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/apic.h> 22 #include <asm/cpu_device_id.h> 23 24 #include "../perf_event.h" 25 26 /* 27 * Intel PerfMon, used on Core and later. 28 */ 29 static u64 intel_perfmon_event_map[PERF_COUNT_HW_MAX] __read_mostly = 30 { 31 [PERF_COUNT_HW_CPU_CYCLES] = 0x003c, 32 [PERF_COUNT_HW_INSTRUCTIONS] = 0x00c0, 33 [PERF_COUNT_HW_CACHE_REFERENCES] = 0x4f2e, 34 [PERF_COUNT_HW_CACHE_MISSES] = 0x412e, 35 [PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x00c4, 36 [PERF_COUNT_HW_BRANCH_MISSES] = 0x00c5, 37 [PERF_COUNT_HW_BUS_CYCLES] = 0x013c, 38 [PERF_COUNT_HW_REF_CPU_CYCLES] = 0x0300, /* pseudo-encoding */ 39 }; 40 41 static struct event_constraint intel_core_event_constraints[] __read_mostly = 42 { 43 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 44 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 45 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 46 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 47 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 48 INTEL_EVENT_CONSTRAINT(0xc1, 0x1), /* FP_COMP_INSTR_RET */ 49 EVENT_CONSTRAINT_END 50 }; 51 52 static struct event_constraint intel_core2_event_constraints[] __read_mostly = 53 { 54 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 55 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 56 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 57 INTEL_EVENT_CONSTRAINT(0x10, 0x1), /* FP_COMP_OPS_EXE */ 58 INTEL_EVENT_CONSTRAINT(0x11, 0x2), /* FP_ASSIST */ 59 INTEL_EVENT_CONSTRAINT(0x12, 0x2), /* MUL */ 60 INTEL_EVENT_CONSTRAINT(0x13, 0x2), /* DIV */ 61 INTEL_EVENT_CONSTRAINT(0x14, 0x1), /* CYCLES_DIV_BUSY */ 62 INTEL_EVENT_CONSTRAINT(0x18, 0x1), /* IDLE_DURING_DIV */ 63 INTEL_EVENT_CONSTRAINT(0x19, 0x2), /* DELAYED_BYPASS */ 64 INTEL_EVENT_CONSTRAINT(0xa1, 0x1), /* RS_UOPS_DISPATCH_CYCLES */ 65 INTEL_EVENT_CONSTRAINT(0xc9, 0x1), /* ITLB_MISS_RETIRED (T30-9) */ 66 INTEL_EVENT_CONSTRAINT(0xcb, 0x1), /* MEM_LOAD_RETIRED */ 67 EVENT_CONSTRAINT_END 68 }; 69 70 static struct event_constraint intel_nehalem_event_constraints[] __read_mostly = 71 { 72 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 73 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 74 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 75 INTEL_EVENT_CONSTRAINT(0x40, 0x3), /* L1D_CACHE_LD */ 76 INTEL_EVENT_CONSTRAINT(0x41, 0x3), /* L1D_CACHE_ST */ 77 INTEL_EVENT_CONSTRAINT(0x42, 0x3), /* L1D_CACHE_LOCK */ 78 INTEL_EVENT_CONSTRAINT(0x43, 0x3), /* L1D_ALL_REF */ 79 INTEL_EVENT_CONSTRAINT(0x48, 0x3), /* L1D_PEND_MISS */ 80 INTEL_EVENT_CONSTRAINT(0x4e, 0x3), /* L1D_PREFETCH */ 81 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 82 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 83 EVENT_CONSTRAINT_END 84 }; 85 86 static struct extra_reg intel_nehalem_extra_regs[] __read_mostly = 87 { 88 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 89 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 90 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 91 EVENT_EXTRA_END 92 }; 93 94 static struct event_constraint intel_westmere_event_constraints[] __read_mostly = 95 { 96 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 97 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 98 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 99 INTEL_EVENT_CONSTRAINT(0x51, 0x3), /* L1D */ 100 INTEL_EVENT_CONSTRAINT(0x60, 0x1), /* OFFCORE_REQUESTS_OUTSTANDING */ 101 INTEL_EVENT_CONSTRAINT(0x63, 0x3), /* CACHE_LOCK_CYCLES */ 102 INTEL_EVENT_CONSTRAINT(0xb3, 0x1), /* SNOOPQ_REQUEST_OUTSTANDING */ 103 EVENT_CONSTRAINT_END 104 }; 105 106 static struct event_constraint intel_snb_event_constraints[] __read_mostly = 107 { 108 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 109 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 110 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 111 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 112 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 113 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 114 INTEL_UEVENT_CONSTRAINT(0x06a3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 115 INTEL_EVENT_CONSTRAINT(0x48, 0x4), /* L1D_PEND_MISS.PENDING */ 116 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 117 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 118 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_DISPATCH */ 119 INTEL_UEVENT_CONSTRAINT(0x02a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 120 121 /* 122 * When HT is off these events can only run on the bottom 4 counters 123 * When HT is on, they are impacted by the HT bug and require EXCL access 124 */ 125 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 126 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 127 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 128 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 129 130 EVENT_CONSTRAINT_END 131 }; 132 133 static struct event_constraint intel_ivb_event_constraints[] __read_mostly = 134 { 135 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 136 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 137 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 138 INTEL_UEVENT_CONSTRAINT(0x0148, 0x4), /* L1D_PEND_MISS.PENDING */ 139 INTEL_UEVENT_CONSTRAINT(0x0279, 0xf), /* IDQ.EMTPY */ 140 INTEL_UEVENT_CONSTRAINT(0x019c, 0xf), /* IDQ_UOPS_NOT_DELIVERED.CORE */ 141 INTEL_UEVENT_CONSTRAINT(0x02a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_LDM_PENDING */ 142 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 143 INTEL_UEVENT_CONSTRAINT(0x05a3, 0xf), /* CYCLE_ACTIVITY.STALLS_L2_PENDING */ 144 INTEL_UEVENT_CONSTRAINT(0x06a3, 0xf), /* CYCLE_ACTIVITY.STALLS_LDM_PENDING */ 145 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 146 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 147 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 148 149 /* 150 * When HT is off these events can only run on the bottom 4 counters 151 * When HT is on, they are impacted by the HT bug and require EXCL access 152 */ 153 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 154 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 155 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 156 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 157 158 EVENT_CONSTRAINT_END 159 }; 160 161 static struct extra_reg intel_westmere_extra_regs[] __read_mostly = 162 { 163 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 164 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffff, RSP_0), 165 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0xffff, RSP_1), 166 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x100b), 167 EVENT_EXTRA_END 168 }; 169 170 static struct event_constraint intel_v1_event_constraints[] __read_mostly = 171 { 172 EVENT_CONSTRAINT_END 173 }; 174 175 static struct event_constraint intel_gen_event_constraints[] __read_mostly = 176 { 177 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 178 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 179 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 180 EVENT_CONSTRAINT_END 181 }; 182 183 static struct event_constraint intel_slm_event_constraints[] __read_mostly = 184 { 185 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 186 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 187 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* pseudo CPU_CLK_UNHALTED.REF */ 188 EVENT_CONSTRAINT_END 189 }; 190 191 static struct event_constraint intel_skl_event_constraints[] = { 192 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 193 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 194 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 195 INTEL_UEVENT_CONSTRAINT(0x1c0, 0x2), /* INST_RETIRED.PREC_DIST */ 196 197 /* 198 * when HT is off, these can only run on the bottom 4 counters 199 */ 200 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 201 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 202 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 203 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 204 INTEL_EVENT_CONSTRAINT(0xc6, 0xf), /* FRONTEND_RETIRED.* */ 205 206 EVENT_CONSTRAINT_END 207 }; 208 209 static struct extra_reg intel_knl_extra_regs[] __read_mostly = { 210 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x799ffbb6e7ull, RSP_0), 211 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x399ffbffe7ull, RSP_1), 212 EVENT_EXTRA_END 213 }; 214 215 static struct extra_reg intel_snb_extra_regs[] __read_mostly = { 216 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 217 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3f807f8fffull, RSP_0), 218 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3f807f8fffull, RSP_1), 219 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 220 EVENT_EXTRA_END 221 }; 222 223 static struct extra_reg intel_snbep_extra_regs[] __read_mostly = { 224 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 225 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 226 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 227 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 228 EVENT_EXTRA_END 229 }; 230 231 static struct extra_reg intel_skl_extra_regs[] __read_mostly = { 232 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffff8fffull, RSP_0), 233 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff8fffull, RSP_1), 234 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 235 /* 236 * Note the low 8 bits eventsel code is not a continuous field, containing 237 * some #GPing bits. These are masked out. 238 */ 239 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 240 EVENT_EXTRA_END 241 }; 242 243 static struct event_constraint intel_icl_event_constraints[] = { 244 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 245 INTEL_UEVENT_CONSTRAINT(0x1c0, 0), /* INST_RETIRED.PREC_DIST */ 246 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 247 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 248 FIXED_EVENT_CONSTRAINT(0x0400, 3), /* SLOTS */ 249 INTEL_EVENT_CONSTRAINT_RANGE(0x03, 0x0a, 0xf), 250 INTEL_EVENT_CONSTRAINT_RANGE(0x1f, 0x28, 0xf), 251 INTEL_EVENT_CONSTRAINT(0x32, 0xf), /* SW_PREFETCH_ACCESS.* */ 252 INTEL_EVENT_CONSTRAINT_RANGE(0x48, 0x54, 0xf), 253 INTEL_EVENT_CONSTRAINT_RANGE(0x60, 0x8b, 0xf), 254 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xff), /* CYCLE_ACTIVITY.STALLS_TOTAL */ 255 INTEL_UEVENT_CONSTRAINT(0x10a3, 0xff), /* CYCLE_ACTIVITY.STALLS_MEM_ANY */ 256 INTEL_EVENT_CONSTRAINT(0xa3, 0xf), /* CYCLE_ACTIVITY.* */ 257 INTEL_EVENT_CONSTRAINT_RANGE(0xa8, 0xb0, 0xf), 258 INTEL_EVENT_CONSTRAINT_RANGE(0xb7, 0xbd, 0xf), 259 INTEL_EVENT_CONSTRAINT_RANGE(0xd0, 0xe6, 0xf), 260 INTEL_EVENT_CONSTRAINT_RANGE(0xf0, 0xf4, 0xf), 261 EVENT_CONSTRAINT_END 262 }; 263 264 static struct extra_reg intel_icl_extra_regs[] __read_mostly = { 265 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x3fffffbfffull, RSP_0), 266 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffffbfffull, RSP_1), 267 INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(0x01cd), 268 INTEL_UEVENT_EXTRA_REG(0x01c6, MSR_PEBS_FRONTEND, 0x7fff17, FE), 269 EVENT_EXTRA_END 270 }; 271 272 EVENT_ATTR_STR(mem-loads, mem_ld_nhm, "event=0x0b,umask=0x10,ldlat=3"); 273 EVENT_ATTR_STR(mem-loads, mem_ld_snb, "event=0xcd,umask=0x1,ldlat=3"); 274 EVENT_ATTR_STR(mem-stores, mem_st_snb, "event=0xcd,umask=0x2"); 275 276 static struct attribute *nhm_mem_events_attrs[] = { 277 EVENT_PTR(mem_ld_nhm), 278 NULL, 279 }; 280 281 /* 282 * topdown events for Intel Core CPUs. 283 * 284 * The events are all in slots, which is a free slot in a 4 wide 285 * pipeline. Some events are already reported in slots, for cycle 286 * events we multiply by the pipeline width (4). 287 * 288 * With Hyper Threading on, topdown metrics are either summed or averaged 289 * between the threads of a core: (count_t0 + count_t1). 290 * 291 * For the average case the metric is always scaled to pipeline width, 292 * so we use factor 2 ((count_t0 + count_t1) / 2 * 4) 293 */ 294 295 EVENT_ATTR_STR_HT(topdown-total-slots, td_total_slots, 296 "event=0x3c,umask=0x0", /* cpu_clk_unhalted.thread */ 297 "event=0x3c,umask=0x0,any=1"); /* cpu_clk_unhalted.thread_any */ 298 EVENT_ATTR_STR_HT(topdown-total-slots.scale, td_total_slots_scale, "4", "2"); 299 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued, 300 "event=0xe,umask=0x1"); /* uops_issued.any */ 301 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired, 302 "event=0xc2,umask=0x2"); /* uops_retired.retire_slots */ 303 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles, 304 "event=0x9c,umask=0x1"); /* idq_uops_not_delivered_core */ 305 EVENT_ATTR_STR_HT(topdown-recovery-bubbles, td_recovery_bubbles, 306 "event=0xd,umask=0x3,cmask=1", /* int_misc.recovery_cycles */ 307 "event=0xd,umask=0x3,cmask=1,any=1"); /* int_misc.recovery_cycles_any */ 308 EVENT_ATTR_STR_HT(topdown-recovery-bubbles.scale, td_recovery_bubbles_scale, 309 "4", "2"); 310 311 static struct attribute *snb_events_attrs[] = { 312 EVENT_PTR(td_slots_issued), 313 EVENT_PTR(td_slots_retired), 314 EVENT_PTR(td_fetch_bubbles), 315 EVENT_PTR(td_total_slots), 316 EVENT_PTR(td_total_slots_scale), 317 EVENT_PTR(td_recovery_bubbles), 318 EVENT_PTR(td_recovery_bubbles_scale), 319 NULL, 320 }; 321 322 static struct attribute *snb_mem_events_attrs[] = { 323 EVENT_PTR(mem_ld_snb), 324 EVENT_PTR(mem_st_snb), 325 NULL, 326 }; 327 328 static struct event_constraint intel_hsw_event_constraints[] = { 329 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 330 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 331 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 332 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 333 INTEL_UEVENT_CONSTRAINT(0x01c0, 0x2), /* INST_RETIRED.PREC_DIST */ 334 INTEL_EVENT_CONSTRAINT(0xcd, 0x8), /* MEM_TRANS_RETIRED.LOAD_LATENCY */ 335 /* CYCLE_ACTIVITY.CYCLES_L1D_PENDING */ 336 INTEL_UEVENT_CONSTRAINT(0x08a3, 0x4), 337 /* CYCLE_ACTIVITY.STALLS_L1D_PENDING */ 338 INTEL_UEVENT_CONSTRAINT(0x0ca3, 0x4), 339 /* CYCLE_ACTIVITY.CYCLES_NO_EXECUTE */ 340 INTEL_UEVENT_CONSTRAINT(0x04a3, 0xf), 341 342 /* 343 * When HT is off these events can only run on the bottom 4 counters 344 * When HT is on, they are impacted by the HT bug and require EXCL access 345 */ 346 INTEL_EXCLEVT_CONSTRAINT(0xd0, 0xf), /* MEM_UOPS_RETIRED.* */ 347 INTEL_EXCLEVT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_UOPS_RETIRED.* */ 348 INTEL_EXCLEVT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_UOPS_LLC_HIT_RETIRED.* */ 349 INTEL_EXCLEVT_CONSTRAINT(0xd3, 0xf), /* MEM_LOAD_UOPS_LLC_MISS_RETIRED.* */ 350 351 EVENT_CONSTRAINT_END 352 }; 353 354 static struct event_constraint intel_bdw_event_constraints[] = { 355 FIXED_EVENT_CONSTRAINT(0x00c0, 0), /* INST_RETIRED.ANY */ 356 FIXED_EVENT_CONSTRAINT(0x003c, 1), /* CPU_CLK_UNHALTED.CORE */ 357 FIXED_EVENT_CONSTRAINT(0x0300, 2), /* CPU_CLK_UNHALTED.REF */ 358 INTEL_UEVENT_CONSTRAINT(0x148, 0x4), /* L1D_PEND_MISS.PENDING */ 359 INTEL_UBIT_EVENT_CONSTRAINT(0x8a3, 0x4), /* CYCLE_ACTIVITY.CYCLES_L1D_MISS */ 360 /* 361 * when HT is off, these can only run on the bottom 4 counters 362 */ 363 INTEL_EVENT_CONSTRAINT(0xd0, 0xf), /* MEM_INST_RETIRED.* */ 364 INTEL_EVENT_CONSTRAINT(0xd1, 0xf), /* MEM_LOAD_RETIRED.* */ 365 INTEL_EVENT_CONSTRAINT(0xd2, 0xf), /* MEM_LOAD_L3_HIT_RETIRED.* */ 366 INTEL_EVENT_CONSTRAINT(0xcd, 0xf), /* MEM_TRANS_RETIRED.* */ 367 EVENT_CONSTRAINT_END 368 }; 369 370 static u64 intel_pmu_event_map(int hw_event) 371 { 372 return intel_perfmon_event_map[hw_event]; 373 } 374 375 /* 376 * Notes on the events: 377 * - data reads do not include code reads (comparable to earlier tables) 378 * - data counts include speculative execution (except L1 write, dtlb, bpu) 379 * - remote node access includes remote memory, remote cache, remote mmio. 380 * - prefetches are not included in the counts. 381 * - icache miss does not include decoded icache 382 */ 383 384 #define SKL_DEMAND_DATA_RD BIT_ULL(0) 385 #define SKL_DEMAND_RFO BIT_ULL(1) 386 #define SKL_ANY_RESPONSE BIT_ULL(16) 387 #define SKL_SUPPLIER_NONE BIT_ULL(17) 388 #define SKL_L3_MISS_LOCAL_DRAM BIT_ULL(26) 389 #define SKL_L3_MISS_REMOTE_HOP0_DRAM BIT_ULL(27) 390 #define SKL_L3_MISS_REMOTE_HOP1_DRAM BIT_ULL(28) 391 #define SKL_L3_MISS_REMOTE_HOP2P_DRAM BIT_ULL(29) 392 #define SKL_L3_MISS (SKL_L3_MISS_LOCAL_DRAM| \ 393 SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 394 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 395 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 396 #define SKL_SPL_HIT BIT_ULL(30) 397 #define SKL_SNOOP_NONE BIT_ULL(31) 398 #define SKL_SNOOP_NOT_NEEDED BIT_ULL(32) 399 #define SKL_SNOOP_MISS BIT_ULL(33) 400 #define SKL_SNOOP_HIT_NO_FWD BIT_ULL(34) 401 #define SKL_SNOOP_HIT_WITH_FWD BIT_ULL(35) 402 #define SKL_SNOOP_HITM BIT_ULL(36) 403 #define SKL_SNOOP_NON_DRAM BIT_ULL(37) 404 #define SKL_ANY_SNOOP (SKL_SPL_HIT|SKL_SNOOP_NONE| \ 405 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 406 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 407 SKL_SNOOP_HITM|SKL_SNOOP_NON_DRAM) 408 #define SKL_DEMAND_READ SKL_DEMAND_DATA_RD 409 #define SKL_SNOOP_DRAM (SKL_SNOOP_NONE| \ 410 SKL_SNOOP_NOT_NEEDED|SKL_SNOOP_MISS| \ 411 SKL_SNOOP_HIT_NO_FWD|SKL_SNOOP_HIT_WITH_FWD| \ 412 SKL_SNOOP_HITM|SKL_SPL_HIT) 413 #define SKL_DEMAND_WRITE SKL_DEMAND_RFO 414 #define SKL_LLC_ACCESS SKL_ANY_RESPONSE 415 #define SKL_L3_MISS_REMOTE (SKL_L3_MISS_REMOTE_HOP0_DRAM| \ 416 SKL_L3_MISS_REMOTE_HOP1_DRAM| \ 417 SKL_L3_MISS_REMOTE_HOP2P_DRAM) 418 419 static __initconst const u64 skl_hw_cache_event_ids 420 [PERF_COUNT_HW_CACHE_MAX] 421 [PERF_COUNT_HW_CACHE_OP_MAX] 422 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 423 { 424 [ C(L1D ) ] = { 425 [ C(OP_READ) ] = { 426 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 427 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 428 }, 429 [ C(OP_WRITE) ] = { 430 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 431 [ C(RESULT_MISS) ] = 0x0, 432 }, 433 [ C(OP_PREFETCH) ] = { 434 [ C(RESULT_ACCESS) ] = 0x0, 435 [ C(RESULT_MISS) ] = 0x0, 436 }, 437 }, 438 [ C(L1I ) ] = { 439 [ C(OP_READ) ] = { 440 [ C(RESULT_ACCESS) ] = 0x0, 441 [ C(RESULT_MISS) ] = 0x283, /* ICACHE_64B.MISS */ 442 }, 443 [ C(OP_WRITE) ] = { 444 [ C(RESULT_ACCESS) ] = -1, 445 [ C(RESULT_MISS) ] = -1, 446 }, 447 [ C(OP_PREFETCH) ] = { 448 [ C(RESULT_ACCESS) ] = 0x0, 449 [ C(RESULT_MISS) ] = 0x0, 450 }, 451 }, 452 [ C(LL ) ] = { 453 [ C(OP_READ) ] = { 454 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 455 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 456 }, 457 [ C(OP_WRITE) ] = { 458 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 459 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 460 }, 461 [ C(OP_PREFETCH) ] = { 462 [ C(RESULT_ACCESS) ] = 0x0, 463 [ C(RESULT_MISS) ] = 0x0, 464 }, 465 }, 466 [ C(DTLB) ] = { 467 [ C(OP_READ) ] = { 468 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_INST_RETIRED.ALL_LOADS */ 469 [ C(RESULT_MISS) ] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 470 }, 471 [ C(OP_WRITE) ] = { 472 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_INST_RETIRED.ALL_STORES */ 473 [ C(RESULT_MISS) ] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 474 }, 475 [ C(OP_PREFETCH) ] = { 476 [ C(RESULT_ACCESS) ] = 0x0, 477 [ C(RESULT_MISS) ] = 0x0, 478 }, 479 }, 480 [ C(ITLB) ] = { 481 [ C(OP_READ) ] = { 482 [ C(RESULT_ACCESS) ] = 0x2085, /* ITLB_MISSES.STLB_HIT */ 483 [ C(RESULT_MISS) ] = 0xe85, /* ITLB_MISSES.WALK_COMPLETED */ 484 }, 485 [ C(OP_WRITE) ] = { 486 [ C(RESULT_ACCESS) ] = -1, 487 [ C(RESULT_MISS) ] = -1, 488 }, 489 [ C(OP_PREFETCH) ] = { 490 [ C(RESULT_ACCESS) ] = -1, 491 [ C(RESULT_MISS) ] = -1, 492 }, 493 }, 494 [ C(BPU ) ] = { 495 [ C(OP_READ) ] = { 496 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 497 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 498 }, 499 [ C(OP_WRITE) ] = { 500 [ C(RESULT_ACCESS) ] = -1, 501 [ C(RESULT_MISS) ] = -1, 502 }, 503 [ C(OP_PREFETCH) ] = { 504 [ C(RESULT_ACCESS) ] = -1, 505 [ C(RESULT_MISS) ] = -1, 506 }, 507 }, 508 [ C(NODE) ] = { 509 [ C(OP_READ) ] = { 510 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 511 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 512 }, 513 [ C(OP_WRITE) ] = { 514 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 515 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 516 }, 517 [ C(OP_PREFETCH) ] = { 518 [ C(RESULT_ACCESS) ] = 0x0, 519 [ C(RESULT_MISS) ] = 0x0, 520 }, 521 }, 522 }; 523 524 static __initconst const u64 skl_hw_cache_extra_regs 525 [PERF_COUNT_HW_CACHE_MAX] 526 [PERF_COUNT_HW_CACHE_OP_MAX] 527 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 528 { 529 [ C(LL ) ] = { 530 [ C(OP_READ) ] = { 531 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 532 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 533 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 534 SKL_L3_MISS|SKL_ANY_SNOOP| 535 SKL_SUPPLIER_NONE, 536 }, 537 [ C(OP_WRITE) ] = { 538 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 539 SKL_LLC_ACCESS|SKL_ANY_SNOOP, 540 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 541 SKL_L3_MISS|SKL_ANY_SNOOP| 542 SKL_SUPPLIER_NONE, 543 }, 544 [ C(OP_PREFETCH) ] = { 545 [ C(RESULT_ACCESS) ] = 0x0, 546 [ C(RESULT_MISS) ] = 0x0, 547 }, 548 }, 549 [ C(NODE) ] = { 550 [ C(OP_READ) ] = { 551 [ C(RESULT_ACCESS) ] = SKL_DEMAND_READ| 552 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 553 [ C(RESULT_MISS) ] = SKL_DEMAND_READ| 554 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 555 }, 556 [ C(OP_WRITE) ] = { 557 [ C(RESULT_ACCESS) ] = SKL_DEMAND_WRITE| 558 SKL_L3_MISS_LOCAL_DRAM|SKL_SNOOP_DRAM, 559 [ C(RESULT_MISS) ] = SKL_DEMAND_WRITE| 560 SKL_L3_MISS_REMOTE|SKL_SNOOP_DRAM, 561 }, 562 [ C(OP_PREFETCH) ] = { 563 [ C(RESULT_ACCESS) ] = 0x0, 564 [ C(RESULT_MISS) ] = 0x0, 565 }, 566 }, 567 }; 568 569 #define SNB_DMND_DATA_RD (1ULL << 0) 570 #define SNB_DMND_RFO (1ULL << 1) 571 #define SNB_DMND_IFETCH (1ULL << 2) 572 #define SNB_DMND_WB (1ULL << 3) 573 #define SNB_PF_DATA_RD (1ULL << 4) 574 #define SNB_PF_RFO (1ULL << 5) 575 #define SNB_PF_IFETCH (1ULL << 6) 576 #define SNB_LLC_DATA_RD (1ULL << 7) 577 #define SNB_LLC_RFO (1ULL << 8) 578 #define SNB_LLC_IFETCH (1ULL << 9) 579 #define SNB_BUS_LOCKS (1ULL << 10) 580 #define SNB_STRM_ST (1ULL << 11) 581 #define SNB_OTHER (1ULL << 15) 582 #define SNB_RESP_ANY (1ULL << 16) 583 #define SNB_NO_SUPP (1ULL << 17) 584 #define SNB_LLC_HITM (1ULL << 18) 585 #define SNB_LLC_HITE (1ULL << 19) 586 #define SNB_LLC_HITS (1ULL << 20) 587 #define SNB_LLC_HITF (1ULL << 21) 588 #define SNB_LOCAL (1ULL << 22) 589 #define SNB_REMOTE (0xffULL << 23) 590 #define SNB_SNP_NONE (1ULL << 31) 591 #define SNB_SNP_NOT_NEEDED (1ULL << 32) 592 #define SNB_SNP_MISS (1ULL << 33) 593 #define SNB_NO_FWD (1ULL << 34) 594 #define SNB_SNP_FWD (1ULL << 35) 595 #define SNB_HITM (1ULL << 36) 596 #define SNB_NON_DRAM (1ULL << 37) 597 598 #define SNB_DMND_READ (SNB_DMND_DATA_RD|SNB_LLC_DATA_RD) 599 #define SNB_DMND_WRITE (SNB_DMND_RFO|SNB_LLC_RFO) 600 #define SNB_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 601 602 #define SNB_SNP_ANY (SNB_SNP_NONE|SNB_SNP_NOT_NEEDED| \ 603 SNB_SNP_MISS|SNB_NO_FWD|SNB_SNP_FWD| \ 604 SNB_HITM) 605 606 #define SNB_DRAM_ANY (SNB_LOCAL|SNB_REMOTE|SNB_SNP_ANY) 607 #define SNB_DRAM_REMOTE (SNB_REMOTE|SNB_SNP_ANY) 608 609 #define SNB_L3_ACCESS SNB_RESP_ANY 610 #define SNB_L3_MISS (SNB_DRAM_ANY|SNB_NON_DRAM) 611 612 static __initconst const u64 snb_hw_cache_extra_regs 613 [PERF_COUNT_HW_CACHE_MAX] 614 [PERF_COUNT_HW_CACHE_OP_MAX] 615 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 616 { 617 [ C(LL ) ] = { 618 [ C(OP_READ) ] = { 619 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_L3_ACCESS, 620 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_L3_MISS, 621 }, 622 [ C(OP_WRITE) ] = { 623 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_L3_ACCESS, 624 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_L3_MISS, 625 }, 626 [ C(OP_PREFETCH) ] = { 627 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_L3_ACCESS, 628 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_L3_MISS, 629 }, 630 }, 631 [ C(NODE) ] = { 632 [ C(OP_READ) ] = { 633 [ C(RESULT_ACCESS) ] = SNB_DMND_READ|SNB_DRAM_ANY, 634 [ C(RESULT_MISS) ] = SNB_DMND_READ|SNB_DRAM_REMOTE, 635 }, 636 [ C(OP_WRITE) ] = { 637 [ C(RESULT_ACCESS) ] = SNB_DMND_WRITE|SNB_DRAM_ANY, 638 [ C(RESULT_MISS) ] = SNB_DMND_WRITE|SNB_DRAM_REMOTE, 639 }, 640 [ C(OP_PREFETCH) ] = { 641 [ C(RESULT_ACCESS) ] = SNB_DMND_PREFETCH|SNB_DRAM_ANY, 642 [ C(RESULT_MISS) ] = SNB_DMND_PREFETCH|SNB_DRAM_REMOTE, 643 }, 644 }, 645 }; 646 647 static __initconst const u64 snb_hw_cache_event_ids 648 [PERF_COUNT_HW_CACHE_MAX] 649 [PERF_COUNT_HW_CACHE_OP_MAX] 650 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 651 { 652 [ C(L1D) ] = { 653 [ C(OP_READ) ] = { 654 [ C(RESULT_ACCESS) ] = 0xf1d0, /* MEM_UOP_RETIRED.LOADS */ 655 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPLACEMENT */ 656 }, 657 [ C(OP_WRITE) ] = { 658 [ C(RESULT_ACCESS) ] = 0xf2d0, /* MEM_UOP_RETIRED.STORES */ 659 [ C(RESULT_MISS) ] = 0x0851, /* L1D.ALL_M_REPLACEMENT */ 660 }, 661 [ C(OP_PREFETCH) ] = { 662 [ C(RESULT_ACCESS) ] = 0x0, 663 [ C(RESULT_MISS) ] = 0x024e, /* HW_PRE_REQ.DL1_MISS */ 664 }, 665 }, 666 [ C(L1I ) ] = { 667 [ C(OP_READ) ] = { 668 [ C(RESULT_ACCESS) ] = 0x0, 669 [ C(RESULT_MISS) ] = 0x0280, /* ICACHE.MISSES */ 670 }, 671 [ C(OP_WRITE) ] = { 672 [ C(RESULT_ACCESS) ] = -1, 673 [ C(RESULT_MISS) ] = -1, 674 }, 675 [ C(OP_PREFETCH) ] = { 676 [ C(RESULT_ACCESS) ] = 0x0, 677 [ C(RESULT_MISS) ] = 0x0, 678 }, 679 }, 680 [ C(LL ) ] = { 681 [ C(OP_READ) ] = { 682 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 683 [ C(RESULT_ACCESS) ] = 0x01b7, 684 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 685 [ C(RESULT_MISS) ] = 0x01b7, 686 }, 687 [ C(OP_WRITE) ] = { 688 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 689 [ C(RESULT_ACCESS) ] = 0x01b7, 690 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 691 [ C(RESULT_MISS) ] = 0x01b7, 692 }, 693 [ C(OP_PREFETCH) ] = { 694 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 695 [ C(RESULT_ACCESS) ] = 0x01b7, 696 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 697 [ C(RESULT_MISS) ] = 0x01b7, 698 }, 699 }, 700 [ C(DTLB) ] = { 701 [ C(OP_READ) ] = { 702 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOP_RETIRED.ALL_LOADS */ 703 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.CAUSES_A_WALK */ 704 }, 705 [ C(OP_WRITE) ] = { 706 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOP_RETIRED.ALL_STORES */ 707 [ C(RESULT_MISS) ] = 0x0149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 708 }, 709 [ C(OP_PREFETCH) ] = { 710 [ C(RESULT_ACCESS) ] = 0x0, 711 [ C(RESULT_MISS) ] = 0x0, 712 }, 713 }, 714 [ C(ITLB) ] = { 715 [ C(OP_READ) ] = { 716 [ C(RESULT_ACCESS) ] = 0x1085, /* ITLB_MISSES.STLB_HIT */ 717 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.CAUSES_A_WALK */ 718 }, 719 [ C(OP_WRITE) ] = { 720 [ C(RESULT_ACCESS) ] = -1, 721 [ C(RESULT_MISS) ] = -1, 722 }, 723 [ C(OP_PREFETCH) ] = { 724 [ C(RESULT_ACCESS) ] = -1, 725 [ C(RESULT_MISS) ] = -1, 726 }, 727 }, 728 [ C(BPU ) ] = { 729 [ C(OP_READ) ] = { 730 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 731 [ C(RESULT_MISS) ] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 732 }, 733 [ C(OP_WRITE) ] = { 734 [ C(RESULT_ACCESS) ] = -1, 735 [ C(RESULT_MISS) ] = -1, 736 }, 737 [ C(OP_PREFETCH) ] = { 738 [ C(RESULT_ACCESS) ] = -1, 739 [ C(RESULT_MISS) ] = -1, 740 }, 741 }, 742 [ C(NODE) ] = { 743 [ C(OP_READ) ] = { 744 [ C(RESULT_ACCESS) ] = 0x01b7, 745 [ C(RESULT_MISS) ] = 0x01b7, 746 }, 747 [ C(OP_WRITE) ] = { 748 [ C(RESULT_ACCESS) ] = 0x01b7, 749 [ C(RESULT_MISS) ] = 0x01b7, 750 }, 751 [ C(OP_PREFETCH) ] = { 752 [ C(RESULT_ACCESS) ] = 0x01b7, 753 [ C(RESULT_MISS) ] = 0x01b7, 754 }, 755 }, 756 757 }; 758 759 /* 760 * Notes on the events: 761 * - data reads do not include code reads (comparable to earlier tables) 762 * - data counts include speculative execution (except L1 write, dtlb, bpu) 763 * - remote node access includes remote memory, remote cache, remote mmio. 764 * - prefetches are not included in the counts because they are not 765 * reliably counted. 766 */ 767 768 #define HSW_DEMAND_DATA_RD BIT_ULL(0) 769 #define HSW_DEMAND_RFO BIT_ULL(1) 770 #define HSW_ANY_RESPONSE BIT_ULL(16) 771 #define HSW_SUPPLIER_NONE BIT_ULL(17) 772 #define HSW_L3_MISS_LOCAL_DRAM BIT_ULL(22) 773 #define HSW_L3_MISS_REMOTE_HOP0 BIT_ULL(27) 774 #define HSW_L3_MISS_REMOTE_HOP1 BIT_ULL(28) 775 #define HSW_L3_MISS_REMOTE_HOP2P BIT_ULL(29) 776 #define HSW_L3_MISS (HSW_L3_MISS_LOCAL_DRAM| \ 777 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 778 HSW_L3_MISS_REMOTE_HOP2P) 779 #define HSW_SNOOP_NONE BIT_ULL(31) 780 #define HSW_SNOOP_NOT_NEEDED BIT_ULL(32) 781 #define HSW_SNOOP_MISS BIT_ULL(33) 782 #define HSW_SNOOP_HIT_NO_FWD BIT_ULL(34) 783 #define HSW_SNOOP_HIT_WITH_FWD BIT_ULL(35) 784 #define HSW_SNOOP_HITM BIT_ULL(36) 785 #define HSW_SNOOP_NON_DRAM BIT_ULL(37) 786 #define HSW_ANY_SNOOP (HSW_SNOOP_NONE| \ 787 HSW_SNOOP_NOT_NEEDED|HSW_SNOOP_MISS| \ 788 HSW_SNOOP_HIT_NO_FWD|HSW_SNOOP_HIT_WITH_FWD| \ 789 HSW_SNOOP_HITM|HSW_SNOOP_NON_DRAM) 790 #define HSW_SNOOP_DRAM (HSW_ANY_SNOOP & ~HSW_SNOOP_NON_DRAM) 791 #define HSW_DEMAND_READ HSW_DEMAND_DATA_RD 792 #define HSW_DEMAND_WRITE HSW_DEMAND_RFO 793 #define HSW_L3_MISS_REMOTE (HSW_L3_MISS_REMOTE_HOP0|\ 794 HSW_L3_MISS_REMOTE_HOP1|HSW_L3_MISS_REMOTE_HOP2P) 795 #define HSW_LLC_ACCESS HSW_ANY_RESPONSE 796 797 #define BDW_L3_MISS_LOCAL BIT(26) 798 #define BDW_L3_MISS (BDW_L3_MISS_LOCAL| \ 799 HSW_L3_MISS_REMOTE_HOP0|HSW_L3_MISS_REMOTE_HOP1| \ 800 HSW_L3_MISS_REMOTE_HOP2P) 801 802 803 static __initconst const u64 hsw_hw_cache_event_ids 804 [PERF_COUNT_HW_CACHE_MAX] 805 [PERF_COUNT_HW_CACHE_OP_MAX] 806 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 807 { 808 [ C(L1D ) ] = { 809 [ C(OP_READ) ] = { 810 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 811 [ C(RESULT_MISS) ] = 0x151, /* L1D.REPLACEMENT */ 812 }, 813 [ C(OP_WRITE) ] = { 814 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 815 [ C(RESULT_MISS) ] = 0x0, 816 }, 817 [ C(OP_PREFETCH) ] = { 818 [ C(RESULT_ACCESS) ] = 0x0, 819 [ C(RESULT_MISS) ] = 0x0, 820 }, 821 }, 822 [ C(L1I ) ] = { 823 [ C(OP_READ) ] = { 824 [ C(RESULT_ACCESS) ] = 0x0, 825 [ C(RESULT_MISS) ] = 0x280, /* ICACHE.MISSES */ 826 }, 827 [ C(OP_WRITE) ] = { 828 [ C(RESULT_ACCESS) ] = -1, 829 [ C(RESULT_MISS) ] = -1, 830 }, 831 [ C(OP_PREFETCH) ] = { 832 [ C(RESULT_ACCESS) ] = 0x0, 833 [ C(RESULT_MISS) ] = 0x0, 834 }, 835 }, 836 [ C(LL ) ] = { 837 [ C(OP_READ) ] = { 838 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 839 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 840 }, 841 [ C(OP_WRITE) ] = { 842 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 843 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 844 }, 845 [ C(OP_PREFETCH) ] = { 846 [ C(RESULT_ACCESS) ] = 0x0, 847 [ C(RESULT_MISS) ] = 0x0, 848 }, 849 }, 850 [ C(DTLB) ] = { 851 [ C(OP_READ) ] = { 852 [ C(RESULT_ACCESS) ] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 853 [ C(RESULT_MISS) ] = 0x108, /* DTLB_LOAD_MISSES.MISS_CAUSES_A_WALK */ 854 }, 855 [ C(OP_WRITE) ] = { 856 [ C(RESULT_ACCESS) ] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 857 [ C(RESULT_MISS) ] = 0x149, /* DTLB_STORE_MISSES.MISS_CAUSES_A_WALK */ 858 }, 859 [ C(OP_PREFETCH) ] = { 860 [ C(RESULT_ACCESS) ] = 0x0, 861 [ C(RESULT_MISS) ] = 0x0, 862 }, 863 }, 864 [ C(ITLB) ] = { 865 [ C(OP_READ) ] = { 866 [ C(RESULT_ACCESS) ] = 0x6085, /* ITLB_MISSES.STLB_HIT */ 867 [ C(RESULT_MISS) ] = 0x185, /* ITLB_MISSES.MISS_CAUSES_A_WALK */ 868 }, 869 [ C(OP_WRITE) ] = { 870 [ C(RESULT_ACCESS) ] = -1, 871 [ C(RESULT_MISS) ] = -1, 872 }, 873 [ C(OP_PREFETCH) ] = { 874 [ C(RESULT_ACCESS) ] = -1, 875 [ C(RESULT_MISS) ] = -1, 876 }, 877 }, 878 [ C(BPU ) ] = { 879 [ C(OP_READ) ] = { 880 [ C(RESULT_ACCESS) ] = 0xc4, /* BR_INST_RETIRED.ALL_BRANCHES */ 881 [ C(RESULT_MISS) ] = 0xc5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 882 }, 883 [ C(OP_WRITE) ] = { 884 [ C(RESULT_ACCESS) ] = -1, 885 [ C(RESULT_MISS) ] = -1, 886 }, 887 [ C(OP_PREFETCH) ] = { 888 [ C(RESULT_ACCESS) ] = -1, 889 [ C(RESULT_MISS) ] = -1, 890 }, 891 }, 892 [ C(NODE) ] = { 893 [ C(OP_READ) ] = { 894 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 895 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 896 }, 897 [ C(OP_WRITE) ] = { 898 [ C(RESULT_ACCESS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 899 [ C(RESULT_MISS) ] = 0x1b7, /* OFFCORE_RESPONSE */ 900 }, 901 [ C(OP_PREFETCH) ] = { 902 [ C(RESULT_ACCESS) ] = 0x0, 903 [ C(RESULT_MISS) ] = 0x0, 904 }, 905 }, 906 }; 907 908 static __initconst const u64 hsw_hw_cache_extra_regs 909 [PERF_COUNT_HW_CACHE_MAX] 910 [PERF_COUNT_HW_CACHE_OP_MAX] 911 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 912 { 913 [ C(LL ) ] = { 914 [ C(OP_READ) ] = { 915 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 916 HSW_LLC_ACCESS, 917 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 918 HSW_L3_MISS|HSW_ANY_SNOOP, 919 }, 920 [ C(OP_WRITE) ] = { 921 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 922 HSW_LLC_ACCESS, 923 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 924 HSW_L3_MISS|HSW_ANY_SNOOP, 925 }, 926 [ C(OP_PREFETCH) ] = { 927 [ C(RESULT_ACCESS) ] = 0x0, 928 [ C(RESULT_MISS) ] = 0x0, 929 }, 930 }, 931 [ C(NODE) ] = { 932 [ C(OP_READ) ] = { 933 [ C(RESULT_ACCESS) ] = HSW_DEMAND_READ| 934 HSW_L3_MISS_LOCAL_DRAM| 935 HSW_SNOOP_DRAM, 936 [ C(RESULT_MISS) ] = HSW_DEMAND_READ| 937 HSW_L3_MISS_REMOTE| 938 HSW_SNOOP_DRAM, 939 }, 940 [ C(OP_WRITE) ] = { 941 [ C(RESULT_ACCESS) ] = HSW_DEMAND_WRITE| 942 HSW_L3_MISS_LOCAL_DRAM| 943 HSW_SNOOP_DRAM, 944 [ C(RESULT_MISS) ] = HSW_DEMAND_WRITE| 945 HSW_L3_MISS_REMOTE| 946 HSW_SNOOP_DRAM, 947 }, 948 [ C(OP_PREFETCH) ] = { 949 [ C(RESULT_ACCESS) ] = 0x0, 950 [ C(RESULT_MISS) ] = 0x0, 951 }, 952 }, 953 }; 954 955 static __initconst const u64 westmere_hw_cache_event_ids 956 [PERF_COUNT_HW_CACHE_MAX] 957 [PERF_COUNT_HW_CACHE_OP_MAX] 958 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 959 { 960 [ C(L1D) ] = { 961 [ C(OP_READ) ] = { 962 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 963 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 964 }, 965 [ C(OP_WRITE) ] = { 966 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 967 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 968 }, 969 [ C(OP_PREFETCH) ] = { 970 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 971 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 972 }, 973 }, 974 [ C(L1I ) ] = { 975 [ C(OP_READ) ] = { 976 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 977 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 978 }, 979 [ C(OP_WRITE) ] = { 980 [ C(RESULT_ACCESS) ] = -1, 981 [ C(RESULT_MISS) ] = -1, 982 }, 983 [ C(OP_PREFETCH) ] = { 984 [ C(RESULT_ACCESS) ] = 0x0, 985 [ C(RESULT_MISS) ] = 0x0, 986 }, 987 }, 988 [ C(LL ) ] = { 989 [ C(OP_READ) ] = { 990 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 991 [ C(RESULT_ACCESS) ] = 0x01b7, 992 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 993 [ C(RESULT_MISS) ] = 0x01b7, 994 }, 995 /* 996 * Use RFO, not WRITEBACK, because a write miss would typically occur 997 * on RFO. 998 */ 999 [ C(OP_WRITE) ] = { 1000 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1001 [ C(RESULT_ACCESS) ] = 0x01b7, 1002 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1003 [ C(RESULT_MISS) ] = 0x01b7, 1004 }, 1005 [ C(OP_PREFETCH) ] = { 1006 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1007 [ C(RESULT_ACCESS) ] = 0x01b7, 1008 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1009 [ C(RESULT_MISS) ] = 0x01b7, 1010 }, 1011 }, 1012 [ C(DTLB) ] = { 1013 [ C(OP_READ) ] = { 1014 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1015 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1016 }, 1017 [ C(OP_WRITE) ] = { 1018 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1019 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1020 }, 1021 [ C(OP_PREFETCH) ] = { 1022 [ C(RESULT_ACCESS) ] = 0x0, 1023 [ C(RESULT_MISS) ] = 0x0, 1024 }, 1025 }, 1026 [ C(ITLB) ] = { 1027 [ C(OP_READ) ] = { 1028 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1029 [ C(RESULT_MISS) ] = 0x0185, /* ITLB_MISSES.ANY */ 1030 }, 1031 [ C(OP_WRITE) ] = { 1032 [ C(RESULT_ACCESS) ] = -1, 1033 [ C(RESULT_MISS) ] = -1, 1034 }, 1035 [ C(OP_PREFETCH) ] = { 1036 [ C(RESULT_ACCESS) ] = -1, 1037 [ C(RESULT_MISS) ] = -1, 1038 }, 1039 }, 1040 [ C(BPU ) ] = { 1041 [ C(OP_READ) ] = { 1042 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1043 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1044 }, 1045 [ C(OP_WRITE) ] = { 1046 [ C(RESULT_ACCESS) ] = -1, 1047 [ C(RESULT_MISS) ] = -1, 1048 }, 1049 [ C(OP_PREFETCH) ] = { 1050 [ C(RESULT_ACCESS) ] = -1, 1051 [ C(RESULT_MISS) ] = -1, 1052 }, 1053 }, 1054 [ C(NODE) ] = { 1055 [ C(OP_READ) ] = { 1056 [ C(RESULT_ACCESS) ] = 0x01b7, 1057 [ C(RESULT_MISS) ] = 0x01b7, 1058 }, 1059 [ C(OP_WRITE) ] = { 1060 [ C(RESULT_ACCESS) ] = 0x01b7, 1061 [ C(RESULT_MISS) ] = 0x01b7, 1062 }, 1063 [ C(OP_PREFETCH) ] = { 1064 [ C(RESULT_ACCESS) ] = 0x01b7, 1065 [ C(RESULT_MISS) ] = 0x01b7, 1066 }, 1067 }, 1068 }; 1069 1070 /* 1071 * Nehalem/Westmere MSR_OFFCORE_RESPONSE bits; 1072 * See IA32 SDM Vol 3B 30.6.1.3 1073 */ 1074 1075 #define NHM_DMND_DATA_RD (1 << 0) 1076 #define NHM_DMND_RFO (1 << 1) 1077 #define NHM_DMND_IFETCH (1 << 2) 1078 #define NHM_DMND_WB (1 << 3) 1079 #define NHM_PF_DATA_RD (1 << 4) 1080 #define NHM_PF_DATA_RFO (1 << 5) 1081 #define NHM_PF_IFETCH (1 << 6) 1082 #define NHM_OFFCORE_OTHER (1 << 7) 1083 #define NHM_UNCORE_HIT (1 << 8) 1084 #define NHM_OTHER_CORE_HIT_SNP (1 << 9) 1085 #define NHM_OTHER_CORE_HITM (1 << 10) 1086 /* reserved */ 1087 #define NHM_REMOTE_CACHE_FWD (1 << 12) 1088 #define NHM_REMOTE_DRAM (1 << 13) 1089 #define NHM_LOCAL_DRAM (1 << 14) 1090 #define NHM_NON_DRAM (1 << 15) 1091 1092 #define NHM_LOCAL (NHM_LOCAL_DRAM|NHM_REMOTE_CACHE_FWD) 1093 #define NHM_REMOTE (NHM_REMOTE_DRAM) 1094 1095 #define NHM_DMND_READ (NHM_DMND_DATA_RD) 1096 #define NHM_DMND_WRITE (NHM_DMND_RFO|NHM_DMND_WB) 1097 #define NHM_DMND_PREFETCH (NHM_PF_DATA_RD|NHM_PF_DATA_RFO) 1098 1099 #define NHM_L3_HIT (NHM_UNCORE_HIT|NHM_OTHER_CORE_HIT_SNP|NHM_OTHER_CORE_HITM) 1100 #define NHM_L3_MISS (NHM_NON_DRAM|NHM_LOCAL_DRAM|NHM_REMOTE_DRAM|NHM_REMOTE_CACHE_FWD) 1101 #define NHM_L3_ACCESS (NHM_L3_HIT|NHM_L3_MISS) 1102 1103 static __initconst const u64 nehalem_hw_cache_extra_regs 1104 [PERF_COUNT_HW_CACHE_MAX] 1105 [PERF_COUNT_HW_CACHE_OP_MAX] 1106 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1107 { 1108 [ C(LL ) ] = { 1109 [ C(OP_READ) ] = { 1110 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_L3_ACCESS, 1111 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_L3_MISS, 1112 }, 1113 [ C(OP_WRITE) ] = { 1114 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_L3_ACCESS, 1115 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_L3_MISS, 1116 }, 1117 [ C(OP_PREFETCH) ] = { 1118 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_L3_ACCESS, 1119 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_L3_MISS, 1120 }, 1121 }, 1122 [ C(NODE) ] = { 1123 [ C(OP_READ) ] = { 1124 [ C(RESULT_ACCESS) ] = NHM_DMND_READ|NHM_LOCAL|NHM_REMOTE, 1125 [ C(RESULT_MISS) ] = NHM_DMND_READ|NHM_REMOTE, 1126 }, 1127 [ C(OP_WRITE) ] = { 1128 [ C(RESULT_ACCESS) ] = NHM_DMND_WRITE|NHM_LOCAL|NHM_REMOTE, 1129 [ C(RESULT_MISS) ] = NHM_DMND_WRITE|NHM_REMOTE, 1130 }, 1131 [ C(OP_PREFETCH) ] = { 1132 [ C(RESULT_ACCESS) ] = NHM_DMND_PREFETCH|NHM_LOCAL|NHM_REMOTE, 1133 [ C(RESULT_MISS) ] = NHM_DMND_PREFETCH|NHM_REMOTE, 1134 }, 1135 }, 1136 }; 1137 1138 static __initconst const u64 nehalem_hw_cache_event_ids 1139 [PERF_COUNT_HW_CACHE_MAX] 1140 [PERF_COUNT_HW_CACHE_OP_MAX] 1141 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1142 { 1143 [ C(L1D) ] = { 1144 [ C(OP_READ) ] = { 1145 [ C(RESULT_ACCESS) ] = 0x010b, /* MEM_INST_RETIRED.LOADS */ 1146 [ C(RESULT_MISS) ] = 0x0151, /* L1D.REPL */ 1147 }, 1148 [ C(OP_WRITE) ] = { 1149 [ C(RESULT_ACCESS) ] = 0x020b, /* MEM_INST_RETURED.STORES */ 1150 [ C(RESULT_MISS) ] = 0x0251, /* L1D.M_REPL */ 1151 }, 1152 [ C(OP_PREFETCH) ] = { 1153 [ C(RESULT_ACCESS) ] = 0x014e, /* L1D_PREFETCH.REQUESTS */ 1154 [ C(RESULT_MISS) ] = 0x024e, /* L1D_PREFETCH.MISS */ 1155 }, 1156 }, 1157 [ C(L1I ) ] = { 1158 [ C(OP_READ) ] = { 1159 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1160 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1161 }, 1162 [ C(OP_WRITE) ] = { 1163 [ C(RESULT_ACCESS) ] = -1, 1164 [ C(RESULT_MISS) ] = -1, 1165 }, 1166 [ C(OP_PREFETCH) ] = { 1167 [ C(RESULT_ACCESS) ] = 0x0, 