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 18 #include <asm/cpufeature.h> 19 #include <asm/hardirq.h> 20 #include <asm/intel-family.h> 21 #include <asm/intel_pt.h> 22 #include <asm/apic.h> 23 #include <asm/cpu_device_id.h> 24 25 #include "../perf_event.h" 26 27 /* 28 * Intel PerfMon, used on Core and later. 29 */ 30 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = 31 { 32 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, 33 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, 34 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, 35 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, 36 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, 37 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, 38 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, 39 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ 40 }; 41 42 static struct event_constraint intel_core_event_constraints[] __read_mostly = 43 { 44 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 45 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 46 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 47 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 48 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 49 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ 50 EVENT_CONSTRAINT_END 51 }; 52 53 static struct event_constraint intel_core2_event_constraints[] __read_mostly = 54 { 55 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 56 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 57 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 58 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ 59 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 60 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 61 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 62 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 63 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ 64 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 65 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ 66 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ 67 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ 68 EVENT_CONSTRAINT_END 69 }; 70 71 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = 72 { 73 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 74 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 75 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 76 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ 77 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ 78 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ 79 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ 80 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ 81 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ 82 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 83 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 84 EVENT_CONSTRAINT_END 85 }; 86 87 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = 88 { 89 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 90 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 91 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 92 EVENT_EXTRA_END 93 }; 94 95 static struct event_constraint intel_westmere_event_constraints[] __read_mostly = 96 { 97 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 98 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 99 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 100 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 101 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ 102 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 103 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ 104 EVENT_CONSTRAINT_END 105 }; 106 107 static struct event_constraint intel_snb_event_constraints[] __read_mostly = 108 { 109 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 110 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 111 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 112 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 113 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 114 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 115 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 116 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ 117 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 118 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 119 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 120 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 121 122 /* 123 * When HT is off these events can only run on the bottom 4 counters 124 * When HT is on, they are impacted by the HT bug and require EXCL access 125 */ 126 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 127 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 128 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 129 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 130 131 EVENT_CONSTRAINT_END 132 }; 133 134 static struct event_constraint intel_ivb_event_constraints[] __read_mostly = 135 { 136 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 137 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 138 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 139 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ 140 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMPTY */ 141 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ 142 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ 143 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 144 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 145 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ 146 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 147 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 148 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 149 150 /* 151 * When HT is off these events can only run on the bottom 4 counters 152 * When HT is on, they are impacted by the HT bug and require EXCL access 153 */ 154 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 155 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 156 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 157 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 158 159 EVENT_CONSTRAINT_END 160 }; 161 162 static struct extra_reg intel_westmere_extra_regs[] __read_mostly = 163 { 164 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 165 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 166 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), 167 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 168 EVENT_EXTRA_END 169 }; 170 171 static struct event_constraint intel_v1_event_constraints[] __read_mostly = 172 { 173 EVENT_CONSTRAINT_END 174 }; 175 176 static struct event_constraint intel_gen_event_constraints[] __read_mostly = 177 { 178 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 179 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 180 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 181 EVENT_CONSTRAINT_END 182 }; 183 184 static struct event_constraint intel_v5_gen_event_constraints[] __read_mostly = 185 { 186 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 187 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 188 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 189 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 190 FIXED_EVENT_CONSTRAINT(0x0500, 4), 191 FIXED_EVENT_CONSTRAINT(0x0600, 5), 192 FIXED_EVENT_CONSTRAINT(0x0700, 6), 193 FIXED_EVENT_CONSTRAINT(0x0800, 7), 194 FIXED_EVENT_CONSTRAINT(0x0900, 8), 195 FIXED_EVENT_CONSTRAINT(0x0a00, 9), 196 FIXED_EVENT_CONSTRAINT(0x0b00, 10), 197 FIXED_EVENT_CONSTRAINT(0x0c00, 11), 198 FIXED_EVENT_CONSTRAINT(0x0d00, 12), 199 FIXED_EVENT_CONSTRAINT(0x0e00, 13), 200 FIXED_EVENT_CONSTRAINT(0x0f00, 14), 201 FIXED_EVENT_CONSTRAINT(0x1000, 15), 202 EVENT_CONSTRAINT_END 203 }; 204 205 static struct event_constraint intel_slm_event_constraints[] __read_mostly = 206 { 207 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 208 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 209 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ 210 EVENT_CONSTRAINT_END 211 }; 212 213 static struct event_constraint intel_skl_event_constraints[] = { 214 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 215 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 216 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 217 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ 218 219 /* 220 * when HT is off, these can only run on the bottom 4 counters 221 */ 222 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 223 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 224 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 225 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 226 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */ 227 228 EVENT_CONSTRAINT_END 229 }; 230 231 static struct extra_reg intel_knl_extra_regs[] __read_mostly = { 232 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0), 233 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1), 234 EVENT_EXTRA_END 235 }; 236 237 static struct extra_reg intel_snb_extra_regs[] __read_mostly = { 238 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 239 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), 240 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), 241 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 242 EVENT_EXTRA_END 243 }; 244 245 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { 246 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 247 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 248 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 249 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 250 EVENT_EXTRA_END 251 }; 252 253 static struct extra_reg intel_skl_extra_regs[] __read_mostly = { 254 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 255 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 256 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 257 /* 258 * Note the low 8 bits eventsel code is not a continuous field, containing 259 * some #GPing bits. These are masked out. 260 */ 261 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 262 EVENT_EXTRA_END 263 }; 264 265 static struct event_constraint intel_icl_event_constraints[] = { 266 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 267 FIXED_EVENT_CONSTRAINT(0x01c0, 0), /* old INST_RETIRED.PREC_DIST */ 268 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 269 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 270 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 271 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 272 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 273 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 274 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 275 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 276 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf), 277 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf), 278 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */ 279 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x54, 0xf), 280 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf), 281 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */ 282 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */ 283 INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */ 284 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */ 285 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf), 286 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf), 287 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf), 288 INTEL_EVENT_CONSTRAINT(0xef, 0xf), 289 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf), 290 EVENT_CONSTRAINT_END 291 }; 292 293 static struct extra_reg intel_icl_extra_regs[] __read_mostly = { 294 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0), 295 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1), 296 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 297 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 298 EVENT_EXTRA_END 299 }; 300 301 static struct extra_reg intel_spr_extra_regs[] __read_mostly = { 302 INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 303 INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 304 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 305 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE), 306 INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE), 307 INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE), 308 EVENT_EXTRA_END 309 }; 310 311 static struct event_constraint intel_spr_event_constraints[] = { 312 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 313 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 314 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 315 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 316 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 317 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 318 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 319 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 320 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 321 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4), 322 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5), 323 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6), 324 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7), 325 326 INTEL_EVENT_CONSTRAINT(0x2e, 0xff), 327 INTEL_EVENT_CONSTRAINT(0x3c, 0xff), 328 /* 329 * Generally event codes < 0x90 are restricted to counters 0-3. 330 * The 0x2E and 0x3C are exception, which has no restriction. 331 */ 332 INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf), 333 334 INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf), 335 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), 336 INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf), 337 INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1), 338 INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1), 339 INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1), 340 INTEL_EVENT_CONSTRAINT(0xce, 0x1), 341 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf), 342 /* 343 * Generally event codes >= 0x90 are likely to have no restrictions. 344 * The exception are defined as above. 345 */ 346 INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff), 347 348 EVENT_CONSTRAINT_END 349 }; 350 351 352 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); 353 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); 354 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); 355 356 static struct attribute *nhm_mem_events_attrs[] = { 357 EVENT_PTR(mem_ld_nhm), 358 NULL, 359 }; 360 361 /* 362 * topdown events for Intel Core CPUs. 363 * 364 * The events are all in slots, which is a free slot in a 4 wide 365 * pipeline. Some events are already reported in slots, for cycle 366 * events we multiply by the pipeline width (4). 367 * 368 * With Hyper Threading on, topdown metrics are either summed or averaged 369 * between the threads of a core: (count_t0 + count_t1). 370 * 371 * For the average case the metric is always scaled to pipeline width, 372 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4) 373 */ 374 375 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots, 376 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */ 377 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */ 378 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2"); 379 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued, 380 "event=0xe,umask=0x1"); /* uops_issued.any */ 381 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired, 382 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */ 383 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles, 384 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */ 385 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles, 386 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */ 387 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */ 388 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale, 389 "4", "2"); 390 391 EVENT_ATTR_STR(slots, slots, "event=0x00,umask=0x4"); 392 EVENT_ATTR_STR(topdown-retiring, td_retiring, "event=0x00,umask=0x80"); 393 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec, "event=0x00,umask=0x81"); 394 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound, "event=0x00,umask=0x82"); 395 EVENT_ATTR_STR(topdown-be-bound, td_be_bound, "event=0x00,umask=0x83"); 396 EVENT_ATTR_STR(topdown-heavy-ops, td_heavy_ops, "event=0x00,umask=0x84"); 397 EVENT_ATTR_STR(topdown-br-mispredict, td_br_mispredict, "event=0x00,umask=0x85"); 398 EVENT_ATTR_STR(topdown-fetch-lat, td_fetch_lat, "event=0x00,umask=0x86"); 399 EVENT_ATTR_STR(topdown-mem-bound, td_mem_bound, "event=0x00,umask=0x87"); 400 401 static struct attribute *snb_events_attrs[] = { 402 EVENT_PTR(td_slots_issued), 403 EVENT_PTR(td_slots_retired), 404 EVENT_PTR(td_fetch_bubbles), 405 EVENT_PTR(td_total_slots), 406 EVENT_PTR(td_total_slots_scale), 407 EVENT_PTR(td_recovery_bubbles), 408 EVENT_PTR(td_recovery_bubbles_scale), 409 NULL, 410 }; 411 412 static struct attribute *snb_mem_events_attrs[] = { 413 EVENT_PTR(mem_ld_snb), 414 EVENT_PTR(mem_st_snb), 415 NULL, 416 }; 417 418 static struct event_constraint intel_hsw_event_constraints[] = { 419 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 420 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 421 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 422 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 423 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 424 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 425 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 426 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), 427 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 428 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), 429 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 430 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), 431 432 /* 433 * When HT is off these events can only run on the bottom 4 counters 434 * When HT is on, they are impacted by the HT bug and require EXCL access 435 */ 436 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 437 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 438 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 439 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 440 441 EVENT_CONSTRAINT_END 442 }; 443 444 static struct event_constraint intel_bdw_event_constraints[] = { 445 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 446 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 447 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 448 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 449 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */ 450 /* 451 * when HT is off, these can only run on the bottom 4 counters 452 */ 453 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 454 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 455 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 456 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 457 EVENT_CONSTRAINT_END 458 }; 459 460 static u64 intel_pmu_event_map(int hw_event) 461 { 462 return intel_perfmon_event_map[hw_event]; 463 } 464 465 static __initconst const u64 spr_hw_cache_event_ids 466 [PERF_COUNT_HW_CACHE_MAX] 467 [PERF_COUNT_HW_CACHE_OP_MAX] 468 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 469 { 470 [ C(L1D ) ] = { 471 [ C(OP_READ) ] = { 472 [ C(RESULT_ACCESS) ] = 0x81d0, 473 [ C(RESULT_MISS) ] = 0xe124, 474 }, 475 [ C(OP_WRITE) ] = { 476 [ C(RESULT_ACCESS) ] = 0x82d0, 477 }, 478 }, 479 [ C(L1I ) ] = { 480 [ C(OP_READ) ] = { 481 [ C(RESULT_MISS) ] = 0xe424, 482 }, 483 [ C(OP_WRITE) ] = { 484 [ C(RESULT_ACCESS) ] = -1, 485 [ C(RESULT_MISS) ] = -1, 486 }, 487 }, 488 [ C(LL ) ] = { 489 [ C(OP_READ) ] = { 490 [ C(RESULT_ACCESS) ] = 0x12a, 491 [ C(RESULT_MISS) ] = 0x12a, 492 }, 493 [ C(OP_WRITE) ] = { 494 [ C(RESULT_ACCESS) ] = 0x12a, 495 [ C(RESULT_MISS) ] = 0x12a, 496 }, 497 }, 498 [ C(DTLB) ] = { 499 [ C(OP_READ) ] = { 500 [ C(RESULT_ACCESS) ] = 0x81d0, 501 [ C(RESULT_MISS) ] = 0xe12, 502 }, 503 [ C(OP_WRITE) ] = { 504 [ C(RESULT_ACCESS) ] = 0x82d0, 505 [ C(RESULT_MISS) ] = 0xe13, 506 }, 507 }, 508 [ C(ITLB) ] = { 509 [ C(OP_READ) ] = { 510 [ C(RESULT_ACCESS) ] = -1, 511 [ C(RESULT_MISS) ] = 0xe11, 512 }, 513 [ C(OP_WRITE) ] = { 514 [ C(RESULT_ACCESS) ] = -1, 515 [ C(RESULT_MISS) ] = -1, 516 }, 517 [ C(OP_PREFETCH) ] = { 518 [ C(RESULT_ACCESS) ] = -1, 519 [ C(RESULT_MISS) ] = -1, 520 }, 521 }, 522 [ C(BPU ) ] = { 523 [ C(OP_READ) ] = { 524 [ C(RESULT_ACCESS) ] = 0x4c4, 525 [ C(RESULT_MISS) ] = 0x4c5, 526 }, 527 [ C(OP_WRITE) ] = { 528 [ C(RESULT_ACCESS) ] = -1, 529 [ C(RESULT_MISS) ] = -1, 530 }, 531 [ C(OP_PREFETCH) ] = { 532 [ C(RESULT_ACCESS) ] = -1, 533 [ C(RESULT_MISS) ] = -1, 534 }, 535 }, 536 [ C(NODE) ] = { 537 [ C(OP_READ) ] = { 538 [ C(RESULT_ACCESS) ] = 0x12a, 539 [ C(RESULT_MISS) ] = 0x12a, 540 }, 541 }, 542 }; 543 544 static __initconst const u64 spr_hw_cache_extra_regs 545 [PERF_COUNT_HW_CACHE_MAX] 546 [PERF_COUNT_HW_CACHE_OP_MAX] 547 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 548 { 549 [ C(LL ) ] = { 550 [ C(OP_READ) ] = { 551 [ C(RESULT_ACCESS) ] = 0x10001, 552 [ C(RESULT_MISS) ] = 0x3fbfc00001, 553 }, 554 [ C(OP_WRITE) ] = { 555 [ C(RESULT_ACCESS) ] = 0x3f3ffc0002, 556 [ C(RESULT_MISS) ] = 0x3f3fc00002, 557 }, 558 }, 559 [ C(NODE) ] = { 560 [ C(OP_READ) ] = { 561 [ C(RESULT_ACCESS) ] = 0x10c000001, 562 [ C(RESULT_MISS) ] = 0x3fb3000001, 563 }, 564 }, 565 }; 566 567 /* 568 * Notes on the events: 569 * - data reads do not include code reads (comparable to earlier tables) 570 * - data counts include speculative execution (except L1 write, dtlb, bpu) 571 * - remote node access includes remote memory, remote cache, remote mmio. 572 * - prefetches are not included in the counts. 573 * - icache miss does not include decoded icache 574 */ 575 576 #define SKL_DEMAND_DATA_RD BIT_ULL(0) 577 #define SKL_DEMAND_RFO BIT_ULL(1) 578 #define SKL_ANY_RESPONSE BIT_ULL(16) 579 #define SKL_SUPPLIER_NONE BIT_ULL(17) 580 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26) 581 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27) 582 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28) 583 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29) 584 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \ 585 SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 586 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 587 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 588 #define SKL_SPL_HIT BIT_ULL(30) 589 #define SKL_SNOOP_NONE BIT_ULL(31) 590 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32) 591 #define SKL_SNOOP_MISS BIT_ULL(33) 592 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34) 593 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35) 594 #define SKL_SNOOP_HITM BIT_ULL(36) 595 #define SKL_SNOOP_NON_DRAM BIT_ULL(37) 596 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \ 597 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 598 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 599 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM) 600 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD 601 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \ 602 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 603 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 604 SKL_SNOOP_HITM|SKL_SPL_HIT) 605 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO 606 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE 607 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 608 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 609 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 610 611 static __initconst const u64 skl_hw_cache_event_ids 612 [PERF_COUNT_HW_CACHE_MAX] 613 [PERF_COUNT_HW_CACHE_OP_MAX] 614 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 615 { 616 [ C(L1D ) ] = { 617 [ C(OP_READ) ] = { 618 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 619 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 620 }, 621 [ C(OP_WRITE) ] = { 622 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 623 [ C(RESULT_MISS) ] = 0x0, 624 }, 625 [ C(OP_PREFETCH) ] = { 626 [ C(RESULT_ACCESS) ] = 0x0, 627 [ C(RESULT_MISS) ] = 0x0, 628 }, 629 }, 630 [ C(L1I ) ] = { 631 [ C(OP_READ) ] = { 632 [ C(RESULT_ACCESS) ] = 0x0, 633 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */ 634 }, 635 [ C(OP_WRITE) ] = { 636 [ C(RESULT_ACCESS) ] = -1, 637 [ C(RESULT_MISS) ] = -1, 638 }, 639 [ C(OP_PREFETCH) ] = { 640 [ C(RESULT_ACCESS) ] = 0x0, 641 [ C(RESULT_MISS) ] = 0x0, 642 }, 643 }, 644 [ C(LL ) ] = { 645 [ C(OP_READ) ] = { 646 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 647 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 648 }, 649 [ C(OP_WRITE) ] = { 650 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 651 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 652 }, 653 [ C(OP_PREFETCH) ] = { 654 [ C(RESULT_ACCESS) ] = 0x0, 655 [ C(RESULT_MISS) ] = 0x0, 656 }, 657 }, 658 [ C(DTLB) ] = { 659 [ C(OP_READ) ] = { 660 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 661 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 662 }, 663 [ C(OP_WRITE) ] = { 664 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 665 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 666 }, 667 [ C(OP_PREFETCH) ] = { 668 [ C(RESULT_ACCESS) ] = 0x0, 669 [ C(RESULT_MISS) ] = 0x0, 670 }, 671 }, 672 [ C(ITLB) ] = { 673 [ C(OP_READ) ] = { 674 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */ 675 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */ 676 }, 677 [ C(OP_WRITE) ] = { 678 [ C(RESULT_ACCESS) ] = -1, 679 [ C(RESULT_MISS) ] = -1, 680 }, 681 [ C(OP_PREFETCH) ] = { 682 [ C(RESULT_ACCESS) ] = -1, 683 [ C(RESULT_MISS) ] = -1, 684 }, 685 }, 686 [ C(BPU ) ] = { 687 [ C(OP_READ) ] = { 688 