arch/powerpc/perf/power9-pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Performance counter support for POWER9 processors.
   4  *
   5  * Copyright 2009 Paul Mackerras, IBM Corporation.
   6  * Copyright 2013 Michael Ellerman, IBM Corporation.
   7  * Copyright 2016 Madhavan Srinivasan, IBM Corporation.
   8  */
   9
  10 #define pr_fmt(fmt)     "power9-pmu: " fmt
  11
  12 #include "isa207-common.h"
  13
  14 /*
  15  * Raw event encoding for Power9:
  16  *
  17  *        60        56        52        48        44        40        36        32
  18  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
  19  *   | | [ ]                       [ ] [      thresh_cmp     ]   [  thresh_ctl   ]
  20  *   | |  |                         |                                     |
  21  *   | |  *- IFM (Linux)            |                  thresh start/stop -*
  22  *   | *- BHRB (Linux)              *sm
  23  *   *- EBB (Linux)
  24  *
  25  *        28        24        20        16        12         8         4         0
  26  * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - |
  27  *   [   ] [  sample ]   [cache]   [ pmc ]   [unit ]   []    m   [    pmcxsel    ]
  28  *     |        |           |                          |     |
  29  *     |        |           |                          |     *- mark
  30  *     |        |           *- L1/L2/L3 cache_sel      |
  31  *     |        |                                      |
  32  *     |        *- sampling mode for marked events     *- combine
  33  *     |
  34  *     *- thresh_sel
  35  *
  36  * Below uses IBM bit numbering.
  37  *
  38  * MMCR1[x:y] = unit    (PMCxUNIT)
  39  * MMCR1[24]   = pmc1combine[0]
  40  * MMCR1[25]   = pmc1combine[1]
  41  * MMCR1[26]   = pmc2combine[0]
  42  * MMCR1[27]   = pmc2combine[1]
  43  * MMCR1[28]   = pmc3combine[0]
  44  * MMCR1[29]   = pmc3combine[1]
  45  * MMCR1[30]   = pmc4combine[0]
  46  * MMCR1[31]   = pmc4combine[1]
  47  *
  48  * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011
  49  *      MMCR1[20:27] = thresh_ctl
  50  * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001
  51  *      MMCR1[20:27] = thresh_ctl
  52  * else
  53  *      MMCRA[48:55] = thresh_ctl   (THRESH START/END)
  54  *
  55  * if thresh_sel:
  56  *      MMCRA[45:47] = thresh_sel
  57  *
  58  * if thresh_cmp:
  59  *      MMCRA[9:11] = thresh_cmp[0:2]
  60  *      MMCRA[12:18] = thresh_cmp[3:9]
  61  *
  62  * MMCR1[16] = cache_sel[2]
  63  * MMCR1[17] = cache_sel[3]
  64  *
  65  * if mark:
  66  *      MMCRA[63]    = 1                (SAMPLE_ENABLE)
  67  *      MMCRA[57:59] = sample[0:2]      (RAND_SAMP_ELIG)
  68  *     MMCRA[61:62] = sample[3:4]      (RAND_SAMP_MODE)
  69  *
  70  * if EBB and BHRB:
  71  *      MMCRA[32:33] = IFM
  72  *
  73  * MMCRA[SDAR_MODE]  = sm
  74  */
  75
  76 /*
  77  * Some power9 event codes.
