1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Per core/cpu state 4 * 5 * Used to coordinate shared registers between HT threads or 6 * among events on a single PMU. 7 */ 8 9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 10 11 #include <linux/stddef.h> 12 #include <linux/types.h> 13 #include <linux/init.h> 14 #include <linux/slab.h> 15 #include <linux/export.h> 16 #include <linux/nmi.h> 17 #include <linux/kvm_host.h> 18 19 #include <asm/cpufeature.h> 20 #include <asm/debugreg.h> 21 #include <asm/hardirq.h> 22 #include <asm/intel-family.h> 23 #include <asm/intel_pt.h> 24 #include <asm/apic.h> 25 #include <asm/cpu_device_id.h> 26 27 #include "../perf_event.h" 28 29 /* 30 * Intel PerfMon, used on Core and later. 31 */ 32 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = 33 { 34 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, 35 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, 36 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, 37 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, 38 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, 39 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, 40 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, 41 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ 42 }; 43 44 static struct event_constraint intel_core_event_constraints[] __read_mostly = 45 { 46 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 47 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 48 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 49 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 50 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 51 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ 52 EVENT_CONSTRAINT_END 53 }; 54 55 static struct event_constraint intel_core2_event_constraints[] __read_mostly = 56 { 57 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 58 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 59 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 60 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ 61 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 62 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 63 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 64 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 65 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ 66 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 67 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ 68 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ 69 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ 70 EVENT_CONSTRAINT_END 71 }; 72 73 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = 74 { 75 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 76 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 77 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 78 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ 79 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ 80 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ 81 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ 82 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ 83 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ 84 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 85 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 86 EVENT_CONSTRAINT_END 87 }; 88 89 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = 90 { 91 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 92 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 93 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 94 EVENT_EXTRA_END 95 }; 96 97 static struct event_constraint intel_westmere_event_constraints[] __read_mostly = 98 { 99 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 100 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 101 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 102 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 103 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ 104 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 105 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ 106 EVENT_CONSTRAINT_END 107 }; 108 109 static struct event_constraint intel_snb_event_constraints[] __read_mostly = 110 { 111 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 112 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 113 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 114 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 115 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 116 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 117 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 118 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ 119 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 120 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 121 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 122 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 123 124 /* 125 * When HT is off these events can only run on the bottom 4 counters 126 * When HT is on, they are impacted by the HT bug and require EXCL access 127 */ 128 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 129 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 130 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 131 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 132 133 EVENT_CONSTRAINT_END 134 }; 135 136 static struct event_constraint intel_ivb_event_constraints[] __read_mostly = 137 { 138 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 139 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 140 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 141 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ 142 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMPTY */ 143 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ 144 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ 145 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 146 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 147 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ 148 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 149 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 150 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 151 152 /* 153 * When HT is off these events can only run on the bottom 4 counters 154 * When HT is on, they are impacted by the HT bug and require EXCL access 155 */ 156 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 157 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 158 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 159 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 160 161 EVENT_CONSTRAINT_END 162 }; 163 164 static struct extra_reg intel_westmere_extra_regs[] __read_mostly = 165 { 166 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 167 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 168 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), 169 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 170 EVENT_EXTRA_END 171 }; 172 173 static struct event_constraint intel_v1_event_constraints[] __read_mostly = 174 { 175 EVENT_CONSTRAINT_END 176 }; 177 178 static struct event_constraint intel_gen_event_constraints[] __read_mostly = 179 { 180 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 181 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 182 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 183 EVENT_CONSTRAINT_END 184 }; 185 186 static struct event_constraint intel_v5_gen_event_constraints[] __read_mostly = 187 { 188 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 189 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 190 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 191 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 192 FIXED_EVENT_CONSTRAINT(0x0500, 4), 193 FIXED_EVENT_CONSTRAINT(0x0600, 5), 194 FIXED_EVENT_CONSTRAINT(0x0700, 6), 195 FIXED_EVENT_CONSTRAINT(0x0800, 7), 196 FIXED_EVENT_CONSTRAINT(0x0900, 8), 197 FIXED_EVENT_CONSTRAINT(0x0a00, 9), 198 FIXED_EVENT_CONSTRAINT(0x0b00, 10), 199 FIXED_EVENT_CONSTRAINT(0x0c00, 11), 200 FIXED_EVENT_CONSTRAINT(0x0d00, 12), 201 FIXED_EVENT_CONSTRAINT(0x0e00, 13), 202 FIXED_EVENT_CONSTRAINT(0x0f00, 14), 203 FIXED_EVENT_CONSTRAINT(0x1000, 15), 204 EVENT_CONSTRAINT_END 205 }; 206 207 static struct event_constraint intel_slm_event_constraints[] __read_mostly = 208 { 209 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 210 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 211 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ 212 EVENT_CONSTRAINT_END 213 }; 214 215 static struct event_constraint intel_grt_event_constraints[] __read_mostly = { 216 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 217 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 218 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ 219 FIXED_EVENT_CONSTRAINT(0x013c, 2), /* CPU_CLK_UNHALTED.REF_TSC_P */ 220 EVENT_CONSTRAINT_END 221 }; 222 223 static struct event_constraint intel_skl_event_constraints[] = { 224 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 225 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 226 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 227 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ 228 229 /* 230 * when HT is off, these can only run on the bottom 4 counters 231 */ 232 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 233 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 234 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 235 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 236 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */ 237 238 EVENT_CONSTRAINT_END 239 }; 240 241 static struct extra_reg intel_knl_extra_regs[] __read_mostly = { 242 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0), 243 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1), 244 EVENT_EXTRA_END 245 }; 246 247 static struct extra_reg intel_snb_extra_regs[] __read_mostly = { 248 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 249 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), 250 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), 251 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 252 EVENT_EXTRA_END 253 }; 254 255 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { 256 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 257 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 258 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 259 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 260 EVENT_EXTRA_END 261 }; 262 263 static struct extra_reg intel_skl_extra_regs[] __read_mostly = { 264 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 265 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 266 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 267 /* 268 * Note the low 8 bits eventsel code is not a continuous field, containing 269 * some #GPing bits. These are masked out. 270 */ 271 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 272 EVENT_EXTRA_END 273 }; 274 275 static struct event_constraint intel_icl_event_constraints[] = { 276 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 277 FIXED_EVENT_CONSTRAINT(0x01c0, 0), /* old INST_RETIRED.PREC_DIST */ 278 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 279 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 280 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 281 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 282 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 283 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 284 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 285 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 286 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf), 287 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf), 288 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */ 289 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x56, 0xf), 290 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf), 291 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */ 292 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */ 293 INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */ 294 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */ 295 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf), 296 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf), 297 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf), 298 INTEL_EVENT_CONSTRAINT(0xef, 0xf), 299 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf), 300 EVENT_CONSTRAINT_END 301 }; 302 303 static struct extra_reg intel_icl_extra_regs[] __read_mostly = { 304 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0), 305 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1), 306 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 307 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 308 EVENT_EXTRA_END 309 }; 310 311 static struct extra_reg intel_glc_extra_regs[] __read_mostly = { 312 INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 313 INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 314 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 315 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE), 316 INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE), 317 INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE), 318 EVENT_EXTRA_END 319 }; 320 321 static struct event_constraint intel_glc_event_constraints[] = { 322 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 323 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 324 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 325 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 326 FIXED_EVENT_CONSTRAINT(0x013c, 2), /* CPU_CLK_UNHALTED.REF_TSC_P */ 327 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 328 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 329 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 330 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 331 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 332 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4), 333 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5), 334 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6), 335 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7), 336 337 INTEL_EVENT_CONSTRAINT(0x2e, 0xff), 338 INTEL_EVENT_CONSTRAINT(0x3c, 0xff), 339 /* 340 * Generally event codes < 0x90 are restricted to counters 0-3. 341 * The 0x2E and 0x3C are exception, which has no restriction. 342 */ 343 INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf), 344 345 INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf), 346 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), 347 INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf), 348 INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1), 349 INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1), 350 INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1), 351 INTEL_EVENT_CONSTRAINT(0xce, 0x1), 352 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf), 353 /* 354 * Generally event codes >= 0x90 are likely to have no restrictions. 355 * The exception are defined as above. 356 */ 357 INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff), 358 359 EVENT_CONSTRAINT_END 360 }; 361 362 static struct extra_reg intel_rwc_extra_regs[] __read_mostly = { 363 INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 364 INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 365 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 366 INTEL_UEVENT_EXTRA_REG(0x02c6, MSR_PEBS_FRONTEND, 0x9, FE), 367 INTEL_UEVENT_EXTRA_REG(0x03c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE), 368 INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE), 369 INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE), 370 EVENT_EXTRA_END 371 }; 372 373 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); 374 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); 375 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); 376 377 static struct attribute *nhm_mem_events_attrs[] = { 378 EVENT_PTR(mem_ld_nhm), 379 NULL, 380 }; 381 382 /* 383 * topdown events for Intel Core CPUs. 384 * 385 * The events are all in slots, which is a free slot in a 4 wide 386 * pipeline. Some events are already reported in slots, for cycle 387 * events we multiply by the pipeline width (4). 388 * 389 * With Hyper Threading on, topdown metrics are either summed or averaged 390 * between the threads of a core: (count_t0 + count_t1). 391 * 392 * For the average case the metric is always scaled to pipeline width, 393 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4) 394 */ 395 396 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots, 397 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */ 398 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */ 399 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2"); 400 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued, 401 "event=0xe,umask=0x1"); /* uops_issued.any */ 402 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired, 403 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */ 404 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles, 405 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */ 406 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles, 407 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */ 408 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */ 409 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale, 410 "4", "2"); 411 412 EVENT_ATTR_STR(slots, slots, "event=0x00,umask=0x4"); 413 EVENT_ATTR_STR(topdown-retiring, td_retiring, "event=0x00,umask=0x80"); 414 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec, "event=0x00,umask=0x81"); 415 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound, "event=0x00,umask=0x82"); 416 EVENT_ATTR_STR(topdown-be-bound, td_be_bound, "event=0x00,umask=0x83"); 417 EVENT_ATTR_STR(topdown-heavy-ops, td_heavy_ops, "event=0x00,umask=0x84"); 418 EVENT_ATTR_STR(topdown-br-mispredict, td_br_mispredict, "event=0x00,umask=0x85"); 419 EVENT_ATTR_STR(topdown-fetch-lat, td_fetch_lat, "event=0x00,umask=0x86"); 420 EVENT_ATTR_STR(topdown-mem-bound, td_mem_bound, "event=0x00,umask=0x87"); 421 422 static struct attribute *snb_events_attrs[] = { 423 EVENT_PTR(td_slots_issued), 424 EVENT_PTR(td_slots_retired), 425 EVENT_PTR(td_fetch_bubbles), 426 EVENT_PTR(td_total_slots), 427 EVENT_PTR(td_total_slots_scale), 428 EVENT_PTR(td_recovery_bubbles), 429 EVENT_PTR(td_recovery_bubbles_scale), 430 NULL, 431 }; 432 433 static struct attribute *snb_mem_events_attrs[] = { 434 EVENT_PTR(mem_ld_snb), 435 EVENT_PTR(mem_st_snb), 436 NULL, 437 }; 438 439 static struct event_constraint intel_hsw_event_constraints[] = { 440 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 441 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 442 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 443 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 444 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 445 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 446 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 447 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), 448 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 449 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), 450 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 451 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), 452 453 /* 454 * When HT is off these events can only run on the bottom 4 counters 455 * When HT is on, they are impacted by the HT bug and require EXCL access 456 */ 457 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 458 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 459 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 460 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 461 462 EVENT_CONSTRAINT_END 463 }; 464 465 static struct event_constraint intel_bdw_event_constraints[] = { 466 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 467 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 468 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 469 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 470 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */ 471 /* 472 * when HT is off, these can only run on the bottom 4 counters 473 */ 474 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 475 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 476 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 477 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 478 EVENT_CONSTRAINT_END 479 }; 480 481 static u64 intel_pmu_event_map(int hw_event) 482 { 483 return intel_perfmon_event_map[hw_event]; 484 } 485 486 static __initconst const u64 glc_hw_cache_event_ids 487 [PERF_COUNT_HW_CACHE_MAX] 488 [PERF_COUNT_HW_CACHE_OP_MAX] 489 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 490 { 491 [ C(L1D ) ] = { 492 [ C(OP_READ) ] = { 493 [ C(RESULT_ACCESS) ] = 0x81d0, 494 [ C(RESULT_MISS) ] = 0xe124, 495 }, 496 [ C(OP_WRITE) ] = { 497 [ C(RESULT_ACCESS) ] = 0x82d0, 498 }, 499 }, 500 [ C(L1I ) ] = { 501 [ C(OP_READ) ] = { 502 [ C(RESULT_MISS) ] = 0xe424, 503 }, 504 [ C(OP_WRITE) ] = { 505 [ C(RESULT_ACCESS) ] = -1, 506 [ C(RESULT_MISS) ] = -1, 507 }, 508 }, 509 [ C(LL ) ] = { 510 [ C(OP_READ) ] = { 511 [ C(RESULT_ACCESS) ] = 0x12a, 512 [ C(RESULT_MISS) ] = 0x12a, 513 }, 514 [ C(OP_WRITE) ] = { 515 [ C(RESULT_ACCESS) ] = 0x12a, 516 [ C(RESULT_MISS) ] = 0x12a, 517 }, 518 }, 519 [ C(DTLB) ] = { 520 [ C(OP_READ) ] = { 521 [ C(RESULT_ACCESS) ] = 0x81d0, 522 [ C(RESULT_MISS) ] = 0xe12, 523 }, 524 [ C(OP_WRITE) ] = { 525 [ C(RESULT_ACCESS) ] = 0x82d0, 526 [ C(RESULT_MISS) ] = 0xe13, 527 }, 528 }, 529 [ C(ITLB) ] = { 530 [ C(OP_READ) ] = { 531 [ C(RESULT_ACCESS) ] = -1, 532 [ C(RESULT_MISS) ] = 0xe11, 533 }, 534 [ C(OP_WRITE) ] = { 535 [ C(RESULT_ACCESS) ] = -1, 536 [ C(RESULT_MISS) ] = -1, 537 }, 538 [ C(OP_PREFETCH) ] = { 539 [ C(RESULT_ACCESS) ] = -1, 540 [ C(RESULT_MISS) ] = -1, 541 }, 542 }, 543 [ C(BPU ) ] = { 544 [ C(OP_READ) ] = { 545 [ C(RESULT_ACCESS) ] = 0x4c4, 546 [ C(RESULT_MISS) ] = 0x4c5, 547 }, 548 [ C(OP_WRITE) ] = { 549 [ C(RESULT_ACCESS) ] = -1, 550 [ C(RESULT_MISS) ] = -1, 551 }, 552 [ C(OP_PREFETCH) ] = { 553 [ C(RESULT_ACCESS) ] = -1, 554 [ C(RESULT_MISS) ] = -1, 555 }, 556 }, 557 [ C(NODE) ] = { 558 [ C(OP_READ) ] = { 559 [ C(RESULT_ACCESS) ] = 0x12a, 560 [ C(RESULT_MISS) ] = 0x12a, 561 }, 562 }, 563 }; 564 565 static __initconst const u64 glc_hw_cache_extra_regs 566 [PERF_COUNT_HW_CACHE_MAX] 567 [PERF_COUNT_HW_CACHE_OP_MAX] 568 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 569 { 570 [ C(LL ) ] = { 571 [ C(OP_READ) ] = { 572 [ C(RESULT_ACCESS) ] = 0x10001, 573 [ C(RESULT_MISS) ] = 0x3fbfc00001, 574 }, 575 [ C(OP_WRITE) ] = { 576 [ C(RESULT_ACCESS) ] = 0x3f3ffc0002, 577 [ C(RESULT_MISS) ] = 0x3f3fc00002, 578 }, 579 }, 580 [ C(NODE) ] = { 581 [ C(OP_READ) ] = { 582 [ C(RESULT_ACCESS) ] = 0x10c000001, 583 [ C(RESULT_MISS) ] = 0x3fb3000001, 584 }, 585 }, 586 }; 587 588 /* 589 * Notes on the events: 590 * - data reads do not include code reads (comparable to earlier tables) 591 * - data counts include speculative execution (except L1 write, dtlb, bpu) 592 * - remote node access includes remote memory, remote cache, remote mmio. 593 * - prefetches are not included in the counts. 594 * - icache miss does not include decoded icache 595 */ 596 597 #define SKL_DEMAND_DATA_RD BIT_ULL(0) 598 #define SKL_DEMAND_RFO BIT_ULL(1) 599 #define SKL_ANY_RESPONSE BIT_ULL(16) 600 #define SKL_SUPPLIER_NONE BIT_ULL(17) 601 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26) 602 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27) 603 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28) 604 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29) 605 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \ 606 SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 607 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 608 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 609 #define SKL_SPL_HIT BIT_ULL(30) 610 #define SKL_SNOOP_NONE BIT_ULL(31) 611 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32) 612 #define SKL_SNOOP_MISS BIT_ULL(33) 613 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34) 614 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35) 615 #define SKL_SNOOP_HITM BIT_ULL(36) 616 #define SKL_SNOOP_NON_DRAM BIT_ULL(37) 617 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \ 618 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 619 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 620 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM) 621 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD 622 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \ 623 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 624 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 625 SKL_SNOOP_HITM|SKL_SPL_HIT) 626 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO 627 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE 628 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 629 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 630 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 631 632 static __initconst const u64 skl_hw_cache_event_ids 633 [PERF_COUNT_HW_CACHE_MAX] 634 [PERF_COUNT_HW_CACHE_OP_MAX] 635 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 636 { 637 [ C(L1D ) ] = { 638 [ C(OP_READ) ] = { 639 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 640 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 641 }, 642 [ C(OP_WRITE) ] = { 643 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 644 [ C(RESULT_MISS) ] = 0x0, 645 }, 646 [ C(OP_PREFETCH) ] = { 647 [ C(RESULT_ACCESS) ] = 0x0, 648 [ C(RESULT_MISS) ] = 0x0, 649 }, 650 }, 651 [ C(L1I ) ] = { 652 [ C(OP_READ) ] = { 653 [ C(RESULT_ACCESS) ] = 0x0, 654 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */ 655 }, 656 [ C(OP_WRITE) ] = { 657 [ C(RESULT_ACCESS) ] = -1, 658 [ C(RESULT_MISS) ] = -1, 659 }, 660 [ C(OP_PREFETCH) ] = { 661 [ C(RESULT_ACCESS) ] = 0x0, 662 [ C(RESULT_MISS) ] = 0x0, 663 }, 664 }, 665 [ C(LL ) ] = { 666 [ C(OP_READ) ] = { 667 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 668 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 669 }, 670 [ C(OP_WRITE) ] = { 671 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 672 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 673 }, 674 [ C(OP_PREFETCH) ] = { 675 [ C(RESULT_ACCESS) ] = 0x0, 676 [ C(RESULT_MISS) ] = 0x0, 677 }, 678 }, 679 [ C(DTLB) ] = { 680 [ C(OP_READ) ] = { 681 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 682 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 683 }, 684 [ C(OP_WRITE) ] = { 685 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 686 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 687 }, 688 [ C(OP_PREFETCH) ] = { 689 [ C(RESULT_ACCESS) ] = 0x0, 690 [ C(RESULT_MISS) ] = 0x0, 691 }, 692 }, 693 [ C(ITLB) ] = { 694 [ C(OP_READ) ] = { 695 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */ 696 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */ 697 }, 698 [ C(OP_WRITE) ] = { 699 [ C(RESULT_ACCESS) ] = -1, 700 [ C(RESULT_MISS) ] = -1, 701 }, 702 [ C(OP_PREFETCH) ] = { 703 [ C(RESULT_ACCESS) ] = -1, 704 [ C(RESULT_MISS) ] = -1, 705 }, 706 }, 707 [ C(BPU ) ] = { 708 [ C(OP_READ) ] = { 709 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 710 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 711 }, 712 [ C(OP_WRITE) ] = { 713 [ C(RESULT_ACCESS) ] = -1, 714 [ C(RESULT_MISS) ] = -1, 715 }, 716 [ C(OP_PREFETCH) ] = { 717 [ C(RESULT_ACCESS) ] = -1, 718 [ C(RESULT_MISS) ] = -1, 719 }, 720 }, 721 [ C(NODE) ] = { 722 [ C(OP_READ) ] = { 723 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 724 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 725 }, 726 [ C(OP_WRITE) ] = { 727 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 728 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 729 }, 730 [ C(OP_PREFETCH) ] = { 731 [ C(RESULT_ACCESS) ] = 0x0, 732 [ C(RESULT_MISS) ] = 0x0, 733 }, 734 }, 735 }; 736 737 static __initconst const u64 skl_hw_cache_extra_regs 738 [PERF_COUNT_HW_CACHE_MAX] 739 [PERF_COUNT_HW_CACHE_OP_MAX] 740 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 741 { 742 [ C(LL ) ] = { 743 [ C(OP_READ) ] = { 744 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 745 