1 /* 2 * Performance counter support for POWER8 processors. 3 * 4 * Copyright 2009 Paul Mackerras, IBM Corporation. 5 * Copyright 2013 Michael Ellerman, IBM Corporation. 6 * 7 * This program is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU General Public License 9 * as published by the Free Software Foundation; either version 10 * 2 of the License, or (at your option) any later version. 11 */ 12 13 #define pr_fmt(fmt) "power8-pmu: " fmt 14 15 #include <linux/kernel.h> 16 #include <linux/perf_event.h> 17 #include <asm/firmware.h> 18 #include <asm/cputable.h> 19 20 21 /* 22 * Some power8 event codes. 23 */ 24 #define PM_CYC 0x0001e 25 #define PM_GCT_NOSLOT_CYC 0x100f8 26 #define PM_CMPLU_STALL 0x4000a 27 #define PM_INST_CMPL 0x00002 28 #define PM_BRU_FIN 0x10068 29 #define PM_BR_MPRED_CMPL 0x400f6 30 31 /* All L1 D cache load references counted at finish, gated by reject */ 32 #define PM_LD_REF_L1 0x100ee 33 /* Load Missed L1 */ 34 #define PM_LD_MISS_L1 0x3e054 35 /* Store Missed L1 */ 36 #define PM_ST_MISS_L1 0x300f0 37 /* L1 cache data prefetches */ 38 #define PM_L1_PREF 0x0d8b8 39 /* Instruction fetches from L1 */ 40 #define PM_INST_FROM_L1 0x04080 41 /* Demand iCache Miss */ 42 #define PM_L1_ICACHE_MISS 0x200fd 43 /* Instruction Demand sectors wriittent into IL1 */ 44 #define PM_L1_DEMAND_WRITE 0x0408c 45 /* Instruction prefetch written into IL1 */ 46 #define PM_IC_PREF_WRITE 0x0408e 47 /* The data cache was reloaded from local core's L3 due to a demand load */ 48 #define PM_DATA_FROM_L3 0x4c042 49 /* Demand LD - L3 Miss (not L2 hit and not L3 hit) */ 50 #define PM_DATA_FROM_L3MISS 0x300fe 51 /* All successful D-side store dispatches for this thread */ 52 #define PM_L2_ST 0x17080 53 /* All successful D-side store dispatches for this thread that were L2 Miss */ 54 #define PM_L2_ST_MISS 0x17082 55 /* Total HW L3 prefetches(Load+store) */ 56 #define PM_L3_PREF_ALL 0x4e052 57 /* Data PTEG reload */ 58 #define PM_DTLB_MISS 0x300fc 59 /* ITLB Reloaded */ 60 #define PM_ITLB_MISS 0x400fc 61 62 63 /* 64 * Raw event encoding for POWER8: 65 * 66 * 60 56 52 48 44 40 36 32 67 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | 68 * | | [ ] [ thresh_cmp ] [ thresh_ctl ] 69 * | | | | 70 * | | *- IFM (Linux) thresh start/stop OR FAB match -* 71 * | *- BHRB (Linux) 72 * *- EBB (Linux) 73 * 74 * 28 24 20 16 12 8 4 0 75 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | 76 * [ ] [ sample ] [cache] [ pmc ] [unit ] c m [ pmcxsel ] 77 * | | | | | 78 * | | | | *- mark 79 * | | *- L1/L2/L3 cache_sel | 80 * | | | 81 * | *- sampling mode for marked events *- combine 82 * | 83 * *- thresh_sel 84 * 85 * Below uses IBM bit numbering. 86 * 87 * MMCR1[x:y] = unit (PMCxUNIT) 88 * MMCR1[x] = combine (PMCxCOMB) 89 * 90 * if pmc == 3 and unit == 0 and pmcxsel[0:6] == 0b0101011 91 * # PM_MRK_FAB_RSP_MATCH 92 * MMCR1[20:27] = thresh_ctl (FAB_CRESP_MATCH / FAB_TYPE_MATCH) 93 * else if pmc == 4 and unit == 0xf and pmcxsel[0:6] == 0b0101001 94 * # PM_MRK_FAB_RSP_MATCH_CYC 95 * MMCR1[20:27] = thresh_ctl (FAB_CRESP_MATCH / FAB_TYPE_MATCH) 96 * else 97 * MMCRA[48:55] = thresh_ctl (THRESH START/END) 98 * 99 * if thresh_sel: 100 * MMCRA[45:47] = thresh_sel 101 * 102 * if thresh_cmp: 103 * MMCRA[22:24] = thresh_cmp[0:2] 104 * MMCRA[25:31] = thresh_cmp[3:9] 105 * 106 * if unit == 6 or unit == 7 107 * MMCRC[53:55] = cache_sel[1:3] (L2EVENT_SEL) 108 * else if unit == 8 or unit == 9: 109 * if cache_sel[0] == 0: # L3 bank 110 * MMCRC[47:49] = cache_sel[1:3] (L3EVENT_SEL0) 111 * else if cache_sel[0] == 1: 112 * MMCRC[50:51] = cache_sel[2:3] (L3EVENT_SEL1) 113 * else if cache_sel[1]: # L1 event 114 * MMCR1[16] = cache_sel[2] 115 * MMCR1[17] = cache_sel[3] 116 * 117 * if mark: 118 * MMCRA[63] = 1 (SAMPLE_ENABLE) 119 * MMCRA[57:59] = sample[0:2] (RAND_SAMP_ELIG) 120 * MMCRA[61:62] = sample[3:4] (RAND_SAMP_MODE) 121 * 122 * if EBB and BHRB: 123 * MMCRA[32:33] = IFM 124 * 125 */ 126 127 #define EVENT_EBB_MASK 1ull 128 #define EVENT_EBB_SHIFT PERF_EVENT_CONFIG_EBB_SHIFT 129 #define EVENT_BHRB_MASK 1ull 130 #define EVENT_BHRB_SHIFT 62 131 #define EVENT_WANTS_BHRB (EVENT_BHRB_MASK << EVENT_BHRB_SHIFT) 132 #define EVENT_IFM_MASK 3ull 133 #define EVENT_IFM_SHIFT 60 134 #define EVENT_THR_CMP_SHIFT 40 /* Threshold CMP value */ 135 #define EVENT_THR_CMP_MASK 0x3ff 136 #define EVENT_THR_CTL_SHIFT 32 /* Threshold control value (start/stop) */ 137 #define EVENT_THR_CTL_MASK 0xffull 138 #define EVENT_THR_SEL_SHIFT 29 /* Threshold select value */ 139 #define EVENT_THR_SEL_MASK 0x7 140 #define EVENT_THRESH_SHIFT 29 /* All threshold bits */ 141 #define EVENT_THRESH_MASK 0x1fffffull 142 #define EVENT_SAMPLE_SHIFT 24 /* Sampling mode & eligibility */ 143 #define EVENT_SAMPLE_MASK 0x1f 144 #define EVENT_CACHE_SEL_SHIFT 20 /* L2/L3 cache select */ 145 #define EVENT_CACHE_SEL_MASK 0xf 146 #define EVENT_IS_L1 (4 << EVENT_CACHE_SEL_SHIFT) 147 #define EVENT_PMC_SHIFT 16 /* PMC number (1-based) */ 148 #define EVENT_PMC_MASK 0xf 149 #define EVENT_UNIT_SHIFT 12 /* Unit */ 150 #define EVENT_UNIT_MASK 0xf 151 #define EVENT_COMBINE_SHIFT 11 /* Combine bit */ 152 #define EVENT_COMBINE_MASK 0x1 153 #define EVENT_MARKED_SHIFT 8 /* Marked bit */ 154 #define EVENT_MARKED_MASK 0x1 155 #define EVENT_IS_MARKED (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT) 156 #define EVENT_PSEL_MASK 0xff /* PMCxSEL value */ 157 158 /* Bits defined by Linux */ 159 #define EVENT_LINUX_MASK \ 160 ((EVENT_EBB_MASK << EVENT_EBB_SHIFT) | \ 161 (EVENT_BHRB_MASK << EVENT_BHRB_SHIFT) | \ 162 (EVENT_IFM_MASK << EVENT_IFM_SHIFT)) 163 164 #define EVENT_VALID_MASK \ 165 ((EVENT_THRESH_MASK << EVENT_THRESH_SHIFT) | \ 166 (EVENT_SAMPLE_MASK << EVENT_SAMPLE_SHIFT) | \ 167 (EVENT_CACHE_SEL_MASK << EVENT_CACHE_SEL_SHIFT) | \ 168 (EVENT_PMC_MASK << EVENT_PMC_SHIFT) | \ 169 (EVENT_UNIT_MASK << EVENT_UNIT_SHIFT) | \ 170 (EVENT_COMBINE_MASK << EVENT_COMBINE_SHIFT) | \ 171 (EVENT_MARKED_MASK << EVENT_MARKED_SHIFT) | \ 172 EVENT_LINUX_MASK | \ 173 EVENT_PSEL_MASK) 174 175 /* MMCRA IFM bits - POWER8 */ 176 #define POWER8_MMCRA_IFM1 0x0000000040000000UL 177 #define POWER8_MMCRA_IFM2 0x0000000080000000UL 178 #define POWER8_MMCRA_IFM3 0x00000000C0000000UL 179 180 #define ONLY_PLM \ 181 (PERF_SAMPLE_BRANCH_USER |\ 182 PERF_SAMPLE_BRANCH_KERNEL |\ 183 PERF_SAMPLE_BRANCH_HV) 184 185 /* 186 * Layout of constraint bits: 187 * 188 * 60 56 52 48 44 40 36 32 189 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | 190 * [ fab_match ] [ thresh_cmp ] [ thresh_ctl ] [ ] 191 * | 192 * thresh_sel -* 193 * 194 * 28 24 20 16 12 8 4 0 195 * | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | - - - - | 196 * [ ] | [ ] [ sample ] [ ] [6] [5] [4] [3] [2] [1] 197 * | | | | 198 * BHRB IFM -* | | | Count of events for each PMC. 199 * EBB -* | | p1, p2, p3, p4, p5, p6. 