arch/powerpc/perf/ppc970-pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Performance counter support for PPC970-family processors.
   4  *
   5  * Copyright 2008-2009 Paul Mackerras, IBM Corporation.
   6  */
   7 #include <linux/string.h>
   8 #include <linux/perf_event.h>
   9 #include <asm/reg.h>
  10 #include <asm/cputable.h>
  11
  12 /*
  13  * Bits in event code for PPC970
  14  */
  15 #define PM_PMC_SH       12      /* PMC number (1-based) for direct events */
  16 #define PM_PMC_MSK      0xf
  17 #define PM_UNIT_SH      8       /* TTMMUX number and setting - unit select */
  18 #define PM_UNIT_MSK     0xf
  19 #define PM_SPCSEL_SH    6
  20 #define PM_SPCSEL_MSK   3
  21 #define PM_BYTE_SH      4       /* Byte number of event bus to use */
  22 #define PM_BYTE_MSK     3
  23 #define PM_PMCSEL_MSK   0xf
  24
  25 /* Values in PM_UNIT field */
  26 #define PM_NONE         0
  27 #define PM_FPU          1
  28 #define PM_VPU          2
  29 #define PM_ISU          3
  30 #define PM_IFU          4
  31 #define PM_IDU          5
  32 #define PM_STS          6
  33 #define PM_LSU0         7
  34 #define PM_LSU1U        8
  35 #define PM_LSU1L        9
  36 #define PM_LASTUNIT     9
  37
  38 /*
  39  * Bits in MMCR0 for PPC970
  40  */
  41 #define MMCR0_PMC1SEL_SH        8
  42 #define MMCR0_PMC2SEL_SH        1
  43 #define MMCR_PMCSEL_MSK         0x1f
  44
  45 /*
  46  * Bits in MMCR1 for PPC970
  47  */
  48 #define MMCR1_TTM0SEL_SH        62
  49 #define MMCR1_TTM1SEL_SH        59
  50 #define MMCR1_TTM3SEL_SH        53
  51 #define MMCR1_TTMSEL_MSK        3
  52 #define MMCR1_TD_CP_DBG0SEL_SH  50
  53 #define MMCR1_TD_CP_DBG1SEL_SH  48
  54 #define MMCR1_TD_CP_DBG2SEL_SH  46
  55 #define MMCR1_TD_CP_DBG3SEL_SH  44
  56 #define MMCR1_PMC1_ADDER_SEL_SH 39
  57 #define MMCR1_PMC2_ADDER_SEL_SH 38
  58 #define MMCR1_PMC6_ADDER_SEL_SH 37
  59 #define MMCR1_PMC5_ADDER_SEL_SH 36
  60 #define MMCR1_PMC8_ADDER_SEL_SH 35
  61 #define MMCR1_PMC7_ADDER_SEL_SH 34
  62 #define MMCR1_PMC3_ADDER_SEL_SH 33
  63 #define MMCR1_PMC4_ADDER_SEL_SH 32
  64 #define MMCR1_PMC3SEL_SH        27
  65 #define MMCR1_PMC4SEL_SH        22
  66 #define MMCR1_PMC5SEL_SH        17
  67 #define MMCR1_PMC6SEL_SH        12
  68 #define MMCR1_PMC7SEL_SH        7
  69 #define MMCR1_PMC8SEL_SH        2
  70
  71 static short mmcr1_adder_bits[8] = {
  72         MMCR1_PMC1_ADDER_SEL_SH,
  73         MMCR1_PMC2_ADDER_SEL_SH,
  74         MMCR1_PMC3_ADDER_SEL_SH,
  75         MMCR1_PMC4_ADDER_SEL_SH,
  76         MMCR1_PMC5_ADDER_SEL_SH,
  77         MMCR1_PMC6_ADDER_SEL_SH,