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 746 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 747 SKL_L3_MISS|SKL_ANY_SNOOP| 748 SKL_SUPPLIER_NONE, 749 }, 750 [ C(OP_WRITE) ] = { 751 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 752 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 753 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 754 SKL_L3_MISS|SKL_ANY_SNOOP| 755 SKL_SUPPLIER_NONE, 756 }, 757 [ C(OP_PREFETCH) ] = { 758 [ C(RESULT_ACCESS) ] = 0x0, 759 [ C(RESULT_MISS) ] = 0x0, 760 }, 761 }, 762 [ C(NODE) ] = { 763 [ C(OP_READ) ] = { 764 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 765 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 766 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 767 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 768 }, 769 [ C(OP_WRITE) ] = { 770 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 771 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 772 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 773 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 774 }, 775 [ C(OP_PREFETCH) ] = { 776 [ C(RESULT_ACCESS) ] = 0x0, 777 [ C(RESULT_MISS) ] = 0x0, 778 }, 779 }, 780 }; 781 782 #define SNB_DMND_DATA_RD (1ULL << 0) 783 #define SNB_DMND_RFO (1ULL << 1) 784 #define SNB_DMND_IFETCH (1ULL << 2) 785 #define SNB_DMND_WB (1ULL << 3) 786 #define SNB_PF_DATA_RD (1ULL << 4) 787 #define SNB_PF_RFO (1ULL << 5) 788 #define SNB_PF_IFETCH (1ULL << 6) 789 #define SNB_LLC_DATA_RD (1ULL << 7) 790 #define SNB_LLC_RFO (1ULL << 8) 791 #define SNB_LLC_IFETCH (1ULL << 9) 792 #define SNB_BUS_LOCKS (1ULL << 10) 793 #define SNB_STRM_ST (1ULL << 11) 794 #define SNB_OTHER (1ULL << 15) 795 #define SNB_RESP_ANY (1ULL << 16) 796 #define SNB_NO_SUPP (1ULL << 17) 797 #define SNB_LLC_HITM (1ULL << 18) 798 #define SNB_LLC_HITE (1ULL << 19) 799 #define SNB_LLC_HITS (1ULL << 20) 800 #define SNB_LLC_HITF (1ULL << 21) 801 #define SNB_LOCAL (1ULL << 22) 802 #define SNB_REMOTE (0xffULL << 23) 803 #define SNB_SNP_NONE (1ULL << 31) 804 #define SNB_SNP_NOT_NEEDED (1ULL << 32) 805 #define SNB_SNP_MISS (1ULL << 33) 806 #define SNB_NO_FWD (1ULL << 34) 807 #define SNB_SNP_FWD (1ULL << 35) 808 #define SNB_HITM (1ULL << 36) 809 #define SNB_NON_DRAM (1ULL << 37) 810 811 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) 812 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) 813 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 814 815 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ 816 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ 817 SNB_HITM) 818 819 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) 820 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) 821 822 #define SNB_L3_ACCESS SNB_RESP_ANY 823 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) 824 825 static __initconst const u64 snb_hw_cache_extra_regs 826 [PERF_COUNT_HW_CACHE_MAX] 827 [PERF_COUNT_HW_CACHE_OP_MAX] 828 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 829 { 830 [ C(LL ) ] = { 831 [ C(OP_READ) ] = { 832 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, 833 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, 834 }, 835 [ C(OP_WRITE) ] = { 836 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, 837 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, 838 }, 839 [ C(OP_PREFETCH) ] = { 840 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, 841 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, 842 }, 843 }, 844 [ C(NODE) ] = { 845 [ C(OP_READ) ] = { 846 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, 847 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, 848 }, 849 [ C(OP_WRITE) ] = { 850 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, 851 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, 852 }, 853 [ C(OP_PREFETCH) ] = { 854 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, 855 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, 856 }, 857 }, 858 }; 859 860 static __initconst const u64 snb_hw_cache_event_ids 861 [PERF_COUNT_HW_CACHE_MAX] 862 [PERF_COUNT_HW_CACHE_OP_MAX] 863 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 864 { 865 [ C(L1D) ] = { 866 [ C(OP_READ) ] = { 867 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ 868 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ 869 }, 870 [ C(OP_WRITE) ] = { 871 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ 872 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ 873 }, 874 [ C(OP_PREFETCH) ] = { 875 [ C(RESULT_ACCESS) ] = 0x0, 876 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ 877 }, 878 }, 879 [ C(L1I ) ] = { 880 [ C(OP_READ) ] = { 881 [ C(RESULT_ACCESS) ] = 0x0, 882 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ 883 }, 884 [ C(OP_WRITE) ] = { 885 [ C(RESULT_ACCESS) ] = -1, 886 [ C(RESULT_MISS) ] = -1, 887 }, 888 [ C(OP_PREFETCH) ] = { 889 [ C(RESULT_ACCESS) ] = 0x0, 890 [ C(RESULT_MISS) ] = 0x0, 891 }, 892 }, 893 [ C(LL ) ] = { 894 [ C(OP_READ) ] = { 895 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 896 [ C(RESULT_ACCESS) ] = 0x01b7, 897 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 898 [ C(RESULT_MISS) ] = 0x01b7, 899 }, 900 [ C(OP_WRITE) ] = { 901 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 902 [ C(RESULT_ACCESS) ] = 0x01b7, 903 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 904 [ C(RESULT_MISS) ] = 0x01b7, 905 }, 906 [ C(OP_PREFETCH) ] = { 907 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 908 [ C(RESULT_ACCESS) ] = 0x01b7, 909 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 910 [ C(RESULT_MISS) ] = 0x01b7, 911 }, 912 }, 913 [ C(DTLB) ] = { 914 [ C(OP_READ) ] = { 915 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ 916 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ 917 }, 918 [ C(OP_WRITE) ] = { 919 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ 920 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 921 }, 922 [ C(OP_PREFETCH) ] = { 923 [ C(RESULT_ACCESS) ] = 0x0, 924 [ C(RESULT_MISS) ] = 0x0, 925 }, 926 }, 927 [ C(ITLB) ] = { 928 [ C(OP_READ) ] = { 929 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ 930 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ 931 }, 932 [ C(OP_WRITE) ] = { 933 [ C(RESULT_ACCESS) ] = -1, 934 [ C(RESULT_MISS) ] = -1, 935 }, 936 [ C(OP_PREFETCH) ] = { 937 [ C(RESULT_ACCESS) ] = -1, 938 [ C(RESULT_MISS) ] = -1, 939 }, 940 }, 941 [ C(BPU ) ] = { 942 [ C(OP_READ) ] = { 943 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 944 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 945 }, 946 [ C(OP_WRITE) ] = { 947 [ C(RESULT_ACCESS) ] = -1, 948 [ C(RESULT_MISS) ] = -1, 949 }, 950 [ C(OP_PREFETCH) ] = { 951 [ C(RESULT_ACCESS) ] = -1, 952 [ C(RESULT_MISS) ] = -1, 953 }, 954 }, 955 [ C(NODE) ] = { 956 [ C(OP_READ) ] = { 957 [ C(RESULT_ACCESS) ] = 0x01b7, 958 [ C(RESULT_MISS) ] = 0x01b7, 959 }, 960 [ C(OP_WRITE) ] = { 961 [ C(RESULT_ACCESS) ] = 0x01b7, 962 [ C(RESULT_MISS) ] = 0x01b7, 963 }, 964 [ C(OP_PREFETCH) ] = { 965 [ C(RESULT_ACCESS) ] = 0x01b7, 966 [ C(RESULT_MISS) ] = 0x01b7, 967 }, 968 }, 969 970 }; 971 972 /* 973 * Notes on the events: 974 * - data reads do not include code reads (comparable to earlier tables) 975 * - data counts include speculative execution (except L1 write, dtlb, bpu) 976 * - remote node access includes remote memory, remote cache, remote mmio. 977 * - prefetches are not included in the counts because they are not 978 * reliably counted. 979 */ 980 981 #define HSW_DEMAND_DATA_RD BIT_ULL(0) 982 #define HSW_DEMAND_RFO BIT_ULL(1) 983 #define HSW_ANY_RESPONSE BIT_ULL(16) 984 #define HSW_SUPPLIER_NONE BIT_ULL(17) 985 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) 986 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) 987 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) 988 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) 989 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ 990 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 991 HSW_L3_MISS_REMOTE_HOP2P) 992 #define HSW_SNOOP_NONE BIT_ULL(31) 993 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) 994 #define HSW_SNOOP_MISS BIT_ULL(33) 995 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) 996 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) 997 #define HSW_SNOOP_HITM BIT_ULL(36) 998 #define HSW_SNOOP_NON_DRAM BIT_ULL(37) 999 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ 1000 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ 1001 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ 1002 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) 1003 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) 1004 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD 1005 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO 1006 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ 1007 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) 1008 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE 1009 1010 #define BDW_L3_MISS_LOCAL BIT(26) 1011 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ 1012 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 1013 HSW_L3_MISS_REMOTE_HOP2P) 1014 1015 1016 static __initconst const u64 hsw_hw_cache_event_ids 1017 [PERF_COUNT_HW_CACHE_MAX] 1018 [PERF_COUNT_HW_CACHE_OP_MAX] 1019 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1020 { 1021 [ C(L1D ) ] = { 1022 [ C(OP_READ) ] = { 1023 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1024 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 1025 }, 1026 [ C(OP_WRITE) ] = { 1027 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1028 [ C(RESULT_MISS) ] = 0x0, 1029 }, 1030 [ C(OP_PREFETCH) ] = { 1031 [ C(RESULT_ACCESS) ] = 0x0, 1032 [ C(RESULT_MISS) ] = 0x0, 1033 }, 1034 }, 1035 [ C(L1I ) ] = { 1036 [ C(OP_READ) ] = { 1037 [ C(RESULT_ACCESS) ] = 0x0, 1038 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ 1039 }, 1040 [ C(OP_WRITE) ] = { 1041 [ C(RESULT_ACCESS) ] = -1, 1042 [ C(RESULT_MISS) ] = -1, 1043 }, 1044 [ C(OP_PREFETCH) ] = { 1045 [ C(RESULT_ACCESS) ] = 0x0, 1046 [ C(RESULT_MISS) ] = 0x0, 1047 }, 1048 }, 1049 [ C(LL ) ] = { 1050 [ C(OP_READ) ] = { 1051 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1052 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1053 }, 1054 [ C(OP_WRITE) ] = { 1055 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1056 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1057 }, 1058 [ C(OP_PREFETCH) ] = { 1059 [ C(RESULT_ACCESS) ] = 0x0, 1060 [ C(RESULT_MISS) ] = 0x0, 1061 }, 1062 }, 1063 [ C(DTLB) ] = { 1064 [ C(OP_READ) ] = { 1065 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1066 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ 1067 }, 1068 [ C(OP_WRITE) ] = { 1069 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1070 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 1071 }, 1072 [ C(OP_PREFETCH) ] = { 1073 [ C(RESULT_ACCESS) ] = 0x0, 1074 [ C(RESULT_MISS) ] = 0x0, 1075 }, 1076 }, 1077 [ C(ITLB) ] = { 1078 [ C(OP_READ) ] = { 1079 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ 1080 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ 1081 }, 1082 [ C(OP_WRITE) ] = { 1083 [ C(RESULT_ACCESS) ] = -1, 1084 [ C(RESULT_MISS) ] = -1, 1085 }, 1086 [ C(OP_PREFETCH) ] = { 1087 [ C(RESULT_ACCESS) ] = -1, 1088 [ C(RESULT_MISS) ] = -1, 1089 }, 1090 }, 1091 [ C(BPU ) ] = { 1092 [ C(OP_READ) ] = { 1093 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1094 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1095 }, 1096 [ C(OP_WRITE) ] = { 1097 [ C(RESULT_ACCESS) ] = -1, 1098 [ C(RESULT_MISS) ] = -1, 1099 }, 1100 [ C(OP_PREFETCH) ] = { 1101 [ C(RESULT_ACCESS) ] = -1, 1102 [ C(RESULT_MISS) ] = -1, 1103 }, 1104 }, 1105 [ C(NODE) ] = { 1106 [ C(OP_READ) ] = { 1107 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1108 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1109 }, 1110 [ C(OP_WRITE) ] = { 1111 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1112 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1113 }, 1114 [ C(OP_PREFETCH) ] = { 1115 [ C(RESULT_ACCESS) ] = 0x0, 1116 [ C(RESULT_MISS) ] = 0x0, 1117 }, 1118 }, 1119 }; 1120 1121 static __initconst const u64 hsw_hw_cache_extra_regs 1122 [PERF_COUNT_HW_CACHE_MAX] 1123 [PERF_COUNT_HW_CACHE_OP_MAX] 1124 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1125 { 1126 [ C(LL ) ] = { 1127 [ C(OP_READ) ] = { 1128 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1129 HSW_LLC_ACCESS, 1130 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1131 HSW_L3_MISS|HSW_ANY_SNOOP, 1132 }, 1133 [ C(OP_WRITE) ] = { 1134 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1135 HSW_LLC_ACCESS, 1136 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1137 HSW_L3_MISS|HSW_ANY_SNOOP, 1138 }, 1139 [ C(OP_PREFETCH) ] = { 1140 [ C(RESULT_ACCESS) ] = 0x0, 1141 [ C(RESULT_MISS) ] = 0x0, 1142 }, 1143 }, 1144 [ C(NODE) ] = { 1145 [ C(OP_READ) ] = { 1146 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1147 HSW_L3_MISS_LOCAL_DRAM| 1148 HSW_SNOOP_DRAM, 1149 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1150 HSW_L3_MISS_REMOTE| 1151 HSW_SNOOP_DRAM, 1152 }, 1153 [ C(OP_WRITE) ] = { 1154 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1155 HSW_L3_MISS_LOCAL_DRAM| 1156 HSW_SNOOP_DRAM, 1157 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1158 HSW_L3_MISS_REMOTE| 1159 HSW_SNOOP_DRAM, 1160 }, 1161 [ C(OP_PREFETCH) ] = { 1162 [ C(RESULT_ACCESS) ] = 0x0, 1163 [ C(RESULT_MISS) ] = 0x0, 1164 }, 1165 }, 1166 }; 1167 1168 static __initconst const u64 westmere_hw_cache_event_ids 1169 [PERF_COUNT_HW_CACHE_MAX] 1170 [PERF_COUNT_HW_CACHE_OP_MAX] 1171 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1172 { 1173 [ C(L1D) ] = { 1174 [ C(OP_READ) ] = { 1175 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1176 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1177 }, 1178 [ C(OP_WRITE) ] = { 1179 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1180 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1181 }, 1182 [ C(OP_PREFETCH) ] = { 1183 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1184 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1185 }, 1186 }, 1187 [ C(L1I ) ] = { 1188 [ C(OP_READ) ] = { 1189 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1190 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1191 }, 1192 [ C(OP_WRITE) ] = { 1193 [ C(RESULT_ACCESS) ] = -1, 1194 [ C(RESULT_MISS) ] = -1, 1195 }, 1196 [ C(OP_PREFETCH) ] = { 1197 [ C(RESULT_ACCESS) ] = 0x0, 1198 [ C(RESULT_MISS) ] = 0x0, 1199 }, 1200 }, 1201 [ C(LL ) ] = { 1202 [ C(OP_READ) ] = { 1203 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1204 [ C(RESULT_ACCESS) ] = 0x01b7, 1205 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1206 [ C(RESULT_MISS) ] = 0x01b7, 1207 }, 1208 /* 1209 * Use RFO, not WRITEBACK, because a write miss would typically occur 1210 * on RFO. 1211 */ 1212 [ C(OP_WRITE) ] = { 1213 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1214 [ C(RESULT_ACCESS) ] = 0x01b7, 1215 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1216 [ C(RESULT_MISS) ] = 0x01b7, 1217 }, 1218 [ C(OP_PREFETCH) ] = { 1219 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1220 [ C(RESULT_ACCESS) ] = 0x01b7, 1221 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1222 [ C(RESULT_MISS) ] = 0x01b7, 1223 }, 1224 }, 1225 [ C(DTLB) ] = { 1226 [ C(OP_READ) ] = { 1227 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1228 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1229 }, 1230 [ C(OP_WRITE) ] = { 1231 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1232 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1233 }, 1234 [ C(OP_PREFETCH) ] = { 1235 [ C(RESULT_ACCESS) ] = 0x0, 1236 [ C(RESULT_MISS) ] = 0x0, 1237 }, 1238 }, 1239 [ C(ITLB) ] = { 1240 [ C(OP_READ) ] = { 1241 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1242 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ 1243 }, 1244 [ C(OP_WRITE) ] = { 1245 [ C(RESULT_ACCESS) ] = -1, 1246 [ C(RESULT_MISS) ] = -1, 1247 }, 1248 [ C(OP_PREFETCH) ] = { 1249 [ C(RESULT_ACCESS) ] = -1, 1250 [ C(RESULT_MISS) ] = -1, 1251 }, 1252 }, 1253 [ C(BPU ) ] = { 1254 [ C(OP_READ) ] = { 1255 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1256 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1257 }, 1258 [ C(OP_WRITE) ] = { 1259 [ C(RESULT_ACCESS) ] = -1, 1260 [ C(RESULT_MISS) ] = -1, 1261 }, 1262 [ C(OP_PREFETCH) ] = { 1263 [ C(RESULT_ACCESS) ] = -1, 1264 [ C(RESULT_MISS) ] = -1, 1265 }, 1266 }, 1267 [ C(NODE) ] = { 1268 [ C(OP_READ) ] = { 1269 [ C(RESULT_ACCESS) ] = 0x01b7, 1270 [ C(RESULT_MISS) ] = 0x01b7, 1271 }, 1272 [ C(OP_WRITE) ] = { 1273 [ C(RESULT_ACCESS) ] = 0x01b7, 1274 [ C(RESULT_MISS) ] = 0x01b7, 1275 }, 1276 [ C(OP_PREFETCH) ] = { 1277 [ C(RESULT_ACCESS) ] = 0x01b7, 1278 [ C(RESULT_MISS) ] = 0x01b7, 1279 }, 1280 }, 1281 }; 1282 1283 /* 1284 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; 1285 * See IA32 SDM Vol 3B 30.6.1.3 1286 */ 1287 1288 #define NHM_DMND_DATA_RD (1 << 0) 1289 #define NHM_DMND_RFO (1 << 1) 1290 #define NHM_DMND_IFETCH (1 << 2) 1291 #define NHM_DMND_WB (1 << 3) 1292 #define NHM_PF_DATA_RD (1 << 4) 1293 #define NHM_PF_DATA_RFO (1 << 5) 1294 #define NHM_PF_IFETCH (1 << 6) 1295 #define NHM_OFFCORE_OTHER (1 << 7) 1296 #define NHM_UNCORE_HIT (1 << 8) 1297 #define NHM_OTHER_CORE_HIT_SNP (1 << 9) 1298 #define NHM_OTHER_CORE_HITM (1 << 10) 1299 /* reserved */ 1300 #define NHM_REMOTE_CACHE_FWD (1 << 12) 1301 #define NHM_REMOTE_DRAM (1 << 13) 1302 #define NHM_LOCAL_DRAM (1 << 14) 1303 #define NHM_NON_DRAM (1 << 15) 1304 1305 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) 1306 #define NHM_REMOTE (NHM_REMOTE_DRAM) 1307 1308 #define NHM_DMND_READ (NHM_DMND_DATA_RD) 1309 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) 1310 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) 1311 1312 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) 1313 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) 1314 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) 1315 1316 static __initconst const u64 nehalem_hw_cache_extra_regs 1317 [PERF_COUNT_HW_CACHE_MAX] 1318 [PERF_COUNT_HW_CACHE_OP_MAX] 1319 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1320 { 1321 [ C(LL ) ] = { 1322 [ C(OP_READ) ] = { 1323 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, 1324 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, 1325 }, 1326 [ C(OP_WRITE) ] = { 1327 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, 1328 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, 1329 }, 1330 [ C(OP_PREFETCH) ] = { 1331 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, 1332 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, 1333 }, 1334 }, 1335 [ C(NODE) ] = { 1336 [ C(OP_READ) ] = { 1337 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, 1338 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, 1339 }, 1340 [ C(OP_WRITE) ] = { 1341 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, 1342 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, 1343 }, 1344 [ C(OP_PREFETCH) ] = { 1345 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, 1346 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, 1347 }, 1348 }, 1349 }; 1350 1351 static __initconst const u64 nehalem_hw_cache_event_ids 1352 [PERF_COUNT_HW_CACHE_MAX] 1353 [PERF_COUNT_HW_CACHE_OP_MAX] 1354 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1355 { 1356 [ C(L1D) ] = { 1357 [ C(OP_READ) ] = { 1358 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1359 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1360 }, 1361 [ C(OP_WRITE) ] = { 1362 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1363 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1364 }, 1365 [ C(OP_PREFETCH) ] = { 1366 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1367 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1368 }, 1369 }, 1370 [ C(L1I ) ] = { 1371 [ C(OP_READ) ] = { 1372 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1373 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1374 }, 1375 [ C(OP_WRITE) ] = { 1376 [ C(RESULT_ACCESS) ] = -1, 1377 [ C(RESULT_MISS) ] = -1, 1378 }, 1379 [ C(OP_PREFETCH) ] = { 1380 [ C(RESULT_ACCESS) ] = 0x0, 1381 [ C(RESULT_MISS) ] = 0x0, 1382 }, 1383 }, 1384 [ C(LL ) ] = { 1385 [ C(OP_READ) ] = { 1386 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1387 [ C(RESULT_ACCESS) ] = 0x01b7, 1388 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1389 [ C(RESULT_MISS) ] = 0x01b7, 1390 }, 1391 /* 1392 * Use RFO, not WRITEBACK, because a write miss would typically occur 1393 * on RFO. 1394 */ 1395 [ C(OP_WRITE) ] = { 1396 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1397 [ C(RESULT_ACCESS) ] = 0x01b7, 1398 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1399 [ C(RESULT_MISS) ] = 0x01b7, 1400 }, 1401 [ C(OP_PREFETCH) ] = { 1402 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1403 [ C(RESULT_ACCESS) ] = 0x01b7, 1404 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1405 [ C(RESULT_MISS) ] = 0x01b7, 1406 }, 1407 }, 1408 [ C(DTLB) ] = { 1409 [ C(OP_READ) ] = { 1410 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1411 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1412 }, 1413 [ C(OP_WRITE) ] = { 1414 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1415 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1416 }, 1417 [ C(OP_PREFETCH) ] = { 1418 [ C(RESULT_ACCESS) ] = 0x0, 1419 [ C(RESULT_MISS) ] = 0x0, 1420 }, 1421 }, 1422 [ C(ITLB) ] = { 1423 [ C(OP_READ) ] = { 1424 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1425 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ 1426 }, 1427 [ C(OP_WRITE) ] = { 1428 [ C(RESULT_ACCESS) ] = -1, 1429 [ C(RESULT_MISS) ] = -1, 1430 }, 1431 [ C(OP_PREFETCH) ] = { 1432 [ C(RESULT_ACCESS) ] = -1, 1433 [ C(RESULT_MISS) ] = -1, 1434 }, 1435 }, 1436 [ C(BPU ) ] = { 1437 [ C(OP_READ) ] = { 1438 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1439 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1440 }, 1441 [ C(OP_WRITE) ] = { 1442 [ C(RESULT_ACCESS) ] = -1, 1443 [ C(RESULT_MISS) ] = -1, 1444 }, 1445 [ C(OP_PREFETCH) ] = { 1446 [ C(RESULT_ACCESS) ] = -1, 1447 [ C(RESULT_MISS) ] = -1, 1448 }, 1449 }, 1450 [ C(NODE) ] = { 1451 [ C(OP_READ) ] = { 1452 [ C(RESULT_ACCESS) ] = 0x01b7, 1453 [ C(RESULT_MISS) ] = 0x01b7, 1454 }, 1455 [ C(OP_WRITE) ] = { 1456 [ C(RESULT_ACCESS) ] = 0x01b7, 1457 [ C(RESULT_MISS) ] = 0x01b7, 1458 }, 1459 [ C(OP_PREFETCH) ] = { 1460 [ C(RESULT_ACCESS) ] = 0x01b7, 1461 [ C(RESULT_MISS) ] = 0x01b7, 1462 }, 1463 }, 1464 }; 1465 1466 static __initconst const u64 core2_hw_cache_event_ids 1467 [PERF_COUNT_HW_CACHE_MAX] 1468 [PERF_COUNT_HW_CACHE_OP_MAX] 1469 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1470 { 1471 [ C(L1D) ] = { 1472 [ C(OP_READ) ] = { 1473 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ 1474 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ 1475 }, 1476 [ C(OP_WRITE) ] = { 1477 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ 1478 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ 1479 }, 1480 [ C(OP_PREFETCH) ] = { 1481 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ 1482 [ C(RESULT_MISS) ] = 0, 1483 }, 1484 }, 1485 [ C(L1I ) ] = { 1486 [ C(OP_READ) ] = { 1487 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ 1488 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ 1489 }, 1490 [ C(OP_WRITE) ] = { 1491 [ C(RESULT_ACCESS) ] = -1, 1492 [ C(RESULT_MISS) ] = -1, 1493 }, 1494 [ C(OP_PREFETCH) ] = { 1495 [ C(RESULT_ACCESS) ] = 0, 1496 [ C(RESULT_MISS) ] = 0, 1497 }, 1498 }, 1499 [ C(LL ) ] = { 1500 [ C(OP_READ) ] = { 1501 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1502 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1503 }, 1504 [ C(OP_WRITE) ] = { 1505 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1506 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1507 }, 1508 [ C(OP_PREFETCH) ] = { 1509 [ C(RESULT_ACCESS) ] = 0, 1510 [ C(RESULT_MISS) ] = 0, 1511 }, 1512 }, 1513 [ C(DTLB) ] = { 1514 [ C(OP_READ) ] = { 1515 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1516 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ 1517 }, 1518 [ C(OP_WRITE) ] = { 1519 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1520 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ 1521 }, 1522 [ C(OP_PREFETCH) ] = { 1523 [ C(RESULT_ACCESS) ] = 0, 1524 [ C(RESULT_MISS) ] = 0, 1525 }, 1526 }, 1527 [ C(ITLB) ] = { 1528 [ C(OP_READ) ] = { 1529 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1530 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ 1531 }, 1532 [ C(OP_WRITE) ] = { 1533 [ C(RESULT_ACCESS) ] = -1, 1534 [ C(RESULT_MISS) ] = -1, 1535 }, 1536 [ C(OP_PREFETCH) ] = { 1537 [ C(RESULT_ACCESS) ] = -1, 1538 [ C(RESULT_MISS) ] = -1, 1539 }, 1540 }, 1541 [ C(BPU ) ] = { 1542 [ C(OP_READ) ] = { 1543 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1544 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1545 }, 1546 [ C(OP_WRITE) ] = { 1547 [ C(RESULT_ACCESS) ] = -1, 1548 [ C(RESULT_MISS) ] = -1, 1549 }, 1550 [ C(OP_PREFETCH) ] = { 1551 [ C(RESULT_ACCESS) ] = -1, 1552 [ C(RESULT_MISS) ] = -1, 1553 }, 1554 }, 1555 }; 1556 1557 static __initconst const u64 atom_hw_cache_event_ids 1558 [PERF_COUNT_HW_CACHE_MAX] 1559 [PERF_COUNT_HW_CACHE_OP_MAX] 1560 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1561 { 1562 [ C(L1D) ] = { 1563 [ C(OP_READ) ] = { 1564 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ 1565 [ C(RESULT_MISS) ] = 0, 1566 }, 1567 [ C(OP_WRITE) ] = { 1568 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ 1569 [ C(RESULT_MISS) ] = 0, 1570 }, 1571 [ C(OP_PREFETCH) ] = { 1572 [ C(RESULT_ACCESS) ] = 0x0, 1573 [ C(RESULT_MISS) ] = 0, 1574 }, 1575 }, 1576 [ C(L1I ) ] = { 1577 [ C(OP_READ) ] = { 1578 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1579 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1580 }, 1581 [ C(OP_WRITE) ] = { 1582 [ C(RESULT_ACCESS) ] = -1, 1583 [ C(RESULT_MISS) ] = -1, 1584 }, 1585 [ C(OP_PREFETCH) ] = { 1586 [ C(RESULT_ACCESS) ] = 0, 1587 [ C(RESULT_MISS) ] = 0, 1588 }, 1589 }, 1590 [ C(LL ) ] = { 1591 [ C(OP_READ) ] = { 1592 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1593 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1594 }, 1595 [ C(OP_WRITE) ] = { 1596 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1597 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1598 }, 1599 [ C(OP_PREFETCH) ] = { 1600 [ C(RESULT_ACCESS) ] = 0, 1601 [ C(RESULT_MISS) ] = 0, 1602 }, 1603 }, 1604 [ C(DTLB) ] = { 1605 [ C(OP_READ) ] = { 1606 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ 1607 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ 1608 }, 1609 [ C(OP_WRITE) ] = { 1610 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ 1611 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ 1612 }, 1613 [ C(OP_PREFETCH) ] = { 1614 [ C(RESULT_ACCESS) ] = 0, 1615 [ C(RESULT_MISS) ] = 0, 1616 }, 1617 }, 1618 [ C(ITLB) ] = { 1619 [ C(OP_READ) ] = { 1620 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1621 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ 1622 }, 1623 [ C(OP_WRITE) ] = { 1624 [ C(RESULT_ACCESS) ] = -1, 1625 [ C(RESULT_MISS) ] = -1, 1626 }, 1627 [ C(OP_PREFETCH) ] = { 1628 [ C(RESULT_ACCESS) ] = -1, 1629 [ C(RESULT_MISS) ] = -1, 1630 }, 1631 }, 1632 [ C(BPU ) ] = { 1633 [ C(OP_READ) ] = { 1634 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1635 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1636 }, 1637 [ C(OP_WRITE) ] = { 1638 [ C(RESULT_ACCESS) ] = -1, 1639 [ C(RESULT_MISS) ] = -1, 1640 }, 1641 [ C(OP_PREFETCH) ] = { 1642 [ C(RESULT_ACCESS) ] = -1, 1643 [ C(RESULT_MISS) ] = -1, 1644 }, 1645 }, 1646 }; 1647 1648 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c"); 1649 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2"); 1650 /* no_alloc_cycles.not_delivered */ 1651 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm, 1652 "event=0xca,umask=0x50"); 1653 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2"); 1654 /* uops_retired.all */ 1655 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm, 1656 "event=0xc2,umask=0x10"); 1657 /* uops_retired.all */ 1658 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm, 1659 "event=0xc2,umask=0x10"); 1660 1661 static struct attribute *slm_events_attrs[] = { 1662 EVENT_PTR(td_total_slots_slm), 1663 EVENT_PTR(td_total_slots_scale_slm), 1664 EVENT_PTR(td_fetch_bubbles_slm), 