arch/powerpc/perf/power9-pmu.c

   1 /*
   2  * Performance counter support for POWER9 processors.
   3  *
   4  * Copyright 2009 Paul Mackerras, IBM Corporation.
   5  * Copyright 2013 Michael Ellerman, IBM Corporation.
   6  * Copyright 2016 Madhavan Srinivasan, IBM Corporation.
   7  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public License
  10  * as published by the Free Software Foundation; either version
  11  * 2 of the License, or later version.
  12  */
  13
  14 #define pr_fmt(fmt)     "power9-pmu: " fmt
  15
  16 #include "isa207-common.h"
  17
  18 /*
  19  * Raw event encoding for Power9:
  20  *
  21  *        60        56        52        48        44        40        36        32
  22  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
  23  *   | | [ ]                       [ ] [      thresh_cmp     ]   [  thresh_ctl   ]
  24  *   | |  |                         |                                     |
  25  *   | |  *- IFM (Linux)            |                  thresh start/stop -*
  26  *   | *- BHRB (Linux)              *sm
  27  *   *- EBB (Linux)
  28  *
  29  *        28        24        20        16        12         8         4         0
  30  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
  31  *   [   ] [  sample ]   [cache]   [ pmc ]   [unit ]   []    m   [    pmcxsel    ]
  32  *     |        |           |                          |     |
  33  *     |        |           |                          |     *- mark
  34  *     |        |           *- L1/L2/L3 cache_sel      |
  35  *     |        |                                      |
  36  *     |        *- sampling mode for marked events     *- combine
  37  *     |
  38  *     *- thresh_sel
  39  *
  40  * Below uses IBM bit numbering.
  41  *
  42  * MMCR1[x:y] = unit    (PMCxUNIT)
  43  * MMCR1[24]   = pmc1combine[0]
  44  * MMCR1[25]   = pmc1combine[1]
  45  * MMCR1[26]   = pmc2combine[0]
  46  * MMCR1[27]   = pmc2combine[1]
  47  * MMCR1[28]   = pmc3combine[0]
  48  * MMCR1[29]   = pmc3combine[1]
  49  * MMCR1[30]   = pmc4combine[0]
  50  * MMCR1[31]   = pmc4combine[1]
  51  *
  52  * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
  53  *      MMCR1[20:27] = thresh_ctl
  54  * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
  55  *      MMCR1[20:27] = thresh_ctl
  56  * else
  57  *      MMCRA[48:55] = thresh_ctl   (THRESH START/END)
  58  *
  59  * if thresh_sel:
  60  *      MMCRA[45:47] = thresh_sel
  61  *
  62  * if thresh_cmp:
  63  *      MMCRA[9:11] = thresh_cmp[0:2]
  64  *      MMCRA[12:18] = thresh_cmp[3:9]
  65  *
  66  * if unit == 6 or unit == 7
  67  *      MMCRC[53:55] = cache_sel[1:3]      (L2EVENT_SEL)
  68  * else if unit == 8 or unit == 9:
  69  *      if cache_sel[0] == 0: # L3 bank
  70  *              MMCRC[47:49] = cache_sel[1:3]  (L3EVENT_SEL0)
  71  *      else if cache_sel[0] == 1:
  72  *              MMCRC[50:51] = cache_sel[2:3]  (L3EVENT_SEL1)
  73  * else if cache_sel[1]: # L1 event
  74  *      MMCR1[16] = cache_sel[2]
  75  *     MMCR1[17] = cache_sel[3]
  76  *
  77  * if mark:
  78  *      MMCRA[63]    = 1                (SAMPLE_ENABLE)
  79  *      MMCRA[57:59] = sample[0:2]      (RAND_SAMP_ELIG)
  80  *     MMCRA[61:62] = sample[3:4]      (RAND_SAMP_MODE)
  81  *
  82  * if EBB and BHRB:
  83  *      MMCRA[32:33] = IFM
  84  *
  85  * MMCRA[SDAR_MODE]  = sm
  86  */
  87
  88 /*
  89  * Some power9 event codes.
