1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Per core/cpu state 4 * 5 * Used to coordinate shared registers between HT threads or 6 * among events on a single PMU. 7 */ 8 9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 10 11 #include <linux/stddef.h> 12 #include <linux/types.h> 13 #include <linux/init.h> 14 #include <linux/slab.h> 15 #include <linux/export.h> 16 #include <linux/nmi.h> 17 #include <linux/kvm_host.h> 18 19 #include <asm/cpufeature.h> 20 #include <asm/hardirq.h> 21 #include <asm/intel-family.h> 22 #include <asm/intel_pt.h> 23 #include <asm/apic.h> 24 #include <asm/cpu_device_id.h> 25 26 #include "../perf_event.h" 27 28 /* 29 * Intel PerfMon, used on Core and later. 30 */ 31 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = 32 { 33 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, 34 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, 35 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, 36 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, 37 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, 38 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, 39 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, 40 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ 41 }; 42 43 static struct event_constraint intel_core_event_constraints[] __read_mostly = 44 { 45 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 46 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 47 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 48 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 49 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 50 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ 51 EVENT_CONSTRAINT_END 52 }; 53 54 static struct event_constraint intel_core2_event_constraints[] __read_mostly = 55 { 56 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 57 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 58 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 59 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ 60 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 61 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 62 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 63 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 64 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ 65 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 66 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ 67 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ 68 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ 69 EVENT_CONSTRAINT_END 70 }; 71 72 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = 73 { 74 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 75 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 76 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 77 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ 78 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ 79 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ 80 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ 81 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ 82 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ 83 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 84 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 85 EVENT_CONSTRAINT_END 86 }; 87 88 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = 89 { 90 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 91 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 92 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 93 EVENT_EXTRA_END 94 }; 95 96 static struct event_constraint intel_westmere_event_constraints[] __read_mostly = 97 { 98 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 99 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 100 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 101 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 102 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ 103 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 104 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ 105 EVENT_CONSTRAINT_END 106 }; 107 108 static struct event_constraint intel_snb_event_constraints[] __read_mostly = 109 { 110 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 111 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 112 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 113 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 114 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 115 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 116 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 117 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ 118 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 119 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 120 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 121 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 122 123 /* 124 * When HT is off these events can only run on the bottom 4 counters 125 * When HT is on, they are impacted by the HT bug and require EXCL access 126 */ 127 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 128 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 129 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 130 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 131 132 EVENT_CONSTRAINT_END 133 }; 134 135 static struct event_constraint intel_ivb_event_constraints[] __read_mostly = 136 { 137 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 138 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 139 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 140 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ 141 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMPTY */ 142 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ 143 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ 144 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 145 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 146 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ 147 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 148 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 149 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 150 151 /* 152 * When HT is off these events can only run on the bottom 4 counters 153 * When HT is on, they are impacted by the HT bug and require EXCL access 154 */ 155 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 156 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 157 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 158 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 159 160 EVENT_CONSTRAINT_END 161 }; 162 163 static struct extra_reg intel_westmere_extra_regs[] __read_mostly = 164 { 165 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 166 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 167 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), 168 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 169 EVENT_EXTRA_END 170 }; 171 172 static struct event_constraint intel_v1_event_constraints[] __read_mostly = 173 { 174 EVENT_CONSTRAINT_END 175 }; 176 177 static struct event_constraint intel_gen_event_constraints[] __read_mostly = 178 { 179 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 180 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 181 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 182 EVENT_CONSTRAINT_END 183 }; 184 185 static struct event_constraint intel_v5_gen_event_constraints[] __read_mostly = 186 { 187 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 188 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 189 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 190 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 191 FIXED_EVENT_CONSTRAINT(0x0500, 4), 192 FIXED_EVENT_CONSTRAINT(0x0600, 5), 193 FIXED_EVENT_CONSTRAINT(0x0700, 6), 194 FIXED_EVENT_CONSTRAINT(0x0800, 7), 195 FIXED_EVENT_CONSTRAINT(0x0900, 8), 196 FIXED_EVENT_CONSTRAINT(0x0a00, 9), 197 FIXED_EVENT_CONSTRAINT(0x0b00, 10), 198 FIXED_EVENT_CONSTRAINT(0x0c00, 11), 199 FIXED_EVENT_CONSTRAINT(0x0d00, 12), 200 FIXED_EVENT_CONSTRAINT(0x0e00, 13), 201 FIXED_EVENT_CONSTRAINT(0x0f00, 14), 202 FIXED_EVENT_CONSTRAINT(0x1000, 15), 203 EVENT_CONSTRAINT_END 204 }; 205 206 static struct event_constraint intel_slm_event_constraints[] __read_mostly = 207 { 208 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 209 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 210 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ 211 EVENT_CONSTRAINT_END 212 }; 213 214 static struct event_constraint intel_skl_event_constraints[] = { 215 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 216 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 217 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 218 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ 219 220 /* 221 * when HT is off, these can only run on the bottom 4 counters 222 */ 223 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 224 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 225 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 226 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 227 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */ 228 229 EVENT_CONSTRAINT_END 230 }; 231 232 static struct extra_reg intel_knl_extra_regs[] __read_mostly = { 233 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0), 234 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1), 235 EVENT_EXTRA_END 236 }; 237 238 static struct extra_reg intel_snb_extra_regs[] __read_mostly = { 239 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 240 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), 241 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), 242 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 243 EVENT_EXTRA_END 244 }; 245 246 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { 247 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 248 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 249 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 250 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 251 EVENT_EXTRA_END 252 }; 253 254 static struct extra_reg intel_skl_extra_regs[] __read_mostly = { 255 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 256 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 257 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 258 /* 259 * Note the low 8 bits eventsel code is not a continuous field, containing 260 * some #GPing bits. These are masked out. 261 */ 262 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 263 EVENT_EXTRA_END 264 }; 265 266 static struct event_constraint intel_icl_event_constraints[] = { 267 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 268 FIXED_EVENT_CONSTRAINT(0x01c0, 0), /* old INST_RETIRED.PREC_DIST */ 269 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 270 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 271 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 272 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 273 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 274 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 275 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 276 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 277 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf), 278 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf), 279 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */ 280 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x56, 0xf), 281 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf), 282 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */ 283 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.CYCLES_MEM_ANY */ 284 INTEL_UEVENT_CONSTRAINT(0x14a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */ 285 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */ 286 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf), 287 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf), 288 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf), 289 INTEL_EVENT_CONSTRAINT(0xef, 0xf), 290 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf), 291 EVENT_CONSTRAINT_END 292 }; 293 294 static struct extra_reg intel_icl_extra_regs[] __read_mostly = { 295 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0), 296 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1), 297 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 298 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 299 EVENT_EXTRA_END 300 }; 301 302 static struct extra_reg intel_spr_extra_regs[] __read_mostly = { 303 INTEL_UEVENT_EXTRA_REG(0x012a, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 304 INTEL_UEVENT_EXTRA_REG(0x012b, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 305 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 306 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff1f, FE), 307 INTEL_UEVENT_EXTRA_REG(0x40ad, MSR_PEBS_FRONTEND, 0x7, FE), 308 INTEL_UEVENT_EXTRA_REG(0x04c2, MSR_PEBS_FRONTEND, 0x8, FE), 309 EVENT_EXTRA_END 310 }; 311 312 static struct event_constraint intel_spr_event_constraints[] = { 313 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 314 FIXED_EVENT_CONSTRAINT(0x0100, 0), /* INST_RETIRED.PREC_DIST */ 315 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 316 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 317 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 318 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_RETIRING, 0), 319 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BAD_SPEC, 1), 320 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FE_BOUND, 2), 321 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BE_BOUND, 3), 322 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_HEAVY_OPS, 4), 323 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_BR_MISPREDICT, 5), 324 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_FETCH_LAT, 6), 325 METRIC_EVENT_CONSTRAINT(INTEL_TD_METRIC_MEM_BOUND, 7), 326 327 INTEL_EVENT_CONSTRAINT(0x2e, 0xff), 328 INTEL_EVENT_CONSTRAINT(0x3c, 0xff), 329 /* 330 * Generally event codes < 0x90 are restricted to counters 0-3. 331 * The 0x2E and 0x3C are exception, which has no restriction. 332 */ 333 INTEL_EVENT_CONSTRAINT_RANGE(0x01, 0x8f, 0xf), 334 335 INTEL_UEVENT_CONSTRAINT(0x01a3, 0xf), 336 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), 337 INTEL_UEVENT_CONSTRAINT(0x08a3, 0xf), 338 INTEL_UEVENT_CONSTRAINT(0x04a4, 0x1), 339 INTEL_UEVENT_CONSTRAINT(0x08a4, 0x1), 340 INTEL_UEVENT_CONSTRAINT(0x02cd, 0x1), 341 INTEL_EVENT_CONSTRAINT(0xce, 0x1), 342 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xdf, 0xf), 343 /* 344 * Generally event codes >= 0x90 are likely to have no restrictions. 345 * The exception are defined as above. 346 */ 347 INTEL_EVENT_CONSTRAINT_RANGE(0x90, 0xfe, 0xff), 348 349 EVENT_CONSTRAINT_END 350 }; 351 352 353 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); 354 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); 355 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); 356 357 static struct attribute *nhm_mem_events_attrs[] = { 358 EVENT_PTR(mem_ld_nhm), 359 NULL, 360 }; 361 362 /* 363 * topdown events for Intel Core CPUs. 364 * 365 * The events are all in slots, which is a free slot in a 4 wide 366 * pipeline. Some events are already reported in slots, for cycle 367 * events we multiply by the pipeline width (4). 368 * 369 * With Hyper Threading on, topdown metrics are either summed or averaged 370 * between the threads of a core: (count_t0 + count_t1). 371 * 372 * For the average case the metric is always scaled to pipeline width, 373 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4) 374 */ 375 376 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots, 377 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */ 378 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */ 379 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2"); 380 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued, 381 "event=0xe,umask=0x1"); /* uops_issued.any */ 382 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired, 383 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */ 384 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles, 385 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */ 386 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles, 387 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */ 388 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */ 389 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale, 390 "4", "2"); 391 392 EVENT_ATTR_STR(slots, slots, "event=0x00,umask=0x4"); 393 EVENT_ATTR_STR(topdown-retiring, td_retiring, "event=0x00,umask=0x80"); 394 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec, "event=0x00,umask=0x81"); 395 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound, "event=0x00,umask=0x82"); 396 EVENT_ATTR_STR(topdown-be-bound, td_be_bound, "event=0x00,umask=0x83"); 397 EVENT_ATTR_STR(topdown-heavy-ops, td_heavy_ops, "event=0x00,umask=0x84"); 398 EVENT_ATTR_STR(topdown-br-mispredict, td_br_mispredict, "event=0x00,umask=0x85"); 399 EVENT_ATTR_STR(topdown-fetch-lat, td_fetch_lat, "event=0x00,umask=0x86"); 400 EVENT_ATTR_STR(topdown-mem-bound, td_mem_bound, "event=0x00,umask=0x87"); 401 402 static struct attribute *snb_events_attrs[] = { 403 EVENT_PTR(td_slots_issued), 404 EVENT_PTR(td_slots_retired), 405 EVENT_PTR(td_fetch_bubbles), 406 EVENT_PTR(td_total_slots), 407 EVENT_PTR(td_total_slots_scale), 408 EVENT_PTR(td_recovery_bubbles), 409 EVENT_PTR(td_recovery_bubbles_scale), 410 NULL, 411 }; 412 413 static struct attribute *snb_mem_events_attrs[] = { 414 EVENT_PTR(mem_ld_snb), 415 EVENT_PTR(mem_st_snb), 416 NULL, 417 }; 418 419 static struct event_constraint intel_hsw_event_constraints[] = { 420 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 421 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 422 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 423 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 424 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 425 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 426 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 427 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), 428 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 429 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), 430 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 431 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), 432 433 /* 434 * When HT is off these events can only run on the bottom 4 counters 435 * When HT is on, they are impacted by the HT bug and require EXCL access 436 */ 437 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 438 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 439 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 440 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 441 442 EVENT_CONSTRAINT_END 443 }; 444 445 static struct event_constraint intel_bdw_event_constraints[] = { 446 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 447 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 448 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 449 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 450 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */ 451 /* 452 * when HT is off, these can only run on the bottom 4 counters 453 */ 454 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 455 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 456 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 457 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 458 EVENT_CONSTRAINT_END 459 }; 460 461 static u64 intel_pmu_event_map(int hw_event) 462 { 463 return intel_perfmon_event_map[hw_event]; 464 } 465 466 static __initconst const u64 spr_hw_cache_event_ids 467 [PERF_COUNT_HW_CACHE_MAX] 468 [PERF_COUNT_HW_CACHE_OP_MAX] 469 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 470 { 471 [ C(L1D ) ] = { 472 [ C(OP_READ) ] = { 473 [ C(RESULT_ACCESS) ] = 0x81d0, 474 [ C(RESULT_MISS) ] = 0xe124, 475 }, 476 [ C(OP_WRITE) ] = { 477 [ C(RESULT_ACCESS) ] = 0x82d0, 478 }, 479 }, 480 [ C(L1I ) ] = { 481 [ C(OP_READ) ] = { 482 [ C(RESULT_MISS) ] = 0xe424, 483 }, 484 [ C(OP_WRITE) ] = { 485 [ C(RESULT_ACCESS) ] = -1, 486 [ C(RESULT_MISS) ] = -1, 487 }, 488 }, 489 [ C(LL ) ] = { 490 [ C(OP_READ) ] = { 491 [ C(RESULT_ACCESS) ] = 0x12a, 492 [ C(RESULT_MISS) ] = 0x12a, 493 }, 494 [ C(OP_WRITE) ] = { 495 [ C(RESULT_ACCESS) ] = 0x12a, 496 [ C(RESULT_MISS) ] = 0x12a, 497 }, 498 }, 499 [ C(DTLB) ] = { 500 [ C(OP_READ) ] = { 501 [ C(RESULT_ACCESS) ] = 0x81d0, 502 [ C(RESULT_MISS) ] = 0xe12, 503 }, 504 [ C(OP_WRITE) ] = { 505 [ C(RESULT_ACCESS) ] = 0x82d0, 506 [ C(RESULT_MISS) ] = 0xe13, 507 }, 508 }, 509 [ C(ITLB) ] = { 510 [ C(OP_READ) ] = { 511 [ C(RESULT_ACCESS) ] = -1, 512 [ C(RESULT_MISS) ] = 0xe11, 513 }, 514 [ C(OP_WRITE) ] = { 515 [ C(RESULT_ACCESS) ] = -1, 516 [ C(RESULT_MISS) ] = -1, 517 }, 518 [ C(OP_PREFETCH) ] = { 519 [ C(RESULT_ACCESS) ] = -1, 520 [ C(RESULT_MISS) ] = -1, 521 }, 522 }, 523 [ C(BPU ) ] = { 524 [ C(OP_READ) ] = { 525 [ C(RESULT_ACCESS) ] = 0x4c4, 526 [ C(RESULT_MISS) ] = 0x4c5, 527 }, 528 [ C(OP_WRITE) ] = { 529 [ C(RESULT_ACCESS) ] = -1, 530 [ C(RESULT_MISS) ] = -1, 531 }, 532 [ C(OP_PREFETCH) ] = { 533 [ C(RESULT_ACCESS) ] = -1, 534 [ C(RESULT_MISS) ] = -1, 535 }, 536 }, 537 [ C(NODE) ] = { 538 [ C(OP_READ) ] = { 539 [ C(RESULT_ACCESS) ] = 0x12a, 540 [ C(RESULT_MISS) ] = 0x12a, 541 }, 542 }, 543 }; 544 545 static __initconst const u64 spr_hw_cache_extra_regs 546 [PERF_COUNT_HW_CACHE_MAX] 547 [PERF_COUNT_HW_CACHE_OP_MAX] 548 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 549 { 550 [ C(LL ) ] = { 551 [ C(OP_READ) ] = { 552 [ C(RESULT_ACCESS) ] = 0x10001, 553 [ C(RESULT_MISS) ] = 0x3fbfc00001, 554 }, 555 [ C(OP_WRITE) ] = { 556 [ C(RESULT_ACCESS) ] = 0x3f3ffc0002, 557 [ C(RESULT_MISS) ] = 0x3f3fc00002, 558 }, 559 }, 560 [ C(NODE) ] = { 561 [ C(OP_READ) ] = { 562 [ C(RESULT_ACCESS) ] = 0x10c000001, 563 [ C(RESULT_MISS) ] = 0x3fb3000001, 564 }, 565 }, 566 }; 567 568 /* 569 * Notes on the events: 570 * - data reads do not include code reads (comparable to earlier tables) 571 * - data counts include speculative execution (except L1 write, dtlb, bpu) 572 * - remote node access includes remote memory, remote cache, remote mmio. 573 * - prefetches are not included in the counts. 574 * - icache miss does not include decoded icache 575 */ 576 577 #define SKL_DEMAND_DATA_RD BIT_ULL(0) 578 #define SKL_DEMAND_RFO BIT_ULL(1) 579 #define SKL_ANY_RESPONSE BIT_ULL(16) 580 #define SKL_SUPPLIER_NONE BIT_ULL(17) 581 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26) 582 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27) 583 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28) 584 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29) 585 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \ 586 SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 587 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 588 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 589 #define SKL_SPL_HIT BIT_ULL(30) 590 #define SKL_SNOOP_NONE BIT_ULL(31) 591 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32) 592 #define SKL_SNOOP_MISS BIT_ULL(33) 593 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34) 594 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35) 595 #define SKL_SNOOP_HITM BIT_ULL(36) 596 #define SKL_SNOOP_NON_DRAM BIT_ULL(37) 597 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \ 598 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 599 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 600 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM) 601 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD 602 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \ 603 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 604 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 605 SKL_SNOOP_HITM|SKL_SPL_HIT) 606 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO 607 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE 608 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 609 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 610 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 611 612 static __initconst const u64 skl_hw_cache_event_ids 613 [PERF_COUNT_HW_CACHE_MAX] 614 [PERF_COUNT_HW_CACHE_OP_MAX] 615 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 616 { 617 [ C(L1D ) ] = { 618 [ C(OP_READ) ] = { 619 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 620 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 621 }, 622 [ C(OP_WRITE) ] = { 623 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 624 [ C(RESULT_MISS) ] = 0x0, 625 }, 626 [ C(OP_PREFETCH) ] = { 627 [ C(RESULT_ACCESS) ] = 0x0, 628 [ C(RESULT_MISS) ] = 0x0, 629 }, 630 }, 631 [ C(L1I ) ] = { 632 [ C(OP_READ) ] = { 633 [ C(RESULT_ACCESS) ] = 0x0, 634 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */ 635 }, 636 [ C(OP_WRITE) ] = { 637 [ C(RESULT_ACCESS) ] = -1, 638 [ C(RESULT_MISS) ] = -1, 639 }, 640 [ C(OP_PREFETCH) ] = { 641 [ C(RESULT_ACCESS) ] = 0x0, 642 [ C(RESULT_MISS) ] = 0x0, 643 }, 644 }, 645 [ C(LL ) ] = { 646 [ C(OP_READ) ] = { 647 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 648 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 649 }, 650 [ C(OP_WRITE) ] = { 651 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 652 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 653 }, 654 [ C(OP_PREFETCH) ] = { 655 [ C(RESULT_ACCESS) ] = 0x0, 656 [ C(RESULT_MISS) ] = 0x0, 657 }, 658 }, 659 [ C(DTLB) ] = { 660 [ C(OP_READ) ] = { 661 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 662 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 663 }, 664 [ C(OP_WRITE) ] = { 665 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 666 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 667 }, 668 [ C(OP_PREFETCH) ] = { 669 [ C(RESULT_ACCESS) ] = 0x0, 670 [ C(RESULT_MISS) ] = 0x0, 671 }, 672 }, 673 [ C(ITLB) ] = { 674 [ C(OP_READ) ] = { 675 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */ 676 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */ 677 }, 678 [ C(OP_WRITE) ] = { 679 [ C(RESULT_ACCESS) ] = -1, 680 [ C(RESULT_MISS) ] = -1, 681 }, 682 [ C(OP_PREFETCH) ] = { 683 [ C(RESULT_ACCESS) ] = -1, 684 [ C(RESULT_MISS) ] = -1, 685 }, 686 }, 687 [ C(BPU ) ] = { 688 [ C(OP_READ) ] = { 689 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 690 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 691 }, 692 [ C(OP_WRITE) ] = { 693 [ C(RESULT_ACCESS) ] = -1, 694 [ C(RESULT_MISS) ] = -1, 695 }, 696 [ C(OP_PREFETCH) ] = { 697 [ C(RESULT_ACCESS) ] = -1, 698 [ C(RESULT_MISS) ] = -1, 699 }, 700 }, 701 [ C(NODE) ] = { 702 [ C(OP_READ) ] = { 703 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 704 