1665 EVENT_PTR(td_fetch_bubbles_scale_slm), 1666 EVENT_PTR(td_slots_issued_slm), 1667 EVENT_PTR(td_slots_retired_slm), 1668 NULL 1669 }; 1670 1671 static struct extra_reg intel_slm_extra_regs[] __read_mostly = 1672 { 1673 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1674 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), 1675 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1), 1676 EVENT_EXTRA_END 1677 }; 1678 1679 #define SLM_DMND_READ SNB_DMND_DATA_RD 1680 #define SLM_DMND_WRITE SNB_DMND_RFO 1681 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1682 1683 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) 1684 #define SLM_LLC_ACCESS SNB_RESP_ANY 1685 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) 1686 1687 static __initconst const u64 slm_hw_cache_extra_regs 1688 [PERF_COUNT_HW_CACHE_MAX] 1689 [PERF_COUNT_HW_CACHE_OP_MAX] 1690 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1691 { 1692 [ C(LL ) ] = { 1693 [ C(OP_READ) ] = { 1694 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, 1695 [ C(RESULT_MISS) ] = 0, 1696 }, 1697 [ C(OP_WRITE) ] = { 1698 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, 1699 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, 1700 }, 1701 [ C(OP_PREFETCH) ] = { 1702 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, 1703 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, 1704 }, 1705 }, 1706 }; 1707 1708 static __initconst const u64 slm_hw_cache_event_ids 1709 [PERF_COUNT_HW_CACHE_MAX] 1710 [PERF_COUNT_HW_CACHE_OP_MAX] 1711 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1712 { 1713 [ C(L1D) ] = { 1714 [ C(OP_READ) ] = { 1715 [ C(RESULT_ACCESS) ] = 0, 1716 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ 1717 }, 1718 [ C(OP_WRITE) ] = { 1719 [ C(RESULT_ACCESS) ] = 0, 1720 [ C(RESULT_MISS) ] = 0, 1721 }, 1722 [ C(OP_PREFETCH) ] = { 1723 [ C(RESULT_ACCESS) ] = 0, 1724 [ C(RESULT_MISS) ] = 0, 1725 }, 1726 }, 1727 [ C(L1I ) ] = { 1728 [ C(OP_READ) ] = { 1729 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ 1730 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ 1731 }, 1732 [ C(OP_WRITE) ] = { 1733 [ C(RESULT_ACCESS) ] = -1, 1734 [ C(RESULT_MISS) ] = -1, 1735 }, 1736 [ C(OP_PREFETCH) ] = { 1737 [ C(RESULT_ACCESS) ] = 0, 1738 [ C(RESULT_MISS) ] = 0, 1739 }, 1740 }, 1741 [ C(LL ) ] = { 1742 [ C(OP_READ) ] = { 1743 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1744 [ C(RESULT_ACCESS) ] = 0x01b7, 1745 [ C(RESULT_MISS) ] = 0, 1746 }, 1747 [ C(OP_WRITE) ] = { 1748 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1749 [ C(RESULT_ACCESS) ] = 0x01b7, 1750 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1751 [ C(RESULT_MISS) ] = 0x01b7, 1752 }, 1753 [ C(OP_PREFETCH) ] = { 1754 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1755 [ C(RESULT_ACCESS) ] = 0x01b7, 1756 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1757 [ C(RESULT_MISS) ] = 0x01b7, 1758 }, 1759 }, 1760 [ C(DTLB) ] = { 1761 [ C(OP_READ) ] = { 1762 [ C(RESULT_ACCESS) ] = 0, 1763 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ 1764 }, 1765 [ C(OP_WRITE) ] = { 1766 [ C(RESULT_ACCESS) ] = 0, 1767 [ C(RESULT_MISS) ] = 0, 1768 }, 1769 [ C(OP_PREFETCH) ] = { 1770 [ C(RESULT_ACCESS) ] = 0, 1771 [ C(RESULT_MISS) ] = 0, 1772 }, 1773 }, 1774 [ C(ITLB) ] = { 1775 [ C(OP_READ) ] = { 1776 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1777 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */ 1778 }, 1779 [ C(OP_WRITE) ] = { 1780 [ C(RESULT_ACCESS) ] = -1, 1781 [ C(RESULT_MISS) ] = -1, 1782 }, 1783 [ C(OP_PREFETCH) ] = { 1784 [ C(RESULT_ACCESS) ] = -1, 1785 [ C(RESULT_MISS) ] = -1, 1786 }, 1787 }, 1788 [ C(BPU ) ] = { 1789 [ C(OP_READ) ] = { 1790 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1791 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1792 }, 1793 [ C(OP_WRITE) ] = { 1794 [ C(RESULT_ACCESS) ] = -1, 1795 [ C(RESULT_MISS) ] = -1, 1796 }, 1797 [ C(OP_PREFETCH) ] = { 1798 [ C(RESULT_ACCESS) ] = -1, 1799 [ C(RESULT_MISS) ] = -1, 1800 }, 1801 }, 1802 }; 1803 1804 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c"); 1805 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3"); 1806 /* UOPS_NOT_DELIVERED.ANY */ 1807 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c"); 1808 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */ 1809 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02"); 1810 /* UOPS_RETIRED.ANY */ 1811 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2"); 1812 /* UOPS_ISSUED.ANY */ 1813 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e"); 1814 1815 static struct attribute *glm_events_attrs[] = { 1816 EVENT_PTR(td_total_slots_glm), 1817 EVENT_PTR(td_total_slots_scale_glm), 1818 EVENT_PTR(td_fetch_bubbles_glm), 1819 EVENT_PTR(td_recovery_bubbles_glm), 1820 EVENT_PTR(td_slots_issued_glm), 1821 EVENT_PTR(td_slots_retired_glm), 1822 NULL 1823 }; 1824 1825 static struct extra_reg intel_glm_extra_regs[] __read_mostly = { 1826 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1827 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0), 1828 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1), 1829 EVENT_EXTRA_END 1830 }; 1831 1832 #define GLM_DEMAND_DATA_RD BIT_ULL(0) 1833 #define GLM_DEMAND_RFO BIT_ULL(1) 1834 #define GLM_ANY_RESPONSE BIT_ULL(16) 1835 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33) 1836 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD 1837 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO 1838 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1839 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE 1840 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM) 1841 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM) 1842 1843 static __initconst const u64 glm_hw_cache_event_ids 1844 [PERF_COUNT_HW_CACHE_MAX] 1845 [PERF_COUNT_HW_CACHE_OP_MAX] 1846 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1847 [C(L1D)] = { 1848 [C(OP_READ)] = { 1849 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1850 [C(RESULT_MISS)] = 0x0, 1851 }, 1852 [C(OP_WRITE)] = { 1853 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1854 [C(RESULT_MISS)] = 0x0, 1855 }, 1856 [C(OP_PREFETCH)] = { 1857 [C(RESULT_ACCESS)] = 0x0, 1858 [C(RESULT_MISS)] = 0x0, 1859 }, 1860 }, 1861 [C(L1I)] = { 1862 [C(OP_READ)] = { 1863 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1864 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1865 }, 1866 [C(OP_WRITE)] = { 1867 [C(RESULT_ACCESS)] = -1, 1868 [C(RESULT_MISS)] = -1, 1869 }, 1870 [C(OP_PREFETCH)] = { 1871 [C(RESULT_ACCESS)] = 0x0, 1872 [C(RESULT_MISS)] = 0x0, 1873 }, 1874 }, 1875 [C(LL)] = { 1876 [C(OP_READ)] = { 1877 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1878 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1879 }, 1880 [C(OP_WRITE)] = { 1881 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1882 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1883 }, 1884 [C(OP_PREFETCH)] = { 1885 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1886 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1887 }, 1888 }, 1889 [C(DTLB)] = { 1890 [C(OP_READ)] = { 1891 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1892 [C(RESULT_MISS)] = 0x0, 1893 }, 1894 [C(OP_WRITE)] = { 1895 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1896 [C(RESULT_MISS)] = 0x0, 1897 }, 1898 [C(OP_PREFETCH)] = { 1899 [C(RESULT_ACCESS)] = 0x0, 1900 [C(RESULT_MISS)] = 0x0, 1901 }, 1902 }, 1903 [C(ITLB)] = { 1904 [C(OP_READ)] = { 1905 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 1906 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 1907 }, 1908 [C(OP_WRITE)] = { 1909 [C(RESULT_ACCESS)] = -1, 1910 [C(RESULT_MISS)] = -1, 1911 }, 1912 [C(OP_PREFETCH)] = { 1913 [C(RESULT_ACCESS)] = -1, 1914 [C(RESULT_MISS)] = -1, 1915 }, 1916 }, 1917 [C(BPU)] = { 1918 [C(OP_READ)] = { 1919 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1920 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1921 }, 1922 [C(OP_WRITE)] = { 1923 [C(RESULT_ACCESS)] = -1, 1924 [C(RESULT_MISS)] = -1, 1925 }, 1926 [C(OP_PREFETCH)] = { 1927 [C(RESULT_ACCESS)] = -1, 1928 [C(RESULT_MISS)] = -1, 1929 }, 1930 }, 1931 }; 1932 1933 static __initconst const u64 glm_hw_cache_extra_regs 1934 [PERF_COUNT_HW_CACHE_MAX] 1935 [PERF_COUNT_HW_CACHE_OP_MAX] 1936 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1937 [C(LL)] = { 1938 [C(OP_READ)] = { 1939 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 1940 GLM_LLC_ACCESS, 1941 [C(RESULT_MISS)] = GLM_DEMAND_READ| 1942 GLM_LLC_MISS, 1943 }, 1944 [C(OP_WRITE)] = { 1945 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 1946 GLM_LLC_ACCESS, 1947 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 1948 GLM_LLC_MISS, 1949 }, 1950 [C(OP_PREFETCH)] = { 1951 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH| 1952 GLM_LLC_ACCESS, 1953 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH| 1954 GLM_LLC_MISS, 1955 }, 1956 }, 1957 }; 1958 1959 static __initconst const u64 glp_hw_cache_event_ids 1960 [PERF_COUNT_HW_CACHE_MAX] 1961 [PERF_COUNT_HW_CACHE_OP_MAX] 1962 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1963 [C(L1D)] = { 1964 [C(OP_READ)] = { 1965 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1966 [C(RESULT_MISS)] = 0x0, 1967 }, 1968 [C(OP_WRITE)] = { 1969 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1970 [C(RESULT_MISS)] = 0x0, 1971 }, 1972 [C(OP_PREFETCH)] = { 1973 [C(RESULT_ACCESS)] = 0x0, 1974 [C(RESULT_MISS)] = 0x0, 1975 }, 1976 }, 1977 [C(L1I)] = { 1978 [C(OP_READ)] = { 1979 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1980 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1981 }, 1982 [C(OP_WRITE)] = { 1983 [C(RESULT_ACCESS)] = -1, 1984 [C(RESULT_MISS)] = -1, 1985 }, 1986 [C(OP_PREFETCH)] = { 1987 [C(RESULT_ACCESS)] = 0x0, 1988 [C(RESULT_MISS)] = 0x0, 1989 }, 1990 }, 1991 [C(LL)] = { 1992 [C(OP_READ)] = { 1993 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1994 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1995 }, 1996 [C(OP_WRITE)] = { 1997 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1998 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1999 }, 2000 [C(OP_PREFETCH)] = { 2001 [C(RESULT_ACCESS)] = 0x0, 2002 [C(RESULT_MISS)] = 0x0, 2003 }, 2004 }, 2005 [C(DTLB)] = { 2006 [C(OP_READ)] = { 2007 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 2008 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 2009 }, 2010 [C(OP_WRITE)] = { 2011 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 2012 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 2013 }, 2014 [C(OP_PREFETCH)] = { 2015 [C(RESULT_ACCESS)] = 0x0, 2016 [C(RESULT_MISS)] = 0x0, 2017 }, 2018 }, 2019 [C(ITLB)] = { 2020 [C(OP_READ)] = { 2021 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 2022 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 2023 }, 2024 [C(OP_WRITE)] = { 2025 [C(RESULT_ACCESS)] = -1, 2026 [C(RESULT_MISS)] = -1, 2027 }, 2028 [C(OP_PREFETCH)] = { 2029 [C(RESULT_ACCESS)] = -1, 2030 [C(RESULT_MISS)] = -1, 2031 }, 2032 }, 2033 [C(BPU)] = { 2034 [C(OP_READ)] = { 2035 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 2036 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 2037 }, 2038 [C(OP_WRITE)] = { 2039 [C(RESULT_ACCESS)] = -1, 2040 [C(RESULT_MISS)] = -1, 2041 }, 2042 [C(OP_PREFETCH)] = { 2043 [C(RESULT_ACCESS)] = -1, 2044 [C(RESULT_MISS)] = -1, 2045 }, 2046 }, 2047 }; 2048 2049 static __initconst const u64 glp_hw_cache_extra_regs 2050 [PERF_COUNT_HW_CACHE_MAX] 2051 [PERF_COUNT_HW_CACHE_OP_MAX] 2052 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2053 [C(LL)] = { 2054 [C(OP_READ)] = { 2055 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 2056 GLM_LLC_ACCESS, 2057 [C(RESULT_MISS)] = GLM_DEMAND_READ| 2058 GLM_LLC_MISS, 2059 }, 2060 [C(OP_WRITE)] = { 2061 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 2062 GLM_LLC_ACCESS, 2063 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 2064 GLM_LLC_MISS, 2065 }, 2066 [C(OP_PREFETCH)] = { 2067 [C(RESULT_ACCESS)] = 0x0, 2068 [C(RESULT_MISS)] = 0x0, 2069 }, 2070 }, 2071 }; 2072 2073 #define TNT_LOCAL_DRAM BIT_ULL(26) 2074 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD 2075 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO 2076 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE 2077 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \ 2078 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM) 2079 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM) 2080 2081 static __initconst const u64 tnt_hw_cache_extra_regs 2082 [PERF_COUNT_HW_CACHE_MAX] 2083 [PERF_COUNT_HW_CACHE_OP_MAX] 2084 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2085 [C(LL)] = { 2086 [C(OP_READ)] = { 2087 [C(RESULT_ACCESS)] = TNT_DEMAND_READ| 2088 TNT_LLC_ACCESS, 2089 [C(RESULT_MISS)] = TNT_DEMAND_READ| 2090 TNT_LLC_MISS, 2091 }, 2092 [C(OP_WRITE)] = { 2093 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE| 2094 TNT_LLC_ACCESS, 2095 [C(RESULT_MISS)] = TNT_DEMAND_WRITE| 2096 TNT_LLC_MISS, 2097 }, 2098 [C(OP_PREFETCH)] = { 2099 [C(RESULT_ACCESS)] = 0x0, 2100 [C(RESULT_MISS)] = 0x0, 2101 }, 2102 }, 2103 }; 2104 2105 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound_tnt, "event=0x71,umask=0x0"); 2106 EVENT_ATTR_STR(topdown-retiring, td_retiring_tnt, "event=0xc2,umask=0x0"); 2107 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec_tnt, "event=0x73,umask=0x6"); 2108 EVENT_ATTR_STR(topdown-be-bound, td_be_bound_tnt, "event=0x74,umask=0x0"); 2109 2110 static struct attribute *tnt_events_attrs[] = { 2111 EVENT_PTR(td_fe_bound_tnt), 2112 EVENT_PTR(td_retiring_tnt), 2113 EVENT_PTR(td_bad_spec_tnt), 2114 EVENT_PTR(td_be_bound_tnt), 2115 NULL, 2116 }; 2117 2118 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = { 2119 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2120 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0), 2121 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1), 2122 EVENT_EXTRA_END 2123 }; 2124 2125 EVENT_ATTR_STR(mem-loads, mem_ld_grt, "event=0xd0,umask=0x5,ldlat=3"); 2126 EVENT_ATTR_STR(mem-stores, mem_st_grt, "event=0xd0,umask=0x6"); 2127 2128 static struct attribute *grt_mem_attrs[] = { 2129 EVENT_PTR(mem_ld_grt), 2130 EVENT_PTR(mem_st_grt), 2131 NULL 2132 }; 2133 2134 static struct extra_reg intel_grt_extra_regs[] __read_mostly = { 2135 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2136 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 2137 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 2138 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0), 2139 EVENT_EXTRA_END 2140 }; 2141 2142 EVENT_ATTR_STR(topdown-retiring, td_retiring_cmt, "event=0x72,umask=0x0"); 2143 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec_cmt, "event=0x73,umask=0x0"); 2144 2145 static struct attribute *cmt_events_attrs[] = { 2146 EVENT_PTR(td_fe_bound_tnt), 2147 EVENT_PTR(td_retiring_cmt), 2148 EVENT_PTR(td_bad_spec_cmt), 2149 EVENT_PTR(td_be_bound_tnt), 2150 NULL 2151 }; 2152 2153 static struct extra_reg intel_cmt_extra_regs[] __read_mostly = { 2154 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2155 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff3ffffffffffull, RSP_0), 2156 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff3ffffffffffull, RSP_1), 2157 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0), 2158 INTEL_UEVENT_EXTRA_REG(0x0127, MSR_SNOOP_RSP_0, 0xffffffffffffffffull, SNOOP_0), 2159 INTEL_UEVENT_EXTRA_REG(0x0227, MSR_SNOOP_RSP_1, 0xffffffffffffffffull, SNOOP_1), 2160 EVENT_EXTRA_END 2161 }; 2162 2163 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */ 2164 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */ 2165 #define KNL_MCDRAM_LOCAL BIT_ULL(21) 2166 #define KNL_MCDRAM_FAR BIT_ULL(22) 2167 #define KNL_DDR_LOCAL BIT_ULL(23) 2168 #define KNL_DDR_FAR BIT_ULL(24) 2169 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \ 2170 KNL_DDR_LOCAL | KNL_DDR_FAR) 2171 #define KNL_L2_READ SLM_DMND_READ 2172 #define KNL_L2_WRITE SLM_DMND_WRITE 2173 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH 2174 #define KNL_L2_ACCESS SLM_LLC_ACCESS 2175 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \ 2176 KNL_DRAM_ANY | SNB_SNP_ANY | \ 2177 SNB_NON_DRAM) 2178 2179 static __initconst const u64 knl_hw_cache_extra_regs 2180 [PERF_COUNT_HW_CACHE_MAX] 2181 [PERF_COUNT_HW_CACHE_OP_MAX] 2182 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2183 [C(LL)] = { 2184 [C(OP_READ)] = { 2185 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS, 2186 [C(RESULT_MISS)] = 0, 2187 }, 2188 [C(OP_WRITE)] = { 2189 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS, 2190 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS, 2191 }, 2192 [C(OP_PREFETCH)] = { 2193 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS, 2194 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS, 2195 }, 2196 }, 2197 }; 2198 2199 /* 2200 * Used from PMIs where the LBRs are already disabled. 2201 * 2202 * This function could be called consecutively. It is required to remain in 2203 * disabled state if called consecutively. 2204 * 2205 * During consecutive calls, the same disable value will be written to related 2206 * registers, so the PMU state remains unchanged. 2207 * 2208 * intel_bts events don't coexist with intel PMU's BTS events because of 2209 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them 2210 * disabled around intel PMU's event batching etc, only inside the PMI handler. 2211 * 2212 * Avoid PEBS_ENABLE MSR access in PMIs. 2213 * The GLOBAL_CTRL has been disabled. All the counters do not count anymore. 2214 * It doesn't matter if the PEBS is enabled or not. 2215 * Usually, the PEBS status are not changed in PMIs. It's unnecessary to 2216 * access PEBS_ENABLE MSR in disable_all()/enable_all(). 2217 * However, there are some cases which may change PEBS status, e.g. PMI 2218 * throttle. The PEBS_ENABLE should be updated where the status changes. 2219 */ 2220 static __always_inline void __intel_pmu_disable_all(bool bts) 2221 { 2222 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2223 2224 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2225 2226 if (bts && test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) 2227 intel_pmu_disable_bts(); 2228 } 2229 2230 static __always_inline void intel_pmu_disable_all(void) 2231 { 2232 __intel_pmu_disable_all(true); 2233 intel_pmu_pebs_disable_all(); 2234 intel_pmu_lbr_disable_all(); 2235 } 2236 2237 static void __intel_pmu_enable_all(int added, bool pmi) 2238 { 2239 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2240 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2241 2242 intel_pmu_lbr_enable_all(pmi); 2243 2244 if (cpuc->fixed_ctrl_val != cpuc->active_fixed_ctrl_val) { 2245 wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, cpuc->fixed_ctrl_val); 2246 cpuc->active_fixed_ctrl_val = cpuc->fixed_ctrl_val; 2247 } 2248 2249 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 2250 intel_ctrl & ~cpuc->intel_ctrl_guest_mask); 2251 2252 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 2253 struct perf_event *event = 2254 cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; 2255 2256 if (WARN_ON_ONCE(!event)) 2257 return; 2258 2259 intel_pmu_enable_bts(event->hw.config); 2260 } 2261 } 2262 2263 static void intel_pmu_enable_all(int added) 2264 { 2265 intel_pmu_pebs_enable_all(); 2266 __intel_pmu_enable_all(added, false); 2267 } 2268 2269 static noinline int 2270 __intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, 2271 unsigned int cnt, unsigned long flags) 2272 { 2273 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2274 2275 intel_pmu_lbr_read(); 2276 cnt = min_t(unsigned int, cnt, x86_pmu.lbr_nr); 2277 2278 memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt); 2279 intel_pmu_enable_all(0); 2280 local_irq_restore(flags); 2281 return cnt; 2282 } 2283 2284 static int 2285 intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2286 { 2287 unsigned long flags; 2288 2289 /* must not have branches... */ 2290 local_irq_save(flags); 2291 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2292 __intel_pmu_lbr_disable(); 2293 /* ... until here */ 2294 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2295 } 2296 2297 static int 2298 intel_pmu_snapshot_arch_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2299 { 2300 unsigned long flags; 2301 2302 /* must not have branches... */ 2303 local_irq_save(flags); 2304 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2305 __intel_pmu_arch_lbr_disable(); 2306 /* ... until here */ 2307 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2308 } 2309 2310 /* 2311 * Workaround for: 2312 * Intel Errata AAK100 (model 26) 2313 * Intel Errata AAP53 (model 30) 2314 * Intel Errata BD53 (model 44) 2315 * 2316 * The official story: 2317 * These chips need to be 'reset' when adding counters by programming the 2318 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either 2319 * in sequence on the same PMC or on different PMCs. 2320 * 2321 * In practice it appears some of these events do in fact count, and 2322 * we need to program all 4 events. 2323 */ 2324 static void intel_pmu_nhm_workaround(void) 2325 { 2326 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2327 static const unsigned long nhm_magic[4] = { 2328 0x4300B5, 2329 0x4300D2, 2330 0x4300B1, 2331 0x4300B1 2332 }; 2333 struct perf_event *event; 2334 int i; 2335 2336 /* 2337 * The Errata requires below steps: 2338 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; 2339 * 2) Configure 4 PERFEVTSELx with the magic events and clear 2340 * the corresponding PMCx; 2341 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; 2342 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; 2343 * 5) Clear 4 pairs of ERFEVTSELx and PMCx; 2344 */ 2345 2346 /* 2347 * The real steps we choose are a little different from above. 2348 * A) To reduce MSR operations, we don't run step 1) as they 2349 * are already cleared before this function is called; 2350 * B) Call x86_perf_event_update to save PMCx before configuring 2351 * PERFEVTSELx with magic number; 2352 * C) With step 5), we do clear only when the PERFEVTSELx is 2353 * not used currently. 2354 * D) Call x86_perf_event_set_period to restore PMCx; 2355 */ 2356 2357 /* We always operate 4 pairs of PERF Counters */ 2358 for (i = 0; i < 4; i++) { 2359 event = cpuc->events[i]; 2360 if (event) 2361 static_call(x86_pmu_update)(event); 2362 } 2363 2364 for (i = 0; i < 4; i++) { 2365 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); 2366 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); 2367 } 2368 2369 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); 2370 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); 2371 2372 for (i = 0; i < 4; i++) { 2373 event = cpuc->events[i]; 2374 2375 if (event) { 2376 static_call(x86_pmu_set_period)(event); 2377 __x86_pmu_enable_event(&event->hw, 2378 ARCH_PERFMON_EVENTSEL_ENABLE); 2379 } else 2380 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); 2381 } 2382 } 2383 2384 static void intel_pmu_nhm_enable_all(int added) 2385 { 2386 if (added) 2387 intel_pmu_nhm_workaround(); 2388 intel_pmu_enable_all(added); 2389 } 2390 2391 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on) 2392 { 2393 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0; 2394 2395 if (cpuc->tfa_shadow != val) { 2396 cpuc->tfa_shadow = val; 2397 wrmsrl(MSR_TSX_FORCE_ABORT, val); 2398 } 2399 } 2400 2401 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2402 { 2403 /* 2404 * We're going to use PMC3, make sure TFA is set before we touch it. 2405 */ 2406 if (cntr == 3) 2407 intel_set_tfa(cpuc, true); 2408 } 2409 2410 static void intel_tfa_pmu_enable_all(int added) 2411 { 2412 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2413 2414 /* 2415 * If we find PMC3 is no longer used when we enable the PMU, we can 2416 * clear TFA. 2417 */ 2418 if (!test_bit(3, cpuc->active_mask)) 2419 intel_set_tfa(cpuc, false); 2420 2421 intel_pmu_enable_all(added); 2422 } 2423 2424 static inline u64 intel_pmu_get_status(void) 2425 { 2426 u64 status; 2427 2428 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); 2429 2430 return status; 2431 } 2432 2433 static inline void intel_pmu_ack_status(u64 ack) 2434 { 2435 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); 2436 } 2437 2438 static inline bool event_is_checkpointed(struct perf_event *event) 2439 { 2440 return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; 2441 } 2442 2443 static inline void intel_set_masks(struct perf_event *event, int idx) 2444 { 2445 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2446 2447 if (event->attr.exclude_host) 2448 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2449 if (event->attr.exclude_guest) 2450 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2451 if (event_is_checkpointed(event)) 2452 __set_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2453 } 2454 2455 static inline void intel_clear_masks(struct perf_event *event, int idx) 2456 { 2457 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2458 2459 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2460 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2461 __clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2462 } 2463 2464 static void intel_pmu_disable_fixed(struct perf_event *event) 2465 { 2466 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2467 struct hw_perf_event *hwc = &event->hw; 2468 int idx = hwc->idx; 2469 u64 mask; 2470 2471 if (is_topdown_idx(idx)) { 2472 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2473 2474 /* 2475 * When there are other active TopDown events, 2476 * don't disable the fixed counter 3. 2477 */ 2478 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2479 return; 2480 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2481 } 2482 2483 intel_clear_masks(event, idx); 2484 2485 mask = intel_fixed_bits_by_idx(idx - INTEL_PMC_IDX_FIXED, INTEL_FIXED_BITS_MASK); 2486 cpuc->fixed_ctrl_val &= ~mask; 2487 } 2488 2489 static void intel_pmu_disable_event(struct perf_event *event) 2490 { 2491 struct hw_perf_event *hwc = &event->hw; 2492 int idx = hwc->idx; 2493 2494 switch (idx) { 2495 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2496 intel_clear_masks(event, idx); 2497 x86_pmu_disable_event(event); 2498 break; 2499 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2500 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2501 intel_pmu_disable_fixed(event); 2502 break; 2503 case INTEL_PMC_IDX_FIXED_BTS: 2504 intel_pmu_disable_bts(); 2505 intel_pmu_drain_bts_buffer(); 2506 return; 2507 case INTEL_PMC_IDX_FIXED_VLBR: 2508 intel_clear_masks(event, idx); 2509 break; 2510 default: 2511 intel_clear_masks(event, idx); 2512 pr_warn("Failed to disable the event with invalid index %d\n", 2513 idx); 2514 return; 2515 } 2516 2517 /* 2518 * Needs to be called after x86_pmu_disable_event, 2519 * so we don't trigger the event without PEBS bit set. 2520 */ 2521 if (unlikely(event->attr.precise_ip)) 2522 intel_pmu_pebs_disable(event); 2523 } 2524 2525 static void intel_pmu_assign_event(struct perf_event *event, int idx) 2526 { 2527 if (is_pebs_pt(event)) 2528 perf_report_aux_output_id(event, idx); 2529 } 2530 2531 static __always_inline bool intel_pmu_needs_branch_stack(struct perf_event *event) 2532 { 2533 return event->hw.flags & PERF_X86_EVENT_NEEDS_BRANCH_STACK; 2534 } 2535 2536 static void intel_pmu_del_event(struct perf_event *event) 2537 { 2538 if (intel_pmu_needs_branch_stack(event)) 2539 intel_pmu_lbr_del(event); 2540 if (event->attr.precise_ip) 2541 intel_pmu_pebs_del(event); 2542 } 2543 2544 static int icl_set_topdown_event_period(struct perf_event *event) 2545 { 2546 struct hw_perf_event *hwc = &event->hw; 2547 s64 left = local64_read(&hwc->period_left); 2548 2549 /* 2550 * The values in PERF_METRICS MSR are derived from fixed counter 3. 2551 * Software should start both registers, PERF_METRICS and fixed 2552 * counter 3, from zero. 2553 * Clear PERF_METRICS and Fixed counter 3 in initialization. 2554 * After that, both MSRs will be cleared for each read. 2555 * Don't need to clear them again. 