1168 [ C(RESULT_MISS) ] = 0x0, 1169 }, 1170 }, 1171 [ C(LL ) ] = { 1172 [ C(OP_READ) ] = { 1173 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1174 [ C(RESULT_ACCESS) ] = 0x01b7, 1175 /* OFFCORE_RESPONSE.ANY_DATA.ANY_LLC_MISS */ 1176 [ C(RESULT_MISS) ] = 0x01b7, 1177 }, 1178 /* 1179 * Use RFO, not WRITEBACK, because a write miss would typically occur 1180 * on RFO. 1181 */ 1182 [ C(OP_WRITE) ] = { 1183 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1184 [ C(RESULT_ACCESS) ] = 0x01b7, 1185 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1186 [ C(RESULT_MISS) ] = 0x01b7, 1187 }, 1188 [ C(OP_PREFETCH) ] = { 1189 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1190 [ C(RESULT_ACCESS) ] = 0x01b7, 1191 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1192 [ C(RESULT_MISS) ] = 0x01b7, 1193 }, 1194 }, 1195 [ C(DTLB) ] = { 1196 [ C(OP_READ) ] = { 1197 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1198 [ C(RESULT_MISS) ] = 0x0108, /* DTLB_LOAD_MISSES.ANY */ 1199 }, 1200 [ C(OP_WRITE) ] = { 1201 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1202 [ C(RESULT_MISS) ] = 0x010c, /* MEM_STORE_RETIRED.DTLB_MISS */ 1203 }, 1204 [ C(OP_PREFETCH) ] = { 1205 [ C(RESULT_ACCESS) ] = 0x0, 1206 [ C(RESULT_MISS) ] = 0x0, 1207 }, 1208 }, 1209 [ C(ITLB) ] = { 1210 [ C(OP_READ) ] = { 1211 [ C(RESULT_ACCESS) ] = 0x01c0, /* INST_RETIRED.ANY_P */ 1212 [ C(RESULT_MISS) ] = 0x20c8, /* ITLB_MISS_RETIRED */ 1213 }, 1214 [ C(OP_WRITE) ] = { 1215 [ C(RESULT_ACCESS) ] = -1, 1216 [ C(RESULT_MISS) ] = -1, 1217 }, 1218 [ C(OP_PREFETCH) ] = { 1219 [ C(RESULT_ACCESS) ] = -1, 1220 [ C(RESULT_MISS) ] = -1, 1221 }, 1222 }, 1223 [ C(BPU ) ] = { 1224 [ C(OP_READ) ] = { 1225 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1226 [ C(RESULT_MISS) ] = 0x03e8, /* BPU_CLEARS.ANY */ 1227 }, 1228 [ C(OP_WRITE) ] = { 1229 [ C(RESULT_ACCESS) ] = -1, 1230 [ C(RESULT_MISS) ] = -1, 1231 }, 1232 [ C(OP_PREFETCH) ] = { 1233 [ C(RESULT_ACCESS) ] = -1, 1234 [ C(RESULT_MISS) ] = -1, 1235 }, 1236 }, 1237 [ C(NODE) ] = { 1238 [ C(OP_READ) ] = { 1239 [ C(RESULT_ACCESS) ] = 0x01b7, 1240 [ C(RESULT_MISS) ] = 0x01b7, 1241 }, 1242 [ C(OP_WRITE) ] = { 1243 [ C(RESULT_ACCESS) ] = 0x01b7, 1244 [ C(RESULT_MISS) ] = 0x01b7, 1245 }, 1246 [ C(OP_PREFETCH) ] = { 1247 [ C(RESULT_ACCESS) ] = 0x01b7, 1248 [ C(RESULT_MISS) ] = 0x01b7, 1249 }, 1250 }, 1251 }; 1252 1253 static __initconst const u64 core2_hw_cache_event_ids 1254 [PERF_COUNT_HW_CACHE_MAX] 1255 [PERF_COUNT_HW_CACHE_OP_MAX] 1256 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1257 { 1258 [ C(L1D) ] = { 1259 [ C(OP_READ) ] = { 1260 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI */ 1261 [ C(RESULT_MISS) ] = 0x0140, /* L1D_CACHE_LD.I_STATE */ 1262 }, 1263 [ C(OP_WRITE) ] = { 1264 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI */ 1265 [ C(RESULT_MISS) ] = 0x0141, /* L1D_CACHE_ST.I_STATE */ 1266 }, 1267 [ C(OP_PREFETCH) ] = { 1268 [ C(RESULT_ACCESS) ] = 0x104e, /* L1D_PREFETCH.REQUESTS */ 1269 [ C(RESULT_MISS) ] = 0, 1270 }, 1271 }, 1272 [ C(L1I ) ] = { 1273 [ C(OP_READ) ] = { 1274 [ C(RESULT_ACCESS) ] = 0x0080, /* L1I.READS */ 1275 [ C(RESULT_MISS) ] = 0x0081, /* L1I.MISSES */ 1276 }, 1277 [ C(OP_WRITE) ] = { 1278 [ C(RESULT_ACCESS) ] = -1, 1279 [ C(RESULT_MISS) ] = -1, 1280 }, 1281 [ C(OP_PREFETCH) ] = { 1282 [ C(RESULT_ACCESS) ] = 0, 1283 [ C(RESULT_MISS) ] = 0, 1284 }, 1285 }, 1286 [ C(LL ) ] = { 1287 [ C(OP_READ) ] = { 1288 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1289 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1290 }, 1291 [ C(OP_WRITE) ] = { 1292 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1293 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1294 }, 1295 [ C(OP_PREFETCH) ] = { 1296 [ C(RESULT_ACCESS) ] = 0, 1297 [ C(RESULT_MISS) ] = 0, 1298 }, 1299 }, 1300 [ C(DTLB) ] = { 1301 [ C(OP_READ) ] = { 1302 [ C(RESULT_ACCESS) ] = 0x0f40, /* L1D_CACHE_LD.MESI (alias) */ 1303 [ C(RESULT_MISS) ] = 0x0208, /* DTLB_MISSES.MISS_LD */ 1304 }, 1305 [ C(OP_WRITE) ] = { 1306 [ C(RESULT_ACCESS) ] = 0x0f41, /* L1D_CACHE_ST.MESI (alias) */ 1307 [ C(RESULT_MISS) ] = 0x0808, /* DTLB_MISSES.MISS_ST */ 1308 }, 1309 [ C(OP_PREFETCH) ] = { 1310 [ C(RESULT_ACCESS) ] = 0, 1311 [ C(RESULT_MISS) ] = 0, 1312 }, 1313 }, 1314 [ C(ITLB) ] = { 1315 [ C(OP_READ) ] = { 1316 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1317 [ C(RESULT_MISS) ] = 0x1282, /* ITLBMISSES */ 1318 }, 1319 [ C(OP_WRITE) ] = { 1320 [ C(RESULT_ACCESS) ] = -1, 1321 [ C(RESULT_MISS) ] = -1, 1322 }, 1323 [ C(OP_PREFETCH) ] = { 1324 [ C(RESULT_ACCESS) ] = -1, 1325 [ C(RESULT_MISS) ] = -1, 1326 }, 1327 }, 1328 [ C(BPU ) ] = { 1329 [ C(OP_READ) ] = { 1330 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1331 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1332 }, 1333 [ C(OP_WRITE) ] = { 1334 [ C(RESULT_ACCESS) ] = -1, 1335 [ C(RESULT_MISS) ] = -1, 1336 }, 1337 [ C(OP_PREFETCH) ] = { 1338 [ C(RESULT_ACCESS) ] = -1, 1339 [ C(RESULT_MISS) ] = -1, 1340 }, 1341 }, 1342 }; 1343 1344 static __initconst const u64 atom_hw_cache_event_ids 1345 [PERF_COUNT_HW_CACHE_MAX] 1346 [PERF_COUNT_HW_CACHE_OP_MAX] 1347 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1348 { 1349 [ C(L1D) ] = { 1350 [ C(OP_READ) ] = { 1351 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE.LD */ 1352 [ C(RESULT_MISS) ] = 0, 1353 }, 1354 [ C(OP_WRITE) ] = { 1355 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE.ST */ 1356 [ C(RESULT_MISS) ] = 0, 1357 }, 1358 [ C(OP_PREFETCH) ] = { 1359 [ C(RESULT_ACCESS) ] = 0x0, 1360 [ C(RESULT_MISS) ] = 0, 1361 }, 1362 }, 1363 [ C(L1I ) ] = { 1364 [ C(OP_READ) ] = { 1365 [ C(RESULT_ACCESS) ] = 0x0380, /* L1I.READS */ 1366 [ C(RESULT_MISS) ] = 0x0280, /* L1I.MISSES */ 1367 }, 1368 [ C(OP_WRITE) ] = { 1369 [ C(RESULT_ACCESS) ] = -1, 1370 [ C(RESULT_MISS) ] = -1, 1371 }, 1372 [ C(OP_PREFETCH) ] = { 1373 [ C(RESULT_ACCESS) ] = 0, 1374 [ C(RESULT_MISS) ] = 0, 1375 }, 1376 }, 1377 [ C(LL ) ] = { 1378 [ C(OP_READ) ] = { 1379 [ C(RESULT_ACCESS) ] = 0x4f29, /* L2_LD.MESI */ 1380 [ C(RESULT_MISS) ] = 0x4129, /* L2_LD.ISTATE */ 1381 }, 1382 [ C(OP_WRITE) ] = { 1383 [ C(RESULT_ACCESS) ] = 0x4f2A, /* L2_ST.MESI */ 1384 [ C(RESULT_MISS) ] = 0x412A, /* L2_ST.ISTATE */ 1385 }, 1386 [ C(OP_PREFETCH) ] = { 1387 [ C(RESULT_ACCESS) ] = 0, 1388 [ C(RESULT_MISS) ] = 0, 1389 }, 1390 }, 1391 [ C(DTLB) ] = { 1392 [ C(OP_READ) ] = { 1393 [ C(RESULT_ACCESS) ] = 0x2140, /* L1D_CACHE_LD.MESI (alias) */ 1394 [ C(RESULT_MISS) ] = 0x0508, /* DTLB_MISSES.MISS_LD */ 1395 }, 1396 [ C(OP_WRITE) ] = { 1397 [ C(RESULT_ACCESS) ] = 0x2240, /* L1D_CACHE_ST.MESI (alias) */ 1398 [ C(RESULT_MISS) ] = 0x0608, /* DTLB_MISSES.MISS_ST */ 1399 }, 1400 [ C(OP_PREFETCH) ] = { 1401 [ C(RESULT_ACCESS) ] = 0, 1402 [ C(RESULT_MISS) ] = 0, 1403 }, 1404 }, 1405 [ C(ITLB) ] = { 1406 [ C(OP_READ) ] = { 1407 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1408 [ C(RESULT_MISS) ] = 0x0282, /* ITLB.MISSES */ 1409 }, 1410 [ C(OP_WRITE) ] = { 1411 [ C(RESULT_ACCESS) ] = -1, 1412 [ C(RESULT_MISS) ] = -1, 1413 }, 1414 [ C(OP_PREFETCH) ] = { 1415 [ C(RESULT_ACCESS) ] = -1, 1416 [ C(RESULT_MISS) ] = -1, 1417 }, 1418 }, 1419 [ C(BPU ) ] = { 1420 [ C(OP_READ) ] = { 1421 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1422 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1423 }, 1424 [ C(OP_WRITE) ] = { 1425 [ C(RESULT_ACCESS) ] = -1, 1426 [ C(RESULT_MISS) ] = -1, 1427 }, 1428 [ C(OP_PREFETCH) ] = { 1429 [ C(RESULT_ACCESS) ] = -1, 1430 [ C(RESULT_MISS) ] = -1, 1431 }, 1432 }, 1433 }; 1434 1435 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_slm, "event=0x3c"); 1436 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_slm, "2"); 1437 /* no_alloc_cycles.not_delivered */ 1438 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_slm, 1439 "event=0xca,umask=0x50"); 1440 EVENT_ATTR_STR(topdown-fetch-bubbles.scale, td_fetch_bubbles_scale_slm, "2"); 1441 /* uops_retired.all */ 1442 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_slm, 1443 "event=0xc2,umask=0x10"); 1444 /* uops_retired.all */ 1445 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_slm, 1446 "event=0xc2,umask=0x10"); 1447 1448 static struct attribute *slm_events_attrs[] = { 1449 EVENT_PTR(td_total_slots_slm), 1450 EVENT_PTR(td_total_slots_scale_slm), 1451 EVENT_PTR(td_fetch_bubbles_slm), 1452 EVENT_PTR(td_fetch_bubbles_scale_slm), 1453 EVENT_PTR(td_slots_issued_slm), 1454 EVENT_PTR(td_slots_retired_slm), 1455 NULL 1456 }; 1457 1458 static struct extra_reg intel_slm_extra_regs[] __read_mostly = 1459 { 1460 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1461 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x768005ffffull, RSP_0), 1462 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x368005ffffull, RSP_1), 1463 EVENT_EXTRA_END 1464 }; 1465 1466 #define SLM_DMND_READ SNB_DMND_DATA_RD 1467 #define SLM_DMND_WRITE SNB_DMND_RFO 1468 #define SLM_DMND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1469 1470 #define SLM_SNP_ANY (SNB_SNP_NONE|SNB_SNP_MISS|SNB_NO_FWD|SNB_HITM) 1471 #define SLM_LLC_ACCESS SNB_RESP_ANY 1472 #define SLM_LLC_MISS (SLM_SNP_ANY|SNB_NON_DRAM) 1473 1474 static __initconst const u64 slm_hw_cache_extra_regs 1475 [PERF_COUNT_HW_CACHE_MAX] 1476 [PERF_COUNT_HW_CACHE_OP_MAX] 1477 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1478 { 1479 [ C(LL ) ] = { 1480 [ C(OP_READ) ] = { 1481 [ C(RESULT_ACCESS) ] = SLM_DMND_READ|SLM_LLC_ACCESS, 1482 [ C(RESULT_MISS) ] = 0, 1483 }, 1484 [ C(OP_WRITE) ] = { 1485 [ C(RESULT_ACCESS) ] = SLM_DMND_WRITE|SLM_LLC_ACCESS, 1486 [ C(RESULT_MISS) ] = SLM_DMND_WRITE|SLM_LLC_MISS, 1487 }, 1488 [ C(OP_PREFETCH) ] = { 1489 [ C(RESULT_ACCESS) ] = SLM_DMND_PREFETCH|SLM_LLC_ACCESS, 1490 [ C(RESULT_MISS) ] = SLM_DMND_PREFETCH|SLM_LLC_MISS, 1491 }, 1492 }, 1493 }; 1494 1495 static __initconst const u64 slm_hw_cache_event_ids 1496 [PERF_COUNT_HW_CACHE_MAX] 1497 [PERF_COUNT_HW_CACHE_OP_MAX] 1498 [PERF_COUNT_HW_CACHE_RESULT_MAX] = 1499 { 1500 [ C(L1D) ] = { 1501 [ C(OP_READ) ] = { 1502 [ C(RESULT_ACCESS) ] = 0, 1503 [ C(RESULT_MISS) ] = 0x0104, /* LD_DCU_MISS */ 1504 }, 1505 [ C(OP_WRITE) ] = { 1506 [ C(RESULT_ACCESS) ] = 0, 1507 [ C(RESULT_MISS) ] = 0, 1508 }, 1509 [ C(OP_PREFETCH) ] = { 1510 [ C(RESULT_ACCESS) ] = 0, 1511 [ C(RESULT_MISS) ] = 0, 1512 }, 1513 }, 1514 [ C(L1I ) ] = { 1515 [ C(OP_READ) ] = { 1516 [ C(RESULT_ACCESS) ] = 0x0380, /* ICACHE.ACCESSES */ 1517 [ C(RESULT_MISS) ] = 0x0280, /* ICACGE.MISSES */ 1518 }, 1519 [ C(OP_WRITE) ] = { 1520 [ C(RESULT_ACCESS) ] = -1, 1521 [ C(RESULT_MISS) ] = -1, 1522 }, 1523 [ C(OP_PREFETCH) ] = { 1524 [ C(RESULT_ACCESS) ] = 0, 1525 [ C(RESULT_MISS) ] = 0, 1526 }, 1527 }, 1528 [ C(LL ) ] = { 1529 [ C(OP_READ) ] = { 1530 /* OFFCORE_RESPONSE.ANY_DATA.LOCAL_CACHE */ 1531 [ C(RESULT_ACCESS) ] = 0x01b7, 1532 [ C(RESULT_MISS) ] = 0, 1533 }, 1534 [ C(OP_WRITE) ] = { 1535 /* OFFCORE_RESPONSE.ANY_RFO.LOCAL_CACHE */ 1536 [ C(RESULT_ACCESS) ] = 0x01b7, 1537 /* OFFCORE_RESPONSE.ANY_RFO.ANY_LLC_MISS */ 1538 [ C(RESULT_MISS) ] = 0x01b7, 1539 }, 1540 [ C(OP_PREFETCH) ] = { 1541 /* OFFCORE_RESPONSE.PREFETCH.LOCAL_CACHE */ 1542 [ C(RESULT_ACCESS) ] = 0x01b7, 1543 /* OFFCORE_RESPONSE.PREFETCH.ANY_LLC_MISS */ 1544 [ C(RESULT_MISS) ] = 0x01b7, 1545 }, 1546 }, 1547 [ C(DTLB) ] = { 1548 [ C(OP_READ) ] = { 1549 [ C(RESULT_ACCESS) ] = 0, 1550 [ C(RESULT_MISS) ] = 0x0804, /* LD_DTLB_MISS */ 1551 }, 1552 [ C(OP_WRITE) ] = { 1553 [ C(RESULT_ACCESS) ] = 0, 1554 [ C(RESULT_MISS) ] = 0, 1555 }, 1556 [ C(OP_PREFETCH) ] = { 1557 [ C(RESULT_ACCESS) ] = 0, 1558 [ C(RESULT_MISS) ] = 0, 1559 }, 1560 }, 1561 [ C(ITLB) ] = { 1562 [ C(OP_READ) ] = { 1563 [ C(RESULT_ACCESS) ] = 0x00c0, /* INST_RETIRED.ANY_P */ 1564 [ C(RESULT_MISS) ] = 0x40205, /* PAGE_WALKS.I_SIDE_WALKS */ 1565 }, 1566 [ C(OP_WRITE) ] = { 1567 [ C(RESULT_ACCESS) ] = -1, 1568 [ C(RESULT_MISS) ] = -1, 1569 }, 1570 [ C(OP_PREFETCH) ] = { 1571 [ C(RESULT_ACCESS) ] = -1, 1572 [ C(RESULT_MISS) ] = -1, 1573 }, 1574 }, 1575 [ C(BPU ) ] = { 1576 [ C(OP_READ) ] = { 1577 [ C(RESULT_ACCESS) ] = 0x00c4, /* BR_INST_RETIRED.ANY */ 1578 [ C(RESULT_MISS) ] = 0x00c5, /* BP_INST_RETIRED.MISPRED */ 1579 }, 1580 [ C(OP_WRITE) ] = { 1581 [ C(RESULT_ACCESS) ] = -1, 1582 [ C(RESULT_MISS) ] = -1, 1583 }, 1584 [ C(OP_PREFETCH) ] = { 1585 [ C(RESULT_ACCESS) ] = -1, 1586 [ C(RESULT_MISS) ] = -1, 1587 }, 1588 }, 1589 }; 1590 1591 EVENT_ATTR_STR(topdown-total-slots, td_total_slots_glm, "event=0x3c"); 1592 EVENT_ATTR_STR(topdown-total-slots.scale, td_total_slots_scale_glm, "3"); 1593 /* UOPS_NOT_DELIVERED.ANY */ 1594 EVENT_ATTR_STR(topdown-fetch-bubbles, td_fetch_bubbles_glm, "event=0x9c"); 1595 /* ISSUE_SLOTS_NOT_CONSUMED.RECOVERY */ 1596 EVENT_ATTR_STR(topdown-recovery-bubbles, td_recovery_bubbles_glm, "event=0xca,umask=0x02"); 1597 /* UOPS_RETIRED.ANY */ 1598 EVENT_ATTR_STR(topdown-slots-retired, td_slots_retired_glm, "event=0xc2"); 1599 /* UOPS_ISSUED.ANY */ 1600 EVENT_ATTR_STR(topdown-slots-issued, td_slots_issued_glm, "event=0x0e"); 1601 1602 static struct attribute *glm_events_attrs[] = { 1603 EVENT_PTR(td_total_slots_glm), 1604 EVENT_PTR(td_total_slots_scale_glm), 1605 EVENT_PTR(td_fetch_bubbles_glm), 1606 EVENT_PTR(td_recovery_bubbles_glm), 1607 EVENT_PTR(td_slots_issued_glm), 1608 EVENT_PTR(td_slots_retired_glm), 1609 NULL 1610 }; 1611 1612 static struct extra_reg intel_glm_extra_regs[] __read_mostly = { 1613 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1614 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0x760005ffbfull, RSP_0), 1615 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0x360005ffbfull, RSP_1), 1616 EVENT_EXTRA_END 1617 }; 1618 1619 #define GLM_DEMAND_DATA_RD BIT_ULL(0) 1620 #define GLM_DEMAND_RFO BIT_ULL(1) 1621 #define GLM_ANY_RESPONSE BIT_ULL(16) 1622 #define GLM_SNP_NONE_OR_MISS BIT_ULL(33) 1623 #define GLM_DEMAND_READ GLM_DEMAND_DATA_RD 1624 #define GLM_DEMAND_WRITE GLM_DEMAND_RFO 1625 #define GLM_DEMAND_PREFETCH (SNB_PF_DATA_RD|SNB_PF_RFO) 1626 #define GLM_LLC_ACCESS GLM_ANY_RESPONSE 1627 #define GLM_SNP_ANY (GLM_SNP_NONE_OR_MISS|SNB_NO_FWD|SNB_HITM) 1628 #define GLM_LLC_MISS (GLM_SNP_ANY|SNB_NON_DRAM) 1629 1630 static __initconst const u64 glm_hw_cache_event_ids 1631 [PERF_COUNT_HW_CACHE_MAX] 1632 [PERF_COUNT_HW_CACHE_OP_MAX] 1633 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1634 [C(L1D)] = { 1635 [C(OP_READ)] = { 1636 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1637 [C(RESULT_MISS)] = 0x0, 1638 }, 1639 [C(OP_WRITE)] = { 1640 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1641 [C(RESULT_MISS)] = 0x0, 1642 }, 1643 [C(OP_PREFETCH)] = { 1644 [C(RESULT_ACCESS)] = 0x0, 1645 [C(RESULT_MISS)] = 0x0, 1646 }, 1647 }, 1648 [C(L1I)] = { 1649 [C(OP_READ)] = { 1650 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1651 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1652 }, 1653 [C(OP_WRITE)] = { 1654 [C(RESULT_ACCESS)] = -1, 1655 [C(RESULT_MISS)] = -1, 1656 }, 1657 [C(OP_PREFETCH)] = { 1658 [C(RESULT_ACCESS)] = 0x0, 1659 [C(RESULT_MISS)] = 0x0, 1660 }, 1661 }, 1662 [C(LL)] = { 1663 [C(OP_READ)] = { 1664 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1665 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1666 }, 1667 [C(OP_WRITE)] = { 1668 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1669 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1670 }, 1671 [C(OP_PREFETCH)] = { 1672 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1673 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1674 }, 1675 }, 1676 [C(DTLB)] = { 1677 [C(OP_READ)] = { 1678 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1679 [C(RESULT_MISS)] = 0x0, 1680 }, 1681 [C(OP_WRITE)] = { 1682 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1683 [C(RESULT_MISS)] = 0x0, 1684 }, 1685 [C(OP_PREFETCH)] = { 1686 [C(RESULT_ACCESS)] = 0x0, 1687 [C(RESULT_MISS)] = 0x0, 1688 }, 1689 }, 1690 [C(ITLB)] = { 1691 [C(OP_READ)] = { 1692 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 1693 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 1694 }, 1695 [C(OP_WRITE)] = { 1696 [C(RESULT_ACCESS)] = -1, 1697 [C(RESULT_MISS)] = -1, 1698 }, 1699 [C(OP_PREFETCH)] = { 1700 [C(RESULT_ACCESS)] = -1, 1701 [C(RESULT_MISS)] = -1, 1702 }, 1703 }, 1704 [C(BPU)] = { 1705 [C(OP_READ)] = { 1706 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1707 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1708 }, 1709 [C(OP_WRITE)] = { 1710 [C(RESULT_ACCESS)] = -1, 1711 [C(RESULT_MISS)] = -1, 1712 }, 1713 [C(OP_PREFETCH)] = { 1714 [C(RESULT_ACCESS)] = -1, 1715 [C(RESULT_MISS)] = -1, 1716 }, 1717 }, 1718 }; 1719 1720 static __initconst const u64 glm_hw_cache_extra_regs 1721 [PERF_COUNT_HW_CACHE_MAX] 1722 [PERF_COUNT_HW_CACHE_OP_MAX] 1723 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1724 [C(LL)] = { 1725 [C(OP_READ)] = { 1726 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 1727 GLM_LLC_ACCESS, 1728 [C(RESULT_MISS)] = GLM_DEMAND_READ| 1729 GLM_LLC_MISS, 1730 }, 1731 [C(OP_WRITE)] = { 1732 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 1733 GLM_LLC_ACCESS, 1734 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 1735 GLM_LLC_MISS, 1736 }, 1737 [C(OP_PREFETCH)] = { 1738 [C(RESULT_ACCESS)] = GLM_DEMAND_PREFETCH| 1739 GLM_LLC_ACCESS, 1740 [C(RESULT_MISS)] = GLM_DEMAND_PREFETCH| 1741 GLM_LLC_MISS, 1742 }, 1743 }, 1744 }; 1745 1746 static __initconst const u64 glp_hw_cache_event_ids 1747 [PERF_COUNT_HW_CACHE_MAX] 1748 [PERF_COUNT_HW_CACHE_OP_MAX] 1749 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1750 [C(L1D)] = { 1751 [C(OP_READ)] = { 1752 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1753 [C(RESULT_MISS)] = 0x0, 1754 }, 1755 [C(OP_WRITE)] = { 1756 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1757 [C(RESULT_MISS)] = 0x0, 1758 }, 1759 [C(OP_PREFETCH)] = { 1760 [C(RESULT_ACCESS)] = 0x0, 1761 [C(RESULT_MISS)] = 0x0, 1762 }, 1763 }, 1764 [C(L1I)] = { 1765 [C(OP_READ)] = { 1766 [C(RESULT_ACCESS)] = 0x0380, /* ICACHE.ACCESSES */ 1767 [C(RESULT_MISS)] = 0x0280, /* ICACHE.MISSES */ 1768 }, 1769 [C(OP_WRITE)] = { 1770 [C(RESULT_ACCESS)] = -1, 1771 [C(RESULT_MISS)] = -1, 1772 }, 1773 [C(OP_PREFETCH)] = { 1774 [C(RESULT_ACCESS)] = 0x0, 1775 [C(RESULT_MISS)] = 0x0, 1776 }, 1777 }, 1778 [C(LL)] = { 1779 [C(OP_READ)] = { 1780 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1781 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1782 }, 1783 [C(OP_WRITE)] = { 1784 [C(RESULT_ACCESS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1785 [C(RESULT_MISS)] = 0x1b7, /* OFFCORE_RESPONSE */ 1786 }, 1787 [C(OP_PREFETCH)] = { 1788 [C(RESULT_ACCESS)] = 0x0, 1789 [C(RESULT_MISS)] = 0x0, 1790 }, 1791 }, 1792 [C(DTLB)] = { 1793 [C(OP_READ)] = { 1794 [C(RESULT_ACCESS)] = 0x81d0, /* MEM_UOPS_RETIRED.ALL_LOADS */ 1795 [C(RESULT_MISS)] = 0xe08, /* DTLB_LOAD_MISSES.WALK_COMPLETED */ 1796 }, 1797 [C(OP_WRITE)] = { 1798 [C(RESULT_ACCESS)] = 0x82d0, /* MEM_UOPS_RETIRED.ALL_STORES */ 1799 [C(RESULT_MISS)] = 0xe49, /* DTLB_STORE_MISSES.WALK_COMPLETED */ 1800 }, 1801 [C(OP_PREFETCH)] = { 1802 [C(RESULT_ACCESS)] = 0x0, 1803 [C(RESULT_MISS)] = 0x0, 1804 }, 1805 }, 1806 [C(ITLB)] = { 1807 [C(OP_READ)] = { 1808 [C(RESULT_ACCESS)] = 0x00c0, /* INST_RETIRED.ANY_P */ 1809 [C(RESULT_MISS)] = 0x0481, /* ITLB.MISS */ 1810 }, 1811 [C(OP_WRITE)] = { 1812 [C(RESULT_ACCESS)] = -1, 1813 [C(RESULT_MISS)] = -1, 1814 }, 1815 [C(OP_PREFETCH)] = { 1816 [C(RESULT_ACCESS)] = -1, 1817 [C(RESULT_MISS)] = -1, 1818 }, 1819 }, 1820 [C(BPU)] = { 1821 [C(OP_READ)] = { 1822 [C(RESULT_ACCESS)] = 0x00c4, /* BR_INST_RETIRED.ALL_BRANCHES */ 1823 [C(RESULT_MISS)] = 0x00c5, /* BR_MISP_RETIRED.ALL_BRANCHES */ 1824 }, 1825 [C(OP_WRITE)] = { 1826 [C(RESULT_ACCESS)] = -1, 1827 [C(RESULT_MISS)] = -1, 1828 }, 1829 [C(OP_PREFETCH)] = { 1830 [C(RESULT_ACCESS)] = -1, 1831 [C(RESULT_MISS)] = -1, 1832 }, 1833 }, 1834 }; 1835 1836 static __initconst const u64 glp_hw_cache_extra_regs 1837 [PERF_COUNT_HW_CACHE_MAX] 1838 [PERF_COUNT_HW_CACHE_OP_MAX] 1839 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1840 [C(LL)] = { 1841 [C(OP_READ)] = { 1842 [C(RESULT_ACCESS)] = GLM_DEMAND_READ| 1843 GLM_LLC_ACCESS, 1844 [C(RESULT_MISS)] = GLM_DEMAND_READ| 1845 GLM_LLC_MISS, 1846 }, 1847 [C(OP_WRITE)] = { 1848 [C(RESULT_ACCESS)] = GLM_DEMAND_WRITE| 1849 GLM_LLC_ACCESS, 1850 [C(RESULT_MISS)] = GLM_DEMAND_WRITE| 1851 GLM_LLC_MISS, 1852 }, 1853 [C(OP_PREFETCH)] = { 1854 [C(RESULT_ACCESS)] = 0x0, 1855 [C(RESULT_MISS)] = 0x0, 1856 }, 1857 }, 1858 }; 1859 1860 #define TNT_LOCAL_DRAM BIT_ULL(26) 1861 #define TNT_DEMAND_READ GLM_DEMAND_DATA_RD 1862 #define TNT_DEMAND_WRITE GLM_DEMAND_RFO 1863 #define TNT_LLC_ACCESS GLM_ANY_RESPONSE 1864 #define TNT_SNP_ANY (SNB_SNP_NOT_NEEDED|SNB_SNP_MISS| \ 1865 SNB_NO_FWD|SNB_SNP_FWD|SNB_HITM) 1866 #define TNT_LLC_MISS (TNT_SNP_ANY|SNB_NON_DRAM|TNT_LOCAL_DRAM) 1867 1868 static __initconst const u64 tnt_hw_cache_extra_regs 1869 [PERF_COUNT_HW_CACHE_MAX] 1870 [PERF_COUNT_HW_CACHE_OP_MAX] 1871 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1872 [C(LL)] = { 1873 [C(OP_READ)] = { 1874 [C(RESULT_ACCESS)] = TNT_DEMAND_READ| 1875 TNT_LLC_ACCESS, 1876 [C(RESULT_MISS)] = TNT_DEMAND_READ| 1877 TNT_LLC_MISS, 1878 }, 1879 [C(OP_WRITE)] = { 1880 [C(RESULT_ACCESS)] = TNT_DEMAND_WRITE| 1881 TNT_LLC_ACCESS, 1882 [C(RESULT_MISS)] = TNT_DEMAND_WRITE| 1883 TNT_LLC_MISS, 1884 }, 1885 [C(OP_PREFETCH)] = { 1886 [C(RESULT_ACCESS)] = 0x0, 1887 [C(RESULT_MISS)] = 0x0, 1888 }, 1889 }, 1890 }; 1891 1892 static struct extra_reg intel_tnt_extra_regs[] __read_mostly = { 1893 /* must define OFFCORE_RSP_X first, see intel_fixup_er() */ 1894 INTEL_UEVENT_EXTRA_REG(0x01b7, MSR_OFFCORE_RSP_0, 0xffffff9fffull, RSP_0), 1895 INTEL_UEVENT_EXTRA_REG(0x02b7, MSR_OFFCORE_RSP_1, 0xffffff9fffull, RSP_1), 1896 EVENT_EXTRA_END 1897 }; 1898 1899 #define KNL_OT_L2_HITE BIT_ULL(19) /* Other Tile L2 Hit */ 1900 #define KNL_OT_L2_HITF BIT_ULL(20) /* Other Tile L2 Hit */ 1901 #define KNL_MCDRAM_LOCAL BIT_ULL(21) 1902 #define KNL_MCDRAM_FAR BIT_ULL(22) 1903 #define KNL_DDR_LOCAL BIT_ULL(23) 1904 #define KNL_DDR_FAR BIT_ULL(24) 1905 #define KNL_DRAM_ANY (KNL_MCDRAM_LOCAL | KNL_MCDRAM_FAR | \ 1906 KNL_DDR_LOCAL | KNL_DDR_FAR) 1907 #define KNL_L2_READ SLM_DMND_READ 1908 #define KNL_L2_WRITE SLM_DMND_WRITE 1909 #define KNL_L2_PREFETCH SLM_DMND_PREFETCH 1910 #define KNL_L2_ACCESS SLM_LLC_ACCESS 1911 #define KNL_L2_MISS (KNL_OT_L2_HITE | KNL_OT_L2_HITF | \ 1912 KNL_DRAM_ANY | SNB_SNP_ANY | \ 1913 SNB_NON_DRAM) 1914 1915 static __initconst const u64 knl_hw_cache_extra_regs 1916 [PERF_COUNT_HW_CACHE_MAX] 1917 [PERF_COUNT_HW_CACHE_OP_MAX] 1918 [PERF_COUNT_HW_CACHE_RESULT_MAX] = { 1919 [C(LL)] = { 1920 [C(OP_READ)] = { 1921 [C(RESULT_ACCESS)] = KNL_L2_READ | KNL_L2_ACCESS, 1922 [C(RESULT_MISS)] = 0, 1923 }, 1924 [C(OP_WRITE)] = { 1925 [C(RESULT_ACCESS)] = KNL_L2_WRITE | KNL_L2_ACCESS, 1926 [C(RESULT_MISS)] = KNL_L2_WRITE | KNL_L2_MISS, 1927 }, 1928 [C(OP_PREFETCH)] = { 1929 [C(RESULT_ACCESS)] = KNL_L2_PREFETCH | KNL_L2_ACCESS, 1930 [C(RESULT_MISS)] = KNL_L2_PREFETCH | KNL_L2_MISS, 1931 }, 1932 }, 1933 }; 1934 1935 /* 1936 * Used from PMIs where the LBRs are already disabled. 