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 689 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 690 }, 691 [ C(OP_WRITE) ] = { 692 [ C(RESULT_ACCESS) ] = -1, 693 [ C(RESULT_MISS) ] = -1, 694 }, 695 [ C(OP_PREFETCH) ] = { 696 [ C(RESULT_ACCESS) ] = -1, 697 [ C(RESULT_MISS) ] = -1, 698 }, 699 }, 700 [ C(NODE) ] = { 701 [ C(OP_READ) ] = { 702 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 703 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 704 }, 705 [ C(OP_WRITE) ] = { 706 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 707 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 708 }, 709 [ C(OP_PREFETCH) ] = { 710 [ C(RESULT_ACCESS) ] = 0x0, 711 [ C(RESULT_MISS) ] = 0x0, 712 }, 713 }, 714 }; 715 716 static __initconst const u64 skl_hw_cache_extra_regs 717 [PERF_COUNT_HW_CACHE_MAX] 718 [PERF_COUNT_HW_CACHE_OP_MAX] 719 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 720 { 721 [ C(LL ) ] = { 722 [ C(OP_READ) ] = { 723 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 724 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 725 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 726 SKL_L3_MISS|SKL_ANY_SNOOP| 727 SKL_SUPPLIER_NONE, 728 }, 729 [ C(OP_WRITE) ] = { 730 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 731 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 732 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 733 SKL_L3_MISS|SKL_ANY_SNOOP| 734 SKL_SUPPLIER_NONE, 735 }, 736 [ C(OP_PREFETCH) ] = { 737 [ C(RESULT_ACCESS) ] = 0x0, 738 [ C(RESULT_MISS) ] = 0x0, 739 }, 740 }, 741 [ C(NODE) ] = { 742 [ C(OP_READ) ] = { 743 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 744 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 745 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 746 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 747 }, 748 [ C(OP_WRITE) ] = { 749 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 750 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 751 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 752 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 753 }, 754 [ C(OP_PREFETCH) ] = { 755 [ C(RESULT_ACCESS) ] = 0x0, 756 [ C(RESULT_MISS) ] = 0x0, 757 }, 758 }, 759 }; 760 761 #define SNB_DMND_DATA_RD (1ULL << 0) 762 #define SNB_DMND_RFO (1ULL << 1) 763 #define SNB_DMND_IFETCH (1ULL << 2) 764 #define SNB_DMND_WB (1ULL << 3) 765 #define SNB_PF_DATA_RD (1ULL << 4) 766 #define SNB_PF_RFO (1ULL << 5) 767 #define SNB_PF_IFETCH (1ULL << 6) 768 #define SNB_LLC_DATA_RD (1ULL << 7) 769 #define SNB_LLC_RFO (1ULL << 8) 770 #define SNB_LLC_IFETCH (1ULL << 9) 771 #define SNB_BUS_LOCKS (1ULL << 10) 772 #define SNB_STRM_ST (1ULL << 11) 773 #define SNB_OTHER (1ULL << 15) 774 #define SNB_RESP_ANY (1ULL << 16) 775 #define SNB_NO_SUPP (1ULL << 17) 776 #define SNB_LLC_HITM (1ULL << 18) 777 #define SNB_LLC_HITE (1ULL << 19) 778 #define SNB_LLC_HITS (1ULL << 20) 779 #define SNB_LLC_HITF (1ULL << 21) 780 #define SNB_LOCAL (1ULL << 22) 781 #define SNB_REMOTE (0xffULL << 23) 782 #define SNB_SNP_NONE (1ULL << 31) 783 #define SNB_SNP_NOT_NEEDED (1ULL << 32) 784 #define SNB_SNP_MISS (1ULL << 33) 785 #define SNB_NO_FWD (1ULL << 34) 786 #define SNB_SNP_FWD (1ULL << 35) 787 #define SNB_HITM (1ULL << 36) 788 #define SNB_NON_DRAM (1ULL << 37) 789 790 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) 791 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) 792 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 793 794 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ 795 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ 796 SNB_HITM) 797 798 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) 799 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) 800 801 #define SNB_L3_ACCESS SNB_RESP_ANY 802 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) 803 804 static __initconst const u64 snb_hw_cache_extra_regs 805 [PERF_COUNT_HW_CACHE_MAX] 806 [PERF_COUNT_HW_CACHE_OP_MAX] 807 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 808 { 809 [ C(LL ) ] = { 810 [ C(OP_READ) ] = { 811 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, 812 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, 813 }, 814 [ C(OP_WRITE) ] = { 815 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, 816 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, 817 }, 818 [ C(OP_PREFETCH) ] = { 819 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, 820 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, 821 }, 822 }, 823 [ C(NODE) ] = { 824 [ C(OP_READ) ] = { 825 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, 826 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, 827 }, 828 [ C(OP_WRITE) ] = { 829 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, 830 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, 831 }, 832 [ C(OP_PREFETCH) ] = { 833 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, 834 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, 835 }, 836 }, 837 }; 838 839 static __initconst const u64 snb_hw_cache_event_ids 840 [PERF_COUNT_HW_CACHE_MAX] 841 [PERF_COUNT_HW_CACHE_OP_MAX] 842 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 843 { 844 [ C(L1D) ] = { 845 [ C(OP_READ) ] = { 846 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ 847 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ 848 }, 849 [ C(OP_WRITE) ] = { 850 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ 851 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ 852 }, 853 [ C(OP_PREFETCH) ] = { 854 [ C(RESULT_ACCESS) ] = 0x0, 855 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ 856 }, 857 }, 858 [ C(L1I ) ] = { 859 [ C(OP_READ) ] = { 860 [ C(RESULT_ACCESS) ] = 0x0, 861 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ 862 }, 863 [ C(OP_WRITE) ] = { 864 [ C(RESULT_ACCESS) ] = -1, 865 [ C(RESULT_MISS) ] = -1, 866 }, 867 [ C(OP_PREFETCH) ] = { 868 [ C(RESULT_ACCESS) ] = 0x0, 869 [ C(RESULT_MISS) ] = 0x0, 870 }, 871 }, 872 [ C(LL ) ] = { 873 [ C(OP_READ) ] = { 874 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 875 [ C(RESULT_ACCESS) ] = 0x01b7, 876 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 877 [ C(RESULT_MISS) ] = 0x01b7, 878 }, 879 [ C(OP_WRITE) ] = { 880 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 881 [ C(RESULT_ACCESS) ] = 0x01b7, 882 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 883 [ C(RESULT_MISS) ] = 0x01b7, 884 }, 885 [ C(OP_PREFETCH) ] = { 886 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 887 [ C(RESULT_ACCESS) ] = 0x01b7, 888 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 889 [ C(RESULT_MISS) ] = 0x01b7, 890 }, 891 }, 892 [ C(DTLB) ] = { 893 [ C(OP_READ) ] = { 894 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ 895 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ 896 }, 897 [ C(OP_WRITE) ] = { 898 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ 899 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 900 }, 901 [ C(OP_PREFETCH) ] = { 902 [ C(RESULT_ACCESS) ] = 0x0, 903 [ C(RESULT_MISS) ] = 0x0, 904 }, 905 }, 906 [ C(ITLB) ] = { 907 [ C(OP_READ) ] = { 908 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ 909 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ 910 }, 911 [ C(OP_WRITE) ] = { 912 [ C(RESULT_ACCESS) ] = -1, 913 [ C(RESULT_MISS) ] = -1, 914 }, 915 [ C(OP_PREFETCH) ] = { 916 [ C(RESULT_ACCESS) ] = -1, 917 [ C(RESULT_MISS) ] = -1, 918 }, 919 }, 920 [ C(BPU ) ] = { 921 [ C(OP_READ) ] = { 922 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 923 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 924 }, 925 [ C(OP_WRITE) ] = { 926 [ C(RESULT_ACCESS) ] = -1, 927 [ C(RESULT_MISS) ] = -1, 928 }, 929 [ C(OP_PREFETCH) ] = { 930 [ C(RESULT_ACCESS) ] = -1, 931 [ C(RESULT_MISS) ] = -1, 932 }, 933 }, 934 [ C(NODE) ] = { 935 [ C(OP_READ) ] = { 936 [ C(RESULT_ACCESS) ] = 0x01b7, 937 [ C(RESULT_MISS) ] = 0x01b7, 938 }, 939 [ C(OP_WRITE) ] = { 940 [ C(RESULT_ACCESS) ] = 0x01b7, 941 [ C(RESULT_MISS) ] = 0x01b7, 942 }, 943 [ C(OP_PREFETCH) ] = { 944 [ C(RESULT_ACCESS) ] = 0x01b7, 945 [ C(RESULT_MISS) ] = 0x01b7, 946 }, 947 }, 948 949 }; 950 951 /* 952 * Notes on the events: 953 * - data reads do not include code reads (comparable to earlier tables) 954 * - data counts include speculative execution (except L1 write, dtlb, bpu) 955 * - remote node access includes remote memory, remote cache, remote mmio. 956 * - prefetches are not included in the counts because they are not 957 * reliably counted. 958 */ 959 960 #define HSW_DEMAND_DATA_RD BIT_ULL(0) 961 #define HSW_DEMAND_RFO BIT_ULL(1) 962 #define HSW_ANY_RESPONSE BIT_ULL(16) 963 #define HSW_SUPPLIER_NONE BIT_ULL(17) 964 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) 965 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) 966 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) 967 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) 968 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ 969 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 970 HSW_L3_MISS_REMOTE_HOP2P) 971 #define HSW_SNOOP_NONE BIT_ULL(31) 972 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) 973 #define HSW_SNOOP_MISS BIT_ULL(33) 974 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) 975 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) 976 #define HSW_SNOOP_HITM BIT_ULL(36) 977 #define HSW_SNOOP_NON_DRAM BIT_ULL(37) 978 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ 979 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ 980 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ 981 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) 982 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) 983 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD 984 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO 985 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ 986 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) 987 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE 988 989 #define BDW_L3_MISS_LOCAL BIT(26) 990 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ 991 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 992 HSW_L3_MISS_REMOTE_HOP2P) 993 994 995 static __initconst const u64 hsw_hw_cache_event_ids 996 [PERF_COUNT_HW_CACHE_MAX] 997 [PERF_COUNT_HW_CACHE_OP_MAX] 998 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 999 { 1000 [ C(L1D ) ] = { 1001 [ C(OP_READ) ] = { 1002 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1003 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 1004 }, 1005 [ C(OP_WRITE) ] = { 1006 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1007 [ C(RESULT_MISS) ] = 0x0, 1008 }, 1009 [ C(OP_PREFETCH) ] = { 1010 [ C(RESULT_ACCESS) ] = 0x0, 1011 [ C(RESULT_MISS) ] = 0x0, 1012 }, 1013 }, 1014 [ C(L1I ) ] = { 1015 [ C(OP_READ) ] = { 1016 [ C(RESULT_ACCESS) ] = 0x0, 1017 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ 1018 }, 1019 [ C(OP_WRITE) ] = { 1020 [ C(RESULT_ACCESS) ] = -1, 1021 [ C(RESULT_MISS) ] = -1, 1022 }, 1023 [ C(OP_PREFETCH) ] = { 1024 [ C(RESULT_ACCESS) ] = 0x0, 1025 [ C(RESULT_MISS) ] = 0x0, 1026 }, 1027 }, 1028 [ C(LL ) ] = { 1029 [ C(OP_READ) ] = { 1030 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1031 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1032 }, 1033 [ C(OP_WRITE) ] = { 1034 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1035 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1036 }, 1037 [ C(OP_PREFETCH) ] = { 1038 [ C(RESULT_ACCESS) ] = 0x0, 1039 [ C(RESULT_MISS) ] = 0x0, 1040 }, 1041 }, 1042 [ C(DTLB) ] = { 1043 [ C(OP_READ) ] = { 1044 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1045 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ 1046 }, 1047 [ C(OP_WRITE) ] = { 1048 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1049 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 1050 }, 1051 [ C(OP_PREFETCH) ] = { 1052 [ C(RESULT_ACCESS) ] = 0x0, 1053 [ C(RESULT_MISS) ] = 0x0, 1054 }, 1055 }, 1056 [ C(ITLB) ] = { 1057 [ C(OP_READ) ] = { 1058 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ 1059 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ 1060 }, 1061 [ C(OP_WRITE) ] = { 1062 [ C(RESULT_ACCESS) ] = -1, 1063 [ C(RESULT_MISS) ] = -1, 1064 }, 1065 [ C(OP_PREFETCH) ] = { 1066 [ C(RESULT_ACCESS) ] = -1, 1067 [ C(RESULT_MISS) ] = -1, 1068 }, 1069 }, 1070 [ C(BPU ) ] = { 1071 [ C(OP_READ) ] = { 1072 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1073 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1074 }, 1075 [ C(OP_WRITE) ] = { 1076 [ C(RESULT_ACCESS) ] = -1, 1077 [ C(RESULT_MISS) ] = -1, 1078 }, 1079 [ C(OP_PREFETCH) ] = { 1080 [ C(RESULT_ACCESS) ] = -1, 1081 [ C(RESULT_MISS) ] = -1, 1082 }, 1083 }, 1084 [ C(NODE) ] = { 1085 [ C(OP_READ) ] = { 1086 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1087 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1088 }, 1089 [ C(OP_WRITE) ] = { 1090 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1091 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1092 }, 1093 [ C(OP_PREFETCH) ] = { 1094 [ C(RESULT_ACCESS) ] = 0x0, 1095 [ C(RESULT_MISS) ] = 0x0, 1096 }, 1097 }, 1098 }; 1099 1100 static __initconst const u64 hsw_hw_cache_extra_regs 1101 [PERF_COUNT_HW_CACHE_MAX] 1102 [PERF_COUNT_HW_CACHE_OP_MAX] 1103 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1104 { 1105 [ C(LL ) ] = { 1106 [ C(OP_READ) ] = { 1107 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1108 HSW_LLC_ACCESS, 1109 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1110 HSW_L3_MISS|HSW_ANY_SNOOP, 1111 }, 1112 [ C(OP_WRITE) ] = { 1113 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1114 HSW_LLC_ACCESS, 1115 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1116 HSW_L3_MISS|HSW_ANY_SNOOP, 1117 }, 1118 [ C(OP_PREFETCH) ] = { 1119 [ C(RESULT_ACCESS) ] = 0x0, 1120 [ C(RESULT_MISS) ] = 0x0, 1121 }, 1122 }, 1123 [ C(NODE) ] = { 1124 [ C(OP_READ) ] = { 1125 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1126 HSW_L3_MISS_LOCAL_DRAM| 1127 HSW_SNOOP_DRAM, 1128 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1129 HSW_L3_MISS_REMOTE| 1130 HSW_SNOOP_DRAM, 1131 }, 1132 [ C(OP_WRITE) ] = { 1133 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1134 HSW_L3_MISS_LOCAL_DRAM| 1135 HSW_SNOOP_DRAM, 1136 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1137 HSW_L3_MISS_REMOTE| 1138 HSW_SNOOP_DRAM, 1139 }, 1140 [ C(OP_PREFETCH) ] = { 1141 [ C(RESULT_ACCESS) ] = 0x0, 1142 [ C(RESULT_MISS) ] = 0x0, 1143 }, 1144 }, 1145 }; 1146 1147 static __initconst const u64 westmere_hw_cache_event_ids 1148 [PERF_COUNT_HW_CACHE_MAX] 1149 [PERF_COUNT_HW_CACHE_OP_MAX] 1150 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1151 { 1152 [ C(L1D) ] = { 1153 [ C(OP_READ) ] = { 1154 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1155 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1156 }, 1157 [ C(OP_WRITE) ] = { 1158 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1159 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1160 }, 1161 [ C(OP_PREFETCH) ] = { 1162 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1163 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1164 }, 1165 }, 1166 [ C(L1I ) ] = { 1167 [ C(OP_READ) ] = { 1168 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1169 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1170 }, 1171 [ C(OP_WRITE) ] = { 1172 [ C(RESULT_ACCESS) ] = -1, 1173 [ C(RESULT_MISS) ] = -1, 1174 }, 1175 [ C(OP_PREFETCH) ] = { 1176 [ C(RESULT_ACCESS) ] = 0x0, 1177 [ C(RESULT_MISS) ] = 0x0, 1178 }, 1179 }, 1180 [ C(LL ) ] = { 1181 [ C(OP_READ) ] = { 1182 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1183 [ C(RESULT_ACCESS) ] = 0x01b7, 1184 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1185 [ C(RESULT_MISS) ] = 0x01b7, 1186 }, 1187 /* 1188 * Use RFO, not WRITEBACK, because a write miss would typically occur 1189 * on RFO. 1190 */ 1191 [ C(OP_WRITE) ] = { 1192 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1193 [ C(RESULT_ACCESS) ] = 0x01b7, 1194 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1195 [ C(RESULT_MISS) ] = 0x01b7, 1196 }, 1197 [ C(OP_PREFETCH) ] = { 1198 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1199 [ C(RESULT_ACCESS) ] = 0x01b7, 1200 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1201 [ C(RESULT_MISS) ] = 0x01b7, 1202 }, 1203 }, 1204 [ C(DTLB) ] = { 1205 [ C(OP_READ) ] = { 1206 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1207 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1208 }, 1209 [ C(OP_WRITE) ] = { 1210 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1211 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1212 }, 1213 [ C(OP_PREFETCH) ] = { 1214 [ C(RESULT_ACCESS) ] = 0x0, 1215 [ C(RESULT_MISS) ] = 0x0, 1216 }, 1217 }, 1218 [ C(ITLB) ] = { 1219 [ C(OP_READ) ] = { 1220 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1221 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ 1222 }, 1223 [ C(OP_WRITE) ] = { 1224 [ C(RESULT_ACCESS) ] = -1, 1225 [ C(RESULT_MISS) ] = -1, 1226 }, 1227 [ C(OP_PREFETCH) ] = { 1228 [ C(RESULT_ACCESS) ] = -1, 1229 [ C(RESULT_MISS) ] = -1, 1230 }, 1231 }, 1232 [ C(BPU ) ] = { 1233 [ C(OP_READ) ] = { 1234 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1235 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1236 }, 1237 [ C(OP_WRITE) ] = { 1238 [ C(RESULT_ACCESS) ] = -1, 1239 [ C(RESULT_MISS) ] = -1, 1240 }, 1241 [ C(OP_PREFETCH) ] = { 1242 [ C(RESULT_ACCESS) ] = -1, 1243 [ C(RESULT_MISS) ] = -1, 1244 }, 1245 }, 1246 [ C(NODE) ] = { 1247 [ C(OP_READ) ] = { 1248 [ C(RESULT_ACCESS) ] = 0x01b7, 1249 [ C(RESULT_MISS) ] = 0x01b7, 1250 }, 1251 [ C(OP_WRITE) ] = { 1252 [ C(RESULT_ACCESS) ] = 0x01b7, 1253 [ C(RESULT_MISS) ] = 0x01b7, 1254 }, 1255 [ C(OP_PREFETCH) ] = { 1256 [ C(RESULT_ACCESS) ] = 0x01b7, 1257 [ C(RESULT_MISS) ] = 0x01b7, 1258 }, 1259 }, 1260 }; 1261 1262 /* 1263 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; 1264 * See IA32 SDM Vol 3B 30.6.1.3 1265 */ 1266 1267 #define NHM_DMND_DATA_RD (1 << 0) 1268 #define NHM_DMND_RFO (1 << 1) 1269 #define NHM_DMND_IFETCH (1 << 2) 1270 #define NHM_DMND_WB (1 << 3) 1271 #define NHM_PF_DATA_RD (1 << 4) 1272 #define NHM_PF_DATA_RFO (1 << 5) 1273 #define NHM_PF_IFETCH (1 << 6) 1274 #define NHM_OFFCORE_OTHER (1 << 7) 1275 #define NHM_UNCORE_HIT (1 << 8) 1276 #define NHM_OTHER_CORE_HIT_SNP (1 << 9) 1277 #define NHM_OTHER_CORE_HITM (1 << 10) 1278 /* reserved */ 1279 #define NHM_REMOTE_CACHE_FWD (1 << 12) 1280 #define NHM_REMOTE_DRAM (1 << 13) 1281 #define NHM_LOCAL_DRAM (1 << 14) 1282 #define NHM_NON_DRAM (1 << 15) 1283 1284 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) 1285 #define NHM_REMOTE (NHM_REMOTE_DRAM) 1286 1287 #define NHM_DMND_READ (NHM_DMND_DATA_RD) 1288 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) 1289 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) 1290 1291 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) 1292 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) 1293 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) 1294 1295 static __initconst const u64 nehalem_hw_cache_extra_regs 1296 [PERF_COUNT_HW_CACHE_MAX] 1297 [PERF_COUNT_HW_CACHE_OP_MAX] 1298 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1299 { 1300 [ C(LL ) ] = { 1301 [ C(OP_READ) ] = { 1302 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, 1303 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, 1304 }, 1305 [ C(OP_WRITE) ] = { 1306 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, 1307 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, 1308 }, 1309 [ C(OP_PREFETCH) ] = { 1310 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, 1311 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, 1312 }, 1313 }, 1314 [ C(NODE) ] = { 1315 [ C(OP_READ) ] = { 1316 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, 1317 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, 1318 }, 1319 [ C(OP_WRITE) ] = { 1320 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, 1321 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, 1322 }, 1323 [ C(OP_PREFETCH) ] = { 1324 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, 1325 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, 1326 }, 1327 }, 1328 }; 1329 1330 static __initconst const u64 nehalem_hw_cache_event_ids 1331 [PERF_COUNT_HW_CACHE_MAX] 1332 [PERF_COUNT_HW_CACHE_OP_MAX] 1333 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1334 { 1335 [ C(L1D) ] = { 1336 [ C(OP_READ) ] = { 1337 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1338 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1339 }, 1340 [ C(OP_WRITE) ] = { 1341 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1342 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1343 }, 1344 [ C(OP_PREFETCH) ] = { 1345 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1346 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1347 }, 1348 }, 1349 [ C(L1I ) ] = { 1350 [ C(OP_READ) ] = { 1351 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1352 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1353 }, 1354 [ C(OP_WRITE) ] = { 1355 [ C(RESULT_ACCESS) ] = -1, 1356 [ C(RESULT_MISS) ] = -1, 1357 }, 1358 [ C(OP_PREFETCH) ] = { 1359 [ C(RESULT_ACCESS) ] = 0x0, 1360 [ C(RESULT_MISS) ] = 0x0, 1361 }, 1362 }, 1363 [ C(LL ) ] = { 1364 [ C(OP_READ) ] = { 1365 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1366 [ C(RESULT_ACCESS) ] = 0x01b7, 1367 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1368 [ C(RESULT_MISS) ] = 0x01b7, 1369 }, 1370 /* 1371 * Use RFO, not WRITEBACK, because a write miss would typically occur 1372 * on RFO. 1373 */ 1374 [ C(OP_WRITE) ] = { 1375 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1376 [ C(RESULT_ACCESS) ] = 0x01b7, 1377 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1378 [ C(RESULT_MISS) ] = 0x01b7, 1379 }, 1380 [ C(OP_PREFETCH) ] = { 1381 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1382 [ C(RESULT_ACCESS) ] = 0x01b7, 1383 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1384 [ C(RESULT_MISS) ] = 0x01b7, 1385 }, 1386 }, 1387 [ C(DTLB) ] = { 1388 [ C(OP_READ) ] = { 1389 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1390 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1391 }, 1392 [ C(OP_WRITE) ] = { 1393 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1394 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1395 }, 1396 [ C(OP_PREFETCH) ] = { 1397 [ C(RESULT_ACCESS) ] = 0x0, 1398 [ C(RESULT_MISS) ] = 0x0, 1399 }, 1400 }, 1401 [ C(ITLB) ] = { 1402 [ C(OP_READ) ] = { 1403 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1404 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ 1405 }, 1406 [ C(OP_WRITE) ] = { 1407 [ C(RESULT_ACCESS) ] = -1, 1408 [ C(RESULT_MISS) ] = -1, 1409 }, 1410 [ C(OP_PREFETCH) ] = { 1411 [ C(RESULT_ACCESS) ] = -1, 1412 [ C(RESULT_MISS) ] = -1, 1413 }, 1414 }, 1415 [ C(BPU ) ] = { 1416 [ C(OP_READ) ] = { 1417 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1418 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1419 }, 1420 [ C(OP_WRITE) ] = { 1421 [ C(RESULT_ACCESS) ] = -1, 1422 [ C(RESULT_MISS) ] = -1, 1423 }, 1424 [ C(OP_PREFETCH) ] = { 1425 [ C(RESULT_ACCESS) ] = -1, 1426 [ C(RESULT_MISS) ] = -1, 1427 }, 1428 }, 1429 [ C(NODE) ] = { 1430 [ C(OP_READ) ] = { 1431 [ C(RESULT_ACCESS) ] = 0x01b7, 1432 [ C(RESULT_MISS) ] = 0x01b7, 1433 }, 1434 [ C(OP_WRITE) ] = { 1435 [ C(RESULT_ACCESS) ] = 0x01b7, 1436 [ C(RESULT_MISS) ] = 0x01b7, 1437 }, 1438 [ C(OP_PREFETCH) ] = { 1439 [ C(RESULT_ACCESS) ] = 0x01b7, 1440 [ C(RESULT_MISS) ] = 0x01b7, 1441 }, 1442 }, 1443 }; 1444 1445 static __initconst const u64 core2_hw_cache_event_ids 1446 [PERF_COUNT_HW_CACHE_MAX] 1447 [PERF_COUNT_HW_CACHE_OP_MAX] 1448 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1449 { 1450 [ C(L1D) ] = { 1451 [ C(OP_READ) ] = { 1452 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ 1453 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ 1454 }, 1455 [ C(OP_WRITE) ] = { 1456 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ 1457 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ 1458 }, 1459 [ C(OP_PREFETCH) ] = { 1460 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ 1461 [ C(RESULT_MISS) ] = 0, 1462 }, 1463 }, 1464 [ C(L1I ) ] = { 1465 [ C(OP_READ) ] = { 1466 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ 1467 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ 1468 }, 1469 [ C(OP_WRITE) ] = { 1470 [ C(RESULT_ACCESS) ] = -1, 1471 [ C(RESULT_MISS) ] = -1, 1472 }, 1473 [ C(OP_PREFETCH) ] = { 1474 [ C(RESULT_ACCESS) ] = 0, 1475 [ C(RESULT_MISS) ] = 0, 1476 }, 1477 }, 1478 [ C(LL ) ] = { 1479 [ C(OP_READ) ] = { 1480 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1481 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1482 }, 1483 [ C(OP_WRITE) ] = { 1484 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1485 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1486 }, 1487 [ C(OP_PREFETCH) ] = { 1488 [ C(RESULT_ACCESS) ] = 0, 1489 [ C(RESULT_MISS) ] = 0, 1490 }, 1491 }, 1492 [ C(DTLB) ] = { 1493 [ C(OP_READ) ] = { 1494 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1495 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ 1496 }, 1497 [ C(OP_WRITE) ] = { 1498 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1499 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ 1500 }, 1501 [ C(OP_PREFETCH) ] = { 1502 [ C(RESULT_ACCESS) ] = 0, 1503 [ C(RESULT_MISS) ] = 0, 1504 }, 1505 }, 1506 [ C(ITLB) ] = { 1507 [ C(OP_READ) ] = { 1508 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1509 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ 1510 }, 1511 [ C(OP_WRITE) ] = { 1512 [ C(RESULT_ACCESS) ] = -1, 1513 [ C(RESULT_MISS) ] = -1, 1514 }, 1515 [ C(OP_PREFETCH) ] = { 1516 [ C(RESULT_ACCESS) ] = -1, 1517 [ C(RESULT_MISS) ] = -1, 1518 }, 1519 }, 1520 [ C(BPU ) ] = { 1521 [ C(OP_READ) ] = { 1522 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1523 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1524 }, 1525 [ C(OP_WRITE) ] = { 1526 [ C(RESULT_ACCESS) ] = -1, 1527 [ C(RESULT_MISS) ] = -1, 1528 }, 1529 [ C(OP_PREFETCH) ] = { 1530 [ C(RESULT_ACCESS) ] = -1, 1531 [ C(RESULT_MISS) ] = -1, 1532 }, 1533 }, 1534 }; 1535 1536 static __initconst const u64 atom_hw_cache_event_ids 1537 [PERF_COUNT_HW_CACHE_MAX] 1538 [PERF_COUNT_HW_CACHE_OP_MAX] 1539 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1540 { 1541 [ C(L1D) ] = { 1542 [ C(OP_READ) ] = { 1543 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ 1544 [ C(RESULT_MISS) ] = 0, 1545 }, 1546 [ C(OP_WRITE) ] = { 1547 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ 1548 [ C(RESULT_MISS) ] = 0, 1549 }, 1550 [ C(OP_PREFETCH) ] = { 1551 [ C(RESULT_ACCESS) ] = 0x0, 1552 [ C(RESULT_MISS) ] = 0, 1553 }, 1554 }, 1555 [ C(L1I ) ] = { 1556 [ C(OP_READ) ] = { 1557 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1558 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1559 }, 1560 [ C(OP_WRITE) ] = { 1561 [ C(RESULT_ACCESS) ] = -1, 1562 [ C(RESULT_MISS) ] = -1, 1563 }, 1564 [ C(OP_PREFETCH) ] = { 1565 [ C(RESULT_ACCESS) ] = 0, 1566 [ C(RESULT_MISS) ] = 0, 1567 }, 1568 }, 1569 [ C(LL ) ] = { 1570 [ C(OP_READ) ] = { 1571 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1572 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1573 }, 1574 [ C(OP_WRITE) ] = { 1575 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1576 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1577 }, 1578 [ C(OP_PREFETCH) ] = { 1579 [ C(RESULT_ACCESS) ] = 0, 1580 [ C(RESULT_MISS) ] = 0, 1581 }, 1582 }, 1583 [ C(DTLB) ] = { 1584 [ C(OP_READ) ] = { 1585 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ 1586 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ 1587 }, 1588 [ C(OP_WRITE) ] = { 1589 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ 1590 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ 1591 }, 1592 [ C(OP_PREFETCH) ] = { 1593 [ C(RESULT_ACCESS) ] = 0, 1594 [ C(RESULT_MISS) ] = 0, 1595 }, 1596 }, 1597 [ C(ITLB) ] = { 1598 [ C(OP_READ) ] = { 1599 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1600 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ 1601 }, 1602 [ C(OP_WRITE) ] = { 1603 [ C(RESULT_ACCESS) ] = -1, 1604 [ C(RESULT_MISS) ] = -1, 1605 }, 1606 [ C(OP_PREFETCH) ] = { 1607 [ C(RESULT_ACCESS) ] = -1, 1608 [ C(RESULT_MISS) ] = -1, 1609 }, 1610 }, 1611 [ C(BPU ) ] = { 1612 [ C(OP_READ) ] = { 1613 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1614 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1615 }, 1616 [ C(OP_WRITE) ] = { 1617 [ C(RESULT_ACCESS) ] = -1, 1618 [ C(RESULT_MISS) ] = -1, 1619 }, 1620 [ C(OP_PREFETCH) ] = { 1621 [ C(RESULT_ACCESS) ] = -1, 1622 [ C(RESULT_MISS) ] = -1, 1623 }, 1624 }, 1625 }; 1626 1627 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c"); 1628 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2"); 1629 /* no_alloc_cycles.not_delivered */ 1630 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm, 1631 "event=0xca,umask=0x50"); 1632 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2"); 1633 /* uops_retired.all */ 1634 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm, 1635 "event=0xc2,umask=0x10"); 1636 /* uops_retired.all */ 1637 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm, 1638 "event=0xc2,umask=0x10"); 1639 1640 static struct attribute *slm_events_attrs[] = { 1641 EVENT_PTR(td_total_slots_slm), 1642 EVENT_PTR(td_total_slots_scale_slm), 1643 EVENT_PTR(td_fetch_bubbles_slm), 1644 EVENT_PTR(td_fetch_bubbles_scale_slm), 1645 EVENT_PTR(td_slots_issued_slm), 1646 EVENT_PTR(td_slots_retired_slm), 1647 NULL 1648 }; 1649 1650 static struct extra_reg intel_slm_extra_regs[] __read_mostly = 1651 { 1652 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1653 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), 1654 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1), 1655 EVENT_EXTRA_END 1656 }; 1657 1658 #define SLM_DMND_READ SNB_DMND_DATA_RD 1659 #define SLM_DMND_WRITE SNB_DMND_RFO 1660 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1661 1662 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) 1663 #define SLM_LLC_ACCESS SNB_RESP_ANY 1664 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) 1665 1666 static __initconst const u64 slm_hw_cache_extra_regs 1667 [PERF_COUNT_HW_CACHE_MAX] 1668 [PERF_COUNT_HW_CACHE_OP_MAX] 1669 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1670 { 1671 [ C(LL ) ] = { 1672 [ C(OP_READ) ] = { 1673 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, 1674 [ C(RESULT_MISS) ] = 0, 1675 }, 1676 [ C(OP_WRITE) ] = { 1677 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, 1678 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, 1679 }, 1680 [ C(OP_PREFETCH) ] = { 1681 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, 1682 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, 1683 }, 1684 }, 1685 }; 1686 1687 static __initconst const u64 slm_hw_cache_event_ids 1688 [PERF_COUNT_HW_CACHE_MAX] 1689 [PERF_COUNT_HW_CACHE_OP_MAX] 1690 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1691 { 1692 [ C(L1D) ] = { 1693 [ C(OP_READ) ] = { 1694 [ C(RESULT_ACCESS) ] = 0, 1695 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ 1696 }, 1697 [ C(OP_WRITE) ] = { 1698 [ C(RESULT_ACCESS) ] = 0, 1699 [ C(RESULT_MISS) ] = 0, 1700 }, 1701 [ C(OP_PREFETCH) ] = { 1702 [ C(RESULT_ACCESS) ] = 0, 1703 [ C(RESULT_MISS) ] = 0, 1704 }, 1705 }, 1706 [ C(L1I ) ] = { 1707 [ C(OP_READ) ] = { 1708 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ 1709 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ 1710 }, 1711 [ C(OP_WRITE) ] = { 1712 [ C(RESULT_ACCESS) ] = -1, 1713 [ C(RESULT_MISS) ] = -1, 1714 }, 1715 [ C(OP_PREFETCH) ] = { 1716 [ C(RESULT_ACCESS) ] = 0, 1717 [ C(RESULT_MISS) ] = 0, 1718 }, 1719 }, 1720 [ C(LL ) ] = { 1721 [ C(OP_READ) ] = { 1722 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1723 [ C(RESULT_ACCESS) ] = 0x01b7, 1724 [ C(RESULT_MISS) ] = 0, 1725 }, 1726 [ C(OP_WRITE) ] = { 1727 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1728 [ C(RESULT_ACCESS) ] = 0x01b7, 1729 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1730 [ C(RESULT_MISS) ] = 0x01b7, 1731 }, 1732 [ C(OP_PREFETCH) ] = { 1733 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1734 [ C(RESULT_ACCESS) ] = 0x01b7, 1735 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1736 [ C(RESULT_MISS) ] = 0x01b7, 1737 }, 1738 }, 1739 [ C(DTLB) ] = { 1740 [ C(OP_READ) ] = { 1741 [ C(RESULT_ACCESS) ] = 0, 1742 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ 1743 }, 1744 [ C(OP_WRITE) ] = { 1745 [ C(RESULT_ACCESS) ] = 0, 1746 [ C(RESULT_MISS) ] = 0, 1747 }, 1748 [ C(OP_PREFETCH) ] = { 1749 [ C(RESULT_ACCESS) ] = 0, 1750 [ C(RESULT_MISS) ] = 0, 1751 }, 1752 }, 1753 [ C(ITLB) ] = { 1754 [ C(OP_READ) ] = { 1755 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1756 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */ 1757 }, 1758 [ C(OP_WRITE) ] = { 1759 [ C(RESULT_ACCESS) ] = -1, 1760 [ C(RESULT_MISS) ] = -1, 1761 }, 1762 [ C(OP_PREFETCH) ] = { 1763 [ C(RESULT_ACCESS) ] = -1, 1764 [ C(RESULT_MISS) ] = -1, 1765 }, 1766 }, 1767 [ C(BPU ) ] = { 1768 [ C(OP_READ) ] = { 1769 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1770 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1771 }, 1772 [ C(OP_WRITE) ] = { 1773 [ C(RESULT_ACCESS) ] = -1, 1774 [ C(RESULT_MISS) ] = -1, 1775 }, 1776 [ C(OP_PREFETCH) ] = { 1777 [ C(RESULT_ACCESS) ] = -1, 1778 [ C(RESULT_MISS) ] = -1, 1779 }, 1780 }, 1781 }; 1782 1783 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c"); 1784 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3"); 1785 /* UOPS_NOT_DELIVERED.ANY */ 1786 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c"); 1787 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */ 1788 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02"); 1789 /* UOPS_RETIRED.ANY */ 1790 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2"); 1791 /* UOPS_ISSUED.ANY */ 1792 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e"); 1793 1794 static struct attribute *glm_events_attrs[] = { 1795 EVENT_PTR(td_total_slots_glm), 1796 EVENT_PTR(td_total_slots_scale_glm), 1797 EVENT_PTR(td_fetch_bubbles_glm), 1798 EVENT_PTR(td_recovery_bubbles_glm), 1799 EVENT_PTR(td_slots_issued_glm), 1800 EVENT_PTR(td_slots_retired_glm), 1801 NULL 1802 }; 1803 1804 static struct extra_reg intel_glm_extra_regs[] __read_mostly = { 1805 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1806 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0), 1807 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1), 1808 EVENT_EXTRA_END 1809 }; 1810 1811 #define GLM_DEMAND_DATA_RD BIT_ULL(0) 1812 #define GLM_DEMAND_RFO BIT_ULL(1) 1813 #define GLM_ANY_RESPONSE BIT_ULL(16) 1814 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33) 1815 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD 1816 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO 1817 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1818 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE 1819 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM) 1820 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM) 1821 1822 static __initconst const u64 glm_hw_cache_event_ids 1823 [PERF_COUNT_HW_CACHE_MAX] 1824 [PERF_COUNT_HW_CACHE_OP_MAX] 1825 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1826 [C(L1D)] = { 1827 [C(OP_READ)] = { 1828 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1829 [C(RESULT_MISS)] = 0x0, 1830 }, 1831 [C(OP_WRITE)] = { 1832 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1833 [C(RESULT_MISS)] = 0x0, 1834 }, 1835 [C(OP_PREFETCH)] = { 1836 [C(RESULT_ACCESS)] = 0x0, 1837 [C(RESULT_MISS)] = 0x0, 1838 }, 1839 }, 1840 [C(L1I)] = { 1841 [C(OP_READ)] = { 1842 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1843 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1844 }, 1845 [C(OP_WRITE)] = { 1846 [C(RESULT_ACCESS)] = -1, 1847 [C(RESULT_MISS)] = -1, 1848 }, 1849 [C(OP_PREFETCH)] = { 1850 [C(RESULT_ACCESS)] = 0x0, 1851 [C(RESULT_MISS)] = 0x0, 1852 }, 1853 }, 1854 [C(LL)] = { 1855 [C(OP_READ)] = { 1856 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1857 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1858 }, 1859 [C(OP_WRITE)] = { 1860 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1861 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1862 }, 1863 [C(OP_PREFETCH)] = { 1864 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1865 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1866 }, 1867 }, 1868 [C(DTLB)] = { 1869 [C(OP_READ)] = { 1870 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1871 [C(RESULT_MISS)] = 0x0, 1872 }, 1873 [C(OP_WRITE)] = { 1874 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1875 [C(RESULT_MISS)] = 0x0, 1876 }, 1877 [C(OP_PREFETCH)] = { 1878 [C(RESULT_ACCESS)] = 0x0, 1879 [C(RESULT_MISS)] = 0x0, 1880 }, 1881 }, 1882 [C(ITLB)] = { 1883 [C(OP_READ)] = { 1884 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 1885 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 1886 }, 1887 [C(OP_WRITE)] = { 1888 [C(RESULT_ACCESS)] = -1, 1889 [C(RESULT_MISS)] = -1, 1890 }, 1891 [C(OP_PREFETCH)] = { 1892 [C(RESULT_ACCESS)] = -1, 1893 [C(RESULT_MISS)] = -1, 1894 }, 1895 }, 1896 [C(BPU)] = { 1897 [C(OP_READ)] = { 1898 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1899 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1900 }, 1901 [C(OP_WRITE)] = { 1902 [C(RESULT_ACCESS)] = -1, 1903 [C(RESULT_MISS)] = -1, 1904 }, 1905 [C(OP_PREFETCH)] = { 1906 [C(RESULT_ACCESS)] = -1, 1907 [C(RESULT_MISS)] = -1, 1908 }, 1909 }, 1910 }; 1911 1912 static __initconst const u64 glm_hw_cache_extra_regs 1913 [PERF_COUNT_HW_CACHE_MAX] 1914 [PERF_COUNT_HW_CACHE_OP_MAX] 1915 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1916 [C(LL)] = { 1917 [C(OP_READ)] = { 1918 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 1919 GLM_LLC_ACCESS, 1920 [C(RESULT_MISS)] = GLM_DEMAND_READ| 1921 GLM_LLC_MISS, 1922 }, 1923 [C(OP_WRITE)] = { 1924 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 1925 GLM_LLC_ACCESS, 1926 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 1927 GLM_LLC_MISS, 1928 }, 1929 [C(OP_PREFETCH)] = { 1930 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH| 1931 GLM_LLC_ACCESS, 1932 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH| 1933 GLM_LLC_MISS, 1934 }, 1935 }, 1936 }; 1937 1938 static __initconst const u64 glp_hw_cache_event_ids 1939 [PERF_COUNT_HW_CACHE_MAX] 1940 [PERF_COUNT_HW_CACHE_OP_MAX] 1941 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1942 [C(L1D)] = { 1943 [C(OP_READ)] = { 1944 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1945 [C(RESULT_MISS)] = 0x0, 1946 }, 1947 [C(OP_WRITE)] = { 1948 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1949 [C(RESULT_MISS)] = 0x0, 1950 }, 1951 [C(OP_PREFETCH)] = { 1952 [C(RESULT_ACCESS)] = 0x0, 1953 [C(RESULT_MISS)] = 0x0, 1954 }, 1955 }, 1956 [C(L1I)] = { 1957 [C(OP_READ)] = { 1958 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1959 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1960 }, 1961 [C(OP_WRITE)] = { 1962 [C(RESULT_ACCESS)] = -1, 1963 [C(RESULT_MISS)] = -1, 1964 }, 1965 [C(OP_PREFETCH)] = { 1966 [C(RESULT_ACCESS)] = 0x0, 1967 [C(RESULT_MISS)] = 0x0, 1968 }, 1969 }, 1970 [C(LL)] = { 1971 [C(OP_READ)] = { 1972 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1973 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1974 }, 1975 [C(OP_WRITE)] = { 1976 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1977 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1978 }, 1979 [C(OP_PREFETCH)] = { 1980 [C(RESULT_ACCESS)] = 0x0, 1981 [C(RESULT_MISS)] = 0x0, 1982 }, 1983 }, 1984 [C(DTLB)] = { 1985 [C(OP_READ)] = { 1986 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1987 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 1988 }, 1989 [C(OP_WRITE)] = { 1990 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1991 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 1992 }, 1993 [C(OP_PREFETCH)] = { 1994 [C(RESULT_ACCESS)] = 0x0, 1995 [C(RESULT_MISS)] = 0x0, 1996 }, 1997 }, 1998 [C(ITLB)] = { 1999 [C(OP_READ)] = { 2000 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 2001 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 2002 }, 2003 [C(OP_WRITE)] = { 2004 [C(RESULT_ACCESS)] = -1, 2005 [C(RESULT_MISS)] = -1, 2006 }, 2007 [C(OP_PREFETCH)] = { 2008 [C(RESULT_ACCESS)] = -1, 2009 [C(RESULT_MISS)] = -1, 2010 }, 2011 }, 2012 [C(BPU)] = { 2013 [C(OP_READ)] = { 2014 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 2015 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 2016 }, 2017 [C(OP_WRITE)] = { 2018 [C(RESULT_ACCESS)] = -1, 2019 [C(RESULT_MISS)] = -1, 2020 }, 2021 [C(OP_PREFETCH)] = { 2022 [C(RESULT_ACCESS)] = -1, 2023 [C(RESULT_MISS)] = -1, 2024 }, 2025 }, 2026 }; 2027 2028 static __initconst const u64 glp_hw_cache_extra_regs 2029 [PERF_COUNT_HW_CACHE_MAX] 2030 [PERF_COUNT_HW_CACHE_OP_MAX] 2031 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2032 [C(LL)] = { 2033 [C(OP_READ)] = { 2034 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 2035 GLM_LLC_ACCESS, 2036 [C(RESULT_MISS)] = GLM_DEMAND_READ| 2037 GLM_LLC_MISS, 2038 }, 2039 [C(OP_WRITE)] = { 2040 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 2041 GLM_LLC_ACCESS, 2042 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 2043 GLM_LLC_MISS, 2044 }, 2045 [C(OP_PREFETCH)] = { 2046 [C(RESULT_ACCESS)] = 0x0, 2047 [C(RESULT_MISS)] = 0x0, 2048 }, 2049 }, 2050 }; 2051 2052 #define TNT_LOCAL_DRAM BIT_ULL(26) 2053 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD 2054 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO 2055 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE 2056 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \ 2057 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM) 2058 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM) 2059 2060 static __initconst const u64 tnt_hw_cache_extra_regs 2061 [PERF_COUNT_HW_CACHE_MAX] 2062 [PERF_COUNT_HW_CACHE_OP_MAX] 2063 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2064 [C(LL)] = { 2065 [C(OP_READ)] = { 2066 [C(RESULT_ACCESS)] = TNT_DEMAND_READ| 2067 TNT_LLC_ACCESS, 2068 [C(RESULT_MISS)] = TNT_DEMAND_READ| 2069 TNT_LLC_MISS, 2070 }, 2071 [C(OP_WRITE)] = { 2072 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE| 2073 TNT_LLC_ACCESS, 2074 [C(RESULT_MISS)] = TNT_DEMAND_WRITE| 2075 TNT_LLC_MISS, 2076 }, 2077 [C(OP_PREFETCH)] = { 2078 [C(RESULT_ACCESS)] = 0x0, 2079 [C(RESULT_MISS)] = 0x0, 2080 }, 2081 }, 2082 }; 2083 2084 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound_tnt, "event=0x71,umask=0x0"); 2085 EVENT_ATTR_STR(topdown-retiring, td_retiring_tnt, "event=0xc2,umask=0x0"); 2086 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec_tnt, "event=0x73,umask=0x6"); 2087 EVENT_ATTR_STR(topdown-be-bound, td_be_bound_tnt, "event=0x74,umask=0x0"); 2088 2089 static struct attribute *tnt_events_attrs[] = { 2090 EVENT_PTR(td_fe_bound_tnt), 2091 EVENT_PTR(td_retiring_tnt), 2092 EVENT_PTR(td_bad_spec_tnt), 2093 EVENT_PTR(td_be_bound_tnt), 2094 NULL, 2095 }; 2096 2097 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = { 2098 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2099 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0), 2100 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1), 2101 EVENT_EXTRA_END 2102 }; 2103 2104 static struct extra_reg intel_grt_extra_regs[] __read_mostly = { 2105 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2106 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 2107 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 2108 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0), 2109 EVENT_EXTRA_END 2110 }; 2111 2112 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */ 2113 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */ 2114 #define KNL_MCDRAM_LOCAL BIT_ULL(21) 2115 #define KNL_MCDRAM_FAR BIT_ULL(22) 2116 #define KNL_DDR_LOCAL BIT_ULL(23) 2117 #define KNL_DDR_FAR BIT_ULL(24) 2118 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \ 2119 KNL_DDR_LOCAL | KNL_DDR_FAR) 2120 #define KNL_L2_READ SLM_DMND_READ 2121 #define KNL_L2_WRITE SLM_DMND_WRITE 2122 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH 2123 #define KNL_L2_ACCESS SLM_LLC_ACCESS 2124 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \ 2125 KNL_DRAM_ANY | SNB_SNP_ANY | \ 2126 SNB_NON_DRAM) 2127 2128 static __initconst const u64 knl_hw_cache_extra_regs 2129 [PERF_COUNT_HW_CACHE_MAX] 2130 [PERF_COUNT_HW_CACHE_OP_MAX] 2131 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2132 [C(LL)] = { 2133 [C(OP_READ)] = { 2134 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS, 2135 [C(RESULT_MISS)] = 0, 2136 }, 2137 [C(OP_WRITE)] = { 2138 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS, 2139 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS, 2140 }, 2141 [C(OP_PREFETCH)] = { 2142 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS, 2143 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS, 2144 }, 2145 }, 2146 }; 2147 2148 /* 2149 * Used from PMIs where the LBRs are already disabled. 2150 * 2151 * This function could be called consecutively. It is required to remain in 2152 * disabled state if called consecutively. 2153 * 2154 * During consecutive calls, the same disable value will be written to related 2155 * registers, so the PMU state remains unchanged. 2156 * 2157 * intel_bts events don't coexist with intel PMU's BTS events because of 2158 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them 2159 * disabled around intel PMU's event batching etc, only inside the PMI handler. 2160 * 2161 * Avoid PEBS_ENABLE MSR access in PMIs. 2162 * The GLOBAL_CTRL has been disabled. All the counters do not count anymore. 2163 * It doesn't matter if the PEBS is enabled or not. 2164 * Usually, the PEBS status are not changed in PMIs. It's unnecessary to 2165 * access PEBS_ENABLE MSR in disable_all()/enable_all(). 2166 * However, there are some cases which may change PEBS status, e.g. PMI 2167 * throttle. The PEBS_ENABLE should be updated where the status changes. 2168 */ 2169 static __always_inline void __intel_pmu_disable_all(bool bts) 2170 { 2171 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2172 2173 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2174 2175 if (bts && test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) 2176 intel_pmu_disable_bts(); 2177 } 2178 2179 static __always_inline void intel_pmu_disable_all(void) 2180 { 2181 __intel_pmu_disable_all(true); 2182 intel_pmu_pebs_disable_all(); 2183 intel_pmu_lbr_disable_all(); 2184 } 2185 2186 static void __intel_pmu_enable_all(int added, bool pmi) 2187 { 2188 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2189 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2190 2191 intel_pmu_lbr_enable_all(pmi); 2192 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 2193 intel_ctrl & ~cpuc->intel_ctrl_guest_mask); 2194 2195 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 2196 struct perf_event *event = 2197 cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; 2198 2199 if (WARN_ON_ONCE(!event)) 2200 return; 2201 2202 intel_pmu_enable_bts(event->hw.config); 2203 } 2204 } 2205 2206 static void intel_pmu_enable_all(int added) 2207 { 2208 intel_pmu_pebs_enable_all(); 2209 __intel_pmu_enable_all(added, false); 2210 } 2211 2212 static noinline int 2213 __intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, 2214 unsigned int cnt, unsigned long flags) 2215 { 2216 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2217 2218 intel_pmu_lbr_read(); 2219 cnt = min_t(unsigned int, cnt, x86_pmu.lbr_nr); 2220 2221 memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt); 2222 intel_pmu_enable_all(0); 2223 local_irq_restore(flags); 2224 return cnt; 2225 } 2226 2227 static int 2228 intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2229 { 2230 unsigned long flags; 2231 2232 /* must not have branches... */ 2233 local_irq_save(flags); 2234 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2235 __intel_pmu_lbr_disable(); 2236 /* ... until here */ 2237 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2238 } 2239 2240 static int 2241 intel_pmu_snapshot_arch_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2242 { 2243 unsigned long flags; 2244 2245 /* must not have branches... */ 2246 local_irq_save(flags); 2247 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2248 __intel_pmu_arch_lbr_disable(); 2249 /* ... until here */ 2250 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2251 } 2252 2253 /* 2254 * Workaround for: 2255 * Intel Errata AAK100 (model 26) 2256 * Intel Errata AAP53 (model 30) 2257 * Intel Errata BD53 (model 44) 2258 * 2259 * The official story: 2260 * These chips need to be 'reset' when adding counters by programming the 2261 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either 2262 * in sequence on the same PMC or on different PMCs. 2263 * 2264 * In practice it appears some of these events do in fact count, and 2265 * we need to program all 4 events. 2266 */ 2267 static void intel_pmu_nhm_workaround(void) 2268 { 2269 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2270 static const unsigned long nhm_magic[4] = { 2271 0x4300B5, 2272 0x4300D2, 2273 0x4300B1, 2274 0x4300B1 2275 }; 2276 struct perf_event *event; 2277 int i; 2278 2279 /* 2280 * The Errata requires below steps: 2281 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; 2282 * 2) Configure 4 PERFEVTSELx with the magic events and clear 2283 * the corresponding PMCx; 2284 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; 2285 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; 2286 * 5) Clear 4 pairs of ERFEVTSELx and PMCx; 2287 */ 2288 2289 /* 2290 * The real steps we choose are a little different from above. 2291 * A) To reduce MSR operations, we don't run step 1) as they 2292 * are already cleared before this function is called; 2293 * B) Call x86_perf_event_update to save PMCx before configuring 2294 * PERFEVTSELx with magic number; 2295 * C) With step 5), we do clear only when the PERFEVTSELx is 2296 * not used currently. 2297 * D) Call x86_perf_event_set_period to restore PMCx; 2298 */ 2299 2300 /* We always operate 4 pairs of PERF Counters */ 2301 for (i = 0; i < 4; i++) { 2302 event = cpuc->events[i]; 2303 if (event) 2304 x86_perf_event_update(event); 2305 } 2306 2307 for (i = 0; i < 4; i++) { 2308 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); 2309 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); 2310 } 2311 2312 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); 2313 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); 2314 2315 for (i = 0; i < 4; i++) { 2316 event = cpuc->events[i]; 2317 2318 if (event) { 2319 x86_perf_event_set_period(event); 2320 __x86_pmu_enable_event(&event->hw, 2321 ARCH_PERFMON_EVENTSEL_ENABLE); 2322 } else 2323 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); 2324 } 2325 } 2326 2327 static void intel_pmu_nhm_enable_all(int added) 2328 { 2329 if (added) 2330 intel_pmu_nhm_workaround(); 2331 intel_pmu_enable_all(added); 2332 } 2333 2334 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on) 2335 { 2336 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0; 2337 2338 if (cpuc->tfa_shadow != val) { 2339 cpuc->tfa_shadow = val; 2340 wrmsrl(MSR_TSX_FORCE_ABORT, val); 2341 } 2342 } 2343 2344 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2345 { 2346 /* 2347 * We're going to use PMC3, make sure TFA is set before we touch it. 2348 */ 2349 if (cntr == 3) 2350 intel_set_tfa(cpuc, true); 2351 } 2352 2353 static void intel_tfa_pmu_enable_all(int added) 2354 { 2355 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2356 2357 /* 2358 * If we find PMC3 is no longer used when we enable the PMU, we can 2359 * clear TFA. 2360 */ 2361 if (!test_bit(3, cpuc->active_mask)) 2362 intel_set_tfa(cpuc, false); 2363 2364 intel_pmu_enable_all(added); 2365 } 2366 2367 static inline u64 intel_pmu_get_status(void) 2368 { 2369 u64 status; 2370 2371 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); 2372 2373 return status; 2374 } 2375 2376 static inline void intel_pmu_ack_status(u64 ack) 2377 { 2378 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); 2379 } 2380 2381 static inline bool event_is_checkpointed(struct perf_event *event) 2382 { 2383 return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; 2384 } 2385 2386 static inline void intel_set_masks(struct perf_event *event, int idx) 2387 { 2388 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2389 2390 if (event->attr.exclude_host) 2391 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2392 if (event->attr.exclude_guest) 2393 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2394 if (event_is_checkpointed(event)) 2395 __set_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2396 } 2397 2398 static inline void intel_clear_masks(struct perf_event *event, int idx) 2399 { 2400 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2401 2402 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2403 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2404 __clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2405 } 2406 2407 static void intel_pmu_disable_fixed(struct perf_event *event) 2408 { 2409 struct hw_perf_event *hwc = &event->hw; 2410 u64 ctrl_val, mask; 2411 int idx = hwc->idx; 2412 2413 if (is_topdown_idx(idx)) { 2414 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2415 2416 /* 2417 * When there are other active TopDown events, 2418 * don't disable the fixed counter 3. 