2556 */ 2557 if (left == x86_pmu.max_period) { 2558 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2559 wrmsrl(MSR_PERF_METRICS, 0); 2560 hwc->saved_slots = 0; 2561 hwc->saved_metric = 0; 2562 } 2563 2564 if ((hwc->saved_slots) && is_slots_event(event)) { 2565 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots); 2566 wrmsrl(MSR_PERF_METRICS, hwc->saved_metric); 2567 } 2568 2569 perf_event_update_userpage(event); 2570 2571 return 0; 2572 } 2573 2574 DEFINE_STATIC_CALL(intel_pmu_set_topdown_event_period, x86_perf_event_set_period); 2575 2576 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx) 2577 { 2578 u32 val; 2579 2580 /* 2581 * The metric is reported as an 8bit integer fraction 2582 * summing up to 0xff. 2583 * slots-in-metric = (Metric / 0xff) * slots 2584 */ 2585 val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff; 2586 return mul_u64_u32_div(slots, val, 0xff); 2587 } 2588 2589 static u64 icl_get_topdown_value(struct perf_event *event, 2590 u64 slots, u64 metrics) 2591 { 2592 int idx = event->hw.idx; 2593 u64 delta; 2594 2595 if (is_metric_idx(idx)) 2596 delta = icl_get_metrics_event_value(metrics, slots, idx); 2597 else 2598 delta = slots; 2599 2600 return delta; 2601 } 2602 2603 static void __icl_update_topdown_event(struct perf_event *event, 2604 u64 slots, u64 metrics, 2605 u64 last_slots, u64 last_metrics) 2606 { 2607 u64 delta, last = 0; 2608 2609 delta = icl_get_topdown_value(event, slots, metrics); 2610 if (last_slots) 2611 last = icl_get_topdown_value(event, last_slots, last_metrics); 2612 2613 /* 2614 * The 8bit integer fraction of metric may be not accurate, 2615 * especially when the changes is very small. 2616 * For example, if only a few bad_spec happens, the fraction 2617 * may be reduced from 1 to 0. If so, the bad_spec event value 2618 * will be 0 which is definitely less than the last value. 2619 * Avoid update event->count for this case. 2620 */ 2621 if (delta > last) { 2622 delta -= last; 2623 local64_add(delta, &event->count); 2624 } 2625 } 2626 2627 static void update_saved_topdown_regs(struct perf_event *event, u64 slots, 2628 u64 metrics, int metric_end) 2629 { 2630 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2631 struct perf_event *other; 2632 int idx; 2633 2634 event->hw.saved_slots = slots; 2635 event->hw.saved_metric = metrics; 2636 2637 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2638 if (!is_topdown_idx(idx)) 2639 continue; 2640 other = cpuc->events[idx]; 2641 other->hw.saved_slots = slots; 2642 other->hw.saved_metric = metrics; 2643 } 2644 } 2645 2646 /* 2647 * Update all active Topdown events. 2648 * 2649 * The PERF_METRICS and Fixed counter 3 are read separately. The values may be 2650 * modify by a NMI. PMU has to be disabled before calling this function. 2651 */ 2652 2653 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end) 2654 { 2655 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2656 struct perf_event *other; 2657 u64 slots, metrics; 2658 bool reset = true; 2659 int idx; 2660 2661 /* read Fixed counter 3 */ 2662 rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots); 2663 if (!slots) 2664 return 0; 2665 2666 /* read PERF_METRICS */ 2667 rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics); 2668 2669 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2670 if (!is_topdown_idx(idx)) 2671 continue; 2672 other = cpuc->events[idx]; 2673 __icl_update_topdown_event(other, slots, metrics, 2674 event ? event->hw.saved_slots : 0, 2675 event ? event->hw.saved_metric : 0); 2676 } 2677 2678 /* 2679 * Check and update this event, which may have been cleared 2680 * in active_mask e.g. x86_pmu_stop() 2681 */ 2682 if (event && !test_bit(event->hw.idx, cpuc->active_mask)) { 2683 __icl_update_topdown_event(event, slots, metrics, 2684 event->hw.saved_slots, 2685 event->hw.saved_metric); 2686 2687 /* 2688 * In x86_pmu_stop(), the event is cleared in active_mask first, 2689 * then drain the delta, which indicates context switch for 2690 * counting. 2691 * Save metric and slots for context switch. 2692 * Don't need to reset the PERF_METRICS and Fixed counter 3. 2693 * Because the values will be restored in next schedule in. 2694 */ 2695 update_saved_topdown_regs(event, slots, metrics, metric_end); 2696 reset = false; 2697 } 2698 2699 if (reset) { 2700 /* The fixed counter 3 has to be written before the PERF_METRICS. */ 2701 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2702 wrmsrl(MSR_PERF_METRICS, 0); 2703 if (event) 2704 update_saved_topdown_regs(event, 0, 0, metric_end); 2705 } 2706 2707 return slots; 2708 } 2709 2710 static u64 icl_update_topdown_event(struct perf_event *event) 2711 { 2712 return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE + 2713 x86_pmu.num_topdown_events - 1); 2714 } 2715 2716 DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, x86_perf_event_update); 2717 2718 static void intel_pmu_read_topdown_event(struct perf_event *event) 2719 { 2720 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2721 2722 /* Only need to call update_topdown_event() once for group read. */ 2723 if ((cpuc->txn_flags & PERF_PMU_TXN_READ) && 2724 !is_slots_event(event)) 2725 return; 2726 2727 perf_pmu_disable(event->pmu); 2728 static_call(intel_pmu_update_topdown_event)(event); 2729 perf_pmu_enable(event->pmu); 2730 } 2731 2732 static void intel_pmu_read_event(struct perf_event *event) 2733 { 2734 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) 2735 intel_pmu_auto_reload_read(event); 2736 else if (is_topdown_count(event)) 2737 intel_pmu_read_topdown_event(event); 2738 else 2739 x86_perf_event_update(event); 2740 } 2741 2742 static void intel_pmu_enable_fixed(struct perf_event *event) 2743 { 2744 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2745 struct hw_perf_event *hwc = &event->hw; 2746 u64 mask, bits = 0; 2747 int idx = hwc->idx; 2748 2749 if (is_topdown_idx(idx)) { 2750 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2751 /* 2752 * When there are other active TopDown events, 2753 * don't enable the fixed counter 3 again. 2754 */ 2755 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2756 return; 2757 2758 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2759 } 2760 2761 intel_set_masks(event, idx); 2762 2763 /* 2764 * Enable IRQ generation (0x8), if not PEBS, 2765 * and enable ring-3 counting (0x2) and ring-0 counting (0x1) 2766 * if requested: 2767 */ 2768 if (!event->attr.precise_ip) 2769 bits |= INTEL_FIXED_0_ENABLE_PMI; 2770 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) 2771 bits |= INTEL_FIXED_0_USER; 2772 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) 2773 bits |= INTEL_FIXED_0_KERNEL; 2774 2775 /* 2776 * ANY bit is supported in v3 and up 2777 */ 2778 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) 2779 bits |= INTEL_FIXED_0_ANYTHREAD; 2780 2781 idx -= INTEL_PMC_IDX_FIXED; 2782 bits = intel_fixed_bits_by_idx(idx, bits); 2783 mask = intel_fixed_bits_by_idx(idx, INTEL_FIXED_BITS_MASK); 2784 2785 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { 2786 bits |= intel_fixed_bits_by_idx(idx, ICL_FIXED_0_ADAPTIVE); 2787 mask |= intel_fixed_bits_by_idx(idx, ICL_FIXED_0_ADAPTIVE); 2788 } 2789 2790 cpuc->fixed_ctrl_val &= ~mask; 2791 cpuc->fixed_ctrl_val |= bits; 2792 } 2793 2794 static void intel_pmu_enable_event(struct perf_event *event) 2795 { 2796 u64 enable_mask = ARCH_PERFMON_EVENTSEL_ENABLE; 2797 struct hw_perf_event *hwc = &event->hw; 2798 int idx = hwc->idx; 2799 2800 if (unlikely(event->attr.precise_ip)) 2801 intel_pmu_pebs_enable(event); 2802 2803 switch (idx) { 2804 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2805 if (branch_sample_counters(event)) 2806 enable_mask |= ARCH_PERFMON_EVENTSEL_BR_CNTR; 2807 intel_set_masks(event, idx); 2808 __x86_pmu_enable_event(hwc, enable_mask); 2809 break; 2810 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2811 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2812 intel_pmu_enable_fixed(event); 2813 break; 2814 case INTEL_PMC_IDX_FIXED_BTS: 2815 if (!__this_cpu_read(cpu_hw_events.enabled)) 2816 return; 2817 intel_pmu_enable_bts(hwc->config); 2818 break; 2819 case INTEL_PMC_IDX_FIXED_VLBR: 2820 intel_set_masks(event, idx); 2821 break; 2822 default: 2823 pr_warn("Failed to enable the event with invalid index %d\n", 2824 idx); 2825 } 2826 } 2827 2828 static void intel_pmu_add_event(struct perf_event *event) 2829 { 2830 if (event->attr.precise_ip) 2831 intel_pmu_pebs_add(event); 2832 if (intel_pmu_needs_branch_stack(event)) 2833 intel_pmu_lbr_add(event); 2834 } 2835 2836 /* 2837 * Save and restart an expired event. Called by NMI contexts, 2838 * so it has to be careful about preempting normal event ops: 2839 */ 2840 int intel_pmu_save_and_restart(struct perf_event *event) 2841 { 2842 static_call(x86_pmu_update)(event); 2843 /* 2844 * For a checkpointed counter always reset back to 0. This 2845 * avoids a situation where the counter overflows, aborts the 2846 * transaction and is then set back to shortly before the 2847 * overflow, and overflows and aborts again. 2848 */ 2849 if (unlikely(event_is_checkpointed(event))) { 2850 /* No race with NMIs because the counter should not be armed */ 2851 wrmsrl(event->hw.event_base, 0); 2852 local64_set(&event->hw.prev_count, 0); 2853 } 2854 return static_call(x86_pmu_set_period)(event); 2855 } 2856 2857 static int intel_pmu_set_period(struct perf_event *event) 2858 { 2859 if (unlikely(is_topdown_count(event))) 2860 return static_call(intel_pmu_set_topdown_event_period)(event); 2861 2862 return x86_perf_event_set_period(event); 2863 } 2864 2865 static u64 intel_pmu_update(struct perf_event *event) 2866 { 2867 if (unlikely(is_topdown_count(event))) 2868 return static_call(intel_pmu_update_topdown_event)(event); 2869 2870 return x86_perf_event_update(event); 2871 } 2872 2873 static void intel_pmu_reset(void) 2874 { 2875 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); 2876 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2877 int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed); 2878 int num_counters = hybrid(cpuc->pmu, num_counters); 2879 unsigned long flags; 2880 int idx; 2881 2882 if (!num_counters) 2883 return; 2884 2885 local_irq_save(flags); 2886 2887 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); 2888 2889 for (idx = 0; idx < num_counters; idx++) { 2890 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); 2891 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); 2892 } 2893 for (idx = 0; idx < num_counters_fixed; idx++) { 2894 if (fixed_counter_disabled(idx, cpuc->pmu)) 2895 continue; 2896 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); 2897 } 2898 2899 if (ds) 2900 ds->bts_index = ds->bts_buffer_base; 2901 2902 /* Ack all overflows and disable fixed counters */ 2903 if (x86_pmu.version >= 2) { 2904 intel_pmu_ack_status(intel_pmu_get_status()); 2905 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2906 } 2907 2908 /* Reset LBRs and LBR freezing */ 2909 if (x86_pmu.lbr_nr) { 2910 update_debugctlmsr(get_debugctlmsr() & 2911 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); 2912 } 2913 2914 local_irq_restore(flags); 2915 } 2916 2917 /* 2918 * We may be running with guest PEBS events created by KVM, and the 2919 * PEBS records are logged into the guest's DS and invisible to host. 2920 * 2921 * In the case of guest PEBS overflow, we only trigger a fake event 2922 * to emulate the PEBS overflow PMI for guest PEBS counters in KVM. 2923 * The guest will then vm-entry and check the guest DS area to read 2924 * the guest PEBS records. 2925 * 2926 * The contents and other behavior of the guest event do not matter. 2927 */ 2928 static void x86_pmu_handle_guest_pebs(struct pt_regs *regs, 2929 struct perf_sample_data *data) 2930 { 2931 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2932 u64 guest_pebs_idxs = cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask; 2933 struct perf_event *event = NULL; 2934 int bit; 2935 2936 if (!unlikely(perf_guest_state())) 2937 return; 2938 2939 if (!x86_pmu.pebs_ept || !x86_pmu.pebs_active || 2940 !guest_pebs_idxs) 2941 return; 2942 2943 for_each_set_bit(bit, (unsigned long *)&guest_pebs_idxs, 2944 INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed) { 2945 event = cpuc->events[bit]; 2946 if (!event->attr.precise_ip) 2947 continue; 2948 2949 perf_sample_data_init(data, 0, event->hw.last_period); 2950 if (perf_event_overflow(event, data, regs)) 2951 x86_pmu_stop(event, 0); 2952 2953 /* Inject one fake event is enough. */ 2954 break; 2955 } 2956 } 2957 2958 static int handle_pmi_common(struct pt_regs *regs, u64 status) 2959 { 2960 struct perf_sample_data data; 2961 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2962 int bit; 2963 int handled = 0; 2964 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2965 2966 inc_irq_stat(apic_perf_irqs); 2967 2968 /* 2969 * Ignore a range of extra bits in status that do not indicate 2970 * overflow by themselves. 2971 */ 2972 status &= ~(GLOBAL_STATUS_COND_CHG | 2973 GLOBAL_STATUS_ASIF | 2974 GLOBAL_STATUS_LBRS_FROZEN); 2975 if (!status) 2976 return 0; 2977 /* 2978 * In case multiple PEBS events are sampled at the same time, 2979 * it is possible to have GLOBAL_STATUS bit 62 set indicating 2980 * PEBS buffer overflow and also seeing at most 3 PEBS counters 2981 * having their bits set in the status register. This is a sign 2982 * that there was at least one PEBS record pending at the time 2983 * of the PMU interrupt. PEBS counters must only be processed 2984 * via the drain_pebs() calls and not via the regular sample 2985 * processing loop coming after that the function, otherwise 2986 * phony regular samples may be generated in the sampling buffer 2987 * not marked with the EXACT tag. Another possibility is to have 2988 * one PEBS event and at least one non-PEBS event which overflows 2989 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will 2990 * not be set, yet the overflow status bit for the PEBS counter will 2991 * be on Skylake. 2992 * 2993 * To avoid this problem, we systematically ignore the PEBS-enabled 2994 * counters from the GLOBAL_STATUS mask and we always process PEBS 2995 * events via drain_pebs(). 2996 */ 2997 status &= ~(cpuc->pebs_enabled & x86_pmu.pebs_capable); 2998 2999 /* 3000 * PEBS overflow sets bit 62 in the global status register 3001 */ 3002 if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) { 3003 u64 pebs_enabled = cpuc->pebs_enabled; 3004 3005 handled++; 3006 x86_pmu_handle_guest_pebs(regs, &data); 3007 x86_pmu.drain_pebs(regs, &data); 3008 status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI; 3009 3010 /* 3011 * PMI throttle may be triggered, which stops the PEBS event. 3012 * Although cpuc->pebs_enabled is updated accordingly, the 3013 * MSR_IA32_PEBS_ENABLE is not updated. Because the 3014 * cpuc->enabled has been forced to 0 in PMI. 3015 * Update the MSR if pebs_enabled is changed. 3016 */ 3017 if (pebs_enabled != cpuc->pebs_enabled) 3018 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled); 3019 } 3020 3021 /* 3022 * Intel PT 3023 */ 3024 if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) { 3025 handled++; 3026 if (!perf_guest_handle_intel_pt_intr()) 3027 intel_pt_interrupt(); 3028 } 3029 3030 /* 3031 * Intel Perf metrics 3032 */ 3033 if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) { 3034 handled++; 3035 static_call(intel_pmu_update_topdown_event)(NULL); 3036 } 3037 3038 /* 3039 * Checkpointed counters can lead to 'spurious' PMIs because the 3040 * rollback caused by the PMI will have cleared the overflow status 3041 * bit. Therefore always force probe these counters. 3042 */ 3043 status |= cpuc->intel_cp_status; 3044 3045 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { 3046 struct perf_event *event = cpuc->events[bit]; 3047 3048 handled++; 3049 3050 if (!test_bit(bit, cpuc->active_mask)) 3051 continue; 3052 3053 if (!intel_pmu_save_and_restart(event)) 3054 continue; 3055 3056 perf_sample_data_init(&data, 0, event->hw.last_period); 3057 3058 if (has_branch_stack(event)) 3059 intel_pmu_lbr_save_brstack(&data, cpuc, event); 3060 3061 if (perf_event_overflow(event, &data, regs)) 3062 x86_pmu_stop(event, 0); 3063 } 3064 3065 return handled; 3066 } 3067 3068 /* 3069 * This handler is triggered by the local APIC, so the APIC IRQ handling 3070 * rules apply: 3071 */ 3072 static int intel_pmu_handle_irq(struct pt_regs *regs) 3073 { 3074 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3075 bool late_ack = hybrid_bit(cpuc->pmu, late_ack); 3076 bool mid_ack = hybrid_bit(cpuc->pmu, mid_ack); 3077 int loops; 3078 u64 status; 3079 int handled; 3080 int pmu_enabled; 3081 3082 /* 3083 * Save the PMU state. 3084 * It needs to be restored when leaving the handler. 3085 */ 3086 pmu_enabled = cpuc->enabled; 3087 /* 3088 * In general, the early ACK is only applied for old platforms. 3089 * For the big core starts from Haswell, the late ACK should be 3090 * applied. 3091 * For the small core after Tremont, we have to do the ACK right 3092 * before re-enabling counters, which is in the middle of the 3093 * NMI handler. 3094 */ 3095 if (!late_ack && !mid_ack) 3096 apic_write(APIC_LVTPC, APIC_DM_NMI); 3097 intel_bts_disable_local(); 3098 cpuc->enabled = 0; 3099 __intel_pmu_disable_all(true); 3100 handled = intel_pmu_drain_bts_buffer(); 3101 handled += intel_bts_interrupt(); 3102 status = intel_pmu_get_status(); 3103 if (!status) 3104 goto done; 3105 3106 loops = 0; 3107 again: 3108 intel_pmu_lbr_read(); 3109 intel_pmu_ack_status(status); 3110 if (++loops > 100) { 3111 static bool warned; 3112 3113 if (!warned) { 3114 WARN(1, "perfevents: irq loop stuck!\n"); 3115 perf_event_print_debug(); 3116 warned = true; 3117 } 3118 intel_pmu_reset(); 3119 goto done; 3120 } 3121 3122 handled += handle_pmi_common(regs, status); 3123 3124 /* 3125 * Repeat if there is more work to be done: 3126 */ 3127 status = intel_pmu_get_status(); 3128 if (status) 3129 goto again; 3130 3131 done: 3132 if (mid_ack) 3133 apic_write(APIC_LVTPC, APIC_DM_NMI); 3134 /* Only restore PMU state when it's active. See x86_pmu_disable(). */ 3135 cpuc->enabled = pmu_enabled; 3136 if (pmu_enabled) 3137 __intel_pmu_enable_all(0, true); 3138 intel_bts_enable_local(); 3139 3140 /* 3141 * Only unmask the NMI after the overflow counters 3142 * have been reset. This avoids spurious NMIs on 3143 * Haswell CPUs. 3144 */ 3145 if (late_ack) 3146 apic_write(APIC_LVTPC, APIC_DM_NMI); 3147 return handled; 3148 } 3149 3150 static struct event_constraint * 3151 intel_bts_constraints(struct perf_event *event) 3152 { 3153 if (unlikely(intel_pmu_has_bts(event))) 3154 return &bts_constraint; 3155 3156 return NULL; 3157 } 3158 3159 /* 3160 * Note: matches a fake event, like Fixed2. 3161 */ 3162 static struct event_constraint * 3163 intel_vlbr_constraints(struct perf_event *event) 3164 { 3165 struct event_constraint *c = &vlbr_constraint; 3166 3167 if (unlikely(constraint_match(c, event->hw.config))) { 3168 event->hw.flags |= c->flags; 3169 return c; 3170 } 3171 3172 return NULL; 3173 } 3174 3175 static int intel_alt_er(struct cpu_hw_events *cpuc, 3176 int idx, u64 config) 3177 { 3178 struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs); 3179 int alt_idx = idx; 3180 3181 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) 3182 return idx; 3183 3184 if (idx == EXTRA_REG_RSP_0) 3185 alt_idx = EXTRA_REG_RSP_1; 3186 3187 if (idx == EXTRA_REG_RSP_1) 3188 alt_idx = EXTRA_REG_RSP_0; 3189 3190 if (config & ~extra_regs[alt_idx].valid_mask) 3191 return idx; 3192 3193 return alt_idx; 3194 } 3195 3196 static void intel_fixup_er(struct perf_event *event, int idx) 3197 { 3198 struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs); 3199 event->hw.extra_reg.idx = idx; 3200 3201 if (idx == EXTRA_REG_RSP_0) { 3202 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3203 event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event; 3204 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; 3205 } else if (idx == EXTRA_REG_RSP_1) { 3206 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3207 event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event; 3208 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; 3209 } 3210 } 3211 3212 /* 3213 * manage allocation of shared extra msr for certain events 3214 * 3215 * sharing can be: 3216 * per-cpu: to be shared between the various events on a single PMU 3217 * per-core: per-cpu + shared by HT threads 3218 */ 3219 static struct event_constraint * 3220 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, 3221 struct perf_event *event, 3222 struct hw_perf_event_extra *reg) 3223 { 3224 struct event_constraint *c = &emptyconstraint; 3225 struct er_account *era; 3226 unsigned long flags; 3227 int idx = reg->idx; 3228 3229 /* 3230 * reg->alloc can be set due to existing state, so for fake cpuc we 3231 * need to ignore this, otherwise we might fail to allocate proper fake 3232 * state for this extra reg constraint. Also see the comment below. 3233 */ 3234 if (reg->alloc && !cpuc->is_fake) 3235 return NULL; /* call x86_get_event_constraint() */ 3236 3237 again: 3238 era = &cpuc->shared_regs->regs[idx]; 3239 /* 3240 * we use spin_lock_irqsave() to avoid lockdep issues when 3241 * passing a fake cpuc 3242 */ 3243 raw_spin_lock_irqsave(&era->lock, flags); 3244 3245 if (!atomic_read(&era->ref) || era->config == reg->config) { 3246 3247 /* 3248 * If its a fake cpuc -- as per validate_{group,event}() we 3249 * shouldn't touch event state and we can avoid doing so 3250 * since both will only call get_event_constraints() once 3251 * on each event, this avoids the need for reg->alloc. 3252 * 3253 * Not doing the ER fixup will only result in era->reg being 3254 * wrong, but since we won't actually try and program hardware 3255 * this isn't a problem either. 3256 */ 3257 if (!cpuc->is_fake) { 3258 if (idx != reg->idx) 3259 intel_fixup_er(event, idx); 3260 3261 /* 3262 * x86_schedule_events() can call get_event_constraints() 3263 * multiple times on events in the case of incremental 3264 * scheduling(). reg->alloc ensures we only do the ER 3265 * allocation once. 3266 */ 3267 reg->alloc = 1; 3268 } 3269 3270 /* lock in msr value */ 3271 era->config = reg->config; 3272 era->reg = reg->reg; 3273 3274 /* one more user */ 3275 atomic_inc(&era->ref); 3276 3277 /* 3278 * need to call x86_get_event_constraint() 3279 * to check if associated event has constraints 3280 */ 3281 c = NULL; 3282 } else { 3283 idx = intel_alt_er(cpuc, idx, reg->config); 3284 if (idx != reg->idx) { 3285 raw_spin_unlock_irqrestore(&era->lock, flags); 3286 goto again; 3287 } 3288 } 3289 raw_spin_unlock_irqrestore(&era->lock, flags); 3290 3291 return c; 3292 } 3293 3294 static void 3295 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, 3296 struct hw_perf_event_extra *reg) 3297 { 3298 struct er_account *era; 3299 3300 /* 3301 * Only put constraint if extra reg was actually allocated. Also takes 3302 * care of event which do not use an extra shared reg. 3303 * 3304 * Also, if this is a fake cpuc we shouldn't touch any event state 3305 * (reg->alloc) and we don't care about leaving inconsistent cpuc state 3306 * either since it'll be thrown out. 3307 */ 3308 if (!reg->alloc || cpuc->is_fake) 3309 return; 3310 3311 era = &cpuc->shared_regs->regs[reg->idx]; 3312 3313 /* one fewer user */ 3314 atomic_dec(&era->ref); 3315 3316 /* allocate again next time */ 3317 reg->alloc = 0; 3318 } 3319 3320 static struct event_constraint * 3321 intel_shared_regs_constraints(struct cpu_hw_events *cpuc, 3322 struct perf_event *event) 3323 { 3324 struct event_constraint *c = NULL, *d; 3325 struct hw_perf_event_extra *xreg, *breg; 3326 3327 xreg = &event->hw.extra_reg; 3328 if (xreg->idx != EXTRA_REG_NONE) { 3329 c = __intel_shared_reg_get_constraints(cpuc, event, xreg); 3330 if (c == &emptyconstraint) 3331 return c; 3332 } 3333 breg = &event->hw.branch_reg; 3334 if (breg->idx != EXTRA_REG_NONE) { 3335 d = __intel_shared_reg_get_constraints(cpuc, event, breg); 3336 if (d == &emptyconstraint) { 3337 __intel_shared_reg_put_constraints(cpuc, xreg); 3338 c = d; 3339 } 3340 } 3341 return c; 3342 } 3343 3344 struct event_constraint * 3345 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3346 struct perf_event *event) 3347 { 3348 struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints); 3349 struct event_constraint *c; 3350 3351 if (event_constraints) { 3352 for_each_event_constraint(c, event_constraints) { 3353 if (constraint_match(c, event->hw.config)) { 3354 event->hw.flags |= c->flags; 3355 return c; 3356 } 3357 } 3358 } 3359 3360 return &hybrid_var(cpuc->pmu, unconstrained); 3361 } 3362 3363 static struct event_constraint * 3364 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3365 struct perf_event *event) 3366 { 3367 struct event_constraint *c; 3368 3369 c = intel_vlbr_constraints(event); 3370 if (c) 3371 return c; 3372 3373 c = intel_bts_constraints(event); 3374 if (c) 3375 return c; 3376 3377 c = intel_shared_regs_constraints(cpuc, event); 3378 if (c) 3379 return c; 3380 3381 c = intel_pebs_constraints(event); 3382 if (c) 3383 return c; 3384 3385 return x86_get_event_constraints(cpuc, idx, event); 3386 } 3387 3388 static void 3389 intel_start_scheduling(struct cpu_hw_events *cpuc) 3390 { 3391 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3392 struct intel_excl_states *xl; 3393 int tid = cpuc->excl_thread_id; 3394 3395 /* 3396 * nothing needed if in group validation mode 3397 */ 3398 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3399 return; 3400 3401 /* 3402 * no exclusion needed 3403 */ 3404 if (WARN_ON_ONCE(!excl_cntrs)) 3405 return; 3406 3407 xl = &excl_cntrs->states[tid]; 3408 3409 xl->sched_started = true; 3410 /* 3411 * lock shared state until we are done scheduling 3412 * in stop_event_scheduling() 3413 * makes scheduling appear as a transaction 3414 */ 3415 raw_spin_lock(&excl_cntrs->lock); 3416 } 3417 3418 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 3419 { 3420 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3421 struct event_constraint *c = cpuc->event_constraint[idx]; 3422 struct intel_excl_states *xl; 3423 int tid = cpuc->excl_thread_id; 3424 3425 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3426 return; 3427 3428 if (WARN_ON_ONCE(!excl_cntrs)) 3429 return; 3430 3431 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) 3432 return; 3433 3434 xl = &excl_cntrs->states[tid]; 3435 3436 lockdep_assert_held(&excl_cntrs->lock); 3437 3438 if (c->flags & PERF_X86_EVENT_EXCL) 3439 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; 3440 else 3441 xl->state[cntr] = INTEL_EXCL_SHARED; 3442 } 3443 3444 static void 3445 intel_stop_scheduling(struct cpu_hw_events *cpuc) 3446 { 3447 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3448 struct intel_excl_states *xl; 3449 int tid = cpuc->excl_thread_id; 3450 3451 /* 3452 * nothing needed if in group validation mode 3453 */ 3454 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3455 return; 3456 /* 3457 * no exclusion needed 3458 */ 3459 if (WARN_ON_ONCE(!excl_cntrs)) 3460 return; 3461 3462 xl = &excl_cntrs->states[tid]; 3463 3464 xl->sched_started = false; 3465 /* 3466 * release shared state lock (acquired in intel_start_scheduling()) 3467 */ 3468 raw_spin_unlock(&excl_cntrs->lock); 3469 } 3470 3471 static struct event_constraint * 3472 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx) 3473 { 3474 WARN_ON_ONCE(!cpuc->constraint_list); 3475 3476 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { 3477 struct event_constraint *cx; 3478 3479 /* 3480 * grab pre-allocated constraint entry 3481 */ 3482 cx = &cpuc->constraint_list[idx]; 3483 3484 /* 3485 * initialize dynamic constraint 3486 * with static constraint 3487 */ 3488 *cx = *c; 3489 3490 /* 3491 * mark constraint as dynamic 3492 */ 3493 cx->flags |= PERF_X86_EVENT_DYNAMIC; 3494 c = cx; 3495 } 3496 3497 return c; 3498 } 3499 3500 static struct event_constraint * 3501 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, 3502 int idx, struct event_constraint *c) 3503 { 3504 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3505 struct intel_excl_states *xlo; 3506 int tid = cpuc->excl_thread_id; 3507 int is_excl, i, w; 3508 3509 /* 3510 * validating a group does not require 3511 * enforcing cross-thread exclusion 3512 */ 3513 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3514 return c; 3515 3516 /* 3517 * no exclusion needed 3518 */ 3519 if (WARN_ON_ONCE(!excl_cntrs)) 3520 return c; 3521 3522 /* 3523 * because we modify the constraint, we need 3524 * to make a copy. Static constraints come 3525 * from static const tables. 