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  * Some power9 event codes.
  20  */
  21 #define EVENT(_name, _code)     _name = _code,
  22
  23 enum {
  24 #include "power9-events-list.h"
  25 };
  26
  27 #undef EVENT
  28
  29 /* MMCRA IFM bits - POWER9 */
  30 #define POWER9_MMCRA_IFM1               0x0000000040000000UL
  31 #define POWER9_MMCRA_IFM2               0x0000000080000000UL
  32 #define POWER9_MMCRA_IFM3               0x00000000C0000000UL
  33
  34 GENERIC_EVENT_ATTR(cpu-cycles,                  PM_CYC);
  35 GENERIC_EVENT_ATTR(stalled-cycles-frontend,     PM_ICT_NOSLOT_CYC);
  36 GENERIC_EVENT_ATTR(stalled-cycles-backend,      PM_CMPLU_STALL);
  37 GENERIC_EVENT_ATTR(instructions,                PM_INST_CMPL);
  38 GENERIC_EVENT_ATTR(branch-instructions,         PM_BRU_CMPL);
  39 GENERIC_EVENT_ATTR(branch-misses,               PM_BR_MPRED_CMPL);
  40 GENERIC_EVENT_ATTR(cache-references,            PM_LD_REF_L1);
  41 GENERIC_EVENT_ATTR(cache-misses,                PM_LD_MISS_L1_FIN);
  42
  43 CACHE_EVENT_ATTR(L1-dcache-load-misses,         PM_LD_MISS_L1_FIN);
  44 CACHE_EVENT_ATTR(L1-dcache-loads,               PM_LD_REF_L1);
  45 CACHE_EVENT_ATTR(L1-dcache-prefetches,          PM_L1_PREF);
  46 CACHE_EVENT_ATTR(L1-dcache-store-misses,        PM_ST_MISS_L1);
  47 CACHE_EVENT_ATTR(L1-icache-load-misses,         PM_L1_ICACHE_MISS);
  48 CACHE_EVENT_ATTR(L1-icache-loads,               PM_INST_FROM_L1);
  49 CACHE_EVENT_ATTR(L1-icache-prefetches,          PM_IC_PREF_WRITE);
  50 CACHE_EVENT_ATTR(LLC-load-misses,               PM_DATA_FROM_L3MISS);
  51 CACHE_EVENT_ATTR(LLC-loads,                     PM_DATA_FROM_L3);
  52 CACHE_EVENT_ATTR(LLC-prefetches,                PM_L3_PREF_ALL);
  53 CACHE_EVENT_ATTR(LLC-store-misses,              PM_L2_ST_MISS);
  54 CACHE_EVENT_ATTR(LLC-stores,                    PM_L2_ST);
  55 CACHE_EVENT_ATTR(branch-load-misses,            PM_BR_MPRED_CMPL);
  56 CACHE_EVENT_ATTR(branch-loads,                  PM_BRU_CMPL);
  57 CACHE_EVENT_ATTR(dTLB-load-misses,              PM_DTLB_MISS);
  58 CACHE_EVENT_ATTR(iTLB-load-misses,              PM_ITLB_MISS);
  59
  60 static struct attribute *power9_events_attr[] = {
  61         GENERIC_EVENT_PTR(PM_CYC),
  62         GENERIC_EVENT_PTR(PM_ICT_NOSLOT_CYC),
  63         GENERIC_EVENT_PTR(PM_CMPLU_STALL),
  64         GENERIC_EVENT_PTR(PM_INST_CMPL),
  65         GENERIC_EVENT_PTR(PM_BRU_CMPL),
  66         GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
  67         GENERIC_EVENT_PTR(PM_LD_REF_L1),
  68         GENERIC_EVENT_PTR(PM_LD_MISS_L1_FIN),
  69         CACHE_EVENT_PTR(PM_LD_MISS_L1_FIN),
  70         CACHE_EVENT_PTR(PM_LD_REF_L1),
  71         CACHE_EVENT_PTR(PM_L1_PREF),
  72         CACHE_EVENT_PTR(PM_ST_MISS_L1),
  73         CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
  74         CACHE_EVENT_PTR(PM_INST_FROM_L1),
  75         CACHE_EVENT_PTR(PM_IC_PREF_WRITE),
  76         CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
  77         CACHE_EVENT_PTR(PM_DATA_FROM_L3),
  78         CACHE_EVENT_PTR(PM_L3_PREF_ALL),
  79         CACHE_EVENT_PTR(PM_L2_ST_MISS),
  80         CACHE_EVENT_PTR(PM_L2_ST),
  81         CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
  82         CACHE_EVENT_PTR(PM_BRU_CMPL),
  83         CACHE_EVENT_PTR(PM_DTLB_MISS),
  84         CACHE_EVENT_PTR(PM_ITLB_MISS),
  85         NULL
  86 };
  87
  88 static struct attribute_group power9_pmu_events_group = {
  89         .name = "events",
  90         .attrs = power9_events_attr,
  91 };
  92
  93 PMU_FORMAT_ATTR(event,          "config:0-49");
  94 PMU_FORMAT_ATTR(pmcxsel,        "config:0-7");
  95 PMU_FORMAT_ATTR(mark,           "config:8");
  96 PMU_FORMAT_ATTR(combine,        "config:11");