2419 */ 2420 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2421 return; 2422 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2423 } 2424 2425 intel_clear_masks(event, idx); 2426 2427 mask = 0xfULL << ((idx - INTEL_PMC_IDX_FIXED) * 4); 2428 rdmsrl(hwc->config_base, ctrl_val); 2429 ctrl_val &= ~mask; 2430 wrmsrl(hwc->config_base, ctrl_val); 2431 } 2432 2433 static void intel_pmu_disable_event(struct perf_event *event) 2434 { 2435 struct hw_perf_event *hwc = &event->hw; 2436 int idx = hwc->idx; 2437 2438 switch (idx) { 2439 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2440 intel_clear_masks(event, idx); 2441 x86_pmu_disable_event(event); 2442 break; 2443 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2444 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2445 intel_pmu_disable_fixed(event); 2446 break; 2447 case INTEL_PMC_IDX_FIXED_BTS: 2448 intel_pmu_disable_bts(); 2449 intel_pmu_drain_bts_buffer(); 2450 return; 2451 case INTEL_PMC_IDX_FIXED_VLBR: 2452 intel_clear_masks(event, idx); 2453 break; 2454 default: 2455 intel_clear_masks(event, idx); 2456 pr_warn("Failed to disable the event with invalid index %d\n", 2457 idx); 2458 return; 2459 } 2460 2461 /* 2462 * Needs to be called after x86_pmu_disable_event, 2463 * so we don't trigger the event without PEBS bit set. 2464 */ 2465 if (unlikely(event->attr.precise_ip)) 2466 intel_pmu_pebs_disable(event); 2467 } 2468 2469 static void intel_pmu_assign_event(struct perf_event *event, int idx) 2470 { 2471 if (is_pebs_pt(event)) 2472 perf_report_aux_output_id(event, idx); 2473 } 2474 2475 static void intel_pmu_del_event(struct perf_event *event) 2476 { 2477 if (needs_branch_stack(event)) 2478 intel_pmu_lbr_del(event); 2479 if (event->attr.precise_ip) 2480 intel_pmu_pebs_del(event); 2481 } 2482 2483 static int icl_set_topdown_event_period(struct perf_event *event) 2484 { 2485 struct hw_perf_event *hwc = &event->hw; 2486 s64 left = local64_read(&hwc->period_left); 2487 2488 /* 2489 * The values in PERF_METRICS MSR are derived from fixed counter 3. 2490 * Software should start both registers, PERF_METRICS and fixed 2491 * counter 3, from zero. 2492 * Clear PERF_METRICS and Fixed counter 3 in initialization. 2493 * After that, both MSRs will be cleared for each read. 2494 * Don't need to clear them again. 2495 */ 2496 if (left == x86_pmu.max_period) { 2497 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2498 wrmsrl(MSR_PERF_METRICS, 0); 2499 hwc->saved_slots = 0; 2500 hwc->saved_metric = 0; 2501 } 2502 2503 if ((hwc->saved_slots) && is_slots_event(event)) { 2504 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots); 2505 wrmsrl(MSR_PERF_METRICS, hwc->saved_metric); 2506 } 2507 2508 perf_event_update_userpage(event); 2509 2510 return 0; 2511 } 2512 2513 static int adl_set_topdown_event_period(struct perf_event *event) 2514 { 2515 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 2516 2517 if (pmu->cpu_type != hybrid_big) 2518 return 0; 2519 2520 return icl_set_topdown_event_period(event); 2521 } 2522 2523 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx) 2524 { 2525 u32 val; 2526 2527 /* 2528 * The metric is reported as an 8bit integer fraction 2529 * summing up to 0xff. 2530 * slots-in-metric = (Metric / 0xff) * slots 2531 */ 2532 val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff; 2533 return mul_u64_u32_div(slots, val, 0xff); 2534 } 2535 2536 static u64 icl_get_topdown_value(struct perf_event *event, 2537 u64 slots, u64 metrics) 2538 { 2539 int idx = event->hw.idx; 2540 u64 delta; 2541 2542 if (is_metric_idx(idx)) 2543 delta = icl_get_metrics_event_value(metrics, slots, idx); 2544 else 2545 delta = slots; 2546 2547 return delta; 2548 } 2549 2550 static void __icl_update_topdown_event(struct perf_event *event, 2551 u64 slots, u64 metrics, 2552 u64 last_slots, u64 last_metrics) 2553 { 2554 u64 delta, last = 0; 2555 2556 delta = icl_get_topdown_value(event, slots, metrics); 2557 if (last_slots) 2558 last = icl_get_topdown_value(event, last_slots, last_metrics); 2559 2560 /* 2561 * The 8bit integer fraction of metric may be not accurate, 2562 * especially when the changes is very small. 2563 * For example, if only a few bad_spec happens, the fraction 2564 * may be reduced from 1 to 0. If so, the bad_spec event value 2565 * will be 0 which is definitely less than the last value. 2566 * Avoid update event->count for this case. 2567 */ 2568 if (delta > last) { 2569 delta -= last; 2570 local64_add(delta, &event->count); 2571 } 2572 } 2573 2574 static void update_saved_topdown_regs(struct perf_event *event, u64 slots, 2575 u64 metrics, int metric_end) 2576 { 2577 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2578 struct perf_event *other; 2579 int idx; 2580 2581 event->hw.saved_slots = slots; 2582 event->hw.saved_metric = metrics; 2583 2584 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2585 if (!is_topdown_idx(idx)) 2586 continue; 2587 other = cpuc->events[idx]; 2588 other->hw.saved_slots = slots; 2589 other->hw.saved_metric = metrics; 2590 } 2591 } 2592 2593 /* 2594 * Update all active Topdown events. 2595 * 2596 * The PERF_METRICS and Fixed counter 3 are read separately. The values may be 2597 * modify by a NMI. PMU has to be disabled before calling this function. 2598 */ 2599 2600 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end) 2601 { 2602 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2603 struct perf_event *other; 2604 u64 slots, metrics; 2605 bool reset = true; 2606 int idx; 2607 2608 /* read Fixed counter 3 */ 2609 rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots); 2610 if (!slots) 2611 return 0; 2612 2613 /* read PERF_METRICS */ 2614 rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics); 2615 2616 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2617 if (!is_topdown_idx(idx)) 2618 continue; 2619 other = cpuc->events[idx]; 2620 __icl_update_topdown_event(other, slots, metrics, 2621 event ? event->hw.saved_slots : 0, 2622 event ? event->hw.saved_metric : 0); 2623 } 2624 2625 /* 2626 * Check and update this event, which may have been cleared 2627 * in active_mask e.g. x86_pmu_stop() 2628 */ 2629 if (event && !test_bit(event->hw.idx, cpuc->active_mask)) { 2630 __icl_update_topdown_event(event, slots, metrics, 2631 event->hw.saved_slots, 2632 event->hw.saved_metric); 2633 2634 /* 2635 * In x86_pmu_stop(), the event is cleared in active_mask first, 2636 * then drain the delta, which indicates context switch for 2637 * counting. 2638 * Save metric and slots for context switch. 2639 * Don't need to reset the PERF_METRICS and Fixed counter 3. 2640 * Because the values will be restored in next schedule in. 2641 */ 2642 update_saved_topdown_regs(event, slots, metrics, metric_end); 2643 reset = false; 2644 } 2645 2646 if (reset) { 2647 /* The fixed counter 3 has to be written before the PERF_METRICS. */ 2648 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2649 wrmsrl(MSR_PERF_METRICS, 0); 2650 if (event) 2651 update_saved_topdown_regs(event, 0, 0, metric_end); 2652 } 2653 2654 return slots; 2655 } 2656 2657 static u64 icl_update_topdown_event(struct perf_event *event) 2658 { 2659 return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE + 2660 x86_pmu.num_topdown_events - 1); 2661 } 2662 2663 static u64 adl_update_topdown_event(struct perf_event *event) 2664 { 2665 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 2666 2667 if (pmu->cpu_type != hybrid_big) 2668 return 0; 2669 2670 return icl_update_topdown_event(event); 2671 } 2672 2673 2674 static void intel_pmu_read_topdown_event(struct perf_event *event) 2675 { 2676 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2677 2678 /* Only need to call update_topdown_event() once for group read. */ 2679 if ((cpuc->txn_flags & PERF_PMU_TXN_READ) && 2680 !is_slots_event(event)) 2681 return; 2682 2683 perf_pmu_disable(event->pmu); 2684 x86_pmu.update_topdown_event(event); 2685 perf_pmu_enable(event->pmu); 2686 } 2687 2688 static void intel_pmu_read_event(struct perf_event *event) 2689 { 2690 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) 2691 intel_pmu_auto_reload_read(event); 2692 else if (is_topdown_count(event) && x86_pmu.update_topdown_event) 2693 intel_pmu_read_topdown_event(event); 2694 else 2695 x86_perf_event_update(event); 2696 } 2697 2698 static void intel_pmu_enable_fixed(struct perf_event *event) 2699 { 2700 struct hw_perf_event *hwc = &event->hw; 2701 u64 ctrl_val, mask, bits = 0; 2702 int idx = hwc->idx; 2703 2704 if (is_topdown_idx(idx)) { 2705 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2706 /* 2707 * When there are other active TopDown events, 2708 * don't enable the fixed counter 3 again. 2709 */ 2710 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2711 return; 2712 2713 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2714 } 2715 2716 intel_set_masks(event, idx); 2717 2718 /* 2719 * Enable IRQ generation (0x8), if not PEBS, 2720 * and enable ring-3 counting (0x2) and ring-0 counting (0x1) 2721 * if requested: 2722 */ 2723 if (!event->attr.precise_ip) 2724 bits |= 0x8; 2725 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) 2726 bits |= 0x2; 2727 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) 2728 bits |= 0x1; 2729 2730 /* 2731 * ANY bit is supported in v3 and up 2732 */ 2733 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) 2734 bits |= 0x4; 2735 2736 idx -= INTEL_PMC_IDX_FIXED; 2737 bits <<= (idx * 4); 2738 mask = 0xfULL << (idx * 4); 2739 2740 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { 2741 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2742 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2743 } 2744 2745 rdmsrl(hwc->config_base, ctrl_val); 2746 ctrl_val &= ~mask; 2747 ctrl_val |= bits; 2748 wrmsrl(hwc->config_base, ctrl_val); 2749 } 2750 2751 static void intel_pmu_enable_event(struct perf_event *event) 2752 { 2753 struct hw_perf_event *hwc = &event->hw; 2754 int idx = hwc->idx; 2755 2756 if (unlikely(event->attr.precise_ip)) 2757 intel_pmu_pebs_enable(event); 2758 2759 switch (idx) { 2760 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2761 intel_set_masks(event, idx); 2762 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 2763 break; 2764 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2765 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2766 intel_pmu_enable_fixed(event); 2767 break; 2768 case INTEL_PMC_IDX_FIXED_BTS: 2769 if (!__this_cpu_read(cpu_hw_events.enabled)) 2770 return; 2771 intel_pmu_enable_bts(hwc->config); 2772 break; 2773 case INTEL_PMC_IDX_FIXED_VLBR: 2774 intel_set_masks(event, idx); 2775 break; 2776 default: 2777 pr_warn("Failed to enable the event with invalid index %d\n", 2778 idx); 2779 } 2780 } 2781 2782 static void intel_pmu_add_event(struct perf_event *event) 2783 { 2784 if (event->attr.precise_ip) 2785 intel_pmu_pebs_add(event); 2786 if (needs_branch_stack(event)) 2787 intel_pmu_lbr_add(event); 2788 } 2789 2790 /* 2791 * Save and restart an expired event. Called by NMI contexts, 2792 * so it has to be careful about preempting normal event ops: 2793 */ 2794 int intel_pmu_save_and_restart(struct perf_event *event) 2795 { 2796 x86_perf_event_update(event); 2797 /* 2798 * For a checkpointed counter always reset back to 0. This 2799 * avoids a situation where the counter overflows, aborts the 2800 * transaction and is then set back to shortly before the 2801 * overflow, and overflows and aborts again. 2802 */ 2803 if (unlikely(event_is_checkpointed(event))) { 2804 /* No race with NMIs because the counter should not be armed */ 2805 wrmsrl(event->hw.event_base, 0); 2806 local64_set(&event->hw.prev_count, 0); 2807 } 2808 return x86_perf_event_set_period(event); 2809 } 2810 2811 static void intel_pmu_reset(void) 2812 { 2813 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); 2814 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2815 int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed); 2816 int num_counters = hybrid(cpuc->pmu, num_counters); 2817 unsigned long flags; 2818 int idx; 2819 2820 if (!num_counters) 2821 return; 2822 2823 local_irq_save(flags); 2824 2825 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); 2826 2827 for (idx = 0; idx < num_counters; idx++) { 2828 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); 2829 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); 2830 } 2831 for (idx = 0; idx < num_counters_fixed; idx++) { 2832 if (fixed_counter_disabled(idx, cpuc->pmu)) 2833 continue; 2834 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); 2835 } 2836 2837 if (ds) 2838 ds->bts_index = ds->bts_buffer_base; 2839 2840 /* Ack all overflows and disable fixed counters */ 2841 if (x86_pmu.version >= 2) { 2842 intel_pmu_ack_status(intel_pmu_get_status()); 2843 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2844 } 2845 2846 /* Reset LBRs and LBR freezing */ 2847 if (x86_pmu.lbr_nr) { 2848 update_debugctlmsr(get_debugctlmsr() & 2849 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); 2850 } 2851 2852 local_irq_restore(flags); 2853 } 2854 2855 static int handle_pmi_common(struct pt_regs *regs, u64 status) 2856 { 2857 struct perf_sample_data data; 2858 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2859 int bit; 2860 int handled = 0; 2861 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2862 2863 inc_irq_stat(apic_perf_irqs); 2864 2865 /* 2866 * Ignore a range of extra bits in status that do not indicate 2867 * overflow by themselves. 2868 */ 2869 status &= ~(GLOBAL_STATUS_COND_CHG | 2870 GLOBAL_STATUS_ASIF | 2871 GLOBAL_STATUS_LBRS_FROZEN); 2872 if (!status) 2873 return 0; 2874 /* 2875 * In case multiple PEBS events are sampled at the same time, 2876 * it is possible to have GLOBAL_STATUS bit 62 set indicating 2877 * PEBS buffer overflow and also seeing at most 3 PEBS counters 2878 * having their bits set in the status register. This is a sign 2879 * that there was at least one PEBS record pending at the time 2880 * of the PMU interrupt. PEBS counters must only be processed 2881 * via the drain_pebs() calls and not via the regular sample 2882 * processing loop coming after that the function, otherwise 2883 * phony regular samples may be generated in the sampling buffer 2884 * not marked with the EXACT tag. Another possibility is to have 2885 * one PEBS event and at least one non-PEBS event which overflows 2886 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will 2887 * not be set, yet the overflow status bit for the PEBS counter will 2888 * be on Skylake. 2889 * 2890 * To avoid this problem, we systematically ignore the PEBS-enabled 2891 * counters from the GLOBAL_STATUS mask and we always process PEBS 2892 * events via drain_pebs(). 2893 */ 2894 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 2895 status &= ~cpuc->pebs_enabled; 2896 else 2897 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 2898 2899 /* 2900 * PEBS overflow sets bit 62 in the global status register 2901 */ 2902 if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) { 2903 u64 pebs_enabled = cpuc->pebs_enabled; 2904 2905 handled++; 2906 x86_pmu.drain_pebs(regs, &data); 2907 status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI; 2908 2909 /* 2910 * PMI throttle may be triggered, which stops the PEBS event. 2911 * Although cpuc->pebs_enabled is updated accordingly, the 2912 * MSR_IA32_PEBS_ENABLE is not updated. Because the 2913 * cpuc->enabled has been forced to 0 in PMI. 2914 * Update the MSR if pebs_enabled is changed. 2915 */ 2916 if (pebs_enabled != cpuc->pebs_enabled) 2917 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled); 2918 } 2919 2920 /* 2921 * Intel PT 2922 */ 2923 if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) { 2924 handled++; 2925 if (!perf_guest_handle_intel_pt_intr()) 2926 intel_pt_interrupt(); 2927 } 2928 2929 /* 2930 * Intel Perf metrics 2931 */ 2932 if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) { 2933 handled++; 2934 if (x86_pmu.update_topdown_event) 2935 x86_pmu.update_topdown_event(NULL); 2936 } 2937 2938 /* 2939 * Checkpointed counters can lead to 'spurious' PMIs because the 2940 * rollback caused by the PMI will have cleared the overflow status 2941 * bit. Therefore always force probe these counters. 2942 */ 2943 status |= cpuc->intel_cp_status; 2944 2945 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { 2946 struct perf_event *event = cpuc->events[bit]; 2947 2948 handled++; 2949 2950 if (!test_bit(bit, cpuc->active_mask)) 2951 continue; 2952 2953 if (!intel_pmu_save_and_restart(event)) 2954 continue; 2955 2956 perf_sample_data_init(&data, 0, event->hw.last_period); 2957 2958 if (has_branch_stack(event)) 2959 data.br_stack = &cpuc->lbr_stack; 2960 2961 if (perf_event_overflow(event, &data, regs)) 2962 x86_pmu_stop(event, 0); 2963 } 2964 2965 return handled; 2966 } 2967 2968 /* 2969 * This handler is triggered by the local APIC, so the APIC IRQ handling 2970 * rules apply: 2971 */ 2972 static int intel_pmu_handle_irq(struct pt_regs *regs) 2973 { 2974 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2975 bool late_ack = hybrid_bit(cpuc->pmu, late_ack); 2976 bool mid_ack = hybrid_bit(cpuc->pmu, mid_ack); 2977 int loops; 2978 u64 status; 2979 int handled; 2980 int pmu_enabled; 2981 2982 /* 2983 * Save the PMU state. 2984 * It needs to be restored when leaving the handler. 2985 */ 2986 pmu_enabled = cpuc->enabled; 2987 /* 2988 * In general, the early ACK is only applied for old platforms. 2989 * For the big core starts from Haswell, the late ACK should be 2990 * applied. 2991 * For the small core after Tremont, we have to do the ACK right 2992 * before re-enabling counters, which is in the middle of the 2993 * NMI handler. 2994 */ 2995 if (!late_ack && !mid_ack) 2996 apic_write(APIC_LVTPC, APIC_DM_NMI); 2997 intel_bts_disable_local(); 2998 cpuc->enabled = 0; 2999 __intel_pmu_disable_all(true); 3000 handled = intel_pmu_drain_bts_buffer(); 3001 handled += intel_bts_interrupt(); 3002 status = intel_pmu_get_status(); 3003 if (!status) 3004 goto done; 3005 3006 loops = 0; 3007 again: 3008 intel_pmu_lbr_read(); 3009 intel_pmu_ack_status(status); 3010 if (++loops > 100) { 3011 static bool warned; 3012 3013 if (!warned) { 3014 WARN(1, "perfevents: irq loop stuck!\n"); 3015 perf_event_print_debug(); 3016 warned = true; 3017 } 3018 intel_pmu_reset(); 3019 goto done; 3020 } 3021 3022 handled += handle_pmi_common(regs, status); 3023 3024 /* 3025 * Repeat if there is more work to be done: 3026 */ 3027 status = intel_pmu_get_status(); 3028 if (status) 3029 goto again; 3030 3031 done: 3032 if (mid_ack) 3033 apic_write(APIC_LVTPC, APIC_DM_NMI); 3034 /* Only restore PMU state when it's active. See x86_pmu_disable(). */ 3035 cpuc->enabled = pmu_enabled; 3036 if (pmu_enabled) 3037 __intel_pmu_enable_all(0, true); 3038 intel_bts_enable_local(); 3039 3040 /* 3041 * Only unmask the NMI after the overflow counters 3042 * have been reset. This avoids spurious NMIs on 3043 * Haswell CPUs. 3044 */ 3045 if (late_ack) 3046 apic_write(APIC_LVTPC, APIC_DM_NMI); 3047 return handled; 3048 } 3049 3050 static struct event_constraint * 3051 intel_bts_constraints(struct perf_event *event) 3052 { 3053 if (unlikely(intel_pmu_has_bts(event))) 3054 return &bts_constraint; 3055 3056 return NULL; 3057 } 3058 3059 /* 3060 * Note: matches a fake event, like Fixed2. 3061 */ 3062 static struct event_constraint * 3063 intel_vlbr_constraints(struct perf_event *event) 3064 { 3065 struct event_constraint *c = &vlbr_constraint; 3066 3067 if (unlikely(constraint_match(c, event->hw.config))) { 3068 event->hw.flags |= c->flags; 3069 return c; 3070 } 3071 3072 return NULL; 3073 } 3074 3075 static int intel_alt_er(struct cpu_hw_events *cpuc, 3076 int idx, u64 config) 3077 { 3078 struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs); 3079 int alt_idx = idx; 3080 3081 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) 3082 return idx; 3083 3084 if (idx == EXTRA_REG_RSP_0) 3085 alt_idx = EXTRA_REG_RSP_1; 3086 3087 if (idx == EXTRA_REG_RSP_1) 3088 alt_idx = EXTRA_REG_RSP_0; 3089 3090 if (config & ~extra_regs[alt_idx].valid_mask) 3091 return idx; 3092 3093 return alt_idx; 3094 } 3095 3096 static void intel_fixup_er(struct perf_event *event, int idx) 3097 { 3098 struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs); 3099 event->hw.extra_reg.idx = idx; 3100 3101 if (idx == EXTRA_REG_RSP_0) { 3102 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3103 event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event; 3104 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; 3105 } else if (idx == EXTRA_REG_RSP_1) { 3106 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3107 event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event; 3108 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; 3109 } 3110 } 3111 3112 /* 3113 * manage allocation of shared extra msr for certain events 3114 * 3115 * sharing can be: 3116 * per-cpu: to be shared between the various events on a single PMU 3117 * per-core: per-cpu + shared by HT threads 3118 */ 3119 static struct event_constraint * 3120 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, 3121 struct perf_event *event, 3122 struct hw_perf_event_extra *reg) 3123 { 3124 struct event_constraint *c = &emptyconstraint; 3125 struct er_account *era; 3126 unsigned long flags; 3127 int idx = reg->idx; 3128 3129 /* 3130 * reg->alloc can be set due to existing state, so for fake cpuc we 3131 * need to ignore this, otherwise we might fail to allocate proper fake 3132 * state for this extra reg constraint. Also see the comment below. 3133 */ 3134 if (reg->alloc && !cpuc->is_fake) 3135 return NULL; /* call x86_get_event_constraint() */ 3136 3137 again: 3138 era = &cpuc->shared_regs->regs[idx]; 3139 /* 3140 * we use spin_lock_irqsave() to avoid lockdep issues when 3141 * passing a fake cpuc 3142 */ 3143 raw_spin_lock_irqsave(&era->lock, flags); 3144 3145 if (!atomic_read(&era->ref) || era->config == reg->config) { 3146 3147 /* 3148 * If its a fake cpuc -- as per validate_{group,event}() we 3149 * shouldn't touch event state and we can avoid doing so 3150 * since both will only call get_event_constraints() once 3151 * on each event, this avoids the need for reg->alloc. 3152 * 3153 * Not doing the ER fixup will only result in era->reg being 3154 * wrong, but since we won't actually try and program hardware 3155 * this isn't a problem either. 