1937 * 1938 * This function could be called consecutively. It is required to remain in 1939 * disabled state if called consecutively. 1940 * 1941 * During consecutive calls, the same disable value will be written to related 1942 * registers, so the PMU state remains unchanged. 1943 * 1944 * intel_bts events don't coexist with intel PMU's BTS events because of 1945 * x86_add_exclusive(x86_lbr_exclusive_lbr); there's no need to keep them 1946 * disabled around intel PMU's event batching etc, only inside the PMI handler. 1947 */ 1948 static void __intel_pmu_disable_all(void) 1949 { 1950 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 1951 1952 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 1953 1954 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) 1955 intel_pmu_disable_bts(); 1956 1957 intel_pmu_pebs_disable_all(); 1958 } 1959 1960 static void intel_pmu_disable_all(void) 1961 { 1962 __intel_pmu_disable_all(); 1963 intel_pmu_lbr_disable_all(); 1964 } 1965 1966 static void __intel_pmu_enable_all(int added, bool pmi) 1967 { 1968 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 1969 1970 intel_pmu_pebs_enable_all(); 1971 intel_pmu_lbr_enable_all(pmi); 1972 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 1973 x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask); 1974 1975 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 1976 struct perf_event *event = 1977 cpuc->events[INTEL_PMC_IDX_FIXED_BTS]; 1978 1979 if (WARN_ON_ONCE(!event)) 1980 return; 1981 1982 intel_pmu_enable_bts(event->hw.config); 1983 } 1984 } 1985 1986 static void intel_pmu_enable_all(int added) 1987 { 1988 __intel_pmu_enable_all(added, false); 1989 } 1990 1991 /* 1992 * Workaround for: 1993 * Intel Errata AAK100 (model 26) 1994 * Intel Errata AAP53 (model 30) 1995 * Intel Errata BD53 (model 44) 1996 * 1997 * The official story: 1998 * These chips need to be 'reset' when adding counters by programming the 1999 * magic three (non-counting) events 0x4300B5, 0x4300D2, and 0x4300B1 either 2000 * in sequence on the same PMC or on different PMCs. 2001 * 2002 * In practise it appears some of these events do in fact count, and 2003 * we need to program all 4 events. 2004 */ 2005 static void intel_pmu_nhm_workaround(void) 2006 { 2007 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2008 static const unsigned long nhm_magic[4] = { 2009 0x4300B5, 2010 0x4300D2, 2011 0x4300B1, 2012 0x4300B1 2013 }; 2014 struct perf_event *event; 2015 int i; 2016 2017 /* 2018 * The Errata requires below steps: 2019 * 1) Clear MSR_IA32_PEBS_ENABLE and MSR_CORE_PERF_GLOBAL_CTRL; 2020 * 2) Configure 4 PERFEVTSELx with the magic events and clear 2021 * the corresponding PMCx; 2022 * 3) set bit0~bit3 of MSR_CORE_PERF_GLOBAL_CTRL; 2023 * 4) Clear MSR_CORE_PERF_GLOBAL_CTRL; 2024 * 5) Clear 4 pairs of ERFEVTSELx and PMCx; 2025 */ 2026 2027 /* 2028 * The real steps we choose are a little different from above. 2029 * A) To reduce MSR operations, we don't run step 1) as they 2030 * are already cleared before this function is called; 2031 * B) Call x86_perf_event_update to save PMCx before configuring 2032 * PERFEVTSELx with magic number; 2033 * C) With step 5), we do clear only when the PERFEVTSELx is 2034 * not used currently. 2035 * D) Call x86_perf_event_set_period to restore PMCx; 2036 */ 2037 2038 /* We always operate 4 pairs of PERF Counters */ 2039 for (i = 0; i < 4; i++) { 2040 event = cpuc->events[i]; 2041 if (event) 2042 x86_perf_event_update(event); 2043 } 2044 2045 for (i = 0; i < 4; i++) { 2046 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, nhm_magic[i]); 2047 wrmsrl(MSR_ARCH_PERFMON_PERFCTR0 + i, 0x0); 2048 } 2049 2050 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0xf); 2051 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0x0); 2052 2053 for (i = 0; i < 4; i++) { 2054 event = cpuc->events[i]; 2055 2056 if (event) { 2057 x86_perf_event_set_period(event); 2058 __x86_pmu_enable_event(&event->hw, 2059 ARCH_PERFMON_EVENTSEL_ENABLE); 2060 } else 2061 wrmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + i, 0x0); 2062 } 2063 } 2064 2065 static void intel_pmu_nhm_enable_all(int added) 2066 { 2067 if (added) 2068 intel_pmu_nhm_workaround(); 2069 intel_pmu_enable_all(added); 2070 } 2071 2072 static void intel_set_tfa(struct cpu_hw_events *cpuc, bool on) 2073 { 2074 u64 val = on ? MSR_TFA_RTM_FORCE_ABORT : 0; 2075 2076 if (cpuc->tfa_shadow != val) { 2077 cpuc->tfa_shadow = val; 2078 wrmsrl(MSR_TSX_FORCE_ABORT, val); 2079 } 2080 } 2081 2082 static void intel_tfa_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2083 { 2084 /* 2085 * We're going to use PMC3, make sure TFA is set before we touch it. 2086 */ 2087 if (cntr == 3) 2088 intel_set_tfa(cpuc, true); 2089 } 2090 2091 static void intel_tfa_pmu_enable_all(int added) 2092 { 2093 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2094 2095 /* 2096 * If we find PMC3 is no longer used when we enable the PMU, we can 2097 * clear TFA. 2098 */ 2099 if (!test_bit(3, cpuc->active_mask)) 2100 intel_set_tfa(cpuc, false); 2101 2102 intel_pmu_enable_all(added); 2103 } 2104 2105 static void enable_counter_freeze(void) 2106 { 2107 update_debugctlmsr(get_debugctlmsr() | 2108 DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI); 2109 } 2110 2111 static void disable_counter_freeze(void) 2112 { 2113 update_debugctlmsr(get_debugctlmsr() & 2114 ~DEBUGCTLMSR_FREEZE_PERFMON_ON_PMI); 2115 } 2116 2117 static inline u64 intel_pmu_get_status(void) 2118 { 2119 u64 status; 2120 2121 rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status); 2122 2123 return status; 2124 } 2125 2126 static inline void intel_pmu_ack_status(u64 ack) 2127 { 2128 wrmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, ack); 2129 } 2130 2131 static void intel_pmu_disable_fixed(struct hw_perf_event *hwc) 2132 { 2133 int idx = hwc->idx - INTEL_PMC_IDX_FIXED; 2134 u64 ctrl_val, mask; 2135 2136 mask = 0xfULL << (idx * 4); 2137 2138 rdmsrl(hwc->config_base, ctrl_val); 2139 ctrl_val &= ~mask; 2140 wrmsrl(hwc->config_base, ctrl_val); 2141 } 2142 2143 static inline bool event_is_checkpointed(struct perf_event *event) 2144 { 2145 return (event->hw.config & HSW_IN_TX_CHECKPOINTED) != 0; 2146 } 2147 2148 static void intel_pmu_disable_event(struct perf_event *event) 2149 { 2150 struct hw_perf_event *hwc = &event->hw; 2151 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2152 2153 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { 2154 intel_pmu_disable_bts(); 2155 intel_pmu_drain_bts_buffer(); 2156 return; 2157 } 2158 2159 cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx); 2160 cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx); 2161 cpuc->intel_cp_status &= ~(1ull << hwc->idx); 2162 2163 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) 2164 intel_pmu_disable_fixed(hwc); 2165 else 2166 x86_pmu_disable_event(event); 2167 2168 /* 2169 * Needs to be called after x86_pmu_disable_event, 2170 * so we don't trigger the event without PEBS bit set. 2171 */ 2172 if (unlikely(event->attr.precise_ip)) 2173 intel_pmu_pebs_disable(event); 2174 } 2175 2176 static void intel_pmu_del_event(struct perf_event *event) 2177 { 2178 if (needs_branch_stack(event)) 2179 intel_pmu_lbr_del(event); 2180 if (event->attr.precise_ip) 2181 intel_pmu_pebs_del(event); 2182 } 2183 2184 static void intel_pmu_read_event(struct perf_event *event) 2185 { 2186 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) 2187 intel_pmu_auto_reload_read(event); 2188 else 2189 x86_perf_event_update(event); 2190 } 2191 2192 static void intel_pmu_enable_fixed(struct perf_event *event) 2193 { 2194 struct hw_perf_event *hwc = &event->hw; 2195 int idx = hwc->idx - INTEL_PMC_IDX_FIXED; 2196 u64 ctrl_val, mask, bits = 0; 2197 2198 /* 2199 * Enable IRQ generation (0x8), if not PEBS, 2200 * and enable ring-3 counting (0x2) and ring-0 counting (0x1) 2201 * if requested: 2202 */ 2203 if (!event->attr.precise_ip) 2204 bits |= 0x8; 2205 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) 2206 bits |= 0x2; 2207 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) 2208 bits |= 0x1; 2209 2210 /* 2211 * ANY bit is supported in v3 and up 2212 */ 2213 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) 2214 bits |= 0x4; 2215 2216 bits <<= (idx * 4); 2217 mask = 0xfULL << (idx * 4); 2218 2219 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { 2220 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2221 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2222 } 2223 2224 rdmsrl(hwc->config_base, ctrl_val); 2225 ctrl_val &= ~mask; 2226 ctrl_val |= bits; 2227 wrmsrl(hwc->config_base, ctrl_val); 2228 } 2229 2230 static void intel_pmu_enable_event(struct perf_event *event) 2231 { 2232 struct hw_perf_event *hwc = &event->hw; 2233 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2234 2235 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { 2236 if (!__this_cpu_read(cpu_hw_events.enabled)) 2237 return; 2238 2239 intel_pmu_enable_bts(hwc->config); 2240 return; 2241 } 2242 2243 if (event->attr.exclude_host) 2244 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); 2245 if (event->attr.exclude_guest) 2246 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); 2247 2248 if (unlikely(event_is_checkpointed(event))) 2249 cpuc->intel_cp_status |= (1ull << hwc->idx); 2250 2251 if (unlikely(event->attr.precise_ip)) 2252 intel_pmu_pebs_enable(event); 2253 2254 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { 2255 intel_pmu_enable_fixed(event); 2256 return; 2257 } 2258 2259 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 2260 } 2261 2262 static void intel_pmu_add_event(struct perf_event *event) 2263 { 2264 if (event->attr.precise_ip) 2265 intel_pmu_pebs_add(event); 2266 if (needs_branch_stack(event)) 2267 intel_pmu_lbr_add(event); 2268 } 2269 2270 /* 2271 * Save and restart an expired event. Called by NMI contexts, 2272 * so it has to be careful about preempting normal event ops: 2273 */ 2274 int intel_pmu_save_and_restart(struct perf_event *event) 2275 { 2276 x86_perf_event_update(event); 2277 /* 2278 * For a checkpointed counter always reset back to 0. This 2279 * avoids a situation where the counter overflows, aborts the 2280 * transaction and is then set back to shortly before the 2281 * overflow, and overflows and aborts again. 2282 */ 2283 if (unlikely(event_is_checkpointed(event))) { 2284 /* No race with NMIs because the counter should not be armed */ 2285 wrmsrl(event->hw.event_base, 0); 2286 local64_set(&event->hw.prev_count, 0); 2287 } 2288 return x86_perf_event_set_period(event); 2289 } 2290 2291 static void intel_pmu_reset(void) 2292 { 2293 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); 2294 unsigned long flags; 2295 int idx; 2296 2297 if (!x86_pmu.num_counters) 2298 return; 2299 2300 local_irq_save(flags); 2301 2302 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); 2303 2304 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 2305 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); 2306 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); 2307 } 2308 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) 2309 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); 2310 2311 if (ds) 2312 ds->bts_index = ds->bts_buffer_base; 2313 2314 /* Ack all overflows and disable fixed counters */ 2315 if (x86_pmu.version >= 2) { 2316 intel_pmu_ack_status(intel_pmu_get_status()); 2317 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2318 } 2319 2320 /* Reset LBRs and LBR freezing */ 2321 if (x86_pmu.lbr_nr) { 2322 update_debugctlmsr(get_debugctlmsr() & 2323 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); 2324 } 2325 2326 local_irq_restore(flags); 2327 } 2328 2329 static int handle_pmi_common(struct pt_regs *regs, u64 status) 2330 { 2331 struct perf_sample_data data; 2332 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2333 int bit; 2334 int handled = 0; 2335 2336 inc_irq_stat(apic_perf_irqs); 2337 2338 /* 2339 * Ignore a range of extra bits in status that do not indicate 2340 * overflow by themselves. 2341 */ 2342 status &= ~(GLOBAL_STATUS_COND_CHG | 2343 GLOBAL_STATUS_ASIF | 2344 GLOBAL_STATUS_LBRS_FROZEN); 2345 if (!status) 2346 return 0; 2347 /* 2348 * In case multiple PEBS events are sampled at the same time, 2349 * it is possible to have GLOBAL_STATUS bit 62 set indicating 2350 * PEBS buffer overflow and also seeing at most 3 PEBS counters 2351 * having their bits set in the status register. This is a sign 2352 * that there was at least one PEBS record pending at the time 2353 * of the PMU interrupt. PEBS counters must only be processed 2354 * via the drain_pebs() calls and not via the regular sample 2355 * processing loop coming after that the function, otherwise 2356 * phony regular samples may be generated in the sampling buffer 2357 * not marked with the EXACT tag. Another possibility is to have 2358 * one PEBS event and at least one non-PEBS event whic hoverflows 2359 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will 2360 * not be set, yet the overflow status bit for the PEBS counter will 2361 * be on Skylake. 2362 * 2363 * To avoid this problem, we systematically ignore the PEBS-enabled 2364 * counters from the GLOBAL_STATUS mask and we always process PEBS 2365 * events via drain_pebs(). 2366 */ 2367 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 2368 status &= ~cpuc->pebs_enabled; 2369 else 2370 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 2371 2372 /* 2373 * PEBS overflow sets bit 62 in the global status register 2374 */ 2375 if (__test_and_clear_bit(62, (unsigned long *)&status)) { 2376 handled++; 2377 x86_pmu.drain_pebs(regs); 2378 status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI; 2379 } 2380 2381 /* 2382 * Intel PT 2383 */ 2384 if (__test_and_clear_bit(55, (unsigned long *)&status)) { 2385 handled++; 2386 if (unlikely(perf_guest_cbs && perf_guest_cbs->is_in_guest() && 2387 perf_guest_cbs->handle_intel_pt_intr)) 2388 perf_guest_cbs->handle_intel_pt_intr(); 2389 else 2390 intel_pt_interrupt(); 2391 } 2392 2393 /* 2394 * Checkpointed counters can lead to 'spurious' PMIs because the 2395 * rollback caused by the PMI will have cleared the overflow status 2396 * bit. Therefore always force probe these counters. 2397 */ 2398 status |= cpuc->intel_cp_status; 2399 2400 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { 2401 struct perf_event *event = cpuc->events[bit]; 2402 2403 handled++; 2404 2405 if (!test_bit(bit, cpuc->active_mask)) 2406 continue; 2407 2408 if (!intel_pmu_save_and_restart(event)) 2409 continue; 2410 2411 perf_sample_data_init(&data, 0, event->hw.last_period); 2412 2413 if (has_branch_stack(event)) 2414 data.br_stack = &cpuc->lbr_stack; 2415 2416 if (perf_event_overflow(event, &data, regs)) 2417 x86_pmu_stop(event, 0); 2418 } 2419 2420 return handled; 2421 } 2422 2423 static bool disable_counter_freezing = true; 2424 static int __init intel_perf_counter_freezing_setup(char *s) 2425 { 2426 bool res; 2427 2428 if (kstrtobool(s, &res)) 2429 return -EINVAL; 2430 2431 disable_counter_freezing = !res; 2432 return 1; 2433 } 2434 __setup("perf_v4_pmi=", intel_perf_counter_freezing_setup); 2435 2436 /* 2437 * Simplified handler for Arch Perfmon v4: 2438 * - We rely on counter freezing/unfreezing to enable/disable the PMU. 2439 * This is done automatically on PMU ack. 2440 * - Ack the PMU only after the APIC. 2441 */ 2442 2443 static int intel_pmu_handle_irq_v4(struct pt_regs *regs) 2444 { 2445 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2446 int handled = 0; 2447 bool bts = false; 2448 u64 status; 2449 int pmu_enabled = cpuc->enabled; 2450 int loops = 0; 2451 2452 /* PMU has been disabled because of counter freezing */ 2453 cpuc->enabled = 0; 2454 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 2455 bts = true; 2456 intel_bts_disable_local(); 2457 handled = intel_pmu_drain_bts_buffer(); 2458 handled += intel_bts_interrupt(); 2459 } 2460 status = intel_pmu_get_status(); 2461 if (!status) 2462 goto done; 2463 again: 2464 intel_pmu_lbr_read(); 2465 if (++loops > 100) { 2466 static bool warned; 2467 2468 if (!warned) { 2469 WARN(1, "perfevents: irq loop stuck!\n"); 2470 perf_event_print_debug(); 2471 warned = true; 2472 } 2473 intel_pmu_reset(); 2474 goto done; 2475 } 2476 2477 2478 handled += handle_pmi_common(regs, status); 2479 done: 2480 /* Ack the PMI in the APIC */ 2481 apic_write(APIC_LVTPC, APIC_DM_NMI); 2482 2483 /* 2484 * The counters start counting immediately while ack the status. 2485 * Make it as close as possible to IRET. This avoids bogus 2486 * freezing on Skylake CPUs. 2487 */ 2488 if (status) { 2489 intel_pmu_ack_status(status); 2490 } else { 2491 /* 2492 * CPU may issues two PMIs very close to each other. 2493 * When the PMI handler services the first one, the 2494 * GLOBAL_STATUS is already updated to reflect both. 2495 * When it IRETs, the second PMI is immediately 2496 * handled and it sees clear status. At the meantime, 2497 * there may be a third PMI, because the freezing bit 2498 * isn't set since the ack in first PMI handlers. 2499 * Double check if there is more work to be done. 2500 */ 2501 status = intel_pmu_get_status(); 2502 if (status) 2503 goto again; 2504 } 2505 2506 if (bts) 2507 intel_bts_enable_local(); 2508 cpuc->enabled = pmu_enabled; 2509 return handled; 2510 } 2511 2512 /* 2513 * This handler is triggered by the local APIC, so the APIC IRQ handling 2514 * rules apply: 2515 */ 2516 static int intel_pmu_handle_irq(struct pt_regs *regs) 2517 { 2518 struct cpu_hw_events *cpuc; 2519 int loops; 2520 u64 status; 2521 int handled; 2522 int pmu_enabled; 2523 2524 cpuc = this_cpu_ptr(&cpu_hw_events); 2525 2526 /* 2527 * Save the PMU state. 2528 * It needs to be restored when leaving the handler. 