git.oblomov.eu Git - linux-2.6/blob - arch/powerpc/kernel/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_counter.h>
  13 #include <asm/reg.h>
  14
  15 /*
  16  * Bits in event code for POWER7
  17  */
  18 #define PM_PMC_SH       16      /* PMC number (1-based) for direct events */
  19 #define PM_PMC_MSK      0xf
  20 #define PM_PMC_MSKS     (PM_PMC_MSK << PM_PMC_SH)
  21 #define PM_UNIT_SH      12      /* TTMMUX number and setting - unit select */
  22 #define PM_UNIT_MSK     0xf
  23 #define PM_COMBINE_SH   11      /* Combined event bit */
  24 #define PM_COMBINE_MSK  1
  25 #define PM_COMBINE_MSKS 0x800
  26 #define PM_L2SEL_SH     8       /* L2 event select */
  27 #define PM_L2SEL_MSK    7
  28 #define PM_PMCSEL_MSK   0xff
  29
  30 /*
  31  * Bits in MMCR1 for POWER7
  32  */
  33 #define MMCR1_TTM0SEL_SH        60
  34 #define MMCR1_TTM1SEL_SH        56
  35 #define MMCR1_TTM2SEL_SH        52
  36 #define MMCR1_TTM3SEL_SH        48
  37 #define MMCR1_TTMSEL_MSK        0xf
  38 #define MMCR1_L2SEL_SH          45
  39 #define MMCR1_L2SEL_MSK         7
  40 #define MMCR1_PMC1_COMBINE_SH   35
  41 #define MMCR1_PMC2_COMBINE_SH   34
  42 #define MMCR1_PMC3_COMBINE_SH   33
  43 #define MMCR1_PMC4_COMBINE_SH   32
  44 #define MMCR1_PMC1SEL_SH        24
  45 #define MMCR1_PMC2SEL_SH        16
  46 #define MMCR1_PMC3SEL_SH        8
  47 #define MMCR1_PMC4SEL_SH        0
  48 #define MMCR1_PMCSEL_SH(n)      (MMCR1_PMC1SEL_SH - (n) * 8)
  49 #define MMCR1_PMCSEL_MSK        0xff
  50
  51 /*
  52  * Bits in MMCRA
  53  */
  54
  55 /*
  56  * Layout of constraint bits:
  57  * 6666555555555544444444443333333333222222222211111111110000000000
  58  * 3210987654321098765432109876543210987654321098765432109876543210
  59  *                                                 [  ><><><><><><>
  60  *                                                  NC P6P5P4P3P2P1
  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;
  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         *maskp = mask;
  96         *valp = value;
  97         return 0;
  98 }
  99
 100 #define MAX_ALT 2       /* at most 2 alternatives for any event */
 101
 102 static const unsigned int event_alternatives[][MAX_ALT] = {
 103         { 0x200f2, 0x300f2 },           /* PM_INST_DISP */
 104         { 0x200f4, 0x600f4 },           /* PM_RUN_CYC */
 105         { 0x400fa, 0x500fa },           /* PM_RUN_INST_CMPL */
 106 };
 107
 108 /*
 109  * Scan the alternatives table for a match and return the
 110  * index into the alternatives table if found, else -1.
 111  */
 112 static int find_alternative(u64 event)
 113 {
 114         int i, j;
 115
 116         for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) {
 117                 if (event < event_alternatives[i][0])
 118                         break;
 119                 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j)
 120                         if (event == event_alternatives[i][j])
 121                                 return i;
 122         }
 123         return -1;
 124 }
 125
 126 static s64 find_alternative_decode(u64 event)
 127 {
 128         int pmc, psel;
 129
 130         /* this only handles the 4x decode events */
 131         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 132         psel = event & PM_PMCSEL_MSK;
 133         if ((pmc == 2 || pmc == 4) && (psel & ~7) == 0x40)
 134                 return event - (1 << PM_PMC_SH) + 8;
 135         if ((pmc == 1 || pmc == 3) && (psel & ~7) == 0x48)
 136                 return event + (1 << PM_PMC_SH) - 8;
 137         return -1;
 138 }
 139
 140 static int power7_get_alternatives(u64 event, unsigned int flags, u64 alt[])
 141 {
 142         int i, j, nalt = 1;
 143         s64 ae;
 144
 145         alt[0] = event;
 146         nalt = 1;
 147         i = find_alternative(event);
 148         if (i >= 0) {
 149                 for (j = 0; j < MAX_ALT; ++j) {
 150                         ae = event_alternatives[i][j];
 151                         if (ae && ae != event)
 152                                 alt[nalt++] = ae;
 153                 }
 154         } else {
 155                 ae = find_alternative_decode(event);
 156                 if (ae > 0)
 157                         alt[nalt++] = ae;
 158         }
 159
 160         if (flags & PPMU_ONLY_COUNT_RUN) {
 161                 /*
 162                  * We're only counting in RUN state,
 163                  * so PM_CYC is equivalent to PM_RUN_CYC
 164                  * and PM_INST_CMPL === PM_RUN_INST_CMPL.