  90  */
  91 #define EVENT(_name, _code)     _name = _code,
  92
  93 enum {
  94 #include "power9-events-list.h"
  95 };
  96
  97 #undef EVENT
  98
  99 /* MMCRA IFM bits - POWER9 */
 100 #define POWER9_MMCRA_IFM1               0x0000000040000000UL
 101 #define POWER9_MMCRA_IFM2               0x0000000080000000UL
 102 #define POWER9_MMCRA_IFM3               0x00000000C0000000UL
 103
 104 /* PowerISA v2.07 format attribute structure*/
 105 extern struct attribute_group isa207_pmu_format_group;
 106
 107 /* Table of alternatives, sorted by column 0 */
 108 static const unsigned int power9_event_alternatives[][MAX_ALT] = {
 109         { PM_INST_DISP,                 PM_INST_DISP_ALT },
 110 };
 111
 112 static int power9_get_alternatives(u64 event, unsigned int flags, u64 alt[])
 113 {
 114         int num_alt = 0;
 115
 116         num_alt = isa207_get_alternatives(event, alt, power9_event_alternatives,
 117                                 (int)ARRAY_SIZE(power9_event_alternatives));
 118
 119         return num_alt;
 120 }
 121
 122 GENERIC_EVENT_ATTR(cpu-cycles,                  PM_CYC);
 123 GENERIC_EVENT_ATTR(stalled-cycles-frontend,     PM_ICT_NOSLOT_CYC);
 124 GENERIC_EVENT_ATTR(stalled-cycles-backend,      PM_CMPLU_STALL);
 125 GENERIC_EVENT_ATTR(instructions,                PM_INST_CMPL);
 126 GENERIC_EVENT_ATTR(branch-instructions,         PM_BRU_CMPL);
 127 GENERIC_EVENT_ATTR(branch-misses,               PM_BR_MPRED_CMPL);
 128 GENERIC_EVENT_ATTR(cache-references,            PM_LD_REF_L1);
 129 GENERIC_EVENT_ATTR(cache-misses,                PM_LD_MISS_L1_FIN);
 130
 131 CACHE_EVENT_ATTR(L1-dcache-load-misses,         PM_LD_MISS_L1_FIN);
 132 CACHE_EVENT_ATTR(L1-dcache-loads,               PM_LD_REF_L1);
 133 CACHE_EVENT_ATTR(L1-dcache-prefetches,          PM_L1_PREF);
 134 CACHE_EVENT_ATTR(L1-dcache-store-misses,        PM_ST_MISS_L1);
 135 CACHE_EVENT_ATTR(L1-icache-load-misses,         PM_L1_ICACHE_MISS);
 136 CACHE_EVENT_ATTR(L1-icache-loads,               PM_INST_FROM_L1);
 137 CACHE_EVENT_ATTR(L1-icache-prefetches,          PM_IC_PREF_WRITE);
 138 CACHE_EVENT_ATTR(LLC-load-misses,               PM_DATA_FROM_L3MISS);
 139 CACHE_EVENT_ATTR(LLC-loads,                     PM_DATA_FROM_L3);
 140 CACHE_EVENT_ATTR(LLC-prefetches,                PM_L3_PREF_ALL);
 141 CACHE_EVENT_ATTR(LLC-store-misses,              PM_L2_ST_MISS);
 142 CACHE_EVENT_ATTR(LLC-stores,                    PM_L2_ST);
 143 CACHE_EVENT_ATTR(branch-load-misses,            PM_BR_MPRED_CMPL);
 144 CACHE_EVENT_ATTR(branch-loads,                  PM_BRU_CMPL);
 145 CACHE_EVENT_ATTR(dTLB-load-misses,              PM_DTLB_MISS);
 146 CACHE_EVENT_ATTR(iTLB-load-misses,              PM_ITLB_MISS);
 147
 148 static struct attribute *power9_events_attr[] = {
 149         GENERIC_EVENT_PTR(PM_CYC),
 150         GENERIC_EVENT_PTR(PM_ICT_NOSLOT_CYC),
 151         GENERIC_EVENT_PTR(PM_CMPLU_STALL),
 152         GENERIC_EVENT_PTR(PM_INST_CMPL),
 153         GENERIC_EVENT_PTR(PM_BRU_CMPL),