  97 PMU_FORMAT_ATTR(unit,           "config:12-15");
  98 PMU_FORMAT_ATTR(pmc,            "config:16-19");
  99 PMU_FORMAT_ATTR(cache_sel,      "config:20-23");
 100 PMU_FORMAT_ATTR(sample_mode,    "config:24-28");
 101 PMU_FORMAT_ATTR(thresh_sel,     "config:29-31");
 102 PMU_FORMAT_ATTR(thresh_stop,    "config:32-35");
 103 PMU_FORMAT_ATTR(thresh_start,   "config:36-39");
 104 PMU_FORMAT_ATTR(thresh_cmp,     "config:40-49");
 105
 106 static struct attribute *power9_pmu_format_attr[] = {
 107         &format_attr_event.attr,
 108         &format_attr_pmcxsel.attr,
 109         &format_attr_mark.attr,
 110         &format_attr_combine.attr,
 111         &format_attr_unit.attr,
 112         &format_attr_pmc.attr,
 113         &format_attr_cache_sel.attr,
 114         &format_attr_sample_mode.attr,
 115         &format_attr_thresh_sel.attr,
 116         &format_attr_thresh_stop.attr,
 117         &format_attr_thresh_start.attr,
 118         &format_attr_thresh_cmp.attr,
 119         NULL,
 120 };
 121
 122 struct attribute_group power9_pmu_format_group = {
 123         .name = "format",
 124         .attrs = power9_pmu_format_attr,
 125 };
 126
 127 static const struct attribute_group *power9_pmu_attr_groups[] = {
 128         &power9_pmu_format_group,
 129         &power9_pmu_events_group,
 130         NULL,
 131 };
 132
 133 static int power9_generic_events[] = {
 134         [PERF_COUNT_HW_CPU_CYCLES] =                    PM_CYC,
 135         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =       PM_ICT_NOSLOT_CYC,
 136         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =        PM_CMPLU_STALL,
 137         [PERF_COUNT_HW_INSTRUCTIONS] =                  PM_INST_CMPL,
 138         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =           PM_BRU_CMPL,
 139         [PERF_COUNT_HW_BRANCH_MISSES] =                 PM_BR_MPRED_CMPL,
 140         [PERF_COUNT_HW_CACHE_REFERENCES] =              PM_LD_REF_L1,
 141         [PERF_COUNT_HW_CACHE_MISSES] =                  PM_LD_MISS_L1_FIN,
 142 };
 143
 144 static u64 power9_bhrb_filter_map(u64 branch_sample_type)
 145 {
 146         u64 pmu_bhrb_filter = 0;
 147
 148         /* BHRB and regular PMU events share the same privilege state
 149          * filter configuration. BHRB is always recorded along with a
 150          * regular PMU event. As the privilege state filter is handled
 151          * in the basic PMC configuration of the accompanying regular
 152          * PMU event, we ignore any separate BHRB specific request.
 153          */
 154
 155         /* No branch filter requested */
 156         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
 157                 return pmu_bhrb_filter;
 158
 159         /* Invalid branch filter options - HW does not support */
 160         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
 161                 return -1;
 162
 163         if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
 164                 return -1;
 165
 166         if (branch_sample_type & PERF_SAMPLE_BRANCH_CALL)
 167                 return -1;
 168
 169         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
 170                 pmu_bhrb_filter |= POWER9_MMCRA_IFM1;
 171                 return pmu_bhrb_filter;
 172         }
 173
 174         /* Every thing else is unsupported */
 175         return -1;
 176 }
 177
 178 static void power9_config_bhrb(u64 pmu_bhrb_filter)
 179 {
 180         /* Enable BHRB filter in PMU */
 181         mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
 182 }
 183
 184 #define C(x)    PERF_COUNT_HW_CACHE_##x
 185
 186 /*
 187  * Table of generalized cache-related events.
 188  * 0 means not supported, -1 means nonsensical, other values
 189  * are event codes.