  78  */
  79 #define EVENT(_name, _code)     _name = _code,
  80
  81 enum {
  82 #include "power9-events-list.h"
  83 };
  84
  85 #undef EVENT
  86
  87 /* MMCRA IFM bits - POWER9 */
  88 #define POWER9_MMCRA_IFM1               0x0000000040000000UL
  89 #define POWER9_MMCRA_IFM2               0x0000000080000000UL
  90 #define POWER9_MMCRA_IFM3               0x00000000C0000000UL
  91 #define POWER9_MMCRA_BHRB_MASK          0x00000000C0000000UL
  92
  93 extern u64 PERF_REG_EXTENDED_MASK;
  94
  95 /* Nasty Power9 specific hack */
  96 #define PVR_POWER9_CUMULUS              0x00002000
  97
  98 /* PowerISA v2.07 format attribute structure*/
  99 extern struct attribute_group isa207_pmu_format_group;
 100
 101 int p9_dd21_bl_ev[] = {
 102         PM_MRK_ST_DONE_L2,
 103         PM_RADIX_PWC_L1_HIT,
 104         PM_FLOP_CMPL,
 105         PM_MRK_NTF_FIN,
 106         PM_RADIX_PWC_L2_HIT,
 107         PM_IFETCH_THROTTLE,
 108         PM_MRK_L2_TM_ST_ABORT_SISTER,
 109         PM_RADIX_PWC_L3_HIT,
 110         PM_RUN_CYC_SMT2_MODE,
 111         PM_TM_TX_PASS_RUN_INST,
 112         PM_DISP_HELD_SYNC_HOLD,
 113 };
 114
 115 int p9_dd22_bl_ev[] = {
 116         PM_DTLB_MISS_16G,
 117         PM_DERAT_MISS_2M,
 118         PM_DTLB_MISS_2M,
 119         PM_MRK_DTLB_MISS_1G,
 120         PM_DTLB_MISS_4K,
 121         PM_DERAT_MISS_1G,
 122         PM_MRK_DERAT_MISS_2M,
 123         PM_MRK_DTLB_MISS_4K,
 124         PM_MRK_DTLB_MISS_16G,
 125         PM_DTLB_MISS_64K,
 126         PM_MRK_DERAT_MISS_1G,
 127         PM_MRK_DTLB_MISS_64K,
 128         PM_DISP_HELD_SYNC_HOLD,
 129         PM_DTLB_MISS_16M,
 130         PM_DTLB_MISS_1G,
 131         PM_MRK_DTLB_MISS_16M,
 132 };
 133
 134 /* Table of alternatives, sorted by column 0 */
 135 static const unsigned int power9_event_alternatives[][MAX_ALT] = {
 136         { PM_INST_DISP,                 PM_INST_DISP_ALT },
 137         { PM_RUN_CYC_ALT,               PM_RUN_CYC },
 138         { PM_RUN_INST_CMPL_ALT,         PM_RUN_INST_CMPL },
 139         { PM_LD_MISS_L1,                PM_LD_MISS_L1_ALT },
 140         { PM_BR_2PATH,                  PM_BR_2PATH_ALT },
 141 };
 142
 143 static int power9_get_alternatives(u64 event, unsigned int flags, u64 alt[])
 144 {
 145         int num_alt = 0;
 146
 147         num_alt = isa207_get_alternatives(event, alt,
 148                                           ARRAY_SIZE(power9_event_alternatives), flags,
 149                                           power9_event_alternatives);
 150
 151         return num_alt;
 152 }
 153
 154 GENERIC_EVENT_ATTR(cpu-cycles,                  PM_CYC);
 155 GENERIC_EVENT_ATTR(stalled-cycles-frontend,     PM_ICT_NOSLOT_CYC);
 156 GENERIC_EVENT_ATTR(stalled-cycles-backend,      PM_CMPLU_STALL);
 157 GENERIC_EVENT_ATTR(instructions,                PM_INST_CMPL);
 158 GENERIC_EVENT_ATTR(branch-instructions,         PM_BR_CMPL);
 159 GENERIC_EVENT_ATTR(branch-misses,               PM_BR_MPRED_CMPL);
 160 GENERIC_EVENT_ATTR(cache-references,            PM_LD_REF_L1);
 161 GENERIC_EVENT_ATTR(cache-misses,                PM_LD_MISS_L1_FIN);
 162 GENERIC_EVENT_ATTR(mem-loads,                   MEM_LOADS);
 163 GENERIC_EVENT_ATTR(mem-stores,                  MEM_STORES);
 164
 165 CACHE_EVENT_ATTR(L1-dcache-load-misses,         PM_LD_MISS_L1_FIN);
 166 CACHE_EVENT_ATTR(L1-dcache-loads,               PM_LD_REF_L1);
 167 CACHE_EVENT_ATTR(L1-dcache-prefetches,          PM_L1_PREF);
 168 CACHE_EVENT_ATTR(L1-dcache-store-misses,        PM_ST_MISS_L1);
 169 CACHE_EVENT_ATTR(L1-icache-load-misses,         PM_L1_ICACHE_MISS);
 170 CACHE_EVENT_ATTR(L1-icache-loads,               PM_INST_FROM_L1);
 171 CACHE_EVENT_ATTR(L1-icache-prefetches,          PM_IC_PREF_WRITE);
 172 CACHE_EVENT_ATTR(LLC-load-misses,               PM_DATA_FROM_L3MISS);
 173 CACHE_EVENT_ATTR(LLC-loads,                     PM_DATA_FROM_L3);
 174 CACHE_EVENT_ATTR(LLC-prefetches,                PM_L3_PREF_ALL);
 175 CACHE_EVENT_ATTR(branch-load-misses,            PM_BR_MPRED_CMPL);
 176 CACHE_EVENT_ATTR(branch-loads,                  PM_BR_CMPL);
 177 CACHE_EVENT_ATTR(dTLB-load-misses,              PM_DTLB_MISS);
 178 CACHE_EVENT_ATTR(iTLB-load-misses,              PM_ITLB_MISS);
