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, 0x3fffff9fffull, RSP_0), 266 INTEL_UEVENT_EXTRA_REG(0x01bb, MSR_OFFCORE_RSP_1, 0x3fffff9fffull, 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 return; 2166 } 2167 2168 x86_pmu_disable_event(event); 2169 2170 /* 2171 * Needs to be called after x86_pmu_disable_event, 2172 * so we don't trigger the event without PEBS bit set. 2173 */ 2174 if (unlikely(event->attr.precise_ip)) 2175 intel_pmu_pebs_disable(event); 2176 } 2177 2178 static void intel_pmu_del_event(struct perf_event *event) 2179 { 2180 if (needs_branch_stack(event)) 2181 intel_pmu_lbr_del(event); 2182 if (event->attr.precise_ip) 2183 intel_pmu_pebs_del(event); 2184 } 2185 2186 static void intel_pmu_read_event(struct perf_event *event) 2187 { 2188 if (event->hw.flags & PERF_X86_EVENT_AUTO_RELOAD) 2189 intel_pmu_auto_reload_read(event); 2190 else 2191 x86_perf_event_update(event); 2192 } 2193 2194 static void intel_pmu_enable_fixed(struct perf_event *event) 2195 { 2196 struct hw_perf_event *hwc = &event->hw; 2197 int idx = hwc->idx - INTEL_PMC_IDX_FIXED; 2198 u64 ctrl_val, mask, bits = 0; 2199 2200 /* 2201 * Enable IRQ generation (0x8), if not PEBS, 2202 * and enable ring-3 counting (0x2) and ring-0 counting (0x1) 2203 * if requested: 2204 */ 2205 if (!event->attr.precise_ip) 2206 bits |= 0x8; 2207 if (hwc->config & ARCH_PERFMON_EVENTSEL_USR) 2208 bits |= 0x2; 2209 if (hwc->config & ARCH_PERFMON_EVENTSEL_OS) 2210 bits |= 0x1; 2211 2212 /* 2213 * ANY bit is supported in v3 and up 2214 */ 2215 if (x86_pmu.version > 2 && hwc->config & ARCH_PERFMON_EVENTSEL_ANY) 2216 bits |= 0x4; 2217 2218 bits <<= (idx * 4); 2219 mask = 0xfULL << (idx * 4); 2220 2221 if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip) { 2222 bits |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2223 mask |= ICL_FIXED_0_ADAPTIVE << (idx * 4); 2224 } 2225 2226 rdmsrl(hwc->config_base, ctrl_val); 2227 ctrl_val &= ~mask; 2228 ctrl_val |= bits; 2229 wrmsrl(hwc->config_base, ctrl_val); 2230 } 2231 2232 static void intel_pmu_enable_event(struct perf_event *event) 2233 { 2234 struct hw_perf_event *hwc = &event->hw; 2235 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2236 2237 if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_BTS)) { 2238 if (!__this_cpu_read(cpu_hw_events.enabled)) 2239 return; 2240 2241 intel_pmu_enable_bts(hwc->config); 2242 return; 2243 } 2244 2245 if (event->attr.exclude_host) 2246 cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx); 2247 if (event->attr.exclude_guest) 2248 cpuc->intel_ctrl_host_mask |= (1ull << hwc->idx); 2249 2250 if (unlikely(event_is_checkpointed(event))) 2251 cpuc->intel_cp_status |= (1ull << hwc->idx); 2252 2253 if (unlikely(event->attr.precise_ip)) 2254 intel_pmu_pebs_enable(event); 2255 2256 if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL)) { 2257 intel_pmu_enable_fixed(event); 2258 return; 2259 } 2260 2261 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 2262 } 2263 2264 static void intel_pmu_add_event(struct perf_event *event) 2265 { 2266 if (event->attr.precise_ip) 2267 intel_pmu_pebs_add(event); 2268 if (needs_branch_stack(event)) 2269 intel_pmu_lbr_add(event); 2270 } 2271 2272 /* 2273 * Save and restart an expired event. Called by NMI contexts, 2274 * so it has to be careful about preempting normal event ops: 2275 */ 2276 int intel_pmu_save_and_restart(struct perf_event *event) 2277 { 2278 x86_perf_event_update(event); 2279 /* 2280 * For a checkpointed counter always reset back to 0. This 2281 * avoids a situation where the counter overflows, aborts the 2282 * transaction and is then set back to shortly before the 2283 * overflow, and overflows and aborts again. 2284 */ 2285 if (unlikely(event_is_checkpointed(event))) { 2286 /* No race with NMIs because the counter should not be armed */ 2287 wrmsrl(event->hw.event_base, 0); 2288 local64_set(&event->hw.prev_count, 0); 2289 } 2290 return x86_perf_event_set_period(event); 2291 } 2292 2293 static void intel_pmu_reset(void) 2294 { 2295 struct debug_store *ds = __this_cpu_read(cpu_hw_events.ds); 2296 unsigned long flags; 2297 int idx; 2298 2299 if (!x86_pmu.num_counters) 2300 return; 2301 2302 local_irq_save(flags); 2303 2304 pr_info("clearing PMU state on CPU#%d\n", smp_processor_id()); 2305 2306 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 2307 wrmsrl_safe(x86_pmu_config_addr(idx), 0ull); 2308 wrmsrl_safe(x86_pmu_event_addr(idx), 0ull); 2309 } 2310 for (idx = 0; idx < x86_pmu.num_counters_fixed; idx++) 2311 wrmsrl_safe(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, 0ull); 2312 2313 if (ds) 2314 ds->bts_index = ds->bts_buffer_base; 2315 2316 /* Ack all overflows and disable fixed counters */ 2317 if (x86_pmu.version >= 2) { 2318 intel_pmu_ack_status(intel_pmu_get_status()); 2319 wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0); 2320 } 2321 2322 /* Reset LBRs and LBR freezing */ 2323 if (x86_pmu.lbr_nr) { 2324 update_debugctlmsr(get_debugctlmsr() & 2325 ~(DEBUGCTLMSR_FREEZE_LBRS_ON_PMI|DEBUGCTLMSR_LBR)); 2326 } 2327 2328 local_irq_restore(flags); 2329 } 2330 2331 static int handle_pmi_common(struct pt_regs *regs, u64 status) 2332 { 2333 struct perf_sample_data data; 2334 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2335 int bit; 2336 int handled = 0; 2337 2338 inc_irq_stat(apic_perf_irqs); 2339 2340 /* 2341 * Ignore a range of extra bits in status that do not indicate 2342 * overflow by themselves. 2343 */ 2344 status &= ~(GLOBAL_STATUS_COND_CHG | 2345 GLOBAL_STATUS_ASIF | 2346 GLOBAL_STATUS_LBRS_FROZEN); 2347 if (!status) 2348 return 0; 2349 /* 2350 * In case multiple PEBS events are sampled at the same time, 2351 * it is possible to have GLOBAL_STATUS bit 62 set indicating 2352 * PEBS buffer overflow and also seeing at most 3 PEBS counters 2353 * having their bits set in the status register. This is a sign 2354 * that there was at least one PEBS record pending at the time 2355 * of the PMU interrupt. PEBS counters must only be processed 2356 * via the drain_pebs() calls and not via the regular sample 2357 * processing loop coming after that the function, otherwise 2358 * phony regular samples may be generated in the sampling buffer 2359 * not marked with the EXACT tag. Another possibility is to have 2360 * one PEBS event and at least one non-PEBS event whic hoverflows 2361 * while PEBS has armed. In this case, bit 62 of GLOBAL_STATUS will 2362 * not be set, yet the overflow status bit for the PEBS counter will 2363 * be on Skylake. 2364 * 2365 * To avoid this problem, we systematically ignore the PEBS-enabled 2366 * counters from the GLOBAL_STATUS mask and we always process PEBS 2367 * events via drain_pebs(). 2368 */ 2369 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 2370 status &= ~cpuc->pebs_enabled; 2371 else 2372 status &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 2373 2374 /* 2375 * PEBS overflow sets bit 62 in the global status register 2376 */ 2377 if (__test_and_clear_bit(62, (unsigned long *)&status)) { 2378 handled++; 2379 x86_pmu.drain_pebs(regs); 2380 status &= x86_pmu.intel_ctrl | GLOBAL_STATUS_TRACE_TOPAPMI; 2381 } 2382 2383 /* 2384 * Intel PT 2385 */ 2386 if (__test_and_clear_bit(55, (unsigned long *)&status)) { 2387 handled++; 2388 if (unlikely(perf_guest_cbs && perf_guest_cbs->is_in_guest() && 2389 perf_guest_cbs->handle_intel_pt_intr)) 2390 perf_guest_cbs->handle_intel_pt_intr(); 2391 else 2392 intel_pt_interrupt(); 2393 } 2394 2395 /* 2396 * Checkpointed counters can lead to 'spurious' PMIs because the 2397 * rollback caused by the PMI will have cleared the overflow status 2398 * bit. Therefore always force probe these counters. 2399 */ 2400 status |= cpuc->intel_cp_status; 2401 2402 for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) { 2403 struct perf_event *event = cpuc->events[bit]; 2404 2405 handled++; 2406 2407 if (!test_bit(bit, cpuc->active_mask)) 2408 continue; 2409 2410 if (!intel_pmu_save_and_restart(event)) 2411 continue; 2412 2413 perf_sample_data_init(&data, 0, event->hw.last_period); 2414 2415 if (has_branch_stack(event)) 2416 data.br_stack = &cpuc->lbr_stack; 2417 2418 if (perf_event_overflow(event, &data, regs)) 2419 x86_pmu_stop(event, 0); 2420 } 2421 2422 return handled; 2423 } 2424 2425 static bool disable_counter_freezing = true; 2426 static int __init intel_perf_counter_freezing_setup(char *s) 2427 { 2428 bool res; 2429 2430 if (kstrtobool(s, &res)) 2431 return -EINVAL; 2432 2433 disable_counter_freezing = !res; 2434 return 1; 2435 } 2436 __setup("perf_v4_pmi=", intel_perf_counter_freezing_setup); 2437 2438 /* 2439 * Simplified handler for Arch Perfmon v4: 2440 * - We rely on counter freezing/unfreezing to enable/disable the PMU. 2441 * This is done automatically on PMU ack. 2442 * - Ack the PMU only after the APIC. 2443 */ 2444 2445 static int intel_pmu_handle_irq_v4(struct pt_regs *regs) 2446 { 2447 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 2448 int handled = 0; 2449 bool bts = false; 2450 u64 status; 2451 int pmu_enabled = cpuc->enabled; 2452 int loops = 0; 2453 2454 /* PMU has been disabled because of counter freezing */ 2455 cpuc->enabled = 0; 2456 if (test_bit(INTEL_PMC_IDX_FIXED_BTS, cpuc->active_mask)) { 2457 bts = true; 2458 intel_bts_disable_local(); 2459 handled = intel_pmu_drain_bts_buffer(); 2460 handled += intel_bts_interrupt(); 2461 } 2462 status = intel_pmu_get_status(); 2463 if (!status) 2464 goto done; 2465 again: 2466 intel_pmu_lbr_read(); 2467 if (++loops > 100) { 2468 static bool warned; 2469 2470 if (!warned) { 2471 WARN(1, "perfevents: irq loop stuck!