 };
 123
 124 static struct event_constraint intel_gen_event_constraints[] __read_mostly =
 125 {
 126         FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */
 127         FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */
 128         /* FIXED_EVENT_CONSTRAINT(0x013c, 2), CPU_CLK_UNHALTED.REF */
 129         EVENT_CONSTRAINT_END
 130 };
 131
 132 static struct extra_reg intel_snb_extra_regs[] __read_mostly = {
 133         INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0),
 134         INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1),
 135         EVENT_EXTRA_END
 136 };
 137
 138 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = {
 139         INTEL_EVENT_EXTRA_REG(0xb7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0),
 140         INTEL_EVENT_EXTRA_REG(0xbb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1),
 141         EVENT_EXTRA_END
 142 };
 143
 144 static u64 intel_pmu_event_map(int hw_event)
 145 {
 146         return intel_perfmon_event_map[hw_event];
 147 }
 148
 149 static __initconst const u64 snb_hw_cache_event_ids
 150                                 [PERF_COUNT_HW_CACHE_MAX]
 151                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 152                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 153 {
 154  [ C(L1D) ] = {
 155         [ C(OP_READ) ] = {
 156                 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS        */
 157                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPLACEMENT              */
 158         },
 159         [ C(OP_WRITE) ] = {
 160                 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES       */
 161                 [ C(RESULT_MISS)   ] = 0x0851, /* L1D.ALL_M_REPLACEMENT        */
 162         },
 163         [ C(OP_PREFETCH) ] = {
 164                 [ C(RESULT_ACCESS) ] = 0x0,
 165                 [ C(RESULT_MISS)   ] = 0x024e, /* HW_PRE_REQ.DL1_MISS          */
 166         },
 167  },
 168  [ C(L1I ) ] = {
 169         [ C(OP_READ) ] = {
 170                 [ C(RESULT_ACCESS) ] = 0x0,
 171                 [ C(RESULT_MISS)   ] = 0x0280, /* ICACHE.MISSES */
 172         },
 173         [ C(OP_WRITE) ] = {
 174                 [ C(RESULT_ACCESS) ] = -1,
 175                 [ C(RESULT_MISS)   ] = -1,
 176         },
 177         [ C(OP_PREFETCH) ] = {
 178                 [ C(RESULT_ACCESS) ] = 0x0,
 179                 [ C(RESULT_MISS)   ] = 0x0,
 180         },
 181  },
 182  [ C(LL  ) ] = {
 183         [ C(OP_READ) ] = {
 184                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
 185                 [ C(RESULT_ACCESS) ] = 0x01b7,
 186                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
 187                 [ C(RESULT_MISS)   ] = 0x01b7,
 188         },
 189         [ C(OP_WRITE) ] = {
 190                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
 191                 [ C(RESULT_ACCESS) ] = 0x01b7,
 192                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
 193                 [ C(RESULT_MISS)   ] = 0x01b7,
 194         },
 195         [ C(OP_PREFETCH) ] = {
 196                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
 197                 [ C(RESULT_ACCESS) ] = 0x01b7,
 198                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
 199                 [ C(RESULT_MISS)   ] = 0x01b7,
 200         },
 201  },
 202  [ C(DTLB) ] = {
 203         [ C(OP_READ) ] = {
 204                 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */
 205                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */
 206         },
 207         [ C(OP_WRITE) ] = {
 208                 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */
 209                 [ C(RESULT_MISS)   ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */
 210         },
 211         [ C(OP_PREFETCH) ] = {
 212                 [ C(RESULT_ACCESS) ] = 0x0,
 213                 [ C(RESULT_MISS)   ] = 0x0,
 214         },
 215  },
 216  [ C(ITLB) ] = {
 217         [ C(OP_READ) ] = {
 218                 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT         */
 219                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK    */
 220         },
 221         [ C(OP_WRITE) ] = {
 222                 [ C(RESULT_ACCESS) ] = -1,
 223                 [ C(RESULT_MISS)   ] = -1,
 224         },
 225         [ C(OP_PREFETCH) ] = {
 226                 [ C(RESULT_ACCESS) ] = -1,
 227                 [ C(RESULT_MISS)   ] = -1,
 228         },
 229  },
 230  [ C(BPU ) ] = {
 231         [ C(OP_READ) ] = {
 232                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
 233                 [ C(RESULT_MISS)   ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */
 234         },
 235         [ C(OP_WRITE) ] = {
 236                 [ C(RESULT_ACCESS) ] = -1,
 237                 [ C(RESULT_MISS)   ] = -1,
 238         },
 239         [ C(OP_PREFETCH) ] = {
 240                 [ C(RESULT_ACCESS) ] = -1,
 241                 [ C(RESULT_MISS)   ] = -1,
 242         },
 243  },
 244  [ C(NODE) ] = {
 245         [ C(OP_READ) ] = {
 246                 [ C(RESULT_ACCESS) ] = -1,
 247                 [ C(RESULT_MISS)   ] = -1,
 248         },
 249         [ C(OP_WRITE) ] = {
 250                 [ C(RESULT_ACCESS) ] = -1,
 251                 [ C(RESULT_MISS)   ] = -1,
 252         },
 253         [ C(OP_PREFETCH) ] = {
 254                 [ C(RESULT_ACCESS) ] = -1,
 255                 [ C(RESULT_MISS)   ] = -1,
 256         },
 257  },
 258
 259 };
 260
 261 static __initconst const u64 westmere_hw_cache_event_ids
 262                                 [PERF_COUNT_HW_CACHE_MAX]
 263                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 264                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 265 {
 266  [ C(L1D) ] = {
 267         [ C(OP_READ) ] = {
 268                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
 269                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
 270         },
 271         [ C(OP_WRITE) ] = {
 272                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
 273                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
 274         },
 275         [ C(OP_PREFETCH) ] = {
 276                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
 277                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
 278         },
 279  },
 280  [ C(L1I ) ] = {
 281         [ C(OP_READ) ] = {
 282                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
 283                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
 284         },
 285         [ C(OP_WRITE) ] = {
 286                 [ C(RESULT_ACCESS) ] = -1,
 287                 [ C(RESULT_MISS)   ] = -1,
 288         },
 289         [ C(OP_PREFETCH) ] = {
 290                 [ C(RESULT_ACCESS) ] = 0x0,
 291                 [ C(RESULT_MISS)   ] = 0x0,
 292         },
 293  },
 294  [ C(LL  ) ] = {
 295         [ C(OP_READ) ] = {
 296                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
 297                 [ C(RESULT_ACCESS) ] = 0x01b7,
 298                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
 299                 [ C(RESULT_MISS)   ] = 0x01b7,
 300         },
 301         /*
 302          * Use RFO, not WRITEBACK, because a write miss would typically occur
 303          * on RFO.