 179
 180 static struct attribute *power9_events_attr[] = {
 181         GENERIC_EVENT_PTR(PM_CYC),
 182         GENERIC_EVENT_PTR(PM_ICT_NOSLOT_CYC),
 183         GENERIC_EVENT_PTR(PM_CMPLU_STALL),
 184         GENERIC_EVENT_PTR(PM_INST_CMPL),
 185         GENERIC_EVENT_PTR(PM_BR_CMPL),
 186         GENERIC_EVENT_PTR(PM_BR_MPRED_CMPL),
 187         GENERIC_EVENT_PTR(PM_LD_REF_L1),
 188         GENERIC_EVENT_PTR(PM_LD_MISS_L1_FIN),
 189         GENERIC_EVENT_PTR(MEM_LOADS),
 190         GENERIC_EVENT_PTR(MEM_STORES),
 191         CACHE_EVENT_PTR(PM_LD_MISS_L1_FIN),
 192         CACHE_EVENT_PTR(PM_LD_REF_L1),
 193         CACHE_EVENT_PTR(PM_L1_PREF),
 194         CACHE_EVENT_PTR(PM_ST_MISS_L1),
 195         CACHE_EVENT_PTR(PM_L1_ICACHE_MISS),
 196         CACHE_EVENT_PTR(PM_INST_FROM_L1),
 197         CACHE_EVENT_PTR(PM_IC_PREF_WRITE),
 198         CACHE_EVENT_PTR(PM_DATA_FROM_L3MISS),
 199         CACHE_EVENT_PTR(PM_DATA_FROM_L3),
 200         CACHE_EVENT_PTR(PM_L3_PREF_ALL),
 201         CACHE_EVENT_PTR(PM_BR_MPRED_CMPL),
 202         CACHE_EVENT_PTR(PM_BR_CMPL),
 203         CACHE_EVENT_PTR(PM_DTLB_MISS),
 204         CACHE_EVENT_PTR(PM_ITLB_MISS),
 205         NULL
 206 };
 207
 208 static struct attribute_group power9_pmu_events_group = {
 209         .name = "events",
 210         .attrs = power9_events_attr,
 211 };
 212
 213 PMU_FORMAT_ATTR(event,          "config:0-51");
 214 PMU_FORMAT_ATTR(pmcxsel,        "config:0-7");
 215 PMU_FORMAT_ATTR(mark,           "config:8");
 216 PMU_FORMAT_ATTR(combine,        "config:10-11");
 217 PMU_FORMAT_ATTR(unit,           "config:12-15");
 218 PMU_FORMAT_ATTR(pmc,            "config:16-19");
 219 PMU_FORMAT_ATTR(cache_sel,      "config:20-23");
 220 PMU_FORMAT_ATTR(sample_mode,    "config:24-28");
 221 PMU_FORMAT_ATTR(thresh_sel,     "config:29-31");
 222 PMU_FORMAT_ATTR(thresh_stop,    "config:32-35");
 223 PMU_FORMAT_ATTR(thresh_start,   "config:36-39");
 224 PMU_FORMAT_ATTR(thresh_cmp,     "config:40-49");
 225 PMU_FORMAT_ATTR(sdar_mode,      "config:50-51");
 226
 227 static struct attribute *power9_pmu_format_attr[] = {
 228         &format_attr_event.attr,
 229         &format_attr_pmcxsel.attr,
 230         &format_attr_mark.attr,
 231         &format_attr_combine.attr,
 232         &format_attr_unit.attr,
 233         &format_attr_pmc.attr,
 234         &format_attr_cache_sel.attr,
 235         &format_attr_sample_mode.attr,
 236         &format_attr_thresh_sel.attr,
 237         &format_attr_thresh_stop.attr,
 238         &format_attr_thresh_start.attr,
 239         &format_attr_thresh_cmp.attr,
 240         &format_attr_sdar_mode.attr,
 241         NULL,
 242 };
 243
 244 static struct attribute_group power9_pmu_format_group = {
 245         .name = "format",
 246         .attrs = power9_pmu_format_attr,
 247 };
 248
 249 static const struct attribute_group *power9_pmu_attr_groups[] = {
 250         &power9_pmu_format_group,
 251         &power9_pmu_events_group,
 252         NULL,
 253 };
 254
 255 static int power9_generic_events[] = {
 256         [PERF_COUNT_HW_CPU_CYCLES] =                    PM_CYC,
 257         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] =       PM_ICT_NOSLOT_CYC,
 258         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] =        PM_CMPLU_STALL,
 259         [PERF_COUNT_HW_INSTRUCTIONS] =                  PM_INST_CMPL,
 260         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] =           PM_BR_CMPL,
 261         [PERF_COUNT_HW_BRANCH_MISSES] =                 PM_BR_MPRED_CMPL,
 262         [PERF_COUNT_HW_CACHE_REFERENCES] =              PM_LD_REF_L1,
 263         [PERF_COUNT_HW_CACHE_MISSES] =                  PM_LD_MISS_L1_FIN,
 264 };
 265
 266 static u64 power9_bhrb_filter_map(u64 branch_sample_type)
 267 {
 268         u64 pmu_bhrb_filter = 0;
 269
 270         /* BHRB and regular PMU events share the same privilege state
 271          * filter configuration. BHRB is always recorded along with a
 272          * regular PMU event. As the privilege state filter is handled
 273          * in the basic PMC configuration of the accompanying regular
 274          * PMU event, we ignore any separate BHRB specific request.