3156 */ 3157 if (!cpuc->is_fake) { 3158 if (idx != reg->idx) 3159 intel_fixup_er(event, idx); 3160 3161 /* 3162 * x86_schedule_events() can call get_event_constraints() 3163 * multiple times on events in the case of incremental 3164 * scheduling(). reg->alloc ensures we only do the ER 3165 * allocation once. 3166 */ 3167 reg->alloc = 1; 3168 } 3169 3170 /* lock in msr value */ 3171 era->config = reg->config; 3172 era->reg = reg->reg; 3173 3174 /* one more user */ 3175 atomic_inc(&era->ref); 3176 3177 /* 3178 * need to call x86_get_event_constraint() 3179 * to check if associated event has constraints 3180 */ 3181 c = NULL; 3182 } else { 3183 idx = intel_alt_er(cpuc, idx, reg->config); 3184 if (idx != reg->idx) { 3185 raw_spin_unlock_irqrestore(&era->lock, flags); 3186 goto again; 3187 } 3188 } 3189 raw_spin_unlock_irqrestore(&era->lock, flags); 3190 3191 return c; 3192 } 3193 3194 static void 3195 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, 3196 struct hw_perf_event_extra *reg) 3197 { 3198 struct er_account *era; 3199 3200 /* 3201 * Only put constraint if extra reg was actually allocated. Also takes 3202 * care of event which do not use an extra shared reg. 3203 * 3204 * Also, if this is a fake cpuc we shouldn't touch any event state 3205 * (reg->alloc) and we don't care about leaving inconsistent cpuc state 3206 * either since it'll be thrown out. 3207 */ 3208 if (!reg->alloc || cpuc->is_fake) 3209 return; 3210 3211 era = &cpuc->shared_regs->regs[reg->idx]; 3212 3213 /* one fewer user */ 3214 atomic_dec(&era->ref); 3215 3216 /* allocate again next time */ 3217 reg->alloc = 0; 3218 } 3219 3220 static struct event_constraint * 3221 intel_shared_regs_constraints(struct cpu_hw_events *cpuc, 3222 struct perf_event *event) 3223 { 3224 struct event_constraint *c = NULL, *d; 3225 struct hw_perf_event_extra *xreg, *breg; 3226 3227 xreg = &event->hw.extra_reg; 3228 if (xreg->idx != EXTRA_REG_NONE) { 3229 c = __intel_shared_reg_get_constraints(cpuc, event, xreg); 3230 if (c == &emptyconstraint) 3231 return c; 3232 } 3233 breg = &event->hw.branch_reg; 3234 if (breg->idx != EXTRA_REG_NONE) { 3235 d = __intel_shared_reg_get_constraints(cpuc, event, breg); 3236 if (d == &emptyconstraint) { 3237 __intel_shared_reg_put_constraints(cpuc, xreg); 3238 c = d; 3239 } 3240 } 3241 return c; 3242 } 3243 3244 struct event_constraint * 3245 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3246 struct perf_event *event) 3247 { 3248 struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints); 3249 struct event_constraint *c; 3250 3251 if (event_constraints) { 3252 for_each_event_constraint(c, event_constraints) { 3253 if (constraint_match(c, event->hw.config)) { 3254 event->hw.flags |= c->flags; 3255 return c; 3256 } 3257 } 3258 } 3259 3260 return &hybrid_var(cpuc->pmu, unconstrained); 3261 } 3262 3263 static struct event_constraint * 3264 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3265 struct perf_event *event) 3266 { 3267 struct event_constraint *c; 3268 3269 c = intel_vlbr_constraints(event); 3270 if (c) 3271 return c; 3272 3273 c = intel_bts_constraints(event); 3274 if (c) 3275 return c; 3276 3277 c = intel_shared_regs_constraints(cpuc, event); 3278 if (c) 3279 return c; 3280 3281 c = intel_pebs_constraints(event); 3282 if (c) 3283 return c; 3284 3285 return x86_get_event_constraints(cpuc, idx, event); 3286 } 3287 3288 static void 3289 intel_start_scheduling(struct cpu_hw_events *cpuc) 3290 { 3291 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3292 struct intel_excl_states *xl; 3293 int tid = cpuc->excl_thread_id; 3294 3295 /* 3296 * nothing needed if in group validation mode 3297 */ 3298 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3299 return; 3300 3301 /* 3302 * no exclusion needed 3303 */ 3304 if (WARN_ON_ONCE(!excl_cntrs)) 3305 return; 3306 3307 xl = &excl_cntrs->states[tid]; 3308 3309 xl->sched_started = true; 3310 /* 3311 * lock shared state until we are done scheduling 3312 * in stop_event_scheduling() 3313 * makes scheduling appear as a transaction 3314 */ 3315 raw_spin_lock(&excl_cntrs->lock); 3316 } 3317 3318 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 3319 { 3320 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3321 struct event_constraint *c = cpuc->event_constraint[idx]; 3322 struct intel_excl_states *xl; 3323 int tid = cpuc->excl_thread_id; 3324 3325 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3326 return; 3327 3328 if (WARN_ON_ONCE(!excl_cntrs)) 3329 return; 3330 3331 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) 3332 return; 3333 3334 xl = &excl_cntrs->states[tid]; 3335 3336 lockdep_assert_held(&excl_cntrs->lock); 3337 3338 if (c->flags & PERF_X86_EVENT_EXCL) 3339 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; 3340 else 3341 xl->state[cntr] = INTEL_EXCL_SHARED; 3342 } 3343 3344 static void 3345 intel_stop_scheduling(struct cpu_hw_events *cpuc) 3346 { 3347 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3348 struct intel_excl_states *xl; 3349 int tid = cpuc->excl_thread_id; 3350 3351 /* 3352 * nothing needed if in group validation mode 3353 */ 3354 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3355 return; 3356 /* 3357 * no exclusion needed 3358 */ 3359 if (WARN_ON_ONCE(!excl_cntrs)) 3360 return; 3361 3362 xl = &excl_cntrs->states[tid]; 3363 3364 xl->sched_started = false; 3365 /* 3366 * release shared state lock (acquired in intel_start_scheduling()) 3367 */ 3368 raw_spin_unlock(&excl_cntrs->lock); 3369 } 3370 3371 static struct event_constraint * 3372 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx) 3373 { 3374 WARN_ON_ONCE(!cpuc->constraint_list); 3375 3376 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { 3377 struct event_constraint *cx; 3378 3379 /* 3380 * grab pre-allocated constraint entry 3381 */ 3382 cx = &cpuc->constraint_list[idx]; 3383 3384 /* 3385 * initialize dynamic constraint 3386 * with static constraint 3387 */ 3388 *cx = *c; 3389 3390 /* 3391 * mark constraint as dynamic 3392 */ 3393 cx->flags |= PERF_X86_EVENT_DYNAMIC; 3394 c = cx; 3395 } 3396 3397 return c; 3398 } 3399 3400 static struct event_constraint * 3401 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, 3402 int idx, struct event_constraint *c) 3403 { 3404 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3405 struct intel_excl_states *xlo; 3406 int tid = cpuc->excl_thread_id; 3407 int is_excl, i, w; 3408 3409 /* 3410 * validating a group does not require 3411 * enforcing cross-thread exclusion 3412 */ 3413 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3414 return c; 3415 3416 /* 3417 * no exclusion needed 3418 */ 3419 if (WARN_ON_ONCE(!excl_cntrs)) 3420 return c; 3421 3422 /* 3423 * because we modify the constraint, we need 3424 * to make a copy. Static constraints come 3425 * from static const tables. 3426 * 3427 * only needed when constraint has not yet 3428 * been cloned (marked dynamic) 3429 */ 3430 c = dyn_constraint(cpuc, c, idx); 3431 3432 /* 3433 * From here on, the constraint is dynamic. 3434 * Either it was just allocated above, or it 3435 * was allocated during a earlier invocation 3436 * of this function 3437 */ 3438 3439 /* 3440 * state of sibling HT 3441 */ 3442 xlo = &excl_cntrs->states[tid ^ 1]; 3443 3444 /* 3445 * event requires exclusive counter access 3446 * across HT threads 3447 */ 3448 is_excl = c->flags & PERF_X86_EVENT_EXCL; 3449 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { 3450 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; 3451 if (!cpuc->n_excl++) 3452 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); 3453 } 3454 3455 /* 3456 * Modify static constraint with current dynamic 3457 * state of thread 3458 * 3459 * EXCLUSIVE: sibling counter measuring exclusive event 3460 * SHARED : sibling counter measuring non-exclusive event 3461 * UNUSED : sibling counter unused 3462 */ 3463 w = c->weight; 3464 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { 3465 /* 3466 * exclusive event in sibling counter 3467 * our corresponding counter cannot be used 3468 * regardless of our event 3469 */ 3470 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) { 3471 __clear_bit(i, c->idxmsk); 3472 w--; 3473 continue; 3474 } 3475 /* 3476 * if measuring an exclusive event, sibling 3477 * measuring non-exclusive, then counter cannot 3478 * be used 3479 */ 3480 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) { 3481 __clear_bit(i, c->idxmsk); 3482 w--; 3483 continue; 3484 } 3485 } 3486 3487 /* 3488 * if we return an empty mask, then switch 3489 * back to static empty constraint to avoid 3490 * the cost of freeing later on 3491 */ 3492 if (!w) 3493 c = &emptyconstraint; 3494 3495 c->weight = w; 3496 3497 return c; 3498 } 3499 3500 static struct event_constraint * 3501 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3502 struct perf_event *event) 3503 { 3504 struct event_constraint *c1, *c2; 3505 3506 c1 = cpuc->event_constraint[idx]; 3507 3508 /* 3509 * first time only 3510 * - static constraint: no change across incremental scheduling calls 3511 * - dynamic constraint: handled by intel_get_excl_constraints() 3512 */ 3513 c2 = __intel_get_event_constraints(cpuc, idx, event); 3514 if (c1) { 3515 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC)); 3516 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); 3517 c1->weight = c2->weight; 3518 c2 = c1; 3519 } 3520 3521 if (cpuc->excl_cntrs) 3522 return intel_get_excl_constraints(cpuc, event, idx, c2); 3523 3524 return c2; 3525 } 3526 3527 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, 3528 struct perf_event *event) 3529 { 3530 struct hw_perf_event *hwc = &event->hw; 3531 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3532 int tid = cpuc->excl_thread_id; 3533 struct intel_excl_states *xl; 3534 3535 /* 3536 * nothing needed if in group validation mode 3537 */ 3538 if (cpuc->is_fake) 3539 return; 3540 3541 if (WARN_ON_ONCE(!excl_cntrs)) 3542 return; 3543 3544 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { 3545 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; 3546 if (!--cpuc->n_excl) 3547 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); 3548 } 3549 3550 /* 3551 * If event was actually assigned, then mark the counter state as 3552 * unused now. 3553 */ 3554 if (hwc->idx >= 0) { 3555 xl = &excl_cntrs->states[tid]; 3556 3557 /* 3558 * put_constraint may be called from x86_schedule_events() 3559 * which already has the lock held so here make locking 3560 * conditional. 3561 */ 3562 if (!xl->sched_started) 3563 raw_spin_lock(&excl_cntrs->lock); 3564 3565 xl->state[hwc->idx] = INTEL_EXCL_UNUSED; 3566 3567 if (!xl->sched_started) 3568 raw_spin_unlock(&excl_cntrs->lock); 3569 } 3570 } 3571 3572 static void 3573 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, 3574 struct perf_event *event) 3575 { 3576 struct hw_perf_event_extra *reg; 3577 3578 reg = &event->hw.extra_reg; 3579 if (reg->idx != EXTRA_REG_NONE) 3580 __intel_shared_reg_put_constraints(cpuc, reg); 3581 3582 reg = &event->hw.branch_reg; 3583 if (reg->idx != EXTRA_REG_NONE) 3584 __intel_shared_reg_put_constraints(cpuc, reg); 3585 } 3586 3587 static void intel_put_event_constraints(struct cpu_hw_events *cpuc, 3588 struct perf_event *event) 3589 { 3590 intel_put_shared_regs_event_constraints(cpuc, event); 3591 3592 /* 3593 * is PMU has exclusive counter restrictions, then 3594 * all events are subject to and must call the 3595 * put_excl_constraints() routine 3596 */ 3597 if (cpuc->excl_cntrs) 3598 intel_put_excl_constraints(cpuc, event); 3599 } 3600 3601 static void intel_pebs_aliases_core2(struct perf_event *event) 3602 { 3603 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3604 /* 3605 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3606 * (0x003c) so that we can use it with PEBS. 3607 * 3608 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3609 * PEBS capable. However we can use INST_RETIRED.ANY_P 3610 * (0x00c0), which is a PEBS capable event, to get the same 3611 * count. 3612 * 3613 * INST_RETIRED.ANY_P counts the number of cycles that retires 3614 * CNTMASK instructions. By setting CNTMASK to a value (16) 3615 * larger than the maximum number of instructions that can be 3616 * retired per cycle (4) and then inverting the condition, we 3617 * count all cycles that retire 16 or less instructions, which 3618 * is every cycle. 3619 * 3620 * Thereby we gain a PEBS capable cycle counter. 3621 */ 3622 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); 3623 3624 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3625 event->hw.config = alt_config; 3626 } 3627 } 3628 3629 static void intel_pebs_aliases_snb(struct perf_event *event) 3630 { 3631 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3632 /* 3633 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3634 * (0x003c) so that we can use it with PEBS. 3635 * 3636 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3637 * PEBS capable. However we can use UOPS_RETIRED.ALL 3638 * (0x01c2), which is a PEBS capable event, to get the same 3639 * count. 3640 * 3641 * UOPS_RETIRED.ALL counts the number of cycles that retires 3642 * CNTMASK micro-ops. By setting CNTMASK to a value (16) 3643 * larger than the maximum number of micro-ops that can be 3644 * retired per cycle (4) and then inverting the condition, we 3645 * count all cycles that retire 16 or less micro-ops, which 3646 * is every cycle. 3647 * 3648 * Thereby we gain a PEBS capable cycle counter. 3649 */ 3650 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); 3651 3652 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3653 event->hw.config = alt_config; 3654 } 3655 } 3656 3657 static void intel_pebs_aliases_precdist(struct perf_event *event) 3658 { 3659 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3660 /* 3661 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3662 * (0x003c) so that we can use it with PEBS. 3663 * 3664 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3665 * PEBS capable. However we can use INST_RETIRED.PREC_DIST 3666 * (0x01c0), which is a PEBS capable event, to get the same 3667 * count. 3668 * 3669 * The PREC_DIST event has special support to minimize sample 3670 * shadowing effects. One drawback is that it can be 3671 * only programmed on counter 1, but that seems like an 3672 * acceptable trade off. 3673 */ 3674 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16); 3675 3676 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3677 event->hw.config = alt_config; 3678 } 3679 } 3680 3681 static void intel_pebs_aliases_ivb(struct perf_event *event) 3682 { 3683 if (event->attr.precise_ip < 3) 3684 return intel_pebs_aliases_snb(event); 3685 return intel_pebs_aliases_precdist(event); 3686 } 3687 3688 static void intel_pebs_aliases_skl(struct perf_event *event) 3689 { 3690 if (event->attr.precise_ip < 3) 3691 return intel_pebs_aliases_core2(event); 3692 return intel_pebs_aliases_precdist(event); 3693 } 3694 3695 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event) 3696 { 3697 unsigned long flags = x86_pmu.large_pebs_flags; 3698 3699 if (event->attr.use_clockid) 3700 flags &= ~PERF_SAMPLE_TIME; 3701 if (!event->attr.exclude_kernel) 3702 flags &= ~PERF_SAMPLE_REGS_USER; 3703 if (event->attr.sample_regs_user & ~PEBS_GP_REGS) 3704 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR); 3705 return flags; 3706 } 3707 3708 static int intel_pmu_bts_config(struct perf_event *event) 3709 { 3710 struct perf_event_attr *attr = &event->attr; 3711 3712 if (unlikely(intel_pmu_has_bts(event))) { 3713 /* BTS is not supported by this architecture. */ 3714 if (!x86_pmu.bts_active) 3715 return -EOPNOTSUPP; 3716 3717 /* BTS is currently only allowed for user-mode. */ 3718 if (!attr->exclude_kernel) 3719 return -EOPNOTSUPP; 3720 3721 /* BTS is not allowed for precise events. */ 3722 if (attr->precise_ip) 3723 return -EOPNOTSUPP; 3724 3725 /* disallow bts if conflicting events are present */ 3726 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3727 return -EBUSY; 3728 3729 event->destroy = hw_perf_lbr_event_destroy; 3730 } 3731 3732 return 0; 3733 } 3734 3735 static int core_pmu_hw_config(struct perf_event *event) 3736 { 3737 int ret = x86_pmu_hw_config(event); 3738 3739 if (ret) 3740 return ret; 3741 3742 return intel_pmu_bts_config(event); 3743 } 3744 3745 #define INTEL_TD_METRIC_AVAILABLE_MAX (INTEL_TD_METRIC_RETIRING + \ 3746 ((x86_pmu.num_topdown_events - 1) << 8)) 3747 3748 static bool is_available_metric_event(struct perf_event *event) 3749 { 3750 return is_metric_event(event) && 3751 event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX; 3752 } 3753 3754 static inline bool is_mem_loads_event(struct perf_event *event) 3755 { 3756 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01); 3757 } 3758 3759 static inline bool is_mem_loads_aux_event(struct perf_event *event) 3760 { 3761 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82); 3762 } 3763 3764 static inline bool require_mem_loads_aux_event(struct perf_event *event) 3765 { 3766 if (!(x86_pmu.flags & PMU_FL_MEM_LOADS_AUX)) 3767 return false; 3768 3769 if (is_hybrid()) 3770 return hybrid_pmu(event->pmu)->cpu_type == hybrid_big; 3771 3772 return true; 3773 } 3774 3775 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx) 3776 { 3777 union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap); 3778 3779 return test_bit(idx, (unsigned long *)&intel_cap->capabilities); 3780 } 3781 3782 static int intel_pmu_hw_config(struct perf_event *event) 3783 { 3784 int ret = x86_pmu_hw_config(event); 3785 3786 if (ret) 3787 return ret; 3788 3789 ret = intel_pmu_bts_config(event); 3790 if (ret) 3791 return ret; 3792 3793 if (event->attr.precise_ip) { 3794 if ((event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_FIXED_VLBR_EVENT) 3795 return -EINVAL; 3796 3797 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) { 3798 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; 3799 if (!(event->attr.sample_type & 3800 ~intel_pmu_large_pebs_flags(event))) { 3801 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS; 3802 event->attach_state |= PERF_ATTACH_SCHED_CB; 3803 } 3804 } 3805 if (x86_pmu.pebs_aliases) 3806 x86_pmu.pebs_aliases(event); 3807 3808 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) 3809 event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY; 3810 } 3811 3812 if (needs_branch_stack(event)) { 3813 ret = intel_pmu_setup_lbr_filter(event); 3814 if (ret) 3815 return ret; 3816 event->attach_state |= PERF_ATTACH_SCHED_CB; 3817 3818 /* 3819 * BTS is set up earlier in this path, so don't account twice 3820 */ 3821 if (!unlikely(intel_pmu_has_bts(event))) { 3822 /* disallow lbr if conflicting events are present */ 3823 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3824 return -EBUSY; 3825 3826 event->destroy = hw_perf_lbr_event_destroy; 3827 } 3828 } 3829 3830 if (event->attr.aux_output) { 3831 if (!event->attr.precise_ip) 3832 return -EINVAL; 3833 3834 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT; 3835 } 3836 3837 if ((event->attr.type == PERF_TYPE_HARDWARE) || 3838 (event->attr.type == PERF_TYPE_HW_CACHE)) 3839 return 0; 3840 3841 /* 3842 * Config Topdown slots and metric events 3843 * 3844 * The slots event on Fixed Counter 3 can support sampling, 3845 * which will be handled normally in x86_perf_event_update(). 3846 * 3847 * Metric events don't support sampling and require being paired 3848 * with a slots event as group leader. When the slots event 3849 * is used in a metrics group, it too cannot support sampling. 3850 */ 3851 if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) { 3852 if (event->attr.config1 || event->attr.config2) 3853 return -EINVAL; 3854 3855 /* 3856 * The TopDown metrics events and slots event don't 3857 * support any filters. 3858 */ 3859 if (event->attr.config & X86_ALL_EVENT_FLAGS) 3860 return -EINVAL; 3861 3862 if (is_available_metric_event(event)) { 3863 struct perf_event *leader = event->group_leader; 3864 3865 /* The metric events don't support sampling. */ 3866 if (is_sampling_event(event)) 3867 return -EINVAL; 3868 3869 /* The metric events require a slots group leader. */ 3870 if (!is_slots_event(leader)) 3871 return -EINVAL; 3872 3873 /* 3874 * The leader/SLOTS must not be a sampling event for 3875 * metric use; hardware requires it starts at 0 when used 3876 * in conjunction with MSR_PERF_METRICS. 3877 */ 3878 if (is_sampling_event(leader)) 3879 return -EINVAL; 3880 3881 event->event_caps |= PERF_EV_CAP_SIBLING; 3882 /* 3883 * Only once we have a METRICs sibling do we 3884 * need TopDown magic. 3885 */ 3886 leader->hw.flags |= PERF_X86_EVENT_TOPDOWN; 3887 event->hw.flags |= PERF_X86_EVENT_TOPDOWN; 3888 } 3889 } 3890 3891 /* 3892 * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR 3893 * doesn't function quite right. As a work-around it needs to always be 3894 * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82). 3895 * The actual count of this second event is irrelevant it just needs 3896 * to be active to make the first event function correctly. 3897 * 3898 * In a group, the auxiliary event must be in front of the load latency 3899 * event. The rule is to simplify the implementation of the check. 3900 * That's because perf cannot have a complete group at the moment. 3901 */ 3902 if (require_mem_loads_aux_event(event) && 3903 (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) && 3904 is_mem_loads_event(event)) { 3905 struct perf_event *leader = event->group_leader; 3906 struct perf_event *sibling = NULL; 3907 3908 if (!is_mem_loads_aux_event(leader)) { 3909 for_each_sibling_event(sibling, leader) { 3910 if (is_mem_loads_aux_event(sibling)) 3911 break; 3912 } 3913 if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list)) 3914 return -ENODATA; 3915 } 3916 } 3917 3918 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) 3919 return 0; 3920 3921 if (x86_pmu.version < 3) 3922 return -EINVAL; 3923 3924 ret = perf_allow_cpu(&event->attr); 3925 if (ret) 3926 return ret; 3927 3928 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; 3929 3930 return 0; 3931 } 3932 3933 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) 3934 { 3935 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3936 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3937 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 3938 3939 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; 3940 arr[0].host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask; 3941 arr[0].guest = intel_ctrl & ~cpuc->intel_ctrl_host_mask; 3942 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 3943 arr[0].guest &= ~cpuc->pebs_enabled; 3944 else 3945 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 3946 *nr = 1; 3947 3948 if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) { 3949 /* 3950 * If PMU counter has PEBS enabled it is not enough to 3951 * disable counter on a guest entry since PEBS memory 3952 * write can overshoot guest entry and corrupt guest 3953 * memory. Disabling PEBS solves the problem. 3954 * 3955 * Don't do this if the CPU already enforces it. 3956 */ 3957 arr[1].msr = MSR_IA32_PEBS_ENABLE; 3958 arr[1].host = cpuc->pebs_enabled; 3959 arr[1].guest = 0; 3960 *nr = 2; 3961 } 3962 3963 return arr; 3964 } 3965 3966 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) 3967 { 3968 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3969 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3970 int idx; 3971 3972 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 3973 struct perf_event *event = cpuc->events[idx]; 3974 3975 arr[idx].msr = x86_pmu_config_addr(idx); 3976 arr[idx].host = arr[idx].guest = 0; 3977 3978 if (!test_bit(idx, cpuc->active_mask)) 3979 continue; 3980 3981 arr[idx].host = arr[idx].guest = 3982 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; 3983 3984 if (event->attr.exclude_host) 3985 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3986 else if (event->attr.exclude_guest) 3987 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3988 } 3989 3990 *nr = x86_pmu.num_counters; 3991 return arr; 3992 } 3993 3994 static void core_pmu_enable_event(struct perf_event *event) 3995 { 3996 if (!event->attr.exclude_host) 3997 x86_pmu_enable_event(event); 3998 } 3999 4000 static void core_pmu_enable_all(int added) 4001 { 4002 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4003 int idx; 4004 4005 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 4006 struct hw_perf_event *hwc = &cpuc->events[idx]->hw; 4007 4008 if (!test_bit(idx, cpuc->active_mask) || 4009 cpuc->events[idx]->attr.exclude_host) 4010 continue; 4011 4012 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 4013 } 4014 } 4015 4016 static int hsw_hw_config(struct perf_event *event) 4017 { 4018 int ret = intel_pmu_hw_config(event); 4019 4020 if (ret) 4021 return ret; 4022 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) 4023 return 0; 4024 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); 4025 4026 /* 4027 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with 4028 * PEBS or in ANY thread mode. Since the results are non-sensical forbid 4029 * this combination. 4030 */ 4031 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && 4032 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || 4033 event->attr.precise_ip > 0)) 4034 return -EOPNOTSUPP; 4035 4036 if (event_is_checkpointed(event)) { 4037 /* 4038 * Sampling of checkpointed events can cause situations where 4039 * the CPU constantly aborts because of a overflow, which is 4040 * then checkpointed back and ignored. Forbid checkpointing 4041 * for sampling. 4042 * 4043 * But still allow a long sampling period, so that perf stat 4044 * from KVM works. 4045 */ 4046 if (event->attr.sample_period > 0 && 4047 event->attr.sample_period < 0x7fffffff) 4048 return -EOPNOTSUPP; 4049 } 4050 return 0; 4051 } 4052 4053 static struct event_constraint counter0_constraint = 4054 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1); 4055 4056 static struct event_constraint counter2_constraint = 4057 EVENT_CONSTRAINT(0, 0x4, 0); 4058 4059 static struct event_constraint fixed0_constraint = 4060 FIXED_EVENT_CONSTRAINT(0x00c0, 0); 4061 4062 static struct event_constraint fixed0_counter0_constraint = 4063 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL); 4064 4065 static struct event_constraint * 4066 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4067 struct perf_event *event) 4068 { 4069 struct event_constraint *c; 4070 4071 c = intel_get_event_constraints(cpuc, idx, event); 4072 4073 /* Handle special quirk on in_tx_checkpointed only in counter 2 */ 4074 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { 4075 if (c->idxmsk64 & (1U << 2)) 4076 return &counter2_constraint; 4077 return &emptyconstraint; 4078 } 4079 4080 return c; 4081 } 4082 4083 static struct event_constraint * 4084 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4085 struct perf_event *event) 4086 { 4087 /* 4088 * Fixed counter 0 has less skid. 4089 * Force instruction:ppp in Fixed counter 0 4090 */ 4091 if ((event->attr.precise_ip == 3) && 4092 constraint_match(&fixed0_constraint, event->hw.config)) 4093 return &fixed0_constraint; 4094 4095 return hsw_get_event_constraints(cpuc, idx, event); 4096 } 4097 4098 static struct event_constraint * 4099 spr_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4100 struct perf_event *event) 4101 { 4102 struct event_constraint *c; 4103 4104 c = icl_get_event_constraints(cpuc, idx, event); 4105 4106 /* 4107 * The :ppp indicates the Precise Distribution (PDist) facility, which 4108 * is only supported on the GP counter 0. If a :ppp event which is not 4109 * available on the GP counter 0, error out. 4110 * Exception: Instruction PDIR is only available on the fixed counter 0. 