 154         GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
 155         GENERIC_EVENT_PTR(PM_LD_REF_L1),
 156         GENERIC_EVENT_PTR(PM_LD_MISS_L1_FIN),
 157         CACHE_EVENT_PTR(PM_LD_MISS_L1_FIN),
 158         CACHE_EVENT_PTR(PM_LD_REF_L1),
 159         CACHE_EVENT_PTR(PM_L1_PREF),
 160         CACHE_EVENT_PTR(PM_ST_MISS_L1),
 161         CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
 162         CACHE_EVENT_PTR(PM_INST_FROM_L1),
 163         CACHE_EVENT_PTR(PM_IC_PREF_WRITE),
 164         CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
 165         CACHE_EVENT_PTR(PM_DATA_FROM_L3),
 166         CACHE_EVENT_PTR(PM_L3_PREF_ALL),
 167         CACHE_EVENT_PTR(PM_L2_ST_MISS),
 168         CACHE_EVENT_PTR(PM_L2_ST),
 169         CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
 170         CACHE_EVENT_PTR(PM_BRU_CMPL),
 171         CACHE_EVENT_PTR(PM_DTLB_MISS),
 172         CACHE_EVENT_PTR(PM_ITLB_MISS),
 173         NULL
 174 };
 175
 176 static struct attribute_group power9_pmu_events_group = {
 177         .name = "events",
 178         .attrs = power9_events_attr,
 179 };
 180
 181 static const struct attribute_group *power9_isa207_pmu_attr_groups[] = {
 182         &isa207_pmu_format_group,
 183         &power9_pmu_events_group,
 184         NULL,
 185 };
 186
 187 PMU_FORMAT_ATTR(event,          "config:0-51");
 188 PMU_FORMAT_ATTR(pmcxsel,        "config:0-7");
 189 PMU_FORMAT_ATTR(mark,           "config:8");
 190 PMU_FORMAT_ATTR(combine,        "config:10-11");
 191 PMU_FORMAT_ATTR(unit,           "config:12-15");
 192 PMU_FORMAT_ATTR(pmc,            "config:16-19");
 193 PMU_FORMAT_ATTR(cache_sel,      "config:20-23");
 194 PMU_FORMAT_ATTR(sample_mode,    "config:24-28");
 195 PMU_FORMAT_ATTR(thresh_sel,     "config:29-31");
 196 PMU_FORMAT_ATTR(thresh_stop,    "config:32-35");
 197 PMU_FORMAT_ATTR(thresh_start,   "config:36-39");
 198 PMU_FORMAT_ATTR(thresh_cmp,     "config:40-49");
 199 PMU_FORMAT_ATTR(sdar_mode,      "config:50-51");
 200
 201 static struct attribute *power9_pmu_format_attr[] = {
 202         &format_attr_event.attr,
 203         &format_attr_pmcxsel.attr,
 204         &format_attr_mark.attr,
 205         &format_attr_combine.attr,
 206         &format_attr_unit.attr,
 207         &format_attr_pmc.attr,
 208         &format_attr_cache_sel.attr,
 209         &format_attr_sample_mode.attr,
 210         &format_attr_thresh_sel.attr,
 211         &format_attr_thresh_stop.attr,
 212         &format_attr_thresh_start.attr,
 213         &format_attr_thresh_cmp.attr,
 214         &format_attr_sdar_mode.attr,
 215         NULL,
 216 };
 217
 218 static struct attribute_group power9_pmu_format_group = {
 219         .name = "format",
 220         .attrs = power9_pmu_format_attr,
 221 };
 222
 223 static const struct attribute_group *power9_pmu_attr_groups[] = {
 224         &power9_pmu_format_group,
 225         &power9_pmu_events_group,
 226         NULL,
 227 };
 228
 229 static int power9_generic_events_dd1[] = {
 230         [PERF_COUNT_HW_CPU_CYCLES] =                    PM_CYC,