 304          */
 305         [ C(OP_WRITE) ] = {
 306                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
 307                 [ C(RESULT_ACCESS) ] = 0x01b7,
 308                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
 309                 [ C(RESULT_MISS)   ] = 0x01b7,
 310         },
 311         [ C(OP_PREFETCH) ] = {
 312                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
 313                 [ C(RESULT_ACCESS) ] = 0x01b7,
 314                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
 315                 [ C(RESULT_MISS)   ] = 0x01b7,
 316         },
 317  },
 318  [ C(DTLB) ] = {
 319         [ C(OP_READ) ] = {
 320                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
 321                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
 322         },
 323         [ C(OP_WRITE) ] = {
 324                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
 325                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
 326         },
 327         [ C(OP_PREFETCH) ] = {
 328                 [ C(RESULT_ACCESS) ] = 0x0,
 329                 [ C(RESULT_MISS)   ] = 0x0,
 330         },
 331  },
 332  [ C(ITLB) ] = {
 333         [ C(OP_READ) ] = {
 334                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
 335                 [ C(RESULT_MISS)   ] = 0x0185, /* ITLB_MISSES.ANY              */
 336         },
 337         [ C(OP_WRITE) ] = {
 338                 [ C(RESULT_ACCESS) ] = -1,
 339                 [ C(RESULT_MISS)   ] = -1,
 340         },
 341         [ C(OP_PREFETCH) ] = {
 342                 [ C(RESULT_ACCESS) ] = -1,
 343                 [ C(RESULT_MISS)   ] = -1,
 344         },
 345  },
 346  [ C(BPU ) ] = {
 347         [ C(OP_READ) ] = {
 348                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
 349                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
 350         },
 351         [ C(OP_WRITE) ] = {
 352                 [ C(RESULT_ACCESS) ] = -1,
 353                 [ C(RESULT_MISS)   ] = -1,
 354         },
 355         [ C(OP_PREFETCH) ] = {
 356                 [ C(RESULT_ACCESS) ] = -1,
 357                 [ C(RESULT_MISS)   ] = -1,
 358         },
 359  },
 360  [ C(NODE) ] = {
 361         [ C(OP_READ) ] = {
 362                 [ C(RESULT_ACCESS) ] = 0x01b7,
 363                 [ C(RESULT_MISS)   ] = 0x01b7,
 364         },
 365         [ C(OP_WRITE) ] = {
 366                 [ C(RESULT_ACCESS) ] = 0x01b7,
 367                 [ C(RESULT_MISS)   ] = 0x01b7,
 368         },
 369         [ C(OP_PREFETCH) ] = {
 370                 [ C(RESULT_ACCESS) ] = 0x01b7,
 371                 [ C(RESULT_MISS)   ] = 0x01b7,
 372         },
 373  },
 374 };
 375
 376 /*
 377  * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits;
 378  * See IA32 SDM Vol 3B 30.6.1.3
 379  */
 380
 381 #define NHM_DMND_DATA_RD        (1 << 0)
 382 #define NHM_DMND_RFO            (1 << 1)
 383 #define NHM_DMND_IFETCH         (1 << 2)
 384 #define NHM_DMND_WB             (1 << 3)
 385 #define NHM_PF_DATA_RD          (1 << 4)
 386 #define NHM_PF_DATA_RFO         (1 << 5)
 387 #define NHM_PF_IFETCH           (1 << 6)
 388 #define NHM_OFFCORE_OTHER       (1 << 7)
 389 #define NHM_UNCORE_HIT          (1 << 8)
 390 #define NHM_OTHER_CORE_HIT_SNP  (1 << 9)
 391 #define NHM_OTHER_CORE_HITM     (1 << 10)
 392                                 /* reserved */
 393 #define NHM_REMOTE_CACHE_FWD    (1 << 12)
 394 #define NHM_REMOTE_DRAM         (1 << 13)
 395 #define NHM_LOCAL_DRAM          (1 << 14)
 396 #define NHM_NON_DRAM            (1 << 15)
 397
 398 #define NHM_LOCAL               (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD)
 399 #define NHM_REMOTE              (NHM_REMOTE_DRAM)
 400
 401 #define NHM_DMND_READ           (NHM_DMND_DATA_RD)
 402 #define NHM_DMND_WRITE          (NHM_DMND_RFO|NHM_DMND_WB)
 403 #define NHM_DMND_PREFETCH       (NHM_PF_DATA_RD|NHM_PF_DATA_RFO)
 404
 405 #define NHM_L3_HIT      (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM)
 406 #define NHM_L3_MISS     (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD)
 407 #define NHM_L3_ACCESS   (NHM_L3_HIT|NHM_L3_MISS)
 408
 409 static __initconst const u64 nehalem_hw_cache_extra_regs
 410                                 [PERF_COUNT_HW_CACHE_MAX]
 411                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 412                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 413 {
 414  [ C(LL  ) ] = {
 415         [ C(OP_READ) ] = {
 416                 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS,
 417                 [ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_L3_MISS,
 418         },
 419         [ C(OP_WRITE) ] = {
 420                 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS,
 421                 [ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_L3_MISS,
 422         },
 423         [ C(OP_PREFETCH) ] = {
 424                 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS,
 425                 [ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_L3_MISS,
 426         },
 427  },
 428  [ C(NODE) ] = {
 429         [ C(OP_READ) ] = {
 430                 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE,
 431                 [ C(RESULT_MISS)   ] = NHM_DMND_READ|NHM_REMOTE,
 432         },
 433         [ C(OP_WRITE) ] = {
 434                 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE,
 435                 [ C(RESULT_MISS)   ] = NHM_DMND_WRITE|NHM_REMOTE,
 436         },
 437         [ C(OP_PREFETCH) ] = {
 438                 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE,
 439                 [ C(RESULT_MISS)   ] = NHM_DMND_PREFETCH|NHM_REMOTE,
 440         },
 441  },
 442 };
 443
 444 static __initconst const u64 nehalem_hw_cache_event_ids
 445                                 [PERF_COUNT_HW_CACHE_MAX]
 446                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 447                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 448 {
 449  [ C(L1D) ] = {
 450         [ C(OP_READ) ] = {
 451                 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS       */
 452                 [ C(RESULT_MISS)   ] = 0x0151, /* L1D.REPL                     */
 453         },
 454         [ C(OP_WRITE) ] = {
 455                 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES      */
 456                 [ C(RESULT_MISS)   ] = 0x0251, /* L1D.M_REPL                   */
 457         },
 458         [ C(OP_PREFETCH) ] = {
 459                 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS        */
 460                 [ C(RESULT_MISS)   ] = 0x024e, /* L1D_PREFETCH.MISS            */
 461         },
 462  },
 463  [ C(L1I ) ] = {
 464         [ C(OP_READ) ] = {
 465                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                    */
 466                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                   */
 467         },
 468         [ C(OP_WRITE) ] = {
 469                 [ C(RESULT_ACCESS) ] = -1,
 470                 [ C(RESULT_MISS)   ] = -1,
 471         },
 472         [ C(OP_PREFETCH) ] = {
 473                 [ C(RESULT_ACCESS) ] = 0x0,
 474                 [ C(RESULT_MISS)   ] = 0x0,
 475         },
 476  },
 477  [ C(LL  ) ] = {
 478         [ C(OP_READ) ] = {
 479                 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */
 480                 [ C(RESULT_ACCESS) ] = 0x01b7,
 481                 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */
 482                 [ C(RESULT_MISS)   ] = 0x01b7,
 483         },
 484         /*
 485          * Use RFO, not WRITEBACK, because a write miss would typically occur
 486          * on RFO.
