arch/x86/kernel/cpu/perf_event_amd.c

   1 #include <linux/perf_event.h>
   2 #include <linux/export.h>
   3 #include <linux/types.h>
   4 #include <linux/init.h>
   5 #include <linux/slab.h>
   6 #include <asm/apicdef.h>
   7
   8 #include "perf_event.h"
   9
  10 static __initconst const u64 amd_hw_cache_event_ids
  11                                 [PERF_COUNT_HW_CACHE_MAX]
  12                                 [PERF_COUNT_HW_CACHE_OP_MAX]
  13                                 [PERF_COUNT_HW_CACHE_RESULT_MAX] =
  14 {
  15  [ C(L1D) ] = {
  16         [ C(OP_READ) ] = {
  17                 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
  18                 [ C(RESULT_MISS)   ] = 0x0141, /* Data Cache Misses          */
  19         },
  20         [ C(OP_WRITE) ] = {
  21                 [ C(RESULT_ACCESS) ] = 0x0142, /* Data Cache Refills :system */
  22                 [ C(RESULT_MISS)   ] = 0,
  23         },
  24         [ C(OP_PREFETCH) ] = {
  25                 [ C(RESULT_ACCESS) ] = 0x0267, /* Data Prefetcher :attempts  */
  26                 [ C(RESULT_MISS)   ] = 0x0167, /* Data Prefetcher :cancelled */
  27         },
  28  },
  29  [ C(L1I ) ] = {
  30         [ C(OP_READ) ] = {
  31                 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction cache fetches  */
  32                 [ C(RESULT_MISS)   ] = 0x0081, /* Instruction cache misses   */
  33         },
  34         [ C(OP_WRITE) ] = {
  35                 [ C(RESULT_ACCESS) ] = -1,
  36                 [ C(RESULT_MISS)   ] = -1,
  37         },
  38         [ C(OP_PREFETCH) ] = {
  39                 [ C(RESULT_ACCESS) ] = 0x014B, /* Prefetch Instructions :Load */
  40                 [ C(RESULT_MISS)   ] = 0,
  41         },
  42  },
  43  [ C(LL  ) ] = {
  44         [ C(OP_READ) ] = {
  45                 [ C(RESULT_ACCESS) ] = 0x037D, /* Requests to L2 Cache :IC+DC */
  46                 [ C(RESULT_MISS)   ] = 0x037E, /* L2 Cache Misses : IC+DC     */
  47         },
  48         [ C(OP_WRITE) ] = {
  49                 [ C(RESULT_ACCESS) ] = 0x017F, /* L2 Fill/Writeback           */
  50                 [ C(RESULT_MISS)   ] = 0,
  51         },
  52         [ C(OP_PREFETCH) ] = {
  53                 [ C(RESULT_ACCESS) ] = 0,
  54                 [ C(RESULT_MISS)   ] = 0,
  55         },
  56  },
  57  [ C(DTLB) ] = {
  58         [ C(OP_READ) ] = {
  59                 [ C(RESULT_ACCESS) ] = 0x0040, /* Data Cache Accesses        */
  60                 [ C(RESULT_MISS)   ] = 0x0746, /* L1_DTLB_AND_L2_DLTB_MISS.ALL */
  61         },
  62         [ C(OP_WRITE) ] = {
  63                 [ C(RESULT_ACCESS) ] = 0,
  64                 [ C(RESULT_MISS)   ] = 0,
  65         },
  66         [ C(OP_PREFETCH) ] = {
  67                 [ C(RESULT_ACCESS) ] = 0,
  68                 [ C(RESULT_MISS)   ] = 0,
  69         },
  70  },
  71  [ C(ITLB) ] = {
  72         [ C(OP_READ) ] = {
  73                 [ C(RESULT_ACCESS) ] = 0x0080, /* Instruction fecthes        */
  74                 [ C(RESULT_MISS)   ] = 0x0385, /* L1_ITLB_AND_L2_ITLB_MISS.ALL */
  75         },
  76         [ C(OP_WRITE) ] = {
  77                 [ C(RESULT_ACCESS) ] = -1,
  78                 [ C(RESULT_MISS)   ] = -1,
  79         },
  80         [ C(OP_PREFETCH) ] = {