 165                  * This doesn't include alternatives that don't provide
 166                  * any extra flexibility in assigning PMCs.
 167                  */
 168                 j = nalt;
 169                 for (i = 0; i < nalt; ++i) {
 170                         switch (alt[i]) {
 171                         case 0x1e:      /* PM_CYC */
 172                                 alt[j++] = 0x600f4;     /* PM_RUN_CYC */
 173                                 break;
 174                         case 0x600f4:   /* PM_RUN_CYC */
 175                                 alt[j++] = 0x1e;
 176                                 break;
 177                         case 0x2:       /* PM_PPC_CMPL */
 178                                 alt[j++] = 0x500fa;     /* PM_RUN_INST_CMPL */
 179                                 break;
 180                         case 0x500fa:   /* PM_RUN_INST_CMPL */
 181                                 alt[j++] = 0x2; /* PM_PPC_CMPL */
 182                                 break;
 183                         }
 184                 }
 185                 nalt = j;
 186         }
 187
 188         return nalt;
 189 }
 190
 191 /*
 192  * Returns 1 if event counts things relating to marked instructions
 193  * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
 194  */
 195 static int power7_marked_instr_event(u64 event)
 196 {
 197         int pmc, psel;
 198         int unit;
 199
 200         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 201         unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
 202         psel = event & PM_PMCSEL_MSK & ~1;      /* trim off edge/level bit */
 203         if (pmc >= 5)
 204                 return 0;
 205
 206         switch (psel >> 4) {
 207         case 2:
 208                 return pmc == 2 || pmc == 4;
 209         case 3:
 210                 if (psel == 0x3c)
 211                         return pmc == 1;
 212                 if (psel == 0x3e)
 213                         return pmc != 2;
 214                 return 1;
 215         case 4:
 216         case 5:
 217                 return unit == 0xd;
 218         case 6:
 219                 if (psel == 0x64)
 220                         return pmc >= 3;
 221         case 8:
 222                 return unit == 0xd;
 223         }
 224         return 0;
 225 }
 226
 227 static int power7_compute_mmcr(u64 event[], int n_ev,
 228                                unsigned int hwc[], unsigned long mmcr[])
 229 {
 230         unsigned long mmcr1 = 0;
 231         unsigned long mmcra = 0;
 232         unsigned int pmc, unit, combine, l2sel, psel;
 233         unsigned int pmc_inuse = 0;
 234         int i;
 235
 236         /* First pass to count resource use */
 237         for (i = 0; i < n_ev; ++i) {
 238                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 239                 if (pmc) {
 240                         if (pmc > 6)
 241                                 return -1;
 242                         if (pmc_inuse & (1 << (pmc - 1)))
 243                                 return -1;
 244                         pmc_inuse |= 1 << (pmc - 1);
 245                 }
 246         }
 247
 248         /* Second pass: assign PMCs, set all MMCR1 fields */
 249         for (i = 0; i < n_ev; ++i) {
 250                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 251                 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
 252                 combine = (event[i] >> PM_COMBINE_SH) & PM_COMBINE_MSK;
 253                 l2sel = (event[i] >> PM_L2SEL_SH) & PM_L2SEL_MSK;
 254                 psel = event[i] & PM_PMCSEL_MSK;
 255                 if (!pmc) {
 256                         /* Bus event or any-PMC direct event */
 257                         for (pmc = 0; pmc < 4; ++pmc) {
 258                                 if (!(pmc_inuse & (1 << pmc)))
 259                                         break;
 260                         }
 261                         if (pmc >= 4)
 262                                 return -1;
 263                         pmc_inuse |= 1 << pmc;
 264                 } else {
 265                         /* Direct or decoded event */
 266                         --pmc;
 267                 }
 268                 if (pmc <= 3) {
 269                         mmcr1 |= (unsigned long) unit
 270                                 << (MMCR1_TTM0SEL_SH - 4 * pmc);