 487          */
 488         [ C(OP_WRITE) ] = {
 489                 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */
 490                 [ C(RESULT_ACCESS) ] = 0x01b7,
 491                 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */
 492                 [ C(RESULT_MISS)   ] = 0x01b7,
 493         },
 494         [ C(OP_PREFETCH) ] = {
 495                 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */
 496                 [ C(RESULT_ACCESS) ] = 0x01b7,
 497                 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */
 498                 [ C(RESULT_MISS)   ] = 0x01b7,
 499         },
 500  },
 501  [ C(DTLB) ] = {
 502         [ C(OP_READ) ] = {
 503                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI   (alias)  */
 504                 [ C(RESULT_MISS)   ] = 0x0108, /* DTLB_LOAD_MISSES.ANY         */
 505         },
 506         [ C(OP_WRITE) ] = {
 507                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI   (alias)  */
 508                 [ C(RESULT_MISS)   ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS  */
 509         },
 510         [ C(OP_PREFETCH) ] = {
 511                 [ C(RESULT_ACCESS) ] = 0x0,
 512                 [ C(RESULT_MISS)   ] = 0x0,
 513         },
 514  },
 515  [ C(ITLB) ] = {
 516         [ C(OP_READ) ] = {
 517                 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P           */
 518                 [ C(RESULT_MISS)   ] = 0x20c8, /* ITLB_MISS_RETIRED            */
 519         },
 520         [ C(OP_WRITE) ] = {
 521                 [ C(RESULT_ACCESS) ] = -1,
 522                 [ C(RESULT_MISS)   ] = -1,
 523         },
 524         [ C(OP_PREFETCH) ] = {
 525                 [ C(RESULT_ACCESS) ] = -1,
 526                 [ C(RESULT_MISS)   ] = -1,
 527         },
 528  },
 529  [ C(BPU ) ] = {
 530         [ C(OP_READ) ] = {
 531                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */
 532                 [ C(RESULT_MISS)   ] = 0x03e8, /* BPU_CLEARS.ANY               */
 533         },
 534         [ C(OP_WRITE) ] = {
 535                 [ C(RESULT_ACCESS) ] = -1,
 536                 [ C(RESULT_MISS)   ] = -1,
 537         },
 538         [ C(OP_PREFETCH) ] = {
 539                 [ C(RESULT_ACCESS) ] = -1,
 540                 [ C(RESULT_MISS)   ] = -1,
 541         },
 542  },
 543  [ C(NODE) ] = {
 544         [ C(OP_READ) ] = {
 545                 [ C(RESULT_ACCESS) ] = 0x01b7,
 546                 [ C(RESULT_MISS)   ] = 0x01b7,
 547         },
 548         [ C(OP_WRITE) ] = {
 549                 [ C(RESULT_ACCESS) ] = 0x01b7,
 550                 [ C(RESULT_MISS)   ] = 0x01b7,
 551         },
 552         [ C(OP_PREFETCH) ] = {
 553                 [ C(RESULT_ACCESS) ] = 0x01b7,
 554                 [ C(RESULT_MISS)   ] = 0x01b7,
 555         },
 556  },
 557 };
 558
 559 static __initconst const u64 core2_hw_cache_event_ids
 560                                 [PERF_COUNT_HW_CACHE_MAX]
 561                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 562                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 563 {
 564  [ C(L1D) ] = {
 565         [ C(OP_READ) ] = {
 566                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI          */
 567                 [ C(RESULT_MISS)   ] = 0x0140, /* L1D_CACHE_LD.I_STATE       */
 568         },
 569         [ C(OP_WRITE) ] = {
 570                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI          */
 571                 [ C(RESULT_MISS)   ] = 0x0141, /* L1D_CACHE_ST.I_STATE       */
 572         },
 573         [ C(OP_PREFETCH) ] = {
 574                 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS      */
 575                 [ C(RESULT_MISS)   ] = 0,
 576         },
 577  },
 578  [ C(L1I ) ] = {
 579         [ C(OP_READ) ] = {
 580                 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS                  */
 581                 [ C(RESULT_MISS)   ] = 0x0081, /* L1I.MISSES                 */
 582         },
 583         [ C(OP_WRITE) ] = {
 584                 [ C(RESULT_ACCESS) ] = -1,
 585                 [ C(RESULT_MISS)   ] = -1,
 586         },
 587         [ C(OP_PREFETCH) ] = {
 588                 [ C(RESULT_ACCESS) ] = 0,
 589                 [ C(RESULT_MISS)   ] = 0,
 590         },
 591  },
 592  [ C(LL  ) ] = {
 593         [ C(OP_READ) ] = {
 594                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
 595                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
 596         },
 597         [ C(OP_WRITE) ] = {
 598                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
 599                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
 600         },
 601         [ C(OP_PREFETCH) ] = {
 602                 [ C(RESULT_ACCESS) ] = 0,
 603                 [ C(RESULT_MISS)   ] = 0,
 604         },
 605  },
 606  [ C(DTLB) ] = {
 607         [ C(OP_READ) ] = {
 608                 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI  (alias) */
 609                 [ C(RESULT_MISS)   ] = 0x0208, /* DTLB_MISSES.MISS_LD        */
 610         },
 611         [ C(OP_WRITE) ] = {
 612                 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI  (alias) */
 613                 [ C(RESULT_MISS)   ] = 0x0808, /* DTLB_MISSES.MISS_ST        */
 614         },
 615         [ C(OP_PREFETCH) ] = {
 616                 [ C(RESULT_ACCESS) ] = 0,
 617                 [ C(RESULT_MISS)   ] = 0,
 618         },
 619  },
 620  [ C(ITLB) ] = {
 621         [ C(OP_READ) ] = {
 622                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
 623                 [ C(RESULT_MISS)   ] = 0x1282, /* ITLBMISSES                 */
 624         },
 625         [ C(OP_WRITE) ] = {
 626                 [ C(RESULT_ACCESS) ] = -1,
 627                 [ C(RESULT_MISS)   ] = -1,
 628         },
 629         [ C(OP_PREFETCH) ] = {
 630                 [ C(RESULT_ACCESS) ] = -1,
 631                 [ C(RESULT_MISS)   ] = -1,
 632         },
 633  },
 634  [ C(BPU ) ] = {
 635         [ C(OP_READ) ] = {
 636                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
 637                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
 638         },
 639         [ C(OP_WRITE) ] = {
 640                 [ C(RESULT_ACCESS) ] = -1,
 641                 [ C(RESULT_MISS)   ] = -1,
 642         },
 643         [ C(OP_PREFETCH) ] = {
 644                 [ C(RESULT_ACCESS) ] = -1,
 645                 [ C(RESULT_MISS)   ] = -1,
 646         },
 647  },
 648 };
 649
 650 static __initconst const u64 atom_hw_cache_event_ids