  81                 [ C(RESULT_ACCESS) ] = -1,
  82                 [ C(RESULT_MISS)   ] = -1,
  83         },
  84  },
  85  [ C(BPU ) ] = {
  86         [ C(OP_READ) ] = {
  87                 [ C(RESULT_ACCESS) ] = 0x00c2, /* Retired Branch Instr.      */
  88                 [ C(RESULT_MISS)   ] = 0x00c3, /* Retired Mispredicted BI    */
  89         },
  90         [ C(OP_WRITE) ] = {
  91                 [ C(RESULT_ACCESS) ] = -1,
  92                 [ C(RESULT_MISS)   ] = -1,
  93         },
  94         [ C(OP_PREFETCH) ] = {
  95                 [ C(RESULT_ACCESS) ] = -1,
  96                 [ C(RESULT_MISS)   ] = -1,
  97         },
  98  },
  99  [ C(NODE) ] = {
 100         [ C(OP_READ) ] = {
 101                 [ C(RESULT_ACCESS) ] = 0xb8e9, /* CPU Request to Memory, l+r */
 102                 [ C(RESULT_MISS)   ] = 0x98e9, /* CPU Request to Memory, r   */
 103         },
 104         [ C(OP_WRITE) ] = {
 105                 [ C(RESULT_ACCESS) ] = -1,
 106                 [ C(RESULT_MISS)   ] = -1,
 107         },
 108         [ C(OP_PREFETCH) ] = {
 109                 [ C(RESULT_ACCESS) ] = -1,
 110                 [ C(RESULT_MISS)   ] = -1,
 111         },
 112  },
 113 };
 114
 115 /*
 116  * AMD Performance Monitor K7 and later.
 117  */
 118 static const u64 amd_perfmon_event_map[] =
 119 {
 120   [PERF_COUNT_HW_CPU_CYCLES]                    = 0x0076,
 121   [PERF_COUNT_HW_INSTRUCTIONS]                  = 0x00c0,
 122   [PERF_COUNT_HW_CACHE_REFERENCES]              = 0x0080,
 123   [PERF_COUNT_HW_CACHE_MISSES]                  = 0x0081,
 124   [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]           = 0x00c2,
 125   [PERF_COUNT_HW_BRANCH_MISSES]                 = 0x00c3,
 126   [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND]       = 0x00d0, /* "Decoder empty" event */
 127   [PERF_COUNT_HW_STALLED_CYCLES_BACKEND]        = 0x00d1, /* "Dispatch stalls" event */
 128 };
 129
 130 static u64 amd_pmu_event_map(int hw_event)
 131 {
 132         return amd_perfmon_event_map[hw_event];
 133 }
 134
 135 static int amd_pmu_hw_config(struct perf_event *event)
 136 {
 137         int ret = x86_pmu_hw_config(event);
 138
 139         if (ret)
 140                 return ret;
 141
 142         if (event->attr.exclude_host && event->attr.exclude_guest)
 143                 /*
 144                  * When HO == GO == 1 the hardware treats that as GO == HO == 0
 145                  * and will count in both modes. We don't want to count in that
 146                  * case so we emulate no-counting by setting US = OS = 0.
 147                  */
 148                 event->hw.config &= ~(ARCH_PERFMON_EVENTSEL_USR |
 149                                       ARCH_PERFMON_EVENTSEL_OS);
 150         else if (event->attr.exclude_host)
 151                 event->hw.config |= AMD_PERFMON_EVENTSEL_GUESTONLY;
 152         else if (event->attr.exclude_guest)
 153                 event->hw.config |= AMD_PERFMON_EVENTSEL_HOSTONLY;
 154
 155         if (event->attr.type != PERF_TYPE_RAW)
 156                 return 0;
 157
 158         event->hw.config |= event->attr.config & AMD64_RAW_EVENT_MASK;
 159
 160         return 0;
 161 }
 162
 163 /*
 164  * AMD64 events are detected based on their event codes.