 271                         mmcr1 |= (unsigned long) combine
 272                                 << (MMCR1_PMC1_COMBINE_SH - pmc);
 273                         mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc);
 274                         if (unit == 6)  /* L2 events */
 275                                 mmcr1 |= (unsigned long) l2sel
 276                                         << MMCR1_L2SEL_SH;
 277                 }
 278                 if (power7_marked_instr_event(event[i]))
 279                         mmcra |= MMCRA_SAMPLE_ENABLE;
 280                 hwc[i] = pmc;
 281         }
 282
 283         /* Return MMCRx values */
 284         mmcr[0] = 0;
 285         if (pmc_inuse & 1)
 286                 mmcr[0] = MMCR0_PMC1CE;
 287         if (pmc_inuse & 0x3e)
 288                 mmcr[0] |= MMCR0_PMCjCE;
 289         mmcr[1] = mmcr1;
 290         mmcr[2] = mmcra;
 291         return 0;
 292 }
 293
 294 static void power7_disable_pmc(unsigned int pmc, unsigned long mmcr[])
 295 {
 296         if (pmc <= 3)
 297                 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc));
 298 }
 299
 300 static int power7_generic_events[] = {
 301         [PERF_COUNT_HW_CPU_CYCLES] = 0x1e,
 302         [PERF_COUNT_HW_INSTRUCTIONS] = 2,
 303         [PERF_COUNT_HW_CACHE_REFERENCES] = 0xc880,      /* LD_REF_L1_LSU*/
 304         [PERF_COUNT_HW_CACHE_MISSES] = 0x400f0,         /* LD_MISS_L1   */
 305         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x10068,  /* BRU_FIN      */
 306         [PERF_COUNT_HW_BRANCH_MISSES] = 0x400f6,        /* BR_MPRED     */
 307 };
 308
 309 #define C(x)    PERF_COUNT_HW_CACHE_##x
 310
 311 /*
 312  * Table of generalized cache-related events.
 313  * 0 means not supported, -1 means nonsensical, other values
 314  * are event codes.
 315  */
 316 static int power7_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 317         [C(L1D)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 318                 [C(OP_READ)] = {        0x400f0,        0xc880  },
 319                 [C(OP_WRITE)] = {       0,              0x300f0 },
 320                 [C(OP_PREFETCH)] = {    0xd8b8,         0       },
 321         },
 322         [C(L1I)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 323                 [C(OP_READ)] = {        0,              0x200fc },
 324                 [C(OP_WRITE)] = {       -1,             -1      },
 325                 [C(OP_PREFETCH)] = {    0x408a,         0       },
 326         },
 327         [C(LL)] = {             /*      RESULT_ACCESS   RESULT_MISS */
 328                 [C(OP_READ)] = {        0x6080,         0x6084  },
 329                 [C(OP_WRITE)] = {       0x6082,         0x6086  },
 330                 [C(OP_PREFETCH)] = {    0,              0       },
 331         },
 332         [C(DTLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 333                 [C(OP_READ)] = {        0,              0x300fc },
 334                 [C(OP_WRITE)] = {       -1,             -1      },
 335                 [C(OP_PREFETCH)] = {    -1,             -1      },
 336         },
 337         [C(ITLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 338                 [C(OP_READ)] = {        0,              0x400fc },
 339                 [C(OP_WRITE)] = {       -1,             -1      },
 340                 [C(OP_PREFETCH)] = {    -1,             -1      },
 341         },
 342         [C(BPU)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 343                 [C(OP_READ)] = {        0x10068,        0x400f6 },
 344                 [C(OP_WRITE)] = {       -1,             -1      },
 345                 [C(OP_PREFETCH)] = {    -1,             -1      },
 346         },
 347 };
 348
 349 struct power_pmu power7_pmu = {
 350         .n_counter              = 6,
 351         .max_alternatives       = MAX_ALT + 1,
 352         .add_fields             = 0x1555ul,
 353         .test_adder             = 0x3000ul,
 354         .compute_mmcr           = power7_compute_mmcr,
 355         .get_constraint         = power7_get_constraint,
 356         .get_alternatives       = power7_get_alternatives,
 357         .disable_pmc            = power7_disable_pmc,
 358         .n_generic              = ARRAY_SIZE(power7_generic_events),
 359         .generic_events         = power7_generic_events,
 360         .cache_events           = &power7_cache_events,
 361 };