 231         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =       PM_ICT_NOSLOT_CYC,
 232         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =        PM_CMPLU_STALL,
 233         [PERF_COUNT_HW_INSTRUCTIONS] =                  PM_INST_DISP,
 234         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =           PM_BR_CMPL_ALT,
 235         [PERF_COUNT_HW_BRANCH_MISSES] =                 PM_BR_MPRED_CMPL,
 236         [PERF_COUNT_HW_CACHE_REFERENCES] =              PM_LD_REF_L1,
 237         [PERF_COUNT_HW_CACHE_MISSES] =                  PM_LD_MISS_L1_FIN,
 238 };
 239
 240 static int power9_generic_events[] = {
 241         [PERF_COUNT_HW_CPU_CYCLES] =                    PM_CYC,
 242         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =       PM_ICT_NOSLOT_CYC,
 243         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =        PM_CMPLU_STALL,
 244         [PERF_COUNT_HW_INSTRUCTIONS] =                  PM_INST_CMPL,
 245         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =           PM_BRU_CMPL,
 246         [PERF_COUNT_HW_BRANCH_MISSES] =                 PM_BR_MPRED_CMPL,
 247         [PERF_COUNT_HW_CACHE_REFERENCES] =              PM_LD_REF_L1,
 248         [PERF_COUNT_HW_CACHE_MISSES] =                  PM_LD_MISS_L1_FIN,
 249 };
 250
 251 static u64 power9_bhrb_filter_map(u64 branch_sample_type)
 252 {
 253         u64 pmu_bhrb_filter = 0;
 254
 255         /* BHRB and regular PMU events share the same privilege state
 256          * filter configuration. BHRB is always recorded along with a
 257          * regular PMU event. As the privilege state filter is handled
 258          * in the basic PMC configuration of the accompanying regular
 259          * PMU event, we ignore any separate BHRB specific request.
 260          */
 261
 262         /* No branch filter requested */
 263         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
 264                 return pmu_bhrb_filter;
 265
 266         /* Invalid branch filter options - HW does not support */
 267         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
 268                 return -1;
 269
 270         if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
 271                 return -1;
 272
 273         if (branch_sample_type & PERF_SAMPLE_BRANCH_CALL)
 274                 return -1;
 275
 276         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
 277                 pmu_bhrb_filter |= POWER9_MMCRA_IFM1;
 278                 return pmu_bhrb_filter;
 279         }
 280
 281         /* Every thing else is unsupported */
 282         return -1;
 283 }
 284
 285 static void power9_config_bhrb(u64 pmu_bhrb_filter)
 286 {
 287         /* Enable BHRB filter in PMU */
 288         mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
 289 }
 290
 291 #define C(x)    PERF_COUNT_HW_CACHE_##x
 292
 293 /*
 294  * Table of generalized cache-related events.
 295  * 0 means not supported, -1 means nonsensical, other values
 296  * are event codes.