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 705 }, 706 [ C(OP_WRITE) ] = { 707 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 708 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 709 }, 710 [ C(OP_PREFETCH) ] = { 711 [ C(RESULT_ACCESS) ] = 0x0, 712 [ C(RESULT_MISS) ] = 0x0, 713 }, 714 }, 715 }; 716 717 static __initconst const u64 skl_hw_cache_extra_regs 718 [PERF_COUNT_HW_CACHE_MAX] 719 [PERF_COUNT_HW_CACHE_OP_MAX] 720 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 721 { 722 [ C(LL ) ] = { 723 [ C(OP_READ) ] = { 724 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 725 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 726 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 727 SKL_L3_MISS|SKL_ANY_SNOOP| 728 SKL_SUPPLIER_NONE, 729 }, 730 [ C(OP_WRITE) ] = { 731 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 732 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 733 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 734 SKL_L3_MISS|SKL_ANY_SNOOP| 735 SKL_SUPPLIER_NONE, 736 }, 737 [ C(OP_PREFETCH) ] = { 738 [ C(RESULT_ACCESS) ] = 0x0, 739 [ C(RESULT_MISS) ] = 0x0, 740 }, 741 }, 742 [ C(NODE) ] = { 743 [ C(OP_READ) ] = { 744 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 745 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 746 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 747 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 748 }, 749 [ C(OP_WRITE) ] = { 750 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 751 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 752 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 753 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 754 }, 755 [ C(OP_PREFETCH) ] = { 756 [ C(RESULT_ACCESS) ] = 0x0, 757 [ C(RESULT_MISS) ] = 0x0, 758 }, 759 }, 760 }; 761 762 #define SNB_DMND_DATA_RD (1ULL << 0) 763 #define SNB_DMND_RFO (1ULL << 1) 764 #define SNB_DMND_IFETCH (1ULL << 2) 765 #define SNB_DMND_WB (1ULL << 3) 766 #define SNB_PF_DATA_RD (1ULL << 4) 767 #define SNB_PF_RFO (1ULL << 5) 768 #define SNB_PF_IFETCH (1ULL << 6) 769 #define SNB_LLC_DATA_RD (1ULL << 7) 770 #define SNB_LLC_RFO (1ULL << 8) 771 #define SNB_LLC_IFETCH (1ULL << 9) 772 #define SNB_BUS_LOCKS (1ULL << 10) 773 #define SNB_STRM_ST (1ULL << 11) 774 #define SNB_OTHER (1ULL << 15) 775 #define SNB_RESP_ANY (1ULL << 16) 776 #define SNB_NO_SUPP (1ULL << 17) 777 #define SNB_LLC_HITM (1ULL << 18) 778 #define SNB_LLC_HITE (1ULL << 19) 779 #define SNB_LLC_HITS (1ULL << 20) 780 #define SNB_LLC_HITF (1ULL << 21) 781 #define SNB_LOCAL (1ULL << 22) 782 #define SNB_REMOTE (0xffULL << 23) 783 #define SNB_SNP_NONE (1ULL << 31) 784 #define SNB_SNP_NOT_NEEDED (1ULL << 32) 785 #define SNB_SNP_MISS (1ULL << 33) 786 #define SNB_NO_FWD (1ULL << 34) 787 #define SNB_SNP_FWD (1ULL << 35) 788 #define SNB_HITM (1ULL << 36) 789 #define SNB_NON_DRAM (1ULL << 37) 790 791 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) 792 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) 793 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 794 795 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ 796 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ 797 SNB_HITM) 798 799 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) 800 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) 801 802 #define SNB_L3_ACCESS SNB_RESP_ANY 803 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) 804 805 static __initconst const u64 snb_hw_cache_extra_regs 806 [PERF_COUNT_HW_CACHE_MAX] 807 [PERF_COUNT_HW_CACHE_OP_MAX] 808 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 809 { 810 [ C(LL ) ] = { 811 [ C(OP_READ) ] = { 812 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, 813 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, 814 }, 815 [ C(OP_WRITE) ] = { 816 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, 817 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, 818 }, 819 [ C(OP_PREFETCH) ] = { 820 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, 821 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, 822 }, 823 }, 824 [ C(NODE) ] = { 825 [ C(OP_READ) ] = { 826 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, 827 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, 828 }, 829 [ C(OP_WRITE) ] = { 830 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, 831 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, 832 }, 833 [ C(OP_PREFETCH) ] = { 834 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, 835 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, 836 }, 837 }, 838 }; 839 840 static __initconst const u64 snb_hw_cache_event_ids 841 [PERF_COUNT_HW_CACHE_MAX] 842 [PERF_COUNT_HW_CACHE_OP_MAX] 843 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 844 { 845 [ C(L1D) ] = { 846 [ C(OP_READ) ] = { 847 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ 848 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ 849 }, 850 [ C(OP_WRITE) ] = { 851 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ 852 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ 853 }, 854 [ C(OP_PREFETCH) ] = { 855 [ C(RESULT_ACCESS) ] = 0x0, 856 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ 857 }, 858 }, 859 [ C(L1I ) ] = { 860 [ C(OP_READ) ] = { 861 [ C(RESULT_ACCESS) ] = 0x0, 862 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ 863 }, 864 [ C(OP_WRITE) ] = { 865 [ C(RESULT_ACCESS) ] = -1, 866 [ C(RESULT_MISS) ] = -1, 867 }, 868 [ C(OP_PREFETCH) ] = { 869 [ C(RESULT_ACCESS) ] = 0x0, 870 [ C(RESULT_MISS) ] = 0x0, 871 }, 872 }, 873 [ C(LL ) ] = { 874 [ C(OP_READ) ] = { 875 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 876 [ C(RESULT_ACCESS) ] = 0x01b7, 877 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 878 [ C(RESULT_MISS) ] = 0x01b7, 879 }, 880 [ C(OP_WRITE) ] = { 881 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 882 [ C(RESULT_ACCESS) ] = 0x01b7, 883 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 884 [ C(RESULT_MISS) ] = 0x01b7, 885 }, 886 [ C(OP_PREFETCH) ] = { 887 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 888 [ C(RESULT_ACCESS) ] = 0x01b7, 889 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 890 [ C(RESULT_MISS) ] = 0x01b7, 891 }, 892 }, 893 [ C(DTLB) ] = { 894 [ C(OP_READ) ] = { 895 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ 896 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ 897 }, 898 [ C(OP_WRITE) ] = { 899 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ 900 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 901 }, 902 [ C(OP_PREFETCH) ] = { 903 [ C(RESULT_ACCESS) ] = 0x0, 904 [ C(RESULT_MISS) ] = 0x0, 905 }, 906 }, 907 [ C(ITLB) ] = { 908 [ C(OP_READ) ] = { 909 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ 910 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ 911 }, 912 [ C(OP_WRITE) ] = { 913 [ C(RESULT_ACCESS) ] = -1, 914 [ C(RESULT_MISS) ] = -1, 915 }, 916 [ C(OP_PREFETCH) ] = { 917 [ C(RESULT_ACCESS) ] = -1, 918 [ C(RESULT_MISS) ] = -1, 919 }, 920 }, 921 [ C(BPU ) ] = { 922 [ C(OP_READ) ] = { 923 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 924 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 925 }, 926 [ C(OP_WRITE) ] = { 927 [ C(RESULT_ACCESS) ] = -1, 928 [ C(RESULT_MISS) ] = -1, 929 }, 930 [ C(OP_PREFETCH) ] = { 931 [ C(RESULT_ACCESS) ] = -1, 932 [ C(RESULT_MISS) ] = -1, 933 }, 934 }, 935 [ C(NODE) ] = { 936 [ C(OP_READ) ] = { 937 [ C(RESULT_ACCESS) ] = 0x01b7, 938 [ C(RESULT_MISS) ] = 0x01b7, 939 }, 940 [ C(OP_WRITE) ] = { 941 [ C(RESULT_ACCESS) ] = 0x01b7, 942 [ C(RESULT_MISS) ] = 0x01b7, 943 }, 944 [ C(OP_PREFETCH) ] = { 945 [ C(RESULT_ACCESS) ] = 0x01b7, 946 [ C(RESULT_MISS) ] = 0x01b7, 947 }, 948 }, 949 950 }; 951 952 /* 953 * Notes on the events: 954 * - data reads do not include code reads (comparable to earlier tables) 955 * - data counts include speculative execution (except L1 write, dtlb, bpu) 956 * - remote node access includes remote memory, remote cache, remote mmio. 957 * - prefetches are not included in the counts because they are not 958 * reliably counted. 959 */ 960 961 #define HSW_DEMAND_DATA_RD BIT_ULL(0) 962 #define HSW_DEMAND_RFO BIT_ULL(1) 963 #define HSW_ANY_RESPONSE BIT_ULL(16) 964 #define HSW_SUPPLIER_NONE BIT_ULL(17) 965 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) 966 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) 967 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) 968 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) 969 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ 970 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 971 HSW_L3_MISS_REMOTE_HOP2P) 972 #define HSW_SNOOP_NONE BIT_ULL(31) 973 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) 974 #define HSW_SNOOP_MISS BIT_ULL(33) 975 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) 976 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) 977 #define HSW_SNOOP_HITM BIT_ULL(36) 978 #define HSW_SNOOP_NON_DRAM BIT_ULL(37) 979 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ 980 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ 981 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ 982 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) 983 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) 984 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD 985 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO 986 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ 987 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) 988 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE 989 990 #define BDW_L3_MISS_LOCAL BIT(26) 991 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ 992 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 993 HSW_L3_MISS_REMOTE_HOP2P) 994 995 996 static __initconst const u64 hsw_hw_cache_event_ids 997 [PERF_COUNT_HW_CACHE_MAX] 998 [PERF_COUNT_HW_CACHE_OP_MAX] 999 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1000 { 1001 [ C(L1D ) ] = { 1002 [ C(OP_READ) ] = { 1003 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1004 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 1005 }, 1006 [ C(OP_WRITE) ] = { 1007 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1008 [ C(RESULT_MISS) ] = 0x0, 1009 }, 1010 [ C(OP_PREFETCH) ] = { 1011 [ C(RESULT_ACCESS) ] = 0x0, 1012 [ C(RESULT_MISS) ] = 0x0, 1013 }, 1014 }, 1015 [ C(L1I ) ] = { 1016 [ C(OP_READ) ] = { 1017 [ C(RESULT_ACCESS) ] = 0x0, 1018 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ 1019 }, 1020 [ C(OP_WRITE) ] = { 1021 [ C(RESULT_ACCESS) ] = -1, 1022 [ C(RESULT_MISS) ] = -1, 1023 }, 1024 [ C(OP_PREFETCH) ] = { 1025 [ C(RESULT_ACCESS) ] = 0x0, 1026 [ C(RESULT_MISS) ] = 0x0, 1027 }, 1028 }, 1029 [ C(LL ) ] = { 1030 [ C(OP_READ) ] = { 1031 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1032 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1033 }, 1034 [ C(OP_WRITE) ] = { 1035 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1036 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1037 }, 1038 [ C(OP_PREFETCH) ] = { 1039 [ C(RESULT_ACCESS) ] = 0x0, 1040 [ C(RESULT_MISS) ] = 0x0, 1041 }, 1042 }, 1043 [ C(DTLB) ] = { 1044 [ C(OP_READ) ] = { 1045 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1046 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ 1047 }, 1048 [ C(OP_WRITE) ] = { 1049 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1050 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 1051 }, 1052 [ C(OP_PREFETCH) ] = { 1053 [ C(RESULT_ACCESS) ] = 0x0, 1054 [ C(RESULT_MISS) ] = 0x0, 1055 }, 1056 }, 1057 [ C(ITLB) ] = { 1058 [ C(OP_READ) ] = { 1059 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ 1060 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ 1061 }, 1062 [ C(OP_WRITE) ] = { 1063 [ C(RESULT_ACCESS) ] = -1, 1064 [ C(RESULT_MISS) ] = -1, 1065 }, 1066 [ C(OP_PREFETCH) ] = { 1067 [ C(RESULT_ACCESS) ] = -1, 1068 [ C(RESULT_MISS) ] = -1, 1069 }, 1070 }, 1071 [ C(BPU ) ] = { 1072 [ C(OP_READ) ] = { 1073 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1074 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1075 }, 1076 [ C(OP_WRITE) ] = { 1077 [ C(RESULT_ACCESS) ] = -1, 1078 [ C(RESULT_MISS) ] = -1, 1079 }, 1080 [ C(OP_PREFETCH) ] = { 1081 [ C(RESULT_ACCESS) ] = -1, 1082 [ C(RESULT_MISS) ] = -1, 1083 }, 1084 }, 1085 [ C(NODE) ] = { 1086 [ C(OP_READ) ] = { 1087 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1088 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1089 }, 1090 [ C(OP_WRITE) ] = { 1091 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1092 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 1093 }, 1094 [ C(OP_PREFETCH) ] = { 1095 [ C(RESULT_ACCESS) ] = 0x0, 1096 [ C(RESULT_MISS) ] = 0x0, 1097 }, 1098 }, 1099 }; 1100 1101 static __initconst const u64 hsw_hw_cache_extra_regs 1102 [PERF_COUNT_HW_CACHE_MAX] 1103 [PERF_COUNT_HW_CACHE_OP_MAX] 1104 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1105 { 1106 [ C(LL ) ] = { 1107 [ C(OP_READ) ] = { 1108 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1109 HSW_LLC_ACCESS, 1110 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1111 HSW_L3_MISS|HSW_ANY_SNOOP, 1112 }, 1113 [ C(OP_WRITE) ] = { 1114 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1115 HSW_LLC_ACCESS, 1116 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1117 HSW_L3_MISS|HSW_ANY_SNOOP, 1118 }, 1119 [ C(OP_PREFETCH) ] = { 1120 [ C(RESULT_ACCESS) ] = 0x0, 1121 [ C(RESULT_MISS) ] = 0x0, 1122 }, 1123 }, 1124 [ C(NODE) ] = { 1125 [ C(OP_READ) ] = { 1126 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 1127 HSW_L3_MISS_LOCAL_DRAM| 1128 HSW_SNOOP_DRAM, 1129 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 1130 HSW_L3_MISS_REMOTE| 1131 HSW_SNOOP_DRAM, 1132 }, 1133 [ C(OP_WRITE) ] = { 1134 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 1135 HSW_L3_MISS_LOCAL_DRAM| 1136 HSW_SNOOP_DRAM, 1137 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 1138 HSW_L3_MISS_REMOTE| 1139 HSW_SNOOP_DRAM, 1140 }, 1141 [ C(OP_PREFETCH) ] = { 1142 [ C(RESULT_ACCESS) ] = 0x0, 1143 [ C(RESULT_MISS) ] = 0x0, 1144 }, 1145 }, 1146 }; 1147 1148 static __initconst const u64 westmere_hw_cache_event_ids 1149 [PERF_COUNT_HW_CACHE_MAX] 1150 [PERF_COUNT_HW_CACHE_OP_MAX] 1151 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1152 { 1153 [ C(L1D) ] = { 1154 [ C(OP_READ) ] = { 1155 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1156 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1157 }, 1158 [ C(OP_WRITE) ] = { 1159 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1160 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1161 }, 1162 [ C(OP_PREFETCH) ] = { 1163 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1164 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1165 }, 1166 }, 1167 [ C(L1I ) ] = { 1168 [ C(OP_READ) ] = { 1169 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1170 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1171 }, 1172 [ C(OP_WRITE) ] = { 1173 [ C(RESULT_ACCESS) ] = -1, 1174 [ C(RESULT_MISS) ] = -1, 1175 }, 1176 [ C(OP_PREFETCH) ] = { 1177 [ C(RESULT_ACCESS) ] = 0x0, 1178 [ C(RESULT_MISS) ] = 0x0, 1179 }, 1180 }, 1181 [ C(LL ) ] = { 1182 [ C(OP_READ) ] = { 1183 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1184 [ C(RESULT_ACCESS) ] = 0x01b7, 1185 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1186 [ C(RESULT_MISS) ] = 0x01b7, 1187 }, 1188 /* 1189 * Use RFO, not WRITEBACK, because a write miss would typically occur 1190 * on RFO. 1191 */ 1192 [ C(OP_WRITE) ] = { 1193 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1194 [ C(RESULT_ACCESS) ] = 0x01b7, 1195 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1196 [ C(RESULT_MISS) ] = 0x01b7, 1197 }, 1198 [ C(OP_PREFETCH) ] = { 1199 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1200 [ C(RESULT_ACCESS) ] = 0x01b7, 1201 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1202 [ C(RESULT_MISS) ] = 0x01b7, 1203 }, 1204 }, 1205 [ C(DTLB) ] = { 1206 [ C(OP_READ) ] = { 1207 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1208 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1209 }, 1210 [ C(OP_WRITE) ] = { 1211 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1212 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1213 }, 1214 [ C(OP_PREFETCH) ] = { 1215 [ C(RESULT_ACCESS) ] = 0x0, 1216 [ C(RESULT_MISS) ] = 0x0, 1217 }, 1218 }, 1219 [ C(ITLB) ] = { 1220 [ C(OP_READ) ] = { 1221 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1222 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ 1223 }, 1224 [ C(OP_WRITE) ] = { 1225 [ C(RESULT_ACCESS) ] = -1, 1226 [ C(RESULT_MISS) ] = -1, 1227 }, 1228 [ C(OP_PREFETCH) ] = { 1229 [ C(RESULT_ACCESS) ] = -1, 1230 [ C(RESULT_MISS) ] = -1, 1231 }, 1232 }, 1233 [ C(BPU ) ] = { 1234 [ C(OP_READ) ] = { 1235 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1236 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1237 }, 1238 [ C(OP_WRITE) ] = { 1239 [ C(RESULT_ACCESS) ] = -1, 1240 [ C(RESULT_MISS) ] = -1, 1241 }, 1242 [ C(OP_PREFETCH) ] = { 1243 [ C(RESULT_ACCESS) ] = -1, 1244 [ C(RESULT_MISS) ] = -1, 1245 }, 1246 }, 1247 [ C(NODE) ] = { 1248 [ C(OP_READ) ] = { 1249 [ C(RESULT_ACCESS) ] = 0x01b7, 1250 [ C(RESULT_MISS) ] = 0x01b7, 1251 }, 1252 [ C(OP_WRITE) ] = { 1253 [ C(RESULT_ACCESS) ] = 0x01b7, 1254 [ C(RESULT_MISS) ] = 0x01b7, 1255 }, 1256 [ C(OP_PREFETCH) ] = { 1257 [ C(RESULT_ACCESS) ] = 0x01b7, 1258 [ C(RESULT_MISS) ] = 0x01b7, 1259 }, 1260 }, 1261 }; 1262 1263 /* 1264 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; 1265 * See IA32 SDM Vol 3B 30.6.1.3 1266 */ 1267 1268 #define NHM_DMND_DATA_RD (1 << 0) 1269 #define NHM_DMND_RFO (1 << 1) 1270 #define NHM_DMND_IFETCH (1 << 2) 1271 #define NHM_DMND_WB (1 << 3) 1272 #define NHM_PF_DATA_RD (1 << 4) 1273 #define NHM_PF_DATA_RFO (1 << 5) 1274 #define NHM_PF_IFETCH (1 << 6) 1275 #define NHM_OFFCORE_OTHER (1 << 7) 1276 #define NHM_UNCORE_HIT (1 << 8) 1277 #define NHM_OTHER_CORE_HIT_SNP (1 << 9) 1278 #define NHM_OTHER_CORE_HITM (1 << 10) 1279 /* reserved */ 1280 #define NHM_REMOTE_CACHE_FWD (1 << 12) 1281 #define NHM_REMOTE_DRAM (1 << 13) 1282 #define NHM_LOCAL_DRAM (1 << 14) 1283 #define NHM_NON_DRAM (1 << 15) 1284 1285 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) 1286 #define NHM_REMOTE (NHM_REMOTE_DRAM) 1287 1288 #define NHM_DMND_READ (NHM_DMND_DATA_RD) 1289 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) 1290 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) 1291 1292 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) 1293 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) 1294 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) 1295 1296 static __initconst const u64 nehalem_hw_cache_extra_regs 1297 [PERF_COUNT_HW_CACHE_MAX] 1298 [PERF_COUNT_HW_CACHE_OP_MAX] 1299 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1300 { 1301 [ C(LL ) ] = { 1302 [ C(OP_READ) ] = { 1303 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, 1304 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, 1305 }, 1306 [ C(OP_WRITE) ] = { 1307 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, 1308 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, 1309 }, 1310 [ C(OP_PREFETCH) ] = { 1311 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, 1312 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, 1313 }, 1314 }, 1315 [ C(NODE) ] = { 1316 [ C(OP_READ) ] = { 1317 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, 1318 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, 1319 }, 1320 [ C(OP_WRITE) ] = { 1321 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, 1322 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, 1323 }, 1324 [ C(OP_PREFETCH) ] = { 1325 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, 1326 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, 1327 }, 1328 }, 1329 }; 1330 1331 static __initconst const u64 nehalem_hw_cache_event_ids 1332 [PERF_COUNT_HW_CACHE_MAX] 1333 [PERF_COUNT_HW_CACHE_OP_MAX] 1334 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1335 { 1336 [ C(L1D) ] = { 1337 [ C(OP_READ) ] = { 1338 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1339 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1340 }, 1341 [ C(OP_WRITE) ] = { 1342 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1343 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1344 }, 1345 [ C(OP_PREFETCH) ] = { 1346 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1347 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1348 }, 1349 }, 1350 [ C(L1I ) ] = { 1351 [ C(OP_READ) ] = { 1352 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1353 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1354 }, 1355 [ C(OP_WRITE) ] = { 1356 [ C(RESULT_ACCESS) ] = -1, 1357 [ C(RESULT_MISS) ] = -1, 1358 }, 1359 [ C(OP_PREFETCH) ] = { 1360 [ C(RESULT_ACCESS) ] = 0x0, 1361 [ C(RESULT_MISS) ] = 0x0, 1362 }, 1363 }, 1364 [ C(LL ) ] = { 1365 [ C(OP_READ) ] = { 1366 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1367 [ C(RESULT_ACCESS) ] = 0x01b7, 1368 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1369 [ C(RESULT_MISS) ] = 0x01b7, 1370 }, 1371 /* 1372 * Use RFO, not WRITEBACK, because a write miss would typically occur 1373 * on RFO. 1374 */ 1375 [ C(OP_WRITE) ] = { 1376 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1377 [ C(RESULT_ACCESS) ] = 0x01b7, 1378 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1379 [ C(RESULT_MISS) ] = 0x01b7, 1380 }, 1381 [ C(OP_PREFETCH) ] = { 1382 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1383 [ C(RESULT_ACCESS) ] = 0x01b7, 1384 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1385 [ C(RESULT_MISS) ] = 0x01b7, 1386 }, 1387 }, 1388 [ C(DTLB) ] = { 1389 [ C(OP_READ) ] = { 1390 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1391 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1392 }, 1393 [ C(OP_WRITE) ] = { 1394 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1395 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1396 }, 1397 [ C(OP_PREFETCH) ] = { 1398 [ C(RESULT_ACCESS) ] = 0x0, 1399 [ C(RESULT_MISS) ] = 0x0, 1400 }, 1401 }, 1402 [ C(ITLB) ] = { 1403 [ C(OP_READ) ] = { 1404 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1405 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ 1406 }, 1407 [ C(OP_WRITE) ] = { 1408 [ C(RESULT_ACCESS) ] = -1, 1409 [ C(RESULT_MISS) ] = -1, 1410 }, 1411 [ C(OP_PREFETCH) ] = { 1412 [ C(RESULT_ACCESS) ] = -1, 1413 [ C(RESULT_MISS) ] = -1, 1414 }, 1415 }, 1416 [ C(BPU ) ] = { 1417 [ C(OP_READ) ] = { 1418 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1419 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1420 }, 1421 [ C(OP_WRITE) ] = { 1422 [ C(RESULT_ACCESS) ] = -1, 1423 [ C(RESULT_MISS) ] = -1, 1424 }, 1425 [ C(OP_PREFETCH) ] = { 1426 [ C(RESULT_ACCESS) ] = -1, 1427 [ C(RESULT_MISS) ] = -1, 1428 }, 1429 }, 1430 [ C(NODE) ] = { 1431 [ C(OP_READ) ] = { 1432 [ C(RESULT_ACCESS) ] = 0x01b7, 1433 [ C(RESULT_MISS) ] = 0x01b7, 1434 }, 1435 [ C(OP_WRITE) ] = { 1436 [ C(RESULT_ACCESS) ] = 0x01b7, 1437 [ C(RESULT_MISS) ] = 0x01b7, 1438 }, 1439 [ C(OP_PREFETCH) ] = { 1440 [ C(RESULT_ACCESS) ] = 0x01b7, 1441 [ C(RESULT_MISS) ] = 0x01b7, 1442 }, 1443 }, 1444 }; 1445 1446 static __initconst const u64 core2_hw_cache_event_ids 1447 [PERF_COUNT_HW_CACHE_MAX] 1448 [PERF_COUNT_HW_CACHE_OP_MAX] 1449 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1450 { 1451 [ C(L1D) ] = { 1452 [ C(OP_READ) ] = { 1453 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ 1454 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ 1455 }, 1456 [ C(OP_WRITE) ] = { 1457 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ 1458 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ 1459 }, 1460 [ C(OP_PREFETCH) ] = { 1461 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ 1462 [ C(RESULT_MISS) ] = 0, 1463 }, 1464 }, 1465 [ C(L1I ) ] = { 1466 [ C(OP_READ) ] = { 1467 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ 1468 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ 1469 }, 1470 [ C(OP_WRITE) ] = { 1471 [ C(RESULT_ACCESS) ] = -1, 1472 [ C(RESULT_MISS) ] = -1, 1473 }, 1474 [ C(OP_PREFETCH) ] = { 1475 [ C(RESULT_ACCESS) ] = 0, 1476 [ C(RESULT_MISS) ] = 0, 1477 }, 1478 }, 1479 [ C(LL ) ] = { 1480 [ C(OP_READ) ] = { 1481 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1482 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1483 }, 1484 [ C(OP_WRITE) ] = { 1485 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1486 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1487 }, 1488 [ C(OP_PREFETCH) ] = { 1489 [ C(RESULT_ACCESS) ] = 0, 1490 [ C(RESULT_MISS) ] = 0, 1491 }, 1492 }, 1493 [ C(DTLB) ] = { 1494 [ C(OP_READ) ] = { 1495 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1496 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ 1497 }, 1498 [ C(OP_WRITE) ] = { 1499 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1500 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ 1501 }, 1502 [ C(OP_PREFETCH) ] = { 1503 [ C(RESULT_ACCESS) ] = 0, 1504 [ C(RESULT_MISS) ] = 0, 1505 }, 1506 }, 1507 [ C(ITLB) ] = { 1508 [ C(OP_READ) ] = { 1509 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1510 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ 1511 }, 1512 [ C(OP_WRITE) ] = { 1513 [ C(RESULT_ACCESS) ] = -1, 1514 [ C(RESULT_MISS) ] = -1, 1515 }, 1516 [ C(OP_PREFETCH) ] = { 1517 [ C(RESULT_ACCESS) ] = -1, 1518 [ C(RESULT_MISS) ] = -1, 1519 }, 1520 }, 1521 [ C(BPU ) ] = { 1522 [ C(OP_READ) ] = { 1523 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1524 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1525 }, 1526 [ C(OP_WRITE) ] = { 1527 [ C(RESULT_ACCESS) ] = -1, 1528 [ C(RESULT_MISS) ] = -1, 1529 }, 1530 [ C(OP_PREFETCH) ] = { 1531 [ C(RESULT_ACCESS) ] = -1, 1532 [ C(RESULT_MISS) ] = -1, 1533 }, 1534 }, 1535 }; 1536 1537 static __initconst const u64 atom_hw_cache_event_ids 1538 [PERF_COUNT_HW_CACHE_MAX] 1539 [PERF_COUNT_HW_CACHE_OP_MAX] 1540 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1541 { 1542 [ C(L1D) ] = { 1543 [ C(OP_READ) ] = { 1544 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ 1545 [ C(RESULT_MISS) ] = 0, 1546 }, 1547 [ C(OP_WRITE) ] = { 1548 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ 1549 [ C(RESULT_MISS) ] = 0, 1550 }, 1551 [ C(OP_PREFETCH) ] = { 1552 [ C(RESULT_ACCESS) ] = 0x0, 1553 [ C(RESULT_MISS) ] = 0, 1554 }, 1555 }, 1556 [ C(L1I ) ] = { 1557 [ C(OP_READ) ] = { 1558 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1559 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1560 }, 1561 [ C(OP_WRITE) ] = { 1562 [ C(RESULT_ACCESS) ] = -1, 1563 [ C(RESULT_MISS) ] = -1, 1564 }, 1565 [ C(OP_PREFETCH) ] = { 1566 [ C(RESULT_ACCESS) ] = 0, 1567 [ C(RESULT_MISS) ] = 0, 1568 }, 1569 }, 1570 [ C(LL ) ] = { 1571 [ C(OP_READ) ] = { 1572 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1573 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1574 }, 1575 [ C(OP_WRITE) ] = { 1576 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1577 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1578 }, 1579 [ C(OP_PREFETCH) ] = { 1580 [ C(RESULT_ACCESS) ] = 0, 1581 [ C(RESULT_MISS) ] = 0, 1582 }, 1583 }, 1584 [ C(DTLB) ] = { 1585 [ C(OP_READ) ] = { 1586 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ 1587 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ 1588 }, 1589 [ C(OP_WRITE) ] = { 1590 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ 1591 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ 1592 }, 1593 [ C(OP_PREFETCH) ] = { 1594 [ C(RESULT_ACCESS) ] = 0, 1595 [ C(RESULT_MISS) ] = 0, 1596 }, 1597 }, 1598 [ C(ITLB) ] = { 1599 [ C(OP_READ) ] = { 1600 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1601 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ 1602 }, 1603 [ C(OP_WRITE) ] = { 1604 [ C(RESULT_ACCESS) ] = -1, 1605 [ C(RESULT_MISS) ] = -1, 1606 }, 1607 [ C(OP_PREFETCH) ] = { 1608 [ C(RESULT_ACCESS) ] = -1, 1609 [ C(RESULT_MISS) ] = -1, 1610 }, 1611 }, 1612 [ C(BPU ) ] = { 1613 [ C(OP_READ) ] = { 1614 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1615 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1616 }, 1617 [ C(OP_WRITE) ] = { 1618 [ C(RESULT_ACCESS) ] = -1, 1619 [ C(RESULT_MISS) ] = -1, 1620 }, 1621 [ C(OP_PREFETCH) ] = { 1622 [ C(RESULT_ACCESS) ] = -1, 1623 [ C(RESULT_MISS) ] = -1, 1624 }, 1625 }, 1626 }; 1627 1628 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c"); 1629 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2"); 1630 /* no_alloc_cycles.not_delivered */ 1631 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm, 1632 "event=0xca,umask=0x50"); 1633 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2"); 1634 /* uops_retired.all */ 1635 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm, 1636 "event=0xc2,umask=0x10"); 1637 /* uops_retired.all */ 1638 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm, 1639 "event=0xc2,umask=0x10"); 1640 1641 static struct attribute *slm_events_attrs[] = { 1642 EVENT_PTR(td_total_slots_slm), 1643 EVENT_PTR(td_total_slots_scale_slm), 1644 EVENT_PTR(td_fetch_bubbles_slm), 1645 EVENT_PTR(td_fetch_bubbles_scale_slm), 1646 EVENT_PTR(td_slots_issued_slm), 1647 EVENT_PTR(td_slots_retired_slm), 1648 NULL 1649 }; 1650 1651 static struct extra_reg intel_slm_extra_regs[] __read_mostly = 1652 { 1653 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1654 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), 1655 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1), 1656 EVENT_EXTRA_END 1657 }; 1658 1659 #define SLM_DMND_READ SNB_DMND_DATA_RD 1660 #define SLM_DMND_WRITE SNB_DMND_RFO 1661 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1662 1663 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) 1664 #define SLM_LLC_ACCESS SNB_RESP_ANY 1665 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) 1666 1667 static __initconst const u64 slm_hw_cache_extra_regs 1668 [PERF_COUNT_HW_CACHE_MAX] 1669 [PERF_COUNT_HW_CACHE_OP_MAX] 1670 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1671 { 1672 [ C(LL ) ] = { 1673 [ C(OP_READ) ] = { 1674 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, 1675 [ C(RESULT_MISS) ] = 0, 1676 }, 1677 [ C(OP_WRITE) ] = { 1678 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, 1679 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, 1680 }, 1681 [ C(OP_PREFETCH) ] = { 1682 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, 1683 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, 1684 }, 1685 }, 1686 }; 1687 1688 static __initconst const u64 slm_hw_cache_event_ids 1689 [PERF_COUNT_HW_CACHE_MAX] 1690 [PERF_COUNT_HW_CACHE_OP_MAX] 1691 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1692 { 1693 [ C(L1D) ] = { 1694 [ C(OP_READ) ] = { 1695 [ C(RESULT_ACCESS) ] = 0, 1696 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ 1697 }, 1698 [ C(OP_WRITE) ] = { 1699 [ C(RESULT_ACCESS) ] = 0, 1700 [ C(RESULT_MISS) ] = 0, 1701 }, 1702 [ C(OP_PREFETCH) ] = { 1703 [ C(RESULT_ACCESS) ] = 0, 1704 [ C(RESULT_MISS) ] = 0, 1705 }, 1706 }, 1707 [ C(L1I ) ] = { 1708 [ C(OP_READ) ] = { 1709 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ 1710 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ 1711 }, 1712 [ C(OP_WRITE) ] = { 1713 [ C(RESULT_ACCESS) ] = -1, 1714 [ C(RESULT_MISS) ] = -1, 1715 }, 1716 [ C(OP_PREFETCH) ] = { 1717 [ C(RESULT_ACCESS) ] = 0, 1718 [ C(RESULT_MISS) ] = 0, 1719 }, 1720 }, 1721 [ C(LL ) ] = { 1722 [ C(OP_READ) ] = { 1723 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1724 [ C(RESULT_ACCESS) ] = 0x01b7, 1725 [ C(RESULT_MISS) ] = 0, 1726 }, 1727 [ C(OP_WRITE) ] = { 1728 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1729 [ C(RESULT_ACCESS) ] = 0x01b7, 1730 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1731 [ C(RESULT_MISS) ] = 0x01b7, 1732 }, 1733 [ C(OP_PREFETCH) ] = { 1734 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1735 [ C(RESULT_ACCESS) ] = 0x01b7, 1736 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1737 [ C(RESULT_MISS) ] = 0x01b7, 1738 }, 1739 }, 1740 [ C(DTLB) ] = { 1741 [ C(OP_READ) ] = { 1742 [ C(RESULT_ACCESS) ] = 0, 1743 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ 1744 }, 1745 [ C(OP_WRITE) ] = { 1746 [ C(RESULT_ACCESS) ] = 0, 1747 [ C(RESULT_MISS) ] = 0, 1748 }, 1749 [ C(OP_PREFETCH) ] = { 1750 [ C(RESULT_ACCESS) ] = 0, 1751 [ C(RESULT_MISS) ] = 0, 1752 }, 1753 }, 1754 [ C(ITLB) ] = { 1755 [ C(OP_READ) ] = { 1756 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1757 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */ 1758 }, 1759 [ C(OP_WRITE) ] = { 1760 [ C(RESULT_ACCESS) ] = -1, 1761 [ C(RESULT_MISS) ] = -1, 1762 }, 1763 [ C(OP_PREFETCH) ] = { 1764 [ C(RESULT_ACCESS) ] = -1, 1765 [ C(RESULT_MISS) ] = -1, 1766 }, 1767 }, 1768 [ C(BPU ) ] = { 1769 [ C(OP_READ) ] = { 1770 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1771 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1772 }, 1773 [ C(OP_WRITE) ] = { 1774 [ C(RESULT_ACCESS) ] = -1, 1775 [ C(RESULT_MISS) ] = -1, 1776 }, 1777 [ C(OP_PREFETCH) ] = { 1778 [ C(RESULT_ACCESS) ] = -1, 1779 [ C(RESULT_MISS) ] = -1, 1780 }, 1781 }, 1782 }; 1783 1784 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c"); 1785 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3"); 1786 /* UOPS_NOT_DELIVERED.ANY */ 1787 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c"); 1788 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */ 1789 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02"); 1790 /* UOPS_RETIRED.ANY */ 1791 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2"); 1792 /* UOPS_ISSUED.ANY */ 1793 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e"); 1794 1795 static struct attribute *glm_events_attrs[] = { 1796 EVENT_PTR(td_total_slots_glm), 1797 EVENT_PTR(td_total_slots_scale_glm), 1798 EVENT_PTR(td_fetch_bubbles_glm), 1799 EVENT_PTR(td_recovery_bubbles_glm), 1800 EVENT_PTR(td_slots_issued_glm), 1801 EVENT_PTR(td_slots_retired_glm), 1802 NULL 1803 }; 1804 1805 static struct extra_reg intel_glm_extra_regs[] __read_mostly = { 1806 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1807 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0), 1808 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1), 1809 EVENT_EXTRA_END 1810 }; 1811 1812 #define GLM_DEMAND_DATA_RD BIT_ULL(0) 1813 #define GLM_DEMAND_RFO BIT_ULL(1) 1814 #define GLM_ANY_RESPONSE BIT_ULL(16) 1815 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33) 1816 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD 1817 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO 1818 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1819 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE 1820 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM) 1821 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM) 1822 1823 static __initconst const u64 glm_hw_cache_event_ids 1824 [PERF_COUNT_HW_CACHE_MAX] 1825 [PERF_COUNT_HW_CACHE_OP_MAX] 1826 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1827 [C(L1D)] = { 1828 [C(OP_READ)] = { 1829 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1830 [C(RESULT_MISS)] = 0x0, 1831 }, 1832 [C(OP_WRITE)] = { 1833 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1834 [C(RESULT_MISS)] = 0x0, 1835 }, 1836 [C(OP_PREFETCH)] = { 1837 [C(RESULT_ACCESS)] = 0x0, 1838 [C(RESULT_MISS)] = 0x0, 1839 }, 1840 }, 1841 [C(L1I)] = { 1842 [C(OP_READ)] = { 1843 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1844 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1845 }, 1846 [C(OP_WRITE)] = { 1847 [C(RESULT_ACCESS)] = -1, 1848 [C(RESULT_MISS)] = -1, 1849 }, 1850 [C(OP_PREFETCH)] = { 1851 [C(RESULT_ACCESS)] = 0x0, 1852 [C(RESULT_MISS)] = 0x0, 1853 }, 1854 }, 1855 [C(LL)] = { 1856 [C(OP_READ)] = { 1857 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1858 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1859 }, 1860 [C(OP_WRITE)] = { 1861 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1862 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1863 }, 1864 [C(OP_PREFETCH)] = { 1865 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1866 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1867 }, 1868 }, 1869 [C(DTLB)] = { 1870 [C(OP_READ)] = { 1871 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1872 [C(RESULT_MISS)] = 0x0, 1873 }, 1874 [C(OP_WRITE)] = { 1875 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1876 [C(RESULT_MISS)] = 0x0, 1877 }, 1878 [C(OP_PREFETCH)] = { 1879 [C(RESULT_ACCESS)] = 0x0, 1880 [C(RESULT_MISS)] = 0x0, 1881 }, 1882 }, 1883 [C(ITLB)] = { 1884 [C(OP_READ)] = { 1885 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 1886 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 1887 }, 1888 [C(OP_WRITE)] = { 1889 [C(RESULT_ACCESS)] = -1, 1890 [C(RESULT_MISS)] = -1, 1891 }, 1892 [C(OP_PREFETCH)] = { 1893 [C(RESULT_ACCESS)] = -1, 1894 [C(RESULT_MISS)] = -1, 1895 }, 1896 }, 1897 [C(BPU)] = { 1898 [C(OP_READ)] = { 1899 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1900 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1901 }, 1902 [C(OP_WRITE)] = { 1903 [C(RESULT_ACCESS)] = -1, 1904 [C(RESULT_MISS)] = -1, 1905 }, 1906 [C(OP_PREFETCH)] = { 1907 [C(RESULT_ACCESS)] = -1, 1908 [C(RESULT_MISS)] = -1, 1909 }, 1910 }, 1911 }; 1912 1913 static __initconst const u64 glm_hw_cache_extra_regs 1914 [PERF_COUNT_HW_CACHE_MAX] 1915 [PERF_COUNT_HW_CACHE_OP_MAX] 1916 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1917 [C(LL)] = { 1918 [C(OP_READ)] = { 1919 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 1920 GLM_LLC_ACCESS, 1921 [C(RESULT_MISS)] = GLM_DEMAND_READ| 1922 GLM_LLC_MISS, 1923 }, 1924 [C(OP_WRITE)] = { 1925 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 1926 GLM_LLC_ACCESS, 1927 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 1928 GLM_LLC_MISS, 1929 }, 1930 [C(OP_PREFETCH)] = { 1931 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH| 1932 GLM_LLC_ACCESS, 1933 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH| 1934 GLM_LLC_MISS, 1935 }, 1936 }, 1937 }; 1938 1939 static __initconst const u64 glp_hw_cache_event_ids 1940 [PERF_COUNT_HW_CACHE_MAX] 1941 [PERF_COUNT_HW_CACHE_OP_MAX] 1942 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1943 [C(L1D)] = { 1944 [C(OP_READ)] = { 1945 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1946 [C(RESULT_MISS)] = 0x0, 1947 }, 1948 [C(OP_WRITE)] = { 1949 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1950 [C(RESULT_MISS)] = 0x0, 1951 }, 1952 [C(OP_PREFETCH)] = { 1953 [C(RESULT_ACCESS)] = 0x0, 1954 [C(RESULT_MISS)] = 0x0, 1955 }, 1956 }, 1957 [C(L1I)] = { 1958 [C(OP_READ)] = { 1959 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1960 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1961 }, 1962 [C(OP_WRITE)] = { 1963 [C(RESULT_ACCESS)] = -1, 1964 [C(RESULT_MISS)] = -1, 1965 }, 1966 [C(OP_PREFETCH)] = { 1967 [C(RESULT_ACCESS)] = 0x0, 1968 [C(RESULT_MISS)] = 0x0, 1969 }, 1970 }, 1971 [C(LL)] = { 1972 [C(OP_READ)] = { 1973 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1974 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1975 }, 1976 [C(OP_WRITE)] = { 1977 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1978 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1979 }, 1980 [C(OP_PREFETCH)] = { 1981 [C(RESULT_ACCESS)] = 0x0, 1982 [C(RESULT_MISS)] = 0x0, 1983 }, 1984 }, 1985 [C(DTLB)] = { 1986 [C(OP_READ)] = { 1987 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1988 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 1989 }, 1990 [C(OP_WRITE)] = { 1991 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1992 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 1993 }, 1994 [C(OP_PREFETCH)] = { 1995 [C(RESULT_ACCESS)] = 0x0, 1996 [C(RESULT_MISS)] = 0x0, 1997 }, 1998 }, 1999 [C(ITLB)] = { 2000 [C(OP_READ)] = { 2001 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 2002 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 2003 }, 2004 [C(OP_WRITE)] = { 2005 [C(RESULT_ACCESS)] = -1, 2006 [C(RESULT_MISS)] = -1, 2007 }, 2008 [C(OP_PREFETCH)] = { 2009 [C(RESULT_ACCESS)] = -1, 2010 [C(RESULT_MISS)] = -1, 2011 }, 2012 }, 2013 [C(BPU)] = { 2014 [C(OP_READ)] = { 2015 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 2016 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 2017 }, 2018 [C(OP_WRITE)] = { 2019 [C(RESULT_ACCESS)] = -1, 2020 [C(RESULT_MISS)] = -1, 2021 }, 2022 [C(OP_PREFETCH)] = { 2023 [C(RESULT_ACCESS)] = -1, 2024 [C(RESULT_MISS)] = -1, 2025 }, 2026 }, 2027 }; 2028 2029 static __initconst const u64 glp_hw_cache_extra_regs 2030 [PERF_COUNT_HW_CACHE_MAX] 2031 [PERF_COUNT_HW_CACHE_OP_MAX] 2032 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2033 [C(LL)] = { 2034 [C(OP_READ)] = { 2035 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 2036 GLM_LLC_ACCESS, 2037 [C(RESULT_MISS)] = GLM_DEMAND_READ| 2038 GLM_LLC_MISS, 2039 }, 2040 [C(OP_WRITE)] = { 2041 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 2042 GLM_LLC_ACCESS, 2043 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 2044 GLM_LLC_MISS, 2045 }, 2046 [C(OP_PREFETCH)] = { 2047 [C(RESULT_ACCESS)] = 0x0, 2048 [C(RESULT_MISS)] = 0x0, 2049 }, 2050 }, 2051 }; 2052 2053 #define TNT_LOCAL_DRAM BIT_ULL(26) 2054 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD 2055 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO 2056 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE 2057 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \ 2058 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM) 2059 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM) 2060 2061 static __initconst const u64 tnt_hw_cache_extra_regs 2062 [PERF_COUNT_HW_CACHE_MAX] 2063 [PERF_COUNT_HW_CACHE_OP_MAX] 2064 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2065 [C(LL)] = { 2066 [C(OP_READ)] = { 2067 [C(RESULT_ACCESS)] = TNT_DEMAND_READ| 2068 TNT_LLC_ACCESS, 2069 [C(RESULT_MISS)] = TNT_DEMAND_READ| 2070 TNT_LLC_MISS, 2071 }, 2072 [C(OP_WRITE)] = { 2073 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE| 2074 TNT_LLC_ACCESS, 2075 [C(RESULT_MISS)] = TNT_DEMAND_WRITE| 2076 TNT_LLC_MISS, 2077 }, 2078 [C(OP_PREFETCH)] = { 2079 [C(RESULT_ACCESS)] = 0x0, 2080 [C(RESULT_MISS)] = 0x0, 2081 }, 2082 }, 2083 }; 2084 2085 EVENT_ATTR_STR(topdown-fe-bound, td_fe_bound_tnt, "event=0x71,umask=0x0"); 2086 EVENT_ATTR_STR(topdown-retiring, td_retiring_tnt, "event=0xc2,umask=0x0"); 2087 EVENT_ATTR_STR(topdown-bad-spec, td_bad_spec_tnt, "event=0x73,umask=0x6"); 2088 EVENT_ATTR_STR(topdown-be-bound, td_be_bound_tnt, "event=0x74,umask=0x0"); 2089 2090 static struct attribute *tnt_events_attrs[] = { 2091 EVENT_PTR(td_fe_bound_tnt), 2092 EVENT_PTR(td_retiring_tnt), 2093 EVENT_PTR(td_bad_spec_tnt), 2094 EVENT_PTR(td_be_bound_tnt), 2095 NULL, 2096 }; 2097 2098 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = { 2099 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2100 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x800ff0ffffff9fffull, RSP_0), 2101 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xff0ffffff9fffull, RSP_1), 2102 EVENT_EXTRA_END 2103 }; 2104 2105 EVENT_ATTR_STR(mem-loads, mem_ld_grt, "event=0xd0,umask=0x5,ldlat=3"); 2106 EVENT_ATTR_STR(mem-stores, mem_st_grt, "event=0xd0,umask=0x6"); 2107 2108 static struct attribute *grt_mem_attrs[] = { 2109 EVENT_PTR(mem_ld_grt), 2110 EVENT_PTR(mem_st_grt), 2111 NULL 2112 }; 2113 2114 static struct extra_reg intel_grt_extra_regs[] __read_mostly = { 2115 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 2116 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffffffull, RSP_0), 2117 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x3fffffffffull, RSP_1), 2118 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x5d0), 2119 EVENT_EXTRA_END 2120 }; 2121 2122 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */ 2123 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */ 2124 #define KNL_MCDRAM_LOCAL BIT_ULL(21) 2125 #define KNL_MCDRAM_FAR BIT_ULL(22) 2126 #define KNL_DDR_LOCAL BIT_ULL(23) 2127 #define KNL_DDR_FAR BIT_ULL(24) 2128 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \ 2129 KNL_DDR_LOCAL | KNL_DDR_FAR) 2130 #define KNL_L2_READ SLM_DMND_READ 2131 #define KNL_L2_WRITE SLM_DMND_WRITE 2132 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH 2133 #define KNL_L2_ACCESS SLM_LLC_ACCESS 2134 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \ 2135 KNL_DRAM_ANY | SNB_SNP_ANY | \ 2136 SNB_NON_DRAM) 2137 2138 static __initconst const u64 knl_hw_cache_extra_regs 2139 [PERF_COUNT_HW_CACHE_MAX] 2140 [PERF_COUNT_HW_CACHE_OP_MAX] 2141 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 2142 [C(LL)] = { 2143 [C(OP_READ)] = { 2144 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS, 2145 [C(RESULT_MISS)] = 0, 2146 }, 2147 [C(OP_WRITE)] = { 2148 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS, 2149 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS, 2150 }, 2151 [C(OP_PREFETCH)] = { 2152 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS, 2153 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS, 2154 }, 2155 }, 2156 }; 2157 2158 /* 2159 * Used from PMIs where the LBRs are already disabled. 2160 * 2161 * This function could be called consecutively. It is required to remain in 2162 * disabled state if called consecutively. 2163 * 2164 * During consecutive calls, the same disable value will be written to related 2165 * registers, so the PMU state remains unchanged. 2166 * 2167 * intel_bts events don't coexist with intel PMU's BTS events because of 2168 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them 2169 * disabled around intel PMU's event batching etc, only inside the PMI handler. 2170 * 2171 * Avoid PEBS_ENABLE MSR access in PMIs. 2172 * The GLOBAL_CTRL has been disabled. All the counters do not count anymore. 2173 * It doesn't matter if the PEBS is enabled or not. 2174 * Usually, the PEBS status are not changed in PMIs. It's unnecessary to 2175 * access PEBS_ENABLE MSR in disable_all()/enable_all(). 2176 * However, there are some cases which may change PEBS status, e.g. PMI 2177 * throttle. The PEBS_ENABLE should be updated where the status changes. 2178 */ 2179 static __always_inline void __intel_pmu_disable_all(bool bts) 2180 { 2181 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2182 2183 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2184 2185 if (bts && test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) 2186 intel_pmu_disable_bts(); 2187 } 2188 2189 static __always_inline void intel_pmu_disable_all(void) 2190 { 2191 __intel_pmu_disable_all(true); 2192 intel_pmu_pebs_disable_all(); 2193 intel_pmu_lbr_disable_all(); 2194 } 2195 2196 static void __intel_pmu_enable_all(int added, bool pmi) 2197 { 2198 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2199 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2200 2201 intel_pmu_lbr_enable_all(pmi); 2202 2203 if (cpuc->fixed_ctrl_val != cpuc->active_fixed_ctrl_val) { 2204 wrmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, cpuc->fixed_ctrl_val); 2205 cpuc->active_fixed_ctrl_val = cpuc->fixed_ctrl_val; 2206 } 2207 2208 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 2209 intel_ctrl & ~cpuc->intel_ctrl_guest_mask); 2210 2211 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 2212 struct perf_event *event = 2213 cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; 2214 2215 if (WARN_ON_ONCE(!event)) 2216 return; 2217 2218 intel_pmu_enable_bts(event->hw.config); 2219 } 2220 } 2221 2222 static void intel_pmu_enable_all(int added) 2223 { 2224 intel_pmu_pebs_enable_all(); 2225 __intel_pmu_enable_all(added, false); 2226 } 2227 2228 static noinline int 2229 __intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, 2230 unsigned int cnt, unsigned long flags) 2231 { 2232 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2233 2234 intel_pmu_lbr_read(); 2235 cnt = min_t(unsigned int, cnt, x86_pmu.lbr_nr); 2236 2237 memcpy(entries, cpuc->lbr_entries, sizeof(struct perf_branch_entry) * cnt); 2238 intel_pmu_enable_all(0); 2239 local_irq_restore(flags); 2240 return cnt; 2241 } 2242 2243 static int 2244 intel_pmu_snapshot_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2245 { 2246 unsigned long flags; 2247 2248 /* must not have branches... */ 2249 local_irq_save(flags); 2250 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2251 __intel_pmu_lbr_disable(); 2252 /* ... until here */ 2253 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2254 } 2255 2256 static int 2257 intel_pmu_snapshot_arch_branch_stack(struct perf_branch_entry *entries, unsigned int cnt) 2258 { 2259 unsigned long flags; 2260 2261 /* must not have branches... */ 2262 local_irq_save(flags); 2263 __intel_pmu_disable_all(false); /* we don't care about BTS */ 2264 __intel_pmu_arch_lbr_disable(); 2265 /* ... until here */ 2266 return __intel_pmu_snapshot_branch_stack(entries, cnt, flags); 2267 } 2268 2269 /* 2270 * Workaround for: 2271 * Intel Errata AAK100 (model 26) 2272 * Intel Errata AAP53 (model 30) 2273 * Intel Errata BD53 (model 44) 2274 * 2275 * The official story: 2276 * These chips need to be 'reset' when adding counters by programming the 2277 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either 2278 * in sequence on the same PMC or on different PMCs. 2279 * 2280 * In practice it appears some of these events do in fact count, and 2281 * we need to program all 4 events. 2282 */ 2283 static void intel_pmu_nhm_workaround(void) 2284 { 2285 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2286 static const unsigned long nhm_magic[4] = { 2287 0x4300B5, 2288 0x4300D2, 2289 0x4300B1, 2290 0x4300B1 2291 }; 2292 struct perf_event *event; 2293 int i; 2294 2295 /* 2296 * The Errata requires below steps: 2297 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; 2298 * 2) Configure 4 PERFEVTSELx with the magic events and clear 2299 * the corresponding PMCx; 2300 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; 2301 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; 2302 * 5) Clear 4 pairs of ERFEVTSELx and PMCx; 2303 */ 2304 2305 /* 2306 * The real steps we choose are a little different from above. 2307 * A) To reduce MSR operations, we don't run step 1) as they 2308 * are already cleared before this function is called; 2309 * B) Call x86_perf_event_update to save PMCx before configuring 2310 * PERFEVTSELx with magic number; 2311 * C) With step 5), we do clear only when the PERFEVTSELx is 2312 * not used currently. 2313 * D) Call x86_perf_event_set_period to restore PMCx; 2314 */ 2315 2316 /* We always operate 4 pairs of PERF Counters */ 2317 for (i = 0; i < 4; i++) { 2318 event = cpuc->events[i]; 2319 if (event) 2320 static_call(x86_pmu_update)(event); 2321 } 2322 2323 for (i = 0; i < 4; i++) { 2324 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); 2325 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); 2326 } 2327 2328 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); 2329 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); 2330 2331 for (i = 0; i < 4; i++) { 2332 event = cpuc->events[i]; 2333 2334 if (event) { 2335 static_call(x86_pmu_set_period)(event); 2336 __x86_pmu_enable_event(&event->hw, 2337 ARCH_PERFMON_EVENTSEL_ENABLE); 2338 } else 2339 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); 2340 } 2341 } 2342 2343 static void intel_pmu_nhm_enable_all(int added) 2344 { 2345 if (added) 2346 intel_pmu_nhm_workaround(); 2347 intel_pmu_enable_all(added); 2348 } 2349 2350 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on) 2351 { 2352 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0; 2353 2354 if (cpuc->tfa_shadow != val) { 2355 cpuc->tfa_shadow = val; 2356 wrmsrl(MSR_TSX_FORCE_ABORT, val); 2357 } 2358 } 2359 2360 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2361 { 2362 /* 2363 * We're going to use PMC3, make sure TFA is set before we touch it. 2364 */ 2365 if (cntr == 3) 2366 intel_set_tfa(cpuc, true); 2367 } 2368 2369 static void intel_tfa_pmu_enable_all(int added) 2370 { 2371 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2372 2373 /* 2374 * If we find PMC3 is no longer used when we enable the PMU, we can 2375 * clear TFA. 2376 */ 2377 if (!test_bit(3, cpuc->active_mask)) 2378 intel_set_tfa(cpuc, false); 2379 2380 intel_pmu_enable_all(added); 2381 } 2382 2383 static inline u64 intel_pmu_get_status(void) 2384 { 2385 u64 status; 2386 2387 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); 2388 2389 return status; 2390 } 2391 2392 static inline void intel_pmu_ack_status(u64 ack) 2393 { 2394 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); 2395 } 2396 2397 static inline bool event_is_checkpointed(struct perf_event *event) 2398 { 2399 return unlikely(event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; 2400 } 2401 2402 static inline void intel_set_masks(struct perf_event *event, int idx) 2403 { 2404 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2405 2406 if (event->attr.exclude_host) 2407 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2408 if (event->attr.exclude_guest) 2409 __set_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2410 if (event_is_checkpointed(event)) 2411 __set_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2412 } 2413 2414 static inline void intel_clear_masks(struct perf_event *event, int idx) 2415 { 2416 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2417 2418 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_guest_mask); 2419 __clear_bit(idx, (unsigned long *)&cpuc->intel_ctrl_host_mask); 2420 __clear_bit(idx, (unsigned long *)&cpuc->intel_cp_status); 2421 } 2422 2423 static void intel_pmu_disable_fixed(struct perf_event *event) 2424 { 2425 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2426 struct hw_perf_event *hwc = &event->hw; 2427 int idx = hwc->idx; 2428 u64 mask; 2429 2430 if (is_topdown_idx(idx)) { 2431 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2432 2433 /* 2434 * When there are other active TopDown events, 2435 * don't disable the fixed counter 3. 