 651                                 [PERF_COUNT_HW_CACHE_MAX]
 652                                 [PERF_COUNT_HW_CACHE_OP_MAX]
 653                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
 654 {
 655  [ C(L1D) ] = {
 656         [ C(OP_READ) ] = {
 657                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD               */
 658                 [ C(RESULT_MISS)   ] = 0,
 659         },
 660         [ C(OP_WRITE) ] = {
 661                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST               */
 662                 [ C(RESULT_MISS)   ] = 0,
 663         },
 664         [ C(OP_PREFETCH) ] = {
 665                 [ C(RESULT_ACCESS) ] = 0x0,
 666                 [ C(RESULT_MISS)   ] = 0,
 667         },
 668  },
 669  [ C(L1I ) ] = {
 670         [ C(OP_READ) ] = {
 671                 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS                  */
 672                 [ C(RESULT_MISS)   ] = 0x0280, /* L1I.MISSES                 */
 673         },
 674         [ C(OP_WRITE) ] = {
 675                 [ C(RESULT_ACCESS) ] = -1,
 676                 [ C(RESULT_MISS)   ] = -1,
 677         },
 678         [ C(OP_PREFETCH) ] = {
 679                 [ C(RESULT_ACCESS) ] = 0,
 680                 [ C(RESULT_MISS)   ] = 0,
 681         },
 682  },
 683  [ C(LL  ) ] = {
 684         [ C(OP_READ) ] = {
 685                 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI                 */
 686                 [ C(RESULT_MISS)   ] = 0x4129, /* L2_LD.ISTATE               */
 687         },
 688         [ C(OP_WRITE) ] = {
 689                 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI                 */
 690                 [ C(RESULT_MISS)   ] = 0x412A, /* L2_ST.ISTATE               */
 691         },
 692         [ C(OP_PREFETCH) ] = {
 693                 [ C(RESULT_ACCESS) ] = 0,
 694                 [ C(RESULT_MISS)   ] = 0,
 695         },
 696  },
 697  [ C(DTLB) ] = {
 698         [ C(OP_READ) ] = {
 699                 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI  (alias) */
 700                 [ C(RESULT_MISS)   ] = 0x0508, /* DTLB_MISSES.MISS_LD        */
 701         },
 702         [ C(OP_WRITE) ] = {
 703                 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI  (alias) */
 704                 [ C(RESULT_MISS)   ] = 0x0608, /* DTLB_MISSES.MISS_ST        */
 705         },
 706         [ C(OP_PREFETCH) ] = {
 707                 [ C(RESULT_ACCESS) ] = 0,
 708                 [ C(RESULT_MISS)   ] = 0,
 709         },
 710  },
 711  [ C(ITLB) ] = {
 712         [ C(OP_READ) ] = {
 713                 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P         */
 714                 [ C(RESULT_MISS)   ] = 0x0282, /* ITLB.MISSES                */
 715         },
 716         [ C(OP_WRITE) ] = {
 717                 [ C(RESULT_ACCESS) ] = -1,
 718                 [ C(RESULT_MISS)   ] = -1,
 719         },
 720         [ C(OP_PREFETCH) ] = {
 721                 [ C(RESULT_ACCESS) ] = -1,
 722                 [ C(RESULT_MISS)   ] = -1,
 723         },
 724  },
 725  [ C(BPU ) ] = {
 726         [ C(OP_READ) ] = {
 727                 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY        */
 728                 [ C(RESULT_MISS)   ] = 0x00c5, /* BP_INST_RETIRED.MISPRED    */
 729         },
 730         [ C(OP_WRITE) ] = {
 731                 [ C(RESULT_ACCESS) ] = -1,
 732                 [ C(RESULT_MISS)   ] = -1,
 733         },
 734         [ C(OP_PREFETCH) ] = {
 735                 [ C(RESULT_ACCESS) ] = -1,
 736                 [ C(RESULT_MISS)   ] = -1,
 737         },
 738  },
 739 };
 740
 741 static void intel_pmu_disable_all(void)
 742 {
 743         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 744
 745         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
 746
 747         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask))
 748                 intel_pmu_disable_bts();
 749
 750         intel_pmu_pebs_disable_all();
 751         intel_pmu_lbr_disable_all();
 752 }
 753
 754 static void intel_pmu_enable_all(int added)
 755 {
 756         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 757
 758         intel_pmu_pebs_enable_all();
 759         intel_pmu_lbr_enable_all();
 760         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
 761                         x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask);
 762
 763         if (test_bit(X86_PMC_IDX_FIXED_BTS, cpuc->active_mask)) {
 764                 struct perf_event *event =
 765                         cpuc->events[X86_PMC_IDX_FIXED_BTS];
 766
 767                 if (WARN_ON_ONCE(!event))
 768                         return;
 769
 770                 intel_pmu_enable_bts(event->hw.config);
 771         }
 772 }
 773
 774 /*
 775  * Workaround for:
 776  *   Intel Errata AAK100 (model 26)
 777  *   Intel Errata AAP53  (model 30)
 778  *   Intel Errata BD53   (model 44)
 779  *
 780  * The official story:
 781  *   These chips need to be 'reset' when adding counters by programming the
 782  *   magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either
 783  *   in sequence on the same PMC or on different PMCs.
 784  *
 785  * In practise it appears some of these events do in fact count, and
 786  * we need to programm all 4 events.
 787  */
 788 static void intel_pmu_nhm_workaround(void)
 789 {
 790         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 791         static const unsigned long nhm_magic[4] = {
 792                 0x4300B5,
 793                 0x4300D2,
 794                 0x4300B1,
 795                 0x4300B1
 796         };
 797         struct perf_event *event;
 798         int i;
 799
 800         /*
 801          * The Errata requires below steps:
 802          * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL;
 803          * 2) Configure 4 PERFEVTSELx with the magic events and clear
 804          *    the corresponding PMCx;
 805          * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL;
 806          * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL;
 807          * 5) Clear 4 pairs of ERFEVTSELx and PMCx;
 808          */
 809
 810         /*
 811          * The real steps we choose are a little different from above.
 812          * A) To reduce MSR operations, we don't run step 1) as they
 813          *    are already cleared before this function is called;
 814          * B) Call x86_perf_event_update to save PMCx before configuring
 815          *    PERFEVTSELx with magic number;
 816          * C) With step 5), we do clear only when the PERFEVTSELx is
 817          *    not used currently.