\n"); 2472 perf_event_print_debug(); 2473 warned = true; 2474 } 2475 intel_pmu_reset(); 2476 goto done; 2477 } 2478 2479 2480 handled += handle_pmi_common(regs, status); 2481 done: 2482 /* Ack the PMI in the APIC */ 2483 apic_write(APIC_LVTPC, APIC_DM_NMI); 2484 2485 /* 2486 * The counters start counting immediately while ack the status. 2487 * Make it as close as possible to IRET. This avoids bogus 2488 * freezing on Skylake CPUs. 2489 */ 2490 if (status) { 2491 intel_pmu_ack_status(status); 2492 } else { 2493 /* 2494 * CPU may issues two PMIs very close to each other. 2495 * When the PMI handler services the first one, the 2496 * GLOBAL_STATUS is already updated to reflect both. 2497 * When it IRETs, the second PMI is immediately 2498 * handled and it sees clear status. At the meantime, 2499 * there may be a third PMI, because the freezing bit 2500 * isn't set since the ack in first PMI handlers. 2501 * Double check if there is more work to be done. 2502 */ 2503 status = intel_pmu_get_status(); 2504 if (status) 2505 goto again; 2506 } 2507 2508 if (bts) 2509 intel_bts_enable_local(); 2510 cpuc->enabled = pmu_enabled; 2511 return handled; 2512 } 2513 2514 /* 2515 * This handler is triggered by the local APIC, so the APIC IRQ handling 2516 * rules apply: 2517 */ 2518 static int intel_pmu_handle_irq(struct pt_regs *regs) 2519 { 2520 struct cpu_hw_events *cpuc; 2521 int loops; 2522 u64 status; 2523 int handled; 2524 int pmu_enabled; 2525 2526 cpuc = this_cpu_ptr(&cpu_hw_events); 2527 2528 /* 2529 * Save the PMU state. 2530 * It needs to be restored when leaving the handler. 2531 */ 2532 pmu_enabled = cpuc->enabled; 2533 /* 2534 * No known reason to not always do late ACK, 2535 * but just in case do it opt-in. 2536 */ 2537 if (!x86_pmu.late_ack) 2538 apic_write(APIC_LVTPC, APIC_DM_NMI); 2539 intel_bts_disable_local(); 2540 cpuc->enabled = 0; 2541 __intel_pmu_disable_all(); 2542 handled = intel_pmu_drain_bts_buffer(); 2543 handled += intel_bts_interrupt(); 2544 status = intel_pmu_get_status(); 2545 if (!status) 2546 goto done; 2547 2548 loops = 0; 2549 again: 2550 intel_pmu_lbr_read(); 2551 intel_pmu_ack_status(status); 2552 if (++loops > 100) { 2553 static bool warned; 2554 2555 if (!warned) { 2556 WARN(1, "perfevents: irq loop stuck!\n"); 2557 perf_event_print_debug(); 2558 warned = true; 2559 } 2560 intel_pmu_reset(); 2561 goto done; 2562 } 2563 2564 handled += handle_pmi_common(regs, status); 2565 2566 /* 2567 * Repeat if there is more work to be done: 2568 */ 2569 status = intel_pmu_get_status(); 2570 if (status) 2571 goto again; 2572 2573 done: 2574 /* Only restore PMU state when it's active. See x86_pmu_disable(). */ 2575 cpuc->enabled = pmu_enabled; 2576 if (pmu_enabled) 2577 __intel_pmu_enable_all(0, true); 2578 intel_bts_enable_local(); 2579 2580 /* 2581 * Only unmask the NMI after the overflow counters 2582 * have been reset. This avoids spurious NMIs on 2583 * Haswell CPUs. 2584 */ 2585 if (x86_pmu.late_ack) 2586 apic_write(APIC_LVTPC, APIC_DM_NMI); 2587 return handled; 2588 } 2589 2590 static struct event_constraint * 2591 intel_bts_constraints(struct perf_event *event) 2592 { 2593 if (unlikely(intel_pmu_has_bts(event))) 2594 return &bts_constraint; 2595 2596 return NULL; 2597 } 2598 2599 static int intel_alt_er(int idx, u64 config) 2600 { 2601 int alt_idx = idx; 2602 2603 if (!(x86_pmu.flags & PMU_FL_HAS_RSP_1)) 2604 return idx; 2605 2606 if (idx == EXTRA_REG_RSP_0) 2607 alt_idx = EXTRA_REG_RSP_1; 2608 2609 if (idx == EXTRA_REG_RSP_1) 2610 alt_idx = EXTRA_REG_RSP_0; 2611 2612 if (config & ~x86_pmu.extra_regs[alt_idx].valid_mask) 2613 return idx; 2614 2615 return alt_idx; 2616 } 2617 2618 static void intel_fixup_er(struct perf_event *event, int idx) 2619 { 2620 event->hw.extra_reg.idx = idx; 2621 2622 if (idx == EXTRA_REG_RSP_0) { 2623 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 2624 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_0].event; 2625 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0; 2626 } else if (idx == EXTRA_REG_RSP_1) { 2627 event->hw.config &= ~INTEL_ARCH_EVENT_MASK; 2628 event->hw.config |= x86_pmu.extra_regs[EXTRA_REG_RSP_1].event; 2629 event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1; 2630 } 2631 } 2632 2633 /* 2634 * manage allocation of shared extra msr for certain events 2635 * 2636 * sharing can be: 2637 * per-cpu: to be shared between the various events on a single PMU 2638 * per-core: per-cpu + shared by HT threads 2639 */ 2640 static struct event_constraint * 2641 __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc, 2642 struct perf_event *event, 2643 struct hw_perf_event_extra *reg) 2644 { 2645 struct event_constraint *c = &emptyconstraint; 2646 struct er_account *era; 2647 unsigned long flags; 2648 int idx = reg->idx; 2649 2650 /* 2651 * reg->alloc can be set due to existing state, so for fake cpuc we 2652 * need to ignore this, otherwise we might fail to allocate proper fake 2653 * state for this extra reg constraint. Also see the comment below. 2654 */ 2655 if (reg->alloc && !cpuc->is_fake) 2656 return NULL; /* call x86_get_event_constraint() */ 2657 2658 again: 2659 era = &cpuc->shared_regs->regs[idx]; 2660 /* 2661 * we use spin_lock_irqsave() to avoid lockdep issues when 2662 * passing a fake cpuc 2663 */ 2664 raw_spin_lock_irqsave(&era->lock, flags); 2665 2666 if (!atomic_read(&era->ref) || era->config == reg->config) { 2667 2668 /* 2669 * If its a fake cpuc -- as per validate_{group,event}() we 2670 * shouldn't touch event state and we can avoid doing so 2671 * since both will only call get_event_constraints() once 2672 * on each event, this avoids the need for reg->alloc. 2673 * 2674 * Not doing the ER fixup will only result in era->reg being 2675 * wrong, but since we won't actually try and program hardware 2676 * this isn't a problem either. 2677 */ 2678 if (!cpuc->is_fake) { 2679 if (idx != reg->idx) 2680 intel_fixup_er(event, idx); 2681 2682 /* 2683 * x86_schedule_events() can call get_event_constraints() 2684 * multiple times on events in the case of incremental 2685 * scheduling(). reg->alloc ensures we only do the ER 2686 * allocation once. 2687 */ 2688 reg->alloc = 1; 2689 } 2690 2691 /* lock in msr value */ 2692 era->config = reg->config; 2693 era->reg = reg->reg; 2694 2695 /* one more user */ 2696 atomic_inc(&era->ref); 2697 2698 /* 2699 * need to call x86_get_event_constraint() 2700 * to check if associated event has constraints 2701 */ 2702 c = NULL; 2703 } else { 2704 idx = intel_alt_er(idx, reg->config); 2705 if (idx != reg->idx) { 2706 raw_spin_unlock_irqrestore(&era->lock, flags); 2707 goto again; 2708 } 2709 } 2710 raw_spin_unlock_irqrestore(&era->lock, flags); 2711 2712 return c; 2713 } 2714 2715 static void 2716 __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc, 2717 struct hw_perf_event_extra *reg) 2718 { 2719 struct er_account *era; 2720 2721 /* 2722 * Only put constraint if extra reg was actually allocated. Also takes 2723 * care of event which do not use an extra shared reg. 2724 * 2725 * Also, if this is a fake cpuc we shouldn't touch any event state 2726 * (reg->alloc) and we don't care about leaving inconsistent cpuc state 2727 * either since it'll be thrown out. 2728 */ 2729 if (!reg->alloc || cpuc->is_fake) 2730 return; 2731 2732 era = &cpuc->shared_regs->regs[reg->idx]; 2733 2734 /* one fewer user */ 2735 atomic_dec(&era->ref); 2736 2737 /* allocate again next time */ 2738 reg->alloc = 0; 2739 } 2740 2741 static struct event_constraint * 2742 intel_shared_regs_constraints(struct cpu_hw_events *cpuc, 2743 struct perf_event *event) 2744 { 2745 struct event_constraint *c = NULL, *d; 2746 struct hw_perf_event_extra *xreg, *breg; 2747 2748 xreg = &event->hw.extra_reg; 2749 if (xreg->idx != EXTRA_REG_NONE) { 2750 c = __intel_shared_reg_get_constraints(cpuc, event, xreg); 2751 if (c == &emptyconstraint) 2752 return c; 2753 } 2754 breg = &event->hw.branch_reg; 2755 if (breg->idx != EXTRA_REG_NONE) { 2756 d = __intel_shared_reg_get_constraints(cpuc, event, breg); 2757 if (d == &emptyconstraint) { 2758 __intel_shared_reg_put_constraints(cpuc, xreg); 2759 c = d; 2760 } 2761 } 2762 return c; 2763 } 2764 2765 struct event_constraint * 2766 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 2767 struct perf_event *event) 2768 { 2769 struct event_constraint *c; 2770 2771 if (x86_pmu.event_constraints) { 2772 for_each_event_constraint(c, x86_pmu.event_constraints) { 2773 if (constraint_match(c, event->hw.config)) { 2774 event->hw.flags |= c->flags; 2775 return c; 2776 } 2777 } 2778 } 2779 2780 return &unconstrained; 2781 } 2782 2783 static struct event_constraint * 2784 __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 2785 struct perf_event *event) 2786 { 2787 struct event_constraint *c; 2788 2789 c = intel_bts_constraints(event); 2790 if (c) 2791 return c; 2792 2793 c = intel_shared_regs_constraints(cpuc, event); 2794 if (c) 2795 return c; 2796 2797 c = intel_pebs_constraints(event); 2798 if (c) 2799 return c; 2800 2801 return x86_get_event_constraints(cpuc, idx, event); 2802 } 2803 2804 static void 2805 