4111 */ 4112 if ((event->attr.precise_ip == 3) && 4113 !constraint_match(&fixed0_constraint, event->hw.config)) { 4114 if (c->idxmsk64 & BIT_ULL(0)) 4115 return &counter0_constraint; 4116 4117 return &emptyconstraint; 4118 } 4119 4120 return c; 4121 } 4122 4123 static struct event_constraint * 4124 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4125 struct perf_event *event) 4126 { 4127 struct event_constraint *c; 4128 4129 /* :ppp means to do reduced skid PEBS which is PMC0 only. */ 4130 if (event->attr.precise_ip == 3) 4131 return &counter0_constraint; 4132 4133 c = intel_get_event_constraints(cpuc, idx, event); 4134 4135 return c; 4136 } 4137 4138 static struct event_constraint * 4139 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4140 struct perf_event *event) 4141 { 4142 struct event_constraint *c; 4143 4144 /* 4145 * :ppp means to do reduced skid PEBS, 4146 * which is available on PMC0 and fixed counter 0. 4147 */ 4148 if (event->attr.precise_ip == 3) { 4149 /* Force instruction:ppp on PMC0 and Fixed counter 0 */ 4150 if (constraint_match(&fixed0_constraint, event->hw.config)) 4151 return &fixed0_counter0_constraint; 4152 4153 return &counter0_constraint; 4154 } 4155 4156 c = intel_get_event_constraints(cpuc, idx, event); 4157 4158 return c; 4159 } 4160 4161 static bool allow_tsx_force_abort = true; 4162 4163 static struct event_constraint * 4164 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4165 struct perf_event *event) 4166 { 4167 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event); 4168 4169 /* 4170 * Without TFA we must not use PMC3. 4171 */ 4172 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) { 4173 c = dyn_constraint(cpuc, c, idx); 4174 c->idxmsk64 &= ~(1ULL << 3); 4175 c->weight--; 4176 } 4177 4178 return c; 4179 } 4180 4181 static struct event_constraint * 4182 adl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4183 struct perf_event *event) 4184 { 4185 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4186 4187 if (pmu->cpu_type == hybrid_big) 4188 return spr_get_event_constraints(cpuc, idx, event); 4189 else if (pmu->cpu_type == hybrid_small) 4190 return tnt_get_event_constraints(cpuc, idx, event); 4191 4192 WARN_ON(1); 4193 return &emptyconstraint; 4194 } 4195 4196 static int adl_hw_config(struct perf_event *event) 4197 { 4198 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4199 4200 if (pmu->cpu_type == hybrid_big) 4201 return hsw_hw_config(event); 4202 else if (pmu->cpu_type == hybrid_small) 4203 return intel_pmu_hw_config(event); 4204 4205 WARN_ON(1); 4206 return -EOPNOTSUPP; 4207 } 4208 4209 static u8 adl_get_hybrid_cpu_type(void) 4210 { 4211 return hybrid_big; 4212 } 4213 4214 /* 4215 * Broadwell: 4216 * 4217 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared 4218 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine 4219 * the two to enforce a minimum period of 128 (the smallest value that has bits 4220 * 0-5 cleared and >= 100). 4221 * 4222 * Because of how the code in x86_perf_event_set_period() works, the truncation 4223 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period 4224 * to make up for the 'lost' events due to carrying the 'error' in period_left. 4225 * 4226 * Therefore the effective (average) period matches the requested period, 4227 * despite coarser hardware granularity. 4228 */ 4229 static u64 bdw_limit_period(struct perf_event *event, u64 left) 4230 { 4231 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == 4232 X86_CONFIG(.event=0xc0, .umask=0x01)) { 4233 if (left < 128) 4234 left = 128; 4235 left &= ~0x3fULL; 4236 } 4237 return left; 4238 } 4239 4240 static u64 nhm_limit_period(struct perf_event *event, u64 left) 4241 { 4242 return max(left, 32ULL); 4243 } 4244 4245 static u64 spr_limit_period(struct perf_event *event, u64 left) 4246 { 4247 if (event->attr.precise_ip == 3) 4248 return max(left, 128ULL); 4249 4250 return left; 4251 } 4252 4253 PMU_FORMAT_ATTR(event, "config:0-7" ); 4254 PMU_FORMAT_ATTR(umask, "config:8-15" ); 4255 PMU_FORMAT_ATTR(edge, "config:18" ); 4256 PMU_FORMAT_ATTR(pc, "config:19" ); 4257 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ 4258 PMU_FORMAT_ATTR(inv, "config:23" ); 4259 PMU_FORMAT_ATTR(cmask, "config:24-31" ); 4260 PMU_FORMAT_ATTR(in_tx, "config:32"); 4261 PMU_FORMAT_ATTR(in_tx_cp, "config:33"); 4262 4263 static struct attribute *intel_arch_formats_attr[] = { 4264 &format_attr_event.attr, 4265 &format_attr_umask.attr, 4266 &format_attr_edge.attr, 4267 &format_attr_pc.attr, 4268 &format_attr_inv.attr, 4269 &format_attr_cmask.attr, 4270 NULL, 4271 }; 4272 4273 ssize_t intel_event_sysfs_show(char *page, u64 config) 4274 { 4275 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); 4276 4277 return x86_event_sysfs_show(page, config, event); 4278 } 4279 4280 static struct intel_shared_regs *allocate_shared_regs(int cpu) 4281 { 4282 struct intel_shared_regs *regs; 4283 int i; 4284 4285 regs = kzalloc_node(sizeof(struct intel_shared_regs), 4286 GFP_KERNEL, cpu_to_node(cpu)); 4287 if (regs) { 4288 /* 4289 * initialize the locks to keep lockdep happy 4290 */ 4291 for (i = 0; i < EXTRA_REG_MAX; i++) 4292 raw_spin_lock_init(®s->regs[i].lock); 4293 4294 regs->core_id = -1; 4295 } 4296 return regs; 4297 } 4298 4299 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) 4300 { 4301 struct intel_excl_cntrs *c; 4302 4303 c = kzalloc_node(sizeof(struct intel_excl_cntrs), 4304 GFP_KERNEL, cpu_to_node(cpu)); 4305 if (c) { 4306 raw_spin_lock_init(&c->lock); 4307 c->core_id = -1; 4308 } 4309 return c; 4310 } 4311 4312 4313 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 4314 { 4315 cpuc->pebs_record_size = x86_pmu.pebs_record_size; 4316 4317 if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { 4318 cpuc->shared_regs = allocate_shared_regs(cpu); 4319 if (!cpuc->shared_regs) 4320 goto err; 4321 } 4322 4323 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) { 4324 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); 4325 4326 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu)); 4327 if (!cpuc->constraint_list) 4328 goto err_shared_regs; 4329 } 4330 4331 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4332 cpuc->excl_cntrs = allocate_excl_cntrs(cpu); 4333 if (!cpuc->excl_cntrs) 4334 goto err_constraint_list; 4335 4336 cpuc->excl_thread_id = 0; 4337 } 4338 4339 return 0; 4340 4341 err_constraint_list: 4342 kfree(cpuc->constraint_list); 4343 cpuc->constraint_list = NULL; 4344 4345 err_shared_regs: 4346 kfree(cpuc->shared_regs); 4347 cpuc->shared_regs = NULL; 4348 4349 err: 4350 return -ENOMEM; 4351 } 4352 4353 static int intel_pmu_cpu_prepare(int cpu) 4354 { 4355 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu); 4356 } 4357 4358 static void flip_smm_bit(void *data) 4359 { 4360 unsigned long set = *(unsigned long *)data; 4361 4362 if (set > 0) { 4363 msr_set_bit(MSR_IA32_DEBUGCTLMSR, 4364 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4365 } else { 4366 msr_clear_bit(MSR_IA32_DEBUGCTLMSR, 4367 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4368 } 4369 } 4370 4371 static bool init_hybrid_pmu(int cpu) 4372 { 4373 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4374 u8 cpu_type = get_this_hybrid_cpu_type(); 4375 struct x86_hybrid_pmu *pmu = NULL; 4376 int i; 4377 4378 if (!cpu_type && x86_pmu.get_hybrid_cpu_type) 4379 cpu_type = x86_pmu.get_hybrid_cpu_type(); 4380 4381 for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) { 4382 if (x86_pmu.hybrid_pmu[i].cpu_type == cpu_type) { 4383 pmu = &x86_pmu.hybrid_pmu[i]; 4384 break; 4385 } 4386 } 4387 if (WARN_ON_ONCE(!pmu || (pmu->pmu.type == -1))) { 4388 cpuc->pmu = NULL; 4389 return false; 4390 } 4391 4392 /* Only check and dump the PMU information for the first CPU */ 4393 if (!cpumask_empty(&pmu->supported_cpus)) 4394 goto end; 4395 4396 if (!check_hw_exists(&pmu->pmu, pmu->num_counters, pmu->num_counters_fixed)) 4397 return false; 4398 4399 pr_info("%s PMU driver: ", pmu->name); 4400 4401 if (pmu->intel_cap.pebs_output_pt_available) 4402 pr_cont("PEBS-via-PT "); 4403 4404 pr_cont("\n"); 4405 4406 x86_pmu_show_pmu_cap(pmu->num_counters, pmu->num_counters_fixed, 4407 pmu->intel_ctrl); 4408 4409 end: 4410 cpumask_set_cpu(cpu, &pmu->supported_cpus); 4411 cpuc->pmu = &pmu->pmu; 4412 4413 x86_pmu_update_cpu_context(&pmu->pmu, cpu); 4414 4415 return true; 4416 } 4417 4418 static void intel_pmu_cpu_starting(int cpu) 4419 { 4420 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4421 int core_id = topology_core_id(cpu); 4422 int i; 4423 4424 if (is_hybrid() && !init_hybrid_pmu(cpu)) 4425 return; 4426 4427 init_debug_store_on_cpu(cpu); 4428 /* 4429 * Deal with CPUs that don't clear their LBRs on power-up. 4430 */ 4431 intel_pmu_lbr_reset(); 4432 4433 cpuc->lbr_sel = NULL; 4434 4435 if (x86_pmu.flags & PMU_FL_TFA) { 4436 WARN_ON_ONCE(cpuc->tfa_shadow); 4437 cpuc->tfa_shadow = ~0ULL; 4438 intel_set_tfa(cpuc, false); 4439 } 4440 4441 if (x86_pmu.version > 1) 4442 flip_smm_bit(&x86_pmu.attr_freeze_on_smi); 4443 4444 /* 4445 * Disable perf metrics if any added CPU doesn't support it. 4446 * 4447 * Turn off the check for a hybrid architecture, because the 4448 * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate 4449 * the architecture features. The perf metrics is a model-specific 4450 * feature for now. The corresponding bit should always be 0 on 4451 * a hybrid platform, e.g., Alder Lake. 4452 */ 4453 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) { 4454 union perf_capabilities perf_cap; 4455 4456 rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities); 4457 if (!perf_cap.perf_metrics) { 4458 x86_pmu.intel_cap.perf_metrics = 0; 4459 x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS); 4460 } 4461 } 4462 4463 if (!cpuc->shared_regs) 4464 return; 4465 4466 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { 4467 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4468 struct intel_shared_regs *pc; 4469 4470 pc = per_cpu(cpu_hw_events, i).shared_regs; 4471 if (pc && pc->core_id == core_id) { 4472 cpuc->kfree_on_online[0] = cpuc->shared_regs; 4473 cpuc->shared_regs = pc; 4474 break; 4475 } 4476 } 4477 cpuc->shared_regs->core_id = core_id; 4478 cpuc->shared_regs->refcnt++; 4479 } 4480 4481 if (x86_pmu.lbr_sel_map) 4482 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; 4483 4484 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4485 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4486 struct cpu_hw_events *sibling; 4487 struct intel_excl_cntrs *c; 4488 4489 sibling = &per_cpu(cpu_hw_events, i); 4490 c = sibling->excl_cntrs; 4491 if (c && c->core_id == core_id) { 4492 cpuc->kfree_on_online[1] = cpuc->excl_cntrs; 4493 cpuc->excl_cntrs = c; 4494 if (!sibling->excl_thread_id) 4495 cpuc->excl_thread_id = 1; 4496 break; 4497 } 4498 } 4499 cpuc->excl_cntrs->core_id = core_id; 4500 cpuc->excl_cntrs->refcnt++; 4501 } 4502 } 4503 4504 static void free_excl_cntrs(struct cpu_hw_events *cpuc) 4505 { 4506 struct intel_excl_cntrs *c; 4507 4508 c = cpuc->excl_cntrs; 4509 if (c) { 4510 if (c->core_id == -1 || --c->refcnt == 0) 4511 kfree(c); 4512 cpuc->excl_cntrs = NULL; 4513 } 4514 4515 kfree(cpuc->constraint_list); 4516 cpuc->constraint_list = NULL; 4517 } 4518 4519 static void intel_pmu_cpu_dying(int cpu) 4520 { 4521 fini_debug_store_on_cpu(cpu); 4522 } 4523 4524 void intel_cpuc_finish(struct cpu_hw_events *cpuc) 4525 { 4526 struct intel_shared_regs *pc; 4527 4528 pc = cpuc->shared_regs; 4529 if (pc) { 4530 if (pc->core_id == -1 || --pc->refcnt == 0) 4531 kfree(pc); 4532 cpuc->shared_regs = NULL; 4533 } 4534 4535 free_excl_cntrs(cpuc); 4536 } 4537 4538 static void intel_pmu_cpu_dead(int cpu) 4539 { 4540 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4541 4542 intel_cpuc_finish(cpuc); 4543 4544 if (is_hybrid() && cpuc->pmu) 4545 cpumask_clear_cpu(cpu, &hybrid_pmu(cpuc->pmu)->supported_cpus); 4546 } 4547 4548 static void intel_pmu_sched_task(struct perf_event_context *ctx, 4549 bool sched_in) 4550 { 4551 intel_pmu_pebs_sched_task(ctx, sched_in); 4552 intel_pmu_lbr_sched_task(ctx, sched_in); 4553 } 4554 4555 static void intel_pmu_swap_task_ctx(struct perf_event_context *prev, 4556 struct perf_event_context *next) 4557 { 4558 intel_pmu_lbr_swap_task_ctx(prev, next); 4559 } 4560 4561 static int intel_pmu_check_period(struct perf_event *event, u64 value) 4562 { 4563 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0; 4564 } 4565 4566 static void intel_aux_output_init(void) 4567 { 4568 /* Refer also intel_pmu_aux_output_match() */ 4569 if (x86_pmu.intel_cap.pebs_output_pt_available) 4570 x86_pmu.assign = intel_pmu_assign_event; 4571 } 4572 4573 static int intel_pmu_aux_output_match(struct perf_event *event) 4574 { 4575 /* intel_pmu_assign_event() is needed, refer intel_aux_output_init() */ 4576 if (!x86_pmu.intel_cap.pebs_output_pt_available) 4577 return 0; 4578 4579 return is_intel_pt_event(event); 4580 } 4581 4582 static int intel_pmu_filter_match(struct perf_event *event) 4583 { 4584 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4585 unsigned int cpu = smp_processor_id(); 4586 4587 return cpumask_test_cpu(cpu, &pmu->supported_cpus); 4588 } 4589 4590 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); 4591 4592 PMU_FORMAT_ATTR(ldlat, "config1:0-15"); 4593 4594 PMU_FORMAT_ATTR(frontend, "config1:0-23"); 4595 4596 static struct attribute *intel_arch3_formats_attr[] = { 4597 &format_attr_event.attr, 4598 &format_attr_umask.attr, 4599 &format_attr_edge.attr, 4600 &format_attr_pc.attr, 4601 &format_attr_any.attr, 4602 &format_attr_inv.attr, 4603 &format_attr_cmask.attr, 4604 NULL, 4605 }; 4606 4607 static struct attribute *hsw_format_attr[] = { 4608 &format_attr_in_tx.attr, 4609 &format_attr_in_tx_cp.attr, 4610 &format_attr_offcore_rsp.attr, 4611 &format_attr_ldlat.attr, 4612 NULL 4613 }; 4614 4615 static struct attribute *nhm_format_attr[] = { 4616 &format_attr_offcore_rsp.attr, 4617 &format_attr_ldlat.attr, 4618 NULL 4619 }; 4620 4621 static struct attribute *slm_format_attr[] = { 4622 &format_attr_offcore_rsp.attr, 4623 NULL 4624 }; 4625 4626 static struct attribute *skl_format_attr[] = { 4627 &format_attr_frontend.attr, 4628 NULL, 4629 }; 4630 4631 static __initconst const struct x86_pmu core_pmu = { 4632 .name = "core", 4633 .handle_irq = x86_pmu_handle_irq, 4634 .disable_all = x86_pmu_disable_all, 4635 .enable_all = core_pmu_enable_all, 4636 .enable = core_pmu_enable_event, 4637 .disable = x86_pmu_disable_event, 4638 .hw_config = core_pmu_hw_config, 4639 .schedule_events = x86_schedule_events, 4640 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 4641 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 4642 .event_map = intel_pmu_event_map, 4643 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 4644 .apic = 1, 4645 .large_pebs_flags = LARGE_PEBS_FLAGS, 4646 4647 /* 4648 * Intel PMCs cannot be accessed sanely above 32-bit width, 4649 * so we install an artificial 1<<31 period regardless of 4650 * the generic event period: 4651 */ 4652 .max_period = (1ULL<<31) - 1, 4653 .get_event_constraints = intel_get_event_constraints, 4654 .put_event_constraints = intel_put_event_constraints, 4655 .event_constraints = intel_core_event_constraints, 4656 .guest_get_msrs = core_guest_get_msrs, 4657 .format_attrs = intel_arch_formats_attr, 4658 .events_sysfs_show = intel_event_sysfs_show, 4659 4660 /* 4661 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs 4662 * together with PMU version 1 and thus be using core_pmu with 4663 * shared_regs. We need following callbacks here to allocate 4664 * it properly. 4665 */ 4666 .cpu_prepare = intel_pmu_cpu_prepare, 4667 .cpu_starting = intel_pmu_cpu_starting, 4668 .cpu_dying = intel_pmu_cpu_dying, 4669 .cpu_dead = intel_pmu_cpu_dead, 4670 4671 .check_period = intel_pmu_check_period, 4672 4673 .lbr_reset = intel_pmu_lbr_reset_64, 4674 .lbr_read = intel_pmu_lbr_read_64, 4675 .lbr_save = intel_pmu_lbr_save, 4676 .lbr_restore = intel_pmu_lbr_restore, 4677 }; 4678 4679 static __initconst const struct x86_pmu intel_pmu = { 4680 .name = "Intel", 4681 .handle_irq = intel_pmu_handle_irq, 4682 .disable_all = intel_pmu_disable_all, 4683 .enable_all = intel_pmu_enable_all, 4684 .enable = intel_pmu_enable_event, 4685 .disable = intel_pmu_disable_event, 4686 .add = intel_pmu_add_event, 4687 .del = intel_pmu_del_event, 4688 .read = intel_pmu_read_event, 4689 .hw_config = intel_pmu_hw_config, 4690 .schedule_events = x86_schedule_events, 4691 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 4692 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 4693 .event_map = intel_pmu_event_map, 4694 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 4695 .apic = 1, 4696 .large_pebs_flags = LARGE_PEBS_FLAGS, 4697 /* 4698 * Intel PMCs cannot be accessed sanely above 32 bit width, 4699 * so we install an artificial 1<<31 period regardless of 4700 * the generic event period: 4701 */ 4702 .max_period = (1ULL << 31) - 1, 4703 .get_event_constraints = intel_get_event_constraints, 4704 .put_event_constraints = intel_put_event_constraints, 4705 .pebs_aliases = intel_pebs_aliases_core2, 4706 4707 .format_attrs = intel_arch3_formats_attr, 4708 .events_sysfs_show = intel_event_sysfs_show, 4709 4710 .cpu_prepare = intel_pmu_cpu_prepare, 4711 .cpu_starting = intel_pmu_cpu_starting, 4712 .cpu_dying = intel_pmu_cpu_dying, 4713 .cpu_dead = intel_pmu_cpu_dead, 4714 4715 .guest_get_msrs = intel_guest_get_msrs, 4716 .sched_task = intel_pmu_sched_task, 4717 .swap_task_ctx = intel_pmu_swap_task_ctx, 4718 4719 .check_period = intel_pmu_check_period, 4720 4721 .aux_output_match = intel_pmu_aux_output_match, 4722 4723 .lbr_reset = intel_pmu_lbr_reset_64, 4724 .lbr_read = intel_pmu_lbr_read_64, 4725 .lbr_save = intel_pmu_lbr_save, 4726 .lbr_restore = intel_pmu_lbr_restore, 4727 4728 /* 4729 * SMM has access to all 4 rings and while traditionally SMM code only 4730 * ran in CPL0, 2021-era firmware is starting to make use of CPL3 in SMM. 4731 * 4732 * Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction 4733 * between SMM or not, this results in what should be pure userspace 4734 * counters including SMM data. 4735 * 4736 * This is a clear privilege issue, therefore globally disable 4737 * counting SMM by default. 4738 */ 4739 .attr_freeze_on_smi = 1, 4740 }; 4741 4742 static __init void intel_clovertown_quirk(void) 4743 { 4744 /* 4745 * PEBS is unreliable due to: 4746 * 4747 * AJ67 - PEBS may experience CPL leaks 4748 * AJ68 - PEBS PMI may be delayed by one event 4749 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] 4750 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS 4751 * 4752 * AJ67 could be worked around by restricting the OS/USR flags. 4753 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. 4754 * 4755 * AJ106 could possibly be worked around by not allowing LBR 4756 * usage from PEBS, including the fixup. 4757 * AJ68 could possibly be worked around by always programming 4758 * a pebs_event_reset[0] value and coping with the lost events. 4759 * 4760 * But taken together it might just make sense to not enable PEBS on 4761 * these chips. 4762 */ 4763 pr_warn("PEBS disabled due to CPU errata\n"); 4764 x86_pmu.pebs = 0; 4765 x86_pmu.pebs_constraints = NULL; 4766 } 4767 4768 static const struct x86_cpu_desc isolation_ucodes[] = { 4769 INTEL_CPU_DESC(INTEL_FAM6_HASWELL, 3, 0x0000001f), 4770 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L, 1, 0x0000001e), 4771 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G, 1, 0x00000015), 4772 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037), 4773 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a), 4774 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL, 4, 0x00000023), 4775 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G, 1, 0x00000014), 4776 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 2, 0x00000010), 4777 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 3, 0x07000009), 4778 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 4, 0x0f000009), 4779 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 5, 0x0e000002), 4780 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 1, 0x0b000014), 4781 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021), 4782 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000), 4783 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 5, 0x00000000), 4784 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 6, 0x00000000), 4785 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 7, 0x00000000), 4786 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L, 3, 0x0000007c), 4787 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE, 3, 0x0000007c), 4788 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 9, 0x0000004e), 4789 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 9, 0x0000004e), 4790 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 10, 0x0000004e), 4791 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 11, 0x0000004e), 4792 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 12, 0x0000004e), 4793 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 10, 0x0000004e), 4794 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 11, 0x0000004e), 4795 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 12, 0x0000004e), 4796 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 13, 0x0000004e), 4797 {} 4798 }; 4799 4800 static void intel_check_pebs_isolation(void) 4801 { 4802 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes); 4803 } 4804 4805 static __init void intel_pebs_isolation_quirk(void) 4806 { 4807 WARN_ON_ONCE(x86_pmu.check_microcode); 4808 x86_pmu.check_microcode = intel_check_pebs_isolation; 4809 intel_check_pebs_isolation(); 4810 } 4811 4812 static const struct x86_cpu_desc pebs_ucodes[] = { 4813 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028), 4814 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618), 4815 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c), 4816 {} 4817 }; 4818 4819 static bool intel_snb_pebs_broken(void) 4820 { 4821 return !x86_cpu_has_min_microcode_rev(pebs_ucodes); 4822 } 4823 4824 static void intel_snb_check_microcode(void) 4825 { 4826 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken) 4827 return; 4828 4829 /* 4830 * Serialized by the microcode lock.. 4831 */ 4832 if (x86_pmu.pebs_broken) { 4833 pr_info("PEBS enabled due to microcode update\n"); 4834 x86_pmu.pebs_broken = 0; 4835 } else { 4836 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); 4837 x86_pmu.pebs_broken = 1; 4838 } 4839 } 4840 4841 static bool is_lbr_from(unsigned long msr) 4842 { 4843 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr; 4844 4845 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr; 4846 } 4847 4848 /* 4849 * Under certain circumstances, access certain MSR may cause #GP. 4850 * The function tests if the input MSR can be safely accessed. 4851 */ 4852 static bool check_msr(unsigned long msr, u64 mask) 4853 { 4854 u64 val_old, val_new, val_tmp; 4855 4856 /* 4857 * Disable the check for real HW, so we don't 4858 * mess with potentially enabled registers: 4859 */ 4860 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) 4861 return true; 4862 4863 /* 4864 * Read the current value, change it and read it back to see if it 4865 * matches, this is needed to detect certain hardware emulators 4866 * (qemu/kvm) that don't trap on the MSR access and always return 0s. 4867 */ 4868 if (rdmsrl_safe(msr, &val_old)) 4869 return false; 4870 4871 /* 4872 * Only change the bits which can be updated by wrmsrl. 4873 */ 4874 val_tmp = val_old ^ mask; 4875 4876 if (is_lbr_from(msr)) 4877 val_tmp = lbr_from_signext_quirk_wr(val_tmp); 4878 4879 if (wrmsrl_safe(msr, val_tmp) || 4880 rdmsrl_safe(msr, &val_new)) 4881 return false; 4882 4883 /* 4884 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value 4885 * should equal rdmsrl()'s even with the quirk. 4886 */ 4887 if (val_new != val_tmp) 4888 return false; 4889 4890 if (is_lbr_from(msr)) 4891 val_old = lbr_from_signext_quirk_wr(val_old); 4892 4893 /* Here it's sure that the MSR can be safely accessed. 4894 * Restore the old value and return. 