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