2436 */ 2437 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2438 return; 2439 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2440 } 2441 2442 intel_clear_masks(event, idx); 2443 2444 mask = 0xfULL << ((idx - INTEL_PMC_IDX_FIXED) * 4); 2445 cpuc->fixed_ctrl_val &= ~mask; 2446 } 2447 2448 static void intel_pmu_disable_event(struct perf_event *event) 2449 { 2450 struct hw_perf_event *hwc = &event->hw; 2451 int idx = hwc->idx; 2452 2453 switch (idx) { 2454 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2455 intel_clear_masks(event, idx); 2456 x86_pmu_disable_event(event); 2457 break; 2458 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2459 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2460 intel_pmu_disable_fixed(event); 2461 break; 2462 case INTEL_PMC_IDX_FIXED_BTS: 2463 intel_pmu_disable_bts(); 2464 intel_pmu_drain_bts_buffer(); 2465 return; 2466 case INTEL_PMC_IDX_FIXED_VLBR: 2467 intel_clear_masks(event, idx); 2468 break; 2469 default: 2470 intel_clear_masks(event, idx); 2471 pr_warn("Failed to disable the event with invalid index %d\n", 2472 idx); 2473 return; 2474 } 2475 2476 /* 2477 * Needs to be called after x86_pmu_disable_event, 2478 * so we don't trigger the event without PEBS bit set. 2479 */ 2480 if (unlikely(event->attr.precise_ip)) 2481 intel_pmu_pebs_disable(event); 2482 } 2483 2484 static void intel_pmu_assign_event(struct perf_event *event, int idx) 2485 { 2486 if (is_pebs_pt(event)) 2487 perf_report_aux_output_id(event, idx); 2488 } 2489 2490 static void intel_pmu_del_event(struct perf_event *event) 2491 { 2492 if (needs_branch_stack(event)) 2493 intel_pmu_lbr_del(event); 2494 if (event->attr.precise_ip) 2495 intel_pmu_pebs_del(event); 2496 } 2497 2498 static int icl_set_topdown_event_period(struct perf_event *event) 2499 { 2500 struct hw_perf_event *hwc = &event->hw; 2501 s64 left = local64_read(&hwc->period_left); 2502 2503 /* 2504 * The values in PERF_METRICS MSR are derived from fixed counter 3. 2505 * Software should start both registers, PERF_METRICS and fixed 2506 * counter 3, from zero. 2507 * Clear PERF_METRICS and Fixed counter 3 in initialization. 2508 * After that, both MSRs will be cleared for each read. 2509 * Don't need to clear them again. 2510 */ 2511 if (left == x86_pmu.max_period) { 2512 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2513 wrmsrl(MSR_PERF_METRICS, 0); 2514 hwc->saved_slots = 0; 2515 hwc->saved_metric = 0; 2516 } 2517 2518 if ((hwc->saved_slots) && is_slots_event(event)) { 2519 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, hwc->saved_slots); 2520 wrmsrl(MSR_PERF_METRICS, hwc->saved_metric); 2521 } 2522 2523 perf_event_update_userpage(event); 2524 2525 return 0; 2526 } 2527 2528 static int adl_set_topdown_event_period(struct perf_event *event) 2529 { 2530 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 2531 2532 if (pmu->cpu_type != hybrid_big) 2533 return 0; 2534 2535 return icl_set_topdown_event_period(event); 2536 } 2537 2538 DEFINE_STATIC_CALL(intel_pmu_set_topdown_event_period, x86_perf_event_set_period); 2539 2540 static inline u64 icl_get_metrics_event_value(u64 metric, u64 slots, int idx) 2541 { 2542 u32 val; 2543 2544 /* 2545 * The metric is reported as an 8bit integer fraction 2546 * summing up to 0xff. 2547 * slots-in-metric = (Metric / 0xff) * slots 2548 */ 2549 val = (metric >> ((idx - INTEL_PMC_IDX_METRIC_BASE) * 8)) & 0xff; 2550 return mul_u64_u32_div(slots, val, 0xff); 2551 } 2552 2553 static u64 icl_get_topdown_value(struct perf_event *event, 2554 u64 slots, u64 metrics) 2555 { 2556 int idx = event->hw.idx; 2557 u64 delta; 2558 2559 if (is_metric_idx(idx)) 2560 delta = icl_get_metrics_event_value(metrics, slots, idx); 2561 else 2562 delta = slots; 2563 2564 return delta; 2565 } 2566 2567 static void __icl_update_topdown_event(struct perf_event *event, 2568 u64 slots, u64 metrics, 2569 u64 last_slots, u64 last_metrics) 2570 { 2571 u64 delta, last = 0; 2572 2573 delta = icl_get_topdown_value(event, slots, metrics); 2574 if (last_slots) 2575 last = icl_get_topdown_value(event, last_slots, last_metrics); 2576 2577 /* 2578 * The 8bit integer fraction of metric may be not accurate, 2579 * especially when the changes is very small. 2580 * For example, if only a few bad_spec happens, the fraction 2581 * may be reduced from 1 to 0. If so, the bad_spec event value 2582 * will be 0 which is definitely less than the last value. 2583 * Avoid update event->count for this case. 2584 */ 2585 if (delta > last) { 2586 delta -= last; 2587 local64_add(delta, &event->count); 2588 } 2589 } 2590 2591 static void update_saved_topdown_regs(struct perf_event *event, u64 slots, 2592 u64 metrics, int metric_end) 2593 { 2594 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2595 struct perf_event *other; 2596 int idx; 2597 2598 event->hw.saved_slots = slots; 2599 event->hw.saved_metric = metrics; 2600 2601 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2602 if (!is_topdown_idx(idx)) 2603 continue; 2604 other = cpuc->events[idx]; 2605 other->hw.saved_slots = slots; 2606 other->hw.saved_metric = metrics; 2607 } 2608 } 2609 2610 /* 2611 * Update all active Topdown events. 2612 * 2613 * The PERF_METRICS and Fixed counter 3 are read separately. The values may be 2614 * modify by a NMI. PMU has to be disabled before calling this function. 2615 */ 2616 2617 static u64 intel_update_topdown_event(struct perf_event *event, int metric_end) 2618 { 2619 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2620 struct perf_event *other; 2621 u64 slots, metrics; 2622 bool reset = true; 2623 int idx; 2624 2625 /* read Fixed counter 3 */ 2626 rdpmcl((3 | INTEL_PMC_FIXED_RDPMC_BASE), slots); 2627 if (!slots) 2628 return 0; 2629 2630 /* read PERF_METRICS */ 2631 rdpmcl(INTEL_PMC_FIXED_RDPMC_METRICS, metrics); 2632 2633 for_each_set_bit(idx, cpuc->active_mask, metric_end + 1) { 2634 if (!is_topdown_idx(idx)) 2635 continue; 2636 other = cpuc->events[idx]; 2637 __icl_update_topdown_event(other, slots, metrics, 2638 event ? event->hw.saved_slots : 0, 2639 event ? event->hw.saved_metric : 0); 2640 } 2641 2642 /* 2643 * Check and update this event, which may have been cleared 2644 * in active_mask e.g. x86_pmu_stop() 2645 */ 2646 if (event && !test_bit(event->hw.idx, cpuc->active_mask)) { 2647 __icl_update_topdown_event(event, slots, metrics, 2648 event->hw.saved_slots, 2649 event->hw.saved_metric); 2650 2651 /* 2652 * In x86_pmu_stop(), the event is cleared in active_mask first, 2653 * then drain the delta, which indicates context switch for 2654 * counting. 2655 * Save metric and slots for context switch. 2656 * Don't need to reset the PERF_METRICS and Fixed counter 3. 2657 * Because the values will be restored in next schedule in. 2658 */ 2659 update_saved_topdown_regs(event, slots, metrics, metric_end); 2660 reset = false; 2661 } 2662 2663 if (reset) { 2664 /* The fixed counter 3 has to be written before the PERF_METRICS. */ 2665 wrmsrl(MSR_CORE_PERF_FIXED_CTR3, 0); 2666 wrmsrl(MSR_PERF_METRICS, 0); 2667 if (event) 2668 update_saved_topdown_regs(event, 0, 0, metric_end); 2669 } 2670 2671 return slots; 2672 } 2673 2674 static u64 icl_update_topdown_event(struct perf_event *event) 2675 { 2676 return intel_update_topdown_event(event, INTEL_PMC_IDX_METRIC_BASE + 2677 x86_pmu.num_topdown_events - 1); 2678 } 2679 2680 static u64 adl_update_topdown_event(struct perf_event *event) 2681 { 2682 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 2683 2684 if (pmu->cpu_type != hybrid_big) 2685 return 0; 2686 2687 return icl_update_topdown_event(event); 2688 } 2689 2690 DEFINE_STATIC_CALL(intel_pmu_update_topdown_event, x86_perf_event_update); 2691 2692 static void intel_pmu_read_topdown_event(struct perf_event *event) 2693 { 2694 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2695 2696 /* Only need to call update_topdown_event() once for group read. */ 2697 if ((cpuc->txn_flags & PERF_PMU_TXN_READ) && 2698 !is_slots_event(event)) 2699 return; 2700 2701 perf_pmu_disable(event->pmu); 2702 static_call(intel_pmu_update_topdown_event)(event); 2703 perf_pmu_enable(event->pmu); 2704 } 2705 2706 static void intel_pmu_read_event(struct perf_event *event) 2707 { 2708 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) 2709 intel_pmu_auto_reload_read(event); 2710 else if (is_topdown_count(event)) 2711 intel_pmu_read_topdown_event(event); 2712 else 2713 x86_perf_event_update(event); 2714 } 2715 2716 static void intel_pmu_enable_fixed(struct perf_event *event) 2717 { 2718 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2719 struct hw_perf_event *hwc = &event->hw; 2720 u64 mask, bits = 0; 2721 int idx = hwc->idx; 2722 2723 if (is_topdown_idx(idx)) { 2724 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2725 /* 2726 * When there are other active TopDown events, 2727 * don't enable the fixed counter 3 again. 2728 */ 2729 if (*(u64 *)cpuc->active_mask & INTEL_PMC_OTHER_TOPDOWN_BITS(idx)) 2730 return; 2731 2732 idx = INTEL_PMC_IDX_FIXED_SLOTS; 2733 } 2734 2735 intel_set_masks(event, idx); 2736 2737 /* 2738 * Enable IRQ generation (0x8), if not PEBS, 2739 * and enable ring-3 counting (0x2) and ring-0 counting (0x1) 2740 * if requested: 2741 */ 2742 if (!event->attr.precise_ip) 2743 bits |= 0x8; 2744 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) 2745 bits |= 0x2; 2746 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) 2747 bits |= 0x1; 2748 2749 /* 2750 * ANY bit is supported in v3 and up 2751 */ 2752 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) 2753 bits |= 0x4; 2754 2755 idx -= INTEL_PMC_IDX_FIXED; 2756 bits <<= (idx * 4); 2757 mask = 0xfULL << (idx * 4); 2758 2759 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { 2760 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2761 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2762 } 2763 2764 cpuc->fixed_ctrl_val &= ~mask; 2765 cpuc->fixed_ctrl_val |= bits; 2766 } 2767 2768 static void intel_pmu_enable_event(struct perf_event *event) 2769 { 2770 struct hw_perf_event *hwc = &event->hw; 2771 int idx = hwc->idx; 2772 2773 if (unlikely(event->attr.precise_ip)) 2774 intel_pmu_pebs_enable(event); 2775 2776 switch (idx) { 2777 case 0 ... INTEL_PMC_IDX_FIXED - 1: 2778 intel_set_masks(event, idx); 2779 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 2780 break; 2781 case INTEL_PMC_IDX_FIXED ... INTEL_PMC_IDX_FIXED_BTS - 1: 2782 case INTEL_PMC_IDX_METRIC_BASE ... INTEL_PMC_IDX_METRIC_END: 2783 intel_pmu_enable_fixed(event); 2784 break; 2785 case INTEL_PMC_IDX_FIXED_BTS: 2786 if (!__this_cpu_read(cpu_hw_events.enabled)) 2787 return; 2788 intel_pmu_enable_bts(hwc->config); 2789 break; 2790 case INTEL_PMC_IDX_FIXED_VLBR: 2791 intel_set_masks(event, idx); 2792 break; 2793 default: 2794 pr_warn("Failed to enable the event with invalid index %d\n", 2795 idx); 2796 } 2797 } 2798 2799 static void intel_pmu_add_event(struct perf_event *event) 2800 { 2801 if (event->attr.precise_ip) 2802 intel_pmu_pebs_add(event); 2803 if (needs_branch_stack(event)) 2804 intel_pmu_lbr_add(event); 2805 } 2806 2807 /* 2808 * Save and restart an expired event. Called by NMI contexts, 2809 * so it has to be careful about preempting normal event ops: 2810 */ 2811 int intel_pmu_save_and_restart(struct perf_event *event) 2812 { 2813 static_call(x86_pmu_update)(event); 2814 /* 2815 * For a checkpointed counter always reset back to 0. This 2816 * avoids a situation where the counter overflows, aborts the 2817 * transaction and is then set back to shortly before the 2818 * overflow, and overflows and aborts again. 2819 */ 2820 if (unlikely(event_is_checkpointed(event))) { 2821 /* No race with NMIs because the counter should not be armed */ 2822 wrmsrl(event->hw.event_base, 0); 2823 local64_set(&event->hw.prev_count, 0); 2824 } 2825 return static_call(x86_pmu_set_period)(event); 2826 } 2827 2828 static int intel_pmu_set_period(struct perf_event *event) 2829 { 2830 if (unlikely(is_topdown_count(event))) 2831 return static_call(intel_pmu_set_topdown_event_period)(event); 2832 2833 return x86_perf_event_set_period(event); 2834 } 2835 2836 static u64 intel_pmu_update(struct perf_event *event) 2837 { 2838 if (unlikely(is_topdown_count(event))) 2839 return static_call(intel_pmu_update_topdown_event)(event); 2840 2841 return x86_perf_event_update(event); 2842 } 2843 2844 static void intel_pmu_reset(void) 2845 { 2846 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); 2847 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2848 int num_counters_fixed = hybrid(cpuc->pmu, num_counters_fixed); 2849 int num_counters = hybrid(cpuc->pmu, num_counters); 2850 unsigned long flags; 2851 int idx; 2852 2853 if (!num_counters) 2854 return; 2855 2856 local_irq_save(flags); 2857 2858 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); 2859 2860 for (idx = 0; idx < num_counters; idx++) { 2861 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); 2862 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); 2863 } 2864 for (idx = 0; idx < num_counters_fixed; idx++) { 2865 if (fixed_counter_disabled(idx, cpuc->pmu)) 2866 continue; 2867 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); 2868 } 2869 2870 if (ds) 2871 ds->bts_index = ds->bts_buffer_base; 2872 2873 /* Ack all overflows and disable fixed counters */ 2874 if (x86_pmu.version >= 2) { 2875 intel_pmu_ack_status(intel_pmu_get_status()); 2876 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2877 } 2878 2879 /* Reset LBRs and LBR freezing */ 2880 if (x86_pmu.lbr_nr) { 2881 update_debugctlmsr(get_debugctlmsr() & 2882 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); 2883 } 2884 2885 local_irq_restore(flags); 2886 } 2887 2888 /* 2889 * We may be running with guest PEBS events created by KVM, and the 2890 * PEBS records are logged into the guest's DS and invisible to host. 2891 * 2892 * In the case of guest PEBS overflow, we only trigger a fake event 2893 * to emulate the PEBS overflow PMI for guest PEBS counters in KVM. 2894 * The guest will then vm-entry and check the guest DS area to read 2895 * the guest PEBS records. 2896 * 2897 * The contents and other behavior of the guest event do not matter. 2898 */ 2899 static void x86_pmu_handle_guest_pebs(struct pt_regs *regs, 2900 struct perf_sample_data *data) 2901 { 2902 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2903 u64 guest_pebs_idxs = cpuc->pebs_enabled & ~cpuc->intel_ctrl_host_mask; 2904 struct perf_event *event = NULL; 2905 int bit; 2906 2907 if (!unlikely(perf_guest_state())) 2908 return; 2909 2910 if (!x86_pmu.pebs_ept || !x86_pmu.pebs_active || 2911 !guest_pebs_idxs) 2912 return; 2913 2914 for_each_set_bit(bit, (unsigned long *)&guest_pebs_idxs, 2915 INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed) { 2916 event = cpuc->events[bit]; 2917 if (!event->attr.precise_ip) 2918 continue; 2919 2920 perf_sample_data_init(data, 0, event->hw.last_period); 2921 if (perf_event_overflow(event, data, regs)) 2922 x86_pmu_stop(event, 0); 2923 2924 /* Inject one fake event is enough. */ 2925 break; 2926 } 2927 } 2928 2929 static int handle_pmi_common(struct pt_regs *regs, u64 status) 2930 { 2931 struct perf_sample_data data; 2932 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2933 int bit; 2934 int handled = 0; 2935 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 2936 2937 inc_irq_stat(apic_perf_irqs); 2938 2939 /* 2940 * Ignore a range of extra bits in status that do not indicate 2941 * overflow by themselves. 2942 */ 2943 status &= ~(GLOBAL_STATUS_COND_CHG | 2944 GLOBAL_STATUS_ASIF | 2945 GLOBAL_STATUS_LBRS_FROZEN); 2946 if (!status) 2947 return 0; 2948 /* 2949 * In case multiple PEBS events are sampled at the same time, 2950 * it is possible to have GLOBAL_STATUS bit 62 set indicating 2951 * PEBS buffer overflow and also seeing at most 3 PEBS counters 2952 * having their bits set in the status register. This is a sign 2953 * that there was at least one PEBS record pending at the time 2954 * of the PMU interrupt. PEBS counters must only be processed 2955 * via the drain_pebs() calls and not via the regular sample 2956 * processing loop coming after that the function, otherwise 2957 * phony regular samples may be generated in the sampling buffer 2958 * not marked with the EXACT tag. Another possibility is to have 2959 * one PEBS event and at least one non-PEBS event which overflows 2960 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will 2961 * not be set, yet the overflow status bit for the PEBS counter will 2962 * be on Skylake. 2963 * 2964 * To avoid this problem, we systematically ignore the PEBS-enabled 2965 * counters from the GLOBAL_STATUS mask and we always process PEBS 2966 * events via drain_pebs(). 2967 */ 2968 status &= ~(cpuc->pebs_enabled & x86_pmu.pebs_capable); 2969 2970 /* 2971 * PEBS overflow sets bit 62 in the global status register 2972 */ 2973 if (__test_and_clear_bit(GLOBAL_STATUS_BUFFER_OVF_BIT, (unsigned long *)&status)) { 2974 u64 pebs_enabled = cpuc->pebs_enabled; 2975 2976 handled++; 2977 x86_pmu_handle_guest_pebs(regs, &data); 2978 x86_pmu.drain_pebs(regs, &data); 2979 status &= intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI; 2980 2981 /* 2982 * PMI throttle may be triggered, which stops the PEBS event. 2983 * Although cpuc->pebs_enabled is updated accordingly, the 2984 * MSR_IA32_PEBS_ENABLE is not updated. Because the 2985 * cpuc->enabled has been forced to 0 in PMI. 2986 * Update the MSR if pebs_enabled is changed. 2987 */ 2988 if (pebs_enabled != cpuc->pebs_enabled) 2989 wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled); 2990 } 2991 2992 /* 2993 * Intel PT 2994 */ 2995 if (__test_and_clear_bit(GLOBAL_STATUS_TRACE_TOPAPMI_BIT, (unsigned long *)&status)) { 2996 handled++; 2997 if (!perf_guest_handle_intel_pt_intr()) 2998 intel_pt_interrupt(); 2999 } 3000 3001 /* 3002 * Intel Perf metrics 3003 */ 3004 if (__test_and_clear_bit(GLOBAL_STATUS_PERF_METRICS_OVF_BIT, (unsigned long *)&status)) { 3005 handled++; 3006 static_call(intel_pmu_update_topdown_event)(NULL); 3007 } 3008 3009 /* 3010 * Checkpointed counters can lead to 'spurious' PMIs because the 3011 * rollback caused by the PMI will have cleared the overflow status 3012 * bit. Therefore always force probe these counters. 3013 */ 3014 status |= cpuc->intel_cp_status; 3015 3016 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { 3017 struct perf_event *event = cpuc->events[bit]; 3018 3019 handled++; 3020 3021 if (!test_bit(bit, cpuc->active_mask)) 3022 continue; 3023 3024 if (!intel_pmu_save_and_restart(event)) 3025 continue; 3026 3027 perf_sample_data_init(&data, 0, event->hw.last_period); 3028 3029 if (has_branch_stack(event)) { 3030 data.br_stack = &cpuc->lbr_stack; 3031 data.sample_flags |= PERF_SAMPLE_BRANCH_STACK; 3032 } 3033 3034 if (perf_event_overflow(event, &data, regs)) 3035 x86_pmu_stop(event, 0); 3036 } 3037 3038 return handled; 3039 } 3040 3041 /* 3042 * This handler is triggered by the local APIC, so the APIC IRQ handling 3043 * rules apply: 3044 */ 3045 static int intel_pmu_handle_irq(struct pt_regs *regs) 3046 { 3047 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3048 bool late_ack = hybrid_bit(cpuc->pmu, late_ack); 3049 bool mid_ack = hybrid_bit(cpuc->pmu, mid_ack); 3050 int loops; 3051 u64 status; 3052 int handled; 3053 int pmu_enabled; 3054 3055 /* 3056 * Save the PMU state. 3057 * It needs to be restored when leaving the handler. 3058 */ 3059 pmu_enabled = cpuc->enabled; 3060 /* 3061 * In general, the early ACK is only applied for old platforms. 3062 * For the big core starts from Haswell, the late ACK should be 3063 * applied. 3064 * For the small core after Tremont, we have to do the ACK right 3065 * before re-enabling counters, which is in the middle of the 3066 * NMI handler. 3067 */ 3068 if (!late_ack && !mid_ack) 3069 apic_write(APIC_LVTPC, APIC_DM_NMI); 3070 intel_bts_disable_local(); 3071 cpuc->enabled = 0; 3072 __intel_pmu_disable_all(true); 3073 handled = intel_pmu_drain_bts_buffer(); 3074 handled += intel_bts_interrupt(); 3075 status = intel_pmu_get_status(); 3076 if (!status) 3077 goto done; 3078 3079 loops = 0; 3080 again: 3081 intel_pmu_lbr_read(); 3082 intel_pmu_ack_status(status); 3083 if (++loops > 100) { 3084 static bool warned; 3085 3086 if (!warned) { 3087 WARN(1, "perfevents: irq loop stuck!\n"); 3088 perf_event_print_debug(); 3089 warned = true; 3090 } 3091 intel_pmu_reset(); 3092 goto done; 3093 } 3094 3095 handled += handle_pmi_common(regs, status); 3096 3097 /* 3098 * Repeat if there is more work to be done: 3099 */ 3100 status = intel_pmu_get_status(); 3101 if (status) 3102 goto again; 3103 3104 done: 3105 if (mid_ack) 3106 apic_write(APIC_LVTPC, APIC_DM_NMI); 3107 /* Only restore PMU state when it's active. See x86_pmu_disable(). */ 3108 cpuc->enabled = pmu_enabled; 3109 if (pmu_enabled) 3110 __intel_pmu_enable_all(0, true); 3111 intel_bts_enable_local(); 3112 3113 /* 3114 * Only unmask the NMI after the overflow counters 3115 * have been reset. This avoids spurious NMIs on 3116 * Haswell CPUs. 3117 */ 3118 if (late_ack) 3119 apic_write(APIC_LVTPC, APIC_DM_NMI); 3120 return handled; 3121 } 3122 3123 static struct event_constraint * 3124 intel_bts_constraints(struct perf_event *event) 3125 { 3126 if (unlikely(intel_pmu_has_bts(event))) 3127 return &bts_constraint; 3128 3129 return NULL; 3130 } 3131 3132 /* 3133 * Note: matches a fake event, like Fixed2. 3134 */ 3135 static struct event_constraint * 3136 intel_vlbr_constraints(struct perf_event *event) 3137 { 3138 struct event_constraint *c = &vlbr_constraint; 3139 3140 if (unlikely(constraint_match(c, event->hw.config))) { 3141 event->hw.flags |= c->flags; 3142 return c; 3143 } 3144 3145 return NULL; 3146 } 3147 3148 static int intel_alt_er(struct cpu_hw_events *cpuc, 3149 int idx, u64 config) 3150 { 3151 struct extra_reg *extra_regs = hybrid(cpuc->pmu, extra_regs); 3152 int alt_idx = idx; 3153 3154 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) 3155 return idx; 3156 3157 if (idx == EXTRA_REG_RSP_0) 3158 alt_idx = EXTRA_REG_RSP_1; 3159 3160 if (idx == EXTRA_REG_RSP_1) 3161 alt_idx = EXTRA_REG_RSP_0; 3162 3163 if (config & ~extra_regs[alt_idx].valid_mask) 3164 return idx; 3165 3166 return alt_idx; 3167 } 3168 3169 static void intel_fixup_er(struct perf_event *event, int idx) 3170 { 3171 struct extra_reg *extra_regs = hybrid(event->pmu, extra_regs); 3172 event->hw.extra_reg.idx = idx; 3173 3174 if (idx == EXTRA_REG_RSP_0) { 3175 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3176 event->hw.config |= extra_regs[EXTRA_REG_RSP_0].event; 3177 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; 3178 } else if (idx == EXTRA_REG_RSP_1) { 3179 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 3180 event->hw.config |= extra_regs[EXTRA_REG_RSP_1].event; 3181 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; 3182 } 3183 } 3184 3185 /* 3186 * manage allocation of shared extra msr for certain events 3187 * 3188 * sharing can be: 3189 * per-cpu: to be shared between the various events on a single PMU 3190 * per-core: per-cpu + shared by HT threads 3191 */ 3192 static struct event_constraint * 3193 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, 3194 struct perf_event *event, 3195 struct hw_perf_event_extra *reg) 3196 { 3197 struct event_constraint *c = &emptyconstraint; 3198 struct er_account *era; 3199 unsigned long flags; 3200 int idx = reg->idx; 3201 3202 /* 3203 * reg->alloc can be set due to existing state, so for fake cpuc we 3204 * need to ignore this, otherwise we might fail to allocate proper fake 3205 * state for this extra reg constraint. Also see the comment below. 3206 */ 3207 if (reg->alloc && !cpuc->is_fake) 3208 return NULL; /* call x86_get_event_constraint() */ 3209 3210 again: 3211 era = &cpuc->shared_regs->regs[idx]; 3212 /* 3213 * we use spin_lock_irqsave() to avoid lockdep issues when 3214 * passing a fake cpuc 3215 */ 3216 raw_spin_lock_irqsave(&era->lock, flags); 3217 3218 if (!atomic_read(&era->ref) || era->config == reg->config) { 3219 3220 /* 3221 * If its a fake cpuc -- as per validate_{group,event}() we 3222 * shouldn't touch event state and we can avoid doing so 3223 * since both will only call get_event_constraints() once 3224 * on each event, this avoids the need for reg->alloc. 3225 * 3226 * Not doing the ER fixup will only result in era->reg being 3227 * wrong, but since we won't actually try and program hardware 3228 * this isn't a problem either. 