intel_start_scheduling(struct cpu_hw_events *cpuc) 2806 { 2807 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2808 struct intel_excl_states *xl; 2809 int tid = cpuc->excl_thread_id; 2810 2811 /* 2812 * nothing needed if in group validation mode 2813 */ 2814 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2815 return; 2816 2817 /* 2818 * no exclusion needed 2819 */ 2820 if (WARN_ON_ONCE(!excl_cntrs)) 2821 return; 2822 2823 xl = &excl_cntrs->states[tid]; 2824 2825 xl->sched_started = true; 2826 /* 2827 * lock shared state until we are done scheduling 2828 * in stop_event_scheduling() 2829 * makes scheduling appear as a transaction 2830 */ 2831 raw_spin_lock(&excl_cntrs->lock); 2832 } 2833 2834 static void intel_commit_scheduling(struct cpu_hw_events *cpuc, int idx, int cntr) 2835 { 2836 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2837 struct event_constraint *c = cpuc->event_constraint[idx]; 2838 struct intel_excl_states *xl; 2839 int tid = cpuc->excl_thread_id; 2840 2841 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2842 return; 2843 2844 if (WARN_ON_ONCE(!excl_cntrs)) 2845 return; 2846 2847 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) 2848 return; 2849 2850 xl = &excl_cntrs->states[tid]; 2851 2852 lockdep_assert_held(&excl_cntrs->lock); 2853 2854 if (c->flags & PERF_X86_EVENT_EXCL) 2855 xl->state[cntr] = INTEL_EXCL_EXCLUSIVE; 2856 else 2857 xl->state[cntr] = INTEL_EXCL_SHARED; 2858 } 2859 2860 static void 2861 intel_stop_scheduling(struct cpu_hw_events *cpuc) 2862 { 2863 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2864 struct intel_excl_states *xl; 2865 int tid = cpuc->excl_thread_id; 2866 2867 /* 2868 * nothing needed if in group validation mode 2869 */ 2870 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2871 return; 2872 /* 2873 * no exclusion needed 2874 */ 2875 if (WARN_ON_ONCE(!excl_cntrs)) 2876 return; 2877 2878 xl = &excl_cntrs->states[tid]; 2879 2880 xl->sched_started = false; 2881 /* 2882 * release shared state lock (acquired in intel_start_scheduling()) 2883 */ 2884 raw_spin_unlock(&excl_cntrs->lock); 2885 } 2886 2887 static struct event_constraint * 2888 dyn_constraint(struct cpu_hw_events *cpuc, struct event_constraint *c, int idx) 2889 { 2890 WARN_ON_ONCE(!cpuc->constraint_list); 2891 2892 if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) { 2893 struct event_constraint *cx; 2894 2895 /* 2896 * grab pre-allocated constraint entry 2897 */ 2898 cx = &cpuc->constraint_list[idx]; 2899 2900 /* 2901 * initialize dynamic constraint 2902 * with static constraint 2903 */ 2904 *cx = *c; 2905 2906 /* 2907 * mark constraint as dynamic 2908 */ 2909 cx->flags |= PERF_X86_EVENT_DYNAMIC; 2910 c = cx; 2911 } 2912 2913 return c; 2914 } 2915 2916 static struct event_constraint * 2917 intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event, 2918 int idx, struct event_constraint *c) 2919 { 2920 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 2921 struct intel_excl_states *xlo; 2922 int tid = cpuc->excl_thread_id; 2923 int is_excl, i, w; 2924 2925 /* 2926 * validating a group does not require 2927 * enforcing cross-thread exclusion 2928 */ 2929 if (cpuc->is_fake || !is_ht_workaround_enabled()) 2930 return c; 2931 2932 /* 2933 * no exclusion needed 2934 */ 2935 if (WARN_ON_ONCE(!excl_cntrs)) 2936 return c; 2937 2938 /* 2939 * because we modify the constraint, we need 2940 * to make a copy. Static constraints come 2941 * from static const tables. 2942 * 2943 * only needed when constraint has not yet 2944 * been cloned (marked dynamic) 2945 */ 2946 c = dyn_constraint(cpuc, c, idx); 2947 2948 /* 2949 * From here on, the constraint is dynamic. 2950 * Either it was just allocated above, or it 2951 * was allocated during a earlier invocation 2952 * of this function 2953 */ 2954 2955 /* 2956 * state of sibling HT 2957 */ 2958 xlo = &excl_cntrs->states[tid ^ 1]; 2959 2960 /* 2961 * event requires exclusive counter access 2962 * across HT threads 2963 */ 2964 is_excl = c->flags & PERF_X86_EVENT_EXCL; 2965 if (is_excl && !(event->hw.flags & PERF_X86_EVENT_EXCL_ACCT)) { 2966 event->hw.flags |= PERF_X86_EVENT_EXCL_ACCT; 2967 if (!cpuc->n_excl++) 2968 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 1); 2969 } 2970 2971 /* 2972 * Modify static constraint with current dynamic 2973 * state of thread 2974 * 2975 * EXCLUSIVE: sibling counter measuring exclusive event 2976 * SHARED : sibling counter measuring non-exclusive event 2977 * UNUSED : sibling counter unused 2978 */ 2979 w = c->weight; 2980 for_each_set_bit(i, c->idxmsk, X86_PMC_IDX_MAX) { 2981 /* 2982 * exclusive event in sibling counter 2983 * our corresponding counter cannot be used 2984 * regardless of our event 2985 */ 2986 if (xlo->state[i] == INTEL_EXCL_EXCLUSIVE) { 2987 __clear_bit(i, c->idxmsk); 2988 w--; 2989 continue; 2990 } 2991 /* 2992 * if measuring an exclusive event, sibling 2993 * measuring non-exclusive, then counter cannot 2994 * be used 2995 */ 2996 if (is_excl && xlo->state[i] == INTEL_EXCL_SHARED) { 2997 __clear_bit(i, c->idxmsk); 2998 w--; 2999 continue; 3000 } 3001 } 3002 3003 /* 3004 * if we return an empty mask, then switch 3005 * back to static empty constraint to avoid 3006 * the cost of freeing later on 3007 */ 3008 if (!w) 3009 c = &emptyconstraint; 3010 3011 c->weight = w; 3012 3013 return c; 3014 } 3015 3016 static struct event_constraint * 3017 intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3018 struct perf_event *event) 3019 { 3020 struct event_constraint *c1, *c2; 3021 3022 c1 = cpuc->event_constraint[idx]; 3023 3024 /* 3025 * first time only 3026 * - static constraint: no change across incremental scheduling calls 3027 * - dynamic constraint: handled by intel_get_excl_constraints() 3028 */ 3029 c2 = __intel_get_event_constraints(cpuc, idx, event); 3030 if (c1) { 3031 WARN_ON_ONCE(!(c1->flags & PERF_X86_EVENT_DYNAMIC)); 3032 bitmap_copy(c1->idxmsk, c2->idxmsk, X86_PMC_IDX_MAX); 3033 c1->weight = c2->weight; 3034 c2 = c1; 3035 } 3036 3037 if (cpuc->excl_cntrs) 3038 return intel_get_excl_constraints(cpuc, event, idx, c2); 3039 3040 return c2; 3041 } 3042 3043 static void intel_put_excl_constraints(struct cpu_hw_events *cpuc, 3044 struct perf_event *event) 3045 { 3046 struct hw_perf_event *hwc = &event->hw; 3047 struct intel_excl_cntrs *excl_cntrs = cpuc->excl_cntrs; 3048 int tid = cpuc->excl_thread_id; 3049 struct intel_excl_states *xl; 3050 3051 /* 3052 * nothing needed if in group validation mode 3053 */ 3054 if (cpuc->is_fake) 3055 return; 3056 3057 if (WARN_ON_ONCE(!excl_cntrs)) 3058 return; 3059 3060 if (hwc->flags & PERF_X86_EVENT_EXCL_ACCT) { 3061 hwc->flags &= ~PERF_X86_EVENT_EXCL_ACCT; 3062 if (!--cpuc->n_excl) 3063 WRITE_ONCE(excl_cntrs->has_exclusive[tid], 0); 3064 } 3065 3066 /* 3067 * If event was actually assigned, then mark the counter state as 3068 * unused now. 3069 */ 3070 if (hwc->idx >= 0) { 3071 xl = &excl_cntrs->states[tid]; 3072 3073 /* 3074 * put_constraint may be called from x86_schedule_events() 3075 * which already has the lock held so here make locking 3076 * conditional. 3077 */ 3078 if (!xl->sched_started) 3079 raw_spin_lock(&excl_cntrs->lock); 3080 3081 xl->state[hwc->idx] = INTEL_EXCL_UNUSED; 3082 3083 if (!xl->sched_started) 3084 raw_spin_unlock(&excl_cntrs->lock); 3085 } 3086 } 3087 3088 static void 3089 intel_put_shared_regs_event_constraints(struct cpu_hw_events *cpuc, 3090 struct perf_event *event) 3091 { 3092 struct hw_perf_event_extra *reg; 3093 3094 reg = &event->hw.extra_reg; 3095 if (reg->idx != EXTRA_REG_NONE) 3096 __intel_shared_reg_put_constraints(cpuc, reg); 3097 3098 reg = &event->hw.branch_reg; 3099 if (reg->idx != EXTRA_REG_NONE) 3100 __intel_shared_reg_put_constraints(cpuc, reg); 3101 } 3102 3103 static void intel_put_event_constraints(struct cpu_hw_events *cpuc, 3104 struct perf_event *event) 3105 { 3106 intel_put_shared_regs_event_constraints(cpuc, event); 3107 3108 /* 3109 * is PMU has exclusive counter restrictions, then 3110 * all events are subject to and must call the 3111 * put_excl_constraints() routine 3112 */ 3113 if (cpuc->excl_cntrs) 3114 intel_put_excl_constraints(cpuc, event); 3115 } 3116 3117 static void intel_pebs_aliases_core2(struct perf_event *event) 3118 { 3119 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3120 /* 3121 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3122 * (0x003c) so that we can use it with PEBS. 3123 * 3124 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3125 * PEBS capable. However we can use INST_RETIRED.ANY_P 3126 * (0x00c0), which is a PEBS capable event, to get the same 3127 * count. 3128 * 3129 * INST_RETIRED.ANY_P counts the number of cycles that retires 3130 * CNTMASK instructions. By setting CNTMASK to a value (16) 3131 * larger than the maximum number of instructions that can be 3132 * retired per cycle (4) and then inverting the condition, we 3133 * count all cycles that retire 16 or less instructions, which 3134 * is every cycle. 3135 * 3136 * Thereby we gain a PEBS capable cycle counter. 3137 */ 3138 u64 alt_config = X86_CONFIG(.event=0xc0, .inv=1, .cmask=16); 3139 3140 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3141 event->hw.config = alt_config; 3142 } 3143 } 3144 3145 static void intel_pebs_aliases_snb(struct perf_event *event) 3146 { 3147 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3148 /* 3149 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3150 * (0x003c) so that we can use it with PEBS. 3151 * 3152 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3153 * PEBS capable. However we can use UOPS_RETIRED.ALL 3154 * (0x01c2), which is a PEBS capable event, to get the same 3155 * count. 