arch/powerpc/perf/power7-pmu.c

   1 /*
   2  * Performance counter support for POWER7 processors.
   3  *
   4  * Copyright 2009 Paul Mackerras, IBM Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/perf_event.h>
  13 #include <linux/string.h>
  14 #include <asm/reg.h>
  15 #include <asm/cputable.h>
  16
  17 /*
  18  * Bits in event code for POWER7
  19  */
  20 #define PM_PMC_SH       16      /* PMC number (1-based) for direct events */
  21 #define PM_PMC_MSK      0xf
  22 #define PM_PMC_MSKS     (PM_PMC_MSK << PM_PMC_SH)
  23 #define PM_UNIT_SH      12      /* TTMMUX number and setting - unit select */
  24 #define PM_UNIT_MSK     0xf
  25 #define PM_COMBINE_SH   11      /* Combined event bit */
  26 #define PM_COMBINE_MSK  1
  27 #define PM_COMBINE_MSKS 0x800
  28 #define PM_L2SEL_SH     8       /* L2 event select */
  29 #define PM_L2SEL_MSK    7
  30 #define PM_PMCSEL_MSK   0xff
  31
  32 /*
  33  * Bits in MMCR1 for POWER7
  34  */
  35 #define MMCR1_TTM0SEL_SH        60
  36 #define MMCR1_TTM1SEL_SH        56
  37 #define MMCR1_TTM2SEL_SH        52
  38 #define MMCR1_TTM3SEL_SH        48
  39 #define MMCR1_TTMSEL_MSK        0xf
  40 #define MMCR1_L2SEL_SH          45
  41 #define MMCR1_L2SEL_MSK         7
  42 #define MMCR1_PMC1_COMBINE_SH   35
  43 #define MMCR1_PMC2_COMBINE_SH   34
  44 #define MMCR1_PMC3_COMBINE_SH   33
  45 #define MMCR1_PMC4_COMBINE_SH   32
  46 #define MMCR1_PMC1SEL_SH        24
  47 #define MMCR1_PMC2SEL_SH        16
  48 #define MMCR1_PMC3SEL_SH        8
  49 #define MMCR1_PMC4SEL_SH        0
  50 #define MMCR1_PMCSEL_SH(n)      (MMCR1_PMC1SEL_SH - (n) * 8)
  51 #define MMCR1_PMCSEL_MSK        0xff
  52
  53 /*
  54  * Layout of constraint bits:
  55  * 6666555555555544444444443333333333222222222211111111110000000000
  56  * 3210987654321098765432109876543210987654321098765432109876543210
  57  *                                              < ><  ><><><><><><>
  58  *                                              L2  NC P6P5P4P3P2P1
  59  *
  60  * L2 - 16-18 - Required L2SEL value (select field)
  61  *
  62  * NC - number of counters
  63  *     15: NC error 0x8000
  64  *     12-14: number of events needing PMC1-4 0x7000
  65  *
  66  * P6
  67  *     11: P6 error 0x800
  68  *     10-11: Count of events needing PMC6
  69  *
  70  * P1..P5
  71  *     0-9: Count of events needing PMC1..PMC5
  72  */
  73
  74 static int power7_get_constraint(u64 event, unsigned long *maskp,
  75                                  unsigned long *valp)
  76 {
  77         int pmc, sh, unit;
  78         unsigned long mask = 0, value = 0;
  79
  80         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
  81         if (pmc) {
  82                 if (pmc > 6)
  83                         return -1;
  84                 sh = (pmc - 1) * 2;
  85                 mask |= 2 << sh;
  86                 value |= 1 << sh;
  87                 if (pmc >= 5 && !(event == 0x500fa || event == 0x600f4))
  88                         return -1;
  89         }
  90         if (pmc < 5) {
  91                 /* need a counter from PMC1-4 set */
  92                 mask  |= 0x8000;
  93                 value |= 0x1000;
  94         }
  95
  96         unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
  97         if (unit == 6) {
  98                 /* L2SEL must be identical across events */
  99                 int l2sel = (event >> PM_L2SEL_SH) & PM_L2SEL_MSK;
 100                 mask  |= 0x7 << 16;
 101                 value |= l2sel << 16;
 102         }
 103
 104         *maskp = mask;
 105         *valp = value;
 106         return 0;
 107 }
 108
 109 #define MAX_ALT 2       /* at most 2 alternatives for any event */
 110
 111 static const unsigned int event_alternatives[][MAX_ALT] = {
 112         { 0x200f2, 0x300f2 },           /* PM_INST_DISP */
 113         { 0x200f4, 0x600f4 },           /* PM_RUN_CYC */
 114         { 0x400fa, 0x500fa },           /* PM_RUN_INST_CMPL */
 115 };
 116
 117 /*
 118  * Scan the alternatives table for a match and return the
 119  * index into the alternatives table if found, else -1.