 165  */
 166 static inline unsigned int amd_get_event_code(struct hw_perf_event *hwc)
 167 {
 168         return ((hwc->config >> 24) & 0x0f00) | (hwc->config & 0x00ff);
 169 }
 170
 171 static inline int amd_is_nb_event(struct hw_perf_event *hwc)
 172 {
 173         return (hwc->config & 0xe0) == 0xe0;
 174 }
 175
 176 static inline int amd_has_nb(struct cpu_hw_events *cpuc)
 177 {
 178         struct amd_nb *nb = cpuc->amd_nb;
 179
 180         return nb && nb->nb_id != -1;
 181 }
 182
 183 static void amd_put_event_constraints(struct cpu_hw_events *cpuc,
 184                                       struct perf_event *event)
 185 {
 186         struct hw_perf_event *hwc = &event->hw;
 187         struct amd_nb *nb = cpuc->amd_nb;
 188         int i;
 189
 190         /*
 191          * only care about NB events
 192          */
 193         if (!(amd_has_nb(cpuc) && amd_is_nb_event(hwc)))
 194                 return;
 195
 196         /*
 197          * need to scan whole list because event may not have
 198          * been assigned during scheduling
 199          *
 200          * no race condition possible because event can only
 201          * be removed on one CPU at a time AND PMU is disabled
 202          * when we come here
 203          */
 204         for (i = 0; i < x86_pmu.num_counters; i++) {
 205                 if (nb->owners[i] == event) {
 206                         cmpxchg(nb->owners+i, event, NULL);
 207                         break;
 208                 }
 209         }
 210 }
 211
 212  /*
 213   * AMD64 NorthBridge events need special treatment because
 214   * counter access needs to be synchronized across all cores
 215   * of a package. Refer to BKDG section 3.12
 216   *
 217   * NB events are events measuring L3 cache, Hypertransport
 218   * traffic. They are identified by an event code >= 0xe00.
 219   * They measure events on the NorthBride which is shared
 220   * by all cores on a package. NB events are counted on a
 221   * shared set of counters. When a NB event is programmed
 222   * in a counter, the data actually comes from a shared
 223   * counter. Thus, access to those counters needs to be
 224   * synchronized.
 225   *
 226   * We implement the synchronization such that no two cores
 227   * can be measuring NB events using the same counters. Thus,
 228   * we maintain a per-NB allocation table. The available slot
 229   * is propagated using the event_constraint structure.
 230   *
 231   * We provide only one choice for each NB event based on
 232   * the fact that only NB events have restrictions. Consequently,
 233   * if a counter is available, there is a guarantee the NB event
 234   * will be assigned to it. If no slot is available, an empty
 235   * constraint is returned and scheduling will eventually fail
 236   * for this event.
 237   *
 238   * Note that all cores attached the same NB compete for the same
 239   * counters to host NB events, this is why we use atomic ops. Some
 240   * multi-chip CPUs may have more than one NB.
 241   *
 242   * Given that resources are allocated (cmpxchg), they must be
 243   * eventually freed for others to use. This is accomplished by
 244   * calling amd_put_event_constraints().
 245   *
 246   * Non NB events are not impacted by this restriction.
