1 /* 2 * ARMv7 Cortex-A8 and Cortex-A9 Performance Events handling code. 3 * 4 * ARMv7 support: Jean Pihet <jpihet@mvista.com> 5 * 2010 (c) MontaVista Software, LLC. 6 * 7 * Copied from ARMv6 code, with the low level code inspired 8 * by the ARMv7 Oprofile code. 9 * 10 * Cortex-A8 has up to 4 configurable performance counters and 11 * a single cycle counter. 12 * Cortex-A9 has up to 31 configurable performance counters and 13 * a single cycle counter. 14 * 15 * All counters can be enabled/disabled and IRQ masked separately. The cycle 16 * counter and all 4 performance counters together can be reset separately. 17 */ 18 19 #ifdef CONFIG_CPU_V7 20 21 static struct arm_pmu armv7pmu; 22 23 /* 24 * Common ARMv7 event types 25 * 26 * Note: An implementation may not be able to count all of these events 27 * but the encodings are considered to be `reserved' in the case that 28 * they are not available. 29 */ 30 enum armv7_perf_types { 31 ARMV7_PERFCTR_PMNC_SW_INCR = 0x00, 32 ARMV7_PERFCTR_L1_ICACHE_REFILL = 0x01, 33 ARMV7_PERFCTR_ITLB_REFILL = 0x02, 34 ARMV7_PERFCTR_L1_DCACHE_REFILL = 0x03, 35 ARMV7_PERFCTR_L1_DCACHE_ACCESS = 0x04, 36 ARMV7_PERFCTR_DTLB_REFILL = 0x05, 37 ARMV7_PERFCTR_MEM_READ = 0x06, 38 ARMV7_PERFCTR_MEM_WRITE = 0x07, 39 ARMV7_PERFCTR_INSTR_EXECUTED = 0x08, 40 ARMV7_PERFCTR_EXC_TAKEN = 0x09, 41 ARMV7_PERFCTR_EXC_EXECUTED = 0x0A, 42 ARMV7_PERFCTR_CID_WRITE = 0x0B, 43 44 /* 45 * ARMV7_PERFCTR_PC_WRITE is equivalent to HW_BRANCH_INSTRUCTIONS. 46 * It counts: 47 * - all (taken) branch instructions, 48 * - instructions that explicitly write the PC, 49 * - exception generating instructions. 50 */ 51 ARMV7_PERFCTR_PC_WRITE = 0x0C, 52 ARMV7_PERFCTR_PC_IMM_BRANCH = 0x0D, 53 ARMV7_PERFCTR_PC_PROC_RETURN = 0x0E, 54 ARMV7_PERFCTR_MEM_UNALIGNED_ACCESS = 0x0F, 55 ARMV7_PERFCTR_PC_BRANCH_MIS_PRED = 0x10, 56 ARMV7_PERFCTR_CLOCK_CYCLES = 0x11, 57 ARMV7_PERFCTR_PC_BRANCH_PRED = 0x12, 58 59 /* These events are defined by the PMUv2 supplement (ARM DDI 0457A). */ 60 ARMV7_PERFCTR_MEM_ACCESS = 0x13, 61 ARMV7_PERFCTR_L1_ICACHE_ACCESS = 0x14, 62 ARMV7_PERFCTR_L1_DCACHE_WB = 0x15, 63 ARMV7_PERFCTR_L2_CACHE_ACCESS = 0x16, 64 ARMV7_PERFCTR_L2_CACHE_REFILL = 0x17, 65 ARMV7_PERFCTR_L2_CACHE_WB = 0x18, 66 ARMV7_PERFCTR_BUS_ACCESS = 0x19, 67 ARMV7_PERFCTR_MEM_ERROR = 0x1A, 68 ARMV7_PERFCTR_INSTR_SPEC = 0x1B, 69 ARMV7_PERFCTR_TTBR_WRITE = 0x1C, 70 ARMV7_PERFCTR_BUS_CYCLES = 0x1D, 71 72 ARMV7_PERFCTR_CPU_CYCLES = 0xFF 73 }; 74 75 /* ARMv7 Cortex-A8 specific event types */ 76 enum armv7_a8_perf_types { 77 ARMV7_A8_PERFCTR_L2_CACHE_ACCESS = 0x43, 78 ARMV7_A8_PERFCTR_L2_CACHE_REFILL = 0x44, 79 ARMV7_A8_PERFCTR_L1_ICACHE_ACCESS = 0x50, 80 ARMV7_A8_PERFCTR_STALL_ISIDE = 0x56, 81 }; 82 83 /* ARMv7 Cortex-A9 specific event types */ 84 enum armv7_a9_perf_types { 85 ARMV7_A9_PERFCTR_INSTR_CORE_RENAME = 0x68, 86 ARMV7_A9_PERFCTR_STALL_ICACHE = 0x60, 87 ARMV7_A9_PERFCTR_STALL_DISPATCH = 0x66, 88 }; 89 90 /* ARMv7 Cortex-A5 specific event types */ 91 enum armv7_a5_perf_types { 92 ARMV7_A5_PERFCTR_PREFETCH_LINEFILL = 0xc2, 93 ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP = 0xc3, 94 }; 95 96 /* ARMv7 Cortex-A15 specific event types */ 97 enum armv7_a15_perf_types { 98 ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_READ = 0x40, 99 ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_WRITE = 0x41, 100 ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_READ = 0x42, 101 ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_WRITE = 0x43, 102 103 ARMV7_A15_PERFCTR_DTLB_REFILL_L1_READ = 0x4C, 104 ARMV7_A15_PERFCTR_DTLB_REFILL_L1_WRITE = 0x4D, 105 106 ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_READ = 0x50, 107 ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_WRITE = 0x51, 108 ARMV7_A15_PERFCTR_L2_CACHE_REFILL_READ = 0x52, 109 ARMV7_A15_PERFCTR_L2_CACHE_REFILL_WRITE = 0x53, 110 111 ARMV7_A15_PERFCTR_PC_WRITE_SPEC = 0x76, 112 }; 113 114 /* 115 * Cortex-A8 HW events mapping 116 * 117 * The hardware events that we support. We do support cache operations but 118 * we have harvard caches and no way to combine instruction and data 119 * accesses/misses in hardware. 120 */ 121 static const unsigned armv7_a8_perf_map[PERF_COUNT_HW_MAX] = { 122 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES, 123 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED, 124 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 125 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 126 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE, 127 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 128 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED, 129 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV7_A8_PERFCTR_STALL_ISIDE, 130 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, 131 }; 132 133 static const unsigned armv7_a8_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] 134 [PERF_COUNT_HW_CACHE_OP_MAX] 135 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 136 [C(L1D)] = { 137 /* 138 * The performance counters don't differentiate between read 139 * and write accesses/misses so this isn't strictly correct, 140 * but it's the best we can do. Writes and reads get 141 * combined. 