 120  */
 121 static int find_alternative(u64 event)
 122 {
 123         int i, j;
 124
 125         for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
 126                 if (event < event_alternatives[i][0])
 127                         break;
 128                 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
 129                         if (event == event_alternatives[i][j])
 130                                 return i;
 131         }
 132         return -1;
 133 }
 134
 135 static s64 find_alternative_decode(u64 event)
 136 {
 137         int pmc, psel;
 138
 139         /* this only handles the 4x decode events */
 140         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 141         psel = event & PM_PMCSEL_MSK;
 142         if ((pmc == 2 || pmc == 4) && (psel & ~7) == 0x40)
 143                 return event - (1 << PM_PMC_SH) + 8;
 144         if ((pmc == 1 || pmc == 3) && (psel & ~7) == 0x48)
 145                 return event + (1 << PM_PMC_SH) - 8;
 146         return -1;
 147 }
 148
 149 static int power7_get_alternatives(u64 event, unsigned int flags, u64 alt[])
 150 {
 151         int i, j, nalt = 1;
 152         s64 ae;
 153
 154         alt[0] = event;
 155         nalt = 1;
 156         i = find_alternative(event);
 157         if (i >= 0) {
 158                 for (j = 0; j < MAX_ALT; ++j) {
 159                         ae = event_alternatives[i][j];
 160                         if (ae && ae != event)
 161                                 alt[nalt++] = ae;
 162                 }
 163         } else {
 164                 ae = find_alternative_decode(event);
 165                 if (ae > 0)
 166                         alt[nalt++] = ae;
 167         }
 168
 169         if (flags & PPMU_ONLY_COUNT_RUN) {
 170                 /*
 171                  * We're only counting in RUN state,
 172                  * so PM_CYC is equivalent to PM_RUN_CYC
 173                  * and PM_INST_CMPL === PM_RUN_INST_CMPL.
 174                  * This doesn't include alternatives that don't provide
 175                  * any extra flexibility in assigning PMCs.