3156 * 3157 * UOPS_RETIRED.ALL counts the number of cycles that retires 3158 * CNTMASK micro-ops. By setting CNTMASK to a value (16) 3159 * larger than the maximum number of micro-ops that can be 3160 * retired per cycle (4) and then inverting the condition, we 3161 * count all cycles that retire 16 or less micro-ops, which 3162 * is every cycle. 3163 * 3164 * Thereby we gain a PEBS capable cycle counter. 3165 */ 3166 u64 alt_config = X86_CONFIG(.event=0xc2, .umask=0x01, .inv=1, .cmask=16); 3167 3168 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3169 event->hw.config = alt_config; 3170 } 3171 } 3172 3173 static void intel_pebs_aliases_precdist(struct perf_event *event) 3174 { 3175 if ((event->hw.config & X86_RAW_EVENT_MASK) == 0x003c) { 3176 /* 3177 * Use an alternative encoding for CPU_CLK_UNHALTED.THREAD_P 3178 * (0x003c) so that we can use it with PEBS. 3179 * 3180 * The regular CPU_CLK_UNHALTED.THREAD_P event (0x003c) isn't 3181 * PEBS capable. However we can use INST_RETIRED.PREC_DIST 3182 * (0x01c0), which is a PEBS capable event, to get the same 3183 * count. 3184 * 3185 * The PREC_DIST event has special support to minimize sample 3186 * shadowing effects. One drawback is that it can be 3187 * only programmed on counter 1, but that seems like an 3188 * acceptable trade off. 3189 */ 3190 u64 alt_config = X86_CONFIG(.event=0xc0, .umask=0x01, .inv=1, .cmask=16); 3191 3192 alt_config |= (event->hw.config & ~X86_RAW_EVENT_MASK); 3193 event->hw.config = alt_config; 3194 } 3195 } 3196 3197 static void intel_pebs_aliases_ivb(struct perf_event *event) 3198 { 3199 if (event->attr.precise_ip < 3) 3200 return intel_pebs_aliases_snb(event); 3201 return intel_pebs_aliases_precdist(event); 3202 } 3203 3204 static void intel_pebs_aliases_skl(struct perf_event *event) 3205 { 3206 if (event->attr.precise_ip < 3) 3207 return intel_pebs_aliases_core2(event); 3208 return intel_pebs_aliases_precdist(event); 3209 } 3210 3211 static unsigned long intel_pmu_large_pebs_flags(struct perf_event *event) 3212 { 3213 unsigned long flags = x86_pmu.large_pebs_flags; 3214 3215 if (event->attr.use_clockid) 3216 flags &= ~PERF_SAMPLE_TIME; 3217 if (!event->attr.exclude_kernel) 3218 flags &= ~PERF_SAMPLE_REGS_USER; 3219 if (event->attr.sample_regs_user & ~PEBS_GP_REGS) 3220 flags &= ~(PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR); 3221 return flags; 3222 } 3223 3224 static int intel_pmu_bts_config(struct perf_event *event) 3225 { 3226 struct perf_event_attr *attr = &event->attr; 3227 3228 if (unlikely(intel_pmu_has_bts(event))) { 3229 /* BTS is not supported by this architecture. */ 3230 if (!x86_pmu.bts_active) 3231 return -EOPNOTSUPP; 3232 3233 /* BTS is currently only allowed for user-mode. */ 3234 if (!attr->exclude_kernel) 3235 return -EOPNOTSUPP; 3236 3237 /* BTS is not allowed for precise events. */ 3238 if (attr->precise_ip) 3239 return -EOPNOTSUPP; 3240 3241 /* disallow bts if conflicting events are present */ 3242 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3243 return -EBUSY; 3244 3245 event->destroy = hw_perf_lbr_event_destroy; 3246 } 3247 3248 return 0; 3249 } 3250 3251 static int core_pmu_hw_config(struct perf_event *event) 3252 { 3253 int ret = x86_pmu_hw_config(event); 3254 3255 if (ret) 3256 return ret; 3257 3258 return intel_pmu_bts_config(event); 3259 } 3260 3261 static int intel_pmu_hw_config(struct perf_event *event) 3262 { 3263 int ret = x86_pmu_hw_config(event); 3264 3265 if (ret) 3266 return ret; 3267 3268 ret = intel_pmu_bts_config(event); 3269 if (ret) 3270 return ret; 3271 3272 if (event->attr.precise_ip) { 3273 if (!(event->attr.freq || (event->attr.wakeup_events && !event->attr.watermark))) { 3274 event->hw.flags |= PERF_X86_EVENT_AUTO_RELOAD; 3275 if (!(event->attr.sample_type & 3276 ~intel_pmu_large_pebs_flags(event))) 3277 event->hw.flags |= PERF_X86_EVENT_LARGE_PEBS; 3278 } 3279 if (x86_pmu.pebs_aliases) 3280 x86_pmu.pebs_aliases(event); 3281 3282 if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) 3283 event->attr.sample_type |= __PERF_SAMPLE_CALLCHAIN_EARLY; 3284 } 3285 3286 if (needs_branch_stack(event)) { 3287 ret = intel_pmu_setup_lbr_filter(event); 3288 if (ret) 3289 return ret; 3290 3291 /* 3292 * BTS is set up earlier in this path, so don't account twice 3293 */ 3294 if (!unlikely(intel_pmu_has_bts(event))) { 3295 /* disallow lbr if conflicting events are present */ 3296 if (x86_add_exclusive(x86_lbr_exclusive_lbr)) 3297 return -EBUSY; 3298 3299 event->destroy = hw_perf_lbr_event_destroy; 3300 } 3301 } 3302 3303 if (event->attr.type != PERF_TYPE_RAW) 3304 return 0; 3305 3306 if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY)) 3307 return 0; 3308 3309 if (x86_pmu.version < 3) 3310 return -EINVAL; 3311 3312 if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN)) 3313 return -EACCES; 3314 3315 event->hw.config |= ARCH_PERFMON_EVENTSEL_ANY; 3316 3317 return 0; 3318 } 3319 3320 struct perf_guest_switch_msr *perf_guest_get_msrs(int *nr) 3321 { 3322 if (x86_pmu.guest_get_msrs) 3323 return x86_pmu.guest_get_msrs(nr); 3324 *nr = 0; 3325 return NULL; 3326 } 3327 EXPORT_SYMBOL_GPL(perf_guest_get_msrs); 3328 3329 static struct perf_guest_switch_msr *intel_guest_get_msrs(int *nr) 3330 { 3331 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3332 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3333 3334 arr[0].msr = MSR_CORE_PERF_GLOBAL_CTRL; 3335 arr[0].host = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_guest_mask; 3336 arr[0].guest = x86_pmu.intel_ctrl & ~cpuc->intel_ctrl_host_mask; 3337 if (x86_pmu.flags & PMU_FL_PEBS_ALL) 3338 arr[0].guest &= ~cpuc->pebs_enabled; 3339 else 3340 arr[0].guest &= ~(cpuc->pebs_enabled & PEBS_COUNTER_MASK); 3341 *nr = 1; 3342 3343 if (x86_pmu.pebs && x86_pmu.pebs_no_isolation) { 3344 /* 3345 * If PMU counter has PEBS enabled it is not enough to 3346 * disable counter on a guest entry since PEBS memory 3347 * write can overshoot guest entry and corrupt guest 3348 * memory. Disabling PEBS solves the problem. 3349 * 3350 * Don't do this if the CPU already enforces it. 3351 */ 3352 arr[1].msr = MSR_IA32_PEBS_ENABLE; 3353 arr[1].host = cpuc->pebs_enabled; 3354 arr[1].guest = 0; 3355 *nr = 2; 3356 } 3357 3358 return arr; 3359 } 3360 3361 static struct perf_guest_switch_msr *core_guest_get_msrs(int *nr) 3362 { 3363 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3364 struct perf_guest_switch_msr *arr = cpuc->guest_switch_msrs; 3365 int idx; 3366 3367 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 3368 struct perf_event *event = cpuc->events[idx]; 3369 3370 arr[idx].msr = x86_pmu_config_addr(idx); 3371 arr[idx].host = arr[idx].guest = 0; 3372 3373 if (!test_bit(idx, cpuc->active_mask)) 3374 continue; 3375 3376 arr[idx].host = arr[idx].guest = 3377 event->hw.config | ARCH_PERFMON_EVENTSEL_ENABLE; 3378 3379 if (event->attr.exclude_host) 3380 arr[idx].host &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3381 else if (event->attr.exclude_guest) 3382 arr[idx].guest &= ~ARCH_PERFMON_EVENTSEL_ENABLE; 3383 } 3384 3385 *nr = x86_pmu.num_counters; 3386 return arr; 3387 } 3388 3389 static void core_pmu_enable_event(struct perf_event *event) 3390 { 3391 if (!event->attr.exclude_host) 3392 x86_pmu_enable_event(event); 3393 } 3394 3395 static void core_pmu_enable_all(int added) 3396 { 3397 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 3398 int idx; 3399 3400 for (idx = 0; idx < x86_pmu.num_counters; idx++) { 3401 struct hw_perf_event *hwc = &cpuc->events[idx]->hw; 3402 3403 if (!test_bit(idx, cpuc->active_mask) || 3404 cpuc->events[idx]->attr.exclude_host) 3405 continue; 3406 3407 __x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE); 3408 } 3409 } 3410 3411 static int hsw_hw_config(struct perf_event *event) 3412 { 3413 int ret = intel_pmu_hw_config(event); 3414 3415 if (ret) 3416 return ret; 3417 if (!boot_cpu_has(X86_FEATURE_RTM) && !boot_cpu_has(X86_FEATURE_HLE)) 3418 return 0; 3419 event->hw.config |= event->attr.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED); 3420 3421 /* 3422 * IN_TX/IN_TX-CP filters are not supported by the Haswell PMU with 3423 * PEBS or in ANY thread mode. Since the results are non-sensical forbid 3424 * this combination. 3425 */ 3426 if ((event->hw.config & (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED)) && 3427 ((event->hw.config & ARCH_PERFMON_EVENTSEL_ANY) || 3428 event->attr.precise_ip > 0)) 3429 return -EOPNOTSUPP; 3430 3431 if (event_is_checkpointed(event)) { 3432 /* 3433 * Sampling of checkpointed events can cause situations where 3434 * the CPU constantly aborts because of a overflow, which is 3435 * then checkpointed back and ignored. Forbid checkpointing 3436 * for sampling. 3437 * 3438 * But still allow a long sampling period, so that perf stat 3439 * from KVM works. 