142 */ 143 [C(OP_READ)] = { 144 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 145 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 146 }, 147 [C(OP_WRITE)] = { 148 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 149 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 150 }, 151 [C(OP_PREFETCH)] = { 152 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 153 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 154 }, 155 }, 156 [C(L1I)] = { 157 [C(OP_READ)] = { 158 [C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L1_ICACHE_ACCESS, 159 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 160 }, 161 [C(OP_WRITE)] = { 162 [C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L1_ICACHE_ACCESS, 163 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 164 }, 165 [C(OP_PREFETCH)] = { 166 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 167 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 168 }, 169 }, 170 [C(LL)] = { 171 [C(OP_READ)] = { 172 [C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L2_CACHE_ACCESS, 173 [C(RESULT_MISS)] = ARMV7_A8_PERFCTR_L2_CACHE_REFILL, 174 }, 175 [C(OP_WRITE)] = { 176 [C(RESULT_ACCESS)] = ARMV7_A8_PERFCTR_L2_CACHE_ACCESS, 177 [C(RESULT_MISS)] = ARMV7_A8_PERFCTR_L2_CACHE_REFILL, 178 }, 179 [C(OP_PREFETCH)] = { 180 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 181 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 182 }, 183 }, 184 [C(DTLB)] = { 185 [C(OP_READ)] = { 186 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 187 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 188 }, 189 [C(OP_WRITE)] = { 190 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 191 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 192 }, 193 [C(OP_PREFETCH)] = { 194 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 195 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 196 }, 197 }, 198 [C(ITLB)] = { 199 [C(OP_READ)] = { 200 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 201 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 202 }, 203 [C(OP_WRITE)] = { 204 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 205 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 206 }, 207 [C(OP_PREFETCH)] = { 208 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 209 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 210 }, 211 }, 212 [C(BPU)] = { 213 [C(OP_READ)] = { 214 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 215 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 216 }, 217 [C(OP_WRITE)] = { 218 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 219 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 220 }, 221 [C(OP_PREFETCH)] = { 222 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 223 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 224 }, 225 }, 226 [C(NODE)] = { 227 [C(OP_READ)] = { 228 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 229 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 230 }, 231 [C(OP_WRITE)] = { 232 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 233 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 234 }, 235 [C(OP_PREFETCH)] = { 236 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 237 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 238 }, 239 }, 240 }; 241 242 /* 243 * Cortex-A9 HW events mapping 244 */ 245 static const unsigned armv7_a9_perf_map[PERF_COUNT_HW_MAX] = { 246 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES, 247 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_A9_PERFCTR_INSTR_CORE_RENAME, 248 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 249 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 250 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE, 251 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 252 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED, 253 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = ARMV7_A9_PERFCTR_STALL_ICACHE, 254 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = ARMV7_A9_PERFCTR_STALL_DISPATCH, 255 }; 256 257 static const unsigned armv7_a9_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] 258 [PERF_COUNT_HW_CACHE_OP_MAX] 259 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 260 [C(L1D)] = { 261 /* 262 * The performance counters don't differentiate between read 263 * and write accesses/misses so this isn't strictly correct, 264 * but it's the best we can do. Writes and reads get 265 * combined. 266 */ 267 [C(OP_READ)] = { 268 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 269 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 270 }, 271 [C(OP_WRITE)] = { 272 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 273 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 274 }, 275 [C(OP_PREFETCH)] = { 276 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 277 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 278 }, 279 }, 280 [C(L1I)] = { 281 [C(OP_READ)] = { 282 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 283 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 284 }, 285 [C(OP_WRITE)] = { 286 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 287 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 288 }, 289 [C(OP_PREFETCH)] = { 290 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 291 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 292 }, 293 }, 294 [C(LL)] = { 295 [C(OP_READ)] = { 296 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 297 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 298 }, 299 [C(OP_WRITE)] = { 300 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 301 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 302 }, 303 [C(OP_PREFETCH)] = { 304 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 305 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 306 }, 307 }, 308 [C(DTLB)] = { 309 [C(OP_READ)] = { 310 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 311 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 312 }, 313 [C(OP_WRITE)] = { 314 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 315 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 316 }, 317 [C(OP_PREFETCH)] = { 318 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 319 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 320 }, 321 }, 322 [C(ITLB)] = { 323 [C(OP_READ)] = { 324 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 325 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 326 }, 327 [C(OP_WRITE)] = { 328 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 329 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 330 }, 331 [C(OP_PREFETCH)] = { 332 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 333 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 334 }, 335 }, 336 [C(BPU)] = { 337 [C(OP_READ)] = { 338 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 339 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 340 }, 341 [C(OP_WRITE)] = { 342 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 343 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 344 }, 345 [C(OP_PREFETCH)] = { 346 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 347 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 348 }, 349 }, 350 [C(NODE)] = { 351 [C(OP_READ)] = { 352 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 353 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 354 }, 355 [C(OP_WRITE)] = { 356 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 357 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 358 }, 359 [C(OP_PREFETCH)] = { 360 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 361 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 362 }, 363 }, 364 }; 365 366 /* 367 * Cortex-A5 HW events mapping 368 */ 369 static const unsigned armv7_a5_perf_map[PERF_COUNT_HW_MAX] = { 370 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES, 371 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED, 372 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 373 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 374 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE, 375 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 376 [PERF_COUNT_HW_BUS_CYCLES] = HW_OP_UNSUPPORTED, 377 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = HW_OP_UNSUPPORTED, 378 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, 379 }; 380 381 static const unsigned armv7_a5_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] 382 [PERF_COUNT_HW_CACHE_OP_MAX] 383 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 384 [C(L1D)] = { 385 [C(OP_READ)] = { 386 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 387 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 388 }, 389 [C(OP_WRITE)] = { 390 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 391 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 392 }, 393 [C(OP_PREFETCH)] = { 394 [C(RESULT_ACCESS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL, 395 [C(RESULT_MISS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP, 396 }, 397 }, 398 [C(L1I)] = { 399 [C(OP_READ)] = { 400 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 401 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 402 }, 403 [C(OP_WRITE)] = { 404 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 405 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 406 }, 407 /* 408 * The prefetch counters don't differentiate between the I 409 * side and the D side. 410 */ 411 [C(OP_PREFETCH)] = { 412 [C(RESULT_ACCESS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL, 413 [C(RESULT_MISS)] = ARMV7_A5_PERFCTR_PREFETCH_LINEFILL_DROP, 414 }, 415 }, 416 [C(LL)] = { 417 [C(OP_READ)] = { 418 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 419 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 420 }, 421 [C(OP_WRITE)] = { 422 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 423 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 424 }, 425 [C(OP_PREFETCH)] = { 426 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 427 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 428 }, 429 }, 430 [C(DTLB)] = { 431 [C(OP_READ)] = { 432 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 433 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 434 }, 435 [C(OP_WRITE)] = { 436 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 437 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 438 }, 439 [C(OP_PREFETCH)] = { 440 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 441 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 442 }, 443 }, 444 [C(ITLB)] = { 445 [C(OP_READ)] = { 446 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 447 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 448 }, 449 [C(OP_WRITE)] = { 450 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 451 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 452 }, 453 [C(OP_PREFETCH)] = { 454 