200 * L1 I/D qualifier -* | 201 * nc - number of counters -* 202 * 203 * The PMC fields P1..P6, and NC, are adder fields. As we accumulate constraints 204 * we want the low bit of each field to be added to any existing value. 205 * 206 * Everything else is a value field. 207 */ 208 209 #define CNST_FAB_MATCH_VAL(v) (((v) & EVENT_THR_CTL_MASK) << 56) 210 #define CNST_FAB_MATCH_MASK CNST_FAB_MATCH_VAL(EVENT_THR_CTL_MASK) 211 212 /* We just throw all the threshold bits into the constraint */ 213 #define CNST_THRESH_VAL(v) (((v) & EVENT_THRESH_MASK) << 32) 214 #define CNST_THRESH_MASK CNST_THRESH_VAL(EVENT_THRESH_MASK) 215 216 #define CNST_EBB_VAL(v) (((v) & EVENT_EBB_MASK) << 24) 217 #define CNST_EBB_MASK CNST_EBB_VAL(EVENT_EBB_MASK) 218 219 #define CNST_IFM_VAL(v) (((v) & EVENT_IFM_MASK) << 25) 220 #define CNST_IFM_MASK CNST_IFM_VAL(EVENT_IFM_MASK) 221 222 #define CNST_L1_QUAL_VAL(v) (((v) & 3) << 22) 223 #define CNST_L1_QUAL_MASK CNST_L1_QUAL_VAL(3) 224 225 #define CNST_SAMPLE_VAL(v) (((v) & EVENT_SAMPLE_MASK) << 16) 226 #define CNST_SAMPLE_MASK CNST_SAMPLE_VAL(EVENT_SAMPLE_MASK) 227 228 /* 229 * For NC we are counting up to 4 events. This requires three bits, and we need 230 * the fifth event to overflow and set the 4th bit. To achieve that we bias the 231 * fields by 3 in test_adder. 232 */ 233 #define CNST_NC_SHIFT 12 234 #define CNST_NC_VAL (1 << CNST_NC_SHIFT) 235 #define CNST_NC_MASK (8 << CNST_NC_SHIFT) 236 #define POWER8_TEST_ADDER (3 << CNST_NC_SHIFT) 237 238 /* 239 * For the per-PMC fields we have two bits. The low bit is added, so if two 240 * events ask for the same PMC the sum will overflow, setting the high bit, 241 * indicating an error. So our mask sets the high bit. 242 */ 243 #define CNST_PMC_SHIFT(pmc) ((pmc - 1) * 2) 244 #define CNST_PMC_VAL(pmc) (1 << CNST_PMC_SHIFT(pmc)) 245 #define CNST_PMC_MASK(pmc) (2 << CNST_PMC_SHIFT(pmc)) 246 247 /* Our add_fields is defined as: */ 248 #define POWER8_ADD_FIELDS \ 249 CNST_PMC_VAL(1) | CNST_PMC_VAL(2) | CNST_PMC_VAL(3) | \ 250 CNST_PMC_VAL(4) | CNST_PMC_VAL(5) | CNST_PMC_VAL(6) | CNST_NC_VAL 251 252 253 /* Bits in MMCR1 for POWER8 */ 254 #define MMCR1_UNIT_SHIFT(pmc) (60 - (4 * ((pmc) - 1))) 255 #define MMCR1_COMBINE_SHIFT(pmc) (35 - ((pmc) - 1)) 256 #define MMCR1_PMCSEL_SHIFT(pmc) (24 - (((pmc) - 1)) * 8) 257 #define MMCR1_FAB_SHIFT 36 258 #define MMCR1_DC_QUAL_SHIFT 47 259 #define MMCR1_IC_QUAL_SHIFT 46 260 261 /* Bits in MMCRA for POWER8 */ 262 #define MMCRA_SAMP_MODE_SHIFT 1 263 #define MMCRA_SAMP_ELIG_SHIFT 4 264 #define MMCRA_THR_CTL_SHIFT 8 265 #define MMCRA_THR_SEL_SHIFT 16 266 #define MMCRA_THR_CMP_SHIFT 32 267 #define MMCRA_SDAR_MODE_TLB (1ull << 42) 268 #define MMCRA_IFM_SHIFT 30 269 270 /* Bits in MMCR2 for POWER8 */ 271 #define MMCR2_FCS(pmc) (1ull << (63 - (((pmc) - 1) * 9))) 272 #define MMCR2_FCP(pmc) (1ull << (62 - (((pmc) - 1) * 9))) 273 #define MMCR2_FCH(pmc) (1ull << (57 - (((pmc) - 1) * 9))) 274 275 276 static inline bool event_is_fab_match(u64 event) 277 { 278 /* Only check pmc, unit and pmcxsel, ignore the edge bit (0) */ 279 event &= 0xff0fe; 280 281 /* PM_MRK_FAB_RSP_MATCH & PM_MRK_FAB_RSP_MATCH_CYC */ 282 return (event == 0x30056 || event == 0x4f052); 283 } 284 285 static int power8_get_constraint(u64 event, unsigned long *maskp, unsigned long *valp) 286 { 287 unsigned int unit, pmc, cache, ebb; 288 unsigned long mask, value; 289 290 mask = value = 0; 291 292 if (event & ~EVENT_VALID_MASK) 293 return -1; 294 295 pmc = (event >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; 296 unit = (event >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK; 297 cache = (event >> EVENT_CACHE_SEL_SHIFT) & EVENT_CACHE_SEL_MASK; 298 ebb = (event >> EVENT_EBB_SHIFT) & EVENT_EBB_MASK; 299 300 if (pmc) { 301 u64 base_event; 302 303 if (pmc > 6) 304 return -1; 305 306 /* Ignore Linux defined bits when checking event below */ 307 base_event = event & ~EVENT_LINUX_MASK; 308 309 if (pmc >= 5 && base_event != 0x500fa && base_event != 0x600f4) 310 return -1; 311 312 mask |= CNST_PMC_MASK(pmc); 313 value |= CNST_PMC_VAL(pmc); 314 } 315 316 if (pmc <= 4) { 317 /* 318 * Add to number of counters in use. Note this includes events with 319 * a PMC of 0 - they still need a PMC, it's just assigned later. 320 * Don't count events on PMC 5 & 6, there is only one valid event 321 * on each of those counters, and they are handled above. 322 */ 323 mask |= CNST_NC_MASK; 324 value |= CNST_NC_VAL; 325 } 326 327 if (unit >= 6 && unit <= 9) { 328 /* 329 * L2/L3 events contain a cache selector field, which is 330 * supposed to be programmed into MMCRC. However MMCRC is only 331 * HV writable, and there is no API for guest kernels to modify 332 * it. The solution is for the hypervisor to initialise the 333 * field to zeroes, and for us to only ever allow events that 334 * have a cache selector of zero. The bank selector (bit 3) is 335 * irrelevant, as long as the rest of the value is 0. 336 */ 337 if (cache & 0x7) 338 return -1; 339 340 } else if (event & EVENT_IS_L1) { 341 mask |= CNST_L1_QUAL_MASK; 342 value |= CNST_L1_QUAL_VAL(cache); 343 } 344 345 if (event & EVENT_IS_MARKED) { 346 mask |= CNST_SAMPLE_MASK; 347 value |= CNST_SAMPLE_VAL(event >> EVENT_SAMPLE_SHIFT); 348 } 349 350 /* 351 * Special case for PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC, 352 * the threshold control bits are used for the match value. 353 */ 354 if (event_is_fab_match(event)) { 355 mask |= CNST_FAB_MATCH_MASK; 356 value |= CNST_FAB_MATCH_VAL(event >> EVENT_THR_CTL_SHIFT); 357 } else { 358 /* 359 * Check the mantissa upper two bits are not zero, unless the 360 * exponent is also zero. See the THRESH_CMP_MANTISSA doc. 361 */ 362 unsigned int cmp, exp; 363 364 cmp = (event >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK; 365 exp = cmp >> 7; 366 367 if (exp && (cmp & 0x60) == 0) 368 return -1; 369 370 mask |= CNST_THRESH_MASK; 371 value |= CNST_THRESH_VAL(event >> EVENT_THRESH_SHIFT); 372 } 373 374 if (!pmc && ebb) 375 /* EBB events must specify the PMC */ 376 return -1; 377 378 if (event & EVENT_WANTS_BHRB) { 379 if (!ebb) 380 /* Only EBB events can