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 * NC P6P5P4P3P2P1 59 * 60 * NC - number of counters 61 * 15: NC error 0x8000 62 * 12-14: number of events needing PMC1-4 0x7000 63 * 64 * P6 65 * 11: P6 error 0x800 66 * 10-11: Count of events needing PMC6 67 * 68 * P1..P5 69 * 0-9: Count of events needing PMC1..PMC5 70 */ 71 72 static int power7_get_constraint(u64 event, unsigned long *maskp, 73 unsigned long *valp) 74 { 75 int pmc, sh; 76 unsigned long mask = 0, value = 0; 77 78 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK; 79 if (pmc) { 80 if (pmc > 6) 81 return -1; 82 sh = (pmc - 1) * 2; 83 mask |= 2 << sh; 84 value |= 1 << sh; 85 if (pmc >= 5 && !(event == 0x500fa || event == 0x600f4)) 86 return -1; 87 } 88 if (pmc < 5) { 89 /* need a counter from PMC1-4 set */ 90 mask |= 0x8000; 91 value |= 0x1000; 92 } 93 *maskp = mask; 94 *valp = value; 95 return 0; 96 } 97 98 #define MAX_ALT 2 /* at most 2 alternatives for any event */ 99 100 static const unsigned int event_alternatives[][MAX_ALT] = { 101 { 0x200f2, 0x300f2 }, /* PM_INST_DISP */ 102 { 0x200f4, 0x600f4 }, /* PM_RUN_CYC */ 103 { 0x400fa, 0x500fa }, /* PM_RUN_INST_CMPL */ 104 }; 105 106 /* 107 * Scan the alternatives table for a match and return the 108 * index into the alternatives table if found, else -1. 109 */ 110 static int find_alternative(u64 event) 111 { 112 int i, j; 113 114 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) { 115 if (event < event_alternatives[i][0]) 116 break; 117 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j) 118 if (event == event_alternatives[i][j]) 119 return i; 120 } 121 return -1; 122 } 123 124 static s64 find_alternative_decode(u64 event) 125 { 126 int pmc, psel; 127 128 /* this only handles the 4x decode events */ 129 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK; 130 psel = event & PM_PMCSEL_MSK; 131 if ((pmc == 2 || pmc == 4) && (psel & ~7) == 0x40) 132 return event - (1 << PM_PMC_SH) + 8; 133 if ((pmc == 1 || pmc == 3) && (psel & ~7) == 0x48) 134 return event + (1 << PM_PMC_SH) - 8; 135 return -1; 136 } 137 138 static int power7_get_alternatives(u64 event, unsigned int flags, u64 alt[]) 139 { 140 int i, j, nalt = 1; 141 s64 ae; 142 143 alt[0] = event; 144 nalt = 1; 145 i = find_alternative(event); 146 if (i >= 0) { 147 for (j = 0; j < MAX_ALT; ++j) { 148 ae = event_alternatives[i][j]; 149 if (ae && ae != event) 150 alt[nalt++] = ae; 151 } 152 } else { 153 ae = find_alternative_decode(event); 154 if (ae > 0) 155 alt[nalt++] = ae; 156 } 157 158 if (flags & PPMU_ONLY_COUNT_RUN) { 159 /* 160 * We're only counting in RUN state, 161 * so PM_CYC is equivalent to PM_RUN_CYC 162 * and PM_INST_CMPL === PM_RUN_INST_CMPL. 163 * This doesn't include alternatives that don't provide 164 * any extra flexibility in assigning PMCs. 165 */ 166 j = nalt; 167 for (i = 0; i < nalt; ++i) { 168 switch (alt[i]) { 169 case 0x1e: /* PM_CYC */ 170 alt[j++] = 0x600f4; /* PM_RUN_CYC */ 171 break; 172 case 0x600f4: /* PM_RUN_CYC */ 173 alt[j++] = 0x1e; 174 break; 175 case 0x2: /* PM_PPC_CMPL */ 176 alt[j++] = 0x500fa; /* PM_RUN_INST_CMPL */ 177 break; 178 case 0x500fa: /* PM_RUN_INST_CMPL */ 179 alt[j++] = 0x2; /* PM_PPC_CMPL */ 180 break; 181 } 182 } 183 nalt = j; 184 } 185 186 return nalt; 187 } 188 189 /* 190 * Returns 1 if event counts things relating to marked instructions 191 * and thus needs the MMCRA_SAMPLE_ENABLE bit set, or 0 if not. 192 */ 193 static int power7_marked_instr_event(u64 event) 194 { 195 int pmc, psel; 196 int unit; 197 198 pmc = (event >> PM_PMC_SH) & PM_PMC_MSK; 199 unit = (event >> PM_UNIT_SH) & PM_UNIT_MSK; 200 psel = event & PM_PMCSEL_MSK & ~1; /* trim off edge/level bit */ 201 if (pmc >= 5) 202 return 0; 203 204 switch (psel >> 4) { 205 case 2: 206 return pmc == 2 || pmc == 4; 