  78         MMCR1_PMC7_ADDER_SEL_SH,
  79         MMCR1_PMC8_ADDER_SEL_SH
  80 };
  81
  82 /*
  83  * Layout of constraint bits:
  84  * 6666555555555544444444443333333333222222222211111111110000000000
  85  * 3210987654321098765432109876543210987654321098765432109876543210
  86  *               <><><>[  >[  >[  ><  ><  ><  ><  ><><><><><><><><>
  87  *               SPT0T1 UC  PS1 PS2 B0  B1  B2  B3 P1P2P3P4P5P6P7P8
  88  *
  89  * SP - SPCSEL constraint
  90  *     48-49: SPCSEL value 0x3_0000_0000_0000
  91  *
  92  * T0 - TTM0 constraint
  93  *     46-47: TTM0SEL value (0=FPU, 2=IFU, 3=VPU) 0xC000_0000_0000
  94  *
  95  * T1 - TTM1 constraint
  96  *     44-45: TTM1SEL value (0=IDU, 3=STS) 0x3000_0000_0000
  97  *
  98  * UC - unit constraint: can't have all three of FPU|IFU|VPU, ISU, IDU|STS
  99  *     43: UC3 error 0x0800_0000_0000
 100  *     42: FPU|IFU|VPU events needed 0x0400_0000_0000
 101  *     41: ISU events needed 0x0200_0000_0000
 102  *     40: IDU|STS events needed 0x0100_0000_0000
 103  *
 104  * PS1
 105  *     39: PS1 error 0x0080_0000_0000
 106  *     36-38: count of events needing PMC1/2/5/6 0x0070_0000_0000
 107  *
 108  * PS2
 109  *     35: PS2 error 0x0008_0000_0000
 110  *     32-34: count of events needing PMC3/4/7/8 0x0007_0000_0000
 111  *
 112  * B0
 113  *     28-31: Byte 0 event source 0xf000_0000
 114  *            Encoding as for the event code
 115  *
 116  * B1, B2, B3
 117  *     24-27, 20-23, 16-19: Byte 1, 2, 3 event sources
 118  *
 119  * P1
 120  *     15: P1 error 0x8000
 121  *     14-15: Count of events needing PMC1
 122  *
 123  * P2..P8
 124  *     0-13: Count of events needing PMC2..PMC8
 125  */
 126
 127 static unsigned char direct_marked_event[8] = {
 128         (1<<2) | (1<<3),        /* PMC1: PM_MRK_GRP_DISP, PM_MRK_ST_CMPL */
 129         (1<<3) | (1<<5),        /* PMC2: PM_THRESH_TIMEO, PM_MRK_BRU_FIN */
 130         (1<<3) | (1<<5),        /* PMC3: PM_MRK_ST_CMPL_INT, PM_MRK_VMX_FIN */
 131         (1<<4) | (1<<5),        /* PMC4: PM_MRK_GRP_CMPL, PM_MRK_CRU_FIN */
 132         (1<<4) | (1<<5),        /* PMC5: PM_GRP_MRK, PM_MRK_GRP_TIMEO */
 133         (1<<3) | (1<<4) | (1<<5),
 134                 /* PMC6: PM_MRK_ST_STS, PM_MRK_FXU_FIN, PM_MRK_GRP_ISSUED */
 135         (1<<4) | (1<<5),        /* PMC7: PM_MRK_FPU_FIN, PM_MRK_INST_FIN */
 136         (1<<4)                  /* PMC8: PM_MRK_LSU_FIN */
 137 };
 138
 139 /*
 140  * Returns 1 if event counts things relating to marked instructions
 141  * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not.