2529 */ 2530 pmu_enabled = cpuc->enabled; 2531 /* 2532 * No known reason to not always do late ACK, 2533 * but just in case do it opt-in. 2534 */ 2535 if (!x86_pmu.late_ack) 2536 apic_write(APIC_LVTPC, APIC_DM_NMI); 2537 intel_bts_disable_local(); 2538 cpuc->enabled = 0; 2539 __intel_pmu_disable_all(); 2540 handled = intel_pmu_drain_bts_buffer(); 2541 handled += intel_bts_interrupt(); 2542 status = intel_pmu_get_status(); 2543 if (!status) 2544 goto done; 2545 2546 loops = 0; 2547 again: 2548 intel_pmu_lbr_read(); 2549 intel_pmu_ack_status(status); 2550 if (++loops > 100) { 2551 static bool warned; 2552 2553 if (!warned) { 2554 WARN(1, "perfevents: irq loop stuck!\n"); 2555 perf_event_print_debug(); 2556 warned = true; 2557 } 2558 intel_pmu_reset(); 2559 goto done; 2560 } 2561 2562 handled += handle_pmi_common(regs, status); 2563 2564 /* 2565 * Repeat if there is more work to be done: 2566 */ 2567 status = intel_pmu_get_status(); 2568 if (status) 2569 goto again; 2570 2571 done: 2572 /* Only restore PMU state when it's active. See x86_pmu_disable(). */ 2573 cpuc->enabled = pmu_enabled; 2574 if (pmu_enabled) 2575 __intel_pmu_enable_all(0, true); 2576 intel_bts_enable_local(); 2577 2578 /* 2579 * Only unmask the NMI after the overflow counters 2580 * have been reset. This avoids spurious NMIs on 2581 * Haswell CPUs. 2582 */ 2583 if (x86_pmu.late_ack) 2584 apic_write(APIC_LVTPC, APIC_DM_NMI); 2585 return handled; 2586 } 2587 2588 static struct event_constraint * 2589 intel_bts_constraints(struct perf_event *event) 2590 { 2591 if (unlikely(intel_pmu_has_bts(event))) 2592 return &bts_constraint; 2593 2594 return NULL; 2595 } 2596 2597 static int intel_alt_er(int idx, u64 config) 2598 { 2599 int alt_idx = idx; 2600 2601 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) 2602 return idx; 2603 2604 if (idx == EXTRA_REG_RSP_0) 2605 alt_idx = EXTRA_REG_RSP_1; 2606 2607 if (idx == EXTRA_REG_RSP_1) 2608 alt_idx = EXTRA_REG_RSP_0; 2609 2610 if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask) 2611 return idx; 2612 2613 return alt_idx; 2614 } 2615 2616 static void intel_fixup_er(struct perf_event *event, int idx) 2617 { 2618 event->hw.extra_reg.idx = idx; 2619 2620 if (idx == EXTRA_REG_RSP_0) { 2621 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 2622 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event; 2623 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; 2624 } else if (idx == EXTRA_REG_RSP_1) { 2625 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 2626 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event; 2627 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; 2628 } 2629 } 2630 2631 /* 2632 * manage allocation of shared extra msr for certain events 2633 * 2634 * sharing can be: 2635 * per-cpu: to be shared between the various events on a single PMU 2636 * per-core: per-cpu + shared by HT threads 2637 */ 2638 static struct event_constraint * 2639 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, 2640 struct perf_event *event, 2641 struct hw_perf_event_extra *reg) 2642 { 2643 struct event_constraint *c = &emptyconstraint; 2644 struct er_account *era; 2645 unsigned long flags; 2646 int idx = reg->idx; 2647 2648 /* 2649 * reg->alloc can be set due to existing state, so for fake cpuc we 2650 * need to ignore this, otherwise we might fail to allocate proper fake 2651 * state for this extra reg constraint. Also see the comment below. 2652 */ 2653 if (reg->alloc && !cpuc->is_fake) 2654 return NULL; /* call x86_get_event_constraint() */ 2655 2656 again: 2657 era = &cpuc->shared_regs->regs[idx]; 2658 /* 2659 * we use spin_lock_irqsave() to avoid lockdep issues when 2660 * passing a fake cpuc 2661 */ 2662 raw_spin_lock_irqsave(&era->lock, flags); 2663 2664 if (!atomic_read(&era->ref) || era->config == reg->config) { 2665 2666 /* 2667 * If its a fake cpuc -- as per validate_{group,event}() we 2668 * shouldn't touch event state and we can avoid doing so 2669 * since both will only call get_event_constraints() once 2670 * on each event, this avoids the need for reg->alloc. 2671 * 2672 * Not doing the ER fixup will only result in era->reg being 2673 * wrong, but since we won't actually try and program hardware 2674 * this isn't a problem either. 2675 */ 2676 if (!cpuc->is_fake) { 2677 if (idx != reg->idx) 2678 intel_fixup_er(event, idx); 2679 2680 /* 2681 * x86_schedule_events() can call get_event_constraints() 2682 * multiple times on events in the case of incremental 2683 * scheduling(). reg->alloc ensures we only do the ER 2684 * allocation once. 2685 */ 2686 reg->alloc = 1; 2687 } 2688 2689 /* lock in msr value */ 2690 era->config = reg->config; 2691 era->reg = reg->reg; 2692 2693 /* one more user */ 2694 atomic_inc(&era->ref); 2695 2696 /* 2697 * need to call x86_get_event_constraint() 2698 * to check if associated event has constraints 2699 */ 2700 c = NULL; 2701 } else { 2702 idx = intel_alt_er(idx, reg->config); 2703 if (idx != reg->idx) { 2704 raw_spin_unlock_irqrestore(&era->lock, flags); 2705 goto again; 2706 } 2707 } 2708 raw_spin_unlock_irqrestore(&era->lock, flags); 2709 2710 return c; 2711 } 2712 2713 static void 2714 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, 2715 struct hw_perf_event_extra *reg) 2716 { 2717 struct er_account *era; 2718 2719 /* 2720 * Only put constraint if extra reg was actually allocated. Also takes 2721 * care of event which do not use an extra shared reg. 2722 * 2723 * Also, if this is a fake cpuc we shouldn't touch any event state 2724 * (reg->alloc) and we don't care about leaving inconsistent cpuc state 2725 * either since it'll be thrown out. 2726 */ 2727 if (!reg->alloc || cpuc->is_fake) 2728 return; 2729 2730 era = &cpuc->shared_regs->regs[reg->idx]; 2731 2732 /* one fewer user */ 2733 atomic_dec(&era->ref); 2734 2735 /* allocate again next time */ 2736 reg->alloc = 0; 2737 } 2738 2739 static struct event_constraint * 2740 intel_shared_regs_constraints(struct cpu_hw_events *cpuc, 2741 struct perf_event *event) 2742 { 2743 struct event_constraint *c = NULL, *d; 2744 struct hw_perf_event_extra *xreg, *breg; 2745 2746 xreg = &event->hw.extra_reg; 2747 if (xreg->idx != EXTRA_REG_NONE) { 2748 c = __intel_shared_reg_get_constraints(cpuc, event, xreg); 2749 if (c == &emptyconstraint) 2750 return c; 2751 } 2752 breg = &event->hw.branch_reg; 2753 if (breg->idx != EXTRA_REG_NONE) { 2754 d = __intel_shared_reg_get_constraints(cpuc, event, breg); 2755 if (d == &emptyconstraint) { 2756 __intel_shared_reg_put_constraints(cpuc, xreg); 2757 c = d; 2758 } 2759 } 2760 return c; 2761 } 2762 2763 struct event_constraint * 2764 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 2765 struct perf_event *event) 2766 { 2767 struct event_constraint *c; 2768 2769 if (x86_pmu.event_constraints) { 2770 for_each_event_constraint(c, x86_pmu.event_constraints) { 2771 if (constraint_match(c, event->hw.config)) { 2772 event->hw.flags |= c->flags; 2773 return c; 2774 } 2775 } 2776 } 2777 2778 return &unconstrained; 2779 } 2780 2781 static struct event_constraint * 2782 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 2783 struct perf_event *event) 2784 { 2785 struct event_constraint *c; 2786 2787 c = intel_bts_constraints(event); 2788 if (c) 2789 return c; 2790 2791 c = intel_shared_regs_constraints(cpuc, event); 2792 if (c) 2793 return c; 2794 2795 c = intel_pebs_constraints(event); 2796 if (c) 2797 return c; 2798 2799 return x86_get_event_constraints(cpuc, idx, event); 2800 } 2801 2802 static void 2803 intel_start_scheduling(struct cpu_hw_events *cpuc) 2804 { 2805 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2806 struct intel_excl_states *xl; 2807 int tid = cpuc->excl_thread_id; 2808 2809 /* 2810 * nothing needed if in group validation mode 2811 */ 2812 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2813 return; 2814 2815 /* 2816 * no exclusion needed 2817 */ 2818 if (WARN_ON_ONCE(!excl_cntrs)) 2819 return; 2820 2821 xl = &excl_cntrs->states[tid]; 2822 2823 xl->sched_started = true; 2824 /* 2825 * lock shared state until we are done scheduling 2826 * in stop_event_scheduling() 2827 * makes scheduling appear as a transaction 2828 */ 2829 raw_spin_lock(&excl_cntrs->lock); 2830 } 2831 2832 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2833 { 2834 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2835 struct event_constraint *c = cpuc->event_constraint[idx]; 2836 struct intel_excl_states *xl; 2837 int tid = cpuc->excl_thread_id; 2838 2839 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2840 return; 2841 2842 if (WARN_ON_ONCE(!excl_cntrs)) 2843 return; 2844 2845 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) 2846 return; 2847 2848 xl = &excl_cntrs->states[tid]; 2849 2850 lockdep_assert_held(&excl_cntrs->lock); 2851 2852 if (c->flags & PERF_X86_EVENT_EXCL) 2853 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; 2854 else 2855 xl->state[cntr] = INTEL_EXCL_SHARED; 2856 } 2857 2858 static void 2859 intel_stop_scheduling(struct cpu_hw_events *cpuc) 2860 { 2861 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2862 struct intel_excl_states *xl; 2863 int tid = cpuc->excl_thread_id; 2864 2865 /* 2866 * nothing needed if in group validation mode 2867 */ 2868 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2869 return; 2870 /* 2871 * no exclusion needed 2872 */ 2873 if (WARN_ON_ONCE(!excl_cntrs)) 2874 return; 2875 2876 xl = &excl_cntrs->states[tid]; 2877 2878 xl->sched_started = false; 2879 /* 2880 * release shared state lock (acquired in intel_start_scheduling()) 2881 */ 2882 raw_spin_unlock(&excl_cntrs->lock); 2883 } 2884 2885 static struct event_constraint * 2886 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx) 2887 { 2888 WARN_ON_ONCE(!cpuc->constraint_list); 2889 2890 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { 2891 struct event_constraint *cx; 2892 2893 /* 2894 * grab pre-allocated constraint entry 2895 */ 2896 cx = &cpuc->constraint_list[idx]; 2897 2898 /* 2899 * initialize dynamic constraint 2900 * with static constraint 2901 */ 2902 *cx = *c; 2903 2904 /* 2905 * mark constraint as dynamic 2906 */ 2907 cx->flags |= PERF_X86_EVENT_DYNAMIC; 2908 c = cx; 2909 } 2910 2911 return c; 2912 } 2913 2914 static struct event_constraint * 2915 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, 2916 int idx, struct event_constraint *c) 2917 { 2918 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2919 struct intel_excl_states *xlo; 2920 int tid = cpuc->excl_thread_id; 2921 int is_excl, i, w; 2922 2923 /* 2924 * validating a group does not require 2925 * enforcing cross-thread exclusion 2926 */ 2927 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2928 return c; 2929 2930 /* 2931 * no exclusion needed 2932 */ 2933 if (WARN_ON_ONCE(!excl_cntrs)) 2934 return c; 2935 2936 /* 2937 * because we modify the constraint, we need 2938 * to make a copy. Static constraints come 2939 * from static const tables. 2940 * 2941 * only needed when constraint has not yet 2942 * been cloned (marked dynamic) 2943 */ 2944 c = dyn_constraint(cpuc, c, idx); 2945 2946 /* 2947 * From here on, the constraint is dynamic. 2948 * Either it was just allocated above, or it 2949 * was allocated during a earlier invocation 2950 * of this function 2951 */ 2952 2953 /* 2954 * state of sibling HT 2955 */ 2956 xlo = &excl_cntrs->states[tid ^ 1]; 2957 2958 /* 2959 * event requires exclusive counter access 2960 * across HT threads 2961 */ 2962 is_excl = c->flags & PERF_X86_EVENT_EXCL; 2963 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { 2964 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; 2965 if (!cpuc->n_excl++) 2966 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); 2967 } 2968 2969 /* 2970 * Modify static constraint with current dynamic 2971 * state of thread 2972 * 2973 * EXCLUSIVE: sibling counter measuring exclusive event 2974 * SHARED : sibling counter measuring non-exclusive event 2975 * UNUSED : sibling counter unused 2976 */ 2977 w = c->weight; 2978 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { 2979 /* 2980 * exclusive event in sibling counter 2981 * our corresponding counter cannot be used 2982 * regardless of our event 2983 */ 2984 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) { 2985 __clear_bit(i, c->idxmsk); 2986 w--; 2987 continue; 2988 } 2989 /* 2990 * if measuring an exclusive event, sibling 2991 * measuring non-exclusive, then counter cannot 2992 * be used 2993 */ 2994 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) { 2995 __clear_bit(i, c->idxmsk); 2996 w--; 2997 continue; 2998 } 2999 } 3000 3001 /* 3002 * if we return an empty mask, then switch 3003 * back to static empty constraint to avoid 3004 * the cost of freeing later on 3005 */ 3006 if (!w) 3007 c = &emptyconstraint; 3008 3009 c->weight = w; 3010 3011 return c; 3012 } 3013 3014 static struct event_constraint * 3015 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3016 struct perf_event *event) 3017 { 3018 struct event_constraint *c1, *c2; 3019 3020 c1 = cpuc->event_constraint[idx]; 3021 3022 /* 3023 * first time only 3024 * - static constraint: no change across incremental scheduling calls 3025 * - dynamic constraint: handled by intel_get_excl_constraints() 3026 */ 3027 c2 = __intel_get_event_constraints(cpuc, idx, event); 3028 if (c1) { 3029 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC)); 3030 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); 3031 c1->weight = c2->weight; 3032 c2 = c1; 3033 } 3034 3035 if (cpuc->excl_cntrs) 3036 return intel_get_excl_constraints(cpuc, event, idx, c2); 3037 3038 return c2; 3039 } 3040 3041 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, 3042 struct perf_event *event) 3043 { 3044 struct hw_perf_event *hwc = &event->hw; 3045 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3046 int tid = cpuc->excl_thread_id; 3047 struct intel_excl_states *xl; 3048 3049 /* 3050 * nothing needed if in group validation mode 3051 */ 3052 if (cpuc->is_fake) 3053 return; 3054 3055 if (WARN_ON_ONCE(!excl_cntrs)) 3056 return; 3057 3058 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { 3059 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; 3060 if (!--cpuc->n_excl) 3061 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); 3062 } 3063 3064 /* 3065 * If event was actually assigned, then mark the counter state as 3066 * unused now. 3067 */ 3068 if (hwc->idx >= 0) { 3069 xl = &excl_cntrs->states[tid]; 3070 3071 /* 3072 * put_constraint may be called from x86_schedule_events() 3073 * which already has the lock held so here make locking 3074 * conditional. 3075 */ 3076 if (!xl->sched_started) 3077 raw_spin_lock(&excl_cntrs->lock); 3078 3079 xl->state[hwc->idx] = INTEL_EXCL_UNUSED; 3080 3081 if (!xl->sched_started) 3082 raw_spin_unlock(&excl_cntrs->lock); 3083 } 3084 } 3085 3086 static void 3087 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, 3088 struct perf_event *event) 3089 { 3090 struct hw_perf_event_extra *reg; 3091 3092 reg = &event->hw.extra_reg; 3093 if (reg->idx != EXTRA_REG_NONE) 3094 __intel_shared_reg_put_constraints(cpuc, reg); 3095 3096 reg = &event->hw.branch_reg; 3097 if (reg->idx != EXTRA_REG_NONE) 3098 __intel_shared_reg_put_constraints(cpuc, reg); 3099 } 3100 3101 static void intel_put_event_constraints(struct cpu_hw_events *cpuc, 3102 struct perf_event *event) 3103 { 3104 intel_put_shared_regs_event_constraints(cpuc, event); 3105 3106 /* 3107 * is PMU has exclusive counter restrictions, then 3108 * all events are subject to and must call the 3109 * put_excl_constraints() routine 3110 */ 3111 if (cpuc->excl_cntrs) 3112 intel_put_excl_constraints(cpuc, event); 3113 } 3114 3115 static void intel_pebs_aliases_core2(struct perf_event *event) 3116 { 3117 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3118 /* 3119 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3120 * (0x003c) so that we can use it with PEBS. 3121 * 3122 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3123 * PEBS capable. However we can use INST_RETIRED.ANY_P 3124 * (0x00c0), which is a PEBS capable event, to get the same 3125 * count. 3126 * 3127 * INST_RETIRED.ANY_P counts the number of cycles that retires 3128 * CNTMASK instructions. By setting CNTMASK to a value (16) 3129 * larger than the maximum number of instructions that can be 3130 * retired per cycle (4) and then inverting the condition, we 3131 * count all cycles that retire 16 or less instructions, which 3132 * is every cycle. 3133 * 3134 * Thereby we gain a PEBS capable cycle counter. 3135 */ 3136 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); 3137 3138 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3139 event->hw.config = alt_config; 3140 } 3141 } 3142 3143 static void intel_pebs_aliases_snb(struct perf_event *event) 3144 { 3145 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3146 /* 3147 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3148 * (0x003c) so that we can use it with PEBS. 3149 * 3150 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3151 * PEBS capable. However we can use UOPS_RETIRED.ALL 3152 * (0x01c2), which is a PEBS capable event, to get the same 3153 * count. 3154 * 3155 * UOPS_RETIRED.ALL counts the number of cycles that retires 3156 * CNTMASK micro-ops. By setting CNTMASK to a value (16) 3157 * larger than the maximum number of micro-ops that can be 3158 * retired per cycle (4) and then inverting the condition, we 3159 * count all cycles that retire 16 or less micro-ops, which 3160 * is every cycle. 3161 * 3162 * Thereby we gain a PEBS capable cycle counter. 3163 */ 3164 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); 3165 3166 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3167 event->hw.config = alt_config; 3168 } 3169 } 3170 3171 static void intel_pebs_aliases_precdist(struct perf_event *event) 3172 { 3173 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3174 /* 3175 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3176 * (0x003c) so that we can use it with PEBS. 3177 * 3178 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3179 * PEBS capable. However we can use INST_RETIRED.PREC_DIST 3180 * (0x01c0), which is a PEBS capable event, to get the same 3181 * count. 3182 * 3183 * The PREC_DIST event has special support to minimize sample 3184 * shadowing effects. One drawback is that it can be 3185 * only programmed on counter 1, but that seems like an 3186 * acceptable trade off. 3187 */ 3188 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16); 3189 3190 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3191 event->hw.config = alt_config; 3192 } 3193 } 3194 3195 static void intel_pebs_aliases_ivb(struct perf_event *event) 3196 { 3197 if (event->attr.precise_ip < 3) 3198 return intel_pebs_aliases_snb(event); 3199 return intel_pebs_aliases_precdist(event); 3200 } 3201 3202 static void intel_pebs_aliases_skl(struct perf_event *event) 3203 { 3204 if (event->attr.precise_ip < 3) 3205 return intel_pebs_aliases_core2(event); 3206 return intel_pebs_aliases_precdist(event); 3207 } 3208 3209 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event) 3210 { 3211 unsigned long flags = x86_pmu.large_pebs_flags; 3212 3213 if (event->attr.use_clockid) 3214 flags &= ~PERF_SAMPLE_TIME; 3215 if (!event->attr.exclude_kernel) 3216 flags &= ~PERF_SAMPLE_REGS_USER; 3217 if (event->attr.sample_regs_user & ~PEBS_GP_REGS) 3218 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR); 3219 return flags; 3220 } 3221 3222 static int intel_pmu_bts_config(struct perf_event *event) 3223 { 3224 struct perf_event_attr *attr = &event->attr; 3225 3226 if (unlikely(intel_pmu_has_bts(event))) { 3227 /* BTS is not supported by this architecture. */ 3228 if (!x86_pmu.bts_active) 3229 return -EOPNOTSUPP; 3230 3231 /* BTS is currently only allowed for user-mode. */ 3232 if (!attr->exclude_kernel) 3233 return -EOPNOTSUPP; 3234 3235 /* BTS is not allowed for precise events. */ 3236 if (attr->precise_ip) 3237 return -EOPNOTSUPP; 3238 3239 /* disallow bts if conflicting events are present */ 3240 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3241 return -EBUSY; 3242 3243 event->destroy = hw_perf_lbr_event_destroy; 3244 } 3245 3246 return 0; 3247 } 3248 3249 static int core_pmu_hw_config(struct perf_event *event) 3250 { 3251 int ret = x86_pmu_hw_config(event); 3252 3253 if (ret) 3254 return ret; 3255 3256 return intel_pmu_bts_config(event); 3257 } 3258 3259 static int intel_pmu_hw_config(struct perf_event *event) 3260 { 3261 int ret = x86_pmu_hw_config(event); 3262 3263 if (ret) 3264 return ret; 3265 3266 ret = intel_pmu_bts_config(event); 3267 if (ret) 3268 return ret; 3269 3270 if (event->attr.precise_ip) { 3271 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) { 3272 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; 3273 if (!