request BHRB */ 381 return -1; 382 383 mask |= CNST_IFM_MASK; 384 value |= CNST_IFM_VAL(event >> EVENT_IFM_SHIFT); 385 } 386 387 /* 388 * All events must agree on EBB, either all request it or none. 389 * EBB events are pinned & exclusive, so this should never actually 390 * hit, but we leave it as a fallback in case. 391 */ 392 mask |= CNST_EBB_VAL(ebb); 393 value |= CNST_EBB_MASK; 394 395 *maskp = mask; 396 *valp = value; 397 398 return 0; 399 } 400 401 static int power8_compute_mmcr(u64 event[], int n_ev, 402 unsigned int hwc[], unsigned long mmcr[], 403 struct perf_event *pevents[]) 404 { 405 unsigned long mmcra, mmcr1, mmcr2, unit, combine, psel, cache, val; 406 unsigned int pmc, pmc_inuse; 407 int i; 408 409 pmc_inuse = 0; 410 411 /* First pass to count resource use */ 412 for (i = 0; i < n_ev; ++i) { 413 pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; 414 if (pmc) 415 pmc_inuse |= 1 << pmc; 416 } 417 418 /* In continous sampling mode, update SDAR on TLB miss */ 419 mmcra = MMCRA_SDAR_MODE_TLB; 420 mmcr1 = mmcr2 = 0; 421 422 /* Second pass: assign PMCs, set all MMCR1 fields */ 423 for (i = 0; i < n_ev; ++i) { 424 pmc = (event[i] >> EVENT_PMC_SHIFT) & EVENT_PMC_MASK; 425 unit = (event[i] >> EVENT_UNIT_SHIFT) & EVENT_UNIT_MASK; 426 combine = (event[i] >> EVENT_COMBINE_SHIFT) & EVENT_COMBINE_MASK; 427 psel = event[i] & EVENT_PSEL_MASK; 428 429 if (!pmc) { 430 for (pmc = 1; pmc <= 4; ++pmc) { 431 if (!(pmc_inuse & (1 << pmc))) 432 break; 433 } 434 435 pmc_inuse |= 1 << pmc; 436 } 437 438 if (pmc <= 4) { 439 mmcr1 |= unit << MMCR1_UNIT_SHIFT(pmc); 440 mmcr1 |= combine << MMCR1_COMBINE_SHIFT(pmc); 441 mmcr1 |= psel << MMCR1_PMCSEL_SHIFT(pmc); 442 } 443 444 if (event[i] & EVENT_IS_L1) { 445 cache = event[i] >> EVENT_CACHE_SEL_SHIFT; 446 mmcr1 |= (cache & 1) << MMCR1_IC_QUAL_SHIFT; 447 cache >>= 1; 448 mmcr1 |= (cache & 1) << MMCR1_DC_QUAL_SHIFT; 449 } 450 451 if (event[i] & EVENT_IS_MARKED) { 452 mmcra |= MMCRA_SAMPLE_ENABLE; 453 454 val = (event[i] >> EVENT_SAMPLE_SHIFT) & EVENT_SAMPLE_MASK; 455 if (val) { 456 mmcra |= (val & 3) << MMCRA_SAMP_MODE_SHIFT; 457 mmcra |= (val >> 2) << MMCRA_SAMP_ELIG_SHIFT; 458 } 459 } 460 461 /* 462 * PM_MRK_FAB_RSP_MATCH and PM_MRK_FAB_RSP_MATCH_CYC, 463 * the threshold bits are used for the match value. 464 */ 465 if (event_is_fab_match(event[i])) { 466 mmcr1 |= ((event[i] >> EVENT_THR_CTL_SHIFT) & 467 EVENT_THR_CTL_MASK) << MMCR1_FAB_SHIFT; 468 } else { 469 val = (event[i] >> EVENT_THR_CTL_SHIFT) & EVENT_THR_CTL_MASK; 470 mmcra |= val << MMCRA_THR_CTL_SHIFT; 471 val = (event[i] >> EVENT_THR_SEL_SHIFT) & EVENT_THR_SEL_MASK; 472 mmcra |= val << MMCRA_THR_SEL_SHIFT; 473 val = (event[i] >> EVENT_THR_CMP_SHIFT) & EVENT_THR_CMP_MASK; 474 mmcra |= val << MMCRA_THR_CMP_SHIFT; 475 } 476 477 if (event[i] & EVENT_WANTS_BHRB) { 478 val = (event[i] >> EVENT_IFM_SHIFT) & EVENT_IFM_MASK; 479 mmcra |= val << MMCRA_IFM_SHIFT; 480 } 481 482 if (pevents[i]->attr.exclude_user) 483 mmcr2 |= MMCR2_FCP(pmc); 484 485 if (pevents[i]->attr.exclude_hv) 486 mmcr2 |= MMCR2_FCH(pmc); 487 488 if (pevents[i]->attr.exclude_kernel) { 489 if (cpu_has_feature(CPU_FTR_HVMODE)) 490 mmcr2 |= MMCR2_FCH(pmc); 491 else 492 mmcr2 |= MMCR2_FCS(pmc); 493 } 494 495 hwc[i] = pmc - 1; 496 } 497 498 /* Return MMCRx values */ 499 mmcr[0] = 0; 500 501 /* pmc_inuse is 1-based */ 502 if (pmc_inuse & 2) 503 mmcr[0] = MMCR0_PMC1CE; 504 505 if (pmc_inuse & 0x7c) 506 mmcr[0] |= MMCR0_PMCjCE; 507 508 /* If we're not using PMC 5 or 6, freeze them */ 509 if (!