4895 */ 4896 wrmsrl(msr, val_old); 4897 4898 return true; 4899 } 4900 4901 static __init void intel_sandybridge_quirk(void) 4902 { 4903 x86_pmu.check_microcode = intel_snb_check_microcode; 4904 cpus_read_lock(); 4905 intel_snb_check_microcode(); 4906 cpus_read_unlock(); 4907 } 4908 4909 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { 4910 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, 4911 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, 4912 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, 4913 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, 4914 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, 4915 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, 4916 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, 4917 }; 4918 4919 static __init void intel_arch_events_quirk(void) 4920 { 4921 int bit; 4922 4923 /* disable event that reported as not present by cpuid */ 4924 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { 4925 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; 4926 pr_warn("CPUID marked event: \'%s\' unavailable\n", 4927 intel_arch_events_map[bit].name); 4928 } 4929 } 4930 4931 static __init void intel_nehalem_quirk(void) 4932 { 4933 union cpuid10_ebx ebx; 4934 4935 ebx.full = x86_pmu.events_maskl; 4936 if (ebx.split.no_branch_misses_retired) { 4937 /* 4938 * Erratum AAJ80 detected, we work it around by using 4939 * the BR_MISP_EXEC.ANY event. This will over-count 4940 * branch-misses, but it's still much better than the 4941 * architectural event which is often completely bogus: 4942 */ 4943 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 4944 ebx.split.no_branch_misses_retired = 0; 4945 x86_pmu.events_maskl = ebx.full; 4946 pr_info("CPU erratum AAJ80 worked around\n"); 4947 } 4948 } 4949 4950 /* 4951 * enable software workaround for errata: 4952 * SNB: BJ122 4953 * IVB: BV98 4954 * HSW: HSD29 4955 * 4956 * Only needed when HT is enabled. However detecting 4957 * if HT is enabled is difficult (model specific). So instead, 4958 * we enable the workaround in the early boot, and verify if 4959 * it is needed in a later initcall phase once we have valid 4960 * topology information to check if HT is actually enabled 4961 */ 4962 static __init void intel_ht_bug(void) 4963 { 4964 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; 4965 4966 x86_pmu.start_scheduling = intel_start_scheduling; 4967 x86_pmu.commit_scheduling = intel_commit_scheduling; 4968 x86_pmu.stop_scheduling = intel_stop_scheduling; 4969 } 4970 4971 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); 4972 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") 4973 4974 /* Haswell special events */ 4975 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); 4976 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); 4977 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); 4978 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); 4979 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); 4980 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); 4981 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); 4982 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); 4983 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); 4984 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); 4985 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); 4986 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); 4987 4988 static struct attribute *hsw_events_attrs[] = { 4989 EVENT_PTR(td_slots_issued), 4990 EVENT_PTR(td_slots_retired), 4991 EVENT_PTR(td_fetch_bubbles), 4992 EVENT_PTR(td_total_slots), 4993 EVENT_PTR(td_total_slots_scale), 4994 EVENT_PTR(td_recovery_bubbles), 4995 EVENT_PTR(td_recovery_bubbles_scale), 4996 NULL 4997 }; 4998 4999 static struct attribute *hsw_mem_events_attrs[] = { 5000 EVENT_PTR(mem_ld_hsw), 5001 EVENT_PTR(mem_st_hsw), 5002 NULL, 5003 }; 5004 5005 static struct attribute *hsw_tsx_events_attrs[] = { 5006 EVENT_PTR(tx_start), 5007 EVENT_PTR(tx_commit), 5008 EVENT_PTR(tx_abort), 5009 EVENT_PTR(tx_capacity), 5010 EVENT_PTR(tx_conflict), 5011 EVENT_PTR(el_start), 5012 EVENT_PTR(el_commit), 5013 EVENT_PTR(el_abort), 5014 EVENT_PTR(el_capacity), 5015 EVENT_PTR(el_conflict), 5016 EVENT_PTR(cycles_t), 5017 EVENT_PTR(cycles_ct), 5018 NULL 5019 }; 5020 5021 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80"); 5022 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2"); 5023 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80"); 5024 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2"); 5025 5026 static struct attribute *icl_events_attrs[] = { 5027 EVENT_PTR(mem_ld_hsw), 5028 EVENT_PTR(mem_st_hsw), 5029 NULL, 5030 }; 5031 5032 static struct attribute *icl_td_events_attrs[] = { 5033 EVENT_PTR(slots), 5034 EVENT_PTR(td_retiring), 5035 EVENT_PTR(td_bad_spec), 5036 EVENT_PTR(td_fe_bound), 5037 EVENT_PTR(td_be_bound), 5038 NULL, 5039 }; 5040 5041 static struct attribute *icl_tsx_events_attrs[] = { 5042 EVENT_PTR(tx_start), 5043 EVENT_PTR(tx_abort), 5044 EVENT_PTR(tx_commit), 5045 EVENT_PTR(tx_capacity_read), 5046 EVENT_PTR(tx_capacity_write), 5047 EVENT_PTR(tx_conflict), 5048 EVENT_PTR(el_start), 5049 EVENT_PTR(el_abort), 5050 EVENT_PTR(el_commit), 5051 EVENT_PTR(el_capacity_read), 5052 EVENT_PTR(el_capacity_write), 5053 EVENT_PTR(el_conflict), 5054 EVENT_PTR(cycles_t), 5055 EVENT_PTR(cycles_ct), 5056 NULL, 5057 }; 5058 5059 5060 EVENT_ATTR_STR(mem-stores, mem_st_spr, "event=0xcd,umask=0x2"); 5061 EVENT_ATTR_STR(mem-loads-aux, mem_ld_aux, "event=0x03,umask=0x82"); 5062 5063 static struct attribute *spr_events_attrs[] = { 5064 EVENT_PTR(mem_ld_hsw), 5065 EVENT_PTR(mem_st_spr), 5066 EVENT_PTR(mem_ld_aux), 5067 NULL, 5068 }; 5069 5070 static struct attribute *spr_td_events_attrs[] = { 5071 EVENT_PTR(slots), 5072 EVENT_PTR(td_retiring), 5073 EVENT_PTR(td_bad_spec), 5074 EVENT_PTR(td_fe_bound), 5075 EVENT_PTR(td_be_bound), 5076 EVENT_PTR(td_heavy_ops), 5077 EVENT_PTR(td_br_mispredict), 5078 EVENT_PTR(td_fetch_lat), 5079 EVENT_PTR(td_mem_bound), 5080 NULL, 5081 }; 5082 5083 static struct attribute *spr_tsx_events_attrs[] = { 5084 EVENT_PTR(tx_start), 5085 EVENT_PTR(tx_abort), 5086 EVENT_PTR(tx_commit), 5087 EVENT_PTR(tx_capacity_read), 5088 EVENT_PTR(tx_capacity_write), 5089 EVENT_PTR(tx_conflict), 5090 EVENT_PTR(cycles_t), 5091 EVENT_PTR(cycles_ct), 5092 NULL, 5093 }; 5094 5095 static ssize_t freeze_on_smi_show(struct device *cdev, 5096 struct device_attribute *attr, 5097 char *buf) 5098 { 5099 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi); 5100 } 5101 5102 static DEFINE_MUTEX(freeze_on_smi_mutex); 5103 5104 static ssize_t freeze_on_smi_store(struct device *cdev, 5105 struct device_attribute *attr, 5106 const char *buf, size_t count) 5107 { 5108 unsigned long val; 5109 ssize_t ret; 5110 5111 ret = kstrtoul(buf, 0, &val); 5112 if (ret) 5113 return ret; 5114 5115 if (val > 1) 5116 return -EINVAL; 5117 5118 mutex_lock(&freeze_on_smi_mutex); 5119 5120 if (x86_pmu.attr_freeze_on_smi == val) 5121 goto done; 5122 5123 x86_pmu.attr_freeze_on_smi = val; 5124 5125 cpus_read_lock(); 5126 on_each_cpu(flip_smm_bit, &val, 1); 5127 cpus_read_unlock(); 5128 done: 5129 mutex_unlock(&freeze_on_smi_mutex); 5130 5131 return count; 5132 } 5133 5134 static void update_tfa_sched(void *ignored) 5135 { 5136 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 5137 5138 /* 5139 * check if PMC3 is used 5140 * and if so force schedule out for all event types all contexts 5141 */ 5142 if (test_bit(3, cpuc->active_mask)) 5143 perf_pmu_resched(x86_get_pmu(smp_processor_id())); 5144 } 5145 5146 static ssize_t show_sysctl_tfa(struct device *cdev, 5147 struct device_attribute *attr, 5148 char *buf) 5149 { 5150 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort); 5151 } 5152 5153 static ssize_t set_sysctl_tfa(struct device *cdev, 5154 struct device_attribute *attr, 5155 const char *buf, size_t count) 5156 { 5157 bool val; 5158 ssize_t ret; 5159 5160 ret = kstrtobool(buf, &val); 5161 if (ret) 5162 return ret; 5163 5164 /* no change */ 5165 if (val == allow_tsx_force_abort) 5166 return count; 5167 5168 allow_tsx_force_abort = val; 5169 5170 cpus_read_lock(); 5171 on_each_cpu(update_tfa_sched, NULL, 1); 5172 cpus_read_unlock(); 5173 5174 return count; 5175 } 5176 5177 5178 static DEVICE_ATTR_RW(freeze_on_smi); 5179 5180 static ssize_t branches_show(struct device *cdev, 5181 struct device_attribute *attr, 5182 char *buf) 5183 { 5184 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr); 5185 } 5186 5187 static DEVICE_ATTR_RO(branches); 5188 5189 static struct attribute *lbr_attrs[] = { 5190 &dev_attr_branches.attr, 5191 NULL 5192 }; 5193 5194 static char pmu_name_str[30]; 5195 5196 static ssize_t pmu_name_show(struct device *cdev, 5197 struct device_attribute *attr, 5198 char *buf) 5199 { 5200 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str); 5201 } 5202 5203 static DEVICE_ATTR_RO(pmu_name); 5204 5205 static struct attribute *intel_pmu_caps_attrs[] = { 5206 &dev_attr_pmu_name.attr, 5207 NULL 5208 }; 5209 5210 static DEVICE_ATTR(allow_tsx_force_abort, 0644, 5211 show_sysctl_tfa, 5212 set_sysctl_tfa); 5213 5214 static struct attribute *intel_pmu_attrs[] = { 5215 &dev_attr_freeze_on_smi.attr, 5216 &dev_attr_allow_tsx_force_abort.attr, 5217 NULL, 5218 }; 5219 5220 static umode_t 5221 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5222 { 5223 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0; 5224 } 5225 5226 static umode_t 5227 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5228 { 5229 return x86_pmu.pebs ? attr->mode : 0; 5230 } 5231 5232 static umode_t 5233 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5234 { 5235 return x86_pmu.lbr_nr ? attr->mode : 0; 5236 } 5237 5238 static umode_t 5239 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5240 { 5241 return x86_pmu.version >= 2 ? attr->mode : 0; 5242 } 5243 5244 static umode_t 5245 default_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5246 { 5247 if (attr == &dev_attr_allow_tsx_force_abort.attr) 5248 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0; 5249 5250 return attr->mode; 5251 } 5252 5253 static struct attribute_group group_events_td = { 5254 .name = "events", 5255 }; 5256 5257 static struct attribute_group group_events_mem = { 5258 .name = "events", 5259 .is_visible = pebs_is_visible, 5260 }; 5261 5262 static struct attribute_group group_events_tsx = { 5263 .name = "events", 5264 .is_visible = tsx_is_visible, 5265 }; 5266 5267 static struct attribute_group group_caps_gen = { 5268 .name = "caps", 5269 .attrs = intel_pmu_caps_attrs, 5270 }; 5271 5272 static struct attribute_group group_caps_lbr = { 5273 .name = "caps", 5274 .attrs = lbr_attrs, 5275 .is_visible = lbr_is_visible, 5276 }; 5277 5278 static struct attribute_group group_format_extra = { 5279 .name = "format", 5280 .is_visible = exra_is_visible, 5281 }; 5282 5283 static struct attribute_group group_format_extra_skl = { 5284 .name = "format", 5285 .is_visible = exra_is_visible, 5286 }; 5287 5288 static struct attribute_group group_default = { 5289 .attrs = intel_pmu_attrs, 5290 .is_visible = default_is_visible, 5291 }; 5292 5293 static const struct attribute_group *attr_update[] = { 5294 &group_events_td, 5295 &group_events_mem, 5296 &group_events_tsx, 5297 &group_caps_gen, 5298 &group_caps_lbr, 5299 &group_format_extra, 5300 &group_format_extra_skl, 5301 &group_default, 5302 NULL, 5303 }; 5304 5305 EVENT_ATTR_STR_HYBRID(slots, slots_adl, "event=0x00,umask=0x4", hybrid_big); 5306 EVENT_ATTR_STR_HYBRID(topdown-retiring, td_retiring_adl, "event=0xc2,umask=0x0;event=0x00,umask=0x80", hybrid_big_small); 5307 EVENT_ATTR_STR_HYBRID(topdown-bad-spec, td_bad_spec_adl, "event=0x73,umask=0x0;event=0x00,umask=0x81", hybrid_big_small); 5308 EVENT_ATTR_STR_HYBRID(topdown-fe-bound, td_fe_bound_adl, "event=0x71,umask=0x0;event=0x00,umask=0x82", hybrid_big_small); 5309 EVENT_ATTR_STR_HYBRID(topdown-be-bound, td_be_bound_adl, "event=0x74,umask=0x0;event=0x00,umask=0x83", hybrid_big_small); 5310 EVENT_ATTR_STR_HYBRID(topdown-heavy-ops, td_heavy_ops_adl, "event=0x00,umask=0x84", hybrid_big); 5311 EVENT_ATTR_STR_HYBRID(topdown-br-mispredict, td_br_mis_adl, "event=0x00,umask=0x85", hybrid_big); 5312 EVENT_ATTR_STR_HYBRID(topdown-fetch-lat, td_fetch_lat_adl, "event=0x00,umask=0x86", hybrid_big); 5313 EVENT_ATTR_STR_HYBRID(topdown-mem-bound, td_mem_bound_adl, "event=0x00,umask=0x87", hybrid_big); 5314 5315 static struct attribute *adl_hybrid_events_attrs[] = { 5316 EVENT_PTR(slots_adl), 5317 EVENT_PTR(td_retiring_adl), 5318 EVENT_PTR(td_bad_spec_adl), 5319 EVENT_PTR(td_fe_bound_adl), 5320 EVENT_PTR(td_be_bound_adl), 5321 EVENT_PTR(td_heavy_ops_adl), 5322 EVENT_PTR(td_br_mis_adl), 5323 EVENT_PTR(td_fetch_lat_adl), 5324 EVENT_PTR(td_mem_bound_adl), 5325 NULL, 5326 }; 5327 5328 /* Must be in IDX order */ 5329 EVENT_ATTR_STR_HYBRID(mem-loads, mem_ld_adl, "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3", hybrid_big_small); 5330 EVENT_ATTR_STR_HYBRID(mem-stores, mem_st_adl, "event=0xd0,umask=0x6;event=0xcd,umask=0x2", hybrid_big_small); 5331 EVENT_ATTR_STR_HYBRID(mem-loads-aux, mem_ld_aux_adl, "event=0x03,umask=0x82", hybrid_big); 5332 5333 static struct attribute *adl_hybrid_mem_attrs[] = { 5334 EVENT_PTR(mem_ld_adl), 5335 EVENT_PTR(mem_st_adl), 5336 EVENT_PTR(mem_ld_aux_adl), 5337 NULL, 5338 }; 5339 5340 EVENT_ATTR_STR_HYBRID(tx-start, tx_start_adl, "event=0xc9,umask=0x1", hybrid_big); 5341 EVENT_ATTR_STR_HYBRID(tx-commit, tx_commit_adl, "event=0xc9,umask=0x2", hybrid_big); 5342 EVENT_ATTR_STR_HYBRID(tx-abort, tx_abort_adl, "event=0xc9,umask=0x4", hybrid_big); 5343 EVENT_ATTR_STR_HYBRID(tx-conflict, tx_conflict_adl, "event=0x54,umask=0x1", hybrid_big); 5344 EVENT_ATTR_STR_HYBRID(cycles-t, cycles_t_adl, "event=0x3c,in_tx=1", hybrid_big); 5345 EVENT_ATTR_STR_HYBRID(cycles-ct, cycles_ct_adl, "event=0x3c,in_tx=1,in_tx_cp=1", hybrid_big); 5346 EVENT_ATTR_STR_HYBRID(tx-capacity-read, tx_capacity_read_adl, "event=0x54,umask=0x80", hybrid_big); 5347 EVENT_ATTR_STR_HYBRID(tx-capacity-write, tx_capacity_write_adl, "event=0x54,umask=0x2", hybrid_big); 5348 5349 static struct attribute *adl_hybrid_tsx_attrs[] = { 5350 EVENT_PTR(tx_start_adl), 5351 EVENT_PTR(tx_abort_adl), 5352 EVENT_PTR(tx_commit_adl), 5353 EVENT_PTR(tx_capacity_read_adl), 5354 EVENT_PTR(tx_capacity_write_adl), 5355 EVENT_PTR(tx_conflict_adl), 5356 EVENT_PTR(cycles_t_adl), 5357 EVENT_PTR(cycles_ct_adl), 5358 NULL, 5359 }; 5360 5361 FORMAT_ATTR_HYBRID(in_tx, hybrid_big); 5362 FORMAT_ATTR_HYBRID(in_tx_cp, hybrid_big); 5363 FORMAT_ATTR_HYBRID(offcore_rsp, hybrid_big_small); 5364 FORMAT_ATTR_HYBRID(ldlat, hybrid_big_small); 5365 FORMAT_ATTR_HYBRID(frontend, hybrid_big); 5366 5367 static struct attribute *adl_hybrid_extra_attr_rtm[] = { 5368 FORMAT_HYBRID_PTR(in_tx), 5369 FORMAT_HYBRID_PTR(in_tx_cp), 5370 FORMAT_HYBRID_PTR(offcore_rsp), 5371 FORMAT_HYBRID_PTR(ldlat), 5372 FORMAT_HYBRID_PTR(frontend), 5373 NULL, 5374 }; 5375 5376 static struct attribute *adl_hybrid_extra_attr[] = { 5377 FORMAT_HYBRID_PTR(offcore_rsp), 5378 FORMAT_HYBRID_PTR(ldlat), 5379 FORMAT_HYBRID_PTR(frontend), 5380 NULL, 5381 }; 5382 5383 static bool is_attr_for_this_pmu(struct kobject *kobj, struct attribute *attr) 5384 { 5385 struct device *dev = kobj_to_dev(kobj); 5386 struct x86_hybrid_pmu *pmu = 5387 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5388 struct perf_pmu_events_hybrid_attr *pmu_attr = 5389 container_of(attr, struct perf_pmu_events_hybrid_attr, attr.attr); 5390 5391 return pmu->cpu_type & pmu_attr->pmu_type; 5392 } 5393 5394 static umode_t hybrid_events_is_visible(struct kobject *kobj, 5395 struct attribute *attr, int i) 5396 { 5397 return is_attr_for_this_pmu(kobj, attr) ? attr->mode : 0; 5398 } 5399 5400 static inline int hybrid_find_supported_cpu(struct x86_hybrid_pmu *pmu) 5401 { 5402 int cpu = cpumask_first(&pmu->supported_cpus); 5403 5404 return (cpu >= nr_cpu_ids) ? -1 : cpu; 5405 } 5406 5407 static umode_t hybrid_tsx_is_visible(struct kobject *kobj, 5408 struct attribute *attr, int i) 5409 { 5410 struct device *dev = kobj_to_dev(kobj); 5411 struct x86_hybrid_pmu *pmu = 5412 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5413 int cpu = hybrid_find_supported_cpu(pmu); 5414 5415 return (cpu >= 0) && is_attr_for_this_pmu(kobj, attr) && cpu_has(&cpu_data(cpu), X86_FEATURE_RTM) ? attr->mode : 0; 5416 } 5417 5418 static umode_t hybrid_format_is_visible(struct kobject *kobj, 5419 struct attribute *attr, int i) 5420 { 5421 struct device *dev = kobj_to_dev(kobj); 5422 struct x86_hybrid_pmu *pmu = 5423 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5424 struct perf_pmu_format_hybrid_attr *pmu_attr = 5425 container_of(attr, struct perf_pmu_format_hybrid_attr, attr.attr); 5426 int cpu = hybrid_find_supported_cpu(pmu); 5427 5428 return (cpu >= 0) && (pmu->cpu_type & pmu_attr->pmu_type) ? attr->mode : 0; 5429 } 5430 5431 static struct attribute_group hybrid_group_events_td = { 5432 .name = "events", 5433 .is_visible = hybrid_events_is_visible, 5434 }; 5435 5436 static struct attribute_group hybrid_group_events_mem = { 5437 .name = "events", 5438 .is_visible = hybrid_events_is_visible, 5439 }; 5440 5441 static struct attribute_group hybrid_group_events_tsx = { 5442 .name = "events", 5443 .is_visible = hybrid_tsx_is_visible, 5444 }; 5445 5446 static struct attribute_group hybrid_group_format_extra = { 5447 .name = "format", 5448 .is_visible = hybrid_format_is_visible, 5449 }; 5450 5451 static ssize_t intel_hybrid_get_attr_cpus(struct device *dev, 5452 struct device_attribute *attr, 5453 char *buf) 5454 { 5455 struct x86_hybrid_pmu *pmu = 5456 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5457 5458 return cpumap_print_to_pagebuf(true, buf, &pmu->supported_cpus); 5459 } 5460 5461 static DEVICE_ATTR(cpus, S_IRUGO, intel_hybrid_get_attr_cpus, NULL); 5462 static struct attribute *intel_hybrid_cpus_attrs[] = { 5463 &dev_attr_cpus.attr, 5464 NULL, 5465 }; 5466 5467 static struct attribute_group hybrid_group_cpus = { 5468 .attrs = intel_hybrid_cpus_attrs, 5469 }; 5470 5471 static const struct attribute_group *hybrid_attr_update[] = { 5472 &hybrid_group_events_td, 5473 &hybrid_group_events_mem, 5474 &hybrid_group_events_tsx, 5475 &group_caps_gen, 5476 &group_caps_lbr, 5477 &hybrid_group_format_extra, 5478 &group_default, 5479 &hybrid_group_cpus, 5480 NULL, 5481 }; 5482 5483 static struct attribute *empty_attrs; 5484 5485 static void intel_pmu_check_num_counters(int *num_counters, 5486 int *num_counters_fixed, 5487 u64 *intel_ctrl, u64 fixed_mask) 5488 { 5489 if (*num_counters > INTEL_PMC_MAX_GENERIC) { 5490 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", 5491 *num_counters, INTEL_PMC_MAX_GENERIC); 5492 *num_counters = INTEL_PMC_MAX_GENERIC; 5493 } 5494 *intel_ctrl = (1ULL << *num_counters) - 1; 5495 5496 if (*num_counters_fixed > INTEL_PMC_MAX_FIXED) { 5497 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", 5498 *num_counters_fixed, INTEL_PMC_MAX_FIXED); 5499 *num_counters_fixed = INTEL_PMC_MAX_FIXED; 5500 } 5501 5502 *intel_ctrl |= fixed_mask << INTEL_PMC_IDX_FIXED; 5503 } 5504 5505 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints, 5506 int num_counters, 5507 int num_counters_fixed, 5508 u64 intel_ctrl) 5509 { 5510 struct event_constraint *c; 5511 5512 if (!event_constraints) 5513 return; 5514 5515 /* 5516 * event on fixed counter2 (REF_CYCLES) only works on this 5517 * counter, so do not extend mask to generic counters 5518 */ 5519 for_each_event_constraint(c, event_constraints) { 5520 /* 5521 * Don't extend the topdown slots and metrics 5522 * events to the generic counters. 5523 */ 5524 if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) { 5525 /* 5526 * Disable topdown slots and metrics events, 5527 * if slots event is not in CPUID. 5528 */ 5529 if (!(INTEL_PMC_MSK_FIXED_SLOTS & intel_ctrl)) 5530 c->idxmsk64 = 0; 5531 c->weight = hweight64(c->idxmsk64); 5532 continue; 5533 } 5534 5535 if (c->cmask == FIXED_EVENT_FLAGS) { 5536 /* Disabled fixed counters which are not in CPUID */ 5537 c->idxmsk64 &= intel_ctrl; 5538 5539 /* 5540 * Don't extend the pseudo-encoding to the 5541 * generic counters 5542 */ 5543 if (!use_fixed_pseudo_encoding(c->code)) 5544 c->idxmsk64 |= (1ULL << num_counters) - 1; 5545 } 5546 c->idxmsk64 &= 5547 ~(~0ULL << (INTEL_PMC_IDX_FIXED + num_counters_fixed)); 5548 c->weight = hweight64(c->idxmsk64); 5549 } 5550 } 5551 5552 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs) 5553 { 5554 struct extra_reg *er; 5555 5556 /* 5557 * Access extra MSR may cause #GP under certain circumstances. 5558 * E.g. KVM doesn't support offcore event 5559 * Check all extra_regs here. 5560 */ 5561 if (!extra_regs) 5562 return; 5563 5564 for (er = extra_regs; er->msr; er++) { 5565 er->extra_msr_access = check_msr(er->msr, 0x11UL); 5566 /* Disable LBR select mapping */ 5567 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) 5568 x86_pmu.lbr_sel_map = NULL; 5569 } 5570 } 5571 5572 static void intel_pmu_check_hybrid_pmus(u64 fixed_mask) 5573 { 5574 struct x86_hybrid_pmu *pmu; 5575 int i; 5576 5577 for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) { 5578 pmu = &x86_pmu.hybrid_pmu[i]; 5579 5580 intel_pmu_check_num_counters(&pmu->num_counters, 5581 &pmu->num_counters_fixed, 5582 &pmu->intel_ctrl, 5583 fixed_mask); 5584 5585 if (pmu->intel_cap.perf_metrics) { 5586 pmu->intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 5587 pmu->intel_ctrl |= INTEL_PMC_MSK_FIXED_SLOTS; 5588 } 5589 5590 if (pmu->intel_cap.pebs_output_pt_available) 5591 pmu->pmu.capabilities |= PERF_PMU_CAP_AUX_OUTPUT; 5592 5593 intel_pmu_check_event_constraints(pmu->event_constraints, 5594 pmu->num_counters, 5595 pmu->num_counters_fixed, 5596 pmu->intel_ctrl); 5597 5598 intel_pmu_check_extra_regs(pmu->extra_regs); 5599 } 5600 } 5601 5602 __init int intel_pmu_init(void) 5603 { 5604 struct attribute **extra_skl_attr = &empty_attrs; 5605 struct attribute **extra_attr = &empty_attrs; 5606 struct attribute **td_attr = &empty_attrs; 5607 struct attribute **mem_attr = &empty_attrs; 5608 struct attribute **tsx_attr = &empty_attrs; 5609 union cpuid10_edx edx; 5610 union cpuid10_eax eax; 5611 union cpuid10_ebx ebx; 5612 unsigned int fixed_mask; 5613 bool pmem = false; 5614 int version, i; 5615 char *name; 5616 struct x86_hybrid_pmu *pmu; 5617 5618 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 5619 switch (boot_cpu_data.x86) { 5620 case 0x6: 5621 return p6_pmu_init(); 5622 case 0xb: 5623 return knc_pmu_init(); 5624 case 0xf: 5625 return p4_pmu_init(); 5626 } 5627 return -ENODEV; 5628 } 5629 5630 /* 5631 * Check whether the Architectural PerfMon supports 5632 * Branch Misses Retired hw_event or not. 5633 */ 5634 cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full); 5635 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) 5636 return -ENODEV; 5637 5638 version = eax.split.version_id; 5639 if (version < 2) 5640 x86_pmu = core_pmu; 5641 else 5642 x86_pmu = intel_pmu; 5643 5644 x86_pmu.version = version; 5645 x86_pmu.num_counters = eax.split.num_counters; 5646 x86_pmu.cntval_bits = eax.split.bit_width; 5647 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; 5648 5649 x86_pmu.events_maskl = ebx.full; 5650 x86_pmu.events_mask_len = eax.split.mask_length; 5651 5652 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); 5653 5654 /* 5655 * Quirk: v2 perfmon does not report fixed-purpose events, so 5656 * assume at least 3 events, when not running in a hypervisor: 5657 */ 5658 if (version > 1 && version < 5) { 5659 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR); 5660 5661 x86_pmu.num_counters_fixed = 5662 max((int)edx.split.num_counters_fixed, assume); 5663 5664 fixed_mask = (1L << x86_pmu.num_counters_fixed) - 1; 5665 } else if (version >= 5) 5666 x86_pmu.num_counters_fixed = fls(fixed_mask); 5667 5668 if (boot_cpu_has(X86_FEATURE_PDCM)) { 5669 u64 capabilities; 5670 5671 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); 5672 x86_pmu.intel_cap.capabilities = capabilities; 5673 } 5674 5675 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) { 5676 x86_pmu.lbr_reset = intel_pmu_lbr_reset_32; 5677 x86_pmu.lbr_read = intel_pmu_lbr_read_32; 5678 } 5679 5680 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) 5681 intel_pmu_arch_lbr_init(); 5682 5683 intel_ds_init(); 5684 5685 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ 5686 5687 if (version >= 5) { 5688 x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated; 5689 if (x86_pmu.intel_cap.anythread_deprecated) 5690 pr_cont(" AnyThread deprecated, "); 5691 } 5692 5693 /* 5694 * Install the hw-cache-events table: 5695 */ 5696 switch (boot_cpu_data.x86_model) { 5697 case INTEL_FAM6_CORE_YONAH: 5698 pr_cont("Core events, "); 5699 name = "core"; 5700 break; 5701 5702 case INTEL_FAM6_CORE2_MEROM: 5703 x86_add_quirk(intel_clovertown_quirk); 5704 fallthrough; 5705 5706 case INTEL_FAM6_CORE2_MEROM_L: 5707 case INTEL_FAM6_CORE2_PENRYN: 5708 case INTEL_FAM6_CORE2_DUNNINGTON: 5709 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, 5710 sizeof(hw_cache_event_ids)); 5711 5712 intel_pmu_lbr_init_core(); 5713 5714 x86_pmu.event_constraints = intel_core2_event_constraints; 5715 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; 5716 pr_cont("Core2 events, "); 5717 name = "core2"; 5718 break; 5719 5720 case INTEL_FAM6_NEHALEM: 5721 case INTEL_FAM6_NEHALEM_EP: 5722 case INTEL_FAM6_NEHALEM_EX: 5723 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, 5724 sizeof(hw_cache_event_ids)); 5725 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 5726 sizeof(hw_cache_extra_regs)); 5727 5728 intel_pmu_lbr_init_nhm(); 5729 5730 x86_pmu.event_constraints = intel_nehalem_event_constraints; 5731 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; 5732 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 5733 x86_pmu.extra_regs = intel_nehalem_extra_regs; 5734 x86_pmu.limit_period = nhm_limit_period; 5735 5736 mem_attr = nhm_mem_events_attrs; 5737 5738 /* UOPS_ISSUED.STALLED_CYCLES */ 5739 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 5740 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 5741 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 5742 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 5743 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 5744 5745 intel_pmu_pebs_data_source_nhm(); 5746 x86_add_quirk(intel_nehalem_quirk); 5747 x86_pmu.pebs_no_tlb = 1; 5748 extra_attr = nhm_format_attr; 5749 5750 pr_cont("Nehalem events, "); 5751 name = "nehalem"; 5752 break; 5753 5754 case INTEL_FAM6_ATOM_BONNELL: 5755 case INTEL_FAM6_ATOM_BONNELL_MID: 5756 case INTEL_FAM6_ATOM_SALTWELL: 5757 case INTEL_FAM6_ATOM_SALTWELL_MID: 5758 case INTEL_FAM6_ATOM_SALTWELL_TABLET: 5759 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, 5760 sizeof(hw_cache_event_ids)); 5761 5762 intel_pmu_lbr_init_atom(); 5763 5764 x86_pmu.event_constraints = intel_gen_event_constraints; 5765 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; 5766 x86_pmu.pebs_aliases = intel_pebs_aliases_core2; 5767 pr_cont("Atom events, "); 5768 name = "bonnell"; 5769 break; 5770 5771 case INTEL_FAM6_ATOM_SILVERMONT: 5772 case INTEL_FAM6_ATOM_SILVERMONT_D: 5773 case INTEL_FAM6_ATOM_SILVERMONT_MID: 5774 case INTEL_FAM6_ATOM_AIRMONT: 5775 case INTEL_FAM6_ATOM_AIRMONT_MID: 5776 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, 5777 sizeof(hw_cache_event_ids)); 5778 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, 5779 sizeof(hw_cache_extra_regs)); 5780 5781 intel_pmu_lbr_init_slm(); 5782 5783 x86_pmu.event_constraints = intel_slm_event_constraints; 5784 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 5785 x86_pmu.extra_regs = intel_slm_extra_regs; 5786 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5787 td_attr = slm_events_attrs; 5788 extra_attr = slm_format_attr; 5789 pr_cont("Silvermont events, "); 5790 name = "silvermont"; 5791 break; 5792 5793 case INTEL_FAM6_ATOM_GOLDMONT: 5794 case INTEL_FAM6_ATOM_GOLDMONT_D: 5795 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids, 5796 sizeof(hw_cache_event_ids)); 5797 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs, 5798 sizeof(hw_cache_extra_regs)); 5799 5800 intel_pmu_lbr_init_skl(); 5801 5802 x86_pmu.event_constraints = intel_slm_event_constraints; 5803 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints; 5804 x86_pmu.extra_regs = intel_glm_extra_regs; 5805 /* 5806 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5807 * for precise cycles. 