 818          * D) Call x86_perf_event_set_period to restore PMCx;
 819          */
 820
 821         /* We always operate 4 pairs of PERF Counters */
 822         for (i = 0; i < 4; i++) {
 823                 event = cpuc->events[i];
 824                 if (event)
 825                         x86_perf_event_update(event);
 826         }
 827
 828         for (i = 0; i < 4; i++) {
 829                 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]);
 830                 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0);
 831         }
 832
 833         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf);
 834         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0);
 835
 836         for (i = 0; i < 4; i++) {
 837                 event = cpuc->events[i];
 838
 839                 if (event) {
 840                         x86_perf_event_set_period(event);
 841                         __x86_pmu_enable_event(&event->hw,
 842                                         ARCH_PERFMON_EVENTSEL_ENABLE);
 843                 } else
 844                         wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0);
 845         }
 846 }
 847
 848 static void intel_pmu_nhm_enable_all(int added)
 849 {
 850         if (added)
 851                 intel_pmu_nhm_workaround();
 852         intel_pmu_enable_all(added);
 853 }
 854
 855 static inline u64 intel_pmu_get_status(void)
 856 {
 857         u64 status;
 858
 859         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
 860
 861         return status;
 862 }
 863
 864 static inline void intel_pmu_ack_status(u64 ack)
 865 {
 866         wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack);
 867 }
 868
 869 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc)
 870 {
 871         int idx = hwc->idx - X86_PMC_IDX_FIXED;
 872         u64 ctrl_val, mask;
 873
 874         mask = 0xfULL << (idx * 4);
 875
 876         rdmsrl(hwc->config_base, ctrl_val);
 877         ctrl_val &= ~mask;
 878         wrmsrl(hwc->config_base, ctrl_val);
 879 }
 880
 881 static void intel_pmu_disable_event(struct perf_event *event)
 882 {
 883         struct hw_perf_event *hwc = &event->hw;
 884         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 885
 886         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
 887                 intel_pmu_disable_bts();
 888                 intel_pmu_drain_bts_buffer();
 889                 return;
 890         }
 891
 892         cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
 893         cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
 894
 895         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 896                 intel_pmu_disable_fixed(hwc);
 897                 return;
 898         }
 899
 900         x86_pmu_disable_event(event);
 901
 902         if (unlikely(event->attr.precise_ip))
 903                 intel_pmu_pebs_disable(event);
 904 }
 905
 906 static void intel_pmu_enable_fixed(struct hw_perf_event *hwc)
 907 {
 908         int idx = hwc->idx - X86_PMC_IDX_FIXED;
 909         u64 ctrl_val, bits, mask;
 910
 911         /*
 912          * Enable IRQ generation (0x8),
 913          * and enable ring-3 counting (0x2) and ring-0 counting (0x1)
 914          * if requested:
 915          */
 916         bits = 0x8ULL;
 917         if (hwc->config & ARCH_PERFMON_EVENTSEL_USR)
 918                 bits |= 0x2;
 919         if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
 920                 bits |= 0x1;
 921
 922         /*
 923          * ANY bit is supported in v3 and up
 924          */
 925         if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY)
 926                 bits |= 0x4;
 927
 928         bits <<= (idx * 4);
 929         mask = 0xfULL << (idx * 4);
 930
 931         rdmsrl(hwc->config_base, ctrl_val);
 932         ctrl_val &= ~mask;
 933         ctrl_val |= bits;
 934         wrmsrl(hwc->config_base, ctrl_val);
 935 }
 936
 937 static void intel_pmu_enable_event(struct perf_event *event)
 938 {
 939         struct hw_perf_event *hwc = &event->hw;
 940         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 941
 942         if (unlikely(hwc->idx == X86_PMC_IDX_FIXED_BTS)) {
 943                 if (!__this_cpu_read(cpu_hw_events.enabled))
 944                         return;
 945
 946                 intel_pmu_enable_bts(hwc->config);
 947                 return;
 948         }
 949
 950         if (event->attr.exclude_host)
 951                 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
 952         if (event->attr.exclude_guest)
 953                 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx);
 954
 955         if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) {
 956                 intel_pmu_enable_fixed(hwc);
 957                 return;
 958         }
 959
 960         if (unlikely(event->attr.precise_ip))
 961                 intel_pmu_pebs_enable(event);
 962
 963         __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
 964 }
 965
 966 /*
 967  * Save and restart an expired event. Called by NMI contexts,
 968  * so it has to be careful about preempting normal event ops:
 969  */
 970 int intel_pmu_save_and_restart(struct perf_event *event)
 971 {
 972         x86_perf_event_update(event);
 973         return x86_perf_event_set_period(event);
 974 }
 975
 976 static void intel_pmu_reset(void)
 977 {
 978         struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds);
 979         unsigned long flags;
 980         int idx;
 981
 982         if (!x86_pmu.num_counters)
 983                 return;
 984
 985         local_irq_save(flags);
 986
 987         printk("clearing PMU state on CPU#%d\n", smp_processor_id());
 988
 989         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
 990                 checking_wrmsrl(x86_pmu_config_addr(idx), 0ull);
 991                 checking_wrmsrl(x86_pmu_event_addr(idx),  0ull);
 992         }
 993         for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++)
 994                 checking_wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull);
 995
 996         if (ds)
 997                 ds->bts_index = ds->bts_buffer_base;
 998
 999         local_irq_restore(flags);
1000 }
1001
1002 /*
1003  * This handler is triggered by the local APIC, so the APIC IRQ handling
1004  * rules apply:
1005  */
1006 static int intel_pmu_handle_irq(struct pt_regs *regs)
1007 {
1008         struct perf_sample_data data;
1009         struct cpu_hw_events *cpuc;
1010         int bit, loops;
1011         u64 status;
1012         int handled;
1013
1014         perf_sample_data_init(&data, 0);
1015
1016         cpuc = &__get_cpu_var(cpu_hw_events);
1017
1018         /*
1019          * Some chipsets need to unmask the LVTPC in a particular spot
1020          * inside the nmi handler.  As a result, the unmasking was pushed
1021          * into all the nmi handlers.
1022          *
1023          * This handler doesn't seem to have any issues with the unmasking
1024          * so it was left at the top.
1025          */
1026         apic_write(APIC_LVTPC, APIC_DM_NMI);
1027
1028         intel_pmu_disable_all();
1029         handled = intel_pmu_drain_bts_buffer();
1030         status = intel_pmu_get_status();
1031         if (!status) {
1032                 intel_pmu_enable_all(0);
1033                 return handled;
1034         }
1035
1036         loops = 0;
1037 again:
1038         intel_pmu_ack_status(status);
1039         if (++loops > 100) {
1040                 WARN_ONCE(1, "perfevents: irq loop stuck!\n");
1041                 perf_event_print_debug();
1042                 intel_pmu_reset();
1043                 goto done;
1044         }
1045
1046         inc_irq_stat(apic_perf_irqs);
1047
1048         intel_pmu_lbr_read();
1049
1050         /*
1051          * PEBS overflow sets bit 62 in the global status register
1052          */
1053         if (__test_and_clear_bit(62, (unsigned long *)&status)) {
1054                 handled++;
1055                 x86_pmu.drain_pebs(regs);
1056         }
1057
1058         for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
1059                 struct perf_event *event = cpuc->events[bit];
1060
1061                 handled++;
1062
1063                 if (!test_bit(bit, cpuc->active_mask))
1064                         continue;
1065
1066                 if (!intel_pmu_save_and_restart(event))
1067                         continue;
1068
1069                 data.period = event->hw.last_period;
1070
1071                 if (perf_event_overflow(event, &data, regs))
1072                         x86_pmu_stop(event, 0);
1073         }
1074
1075         /*
1076          * Repeat if there is more work to be done:
1077          */
1078         status = intel_pmu_get_status();
1079         if (status)
1080                 goto again;
1081
1082 done:
1083         intel_pmu_enable_all(0);
1084         return handled;
1085 }
1086
1087 static struct event_constraint *
1088 intel_bts_constraints(struct perf_event *event)
1089 {
1090         struct hw_perf_event *hwc = &event->hw;
1091         unsigned int hw_event, bts_event;
1092
1093         if (event->attr.freq)
1094                 return NULL;
1095
1096         hw_event = hwc->config & INTEL_ARCH_EVENT_MASK;
1097         bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
1098
1099         if (unlikely(hw_event == bts_event && hwc->sample_period == 1))
1100                 return &bts_constraint;
1101
1102         return NULL;
1103 }
1104
1105 static bool intel_try_alt_er(struct perf_event *event, int orig_idx)
1106 {
1107         if (!(x86_pmu.er_flags & ERF_HAS_RSP_1))
1108                 return false;
1109
1110         if (event->hw.extra_reg.idx == EXTRA_REG_RSP_0) {
1111                 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1112                 event->hw.config |= 0x01bb;
1113                 event->hw.extra_reg.idx = EXTRA_REG_RSP_1;
1114                 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
1115         } else if (event->hw.extra_reg.idx == EXTRA_REG_RSP_1) {
1116                 event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
1117                 event->hw.config |= 0x01b7;
1118                 event->hw.extra_reg.idx = EXTRA_REG_RSP_0;
1119                 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
1120         }
1121
1122         if (event->hw.extra_reg.idx == orig_idx)
1123                 return false;
1124
1125         return true;
1126 }
1127
1128 /*
1129  * manage allocation of shared extra msr for certain events
1130  *
1131  * sharing can be:
1132  * per-cpu: to be shared between the various events on a single PMU
1133  * per-core: per-cpu + shared by HT threads
1134  */
1135 static struct event_constraint *
1136 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
1137                                    struct perf_event *event)
1138 {
1139         struct event_constraint *c = &emptyconstraint;
1140         struct hw_perf_event_extra *reg = &event->hw.extra_reg;
1141         struct er_account *era;
1142         unsigned long flags;
1143         int orig_idx = reg->idx;
1144
1145         /* already allocated shared msr */
1146         if (reg->alloc)
1147                 return &unconstrained;
1148
1149 again:
1150         era = &cpuc->shared_regs->regs[reg->idx];
1151         /*
1152          * we use spin_lock_irqsave() to avoid lockdep issues when
1153          * passing a fake cpuc
1154          */
1155         raw_spin_lock_irqsave(&era->lock, flags);
1156
1157         if (!atomic_read(&era->ref) || era->config == reg->config) {
1158
1159                 /* lock in msr value */
1160                 era->config = reg->config;
1161                 era->reg = reg->reg;
1162
1163                 /* one more user */
1164                 atomic_inc(&era->ref);
1165
1166                 /* no need to reallocate during incremental event scheduling */
1167                 reg->alloc = 1;
1168
1169                 /*
1170                  * All events using extra_reg are unconstrained.