 247   */
 248 static struct event_constraint *
 249 amd_get_event_constraints(struct cpu_hw_events *cpuc, struct perf_event *event)
 250 {
 251         struct hw_perf_event *hwc = &event->hw;
 252         struct amd_nb *nb = cpuc->amd_nb;
 253         struct perf_event *old = NULL;
 254         int max = x86_pmu.num_counters;
 255         int i, j, k = -1;
 256
 257         /*
 258          * if not NB event or no NB, then no constraints
 259          */
 260         if (!(amd_has_nb(cpuc) && amd_is_nb_event(hwc)))
 261                 return &unconstrained;
 262
 263         /*
 264          * detect if already present, if so reuse
 265          *
 266          * cannot merge with actual allocation
 267          * because of possible holes
 268          *
 269          * event can already be present yet not assigned (in hwc->idx)
 270          * because of successive calls to x86_schedule_events() from
 271          * hw_perf_group_sched_in() without hw_perf_enable()
 272          */
 273         for (i = 0; i < max; i++) {
 274                 /*
 275                  * keep track of first free slot
 276                  */
 277                 if (k == -1 && !nb->owners[i])
 278                         k = i;
 279
 280                 /* already present, reuse */
 281                 if (nb->owners[i] == event)
 282                         goto done;
 283         }
 284         /*
 285          * not present, so grab a new slot
 286          * starting either at:
 287          */
 288         if (hwc->idx != -1) {
 289                 /* previous assignment */
 290                 i = hwc->idx;
 291         } else if (k != -1) {
 292                 /* start from free slot found */
 293                 i = k;
 294         } else {
 295                 /*
 296                  * event not found, no slot found in
 297                  * first pass, try again from the
 298                  * beginning
 299                  */
 300                 i = 0;
 301         }
 302         j = i;
 303         do {
 304                 old = cmpxchg(nb->owners+i, NULL, event);
 305                 if (!old)
 306                         break;
 307                 if (++i == max)
 308                         i = 0;
 309         } while (i != j);
 310 done:
 311         if (!old)
 312                 return &nb->event_constraints[i];
 313
 314         return &emptyconstraint;
 315 }
 316
 317 static struct amd_nb *amd_alloc_nb(int cpu)
 318 {
 319         struct amd_nb *nb;
 320         int i;
 321
 322         nb = kmalloc_node(sizeof(struct amd_nb), GFP_KERNEL | __GFP_ZERO,
 323                           cpu_to_node(cpu));
 324         if (!nb)
 325                 return NULL;
 326
 327         nb->nb_id = -1;
 328
 329         /*
 330          * initialize all possible NB constraints
 331          */
 332         for (i = 0; i < x86_pmu.num_counters; i++) {
 333                 __set_bit(i, nb->event_constraints[i].idxmsk);
 334                 nb->event_constraints[i].weight = 1;
 335         }
 336         return nb;
 337 }
 338
 339 static int amd_pmu_cpu_prepare(int cpu)
 340 {
 341         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
 342
 343         WARN_ON_ONCE(cpuc->amd_nb);
 344
 345         if (boot_cpu_data.x86_max_cores < 2)
 346                 return NOTIFY_OK;
 347
 348         cpuc->amd_nb = amd_alloc_nb(cpu);
 349         if (!cpuc->amd_nb)
 350                 return NOTIFY_BAD;
 351
 352         return NOTIFY_OK;
 353 }
 354
 355 static void amd_pmu_cpu_starting(int cpu)
 356 {
 357         struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu);
 358         struct amd_nb *nb;
 359         int i, nb_id;
 360
 361         cpuc->perf_ctr_virt_mask = AMD_PERFMON_EVENTSEL_HOSTONLY;
 362
 363         if (boot_cpu_data.x86_max_cores < 2 || boot_cpu_data.x86 == 0x15)
 364                 return;
 365
