arch/x86/kernel/cpu/perf_event_amd.c

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