3440 */ 3441 if (event->attr.sample_period > 0 && 3442 event->attr.sample_period < 0x7fffffff) 3443 return -EOPNOTSUPP; 3444 } 3445 return 0; 3446 } 3447 3448 static struct event_constraint counter0_constraint = 3449 INTEL_ALL_EVENT_CONSTRAINT(0, 0x1); 3450 3451 static struct event_constraint counter2_constraint = 3452 EVENT_CONSTRAINT(0, 0x4, 0); 3453 3454 static struct event_constraint fixed0_constraint = 3455 FIXED_EVENT_CONSTRAINT(0x00c0, 0); 3456 3457 static struct event_constraint fixed0_counter0_constraint = 3458 INTEL_ALL_EVENT_CONSTRAINT(0, 0x100000001ULL); 3459 3460 static struct event_constraint * 3461 hsw_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3462 struct perf_event *event) 3463 { 3464 struct event_constraint *c; 3465 3466 c = intel_get_event_constraints(cpuc, idx, event); 3467 3468 /* Handle special quirk on in_tx_checkpointed only in counter 2 */ 3469 if (event->hw.config & HSW_IN_TX_CHECKPOINTED) { 3470 if (c->idxmsk64 & (1U << 2)) 3471 return &counter2_constraint; 3472 return &emptyconstraint; 3473 } 3474 3475 return c; 3476 } 3477 3478 static struct event_constraint * 3479 icl_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3480 struct perf_event *event) 3481 { 3482 /* 3483 * Fixed counter 0 has less skid. 3484 * Force instruction:ppp in Fixed counter 0 3485 */ 3486 if ((event->attr.precise_ip == 3) && 3487 constraint_match(&fixed0_constraint, event->hw.config)) 3488 return &fixed0_constraint; 3489 3490 return hsw_get_event_constraints(cpuc, idx, event); 3491 } 3492 3493 static struct event_constraint * 3494 glp_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3495 struct perf_event *event) 3496 { 3497 struct event_constraint *c; 3498 3499 /* :ppp means to do reduced skid PEBS which is PMC0 only. */ 3500 if (event->attr.precise_ip == 3) 3501 return &counter0_constraint; 3502 3503 c = intel_get_event_constraints(cpuc, idx, event); 3504 3505 return c; 3506 } 3507 3508 static struct event_constraint * 3509 tnt_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3510 struct perf_event *event) 3511 { 3512 struct event_constraint *c; 3513 3514 /* 3515 * :ppp means to do reduced skid PEBS, 3516 * which is available on PMC0 and fixed counter 0. 3517 */ 3518 if (event->attr.precise_ip == 3) { 3519 /* Force instruction:ppp on PMC0 and Fixed counter 0 */ 3520 if (constraint_match(&fixed0_constraint, event->hw.config)) 3521 return &fixed0_counter0_constraint; 3522 3523 return &counter0_constraint; 3524 } 3525 3526 c = intel_get_event_constraints(cpuc, idx, event); 3527 3528 return c; 3529 } 3530 3531 static bool allow_tsx_force_abort = true; 3532 3533 static struct event_constraint * 3534 tfa_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 3535 struct perf_event *event) 3536 { 3537 struct event_constraint *c = hsw_get_event_constraints(cpuc, idx, event); 3538 3539 /* 3540 * Without TFA we must not use PMC3. 3541 */ 3542 if (!allow_tsx_force_abort && test_bit(3, c->idxmsk)) { 3543 c = dyn_constraint(cpuc, c, idx); 3544 c->idxmsk64 &= ~(1ULL << 3); 3545 c->weight--; 3546 } 3547 3548 return c; 3549 } 3550 3551 /* 3552 * Broadwell: 3553 * 3554 * The INST_RETIRED.ALL period always needs to have lowest 6 bits cleared 3555 * (BDM55) and it must not use a period smaller than 100 (BDM11). We combine 3556 * the two to enforce a minimum period of 128 (the smallest value that has bits 3557 * 0-5 cleared and >= 100). 3558 * 3559 * Because of how the code in x86_perf_event_set_period() works, the truncation 3560 * of the lower 6 bits is 'harmless' as we'll occasionally add a longer period 3561 * to make up for the 'lost' events due to carrying the 'error' in period_left. 3562 * 3563 * Therefore the effective (average) period matches the requested period, 3564 * despite coarser hardware granularity. 3565 */ 3566 static u64 bdw_limit_period(struct perf_event *event, u64 left) 3567 { 3568 if ((event->hw.config & INTEL_ARCH_EVENT_MASK) == 3569 X86_CONFIG(.event=0xc0, .umask=0x01)) { 3570 if (left < 128) 3571 left = 128; 3572 left &= ~0x3fULL; 3573 } 3574 return left; 3575 } 3576 3577 PMU_FORMAT_ATTR(event, "config:0-7" ); 3578 PMU_FORMAT_ATTR(umask, "config:8-15" ); 3579 PMU_FORMAT_ATTR(edge, "config:18" ); 3580 PMU_FORMAT_ATTR(pc, "config:19" ); 3581 PMU_FORMAT_ATTR(any, "config:21" ); /* v3 + */ 3582 PMU_FORMAT_ATTR(inv, "config:23" ); 3583 PMU_FORMAT_ATTR(cmask, "config:24-31" ); 3584 PMU_FORMAT_ATTR(in_tx, "config:32"); 3585 PMU_FORMAT_ATTR(in_tx_cp, "config:33"); 3586 3587 static struct attribute *intel_arch_formats_attr[] = { 3588 &format_attr_event.attr, 3589 &format_attr_umask.attr, 3590 &format_attr_edge.attr, 3591 &format_attr_pc.attr, 3592 &format_attr_inv.attr, 3593 &format_attr_cmask.attr, 3594 NULL, 3595 }; 3596 3597 ssize_t intel_event_sysfs_show(char *page, u64 config) 3598 { 3599 u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT); 3600 3601 return x86_event_sysfs_show(page, config, event); 3602 } 3603 3604 static struct intel_shared_regs *allocate_shared_regs(int cpu) 3605 { 3606 struct intel_shared_regs *regs; 3607 int i; 3608 3609 regs = kzalloc_node(sizeof(struct intel_shared_regs), 3610 GFP_KERNEL, cpu_to_node(cpu)); 3611 if (regs) { 3612 /* 3613 * initialize the locks to keep lockdep happy 3614 */ 3615 for (i = 0; i < EXTRA_REG_MAX; i++) 3616 raw_spin_lock_init(®s->regs[i].lock); 3617 3618 regs->core_id = -1; 3619 } 3620 return regs; 3621 } 3622 3623 static struct intel_excl_cntrs *allocate_excl_cntrs(int cpu) 3624 { 3625 struct intel_excl_cntrs *c; 3626 3627 c = kzalloc_node(sizeof(struct intel_excl_cntrs), 3628 GFP_KERNEL, cpu_to_node(cpu)); 3629 if (c) { 3630 raw_spin_lock_init(&c->lock); 3631 c->core_id = -1; 3632 } 3633 return c; 3634 } 3635 3636 3637 int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 3638 { 3639 cpuc->pebs_record_size = x86_pmu.pebs_record_size; 3640 3641 if (x86_pmu.extra_regs || x86_pmu.lbr_sel_map) { 3642 cpuc->shared_regs = allocate_shared_regs(cpu); 3643 if (!cpuc->shared_regs) 3644 goto err; 3645 } 3646 3647 if (x86_pmu.flags & (PMU_FL_EXCL_CNTRS | PMU_FL_TFA)) { 3648 size_t sz = X86_PMC_IDX_MAX * sizeof(struct event_constraint); 3649 3650 cpuc->constraint_list = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu)); 3651 if (!cpuc->constraint_list) 3652 goto err_shared_regs; 3653 } 3654 3655 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 3656 cpuc->excl_cntrs = allocate_excl_cntrs(cpu); 3657 if (!cpuc->excl_cntrs) 3658 goto err_constraint_list; 3659 3660 cpuc->excl_thread_id = 0; 3661 } 3662 3663 return 0; 3664 3665 err_constraint_list: 3666 kfree(cpuc->constraint_list); 3667 cpuc->constraint_list = NULL; 3668 3669 err_shared_regs: 3670 kfree(cpuc->shared_regs); 3671 cpuc->shared_regs = NULL; 3672 3673 err: 3674 return -ENOMEM; 3675 } 3676 3677 static int intel_pmu_cpu_prepare(int cpu) 3678 { 3679 return intel_cpuc_prepare(&per_cpu(cpu_hw_events, cpu), cpu); 3680 } 3681 3682 static void flip_smm_bit(void *data) 3683 { 3684 unsigned long set = *(unsigned long *)data; 3685 3686 if (set > 0) { 3687 msr_set_bit(MSR_IA32_DEBUGCTLMSR, 3688 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 3689 } else { 3690 msr_clear_bit(MSR_IA32_DEBUGCTLMSR, 3691 DEBUGCTLMSR_FREEZE_IN_SMM_BIT); 3692 } 3693 } 3694 3695 static void intel_pmu_cpu_starting(int cpu) 3696 { 3697 struct cpu_hw_events *cpuc = &per_cpu(cpu_hw_events, cpu); 3698 int core_id = topology_core_id(cpu); 3699 int i; 3700 3701 init_debug_store_on_cpu(cpu); 3702 /* 3703 * Deal with CPUs that don't clear their LBRs on power-up. 3704 */ 3705 intel_pmu_lbr_reset(); 3706 3707 cpuc->lbr_sel = NULL; 3708 3709 if (x86_pmu.flags & PMU_FL_TFA) { 3710 WARN_ON_ONCE(cpuc->tfa_shadow); 3711 cpuc->tfa_shadow = ~0ULL; 3712 intel_set_tfa(cpuc, false); 3713 } 3714 3715 if (x86_pmu.version > 1) 3716 flip_smm_bit(&x86_pmu.attr_freeze_on_smi); 3717 3718 if (x86_pmu.counter_freezing) 3719 enable_counter_freeze(); 3720 3721 if (!cpuc->shared_regs) 3722 return; 3723 3724 if (!(x86_pmu.flags & PMU_FL_NO_HT_SHARING)) { 3725 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 3726 struct intel_shared_regs *pc; 3727 3728 pc = per_cpu(cpu_hw_events, i).shared_regs; 3729 if (pc && pc->core_id == core_id) { 3730 cpuc->kfree_on_online[0] = cpuc->shared_regs; 3731 cpuc->shared_regs = pc; 3732 break; 3733 } 3734 } 3735 cpuc->shared_regs->core_id = core_id; 3736 cpuc->shared_regs->refcnt++; 3737 } 3738 3739 if (x86_pmu.lbr_sel_map) 3740 cpuc->lbr_sel = &cpuc->shared_regs->regs[EXTRA_REG_LBR]; 3741 3742 if (x86_pmu.flags & PMU_FL_EXCL_CNTRS) { 3743 for_each_cpu(i, topology_sibling_cpumask(cpu)) { 3744 struct cpu_hw_events *sibling; 3745 struct intel_excl_cntrs *c; 3746 3747 sibling = &per_cpu(cpu_hw_events, i); 3748 c = sibling->excl_cntrs; 3749 if (c && c->core_id == core_id) { 3750 cpuc->kfree_on_online[1] = cpuc->excl_cntrs; 3751 cpuc->excl_cntrs = c; 3752 if (!sibling->excl_thread_id) 3753 cpuc->excl_thread_id = 1; 3754 break; 3755 } 3756 } 3757 cpuc->excl_cntrs->core_id = core_id; 3758 cpuc->excl_cntrs->refcnt++; 3759 } 3760 } 3761 3762 static void free_excl_cntrs(struct cpu_hw_events *cpuc) 3763 { 3764 struct intel_excl_cntrs *c; 3765 3766 c = cpuc->excl_cntrs; 3767 if (c) { 3768 if (c->core_id == -1 || --c->refcnt == 0) 3769 kfree(c); 3770 cpuc->excl_cntrs = NULL; 3771 } 3772 3773 kfree(cpuc->constraint_list); 3774 cpuc->constraint_list = NULL; 3775 } 3776 3777 static void intel_pmu_cpu_dying(int cpu) 3778 { 3779 fini_debug_store_on_cpu(cpu); 3780 3781 if (x86_pmu.counter_freezing) 3782 disable_counter_freeze(); 3783 } 3784 3785 void intel_cpuc_finish(struct cpu_hw_events *cpuc) 3786 { 3787 