 142  */
 143 static int p970_marked_instr_event(u64 event)
 144 {
 145         int pmc, psel, unit, byte, bit;
 146         unsigned int mask;
 147
 148         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 149         psel = event & PM_PMCSEL_MSK;
 150         if (pmc) {
 151                 if (direct_marked_event[pmc - 1] & (1 << psel))
 152                         return 1;
 153                 if (psel == 0)          /* add events */
 154                         bit = (pmc <= 4)? pmc - 1: 8 - pmc;
 155                 else if (psel == 7 || psel == 13)       /* decode events */
 156                         bit = 4;
 157                 else
 158                         return 0;
 159         } else
 160                 bit = psel;
 161
 162         byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
 163         unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
 164         mask = 0;
 165         switch (unit) {
 166         case PM_VPU:
 167                 mask = 0x4c;            /* byte 0 bits 2,3,6 */
 168                 break;
 169         case PM_LSU0:
 170                 /* byte 2 bits 0,2,3,4,6; all of byte 1 */
 171                 mask = 0x085dff00;
 172                 break;
 173         case PM_LSU1L:
 174                 mask = 0x50 << 24;      /* byte 3 bits 4,6 */
 175                 break;
 176         }
 177         return (mask >> (byte * 8 + bit)) & 1;
 178 }
 179
 180 /* Masks and values for using events from the various units */
 181 static unsigned long unit_cons[PM_LASTUNIT+1][2] = {
 182         [PM_FPU] =   { 0xc80000000000ull, 0x040000000000ull },
 183         [PM_VPU] =   { 0xc80000000000ull, 0xc40000000000ull },
 184         [PM_ISU] =   { 0x080000000000ull, 0x020000000000ull },
 185         [PM_IFU] =   { 0xc80000000000ull, 0x840000000000ull },
 186         [PM_IDU] =   { 0x380000000000ull, 0x010000000000ull },
 187         [PM_STS] =   { 0x380000000000ull, 0x310000000000ull },
 188 };
 189
 190 static int p970_get_constraint(u64 event, unsigned long *maskp,
 191                                unsigned long *valp)
 192 {
 193         int pmc, byte, unit, sh, spcsel;
 194         unsigned long mask = 0, value = 0;
 195         int grp = -1;
 196
 197         pmc = (event >> PM_PMC_SH) & PM_PMC_MSK;
 198         if (pmc) {
 199                 if (pmc > 8)
 200                         return -1;
 201                 sh = (pmc - 1) * 2;
 202                 mask |= 2 << sh;
 203                 value |= 1 << sh;
 204                 grp = ((pmc - 1) >> 1) & 1;
 205         }
 206         unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK;
 207         if (unit) {
 208                 if (unit > PM_LASTUNIT)
 209                         return -1;
 210                 mask |= unit_cons[unit][0];
 211                 value |= unit_cons[unit][1];
 212                 byte = (event >> PM_BYTE_SH) & PM_BYTE_MSK;
 213                 /*
 214                  * Bus events on bytes 0 and 2 can be counted
 215                  * on PMC1/2/5/6; bytes 1 and 3 on PMC3/4/7/8.
 216                  */
 217                 if (!pmc)
 218                         grp = byte & 1;
 219                 /* Set byte lane select field */
 220                 mask  |= 0xfULL << (28 - 4 * byte);
 221                 value |= (unsigned long)unit << (28 - 4 * byte);
 222         }
 223         if (grp == 0) {
 224                 /* increment PMC1/2/5/6 field */
 225                 mask  |= 0x8000000000ull;
 226                 value |= 0x1000000000ull;
 227         } else if (grp == 1) {
 228                 /* increment PMC3/4/7/8 field */
 229                 mask  |= 0x800000000ull;
 230                 value |= 0x100000000ull;
 231         }
 232         spcsel = (event >> PM_SPCSEL_SH) & PM_SPCSEL_MSK;
 233         if (spcsel) {
 234                 mask  |= 3ull << 48;