(pmc_inuse & 0x60)) 510 mmcr[0] |= MMCR0_FC56; 511 512 mmcr[1] = mmcr1; 513 mmcr[2] = mmcra; 514 mmcr[3] = mmcr2; 515 516 return 0; 517 } 518 519 #define MAX_ALT 2 520 521 /* Table of alternatives, sorted by column 0 */ 522 static const unsigned int event_alternatives[][MAX_ALT] = { 523 { 0x10134, 0x301e2 }, /* PM_MRK_ST_CMPL */ 524 { 0x10138, 0x40138 }, /* PM_BR_MRK_2PATH */ 525 { 0x18082, 0x3e05e }, /* PM_L3_CO_MEPF */ 526 { 0x1d14e, 0x401e8 }, /* PM_MRK_DATA_FROM_L2MISS */ 527 { 0x1e054, 0x4000a }, /* PM_CMPLU_STALL */ 528 { 0x20036, 0x40036 }, /* PM_BR_2PATH */ 529 { 0x200f2, 0x300f2 }, /* PM_INST_DISP */ 530 { 0x200f4, 0x600f4 }, /* PM_RUN_CYC */ 531 { 0x2013c, 0x3012e }, /* PM_MRK_FILT_MATCH */ 532 { 0x3e054, 0x400f0 }, /* PM_LD_MISS_L1 */ 533 { 0x400fa, 0x500fa }, /* PM_RUN_INST_CMPL */ 534 }; 535 536 /* 537 * Scan the alternatives table for a match and return the 538 * index into the alternatives table if found, else -1. 539 */ 540 static int find_alternative(u64 event) 541 { 542 int i, j; 543 544 for (i = 0; i < ARRAY_SIZE(event_alternatives); ++i) { 545 if (event < event_alternatives[i][0]) 546 break; 547 548 for (j = 0; j < MAX_ALT && event_alternatives[i][j]; ++j) 549 if (event == event_alternatives[i][j]) 550 return i; 551 } 552 553 return -1; 554 } 555 556 static int power8_get_alternatives(u64 event, unsigned int flags, u64 alt[]) 557 { 558 int i, j, num_alt = 0; 559 u64 alt_event; 560 561 alt[num_alt++] = event; 562 563 i = find_alternative(event); 564 if (i >= 0) { 565 /* Filter out the original event, it's already in alt[0] */ 566 for (j = 0; j < MAX_ALT; ++j) { 567 alt_event = event_alternatives[i][j]; 568 if (alt_event && alt_event != event) 569 alt[num_alt++] = alt_event; 570 } 571 } 572 573 if (flags & PPMU_ONLY_COUNT_RUN) { 574 /* 575 * We're only counting in RUN state, so PM_CYC is equivalent to 576 * PM_RUN_CYC and PM_INST_CMPL === PM_RUN_INST_CMPL. 577 */ 578 j = num_alt; 579 for (i = 0; i < num_alt; ++i) { 580 switch (alt[i]) { 581 case 0x1e: /* PM_CYC */ 582 alt[j++] = 0x600f4; /* PM_RUN_CYC */ 583 break; 584 case 0x600f4: /* PM_RUN_CYC */ 585 alt[j++] = 0x1e; 586 break; 587 case 0x2: /* PM_PPC_CMPL */ 588 alt[j++] = 0x500fa; /* PM_RUN_INST_CMPL */ 589 break; 590 case 0x500fa: /* PM_RUN_INST_CMPL */ 591 alt[j++] = 0x2; /* PM_PPC_CMPL */ 592 break; 593 } 594 } 595 num_alt = j; 596 } 597 598 return num_alt; 599 } 600 601 static void power8_disable_pmc(unsigned int pmc, unsigned long mmcr[]) 602 { 603 if (pmc <= 3) 604 mmcr[1] &= ~(0xffUL << MMCR1_PMCSEL_SHIFT(pmc + 1)); 605 } 606 607 PMU_FORMAT_ATTR(event, "config:0-49"); 608 PMU_FORMAT_ATTR(pmcxsel, "config:0-7"); 609 PMU_FORMAT_ATTR(mark, "config:8"); 610 PMU_FORMAT_ATTR(combine, "config:11"); 611 PMU_FORMAT_ATTR(unit, "config:12-15"); 612 PMU_FORMAT_ATTR(pmc, "config:16-19"); 613 PMU_FORMAT_ATTR(cache_sel, "config:20-23"); 614 PMU_FORMAT_ATTR(sample_mode, "config:24-28"); 615 PMU_FORMAT_ATTR(thresh_sel, "config:29-31"); 