struct intel_shared_regs *pc; 3788 3789 pc = cpuc->shared_regs; 3790 if (pc) { 3791 if (pc->core_id == -1 || --pc->refcnt == 0) 3792 kfree(pc); 3793 cpuc->shared_regs = NULL; 3794 } 3795 3796 free_excl_cntrs(cpuc); 3797 } 3798 3799 static void intel_pmu_cpu_dead(int cpu) 3800 { 3801 intel_cpuc_finish(&per_cpu(cpu_hw_events, cpu)); 3802 } 3803 3804 static void intel_pmu_sched_task(struct perf_event_context *ctx, 3805 bool sched_in) 3806 { 3807 intel_pmu_pebs_sched_task(ctx, sched_in); 3808 intel_pmu_lbr_sched_task(ctx, sched_in); 3809 } 3810 3811 static int intel_pmu_check_period(struct perf_event *event, u64 value) 3812 { 3813 return intel_pmu_has_bts_period(event, value) ? -EINVAL : 0; 3814 } 3815 3816 PMU_FORMAT_ATTR(offcore_rsp, "config1:0-63"); 3817 3818 PMU_FORMAT_ATTR(ldlat, "config1:0-15"); 3819 3820 PMU_FORMAT_ATTR(frontend, "config1:0-23"); 3821 3822 static struct attribute *intel_arch3_formats_attr[] = { 3823 &format_attr_event.attr, 3824 &format_attr_umask.attr, 3825 &format_attr_edge.attr, 3826 &format_attr_pc.attr, 3827 &format_attr_any.attr, 3828 &format_attr_inv.attr, 3829 &format_attr_cmask.attr, 3830 NULL, 3831 }; 3832 3833 static struct attribute *hsw_format_attr[] = { 3834 &format_attr_in_tx.attr, 3835 &format_attr_in_tx_cp.attr, 3836 &format_attr_offcore_rsp.attr, 3837 &format_attr_ldlat.attr, 3838 NULL 3839 }; 3840 3841 static struct attribute *nhm_format_attr[] = { 3842 &format_attr_offcore_rsp.attr, 3843 &format_attr_ldlat.attr, 3844 NULL 3845 }; 3846 3847 static struct attribute *slm_format_attr[] = { 3848 &format_attr_offcore_rsp.attr, 3849 NULL 3850 }; 3851 3852 static struct attribute *skl_format_attr[] = { 3853 &format_attr_frontend.attr, 3854 NULL, 3855 }; 3856 3857 static __initconst const struct x86_pmu core_pmu = { 3858 .name = "core", 3859 .handle_irq = x86_pmu_handle_irq, 3860 .disable_all = x86_pmu_disable_all, 3861 .enable_all = core_pmu_enable_all, 3862 .enable = core_pmu_enable_event, 3863 .disable = x86_pmu_disable_event, 3864 .hw_config = core_pmu_hw_config, 3865 .schedule_events = x86_schedule_events, 3866 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 3867 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 3868 .event_map = intel_pmu_event_map, 3869 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 3870 .apic = 1, 3871 .large_pebs_flags = LARGE_PEBS_FLAGS, 3872 3873 /* 3874 * Intel PMCs cannot be accessed sanely above 32-bit width, 3875 * so we install an artificial 1<<31 period regardless of 3876 * the generic event period: 3877 */ 3878 .max_period = (1ULL<<31) - 1, 3879 .get_event_constraints = intel_get_event_constraints, 3880 .put_event_constraints = intel_put_event_constraints, 3881 .event_constraints = intel_core_event_constraints, 3882 .guest_get_msrs = core_guest_get_msrs, 3883 .format_attrs = intel_arch_formats_attr, 3884 .events_sysfs_show = intel_event_sysfs_show, 3885 3886 /* 3887 * Virtual (or funny metal) CPU can define x86_pmu.extra_regs 3888 * together with PMU version 1 and thus be using core_pmu with 3889 * shared_regs. We need following callbacks here to allocate 3890 * it properly. 3891 */ 3892 .cpu_prepare = intel_pmu_cpu_prepare, 3893 .cpu_starting = intel_pmu_cpu_starting, 3894 .cpu_dying = intel_pmu_cpu_dying, 3895 .cpu_dead = intel_pmu_cpu_dead, 3896 3897 .check_period = intel_pmu_check_period, 3898 }; 3899 3900 static struct attribute *intel_pmu_attrs[]; 3901 3902 static __initconst const struct x86_pmu intel_pmu = { 3903 .name = "Intel", 3904 .handle_irq = intel_pmu_handle_irq, 3905 .disable_all = intel_pmu_disable_all, 3906 .enable_all = intel_pmu_enable_all, 3907 .enable = intel_pmu_enable_event, 3908 .disable = intel_pmu_disable_event, 3909 .add = intel_pmu_add_event, 3910 .del = intel_pmu_del_event, 3911 .read = intel_pmu_read_event, 3912 .hw_config = intel_pmu_hw_config, 3913 .schedule_events = x86_schedule_events, 3914 .eventsel = MSR_ARCH_PERFMON_EVENTSEL0, 3915 .perfctr = MSR_ARCH_PERFMON_PERFCTR0, 3916 .event_map = intel_pmu_event_map, 3917 .max_events = ARRAY_SIZE(intel_perfmon_event_map), 3918 .apic = 1, 3919 .large_pebs_flags = LARGE_PEBS_FLAGS, 3920 /* 3921 * Intel PMCs cannot be accessed sanely above 32 bit width, 3922 * so we install an artificial 1<<31 period regardless of 3923 * the generic event period: 3924 */ 3925 .max_period = (1ULL << 31) - 1, 3926 .get_event_constraints = intel_get_event_constraints, 3927 .put_event_constraints = intel_put_event_constraints, 3928 .pebs_aliases = intel_pebs_aliases_core2, 3929 3930 .format_attrs = intel_arch3_formats_attr, 3931 .events_sysfs_show = intel_event_sysfs_show, 3932 3933 .attrs = intel_pmu_attrs, 3934 3935 .cpu_prepare = intel_pmu_cpu_prepare, 3936 .cpu_starting = intel_pmu_cpu_starting, 3937 .cpu_dying = intel_pmu_cpu_dying, 3938 .cpu_dead = intel_pmu_cpu_dead, 3939 3940 .guest_get_msrs = intel_guest_get_msrs, 3941 .sched_task = intel_pmu_sched_task, 3942 3943 .check_period = intel_pmu_check_period, 3944 }; 3945 3946 static __init void intel_clovertown_quirk(void) 3947 { 3948 /* 3949 * PEBS is unreliable due to: 3950 * 3951 * AJ67 - PEBS may experience CPL leaks 3952 * AJ68 - PEBS PMI may be delayed by one event 3953 * AJ69 - GLOBAL_STATUS[62] will only be set when DEBUGCTL[12] 3954 * AJ106 - FREEZE_LBRS_ON_PMI doesn't work in combination with PEBS 3955 * 3956 * AJ67 could be worked around by restricting the OS/USR flags. 3957 * AJ69 could be worked around by setting PMU_FREEZE_ON_PMI. 3958 * 3959 * AJ106 could possibly be worked around by not allowing LBR 3960 * usage from PEBS, including the fixup. 3961 * AJ68 could possibly be worked around by always programming 3962 * a pebs_event_reset[0] value and coping with the lost events. 3963 * 3964 * But taken together it might just make sense to not enable PEBS on 3965 * these chips. 3966 */ 3967 pr_warn("PEBS disabled due to CPU errata\n"); 3968 x86_pmu.pebs = 0; 3969 x86_pmu.pebs_constraints = NULL; 3970 } 3971 3972 static const struct x86_cpu_desc isolation_ucodes[] = { 3973 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_CORE, 3, 0x0000001f), 3974 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_ULT, 1, 0x0000001e), 3975 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_GT3E, 1, 0x00000015), 3976 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 2, 0x00000037), 3977 INTEL_CPU_DESC(INTEL_FAM6_HASWELL_X, 4, 0x0000000a), 3978 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_CORE, 4, 0x00000023), 3979 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_GT3E, 1, 0x00000014), 3980 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 2, 0x00000010), 3981 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 3, 0x07000009), 3982 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 4, 0x0f000009), 3983 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_XEON_D, 5, 0x0e000002), 3984 INTEL_CPU_DESC(INTEL_FAM6_BROADWELL_X, 2, 0x0b000014), 3985 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 3, 0x00000021), 3986 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_X, 4, 0x00000000), 3987 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_MOBILE, 3, 0x0000007c), 3988 INTEL_CPU_DESC(INTEL_FAM6_SKYLAKE_DESKTOP, 3, 0x0000007c), 3989 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 9, 0x0000004e), 3990 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 9, 0x0000004e), 3991 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 10, 0x0000004e), 3992 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 11, 0x0000004e), 3993 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_MOBILE, 12, 0x0000004e), 3994 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 10, 0x0000004e), 3995 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 11, 0x0000004e), 3996 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 12, 0x0000004e), 3997 INTEL_CPU_DESC(INTEL_FAM6_KABYLAKE_DESKTOP, 13, 0x0000004e), 3998 {} 3999 }; 4000 4001 static void intel_check_pebs_isolation(void) 4002 { 4003 x86_pmu.pebs_no_isolation = !x86_cpu_has_min_microcode_rev(isolation_ucodes); 4004 } 4005 4006 static __init void intel_pebs_isolation_quirk(void) 4007 { 4008 WARN_ON_ONCE(x86_pmu.check_microcode); 4009 x86_pmu.check_microcode = intel_check_pebs_isolation; 4010 intel_check_pebs_isolation(); 4011 } 4012 4013 static const struct x86_cpu_desc pebs_ucodes[] = { 4014 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE, 7, 0x00000028), 4015 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 6, 0x00000618), 4016 INTEL_CPU_DESC(INTEL_FAM6_SANDYBRIDGE_X, 7, 0x0000070c), 4017 {} 4018 }; 4019 4020 static bool intel_snb_pebs_broken(void) 4021 { 4022 return !x86_cpu_has_min_microcode_rev(pebs_ucodes); 4023 } 4024 4025 static void intel_snb_check_microcode(void) 4026 { 4027 if (intel_snb_pebs_broken() == x86_pmu.pebs_broken) 4028 return; 4029 4030 /* 4031 * Serialized by the microcode lock.. 4032 */ 4033 if (x86_pmu.pebs_broken) { 4034 pr_info("PEBS enabled due to microcode update\n"); 4035 x86_pmu.pebs_broken = 0; 4036 } else { 4037 pr_info("PEBS disabled due to CPU errata, please upgrade microcode\n"); 4038 x86_pmu.pebs_broken = 1; 4039 } 4040 } 4041 4042 static bool is_lbr_from(unsigned long msr) 4043 { 4044 unsigned long lbr_from_nr = x86_pmu.lbr_from + x86_pmu.lbr_nr; 4045 4046 return x86_pmu.lbr_from <= msr && msr < lbr_from_nr; 4047 } 4048 4049 /* 4050 * Under certain circumstances, access certain MSR may cause #GP. 4051 * The function tests if the input MSR can be safely accessed. 