 235                 value |= (unsigned long)spcsel << 48;
 236         }
 237         *maskp = mask;
 238         *valp = value;
 239         return 0;
 240 }
 241
 242 static int p970_get_alternatives(u64 event, unsigned int flags, u64 alt[])
 243 {
 244         alt[0] = event;
 245
 246         /* 2 alternatives for LSU empty */
 247         if (event == 0x2002 || event == 0x3002) {
 248                 alt[1] = event ^ 0x1000;
 249                 return 2;
 250         }
 251
 252         return 1;
 253 }
 254
 255 static int p970_compute_mmcr(u64 event[], int n_ev,
 256                              unsigned int hwc[], unsigned long mmcr[], struct perf_event *pevents[])
 257 {
 258         unsigned long mmcr0 = 0, mmcr1 = 0, mmcra = 0;
 259         unsigned int pmc, unit, byte, psel;
 260         unsigned int ttm, grp;
 261         unsigned int pmc_inuse = 0;
 262         unsigned int pmc_grp_use[2];
 263         unsigned char busbyte[4];
 264         unsigned char unituse[16];
 265         unsigned char unitmap[] = { 0, 0<<3, 3<<3, 1<<3, 2<<3, 0|4, 3|4 };
 266         unsigned char ttmuse[2];
 267         unsigned char pmcsel[8];
 268         int i;
 269         int spcsel;
 270
 271         if (n_ev > 8)
 272                 return -1;
 273
 274         /* First pass to count resource use */
 275         pmc_grp_use[0] = pmc_grp_use[1] = 0;
 276         memset(busbyte, 0, sizeof(busbyte));
 277         memset(unituse, 0, sizeof(unituse));
 278         for (i = 0; i < n_ev; ++i) {
 279                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 280                 if (pmc) {
 281                         if (pmc_inuse & (1 << (pmc - 1)))
 282                                 return -1;
 283                         pmc_inuse |= 1 << (pmc - 1);
 284                         /* count 1/2/5/6 vs 3/4/7/8 use */
 285                         ++pmc_grp_use[((pmc - 1) >> 1) & 1];
 286                 }
 287                 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
 288                 byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
 289                 if (unit) {
 290                         if (unit > PM_LASTUNIT)
 291                                 return -1;
 292                         if (!pmc)
 293                                 ++pmc_grp_use[byte & 1];
 294                         if (busbyte[byte] && busbyte[byte] != unit)
 295                                 return -1;
 296                         busbyte[byte] = unit;
 297                         unituse[unit] = 1;
 298                 }
 299         }
 300         if (pmc_grp_use[0] > 4 || pmc_grp_use[1] > 4)
 301                 return -1;
 302
 303         /*
 304          * Assign resources and set multiplexer selects.
 305          *
 306          * PM_ISU can go either on TTM0 or TTM1, but that's the only
 307          * choice we have to deal with.
 308          */
 309         if (unituse[PM_ISU] &
 310             (unituse[PM_FPU] | unituse[PM_IFU] | unituse[PM_VPU]))
 311                 unitmap[PM_ISU] = 2 | 4;        /* move ISU to TTM1 */
 312         /* Set TTM[01]SEL fields. */
 313         ttmuse[0] = ttmuse[1] = 0;
 314         for (i = PM_FPU; i <= PM_STS; ++i) {
 315                 if (!unituse[i])
 316                         continue;
 317                 ttm = unitmap[i];
 318                 ++ttmuse[(ttm >> 2) & 1];
 319                 mmcr1 |= (unsigned long)(ttm & ~4) << MMCR1_TTM1SEL_SH;
 320         }
 321         /* Check only one unit per TTMx */
 322         if (ttmuse[0] > 1 || ttmuse[1] > 1)
 323                 return -1;
 324
 325         /* Set byte lane select fields and TTM3SEL. */
 326         for (byte = 0; byte < 4; ++byte) {
 327                 unit = busbyte[byte];
 328                 if (!unit)
 329                         continue;