 190  */
 191 static int power9_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 192         [ C(L1D) ] = {
 193                 [ C(OP_READ) ] = {
 194                         [ C(RESULT_ACCESS) ] = PM_LD_REF_L1,
 195                         [ C(RESULT_MISS)   ] = PM_LD_MISS_L1_FIN,
 196                 },
 197                 [ C(OP_WRITE) ] = {
 198                         [ C(RESULT_ACCESS) ] = 0,
 199                         [ C(RESULT_MISS)   ] = PM_ST_MISS_L1,
 200                 },
 201                 [ C(OP_PREFETCH) ] = {
 202                         [ C(RESULT_ACCESS) ] = PM_L1_PREF,
 203                         [ C(RESULT_MISS)   ] = 0,
 204                 },
 205         },
 206         [ C(L1I) ] = {
 207                 [ C(OP_READ) ] = {
 208                         [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1,
 209                         [ C(RESULT_MISS)   ] = PM_L1_ICACHE_MISS,
 210                 },
 211                 [ C(OP_WRITE) ] = {
 212                         [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE,
 213                         [ C(RESULT_MISS)   ] = -1,
 214                 },
 215                 [ C(OP_PREFETCH) ] = {
 216                         [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE,
 217                         [ C(RESULT_MISS)   ] = 0,
 218                 },
 219         },
 220         [ C(LL) ] = {
 221                 [ C(OP_READ) ] = {
 222                         [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3,
 223                         [ C(RESULT_MISS)   ] = PM_DATA_FROM_L3MISS,
 224                 },
 225                 [ C(OP_WRITE) ] = {
 226                         [ C(RESULT_ACCESS) ] = PM_L2_ST,
 227                         [ C(RESULT_MISS)   ] = PM_L2_ST_MISS,
 228                 },
 229                 [ C(OP_PREFETCH) ] = {
 230                         [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL,
 231                         [ C(RESULT_MISS)   ] = 0,
 232                 },
 233         },
 234         [ C(DTLB) ] = {
 235                 [ C(OP_READ) ] = {
 236                         [ C(RESULT_ACCESS) ] = 0,
 237                         [ C(RESULT_MISS)   ] = PM_DTLB_MISS,
 238                 },
 239                 [ C(OP_WRITE) ] = {
 240                         [ C(RESULT_ACCESS) ] = -1,
 241                         [ C(RESULT_MISS)   ] = -1,
 242                 },
 243                 [ C(OP_PREFETCH) ] = {
 244                         [ C(RESULT_ACCESS) ] = -1,
 245                         [ C(RESULT_MISS)   ] = -1,
 246                 },
 247         },
 248         [ C(ITLB) ] = {
 249                 [ C(OP_READ) ] = {
 250                         [ C(RESULT_ACCESS) ] = 0,
 251                         [ C(RESULT_MISS)   ] = PM_ITLB_MISS,
 252                 },
 253                 [ C(OP_WRITE) ] = {
 254                         [ C(RESULT_ACCESS) ] = -1,
 255                         [ C(RESULT_MISS)   ] = -1,
 256                 },
 257                 [ C(OP_PREFETCH) ] = {
 258                         [ C(RESULT_ACCESS) ] = -1,
 259                         [ C(RESULT_MISS)   ] = -1,
 260                 },
 261         },
 262         [ C(BPU) ] = {
 263                 [ C(OP_READ) ] = {
 264                         [ C(RESULT_ACCESS) ] = PM_BRU_CMPL,
 265                         [ C(RESULT_MISS)   ] = PM_BR_MPRED_CMPL,
 266                 },
 267                 [ C(OP_WRITE) ] = {
 268                         [ C(RESULT_ACCESS) ] = -1,
 269                         [ C(RESULT_MISS)   ] = -1,
 270                 },
 271                 [ C(OP_PREFETCH) ] = {
 272                         [ C(RESULT_ACCESS) ] = -1,
 273                         [ C(RESULT_MISS)   ] = -1,
 274                 },
 275         },
 276         [ C(NODE) ] = {
 277                 [ C(OP_READ) ] = {
 278                         [ C(RESULT_ACCESS) ] = -1,
 279                         [ C(RESULT_MISS)   ] = -1,
 280                 },
 281                 [ C(OP_WRITE) ] = {
 282                         [ C(RESULT_ACCESS) ] = -1,
 283                         [ C(RESULT_MISS)   ] = -1,
 284                 },
 285                 [ C(OP_PREFETCH) ] = {
 286                         [ C(RESULT_ACCESS) ] = -1,
 287                         [ C(RESULT_MISS)   ] = -1,
 288                 },
 289         },
 290 };
 291
 292 #undef C
 293
 294 static struct power_pmu power9_pmu = {
 295         .name                   = "POWER9",
 296         .n_counter              = MAX_PMU_COUNTERS,
 297         .add_fields             = ISA207_ADD_FIELDS,
 298         .test_adder             = ISA207_TEST_ADDER,
 299         .compute_mmcr           = isa207_compute_mmcr,
 300         .config_bhrb            = power9_config_bhrb,
 301         .bhrb_filter_map        = power9_bhrb_filter_map,
 302         .get_constraint         = isa207_get_constraint,
 303         .disable_pmc            = isa207_disable_pmc,
 304         .flags                  = PPMU_HAS_SIER | PPMU_ARCH_207S,
 305         .n_generic              = ARRAY_SIZE(power9_generic_events),
 306         .generic_events         = power9_generic_events,
 307         .cache_events           = &power9_cache_events,
 308         .attr_groups            = power9_pmu_attr_groups,
 309         .bhrb_nr                = 32,
 310 };
 311
 312 static int __init init_power9_pmu(void)
 313 {
 314         int rc;
 315
 316         /* Comes from cpu_specs[] */
 317         if (!cur_cpu_spec->oprofile_cpu_type ||
 318             strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power9"))
 319                 return -ENODEV;
 320
 321         rc = register_power_pmu(&power9_pmu);
 322         if (rc)
 323                 return rc;
 324
 325         /* Tell userspace that EBB is supported */
 326         cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB;
 327
 328         return 0;
 329 }
 330 early_initcall(init_power9_pmu);