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 455 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 456 }, 457 }, 458 [C(BPU)] = { 459 [C(OP_READ)] = { 460 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 461 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 462 }, 463 [C(OP_WRITE)] = { 464 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 465 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 466 }, 467 [C(OP_PREFETCH)] = { 468 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 469 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 470 }, 471 }, 472 [C(NODE)] = { 473 [C(OP_READ)] = { 474 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 475 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 476 }, 477 [C(OP_WRITE)] = { 478 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 479 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 480 }, 481 [C(OP_PREFETCH)] = { 482 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 483 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 484 }, 485 }, 486 }; 487 488 /* 489 * Cortex-A15 HW events mapping 490 */ 491 static const unsigned armv7_a15_perf_map[PERF_COUNT_HW_MAX] = { 492 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES, 493 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED, 494 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 495 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 496 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_A15_PERFCTR_PC_WRITE_SPEC, 497 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 498 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES, 499 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = HW_OP_UNSUPPORTED, 500 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, 501 }; 502 503 static const unsigned armv7_a15_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] 504 [PERF_COUNT_HW_CACHE_OP_MAX] 505 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 506 [C(L1D)] = { 507 [C(OP_READ)] = { 508 [C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_READ, 509 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_READ, 510 }, 511 [C(OP_WRITE)] = { 512 [C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L1_DCACHE_ACCESS_WRITE, 513 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L1_DCACHE_REFILL_WRITE, 514 }, 515 [C(OP_PREFETCH)] = { 516 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 517 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 518 }, 519 }, 520 [C(L1I)] = { 521 /* 522 * Not all performance counters differentiate between read 523 * and write accesses/misses so we're not always strictly 524 * correct, but it's the best we can do. Writes and reads get 525 * combined in these cases. 526 */ 527 [C(OP_READ)] = { 528 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 529 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 530 }, 531 [C(OP_WRITE)] = { 532 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 533 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 534 }, 535 [C(OP_PREFETCH)] = { 536 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 537 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 538 }, 539 }, 540 [C(LL)] = { 541 [C(OP_READ)] = { 542 [C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_READ, 543 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L2_CACHE_REFILL_READ, 544 }, 545 [C(OP_WRITE)] = { 546 [C(RESULT_ACCESS)] = ARMV7_A15_PERFCTR_L2_CACHE_ACCESS_WRITE, 547 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_L2_CACHE_REFILL_WRITE, 548 }, 549 [C(OP_PREFETCH)] = { 550 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 551 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 552 }, 553 }, 554 [C(DTLB)] = { 555 [C(OP_READ)] = { 556 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 557 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_DTLB_REFILL_L1_READ, 558 }, 559 [C(OP_WRITE)] = { 560 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 561 [C(RESULT_MISS)] = ARMV7_A15_PERFCTR_DTLB_REFILL_L1_WRITE, 562 }, 563 [C(OP_PREFETCH)] = { 564 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 565 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 566 }, 567 }, 568 [C(ITLB)] = { 569 [C(OP_READ)] = { 570 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 571 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 572 }, 573 [C(OP_WRITE)] = { 574 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 575 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 576 }, 577 [C(OP_PREFETCH)] = { 578 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 579 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 580 }, 581 }, 582 [C(BPU)] = { 583 [C(OP_READ)] = { 584 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 585 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 586 }, 587 [C(OP_WRITE)] = { 588 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 589 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 590 }, 591 [C(OP_PREFETCH)] = { 592 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 593 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 594 }, 595 }, 596 [C(NODE)] = { 597 [C(OP_READ)] = { 598 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 599 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 600 }, 601 [C(OP_WRITE)] = { 602 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 603 