 275          */
 276
 277         /* No branch filter requested */
 278         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY)
 279                 return pmu_bhrb_filter;
 280
 281         /* Invalid branch filter options - HW does not support */
 282         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN)
 283                 return -1;
 284
 285         if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL)
 286                 return -1;
 287
 288         if (branch_sample_type & PERF_SAMPLE_BRANCH_CALL)
 289                 return -1;
 290
 291         if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) {
 292                 pmu_bhrb_filter |= POWER9_MMCRA_IFM1;
 293                 return pmu_bhrb_filter;
 294         }
 295
 296         /* Every thing else is unsupported */
 297         return -1;
 298 }
 299
 300 static void power9_config_bhrb(u64 pmu_bhrb_filter)
 301 {
 302         pmu_bhrb_filter &= POWER9_MMCRA_BHRB_MASK;
 303
 304         /* Enable BHRB filter in PMU */
 305         mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter));
 306 }
 307
 308 #define C(x)    PERF_COUNT_HW_CACHE_##x
 309
 310 /*
 311  * Table of generalized cache-related events.
 312  * 0 means not supported, -1 means nonsensical, other values
 313  * are event codes.
 314  */
 315 static u64 power9_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 316         [ C(L1D) ] = {
 317                 [ C(OP_READ) ] = {
 318                         [ C(RESULT_ACCESS) ] = PM_LD_REF_L1,
 319                         [ C(RESULT_MISS)   ] = PM_LD_MISS_L1_FIN,
 320                 },
 321                 [ C(OP_WRITE) ] = {
 322                         [ C(RESULT_ACCESS) ] = 0,
 323                         [ C(RESULT_MISS)   ] = PM_ST_MISS_L1,
 324                 },
 325                 [ C(OP_PREFETCH) ] = {
 326                         [ C(RESULT_ACCESS) ] = PM_L1_PREF,
 327                         [ C(RESULT_MISS)   ] = 0,
 328                 },
 329         },
 330         [ C(L1I) ] = {
 331                 [ C(OP_READ) ] = {
 332                         [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1,
 333                         [ C(RESULT_MISS)   ] = PM_L1_ICACHE_MISS,
 334                 },
 335                 [ C(OP_WRITE) ] = {
 336                         [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE,
 337                         [ C(RESULT_MISS)   ] = -1,
 338                 },
 339                 [ C(OP_PREFETCH) ] = {
 340                         [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE,
 341                         [ C(RESULT_MISS)   ] = 0,
 342                 },
 343         },
 344         [ C(LL) ] = {
 345                 [ C(OP_READ) ] = {
 346                         [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3,
 347                         [ C(RESULT_MISS)   ] = PM_DATA_FROM_L3MISS,
 348                 },
 349                 [ C(OP_WRITE) ] = {
 350                         [ C(RESULT_ACCESS) ] = 0,
 351                         [ C(RESULT_MISS)   ] = 0,
 352                 },
 353                 [ C(OP_PREFETCH) ] = {
 354                         [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL,
 355                         [ C(RESULT_MISS)   ] = 0,
 356                 },
 357         },
 358         [ C(DTLB) ] = {
 359                 [ C(OP_READ) ] = {
 360                         [ C(RESULT_ACCESS) ] = 0,
 361                         [ C(RESULT_MISS)   ] = PM_DTLB_MISS,
 362                 },
 363                 [ C(OP_WRITE) ] = {
 364                         [ C(RESULT_ACCESS) ] = -1,
 365                         [ C(RESULT_MISS)   ] = -1,