4052 */ 4053 static bool check_msr(unsigned long msr, u64 mask) 4054 { 4055 u64 val_old, val_new, val_tmp; 4056 4057 /* 4058 * Read the current value, change it and read it back to see if it 4059 * matches, this is needed to detect certain hardware emulators 4060 * (qemu/kvm) that don't trap on the MSR access and always return 0s. 4061 */ 4062 if (rdmsrl_safe(msr, &val_old)) 4063 return false; 4064 4065 /* 4066 * Only change the bits which can be updated by wrmsrl. 4067 */ 4068 val_tmp = val_old ^ mask; 4069 4070 if (is_lbr_from(msr)) 4071 val_tmp = lbr_from_signext_quirk_wr(val_tmp); 4072 4073 if (wrmsrl_safe(msr, val_tmp) || 4074 rdmsrl_safe(msr, &val_new)) 4075 return false; 4076 4077 /* 4078 * Quirk only affects validation in wrmsr(), so wrmsrl()'s value 4079 * should equal rdmsrl()'s even with the quirk. 4080 */ 4081 if (val_new != val_tmp) 4082 return false; 4083 4084 if (is_lbr_from(msr)) 4085 val_old = lbr_from_signext_quirk_wr(val_old); 4086 4087 /* Here it's sure that the MSR can be safely accessed. 4088 * Restore the old value and return. 4089 */ 4090 wrmsrl(msr, val_old); 4091 4092 return true; 4093 } 4094 4095 static __init void intel_sandybridge_quirk(void) 4096 { 4097 x86_pmu.check_microcode = intel_snb_check_microcode; 4098 cpus_read_lock(); 4099 intel_snb_check_microcode(); 4100 cpus_read_unlock(); 4101 } 4102 4103 static const struct { int id; char *name; } intel_arch_events_map[] __initconst = { 4104 { PERF_COUNT_HW_CPU_CYCLES, "cpu cycles" }, 4105 { PERF_COUNT_HW_INSTRUCTIONS, "instructions" }, 4106 { PERF_COUNT_HW_BUS_CYCLES, "bus cycles" }, 4107 { PERF_COUNT_HW_CACHE_REFERENCES, "cache references" }, 4108 { PERF_COUNT_HW_CACHE_MISSES, "cache misses" }, 4109 { PERF_COUNT_HW_BRANCH_INSTRUCTIONS, "branch instructions" }, 4110 { PERF_COUNT_HW_BRANCH_MISSES, "branch misses" }, 4111 }; 4112 4113 static __init void intel_arch_events_quirk(void) 4114 { 4115 int bit; 4116 4117 /* disable event that reported as not presend by cpuid */ 4118 for_each_set_bit(bit, x86_pmu.events_mask, ARRAY_SIZE(intel_arch_events_map)) { 4119 intel_perfmon_event_map[intel_arch_events_map[bit].id] = 0; 4120 pr_warn("CPUID marked event: \'%s\' unavailable\n", 4121 intel_arch_events_map[bit].name); 4122 } 4123 } 4124 4125 static __init void intel_nehalem_quirk(void) 4126 { 4127 union cpuid10_ebx ebx; 4128 4129 ebx.full = x86_pmu.events_maskl; 4130 if (ebx.split.no_branch_misses_retired) { 4131 /* 4132 * Erratum AAJ80 detected, we work it around by using 4133 * the BR_MISP_EXEC.ANY event. This will over-count 4134 * branch-misses, but it's still much better than the 4135 * architectural event which is often completely bogus: 4136 */ 4137 intel_perfmon_event_map[PERF_COUNT_HW_BRANCH_MISSES] = 0x7f89; 4138 ebx.split.no_branch_misses_retired = 0; 4139 x86_pmu.events_maskl = ebx.full; 4140 pr_info("CPU erratum AAJ80 worked around\n"); 4141 } 4142 } 4143 4144 static const struct x86_cpu_desc counter_freezing_ucodes[] = { 4145 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 2, 0x0000000e), 4146 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 9, 0x0000002e), 4147 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT, 10, 0x00000008), 4148 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_X, 1, 0x00000028), 4149 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 1, 0x00000028), 4150 INTEL_CPU_DESC(INTEL_FAM6_ATOM_GOLDMONT_PLUS, 8, 0x00000006), 4151 {} 4152 }; 4153 4154 static bool intel_counter_freezing_broken(void) 4155 { 4156 return !x86_cpu_has_min_microcode_rev(counter_freezing_ucodes); 4157 } 4158 4159 static __init void intel_counter_freezing_quirk(void) 4160 { 4161 /* Check if it's already disabled */ 4162 if (disable_counter_freezing) 4163 return; 4164 4165 /* 4166 * If the system starts with the wrong ucode, leave the 4167 * counter-freezing feature permanently disabled. 4168 */ 4169 if (intel_counter_freezing_broken()) { 4170 pr_info("PMU counter freezing disabled due to CPU errata," 4171 "please upgrade microcode\n"); 4172 x86_pmu.counter_freezing = false; 4173 x86_pmu.handle_irq = intel_pmu_handle_irq; 4174 } 4175 } 4176 4177 /* 4178 * enable software workaround for errata: 4179 * SNB: BJ122 4180 * IVB: BV98 4181 * HSW: HSD29 4182 * 4183 * Only needed when HT is enabled. However detecting 4184 * if HT is enabled is difficult (model specific). So instead, 4185 * we enable the workaround in the early boot, and verify if 4186 * it is needed in a later initcall phase once we have valid 4187 * topology information to check if HT is actually enabled 4188 */ 4189 static __init void intel_ht_bug(void) 4190 { 4191 x86_pmu.flags |= PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED; 4192 4193 x86_pmu.start_scheduling = intel_start_scheduling; 4194 x86_pmu.commit_scheduling = intel_commit_scheduling; 4195 x86_pmu.stop_scheduling = intel_stop_scheduling; 4196 } 4197 4198 EVENT_ATTR_STR(mem-loads, mem_ld_hsw, "event=0xcd,umask=0x1,ldlat=3"); 4199 EVENT_ATTR_STR(mem-stores, mem_st_hsw, "event=0xd0,umask=0x82") 4200 4201 /* Haswell special events */ 4202 EVENT_ATTR_STR(tx-start, tx_start, "event=0xc9,umask=0x1"); 4203 EVENT_ATTR_STR(tx-commit, tx_commit, "event=0xc9,umask=0x2"); 4204 EVENT_ATTR_STR(tx-abort, tx_abort, "event=0xc9,umask=0x4"); 4205 EVENT_ATTR_STR(tx-capacity, tx_capacity, "event=0x54,umask=0x2"); 4206 EVENT_ATTR_STR(tx-conflict, tx_conflict, "event=0x54,umask=0x1"); 4207 EVENT_ATTR_STR(el-start, el_start, "event=0xc8,umask=0x1"); 4208 EVENT_ATTR_STR(el-commit, el_commit, "event=0xc8,umask=0x2"); 4209 EVENT_ATTR_STR(el-abort, el_abort, "event=0xc8,umask=0x4"); 4210 EVENT_ATTR_STR(el-capacity, el_capacity, "event=0x54,umask=0x2"); 4211 EVENT_ATTR_STR(el-conflict, el_conflict, "event=0x54,umask=0x1"); 4212 EVENT_ATTR_STR(cycles-t, cycles_t, "event=0x3c,in_tx=1"); 4213 EVENT_ATTR_STR(cycles-ct, cycles_ct, "event=0x3c,in_tx=1,in_tx_cp=1"); 4214 4215 static struct attribute *hsw_events_attrs[] = { 4216 EVENT_PTR(td_slots_issued), 4217 EVENT_PTR(td_slots_retired), 4218 EVENT_PTR(td_fetch_bubbles), 4219 EVENT_PTR(td_total_slots), 4220 EVENT_PTR(td_total_slots_scale), 4221 EVENT_PTR(td_recovery_bubbles), 4222 EVENT_PTR(td_recovery_bubbles_scale), 4223 NULL 4224 }; 4225 4226 static struct attribute *hsw_mem_events_attrs[] = { 4227 EVENT_PTR(mem_ld_hsw), 4228 EVENT_PTR(mem_st_hsw), 4229 NULL, 4230 }; 4231 4232 static struct attribute *hsw_tsx_events_attrs[] = { 4233 EVENT_PTR(tx_start), 4234 EVENT_PTR(tx_commit), 4235 EVENT_PTR(tx_abort), 4236 EVENT_PTR(tx_capacity), 4237 EVENT_PTR(tx_conflict), 4238 EVENT_PTR(el_start), 4239 EVENT_PTR(el_commit), 4240 EVENT_PTR(el_abort), 4241 EVENT_PTR(el_capacity), 4242 EVENT_PTR(el_conflict), 4243 EVENT_PTR(cycles_t), 4244 EVENT_PTR(cycles_ct), 4245 NULL 4246 }; 4247 4248 EVENT_ATTR_STR(tx-capacity-read, tx_capacity_read, "event=0x54,umask=0x80"); 4249 EVENT_ATTR_STR(tx-capacity-write, tx_capacity_write, "event=0x54,umask=0x2"); 4250 EVENT_ATTR_STR(el-capacity-read, el_capacity_read, "event=0x54,umask=0x80"); 4251 EVENT_ATTR_STR(el-capacity-write, el_capacity_write, "event=0x54,umask=0x2"); 4252 4253 static struct attribute *icl_events_attrs[] = { 4254 EVENT_PTR(mem_ld_hsw), 4255 EVENT_PTR(mem_st_hsw), 4256 NULL, 4257 }; 4258 4259 static struct attribute *icl_tsx_events_attrs[] = { 4260 EVENT_PTR(tx_start), 4261 EVENT_PTR(tx_abort), 4262 EVENT_PTR(tx_commit), 4263 EVENT_PTR(tx_capacity_read), 4264 EVENT_PTR(tx_capacity_write), 4265 EVENT_PTR(tx_conflict), 4266 EVENT_PTR(el_start), 4267 EVENT_PTR(el_abort), 4268 EVENT_PTR(el_commit), 4269 EVENT_PTR(el_capacity_read), 4270 EVENT_PTR(el_capacity_write), 4271 EVENT_PTR(el_conflict), 4272 EVENT_PTR(cycles_t), 4273 EVENT_PTR(cycles_ct), 4274 NULL, 4275 }; 4276 4277 static __init struct attribute **get_icl_events_attrs(void) 4278 { 4279 return boot_cpu_has(X86_FEATURE_RTM) ? 4280 merge_attr(icl_events_attrs, icl_tsx_events_attrs) : 4281 icl_events_attrs; 4282 } 4283 4284 static ssize_t freeze_on_smi_show(struct device *cdev, 4285 struct device_attribute *attr, 4286 char *buf) 4287 { 4288 return sprintf(buf, "%lu\n", x86_pmu.attr_freeze_on_smi); 4289 } 4290 4291 static DEFINE_MUTEX(freeze_on_smi_mutex); 4292 4293 static ssize_t freeze_on_smi_store(struct device *cdev, 4294 struct device_attribute *attr, 4295 const char *buf, size_t count) 4296 { 4297 unsigned long val; 4298 ssize_t ret; 4299 4300 ret = kstrtoul(buf, 0, &val); 4301 if (ret) 4302 return ret; 4303 4304 if (val > 1) 4305 return -EINVAL; 4306 4307 mutex_lock(&freeze_on_smi_mutex); 4308 4309 if (x86_pmu.attr_freeze_on_smi == val) 4310 goto done; 4311 4312 x86_pmu.attr_freeze_on_smi = val; 4313 4314 get_online_cpus(); 4315 on_each_cpu(flip_smm_bit, &val, 1); 4316 put_online_cpus(); 4317 done: 4318 mutex_unlock(&freeze_on_smi_mutex); 4319 4320 return count; 4321 } 4322 4323 static void update_tfa_sched(void *ignored) 4324 { 4325 struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); 4326 4327 /* 4328 * check if PMC3 is used 4329 * and if so force schedule out for all event types all contexts 4330 */ 4331 if (test_bit(3, cpuc->active_mask)) 4332 perf_pmu_resched(x86_get_pmu()); 4333 } 4334 4335 static ssize_t show_sysctl_tfa(struct device *cdev, 4336 struct device_attribute *attr, 4337 char *buf) 4338 { 