(event->attr.sample_type & 3274 ~intel_pmu_large_pebs_flags(event))) 3275 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS; 3276 } 3277 if (x86_pmu.pebs_aliases) 3278 x86_pmu.pebs_aliases(event); 3279 3280 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) 3281 event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY; 3282 } 3283 3284 if (needs_branch_stack(event)) { 3285 ret = intel_pmu_setup_lbr_filter(event); 3286 if (ret) 3287 return ret; 3288 3289 /* 3290 * BTS is set up earlier in this path, so don't account twice 3291 */ 3292 if (!unlikely(intel_pmu_has_bts(event))) { 3293 /* disallow lbr if conflicting events are present */ 3294 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3295 return -EBUSY; 3296 3297 event->destroy = hw_perf_lbr_event_destroy; 3298 } 3299 } 3300 3301 if (event->attr.type != PERF_TYPE_RAW) 3302 return 0; 3303 3304 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) 3305 return 0; 3306 3307 if (x86_pmu.version < 3) 3308 return -EINVAL; 3309 3310 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) 3311 return -EACCES; 3312 3313 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; 3314 3315 return 0; 3316 } 3317 3318 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) 3319 { 3320 if (x86_pmu.guest_get_msrs) 3321 return x86_pmu.guest_get_msrs(nr); 3322 *nr = 0; 3323 return NULL; 3324 } 3325 EXPORT_SYMBOL_GPL(perf_guest_get_msrs); 3326 3327 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) 3328 { 3329 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3330 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3331 3332 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; 3333 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; 3334 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; 3335 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 3336 arr[0].guest &= ~cpuc->pebs_enabled; 3337 else 3338 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 3339 *nr = 1; 3340 3341 if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) { 3342 /* 3343 * If PMU counter has PEBS enabled it is not enough to 3344 * disable counter on a guest entry since PEBS memory 3345 * write can overshoot guest entry and corrupt guest 3346 * memory. Disabling PEBS solves the problem. 3347 * 3348 * Don't do this if the CPU already enforces it. 3349 */ 3350 arr[1].msr = MSR_IA32_PEBS_ENABLE; 3351 arr[1].host = cpuc->pebs_enabled; 3352 arr[1].guest = 0; 3353 *nr = 2; 3354 } 3355 3356 return arr; 3357 } 3358 3359 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) 3360 { 3361 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3362 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3363 int idx; 3364 3365 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 3366 struct perf_event *event = cpuc->events[idx]; 3367 3368 arr[idx].msr = x86_pmu_config_addr(idx); 3369 arr[idx].host = arr[idx].guest = 0; 3370 3371 if (!test_bit(idx, cpuc->active_mask)) 3372 continue; 3373 3374 arr[idx].host = arr[idx].guest = 3375 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; 3376 3377 if (event->attr.exclude_host) 3378 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3379 else if (event->attr.exclude_guest) 3380 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3381 } 3382 3383 *nr = x86_pmu.num_counters; 3384 return arr; 3385 } 3386 3387 static void core_pmu_enable_event(struct perf_event *event) 3388 { 3389 if (!event->attr.exclude_host) 3390 x86_pmu_enable_event(event); 3391 } 3392 3393 static void core_pmu_enable_all(int added) 3394 { 3395 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3396 int idx; 3397 3398 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 3399 struct hw_perf_event *hwc = &cpuc->events[idx]->hw; 3400 3401 if (!test_bit(idx, cpuc->active_mask) || 3402 cpuc->events[idx]->attr.exclude_host) 3403 continue; 3404 3405 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 3406 } 3407 } 3408 3409 static int hsw_hw_config(struct perf_event *event) 3410 { 3411 int ret = intel_pmu_hw_config(event); 3412 3413 if (ret) 3414 return ret; 3415 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) 3416 return 0; 3417 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); 3418 3419 /* 3420 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with 3421 * PEBS or in ANY thread mode. Since the results are non-sensical forbid 3422 * this combination. 3423 */ 3424 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && 3425 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || 3426 event->attr.precise_ip > 0)) 3427 return -EOPNOTSUPP; 3428 3429 if (event_is_checkpointed(event)) { 3430 /* 3431 * Sampling of checkpointed events can cause situations where 3432 * the CPU constantly aborts because of a overflow, which is 3433 * then checkpointed back and ignored. Forbid checkpointing 3434 * for sampling. 3435 * 3436 * But still allow a long sampling period, so that perf stat 3437 * from KVM works. 3438 */ 3439 if (event->attr.sample_period > 0 && 3440 event->attr.sample_period < 0x7fffffff) 3441 return -EOPNOTSUPP; 3442 } 3443 return 0; 3444 } 3445 3446 static struct event_constraint counter0_constraint = 3447 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1); 3448 3449 static struct event_constraint counter2_constraint = 3450 EVENT_CONSTRAINT(0, 0x4, 0); 3451 3452 static struct event_constraint fixed0_constraint = 3453 FIXED_EVENT_CONSTRAINT(0x00c0, 0); 3454 3455 static struct event_constraint fixed0_counter0_constraint = 3456 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL); 3457 3458 static struct event_constraint * 3459 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3460 struct perf_event *event) 3461 { 3462 struct event_constraint *c; 3463 3464 c = intel_get_event_constraints(cpuc, idx, event); 3465 3466 /* Handle special quirk on in_tx_checkpointed only in counter 2 */ 3467 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { 3468 if (c->idxmsk64 & (1U << 2)) 3469 return &counter2_constraint; 3470 return &emptyconstraint; 3471 } 3472 3473 return c; 3474 } 3475 3476 static struct event_constraint * 3477 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3478 struct perf_event *event) 3479 { 3480 /* 3481 * Fixed counter 0 has less skid. 3482 * Force instruction:ppp in Fixed counter 0 3483 */ 3484 if ((event->attr.precise_ip == 3) && 3485 constraint_match(&fixed0_constraint, event->hw.config)) 3486 return &fixed0_constraint; 3487 3488 return hsw_get_event_constraints(cpuc, idx, event); 3489 } 3490 3491 static struct event_constraint * 3492 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3493 struct perf_event *event) 3494 { 3495 struct event_constraint *c; 3496 3497 /* :ppp means to do reduced skid PEBS which is PMC0 only. */ 3498 if (event->attr.precise_ip == 3) 3499 return &counter0_constraint; 3500 3501 c = intel_get_event_constraints(cpuc, idx, event); 3502 3503 return c; 3504 } 3505 3506 static struct event_constraint * 3507 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3508 struct perf_event *event) 3509 { 3510 struct event_constraint *c; 3511 3512 /* 3513 * :ppp means to do reduced skid PEBS, 3514 * which is available on PMC0 and fixed counter 0. 3515 */ 3516 if (event->attr.precise_ip == 3) { 3517 /* Force instruction:ppp on PMC0 and Fixed counter 0 */ 3518 if (constraint_match(&fixed0_constraint, event->hw.config)) 3519 return &fixed0_counter0_constraint; 3520 3521 return &counter0_constraint; 3522 } 3523 3524 c = intel_get_event_constraints(cpuc, idx, event); 3525 3526 return c; 3527 } 3528 3529 static bool allow_tsx_force_abort = true; 3530 3531 static struct event_constraint * 3532 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3533 struct perf_event *event) 3534 { 3535 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event); 3536 3537 /* 3538 * Without TFA we must not use PMC3. 3539 */ 3540 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) { 3541 c = dyn_constraint(cpuc, c, idx); 3542 c->idxmsk64 &= ~(1ULL << 3); 3543 c->weight--; 3544 } 3545 3546 return c; 3547 } 3548 3549 /* 3550 * Broadwell: 3551 * 3552 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared 3553 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine 3554 * the two to enforce a minimum period of 128 (the smallest value that has bits 3555 * 0-5 cleared and >= 100). 3556 * 3557 * Because of how the code in x86_perf_event_set_period() works, the truncation 3558 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period 3559 * to make up for the 'lost' events due to carrying the 'error' in period_left. 3560 * 3561 * Therefore the effective (average) period matches the requested period, 3562 * despite coarser hardware granularity. 3563 */ 3564 static u64 bdw_limit_period(struct perf_event *event, u64 left) 3565 { 3566 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == 3567 X86_CONFIG(.event=0xc0, .umask=0x01)) { 3568 if (left < 128) 3569 left = 128; 3570 left &= ~0x3fULL; 3571 } 3572 return left; 3573 } 3574 3575 PMU_FORMAT_ATTR(event, "config:0-7" ); 3576 PMU_FORMAT_ATTR(umask, "config:8-15" ); 3577 PMU_FORMAT_ATTR(edge, "config:18" ); 3578 PMU_FORMAT_ATTR(pc, "config:19" ); 3579 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ 3580 PMU_FORMAT_ATTR(inv, "config:23" ); 3581 PMU_FORMAT_ATTR(cmask, "config:24-31" ); 3582 PMU_FORMAT_ATTR(in_tx, "config:32"); 3583 PMU_FORMAT_ATTR(in_tx_cp, "config:33"); 3584 3585 static struct attribute *intel_arch_formats_attr[] = { 3586 &format_attr_event.attr, 3587 &format_attr_umask.attr, 3588 &format_attr_edge.attr, 3589 &format_attr_pc.attr, 3590 &format_attr_inv.attr, 3591 &format_attr_cmask.attr, 3592 NULL, 3593 }; 3594 3595 ssize_t intel_event_sysfs_show(char *page, u64 config) 3596 { 3597 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); 3598 3599 return x86_event_sysfs_show(page, config, event); 3600 } 3601 3602 static struct intel_shared_regs *allocate_shared_regs(int cpu) 3603 { 3604 struct intel_shared_regs *regs; 3605 int i; 3606 3607 regs = kzalloc_node(sizeof(struct intel_shared_regs), 3608 GFP_KERNEL, cpu_to_node(cpu)); 3609 if (regs) { 3610 /* 3611 * initialize the locks to keep lockdep happy 3612 */ 3613 for (i = 0; i < EXTRA_REG_MAX; i++) 3614 raw_spin_lock_init(®s->regs[i].lock); 3615 3616 regs->core_id = -1; 3617 } 3618 return regs; 3619 } 3620 3621 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) 3622 { 3623 struct intel_excl_cntrs *c; 3624 3625 c = kzalloc_node(sizeof(struct intel_excl_cntrs), 3626 GFP_KERNEL, cpu_to_node(cpu)); 3627 if (c) { 3628 raw_spin_lock_init(&c->lock); 3629 c->core_id = -1; 3630 } 3631 return c; 3632 } 3633 3634 3635 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 3636 { 3637 cpuc->pebs_record_size = x86_pmu.pebs_record_size; 3638 3639 if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { 3640 cpuc->shared_regs = allocate_shared_regs(cpu); 3641 if (!cpuc->shared_regs) 3642 goto err; 3643 } 3644 3645 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) { 3646 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); 3647 3648 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu)); 3649 if (!cpuc->constraint_list) 3650 goto err_shared_regs; 3651 } 3652 3653 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 3654 cpuc->excl_cntrs = allocate_excl_cntrs(cpu); 3655 if (!cpuc->excl_cntrs) 3656 goto err_constraint_list; 3657 3658 cpuc->excl_thread_id = 0; 3659 } 3660 3661 return 0; 3662 3663 err_constraint_list: 3664 kfree(cpuc->constraint_list); 3665 cpuc->constraint_list = NULL; 3666 3667 err_shared_regs: 3668 kfree(cpuc->shared_regs); 3669 cpuc->shared_regs = NULL; 3670 3671 err: 3672 return -ENOMEM; 3673 } 3674 3675 static int intel_pmu_cpu_prepare(int cpu) 3676 { 3677 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu); 3678 } 3679 3680 static void flip_smm_bit(void *data) 3681 { 3682 unsigned long set = *(unsigned long *)data; 3683 3684 if (set > 0) { 3685 msr_set_bit(MSR_IA32_DEBUGCTLMSR, 3686 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 3687 } else { 3688 msr_clear_bit(MSR_IA32_DEBUGCTLMSR, 3689 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 3690 } 3691 } 3692 3693 static void intel_pmu_cpu_starting(int cpu) 3694 { 3695 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 3696 int core_id = topology_core_id(cpu); 3697 int i; 3698 3699 init_debug_store_on_cpu(cpu); 3700 /* 3701 * Deal with CPUs that don't clear their LBRs on power-up. 3702 */ 3703 intel_pmu_lbr_reset(); 3704 3705 cpuc->lbr_sel = NULL; 3706 3707 if (x86_pmu.flags & PMU_FL_TFA) { 3708 WARN_ON_ONCE(cpuc->tfa_shadow); 3709 cpuc->tfa_shadow = ~0ULL; 3710 intel_set_tfa(cpuc, false); 3711 } 3712 3713 if (x86_pmu.version > 1) 3714 flip_smm_bit(&x86_pmu.attr_freeze_on_smi); 3715 3716 if (x86_pmu.counter_freezing) 3717 enable_counter_freeze(); 3718 3719 if (!cpuc->shared_regs) 3720 return; 3721 3722 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { 3723 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 3724 struct intel_shared_regs *pc; 3725 3726 pc = per_cpu(cpu_hw_events, i).shared_regs; 3727 if (pc && pc->core_id == core_id) { 3728 cpuc->kfree_on_online[0] = cpuc->shared_regs; 3729 cpuc->shared_regs = pc; 3730 break; 3731 } 3732 } 3733 cpuc->shared_regs->core_id = core_id; 3734 cpuc->shared_regs->refcnt++; 3735 } 3736 3737 if (x86_pmu.lbr_sel_map) 3738 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; 3739 3740 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 3741 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 3742 struct cpu_hw_events *sibling; 3743 struct intel_excl_cntrs *c; 3744 3745 sibling = &per_cpu(cpu_hw_events, i); 3746 c = sibling->excl_cntrs; 3747 if (c && c->core_id == core_id) { 3748 cpuc->kfree_on_online[1] = cpuc->excl_cntrs; 3749 cpuc->excl_cntrs = c; 3750 if (!sibling->excl_thread_id) 3751 cpuc->excl_thread_id = 1; 3752 break; 3753 } 3754 } 3755 cpuc->excl_cntrs->core_id = core_id; 3756 cpuc->excl_cntrs->refcnt++; 3757 } 3758 } 3759 3760 static void free_excl_cntrs(struct cpu_hw_events *cpuc) 3761 { 3762 struct intel_excl_cntrs *c; 3763 3764 c = cpuc->excl_cntrs; 3765 if (c) { 3766 if (c->core_id == -1 || --c->refcnt == 0) 3767 kfree(c); 3768 cpuc->excl_cntrs = NULL; 3769 } 3770 3771 kfree(cpuc->constraint_list); 3772 cpuc->constraint_list = NULL; 3773 } 3774 3775 static void intel_pmu_cpu_dying(int cpu) 3776 { 3777 fini_debug_store_on_cpu(cpu); 3778 3779 if (x86_pmu.counter_freezing) 3780 disable_counter_freeze(); 3781 } 3782 3783 void intel_cpuc_finish(struct cpu_hw_events *cpuc) 3784 { 3785 struct intel_shared_regs *pc; 3786 3787 pc = cpuc->shared_regs; 3788 if (pc) { 3789 if (pc->core_id == -1 || --pc->refcnt == 0) 3790 kfree(pc); 3791 cpuc->shared_regs = NULL; 3792 } 3793 3794 free_excl_cntrs(cpuc); 3795 } 3796 3797 static void intel_pmu_cpu_dead(int cpu) 3798 { 3799 intel_cpuc_finish(&per_cpu(cpu_hw_events, cpu)); 3800 } 3801 3802 static void intel_pmu_sched_task(struct perf_event_context *ctx, 3803 bool sched_in) 3804 { 3805 intel_pmu_pebs_sched_task(ctx, sched_in); 3806 intel_pmu_lbr_sched_task(ctx, sched_in); 3807 } 3808 3809 static int intel_pmu_check_period(struct perf_event *event, u64 value) 3810 { 3811 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0; 3812 } 3813 3814 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); 3815 3816 PMU_FORMAT_ATTR(ldlat, "config1:0-15"); 3817 3818 PMU_FORMAT_ATTR(frontend, "config1:0-23"); 3819 3820 static struct attribute *intel_arch3_formats_attr[] = { 3821 &format_attr_event.attr, 3822 &format_attr_umask.attr, 3823 &format_attr_edge.attr, 3824 &format_attr_pc.attr, 3825 &format_attr_any.attr, 3826 &format_attr_inv.attr, 3827 &format_attr_cmask.attr, 3828 NULL, 3829 }; 3830 3831 static struct attribute *hsw_format_attr[] = { 3832 &format_attr_in_tx.attr, 3833 &format_attr_in_tx_cp.attr, 3834 &format_attr_offcore_rsp.attr, 3835 &format_attr_ldlat.attr, 3836 NULL 3837 }; 3838 3839 static struct attribute *nhm_format_attr[] = { 3840 &format_attr_offcore_rsp.attr, 3841 &format_attr_ldlat.attr, 3842 NULL 3843 }; 3844 3845 static struct attribute *slm_format_attr[] = { 3846 &format_attr_offcore_rsp.attr, 3847 NULL 3848 }; 3849 3850 static struct attribute *skl_format_attr[] = { 3851 &format_attr_frontend.attr, 3852 NULL, 3853 }; 3854 3855 static __initconst const struct x86_pmu core_pmu = { 3856 .name = "core", 3857 .handle_irq = x86_pmu_handle_irq, 3858 .disable_all = x86_pmu_disable_all, 3859 .enable_all = core_pmu_enable_all, 3860 .enable = core_pmu_enable_event, 3861 .disable = x86_pmu_disable_event, 3862 .hw_config = core_pmu_hw_config, 3863 .schedule_events = x86_schedule_events, 3864 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 3865 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 3866 .event_map = intel_pmu_event_map, 3867 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 3868 .apic = 1, 3869 .large_pebs_flags = LARGE_PEBS_FLAGS, 3870 3871 /* 3872 * Intel PMCs cannot be accessed sanely above 32-bit width, 3873 * so we install an artificial 1<<31 period regardless of 3874 * the generic event period: 3875 */ 3876 .max_period = (1ULL<<31) - 1, 3877 .get_event_constraints = intel_get_event_constraints, 3878 .put_event_constraints = intel_put_event_constraints, 3879 .event_constraints = intel_core_event_constraints, 3880 .guest_get_msrs = core_guest_get_msrs, 3881 .format_attrs = intel_arch_formats_attr, 3882 .events_sysfs_show = intel_event_sysfs_show, 3883 3884 /* 3885 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs 3886 * together with PMU version 1 and thus be using core_pmu with 3887 * shared_regs. We need following callbacks here to allocate 3888 * it properly. 3889 */ 3890 .cpu_prepare = intel_pmu_cpu_prepare, 3891 .cpu_starting = intel_pmu_cpu_starting, 3892 .cpu_dying = intel_pmu_cpu_dying, 3893 .cpu_dead = intel_pmu_cpu_dead, 3894 3895 .check_period = intel_pmu_check_period, 3896 }; 3897 3898 static __initconst const struct x86_pmu intel_pmu = { 3899 .name = "Intel", 3900 .handle_irq = intel_pmu_handle_irq, 3901 .disable_all = intel_pmu_disable_all, 3902 .enable_all = intel_pmu_enable_all, 3903 .enable = intel_pmu_enable_event, 3904 .disable = intel_pmu_disable_event, 3905 .add = intel_pmu_add_event, 3906 .del = intel_pmu_del_event, 3907 .read = intel_pmu_read_event, 3908 .hw_config = intel_pmu_hw_config, 3909 .schedule_events = x86_schedule_events, 3910 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 3911 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 3912 .event_map = intel_pmu_event_map, 3913 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 3914 .apic = 1, 3915 .large_pebs_flags = LARGE_PEBS_FLAGS, 3916 /* 3917 * Intel PMCs cannot be accessed sanely above 32 bit width, 3918 * so we install an artificial 1<<31 period regardless of 3919 * the generic event period: 3920 */ 3921 .max_period = (1ULL << 31) - 1, 3922 .get_event_constraints = intel_get_event_constraints, 3923 .put_event_constraints = intel_put_event_constraints, 3924 .pebs_aliases = intel_pebs_aliases_core2, 3925 3926 .format_attrs = intel_arch3_formats_attr, 3927 .events_sysfs_show = intel_event_sysfs_show, 3928 3929 .cpu_prepare = intel_pmu_cpu_prepare, 3930 .cpu_starting = intel_pmu_cpu_starting, 3931 .cpu_dying = intel_pmu_cpu_dying, 3932 .cpu_dead = intel_pmu_cpu_dead, 3933 3934 .guest_get_msrs = intel_guest_get_msrs, 3935 .sched_task = intel_pmu_sched_task, 3936 3937 .check_period = intel_pmu_check_period, 3938 }; 3939 3940 static __init void intel_clovertown_quirk(void) 3941 { 3942 /* 3943 * PEBS is unreliable due to: 3944 * 3945 * AJ67 - PEBS may experience CPL leaks 3946 * AJ68 - PEBS PMI may be delayed by one event 3947 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] 3948 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS 3949 * 3950 * AJ67 could be worked around by restricting the OS/USR flags. 3951 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. 3952 * 3953 * AJ106 could possibly be worked around by not allowing LBR 3954 * usage from PEBS, including the fixup. 