616 PMU_FORMAT_ATTR(thresh_stop, "config:32-35"); 617 PMU_FORMAT_ATTR(thresh_start, "config:36-39"); 618 PMU_FORMAT_ATTR(thresh_cmp, "config:40-49"); 619 620 static struct attribute *power8_pmu_format_attr[] = { 621 &format_attr_event.attr, 622 &format_attr_pmcxsel.attr, 623 &format_attr_mark.attr, 624 &format_attr_combine.attr, 625 &format_attr_unit.attr, 626 &format_attr_pmc.attr, 627 &format_attr_cache_sel.attr, 628 &format_attr_sample_mode.attr, 629 &format_attr_thresh_sel.attr, 630 &format_attr_thresh_stop.attr, 631 &format_attr_thresh_start.attr, 632 &format_attr_thresh_cmp.attr, 633 NULL, 634 }; 635 636 struct attribute_group power8_pmu_format_group = { 637 .name = "format", 638 .attrs = power8_pmu_format_attr, 639 }; 640 641 static const struct attribute_group *power8_pmu_attr_groups[] = { 642 &power8_pmu_format_group, 643 NULL, 644 }; 645 646 static int power8_generic_events[] = { 647 [PERF_COUNT_HW_CPU_CYCLES] = PM_CYC, 648 [PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = PM_GCT_NOSLOT_CYC, 649 [PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = PM_CMPLU_STALL, 650 [PERF_COUNT_HW_INSTRUCTIONS] = PM_INST_CMPL, 651 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = PM_BRU_FIN, 652 [PERF_COUNT_HW_BRANCH_MISSES] = PM_BR_MPRED_CMPL, 653 [PERF_COUNT_HW_CACHE_REFERENCES] = PM_LD_REF_L1, 654 [PERF_COUNT_HW_CACHE_MISSES] = PM_LD_MISS_L1, 655 }; 656 657 static u64 power8_bhrb_filter_map(u64 branch_sample_type) 658 { 659 u64 pmu_bhrb_filter = 0; 660 661 /* BHRB and regular PMU events share the same privilege state 662 * filter configuration. BHRB is always recorded along with a 663 * regular PMU event. As the privilege state filter is handled 664 * in the basic PMC configuration of the accompanying regular 665 * PMU event, we ignore any separate BHRB specific request. 666 */ 667 668 /* No branch filter requested */ 669 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY) 670 return pmu_bhrb_filter; 671 672 /* Invalid branch filter options - HW does not support */ 673 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_RETURN) 674 return -1; 675 676 if (branch_sample_type & PERF_SAMPLE_BRANCH_IND_CALL) 677 return -1; 678 679 if (branch_sample_type & PERF_SAMPLE_BRANCH_ANY_CALL) { 680 pmu_bhrb_filter |= POWER8_MMCRA_IFM1; 681 return pmu_bhrb_filter; 682 } 683 684 /* Every thing else is unsupported */ 685 return -1; 686 } 687 688 static void power8_config_bhrb(u64 pmu_bhrb_filter) 689 { 690 /* Enable BHRB filter in PMU */ 691 mtspr(SPRN_MMCRA, (mfspr(SPRN_MMCRA) | pmu_bhrb_filter)); 692 } 693 694 #define C(x) PERF_COUNT_HW_CACHE_##x 695 696 /* 697 * Table of generalized cache-related events. 698 * 0 means not supported, -1 means nonsensical, other values 699 * are event codes. 