1171                  * Avoids calling x86_get_event_constraints()
1172                  *
1173                  * Must revisit if extra_reg controlling events
1174                  * ever have constraints. Worst case we go through
1175                  * the regular event constraint table.
1176                  */
1177                 c = &unconstrained;
1178         } else if (intel_try_alt_er(event, orig_idx)) {
1179                 raw_spin_unlock_irqrestore(&era->lock, flags);
1180                 goto again;
1181         }
1182         raw_spin_unlock_irqrestore(&era->lock, flags);
1183
1184         return c;
1185 }
1186
1187 static void
1188 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
1189                                    struct hw_perf_event_extra *reg)
1190 {
1191         struct er_account *era;
1192
1193         /*
1194          * only put constraint if extra reg was actually
1195          * allocated. Also takes care of event which do
1196          * not use an extra shared reg
1197          */
1198         if (!reg->alloc)
1199                 return;
1200
1201         era = &cpuc->shared_regs->regs[reg->idx];
1202
1203         /* one fewer user */
1204         atomic_dec(&era->ref);
1205
1206         /* allocate again next time */
1207         reg->alloc = 0;
1208 }
1209
1210 static struct event_constraint *
1211 intel_shared_regs_constraints(struct cpu_hw_events *cpuc,
1212                               struct perf_event *event)
1213 {
1214         struct event_constraint *c = NULL;
1215
1216         if (event->hw.extra_reg.idx != EXTRA_REG_NONE)
1217                 c = __intel_shared_reg_get_constraints(cpuc, event);
1218
1219         return c;
1220 }
1221
1222 struct event_constraint *
1223 x86_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1224 {
1225         struct event_constraint *c;
1226
1227         if (x86_pmu.event_constraints) {
1228                 for_each_event_constraint(c, x86_pmu.event_constraints) {
1229                         if ((event->hw.config & c->cmask) == c->code)
1230                                 return c;
1231                 }
1232         }
1233
1234         return &unconstrained;
1235 }
1236
1237 static struct event_constraint *
1238 intel_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
1239 {
1240         struct event_constraint *c;
1241
1242         c = intel_bts_constraints(event);
1243         if (c)
1244                 return c;
1245
1246         c = intel_pebs_constraints(event);
1247         if (c)
1248                 return c;
1249
1250         c = intel_shared_regs_constraints(cpuc, event);
1251         if (c)
1252                 return c;
1253
1254         return x86_get_event_constraints(cpuc, event);
1255 }
1256
1257 static void
1258 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc,
1259                                         struct perf_event *event)
1260 {
1261         struct hw_perf_event_extra *reg;
1262
1263         reg = &event->hw.extra_reg;
1264         if (reg->idx != EXTRA_REG_NONE)
1265                 __intel_shared_reg_put_constraints(cpuc, reg);
1266 }
1267
1268 static void intel_put_event_constraints(struct cpu_hw_events *cpuc,
1269                                         struct perf_event *event)
1270 {
1271         intel_put_shared_regs_event_constraints(cpuc, event);
1272 }
1273
1274 static int intel_pmu_hw_config(struct perf_event *event)
1275 {
1276         int ret = x86_pmu_hw_config(event);
1277
1278         if (ret)
1279                 return ret;
1280
1281         if (event->attr.precise_ip &&
1282             (event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) {
1283                 /*
1284                  * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P
1285                  * (0x003c) so that we can use it with PEBS.
1286                  *
1287                  * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't
1288                  * PEBS capable. However we can use INST_RETIRED.ANY_P
1289                  * (0x00c0), which is a PEBS capable event, to get the same
1290                  * count.
1291                  *
1292                  * INST_RETIRED.ANY_P counts the number of cycles that retires
1293                  * CNTMASK instructions. By setting CNTMASK to a value (16)
1294                  * larger than the maximum number of instructions that can be
1295                  * retired per cycle (4) and then inverting the condition, we
1296                  * count all cycles that retire 16 or less instructions, which
1297                  * is every cycle.
1298                  *
1299                  * Thereby we gain a PEBS capable cycle counter.