 176                  */
 177                 j = nalt;
 178                 for (i = 0; i < nalt; ++i) {
 179                         switch (alt[i]) {
 180                         case 0x1e:      /* PM_CYC */
 181                                 alt[j++] = 0x600f4;     /* PM_RUN_CYC */
 182                                 break;
 183                         case 0x600f4:   /* PM_RUN_CYC */
 184                                 alt[j++] = 0x1e;
 185                                 break;
 186                         case 0x2:       /* PM_PPC_CMPL */
 187                                 alt[j++] = 0x500fa;     /* PM_RUN_INST_CMPL */
 188                                 break;
 189                         case 0x500fa:   /* PM_RUN_INST_CMPL */
 190                                 alt[j++] = 0x2; /* PM_PPC_CMPL */
 191                                 break;
 192                         }
 193                 }
 194                 nalt = j;
 195         }
 196
 197         return nalt;
 198 }
 199
 200 /*
 201  * Returns 1 if event counts things relating to marked instructions
 202  * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
 203  */
 204 static int power7_marked_instr_event(u64 event)
 205 {
 206         int pmc, psel;
 207         int unit;
 208
 209         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 210         unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
 211         psel = event & PM_PMCSEL_MSK & ~1;      /* trim off edge/level bit */
 212         if (pmc >= 5)
 213                 return 0;
 214
 215         switch (psel >> 4) {
 216         case 2:
 217                 return pmc == 2 || pmc == 4;
 218         case 3:
 219                 if (psel == 0x3c)
 220                         return pmc == 1;
 221                 if (psel == 0x3e)
 222                         return pmc != 2;
 223                 return 1;
 224         case 4:
 225         case 5:
 226                 return unit == 0xd;
 227         case 6:
 228                 if (psel == 0x64)
 229                         return pmc >= 3;
 230         case 8:
 231                 return unit == 0xd;
 232         }
 233         return 0;
 234 }
 235
 236 static int power7_compute_mmcr(u64 event[], int n_ev,
 237                                unsigned int hwc[], unsigned long mmcr[])
 238 {
 239         unsigned long mmcr1 = 0;
 240         unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS;
 241         unsigned int pmc, unit, combine, l2sel, psel;
 242         unsigned int pmc_inuse = 0;
 243         int i;
 244
 245         /* First pass to count resource use */
 246         for (i = 0; i < n_ev; ++i) {
 247                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 248                 if (pmc) {
 249                         if (pmc > 6)
 250                                 return -1;
 251                         if (pmc_inuse & (1 << (pmc - 1)))
 252                                 return -1;
 253                         pmc_inuse |= 1 << (pmc - 1);
 254                 }
 255         }
 256
 257         /* Second pass: assign PMCs, set all MMCR1 fields */
 258         for (i = 0; i < n_ev; ++i) {
 259                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 260                 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
 261                 combine = (event[i] >> PM_COMBINE_SH) & PM_COMBINE_MSK;
 262                 l2sel = (event[i] >> PM_L2SEL_SH) & PM_L2SEL_MSK;
 263                 psel = event[i] & PM_PMCSEL_MSK;
 264                 if (!pmc) {
 265                         /* Bus event or any-PMC direct event */
 266                         for (pmc = 0; pmc < 4; ++pmc) {
 267                                 if (!(pmc_inuse & (1 << pmc)))
 268                                         break;
 269                         }
 270                         if (pmc >= 4)
 271                                 return -1;
 272                         pmc_inuse |= 1 << pmc;
 273                 } else {
 274                         /* Direct or decoded event */
 275                         --pmc;
 276                 }
 277                 if (pmc <= 3) {
 278                         mmcr1 |= (unsigned long) unit
 279                                 << (MMCR1_TTM0SEL_SH - 4 * pmc);
 280                         mmcr1 |= (unsigned long) combine
 281                                 << (MMCR1_PMC1_COMBINE_SH - pmc);
 282                         mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
 283                         if (unit == 6)  /* L2 events */
 284                                 mmcr1 |= (unsigned long) l2sel
 285                                         << MMCR1_L2SEL_SH;
 286                 }
 287                 if (power7_marked_instr_event(event[i]))
 288                         mmcra |= MMCRA_SAMPLE_ENABLE;
 289                 hwc[i] = pmc;
 290         }
 291
 292         /* Return MMCRx values */
 293         mmcr[0] = 0;
 294         if (pmc_inuse & 1)
 295                 mmcr[0] = MMCR0_PMC1CE;
 296         if (pmc_inuse & 0x3e)
 297                 mmcr[0] |= MMCR0_PMCjCE;
 298         mmcr[1] = mmcr1;
 299         mmcr[2] = mmcra;
 300         return 0;
 301 }
 302
 303 static void power7_disable_pmc(unsigned int pmc, unsigned long mmcr[])
 304 {
 305         if (pmc <= 3)
 306                 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
 307 }
 308
 309 static int power7_generic_events[] = {
 310         [PERF_COUNT_HW_CPU_CYCLES] = 0x1e,
 311         [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x100f8, /* GCT_NOSLOT_CYC */
 312         [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x4000a,  /* CMPLU_STALL */
 313         [PERF_COUNT_HW_INSTRUCTIONS] = 2,
 314         [PERF_COUNT_HW_CACHE_REFERENCES] = 0xc880,      /* LD_REF_L1_LSU*/
 315         [PERF_COUNT_HW_CACHE_MISSES] = 0x400f0,         /* LD_MISS_L1   */
 316         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x10068,  /* BRU_FIN      */
 317         [PERF_COUNT_HW_BRANCH_MISSES] = 0x400f6,        /* BR_MPRED     */
 318 };
 319
 320 #define C(x)    PERF_COUNT_HW_CACHE_##x
 321
 322 /*
 323  * Table of generalized cache-related events.