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 604 }, 605 [C(OP_PREFETCH)] = { 606 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 607 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 608 }, 609 }, 610 }; 611 612 /* 613 * Cortex-A7 HW events mapping 614 */ 615 static const unsigned armv7_a7_perf_map[PERF_COUNT_HW_MAX] = { 616 [PERF_COUNT_HW_CPU_CYCLES] = ARMV7_PERFCTR_CPU_CYCLES, 617 [PERF_COUNT_HW_INSTRUCTIONS] = ARMV7_PERFCTR_INSTR_EXECUTED, 618 [PERF_COUNT_HW_CACHE_REFERENCES] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 619 [PERF_COUNT_HW_CACHE_MISSES] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 620 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = ARMV7_PERFCTR_PC_WRITE, 621 [PERF_COUNT_HW_BRANCH_MISSES] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 622 [PERF_COUNT_HW_BUS_CYCLES] = ARMV7_PERFCTR_BUS_CYCLES, 623 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = HW_OP_UNSUPPORTED, 624 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = HW_OP_UNSUPPORTED, 625 }; 626 627 static const unsigned armv7_a7_perf_cache_map[PERF_COUNT_HW_CACHE_MAX] 628 [PERF_COUNT_HW_CACHE_OP_MAX] 629 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 630 [C(L1D)] = { 631 /* 632 * The performance counters don't differentiate between read 633 * and write accesses/misses so this isn't strictly correct, 634 * but it's the best we can do. Writes and reads get 635 * combined. 636 */ 637 [C(OP_READ)] = { 638 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 639 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 640 }, 641 [C(OP_WRITE)] = { 642 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_DCACHE_ACCESS, 643 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_DCACHE_REFILL, 644 }, 645 [C(OP_PREFETCH)] = { 646 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 647 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 648 }, 649 }, 650 [C(L1I)] = { 651 [C(OP_READ)] = { 652 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 653 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 654 }, 655 [C(OP_WRITE)] = { 656 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L1_ICACHE_ACCESS, 657 [C(RESULT_MISS)] = ARMV7_PERFCTR_L1_ICACHE_REFILL, 658 }, 659 [C(OP_PREFETCH)] = { 660 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 661 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 662 }, 663 }, 664 [C(LL)] = { 665 [C(OP_READ)] = { 666 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_CACHE_ACCESS, 667 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL, 668 }, 669 [C(OP_WRITE)] = { 670 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_L2_CACHE_ACCESS, 671 [C(RESULT_MISS)] = ARMV7_PERFCTR_L2_CACHE_REFILL, 672 }, 673 [C(OP_PREFETCH)] = { 674 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 675 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 676 }, 677 }, 678 [C(DTLB)] = { 679 [C(OP_READ)] = { 680 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 681 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 682 }, 683 [C(OP_WRITE)] = { 684 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 685 [C(RESULT_MISS)] = ARMV7_PERFCTR_DTLB_REFILL, 686 }, 687 [C(OP_PREFETCH)] = { 688 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 689 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 690 }, 691 }, 692 [C(ITLB)] = { 693 [C(OP_READ)] = { 694 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 695 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 696 }, 697 [C(OP_WRITE)] = { 698 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 699 [C(RESULT_MISS)] = ARMV7_PERFCTR_ITLB_REFILL, 700 }, 701 [C(OP_PREFETCH)] = { 702 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 703 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 704 }, 705 }, 706 [C(BPU)] = { 707 [C(OP_READ)] = { 708 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 709 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 710 }, 711 [C(OP_WRITE)] = { 712 [C(RESULT_ACCESS)] = ARMV7_PERFCTR_PC_BRANCH_PRED, 713 [C(RESULT_MISS)] = ARMV7_PERFCTR_PC_BRANCH_MIS_PRED, 714 }, 715 [C(OP_PREFETCH)] = { 716 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 717 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 718 }, 719 }, 720 [C(NODE)] = { 721 [C(OP_READ)] = { 722 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 723 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 724 }, 725 [C(OP_WRITE)] = { 726 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 727 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 728 }, 729 [C(OP_PREFETCH)] = { 730 [C(RESULT_ACCESS)] = CACHE_OP_UNSUPPORTED, 731 [C(RESULT_MISS)] = CACHE_OP_UNSUPPORTED, 732 }, 733 }, 734 }; 735 736 /* 737 * Perf Events' indices 738 */ 739 #define ARMV7_IDX_CYCLE_COUNTER 0 740 #define ARMV7_IDX_COUNTER0 1 741 #define ARMV7_IDX_COUNTER_LAST (ARMV7_IDX_CYCLE_COUNTER + cpu_pmu->num_events - 1) 742 743 #define ARMV7_MAX_COUNTERS 32 744 #define ARMV7_COUNTER_MASK (ARMV7_MAX_COUNTERS - 1) 745 746 /* 747 * ARMv7 low level PMNC access 748 */ 749 750 /* 751 * Perf Event to low level counters mapping 752 */ 753 #define ARMV7_IDX_TO_COUNTER(x) \ 754 (((x) - ARMV7_IDX_COUNTER0) & ARMV7_COUNTER_MASK) 755 756 /* 757 * Per-CPU PMNC: config reg 758 */ 759 #define ARMV7_PMNC_E (1 << 0) /* Enable all counters */ 760 #define ARMV7_PMNC_P (1 << 1) /* Reset all counters */ 761 #define