5808 * :pp is identical to :ppp 5809 */ 5810 x86_pmu.pebs_aliases = NULL; 5811 x86_pmu.pebs_prec_dist = true; 5812 x86_pmu.lbr_pt_coexist = true; 5813 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5814 td_attr = glm_events_attrs; 5815 extra_attr = slm_format_attr; 5816 pr_cont("Goldmont events, "); 5817 name = "goldmont"; 5818 break; 5819 5820 case INTEL_FAM6_ATOM_GOLDMONT_PLUS: 5821 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 5822 sizeof(hw_cache_event_ids)); 5823 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs, 5824 sizeof(hw_cache_extra_regs)); 5825 5826 intel_pmu_lbr_init_skl(); 5827 5828 x86_pmu.event_constraints = intel_slm_event_constraints; 5829 x86_pmu.extra_regs = intel_glm_extra_regs; 5830 /* 5831 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5832 * for precise cycles. 5833 */ 5834 x86_pmu.pebs_aliases = NULL; 5835 x86_pmu.pebs_prec_dist = true; 5836 x86_pmu.lbr_pt_coexist = true; 5837 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5838 x86_pmu.flags |= PMU_FL_PEBS_ALL; 5839 x86_pmu.get_event_constraints = glp_get_event_constraints; 5840 td_attr = glm_events_attrs; 5841 /* Goldmont Plus has 4-wide pipeline */ 5842 event_attr_td_total_slots_scale_glm.event_str = "4"; 5843 extra_attr = slm_format_attr; 5844 pr_cont("Goldmont plus events, "); 5845 name = "goldmont_plus"; 5846 break; 5847 5848 case INTEL_FAM6_ATOM_TREMONT_D: 5849 case INTEL_FAM6_ATOM_TREMONT: 5850 case INTEL_FAM6_ATOM_TREMONT_L: 5851 x86_pmu.late_ack = true; 5852 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 5853 sizeof(hw_cache_event_ids)); 5854 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 5855 sizeof(hw_cache_extra_regs)); 5856 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 5857 5858 intel_pmu_lbr_init_skl(); 5859 5860 x86_pmu.event_constraints = intel_slm_event_constraints; 5861 x86_pmu.extra_regs = intel_tnt_extra_regs; 5862 /* 5863 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5864 * for precise cycles. 5865 */ 5866 x86_pmu.pebs_aliases = NULL; 5867 x86_pmu.pebs_prec_dist = true; 5868 x86_pmu.lbr_pt_coexist = true; 5869 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5870 x86_pmu.get_event_constraints = tnt_get_event_constraints; 5871 td_attr = tnt_events_attrs; 5872 extra_attr = slm_format_attr; 5873 pr_cont("Tremont events, "); 5874 name = "Tremont"; 5875 break; 5876 5877 case INTEL_FAM6_WESTMERE: 5878 case INTEL_FAM6_WESTMERE_EP: 5879 case INTEL_FAM6_WESTMERE_EX: 5880 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, 5881 sizeof(hw_cache_event_ids)); 5882 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 5883 sizeof(hw_cache_extra_regs)); 5884 5885 intel_pmu_lbr_init_nhm(); 5886 5887 x86_pmu.event_constraints = intel_westmere_event_constraints; 5888 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 5889 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; 5890 x86_pmu.extra_regs = intel_westmere_extra_regs; 5891 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5892 5893 mem_attr = nhm_mem_events_attrs; 5894 5895 /* UOPS_ISSUED.STALLED_CYCLES */ 5896 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 5897 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 5898 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 5899 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 5900 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 5901 5902 intel_pmu_pebs_data_source_nhm(); 5903 extra_attr = nhm_format_attr; 5904 pr_cont("Westmere events, "); 5905 name = "westmere"; 5906 break; 5907 5908 case INTEL_FAM6_SANDYBRIDGE: 5909 case INTEL_FAM6_SANDYBRIDGE_X: 5910 x86_add_quirk(intel_sandybridge_quirk); 5911 x86_add_quirk(intel_ht_bug); 5912 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 5913 sizeof(hw_cache_event_ids)); 5914 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 5915 sizeof(hw_cache_extra_regs)); 5916 5917 intel_pmu_lbr_init_snb(); 5918 5919 x86_pmu.event_constraints = intel_snb_event_constraints; 5920 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; 5921 x86_pmu.pebs_aliases = intel_pebs_aliases_snb; 5922 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X) 5923 x86_pmu.extra_regs = intel_snbep_extra_regs; 5924 else 5925 x86_pmu.extra_regs = intel_snb_extra_regs; 5926 5927 5928 /* all extra regs are per-cpu when HT is on */ 5929 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5930 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 5931 5932 td_attr = snb_events_attrs; 5933 mem_attr = snb_mem_events_attrs; 5934 5935 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 5936 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 5937 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 5938 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ 5939 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 5940 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); 5941 5942 extra_attr = nhm_format_attr; 5943 5944 pr_cont("SandyBridge events, "); 5945 name = "sandybridge"; 5946 break; 5947 5948 case INTEL_FAM6_IVYBRIDGE: 5949 case INTEL_FAM6_IVYBRIDGE_X: 5950 x86_add_quirk(intel_ht_bug); 5951 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 5952 sizeof(hw_cache_event_ids)); 5953 /* dTLB-load-misses on IVB is different than SNB */ 5954 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ 5955 5956 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 5957 sizeof(hw_cache_extra_regs)); 5958 5959 intel_pmu_lbr_init_snb(); 5960 5961 x86_pmu.event_constraints = intel_ivb_event_constraints; 5962 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; 5963 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 5964 x86_pmu.pebs_prec_dist = true; 5965 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X) 5966 x86_pmu.extra_regs = intel_snbep_extra_regs; 5967 else 5968 x86_pmu.extra_regs = intel_snb_extra_regs; 5969 /* all extra regs are per-cpu when HT is on */ 5970 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5971 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 5972 5973 td_attr = snb_events_attrs; 5974 mem_attr = snb_mem_events_attrs; 5975 5976 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 5977 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 5978 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 5979 5980 extra_attr = nhm_format_attr; 5981 5982 pr_cont("IvyBridge events, "); 5983 name = "ivybridge"; 5984 break; 5985 5986 5987 case INTEL_FAM6_HASWELL: 5988 case INTEL_FAM6_HASWELL_X: 5989 case INTEL_FAM6_HASWELL_L: 5990 case INTEL_FAM6_HASWELL_G: 5991 x86_add_quirk(intel_ht_bug); 5992 x86_add_quirk(intel_pebs_isolation_quirk); 5993 x86_pmu.late_ack = true; 5994 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 5995 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 5996 5997 intel_pmu_lbr_init_hsw(); 5998 5999 x86_pmu.event_constraints = intel_hsw_event_constraints; 6000 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; 6001 x86_pmu.extra_regs = intel_snbep_extra_regs; 6002 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6003 x86_pmu.pebs_prec_dist = true; 6004 /* all extra regs are per-cpu when HT is on */ 6005 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6006 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6007 6008 x86_pmu.hw_config = hsw_hw_config; 6009 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6010 x86_pmu.lbr_double_abort = true; 6011 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6012 hsw_format_attr : nhm_format_attr; 6013 td_attr = hsw_events_attrs; 6014 mem_attr = hsw_mem_events_attrs; 6015 tsx_attr = hsw_tsx_events_attrs; 6016 pr_cont("Haswell events, "); 6017 name = "haswell"; 6018 break; 6019 6020 case INTEL_FAM6_BROADWELL: 6021 case INTEL_FAM6_BROADWELL_D: 6022 case INTEL_FAM6_BROADWELL_G: 6023 case INTEL_FAM6_BROADWELL_X: 6024 x86_add_quirk(intel_pebs_isolation_quirk); 6025 x86_pmu.late_ack = true; 6026 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6027 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6028 6029 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ 6030 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | 6031 BDW_L3_MISS|HSW_SNOOP_DRAM; 6032 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| 6033 HSW_SNOOP_DRAM; 6034 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| 6035 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6036 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| 6037 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6038 6039 intel_pmu_lbr_init_hsw(); 6040 6041 x86_pmu.event_constraints = intel_bdw_event_constraints; 6042 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints; 6043 x86_pmu.extra_regs = intel_snbep_extra_regs; 6044 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6045 x86_pmu.pebs_prec_dist = true; 6046 /* all extra regs are per-cpu when HT is on */ 6047 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6048 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6049 6050 x86_pmu.hw_config = hsw_hw_config; 6051 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6052 x86_pmu.limit_period = bdw_limit_period; 6053 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6054 hsw_format_attr : nhm_format_attr; 6055 td_attr = hsw_events_attrs; 6056 mem_attr = hsw_mem_events_attrs; 6057 tsx_attr = hsw_tsx_events_attrs; 6058 pr_cont("Broadwell events, "); 6059 name = "broadwell"; 6060 break; 6061 6062 case INTEL_FAM6_XEON_PHI_KNL: 6063 case INTEL_FAM6_XEON_PHI_KNM: 6064 memcpy(hw_cache_event_ids, 6065 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6066 memcpy(hw_cache_extra_regs, 6067 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6068 intel_pmu_lbr_init_knl(); 6069 6070 x86_pmu.event_constraints = intel_slm_event_constraints; 6071 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 6072 x86_pmu.extra_regs = intel_knl_extra_regs; 6073 6074 /* all extra regs are per-cpu when HT is on */ 6075 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6076 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6077 extra_attr = slm_format_attr; 6078 pr_cont("Knights Landing/Mill events, "); 6079 name = "knights-landing"; 6080 break; 6081 6082 case INTEL_FAM6_SKYLAKE_X: 6083 pmem = true; 6084 fallthrough; 6085 case INTEL_FAM6_SKYLAKE_L: 6086 case INTEL_FAM6_SKYLAKE: 6087 case INTEL_FAM6_KABYLAKE_L: 6088 case INTEL_FAM6_KABYLAKE: 6089 case INTEL_FAM6_COMETLAKE_L: 6090 case INTEL_FAM6_COMETLAKE: 6091 x86_add_quirk(intel_pebs_isolation_quirk); 6092 x86_pmu.late_ack = true; 6093 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6094 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6095 intel_pmu_lbr_init_skl(); 6096 6097 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */ 6098 event_attr_td_recovery_bubbles.event_str_noht = 6099 "event=0xd,umask=0x1,cmask=1"; 6100 event_attr_td_recovery_bubbles.event_str_ht = 6101 "event=0xd,umask=0x1,cmask=1,any=1"; 6102 6103 x86_pmu.event_constraints = intel_skl_event_constraints; 6104 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; 6105 x86_pmu.extra_regs = intel_skl_extra_regs; 6106 x86_pmu.pebs_aliases = intel_pebs_aliases_skl; 6107 x86_pmu.pebs_prec_dist = true; 6108 /* all extra regs are per-cpu when HT is on */ 6109 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6110 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6111 6112 x86_pmu.hw_config = hsw_hw_config; 6113 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6114 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6115 hsw_format_attr : nhm_format_attr; 6116 extra_skl_attr = skl_format_attr; 6117 td_attr = hsw_events_attrs; 6118 mem_attr = hsw_mem_events_attrs; 6119 tsx_attr = hsw_tsx_events_attrs; 6120 intel_pmu_pebs_data_source_skl(pmem); 6121 6122 /* 6123 * Processors with CPUID.RTM_ALWAYS_ABORT have TSX deprecated by default. 6124 * TSX force abort hooks are not required on these systems. Only deploy 6125 * workaround when microcode has not enabled X86_FEATURE_RTM_ALWAYS_ABORT. 6126 */ 6127 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT) && 6128 !boot_cpu_has(X86_FEATURE_RTM_ALWAYS_ABORT)) { 6129 x86_pmu.flags |= PMU_FL_TFA; 6130 x86_pmu.get_event_constraints = tfa_get_event_constraints; 6131 x86_pmu.enable_all = intel_tfa_pmu_enable_all; 6132 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; 6133 } 6134 6135 pr_cont("Skylake events, "); 6136 name = "skylake"; 6137 break; 6138 6139 case INTEL_FAM6_ICELAKE_X: 6140 case INTEL_FAM6_ICELAKE_D: 6141 pmem = true; 6142 fallthrough; 6143 case INTEL_FAM6_ICELAKE_L: 6144 case INTEL_FAM6_ICELAKE: 6145 case INTEL_FAM6_TIGERLAKE_L: 6146 case INTEL_FAM6_TIGERLAKE: 6147 case INTEL_FAM6_ROCKETLAKE: 6148 x86_pmu.late_ack = true; 6149 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6150 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6151 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6152 intel_pmu_lbr_init_skl(); 6153 6154 x86_pmu.event_constraints = intel_icl_event_constraints; 6155 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints; 6156 x86_pmu.extra_regs = intel_icl_extra_regs; 6157 x86_pmu.pebs_aliases = NULL; 6158 x86_pmu.pebs_prec_dist = true; 6159 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6160 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6161 6162 x86_pmu.hw_config = hsw_hw_config; 6163 x86_pmu.get_event_constraints = icl_get_event_constraints; 6164 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6165 hsw_format_attr : nhm_format_attr; 6166 extra_skl_attr = skl_format_attr; 6167 mem_attr = icl_events_attrs; 6168 td_attr = icl_td_events_attrs; 6169 tsx_attr = icl_tsx_events_attrs; 6170 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6171 x86_pmu.lbr_pt_coexist = true; 6172 intel_pmu_pebs_data_source_skl(pmem); 6173 x86_pmu.num_topdown_events = 4; 6174 x86_pmu.update_topdown_event = icl_update_topdown_event; 6175 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; 6176 pr_cont("Icelake events, "); 6177 name = "icelake"; 6178 break; 6179 6180 case INTEL_FAM6_SAPPHIRERAPIDS_X: 6181 pmem = true; 6182 x86_pmu.late_ack = true; 6183 memcpy(hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6184 memcpy(hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6185 6186 x86_pmu.event_constraints = intel_spr_event_constraints; 6187 x86_pmu.pebs_constraints = intel_spr_pebs_event_constraints; 6188 x86_pmu.extra_regs = intel_spr_extra_regs; 6189 x86_pmu.limit_period = spr_limit_period; 6190 x86_pmu.pebs_aliases = NULL; 6191 x86_pmu.pebs_prec_dist = true; 6192 x86_pmu.pebs_block = true; 6193 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6194 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6195 x86_pmu.flags |= PMU_FL_PEBS_ALL; 6196 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6197 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6198 6199 x86_pmu.hw_config = hsw_hw_config; 6200 x86_pmu.get_event_constraints = spr_get_event_constraints; 6201 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6202 hsw_format_attr : nhm_format_attr; 6203 extra_skl_attr = skl_format_attr; 6204 mem_attr = spr_events_attrs; 6205 td_attr = spr_td_events_attrs; 6206 tsx_attr = spr_tsx_events_attrs; 6207 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6208 x86_pmu.lbr_pt_coexist = true; 6209 intel_pmu_pebs_data_source_skl(pmem); 6210 x86_pmu.num_topdown_events = 8; 6211 x86_pmu.update_topdown_event = icl_update_topdown_event; 6212 x86_pmu.set_topdown_event_period = icl_set_topdown_event_period; 6213 pr_cont("Sapphire Rapids events, "); 6214 name = "sapphire_rapids"; 6215 break; 6216 6217 case INTEL_FAM6_ALDERLAKE: 6218 case INTEL_FAM6_ALDERLAKE_L: 6219 case INTEL_FAM6_RAPTORLAKE: 6220 /* 6221 * Alder Lake has 2 types of CPU, core and atom. 6222 * 6223 * Initialize the common PerfMon capabilities here. 6224 */ 6225 x86_pmu.hybrid_pmu = kcalloc(X86_HYBRID_NUM_PMUS, 6226 sizeof(struct x86_hybrid_pmu), 6227 GFP_KERNEL); 6228 if (!x86_pmu.hybrid_pmu) 6229 return -ENOMEM; 6230 static_branch_enable(&perf_is_hybrid); 6231 x86_pmu.num_hybrid_pmus = X86_HYBRID_NUM_PMUS; 6232 6233 x86_pmu.pebs_aliases = NULL; 6234 x86_pmu.pebs_prec_dist = true; 6235 x86_pmu.pebs_block = true; 6236 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6237 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6238 x86_pmu.flags |= PMU_FL_PEBS_ALL; 6239 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6240 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6241 x86_pmu.lbr_pt_coexist = true; 6242 intel_pmu_pebs_data_source_skl(false); 6243 x86_pmu.num_topdown_events = 8; 6244 x86_pmu.update_topdown_event = adl_update_topdown_event; 6245 x86_pmu.set_topdown_event_period = adl_set_topdown_event_period; 6246 6247 x86_pmu.filter_match = intel_pmu_filter_match; 6248 x86_pmu.get_event_constraints = adl_get_event_constraints; 6249 x86_pmu.hw_config = adl_hw_config; 6250 x86_pmu.limit_period = spr_limit_period; 6251 x86_pmu.get_hybrid_cpu_type = adl_get_hybrid_cpu_type; 6252 /* 6253 * The rtm_abort_event is used to check whether to enable GPRs 6254 * for the RTM abort event. Atom doesn't have the RTM abort 6255 * event. There is no harmful to set it in the common 6256 * x86_pmu.rtm_abort_event. 6257 */ 6258 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6259 6260 td_attr = adl_hybrid_events_attrs; 6261 mem_attr = adl_hybrid_mem_attrs; 6262 tsx_attr = adl_hybrid_tsx_attrs; 6263 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6264 adl_hybrid_extra_attr_rtm : adl_hybrid_extra_attr; 6265 6266 /* Initialize big core specific PerfMon capabilities.*/ 6267 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX]; 6268 pmu->name = "cpu_core"; 6269 pmu->cpu_type = hybrid_big; 6270 pmu->late_ack = true; 6271 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) { 6272 pmu->num_counters = x86_pmu.num_counters + 2; 6273 pmu->num_counters_fixed = x86_pmu.num_counters_fixed + 1; 6274 } else { 6275 pmu->num_counters = x86_pmu.num_counters; 6276 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6277 } 6278 6279 /* 6280 * Quirk: For some Alder Lake machine, when all E-cores are disabled in 6281 * a BIOS, the leaf 0xA will enumerate all counters of P-cores. However, 6282 * the X86_FEATURE_HYBRID_CPU is still set. The above codes will 6283 * mistakenly add extra counters for P-cores. Correct the number of 6284 * counters here. 6285 */ 6286 if ((pmu->num_counters > 8) || (pmu->num_counters_fixed > 4)) { 6287 pmu->num_counters = x86_pmu.num_counters; 6288 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6289 } 6290 6291 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters); 6292 pmu->unconstrained = (struct event_constraint) 6293 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6294 0, pmu->num_counters, 0, 0); 6295 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; 6296 pmu->intel_cap.perf_metrics = 1; 6297 pmu->intel_cap.pebs_output_pt_available = 0; 6298 6299 memcpy(pmu->hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids)); 6300 memcpy(pmu->hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs)); 6301 pmu->event_constraints = intel_spr_event_constraints; 6302 pmu->pebs_constraints = intel_spr_pebs_event_constraints; 6303 pmu->extra_regs = intel_spr_extra_regs; 6304 6305 /* Initialize Atom core specific PerfMon capabilities.*/ 6306 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX]; 6307 pmu->name = "cpu_atom"; 6308 pmu->cpu_type = hybrid_small; 6309 pmu->mid_ack = true; 6310 pmu->num_counters = x86_pmu.num_counters; 6311 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6312 pmu->max_pebs_events = x86_pmu.max_pebs_events; 6313 pmu->unconstrained = (struct event_constraint) 6314 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6315 0, pmu->num_counters, 0, 0); 6316 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; 6317 pmu->intel_cap.perf_metrics = 0; 6318 pmu->intel_cap.pebs_output_pt_available = 1; 6319 6320 memcpy(pmu->hw_cache_event_ids, glp_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids)); 6321 memcpy(pmu->hw_cache_extra_regs, tnt_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs)); 6322 pmu->hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6323 pmu->event_constraints = intel_slm_event_constraints; 6324 pmu->pebs_constraints = intel_grt_pebs_event_constraints; 6325 pmu->extra_regs = intel_grt_extra_regs; 6326 pr_cont("Alderlake Hybrid events, "); 6327 name = "alderlake_hybrid"; 6328 break; 6329 6330 default: 6331 switch (x86_pmu.version) { 6332 case 1: 6333 x86_pmu.event_constraints = intel_v1_event_constraints; 6334 pr_cont("generic architected perfmon v1, "); 6335 name = "generic_arch_v1"; 6336 break; 6337 case 2: 6338 case 3: 6339 case 4: 6340 /* 6341 * default constraints for v2 and up 6342 */ 6343 x86_pmu.event_constraints = intel_gen_event_constraints; 6344 pr_cont("generic architected perfmon, "); 6345 name = "generic_arch_v2+"; 6346 break; 6347 default: 6348 /* 6349 * The default constraints for v5 and up can support up to 6350 * 16 fixed counters. For the fixed counters 4 and later, 6351 * the pseudo-encoding is applied. 6352 * The constraints may be cut according to the CPUID enumeration 6353 * by inserting the EVENT_CONSTRAINT_END. 6354 */ 6355 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) 6356 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; 6357 intel_v5_gen_event_constraints[x86_pmu.num_counters_fixed].weight = -1; 6358 x86_pmu.event_constraints = intel_v5_gen_event_constraints; 6359 pr_cont("generic architected perfmon, "); 6360 name = "generic_arch_v5+"; 6361 break; 6362 } 6363 } 6364 6365 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name); 6366 6367 if (!is_hybrid()) { 6368 group_events_td.attrs = td_attr; 6369 group_events_mem.attrs = mem_attr; 6370 group_events_tsx.attrs = tsx_attr; 6371 group_format_extra.attrs = extra_attr; 6372 group_format_extra_skl.attrs = extra_skl_attr; 6373 6374 x86_pmu.attr_update = attr_update; 6375 } else { 6376 hybrid_group_events_td.attrs = td_attr; 6377 hybrid_group_events_mem.attrs = mem_attr; 6378 hybrid_group_events_tsx.attrs = tsx_attr; 6379 hybrid_group_format_extra.attrs = extra_attr; 6380 6381 x86_pmu.attr_update = hybrid_attr_update; 6382 } 6383 6384 intel_pmu_check_num_counters(&x86_pmu.num_counters, 6385 &x86_pmu.num_counters_fixed, 6386 &x86_pmu.intel_ctrl, 6387 (u64)fixed_mask); 6388 6389 /* AnyThread may be deprecated on arch perfmon v5 or later */ 6390 if (x86_pmu.intel_cap.anythread_deprecated) 6391 x86_pmu.format_attrs = intel_arch_formats_attr; 6392 6393 intel_pmu_check_event_constraints(x86_pmu.event_constraints, 6394 x86_pmu.num_counters, 6395 x86_pmu.num_counters_fixed, 6396 x86_pmu.intel_ctrl); 6397 /* 6398 * Access LBR MSR may cause #GP under certain circumstances. 6399 * E.g. KVM doesn't support LBR MSR 6400 * Check all LBT MSR here. 6401 * Disable LBR access if any LBR MSRs can not be accessed. 6402 */ 6403 if (x86_pmu.lbr_tos && !check_msr(x86_pmu.lbr_tos, 0x3UL)) 6404 x86_pmu.lbr_nr = 0; 6405 for (i = 0; i < x86_pmu.lbr_nr; i++) { 6406 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && 6407 check_msr(x86_pmu.lbr_to + i, 0xffffUL))) 6408 x86_pmu.lbr_nr = 0; 6409 } 6410 6411 if (x86_pmu.lbr_nr) { 6412 intel_pmu_lbr_init(); 6413 6414 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr); 6415 6416 /* only support branch_stack snapshot for perfmon >= v2 */ 6417 if (x86_pmu.disable_all == intel_pmu_disable_all) { 6418 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) { 6419 static_call_update(perf_snapshot_branch_stack, 6420 intel_pmu_snapshot_arch_branch_stack); 6421 } else { 6422 static_call_update(perf_snapshot_branch_stack, 6423 intel_pmu_snapshot_branch_stack); 6424 } 6425 } 6426 } 6427 6428 intel_pmu_check_extra_regs(x86_pmu.extra_regs); 6429 6430 /* Support full width counters using alternative MSR range */ 6431 if (x86_pmu.intel_cap.full_width_write) { 6432 x86_pmu.max_period = x86_pmu.cntval_mask >> 1; 6433 x86_pmu.perfctr = MSR_IA32_PMC0; 6434 pr_cont("full-width counters, "); 6435 } 6436 6437 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) 6438 x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 6439 6440 if (is_hybrid()) 6441 intel_pmu_check_hybrid_pmus((u64)fixed_mask); 6442 6443 intel_aux_output_init(); 6444 6445 return 0; 6446 } 6447 6448 /* 6449 * HT bug: phase 2 init 6450 * Called once we have valid topology information to check 6451 * whether or not HT is enabled 6452 * If HT is off, then we disable the workaround 6453 */ 6454 static __init int fixup_ht_bug(void) 6455 { 6456 int c; 6457 /* 6458 * problem not present on this CPU model, nothing to do 6459 */ 6460 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) 6461 return 0; 6462 6463 if (topology_max_smt_threads() > 1) { 6464 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); 6465 return 0; 6466 } 6467 6468 cpus_read_lock(); 6469 6470 hardlockup_detector_perf_stop(); 6471 6472 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); 6473 6474 x86_pmu.start_scheduling = NULL; 6475 x86_pmu.commit_scheduling = NULL; 6476 x86_pmu.stop_scheduling = NULL; 6477 6478 hardlockup_detector_perf_restart(); 6479 6480 for_each_online_cpu(c) 6481 free_excl_cntrs(&per_cpu(cpu_hw_events, c)); 6482 6483 cpus_read_unlock(); 6484 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); 6485 return 0; 6486 } 6487 subsys_initcall(fixup_ht_bug) 6488