3526 * 3527 * only needed when constraint has not yet 3528 * been cloned (marked dynamic) 3529 */ 3530 c = dyn_constraint(cpuc, c, idx); 3531 3532 /* 3533 * From here on, the constraint is dynamic. 3534 * Either it was just allocated above, or it 3535 * was allocated during a earlier invocation 3536 * of this function 3537 */ 3538 3539 /* 3540 * state of sibling HT 3541 */ 3542 xlo = &excl_cntrs->states[tid ^ 1]; 3543 3544 /* 3545 * event requires exclusive counter access 3546 * across HT threads 3547 */ 3548 is_excl = c->flags & PERF_X86_EVENT_EXCL; 3549 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { 3550 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; 3551 if (!cpuc->n_excl++) 3552 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); 3553 } 3554 3555 /* 3556 * Modify static constraint with current dynamic 3557 * state of thread 3558 * 3559 * EXCLUSIVE: sibling counter measuring exclusive event 3560 * SHARED : sibling counter measuring non-exclusive event 3561 * UNUSED : sibling counter unused 3562 */ 3563 w = c->weight; 3564 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { 3565 /* 3566 * exclusive event in sibling counter 3567 * our corresponding counter cannot be used 3568 * regardless of our event 3569 */ 3570 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) { 3571 __clear_bit(i, c->idxmsk); 3572 w--; 3573 continue; 3574 } 3575 /* 3576 * if measuring an exclusive event, sibling 3577 * measuring non-exclusive, then counter cannot 3578 * be used 3579 */ 3580 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) { 3581 __clear_bit(i, c->idxmsk); 3582 w--; 3583 continue; 3584 } 3585 } 3586 3587 /* 3588 * if we return an empty mask, then switch 3589 * back to static empty constraint to avoid 3590 * the cost of freeing later on 3591 */ 3592 if (!w) 3593 c = &emptyconstraint; 3594 3595 c->weight = w; 3596 3597 return c; 3598 } 3599 3600 static struct event_constraint * 3601 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3602 struct perf_event *event) 3603 { 3604 struct event_constraint *c1, *c2; 3605 3606 c1 = cpuc->event_constraint[idx]; 3607 3608 /* 3609 * first time only 3610 * - static constraint: no change across incremental scheduling calls 3611 * - dynamic constraint: handled by intel_get_excl_constraints() 3612 */ 3613 c2 = __intel_get_event_constraints(cpuc, idx, event); 3614 if (c1) { 3615 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC)); 3616 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); 3617 c1->weight = c2->weight; 3618 c2 = c1; 3619 } 3620 3621 if (cpuc->excl_cntrs) 3622 return intel_get_excl_constraints(cpuc, event, idx, c2); 3623 3624 /* Not all counters support the branch counter feature. */ 3625 if (branch_sample_counters(event)) { 3626 c2 = dyn_constraint(cpuc, c2, idx); 3627 c2->idxmsk64 &= x86_pmu.lbr_counters; 3628 c2->weight = hweight64(c2->idxmsk64); 3629 } 3630 3631 return c2; 3632 } 3633 3634 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, 3635 struct perf_event *event) 3636 { 3637 struct hw_perf_event *hwc = &event->hw; 3638 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3639 int tid = cpuc->excl_thread_id; 3640 struct intel_excl_states *xl; 3641 3642 /* 3643 * nothing needed if in group validation mode 3644 */ 3645 if (cpuc->is_fake) 3646 return; 3647 3648 if (WARN_ON_ONCE(!excl_cntrs)) 3649 return; 3650 3651 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { 3652 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; 3653 if (!--cpuc->n_excl) 3654 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); 3655 } 3656 3657 /* 3658 * If event was actually assigned, then mark the counter state as 3659 * unused now. 3660 */ 3661 if (hwc->idx >= 0) { 3662 xl = &excl_cntrs->states[tid]; 3663 3664 /* 3665 * put_constraint may be called from x86_schedule_events() 3666 * which already has the lock held so here make locking 3667 * conditional. 3668 */ 3669 if (!xl->sched_started) 3670 raw_spin_lock(&excl_cntrs->lock); 3671 3672 xl->state[hwc->idx] = INTEL_EXCL_UNUSED; 3673 3674 if (!xl->sched_started) 3675 raw_spin_unlock(&excl_cntrs->lock); 3676 } 3677 } 3678 3679 static void 3680 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, 3681 struct perf_event *event) 3682 { 3683 struct hw_perf_event_extra *reg; 3684 3685 reg = &event->hw.extra_reg; 3686 if (reg->idx != EXTRA_REG_NONE) 3687 __intel_shared_reg_put_constraints(cpuc, reg); 3688 3689 reg = &event->hw.branch_reg; 3690 if (reg->idx != EXTRA_REG_NONE) 3691 __intel_shared_reg_put_constraints(cpuc, reg); 3692 } 3693 3694 static void intel_put_event_constraints(struct cpu_hw_events *cpuc, 3695 struct perf_event *event) 3696 { 3697 intel_put_shared_regs_event_constraints(cpuc, event); 3698 3699 /* 3700 * is PMU has exclusive counter restrictions, then 3701 * all events are subject to and must call the 3702 * put_excl_constraints() routine 3703 */ 3704 if (cpuc->excl_cntrs) 3705 intel_put_excl_constraints(cpuc, event); 3706 } 3707 3708 static void intel_pebs_aliases_core2(struct perf_event *event) 3709 { 3710 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3711 /* 3712 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3713 * (0x003c) so that we can use it with PEBS. 3714 * 3715 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3716 * PEBS capable. However we can use INST_RETIRED.ANY_P 3717 * (0x00c0), which is a PEBS capable event, to get the same 3718 * count. 3719 * 3720 * INST_RETIRED.ANY_P counts the number of cycles that retires 3721 * CNTMASK instructions. By setting CNTMASK to a value (16) 3722 * larger than the maximum number of instructions that can be 3723 * retired per cycle (4) and then inverting the condition, we 3724 * count all cycles that retire 16 or less instructions, which 3725 * is every cycle. 3726 * 3727 * Thereby we gain a PEBS capable cycle counter. 3728 */ 3729 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); 3730 3731 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3732 event->hw.config = alt_config; 3733 } 3734 } 3735 3736 static void intel_pebs_aliases_snb(struct perf_event *event) 3737 { 3738 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3739 /* 3740 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3741 * (0x003c) so that we can use it with PEBS. 3742 * 3743 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3744 * PEBS capable. However we can use UOPS_RETIRED.ALL 3745 * (0x01c2), which is a PEBS capable event, to get the same 3746 * count. 3747 * 3748 * UOPS_RETIRED.ALL counts the number of cycles that retires 3749 * CNTMASK micro-ops. By setting CNTMASK to a value (16) 3750 * larger than the maximum number of micro-ops that can be 3751 * retired per cycle (4) and then inverting the condition, we 3752 * count all cycles that retire 16 or less micro-ops, which 3753 * is every cycle. 3754 * 3755 * Thereby we gain a PEBS capable cycle counter. 3756 */ 3757 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); 3758 3759 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3760 event->hw.config = alt_config; 3761 } 3762 } 3763 3764 static void intel_pebs_aliases_precdist(struct perf_event *event) 3765 { 3766 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3767 /* 3768 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3769 * (0x003c) so that we can use it with PEBS. 3770 * 3771 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3772 * PEBS capable. However we can use INST_RETIRED.PREC_DIST 3773 * (0x01c0), which is a PEBS capable event, to get the same 3774 * count. 3775 * 3776 * The PREC_DIST event has special support to minimize sample 3777 * shadowing effects. One drawback is that it can be 3778 * only programmed on counter 1, but that seems like an 3779 * acceptable trade off. 3780 */ 3781 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16); 3782 3783 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3784 event->hw.config = alt_config; 3785 } 3786 } 3787 3788 static void intel_pebs_aliases_ivb(struct perf_event *event) 3789 { 3790 if (event->attr.precise_ip < 3) 3791 return intel_pebs_aliases_snb(event); 3792 return intel_pebs_aliases_precdist(event); 3793 } 3794 3795 static void intel_pebs_aliases_skl(struct perf_event *event) 3796 { 3797 if (event->attr.precise_ip < 3) 3798 return intel_pebs_aliases_core2(event); 3799 return intel_pebs_aliases_precdist(event); 3800 } 3801 3802 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event) 3803 { 3804 unsigned long flags = x86_pmu.large_pebs_flags; 3805 3806 if (event->attr.use_clockid) 3807 flags &= ~PERF_SAMPLE_TIME; 3808 if (!event->attr.exclude_kernel) 3809 flags &= ~PERF_SAMPLE_REGS_USER; 3810 if (event->attr.sample_regs_user & ~PEBS_GP_REGS) 3811 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR); 3812 return flags; 3813 } 3814 3815 static int intel_pmu_bts_config(struct perf_event *event) 3816 { 3817 struct perf_event_attr *attr = &event->attr; 3818 3819 if (unlikely(intel_pmu_has_bts(event))) { 3820 /* BTS is not supported by this architecture. */ 3821 if (!x86_pmu.bts_active) 3822 return -EOPNOTSUPP; 3823 3824 /* BTS is currently only allowed for user-mode. */ 3825 if (!attr->exclude_kernel) 3826 return -EOPNOTSUPP; 3827 3828 /* BTS is not allowed for precise events. */ 3829 if (attr->precise_ip) 3830 return -EOPNOTSUPP; 3831 3832 /* disallow bts if conflicting events are present */ 3833 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3834 return -EBUSY; 3835 3836 event->destroy = hw_perf_lbr_event_destroy; 3837 } 3838 3839 return 0; 3840 } 3841 3842 static int core_pmu_hw_config(struct perf_event *event) 3843 { 3844 int ret = x86_pmu_hw_config(event); 3845 3846 if (ret) 3847 return ret; 3848 3849 return intel_pmu_bts_config(event); 3850 } 3851 3852 #define INTEL_TD_METRIC_AVAILABLE_MAX (INTEL_TD_METRIC_RETIRING + \ 3853 ((x86_pmu.num_topdown_events - 1) << 8)) 3854 3855 static bool is_available_metric_event(struct perf_event *event) 3856 { 3857 return is_metric_event(event) && 3858 event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX; 3859 } 3860 3861 static inline bool is_mem_loads_event(struct perf_event *event) 3862 { 3863 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01); 3864 } 3865 3866 static inline bool is_mem_loads_aux_event(struct perf_event *event) 3867 { 3868 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82); 3869 } 3870 3871 static inline bool require_mem_loads_aux_event(struct perf_event *event) 3872 { 3873 if (!(x86_pmu.flags & PMU_FL_MEM_LOADS_AUX)) 3874 return false; 3875 3876 if (is_hybrid()) 3877 return hybrid_pmu(event->pmu)->pmu_type == hybrid_big; 3878 3879 return true; 3880 } 3881 3882 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx) 3883 { 3884 union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap); 3885 3886 return test_bit(idx, (unsigned long *)&intel_cap->capabilities); 3887 } 3888 3889 static int intel_pmu_hw_config(struct perf_event *event) 3890 { 3891 int ret = x86_pmu_hw_config(event); 3892 3893 if (ret) 3894 return ret; 3895 3896 ret = intel_pmu_bts_config(event); 3897 if (ret) 3898 return ret; 3899 3900 if (event->attr.precise_ip) { 3901 if ((event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_FIXED_VLBR_EVENT) 3902 return -EINVAL; 3903 3904 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) { 3905 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; 3906 if (!(event->attr.sample_type & 3907 ~intel_pmu_large_pebs_flags(event))) { 3908 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS; 3909 event->attach_state |= PERF_ATTACH_SCHED_CB; 3910 } 3911 } 3912 if (x86_pmu.pebs_aliases) 3913 x86_pmu.pebs_aliases(event); 3914 } 3915 3916 if (needs_branch_stack(event) && is_sampling_event(event)) 3917 event->hw.flags |= PERF_X86_EVENT_NEEDS_BRANCH_STACK; 3918 3919 if (branch_sample_counters(event)) { 3920 struct perf_event *leader, *sibling; 3921 int num = 0; 3922 3923 if (!(x86_pmu.flags & PMU_FL_BR_CNTR) || 3924 (event->attr.config & ~INTEL_ARCH_EVENT_MASK)) 3925 return -EINVAL; 3926 3927 /* 3928 * The branch counter logging is not supported in the call stack 3929 * mode yet, since we cannot simply flush the LBR during e.g., 3930 * multiplexing. Also, there is no obvious usage with the call 3931 * stack mode. Simply forbids it for now. 3932 * 3933 * If any events in the group enable the branch counter logging 3934 * feature, the group is treated as a branch counter logging 3935 * group, which requires the extra space to store the counters. 3936 */ 3937 leader = event->group_leader; 3938 if (branch_sample_call_stack(leader)) 3939 return -EINVAL; 3940 if (branch_sample_counters(leader)) 3941 num++; 3942 leader->hw.flags |= PERF_X86_EVENT_BRANCH_COUNTERS; 3943 3944 for_each_sibling_event(sibling, leader) { 3945 if (branch_sample_call_stack(sibling)) 3946 return -EINVAL; 3947 if (branch_sample_counters(sibling)) 3948 num++; 3949 } 3950 3951 if (num > fls(x86_pmu.lbr_counters)) 3952 return -EINVAL; 3953 /* 3954 * Only applying the PERF_SAMPLE_BRANCH_COUNTERS doesn't 3955 * require any branch stack setup. 3956 * Clear the bit to avoid unnecessary branch stack setup. 3957 */ 3958 if (0 == (event->attr.branch_sample_type & 3959 ~(PERF_SAMPLE_BRANCH_PLM_ALL | 3960 PERF_SAMPLE_BRANCH_COUNTERS))) 3961 event->hw.flags &= ~PERF_X86_EVENT_NEEDS_BRANCH_STACK; 3962 3963 /* 3964 * Force the leader to be a LBR event. So LBRs can be reset 3965 * with the leader event. See intel_pmu_lbr_del() for details. 3966 */ 3967 if (!intel_pmu_needs_branch_stack(leader)) 3968 return -EINVAL; 3969 } 3970 3971 if (intel_pmu_needs_branch_stack(event)) { 3972 ret = intel_pmu_setup_lbr_filter(event); 3973 if (ret) 3974 return ret; 3975 event->attach_state |= PERF_ATTACH_SCHED_CB; 3976 3977 /* 3978 * BTS is set up earlier in this path, so don't account twice 3979 */ 3980 if (!unlikely(intel_pmu_has_bts(event))) { 3981 /* disallow lbr if conflicting events are present */ 3982 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3983 return -EBUSY; 3984 3985 event->destroy = hw_perf_lbr_event_destroy; 3986 } 3987 } 3988 3989 if (event->attr.aux_output) { 3990 if (!event->attr.precise_ip) 3991 return -EINVAL; 3992 3993 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT; 3994 } 3995 3996 if ((event->attr.type == PERF_TYPE_HARDWARE) || 3997 (event->attr.type == PERF_TYPE_HW_CACHE)) 3998 return 0; 3999 4000 /* 4001 * Config Topdown slots and metric events 4002 * 4003 * The slots event on Fixed Counter 3 can support sampling, 4004 * which will be handled normally in x86_perf_event_update(). 4005 * 4006 * Metric events don't support sampling and require being paired 4007 * with a slots event as group leader. When the slots event 4008 * is used in a metrics group, it too cannot support sampling. 4009 */ 4010 if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) { 4011 if (event->attr.config1 || event->attr.config2) 4012 return -EINVAL; 4013 4014 /* 4015 * The TopDown metrics events and slots event don't 4016 * support any filters. 4017 */ 4018 if (event->attr.config & X86_ALL_EVENT_FLAGS) 4019 return -EINVAL; 4020 4021 if (is_available_metric_event(event)) { 4022 struct perf_event *leader = event->group_leader; 4023 4024 /* The metric events don't support sampling. */ 4025 if (is_sampling_event(event)) 4026 return -EINVAL; 4027 4028 /* The metric events require a slots group leader. */ 4029 if (!is_slots_event(leader)) 4030 return -EINVAL; 4031 4032 /* 4033 * The leader/SLOTS must not be a sampling event for 4034 * metric use; hardware requires it starts at 0 when used 4035 * in conjunction with MSR_PERF_METRICS. 4036 */ 4037 if (is_sampling_event(leader)) 4038 return -EINVAL; 4039 4040 event->event_caps |= PERF_EV_CAP_SIBLING; 4041 /* 4042 * Only once we have a METRICs sibling do we 4043 * need TopDown magic. 4044 */ 4045 leader->hw.flags |= PERF_X86_EVENT_TOPDOWN; 4046 event->hw.flags |= PERF_X86_EVENT_TOPDOWN; 4047 } 4048 } 4049 4050 /* 4051 * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR 4052 * doesn't function quite right. As a work-around it needs to always be 4053 * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82). 4054 * The actual count of this second event is irrelevant it just needs 4055 * to be active to make the first event function correctly. 4056 * 4057 * In a group, the auxiliary event must be in front of the load latency 4058 * event. The rule is to simplify the implementation of the check. 4059 * That's because perf cannot have a complete group at the moment. 4060 */ 4061 if (require_mem_loads_aux_event(event) && 4062 (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) && 4063 is_mem_loads_event(event)) { 4064 struct perf_event *leader = event->group_leader; 4065 struct perf_event *sibling = NULL; 4066 4067 /* 4068 * When this memload event is also the first event (no group 4069 * exists yet), then there is no aux event before it. 4070 */ 4071 if (leader == event) 4072 return -ENODATA; 4073 4074 if (!is_mem_loads_aux_event(leader)) { 4075 for_each_sibling_event(sibling, leader) { 4076 if (is_mem_loads_aux_event(sibling)) 4077 break; 4078 } 4079 if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list)) 4080 return -ENODATA; 4081 } 4082 } 4083 4084 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) 4085 return 0; 4086 4087 if (x86_pmu.version < 3) 4088 return -EINVAL; 4089 4090 ret = perf_allow_cpu(&event->attr); 4091 if (ret) 4092 return ret; 4093 4094 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; 4095 4096 return 0; 4097 } 4098 4099 /* 4100 * Currently, the only caller of this function is the atomic_switch_perf_msrs(). 4101 * The host perf context helps to prepare the values of the real hardware for 4102 * a set of msrs that need to be switched atomically in a vmx transaction. 4103 * 4104 * For example, the pseudocode needed to add a new msr should look like: 4105 * 4106 * arr[(*nr)++] = (struct perf_guest_switch_msr){ 4107 * .msr = the hardware msr address, 4108 * .host = the value the hardware has when it doesn't run a guest, 4109 * .guest = the value the hardware has when it runs a guest, 4110 * }; 4111 * 4112 * These values have nothing to do with the emulated values the guest sees 4113 * when it uses {RD,WR}MSR, which should be handled by the KVM context, 4114 * specifically in the intel_pmu_{get,set}_msr(). 4115 */ 4116 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr, void *data) 4117 { 4118 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4119 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 4120 struct kvm_pmu *kvm_pmu = (struct kvm_pmu *)data; 4121 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 4122 u64 pebs_mask = cpuc->pebs_enabled & x86_pmu.pebs_capable; 4123 int global_ctrl, pebs_enable; 4124 4125 /* 4126 * In addition to obeying exclude_guest/exclude_host, remove bits being 4127 * used for PEBS when running a guest, because PEBS writes to virtual 4128 * addresses (not physical addresses). 4129 */ 4130 *nr = 0; 4131 global_ctrl = (*nr)++; 4132 arr[global_ctrl] = (struct perf_guest_switch_msr){ 4133 .msr = MSR_CORE_PERF_GLOBAL_CTRL, 4134 .host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask, 4135 .guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask & ~pebs_mask, 4136 }; 4137 4138 if (!x86_pmu.pebs) 4139 return arr; 4140 4141 /* 4142 * If PMU counter has PEBS enabled it is not enough to 4143 * disable counter on a guest entry since PEBS memory 4144 * write can overshoot guest entry and corrupt guest 4145 * memory. Disabling PEBS solves the problem. 4146 * 4147 * Don't do this if the CPU already enforces it. 4148 */ 4149 if (x86_pmu.pebs_no_isolation) { 4150 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4151 .msr = MSR_IA32_PEBS_ENABLE, 4152 .host = cpuc->pebs_enabled, 4153 .guest = 0, 4154 }; 4155 return arr; 4156 } 4157 4158 if (!kvm_pmu || !x86_pmu.pebs_ept) 4159 return arr; 4160 4161 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4162 .msr = MSR_IA32_DS_AREA, 4163 .host = (unsigned long)cpuc->ds, 4164 .guest = kvm_pmu->ds_area, 4165 }; 4166 4167 if (x86_pmu.intel_cap.pebs_baseline) { 4168 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4169 .msr = MSR_PEBS_DATA_CFG, 4170 .host = cpuc->active_pebs_data_cfg, 4171 .guest = kvm_pmu->pebs_data_cfg, 4172 }; 4173 } 4174 4175 pebs_enable = (*nr)++; 4176 arr[pebs_enable] = (struct perf_guest_switch_msr){ 4177 .msr = MSR_IA32_PEBS_ENABLE, 4178 .host = cpuc->pebs_enabled & ~cpuc->intel_ctrl_guest_mask, 4179 .guest = pebs_mask & ~cpuc->intel_ctrl_host_mask, 4180 }; 4181 4182 if (arr[pebs_enable].host) { 4183 /* Disable guest PEBS if host PEBS is enabled. */ 4184 arr[pebs_enable].guest = 0; 4185 } else { 4186 /* Disable guest PEBS thoroughly for cross-mapped PEBS counters. */ 4187 arr[pebs_enable].guest &= ~kvm_pmu->host_cross_mapped_mask; 4188 arr[global_ctrl].guest &= ~kvm_pmu->host_cross_mapped_mask; 4189 /* Set hw GLOBAL_CTRL bits for PEBS counter when it runs for guest */ 4190 arr[global_ctrl].guest |= arr[pebs_enable].guest; 4191 } 4192 4193 return arr; 4194 } 4195 4196 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr, void *data) 4197 { 4198 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4199 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 4200 int idx; 4201 4202 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 4203 struct perf_event *event = cpuc->events[idx]; 4204 4205 arr[idx].msr = x86_pmu_config_addr(idx); 4206 arr[idx].host = arr[idx].guest = 0; 4207 4208 if (!test_bit(idx, cpuc->active_mask)) 4209 continue; 4210 4211 arr[idx].host = arr[idx].guest = 4212 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; 4213 4214 if (event->attr.exclude_host) 4215 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 4216 else if (event->attr.exclude_guest) 4217 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 4218 } 4219 4220 *nr = x86_pmu.num_counters; 4221 return arr; 4222 } 4223 4224 static void core_pmu_enable_event(struct perf_event *event) 4225 { 4226 if (!event->attr.exclude_host) 4227 x86_pmu_enable_event(event); 4228 } 4229 4230 static void core_pmu_enable_all(int added) 4231 { 4232 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4233 int idx; 4234 4235 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 4236 struct hw_perf_event *hwc = &cpuc->events[idx]->hw; 4237 4238 if (!test_bit(idx, cpuc->active_mask) || 4239 cpuc->events[idx]->attr.exclude_host) 4240 continue; 4241 4242 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 4243 } 4244 } 4245 4246 static int hsw_hw_config(struct perf_event *event) 4247 { 4248 int ret = intel_pmu_hw_config(event); 4249 4250 if (ret) 4251 return ret; 4252 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) 4253 return 0; 4254 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); 4255 4256 /* 4257 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with 4258 * PEBS or in ANY thread mode. Since the results are non-sensical forbid 4259 * this combination. 4260 */ 4261 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && 4262 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || 4263 event->attr.precise_ip > 0)) 4264 return -EOPNOTSUPP; 4265 4266 if (event_is_checkpointed(event)) { 4267 /* 4268 * Sampling of checkpointed events can cause situations where 4269 * the CPU constantly aborts because of a overflow, which is 4270 * then checkpointed back and ignored. Forbid checkpointing 4271 * for sampling. 4272 * 4273 * But still allow a long sampling period, so that perf stat 4274 * from KVM works. 4275 */ 4276 if (event->attr.sample_period > 0 && 4277 event->attr.sample_period < 0x7fffffff) 4278 return -EOPNOTSUPP; 4279 } 4280 return 0; 4281 } 4282 4283 static struct event_constraint counter0_constraint = 4284 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1); 4285 4286 static struct event_constraint counter1_constraint = 4287 INTEL_ALL_EVENT_CONSTRAINT(0, 0x2); 4288 4289 static struct event_constraint counter0_1_constraint = 4290 INTEL_ALL_EVENT_CONSTRAINT(0, 0x3); 4291 4292 static struct event_constraint counter2_constraint = 4293 EVENT_CONSTRAINT(0, 0x4, 0); 4294 4295 static struct event_constraint fixed0_constraint = 4296 FIXED_EVENT_CONSTRAINT(0x00c0, 0); 4297 4298 static struct event_constraint fixed0_counter0_constraint = 4299 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL); 4300 4301 static struct event_constraint fixed0_counter0_1_constraint = 4302 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000003ULL); 4303 4304 static struct event_constraint counters_1_7_constraint = 4305 INTEL_ALL_EVENT_CONSTRAINT(0, 0xfeULL); 4306 4307 static struct event_constraint * 4308 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4309 struct perf_event *event) 4310 { 4311 struct event_constraint *c; 4312 4313 c = intel_get_event_constraints(cpuc, idx, event); 4314 4315 /* Handle special quirk on in_tx_checkpointed only in counter 2 */ 4316 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { 4317 if (c->idxmsk64 & (1U << 2)) 4318 return &counter2_constraint; 4319 return &emptyconstraint; 4320 } 4321 4322 return c; 4323 } 4324 4325 static struct event_constraint * 4326 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4327 struct perf_event *event) 4328 { 4329 /* 4330 * Fixed counter 0 has less skid. 4331 * Force instruction:ppp in Fixed counter 0 4332 */ 4333 if ((event->attr.precise_ip == 3) && 4334 constraint_match(&fixed0_constraint, event->hw.config)) 4335 return &fixed0_constraint; 4336 4337 return hsw_get_event_constraints(cpuc, idx, event); 4338 } 4339 4340 static struct event_constraint * 4341 glc_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4342 struct perf_event *event) 4343 { 4344 struct event_constraint *c; 4345 4346 c = icl_get_event_constraints(cpuc, idx, event); 4347 4348 /* 4349 * The :ppp indicates the Precise Distribution (PDist) facility, which 4350 * is only supported on the GP counter 0. If a :ppp event which is not 4351 * available on the GP counter 0, error out. 4352 * Exception: Instruction PDIR is only available on the fixed counter 0. 4353 */ 4354 if ((event->attr.precise_ip == 3) && 4355 !constraint_match(&fixed0_constraint, event->hw.config)) { 4356 if (c->idxmsk64 & BIT_ULL(0)) 4357 return &counter0_constraint; 4358 4359 return &emptyconstraint; 4360 } 4361 4362 return c; 4363 } 4364 4365 static struct event_constraint * 4366 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4367 struct perf_event *event) 4368 { 4369 struct event_constraint *c; 4370 4371 /* :ppp means to do reduced skid PEBS which is PMC0 only. */ 4372 if (event->attr.precise_ip == 3) 4373 return &counter0_constraint; 4374 4375 c = intel_get_event_constraints(cpuc, idx, event); 4376 4377 return c; 4378 } 4379 4380 static struct event_constraint * 4381 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4382 struct perf_event *event) 4383 { 4384 struct event_constraint *c; 4385 4386 c = intel_get_event_constraints(cpuc, idx, event); 4387 4388 /* 4389 * :ppp means to do reduced skid PEBS, 4390 * which is available on PMC0 and fixed counter 0. 4391 */ 4392 if (event->attr.precise_ip == 3) { 4393 /* Force instruction:ppp on PMC0 and Fixed counter 0 */ 4394 if (constraint_match(&fixed0_constraint, event->hw.config)) 4395 return &fixed0_counter0_constraint; 4396 4397 return &counter0_constraint; 4398 } 4399 4400 return c; 4401 } 4402 4403 static bool allow_tsx_force_abort = true; 4404 4405 static struct event_constraint * 4406 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4407 struct perf_event *event) 4408 { 4409 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event); 4410 4411 /* 4412 * Without TFA we must not use PMC3. 4413 */ 4414 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) { 4415 c = dyn_constraint(cpuc, c, idx); 4416 c->idxmsk64 &= ~(1ULL << 3); 4417 c->weight--; 4418 } 4419 4420 return c; 4421 } 4422 4423 static struct event_constraint * 4424 adl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4425 struct perf_event *event) 4426 { 4427 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4428 4429 if (pmu->pmu_type == hybrid_big) 4430 return glc_get_event_constraints(cpuc, idx, event); 4431 else if (pmu->pmu_type == hybrid_small) 4432 return tnt_get_event_constraints(cpuc, idx, event); 4433 4434 WARN_ON(1); 4435 return &emptyconstraint; 4436 } 4437 4438 static struct event_constraint * 4439 cmt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4440 struct perf_event *event) 4441 { 4442 struct event_constraint *c; 4443 4444 c = intel_get_event_constraints(cpuc, idx, event); 4445 4446 /* 4447 * The :ppp indicates the Precise Distribution (PDist) facility, which 4448 * is only supported on the GP counter 0 & 1 and Fixed counter 0. 