700 */ 701 static int power8_cache_events[C(MAX)][C(OP_MAX)][C(RESULT_MAX)] = { 702 [ C(L1D) ] = { 703 [ C(OP_READ) ] = { 704 [ C(RESULT_ACCESS) ] = PM_LD_REF_L1, 705 [ C(RESULT_MISS) ] = PM_LD_MISS_L1, 706 }, 707 [ C(OP_WRITE) ] = { 708 [ C(RESULT_ACCESS) ] = 0, 709 [ C(RESULT_MISS) ] = PM_ST_MISS_L1, 710 }, 711 [ C(OP_PREFETCH) ] = { 712 [ C(RESULT_ACCESS) ] = PM_L1_PREF, 713 [ C(RESULT_MISS) ] = 0, 714 }, 715 }, 716 [ C(L1I) ] = { 717 [ C(OP_READ) ] = { 718 [ C(RESULT_ACCESS) ] = PM_INST_FROM_L1, 719 [ C(RESULT_MISS) ] = PM_L1_ICACHE_MISS, 720 }, 721 [ C(OP_WRITE) ] = { 722 [ C(RESULT_ACCESS) ] = PM_L1_DEMAND_WRITE, 723 [ C(RESULT_MISS) ] = -1, 724 }, 725 [ C(OP_PREFETCH) ] = { 726 [ C(RESULT_ACCESS) ] = PM_IC_PREF_WRITE, 727 [ C(RESULT_MISS) ] = 0, 728 }, 729 }, 730 [ C(LL) ] = { 731 [ C(OP_READ) ] = { 732 [ C(RESULT_ACCESS) ] = PM_DATA_FROM_L3, 733 [ C(RESULT_MISS) ] = PM_DATA_FROM_L3MISS, 734 }, 735 [ C(OP_WRITE) ] = { 736 [ C(RESULT_ACCESS) ] = PM_L2_ST, 737 [ C(RESULT_MISS) ] = PM_L2_ST_MISS, 738 }, 739 [ C(OP_PREFETCH) ] = { 740 [ C(RESULT_ACCESS) ] = PM_L3_PREF_ALL, 741 [ C(RESULT_MISS) ] = 0, 742 }, 743 }, 744 [ C(DTLB) ] = { 745 [ C(OP_READ) ] = { 746 [ C(RESULT_ACCESS) ] = 0, 747 [ C(RESULT_MISS) ] = PM_DTLB_MISS, 748 }, 749 [ C(OP_WRITE) ] = { 750 [ C(RESULT_ACCESS) ] = -1, 751 [ C(RESULT_MISS) ] = -1, 752 }, 753 [ C(OP_PREFETCH) ] = { 754 [ C(RESULT_ACCESS) ] = -1, 755 [ C(RESULT_MISS) ] = -1, 756 }, 757 }, 758 [ C(ITLB) ] = { 759 [ C(OP_READ) ] = { 760 [ C(RESULT_ACCESS) ] = 0, 761 [ C(RESULT_MISS) ] = PM_ITLB_MISS, 762 }, 763 [ C(OP_WRITE) ] = { 764 [ C(RESULT_ACCESS) ] = -1, 765 [ C(RESULT_MISS) ] = -1, 766 }, 767 [ C(OP_PREFETCH) ] = { 768 [ C(RESULT_ACCESS) ] = -1, 769 [ C(RESULT_MISS) ] = -1, 770 }, 771 }, 772 [ C(BPU) ] = { 773 [ C(OP_READ) ] = { 774 [ C(RESULT_ACCESS) ] = PM_BRU_FIN, 775 [ C(RESULT_MISS) ] = PM_BR_MPRED_CMPL, 776 }, 777 [ C(OP_WRITE) ] = { 778 [ C(RESULT_ACCESS) ] = -1, 779 [ C(RESULT_MISS) ] = -1, 780 }, 781 [ C(OP_PREFETCH) ] = { 782 [ C(RESULT_ACCESS) ] = -1, 783 [ C(RESULT_MISS) ] = -1, 784 }, 785 }, 786 [ C(NODE) ] = { 787 [ C(OP_READ) ] = { 788 [ C(RESULT_ACCESS) ] = -1, 789 [ C(RESULT_MISS) ] = -1, 790 }, 791 [ C(OP_WRITE) ] = { 792 [ C(RESULT_ACCESS) ] = -1, 793 [ C(RESULT_MISS) ] = -1, 794 }, 795 [ C(OP_PREFETCH) ] = { 796 [ C(RESULT_ACCESS) ] = -1, 797 [ C(RESULT_MISS) ] = -1, 798 }, 799 }, 800 }; 801 802 #undef C 803 804 static struct power_pmu power8_pmu = { 805 .name = "POWER8", 806 .n_counter = 6, 807 .max_alternatives = MAX_ALT + 1, 808 .add_fields = POWER8_ADD_FIELDS, 809 .test_adder = POWER8_TEST_ADDER, 810 .compute_mmcr = power8_compute_mmcr, 811 .config_bhrb = power8_config_bhrb, 812 .bhrb_filter_map = power8_bhrb_filter_map, 813 .get_constraint = power8_get_constraint, 814 .get_alternatives = power8_get_alternatives, 815 .disable_pmc = power8_disable_pmc, 816 .flags = PPMU_HAS_SSLOT | PPMU_HAS_SIER | PPMU_ARCH_207S, 817 .n_generic = ARRAY_SIZE(power8_generic_events), 818 .generic_events = power8_generic_events, 819 .cache_events = &power8_cache_events, 820 .attr_groups = power8_pmu_attr_groups, 821 .bhrb_nr = 32, 822 }; 823 824 static int __init init_power8_pmu(void) 825 { 826 int rc; 827 828 if (!cur_cpu_spec->oprofile_cpu_type || 829 strcmp(cur_cpu_spec->oprofile_cpu_type, "ppc64/power8")) 830 return -ENODEV; 831 832 rc = register_power_pmu(&power8_pmu); 833 if (rc) 834 return rc; 835 836 /* Tell userspace that EBB is supported */ 837 cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_EBB; 838 839 if (cpu_has_feature(CPU_FTR_PMAO_BUG)) 840 pr_info("PMAO restore workaround active.\n"); 841 842 return 0; 843 } 844 early_initcall(init_power8_pmu); 845