3229 */ 3230 if (!cpuc->is_fake) { 3231 if (idx != reg->idx) 3232 intel_fixup_er(event, idx); 3233 3234 /* 3235 * x86_schedule_events() can call get_event_constraints() 3236 * multiple times on events in the case of incremental 3237 * scheduling(). reg->alloc ensures we only do the ER 3238 * allocation once. 3239 */ 3240 reg->alloc = 1; 3241 } 3242 3243 /* lock in msr value */ 3244 era->config = reg->config; 3245 era->reg = reg->reg; 3246 3247 /* one more user */ 3248 atomic_inc(&era->ref); 3249 3250 /* 3251 * need to call x86_get_event_constraint() 3252 * to check if associated event has constraints 3253 */ 3254 c = NULL; 3255 } else { 3256 idx = intel_alt_er(cpuc, idx, reg->config); 3257 if (idx != reg->idx) { 3258 raw_spin_unlock_irqrestore(&era->lock, flags); 3259 goto again; 3260 } 3261 } 3262 raw_spin_unlock_irqrestore(&era->lock, flags); 3263 3264 return c; 3265 } 3266 3267 static void 3268 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, 3269 struct hw_perf_event_extra *reg) 3270 { 3271 struct er_account *era; 3272 3273 /* 3274 * Only put constraint if extra reg was actually allocated. Also takes 3275 * care of event which do not use an extra shared reg. 3276 * 3277 * Also, if this is a fake cpuc we shouldn't touch any event state 3278 * (reg->alloc) and we don't care about leaving inconsistent cpuc state 3279 * either since it'll be thrown out. 3280 */ 3281 if (!reg->alloc || cpuc->is_fake) 3282 return; 3283 3284 era = &cpuc->shared_regs->regs[reg->idx]; 3285 3286 /* one fewer user */ 3287 atomic_dec(&era->ref); 3288 3289 /* allocate again next time */ 3290 reg->alloc = 0; 3291 } 3292 3293 static struct event_constraint * 3294 intel_shared_regs_constraints(struct cpu_hw_events *cpuc, 3295 struct perf_event *event) 3296 { 3297 struct event_constraint *c = NULL, *d; 3298 struct hw_perf_event_extra *xreg, *breg; 3299 3300 xreg = &event->hw.extra_reg; 3301 if (xreg->idx != EXTRA_REG_NONE) { 3302 c = __intel_shared_reg_get_constraints(cpuc, event, xreg); 3303 if (c == &emptyconstraint) 3304 return c; 3305 } 3306 breg = &event->hw.branch_reg; 3307 if (breg->idx != EXTRA_REG_NONE) { 3308 d = __intel_shared_reg_get_constraints(cpuc, event, breg); 3309 if (d == &emptyconstraint) { 3310 __intel_shared_reg_put_constraints(cpuc, xreg); 3311 c = d; 3312 } 3313 } 3314 return c; 3315 } 3316 3317 struct event_constraint * 3318 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3319 struct perf_event *event) 3320 { 3321 struct event_constraint *event_constraints = hybrid(cpuc->pmu, event_constraints); 3322 struct event_constraint *c; 3323 3324 if (event_constraints) { 3325 for_each_event_constraint(c, event_constraints) { 3326 if (constraint_match(c, event->hw.config)) { 3327 event->hw.flags |= c->flags; 3328 return c; 3329 } 3330 } 3331 } 3332 3333 return &hybrid_var(cpuc->pmu, unconstrained); 3334 } 3335 3336 static struct event_constraint * 3337 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3338 struct perf_event *event) 3339 { 3340 struct event_constraint *c; 3341 3342 c = intel_vlbr_constraints(event); 3343 if (c) 3344 return c; 3345 3346 c = intel_bts_constraints(event); 3347 if (c) 3348 return c; 3349 3350 c = intel_shared_regs_constraints(cpuc, event); 3351 if (c) 3352 return c; 3353 3354 c = intel_pebs_constraints(event); 3355 if (c) 3356 return c; 3357 3358 return x86_get_event_constraints(cpuc, idx, event); 3359 } 3360 3361 static void 3362 intel_start_scheduling(struct cpu_hw_events *cpuc) 3363 { 3364 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3365 struct intel_excl_states *xl; 3366 int tid = cpuc->excl_thread_id; 3367 3368 /* 3369 * nothing needed if in group validation mode 3370 */ 3371 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3372 return; 3373 3374 /* 3375 * no exclusion needed 3376 */ 3377 if (WARN_ON_ONCE(!excl_cntrs)) 3378 return; 3379 3380 xl = &excl_cntrs->states[tid]; 3381 3382 xl->sched_started = true; 3383 /* 3384 * lock shared state until we are done scheduling 3385 * in stop_event_scheduling() 3386 * makes scheduling appear as a transaction 3387 */ 3388 raw_spin_lock(&excl_cntrs->lock); 3389 } 3390 3391 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 3392 { 3393 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3394 struct event_constraint *c = cpuc->event_constraint[idx]; 3395 struct intel_excl_states *xl; 3396 int tid = cpuc->excl_thread_id; 3397 3398 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3399 return; 3400 3401 if (WARN_ON_ONCE(!excl_cntrs)) 3402 return; 3403 3404 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) 3405 return; 3406 3407 xl = &excl_cntrs->states[tid]; 3408 3409 lockdep_assert_held(&excl_cntrs->lock); 3410 3411 if (c->flags & PERF_X86_EVENT_EXCL) 3412 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; 3413 else 3414 xl->state[cntr] = INTEL_EXCL_SHARED; 3415 } 3416 3417 static void 3418 intel_stop_scheduling(struct cpu_hw_events *cpuc) 3419 { 3420 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3421 struct intel_excl_states *xl; 3422 int tid = cpuc->excl_thread_id; 3423 3424 /* 3425 * nothing needed if in group validation mode 3426 */ 3427 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3428 return; 3429 /* 3430 * no exclusion needed 3431 */ 3432 if (WARN_ON_ONCE(!excl_cntrs)) 3433 return; 3434 3435 xl = &excl_cntrs->states[tid]; 3436 3437 xl->sched_started = false; 3438 /* 3439 * release shared state lock (acquired in intel_start_scheduling()) 3440 */ 3441 raw_spin_unlock(&excl_cntrs->lock); 3442 } 3443 3444 static struct event_constraint * 3445 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx) 3446 { 3447 WARN_ON_ONCE(!cpuc->constraint_list); 3448 3449 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { 3450 struct event_constraint *cx; 3451 3452 /* 3453 * grab pre-allocated constraint entry 3454 */ 3455 cx = &cpuc->constraint_list[idx]; 3456 3457 /* 3458 * initialize dynamic constraint 3459 * with static constraint 3460 */ 3461 *cx = *c; 3462 3463 /* 3464 * mark constraint as dynamic 3465 */ 3466 cx->flags |= PERF_X86_EVENT_DYNAMIC; 3467 c = cx; 3468 } 3469 3470 return c; 3471 } 3472 3473 static struct event_constraint * 3474 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, 3475 int idx, struct event_constraint *c) 3476 { 3477 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3478 struct intel_excl_states *xlo; 3479 int tid = cpuc->excl_thread_id; 3480 int is_excl, i, w; 3481 3482 /* 3483 * validating a group does not require 3484 * enforcing cross-thread exclusion 3485 */ 3486 if (cpuc->is_fake || !is_ht_workaround_enabled()) 3487 return c; 3488 3489 /* 3490 * no exclusion needed 3491 */ 3492 if (WARN_ON_ONCE(!excl_cntrs)) 3493 return c; 3494 3495 /* 3496 * because we modify the constraint, we need 3497 * to make a copy. Static constraints come 3498 * from static const tables. 3499 * 3500 * only needed when constraint has not yet 3501 * been cloned (marked dynamic) 3502 */ 3503 c = dyn_constraint(cpuc, c, idx); 3504 3505 /* 3506 * From here on, the constraint is dynamic. 3507 * Either it was just allocated above, or it 3508 * was allocated during a earlier invocation 3509 * of this function 3510 */ 3511 3512 /* 3513 * state of sibling HT 3514 */ 3515 xlo = &excl_cntrs->states[tid ^ 1]; 3516 3517 /* 3518 * event requires exclusive counter access 3519 * across HT threads 3520 */ 3521 is_excl = c->flags & PERF_X86_EVENT_EXCL; 3522 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { 3523 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; 3524 if (!cpuc->n_excl++) 3525 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); 3526 } 3527 3528 /* 3529 * Modify static constraint with current dynamic 3530 * state of thread 3531 * 3532 * EXCLUSIVE: sibling counter measuring exclusive event 3533 * SHARED : sibling counter measuring non-exclusive event 3534 * UNUSED : sibling counter unused 3535 */ 3536 w = c->weight; 3537 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { 3538 /* 3539 * exclusive event in sibling counter 3540 * our corresponding counter cannot be used 3541 * regardless of our event 3542 */ 3543 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) { 3544 __clear_bit(i, c->idxmsk); 3545 w--; 3546 continue; 3547 } 3548 /* 3549 * if measuring an exclusive event, sibling 3550 * measuring non-exclusive, then counter cannot 3551 * be used 3552 */ 3553 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) { 3554 __clear_bit(i, c->idxmsk); 3555 w--; 3556 continue; 3557 } 3558 } 3559 3560 /* 3561 * if we return an empty mask, then switch 3562 * back to static empty constraint to avoid 3563 * the cost of freeing later on 3564 */ 3565 if (!w) 3566 c = &emptyconstraint; 3567 3568 c->weight = w; 3569 3570 return c; 3571 } 3572 3573 static struct event_constraint * 3574 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3575 struct perf_event *event) 3576 { 3577 struct event_constraint *c1, *c2; 3578 3579 c1 = cpuc->event_constraint[idx]; 3580 3581 /* 3582 * first time only 3583 * - static constraint: no change across incremental scheduling calls 3584 * - dynamic constraint: handled by intel_get_excl_constraints() 3585 */ 3586 c2 = __intel_get_event_constraints(cpuc, idx, event); 3587 if (c1) { 3588 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC)); 3589 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); 3590 c1->weight = c2->weight; 3591 c2 = c1; 3592 } 3593 3594 if (cpuc->excl_cntrs) 3595 return intel_get_excl_constraints(cpuc, event, idx, c2); 3596 3597 return c2; 3598 } 3599 3600 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, 3601 struct perf_event *event) 3602 { 3603 struct hw_perf_event *hwc = &event->hw; 3604 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3605 int tid = cpuc->excl_thread_id; 3606 struct intel_excl_states *xl; 3607 3608 /* 3609 * nothing needed if in group validation mode 3610 */ 3611 if (cpuc->is_fake) 3612 return; 3613 3614 if (WARN_ON_ONCE(!excl_cntrs)) 3615 return; 3616 3617 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { 3618 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; 3619 if (!--cpuc->n_excl) 3620 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); 3621 } 3622 3623 /* 3624 * If event was actually assigned, then mark the counter state as 3625 * unused now. 3626 */ 3627 if (hwc->idx >= 0) { 3628 xl = &excl_cntrs->states[tid]; 3629 3630 /* 3631 * put_constraint may be called from x86_schedule_events() 3632 * which already has the lock held so here make locking 3633 * conditional. 3634 */ 3635 if (!xl->sched_started) 3636 raw_spin_lock(&excl_cntrs->lock); 3637 3638 xl->state[hwc->idx] = INTEL_EXCL_UNUSED; 3639 3640 if (!xl->sched_started) 3641 raw_spin_unlock(&excl_cntrs->lock); 3642 } 3643 } 3644 3645 static void 3646 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, 3647 struct perf_event *event) 3648 { 3649 struct hw_perf_event_extra *reg; 3650 3651 reg = &event->hw.extra_reg; 3652 if (reg->idx != EXTRA_REG_NONE) 3653 __intel_shared_reg_put_constraints(cpuc, reg); 3654 3655 reg = &event->hw.branch_reg; 3656 if (reg->idx != EXTRA_REG_NONE) 3657 __intel_shared_reg_put_constraints(cpuc, reg); 3658 } 3659 3660 static void intel_put_event_constraints(struct cpu_hw_events *cpuc, 3661 struct perf_event *event) 3662 { 3663 intel_put_shared_regs_event_constraints(cpuc, event); 3664 3665 /* 3666 * is PMU has exclusive counter restrictions, then 3667 * all events are subject to and must call the 3668 * put_excl_constraints() routine 3669 */ 3670 if (cpuc->excl_cntrs) 3671 intel_put_excl_constraints(cpuc, event); 3672 } 3673 3674 static void intel_pebs_aliases_core2(struct perf_event *event) 3675 { 3676 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3677 /* 3678 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3679 * (0x003c) so that we can use it with PEBS. 3680 * 3681 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3682 * PEBS capable. However we can use INST_RETIRED.ANY_P 3683 * (0x00c0), which is a PEBS capable event, to get the same 3684 * count. 3685 * 3686 * INST_RETIRED.ANY_P counts the number of cycles that retires 3687 * CNTMASK instructions. By setting CNTMASK to a value (16) 3688 * larger than the maximum number of instructions that can be 3689 * retired per cycle (4) and then inverting the condition, we 3690 * count all cycles that retire 16 or less instructions, which 3691 * is every cycle. 3692 * 3693 * Thereby we gain a PEBS capable cycle counter. 3694 */ 3695 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); 3696 3697 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3698 event->hw.config = alt_config; 3699 } 3700 } 3701 3702 static void intel_pebs_aliases_snb(struct perf_event *event) 3703 { 3704 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3705 /* 3706 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3707 * (0x003c) so that we can use it with PEBS. 3708 * 3709 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3710 * PEBS capable. However we can use UOPS_RETIRED.ALL 3711 * (0x01c2), which is a PEBS capable event, to get the same 3712 * count. 3713 * 3714 * UOPS_RETIRED.ALL counts the number of cycles that retires 3715 * CNTMASK micro-ops. By setting CNTMASK to a value (16) 3716 * larger than the maximum number of micro-ops that can be 3717 * retired per cycle (4) and then inverting the condition, we 3718 * count all cycles that retire 16 or less micro-ops, which 3719 * is every cycle. 3720 * 3721 * Thereby we gain a PEBS capable cycle counter. 3722 */ 3723 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); 3724 3725 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3726 event->hw.config = alt_config; 3727 } 3728 } 3729 3730 static void intel_pebs_aliases_precdist(struct perf_event *event) 3731 { 3732 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3733 /* 3734 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3735 * (0x003c) so that we can use it with PEBS. 3736 * 3737 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3738 * PEBS capable. However we can use INST_RETIRED.PREC_DIST 3739 * (0x01c0), which is a PEBS capable event, to get the same 3740 * count. 3741 * 3742 * The PREC_DIST event has special support to minimize sample 3743 * shadowing effects. One drawback is that it can be 3744 * only programmed on counter 1, but that seems like an 3745 * acceptable trade off. 3746 */ 3747 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16); 3748 3749 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3750 event->hw.config = alt_config; 3751 } 3752 } 3753 3754 static void intel_pebs_aliases_ivb(struct perf_event *event) 3755 { 3756 if (event->attr.precise_ip < 3) 3757 return intel_pebs_aliases_snb(event); 3758 return intel_pebs_aliases_precdist(event); 3759 } 3760 3761 static void intel_pebs_aliases_skl(struct perf_event *event) 3762 { 3763 if (event->attr.precise_ip < 3) 3764 return intel_pebs_aliases_core2(event); 3765 return intel_pebs_aliases_precdist(event); 3766 } 3767 3768 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event) 3769 { 3770 unsigned long flags = x86_pmu.large_pebs_flags; 3771 3772 if (event->attr.use_clockid) 3773 flags &= ~PERF_SAMPLE_TIME; 3774 if (!event->attr.exclude_kernel) 3775 flags &= ~PERF_SAMPLE_REGS_USER; 3776 if (event->attr.sample_regs_user & ~PEBS_GP_REGS) 3777 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR); 3778 return flags; 3779 } 3780 3781 static int intel_pmu_bts_config(struct perf_event *event) 3782 { 3783 struct perf_event_attr *attr = &event->attr; 3784 3785 if (unlikely(intel_pmu_has_bts(event))) { 3786 /* BTS is not supported by this architecture. */ 3787 if (!x86_pmu.bts_active) 3788 return -EOPNOTSUPP; 3789 3790 /* BTS is currently only allowed for user-mode. */ 3791 if (!attr->exclude_kernel) 3792 return -EOPNOTSUPP; 3793 3794 /* BTS is not allowed for precise events. */ 3795 if (attr->precise_ip) 3796 return -EOPNOTSUPP; 3797 3798 /* disallow bts if conflicting events are present */ 3799 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3800 return -EBUSY; 3801 3802 event->destroy = hw_perf_lbr_event_destroy; 3803 } 3804 3805 return 0; 3806 } 3807 3808 static int core_pmu_hw_config(struct perf_event *event) 3809 { 3810 int ret = x86_pmu_hw_config(event); 3811 3812 if (ret) 3813 return ret; 3814 3815 return intel_pmu_bts_config(event); 3816 } 3817 3818 #define INTEL_TD_METRIC_AVAILABLE_MAX (INTEL_TD_METRIC_RETIRING + \ 3819 ((x86_pmu.num_topdown_events - 1) << 8)) 3820 3821 static bool is_available_metric_event(struct perf_event *event) 3822 { 3823 return is_metric_event(event) && 3824 event->attr.config <= INTEL_TD_METRIC_AVAILABLE_MAX; 3825 } 3826 3827 static inline bool is_mem_loads_event(struct perf_event *event) 3828 { 3829 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0xcd, .umask=0x01); 3830 } 3831 3832 static inline bool is_mem_loads_aux_event(struct perf_event *event) 3833 { 3834 return (event->attr.config & INTEL_ARCH_EVENT_MASK) == X86_CONFIG(.event=0x03, .umask=0x82); 3835 } 3836 3837 static inline bool require_mem_loads_aux_event(struct perf_event *event) 3838 { 3839 if (!(x86_pmu.flags & PMU_FL_MEM_LOADS_AUX)) 3840 return false; 3841 3842 if (is_hybrid()) 3843 return hybrid_pmu(event->pmu)->cpu_type == hybrid_big; 3844 3845 return true; 3846 } 3847 3848 static inline bool intel_pmu_has_cap(struct perf_event *event, int idx) 3849 { 3850 union perf_capabilities *intel_cap = &hybrid(event->pmu, intel_cap); 3851 3852 return test_bit(idx, (unsigned long *)&intel_cap->capabilities); 3853 } 3854 3855 static int intel_pmu_hw_config(struct perf_event *event) 3856 { 3857 int ret = x86_pmu_hw_config(event); 3858 3859 if (ret) 3860 return ret; 3861 3862 ret = intel_pmu_bts_config(event); 3863 if (ret) 3864 return ret; 3865 3866 if (event->attr.precise_ip) { 3867 if ((event->attr.config & INTEL_ARCH_EVENT_MASK) == INTEL_FIXED_VLBR_EVENT) 3868 return -EINVAL; 3869 3870 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) { 3871 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; 3872 if (!(event->attr.sample_type & 3873 ~intel_pmu_large_pebs_flags(event))) { 3874 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS; 3875 event->attach_state |= PERF_ATTACH_SCHED_CB; 3876 } 3877 } 3878 if (x86_pmu.pebs_aliases) 3879 x86_pmu.pebs_aliases(event); 3880 } 3881 3882 if (needs_branch_stack(event)) { 3883 ret = intel_pmu_setup_lbr_filter(event); 3884 if (ret) 3885 return ret; 3886 event->attach_state |= PERF_ATTACH_SCHED_CB; 3887 3888 /* 3889 * BTS is set up earlier in this path, so don't account twice 3890 */ 3891 if (!unlikely(intel_pmu_has_bts(event))) { 3892 /* disallow lbr if conflicting events are present */ 3893 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3894 return -EBUSY; 3895 3896 event->destroy = hw_perf_lbr_event_destroy; 3897 } 3898 } 3899 3900 if (event->attr.aux_output) { 3901 if (!event->attr.precise_ip) 3902 return -EINVAL; 3903 3904 event->hw.flags |= PERF_X86_EVENT_PEBS_VIA_PT; 3905 } 3906 3907 if ((event->attr.type == PERF_TYPE_HARDWARE) || 3908 (event->attr.type == PERF_TYPE_HW_CACHE)) 3909 return 0; 3910 3911 /* 3912 * Config Topdown slots and metric events 3913 * 3914 * The slots event on Fixed Counter 3 can support sampling, 3915 * which will be handled normally in x86_perf_event_update(). 3916 * 3917 * Metric events don't support sampling and require being paired 3918 * with a slots event as group leader. When the slots event 3919 * is used in a metrics group, it too cannot support sampling. 3920 */ 3921 if (intel_pmu_has_cap(event, PERF_CAP_METRICS_IDX) && is_topdown_event(event)) { 3922 if (event->attr.config1 || event->attr.config2) 3923 return -EINVAL; 3924 3925 /* 3926 * The TopDown metrics events and slots event don't 3927 * support any filters. 3928 */ 3929 if (event->attr.config & X86_ALL_EVENT_FLAGS) 3930 return -EINVAL; 3931 3932 if (is_available_metric_event(event)) { 3933 struct perf_event *leader = event->group_leader; 3934 3935 /* The metric events don't support sampling. */ 3936 if (is_sampling_event(event)) 3937 return -EINVAL; 3938 3939 /* The metric events require a slots group leader. */ 3940 if (!is_slots_event(leader)) 3941 return -EINVAL; 3942 3943 /* 3944 * The leader/SLOTS must not be a sampling event for 3945 * metric use; hardware requires it starts at 0 when used 3946 * in conjunction with MSR_PERF_METRICS. 3947 */ 3948 if (is_sampling_event(leader)) 3949 return -EINVAL; 3950 3951 event->event_caps |= PERF_EV_CAP_SIBLING; 3952 /* 3953 * Only once we have a METRICs sibling do we 3954 * need TopDown magic. 3955 */ 3956 leader->hw.flags |= PERF_X86_EVENT_TOPDOWN; 3957 event->hw.flags |= PERF_X86_EVENT_TOPDOWN; 3958 } 3959 } 3960 3961 /* 3962 * The load latency event X86_CONFIG(.event=0xcd, .umask=0x01) on SPR 3963 * doesn't function quite right. As a work-around it needs to always be 3964 * co-scheduled with a auxiliary event X86_CONFIG(.event=0x03, .umask=0x82). 3965 * The actual count of this second event is irrelevant it just needs 3966 * to be active to make the first event function correctly. 3967 * 3968 * In a group, the auxiliary event must be in front of the load latency 3969 * event. The rule is to simplify the implementation of the check. 3970 * That's because perf cannot have a complete group at the moment. 3971 */ 3972 if (require_mem_loads_aux_event(event) && 3973 (event->attr.sample_type & PERF_SAMPLE_DATA_SRC) && 3974 is_mem_loads_event(event)) { 3975 struct perf_event *leader = event->group_leader; 3976 struct perf_event *sibling = NULL; 3977 3978 if (!is_mem_loads_aux_event(leader)) { 3979 for_each_sibling_event(sibling, leader) { 3980 if (is_mem_loads_aux_event(sibling)) 3981 break; 3982 } 3983 if (list_entry_is_head(sibling, &leader->sibling_list, sibling_list)) 3984 return -ENODATA; 3985 } 3986 } 3987 3988 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) 3989 return 0; 3990 3991 if (x86_pmu.version < 3) 3992 return -EINVAL; 3993 3994 ret = perf_allow_cpu(&event->attr); 3995 if (ret) 3996 return ret; 3997 3998 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; 3999 4000 return 0; 4001 } 4002 4003 /* 4004 * Currently, the only caller of this function is the atomic_switch_perf_msrs(). 4005 * The host perf conext helps to prepare the values of the real hardware for 4006 * a set of msrs that need to be switched atomically in a vmx transaction. 4007 * 4008 * For example, the pseudocode needed to add a new msr should look like: 4009 * 4010 * arr[(*nr)++] = (struct perf_guest_switch_msr){ 4011 * .msr = the hardware msr address, 4012 * .host = the value the hardware has when it doesn't run a guest, 4013 * .guest = the value the hardware has when it runs a guest, 4014 * }; 4015 * 4016 * These values have nothing to do with the emulated values the guest sees 4017 * when it uses {RD,WR}MSR, which should be handled by the KVM context, 4018 * specifically in the intel_pmu_{get,set}_msr(). 4019 */ 4020 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr, void *data) 4021 { 4022 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4023 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 4024 struct kvm_pmu *kvm_pmu = (struct kvm_pmu *)data; 4025 u64 intel_ctrl = hybrid(cpuc->pmu, intel_ctrl); 4026 u64 pebs_mask = cpuc->pebs_enabled & x86_pmu.pebs_capable; 4027 int global_ctrl, pebs_enable; 4028 4029 *nr = 0; 4030 global_ctrl = (*nr)++; 4031 arr[global_ctrl] = (struct perf_guest_switch_msr){ 4032 .msr = MSR_CORE_PERF_GLOBAL_CTRL, 4033 .host = intel_ctrl & ~cpuc->intel_ctrl_guest_mask, 4034 .guest = intel_ctrl & (~cpuc->intel_ctrl_host_mask | ~pebs_mask), 4035 }; 4036 4037 if (!x86_pmu.pebs) 4038 return arr; 4039 4040 /* 4041 * If PMU counter has PEBS enabled it is not enough to 4042 * disable counter on a guest entry since PEBS memory 4043 * write can overshoot guest entry and corrupt guest 4044 * memory. Disabling PEBS solves the problem. 4045 * 4046 * Don't do this if the CPU already enforces it. 4047 */ 4048 if (x86_pmu.pebs_no_isolation) { 4049 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4050 .msr = MSR_IA32_PEBS_ENABLE, 4051 .host = cpuc->pebs_enabled, 4052 .guest = 0, 4053 }; 4054 return arr; 4055 } 4056 4057 if (!kvm_pmu || !x86_pmu.pebs_ept) 4058 return arr; 4059 4060 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4061 .msr = MSR_IA32_DS_AREA, 4062 .host = (unsigned long)cpuc->ds, 4063 .guest = kvm_pmu->ds_area, 4064 }; 4065 4066 if (x86_pmu.intel_cap.pebs_baseline) { 4067 arr[(*nr)++] = (struct perf_guest_switch_msr){ 4068 .msr = MSR_PEBS_DATA_CFG, 4069 .host = cpuc->pebs_data_cfg, 4070 .guest = kvm_pmu->pebs_data_cfg, 4071 }; 4072 } 4073 4074 pebs_enable = (*nr)++; 4075 arr[pebs_enable] = (struct perf_guest_switch_msr){ 4076 .msr = MSR_IA32_PEBS_ENABLE, 4077 .host = cpuc->pebs_enabled & ~cpuc->intel_ctrl_guest_mask, 4078 .guest = pebs_mask & ~cpuc->intel_ctrl_host_mask, 4079 }; 4080 4081 if (arr[pebs_enable].host) { 4082 /* Disable guest PEBS if host PEBS is enabled. */ 4083 arr[pebs_enable].guest = 0; 4084 } else { 4085 /* Disable guest PEBS thoroughly for cross-mapped PEBS counters. */ 4086 arr[pebs_enable].guest &= ~kvm_pmu->host_cross_mapped_mask; 4087 arr[global_ctrl].guest &= ~kvm_pmu->host_cross_mapped_mask; 4088 /* Set hw GLOBAL_CTRL bits for PEBS counter when it runs for guest */ 4089 arr[global_ctrl].guest |= arr[pebs_enable].guest; 4090 } 4091 4092 return arr; 4093 } 4094 4095 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr, void *data) 4096 { 4097 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4098 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 4099 int idx; 4100 4101 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 4102 struct perf_event *event = cpuc->events[idx]; 4103 4104 arr[idx].msr = x86_pmu_config_addr(idx); 4105 arr[idx].host = arr[idx].guest = 0; 4106 4107 if (!test_bit(idx, cpuc->active_mask)) 4108 continue; 4109 4110 arr[idx].host = arr[idx].guest = 4111 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; 4112 4113 if (event->attr.exclude_host) 4114 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 4115 else if (event->attr.exclude_guest) 4116 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 4117 } 4118 4119 *nr = x86_pmu.num_counters; 4120 return arr; 4121 } 4122 4123 static void core_pmu_enable_event(struct perf_event *event) 4124 { 4125 if (!event->attr.exclude_host) 4126 x86_pmu_enable_event(event); 4127 } 4128 4129 static void core_pmu_enable_all(int added) 4130 { 4131 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4132 int idx; 4133 4134 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 4135 struct hw_perf_event *hwc = &cpuc->events[idx]->hw; 4136 4137 if (!test_bit(idx, cpuc->active_mask) || 4138 cpuc->events[idx]->attr.exclude_host) 4139 continue; 4140 4141 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 4142 } 4143 } 4144 4145 static int hsw_hw_config(struct perf_event *event) 4146 { 4147 int ret = intel_pmu_hw_config(event); 4148 4149 if (ret) 4150 return ret; 4151 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) 4152 return 0; 4153 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); 4154 4155 /* 4156 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with 4157 * PEBS or in ANY thread mode. Since the results are non-sensical forbid 4158 * this combination. 4159 */ 4160 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && 4161 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || 4162 event->attr.precise_ip > 0)) 4163 return -EOPNOTSUPP; 4164 4165 if (event_is_checkpointed(event)) { 4166 /* 4167 * Sampling of checkpointed events can cause situations where 4168 * the CPU constantly aborts because of a overflow, which is 4169 * then checkpointed back and ignored. Forbid checkpointing 4170 * for sampling. 4171 * 4172 * But still allow a long sampling period, so that perf stat 4173 * from KVM works. 4174 */ 4175 if (event->attr.sample_period > 0 && 4176 event->attr.sample_period < 0x7fffffff) 4177 return -EOPNOTSUPP; 4178 } 4179 return 0; 4180 } 4181 4182 static struct event_constraint counter0_constraint = 4183 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1); 4184 4185 static struct event_constraint counter2_constraint = 4186 EVENT_CONSTRAINT(0, 0x4, 0); 4187 4188 static struct event_constraint fixed0_constraint = 4189 FIXED_EVENT_CONSTRAINT(0x00c0, 0); 4190 4191 static struct event_constraint fixed0_counter0_constraint = 4192 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL); 4193 4194 static struct event_constraint * 4195 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4196 struct perf_event *event) 4197 { 4198 struct event_constraint *c; 4199 4200 c = intel_get_event_constraints(cpuc, idx, event); 4201 4202 /* Handle special quirk on in_tx_checkpointed only in counter 2 */ 4203 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { 4204 if (c->idxmsk64 & (1U << 2)) 4205 return &counter2_constraint; 4206 return &emptyconstraint; 4207 } 4208 4209 return c; 4210 } 4211 4212 static struct event_constraint * 4213 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4214 struct perf_event *event) 4215 { 4216 /* 4217 * Fixed counter 0 has less skid. 4218 * Force instruction:ppp in Fixed counter 0 4219 */ 4220 if ((event->attr.precise_ip == 3) && 4221 constraint_match(&fixed0_constraint, event->hw.config)) 4222 return &fixed0_constraint; 4223 4224 return hsw_get_event_constraints(cpuc, idx, event); 4225 } 4226 4227 static struct event_constraint * 4228 spr_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4229 struct perf_event *event) 4230 { 4231 struct event_constraint *c; 4232 4233 c = icl_get_event_constraints(cpuc, idx, event); 4234 4235 /* 4236 * The :ppp indicates the Precise Distribution (PDist) facility, which 4237 * is only supported on the GP counter 0. If a :ppp event which is not 4238 * available on the GP counter 0, error out. 4239 * Exception: Instruction PDIR is only available on the fixed counter 0. 