 297  */
 298 static int power9_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 299         [ C(L1D) ] = {
 300                 [ C(OP_READ) ] = {
 301                         [ C(RESULT_ACCESS) ] = PM_LD_REF_L1,
 302                         [ C(RESULT_MISS)   ] = PM_LD_MISS_L1_FIN,
 303                 },
 304                 [ C(OP_WRITE) ] = {
 305                         [ C(RESULT_ACCESS) ] = 0,
 306                         [ C(RESULT_MISS)   ] = PM_ST_MISS_L1,
 307                 },
 308                 [ C(OP_PREFETCH) ] = {
 309                         [ C(RESULT_ACCESS) ] = PM_L1_PREF,
 310                         [ C(RESULT_MISS)   ] = 0,
 311                 },
 312         },
 313         [ C(L1I) ] = {
 314                 [ C(OP_READ) ] = {
 315                         [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1,
 316                         [ C(RESULT_MISS)   ] = PM_L1_ICACHE_MISS,
 317                 },
 318                 [ C(OP_WRITE) ] = {
 319                         [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE,
 320                         [ C(RESULT_MISS)   ] = -1,
 321                 },
 322                 [ C(OP_PREFETCH) ] = {
 323                         [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE,
 324                         [ C(RESULT_MISS)   ] = 0,
 325                 },
 326         },
 327         [ C(LL) ] = {
 328                 [ C(OP_READ) ] = {
 329                         [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3,
 330                         [ C(RESULT_MISS)   ] = PM_DATA_FROM_L3MISS,
 331                 },
 332                 [ C(OP_WRITE) ] = {
 333                         [ C(RESULT_ACCESS) ] = PM_L2_ST,
 334                         [ C(RESULT_MISS)   ] = PM_L2_ST_MISS,
 335                 },
 336                 [ C(OP_PREFETCH) ] = {
 337                         [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL,
 338                         [ C(RESULT_MISS)   ] = 0,
 339                 },
 340         },
 341         [ C(DTLB) ] = {
 342                 [ C(OP_READ) ] = {
 343                         [ C(RESULT_ACCESS) ] = 0,
 344                         [ C(RESULT_MISS)   ] = PM_DTLB_MISS,
 345                 },
 346                 [ C(OP_WRITE) ] = {
 347                         [ C(RESULT_ACCESS) ] = -1,
 348                         [ C(RESULT_MISS)   ] = -1,
 349                 },
 350                 [ C(OP_PREFETCH) ] = {
 351                         [ C(RESULT_ACCESS) ] = -1,
 352                         [ C(RESULT_MISS)   ] = -1,
 353                 },
 354         },
 355         [ C(ITLB) ] = {
 356                 [ C(OP_READ) ] = {
 357                         [ C(RESULT_ACCESS) ] = 0,
 358                         [ C(RESULT_MISS)   ] = PM_ITLB_MISS,
 359                 },
 360                 [ C(OP_WRITE) ] = {
 361                         [ C(RESULT_ACCESS) ] = -1,
 362                         [ C(RESULT_MISS)   ] = -1,
 363                 },
 364                 [ C(OP_PREFETCH) ] = {
 365                         [ C(RESULT_ACCESS) ] = -1,
 366                         [ C(RESULT_MISS)   ] = -1,
 367                 },
 368         },
 369         [ C(BPU) ] = {
 370                 [ C(OP_READ) ] = {
 371                         [ C(RESULT_ACCESS) ] = PM_BRU_CMPL,
 372                         [ C(RESULT_MISS)   ] = PM_BR_MPRED_CMPL,
 373                 },
 374                 [ C(OP_WRITE) ] = {
 375                         [ C(RESULT_ACCESS) ] = -1,
 376                         [ C(RESULT_MISS)   ] = -1,
 377                 },
 378                 [ C(OP_PREFETCH) ] = {
 379                         [ C(RESULT_ACCESS) ] = -1,
 380                         [ C(RESULT_MISS)   ] = -1,
 381                 },
 382         },
 383         [ C(NODE) ] = {
 384                 [ C(OP_READ) ] = {
 385                         [ C(RESULT_ACCESS) ] = -1,
 386                         [ C(RESULT_MISS)   ] = -1,