 330                 if (unit <= PM_STS)
 331                         ttm = (unitmap[unit] >> 2) & 1;
 332                 else if (unit == PM_LSU0)
 333                         ttm = 2;
 334                 else {
 335                         ttm = 3;
 336                         if (unit == PM_LSU1L && byte >= 2)
 337                                 mmcr1 |= 1ull << (MMCR1_TTM3SEL_SH + 3 - byte);
 338                 }
 339                 mmcr1 |= (unsigned long)ttm
 340                         << (MMCR1_TD_CP_DBG0SEL_SH - 2 * byte);
 341         }
 342
 343         /* Second pass: assign PMCs, set PMCxSEL and PMCx_ADDER_SEL fields */
 344         memset(pmcsel, 0x8, sizeof(pmcsel));    /* 8 means don't count */
 345         for (i = 0; i < n_ev; ++i) {
 346                 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK;
 347                 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK;
 348                 byte = (event[i] >> PM_BYTE_SH) & PM_BYTE_MSK;
 349                 psel = event[i] & PM_PMCSEL_MSK;
 350                 if (!pmc) {
 351                         /* Bus event or any-PMC direct event */
 352                         if (unit)
 353                                 psel |= 0x10 | ((byte & 2) << 2);
 354                         else
 355                                 psel |= 8;
 356                         for (pmc = 0; pmc < 8; ++pmc) {
 357                                 if (pmc_inuse & (1 << pmc))
 358                                         continue;
 359                                 grp = (pmc >> 1) & 1;
 360                                 if (unit) {
 361                                         if (grp == (byte & 1))
 362                                                 break;
 363                                 } else if (pmc_grp_use[grp] < 4) {
 364                                         ++pmc_grp_use[grp];
 365                                         break;
 366                                 }
 367                         }
 368                         pmc_inuse |= 1 << pmc;
 369                 } else {
 370                         /* Direct event */
 371                         --pmc;
 372                         if (psel == 0 && (byte & 2))
 373                                 /* add events on higher-numbered bus */
 374                                 mmcr1 |= 1ull << mmcr1_adder_bits[pmc];
 375                 }
 376                 pmcsel[pmc] = psel;
 377                 hwc[i] = pmc;
 378                 spcsel = (event[i] >> PM_SPCSEL_SH) & PM_SPCSEL_MSK;
 379                 mmcr1 |= spcsel;
 380                 if (p970_marked_instr_event(event[i]))
 381                         mmcra |= MMCRA_SAMPLE_ENABLE;
 382         }
 383         for (pmc = 0; pmc < 2; ++pmc)
 384                 mmcr0 |= pmcsel[pmc] << (MMCR0_PMC1SEL_SH - 7 * pmc);
 385         for (; pmc < 8; ++pmc)
 386                 mmcr1 |= (unsigned long)pmcsel[pmc]
 387                         << (MMCR1_PMC3SEL_SH - 5 * (pmc - 2));
 388         if (pmc_inuse & 1)
 389                 mmcr0 |= MMCR0_PMC1CE;
 390         if (pmc_inuse & 0xfe)
 391                 mmcr0 |= MMCR0_PMCjCE;
 392
 393         mmcra |= 0x2000;        /* mark only one IOP per PPC instruction */
 394
 395         /* Return MMCRx values */
 396         mmcr[0] = mmcr0;
 397         mmcr[1] = mmcr1;
 398         mmcr[2] = mmcra;
 399         return 0;
 400 }
 401
 402 static void p970_disable_pmc(unsigned int pmc, unsigned long mmcr[])
 403 {
 404         int shift, i;
 405
 406         if (pmc <= 1) {
 407                 shift = MMCR0_PMC1SEL_SH - 7 * pmc;
 408                 i = 0;
 409         } else {
 410                 shift = MMCR1_PMC3SEL_SH - 5 * (pmc - 2);
 411                 i = 1;
 412         }
 413         /*
 414          * Setting the PMCxSEL field to 0x08 disables PMC x.