4449 * If a :ppp event which is not available on the above eligible counters, 4450 * error out. 4451 */ 4452 if (event->attr.precise_ip == 3) { 4453 /* Force instruction:ppp on PMC0, 1 and Fixed counter 0 */ 4454 if (constraint_match(&fixed0_constraint, event->hw.config)) { 4455 /* The fixed counter 0 doesn't support LBR event logging. */ 4456 if (branch_sample_counters(event)) 4457 return &counter0_1_constraint; 4458 else 4459 return &fixed0_counter0_1_constraint; 4460 } 4461 4462 switch (c->idxmsk64 & 0x3ull) { 4463 case 0x1: 4464 return &counter0_constraint; 4465 case 0x2: 4466 return &counter1_constraint; 4467 case 0x3: 4468 return &counter0_1_constraint; 4469 } 4470 return &emptyconstraint; 4471 } 4472 4473 return c; 4474 } 4475 4476 static struct event_constraint * 4477 rwc_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4478 struct perf_event *event) 4479 { 4480 struct event_constraint *c; 4481 4482 c = glc_get_event_constraints(cpuc, idx, event); 4483 4484 /* The Retire Latency is not supported by the fixed counter 0. */ 4485 if (event->attr.precise_ip && 4486 (event->attr.sample_type & PERF_SAMPLE_WEIGHT_TYPE) && 4487 constraint_match(&fixed0_constraint, event->hw.config)) { 4488 /* 4489 * The Instruction PDIR is only available 4490 * on the fixed counter 0. Error out for this case. 4491 */ 4492 if (event->attr.precise_ip == 3) 4493 return &emptyconstraint; 4494 return &counters_1_7_constraint; 4495 } 4496 4497 return c; 4498 } 4499 4500 static struct event_constraint * 4501 mtl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4502 struct perf_event *event) 4503 { 4504 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4505 4506 if (pmu->pmu_type == hybrid_big) 4507 return rwc_get_event_constraints(cpuc, idx, event); 4508 if (pmu->pmu_type == hybrid_small) 4509 return cmt_get_event_constraints(cpuc, idx, event); 4510 4511 WARN_ON(1); 4512 return &emptyconstraint; 4513 } 4514 4515 static int adl_hw_config(struct perf_event *event) 4516 { 4517 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4518 4519 if (pmu->pmu_type == hybrid_big) 4520 return hsw_hw_config(event); 4521 else if (pmu->pmu_type == hybrid_small) 4522 return intel_pmu_hw_config(event); 4523 4524 WARN_ON(1); 4525 return -EOPNOTSUPP; 4526 } 4527 4528 static enum hybrid_cpu_type adl_get_hybrid_cpu_type(void) 4529 { 4530 return HYBRID_INTEL_CORE; 4531 } 4532 4533 /* 4534 * Broadwell: 4535 * 4536 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared 4537 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine 4538 * the two to enforce a minimum period of 128 (the smallest value that has bits 4539 * 0-5 cleared and >= 100). 4540 * 4541 * Because of how the code in x86_perf_event_set_period() works, the truncation 4542 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period 4543 * to make up for the 'lost' events due to carrying the 'error' in period_left. 4544 * 4545 * Therefore the effective (average) period matches the requested period, 4546 * despite coarser hardware granularity. 4547 */ 4548 static void bdw_limit_period(struct perf_event *event, s64 *left) 4549 { 4550 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == 4551 X86_CONFIG(.event=0xc0, .umask=0x01)) { 4552 if (*left < 128) 4553 *left = 128; 4554 *left &= ~0x3fULL; 4555 } 4556 } 4557 4558 static void nhm_limit_period(struct perf_event *event, s64 *left) 4559 { 4560 *left = max(*left, 32LL); 4561 } 4562 4563 static void glc_limit_period(struct perf_event *event, s64 *left) 4564 { 4565 if (event->attr.precise_ip == 3) 4566 *left = max(*left, 128LL); 4567 } 4568 4569 PMU_FORMAT_ATTR(event, "config:0-7" ); 4570 PMU_FORMAT_ATTR(umask, "config:8-15" ); 4571 PMU_FORMAT_ATTR(edge, "config:18" ); 4572 PMU_FORMAT_ATTR(pc, "config:19" ); 4573 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ 4574 PMU_FORMAT_ATTR(inv, "config:23" ); 4575 PMU_FORMAT_ATTR(cmask, "config:24-31" ); 4576 PMU_FORMAT_ATTR(in_tx, "config:32"); 4577 PMU_FORMAT_ATTR(in_tx_cp, "config:33"); 4578 4579 static struct attribute *intel_arch_formats_attr[] = { 4580 &format_attr_event.attr, 4581 &format_attr_umask.attr, 4582 &format_attr_edge.attr, 4583 &format_attr_pc.attr, 4584 &format_attr_inv.attr, 4585 &format_attr_cmask.attr, 4586 NULL, 4587 }; 4588 4589 ssize_t intel_event_sysfs_show(char *page, u64 config) 4590 { 4591 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); 4592 4593 return x86_event_sysfs_show(page, config, event); 4594 } 4595 4596 static struct intel_shared_regs *allocate_shared_regs(int cpu) 4597 { 4598 struct intel_shared_regs *regs; 4599 int i; 4600 4601 regs = kzalloc_node(sizeof(struct intel_shared_regs), 4602 GFP_KERNEL, cpu_to_node(cpu)); 4603 if (regs) { 4604 /* 4605 * initialize the locks to keep lockdep happy 4606 */ 4607 for (i = 0; i < EXTRA_REG_MAX; i++) 4608 raw_spin_lock_init(®s->regs[i].lock); 4609 4610 regs->core_id = -1; 4611 } 4612 return regs; 4613 } 4614 4615 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) 4616 { 4617 struct intel_excl_cntrs *c; 4618 4619 c = kzalloc_node(sizeof(struct intel_excl_cntrs), 4620 GFP_KERNEL, cpu_to_node(cpu)); 4621 if (c) { 4622 raw_spin_lock_init(&c->lock); 4623 c->core_id = -1; 4624 } 4625 return c; 4626 } 4627 4628 4629 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 4630 { 4631 cpuc->pebs_record_size = x86_pmu.pebs_record_size; 4632 4633 if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { 4634 cpuc->shared_regs = allocate_shared_regs(cpu); 4635 if (!cpuc->shared_regs) 4636 goto err; 4637 } 4638 4639 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA | PMU_FL_BR_CNTR)) { 4640 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); 4641 4642 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu)); 4643 if (!cpuc->constraint_list) 4644 goto err_shared_regs; 4645 } 4646 4647 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4648 cpuc->excl_cntrs = allocate_excl_cntrs(cpu); 4649 if (!cpuc->excl_cntrs) 4650 goto err_constraint_list; 4651 4652 cpuc->excl_thread_id = 0; 4653 } 4654 4655 return 0; 4656 4657 err_constraint_list: 4658 kfree(cpuc->constraint_list); 4659 cpuc->constraint_list = NULL; 4660 4661 err_shared_regs: 4662 kfree(cpuc->shared_regs); 4663 cpuc->shared_regs = NULL; 4664 4665 err: 4666 return -ENOMEM; 4667 } 4668 4669 static int intel_pmu_cpu_prepare(int cpu) 4670 { 4671 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu); 4672 } 4673 4674 static void flip_smm_bit(void *data) 4675 { 4676 unsigned long set = *(unsigned long *)data; 4677 4678 if (set > 0) { 4679 msr_set_bit(MSR_IA32_DEBUGCTLMSR, 4680 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4681 } else { 4682 msr_clear_bit(MSR_IA32_DEBUGCTLMSR, 4683 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4684 } 4685 } 4686 4687 static void intel_pmu_check_num_counters(int *num_counters, 4688 int *num_counters_fixed, 4689 u64 *intel_ctrl, u64 fixed_mask); 4690 4691 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints, 4692 int num_counters, 4693 int num_counters_fixed, 4694 u64 intel_ctrl); 4695 4696 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs); 4697 4698 static inline bool intel_pmu_broken_perf_cap(void) 4699 { 4700 /* The Perf Metric (Bit 15) is always cleared */ 4701 if ((boot_cpu_data.x86_model == INTEL_FAM6_METEORLAKE) || 4702 (boot_cpu_data.x86_model == INTEL_FAM6_METEORLAKE_L)) 4703 return true; 4704 4705 return false; 4706 } 4707 4708 static void update_pmu_cap(struct x86_hybrid_pmu *pmu) 4709 { 4710 unsigned int sub_bitmaps = cpuid_eax(ARCH_PERFMON_EXT_LEAF); 4711 unsigned int eax, ebx, ecx, edx; 4712 4713 if (sub_bitmaps & ARCH_PERFMON_NUM_COUNTER_LEAF_BIT) { 4714 cpuid_count(ARCH_PERFMON_EXT_LEAF, ARCH_PERFMON_NUM_COUNTER_LEAF, 4715 &eax, &ebx, &ecx, &edx); 4716 pmu->num_counters = fls(eax); 4717 pmu->num_counters_fixed = fls(ebx); 4718 } 4719 4720 4721 if (!intel_pmu_broken_perf_cap()) { 4722 /* Perf Metric (Bit 15) and PEBS via PT (Bit 16) are hybrid enumeration */ 4723 rdmsrl(MSR_IA32_PERF_CAPABILITIES, pmu->intel_cap.capabilities); 4724 } 4725 } 4726 4727 static void intel_pmu_check_hybrid_pmus(struct x86_hybrid_pmu *pmu) 4728 { 4729 intel_pmu_check_num_counters(&pmu->num_counters, &pmu->num_counters_fixed, 4730 &pmu->intel_ctrl, (1ULL << pmu->num_counters_fixed) - 1); 4731 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters); 4732 pmu->unconstrained = (struct event_constraint) 4733 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 4734 0, pmu->num_counters, 0, 0); 4735 4736 if (pmu->intel_cap.perf_metrics) 4737 pmu->intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 4738 else 4739 pmu->intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS); 4740 4741 if (pmu->intel_cap.pebs_output_pt_available) 4742 pmu->pmu.capabilities |= PERF_PMU_CAP_AUX_OUTPUT; 4743 else 4744 pmu->pmu.capabilities &= ~PERF_PMU_CAP_AUX_OUTPUT; 4745 4746 intel_pmu_check_event_constraints(pmu->event_constraints, 4747 pmu->num_counters, 4748 pmu->num_counters_fixed, 4749 pmu->intel_ctrl); 4750 4751 intel_pmu_check_extra_regs(pmu->extra_regs); 4752 } 4753 4754 static struct x86_hybrid_pmu *find_hybrid_pmu_for_cpu(void) 4755 { 4756 u8 cpu_type = get_this_hybrid_cpu_type(); 4757 int i; 4758 4759 /* 4760 * This is running on a CPU model that is known to have hybrid 4761 * configurations. But the CPU told us it is not hybrid, shame 4762 * on it. There should be a fixup function provided for these 4763 * troublesome CPUs (->get_hybrid_cpu_type). 4764 */ 4765 if (cpu_type == HYBRID_INTEL_NONE) { 4766 if (x86_pmu.get_hybrid_cpu_type) 4767 cpu_type = x86_pmu.get_hybrid_cpu_type(); 4768 else 4769 return NULL; 4770 } 4771 4772 /* 4773 * This essentially just maps between the 'hybrid_cpu_type' 4774 * and 'hybrid_pmu_type' enums: 4775 */ 4776 for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) { 4777 enum hybrid_pmu_type pmu_type = x86_pmu.hybrid_pmu[i].pmu_type; 4778 4779 if (cpu_type == HYBRID_INTEL_CORE && 4780 pmu_type == hybrid_big) 4781 return &x86_pmu.hybrid_pmu[i]; 4782 if (cpu_type == HYBRID_INTEL_ATOM && 4783 pmu_type == hybrid_small) 4784 return &x86_pmu.hybrid_pmu[i]; 4785 } 4786 4787 return NULL; 4788 } 4789 4790 static bool init_hybrid_pmu(int cpu) 4791 { 4792 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4793 struct x86_hybrid_pmu *pmu = find_hybrid_pmu_for_cpu(); 4794 4795 if (WARN_ON_ONCE(!pmu || (pmu->pmu.type == -1))) { 4796 cpuc->pmu = NULL; 4797 return false; 4798 } 4799 4800 /* Only check and dump the PMU information for the first CPU */ 4801 if (!cpumask_empty(&pmu->supported_cpus)) 4802 goto end; 4803 4804 if (this_cpu_has(X86_FEATURE_ARCH_PERFMON_EXT)) 4805 update_pmu_cap(pmu); 4806 4807 intel_pmu_check_hybrid_pmus(pmu); 4808 4809 if (!check_hw_exists(&pmu->pmu, pmu->num_counters, pmu->num_counters_fixed)) 4810 return false; 4811 4812 pr_info("%s PMU driver: ", pmu->name); 4813 4814 if (pmu->intel_cap.pebs_output_pt_available) 4815 pr_cont("PEBS-via-PT "); 4816 4817 pr_cont("\n"); 4818 4819 x86_pmu_show_pmu_cap(pmu->num_counters, pmu->num_counters_fixed, 4820 pmu->intel_ctrl); 4821 4822 end: 4823 cpumask_set_cpu(cpu, &pmu->supported_cpus); 4824 cpuc->pmu = &pmu->pmu; 4825 4826 return true; 4827 } 4828 4829 static void intel_pmu_cpu_starting(int cpu) 4830 { 4831 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4832 int core_id = topology_core_id(cpu); 4833 int i; 4834 4835 if (is_hybrid() && !init_hybrid_pmu(cpu)) 4836 return; 4837 4838 init_debug_store_on_cpu(cpu); 4839 /* 4840 * Deal with CPUs that don't clear their LBRs on power-up. 4841 */ 4842 intel_pmu_lbr_reset(); 4843 4844 cpuc->lbr_sel = NULL; 4845 4846 if (x86_pmu.flags & PMU_FL_TFA) { 4847 WARN_ON_ONCE(cpuc->tfa_shadow); 4848 cpuc->tfa_shadow = ~0ULL; 4849 intel_set_tfa(cpuc, false); 4850 } 4851 4852 if (x86_pmu.version > 1) 4853 flip_smm_bit(&x86_pmu.attr_freeze_on_smi); 4854 4855 /* 4856 * Disable perf metrics if any added CPU doesn't support it. 4857 * 4858 * Turn off the check for a hybrid architecture, because the 4859 * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate 4860 * the architecture features. The perf metrics is a model-specific 4861 * feature for now. The corresponding bit should always be 0 on 4862 * a hybrid platform, e.g., Alder Lake. 4863 */ 4864 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) { 4865 union perf_capabilities perf_cap; 4866 4867 rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities); 4868 if (!perf_cap.perf_metrics) { 4869 x86_pmu.intel_cap.perf_metrics = 0; 4870 x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS); 4871 } 4872 } 4873 4874 if (!cpuc->shared_regs) 4875 return; 4876 4877 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { 4878 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4879 struct intel_shared_regs *pc; 4880 4881 pc = per_cpu(cpu_hw_events, i).shared_regs; 4882 if (pc && pc->core_id == core_id) { 4883 cpuc->kfree_on_online[0] = cpuc->shared_regs; 4884 cpuc->shared_regs = pc; 4885 break; 4886 } 4887 } 4888 cpuc->shared_regs->core_id = core_id; 4889 cpuc->shared_regs->refcnt++; 4890 } 4891 4892 if (x86_pmu.lbr_sel_map) 4893 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; 4894 4895 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4896 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4897 struct cpu_hw_events *sibling; 4898 struct intel_excl_cntrs *c; 4899 4900 sibling = &per_cpu(cpu_hw_events, i); 4901 c = sibling->excl_cntrs; 4902 if (c && c->core_id == core_id) { 4903 cpuc->kfree_on_online[1] = cpuc->excl_cntrs; 4904 cpuc->excl_cntrs = c; 4905 if (!sibling->excl_thread_id) 4906 cpuc->excl_thread_id = 1; 4907 break; 4908 } 4909 } 4910 cpuc->excl_cntrs->core_id = core_id; 4911 cpuc->excl_cntrs->refcnt++; 4912 } 4913 } 4914 4915 static void free_excl_cntrs(struct cpu_hw_events *cpuc) 4916 { 4917 struct intel_excl_cntrs *c; 4918 4919 c = cpuc->excl_cntrs; 4920 if (c) { 4921 if (c->core_id == -1 || --c->refcnt == 0) 4922 kfree(c); 4923 cpuc->excl_cntrs = NULL; 4924 } 4925 4926 kfree(cpuc->constraint_list); 4927 cpuc->constraint_list = NULL; 4928 } 4929 4930 static void intel_pmu_cpu_dying(int cpu) 4931 { 4932 fini_debug_store_on_cpu(cpu); 4933 } 4934 4935 void intel_cpuc_finish(struct cpu_hw_events *cpuc) 4936 { 4937 struct intel_shared_regs *pc; 4938 4939 pc = cpuc->shared_regs; 4940 if (pc) { 4941 if (pc->core_id == -1 || --pc->refcnt == 0) 4942 kfree(pc); 4943 cpuc->shared_regs = NULL; 4944 } 4945 4946 free_excl_cntrs(cpuc); 4947 } 4948 4949 static void intel_pmu_cpu_dead(int cpu) 4950 { 4951 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4952 4953 intel_cpuc_finish(cpuc); 4954 4955 if (is_hybrid() && cpuc->pmu) 4956 cpumask_clear_cpu(cpu, &hybrid_pmu(cpuc->pmu)->supported_cpus); 4957 } 4958 4959 static void intel_pmu_sched_task(struct perf_event_pmu_context *pmu_ctx, 4960 bool sched_in) 4961 { 4962 intel_pmu_pebs_sched_task(pmu_ctx, sched_in); 4963 intel_pmu_lbr_sched_task(pmu_ctx, sched_in); 4964 } 4965 4966 static void intel_pmu_swap_task_ctx(struct perf_event_pmu_context *prev_epc, 4967 struct perf_event_pmu_context *next_epc) 4968 { 4969 intel_pmu_lbr_swap_task_ctx(prev_epc, next_epc); 4970 } 4971 4972 static int intel_pmu_check_period(struct perf_event *event, u64 value) 4973 { 4974 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0; 4975 } 4976 4977 static void intel_aux_output_init(void) 4978 { 4979 /* Refer also intel_pmu_aux_output_match() */ 4980 if (x86_pmu.intel_cap.pebs_output_pt_available) 4981 x86_pmu.assign = intel_pmu_assign_event; 4982 } 4983 4984 static int intel_pmu_aux_output_match(struct perf_event *event) 4985 { 4986 /* intel_pmu_assign_event() is needed, refer intel_aux_output_init() */ 4987 if (!x86_pmu.intel_cap.pebs_output_pt_available) 4988 return 0; 4989 4990 return is_intel_pt_event(event); 4991 } 4992 4993 static void intel_pmu_filter(struct pmu *pmu, int cpu, bool *ret) 4994 { 4995 struct x86_hybrid_pmu *hpmu = hybrid_pmu(pmu); 4996 4997 *ret = !cpumask_test_cpu(cpu, &hpmu->supported_cpus); 4998 } 4999 5000 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); 5001 5002 PMU_FORMAT_ATTR(ldlat, "config1:0-15"); 5003 5004 PMU_FORMAT_ATTR(frontend, "config1:0-23"); 5005 5006 PMU_FORMAT_ATTR(snoop_rsp, "config1:0-63"); 5007 5008 static struct attribute *intel_arch3_formats_attr[] = { 5009 &format_attr_event.attr, 5010 &format_attr_umask.attr, 5011 &format_attr_edge.attr, 5012 &format_attr_pc.attr, 5013 &format_attr_any.attr, 5014 &format_attr_inv.attr, 5015 &format_attr_cmask.attr, 5016 NULL, 5017 }; 5018 5019 static struct attribute *hsw_format_attr[] = { 5020 &format_attr_in_tx.attr, 5021 &format_attr_in_tx_cp.attr, 5022 &format_attr_offcore_rsp.attr, 5023 &format_attr_ldlat.attr, 5024 NULL 5025 }; 5026 5027 static struct attribute *nhm_format_attr[] = { 5028 &format_attr_offcore_rsp.attr, 5029 &format_attr_ldlat.attr, 5030 NULL 5031 }; 5032 5033 static struct attribute *slm_format_attr[] = { 5034 &format_attr_offcore_rsp.attr, 5035 NULL 5036 }; 5037 5038 static struct attribute *cmt_format_attr[] = { 5039 &format_attr_offcore_rsp.attr, 5040 &format_attr_ldlat.attr, 5041 &format_attr_snoop_rsp.attr, 5042 NULL 5043 }; 5044 5045 static struct attribute *skl_format_attr[] = { 5046 &format_attr_frontend.attr, 5047 NULL, 5048 }; 5049 5050 static __initconst const struct x86_pmu core_pmu = { 5051 .name = "core", 5052 .handle_irq = x86_pmu_handle_irq, 5053 .disable_all = x86_pmu_disable_all, 5054 .enable_all = core_pmu_enable_all, 5055 .enable = core_pmu_enable_event, 5056 .disable = x86_pmu_disable_event, 5057 .hw_config = core_pmu_hw_config, 5058 .schedule_events = x86_schedule_events, 5059 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 5060 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 5061 .event_map = intel_pmu_event_map, 5062 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 5063 .apic = 1, 5064 .large_pebs_flags = LARGE_PEBS_FLAGS, 5065 5066 /* 5067 * Intel PMCs cannot be accessed sanely above 32-bit width, 5068 * so we install an artificial 1<<31 period regardless of 5069 * the generic event period: 5070 */ 5071 .max_period = (1ULL<<31) - 1, 5072 .get_event_constraints = intel_get_event_constraints, 5073 .put_event_constraints = intel_put_event_constraints, 5074 .event_constraints = intel_core_event_constraints, 5075 .guest_get_msrs = core_guest_get_msrs, 5076 .format_attrs = intel_arch_formats_attr, 5077 .events_sysfs_show = intel_event_sysfs_show, 5078 5079 /* 5080 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs 5081 * together with PMU version 1 and thus be using core_pmu with 5082 * shared_regs. We need following callbacks here to allocate 5083 * it properly. 5084 */ 5085 .cpu_prepare = intel_pmu_cpu_prepare, 5086 .cpu_starting = intel_pmu_cpu_starting, 5087 .cpu_dying = intel_pmu_cpu_dying, 5088 .cpu_dead = intel_pmu_cpu_dead, 5089 5090 .check_period = intel_pmu_check_period, 5091 5092 .lbr_reset = intel_pmu_lbr_reset_64, 5093 .lbr_read = intel_pmu_lbr_read_64, 5094 .lbr_save = intel_pmu_lbr_save, 5095 .lbr_restore = intel_pmu_lbr_restore, 5096 }; 5097 5098 static __initconst const struct x86_pmu intel_pmu = { 5099 .name = "Intel", 5100 .handle_irq = intel_pmu_handle_irq, 5101 .disable_all = intel_pmu_disable_all, 5102 .enable_all = intel_pmu_enable_all, 5103 .enable = intel_pmu_enable_event, 5104 .disable = intel_pmu_disable_event, 5105 .add = intel_pmu_add_event, 5106 .del = intel_pmu_del_event, 5107 .read = intel_pmu_read_event, 5108 .set_period = intel_pmu_set_period, 5109 .update = intel_pmu_update, 5110 .hw_config = intel_pmu_hw_config, 5111 .schedule_events = x86_schedule_events, 5112 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 5113 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 5114 .event_map = intel_pmu_event_map, 5115 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 5116 .apic = 1, 5117 .large_pebs_flags = LARGE_PEBS_FLAGS, 5118 /* 5119 * Intel PMCs cannot be accessed sanely above 32 bit width, 5120 * so we install an artificial 1<<31 period regardless of 5121 * the generic event period: 5122 */ 5123 .max_period = (1ULL << 31) - 1, 5124 .get_event_constraints = intel_get_event_constraints, 5125 .put_event_constraints = intel_put_event_constraints, 5126 .pebs_aliases = intel_pebs_aliases_core2, 5127 5128 .format_attrs = intel_arch3_formats_attr, 5129 .events_sysfs_show = intel_event_sysfs_show, 5130 5131 .cpu_prepare = intel_pmu_cpu_prepare, 5132 .cpu_starting = intel_pmu_cpu_starting, 5133 .cpu_dying = intel_pmu_cpu_dying, 5134 .cpu_dead = intel_pmu_cpu_dead, 5135 5136 .guest_get_msrs = intel_guest_get_msrs, 5137 .sched_task = intel_pmu_sched_task, 5138 .swap_task_ctx = intel_pmu_swap_task_ctx, 5139 5140 .check_period = intel_pmu_check_period, 5141 5142 .aux_output_match = intel_pmu_aux_output_match, 5143 5144 .lbr_reset = intel_pmu_lbr_reset_64, 5145 .lbr_read = intel_pmu_lbr_read_64, 5146 .lbr_save = intel_pmu_lbr_save, 5147 .lbr_restore = intel_pmu_lbr_restore, 5148 5149 /* 5150 * SMM has access to all 4 rings and while traditionally SMM code only 5151 * ran in CPL0, 2021-era firmware is starting to make use of CPL3 in SMM. 5152 * 5153 * Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction 5154 * between SMM or not, this results in what should be pure userspace 5155 * counters including SMM data. 5156 * 5157 * This is a clear privilege issue, therefore globally disable 5158 * counting SMM by default. 5159 */ 5160 .attr_freeze_on_smi = 1, 5161 }; 5162 5163 static __init void intel_clovertown_quirk(void) 5164 { 5165 /* 5166 * PEBS is unreliable due to: 5167 * 5168 * AJ67 - PEBS may experience CPL leaks 5169 * AJ68 - PEBS PMI may be delayed by one event 5170 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] 5171 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS 5172 * 5173 * AJ67 could be worked around by restricting the OS/USR flags. 5174 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. 5175 * 5176 * AJ106 could possibly be worked around by not allowing LBR 5177 * usage from PEBS, including the fixup. 5178 * AJ68 could possibly be worked around by always programming 5179 * a pebs_event_reset[0] value and coping with the lost events. 5180 * 5181 * But taken together it might just make sense to not enable PEBS on 5182 * these chips. 5183 */ 5184 pr_warn("PEBS disabled due to CPU errata\n"); 5185 x86_pmu.pebs = 0; 5186 x86_pmu.pebs_constraints = NULL; 5187 } 5188 5189 static const struct x86_cpu_desc isolation_ucodes[] = { 5190 INTEL_CPU_DESC(INTEL_FAM6_HASWELL, 3, 0x0000001f), 5191 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L, 1, 0x0000001e), 5192 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G, 1, 0x00000015), 5193 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037), 5194 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a), 5195 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL, 4, 0x00000023), 5196 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G, 1, 0x00000014), 5197 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 2, 0x00000010), 5198 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 3, 0x07000009), 5199 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 4, 0x0f000009), 5200 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 5, 0x0e000002), 5201 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 1, 0x0b000014), 5202 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021), 5203 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000), 5204 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 5, 0x00000000), 5205 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 6, 0x00000000), 5206 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 7, 0x00000000), 5207 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 11, 0x00000000), 5208 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L, 3, 0x0000007c), 5209 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE, 3, 0x0000007c), 5210 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 9, 0x0000004e), 5211 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 9, 0x0000004e), 5212 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 10, 0x0000004e), 5213 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 11, 0x0000004e), 5214 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 12, 0x0000004e), 5215 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 10, 0x0000004e), 5216 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 11, 0x0000004e), 5217 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 12, 0x0000004e), 5218 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 13, 0x0000004e), 5219 {} 5220 }; 5221 5222 static void intel_check_pebs_isolation(void) 5223 { 5224 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes); 5225 } 5226 5227 static __init void intel_pebs_isolation_quirk(void) 5228 { 5229 WARN_ON_ONCE(x86_pmu.check_microcode); 5230 x86_pmu.check_microcode = intel_check_pebs_isolation; 5231 intel_check_pebs_isolation(); 5232 } 5233 5234 static const struct x86_cpu_desc pebs_ucodes[] = { 5235 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028), 5236 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618), 5237 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c), 5238 {} 5239 }; 5240 5241 static bool intel_snb_pebs_broken(void) 5242 { 5243 return !x86_cpu_has_min_microcode_rev(pebs_ucodes); 5244 } 5245 5246 static void intel_snb_check_microcode(void) 5247 { 5248 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken) 5249 return; 5250 5251 /* 5252 * Serialized by the microcode lock.. 5253 */ 5254 if (x86_pmu.pebs_broken) { 5255 pr_info("PEBS enabled due to microcode update\n"); 5256 x86_pmu.pebs_broken = 0; 5257 } else { 5258 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); 5259 x86_pmu.pebs_broken = 1; 5260 } 5261 } 5262 5263 static bool is_lbr_from(unsigned long msr) 5264 { 5265 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr; 5266 5267 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr; 5268 } 5269 5270 /* 5271 * Under certain circumstances, access certain MSR may cause #GP. 5272 * The function tests if the input MSR can be safely accessed. 5273 */ 5274 static bool check_msr(unsigned long msr, u64 mask) 5275 { 5276 u64 val_old, val_new, val_tmp; 5277 5278 /* 5279 * Disable the check for real HW, so we don't 5280 * mess with potentially enabled registers: 5281 */ 5282 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) 5283 return true; 5284 5285 /* 5286 * Read the current value, change it and read it back to see if it 5287 * matches, this is needed to detect certain hardware emulators 5288 * (qemu/kvm) that don't trap on the MSR access and always return 0s. 5289 */ 5290 if (rdmsrl_safe(msr, &val_old)) 5291 return false; 5292 5293 /* 5294 * Only change the bits which can be updated by wrmsrl. 5295 */ 5296 val_tmp = val_old ^ mask; 5297 5298 if (is_lbr_from(msr)) 5299 val_tmp = lbr_from_signext_quirk_wr(val_tmp); 5300 5301 if (wrmsrl_safe(msr, val_tmp) || 5302 rdmsrl_safe(msr, &val_new)) 5303 return false; 5304 5305 /* 5306 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value 5307 * should equal rdmsrl()'s even with the quirk. 5308 */ 5309 if (val_new != val_tmp) 5310 return false; 5311 5312 if (is_lbr_from(msr)) 5313 val_old = lbr_from_signext_quirk_wr(val_old); 5314 5315 /* Here it's sure that the MSR can be safely accessed. 5316 * Restore the old value and return. 