4240 */ 4241 if ((event->attr.precise_ip == 3) && 4242 !constraint_match(&fixed0_constraint, event->hw.config)) { 4243 if (c->idxmsk64 & BIT_ULL(0)) 4244 return &counter0_constraint; 4245 4246 return &emptyconstraint; 4247 } 4248 4249 return c; 4250 } 4251 4252 static struct event_constraint * 4253 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4254 struct perf_event *event) 4255 { 4256 struct event_constraint *c; 4257 4258 /* :ppp means to do reduced skid PEBS which is PMC0 only. */ 4259 if (event->attr.precise_ip == 3) 4260 return &counter0_constraint; 4261 4262 c = intel_get_event_constraints(cpuc, idx, event); 4263 4264 return c; 4265 } 4266 4267 static struct event_constraint * 4268 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4269 struct perf_event *event) 4270 { 4271 struct event_constraint *c; 4272 4273 c = intel_get_event_constraints(cpuc, idx, event); 4274 4275 /* 4276 * :ppp means to do reduced skid PEBS, 4277 * which is available on PMC0 and fixed counter 0. 4278 */ 4279 if (event->attr.precise_ip == 3) { 4280 /* Force instruction:ppp on PMC0 and Fixed counter 0 */ 4281 if (constraint_match(&fixed0_constraint, event->hw.config)) 4282 return &fixed0_counter0_constraint; 4283 4284 return &counter0_constraint; 4285 } 4286 4287 return c; 4288 } 4289 4290 static bool allow_tsx_force_abort = true; 4291 4292 static struct event_constraint * 4293 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4294 struct perf_event *event) 4295 { 4296 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event); 4297 4298 /* 4299 * Without TFA we must not use PMC3. 4300 */ 4301 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) { 4302 c = dyn_constraint(cpuc, c, idx); 4303 c->idxmsk64 &= ~(1ULL << 3); 4304 c->weight--; 4305 } 4306 4307 return c; 4308 } 4309 4310 static struct event_constraint * 4311 adl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 4312 struct perf_event *event) 4313 { 4314 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4315 4316 if (pmu->cpu_type == hybrid_big) 4317 return spr_get_event_constraints(cpuc, idx, event); 4318 else if (pmu->cpu_type == hybrid_small) 4319 return tnt_get_event_constraints(cpuc, idx, event); 4320 4321 WARN_ON(1); 4322 return &emptyconstraint; 4323 } 4324 4325 static int adl_hw_config(struct perf_event *event) 4326 { 4327 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4328 4329 if (pmu->cpu_type == hybrid_big) 4330 return hsw_hw_config(event); 4331 else if (pmu->cpu_type == hybrid_small) 4332 return intel_pmu_hw_config(event); 4333 4334 WARN_ON(1); 4335 return -EOPNOTSUPP; 4336 } 4337 4338 static u8 adl_get_hybrid_cpu_type(void) 4339 { 4340 return hybrid_big; 4341 } 4342 4343 /* 4344 * Broadwell: 4345 * 4346 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared 4347 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine 4348 * the two to enforce a minimum period of 128 (the smallest value that has bits 4349 * 0-5 cleared and >= 100). 4350 * 4351 * Because of how the code in x86_perf_event_set_period() works, the truncation 4352 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period 4353 * to make up for the 'lost' events due to carrying the 'error' in period_left. 4354 * 4355 * Therefore the effective (average) period matches the requested period, 4356 * despite coarser hardware granularity. 4357 */ 4358 static void bdw_limit_period(struct perf_event *event, s64 *left) 4359 { 4360 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == 4361 X86_CONFIG(.event=0xc0, .umask=0x01)) { 4362 if (*left < 128) 4363 *left = 128; 4364 *left &= ~0x3fULL; 4365 } 4366 } 4367 4368 static void nhm_limit_period(struct perf_event *event, s64 *left) 4369 { 4370 *left = max(*left, 32LL); 4371 } 4372 4373 static void spr_limit_period(struct perf_event *event, s64 *left) 4374 { 4375 if (event->attr.precise_ip == 3) 4376 *left = max(*left, 128LL); 4377 } 4378 4379 PMU_FORMAT_ATTR(event, "config:0-7" ); 4380 PMU_FORMAT_ATTR(umask, "config:8-15" ); 4381 PMU_FORMAT_ATTR(edge, "config:18" ); 4382 PMU_FORMAT_ATTR(pc, "config:19" ); 4383 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ 4384 PMU_FORMAT_ATTR(inv, "config:23" ); 4385 PMU_FORMAT_ATTR(cmask, "config:24-31" ); 4386 PMU_FORMAT_ATTR(in_tx, "config:32"); 4387 PMU_FORMAT_ATTR(in_tx_cp, "config:33"); 4388 4389 static struct attribute *intel_arch_formats_attr[] = { 4390 &format_attr_event.attr, 4391 &format_attr_umask.attr, 4392 &format_attr_edge.attr, 4393 &format_attr_pc.attr, 4394 &format_attr_inv.attr, 4395 &format_attr_cmask.attr, 4396 NULL, 4397 }; 4398 4399 ssize_t intel_event_sysfs_show(char *page, u64 config) 4400 { 4401 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); 4402 4403 return x86_event_sysfs_show(page, config, event); 4404 } 4405 4406 static struct intel_shared_regs *allocate_shared_regs(int cpu) 4407 { 4408 struct intel_shared_regs *regs; 4409 int i; 4410 4411 regs = kzalloc_node(sizeof(struct intel_shared_regs), 4412 GFP_KERNEL, cpu_to_node(cpu)); 4413 if (regs) { 4414 /* 4415 * initialize the locks to keep lockdep happy 4416 */ 4417 for (i = 0; i < EXTRA_REG_MAX; i++) 4418 raw_spin_lock_init(®s->regs[i].lock); 4419 4420 regs->core_id = -1; 4421 } 4422 return regs; 4423 } 4424 4425 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) 4426 { 4427 struct intel_excl_cntrs *c; 4428 4429 c = kzalloc_node(sizeof(struct intel_excl_cntrs), 4430 GFP_KERNEL, cpu_to_node(cpu)); 4431 if (c) { 4432 raw_spin_lock_init(&c->lock); 4433 c->core_id = -1; 4434 } 4435 return c; 4436 } 4437 4438 4439 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 4440 { 4441 cpuc->pebs_record_size = x86_pmu.pebs_record_size; 4442 4443 if (is_hybrid() || x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { 4444 cpuc->shared_regs = allocate_shared_regs(cpu); 4445 if (!cpuc->shared_regs) 4446 goto err; 4447 } 4448 4449 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) { 4450 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); 4451 4452 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu)); 4453 if (!cpuc->constraint_list) 4454 goto err_shared_regs; 4455 } 4456 4457 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4458 cpuc->excl_cntrs = allocate_excl_cntrs(cpu); 4459 if (!cpuc->excl_cntrs) 4460 goto err_constraint_list; 4461 4462 cpuc->excl_thread_id = 0; 4463 } 4464 4465 return 0; 4466 4467 err_constraint_list: 4468 kfree(cpuc->constraint_list); 4469 cpuc->constraint_list = NULL; 4470 4471 err_shared_regs: 4472 kfree(cpuc->shared_regs); 4473 cpuc->shared_regs = NULL; 4474 4475 err: 4476 return -ENOMEM; 4477 } 4478 4479 static int intel_pmu_cpu_prepare(int cpu) 4480 { 4481 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu); 4482 } 4483 4484 static void flip_smm_bit(void *data) 4485 { 4486 unsigned long set = *(unsigned long *)data; 4487 4488 if (set > 0) { 4489 msr_set_bit(MSR_IA32_DEBUGCTLMSR, 4490 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4491 } else { 4492 msr_clear_bit(MSR_IA32_DEBUGCTLMSR, 4493 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 4494 } 4495 } 4496 4497 static bool init_hybrid_pmu(int cpu) 4498 { 4499 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4500 u8 cpu_type = get_this_hybrid_cpu_type(); 4501 struct x86_hybrid_pmu *pmu = NULL; 4502 int i; 4503 4504 if (!cpu_type && x86_pmu.get_hybrid_cpu_type) 4505 cpu_type = x86_pmu.get_hybrid_cpu_type(); 4506 4507 for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) { 4508 if (x86_pmu.hybrid_pmu[i].cpu_type == cpu_type) { 4509 pmu = &x86_pmu.hybrid_pmu[i]; 4510 break; 4511 } 4512 } 4513 if (WARN_ON_ONCE(!pmu || (pmu->pmu.type == -1))) { 4514 cpuc->pmu = NULL; 4515 return false; 4516 } 4517 4518 /* Only check and dump the PMU information for the first CPU */ 4519 if (!cpumask_empty(&pmu->supported_cpus)) 4520 goto end; 4521 4522 if (!check_hw_exists(&pmu->pmu, pmu->num_counters, pmu->num_counters_fixed)) 4523 return false; 4524 4525 pr_info("%s PMU driver: ", pmu->name); 4526 4527 if (pmu->intel_cap.pebs_output_pt_available) 4528 pr_cont("PEBS-via-PT "); 4529 4530 pr_cont("\n"); 4531 4532 x86_pmu_show_pmu_cap(pmu->num_counters, pmu->num_counters_fixed, 4533 pmu->intel_ctrl); 4534 4535 end: 4536 cpumask_set_cpu(cpu, &pmu->supported_cpus); 4537 cpuc->pmu = &pmu->pmu; 4538 4539 x86_pmu_update_cpu_context(&pmu->pmu, cpu); 4540 4541 return true; 4542 } 4543 4544 static void intel_pmu_cpu_starting(int cpu) 4545 { 4546 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4547 int core_id = topology_core_id(cpu); 4548 int i; 4549 4550 if (is_hybrid() && !init_hybrid_pmu(cpu)) 4551 return; 4552 4553 init_debug_store_on_cpu(cpu); 4554 /* 4555 * Deal with CPUs that don't clear their LBRs on power-up. 4556 */ 4557 intel_pmu_lbr_reset(); 4558 4559 cpuc->lbr_sel = NULL; 4560 4561 if (x86_pmu.flags & PMU_FL_TFA) { 4562 WARN_ON_ONCE(cpuc->tfa_shadow); 4563 cpuc->tfa_shadow = ~0ULL; 4564 intel_set_tfa(cpuc, false); 4565 } 4566 4567 if (x86_pmu.version > 1) 4568 flip_smm_bit(&x86_pmu.attr_freeze_on_smi); 4569 4570 /* 4571 * Disable perf metrics if any added CPU doesn't support it. 4572 * 4573 * Turn off the check for a hybrid architecture, because the 4574 * architecture MSR, MSR_IA32_PERF_CAPABILITIES, only indicate 4575 * the architecture features. The perf metrics is a model-specific 4576 * feature for now. The corresponding bit should always be 0 on 4577 * a hybrid platform, e.g., Alder Lake. 4578 */ 4579 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) { 4580 union perf_capabilities perf_cap; 4581 4582 rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_cap.capabilities); 4583 if (!perf_cap.perf_metrics) { 4584 x86_pmu.intel_cap.perf_metrics = 0; 4585 x86_pmu.intel_ctrl &= ~(1ULL << GLOBAL_CTRL_EN_PERF_METRICS); 4586 } 4587 } 4588 4589 if (!cpuc->shared_regs) 4590 return; 4591 4592 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { 4593 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4594 struct intel_shared_regs *pc; 4595 4596 pc = per_cpu(cpu_hw_events, i).shared_regs; 4597 if (pc && pc->core_id == core_id) { 4598 cpuc->kfree_on_online[0] = cpuc->shared_regs; 4599 cpuc->shared_regs = pc; 4600 break; 4601 } 4602 } 4603 cpuc->shared_regs->core_id = core_id; 4604 cpuc->shared_regs->refcnt++; 4605 } 4606 4607 if (x86_pmu.lbr_sel_map) 4608 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; 4609 4610 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 4611 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 4612 struct cpu_hw_events *sibling; 4613 struct intel_excl_cntrs *c; 4614 4615 sibling = &per_cpu(cpu_hw_events, i); 4616 c = sibling->excl_cntrs; 4617 if (c && c->core_id == core_id) { 4618 cpuc->kfree_on_online[1] = cpuc->excl_cntrs; 4619 cpuc->excl_cntrs = c; 4620 if (!sibling->excl_thread_id) 4621 cpuc->excl_thread_id = 1; 4622 break; 4623 } 4624 } 4625 cpuc->excl_cntrs->core_id = core_id; 4626 cpuc->excl_cntrs->refcnt++; 4627 } 4628 } 4629 4630 static void free_excl_cntrs(struct cpu_hw_events *cpuc) 4631 { 4632 struct intel_excl_cntrs *c; 4633 4634 c = cpuc->excl_cntrs; 4635 if (c) { 4636 if (c->core_id == -1 || --c->refcnt == 0) 4637 kfree(c); 4638 cpuc->excl_cntrs = NULL; 4639 } 4640 4641 kfree(cpuc->constraint_list); 4642 cpuc->constraint_list = NULL; 4643 } 4644 4645 static void intel_pmu_cpu_dying(int cpu) 4646 { 4647 fini_debug_store_on_cpu(cpu); 4648 } 4649 4650 void intel_cpuc_finish(struct cpu_hw_events *cpuc) 4651 { 4652 struct intel_shared_regs *pc; 4653 4654 pc = cpuc->shared_regs; 4655 if (pc) { 4656 if (pc->core_id == -1 || --pc->refcnt == 0) 4657 kfree(pc); 4658 cpuc->shared_regs = NULL; 4659 } 4660 4661 free_excl_cntrs(cpuc); 4662 } 4663 4664 static void intel_pmu_cpu_dead(int cpu) 4665 { 4666 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 4667 4668 intel_cpuc_finish(cpuc); 4669 4670 if (is_hybrid() && cpuc->pmu) 4671 cpumask_clear_cpu(cpu, &hybrid_pmu(cpuc->pmu)->supported_cpus); 4672 } 4673 4674 static void intel_pmu_sched_task(struct perf_event_context *ctx, 4675 bool sched_in) 4676 { 4677 intel_pmu_pebs_sched_task(ctx, sched_in); 4678 intel_pmu_lbr_sched_task(ctx, sched_in); 4679 } 4680 4681 static void intel_pmu_swap_task_ctx(struct perf_event_context *prev, 4682 struct perf_event_context *next) 4683 { 4684 intel_pmu_lbr_swap_task_ctx(prev, next); 4685 } 4686 4687 static int intel_pmu_check_period(struct perf_event *event, u64 value) 4688 { 4689 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0; 4690 } 4691 4692 static void intel_aux_output_init(void) 4693 { 4694 /* Refer also intel_pmu_aux_output_match() */ 4695 if (x86_pmu.intel_cap.pebs_output_pt_available) 4696 x86_pmu.assign = intel_pmu_assign_event; 4697 } 4698 4699 static int intel_pmu_aux_output_match(struct perf_event *event) 4700 { 4701 /* intel_pmu_assign_event() is needed, refer intel_aux_output_init() */ 4702 if (!x86_pmu.intel_cap.pebs_output_pt_available) 4703 return 0; 4704 4705 return is_intel_pt_event(event); 4706 } 4707 4708 static int intel_pmu_filter_match(struct perf_event *event) 4709 { 4710 struct x86_hybrid_pmu *pmu = hybrid_pmu(event->pmu); 4711 unsigned int cpu = smp_processor_id(); 4712 4713 return cpumask_test_cpu(cpu, &pmu->supported_cpus); 4714 } 4715 4716 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); 4717 4718 PMU_FORMAT_ATTR(ldlat, "config1:0-15"); 4719 4720 PMU_FORMAT_ATTR(frontend, "config1:0-23"); 4721 4722 static struct attribute *intel_arch3_formats_attr[] = { 4723 &format_attr_event.attr, 4724 &format_attr_umask.attr, 4725 &format_attr_edge.attr, 4726 &format_attr_pc.attr, 4727 &format_attr_any.attr, 4728 &format_attr_inv.attr, 4729 &format_attr_cmask.attr, 4730 NULL, 4731 }; 4732 4733 static struct attribute *hsw_format_attr[] = { 4734 &format_attr_in_tx.attr, 4735 &format_attr_in_tx_cp.attr, 4736 &format_attr_offcore_rsp.attr, 4737 &format_attr_ldlat.attr, 4738 NULL 4739 }; 4740 4741 static struct attribute *nhm_format_attr[] = { 4742 &format_attr_offcore_rsp.attr, 4743 &format_attr_ldlat.attr, 4744 NULL 4745 }; 4746 4747 static struct attribute *slm_format_attr[] = { 4748 &format_attr_offcore_rsp.attr, 4749 NULL 4750 }; 4751 4752 static struct attribute *skl_format_attr[] = { 4753 &format_attr_frontend.attr, 4754 NULL, 4755 }; 4756 4757 static __initconst const struct x86_pmu core_pmu = { 4758 .name = "core", 4759 .handle_irq = x86_pmu_handle_irq, 4760 .disable_all = x86_pmu_disable_all, 4761 .enable_all = core_pmu_enable_all, 4762 .enable = core_pmu_enable_event, 4763 .disable = x86_pmu_disable_event, 4764 .hw_config = core_pmu_hw_config, 4765 .schedule_events = x86_schedule_events, 4766 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 4767 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 4768 .event_map = intel_pmu_event_map, 4769 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 4770 .apic = 1, 4771 .large_pebs_flags = LARGE_PEBS_FLAGS, 4772 4773 /* 4774 * Intel PMCs cannot be accessed sanely above 32-bit width, 4775 * so we install an artificial 1<<31 period regardless of 4776 * the generic event period: 4777 */ 4778 .max_period = (1ULL<<31) - 1, 4779 .get_event_constraints = intel_get_event_constraints, 4780 .put_event_constraints = intel_put_event_constraints, 4781 .event_constraints = intel_core_event_constraints, 4782 .guest_get_msrs = core_guest_get_msrs, 4783 .format_attrs = intel_arch_formats_attr, 4784 .events_sysfs_show = intel_event_sysfs_show, 4785 4786 /* 4787 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs 4788 * together with PMU version 1 and thus be using core_pmu with 4789 * shared_regs. We need following callbacks here to allocate 4790 * it properly. 4791 */ 4792 .cpu_prepare = intel_pmu_cpu_prepare, 4793 .cpu_starting = intel_pmu_cpu_starting, 4794 .cpu_dying = intel_pmu_cpu_dying, 4795 .cpu_dead = intel_pmu_cpu_dead, 4796 4797 .check_period = intel_pmu_check_period, 4798 4799 .lbr_reset = intel_pmu_lbr_reset_64, 4800 .lbr_read = intel_pmu_lbr_read_64, 4801 .lbr_save = intel_pmu_lbr_save, 4802 .lbr_restore = intel_pmu_lbr_restore, 4803 }; 4804 4805 static __initconst const struct x86_pmu intel_pmu = { 4806 .name = "Intel", 4807 .handle_irq = intel_pmu_handle_irq, 4808 .disable_all = intel_pmu_disable_all, 4809 .enable_all = intel_pmu_enable_all, 4810 .enable = intel_pmu_enable_event, 4811 .disable = intel_pmu_disable_event, 4812 .add = intel_pmu_add_event, 4813 .del = intel_pmu_del_event, 4814 .read = intel_pmu_read_event, 4815 .set_period = intel_pmu_set_period, 4816 .update = intel_pmu_update, 4817 .hw_config = intel_pmu_hw_config, 4818 .schedule_events = x86_schedule_events, 4819 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 4820 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 4821 .event_map = intel_pmu_event_map, 4822 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 4823 .apic = 1, 4824 .large_pebs_flags = LARGE_PEBS_FLAGS, 4825 /* 4826 * Intel PMCs cannot be accessed sanely above 32 bit width, 4827 * so we install an artificial 1<<31 period regardless of 4828 * the generic event period: 4829 */ 4830 .max_period = (1ULL << 31) - 1, 4831 .get_event_constraints = intel_get_event_constraints, 4832 .put_event_constraints = intel_put_event_constraints, 4833 .pebs_aliases = intel_pebs_aliases_core2, 4834 4835 .format_attrs = intel_arch3_formats_attr, 4836 .events_sysfs_show = intel_event_sysfs_show, 4837 4838 .cpu_prepare = intel_pmu_cpu_prepare, 4839 .cpu_starting = intel_pmu_cpu_starting, 4840 .cpu_dying = intel_pmu_cpu_dying, 4841 .cpu_dead = intel_pmu_cpu_dead, 4842 4843 .guest_get_msrs = intel_guest_get_msrs, 4844 .sched_task = intel_pmu_sched_task, 4845 .swap_task_ctx = intel_pmu_swap_task_ctx, 4846 4847 .check_period = intel_pmu_check_period, 4848 4849 .aux_output_match = intel_pmu_aux_output_match, 4850 4851 .lbr_reset = intel_pmu_lbr_reset_64, 4852 .lbr_read = intel_pmu_lbr_read_64, 4853 .lbr_save = intel_pmu_lbr_save, 4854 .lbr_restore = intel_pmu_lbr_restore, 4855 4856 /* 4857 * SMM has access to all 4 rings and while traditionally SMM code only 4858 * ran in CPL0, 2021-era firmware is starting to make use of CPL3 in SMM. 4859 * 4860 * Since the EVENTSEL.{USR,OS} CPL filtering makes no distinction 4861 * between SMM or not, this results in what should be pure userspace 4862 * counters including SMM data. 4863 * 4864 * This is a clear privilege issue, therefore globally disable 4865 * counting SMM by default. 4866 */ 4867 .attr_freeze_on_smi = 1, 4868 }; 4869 4870 static __init void intel_clovertown_quirk(void) 4871 { 4872 /* 4873 * PEBS is unreliable due to: 4874 * 4875 * AJ67 - PEBS may experience CPL leaks 4876 * AJ68 - PEBS PMI may be delayed by one event 4877 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] 4878 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS 4879 * 4880 * AJ67 could be worked around by restricting the OS/USR flags. 4881 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. 4882 * 4883 * AJ106 could possibly be worked around by not allowing LBR 4884 * usage from PEBS, including the fixup. 4885 * AJ68 could possibly be worked around by always programming 4886 * a pebs_event_reset[0] value and coping with the lost events. 4887 * 4888 * But taken together it might just make sense to not enable PEBS on 4889 * these chips. 4890 */ 4891 pr_warn("PEBS disabled due to CPU errata\n"); 4892 x86_pmu.pebs = 0; 4893 x86_pmu.pebs_constraints = NULL; 4894 } 4895 4896 static const struct x86_cpu_desc isolation_ucodes[] = { 4897 INTEL_CPU_DESC(INTEL_FAM6_HASWELL, 3, 0x0000001f), 4898 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_L, 1, 0x0000001e), 4899 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_G, 1, 0x00000015), 4900 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037), 4901 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a), 4902 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL, 4, 0x00000023), 4903 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_G, 1, 0x00000014), 4904 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 2, 0x00000010), 4905 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 3, 0x07000009), 4906 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 4, 0x0f000009), 4907 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_D, 5, 0x0e000002), 4908 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 1, 0x0b000014), 4909 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021), 4910 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000), 4911 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 5, 0x00000000), 4912 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 6, 0x00000000), 4913 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 7, 0x00000000), 4914 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_L, 3, 0x0000007c), 4915 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE, 3, 0x0000007c), 4916 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 9, 0x0000004e), 4917 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 9, 0x0000004e), 4918 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 10, 0x0000004e), 4919 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 11, 0x0000004e), 4920 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_L, 12, 0x0000004e), 4921 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 10, 0x0000004e), 4922 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 11, 0x0000004e), 4923 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 12, 0x0000004e), 4924 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE, 13, 0x0000004e), 4925 {} 4926 }; 4927 4928 static void intel_check_pebs_isolation(void) 4929 { 4930 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes); 4931 } 4932 4933 static __init void intel_pebs_isolation_quirk(void) 4934 { 4935 WARN_ON_ONCE(x86_pmu.check_microcode); 4936 x86_pmu.check_microcode = intel_check_pebs_isolation; 4937 intel_check_pebs_isolation(); 4938 } 4939 4940 static const struct x86_cpu_desc pebs_ucodes[] = { 4941 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028), 4942 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618), 4943 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c), 4944 {} 4945 }; 4946 4947 static bool intel_snb_pebs_broken(void) 4948 { 4949 return !x86_cpu_has_min_microcode_rev(pebs_ucodes); 4950 } 4951 4952 static void intel_snb_check_microcode(void) 4953 { 4954 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken) 4955 return; 4956 4957 /* 4958 * Serialized by the microcode lock.. 4959 */ 4960 if (x86_pmu.pebs_broken) { 4961 pr_info("PEBS enabled due to microcode update\n"); 4962 x86_pmu.pebs_broken = 0; 4963 } else { 4964 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); 4965 x86_pmu.pebs_broken = 1; 4966 } 4967 } 4968 4969 static bool is_lbr_from(unsigned long msr) 4970 { 4971 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr; 4972 4973 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr; 4974 } 4975 4976 /* 4977 * Under certain circumstances, access certain MSR may cause #GP. 4978 * The function tests if the input MSR can be safely accessed. 4979 */ 4980 static bool check_msr(unsigned long msr, u64 mask) 4981 { 4982 u64 val_old, val_new, val_tmp; 4983 4984 /* 4985 * Disable the check for real HW, so we don't 4986 * mess with potentially enabled registers: 4987 */ 4988 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) 4989 return true; 4990 4991 /* 4992 * Read the current value, change it and read it back to see if it 4993 * matches, this is needed to detect certain hardware emulators 4994 * (qemu/kvm) that don't trap on the MSR access and always return 0s. 4995 */ 4996 if (rdmsrl_safe(msr, &val_old)) 4997 return false; 4998 4999 /* 5000 * Only change the bits which can be updated by wrmsrl. 5001 */ 5002 val_tmp = val_old ^ mask; 5003 5004 if (is_lbr_from(msr)) 5005 val_tmp = lbr_from_signext_quirk_wr(val_tmp); 5006 5007 if (wrmsrl_safe(msr, val_tmp) || 5008 rdmsrl_safe(msr, &val_new)) 5009 return false; 5010 5011 /* 5012 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value 5013 * should equal rdmsrl()'s even with the quirk. 5014 */ 5015 if (val_new != val_tmp) 5016 return false; 5017 5018 if (is_lbr_from(msr)) 5019 val_old = lbr_from_signext_quirk_wr(val_old); 5020 5021 /* Here it's sure that the MSR can be safely accessed. 5022 * Restore the old value and return. 