1300                  */
1301                 u64 alt_config = 0x108000c0; /* INST_RETIRED.TOTAL_CYCLES */
1302
1303                 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK);
1304                 event->hw.config = alt_config;
1305         }
1306
1307         if (event->attr.type != PERF_TYPE_RAW)
1308                 return 0;
1309
1310         if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
1311                 return 0;
1312
1313         if (x86_pmu.version < 3)
1314                 return -EINVAL;
1315
1316         if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
1317                 return -EACCES;
1318
1319         event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY;
1320
1321         return 0;
1322 }
1323
1324 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr)
1325 {
1326         if (x86_pmu.guest_get_msrs)
1327                 return x86_pmu.guest_get_msrs(nr);
1328         *nr = 0;
1329         return NULL;
1330 }
1331 EXPORT_SYMBOL_GPL(perf_guest_get_msrs);
1332
1333 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr)
1334 {
1335         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1336         struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1337
1338         arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL;
1339         arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask;
1340         arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask;
1341
1342         *nr = 1;
1343         return arr;
1344 }
1345
1346 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr)
1347 {
1348         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1349         struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs;
1350         int idx;
1351
1352         for (idx = 0; idx < x86_pmu.num_counters; idx++)  {
1353                 struct perf_event *event = cpuc->events[idx];
1354
1355                 arr[idx].msr = x86_pmu_config_addr(idx);
1356                 arr[idx].host = arr[idx].guest = 0;
1357
1358                 if (!test_bit(idx, cpuc->active_mask))
1359                         continue;
1360
1361                 arr[idx].host = arr[idx].guest =
1362                         event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE;
1363
1364                 if (event->attr.exclude_host)
1365                         arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1366                 else if (event->attr.exclude_guest)
1367                         arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE;
1368         }
1369
1370         *nr = x86_pmu.num_counters;
1371         return arr;
1372 }
1373
1374 static void core_pmu_enable_event(struct perf_event *event)
1375 {
1376         if (!event->attr.exclude_host)
1377                 x86_pmu_enable_event(event);
1378 }
1379
1380 static void core_pmu_enable_all(int added)
1381 {
1382         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
1383         int idx;
1384
1385         for (idx = 0; idx < x86_pmu.num_counters; idx++) {
1386                 struct hw_perf_event *hwc = &cpuc->events[idx]->hw;
1387
1388                 if (!test_bit(idx, cpuc->active_mask) ||
1389                                 cpuc->events[idx]->attr.exclude_host)
1390                         continue;
1391
1392                 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
1393         }
1394 }
1395
1396 static __initconst const struct x86_pmu core_pmu = {
1397         .name                   = "core",
1398         .handle_irq             = x86_pmu_handle_irq,
1399         .disable_all            = x86_pmu_disable_all,
1400         .enable_all             = core_pmu_enable_all,
1401         .enable                 = core_pmu_enable_event,
1402         .disable                = x86_pmu_disable_event,
1403         .hw_config              = x86_pmu_hw_config,
1404         .schedule_events        = x86_schedule_events,
1405         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
1406         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
1407         .event_map              = intel_pmu_event_map,
1408         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
1409         .apic                   = 1,
1410         /*
1411          * Intel PMCs cannot be accessed sanely above 32 bit width,
1412          * so we install an artificial 1<<31 period regardless of
1413          * the generic event period:
1414          */
1415         .max_period             = (1ULL << 31) - 1,
1416         .get_event_constraints  = intel_get_event_constraints,
1417         .put_event_constraints  = intel_put_event_constraints,
1418         .event_constraints      = intel_core_event_constraints,
1419         .guest_get_msrs         = core_guest_get_msrs,
1420 };
1421
1422 struct intel_shared_regs *allocate_shared_regs(int cpu)
1423 {
1424         struct intel_shared_regs *regs;
1425         int i;
1426
1427         regs = kzalloc_node(sizeof(struct intel_shared_regs),
1428                             GFP_KERNEL, cpu_to_node(cpu));
1429         if (regs) {
1430                 /*
1431                  * initialize the locks to keep lockdep happy
1432                  */
1433                 for (i = 0; i < EXTRA_REG_MAX; i++)
1434                         raw_spin_lock_init(&regs->regs[i].lock);
1435
1436                 regs->core_id = -1;
1437         }
1438         return regs;
1439 }
1440
1441 static int intel_pmu_cpu_prepare(int cpu)
1442 {
1443         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1444
1445         if (!x86_pmu.extra_regs)
1446                 return NOTIFY_OK;
1447
1448         cpuc->shared_regs = allocate_shared_regs(cpu);
1449         if (!cpuc->shared_regs)
1450                 return NOTIFY_BAD;
1451
1452         return NOTIFY_OK;
1453 }
1454
1455 static void intel_pmu_cpu_starting(int cpu)
1456 {
1457         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1458         int core_id = topology_core_id(cpu);
1459         int i;
1460
1461         init_debug_store_on_cpu(cpu);
1462         /*
1463          * Deal with CPUs that don't clear their LBRs on power-up.
1464          */
1465         intel_pmu_lbr_reset();
1466
1467         if (!cpuc->shared_regs || (x86_pmu.er_flags & ERF_NO_HT_SHARING))
1468                 return;
1469
1470         for_each_cpu(i, topology_thread_cpumask(cpu)) {
1471                 struct intel_shared_regs *pc;
1472
1473                 pc = per_cpu(cpu_hw_events, i).shared_regs;
1474                 if (pc && pc->core_id == core_id) {
1475                         cpuc->kfree_on_online = cpuc->shared_regs;
1476                         cpuc->shared_regs = pc;
1477                         break;
1478                 }
1479         }
1480
1481         cpuc->shared_regs->core_id = core_id;
1482         cpuc->shared_regs->refcnt++;
1483 }
1484
1485 static void intel_pmu_cpu_dying(int cpu)
1486 {
1487         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
1488         struct intel_shared_regs *pc;
1489
1490         pc = cpuc->shared_regs;
1491         if (pc) {
1492                 if (pc->core_id == -1 || --pc->refcnt == 0)
1493                         kfree(pc);
1494                 cpuc->shared_regs = NULL;
1495         }
1496
1497         fini_debug_store_on_cpu(cpu);
1498 }
1499
1500 static __initconst const struct x86_pmu intel_pmu = {
1501         .name                   = "Intel",
1502         .handle_irq             = intel_pmu_handle_irq,
1503         .disable_all            = intel_pmu_disable_all,
1504         .enable_all             = intel_pmu_enable_all,
1505         .enable                 = intel_pmu_enable_event,
1506         .disable                = intel_pmu_disable_event,
1507         .hw_config              = intel_pmu_hw_config,
1508         .schedule_events        = x86_schedule_events,
1509         .eventsel               = MSR_ARCH_PERFMON_EVENTSEL0,
1510         .perfctr                = MSR_ARCH_PERFMON_PERFCTR0,
1511         .event_map              = intel_pmu_event_map,
1512         .max_events             = ARRAY_SIZE(intel_perfmon_event_map),
1513         .apic                   = 1,
1514         /*
1515          * Intel PMCs cannot be accessed sanely above 32 bit width,
1516          * so we install an artificial 1<<31 period regardless of
1517          * the generic event period:
1518          */
1519         .max_period             = (1ULL << 31) - 1,
1520         .get_event_constraints  = intel_get_event_constraints,
1521         .put_event_constraints  = intel_put_event_constraints,
1522
1523         .cpu_prepare            = intel_pmu_cpu_prepare,
1524         .cpu_starting           = intel_pmu_cpu_starting,
1525         .cpu_dying              = intel_pmu_cpu_dying,
1526         .guest_get_msrs         = intel_guest_get_msrs,
1527 };
1528
1529 static void intel_clovertown_quirks(void)
1530 {
1531         /*
1532          * PEBS is unreliable due to:
1533          *
1534          *   AJ67  - PEBS may experience CPL leaks
1535          *   AJ68  - PEBS PMI may be delayed by one event
1536          *   AJ69  - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12]
1537          *   AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS
1538          *
1539          * AJ67 could be worked around by restricting the OS/USR flags.
1540          * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI.
1541          *
1542          * AJ106 could possibly be worked around by not allowing LBR
1543          *       usage from PEBS, including the fixup.
1544          * AJ68  could possibly be worked around by always programming
1545          *       a pebs_event_reset[0] value and coping with the lost events.
1546          *
1547          * But taken together it might just make sense to not enable PEBS on
1548          * these chips.