 366                 },
 367                 [ C(OP_PREFETCH) ] = {
 368                         [ C(RESULT_ACCESS) ] = -1,
 369                         [ C(RESULT_MISS)   ] = -1,
 370                 },
 371         },
 372         [ C(ITLB) ] = {
 373                 [ C(OP_READ) ] = {
 374                         [ C(RESULT_ACCESS) ] = 0,
 375                         [ C(RESULT_MISS)   ] = PM_ITLB_MISS,
 376                 },
 377                 [ C(OP_WRITE) ] = {
 378                         [ C(RESULT_ACCESS) ] = -1,
 379                         [ C(RESULT_MISS)   ] = -1,
 380                 },
 381                 [ C(OP_PREFETCH) ] = {
 382                         [ C(RESULT_ACCESS) ] = -1,
 383                         [ C(RESULT_MISS)   ] = -1,
 384                 },
 385         },
 386         [ C(BPU) ] = {
 387                 [ C(OP_READ) ] = {
 388                         [ C(RESULT_ACCESS) ] = PM_BR_CMPL,
 389                         [ C(RESULT_MISS)   ] = PM_BR_MPRED_CMPL,
 390                 },
 391                 [ C(OP_WRITE) ] = {
 392                         [ C(RESULT_ACCESS) ] = -1,
 393                         [ C(RESULT_MISS)   ] = -1,
 394                 },
 395                 [ C(OP_PREFETCH) ] = {
 396                         [ C(RESULT_ACCESS) ] = -1,
 397                         [ C(RESULT_MISS)   ] = -1,
 398                 },
 399         },
 400         [ C(NODE) ] = {
 401                 [ C(OP_READ) ] = {
 402                         [ C(RESULT_ACCESS) ] = -1,
 403                         [ C(RESULT_MISS)   ] = -1,
 404                 },
 405                 [ C(OP_WRITE) ] = {
 406                         [ C(RESULT_ACCESS) ] = -1,
 407                         [ C(RESULT_MISS)   ] = -1,
 408                 },
 409                 [ C(OP_PREFETCH) ] = {
 410                         [ C(RESULT_ACCESS) ] = -1,
 411                         [ C(RESULT_MISS)   ] = -1,
 412                 },
 413         },
 414 };
 415
 416 #undef C
 417
 418 static struct power_pmu power9_pmu = {
 419         .name                   = "POWER9",
 420         .n_counter              = MAX_PMU_COUNTERS,
 421         .add_fields             = ISA207_ADD_FIELDS,
 422         .test_adder             = ISA207_TEST_ADDER,
 423         .group_constraint_mask  = CNST_CACHE_PMC4_MASK,
 424         .group_constraint_val   = CNST_CACHE_PMC4_VAL,
 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         .capabilities           = PERF_PMU_CAP_EXTENDED_REGS,
 440 };
 441
 442 int init_power9_pmu(void)
 443 {
 444         int rc = 0;
 445         unsigned int pvr = mfspr(SPRN_PVR);
 446
 447         /* Comes from cpu_specs[] */
 448         if (!cur_cpu_spec->oprofile_cpu_type ||
 449             strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power9"))
 450                 return -ENODEV;
 451
 452         /* Blacklist events */
 453         if (!(pvr & PVR_POWER9_CUMULUS)) {
 454                 if ((PVR_CFG(pvr) == 2) && (PVR_MIN(pvr) == 1)) {
 455                         power9_pmu.blacklist_ev = p9_dd21_bl_ev;
 456                         power9_pmu.n_blacklist_ev = ARRAY_SIZE(p9_dd21_bl_ev);
 457                 } else if ((PVR_CFG(pvr) == 2) && (PVR_MIN(pvr) == 2)) {
 458                         power9_pmu.blacklist_ev = p9_dd22_bl_ev;
 459                         power9_pmu.n_blacklist_ev = ARRAY_SIZE(p9_dd22_bl_ev);
 460                 }
 461         }
 462
 463         /* Set the PERF_REG_EXTENDED_MASK here */
 464         PERF_REG_EXTENDED_MASK = PERF_REG_PMU_MASK_300;
 465
 466         rc = register_power_pmu(&power9_pmu);
 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 }