5317 */ 5318 wrmsrl(msr, val_old); 5319 5320 return true; 5321 } 5322 5323 static __init void intel_sandybridge_quirk(void) 5324 { 5325 x86_pmu.check_microcode = intel_snb_check_microcode; 5326 cpus_read_lock(); 5327 intel_snb_check_microcode(); 5328 cpus_read_unlock(); 5329 } 5330 5331 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { 5332 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, 5333 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, 5334 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, 5335 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, 5336 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, 5337 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, 5338 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, 5339 }; 5340 5341 static __init void intel_arch_events_quirk(void) 5342 { 5343 int bit; 5344 5345 /* disable event that reported as not present by cpuid */ 5346 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { 5347 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; 5348 pr_warn("CPUID marked event: \'%s\' unavailable\n", 5349 intel_arch_events_map[bit].name); 5350 } 5351 } 5352 5353 static __init void intel_nehalem_quirk(void) 5354 { 5355 union cpuid10_ebx ebx; 5356 5357 ebx.full = x86_pmu.events_maskl; 5358 if (ebx.split.no_branch_misses_retired) { 5359 /* 5360 * Erratum AAJ80 detected, we work it around by using 5361 * the BR_MISP_EXEC.ANY event. This will over-count 5362 * branch-misses, but it's still much better than the 5363 * architectural event which is often completely bogus: 5364 */ 5365 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 5366 ebx.split.no_branch_misses_retired = 0; 5367 x86_pmu.events_maskl = ebx.full; 5368 pr_info("CPU erratum AAJ80 worked around\n"); 5369 } 5370 } 5371 5372 /* 5373 * enable software workaround for errata: 5374 * SNB: BJ122 5375 * IVB: BV98 5376 * HSW: HSD29 5377 * 5378 * Only needed when HT is enabled. However detecting 5379 * if HT is enabled is difficult (model specific). So instead, 5380 * we enable the workaround in the early boot, and verify if 5381 * it is needed in a later initcall phase once we have valid 5382 * topology information to check if HT is actually enabled 5383 */ 5384 static __init void intel_ht_bug(void) 5385 { 5386 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; 5387 5388 x86_pmu.start_scheduling = intel_start_scheduling; 5389 x86_pmu.commit_scheduling = intel_commit_scheduling; 5390 x86_pmu.stop_scheduling = intel_stop_scheduling; 5391 } 5392 5393 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); 5394 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") 5395 5396 /* Haswell special events */ 5397 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); 5398 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); 5399 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); 5400 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); 5401 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); 5402 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); 5403 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); 5404 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); 5405 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); 5406 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); 5407 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); 5408 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); 5409 5410 static struct attribute *hsw_events_attrs[] = { 5411 EVENT_PTR(td_slots_issued), 5412 EVENT_PTR(td_slots_retired), 5413 EVENT_PTR(td_fetch_bubbles), 5414 EVENT_PTR(td_total_slots), 5415 EVENT_PTR(td_total_slots_scale), 5416 EVENT_PTR(td_recovery_bubbles), 5417 EVENT_PTR(td_recovery_bubbles_scale), 5418 NULL 5419 }; 5420 5421 static struct attribute *hsw_mem_events_attrs[] = { 5422 EVENT_PTR(mem_ld_hsw), 5423 EVENT_PTR(mem_st_hsw), 5424 NULL, 5425 }; 5426 5427 static struct attribute *hsw_tsx_events_attrs[] = { 5428 EVENT_PTR(tx_start), 5429 EVENT_PTR(tx_commit), 5430 EVENT_PTR(tx_abort), 5431 EVENT_PTR(tx_capacity), 5432 EVENT_PTR(tx_conflict), 5433 EVENT_PTR(el_start), 5434 EVENT_PTR(el_commit), 5435 EVENT_PTR(el_abort), 5436 EVENT_PTR(el_capacity), 5437 EVENT_PTR(el_conflict), 5438 EVENT_PTR(cycles_t), 5439 EVENT_PTR(cycles_ct), 5440 NULL 5441 }; 5442 5443 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80"); 5444 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2"); 5445 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80"); 5446 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2"); 5447 5448 static struct attribute *icl_events_attrs[] = { 5449 EVENT_PTR(mem_ld_hsw), 5450 EVENT_PTR(mem_st_hsw), 5451 NULL, 5452 }; 5453 5454 static struct attribute *icl_td_events_attrs[] = { 5455 EVENT_PTR(slots), 5456 EVENT_PTR(td_retiring), 5457 EVENT_PTR(td_bad_spec), 5458 EVENT_PTR(td_fe_bound), 5459 EVENT_PTR(td_be_bound), 5460 NULL, 5461 }; 5462 5463 static struct attribute *icl_tsx_events_attrs[] = { 5464 EVENT_PTR(tx_start), 5465 EVENT_PTR(tx_abort), 5466 EVENT_PTR(tx_commit), 5467 EVENT_PTR(tx_capacity_read), 5468 EVENT_PTR(tx_capacity_write), 5469 EVENT_PTR(tx_conflict), 5470 EVENT_PTR(el_start), 5471 EVENT_PTR(el_abort), 5472 EVENT_PTR(el_commit), 5473 EVENT_PTR(el_capacity_read), 5474 EVENT_PTR(el_capacity_write), 5475 EVENT_PTR(el_conflict), 5476 EVENT_PTR(cycles_t), 5477 EVENT_PTR(cycles_ct), 5478 NULL, 5479 }; 5480 5481 5482 EVENT_ATTR_STR(mem-stores, mem_st_spr, "event=0xcd,umask=0x2"); 5483 EVENT_ATTR_STR(mem-loads-aux, mem_ld_aux, "event=0x03,umask=0x82"); 5484 5485 static struct attribute *glc_events_attrs[] = { 5486 EVENT_PTR(mem_ld_hsw), 5487 EVENT_PTR(mem_st_spr), 5488 EVENT_PTR(mem_ld_aux), 5489 NULL, 5490 }; 5491 5492 static struct attribute *glc_td_events_attrs[] = { 5493 EVENT_PTR(slots), 5494 EVENT_PTR(td_retiring), 5495 EVENT_PTR(td_bad_spec), 5496 EVENT_PTR(td_fe_bound), 5497 EVENT_PTR(td_be_bound), 5498 EVENT_PTR(td_heavy_ops), 5499 EVENT_PTR(td_br_mispredict), 5500 EVENT_PTR(td_fetch_lat), 5501 EVENT_PTR(td_mem_bound), 5502 NULL, 5503 }; 5504 5505 static struct attribute *glc_tsx_events_attrs[] = { 5506 EVENT_PTR(tx_start), 5507 EVENT_PTR(tx_abort), 5508 EVENT_PTR(tx_commit), 5509 EVENT_PTR(tx_capacity_read), 5510 EVENT_PTR(tx_capacity_write), 5511 EVENT_PTR(tx_conflict), 5512 EVENT_PTR(cycles_t), 5513 EVENT_PTR(cycles_ct), 5514 NULL, 5515 }; 5516 5517 static ssize_t freeze_on_smi_show(struct device *cdev, 5518 struct device_attribute *attr, 5519 char *buf) 5520 { 5521 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi); 5522 } 5523 5524 static DEFINE_MUTEX(freeze_on_smi_mutex); 5525 5526 static ssize_t freeze_on_smi_store(struct device *cdev, 5527 struct device_attribute *attr, 5528 const char *buf, size_t count) 5529 { 5530 unsigned long val; 5531 ssize_t ret; 5532 5533 ret = kstrtoul(buf, 0, &val); 5534 if (ret) 5535 return ret; 5536 5537 if (val > 1) 5538 return -EINVAL; 5539 5540 mutex_lock(&freeze_on_smi_mutex); 5541 5542 if (x86_pmu.attr_freeze_on_smi == val) 5543 goto done; 5544 5545 x86_pmu.attr_freeze_on_smi = val; 5546 5547 cpus_read_lock(); 5548 on_each_cpu(flip_smm_bit, &val, 1); 5549 cpus_read_unlock(); 5550 done: 5551 mutex_unlock(&freeze_on_smi_mutex); 5552 5553 return count; 5554 } 5555 5556 static void update_tfa_sched(void *ignored) 5557 { 5558 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 5559 5560 /* 5561 * check if PMC3 is used 5562 * and if so force schedule out for all event types all contexts 5563 */ 5564 if (test_bit(3, cpuc->active_mask)) 5565 perf_pmu_resched(x86_get_pmu(smp_processor_id())); 5566 } 5567 5568 static ssize_t show_sysctl_tfa(struct device *cdev, 5569 struct device_attribute *attr, 5570 char *buf) 5571 { 5572 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort); 5573 } 5574 5575 static ssize_t set_sysctl_tfa(struct device *cdev, 5576 struct device_attribute *attr, 5577 const char *buf, size_t count) 5578 { 5579 bool val; 5580 ssize_t ret; 5581 5582 ret = kstrtobool(buf, &val); 5583 if (ret) 5584 return ret; 5585 5586 /* no change */ 5587 if (val == allow_tsx_force_abort) 5588 return count; 5589 5590 allow_tsx_force_abort = val; 5591 5592 cpus_read_lock(); 5593 on_each_cpu(update_tfa_sched, NULL, 1); 5594 cpus_read_unlock(); 5595 5596 return count; 5597 } 5598 5599 5600 static DEVICE_ATTR_RW(freeze_on_smi); 5601 5602 static ssize_t branches_show(struct device *cdev, 5603 struct device_attribute *attr, 5604 char *buf) 5605 { 5606 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr); 5607 } 5608 5609 static DEVICE_ATTR_RO(branches); 5610 5611 static ssize_t branch_counter_nr_show(struct device *cdev, 5612 struct device_attribute *attr, 5613 char *buf) 5614 { 5615 return snprintf(buf, PAGE_SIZE, "%d\n", fls(x86_pmu.lbr_counters)); 5616 } 5617 5618 static DEVICE_ATTR_RO(branch_counter_nr); 5619 5620 static ssize_t branch_counter_width_show(struct device *cdev, 5621 struct device_attribute *attr, 5622 char *buf) 5623 { 5624 return snprintf(buf, PAGE_SIZE, "%d\n", LBR_INFO_BR_CNTR_BITS); 5625 } 5626 5627 static DEVICE_ATTR_RO(branch_counter_width); 5628 5629 static struct attribute *lbr_attrs[] = { 5630 &dev_attr_branches.attr, 5631 &dev_attr_branch_counter_nr.attr, 5632 &dev_attr_branch_counter_width.attr, 5633 NULL 5634 }; 5635 5636 static umode_t 5637 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5638 { 5639 /* branches */ 5640 if (i == 0) 5641 return x86_pmu.lbr_nr ? attr->mode : 0; 5642 5643 return (x86_pmu.flags & PMU_FL_BR_CNTR) ? attr->mode : 0; 5644 } 5645 5646 static char pmu_name_str[30]; 5647 5648 static ssize_t pmu_name_show(struct device *cdev, 5649 struct device_attribute *attr, 5650 char *buf) 5651 { 5652 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str); 5653 } 5654 5655 static DEVICE_ATTR_RO(pmu_name); 5656 5657 static struct attribute *intel_pmu_caps_attrs[] = { 5658 &dev_attr_pmu_name.attr, 5659 NULL 5660 }; 5661 5662 static DEVICE_ATTR(allow_tsx_force_abort, 0644, 5663 show_sysctl_tfa, 5664 set_sysctl_tfa); 5665 5666 static struct attribute *intel_pmu_attrs[] = { 5667 &dev_attr_freeze_on_smi.attr, 5668 &dev_attr_allow_tsx_force_abort.attr, 5669 NULL, 5670 }; 5671 5672 static umode_t 5673 default_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5674 { 5675 if (attr == &dev_attr_allow_tsx_force_abort.attr) 5676 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0; 5677 5678 return attr->mode; 5679 } 5680 5681 static umode_t 5682 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5683 { 5684 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0; 5685 } 5686 5687 static umode_t 5688 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5689 { 5690 return x86_pmu.pebs ? attr->mode : 0; 5691 } 5692 5693 static umode_t 5694 mem_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5695 { 5696 if (attr == &event_attr_mem_ld_aux.attr.attr) 5697 return x86_pmu.flags & PMU_FL_MEM_LOADS_AUX ? attr->mode : 0; 5698 5699 return pebs_is_visible(kobj, attr, i); 5700 } 5701 5702 static umode_t 5703 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5704 { 5705 return x86_pmu.version >= 2 ? attr->mode : 0; 5706 } 5707 5708 static struct attribute_group group_events_td = { 5709 .name = "events", 5710 }; 5711 5712 static struct attribute_group group_events_mem = { 5713 .name = "events", 5714 .is_visible = mem_is_visible, 5715 }; 5716 5717 static struct attribute_group group_events_tsx = { 5718 .name = "events", 5719 .is_visible = tsx_is_visible, 5720 }; 5721 5722 static struct attribute_group group_caps_gen = { 5723 .name = "caps", 5724 .attrs = intel_pmu_caps_attrs, 5725 }; 5726 5727 static struct attribute_group group_caps_lbr = { 5728 .name = "caps", 5729 .attrs = lbr_attrs, 5730 .is_visible = lbr_is_visible, 5731 }; 5732 5733 static struct attribute_group group_format_extra = { 5734 .name = "format", 5735 .is_visible = exra_is_visible, 5736 }; 5737 5738 static struct attribute_group group_format_extra_skl = { 5739 .name = "format", 5740 .is_visible = exra_is_visible, 5741 }; 5742 5743 static struct attribute_group group_default = { 5744 .attrs = intel_pmu_attrs, 5745 .is_visible = default_is_visible, 5746 }; 5747 5748 static const struct attribute_group *attr_update[] = { 5749 &group_events_td, 5750 &group_events_mem, 5751 &group_events_tsx, 5752 &group_caps_gen, 5753 &group_caps_lbr, 5754 &group_format_extra, 5755 &group_format_extra_skl, 5756 &group_default, 5757 NULL, 5758 }; 5759 5760 EVENT_ATTR_STR_HYBRID(slots, slots_adl, "event=0x00,umask=0x4", hybrid_big); 5761 EVENT_ATTR_STR_HYBRID(topdown-retiring, td_retiring_adl, "event=0xc2,umask=0x0;event=0x00,umask=0x80", hybrid_big_small); 5762 EVENT_ATTR_STR_HYBRID(topdown-bad-spec, td_bad_spec_adl, "event=0x73,umask=0x0;event=0x00,umask=0x81", hybrid_big_small); 5763 EVENT_ATTR_STR_HYBRID(topdown-fe-bound, td_fe_bound_adl, "event=0x71,umask=0x0;event=0x00,umask=0x82", hybrid_big_small); 5764 EVENT_ATTR_STR_HYBRID(topdown-be-bound, td_be_bound_adl, "event=0x74,umask=0x0;event=0x00,umask=0x83", hybrid_big_small); 5765 EVENT_ATTR_STR_HYBRID(topdown-heavy-ops, td_heavy_ops_adl, "event=0x00,umask=0x84", hybrid_big); 5766 EVENT_ATTR_STR_HYBRID(topdown-br-mispredict, td_br_mis_adl, "event=0x00,umask=0x85", hybrid_big); 5767 EVENT_ATTR_STR_HYBRID(topdown-fetch-lat, td_fetch_lat_adl, "event=0x00,umask=0x86", hybrid_big); 5768 EVENT_ATTR_STR_HYBRID(topdown-mem-bound, td_mem_bound_adl, "event=0x00,umask=0x87", hybrid_big); 5769 5770 static struct attribute *adl_hybrid_events_attrs[] = { 5771 EVENT_PTR(slots_adl), 5772 EVENT_PTR(td_retiring_adl), 5773 EVENT_PTR(td_bad_spec_adl), 5774 EVENT_PTR(td_fe_bound_adl), 5775 EVENT_PTR(td_be_bound_adl), 5776 EVENT_PTR(td_heavy_ops_adl), 5777 EVENT_PTR(td_br_mis_adl), 5778 EVENT_PTR(td_fetch_lat_adl), 5779 EVENT_PTR(td_mem_bound_adl), 5780 NULL, 5781 }; 5782 5783 /* Must be in IDX order */ 5784 EVENT_ATTR_STR_HYBRID(mem-loads, mem_ld_adl, "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3", hybrid_big_small); 5785 EVENT_ATTR_STR_HYBRID(mem-stores, mem_st_adl, "event=0xd0,umask=0x6;event=0xcd,umask=0x2", hybrid_big_small); 5786 EVENT_ATTR_STR_HYBRID(mem-loads-aux, mem_ld_aux_adl, "event=0x03,umask=0x82", hybrid_big); 5787 5788 static struct attribute *adl_hybrid_mem_attrs[] = { 5789 EVENT_PTR(mem_ld_adl), 5790 EVENT_PTR(mem_st_adl), 5791 EVENT_PTR(mem_ld_aux_adl), 5792 NULL, 5793 }; 5794 5795 static struct attribute *mtl_hybrid_mem_attrs[] = { 5796 EVENT_PTR(mem_ld_adl), 5797 EVENT_PTR(mem_st_adl), 5798 NULL 5799 }; 5800 5801 EVENT_ATTR_STR_HYBRID(tx-start, tx_start_adl, "event=0xc9,umask=0x1", hybrid_big); 5802 EVENT_ATTR_STR_HYBRID(tx-commit, tx_commit_adl, "event=0xc9,umask=0x2", hybrid_big); 5803 EVENT_ATTR_STR_HYBRID(tx-abort, tx_abort_adl, "event=0xc9,umask=0x4", hybrid_big); 5804 EVENT_ATTR_STR_HYBRID(tx-conflict, tx_conflict_adl, "event=0x54,umask=0x1", hybrid_big); 5805 EVENT_ATTR_STR_HYBRID(cycles-t, cycles_t_adl, "event=0x3c,in_tx=1", hybrid_big); 5806 EVENT_ATTR_STR_HYBRID(cycles-ct, cycles_ct_adl, "event=0x3c,in_tx=1,in_tx_cp=1", hybrid_big); 5807 EVENT_ATTR_STR_HYBRID(tx-capacity-read, tx_capacity_read_adl, "event=0x54,umask=0x80", hybrid_big); 5808 EVENT_ATTR_STR_HYBRID(tx-capacity-write, tx_capacity_write_adl, "event=0x54,umask=0x2", hybrid_big); 5809 5810 static struct attribute *adl_hybrid_tsx_attrs[] = { 5811 EVENT_PTR(tx_start_adl), 5812 EVENT_PTR(tx_abort_adl), 5813 EVENT_PTR(tx_commit_adl), 5814 EVENT_PTR(tx_capacity_read_adl), 5815 EVENT_PTR(tx_capacity_write_adl), 5816 EVENT_PTR(tx_conflict_adl), 5817 EVENT_PTR(cycles_t_adl), 5818 EVENT_PTR(cycles_ct_adl), 5819 NULL, 5820 }; 5821 5822 FORMAT_ATTR_HYBRID(in_tx, hybrid_big); 5823 FORMAT_ATTR_HYBRID(in_tx_cp, hybrid_big); 5824 FORMAT_ATTR_HYBRID(offcore_rsp, hybrid_big_small); 5825 FORMAT_ATTR_HYBRID(ldlat, hybrid_big_small); 5826 FORMAT_ATTR_HYBRID(frontend, hybrid_big); 5827 5828 #define ADL_HYBRID_RTM_FORMAT_ATTR \ 5829 FORMAT_HYBRID_PTR(in_tx), \ 5830 FORMAT_HYBRID_PTR(in_tx_cp) 5831 5832 #define ADL_HYBRID_FORMAT_ATTR \ 5833 FORMAT_HYBRID_PTR(offcore_rsp), \ 5834 FORMAT_HYBRID_PTR(ldlat), \ 5835 FORMAT_HYBRID_PTR(frontend) 5836 5837 static struct attribute *adl_hybrid_extra_attr_rtm[] = { 5838 ADL_HYBRID_RTM_FORMAT_ATTR, 5839 ADL_HYBRID_FORMAT_ATTR, 5840 NULL 5841 }; 5842 5843 static struct attribute *adl_hybrid_extra_attr[] = { 5844 ADL_HYBRID_FORMAT_ATTR, 5845 NULL 5846 }; 5847 5848 FORMAT_ATTR_HYBRID(snoop_rsp, hybrid_small); 5849 5850 static struct attribute *mtl_hybrid_extra_attr_rtm[] = { 5851 ADL_HYBRID_RTM_FORMAT_ATTR, 5852 ADL_HYBRID_FORMAT_ATTR, 5853 FORMAT_HYBRID_PTR(snoop_rsp), 5854 NULL 5855 }; 5856 5857 static struct attribute *mtl_hybrid_extra_attr[] = { 5858 ADL_HYBRID_FORMAT_ATTR, 5859 FORMAT_HYBRID_PTR(snoop_rsp), 5860 NULL 5861 }; 5862 5863 static bool is_attr_for_this_pmu(struct kobject *kobj, struct attribute *attr) 5864 { 5865 struct device *dev = kobj_to_dev(kobj); 5866 struct x86_hybrid_pmu *pmu = 5867 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5868 struct perf_pmu_events_hybrid_attr *pmu_attr = 5869 container_of(attr, struct perf_pmu_events_hybrid_attr, attr.attr); 5870 5871 return pmu->pmu_type & pmu_attr->pmu_type; 5872 } 5873 5874 static umode_t hybrid_events_is_visible(struct kobject *kobj, 5875 struct attribute *attr, int i) 5876 { 5877 return is_attr_for_this_pmu(kobj, attr) ? attr->mode : 0; 5878 } 5879 5880 static inline int hybrid_find_supported_cpu(struct x86_hybrid_pmu *pmu) 5881 { 5882 int cpu = cpumask_first(&pmu->supported_cpus); 5883 5884 return (cpu >= nr_cpu_ids) ? -1 : cpu; 5885 } 5886 5887 static umode_t hybrid_tsx_is_visible(struct kobject *kobj, 5888 struct attribute *attr, int i) 5889 { 5890 struct device *dev = kobj_to_dev(kobj); 5891 struct x86_hybrid_pmu *pmu = 5892 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5893 int cpu = hybrid_find_supported_cpu(pmu); 5894 5895 return (cpu >= 0) && is_attr_for_this_pmu(kobj, attr) && cpu_has(&cpu_data(cpu), X86_FEATURE_RTM) ? attr->mode : 0; 5896 } 5897 5898 static umode_t hybrid_format_is_visible(struct kobject *kobj, 5899 struct attribute *attr, int i) 5900 { 5901 struct device *dev = kobj_to_dev(kobj); 5902 struct x86_hybrid_pmu *pmu = 5903 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5904 struct perf_pmu_format_hybrid_attr *pmu_attr = 5905 container_of(attr, struct perf_pmu_format_hybrid_attr, attr.attr); 5906 int cpu = hybrid_find_supported_cpu(pmu); 5907 5908 return (cpu >= 0) && (pmu->pmu_type & pmu_attr->pmu_type) ? attr->mode : 0; 5909 } 5910 5911 static struct attribute_group hybrid_group_events_td = { 5912 .name = "events", 5913 .is_visible = hybrid_events_is_visible, 5914 }; 5915 5916 static struct attribute_group hybrid_group_events_mem = { 5917 .name = "events", 5918 .is_visible = hybrid_events_is_visible, 5919 }; 5920 5921 static struct attribute_group hybrid_group_events_tsx = { 5922 .name = "events", 5923 .is_visible = hybrid_tsx_is_visible, 5924 }; 5925 5926 static struct attribute_group hybrid_group_format_extra = { 5927 .name = "format", 5928 .is_visible = hybrid_format_is_visible, 5929 }; 5930 5931 static ssize_t intel_hybrid_get_attr_cpus(struct device *dev, 5932 struct device_attribute *attr, 5933 char *buf) 5934 { 5935 struct x86_hybrid_pmu *pmu = 5936 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5937 5938 return cpumap_print_to_pagebuf(true, buf, &pmu->supported_cpus); 5939 } 5940 5941 static DEVICE_ATTR(cpus, S_IRUGO, intel_hybrid_get_attr_cpus, NULL); 5942 static struct attribute *intel_hybrid_cpus_attrs[] = { 5943 &dev_attr_cpus.attr, 5944 NULL, 5945 }; 5946 5947 static struct attribute_group hybrid_group_cpus = { 5948 .attrs = intel_hybrid_cpus_attrs, 5949 }; 5950 5951 static const struct attribute_group *hybrid_attr_update[] = { 5952 &hybrid_group_events_td, 5953 &hybrid_group_events_mem, 5954 &hybrid_group_events_tsx, 5955 &group_caps_gen, 5956 &group_caps_lbr, 5957 &hybrid_group_format_extra, 5958 &group_default, 5959 &hybrid_group_cpus, 5960 NULL, 5961 }; 5962 5963 static struct attribute *empty_attrs; 5964 5965 static void intel_pmu_check_num_counters(int *num_counters, 5966 int *num_counters_fixed, 5967 u64 *intel_ctrl, u64 fixed_mask) 5968 { 5969 if (*num_counters > INTEL_PMC_MAX_GENERIC) { 5970 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", 5971 *num_counters, INTEL_PMC_MAX_GENERIC); 5972 *num_counters = INTEL_PMC_MAX_GENERIC; 5973 } 5974 *intel_ctrl = (1ULL << *num_counters) - 1; 5975 5976 if (*num_counters_fixed > INTEL_PMC_MAX_FIXED) { 5977 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", 5978 *num_counters_fixed, INTEL_PMC_MAX_FIXED); 5979 *num_counters_fixed = INTEL_PMC_MAX_FIXED; 5980 } 5981 5982 *intel_ctrl |= fixed_mask << INTEL_PMC_IDX_FIXED; 5983 } 5984 5985 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints, 5986 int num_counters, 5987 int num_counters_fixed, 5988 u64 intel_ctrl) 5989 { 5990 struct event_constraint *c; 5991 5992 if (!event_constraints) 5993 return; 5994 5995 /* 5996 * event on fixed counter2 (REF_CYCLES) only works on this 5997 * counter, so do not extend mask to generic counters 5998 */ 5999 for_each_event_constraint(c, event_constraints) { 6000 /* 6001 * Don't extend the topdown slots and metrics 6002 * events to the generic counters. 6003 */ 6004 if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) { 6005 /* 6006 * Disable topdown slots and metrics events, 6007 * if slots event is not in CPUID. 6008 */ 6009 if (!(INTEL_PMC_MSK_FIXED_SLOTS & intel_ctrl)) 6010 c->idxmsk64 = 0; 6011 c->weight = hweight64(c->idxmsk64); 6012 continue; 6013 } 6014 6015 if (c->cmask == FIXED_EVENT_FLAGS) { 6016 /* Disabled fixed counters which are not in CPUID */ 6017 c->idxmsk64 &= intel_ctrl; 6018 6019 /* 6020 * Don't extend the pseudo-encoding to the 6021 * generic counters 6022 */ 6023 if (!use_fixed_pseudo_encoding(c->code)) 6024 c->idxmsk64 |= (1ULL << num_counters) - 1; 6025 } 6026 c->idxmsk64 &= 6027 ~(~0ULL << (INTEL_PMC_IDX_FIXED + num_counters_fixed)); 6028 c->weight = hweight64(c->idxmsk64); 6029 } 6030 } 6031 6032 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs) 6033 { 6034 struct extra_reg *er; 6035 6036 /* 6037 * Access extra MSR may cause #GP under certain circumstances. 6038 * E.g. KVM doesn't support offcore event 6039 * Check all extra_regs here. 6040 */ 6041 if (!extra_regs) 6042 return; 6043 6044 for (er = extra_regs; er->msr; er++) { 6045 er->extra_msr_access = check_msr(er->msr, 0x11UL); 6046 /* Disable LBR select mapping */ 6047 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) 6048 x86_pmu.lbr_sel_map = NULL; 6049 } 6050 } 6051 6052 static const struct { enum hybrid_pmu_type id; char *name; } intel_hybrid_pmu_type_map[] __initconst = { 6053 { hybrid_small, "cpu_atom" }, 6054 { hybrid_big, "cpu_core" }, 6055 }; 6056 6057 static __always_inline int intel_pmu_init_hybrid(enum hybrid_pmu_type pmus) 6058 { 6059 unsigned long pmus_mask = pmus; 6060 struct x86_hybrid_pmu *pmu; 6061 int idx = 0, bit; 6062 6063 x86_pmu.num_hybrid_pmus = hweight_long(pmus_mask); 6064 x86_pmu.hybrid_pmu = kcalloc(x86_pmu.num_hybrid_pmus, 6065 sizeof(struct x86_hybrid_pmu), 6066 GFP_KERNEL); 6067 if (!x86_pmu.hybrid_pmu) 6068 return -ENOMEM; 6069 6070 static_branch_enable(&perf_is_hybrid); 6071 x86_pmu.filter = intel_pmu_filter; 6072 6073 for_each_set_bit(bit, &pmus_mask, ARRAY_SIZE(intel_hybrid_pmu_type_map)) { 6074 pmu = &x86_pmu.hybrid_pmu[idx++]; 6075 pmu->pmu_type = intel_hybrid_pmu_type_map[bit].id; 6076 pmu->name = intel_hybrid_pmu_type_map[bit].name; 6077 6078 pmu->num_counters = x86_pmu.num_counters; 6079 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6080 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters); 6081 pmu->unconstrained = (struct event_constraint) 6082 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6083 0, pmu->num_counters, 0, 0); 6084 6085 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; 6086 if (pmu->pmu_type & hybrid_small) { 6087 pmu->intel_cap.perf_metrics = 0; 6088 pmu->intel_cap.pebs_output_pt_available = 1; 6089 pmu->mid_ack = true; 6090 } else if (pmu->pmu_type & hybrid_big) { 6091 pmu->intel_cap.perf_metrics = 1; 6092 pmu->intel_cap.pebs_output_pt_available = 0; 6093 pmu->late_ack = true; 6094 } 6095 } 6096 6097 return 0; 6098 } 6099 6100 static __always_inline void intel_pmu_ref_cycles_ext(void) 6101 { 6102 if (!(x86_pmu.events_maskl & (INTEL_PMC_MSK_FIXED_REF_CYCLES >> INTEL_PMC_IDX_FIXED))) 6103 intel_perfmon_event_map[PERF_COUNT_HW_REF_CPU_CYCLES] = 0x013c; 6104 } 6105 6106 static __always_inline void intel_pmu_init_glc(struct pmu *pmu) 6107 { 6108 x86_pmu.late_ack = true; 6109 x86_pmu.limit_period = glc_limit_period; 6110 x86_pmu.pebs_aliases = NULL; 6111 x86_pmu.pebs_prec_dist = true; 6112 x86_pmu.pebs_block = true; 6113 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6114 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6115 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6116 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6117 x86_pmu.lbr_pt_coexist = true; 6118 x86_pmu.num_topdown_events = 8; 6119 static_call_update(intel_pmu_update_topdown_event, 6120 &icl_update_topdown_event); 6121 static_call_update(intel_pmu_set_topdown_event_period, 6122 &icl_set_topdown_event_period); 6123 6124 memcpy(hybrid_var(pmu, hw_cache_event_ids), glc_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6125 memcpy(hybrid_var(pmu, hw_cache_extra_regs), glc_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6126 hybrid(pmu, event_constraints) = intel_glc_event_constraints; 6127 hybrid(pmu, pebs_constraints) = intel_glc_pebs_event_constraints; 6128 6129 intel_pmu_ref_cycles_ext(); 6130 } 6131 6132 static __always_inline void intel_pmu_init_grt(struct pmu *pmu) 6133 { 6134 x86_pmu.mid_ack = true; 6135 x86_pmu.limit_period = glc_limit_period; 6136 x86_pmu.pebs_aliases = NULL; 6137 x86_pmu.pebs_prec_dist = true; 6138 x86_pmu.pebs_block = true; 6139 x86_pmu.lbr_pt_coexist = true; 6140 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6141 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6142 6143 memcpy(hybrid_var(pmu, hw_cache_event_ids), glp_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6144 memcpy(hybrid_var(pmu, hw_cache_extra_regs), tnt_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6145 hybrid_var(pmu, hw_cache_event_ids)[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6146 hybrid(pmu, event_constraints) = intel_grt_event_constraints; 6147 hybrid(pmu, pebs_constraints) = intel_grt_pebs_event_constraints; 6148 hybrid(pmu, extra_regs) = intel_grt_extra_regs; 6149 6150 intel_pmu_ref_cycles_ext(); 6151 } 6152 6153 __init int intel_pmu_init(void) 6154 { 6155 struct attribute **extra_skl_attr = &empty_attrs; 6156 struct attribute **extra_attr = &empty_attrs; 6157 struct attribute **td_attr = &empty_attrs; 6158 struct attribute **mem_attr = &empty_attrs; 6159 struct attribute **tsx_attr = &empty_attrs; 6160 union cpuid10_edx edx; 6161 union cpuid10_eax eax; 6162 union cpuid10_ebx ebx; 6163 unsigned int fixed_mask; 6164 bool pmem = false; 6165 int version, i; 6166 char *name; 6167 struct x86_hybrid_pmu *pmu; 6168 6169 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 6170 switch (boot_cpu_data.x86) { 6171 case 0x6: 6172 return p6_pmu_init(); 6173 case 0xb: 6174 return knc_pmu_init(); 6175 case 0xf: 6176 return p4_pmu_init(); 6177 } 6178 return -ENODEV; 6179 } 6180 6181 /* 6182 * Check whether the Architectural PerfMon supports 6183 * Branch Misses Retired hw_event or not. 6184 */ 6185 cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full); 6186 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) 6187 return -ENODEV; 6188 6189 version = eax.split.version_id; 6190 if (version < 2) 6191 x86_pmu = core_pmu; 6192 else 6193 x86_pmu = intel_pmu; 6194 6195 x86_pmu.version = version; 6196 x86_pmu.num_counters = eax.split.num_counters; 6197 x86_pmu.cntval_bits = eax.split.bit_width; 6198 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; 6199 6200 x86_pmu.events_maskl = ebx.full; 6201 x86_pmu.events_mask_len = eax.split.mask_length; 6202 6203 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); 6204 x86_pmu.pebs_capable = PEBS_COUNTER_MASK; 6205 6206 /* 6207 * Quirk: v2 perfmon does not report fixed-purpose events, so 6208 * assume at least 3 events, when not running in a hypervisor: 6209 */ 6210 if (version > 1 && version < 5) { 6211 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR); 6212 6213 x86_pmu.num_counters_fixed = 6214 max((int)edx.split.num_counters_fixed, assume); 6215 6216 fixed_mask = (1L << x86_pmu.num_counters_fixed) - 1; 6217 } else if (version >= 5) 6218 x86_pmu.num_counters_fixed = fls(fixed_mask); 6219 6220 if (boot_cpu_has(X86_FEATURE_PDCM)) { 6221 u64 capabilities; 6222 6223 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); 6224 x86_pmu.intel_cap.capabilities = capabilities; 6225 } 6226 6227 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) { 6228 x86_pmu.lbr_reset = intel_pmu_lbr_reset_32; 6229 x86_pmu.lbr_read = intel_pmu_lbr_read_32; 6230 } 6231 6232 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) 6233 intel_pmu_arch_lbr_init(); 6234 6235 intel_ds_init(); 6236 6237 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ 6238 6239 if (version >= 5) { 6240 x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated; 6241 if (x86_pmu.intel_cap.anythread_deprecated) 6242 pr_cont(" AnyThread deprecated, "); 6243 } 6244 6245 /* 6246 * Install the hw-cache-events table: 6247 */ 6248 switch (boot_cpu_data.x86_model) { 6249 case INTEL_FAM6_CORE_YONAH: 6250 pr_cont("Core events, "); 6251 name = "core"; 6252 break; 6253 6254 case INTEL_FAM6_CORE2_MEROM: 6255 x86_add_quirk(intel_clovertown_quirk); 6256 fallthrough; 6257 6258 case INTEL_FAM6_CORE2_MEROM_L: 6259 case INTEL_FAM6_CORE2_PENRYN: 6260 case INTEL_FAM6_CORE2_DUNNINGTON: 6261 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, 6262 sizeof(hw_cache_event_ids)); 6263 6264 intel_pmu_lbr_init_core(); 6265 6266 x86_pmu.event_constraints = intel_core2_event_constraints; 6267 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; 6268 pr_cont("Core2 events, "); 6269 name = "core2"; 6270 break; 6271 6272 case INTEL_FAM6_NEHALEM: 6273 case INTEL_FAM6_NEHALEM_EP: 6274 case INTEL_FAM6_NEHALEM_EX: 6275 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, 6276 sizeof(hw_cache_event_ids)); 6277 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 6278 sizeof(hw_cache_extra_regs)); 6279 6280 intel_pmu_lbr_init_nhm(); 6281 6282 x86_pmu.event_constraints = intel_nehalem_event_constraints; 6283 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; 6284 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 6285 x86_pmu.extra_regs = intel_nehalem_extra_regs; 6286 x86_pmu.limit_period = nhm_limit_period; 6287 6288 mem_attr = nhm_mem_events_attrs; 6289 6290 /* UOPS_ISSUED.STALLED_CYCLES */ 6291 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6292 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6293 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 6294 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 6295 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 6296 6297 intel_pmu_pebs_data_source_nhm(); 6298 x86_add_quirk(intel_nehalem_quirk); 6299 x86_pmu.pebs_no_tlb = 1; 6300 extra_attr = nhm_format_attr; 6301 6302 pr_cont("Nehalem events, "); 6303 name = "nehalem"; 6304 break; 6305 6306 case INTEL_FAM6_ATOM_BONNELL: 6307 case INTEL_FAM6_ATOM_BONNELL_MID: 6308 case INTEL_FAM6_ATOM_SALTWELL: 6309 case INTEL_FAM6_ATOM_SALTWELL_MID: 6310 case INTEL_FAM6_ATOM_SALTWELL_TABLET: 6311 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, 6312 sizeof(hw_cache_event_ids)); 6313 6314 intel_pmu_lbr_init_atom(); 6315 6316 x86_pmu.event_constraints = intel_gen_event_constraints; 6317 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; 6318 x86_pmu.pebs_aliases = intel_pebs_aliases_core2; 6319 pr_cont("Atom events, "); 6320 name = "bonnell"; 6321 break; 6322 6323 case INTEL_FAM6_ATOM_SILVERMONT: 6324 case INTEL_FAM6_ATOM_SILVERMONT_D: 6325 case INTEL_FAM6_ATOM_SILVERMONT_MID: 6326 case INTEL_FAM6_ATOM_AIRMONT: 6327 case INTEL_FAM6_ATOM_AIRMONT_MID: 6328 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, 6329 sizeof(hw_cache_event_ids)); 6330 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, 6331 sizeof(hw_cache_extra_regs)); 6332 6333 intel_pmu_lbr_init_slm(); 6334 6335 x86_pmu.event_constraints = intel_slm_event_constraints; 6336 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 6337 x86_pmu.extra_regs = intel_slm_extra_regs; 6338 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6339 td_attr = slm_events_attrs; 6340 extra_attr = slm_format_attr; 6341 pr_cont("Silvermont events, "); 6342 name = "silvermont"; 6343 break; 6344 6345 case INTEL_FAM6_ATOM_GOLDMONT: 6346 case INTEL_FAM6_ATOM_GOLDMONT_D: 6347 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids, 6348 sizeof(hw_cache_event_ids)); 6349 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs, 6350 sizeof(hw_cache_extra_regs)); 6351 6352 intel_pmu_lbr_init_skl(); 6353 6354 x86_pmu.event_constraints = intel_slm_event_constraints; 6355 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints; 6356 x86_pmu.extra_regs = intel_glm_extra_regs; 6357 /* 6358 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 6359 * for precise cycles. 