1549          */
1550         printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
1551         x86_pmu.pebs = 0;
1552         x86_pmu.pebs_constraints = NULL;
1553 }
1554
1555 static void intel_sandybridge_quirks(void)
1556 {
1557         printk(KERN_WARNING "PEBS disabled due to CPU errata.\n");
1558         x86_pmu.pebs = 0;
1559         x86_pmu.pebs_constraints = NULL;
1560 }
1561
1562 __init int intel_pmu_init(void)
1563 {
1564         union cpuid10_edx edx;
1565         union cpuid10_eax eax;
1566         unsigned int unused;
1567         unsigned int ebx;
1568         int version;
1569
1570         if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) {
1571                 switch (boot_cpu_data.x86) {
1572                 case 0x6:
1573                         return p6_pmu_init();
1574                 case 0xf:
1575                         return p4_pmu_init();
1576                 }
1577                 return -ENODEV;
1578         }
1579
1580         /*
1581          * Check whether the Architectural PerfMon supports
1582          * Branch Misses Retired hw_event or not.
1583          */
1584         cpuid(10, &eax.full, &ebx, &unused, &edx.full);
1585         if (eax.split.mask_length <= ARCH_PERFMON_BRANCH_MISSES_RETIRED)
1586                 return -ENODEV;
1587
1588         version = eax.split.version_id;
1589         if (version < 2)
1590                 x86_pmu = core_pmu;
1591         else
1592                 x86_pmu = intel_pmu;
1593
1594         x86_pmu.version                 = version;
1595         x86_pmu.num_counters            = eax.split.num_counters;
1596         x86_pmu.cntval_bits             = eax.split.bit_width;
1597         x86_pmu.cntval_mask             = (1ULL << eax.split.bit_width) - 1;
1598
1599         /*
1600          * Quirk: v2 perfmon does not report fixed-purpose events, so
1601          * assume at least 3 events:
1602          */
1603         if (version > 1)
1604                 x86_pmu.num_counters_fixed = max((int)edx.split.num_counters_fixed, 3);
1605
1606         /*
1607          * v2 and above have a perf capabilities MSR
1608          */
1609         if (version > 1) {
1610                 u64 capabilities;
1611
1612                 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities);
1613                 x86_pmu.intel_cap.capabilities = capabilities;
1614         }
1615
1616         intel_ds_init();
1617
1618         /*
1619          * Install the hw-cache-events table:
1620          */
1621         switch (boot_cpu_data.x86_model) {
1622         case 14: /* 65 nm core solo/duo, "Yonah" */
1623                 pr_cont("Core events, ");
1624                 break;
1625
1626         case 15: /* original 65 nm celeron/pentium/core2/xeon, "Merom"/"Conroe" */
1627                 x86_pmu.quirks = intel_clovertown_quirks;
1628         case 22: /* single-core 65 nm celeron/core2solo "Merom-L"/"Conroe-L" */
1629         case 23: /* current 45 nm celeron/core2/xeon "Penryn"/"Wolfdale" */
1630         case 29: /* six-core 45 nm xeon "Dunnington" */
1631                 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids,
1632                        sizeof(hw_cache_event_ids));
1633
1634                 intel_pmu_lbr_init_core();
1635
1636                 x86_pmu.event_constraints = intel_core2_event_constraints;
1637                 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints;
1638                 pr_cont("Core2 events, ");
1639                 break;
1640
1641         case 26: /* 45 nm nehalem, "Bloomfield" */
1642         case 30: /* 45 nm nehalem, "Lynnfield" */
1643         case 46: /* 45 nm nehalem-ex, "Beckton" */
1644                 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids,
1645                        sizeof(hw_cache_event_ids));
1646                 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1647                        sizeof(hw_cache_extra_regs));
1648
1649                 intel_pmu_lbr_init_nhm();
1650
1651                 x86_pmu.event_constraints = intel_nehalem_event_constraints;
1652                 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints;
1653                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
1654                 x86_pmu.extra_regs = intel_nehalem_extra_regs;
1655
1656                 /* UOPS_ISSUED.STALLED_CYCLES */
1657                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1658                 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
1659                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
1660
1661                 if (ebx & 0x40) {
1662                         /*
1663                          * Erratum AAJ80 detected, we work it around by using
1664                          * the BR_MISP_EXEC.ANY event. This will over-count
1665                          * branch-misses, but it's still much better than the
1666                          * architectural event which is often completely bogus:
1667                          */
1668                         intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89;
1669
1670                         pr_cont("erratum AAJ80 worked around, ");
1671                 }
1672                 pr_cont("Nehalem events, ");
1673                 break;
1674
1675         case 28: /* Atom */
1676                 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids,
1677                        sizeof(hw_cache_event_ids));
1678
1679                 intel_pmu_lbr_init_atom();
1680
1681                 x86_pmu.event_constraints = intel_gen_event_constraints;
1682                 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints;
1683                 pr_cont("Atom events, ");
1684                 break;
1685
1686         case 37: /* 32 nm nehalem, "Clarkdale" */
1687         case 44: /* 32 nm nehalem, "Gulftown" */
1688         case 47: /* 32 nm Xeon E7 */
1689                 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids,
1690                        sizeof(hw_cache_event_ids));
1691                 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs,
1692                        sizeof(hw_cache_extra_regs));
1693
1694                 intel_pmu_lbr_init_nhm();
1695
1696                 x86_pmu.event_constraints = intel_westmere_event_constraints;
1697                 x86_pmu.enable_all = intel_pmu_nhm_enable_all;
1698                 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints;
1699                 x86_pmu.extra_regs = intel_westmere_extra_regs;
1700                 x86_pmu.er_flags |= ERF_HAS_RSP_1;
1701
1702                 /* UOPS_ISSUED.STALLED_CYCLES */
1703                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1704                 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */
1705                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x1803fb1;
1706
1707                 pr_cont("Westmere events, ");
1708                 break;
1709
1710         case 42: /* SandyBridge */
1711                 x86_pmu.quirks = intel_sandybridge_quirks;
1712         case 45: /* SandyBridge, "Romely-EP" */
1713                 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids,
1714                        sizeof(hw_cache_event_ids));
1715
1716                 intel_pmu_lbr_init_nhm();
1717
1718                 x86_pmu.event_constraints = intel_snb_event_constraints;
1719                 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints;
1720                 if (boot_cpu_data.x86_model == 45)
1721                         x86_pmu.extra_regs = intel_snbep_extra_regs;
1722                 else
1723                         x86_pmu.extra_regs = intel_snb_extra_regs;
1724                 /* all extra regs are per-cpu when HT is on */
1725                 x86_pmu.er_flags |= ERF_HAS_RSP_1;
1726                 x86_pmu.er_flags |= ERF_NO_HT_SHARING;
1727
1728                 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */
1729                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x180010e;
1730                 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/
1731                 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x18001b1;
1732
1733                 pr_cont("SandyBridge events, ");
1734                 break;
1735
1736         default:
1737                 switch (x86_pmu.version) {
1738                 case 1:
1739                         x86_pmu.event_constraints = intel_v1_event_constraints;
1740                         pr_cont("generic architected perfmon v1, ");
1741                         break;
1742                 default:
1743                         /*
1744                          * default constraints for v2 and up
1745                          */
1746                         x86_pmu.event_constraints = intel_gen_event_constraints;
1747                         pr_cont("generic architected perfmon, ");
1748                         break;
1749                 }
1750         }
1751         return 0;
1752 }