207 case 3: 208 if (psel == 0x3c) 209 return pmc == 1; 210 if (psel == 0x3e) 211 return pmc != 2; 212 return 1; 213 case 4: 214 case 5: 215 return unit == 0xd; 216 case 6: 217 if (psel == 0x64) 218 return pmc >= 3; 219 case 8: 220 return unit == 0xd; 221 } 222 return 0; 223 } 224 225 static int power7_compute_mmcr(u64 event[], int n_ev, 226 unsigned int hwc[], unsigned long mmcr[]) 227 { 228 unsigned long mmcr1 = 0; 229 unsigned long mmcra = MMCRA_SDAR_DCACHE_MISS | MMCRA_SDAR_ERAT_MISS; 230 unsigned int pmc, unit, combine, l2sel, psel; 231 unsigned int pmc_inuse = 0; 232 int i; 233 234 /* First pass to count resource use */ 235 for (i = 0; i < n_ev; ++i) { 236 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK; 237 if (pmc) { 238 if (pmc > 6) 239 return -1; 240 if (pmc_inuse & (1 << (pmc - 1))) 241 return -1; 242 pmc_inuse |= 1 << (pmc - 1); 243 } 244 } 245 246 /* Second pass: assign PMCs, set all MMCR1 fields */ 247 for (i = 0; i < n_ev; ++i) { 248 pmc = (event[i] >> PM_PMC_SH) & PM_PMC_MSK; 249 unit = (event[i] >> PM_UNIT_SH) & PM_UNIT_MSK; 250 combine = (event[i] >> PM_COMBINE_SH) & PM_COMBINE_MSK; 251 l2sel = (event[i] >> PM_L2SEL_SH) & PM_L2SEL_MSK; 252 psel = event[i] & PM_PMCSEL_MSK; 253 if (!pmc) { 254 /* Bus event or any-PMC direct event */ 255 for (pmc = 0; pmc < 4; ++pmc) { 256 if (!(pmc_inuse & (1 << pmc))) 257 break; 258 } 259 if (pmc >= 4) 260 return -1; 261 pmc_inuse |= 1 << pmc; 262 } else { 263 /* Direct or decoded event */ 264 --pmc; 265 } 266 if (pmc <= 3) { 267 mmcr1 |= (unsigned long) unit 268 << (MMCR1_TTM0SEL_SH - 4 * pmc); 269 mmcr1 |= (unsigned long) combine 270 << (MMCR1_PMC1_COMBINE_SH - pmc); 271 mmcr1 |= psel << MMCR1_PMCSEL_SH(pmc); 272 if (unit == 6) /* L2 events */ 273 mmcr1 |= (unsigned long) l2sel 274 << MMCR1_L2SEL_SH; 275 } 276 if (power7_marked_instr_event(event[i])) 277 mmcra |= MMCRA_SAMPLE_ENABLE; 278 hwc[i] = pmc; 279 } 280 281 /* Return MMCRx values */ 282 mmcr[0] = 0; 283 if (pmc_inuse & 1) 284 mmcr[0] = MMCR0_PMC1CE; 285 if (pmc_inuse & 0x3e) 286 mmcr[0] |= MMCR0_PMCjCE; 287 mmcr[1] = mmcr1; 288 mmcr[2] = mmcra; 289 return 0; 290 } 291 292 static void power7_disable_pmc(unsigned int pmc, unsigned long mmcr[]) 293 { 294 if (pmc <= 3) 295 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SH(pmc)); 296 } 297 298 static int power7_generic_events[] = { 299 [PERF_COUNT_HW_CPU_CYCLES] = 0x1e, 300 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 0x100f8, /* GCT_NOSLOT_CYC */ 301 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 0x4000a, /* CMPLU_STALL */ 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)] = { 0xc880, 0x400f0 }, 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)] = { 0x16080, 0x26080 }, 329 [C(OP_WRITE)] = { 0x16082, 0x26082 }, 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 [C(NODE)] = { /* RESULT_ACCESS RESULT_MISS */ 348 [C(OP_READ)] = { -1, -1 }, 349 [C(OP_WRITE)] = { -1, -1 }, 350 [C(OP_PREFETCH)] = { -1, -1 }, 351 }, 352 }; 353 354 static struct power_pmu power7_pmu = { 355 .name = "POWER7", 356 .n_counter = 6, 357 .max_alternatives = MAX_ALT + 1, 358 .add_fields = 0x1555ul, 359 .test_adder = 0x3000ul, 360 .compute_mmcr = power7_compute_mmcr, 361 .get_constraint = power7_get_constraint, 362 .get_alternatives = power7_get_alternatives, 363 .disable_pmc = power7_disable_pmc, 364 .flags = PPMU_ALT_SIPR, 365 .n_generic = ARRAY_SIZE(power7_generic_events), 366 .generic_events = power7_generic_events, 367 .cache_events = &power7_cache_events, 368 }; 369 370 static int __init init_power7_pmu(void) 371 { 372 if (!cur_cpu_spec->oprofile_cpu_type || 373 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power7")) 374 return -ENODEV; 375 376 return register_power_pmu(&power7_pmu); 377 } 378 379 early_initcall(init_power7_pmu); 380