5023 */ 5024 wrmsrl(msr, val_old); 5025 5026 return true; 5027 } 5028 5029 static __init void intel_sandybridge_quirk(void) 5030 { 5031 x86_pmu.check_microcode = intel_snb_check_microcode; 5032 cpus_read_lock(); 5033 intel_snb_check_microcode(); 5034 cpus_read_unlock(); 5035 } 5036 5037 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { 5038 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, 5039 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, 5040 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, 5041 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, 5042 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, 5043 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, 5044 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, 5045 }; 5046 5047 static __init void intel_arch_events_quirk(void) 5048 { 5049 int bit; 5050 5051 /* disable event that reported as not present by cpuid */ 5052 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { 5053 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; 5054 pr_warn("CPUID marked event: \'%s\' unavailable\n", 5055 intel_arch_events_map[bit].name); 5056 } 5057 } 5058 5059 static __init void intel_nehalem_quirk(void) 5060 { 5061 union cpuid10_ebx ebx; 5062 5063 ebx.full = x86_pmu.events_maskl; 5064 if (ebx.split.no_branch_misses_retired) { 5065 /* 5066 * Erratum AAJ80 detected, we work it around by using 5067 * the BR_MISP_EXEC.ANY event. This will over-count 5068 * branch-misses, but it's still much better than the 5069 * architectural event which is often completely bogus: 5070 */ 5071 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 5072 ebx.split.no_branch_misses_retired = 0; 5073 x86_pmu.events_maskl = ebx.full; 5074 pr_info("CPU erratum AAJ80 worked around\n"); 5075 } 5076 } 5077 5078 /* 5079 * enable software workaround for errata: 5080 * SNB: BJ122 5081 * IVB: BV98 5082 * HSW: HSD29 5083 * 5084 * Only needed when HT is enabled. However detecting 5085 * if HT is enabled is difficult (model specific). So instead, 5086 * we enable the workaround in the early boot, and verify if 5087 * it is needed in a later initcall phase once we have valid 5088 * topology information to check if HT is actually enabled 5089 */ 5090 static __init void intel_ht_bug(void) 5091 { 5092 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; 5093 5094 x86_pmu.start_scheduling = intel_start_scheduling; 5095 x86_pmu.commit_scheduling = intel_commit_scheduling; 5096 x86_pmu.stop_scheduling = intel_stop_scheduling; 5097 } 5098 5099 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); 5100 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") 5101 5102 /* Haswell special events */ 5103 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); 5104 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); 5105 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); 5106 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); 5107 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); 5108 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); 5109 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); 5110 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); 5111 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); 5112 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); 5113 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); 5114 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); 5115 5116 static struct attribute *hsw_events_attrs[] = { 5117 EVENT_PTR(td_slots_issued), 5118 EVENT_PTR(td_slots_retired), 5119 EVENT_PTR(td_fetch_bubbles), 5120 EVENT_PTR(td_total_slots), 5121 EVENT_PTR(td_total_slots_scale), 5122 EVENT_PTR(td_recovery_bubbles), 5123 EVENT_PTR(td_recovery_bubbles_scale), 5124 NULL 5125 }; 5126 5127 static struct attribute *hsw_mem_events_attrs[] = { 5128 EVENT_PTR(mem_ld_hsw), 5129 EVENT_PTR(mem_st_hsw), 5130 NULL, 5131 }; 5132 5133 static struct attribute *hsw_tsx_events_attrs[] = { 5134 EVENT_PTR(tx_start), 5135 EVENT_PTR(tx_commit), 5136 EVENT_PTR(tx_abort), 5137 EVENT_PTR(tx_capacity), 5138 EVENT_PTR(tx_conflict), 5139 EVENT_PTR(el_start), 5140 EVENT_PTR(el_commit), 5141 EVENT_PTR(el_abort), 5142 EVENT_PTR(el_capacity), 5143 EVENT_PTR(el_conflict), 5144 EVENT_PTR(cycles_t), 5145 EVENT_PTR(cycles_ct), 5146 NULL 5147 }; 5148 5149 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80"); 5150 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2"); 5151 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80"); 5152 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2"); 5153 5154 static struct attribute *icl_events_attrs[] = { 5155 EVENT_PTR(mem_ld_hsw), 5156 EVENT_PTR(mem_st_hsw), 5157 NULL, 5158 }; 5159 5160 static struct attribute *icl_td_events_attrs[] = { 5161 EVENT_PTR(slots), 5162 EVENT_PTR(td_retiring), 5163 EVENT_PTR(td_bad_spec), 5164 EVENT_PTR(td_fe_bound), 5165 EVENT_PTR(td_be_bound), 5166 NULL, 5167 }; 5168 5169 static struct attribute *icl_tsx_events_attrs[] = { 5170 EVENT_PTR(tx_start), 5171 EVENT_PTR(tx_abort), 5172 EVENT_PTR(tx_commit), 5173 EVENT_PTR(tx_capacity_read), 5174 EVENT_PTR(tx_capacity_write), 5175 EVENT_PTR(tx_conflict), 5176 EVENT_PTR(el_start), 5177 EVENT_PTR(el_abort), 5178 EVENT_PTR(el_commit), 5179 EVENT_PTR(el_capacity_read), 5180 EVENT_PTR(el_capacity_write), 5181 EVENT_PTR(el_conflict), 5182 EVENT_PTR(cycles_t), 5183 EVENT_PTR(cycles_ct), 5184 NULL, 5185 }; 5186 5187 5188 EVENT_ATTR_STR(mem-stores, mem_st_spr, "event=0xcd,umask=0x2"); 5189 EVENT_ATTR_STR(mem-loads-aux, mem_ld_aux, "event=0x03,umask=0x82"); 5190 5191 static struct attribute *spr_events_attrs[] = { 5192 EVENT_PTR(mem_ld_hsw), 5193 EVENT_PTR(mem_st_spr), 5194 EVENT_PTR(mem_ld_aux), 5195 NULL, 5196 }; 5197 5198 static struct attribute *spr_td_events_attrs[] = { 5199 EVENT_PTR(slots), 5200 EVENT_PTR(td_retiring), 5201 EVENT_PTR(td_bad_spec), 5202 EVENT_PTR(td_fe_bound), 5203 EVENT_PTR(td_be_bound), 5204 EVENT_PTR(td_heavy_ops), 5205 EVENT_PTR(td_br_mispredict), 5206 EVENT_PTR(td_fetch_lat), 5207 EVENT_PTR(td_mem_bound), 5208 NULL, 5209 }; 5210 5211 static struct attribute *spr_tsx_events_attrs[] = { 5212 EVENT_PTR(tx_start), 5213 EVENT_PTR(tx_abort), 5214 EVENT_PTR(tx_commit), 5215 EVENT_PTR(tx_capacity_read), 5216 EVENT_PTR(tx_capacity_write), 5217 EVENT_PTR(tx_conflict), 5218 EVENT_PTR(cycles_t), 5219 EVENT_PTR(cycles_ct), 5220 NULL, 5221 }; 5222 5223 static ssize_t freeze_on_smi_show(struct device *cdev, 5224 struct device_attribute *attr, 5225 char *buf) 5226 { 5227 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi); 5228 } 5229 5230 static DEFINE_MUTEX(freeze_on_smi_mutex); 5231 5232 static ssize_t freeze_on_smi_store(struct device *cdev, 5233 struct device_attribute *attr, 5234 const char *buf, size_t count) 5235 { 5236 unsigned long val; 5237 ssize_t ret; 5238 5239 ret = kstrtoul(buf, 0, &val); 5240 if (ret) 5241 return ret; 5242 5243 if (val > 1) 5244 return -EINVAL; 5245 5246 mutex_lock(&freeze_on_smi_mutex); 5247 5248 if (x86_pmu.attr_freeze_on_smi == val) 5249 goto done; 5250 5251 x86_pmu.attr_freeze_on_smi = val; 5252 5253 cpus_read_lock(); 5254 on_each_cpu(flip_smm_bit, &val, 1); 5255 cpus_read_unlock(); 5256 done: 5257 mutex_unlock(&freeze_on_smi_mutex); 5258 5259 return count; 5260 } 5261 5262 static void update_tfa_sched(void *ignored) 5263 { 5264 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 5265 5266 /* 5267 * check if PMC3 is used 5268 * and if so force schedule out for all event types all contexts 5269 */ 5270 if (test_bit(3, cpuc->active_mask)) 5271 perf_pmu_resched(x86_get_pmu(smp_processor_id())); 5272 } 5273 5274 static ssize_t show_sysctl_tfa(struct device *cdev, 5275 struct device_attribute *attr, 5276 char *buf) 5277 { 5278 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort); 5279 } 5280 5281 static ssize_t set_sysctl_tfa(struct device *cdev, 5282 struct device_attribute *attr, 5283 const char *buf, size_t count) 5284 { 5285 bool val; 5286 ssize_t ret; 5287 5288 ret = kstrtobool(buf, &val); 5289 if (ret) 5290 return ret; 5291 5292 /* no change */ 5293 if (val == allow_tsx_force_abort) 5294 return count; 5295 5296 allow_tsx_force_abort = val; 5297 5298 cpus_read_lock(); 5299 on_each_cpu(update_tfa_sched, NULL, 1); 5300 cpus_read_unlock(); 5301 5302 return count; 5303 } 5304 5305 5306 static DEVICE_ATTR_RW(freeze_on_smi); 5307 5308 static ssize_t branches_show(struct device *cdev, 5309 struct device_attribute *attr, 5310 char *buf) 5311 { 5312 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr); 5313 } 5314 5315 static DEVICE_ATTR_RO(branches); 5316 5317 static struct attribute *lbr_attrs[] = { 5318 &dev_attr_branches.attr, 5319 NULL 5320 }; 5321 5322 static char pmu_name_str[30]; 5323 5324 static ssize_t pmu_name_show(struct device *cdev, 5325 struct device_attribute *attr, 5326 char *buf) 5327 { 5328 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str); 5329 } 5330 5331 static DEVICE_ATTR_RO(pmu_name); 5332 5333 static struct attribute *intel_pmu_caps_attrs[] = { 5334 &dev_attr_pmu_name.attr, 5335 NULL 5336 }; 5337 5338 static DEVICE_ATTR(allow_tsx_force_abort, 0644, 5339 show_sysctl_tfa, 5340 set_sysctl_tfa); 5341 5342 static struct attribute *intel_pmu_attrs[] = { 5343 &dev_attr_freeze_on_smi.attr, 5344 &dev_attr_allow_tsx_force_abort.attr, 5345 NULL, 5346 }; 5347 5348 static umode_t 5349 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5350 { 5351 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0; 5352 } 5353 5354 static umode_t 5355 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5356 { 5357 return x86_pmu.pebs ? attr->mode : 0; 5358 } 5359 5360 static umode_t 5361 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5362 { 5363 return x86_pmu.lbr_nr ? attr->mode : 0; 5364 } 5365 5366 static umode_t 5367 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5368 { 5369 return x86_pmu.version >= 2 ? attr->mode : 0; 5370 } 5371 5372 static umode_t 5373 default_is_visible(struct kobject *kobj, struct attribute *attr, int i) 5374 { 5375 if (attr == &dev_attr_allow_tsx_force_abort.attr) 5376 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0; 5377 5378 return attr->mode; 5379 } 5380 5381 static struct attribute_group group_events_td = { 5382 .name = "events", 5383 }; 5384 5385 static struct attribute_group group_events_mem = { 5386 .name = "events", 5387 .is_visible = pebs_is_visible, 5388 }; 5389 5390 static struct attribute_group group_events_tsx = { 5391 .name = "events", 5392 .is_visible = tsx_is_visible, 5393 }; 5394 5395 static struct attribute_group group_caps_gen = { 5396 .name = "caps", 5397 .attrs = intel_pmu_caps_attrs, 5398 }; 5399 5400 static struct attribute_group group_caps_lbr = { 5401 .name = "caps", 5402 .attrs = lbr_attrs, 5403 .is_visible = lbr_is_visible, 5404 }; 5405 5406 static struct attribute_group group_format_extra = { 5407 .name = "format", 5408 .is_visible = exra_is_visible, 5409 }; 5410 5411 static struct attribute_group group_format_extra_skl = { 5412 .name = "format", 5413 .is_visible = exra_is_visible, 5414 }; 5415 5416 static struct attribute_group group_default = { 5417 .attrs = intel_pmu_attrs, 5418 .is_visible = default_is_visible, 5419 }; 5420 5421 static const struct attribute_group *attr_update[] = { 5422 &group_events_td, 5423 &group_events_mem, 5424 &group_events_tsx, 5425 &group_caps_gen, 5426 &group_caps_lbr, 5427 &group_format_extra, 5428 &group_format_extra_skl, 5429 &group_default, 5430 NULL, 5431 }; 5432 5433 EVENT_ATTR_STR_HYBRID(slots, slots_adl, "event=0x00,umask=0x4", hybrid_big); 5434 EVENT_ATTR_STR_HYBRID(topdown-retiring, td_retiring_adl, "event=0xc2,umask=0x0;event=0x00,umask=0x80", hybrid_big_small); 5435 EVENT_ATTR_STR_HYBRID(topdown-bad-spec, td_bad_spec_adl, "event=0x73,umask=0x0;event=0x00,umask=0x81", hybrid_big_small); 5436 EVENT_ATTR_STR_HYBRID(topdown-fe-bound, td_fe_bound_adl, "event=0x71,umask=0x0;event=0x00,umask=0x82", hybrid_big_small); 5437 EVENT_ATTR_STR_HYBRID(topdown-be-bound, td_be_bound_adl, "event=0x74,umask=0x0;event=0x00,umask=0x83", hybrid_big_small); 5438 EVENT_ATTR_STR_HYBRID(topdown-heavy-ops, td_heavy_ops_adl, "event=0x00,umask=0x84", hybrid_big); 5439 EVENT_ATTR_STR_HYBRID(topdown-br-mispredict, td_br_mis_adl, "event=0x00,umask=0x85", hybrid_big); 5440 EVENT_ATTR_STR_HYBRID(topdown-fetch-lat, td_fetch_lat_adl, "event=0x00,umask=0x86", hybrid_big); 5441 EVENT_ATTR_STR_HYBRID(topdown-mem-bound, td_mem_bound_adl, "event=0x00,umask=0x87", hybrid_big); 5442 5443 static struct attribute *adl_hybrid_events_attrs[] = { 5444 EVENT_PTR(slots_adl), 5445 EVENT_PTR(td_retiring_adl), 5446 EVENT_PTR(td_bad_spec_adl), 5447 EVENT_PTR(td_fe_bound_adl), 5448 EVENT_PTR(td_be_bound_adl), 5449 EVENT_PTR(td_heavy_ops_adl), 5450 EVENT_PTR(td_br_mis_adl), 5451 EVENT_PTR(td_fetch_lat_adl), 5452 EVENT_PTR(td_mem_bound_adl), 5453 NULL, 5454 }; 5455 5456 /* Must be in IDX order */ 5457 EVENT_ATTR_STR_HYBRID(mem-loads, mem_ld_adl, "event=0xd0,umask=0x5,ldlat=3;event=0xcd,umask=0x1,ldlat=3", hybrid_big_small); 5458 EVENT_ATTR_STR_HYBRID(mem-stores, mem_st_adl, "event=0xd0,umask=0x6;event=0xcd,umask=0x2", hybrid_big_small); 5459 EVENT_ATTR_STR_HYBRID(mem-loads-aux, mem_ld_aux_adl, "event=0x03,umask=0x82", hybrid_big); 5460 5461 static struct attribute *adl_hybrid_mem_attrs[] = { 5462 EVENT_PTR(mem_ld_adl), 5463 EVENT_PTR(mem_st_adl), 5464 EVENT_PTR(mem_ld_aux_adl), 5465 NULL, 5466 }; 5467 5468 EVENT_ATTR_STR_HYBRID(tx-start, tx_start_adl, "event=0xc9,umask=0x1", hybrid_big); 5469 EVENT_ATTR_STR_HYBRID(tx-commit, tx_commit_adl, "event=0xc9,umask=0x2", hybrid_big); 5470 EVENT_ATTR_STR_HYBRID(tx-abort, tx_abort_adl, "event=0xc9,umask=0x4", hybrid_big); 5471 EVENT_ATTR_STR_HYBRID(tx-conflict, tx_conflict_adl, "event=0x54,umask=0x1", hybrid_big); 5472 EVENT_ATTR_STR_HYBRID(cycles-t, cycles_t_adl, "event=0x3c,in_tx=1", hybrid_big); 5473 EVENT_ATTR_STR_HYBRID(cycles-ct, cycles_ct_adl, "event=0x3c,in_tx=1,in_tx_cp=1", hybrid_big); 5474 EVENT_ATTR_STR_HYBRID(tx-capacity-read, tx_capacity_read_adl, "event=0x54,umask=0x80", hybrid_big); 5475 EVENT_ATTR_STR_HYBRID(tx-capacity-write, tx_capacity_write_adl, "event=0x54,umask=0x2", hybrid_big); 5476 5477 static struct attribute *adl_hybrid_tsx_attrs[] = { 5478 EVENT_PTR(tx_start_adl), 5479 EVENT_PTR(tx_abort_adl), 5480 EVENT_PTR(tx_commit_adl), 5481 EVENT_PTR(tx_capacity_read_adl), 5482 EVENT_PTR(tx_capacity_write_adl), 5483 EVENT_PTR(tx_conflict_adl), 5484 EVENT_PTR(cycles_t_adl), 5485 EVENT_PTR(cycles_ct_adl), 5486 NULL, 5487 }; 5488 5489 FORMAT_ATTR_HYBRID(in_tx, hybrid_big); 5490 FORMAT_ATTR_HYBRID(in_tx_cp, hybrid_big); 5491 FORMAT_ATTR_HYBRID(offcore_rsp, hybrid_big_small); 5492 FORMAT_ATTR_HYBRID(ldlat, hybrid_big_small); 5493 FORMAT_ATTR_HYBRID(frontend, hybrid_big); 5494 5495 static struct attribute *adl_hybrid_extra_attr_rtm[] = { 5496 FORMAT_HYBRID_PTR(in_tx), 5497 FORMAT_HYBRID_PTR(in_tx_cp), 5498 FORMAT_HYBRID_PTR(offcore_rsp), 5499 FORMAT_HYBRID_PTR(ldlat), 5500 FORMAT_HYBRID_PTR(frontend), 5501 NULL, 5502 }; 5503 5504 static struct attribute *adl_hybrid_extra_attr[] = { 5505 FORMAT_HYBRID_PTR(offcore_rsp), 5506 FORMAT_HYBRID_PTR(ldlat), 5507 FORMAT_HYBRID_PTR(frontend), 5508 NULL, 5509 }; 5510 5511 static bool is_attr_for_this_pmu(struct kobject *kobj, struct attribute *attr) 5512 { 5513 struct device *dev = kobj_to_dev(kobj); 5514 struct x86_hybrid_pmu *pmu = 5515 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5516 struct perf_pmu_events_hybrid_attr *pmu_attr = 5517 container_of(attr, struct perf_pmu_events_hybrid_attr, attr.attr); 5518 5519 return pmu->cpu_type & pmu_attr->pmu_type; 5520 } 5521 5522 static umode_t hybrid_events_is_visible(struct kobject *kobj, 5523 struct attribute *attr, int i) 5524 { 5525 return is_attr_for_this_pmu(kobj, attr) ? attr->mode : 0; 5526 } 5527 5528 static inline int hybrid_find_supported_cpu(struct x86_hybrid_pmu *pmu) 5529 { 5530 int cpu = cpumask_first(&pmu->supported_cpus); 5531 5532 return (cpu >= nr_cpu_ids) ? -1 : cpu; 5533 } 5534 5535 static umode_t hybrid_tsx_is_visible(struct kobject *kobj, 5536 struct attribute *attr, int i) 5537 { 5538 struct device *dev = kobj_to_dev(kobj); 5539 struct x86_hybrid_pmu *pmu = 5540 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5541 int cpu = hybrid_find_supported_cpu(pmu); 5542 5543 return (cpu >= 0) && is_attr_for_this_pmu(kobj, attr) && cpu_has(&cpu_data(cpu), X86_FEATURE_RTM) ? attr->mode : 0; 5544 } 5545 5546 static umode_t hybrid_format_is_visible(struct kobject *kobj, 5547 struct attribute *attr, int i) 5548 { 5549 struct device *dev = kobj_to_dev(kobj); 5550 struct x86_hybrid_pmu *pmu = 5551 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5552 struct perf_pmu_format_hybrid_attr *pmu_attr = 5553 container_of(attr, struct perf_pmu_format_hybrid_attr, attr.attr); 5554 int cpu = hybrid_find_supported_cpu(pmu); 5555 5556 return (cpu >= 0) && (pmu->cpu_type & pmu_attr->pmu_type) ? attr->mode : 0; 5557 } 5558 5559 static struct attribute_group hybrid_group_events_td = { 5560 .name = "events", 5561 .is_visible = hybrid_events_is_visible, 5562 }; 5563 5564 static struct attribute_group hybrid_group_events_mem = { 5565 .name = "events", 5566 .is_visible = hybrid_events_is_visible, 5567 }; 5568 5569 static struct attribute_group hybrid_group_events_tsx = { 5570 .name = "events", 5571 .is_visible = hybrid_tsx_is_visible, 5572 }; 5573 5574 static struct attribute_group hybrid_group_format_extra = { 5575 .name = "format", 5576 .is_visible = hybrid_format_is_visible, 5577 }; 5578 5579 static ssize_t intel_hybrid_get_attr_cpus(struct device *dev, 5580 struct device_attribute *attr, 5581 char *buf) 5582 { 5583 struct x86_hybrid_pmu *pmu = 5584 container_of(dev_get_drvdata(dev), struct x86_hybrid_pmu, pmu); 5585 5586 return cpumap_print_to_pagebuf(true, buf, &pmu->supported_cpus); 5587 } 5588 5589 static DEVICE_ATTR(cpus, S_IRUGO, intel_hybrid_get_attr_cpus, NULL); 5590 static struct attribute *intel_hybrid_cpus_attrs[] = { 5591 &dev_attr_cpus.attr, 5592 NULL, 5593 }; 5594 5595 static struct attribute_group hybrid_group_cpus = { 5596 .attrs = intel_hybrid_cpus_attrs, 5597 }; 5598 5599 static const struct attribute_group *hybrid_attr_update[] = { 5600 &hybrid_group_events_td, 5601 &hybrid_group_events_mem, 5602 &hybrid_group_events_tsx, 5603 &group_caps_gen, 5604 &group_caps_lbr, 5605 &hybrid_group_format_extra, 5606 &group_default, 5607 &hybrid_group_cpus, 5608 NULL, 5609 }; 5610 5611 static struct attribute *empty_attrs; 5612 5613 static void intel_pmu_check_num_counters(int *num_counters, 5614 int *num_counters_fixed, 5615 u64 *intel_ctrl, u64 fixed_mask) 5616 { 5617 if (*num_counters > INTEL_PMC_MAX_GENERIC) { 5618 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", 5619 *num_counters, INTEL_PMC_MAX_GENERIC); 5620 *num_counters = INTEL_PMC_MAX_GENERIC; 5621 } 5622 *intel_ctrl = (1ULL << *num_counters) - 1; 5623 5624 if (*num_counters_fixed > INTEL_PMC_MAX_FIXED) { 5625 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", 5626 *num_counters_fixed, INTEL_PMC_MAX_FIXED); 5627 *num_counters_fixed = INTEL_PMC_MAX_FIXED; 5628 } 5629 5630 *intel_ctrl |= fixed_mask << INTEL_PMC_IDX_FIXED; 5631 } 5632 5633 static void intel_pmu_check_event_constraints(struct event_constraint *event_constraints, 5634 int num_counters, 5635 int num_counters_fixed, 5636 u64 intel_ctrl) 5637 { 5638 struct event_constraint *c; 5639 5640 if (!event_constraints) 5641 return; 5642 5643 /* 5644 * event on fixed counter2 (REF_CYCLES) only works on this 5645 * counter, so do not extend mask to generic counters 5646 */ 5647 for_each_event_constraint(c, event_constraints) { 5648 /* 5649 * Don't extend the topdown slots and metrics 5650 * events to the generic counters. 5651 */ 5652 if (c->idxmsk64 & INTEL_PMC_MSK_TOPDOWN) { 5653 /* 5654 * Disable topdown slots and metrics events, 5655 * if slots event is not in CPUID. 5656 */ 5657 if (!(INTEL_PMC_MSK_FIXED_SLOTS & intel_ctrl)) 5658 c->idxmsk64 = 0; 5659 c->weight = hweight64(c->idxmsk64); 5660 continue; 5661 } 5662 5663 if (c->cmask == FIXED_EVENT_FLAGS) { 5664 /* Disabled fixed counters which are not in CPUID */ 5665 c->idxmsk64 &= intel_ctrl; 5666 5667 /* 5668 * Don't extend the pseudo-encoding to the 5669 * generic counters 5670 */ 5671 if (!use_fixed_pseudo_encoding(c->code)) 5672 c->idxmsk64 |= (1ULL << num_counters) - 1; 5673 } 5674 c->idxmsk64 &= 5675 ~(~0ULL << (INTEL_PMC_IDX_FIXED + num_counters_fixed)); 5676 c->weight = hweight64(c->idxmsk64); 5677 } 5678 } 5679 5680 static void intel_pmu_check_extra_regs(struct extra_reg *extra_regs) 5681 { 5682 struct extra_reg *er; 5683 5684 /* 5685 * Access extra MSR may cause #GP under certain circumstances. 5686 * E.g. KVM doesn't support offcore event 5687 * Check all extra_regs here. 5688 */ 5689 if (!extra_regs) 5690 return; 5691 5692 for (er = extra_regs; er->msr; er++) { 5693 er->extra_msr_access = check_msr(er->msr, 0x11UL); 5694 /* Disable LBR select mapping */ 5695 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) 5696 x86_pmu.lbr_sel_map = NULL; 5697 } 5698 } 5699 5700 static void intel_pmu_check_hybrid_pmus(u64 fixed_mask) 5701 { 5702 struct x86_hybrid_pmu *pmu; 5703 int i; 5704 5705 for (i = 0; i < x86_pmu.num_hybrid_pmus; i++) { 5706 pmu = &x86_pmu.hybrid_pmu[i]; 5707 5708 intel_pmu_check_num_counters(&pmu->num_counters, 5709 &pmu->num_counters_fixed, 5710 &pmu->intel_ctrl, 5711 fixed_mask); 5712 5713 if (pmu->intel_cap.perf_metrics) { 5714 pmu->intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 5715 pmu->intel_ctrl |= INTEL_PMC_MSK_FIXED_SLOTS; 5716 } 5717 5718 if (pmu->intel_cap.pebs_output_pt_available) 5719 pmu->pmu.capabilities |= PERF_PMU_CAP_AUX_OUTPUT; 5720 5721 intel_pmu_check_event_constraints(pmu->event_constraints, 5722 pmu->num_counters, 5723 pmu->num_counters_fixed, 5724 pmu->intel_ctrl); 5725 5726 intel_pmu_check_extra_regs(pmu->extra_regs); 5727 } 5728 } 5729 5730 __init int intel_pmu_init(void) 5731 { 5732 struct attribute **extra_skl_attr = &empty_attrs; 5733 struct attribute **extra_attr = &empty_attrs; 5734 struct attribute **td_attr = &empty_attrs; 5735 struct attribute **mem_attr = &empty_attrs; 5736 struct attribute **tsx_attr = &empty_attrs; 5737 union cpuid10_edx edx; 5738 union cpuid10_eax eax; 5739 union cpuid10_ebx ebx; 5740 unsigned int fixed_mask; 5741 bool pmem = false; 5742 int version, i; 5743 char *name; 5744 struct x86_hybrid_pmu *pmu; 5745 5746 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 5747 switch (boot_cpu_data.x86) { 5748 case 0x6: 5749 return p6_pmu_init(); 5750 case 0xb: 5751 return knc_pmu_init(); 5752 case 0xf: 5753 return p4_pmu_init(); 5754 } 5755 return -ENODEV; 5756 } 5757 5758 /* 5759 * Check whether the Architectural PerfMon supports 5760 * Branch Misses Retired hw_event or not. 5761 */ 5762 cpuid(10, &eax.full, &ebx.full, &fixed_mask, &edx.full); 5763 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) 5764 return -ENODEV; 5765 5766 version = eax.split.version_id; 5767 if (version < 2) 5768 x86_pmu = core_pmu; 5769 else 5770 x86_pmu = intel_pmu; 5771 5772 x86_pmu.version = version; 5773 x86_pmu.num_counters = eax.split.num_counters; 5774 x86_pmu.cntval_bits = eax.split.bit_width; 5775 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; 5776 5777 x86_pmu.events_maskl = ebx.full; 5778 x86_pmu.events_mask_len = eax.split.mask_length; 5779 5780 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); 5781 x86_pmu.pebs_capable = PEBS_COUNTER_MASK; 5782 5783 /* 5784 * Quirk: v2 perfmon does not report fixed-purpose events, so 5785 * assume at least 3 events, when not running in a hypervisor: 5786 */ 5787 if (version > 1 && version < 5) { 5788 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR); 5789 5790 x86_pmu.num_counters_fixed = 5791 max((int)edx.split.num_counters_fixed, assume); 5792 5793 fixed_mask = (1L << x86_pmu.num_counters_fixed) - 1; 5794 } else if (version >= 5) 5795 x86_pmu.num_counters_fixed = fls(fixed_mask); 5796 5797 if (boot_cpu_has(X86_FEATURE_PDCM)) { 5798 u64 capabilities; 5799 5800 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); 5801 x86_pmu.intel_cap.capabilities = capabilities; 5802 } 5803 5804 if (x86_pmu.intel_cap.lbr_format == LBR_FORMAT_32) { 5805 x86_pmu.lbr_reset = intel_pmu_lbr_reset_32; 5806 x86_pmu.lbr_read = intel_pmu_lbr_read_32; 5807 } 5808 5809 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) 5810 intel_pmu_arch_lbr_init(); 5811 5812 intel_ds_init(); 5813 5814 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ 5815 5816 if (version >= 5) { 5817 x86_pmu.intel_cap.anythread_deprecated = edx.split.anythread_deprecated; 5818 if (x86_pmu.intel_cap.anythread_deprecated) 5819 pr_cont(" AnyThread deprecated, "); 5820 } 5821 5822 /* 5823 * Install the hw-cache-events table: 5824 */ 5825 switch (boot_cpu_data.x86_model) { 5826 case INTEL_FAM6_CORE_YONAH: 5827 pr_cont("Core events, "); 5828 name = "core"; 5829 break; 5830 5831 case INTEL_FAM6_CORE2_MEROM: 5832 x86_add_quirk(intel_clovertown_quirk); 5833 fallthrough; 5834 5835 case INTEL_FAM6_CORE2_MEROM_L: 5836 case INTEL_FAM6_CORE2_PENRYN: 5837 case INTEL_FAM6_CORE2_DUNNINGTON: 5838 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, 5839 sizeof(hw_cache_event_ids)); 5840 5841 intel_pmu_lbr_init_core(); 5842 5843 x86_pmu.event_constraints = intel_core2_event_constraints; 5844 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; 5845 pr_cont("Core2 events, "); 5846 name = "core2"; 5847 break; 5848 5849 case INTEL_FAM6_NEHALEM: 5850 case INTEL_FAM6_NEHALEM_EP: 5851 case INTEL_FAM6_NEHALEM_EX: 5852 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, 5853 sizeof(hw_cache_event_ids)); 5854 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 5855 sizeof(hw_cache_extra_regs)); 5856 5857 intel_pmu_lbr_init_nhm(); 5858 5859 x86_pmu.event_constraints = intel_nehalem_event_constraints; 5860 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; 5861 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 5862 x86_pmu.extra_regs = intel_nehalem_extra_regs; 5863 x86_pmu.limit_period = nhm_limit_period; 5864 5865 mem_attr = nhm_mem_events_attrs; 5866 5867 /* UOPS_ISSUED.STALLED_CYCLES */ 5868 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 5869 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 5870 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 5871 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 5872 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 5873 5874 intel_pmu_pebs_data_source_nhm(); 5875 x86_add_quirk(intel_nehalem_quirk); 5876 x86_pmu.pebs_no_tlb = 1; 5877 extra_attr = nhm_format_attr; 5878 5879 pr_cont("Nehalem events, "); 5880 name = "nehalem"; 5881 break; 5882 5883 case INTEL_FAM6_ATOM_BONNELL: 5884 case INTEL_FAM6_ATOM_BONNELL_MID: 5885 case INTEL_FAM6_ATOM_SALTWELL: 5886 case INTEL_FAM6_ATOM_SALTWELL_MID: 5887 case INTEL_FAM6_ATOM_SALTWELL_TABLET: 5888 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, 5889 sizeof(hw_cache_event_ids)); 5890 5891 intel_pmu_lbr_init_atom(); 5892 5893 x86_pmu.event_constraints = intel_gen_event_constraints; 5894 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; 5895 x86_pmu.pebs_aliases = intel_pebs_aliases_core2; 5896 pr_cont("Atom events, "); 5897 name = "bonnell"; 5898 break; 5899 5900 case INTEL_FAM6_ATOM_SILVERMONT: 5901 case INTEL_FAM6_ATOM_SILVERMONT_D: 5902 case INTEL_FAM6_ATOM_SILVERMONT_MID: 5903 case INTEL_FAM6_ATOM_AIRMONT: 5904 case INTEL_FAM6_ATOM_AIRMONT_MID: 5905 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, 5906 sizeof(hw_cache_event_ids)); 5907 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, 5908 sizeof(hw_cache_extra_regs)); 5909 5910 intel_pmu_lbr_init_slm(); 5911 5912 x86_pmu.event_constraints = intel_slm_event_constraints; 5913 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 5914 x86_pmu.extra_regs = intel_slm_extra_regs; 5915 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5916 td_attr = slm_events_attrs; 5917 extra_attr = slm_format_attr; 5918 pr_cont("Silvermont events, "); 5919 name = "silvermont"; 5920 break; 5921 5922 case INTEL_FAM6_ATOM_GOLDMONT: 5923 case INTEL_FAM6_ATOM_GOLDMONT_D: 5924 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids, 5925 sizeof(hw_cache_event_ids)); 5926 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs, 5927 sizeof(hw_cache_extra_regs)); 5928 5929 intel_pmu_lbr_init_skl(); 5930 5931 x86_pmu.event_constraints = intel_slm_event_constraints; 5932 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints; 5933 x86_pmu.extra_regs = intel_glm_extra_regs; 5934 /* 5935 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5936 * for precise cycles. 