 324  * 0 means not supported, -1 means nonsensical, other values
 325  * are event codes.
 326  */
 327 static int power7_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 328         [C(L1D)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 329                 [C(OP_READ)] = {        0xc880,         0x400f0 },
 330                 [C(OP_WRITE)] = {       0,              0x300f0 },
 331                 [C(OP_PREFETCH)] = {    0xd8b8,         0       },
 332         },
 333         [C(L1I)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 334                 [C(OP_READ)] = {        0,              0x200fc },
 335                 [C(OP_WRITE)] = {       -1,             -1      },
 336                 [C(OP_PREFETCH)] = {    0x408a,         0       },
 337         },
 338         [C(LL)] = {             /*      RESULT_ACCESS   RESULT_MISS */
 339                 [C(OP_READ)] = {        0x16080,        0x26080 },
 340                 [C(OP_WRITE)] = {       0x16082,        0x26082 },
 341                 [C(OP_PREFETCH)] = {    0,              0       },
 342         },
 343         [C(DTLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 344                 [C(OP_READ)] = {        0,              0x300fc },
 345                 [C(OP_WRITE)] = {       -1,             -1      },
 346                 [C(OP_PREFETCH)] = {    -1,             -1      },
 347         },
 348         [C(ITLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 349                 [C(OP_READ)] = {        0,              0x400fc },
 350                 [C(OP_WRITE)] = {       -1,             -1      },
 351                 [C(OP_PREFETCH)] = {    -1,             -1      },
 352         },
 353         [C(BPU)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 354                 [C(OP_READ)] = {        0x10068,        0x400f6 },
 355                 [C(OP_WRITE)] = {       -1,             -1      },
 356                 [C(OP_PREFETCH)] = {    -1,             -1      },
 357         },
 358         [C(NODE)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 359                 [C(OP_READ)] = {        -1,             -1      },
 360                 [C(OP_WRITE)] = {       -1,             -1      },
 361                 [C(OP_PREFETCH)] = {    -1,             -1      },
 362         },
 363 };
 364
 365 static struct power_pmu power7_pmu = {
 366         .name                   = "POWER7",
 367         .n_counter              = 6,
 368         .max_alternatives       = MAX_ALT + 1,
 369         .add_fields             = 0x1555ul,
 370         .test_adder             = 0x3000ul,
 371         .compute_mmcr           = power7_compute_mmcr,
 372         .get_constraint         = power7_get_constraint,
 373         .get_alternatives       = power7_get_alternatives,
 374         .disable_pmc            = power7_disable_pmc,
 375         .flags                  = PPMU_ALT_SIPR,
 376         .n_generic              = ARRAY_SIZE(power7_generic_events),
 377         .generic_events         = power7_generic_events,
 378         .cache_events           = &power7_cache_events,
 379 };
 380
 381 static int __init init_power7_pmu(void)
 382 {
 383         if (!cur_cpu_spec->oprofile_cpu_type ||
 384             strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power7"))
 385                 return -ENODEV;
 386
 387         if (pvr_version_is(PVR_POWER7p))
 388                 power7_pmu.flags |= PPMU_SIAR_VALID;
 389
 390         return register_power_pmu(&power7_pmu);
 391 }
 392
 393 early_initcall(init_power7_pmu);