 366         nb_id = amd_get_nb_id(cpu);
 367         WARN_ON_ONCE(nb_id == BAD_APICID);
 368
 369         for_each_online_cpu(i) {
 370                 nb = per_cpu(cpu_hw_events, i).amd_nb;
 371                 if (WARN_ON_ONCE(!nb))
 372                         continue;
 373
 374                 if (nb->nb_id == nb_id) {
 375                         cpuc->kfree_on_online = cpuc->amd_nb;
 376                         cpuc->amd_nb = nb;
 377                         break;
 378                 }
 379         }
 380
 381         cpuc->amd_nb->nb_id = nb_id;
 382         cpuc->amd_nb->refcnt++;
 383 }
 384
 385 static void amd_pmu_cpu_dead(int cpu)
 386 {
 387         struct cpu_hw_events *cpuhw;
 388
 389         if (boot_cpu_data.x86_max_cores < 2)
 390                 return;
 391
 392         cpuhw = &per_cpu(cpu_hw_events, cpu);
 393
 394         if (cpuhw->amd_nb) {
 395                 struct amd_nb *nb = cpuhw->amd_nb;
 396
 397                 if (nb->nb_id == -1 || --nb->refcnt == 0)
 398                         kfree(nb);
 399
 400                 cpuhw->amd_nb = NULL;
 401         }
 402 }
 403
 404 static __initconst const struct x86_pmu amd_pmu = {
 405         .name                   = "AMD",
 406         .handle_irq             = x86_pmu_handle_irq,
 407         .disable_all            = x86_pmu_disable_all,
 408         .enable_all             = x86_pmu_enable_all,
 409         .enable                 = x86_pmu_enable_event,
 410         .disable                = x86_pmu_disable_event,
 411         .hw_config              = amd_pmu_hw_config,
 412         .schedule_events        = x86_schedule_events,
 413         .eventsel               = MSR_K7_EVNTSEL0,
 414         .perfctr                = MSR_K7_PERFCTR0,
 415         .event_map              = amd_pmu_event_map,
 416         .max_events             = ARRAY_SIZE(amd_perfmon_event_map),
 417         .num_counters           = AMD64_NUM_COUNTERS,
 418         .cntval_bits            = 48,
 419         .cntval_mask            = (1ULL << 48) - 1,
 420         .apic                   = 1,
 421         /* use highest bit to detect overflow */
 422         .max_period             = (1ULL << 47) - 1,
 423         .get_event_constraints  = amd_get_event_constraints,
 424         .put_event_constraints  = amd_put_event_constraints,
 425
 426         .cpu_prepare            = amd_pmu_cpu_prepare,
 427         .cpu_starting           = amd_pmu_cpu_starting,
 428         .cpu_dead               = amd_pmu_cpu_dead,
 429 };
 430
 431 /* AMD Family 15h */
 432
 433 #define AMD_EVENT_TYPE_MASK     0x000000F0ULL
 434
 435 #define AMD_EVENT_FP            0x00000000ULL ... 0x00000010ULL
 436 #define AMD_EVENT_LS            0x00000020ULL ... 0x00000030ULL
 437 #define AMD_EVENT_DC            0x00000040ULL ... 0x00000050ULL
 438 #define AMD_EVENT_CU            0x00000060ULL ... 0x00000070ULL
 439 #define AMD_EVENT_IC_DE         0x00000080ULL ... 0x00000090ULL
 440 #define AMD_EVENT_EX_LS         0x000000C0ULL
 441 #define AMD_EVENT_DE            0x000000D0ULL
 442 #define AMD_EVENT_NB            0x000000E0ULL ... 0x000000F0ULL
 443
 444 /*
 445  * AMD family 15h event code/PMC mappings:
 446  *
 447  * type = event_code & 0x0F0:
 448  *
 449  * 0x000        FP      PERF_CTL[5:3]
 450  * 0x010        FP      PERF_CTL[5:3]
 451  * 0x020        LS      PERF_CTL[5:0]
 452  * 0x030        LS      PERF_CTL[5:0]
 453  * 0x040        DC      PERF_CTL[5:0]
 454  * 0x050        DC      PERF_CTL[5:0]
 455  * 0x060        CU      PERF_CTL[2:0]
 456  * 0x070        CU      PERF_CTL[2:0]
 457  * 0x080        IC/DE   PERF_CTL[2:0]
 458  * 0x090        IC/DE   PERF_CTL[2:0]
 459  * 0x0A0        ---
 460  * 0x0B0        ---
 461  * 0x0C0        EX/LS   PERF_CTL[5:0]
 462  * 0x0D0        DE      PERF_CTL[2:0]
 463  * 0x0E0        NB      NB_PERF_CTL[3:0]
 464  * 0x0F0        NB      NB_PERF_CTL[3:0]