3955 * AJ68 could possibly be worked around by always programming 3956 * a pebs_event_reset[0] value and coping with the lost events. 3957 * 3958 * But taken together it might just make sense to not enable PEBS on 3959 * these chips. 3960 */ 3961 pr_warn("PEBS disabled due to CPU errata\n"); 3962 x86_pmu.pebs = 0; 3963 x86_pmu.pebs_constraints = NULL; 3964 } 3965 3966 static const struct x86_cpu_desc isolation_ucodes[] = { 3967 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_CORE, 3, 0x0000001f), 3968 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_ULT, 1, 0x0000001e), 3969 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_GT3E, 1, 0x00000015), 3970 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037), 3971 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a), 3972 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_CORE, 4, 0x00000023), 3973 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_GT3E, 1, 0x00000014), 3974 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 2, 0x00000010), 3975 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 3, 0x07000009), 3976 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 4, 0x0f000009), 3977 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 5, 0x0e000002), 3978 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 2, 0x0b000014), 3979 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021), 3980 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000), 3981 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_MOBILE, 3, 0x0000007c), 3982 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_DESKTOP, 3, 0x0000007c), 3983 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 9, 0x0000004e), 3984 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 9, 0x0000004e), 3985 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 10, 0x0000004e), 3986 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 11, 0x0000004e), 3987 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 12, 0x0000004e), 3988 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 10, 0x0000004e), 3989 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 11, 0x0000004e), 3990 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 12, 0x0000004e), 3991 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 13, 0x0000004e), 3992 {} 3993 }; 3994 3995 static void intel_check_pebs_isolation(void) 3996 { 3997 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes); 3998 } 3999 4000 static __init void intel_pebs_isolation_quirk(void) 4001 { 4002 WARN_ON_ONCE(x86_pmu.check_microcode); 4003 x86_pmu.check_microcode = intel_check_pebs_isolation; 4004 intel_check_pebs_isolation(); 4005 } 4006 4007 static const struct x86_cpu_desc pebs_ucodes[] = { 4008 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028), 4009 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618), 4010 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c), 4011 {} 4012 }; 4013 4014 static bool intel_snb_pebs_broken(void) 4015 { 4016 return !x86_cpu_has_min_microcode_rev(pebs_ucodes); 4017 } 4018 4019 static void intel_snb_check_microcode(void) 4020 { 4021 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken) 4022 return; 4023 4024 /* 4025 * Serialized by the microcode lock.. 4026 */ 4027 if (x86_pmu.pebs_broken) { 4028 pr_info("PEBS enabled due to microcode update\n"); 4029 x86_pmu.pebs_broken = 0; 4030 } else { 4031 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); 4032 x86_pmu.pebs_broken = 1; 4033 } 4034 } 4035 4036 static bool is_lbr_from(unsigned long msr) 4037 { 4038 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr; 4039 4040 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr; 4041 } 4042 4043 /* 4044 * Under certain circumstances, access certain MSR may cause #GP. 4045 * The function tests if the input MSR can be safely accessed. 4046 */ 4047 static bool check_msr(unsigned long msr, u64 mask) 4048 { 4049 u64 val_old, val_new, val_tmp; 4050 4051 /* 4052 * Disable the check for real HW, so we don't 4053 * mess with potentionaly enabled registers: 4054 */ 4055 if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) 4056 return true; 4057 4058 /* 4059 * Read the current value, change it and read it back to see if it 4060 * matches, this is needed to detect certain hardware emulators 4061 * (qemu/kvm) that don't trap on the MSR access and always return 0s. 4062 */ 4063 if (rdmsrl_safe(msr, &val_old)) 4064 return false; 4065 4066 /* 4067 * Only change the bits which can be updated by wrmsrl. 4068 */ 4069 val_tmp = val_old ^ mask; 4070 4071 if (is_lbr_from(msr)) 4072 val_tmp = lbr_from_signext_quirk_wr(val_tmp); 4073 4074 if (wrmsrl_safe(msr, val_tmp) || 4075 rdmsrl_safe(msr, &val_new)) 4076 return false; 4077 4078 /* 4079 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value 4080 * should equal rdmsrl()'s even with the quirk. 4081 */ 4082 if (val_new != val_tmp) 4083 return false; 4084 4085 if (is_lbr_from(msr)) 4086 val_old = lbr_from_signext_quirk_wr(val_old); 4087 4088 /* Here it's sure that the MSR can be safely accessed. 4089 * Restore the old value and return. 4090 */ 4091 wrmsrl(msr, val_old); 4092 4093 return true; 4094 } 4095 4096 static __init void intel_sandybridge_quirk(void) 4097 { 4098 x86_pmu.check_microcode = intel_snb_check_microcode; 4099 cpus_read_lock(); 4100 intel_snb_check_microcode(); 4101 cpus_read_unlock(); 4102 } 4103 4104 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { 4105 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, 4106 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, 4107 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, 4108 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, 4109 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, 4110 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, 4111 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, 4112 }; 4113 4114 static __init void intel_arch_events_quirk(void) 4115 { 4116 int bit; 4117 4118 /* disable event that reported as not presend by cpuid */ 4119 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { 4120 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; 4121 pr_warn("CPUID marked event: \'%s\' unavailable\n", 4122 intel_arch_events_map[bit].name); 4123 } 4124 } 4125 4126 static __init void intel_nehalem_quirk(void) 4127 { 4128 union cpuid10_ebx ebx; 4129 4130 ebx.full = x86_pmu.events_maskl; 4131 if (ebx.split.no_branch_misses_retired) { 4132 /* 4133 * Erratum AAJ80 detected, we work it around by using 4134 * the BR_MISP_EXEC.ANY event. This will over-count 4135 * branch-misses, but it's still much better than the 4136 * architectural event which is often completely bogus: 4137 */ 4138 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 4139 ebx.split.no_branch_misses_retired = 0; 4140 x86_pmu.events_maskl = ebx.full; 4141 pr_info("CPU erratum AAJ80 worked around\n"); 4142 } 4143 } 4144 4145 static const struct x86_cpu_desc counter_freezing_ucodes[] = { 4146 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 2, 0x0000000e), 4147 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 9, 0x0000002e), 4148 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 10, 0x00000008), 4149 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_X, 1, 0x00000028), 4150 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 1, 0x00000028), 4151 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 8, 0x00000006), 4152 {} 4153 }; 4154 4155 static bool intel_counter_freezing_broken(void) 4156 { 4157 return !x86_cpu_has_min_microcode_rev(counter_freezing_ucodes); 4158 } 4159 4160 static __init void intel_counter_freezing_quirk(void) 4161 { 4162 /* Check if it's already disabled */ 4163 if (disable_counter_freezing) 4164 return; 4165 4166 /* 4167 * If the system starts with the wrong ucode, leave the 4168 * counter-freezing feature permanently disabled. 4169 */ 4170 if (intel_counter_freezing_broken()) { 4171 pr_info("PMU counter freezing disabled due to CPU errata," 4172 "please upgrade microcode\n"); 4173 x86_pmu.counter_freezing = false; 4174 x86_pmu.handle_irq = intel_pmu_handle_irq; 4175 } 4176 } 4177 4178 /* 4179 * enable software workaround for errata: 4180 * SNB: BJ122 4181 * IVB: BV98 4182 * HSW: HSD29 4183 * 4184 * Only needed when HT is enabled. However detecting 4185 * if HT is enabled is difficult (model specific). So instead, 4186 * we enable the workaround in the early boot, and verify if 4187 * it is needed in a later initcall phase once we have valid 4188 * topology information to check if HT is actually enabled 4189 */ 4190 static __init void intel_ht_bug(void) 4191 { 4192 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; 4193 4194 x86_pmu.start_scheduling = intel_start_scheduling; 4195 x86_pmu.commit_scheduling = intel_commit_scheduling; 4196 x86_pmu.stop_scheduling = intel_stop_scheduling; 4197 } 4198 4199 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); 4200 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") 4201 4202 /* Haswell special events */ 4203 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); 4204 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); 4205 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); 4206 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); 4207 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); 4208 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); 4209 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); 4210 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); 4211 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); 4212 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); 4213 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); 4214 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); 4215 4216 static struct attribute *hsw_events_attrs[] = { 4217 EVENT_PTR(td_slots_issued), 4218 EVENT_PTR(td_slots_retired), 4219 EVENT_PTR(td_fetch_bubbles), 4220 EVENT_PTR(td_total_slots), 4221 EVENT_PTR(td_total_slots_scale), 4222 EVENT_PTR(td_recovery_bubbles), 4223 EVENT_PTR(td_recovery_bubbles_scale), 4224 NULL 4225 }; 4226 4227 static struct attribute *hsw_mem_events_attrs[] = { 4228 EVENT_PTR(mem_ld_hsw), 4229 EVENT_PTR(mem_st_hsw), 4230 NULL, 4231 }; 4232 4233 static struct attribute *hsw_tsx_events_attrs[] = { 4234 EVENT_PTR(tx_start), 4235 EVENT_PTR(tx_commit), 4236 EVENT_PTR(tx_abort), 4237 EVENT_PTR(tx_capacity), 4238 EVENT_PTR(tx_conflict), 4239 EVENT_PTR(el_start), 4240 EVENT_PTR(el_commit), 4241 EVENT_PTR(el_abort), 4242 EVENT_PTR(el_capacity), 4243 EVENT_PTR(el_conflict), 4244 EVENT_PTR(cycles_t), 4245 EVENT_PTR(cycles_ct), 4246 NULL 4247 }; 4248 4249 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80"); 4250 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2"); 4251 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80"); 4252 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2"); 4253 4254 static struct attribute *icl_events_attrs[] = { 4255 EVENT_PTR(mem_ld_hsw), 4256 EVENT_PTR(mem_st_hsw), 4257 NULL, 4258 }; 4259 4260 static struct attribute *icl_tsx_events_attrs[] = { 4261 EVENT_PTR(tx_start), 4262 EVENT_PTR(tx_abort), 4263 EVENT_PTR(tx_commit), 4264 EVENT_PTR(tx_capacity_read), 4265 EVENT_PTR(tx_capacity_write), 4266 EVENT_PTR(tx_conflict), 4267 EVENT_PTR(el_start), 4268 EVENT_PTR(el_abort), 4269 EVENT_PTR(el_commit), 4270 EVENT_PTR(el_capacity_read), 4271 EVENT_PTR(el_capacity_write), 4272 EVENT_PTR(el_conflict), 4273 EVENT_PTR(cycles_t), 4274 EVENT_PTR(cycles_ct), 4275 NULL, 4276 }; 4277 4278 static ssize_t freeze_on_smi_show(struct device *cdev, 4279 struct device_attribute *attr, 4280 char *buf) 4281 { 4282 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi); 4283 } 4284 4285 static DEFINE_MUTEX(freeze_on_smi_mutex); 4286 4287 static ssize_t freeze_on_smi_store(struct device *cdev, 4288 struct device_attribute *attr, 4289 const char *buf, size_t count) 4290 { 4291 unsigned long val; 4292 ssize_t ret; 4293 4294 ret = kstrtoul(buf, 0, &val); 4295 if (ret) 4296 return ret; 4297 4298 if (val > 1) 4299 return -EINVAL; 4300 4301 mutex_lock(&freeze_on_smi_mutex); 4302 4303 if (x86_pmu.attr_freeze_on_smi == val) 4304 goto done; 4305 4306 x86_pmu.attr_freeze_on_smi = val; 4307 4308 get_online_cpus(); 4309 on_each_cpu(flip_smm_bit, &val, 1); 4310 put_online_cpus(); 4311 done: 4312 mutex_unlock(&freeze_on_smi_mutex); 4313 4314 return count; 4315 } 4316 4317 static void update_tfa_sched(void *ignored) 4318 { 4319 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4320 4321 /* 4322 * check if PMC3 is used 4323 * and if so force schedule out for all event types all contexts 4324 */ 4325 if (test_bit(3, cpuc->active_mask)) 4326 perf_pmu_resched(x86_get_pmu()); 4327 } 4328 4329 static ssize_t show_sysctl_tfa(struct device *cdev, 4330 struct device_attribute *attr, 4331 char *buf) 4332 { 4333 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort); 4334 } 4335 4336 static ssize_t set_sysctl_tfa(struct device *cdev, 4337 struct device_attribute *attr, 4338 const char *buf, size_t count) 4339 { 4340 bool val; 4341 ssize_t ret; 4342 4343 ret = kstrtobool(buf, &val); 4344 if (ret) 4345 return ret; 4346 4347 /* no change */ 4348 if (val == allow_tsx_force_abort) 4349 return count; 4350 4351 allow_tsx_force_abort = val; 4352 4353 get_online_cpus(); 4354 on_each_cpu(update_tfa_sched, NULL, 1); 4355 put_online_cpus(); 4356 4357 return count; 4358 } 4359 4360 4361 static DEVICE_ATTR_RW(freeze_on_smi); 4362 4363 static ssize_t branches_show(struct device *cdev, 4364 struct device_attribute *attr, 4365 char *buf) 4366 { 4367 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr); 4368 } 4369 4370 static DEVICE_ATTR_RO(branches); 4371 4372 static struct attribute *lbr_attrs[] = { 4373 &dev_attr_branches.attr, 4374 NULL 4375 }; 4376 4377 static char pmu_name_str[30]; 4378 4379 static ssize_t pmu_name_show(struct device *cdev, 4380 struct device_attribute *attr, 4381 char *buf) 4382 { 4383 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str); 4384 } 4385 4386 static DEVICE_ATTR_RO(pmu_name); 4387 4388 static struct attribute *intel_pmu_caps_attrs[] = { 4389 &dev_attr_pmu_name.attr, 4390 NULL 4391 }; 4392 4393 static DEVICE_ATTR(allow_tsx_force_abort, 0644, 4394 show_sysctl_tfa, 4395 set_sysctl_tfa); 4396 4397 static struct attribute *intel_pmu_attrs[] = { 4398 &dev_attr_freeze_on_smi.attr, 4399 &dev_attr_allow_tsx_force_abort.attr, 4400 NULL, 4401 }; 4402 4403 static umode_t 4404 tsx_is_visible(struct kobject *kobj, struct attribute *attr, int i) 4405 { 4406 return boot_cpu_has(X86_FEATURE_RTM) ? attr->mode : 0; 4407 } 4408 4409 static umode_t 4410 pebs_is_visible(struct kobject *kobj, struct attribute *attr, int i) 4411 { 4412 return x86_pmu.pebs ? attr->mode : 0; 4413 } 4414 4415 static umode_t 4416 lbr_is_visible(struct kobject *kobj, struct attribute *attr, int i) 4417 { 4418 return x86_pmu.lbr_nr ? attr->mode : 0; 4419 } 4420 4421 static umode_t 4422 exra_is_visible(struct kobject *kobj, struct attribute *attr, int i) 4423 { 4424 return x86_pmu.version >= 2 ? attr->mode : 0; 4425 } 4426 4427 static umode_t 4428 default_is_visible(struct kobject *kobj, struct attribute *attr, int i) 4429 { 4430 if (attr == &dev_attr_allow_tsx_force_abort.attr) 4431 return x86_pmu.flags & PMU_FL_TFA ? attr->mode : 0; 4432 4433 return attr->mode; 4434 } 4435 4436 static struct attribute_group group_events_td = { 4437 .name = "events", 4438 }; 4439 4440 static struct attribute_group group_events_mem = { 4441 .name = "events", 4442 .is_visible = pebs_is_visible, 4443 }; 4444 4445 static struct attribute_group group_events_tsx = { 4446 .name = "events", 4447 .is_visible = tsx_is_visible, 4448 }; 4449 4450 static struct attribute_group group_caps_gen = { 4451 .name = "caps", 4452 .attrs = intel_pmu_caps_attrs, 4453 }; 4454 4455 static struct attribute_group group_caps_lbr = { 4456 .name = "caps", 4457 .attrs = lbr_attrs, 4458 .is_visible = lbr_is_visible, 4459 }; 4460 4461 static struct attribute_group group_format_extra = { 4462 .name = "format", 4463 .is_visible = exra_is_visible, 4464 }; 4465 4466 static struct attribute_group group_format_extra_skl = { 4467 .name = "format", 4468 .is_visible = exra_is_visible, 4469 }; 4470 4471 static struct attribute_group group_default = { 4472 .attrs = intel_pmu_attrs, 4473 .is_visible = default_is_visible, 4474 }; 4475 4476 static const struct attribute_group *attr_update[] = { 4477 &group_events_td, 4478 &group_events_mem, 4479 &group_events_tsx, 4480 &group_caps_gen, 4481 &group_caps_lbr, 4482 &group_format_extra, 4483 &group_format_extra_skl, 4484 &group_default, 4485 NULL, 4486 }; 4487 4488 static struct attribute *empty_attrs; 4489 4490 __init int intel_pmu_init(void) 4491 { 4492 struct attribute **extra_skl_attr = &empty_attrs; 4493 struct attribute **extra_attr = &empty_attrs; 4494 struct attribute **td_attr = &empty_attrs; 4495 struct attribute **mem_attr = &empty_attrs; 4496 struct attribute **tsx_attr = &empty_attrs; 4497 union cpuid10_edx edx; 4498 union cpuid10_eax eax; 4499 union cpuid10_ebx ebx; 4500 struct event_constraint *c; 4501 unsigned int unused; 4502 struct extra_reg *er; 4503 bool pmem = false; 4504 int version, i; 4505 char *name; 4506 4507 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 4508 switch (boot_cpu_data.x86) { 4509 case 0x6: 4510 return p6_pmu_init(); 4511 case 0xb: 4512 return knc_pmu_init(); 4513 case 0xf: 4514 return p4_pmu_init(); 4515 } 4516 return -ENODEV; 4517 } 4518 4519 /* 4520 * Check whether the Architectural PerfMon supports 4521 * Branch Misses Retired hw_event or not. 4522 */ 4523 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); 4524 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) 4525 return -ENODEV; 4526 4527 version = eax.split.version_id; 4528 if (version < 2) 4529 x86_pmu = core_pmu; 4530 else 4531 x86_pmu = intel_pmu; 4532 4533 x86_pmu.version = version; 4534 x86_pmu.num_counters = eax.split.num_counters; 4535 x86_pmu.cntval_bits = eax.split.bit_width; 4536 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; 4537 4538 x86_pmu.events_maskl = ebx.full; 4539 x86_pmu.events_mask_len = eax.split.mask_length; 4540 4541 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); 4542 4543 /* 4544 * Quirk: v2 perfmon does not report fixed-purpose events, so 4545 * assume at least 3 events, when not running in a hypervisor: 4546 */ 4547 if (version > 1) { 4548 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR); 4549 4550 x86_pmu.num_counters_fixed = 4551 max((int)edx.split.num_counters_fixed, assume); 4552 } 4553 4554 if (version >= 4) 4555 x86_pmu.counter_freezing = !disable_counter_freezing; 4556 4557 if (boot_cpu_has(X86_FEATURE_PDCM)) { 4558 u64 capabilities; 4559 4560 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); 4561 x86_pmu.intel_cap.capabilities = capabilities; 4562 } 4563 4564 intel_ds_init(); 4565 4566 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ 4567 4568 /* 4569 * Install the hw-cache-events table: 4570 */ 4571 switch (boot_cpu_data.x86_model) { 4572 case INTEL_FAM6_CORE_YONAH: 4573 pr_cont("Core events, "); 4574 name = "core"; 4575 break; 4576 4577 case INTEL_FAM6_CORE2_MEROM: 4578 x86_add_quirk(intel_clovertown_quirk); 4579 /* fall through */ 4580 4581 case INTEL_FAM6_CORE2_MEROM_L: 4582 case INTEL_FAM6_CORE2_PENRYN: 4583 case INTEL_FAM6_CORE2_DUNNINGTON: 4584 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, 4585 sizeof(hw_cache_event_ids)); 4586 4587 intel_pmu_lbr_init_core(); 4588 4589 x86_pmu.event_constraints = intel_core2_event_constraints; 4590 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; 4591 pr_cont("Core2 events, "); 4592 name = "core2"; 4593 break; 4594 4595 case INTEL_FAM6_NEHALEM: 4596 case INTEL_FAM6_NEHALEM_EP: 4597 case INTEL_FAM6_NEHALEM_EX: 4598 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, 4599 sizeof(hw_cache_event_ids)); 4600 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 4601 sizeof(hw_cache_extra_regs)); 4602 4603 intel_pmu_lbr_init_nhm(); 4604 4605 x86_pmu.event_constraints = intel_nehalem_event_constraints; 4606 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; 4607 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 4608 x86_pmu.extra_regs = intel_nehalem_extra_regs; 4609 4610 mem_attr = nhm_mem_events_attrs; 4611 4612 /* UOPS_ISSUED.STALLED_CYCLES */ 4613 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4614 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4615 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 4616 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4617 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 4618 4619 intel_pmu_pebs_data_source_nhm(); 4620 x86_add_quirk(intel_nehalem_quirk); 4621 x86_pmu.pebs_no_tlb = 1; 4622 extra_attr = nhm_format_attr; 4623 4624 pr_cont("Nehalem events, "); 4625 name = "nehalem"; 4626 break; 4627 4628 case INTEL_FAM6_ATOM_BONNELL: 4629 case INTEL_FAM6_ATOM_BONNELL_MID: 4630 case INTEL_FAM6_ATOM_SALTWELL: 4631 case INTEL_FAM6_ATOM_SALTWELL_MID: 4632 case INTEL_FAM6_ATOM_SALTWELL_TABLET: 4633 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, 4634 sizeof(hw_cache_event_ids)); 4635 4636 intel_pmu_lbr_init_atom(); 4637 4638 x86_pmu.event_constraints = intel_gen_event_constraints; 4639 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; 4640 x86_pmu.pebs_aliases = intel_pebs_aliases_core2; 4641 pr_cont("Atom events, "); 4642 name = "bonnell"; 4643 break; 4644 4645 case INTEL_FAM6_ATOM_SILVERMONT: 4646 case INTEL_FAM6_ATOM_SILVERMONT_X: 4647 case INTEL_FAM6_ATOM_SILVERMONT_MID: 4648 case INTEL_FAM6_ATOM_AIRMONT: 4649 case INTEL_FAM6_ATOM_AIRMONT_MID: 4650 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, 4651 sizeof(hw_cache_event_ids)); 4652 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, 4653 sizeof(hw_cache_extra_regs)); 4654 4655 intel_pmu_lbr_init_slm(); 4656 4657 x86_pmu.event_constraints = intel_slm_event_constraints; 4658 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 4659 x86_pmu.extra_regs = intel_slm_extra_regs; 4660 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4661 td_attr = slm_events_attrs; 4662 extra_attr = slm_format_attr; 4663 pr_cont("Silvermont events, "); 4664 name = "silvermont"; 4665 break; 4666 4667 case INTEL_FAM6_ATOM_GOLDMONT: 4668 case INTEL_FAM6_ATOM_GOLDMONT_X: 4669 x86_add_quirk(intel_counter_freezing_quirk); 4670 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids, 4671 sizeof(hw_cache_event_ids)); 4672 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs, 4673 sizeof(hw_cache_extra_regs)); 4674 4675 intel_pmu_lbr_init_skl(); 4676 4677 x86_pmu.event_constraints = intel_slm_event_constraints; 4678 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints; 4679 x86_pmu.extra_regs = intel_glm_extra_regs; 4680 /* 4681 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4682 * for precise cycles. 