5937 * :pp is identical to :ppp 5938 */ 5939 x86_pmu.pebs_aliases = NULL; 5940 x86_pmu.pebs_prec_dist = true; 5941 x86_pmu.lbr_pt_coexist = true; 5942 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5943 td_attr = glm_events_attrs; 5944 extra_attr = slm_format_attr; 5945 pr_cont("Goldmont events, "); 5946 name = "goldmont"; 5947 break; 5948 5949 case INTEL_FAM6_ATOM_GOLDMONT_PLUS: 5950 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 5951 sizeof(hw_cache_event_ids)); 5952 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs, 5953 sizeof(hw_cache_extra_regs)); 5954 5955 intel_pmu_lbr_init_skl(); 5956 5957 x86_pmu.event_constraints = intel_slm_event_constraints; 5958 x86_pmu.extra_regs = intel_glm_extra_regs; 5959 /* 5960 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5961 * for precise cycles. 5962 */ 5963 x86_pmu.pebs_aliases = NULL; 5964 x86_pmu.pebs_prec_dist = true; 5965 x86_pmu.lbr_pt_coexist = true; 5966 x86_pmu.pebs_capable = ~0ULL; 5967 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5968 x86_pmu.flags |= PMU_FL_PEBS_ALL; 5969 x86_pmu.get_event_constraints = glp_get_event_constraints; 5970 td_attr = glm_events_attrs; 5971 /* Goldmont Plus has 4-wide pipeline */ 5972 event_attr_td_total_slots_scale_glm.event_str = "4"; 5973 extra_attr = slm_format_attr; 5974 pr_cont("Goldmont plus events, "); 5975 name = "goldmont_plus"; 5976 break; 5977 5978 case INTEL_FAM6_ATOM_TREMONT_D: 5979 case INTEL_FAM6_ATOM_TREMONT: 5980 case INTEL_FAM6_ATOM_TREMONT_L: 5981 x86_pmu.late_ack = true; 5982 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 5983 sizeof(hw_cache_event_ids)); 5984 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 5985 sizeof(hw_cache_extra_regs)); 5986 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 5987 5988 intel_pmu_lbr_init_skl(); 5989 5990 x86_pmu.event_constraints = intel_slm_event_constraints; 5991 x86_pmu.extra_regs = intel_tnt_extra_regs; 5992 /* 5993 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 5994 * for precise cycles. 5995 */ 5996 x86_pmu.pebs_aliases = NULL; 5997 x86_pmu.pebs_prec_dist = true; 5998 x86_pmu.lbr_pt_coexist = true; 5999 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6000 x86_pmu.get_event_constraints = tnt_get_event_constraints; 6001 td_attr = tnt_events_attrs; 6002 extra_attr = slm_format_attr; 6003 pr_cont("Tremont events, "); 6004 name = "Tremont"; 6005 break; 6006 6007 case INTEL_FAM6_ALDERLAKE_N: 6008 x86_pmu.mid_ack = true; 6009 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 6010 sizeof(hw_cache_event_ids)); 6011 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 6012 sizeof(hw_cache_extra_regs)); 6013 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6014 6015 x86_pmu.event_constraints = intel_slm_event_constraints; 6016 x86_pmu.pebs_constraints = intel_grt_pebs_event_constraints; 6017 x86_pmu.extra_regs = intel_grt_extra_regs; 6018 6019 x86_pmu.pebs_aliases = NULL; 6020 x86_pmu.pebs_prec_dist = true; 6021 x86_pmu.pebs_block = true; 6022 x86_pmu.lbr_pt_coexist = true; 6023 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6024 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6025 6026 intel_pmu_pebs_data_source_grt(); 6027 x86_pmu.pebs_latency_data = adl_latency_data_small; 6028 x86_pmu.get_event_constraints = tnt_get_event_constraints; 6029 x86_pmu.limit_period = spr_limit_period; 6030 td_attr = tnt_events_attrs; 6031 mem_attr = grt_mem_attrs; 6032 extra_attr = nhm_format_attr; 6033 pr_cont("Gracemont events, "); 6034 name = "gracemont"; 6035 break; 6036 6037 case INTEL_FAM6_WESTMERE: 6038 case INTEL_FAM6_WESTMERE_EP: 6039 case INTEL_FAM6_WESTMERE_EX: 6040 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, 6041 sizeof(hw_cache_event_ids)); 6042 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 6043 sizeof(hw_cache_extra_regs)); 6044 6045 intel_pmu_lbr_init_nhm(); 6046 6047 x86_pmu.event_constraints = intel_westmere_event_constraints; 6048 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 6049 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; 6050 x86_pmu.extra_regs = intel_westmere_extra_regs; 6051 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6052 6053 mem_attr = nhm_mem_events_attrs; 6054 6055 /* UOPS_ISSUED.STALLED_CYCLES */ 6056 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6057 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6058 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 6059 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 6060 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 6061 6062 intel_pmu_pebs_data_source_nhm(); 6063 extra_attr = nhm_format_attr; 6064 pr_cont("Westmere events, "); 6065 name = "westmere"; 6066 break; 6067 6068 case INTEL_FAM6_SANDYBRIDGE: 6069 case INTEL_FAM6_SANDYBRIDGE_X: 6070 x86_add_quirk(intel_sandybridge_quirk); 6071 x86_add_quirk(intel_ht_bug); 6072 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 6073 sizeof(hw_cache_event_ids)); 6074 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 6075 sizeof(hw_cache_extra_regs)); 6076 6077 intel_pmu_lbr_init_snb(); 6078 6079 x86_pmu.event_constraints = intel_snb_event_constraints; 6080 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; 6081 x86_pmu.pebs_aliases = intel_pebs_aliases_snb; 6082 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X) 6083 x86_pmu.extra_regs = intel_snbep_extra_regs; 6084 else 6085 x86_pmu.extra_regs = intel_snb_extra_regs; 6086 6087 6088 /* all extra regs are per-cpu when HT is on */ 6089 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6090 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6091 6092 td_attr = snb_events_attrs; 6093 mem_attr = snb_mem_events_attrs; 6094 6095 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 6096 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6097 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6098 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ 6099 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 6100 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); 6101 6102 extra_attr = nhm_format_attr; 6103 6104 pr_cont("SandyBridge events, "); 6105 name = "sandybridge"; 6106 break; 6107 6108 case INTEL_FAM6_IVYBRIDGE: 6109 case INTEL_FAM6_IVYBRIDGE_X: 6110 x86_add_quirk(intel_ht_bug); 6111 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 6112 sizeof(hw_cache_event_ids)); 6113 /* dTLB-load-misses on IVB is different than SNB */ 6114 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ 6115 6116 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 6117 sizeof(hw_cache_extra_regs)); 6118 6119 intel_pmu_lbr_init_snb(); 6120 6121 x86_pmu.event_constraints = intel_ivb_event_constraints; 6122 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; 6123 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6124 x86_pmu.pebs_prec_dist = true; 6125 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X) 6126 x86_pmu.extra_regs = intel_snbep_extra_regs; 6127 else 6128 x86_pmu.extra_regs = intel_snb_extra_regs; 6129 /* all extra regs are per-cpu when HT is on */ 6130 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6131 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6132 6133 td_attr = snb_events_attrs; 6134 mem_attr = snb_mem_events_attrs; 6135 6136 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 6137 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 6138 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 6139 6140 extra_attr = nhm_format_attr; 6141 6142 pr_cont("IvyBridge events, "); 6143 name = "ivybridge"; 6144 break; 6145 6146 6147 case INTEL_FAM6_HASWELL: 6148 case INTEL_FAM6_HASWELL_X: 6149 case INTEL_FAM6_HASWELL_L: 6150 case INTEL_FAM6_HASWELL_G: 6151 x86_add_quirk(intel_ht_bug); 6152 x86_add_quirk(intel_pebs_isolation_quirk); 6153 x86_pmu.late_ack = true; 6154 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6155 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6156 6157 intel_pmu_lbr_init_hsw(); 6158 6159 x86_pmu.event_constraints = intel_hsw_event_constraints; 6160 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; 6161 x86_pmu.extra_regs = intel_snbep_extra_regs; 6162 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6163 x86_pmu.pebs_prec_dist = true; 6164 /* all extra regs are per-cpu when HT is on */ 6165 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6166 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6167 6168 x86_pmu.hw_config = hsw_hw_config; 6169 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6170 x86_pmu.lbr_double_abort = true; 6171 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6172 hsw_format_attr : nhm_format_attr; 6173 td_attr = hsw_events_attrs; 6174 mem_attr = hsw_mem_events_attrs; 6175 tsx_attr = hsw_tsx_events_attrs; 6176 pr_cont("Haswell events, "); 6177 name = "haswell"; 6178 break; 6179 6180 case INTEL_FAM6_BROADWELL: 6181 case INTEL_FAM6_BROADWELL_D: 6182 case INTEL_FAM6_BROADWELL_G: 6183 case INTEL_FAM6_BROADWELL_X: 6184 x86_add_quirk(intel_pebs_isolation_quirk); 6185 x86_pmu.late_ack = true; 6186 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6187 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6188 6189 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ 6190 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | 6191 BDW_L3_MISS|HSW_SNOOP_DRAM; 6192 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| 6193 HSW_SNOOP_DRAM; 6194 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| 6195 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6196 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| 6197 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 6198 6199 intel_pmu_lbr_init_hsw(); 6200 6201 x86_pmu.event_constraints = intel_bdw_event_constraints; 6202 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints; 6203 x86_pmu.extra_regs = intel_snbep_extra_regs; 6204 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 6205 x86_pmu.pebs_prec_dist = true; 6206 /* all extra regs are per-cpu when HT is on */ 6207 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6208 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6209 6210 x86_pmu.hw_config = hsw_hw_config; 6211 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6212 x86_pmu.limit_period = bdw_limit_period; 6213 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6214 hsw_format_attr : nhm_format_attr; 6215 td_attr = hsw_events_attrs; 6216 mem_attr = hsw_mem_events_attrs; 6217 tsx_attr = hsw_tsx_events_attrs; 6218 pr_cont("Broadwell events, "); 6219 name = "broadwell"; 6220 break; 6221 6222 case INTEL_FAM6_XEON_PHI_KNL: 6223 case INTEL_FAM6_XEON_PHI_KNM: 6224 memcpy(hw_cache_event_ids, 6225 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6226 memcpy(hw_cache_extra_regs, 6227 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6228 intel_pmu_lbr_init_knl(); 6229 6230 x86_pmu.event_constraints = intel_slm_event_constraints; 6231 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 6232 x86_pmu.extra_regs = intel_knl_extra_regs; 6233 6234 /* all extra regs are per-cpu when HT is on */ 6235 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6236 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6237 extra_attr = slm_format_attr; 6238 pr_cont("Knights Landing/Mill events, "); 6239 name = "knights-landing"; 6240 break; 6241 6242 case INTEL_FAM6_SKYLAKE_X: 6243 pmem = true; 6244 fallthrough; 6245 case INTEL_FAM6_SKYLAKE_L: 6246 case INTEL_FAM6_SKYLAKE: 6247 case INTEL_FAM6_KABYLAKE_L: 6248 case INTEL_FAM6_KABYLAKE: 6249 case INTEL_FAM6_COMETLAKE_L: 6250 case INTEL_FAM6_COMETLAKE: 6251 x86_add_quirk(intel_pebs_isolation_quirk); 6252 x86_pmu.late_ack = true; 6253 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6254 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6255 intel_pmu_lbr_init_skl(); 6256 6257 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */ 6258 event_attr_td_recovery_bubbles.event_str_noht = 6259 "event=0xd,umask=0x1,cmask=1"; 6260 event_attr_td_recovery_bubbles.event_str_ht = 6261 "event=0xd,umask=0x1,cmask=1,any=1"; 6262 6263 x86_pmu.event_constraints = intel_skl_event_constraints; 6264 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; 6265 x86_pmu.extra_regs = intel_skl_extra_regs; 6266 x86_pmu.pebs_aliases = intel_pebs_aliases_skl; 6267 x86_pmu.pebs_prec_dist = true; 6268 /* all extra regs are per-cpu when HT is on */ 6269 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6270 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6271 6272 x86_pmu.hw_config = hsw_hw_config; 6273 x86_pmu.get_event_constraints = hsw_get_event_constraints; 6274 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6275 hsw_format_attr : nhm_format_attr; 6276 extra_skl_attr = skl_format_attr; 6277 td_attr = hsw_events_attrs; 6278 mem_attr = hsw_mem_events_attrs; 6279 tsx_attr = hsw_tsx_events_attrs; 6280 intel_pmu_pebs_data_source_skl(pmem); 6281 6282 /* 6283 * Processors with CPUID.RTM_ALWAYS_ABORT have TSX deprecated by default. 6284 * TSX force abort hooks are not required on these systems. Only deploy 6285 * workaround when microcode has not enabled X86_FEATURE_RTM_ALWAYS_ABORT. 6286 */ 6287 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT) && 6288 !boot_cpu_has(X86_FEATURE_RTM_ALWAYS_ABORT)) { 6289 x86_pmu.flags |= PMU_FL_TFA; 6290 x86_pmu.get_event_constraints = tfa_get_event_constraints; 6291 x86_pmu.enable_all = intel_tfa_pmu_enable_all; 6292 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; 6293 } 6294 6295 pr_cont("Skylake events, "); 6296 name = "skylake"; 6297 break; 6298 6299 case INTEL_FAM6_ICELAKE_X: 6300 case INTEL_FAM6_ICELAKE_D: 6301 x86_pmu.pebs_ept = 1; 6302 pmem = true; 6303 fallthrough; 6304 case INTEL_FAM6_ICELAKE_L: 6305 case INTEL_FAM6_ICELAKE: 6306 case INTEL_FAM6_TIGERLAKE_L: 6307 case INTEL_FAM6_TIGERLAKE: 6308 case INTEL_FAM6_ROCKETLAKE: 6309 x86_pmu.late_ack = true; 6310 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6311 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6312 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6313 intel_pmu_lbr_init_skl(); 6314 6315 x86_pmu.event_constraints = intel_icl_event_constraints; 6316 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints; 6317 x86_pmu.extra_regs = intel_icl_extra_regs; 6318 x86_pmu.pebs_aliases = NULL; 6319 x86_pmu.pebs_prec_dist = true; 6320 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6321 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6322 6323 x86_pmu.hw_config = hsw_hw_config; 6324 x86_pmu.get_event_constraints = icl_get_event_constraints; 6325 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6326 hsw_format_attr : nhm_format_attr; 6327 extra_skl_attr = skl_format_attr; 6328 mem_attr = icl_events_attrs; 6329 td_attr = icl_td_events_attrs; 6330 tsx_attr = icl_tsx_events_attrs; 6331 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6332 x86_pmu.lbr_pt_coexist = true; 6333 intel_pmu_pebs_data_source_skl(pmem); 6334 x86_pmu.num_topdown_events = 4; 6335 static_call_update(intel_pmu_update_topdown_event, 6336 &icl_update_topdown_event); 6337 static_call_update(intel_pmu_set_topdown_event_period, 6338 &icl_set_topdown_event_period); 6339 pr_cont("Icelake events, "); 6340 name = "icelake"; 6341 break; 6342 6343 case INTEL_FAM6_SAPPHIRERAPIDS_X: 6344 pmem = true; 6345 x86_pmu.late_ack = true; 6346 memcpy(hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 6347 memcpy(hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 6348 6349 x86_pmu.event_constraints = intel_spr_event_constraints; 6350 x86_pmu.pebs_constraints = intel_spr_pebs_event_constraints; 6351 x86_pmu.extra_regs = intel_spr_extra_regs; 6352 x86_pmu.limit_period = spr_limit_period; 6353 x86_pmu.pebs_aliases = NULL; 6354 x86_pmu.pebs_prec_dist = true; 6355 x86_pmu.pebs_block = true; 6356 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6357 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6358 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6359 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6360 6361 x86_pmu.hw_config = hsw_hw_config; 6362 x86_pmu.get_event_constraints = spr_get_event_constraints; 6363 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6364 hsw_format_attr : nhm_format_attr; 6365 extra_skl_attr = skl_format_attr; 6366 mem_attr = spr_events_attrs; 6367 td_attr = spr_td_events_attrs; 6368 tsx_attr = spr_tsx_events_attrs; 6369 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6370 x86_pmu.lbr_pt_coexist = true; 6371 intel_pmu_pebs_data_source_skl(pmem); 6372 x86_pmu.num_topdown_events = 8; 6373 static_call_update(intel_pmu_update_topdown_event, 6374 &icl_update_topdown_event); 6375 static_call_update(intel_pmu_set_topdown_event_period, 6376 &icl_set_topdown_event_period); 6377 pr_cont("Sapphire Rapids events, "); 6378 name = "sapphire_rapids"; 6379 break; 6380 6381 case INTEL_FAM6_ALDERLAKE: 6382 case INTEL_FAM6_ALDERLAKE_L: 6383 case INTEL_FAM6_RAPTORLAKE: 6384 case INTEL_FAM6_RAPTORLAKE_P: 6385 case INTEL_FAM6_RAPTORLAKE_S: 6386 /* 6387 * Alder Lake has 2 types of CPU, core and atom. 6388 * 6389 * Initialize the common PerfMon capabilities here. 6390 */ 6391 x86_pmu.hybrid_pmu = kcalloc(X86_HYBRID_NUM_PMUS, 6392 sizeof(struct x86_hybrid_pmu), 6393 GFP_KERNEL); 6394 if (!x86_pmu.hybrid_pmu) 6395 return -ENOMEM; 6396 static_branch_enable(&perf_is_hybrid); 6397 x86_pmu.num_hybrid_pmus = X86_HYBRID_NUM_PMUS; 6398 6399 x86_pmu.pebs_aliases = NULL; 6400 x86_pmu.pebs_prec_dist = true; 6401 x86_pmu.pebs_block = true; 6402 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 6403 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 6404 x86_pmu.flags |= PMU_FL_INSTR_LATENCY; 6405 x86_pmu.flags |= PMU_FL_MEM_LOADS_AUX; 6406 x86_pmu.lbr_pt_coexist = true; 6407 intel_pmu_pebs_data_source_adl(); 6408 x86_pmu.pebs_latency_data = adl_latency_data_small; 6409 x86_pmu.num_topdown_events = 8; 6410 static_call_update(intel_pmu_update_topdown_event, 6411 &adl_update_topdown_event); 6412 static_call_update(intel_pmu_set_topdown_event_period, 6413 &adl_set_topdown_event_period); 6414 6415 x86_pmu.filter_match = intel_pmu_filter_match; 6416 x86_pmu.get_event_constraints = adl_get_event_constraints; 6417 x86_pmu.hw_config = adl_hw_config; 6418 x86_pmu.limit_period = spr_limit_period; 6419 x86_pmu.get_hybrid_cpu_type = adl_get_hybrid_cpu_type; 6420 /* 6421 * The rtm_abort_event is used to check whether to enable GPRs 6422 * for the RTM abort event. Atom doesn't have the RTM abort 6423 * event. There is no harmful to set it in the common 6424 * x86_pmu.rtm_abort_event. 6425 */ 6426 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xc9, .umask=0x04); 6427 6428 td_attr = adl_hybrid_events_attrs; 6429 mem_attr = adl_hybrid_mem_attrs; 6430 tsx_attr = adl_hybrid_tsx_attrs; 6431 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 6432 adl_hybrid_extra_attr_rtm : adl_hybrid_extra_attr; 6433 6434 /* Initialize big core specific PerfMon capabilities.*/ 6435 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_CORE_IDX]; 6436 pmu->name = "cpu_core"; 6437 pmu->cpu_type = hybrid_big; 6438 pmu->late_ack = true; 6439 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) { 6440 pmu->num_counters = x86_pmu.num_counters + 2; 6441 pmu->num_counters_fixed = x86_pmu.num_counters_fixed + 1; 6442 } else { 6443 pmu->num_counters = x86_pmu.num_counters; 6444 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6445 } 6446 6447 /* 6448 * Quirk: For some Alder Lake machine, when all E-cores are disabled in 6449 * a BIOS, the leaf 0xA will enumerate all counters of P-cores. However, 6450 * the X86_FEATURE_HYBRID_CPU is still set. The above codes will 6451 * mistakenly add extra counters for P-cores. Correct the number of 6452 * counters here. 6453 */ 6454 if ((pmu->num_counters > 8) || (pmu->num_counters_fixed > 4)) { 6455 pmu->num_counters = x86_pmu.num_counters; 6456 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6457 } 6458 6459 pmu->max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, pmu->num_counters); 6460 pmu->unconstrained = (struct event_constraint) 6461 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6462 0, pmu->num_counters, 0, 0); 6463 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; 6464 pmu->intel_cap.perf_metrics = 1; 6465 pmu->intel_cap.pebs_output_pt_available = 0; 6466 6467 memcpy(pmu->hw_cache_event_ids, spr_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids)); 6468 memcpy(pmu->hw_cache_extra_regs, spr_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs)); 6469 pmu->event_constraints = intel_spr_event_constraints; 6470 pmu->pebs_constraints = intel_spr_pebs_event_constraints; 6471 pmu->extra_regs = intel_spr_extra_regs; 6472 6473 /* Initialize Atom core specific PerfMon capabilities.*/ 6474 pmu = &x86_pmu.hybrid_pmu[X86_HYBRID_PMU_ATOM_IDX]; 6475 pmu->name = "cpu_atom"; 6476 pmu->cpu_type = hybrid_small; 6477 pmu->mid_ack = true; 6478 pmu->num_counters = x86_pmu.num_counters; 6479 pmu->num_counters_fixed = x86_pmu.num_counters_fixed; 6480 pmu->max_pebs_events = x86_pmu.max_pebs_events; 6481 pmu->unconstrained = (struct event_constraint) 6482 __EVENT_CONSTRAINT(0, (1ULL << pmu->num_counters) - 1, 6483 0, pmu->num_counters, 0, 0); 6484 pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; 6485 pmu->intel_cap.perf_metrics = 0; 6486 pmu->intel_cap.pebs_output_pt_available = 1; 6487 6488 memcpy(pmu->hw_cache_event_ids, glp_hw_cache_event_ids, sizeof(pmu->hw_cache_event_ids)); 6489 memcpy(pmu->hw_cache_extra_regs, tnt_hw_cache_extra_regs, sizeof(pmu->hw_cache_extra_regs)); 6490 pmu->hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 6491 pmu->event_constraints = intel_slm_event_constraints; 6492 pmu->pebs_constraints = intel_grt_pebs_event_constraints; 6493 pmu->extra_regs = intel_grt_extra_regs; 6494 pr_cont("Alderlake Hybrid events, "); 6495 name = "alderlake_hybrid"; 6496 break; 6497 6498 default: 6499 switch (x86_pmu.version) { 6500 case 1: 6501 x86_pmu.event_constraints = intel_v1_event_constraints; 6502 pr_cont("generic architected perfmon v1, "); 6503 name = "generic_arch_v1"; 6504 break; 6505 case 2: 6506 case 3: 6507 case 4: 6508 /* 6509 * default constraints for v2 and up 6510 */ 6511 x86_pmu.event_constraints = intel_gen_event_constraints; 6512 pr_cont("generic architected perfmon, "); 6513 name = "generic_arch_v2+"; 6514 break; 6515 default: 6516 /* 6517 * The default constraints for v5 and up can support up to 6518 * 16 fixed counters. For the fixed counters 4 and later, 6519 * the pseudo-encoding is applied. 6520 * The constraints may be cut according to the CPUID enumeration 6521 * by inserting the EVENT_CONSTRAINT_END. 6522 */ 6523 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) 6524 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; 6525 intel_v5_gen_event_constraints[x86_pmu.num_counters_fixed].weight = -1; 6526 x86_pmu.event_constraints = intel_v5_gen_event_constraints; 6527 pr_cont("generic architected perfmon, "); 6528 name = "generic_arch_v5+"; 6529 break; 6530 } 6531 } 6532 6533 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name); 6534 6535 if (!is_hybrid()) { 6536 group_events_td.attrs = td_attr; 6537 group_events_mem.attrs = mem_attr; 6538 group_events_tsx.attrs = tsx_attr; 6539 group_format_extra.attrs = extra_attr; 6540 group_format_extra_skl.attrs = extra_skl_attr; 6541 6542 x86_pmu.attr_update = attr_update; 6543 } else { 6544 hybrid_group_events_td.attrs = td_attr; 6545 hybrid_group_events_mem.attrs = mem_attr; 6546 hybrid_group_events_tsx.attrs = tsx_attr; 6547 hybrid_group_format_extra.attrs = extra_attr; 6548 6549 x86_pmu.attr_update = hybrid_attr_update; 6550 } 6551 6552 intel_pmu_check_num_counters(&x86_pmu.num_counters, 6553 &x86_pmu.num_counters_fixed, 6554 &x86_pmu.intel_ctrl, 6555 (u64)fixed_mask); 6556 6557 /* AnyThread may be deprecated on arch perfmon v5 or later */ 6558 if (x86_pmu.intel_cap.anythread_deprecated) 6559 x86_pmu.format_attrs = intel_arch_formats_attr; 6560 6561 intel_pmu_check_event_constraints(x86_pmu.event_constraints, 6562 x86_pmu.num_counters, 6563 x86_pmu.num_counters_fixed, 6564 x86_pmu.intel_ctrl); 6565 /* 6566 * Access LBR MSR may cause #GP under certain circumstances. 6567 * Check all LBR MSR here. 6568 * Disable LBR access if any LBR MSRs can not be accessed. 6569 */ 6570 if (x86_pmu.lbr_tos && !check_msr(x86_pmu.lbr_tos, 0x3UL)) 6571 x86_pmu.lbr_nr = 0; 6572 for (i = 0; i < x86_pmu.lbr_nr; i++) { 6573 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && 6574 check_msr(x86_pmu.lbr_to + i, 0xffffUL))) 6575 x86_pmu.lbr_nr = 0; 6576 } 6577 6578 if (x86_pmu.lbr_nr) { 6579 intel_pmu_lbr_init(); 6580 6581 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr); 6582 6583 /* only support branch_stack snapshot for perfmon >= v2 */ 6584 if (x86_pmu.disable_all == intel_pmu_disable_all) { 6585 if (boot_cpu_has(X86_FEATURE_ARCH_LBR)) { 6586 static_call_update(perf_snapshot_branch_stack, 6587 intel_pmu_snapshot_arch_branch_stack); 6588 } else { 6589 static_call_update(perf_snapshot_branch_stack, 6590 intel_pmu_snapshot_branch_stack); 6591 } 6592 } 6593 } 6594 6595 intel_pmu_check_extra_regs(x86_pmu.extra_regs); 6596 6597 /* Support full width counters using alternative MSR range */ 6598 if (x86_pmu.intel_cap.full_width_write) { 6599 x86_pmu.max_period = x86_pmu.cntval_mask >> 1; 6600 x86_pmu.perfctr = MSR_IA32_PMC0; 6601 pr_cont("full-width counters, "); 6602 } 6603 6604 if (!is_hybrid() && x86_pmu.intel_cap.perf_metrics) 6605 x86_pmu.intel_ctrl |= 1ULL << GLOBAL_CTRL_EN_PERF_METRICS; 6606 6607 if (is_hybrid()) 6608 intel_pmu_check_hybrid_pmus((u64)fixed_mask); 6609 6610 intel_aux_output_init(); 6611 6612 return 0; 6613 } 6614 6615 /* 6616 * HT bug: phase 2 init 6617 * Called once we have valid topology information to check 6618 * whether or not HT is enabled 6619 * If HT is off, then we disable the workaround 6620 */ 6621 static __init int fixup_ht_bug(void) 6622 { 6623 int c; 6624 /* 6625 * problem not present on this CPU model, nothing to do 6626 */ 6627 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) 6628 return 0; 6629 6630 if (topology_max_smt_threads() > 1) { 6631 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); 6632 return 0; 6633 } 6634 6635 cpus_read_lock(); 6636 6637 hardlockup_detector_perf_stop(); 6638 6639 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); 6640 6641 x86_pmu.start_scheduling = NULL; 6642 x86_pmu.commit_scheduling = NULL; 6643 x86_pmu.stop_scheduling = NULL; 6644 6645 hardlockup_detector_perf_restart(); 6646 6647 for_each_online_cpu(c) 6648 free_excl_cntrs(&per_cpu(cpu_hw_events, c)); 6649 6650 cpus_read_unlock(); 6651 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); 6652 return 0; 6653 } 6654 subsys_initcall(fixup_ht_bug) 6655