6360 * :pp is identical to :ppp 6361 */ 6362 x86_pmu.pebs_aliases = NULL; 6363 x86_pmu.pebs_prec_dist = true; 6364 x86_pmu.lbr_pt_coexist = true; 6365 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6366 td_attr = glm_events_attrs; 6367 extra_attr = slm_format_attr; 6368 pr_cont("Goldmont events, "); 6369 name = "goldmont"; 6370 break; 6371 6372 case INTEL_FAM6_ATOM_GOLDMONT_PLUS: 6373 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 6374 sizeof(hw_cache_event_ids)); 6375 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs, 6376 sizeof(hw_cache_extra_regs)); 6377 6378 intel_pmu_lbr_init_skl(); 6379 6380 x86_pmu.event_constraints = intel_slm_event_constraints; 6381 x86_pmu.extra_regs = intel_glm_extra_regs; 6382 /* 6383 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 6384 * for precise cycles. 6385 */ 6386 x86_pmu.pebs_aliases = NULL; 6387 x86_pmu.pebs_prec_dist = true; 6388 x86_pmu.lbr_pt_coexist = true; 6389 x86_pmu.pebs_capable = ~0ULL; 6390 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6391 x86_pmu.flags |= PMU_FL_PEBS_ALL; 6392 x86_pmu.get_event_constraints = glp_get_event_constraints; 6393 td_attr = glm_events_attrs; 6394 /* Goldmont Plus has 4-wide pipeline */ 6395 event_attr_td_total_slots_scale_glm.event_str = "4"; 6396 extra_attr = slm_format_attr; 6397 pr_cont("Goldmont plus events, "); 6398 name = "goldmont_plus"; 6399 break; 6400 6401 case INTEL_FAM6_ATOM_TREMONT_D: 6402 case INTEL_FAM6_ATOM_TREMONT: 6403 case INTEL_FAM6_ATOM_TREMONT_L: 6404 x86_pmu.late_ack = true; 6405 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 6406 sizeof(hw_cache_event_ids)); 6407 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 6408 sizeof(hw_cache_extra_regs)); 6409 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6410 6411 intel_pmu_lbr_init_skl(); 6412 6413 x86_pmu.event_constraints = intel_slm_event_constraints; 6414 x86_pmu.extra_regs = intel_tnt_extra_regs; 6415 /* 6416 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 6417 * for precise cycles. 6418 */ 6419 x86_pmu.pebs_aliases = NULL; 6420 x86_pmu.pebs_prec_dist = true; 6421 x86_pmu.lbr_pt_coexist = true; 6422 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6423 x86_pmu.get_event_constraints = tnt_get_event_constraints; 6424 td_attr = tnt_events_attrs; 6425 extra_attr = slm_format_attr; 6426 pr_cont("Tremont events, "); 6427 name = "Tremont"; 6428 break; 6429 6430 case INTEL_FAM6_ATOM_GRACEMONT: 6431 intel_pmu_init_grt(NULL); 6432 intel_pmu_pebs_data_source_grt(); 6433 x86_pmu.pebs_latency_data = adl_latency_data_small; 6434 x86_pmu.get_event_constraints = tnt_get_event_constraints; 6435 td_attr = tnt_events_attrs; 6436 mem_attr = grt_mem_attrs; 6437 extra_attr = nhm_format_attr; 6438 pr_cont("Gracemont events, "); 6439 name = "gracemont"; 6440 break; 6441 6442 case INTEL_FAM6_ATOM_CRESTMONT: 6443 case INTEL_FAM6_ATOM_CRESTMONT_X: 6444 intel_pmu_init_grt(NULL); 6445 x86_pmu.extra_regs = intel_cmt_extra_regs; 6446 intel_pmu_pebs_data_source_cmt(); 6447 x86_pmu.pebs_latency_data = mtl_latency_data_small; 6448 x86_pmu.get_event_constraints = cmt_get_event_constraints; 6449 td_attr = cmt_events_attrs; 6450 mem_attr = grt_mem_attrs; 6451 extra_attr = cmt_format_attr; 6452 pr_cont("Crestmont events, "); 6453 name = "crestmont"; 6454 break; 6455 6456 case INTEL_FAM6_WESTMERE: 6457 case INTEL_FAM6_WESTMERE_EP: 6458 case INTEL_FAM6_WESTMERE_EX: 6459 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, 6460 sizeof(hw_cache_event_ids)); 6461 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 6462 sizeof(hw_cache_extra_regs)); 6463 6464 intel_pmu_lbr_init_nhm(); 6465 6466 x86_pmu.event_constraints = intel_westmere_event_constraints; 6467 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 6468 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; 6469 x86_pmu.extra_regs = intel_westmere_extra_regs; 6470 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6471 6472 mem_attr = nhm_mem_events_attrs; 6473 6474 /* UOPS_ISSUED.STALLED_CYCLES */ 6475 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6476 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6477 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 6478 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 6479 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 6480 6481 intel_pmu_pebs_data_source_nhm(); 6482 extra_attr = nhm_format_attr; 6483 pr_cont("Westmere events, "); 6484 name = "westmere"; 6485 break; 6486 6487 case INTEL_FAM6_SANDYBRIDGE: 6488 case INTEL_FAM6_SANDYBRIDGE_X: 6489 x86_add_quirk(intel_sandybridge_quirk); 6490 x86_add_quirk(intel_ht_bug); 6491 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 6492 sizeof(hw_cache_event_ids)); 6493 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 6494 sizeof(hw_cache_extra_regs)); 6495 6496 intel_pmu_lbr_init_snb(); 6497 6498 x86_pmu.event_constraints = intel_snb_event_constraints; 6499 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; 6500 x86_pmu.pebs_aliases = intel_pebs_aliases_snb; 6501 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X) 6502 x86_pmu.extra_regs = intel_snbep_extra_regs; 6503 else 6504 x86_pmu.extra_regs = intel_snb_extra_regs; 6505 6506 6507 /* all extra regs are per-cpu when HT is on */ 6508 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6509 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6510 6511 td_attr = snb_events_attrs; 6512 mem_attr = snb_mem_events_attrs; 6513 6514 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 6515 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6516 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6517 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ 6518 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 6519 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); 6520 6521 extra_attr = nhm_format_attr; 6522 6523 pr_cont("SandyBridge events, "); 6524 name = "sandybridge"; 6525 break; 6526 6527 case INTEL_FAM6_IVYBRIDGE: 6528 case INTEL_FAM6_IVYBRIDGE_X: 6529 x86_add_quirk(intel_ht_bug); 6530 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 6531 sizeof(hw_cache_event_ids)); 6532 /* dTLB-load-misses on IVB is different than SNB */ 6533 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ 6534 6535 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 6536 sizeof(hw_cache_extra_regs)); 6537 6538 intel_pmu_lbr_init_snb(); 6539 6540 x86_pmu.event_constraints = intel_ivb_event_constraints; 6541 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; 6542 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6543 x86_pmu.pebs_prec_dist = true; 6544 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X) 6545 x86_pmu.extra_regs = intel_snbep_extra_regs; 6546 else 6547 x86_pmu.extra_regs = intel_snb_extra_regs; 6548 /* all extra regs are per-cpu when HT is on */ 6549 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6550 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6551 6552 td_attr = snb_events_attrs; 6553 mem_attr = snb_mem_events_attrs; 6554 6555 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 6556 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6557 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6558 6559 extra_attr = nhm_format_attr; 6560 6561 pr_cont("IvyBridge events, "); 6562 name = "ivybridge"; 6563 break; 6564 6565 6566 case INTEL_FAM6_HASWELL: 6567 case INTEL_FAM6_HASWELL_X: 6568 case INTEL_FAM6_HASWELL_L: 6569 case INTEL_FAM6_HASWELL_G: 6570 x86_add_quirk(intel_ht_bug); 6571 x86_add_quirk(intel_pebs_isolation_quirk); 6572 x86_pmu.late_ack = true; 6573 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6574 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6575 6576 intel_pmu_lbr_init_hsw(); 6577 6578 x86_pmu.event_constraints = intel_hsw_event_constraints; 6579 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; 6580 x86_pmu.extra_regs = intel_snbep_extra_regs; 6581 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6582 x86_pmu.pebs_prec_dist = true; 6583 /* all extra regs are per-cpu when HT is on */ 6584 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6585 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6586 6587 x86_pmu.hw_config = hsw_hw_config; 6588 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6589 x86_pmu.lbr_double_abort = true; 6590 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6591 hsw_format_attr : nhm_format_attr; 6592 td_attr = hsw_events_attrs; 6593 mem_attr = hsw_mem_events_attrs; 6594 tsx_attr = hsw_tsx_events_attrs; 6595 pr_cont("Haswell events, "); 6596 name = "haswell"; 6597 break; 6598 6599 case INTEL_FAM6_BROADWELL: 6600 case INTEL_FAM6_BROADWELL_D: 6601 case INTEL_FAM6_BROADWELL_G: 6602 case INTEL_FAM6_BROADWELL_X: 6603 x86_add_quirk(intel_pebs_isolation_quirk); 6604 x86_pmu.late_ack = true; 6605 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6606 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6607 6608 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ 6609 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | 6610 BDW_L3_MISS|HSW_SNOOP_DRAM; 6611 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| 6612 HSW_SNOOP_DRAM; 6613 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| 6614 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6615 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| 6616 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6617 6618 intel_pmu_lbr_init_hsw(); 6619 6620 x86_pmu.event_constraints = intel_bdw_event_constraints; 6621 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints; 6622 x86_pmu.extra_regs = intel_snbep_extra_regs; 6623 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6624 x86_pmu.pebs_prec_dist = true; 6625 /* all extra regs are per-cpu when HT is on */ 6626 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6627 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6628 6629 x86_pmu.hw_config = hsw_hw_config; 6630 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6631 x86_pmu.limit_period = bdw_limit_period; 6632 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6633 hsw_format_attr : nhm_format_attr; 6634 td_attr = hsw_events_attrs; 6635 mem_attr = hsw_mem_events_attrs; 6636 tsx_attr = hsw_tsx_events_attrs; 6637 pr_cont("Broadwell events, "); 6638 name = "broadwell"; 6639 break; 6640 6641 case INTEL_FAM6_XEON_PHI_KNL: 6642 case INTEL_FAM6_XEON_PHI_KNM: 6643 memcpy(hw_cache_event_ids, 6644 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6645 memcpy(hw_cache_extra_regs, 6646 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6647 intel_pmu_lbr_init_knl(); 6648 6649 x86_pmu.event_constraints = intel_slm_event_constraints; 6650 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 6651 x86_pmu.extra_regs = intel_knl_extra_regs; 6652 6653 /* all extra regs are per-cpu when HT is on */ 6654 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6655 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6656 extra_attr = slm_format_attr; 6657 pr_cont("Knights Landing/Mill events, "); 6658 name = "knights-landing"; 6659 break; 6660 6661 case INTEL_FAM6_SKYLAKE_X: 6662 pmem = true; 6663 fallthrough; 6664 case INTEL_FAM6_SKYLAKE_L: 6665 case INTEL_FAM6_SKYLAKE: 6666 case INTEL_FAM6_KABYLAKE_L: 6667 case INTEL_FAM6_KABYLAKE: 6668 case INTEL_FAM6_COMETLAKE_L: 6669 case INTEL_FAM6_COMETLAKE: 6670 x86_add_quirk(intel_pebs_isolation_quirk); 6671 x86_pmu.late_ack = true; 6672 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6673 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6674 intel_pmu_lbr_init_skl(); 6675 6676 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */ 6677 event_attr_td_recovery_bubbles.event_str_noht = 6678 "event=0xd,umask=0x1,cmask=1"; 6679 event_attr_td_recovery_bubbles.event_str_ht = 6680 "event=0xd,umask=0x1,cmask=1,any=1"; 6681 6682 x86_pmu.event_constraints = intel_skl_event_constraints; 6683 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; 6684 x86_pmu.extra_regs = intel_skl_extra_regs; 6685 x86_pmu.pebs_aliases = intel_pebs_aliases_skl; 6686 x86_pmu.pebs_prec_dist = true; 6687 /* all extra regs are per-cpu when HT is on */ 6688 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6689 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6690 6691 x86_pmu.hw_config = hsw_hw_config; 6692 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6693 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6694 hsw_format_attr : nhm_format_attr; 6695 extra_skl_attr = skl_format_attr; 6696 td_attr = hsw_events_attrs; 6697 mem_attr = hsw_mem_events_attrs; 6698 tsx_attr = hsw_tsx_events_attrs; 6699 intel_pmu_pebs_data_source_skl(pmem); 6700 6701 /* 6702 * Processors with CPUID.RTM_ALWAYS_ABORT have TSX deprecated by default. 6703 * TSX force abort hooks are not required on these systems. Only deploy 6704 * workaround when microcode has not enabled X86_FEATURE_RTM_ALWAYS_ABORT. 6705 */ 6706 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT) && 6707 !boot_cpu_has(X86_FEATURE_RTM_ALWAYS_ABORT)) { 6708 x86_pmu.flags |= PMU_FL_TFA; 6709 x86_pmu.get_event_constraints = tfa_get_event_constraints; 6710 x86_pmu.enable_all = intel_tfa_pmu_enable_all; 6711 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; 6712 } 6713 6714 pr_cont("Skylake events, "); 6715 name = "skylake"; 6716 break; 6717 6718 case INTEL_FAM6_ICELAKE_X: 6719 case INTEL_FAM6_ICELAKE_D: 6720 x86_pmu.pebs_ept = 1; 6721 pmem = true; 6722 fallthrough; 6723 case INTEL_FAM6_ICELAKE_L: 6724 case INTEL_FAM6_ICELAKE: 6725 case INTEL_FAM6_TIGERLAKE_L: 6726 case INTEL_FAM6_TIGERLAKE: 6727 case INTEL_FAM6_ROCKETLAKE: 6728 x86_pmu.late_ack = true; 6729 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6730 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6731 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6732 intel_pmu_lbr_init_skl(); 6733 6734 x86_pmu.event_constraints = intel_icl_event_constraints; 6735 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints; 6736 x86_pmu.extra_regs = intel_icl_extra_regs; 6737 x86_pmu.pebs_aliases = NULL; 6738 x86_pmu.pebs_prec_dist = true; 6739 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6740 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6741 6742 x86_pmu.hw_config = hsw_hw_config; 6743 x86_pmu.get_event_constraints = icl_get_event_constraints; 6744 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6745 hsw_format_attr : nhm_format_attr; 6746 extra_skl_attr = skl_format_attr; 6747 mem_attr = icl_events_attrs; 6748 td_attr = icl_td_events_attrs; 6749 tsx_attr = icl_tsx_events_attrs; 6750 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6751 x86_pmu.lbr_pt_coexist = true; 6752 intel_pmu_pebs_data_source_skl(pmem); 6753 x86_pmu.num_topdown_events = 4; 6754 static_call_update(intel_pmu_update_topdown_event, 6755 &icl_update_topdown_event); 6756 static_call_update(intel_pmu_set_topdown_event_period, 6757 &icl_set_topdown_event_period); 6758 pr_cont("Icelake events, "); 6759 name = "icelake"; 6760 break; 6761 6762 case INTEL_FAM6_SAPPHIRERAPIDS_X: 6763 case INTEL_FAM6_EMERALDRAPIDS_X: 6764 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6765 x86_pmu.extra_regs = intel_glc_extra_regs; 6766 fallthrough; 6767 case INTEL_FAM6_GRANITERAPIDS_X: 6768 case INTEL_FAM6_GRANITERAPIDS_D: 6769 intel_pmu_init_glc(NULL); 6770 if (!x86_pmu.extra_regs) 6771 x86_pmu.extra_regs = intel_rwc_extra_regs; 6772 x86_pmu.pebs_ept = 1; 6773 x86_pmu.hw_config = hsw_hw_config; 6774 x86_pmu.get_event_constraints = glc_get_event_constraints; 6775 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6776 hsw_format_attr : nhm_format_attr; 6777 extra_skl_attr = skl_format_attr; 6778 mem_attr = glc_events_attrs; 6779 td_attr = glc_td_events_attrs; 6780 tsx_attr = glc_tsx_events_attrs; 6781 intel_pmu_pebs_data_source_skl(true); 6782 pr_cont("Sapphire Rapids events, "); 6783 name = "sapphire_rapids"; 6784 break; 6785 6786 case INTEL_FAM6_ALDERLAKE: 6787 case INTEL_FAM6_ALDERLAKE_L: 6788 case INTEL_FAM6_RAPTORLAKE: 6789 case INTEL_FAM6_RAPTORLAKE_P: 6790 case INTEL_FAM6_RAPTORLAKE_S: 6791 /* 6792 * Alder Lake has 2 types of CPU, core and atom. 6793 * 6794 * Initialize the common PerfMon capabilities here. 6795 */ 6796 intel_pmu_init_hybrid(hybrid_big_small); 6797 6798 x86_pmu.pebs_latency_data = adl_latency_data_small; 6799 x86_pmu.get_event_constraints = adl_get_event_constraints; 6800 x86_pmu.hw_config = adl_hw_config; 6801 x86_pmu.get_hybrid_cpu_type = adl_get_hybrid_cpu_type; 6802 6803 td_attr = adl_hybrid_events_attrs; 6804 mem_attr = adl_hybrid_mem_attrs; 6805 tsx_attr = adl_hybrid_tsx_attrs; 6806 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6807 adl_hybrid_extra_attr_rtm : adl_hybrid_extra_attr; 6808 6809 /* Initialize big core specific PerfMon capabilities.*/ 6810 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX]; 6811 intel_pmu_init_glc(&pmu->pmu); 6812 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) { 6813 pmu->num_counters = x86_pmu.num_counters + 2; 6814 pmu->num_counters_fixed = x86_pmu.num_counters_fixed + 1; 6815 } else { 6816 pmu->num_counters = x86_pmu.num_counters; 6817 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6818 } 6819 6820 /* 6821 * Quirk: For some Alder Lake machine, when all E-cores are disabled in 6822 * a BIOS, the leaf 0xA will enumerate all counters of P-cores. However, 6823 * the X86_FEATURE_HYBRID_CPU is still set. The above codes will 6824 * mistakenly add extra counters for P-cores. Correct the number of 6825 * counters here. 6826 */ 6827 if ((pmu->num_counters > 8) || (pmu->num_counters_fixed > 4)) { 6828 pmu->num_counters = x86_pmu.num_counters; 6829 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6830 } 6831 6832 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters); 6833 pmu->unconstrained = (struct event_constraint) 6834 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6835 0, pmu->num_counters, 0, 0); 6836 pmu->extra_regs = intel_glc_extra_regs; 6837 6838 /* Initialize Atom core specific PerfMon capabilities.*/ 6839 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX]; 6840 intel_pmu_init_grt(&pmu->pmu); 6841 6842 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6843 intel_pmu_pebs_data_source_adl(); 6844 pr_cont("Alderlake Hybrid events, "); 6845 name = "alderlake_hybrid"; 6846 break; 6847 6848 case INTEL_FAM6_METEORLAKE: 6849 case INTEL_FAM6_METEORLAKE_L: 6850 intel_pmu_init_hybrid(hybrid_big_small); 6851 6852 x86_pmu.pebs_latency_data = mtl_latency_data_small; 6853 x86_pmu.get_event_constraints = mtl_get_event_constraints; 6854 x86_pmu.hw_config = adl_hw_config; 6855 6856 td_attr = adl_hybrid_events_attrs; 6857 mem_attr = mtl_hybrid_mem_attrs; 6858 tsx_attr = adl_hybrid_tsx_attrs; 6859 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6860 mtl_hybrid_extra_attr_rtm : mtl_hybrid_extra_attr; 6861 6862 /* Initialize big core specific PerfMon capabilities.*/ 6863 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX]; 6864 intel_pmu_init_glc(&pmu->pmu); 6865 pmu->extra_regs = intel_rwc_extra_regs; 6866 6867 /* Initialize Atom core specific PerfMon capabilities.*/ 6868 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX]; 6869 intel_pmu_init_grt(&pmu->pmu); 6870 pmu->extra_regs = intel_cmt_extra_regs; 6871 6872 intel_pmu_pebs_data_source_mtl(); 6873 pr_cont("Meteorlake Hybrid events, "); 6874 name = "meteorlake_hybrid"; 6875 break; 6876 6877 default: 6878 switch (x86_pmu.version) { 6879 case 1: 6880 x86_pmu.event_constraints = intel_v1_event_constraints; 6881 pr_cont("generic architected perfmon v1, "); 6882 name = "generic_arch_v1"; 6883 break; 6884 case 2: 6885 case 3: 6886 case 4: 6887 /* 6888 * default constraints for v2 and up 6889 */ 6890 x86_pmu.event_constraints = intel_gen_event_constraints; 6891 pr_cont("generic architected perfmon, "); 6892 name = "generic_arch_v2+"; 6893 break; 6894 default: 6895 /* 6896 * The default constraints for v5 and up can support up to 6897 * 16 fixed counters. For the fixed counters 4 and later, 6898 * the pseudo-encoding is applied. 6899 * The constraints may be cut according to the CPUID enumeration 6900 * by inserting the EVENT_CONSTRAINT_END. 6901 */ 6902 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) 6903 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; 6904 intel_v5_gen_event_constraints[x86_pmu.num_counters_fixed].weight = -1; 6905 x86_pmu.event_constraints = intel_v5_gen_event_constraints; 6906 pr_cont("generic architected perfmon, "); 6907 name = "generic_arch_v5+"; 6908 break; 6909 } 6910 } 6911 6912 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name); 6913 6914 if (!is_hybrid()) { 6915 group_events_td.attrs = td_attr; 6916 group_events_mem.attrs = mem_attr; 6917 group_events_tsx.attrs = tsx_attr; 6918 group_format_extra.attrs = extra_attr; 6919 group_format_extra_skl.attrs = extra_skl_attr; 6920 6921 x86_pmu.attr_update = attr_update; 6922 } else { 6923 hybrid_group_events_td.attrs = td_attr; 6924 hybrid_group_events_mem.attrs = mem_attr; 6925 hybrid_group_events_tsx.attrs = tsx_attr; 6926 hybrid_group_format_extra.attrs = extra_attr; 6927 6928 x86_pmu.attr_update = hybrid_attr_update; 6929 } 6930 6931 intel_pmu_check_num_counters(&x86_pmu.num_counters, 6932 &x86_pmu.num_counters_fixed, 6933 &x86_pmu.intel_ctrl, 6934 (u64)fixed_mask); 6935 6936 /* AnyThread may be deprecated on arch perfmon v5 or later */ 6937 if (x86_pmu.intel_cap.anythread_deprecated) 6938 x86_pmu.format_attrs = intel_arch_formats_attr; 6939 6940 intel_pmu_check_event_constraints(x86_pmu.event_constraints, 6941 x86_pmu.num_counters, 6942 x86_pmu.num_counters_fixed, 6943 x86_pmu.intel_ctrl); 6944 /* 6945 * Access LBR MSR may cause #GP under certain circumstances. 6946 * Check all LBR MSR here. 6947 * Disable LBR access if any LBR MSRs can not be accessed. 6948 */ 6949 if (x86_pmu.lbr_tos && !check_msr(x86_pmu.lbr_tos, 0x3UL)) 6950 x86_pmu.lbr_nr = 0; 6951 for (i = 0; i < x86_pmu.lbr_nr; i++) { 6952 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && 6953 check_msr(x86_pmu.lbr_to + i, 0xffffUL))) 6954 x86_pmu.lbr_nr = 0; 6955 } 6956 6957 if (x86_pmu.lbr_nr) { 6958 intel_pmu_lbr_init(); 6959 6960 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr); 6961 6962 /* only support branch_stack snapshot for perfmon >= v2 */ 6963 if (x86_pmu.disable_all == intel_pmu_disable_all) { 6964 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) { 6965 static_call_update(perf_snapshot_branch_stack, 6966 intel_pmu_snapshot_arch_branch_stack); 6967 } else { 6968 static_call_update(perf_snapshot_branch_stack, 6969 intel_pmu_snapshot_branch_stack); 6970 } 6971 } 6972 } 6973 6974 intel_pmu_check_extra_regs(x86_pmu.extra_regs); 6975 6976 /* Support full width counters using alternative MSR range */ 6977 if (x86_pmu.intel_cap.full_width_write) { 6978 x86_pmu.max_period = x86_pmu.cntval_mask >> 1; 6979 x86_pmu.perfctr = MSR_IA32_PMC0; 6980 pr_cont("full-width counters, "); 6981 } 6982 6983 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) 6984 x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 6985 6986 if (x86_pmu.intel_cap.pebs_timing_info) 6987 x86_pmu.flags |= PMU_FL_RETIRE_LATENCY; 6988 6989 intel_aux_output_init(); 6990 6991 return 0; 6992 } 6993 6994 /* 6995 * HT bug: phase 2 init 6996 * Called once we have valid topology information to check 6997 * whether or not HT is enabled 6998 * If HT is off, then we disable the workaround 6999 */ 7000 static __init int fixup_ht_bug(void) 7001 { 7002 int c; 7003 /* 7004 * problem not present on this CPU model, nothing to do 7005 */ 7006 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) 7007 return 0; 7008 7009 if (topology_max_smt_threads() > 1) { 7010 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); 7011 return 0; 7012 } 7013 7014 cpus_read_lock(); 7015 7016 hardlockup_detector_perf_stop(); 7017 7018 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); 7019 7020 x86_pmu.start_scheduling = NULL; 7021 x86_pmu.commit_scheduling = NULL; 7022 x86_pmu.stop_scheduling = NULL; 7023 7024 hardlockup_detector_perf_restart(); 7025 7026 for_each_online_cpu(c) 7027 free_excl_cntrs(&per_cpu(cpu_hw_events, c)); 7028 7029 cpus_read_unlock(); 7030 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); 7031 return 0; 7032 } 7033 subsys_initcall(fixup_ht_bug) 7034