 415          */
 416         mmcr[i] = (mmcr[i] & ~(0x1fUL << shift)) | (0x08UL << shift);
 417 }
 418
 419 static int ppc970_generic_events[] = {
 420         [PERF_COUNT_HW_CPU_CYCLES]              = 7,
 421         [PERF_COUNT_HW_INSTRUCTIONS]            = 1,
 422         [PERF_COUNT_HW_CACHE_REFERENCES]        = 0x8810, /* PM_LD_REF_L1 */
 423         [PERF_COUNT_HW_CACHE_MISSES]            = 0x3810, /* PM_LD_MISS_L1 */
 424         [PERF_COUNT_HW_BRANCH_INSTRUCTIONS]     = 0x431,  /* PM_BR_ISSUED */
 425         [PERF_COUNT_HW_BRANCH_MISSES]           = 0x327,  /* PM_GRP_BR_MPRED */
 426 };
 427
 428 #define C(x)    PERF_COUNT_HW_CACHE_##x
 429
 430 /*
 431  * Table of generalized cache-related events.
 432  * 0 means not supported, -1 means nonsensical, other values
 433  * are event codes.
 434  */
 435 static int ppc970_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = {
 436         [C(L1D)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 437                 [C(OP_READ)] = {        0x8810,         0x3810  },
 438                 [C(OP_WRITE)] = {       0x7810,         0x813   },
 439                 [C(OP_PREFETCH)] = {    0x731,          0       },
 440         },
 441         [C(L1I)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 442                 [C(OP_READ)] = {        0,              0       },
 443                 [C(OP_WRITE)] = {       -1,             -1      },
 444                 [C(OP_PREFETCH)] = {    0,              0       },
 445         },
 446         [C(LL)] = {             /*      RESULT_ACCESS   RESULT_MISS */
 447                 [C(OP_READ)] = {        0,              0       },
 448                 [C(OP_WRITE)] = {       0,              0       },
 449                 [C(OP_PREFETCH)] = {    0x733,          0       },
 450         },
 451         [C(DTLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 452                 [C(OP_READ)] = {        0,              0x704   },
 453                 [C(OP_WRITE)] = {       -1,             -1      },
 454                 [C(OP_PREFETCH)] = {    -1,             -1      },
 455         },
 456         [C(ITLB)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 457                 [C(OP_READ)] = {        0,              0x700   },
 458                 [C(OP_WRITE)] = {       -1,             -1      },
 459                 [C(OP_PREFETCH)] = {    -1,             -1      },
 460         },
 461         [C(BPU)] = {            /*      RESULT_ACCESS   RESULT_MISS */
 462                 [C(OP_READ)] = {        0x431,          0x327   },
 463                 [C(OP_WRITE)] = {       -1,             -1      },
 464                 [C(OP_PREFETCH)] = {    -1,             -1      },
 465         },
 466         [C(NODE)] = {           /*      RESULT_ACCESS   RESULT_MISS */
 467                 [C(OP_READ)] = {        -1,             -1      },
 468                 [C(OP_WRITE)] = {       -1,             -1      },
 469                 [C(OP_PREFETCH)] = {    -1,             -1      },
 470         },
 471 };
 472
 473 static struct power_pmu ppc970_pmu = {
 474         .name                   = "PPC970/FX/MP",
 475         .n_counter              = 8,
 476         .max_alternatives       = 2,
 477         .add_fields             = 0x001100005555ull,
 478         .test_adder             = 0x013300000000ull,
 479         .compute_mmcr           = p970_compute_mmcr,
 480         .get_constraint         = p970_get_constraint,
 481         .get_alternatives       = p970_get_alternatives,
 482         .disable_pmc            = p970_disable_pmc,
 483         .n_generic              = ARRAY_SIZE(ppc970_generic_events),
 484         .generic_events         = ppc970_generic_events,
 485         .cache_events           = &ppc970_cache_events,
 486         .flags                  = PPMU_NO_SIPR | PPMU_NO_CONT_SAMPLING,
 487 };
 488
 489 int init_ppc970_pmu(void)
 490 {
 491         if (!cur_cpu_spec->oprofile_cpu_type ||
 492             (strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/970")
 493              && strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/970MP")))
 494                 return -ENODEV;
 495
 496         return register_power_pmu(&ppc970_pmu);
 497 }