ARMV7_PMNC_C (1 << 2) /* Cycle counter reset */ 762 #define ARMV7_PMNC_D (1 << 3) /* CCNT counts every 64th cpu cycle */ 763 #define ARMV7_PMNC_X (1 << 4) /* Export to ETM */ 764 #define ARMV7_PMNC_DP (1 << 5) /* Disable CCNT if non-invasive debug*/ 765 #define ARMV7_PMNC_N_SHIFT 11 /* Number of counters supported */ 766 #define ARMV7_PMNC_N_MASK 0x1f 767 #define ARMV7_PMNC_MASK 0x3f /* Mask for writable bits */ 768 769 /* 770 * FLAG: counters overflow flag status reg 771 */ 772 #define ARMV7_FLAG_MASK 0xffffffff /* Mask for writable bits */ 773 #define ARMV7_OVERFLOWED_MASK ARMV7_FLAG_MASK 774 775 /* 776 * PMXEVTYPER: Event selection reg 777 */ 778 #define ARMV7_EVTYPE_MASK 0xc00000ff /* Mask for writable bits */ 779 #define ARMV7_EVTYPE_EVENT 0xff /* Mask for EVENT bits */ 780 781 /* 782 * Event filters for PMUv2 783 */ 784 #define ARMV7_EXCLUDE_PL1 (1 << 31) 785 #define ARMV7_EXCLUDE_USER (1 << 30) 786 #define ARMV7_INCLUDE_HYP (1 << 27) 787 788 static inline u32 armv7_pmnc_read(void) 789 { 790 u32 val; 791 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(val)); 792 return val; 793 } 794 795 static inline void armv7_pmnc_write(u32 val) 796 { 797 val &= ARMV7_PMNC_MASK; 798 isb(); 799 asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r"(val)); 800 } 801 802 static inline int armv7_pmnc_has_overflowed(u32 pmnc) 803 { 804 return pmnc & ARMV7_OVERFLOWED_MASK; 805 } 806 807 static inline int armv7_pmnc_counter_valid(int idx) 808 { 809 return idx >= ARMV7_IDX_CYCLE_COUNTER && idx <= ARMV7_IDX_COUNTER_LAST; 810 } 811 812 static inline int armv7_pmnc_counter_has_overflowed(u32 pmnc, int idx) 813 { 814 int ret = 0; 815 u32 counter; 816 817 if (!armv7_pmnc_counter_valid(idx)) { 818 pr_err("CPU%u checking wrong counter %d overflow status\n", 819 smp_processor_id(), idx); 820 } else { 821 counter = ARMV7_IDX_TO_COUNTER(idx); 822 ret = pmnc & BIT(counter); 823 } 824 825 return ret; 826 } 827 828 static inline int armv7_pmnc_select_counter(int idx) 829 { 830 u32 counter; 831 832 if (!armv7_pmnc_counter_valid(idx)) { 833 pr_err("CPU%u selecting wrong PMNC counter %d\n", 834 smp_processor_id(), idx); 835 return -EINVAL; 836 } 837 838 counter = ARMV7_IDX_TO_COUNTER(idx); 839 asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (counter)); 840 isb(); 841 842 return idx; 843 } 844 845 static inline u32 armv7pmu_read_counter(int idx) 846 { 847 u32 value = 0; 848 849 if (!armv7_pmnc_counter_valid(idx)) 850 pr_err("CPU%u reading wrong counter %d\n", 851 smp_processor_id(), idx); 852 else if (idx == ARMV7_IDX_CYCLE_COUNTER) 853 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (value)); 854 else if (armv7_pmnc_select_counter(idx) == idx) 855 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (value)); 856 857 return value; 858 } 859 860 static inline void armv7pmu_write_counter(int idx, u32 value) 861 { 862 if (!armv7_pmnc_counter_valid(idx)) 863 pr_err("CPU%u writing wrong counter %d\n", 864 smp_processor_id(), idx); 865 else if (idx == ARMV7_IDX_CYCLE_COUNTER) 866 asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (value)); 867 else if (armv7_pmnc_select_counter(idx) == idx) 868 asm volatile("mcr p15, 0, %0, c9, c13, 2" : : "r" (value)); 869 } 870 871 static inline void armv7_pmnc_write_evtsel(int idx, u32 val) 872 { 873 if (armv7_pmnc_select_counter(idx) == idx) { 874 val &= ARMV7_EVTYPE_MASK; 875 asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (val)); 876 } 877 } 878 879 static inline int armv7_pmnc_enable_counter(int idx) 880 { 881 u32 counter; 882 883 if (!armv7_pmnc_counter_valid(idx)) { 884 pr_err("CPU%u enabling wrong PMNC counter %d\n", 885 smp_processor_id(), idx); 886 return -EINVAL; 887 } 888 889 counter = ARMV7_IDX_TO_COUNTER(idx); 890 asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (BIT(counter))); 891 return idx; 892 } 893 894 static inline int armv7_pmnc_disable_counter(int idx) 895 { 896 u32 counter; 897 898 if (!armv7_pmnc_counter_valid(idx)) { 899 pr_err("CPU%u disabling wrong PMNC counter %d\n", 900 smp_processor_id(), idx); 901 return -EINVAL; 902 } 903 904 counter = ARMV7_IDX_TO_COUNTER(idx); 905 asm volatile("mcr p15, 0, %0, c9, c12, 2" : : "r" (BIT(counter))); 906 return idx; 907 } 908 909 static inline int armv7_pmnc_enable_intens(int idx) 910 { 911 u32 counter; 912 913 if (!armv7_pmnc_counter_valid(idx)) { 914 pr_err("CPU%u enabling wrong PMNC counter IRQ enable %d\n", 915 smp_processor_id(), idx); 916 return -EINVAL; 917 } 918 919 counter = ARMV7_IDX_TO_COUNTER(idx); 920 asm volatile("mcr p15, 0, %0, c9, c14, 1" : : "r" (BIT(counter))); 921 return idx; 922 } 923 924 static inline int armv7_pmnc_disable_intens(int idx) 925 { 926 u32 counter; 927 928 if (!armv7_pmnc_counter_valid(idx)) { 929 pr_err("CPU%u disabling wrong PMNC counter IRQ enable %d\n", 930 smp_processor_id(), idx); 931 return -EINVAL; 932 } 933 934 counter = ARMV7_IDX_TO_COUNTER(idx); 935 asm volatile("mcr p15, 0, %0, c9, c14, 2" : : "r" (BIT(counter))); 936 isb(); 937 /* Clear the overflow flag in case an interrupt is pending. */ 938 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (BIT(counter))); 939 isb(); 940 941 return idx; 942 } 943 944 static inline u32 armv7_pmnc_getreset_flags(void) 945 { 946 u32 val; 947 948 /* Read */ 949 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val)); 950 951 /* Write to clear flags */ 952 val &= ARMV7_FLAG_MASK; 