 465  *
 466  * Exceptions:
 467  *
 468  * 0x000        FP      PERF_CTL[3], PERF_CTL[5:3] (*)
 469  * 0x003        FP      PERF_CTL[3]
 470  * 0x004        FP      PERF_CTL[3], PERF_CTL[5:3] (*)
 471  * 0x00B        FP      PERF_CTL[3]
 472  * 0x00D        FP      PERF_CTL[3]
 473  * 0x023        DE      PERF_CTL[2:0]
 474  * 0x02D        LS      PERF_CTL[3]
 475  * 0x02E        LS      PERF_CTL[3,0]
 476  * 0x043        CU      PERF_CTL[2:0]
 477  * 0x045        CU      PERF_CTL[2:0]
 478  * 0x046        CU      PERF_CTL[2:0]
 479  * 0x054        CU      PERF_CTL[2:0]
 480  * 0x055        CU      PERF_CTL[2:0]
 481  * 0x08F        IC      PERF_CTL[0]
 482  * 0x187        DE      PERF_CTL[0]
 483  * 0x188        DE      PERF_CTL[0]
 484  * 0x0DB        EX      PERF_CTL[5:0]
 485  * 0x0DC        LS      PERF_CTL[5:0]
 486  * 0x0DD        LS      PERF_CTL[5:0]
 487  * 0x0DE        LS      PERF_CTL[5:0]
 488  * 0x0DF        LS      PERF_CTL[5:0]
 489  * 0x1D6        EX      PERF_CTL[5:0]
 490  * 0x1D8        EX      PERF_CTL[5:0]
 491  *
 492  * (*) depending on the umask all FPU counters may be used
 493  */
 494
 495 static struct event_constraint amd_f15_PMC0  = EVENT_CONSTRAINT(0, 0x01, 0);
 496 static struct event_constraint amd_f15_PMC20 = EVENT_CONSTRAINT(0, 0x07, 0);
 497 static struct event_constraint amd_f15_PMC3  = EVENT_CONSTRAINT(0, 0x08, 0);
 498 static struct event_constraint amd_f15_PMC30 = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0);
 499 static struct event_constraint amd_f15_PMC50 = EVENT_CONSTRAINT(0, 0x3F, 0);
 500 static struct event_constraint amd_f15_PMC53 = EVENT_CONSTRAINT(0, 0x38, 0);
 501
 502 static struct event_constraint *
 503 amd_get_event_constraints_f15h(struct cpu_hw_events *cpuc, struct perf_event *event)
 504 {
 505         struct hw_perf_event *hwc = &event->hw;
 506         unsigned int event_code = amd_get_event_code(hwc);
 507
 508         switch (event_code & AMD_EVENT_TYPE_MASK) {
 509         case AMD_EVENT_FP:
 510                 switch (event_code) {
 511                 case 0x000:
 512                         if (!(hwc->config & 0x0000F000ULL))
 513                                 break;
 514                         if (!(hwc->config & 0x00000F00ULL))
 515                                 break;
 516                         return &amd_f15_PMC3;
 517                 case 0x004:
 518                         if (hweight_long(hwc->config & ARCH_PERFMON_EVENTSEL_UMASK) <= 1)
 519                                 break;
 520                         return &amd_f15_PMC3;
 521                 case 0x003:
 522                 case 0x00B:
 523                 case 0x00D:
 524                         return &amd_f15_PMC3;
 525                 }
 526                 return &amd_f15_PMC53;
 527         case AMD_EVENT_LS:
 528         case AMD_EVENT_DC:
 529         case AMD_EVENT_EX_LS:
 530                 switch (event_code) {
 531                 case 0x023:
 532                 case 0x043:
 533                 case 0x045:
 534                 case 0x046:
 535                 case 0x054:
 536                 case 0x055:
 537                         return &amd_f15_PMC20;
 538                 case 0x02D:
 539                         return &amd_f15_PMC3;
 540                 case 0x02E:
 541                         return &amd_f15_PMC30;
 542                 default:
 543                         return &amd_f15_PMC50;
 544                 }
 545         case AMD_EVENT_CU:
 546         case AMD_EVENT_IC_DE:
 547         case AMD_EVENT_DE:
 548                 switch (event_code) {
 549                 case 0x08F:
 550                 case 0x187:
 551                 case 0x188:
 552                         return &amd_f15_PMC0;
 553                 case 0x0DB ... 0x0DF:
 554                 case 0x1D6:
 555                 case 0x1D8:
 556                         return &amd_f15_PMC50;
 557                 default:
 558                         return &amd_f15_PMC20;
 559                 }
 560         case AMD_EVENT_NB:
 561                 /* not yet implemented */
 562                 return &emptyconstraint;
 563         default:
 564                 return &emptyconstraint;
 565         }
 566 }
 567
 568 static __initconst const struct x86_pmu amd_pmu_f15h = {
 569         .name                   = "AMD Family 15h",
 570         .handle_irq             = x86_pmu_handle_irq,
 571         .disable_all            = x86_pmu_disable_all,
 572         .enable_all             = x86_pmu_enable_all,
 573         .enable                 = x86_pmu_enable_event,
 574         .disable                = x86_pmu_disable_event,
 575         .hw_config              = amd_pmu_hw_config,
 576         .schedule_events        = x86_schedule_events,
 577         .eventsel               = MSR_F15H_PERF_CTL,
 578         .perfctr                = MSR_F15H_PERF_CTR,
 579         .event_map              = amd_pmu_event_map,
 580         .max_events             = ARRAY_SIZE(amd_perfmon_event_map),
 581         .num_counters           = AMD64_NUM_COUNTERS_F15H,
 582         .cntval_bits            = 48,
 583         .cntval_mask            = (1ULL << 48) - 1,
 584         .apic                   = 1,
 585         /* use highest bit to detect overflow */
 586         .max_period             = (1ULL << 47) - 1,
 587         .get_event_constraints  = amd_get_event_constraints_f15h,
 588         /* nortbridge counters not yet implemented: */
 589 #if 0
 590         .put_event_constraints  = amd_put_event_constraints,
 591
 592         .cpu_prepare            = amd_pmu_cpu_prepare,
 593         .cpu_dead               = amd_pmu_cpu_dead,
 594 #endif
 595         .cpu_starting           = amd_pmu_cpu_starting,
 596 };
 597
 598 __init int amd_pmu_init(void)
 599 {
 600         /* Performance-monitoring supported from K7 and later: */
 601         if (boot_cpu_data.x86 < 6)
 602                 return -ENODEV;
 603
 604         /*
 605          * If core performance counter extensions exists, it must be
 606          * family 15h, otherwise fail. See x86_pmu_addr_offset().
 607          */
 608         switch (boot_cpu_data.x86) {
 609         case 0x15:
 610                 if (!cpu_has_perfctr_core)
 611                         return -ENODEV;
 612                 x86_pmu = amd_pmu_f15h;
 613                 break;
 614         default:
 615                 if (cpu_has_perfctr_core)
 616                         return -ENODEV;
 617                 x86_pmu = amd_pmu;
 618                 break;
 619         }
 620
 621         /* Events are common for all AMDs */
 622         memcpy(hw_cache_event_ids, amd_hw_cache_event_ids,
 623                sizeof(hw_cache_event_ids));
 624
 625         return 0;
 626 }
 627
 628 void amd_pmu_enable_virt(void)
 629 {
 630         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 631
 632         cpuc->perf_ctr_virt_mask = 0;
 633
 634         /* Reload all events */
 635         x86_pmu_disable_all();
 636         x86_pmu_enable_all(0);
 637 }
 638 EXPORT_SYMBOL_GPL(amd_pmu_enable_virt);
 639
 640 void amd_pmu_disable_virt(void)
 641 {
 642         struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
 643
 644         /*
 645          * We only mask out the Host-only bit so that host-only counting works
 646          * when SVM is disabled. If someone sets up a guest-only counter when
 647          * SVM is disabled the Guest-only bits still gets set and the counter
 648          * will not count anything.
 649          */
 650         cpuc->perf_ctr_virt_mask = AMD_PERFMON_EVENTSEL_HOSTONLY;
 651
 652         /* Reload all events */
 653         x86_pmu_disable_all();
 654         x86_pmu_enable_all(0);
 655 }
 656 EXPORT_SYMBOL_GPL(amd_pmu_disable_virt);