 387                 },
 388                 [ C(OP_WRITE) ] = {
 389                         [ C(RESULT_ACCESS) ] = -1,
 390                         [ C(RESULT_MISS)   ] = -1,
 391                 },
 392                 [ C(OP_PREFETCH) ] = {
 393                         [ C(RESULT_ACCESS) ] = -1,
 394                         [ C(RESULT_MISS)   ] = -1,
 395                 },
 396         },
 397 };
 398
 399 #undef C
 400
 401 static struct power_pmu power9_isa207_pmu = {
 402         .name                   = "POWER9",
 403         .n_counter              = MAX_PMU_COUNTERS,
 404         .add_fields             = ISA207_ADD_FIELDS,
 405         .test_adder             = P9_DD1_TEST_ADDER,
 406         .compute_mmcr           = isa207_compute_mmcr,
 407         .config_bhrb            = power9_config_bhrb,
 408         .bhrb_filter_map        = power9_bhrb_filter_map,
 409         .get_constraint         = isa207_get_constraint,
 410         .get_alternatives       = power9_get_alternatives,
 411         .disable_pmc            = isa207_disable_pmc,
 412         .flags                  = PPMU_NO_SIAR | PPMU_ARCH_207S,
 413         .n_generic              = ARRAY_SIZE(power9_generic_events_dd1),
 414         .generic_events         = power9_generic_events_dd1,
 415         .cache_events           = &power9_cache_events,
 416         .attr_groups            = power9_isa207_pmu_attr_groups,
 417         .bhrb_nr                = 32,
 418 };
 419
 420 static struct power_pmu power9_pmu = {
 421         .name                   = "POWER9",
 422         .n_counter              = MAX_PMU_COUNTERS,
 423         .add_fields             = ISA207_ADD_FIELDS,
 424         .test_adder             = ISA207_TEST_ADDER,
 425         .compute_mmcr           = isa207_compute_mmcr,
 426         .config_bhrb            = power9_config_bhrb,
 427         .bhrb_filter_map        = power9_bhrb_filter_map,
 428         .get_constraint         = isa207_get_constraint,
 429         .get_alternatives       = power9_get_alternatives,
 430         .get_mem_data_src       = isa207_get_mem_data_src,
 431         .get_mem_weight         = isa207_get_mem_weight,
 432         .disable_pmc            = isa207_disable_pmc,
 433         .flags                  = PPMU_HAS_SIER | PPMU_ARCH_207S,
 434         .n_generic              = ARRAY_SIZE(power9_generic_events),
 435         .generic_events         = power9_generic_events,
 436         .cache_events           = &power9_cache_events,
 437         .attr_groups            = power9_pmu_attr_groups,
 438         .bhrb_nr                = 32,
 439 };
 440
 441 static int __init init_power9_pmu(void)
 442 {
 443         int rc = 0;
 444
 445         /* Comes from cpu_specs[] */
 446         if (!cur_cpu_spec->oprofile_cpu_type ||
 447             strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power9"))
 448                 return -ENODEV;
 449
 450         if (cpu_has_feature(CPU_FTR_POWER9_DD1)) {
 451                 /*
 452                  * Since PM_INST_CMPL may not provide right counts in all
 453                  * sampling scenarios in power9 DD1, instead use PM_INST_DISP.
 454                  */
 455                 EVENT_VAR(PM_INST_CMPL, _g).id = PM_INST_DISP;
 456                 /*
 457                  * Power9 DD1 should use PM_BR_CMPL_ALT event code for
 458                  * "branches" to provide correct counter value.
 459                  */
 460                 EVENT_VAR(PM_BRU_CMPL, _g).id = PM_BR_CMPL_ALT;
 461                 EVENT_VAR(PM_BRU_CMPL, _c).id = PM_BR_CMPL_ALT;
 462                 rc = register_power_pmu(&power9_isa207_pmu);
 463         } else {
 464                 rc = register_power_pmu(&power9_pmu);
 465         }
 466
 467         if (rc)
 468                 return rc;
 469
 470         /* Tell userspace that EBB is supported */
 471         cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
 472
 473         return 0;
 474 }
 475 early_initcall(init_power9_pmu);