953 asm volatile("mcr p15, 0, %0, c9, c12, 3" : : "r" (val)); 954 955 return val; 956 } 957 958 #ifdef DEBUG 959 static void armv7_pmnc_dump_regs(void) 960 { 961 u32 val; 962 unsigned int cnt; 963 964 printk(KERN_INFO "PMNC registers dump:\n"); 965 966 asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (val)); 967 printk(KERN_INFO "PMNC =0x%08x\n", val); 968 969 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r" (val)); 970 printk(KERN_INFO "CNTENS=0x%08x\n", val); 971 972 asm volatile("mrc p15, 0, %0, c9, c14, 1" : "=r" (val)); 973 printk(KERN_INFO "INTENS=0x%08x\n", val); 974 975 asm volatile("mrc p15, 0, %0, c9, c12, 3" : "=r" (val)); 976 printk(KERN_INFO "FLAGS =0x%08x\n", val); 977 978 asm volatile("mrc p15, 0, %0, c9, c12, 5" : "=r" (val)); 979 printk(KERN_INFO "SELECT=0x%08x\n", val); 980 981 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (val)); 982 printk(KERN_INFO "CCNT =0x%08x\n", val); 983 984 for (cnt = ARMV7_IDX_COUNTER0; cnt <= ARMV7_IDX_COUNTER_LAST; cnt++) { 985 armv7_pmnc_select_counter(cnt); 986 asm volatile("mrc p15, 0, %0, c9, c13, 2" : "=r" (val)); 987 printk(KERN_INFO "CNT[%d] count =0x%08x\n", 988 ARMV7_IDX_TO_COUNTER(cnt), val); 989 asm volatile("mrc p15, 0, %0, c9, c13, 1" : "=r" (val)); 990 printk(KERN_INFO "CNT[%d] evtsel=0x%08x\n", 991 ARMV7_IDX_TO_COUNTER(cnt), val); 992 } 993 } 994 #endif 995 996 static void armv7pmu_enable_event(struct hw_perf_event *hwc, int idx) 997 { 998 unsigned long flags; 999 struct pmu_hw_events *events = cpu_pmu->get_hw_events(); 1000 1001 /* 1002 * Enable counter and interrupt, and set the counter to count 1003 * the event that we're interested in. 1004 */ 1005 raw_spin_lock_irqsave(&events->pmu_lock, flags); 1006 1007 /* 1008 * Disable counter 1009 */ 1010 armv7_pmnc_disable_counter(idx); 1011 1012 /* 1013 * Set event (if destined for PMNx counters) 1014 * We only need to set the event for the cycle counter if we 1015 * have the ability to perform event filtering. 1016 */ 1017 if (armv7pmu.set_event_filter || idx != ARMV7_IDX_CYCLE_COUNTER) 1018 armv7_pmnc_write_evtsel(idx, hwc->config_base); 1019 1020 /* 1021 * Enable interrupt for this counter 1022 */ 1023 armv7_pmnc_enable_intens(idx); 1024 1025 /* 1026 * Enable counter 1027 */ 1028 armv7_pmnc_enable_counter(idx); 1029 1030 raw_spin_unlock_irqrestore(&events->pmu_lock, flags); 1031 } 1032 1033 static void armv7pmu_disable_event(struct hw_perf_event *hwc, int idx) 1034 { 1035 unsigned long flags; 1036 struct pmu_hw_events *events = cpu_pmu->get_hw_events(); 1037 1038 /* 1039 * Disable counter and interrupt 1040 */ 1041 raw_spin_lock_irqsave(&events->pmu_lock, flags); 1042 1043 /* 1044 * Disable counter 1045 */ 1046 armv7_pmnc_disable_counter(idx); 1047 1048 /* 1049 * Disable interrupt for this counter 1050 */ 1051 armv7_pmnc_disable_intens(idx); 1052 1053 raw_spin_unlock_irqrestore(&events->pmu_lock, flags); 1054 } 1055 1056 static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev) 1057 { 1058 u32 pmnc; 1059 struct perf_sample_data data; 1060 struct pmu_hw_events *cpuc; 1061 struct pt_regs *regs; 1062 int idx; 1063 1064 /* 1065 * Get and reset the IRQ flags 1066 */ 1067 pmnc = armv7_pmnc_getreset_flags(); 1068 1069 /* 1070 * Did an overflow occur? 1071 */ 1072 if (!armv7_pmnc_has_overflowed(pmnc)) 1073 return IRQ_NONE; 1074 1075 /* 1076 * Handle the counter(s) overflow(s) 1077 */ 1078 regs = get_irq_regs(); 1079 1080 cpuc = &__get_cpu_var(cpu_hw_events); 1081 for (idx = 0; idx < cpu_pmu->num_events; ++idx) { 1082 struct perf_event *event = cpuc->events[idx]; 1083 struct hw_perf_event *hwc; 1084 1085 /* Ignore if we don't have an event. */ 1086 if (!event) 1087 continue; 1088 1089 /* 1090 * We have a single interrupt for all counters. Check that 1091 * each counter has overflowed before we process it. 1092 */ 1093 if (!armv7_pmnc_counter_has_overflowed(pmnc, idx)) 1094 continue; 1095 1096 hwc = &event->hw; 1097 armpmu_event_update(event, hwc, idx); 1098 perf_sample_data_init(&data, 0, hwc->last_period); 1099 if (!armpmu_event_set_period(event, hwc, idx)) 1100 continue; 1101 1102 if (perf_event_overflow(event, &data, regs)) 1103 cpu_pmu->disable(hwc, idx); 1104 } 1105 1106 /* 1107 * Handle the pending perf events. 1108 * 1109 * Note: this call *must* be run with interrupts disabled. For 1110 * platforms that can have the PMU interrupts raised as an NMI, this 1111 * will not work. 1112 */ 1113 irq_work_run(); 1114 1115 return IRQ_HANDLED; 1116 } 1117 1118 static void armv7pmu_start(void) 1119 { 1120 unsigned long flags; 1121 struct pmu_hw_events *events = cpu_pmu->get_hw_events(); 1122 1123 raw_spin_lock_irqsave(&events->pmu_lock, flags); 1124 /* Enable all counters */ 1125 armv7_pmnc_write(armv7_pmnc_read() | ARMV7_PMNC_E); 1126 raw_spin_unlock_irqrestore(&events->pmu_lock, flags); 1127 } 1128 1129 static void armv7pmu_stop(void) 1130 { 1131 unsigned long flags; 1132 struct pmu_hw_events *events = cpu_pmu->get_hw_events(); 1133 1134 raw_spin_lock_irqsave(&events->pmu_lock, flags); 1135 /* Disable all counters */ 1136 armv7_pmnc_write(armv7_pmnc_read() & ~ARMV7_PMNC_E); 1137 raw_spin_unlock_irqrestore(&events->pmu_lock, flags); 1138 } 1139 1140 static int armv7pmu_get_event_idx(struct pmu_hw_events *cpuc, 1141 struct hw_perf_event *event) 1142 { 1143 int idx; 1144 unsigned long evtype = event->config_base & ARMV7_EVTYPE_EVENT; 1145 1146 /* Always place a cycle counter into the cycle counter. */ 1147 if (evtype == ARMV7_PERFCTR_CPU_CYCLES) { 1148 if (test_and_set_bit(ARMV7_IDX_CYCLE_COUNTER, cpuc->used_mask)) 1149 return -EAGAIN; 1150 1151 return ARMV7_IDX_CYCLE_COUNTER; 1152 } 1153 1154 /* 1155 * For anything other than a cycle counter, try and use 1156 * the events counters 1157 */ 1158 for (idx = ARMV7_IDX_COUNTER0; idx < cpu_pmu->num_events; ++idx) { 1159 if (!test_and_set_bit(idx, cpuc->used_mask)) 1160 return idx; 1161 } 1162 1163 /* The counters are all in use. */ 1164 return -EAGAIN; 1165 } 1166 1167 /* 1168 * Add an event filter to a given event. This will only work for PMUv2 PMUs. 1169 */ 1170 static int armv7pmu_set_event_filter(struct hw_perf_event *event, 1171 struct perf_event_attr *attr) 1172 { 1173 unsigned long config_base = 0; 1174 1175 if (attr->exclude_idle) 1176 return -EPERM; 1177 if (attr->exclude_user) 1178 config_base |= ARMV7_EXCLUDE_USER; 1179 if (attr->exclude_kernel) 1180 config_base |= ARMV7_EXCLUDE_PL1; 1181 if (!attr->exclude_hv) 1182 config_base |= ARMV7_INCLUDE_HYP; 1183 1184 /* 1185 * Install the filter into config_base as this is used to 1186 * construct the event type. 1187 */ 1188 event->config_base = config_base; 1189 1190 return 0; 1191 } 1192 1193 static void armv7pmu_reset(void *info) 1194 { 1195 u32 idx, nb_cnt = cpu_pmu->num_events; 1196 1197 /* The counter and interrupt enable registers are unknown at reset. */ 1198 for (idx = ARMV7_IDX_CYCLE_COUNTER; idx < nb_cnt; ++idx) 1199 armv7pmu_disable_event(NULL, idx); 1200 1201 /* Initialize & Reset PMNC: C and P bits */ 1202 armv7_pmnc_write(ARMV7_PMNC_P | ARMV7_PMNC_C); 1203 } 1204 1205 static int armv7_a8_map_event(struct perf_event *event) 1206 { 1207 return map_cpu_event(event, &armv7_a8_perf_map, 1208 &armv7_a8_perf_cache_map, 0xFF); 1209 } 1210 1211 static int armv7_a9_map_event(struct perf_event *event) 1212 { 1213 return map_cpu_event(event, &armv7_a9_perf_map, 1214 &armv7_a9_perf_cache_map, 0xFF); 1215 } 1216 1217 static int armv7_a5_map_event(struct perf_event *event) 1218 { 1219 return map_cpu_event(event, &armv7_a5_perf_map, 1220 &armv7_a5_perf_cache_map, 0xFF); 1221 } 1222 1223 static int armv7_a15_map_event(struct perf_event *event) 1224 { 1225 return map_cpu_event(event, &armv7_a15_perf_map, 1226 &armv7_a15_perf_cache_map, 0xFF); 1227 } 1228 1229 static int armv7_a7_map_event(struct perf_event *event) 1230 { 1231 return map_cpu_event(event, &armv7_a7_perf_map, 1232 &armv7_a7_perf_cache_map, 0xFF); 1233 } 1234 1235 static struct arm_pmu armv7pmu = { 1236 .handle_irq = armv7pmu_handle_irq, 1237 .enable = armv7pmu_enable_event, 1238 .disable = armv7pmu_disable_event, 1239 .read_counter = armv7pmu_read_counter, 1240 .write_counter = armv7pmu_write_counter, 1241 .get_event_idx = armv7pmu_get_event_idx, 1242 .start = armv7pmu_start, 1243 .stop = armv7pmu_stop, 1244 .reset = armv7pmu_reset, 1245 .max_period = (1LLU << 32) - 1, 1246 }; 1247 1248 static u32 __init armv7_read_num_pmnc_events(void) 1249 { 1250 u32 nb_cnt; 1251 1252 /* Read the nb of CNTx counters supported from PMNC */ 1253 nb_cnt = (armv7_pmnc_read() >> ARMV7_PMNC_N_SHIFT) & ARMV7_PMNC_N_MASK; 1254 1255 /* Add the CPU cycles counter and return */ 1256 return nb_cnt + 1; 1257 } 1258 1259 static struct arm_pmu *__init armv7_a8_pmu_init(void) 1260 { 1261 armv7pmu.id = ARM_PERF_PMU_ID_CA8; 1262 armv7pmu.name = "ARMv7 Cortex-A8"; 1263 armv7pmu.map_event = armv7_a8_map_event; 1264 armv7pmu.num_events = armv7_read_num_pmnc_events(); 1265 return &armv7pmu; 1266 } 1267 1268 static struct arm_pmu *__init armv7_a9_pmu_init(void) 1269 { 1270 armv7pmu.id = ARM_PERF_PMU_ID_CA9; 1271 armv7pmu.name = "ARMv7 Cortex-A9"; 1272 armv7pmu.map_event = armv7_a9_map_event; 1273 armv7pmu.num_events = armv7_read_num_pmnc_events(); 1274 return &armv7pmu; 1275 } 1276 1277 static struct arm_pmu *__init armv7_a5_pmu_init(void) 1278 { 1279 armv7pmu.id = ARM_PERF_PMU_ID_CA5; 1280 armv7pmu.name = "ARMv7 Cortex-A5"; 1281 armv7pmu.map_event = armv7_a5_map_event; 1282 armv7pmu.num_events = armv7_read_num_pmnc_events(); 1283 return &armv7pmu; 1284 } 1285 1286 static struct arm_pmu *__init armv7_a15_pmu_init(void) 1287 { 1288 armv7pmu.id = ARM_PERF_PMU_ID_CA15; 1289 armv7pmu.name = "ARMv7 Cortex-A15"; 1290 armv7pmu.map_event = armv7_a15_map_event; 1291 armv7pmu.num_events = armv7_read_num_pmnc_events(); 1292 armv7pmu.set_event_filter = armv7pmu_set_event_filter; 1293 return &armv7pmu; 1294 } 1295 1296 static struct arm_pmu *__init armv7_a7_pmu_init(void) 1297 { 1298 armv7pmu.id = ARM_PERF_PMU_ID_CA7; 1299 armv7pmu.name = "ARMv7 Cortex-A7"; 1300 armv7pmu.map_event = armv7_a7_map_event; 1301 armv7pmu.num_events = armv7_read_num_pmnc_events(); 1302 armv7pmu.set_event_filter = armv7pmu_set_event_filter; 1303 return &armv7pmu; 1304 } 1305 #else 1306 static struct arm_pmu *__init armv7_a8_pmu_init(void) 1307 { 1308 return NULL; 1309 } 1310 1311 static struct arm_pmu *__init armv7_a9_pmu_init(void) 1312 { 1313 return NULL; 1314 } 1315 1316 static struct arm_pmu *__init armv7_a5_pmu_init(void) 1317 { 1318 return NULL; 1319 } 1320 1321 static struct arm_pmu *__init armv7_a15_pmu_init(void) 1322 { 1323 return NULL; 1324 } 1325 1326 static struct arm_pmu *__init armv7_a7_pmu_init(void) 1327 { 1328 return NULL; 1329 } 1330 #endif /* CONFIG_CPU_V7 */ 1331