4683 * :pp is identical to :ppp 4684 */ 4685 x86_pmu.pebs_aliases = NULL; 4686 x86_pmu.pebs_prec_dist = true; 4687 x86_pmu.lbr_pt_coexist = true; 4688 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4689 td_attr = glm_events_attrs; 4690 extra_attr = slm_format_attr; 4691 pr_cont("Goldmont events, "); 4692 name = "goldmont"; 4693 break; 4694 4695 case INTEL_FAM6_ATOM_GOLDMONT_PLUS: 4696 x86_add_quirk(intel_counter_freezing_quirk); 4697 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 4698 sizeof(hw_cache_event_ids)); 4699 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs, 4700 sizeof(hw_cache_extra_regs)); 4701 4702 intel_pmu_lbr_init_skl(); 4703 4704 x86_pmu.event_constraints = intel_slm_event_constraints; 4705 x86_pmu.extra_regs = intel_glm_extra_regs; 4706 /* 4707 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4708 * for precise cycles. 4709 */ 4710 x86_pmu.pebs_aliases = NULL; 4711 x86_pmu.pebs_prec_dist = true; 4712 x86_pmu.lbr_pt_coexist = true; 4713 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4714 x86_pmu.flags |= PMU_FL_PEBS_ALL; 4715 x86_pmu.get_event_constraints = glp_get_event_constraints; 4716 td_attr = glm_events_attrs; 4717 /* Goldmont Plus has 4-wide pipeline */ 4718 event_attr_td_total_slots_scale_glm.event_str = "4"; 4719 extra_attr = slm_format_attr; 4720 pr_cont("Goldmont plus events, "); 4721 name = "goldmont_plus"; 4722 break; 4723 4724 case INTEL_FAM6_ATOM_TREMONT_X: 4725 x86_pmu.late_ack = true; 4726 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 4727 sizeof(hw_cache_event_ids)); 4728 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 4729 sizeof(hw_cache_extra_regs)); 4730 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 4731 4732 intel_pmu_lbr_init_skl(); 4733 4734 x86_pmu.event_constraints = intel_slm_event_constraints; 4735 x86_pmu.extra_regs = intel_tnt_extra_regs; 4736 /* 4737 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4738 * for precise cycles. 4739 */ 4740 x86_pmu.pebs_aliases = NULL; 4741 x86_pmu.pebs_prec_dist = true; 4742 x86_pmu.lbr_pt_coexist = true; 4743 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4744 x86_pmu.get_event_constraints = tnt_get_event_constraints; 4745 extra_attr = slm_format_attr; 4746 pr_cont("Tremont events, "); 4747 name = "Tremont"; 4748 break; 4749 4750 case INTEL_FAM6_WESTMERE: 4751 case INTEL_FAM6_WESTMERE_EP: 4752 case INTEL_FAM6_WESTMERE_EX: 4753 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, 4754 sizeof(hw_cache_event_ids)); 4755 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 4756 sizeof(hw_cache_extra_regs)); 4757 4758 intel_pmu_lbr_init_nhm(); 4759 4760 x86_pmu.event_constraints = intel_westmere_event_constraints; 4761 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 4762 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; 4763 x86_pmu.extra_regs = intel_westmere_extra_regs; 4764 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4765 4766 mem_attr = nhm_mem_events_attrs; 4767 4768 /* UOPS_ISSUED.STALLED_CYCLES */ 4769 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4770 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4771 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 4772 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4773 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 4774 4775 intel_pmu_pebs_data_source_nhm(); 4776 extra_attr = nhm_format_attr; 4777 pr_cont("Westmere events, "); 4778 name = "westmere"; 4779 break; 4780 4781 case INTEL_FAM6_SANDYBRIDGE: 4782 case INTEL_FAM6_SANDYBRIDGE_X: 4783 x86_add_quirk(intel_sandybridge_quirk); 4784 x86_add_quirk(intel_ht_bug); 4785 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 4786 sizeof(hw_cache_event_ids)); 4787 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 4788 sizeof(hw_cache_extra_regs)); 4789 4790 intel_pmu_lbr_init_snb(); 4791 4792 x86_pmu.event_constraints = intel_snb_event_constraints; 4793 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; 4794 x86_pmu.pebs_aliases = intel_pebs_aliases_snb; 4795 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X) 4796 x86_pmu.extra_regs = intel_snbep_extra_regs; 4797 else 4798 x86_pmu.extra_regs = intel_snb_extra_regs; 4799 4800 4801 /* all extra regs are per-cpu when HT is on */ 4802 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4803 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4804 4805 td_attr = snb_events_attrs; 4806 mem_attr = snb_mem_events_attrs; 4807 4808 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 4809 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4810 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4811 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ 4812 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4813 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); 4814 4815 extra_attr = nhm_format_attr; 4816 4817 pr_cont("SandyBridge events, "); 4818 name = "sandybridge"; 4819 break; 4820 4821 case INTEL_FAM6_IVYBRIDGE: 4822 case INTEL_FAM6_IVYBRIDGE_X: 4823 x86_add_quirk(intel_ht_bug); 4824 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 4825 sizeof(hw_cache_event_ids)); 4826 /* dTLB-load-misses on IVB is different than SNB */ 4827 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ 4828 4829 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 4830 sizeof(hw_cache_extra_regs)); 4831 4832 intel_pmu_lbr_init_snb(); 4833 4834 x86_pmu.event_constraints = intel_ivb_event_constraints; 4835 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; 4836 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4837 x86_pmu.pebs_prec_dist = true; 4838 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X) 4839 x86_pmu.extra_regs = intel_snbep_extra_regs; 4840 else 4841 x86_pmu.extra_regs = intel_snb_extra_regs; 4842 /* all extra regs are per-cpu when HT is on */ 4843 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4844 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4845 4846 td_attr = snb_events_attrs; 4847 mem_attr = snb_mem_events_attrs; 4848 4849 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 4850 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4851 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4852 4853 extra_attr = nhm_format_attr; 4854 4855 pr_cont("IvyBridge events, "); 4856 name = "ivybridge"; 4857 break; 4858 4859 4860 case INTEL_FAM6_HASWELL_CORE: 4861 case INTEL_FAM6_HASWELL_X: 4862 case INTEL_FAM6_HASWELL_ULT: 4863 case INTEL_FAM6_HASWELL_GT3E: 4864 x86_add_quirk(intel_ht_bug); 4865 x86_add_quirk(intel_pebs_isolation_quirk); 4866 x86_pmu.late_ack = true; 4867 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4868 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4869 4870 intel_pmu_lbr_init_hsw(); 4871 4872 x86_pmu.event_constraints = intel_hsw_event_constraints; 4873 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; 4874 x86_pmu.extra_regs = intel_snbep_extra_regs; 4875 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4876 x86_pmu.pebs_prec_dist = true; 4877 /* all extra regs are per-cpu when HT is on */ 4878 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4879 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4880 4881 x86_pmu.hw_config = hsw_hw_config; 4882 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4883 x86_pmu.lbr_double_abort = true; 4884 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4885 hsw_format_attr : nhm_format_attr; 4886 td_attr = hsw_events_attrs; 4887 mem_attr = hsw_mem_events_attrs; 4888 tsx_attr = hsw_tsx_events_attrs; 4889 pr_cont("Haswell events, "); 4890 name = "haswell"; 4891 break; 4892 4893 case INTEL_FAM6_BROADWELL_CORE: 4894 case INTEL_FAM6_BROADWELL_XEON_D: 4895 case INTEL_FAM6_BROADWELL_GT3E: 4896 case INTEL_FAM6_BROADWELL_X: 4897 x86_add_quirk(intel_pebs_isolation_quirk); 4898 x86_pmu.late_ack = true; 4899 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4900 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4901 4902 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ 4903 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | 4904 BDW_L3_MISS|HSW_SNOOP_DRAM; 4905 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| 4906 HSW_SNOOP_DRAM; 4907 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| 4908 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 4909 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| 4910 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 4911 4912 intel_pmu_lbr_init_hsw(); 4913 4914 x86_pmu.event_constraints = intel_bdw_event_constraints; 4915 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints; 4916 x86_pmu.extra_regs = intel_snbep_extra_regs; 4917 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4918 x86_pmu.pebs_prec_dist = true; 4919 /* all extra regs are per-cpu when HT is on */ 4920 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4921 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4922 4923 x86_pmu.hw_config = hsw_hw_config; 4924 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4925 x86_pmu.limit_period = bdw_limit_period; 4926 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4927 hsw_format_attr : nhm_format_attr; 4928 td_attr = hsw_events_attrs; 4929 mem_attr = hsw_mem_events_attrs; 4930 tsx_attr = hsw_tsx_events_attrs; 4931 pr_cont("Broadwell events, "); 4932 name = "broadwell"; 4933 break; 4934 4935 case INTEL_FAM6_XEON_PHI_KNL: 4936 case INTEL_FAM6_XEON_PHI_KNM: 4937 memcpy(hw_cache_event_ids, 4938 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4939 memcpy(hw_cache_extra_regs, 4940 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4941 intel_pmu_lbr_init_knl(); 4942 4943 x86_pmu.event_constraints = intel_slm_event_constraints; 4944 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 4945 x86_pmu.extra_regs = intel_knl_extra_regs; 4946 4947 /* all extra regs are per-cpu when HT is on */ 4948 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4949 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4950 extra_attr = slm_format_attr; 4951 pr_cont("Knights Landing/Mill events, "); 4952 name = "knights-landing"; 4953 break; 4954 4955 case INTEL_FAM6_SKYLAKE_X: 4956 pmem = true; 4957 /* fall through */ 4958 case INTEL_FAM6_SKYLAKE_MOBILE: 4959 case INTEL_FAM6_SKYLAKE_DESKTOP: 4960 case INTEL_FAM6_KABYLAKE_MOBILE: 4961 case INTEL_FAM6_KABYLAKE_DESKTOP: 4962 x86_add_quirk(intel_pebs_isolation_quirk); 4963 x86_pmu.late_ack = true; 4964 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4965 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4966 intel_pmu_lbr_init_skl(); 4967 4968 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */ 4969 event_attr_td_recovery_bubbles.event_str_noht = 4970 "event=0xd,umask=0x1,cmask=1"; 4971 event_attr_td_recovery_bubbles.event_str_ht = 4972 "event=0xd,umask=0x1,cmask=1,any=1"; 4973 4974 x86_pmu.event_constraints = intel_skl_event_constraints; 4975 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; 4976 x86_pmu.extra_regs = intel_skl_extra_regs; 4977 x86_pmu.pebs_aliases = intel_pebs_aliases_skl; 4978 x86_pmu.pebs_prec_dist = true; 4979 /* all extra regs are per-cpu when HT is on */ 4980 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4981 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4982 4983 x86_pmu.hw_config = hsw_hw_config; 4984 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4985 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4986 hsw_format_attr : nhm_format_attr; 4987 extra_skl_attr = skl_format_attr; 4988 td_attr = hsw_events_attrs; 4989 mem_attr = hsw_mem_events_attrs; 4990 tsx_attr = hsw_tsx_events_attrs; 4991 intel_pmu_pebs_data_source_skl(pmem); 4992 4993 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT)) { 4994 x86_pmu.flags |= PMU_FL_TFA; 4995 x86_pmu.get_event_constraints = tfa_get_event_constraints; 4996 x86_pmu.enable_all = intel_tfa_pmu_enable_all; 4997 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; 4998 } 4999 5000 pr_cont("Skylake events, "); 5001 name = "skylake"; 5002 break; 5003 5004 case INTEL_FAM6_ICELAKE_X: 5005 case INTEL_FAM6_ICELAKE_XEON_D: 5006 pmem = true; 5007 /* fall through */ 5008 case INTEL_FAM6_ICELAKE_MOBILE: 5009 case INTEL_FAM6_ICELAKE_DESKTOP: 5010 x86_pmu.late_ack = true; 5011 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 5012 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 5013 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 5014 intel_pmu_lbr_init_skl(); 5015 5016 x86_pmu.event_constraints = intel_icl_event_constraints; 5017 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints; 5018 x86_pmu.extra_regs = intel_icl_extra_regs; 5019 x86_pmu.pebs_aliases = NULL; 5020 x86_pmu.pebs_prec_dist = true; 5021 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 5022 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 5023 5024 x86_pmu.hw_config = hsw_hw_config; 5025 x86_pmu.get_event_constraints = icl_get_event_constraints; 5026 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 5027 hsw_format_attr : nhm_format_attr; 5028 extra_skl_attr = skl_format_attr; 5029 mem_attr = icl_events_attrs; 5030 tsx_attr = icl_tsx_events_attrs; 5031 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xca, .umask=0x02); 5032 x86_pmu.lbr_pt_coexist = true; 5033 intel_pmu_pebs_data_source_skl(pmem); 5034 pr_cont("Icelake events, "); 5035 name = "icelake"; 5036 break; 5037 5038 default: 5039 switch (x86_pmu.version) { 5040 case 1: 5041 x86_pmu.event_constraints = intel_v1_event_constraints; 5042 pr_cont("generic architected perfmon v1, "); 5043 name = "generic_arch_v1"; 5044 break; 5045 default: 5046 /* 5047 * default constraints for v2 and up 5048 */ 5049 x86_pmu.event_constraints = intel_gen_event_constraints; 5050 pr_cont("generic architected perfmon, "); 5051 name = "generic_arch_v2+"; 5052 break; 5053 } 5054 } 5055 5056 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name); 5057 5058 5059 group_events_td.attrs = td_attr; 5060 group_events_mem.attrs = mem_attr; 5061 group_events_tsx.attrs = tsx_attr; 5062 group_format_extra.attrs = extra_attr; 5063 group_format_extra_skl.attrs = extra_skl_attr; 5064 5065 x86_pmu.attr_update = attr_update; 5066 5067 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { 5068 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", 5069 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); 5070 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; 5071 } 5072 x86_pmu.intel_ctrl = (1ULL << x86_pmu.num_counters) - 1; 5073 5074 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { 5075 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", 5076 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); 5077 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; 5078 } 5079 5080 x86_pmu.intel_ctrl |= 5081 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; 5082 5083 if (x86_pmu.event_constraints) { 5084 /* 5085 * event on fixed counter2 (REF_CYCLES) only works on this 5086 * counter, so do not extend mask to generic counters 5087 */ 5088 for_each_event_constraint(c, x86_pmu.event_constraints) { 5089 if (c->cmask == FIXED_EVENT_FLAGS 5090 && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) { 5091 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; 5092 } 5093 c->idxmsk64 &= 5094 ~(~0ULL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed)); 5095 c->weight = hweight64(c->idxmsk64); 5096 } 5097 } 5098 5099 /* 5100 * Access LBR MSR may cause #GP under certain circumstances. 5101 * E.g. KVM doesn't support LBR MSR 5102 * Check all LBT MSR here. 5103 * Disable LBR access if any LBR MSRs can not be accessed. 5104 */ 5105 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL)) 5106 x86_pmu.lbr_nr = 0; 5107 for (i = 0; i < x86_pmu.lbr_nr; i++) { 5108 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && 5109 check_msr(x86_pmu.lbr_to + i, 0xffffUL))) 5110 x86_pmu.lbr_nr = 0; 5111 } 5112 5113 if (x86_pmu.lbr_nr) 5114 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr); 5115 5116 /* 5117 * Access extra MSR may cause #GP under certain circumstances. 5118 * E.g. KVM doesn't support offcore event 5119 * Check all extra_regs here. 5120 */ 5121 if (x86_pmu.extra_regs) { 5122 for (er = x86_pmu.extra_regs; er->msr; er++) { 5123 er->extra_msr_access = check_msr(er->msr, 0x11UL); 5124 /* Disable LBR select mapping */ 5125 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) 5126 x86_pmu.lbr_sel_map = NULL; 5127 } 5128 } 5129 5130 /* Support full width counters using alternative MSR range */ 5131 if (x86_pmu.intel_cap.full_width_write) { 5132 x86_pmu.max_period = x86_pmu.cntval_mask >> 1; 5133 x86_pmu.perfctr = MSR_IA32_PMC0; 5134 pr_cont("full-width counters, "); 5135 } 5136 5137 /* 5138 * For arch perfmon 4 use counter freezing to avoid 5139 * several MSR accesses in the PMI. 5140 */ 5141 if (x86_pmu.counter_freezing) 5142 x86_pmu.handle_irq = intel_pmu_handle_irq_v4; 5143 5144 return 0; 5145 } 5146 5147 /* 5148 * HT bug: phase 2 init 5149 * Called once we have valid topology information to check 5150 * whether or not HT is enabled 5151 * If HT is off, then we disable the workaround 5152 */ 5153 static __init int fixup_ht_bug(void) 5154 { 5155 int c; 5156 /* 5157 * problem not present on this CPU model, nothing to do 5158 */ 5159 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) 5160 return 0; 5161 5162 if (topology_max_smt_threads() > 1) { 5163 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); 5164 return 0; 5165 } 5166 5167 cpus_read_lock(); 5168 5169 hardlockup_detector_perf_stop(); 5170 5171 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); 5172 5173 x86_pmu.start_scheduling = NULL; 5174 x86_pmu.commit_scheduling = NULL; 5175 x86_pmu.stop_scheduling = NULL; 5176 5177 hardlockup_detector_perf_restart(); 5178 5179 for_each_online_cpu(c) 5180 free_excl_cntrs(&per_cpu(cpu_hw_events, c)); 5181 5182 cpus_read_unlock(); 5183 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); 5184 return 0; 5185 } 5186 subsys_initcall(fixup_ht_bug) 5187