4339 return snprintf(buf, 40, "%d\n", allow_tsx_force_abort); 4340 } 4341 4342 static ssize_t set_sysctl_tfa(struct device *cdev, 4343 struct device_attribute *attr, 4344 const char *buf, size_t count) 4345 { 4346 bool val; 4347 ssize_t ret; 4348 4349 ret = kstrtobool(buf, &val); 4350 if (ret) 4351 return ret; 4352 4353 /* no change */ 4354 if (val == allow_tsx_force_abort) 4355 return count; 4356 4357 allow_tsx_force_abort = val; 4358 4359 get_online_cpus(); 4360 on_each_cpu(update_tfa_sched, NULL, 1); 4361 put_online_cpus(); 4362 4363 return count; 4364 } 4365 4366 4367 static DEVICE_ATTR_RW(freeze_on_smi); 4368 4369 static ssize_t branches_show(struct device *cdev, 4370 struct device_attribute *attr, 4371 char *buf) 4372 { 4373 return snprintf(buf, PAGE_SIZE, "%d\n", x86_pmu.lbr_nr); 4374 } 4375 4376 static DEVICE_ATTR_RO(branches); 4377 4378 static struct attribute *lbr_attrs[] = { 4379 &dev_attr_branches.attr, 4380 NULL 4381 }; 4382 4383 static char pmu_name_str[30]; 4384 4385 static ssize_t pmu_name_show(struct device *cdev, 4386 struct device_attribute *attr, 4387 char *buf) 4388 { 4389 return snprintf(buf, PAGE_SIZE, "%s\n", pmu_name_str); 4390 } 4391 4392 static DEVICE_ATTR_RO(pmu_name); 4393 4394 static struct attribute *intel_pmu_caps_attrs[] = { 4395 &dev_attr_pmu_name.attr, 4396 NULL 4397 }; 4398 4399 static DEVICE_ATTR(allow_tsx_force_abort, 0644, 4400 show_sysctl_tfa, 4401 set_sysctl_tfa); 4402 4403 static struct attribute *intel_pmu_attrs[] = { 4404 &dev_attr_freeze_on_smi.attr, 4405 NULL, /* &dev_attr_allow_tsx_force_abort.attr.attr */ 4406 NULL, 4407 }; 4408 4409 static __init struct attribute ** 4410 get_events_attrs(struct attribute **base, 4411 struct attribute **mem, 4412 struct attribute **tsx) 4413 { 4414 struct attribute **attrs = base; 4415 struct attribute **old; 4416 4417 if (mem && x86_pmu.pebs) 4418 attrs = merge_attr(attrs, mem); 4419 4420 if (tsx && boot_cpu_has(X86_FEATURE_RTM)) { 4421 old = attrs; 4422 attrs = merge_attr(attrs, tsx); 4423 if (old != base) 4424 kfree(old); 4425 } 4426 4427 return attrs; 4428 } 4429 4430 __init int intel_pmu_init(void) 4431 { 4432 struct attribute **extra_attr = NULL; 4433 struct attribute **mem_attr = NULL; 4434 struct attribute **tsx_attr = NULL; 4435 struct attribute **to_free = NULL; 4436 union cpuid10_edx edx; 4437 union cpuid10_eax eax; 4438 union cpuid10_ebx ebx; 4439 struct event_constraint *c; 4440 unsigned int unused; 4441 struct extra_reg *er; 4442 int version, i; 4443 char *name; 4444 4445 if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON)) { 4446 switch (boot_cpu_data.x86) { 4447 case 0x6: 4448 return p6_pmu_init(); 4449 case 0xb: 4450 return knc_pmu_init(); 4451 case 0xf: 4452 return p4_pmu_init(); 4453 } 4454 return -ENODEV; 4455 } 4456 4457 /* 4458 * Check whether the Architectural PerfMon supports 4459 * Branch Misses Retired hw_event or not. 4460 */ 4461 cpuid(10, &eax.full, &ebx.full, &unused, &edx.full); 4462 if (eax.split.mask_length < ARCH_PERFMON_EVENTS_COUNT) 4463 return -ENODEV; 4464 4465 version = eax.split.version_id; 4466 if (version < 2) 4467 x86_pmu = core_pmu; 4468 else 4469 x86_pmu = intel_pmu; 4470 4471 x86_pmu.version = version; 4472 x86_pmu.num_counters = eax.split.num_counters; 4473 x86_pmu.cntval_bits = eax.split.bit_width; 4474 x86_pmu.cntval_mask = (1ULL << eax.split.bit_width) - 1; 4475 4476 x86_pmu.events_maskl = ebx.full; 4477 x86_pmu.events_mask_len = eax.split.mask_length; 4478 4479 x86_pmu.max_pebs_events = min_t(unsigned, MAX_PEBS_EVENTS, x86_pmu.num_counters); 4480 4481 /* 4482 * Quirk: v2 perfmon does not report fixed-purpose events, so 4483 * assume at least 3 events, when not running in a hypervisor: 4484 */ 4485 if (version > 1) { 4486 int assume = 3 * !boot_cpu_has(X86_FEATURE_HYPERVISOR); 4487 4488 x86_pmu.num_counters_fixed = 4489 max((int)edx.split.num_counters_fixed, assume); 4490 } 4491 4492 if (version >= 4) 4493 x86_pmu.counter_freezing = !disable_counter_freezing; 4494 4495 if (boot_cpu_has(X86_FEATURE_PDCM)) { 4496 u64 capabilities; 4497 4498 rdmsrl(MSR_IA32_PERF_CAPABILITIES, capabilities); 4499 x86_pmu.intel_cap.capabilities = capabilities; 4500 } 4501 4502 intel_ds_init(); 4503 4504 x86_add_quirk(intel_arch_events_quirk); /* Install first, so it runs last */ 4505 4506 /* 4507 * Install the hw-cache-events table: 4508 */ 4509 switch (boot_cpu_data.x86_model) { 4510 case INTEL_FAM6_CORE_YONAH: 4511 pr_cont("Core events, "); 4512 name = "core"; 4513 break; 4514 4515 case INTEL_FAM6_CORE2_MEROM: 4516 x86_add_quirk(intel_clovertown_quirk); 4517 /* fall through */ 4518 4519 case INTEL_FAM6_CORE2_MEROM_L: 4520 case INTEL_FAM6_CORE2_PENRYN: 4521 case INTEL_FAM6_CORE2_DUNNINGTON: 4522 memcpy(hw_cache_event_ids, core2_hw_cache_event_ids, 4523 sizeof(hw_cache_event_ids)); 4524 4525 intel_pmu_lbr_init_core(); 4526 4527 x86_pmu.event_constraints = intel_core2_event_constraints; 4528 x86_pmu.pebs_constraints = intel_core2_pebs_event_constraints; 4529 pr_cont("Core2 events, "); 4530 name = "core2"; 4531 break; 4532 4533 case INTEL_FAM6_NEHALEM: 4534 case INTEL_FAM6_NEHALEM_EP: 4535 case INTEL_FAM6_NEHALEM_EX: 4536 memcpy(hw_cache_event_ids, nehalem_hw_cache_event_ids, 4537 sizeof(hw_cache_event_ids)); 4538 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 4539 sizeof(hw_cache_extra_regs)); 4540 4541 intel_pmu_lbr_init_nhm(); 4542 4543 x86_pmu.event_constraints = intel_nehalem_event_constraints; 4544 x86_pmu.pebs_constraints = intel_nehalem_pebs_event_constraints; 4545 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 4546 x86_pmu.extra_regs = intel_nehalem_extra_regs; 4547 4548 mem_attr = nhm_mem_events_attrs; 4549 4550 /* UOPS_ISSUED.STALLED_CYCLES */ 4551 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4552 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4553 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 4554 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4555 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 4556 4557 intel_pmu_pebs_data_source_nhm(); 4558 x86_add_quirk(intel_nehalem_quirk); 4559 x86_pmu.pebs_no_tlb = 1; 4560 extra_attr = nhm_format_attr; 4561 4562 pr_cont("Nehalem events, "); 4563 name = "nehalem"; 4564 break; 4565 4566 case INTEL_FAM6_ATOM_BONNELL: 4567 case INTEL_FAM6_ATOM_BONNELL_MID: 4568 case INTEL_FAM6_ATOM_SALTWELL: 4569 case INTEL_FAM6_ATOM_SALTWELL_MID: 4570 case INTEL_FAM6_ATOM_SALTWELL_TABLET: 4571 memcpy(hw_cache_event_ids, atom_hw_cache_event_ids, 4572 sizeof(hw_cache_event_ids)); 4573 4574 intel_pmu_lbr_init_atom(); 4575 4576 x86_pmu.event_constraints = intel_gen_event_constraints; 4577 x86_pmu.pebs_constraints = intel_atom_pebs_event_constraints; 4578 x86_pmu.pebs_aliases = intel_pebs_aliases_core2; 4579 pr_cont("Atom events, "); 4580 name = "bonnell"; 4581 break; 4582 4583 case INTEL_FAM6_ATOM_SILVERMONT: 4584 case INTEL_FAM6_ATOM_SILVERMONT_X: 4585 case INTEL_FAM6_ATOM_SILVERMONT_MID: 4586 case INTEL_FAM6_ATOM_AIRMONT: 4587 case INTEL_FAM6_ATOM_AIRMONT_MID: 4588 memcpy(hw_cache_event_ids, slm_hw_cache_event_ids, 4589 sizeof(hw_cache_event_ids)); 4590 memcpy(hw_cache_extra_regs, slm_hw_cache_extra_regs, 4591 sizeof(hw_cache_extra_regs)); 4592 4593 intel_pmu_lbr_init_slm(); 4594 4595 x86_pmu.event_constraints = intel_slm_event_constraints; 4596 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 4597 x86_pmu.extra_regs = intel_slm_extra_regs; 4598 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4599 x86_pmu.cpu_events = slm_events_attrs; 4600 extra_attr = slm_format_attr; 4601 pr_cont("Silvermont events, "); 4602 name = "silvermont"; 4603 break; 4604 4605 case INTEL_FAM6_ATOM_GOLDMONT: 4606 case INTEL_FAM6_ATOM_GOLDMONT_X: 4607 x86_add_quirk(intel_counter_freezing_quirk); 4608 memcpy(hw_cache_event_ids, glm_hw_cache_event_ids, 4609 sizeof(hw_cache_event_ids)); 4610 memcpy(hw_cache_extra_regs, glm_hw_cache_extra_regs, 4611 sizeof(hw_cache_extra_regs)); 4612 4613 intel_pmu_lbr_init_skl(); 4614 4615 x86_pmu.event_constraints = intel_slm_event_constraints; 4616 x86_pmu.pebs_constraints = intel_glm_pebs_event_constraints; 4617 x86_pmu.extra_regs = intel_glm_extra_regs; 4618 /* 4619 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4620 * for precise cycles. 4621 * :pp is identical to :ppp 4622 */ 4623 x86_pmu.pebs_aliases = NULL; 4624 x86_pmu.pebs_prec_dist = true; 4625 x86_pmu.lbr_pt_coexist = true; 4626 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4627 x86_pmu.cpu_events = glm_events_attrs; 4628 extra_attr = slm_format_attr; 4629 pr_cont("Goldmont events, "); 4630 name = "goldmont"; 4631 break; 4632 4633 case INTEL_FAM6_ATOM_GOLDMONT_PLUS: 4634 x86_add_quirk(intel_counter_freezing_quirk); 4635 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 4636 sizeof(hw_cache_event_ids)); 4637 memcpy(hw_cache_extra_regs, glp_hw_cache_extra_regs, 4638 sizeof(hw_cache_extra_regs)); 4639 4640 intel_pmu_lbr_init_skl(); 4641 4642 x86_pmu.event_constraints = intel_slm_event_constraints; 4643 x86_pmu.extra_regs = intel_glm_extra_regs; 4644 /* 4645 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4646 * for precise cycles. 4647 */ 4648 x86_pmu.pebs_aliases = NULL; 4649 x86_pmu.pebs_prec_dist = true; 4650 x86_pmu.lbr_pt_coexist = true; 4651 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4652 x86_pmu.flags |= PMU_FL_PEBS_ALL; 4653 x86_pmu.get_event_constraints = glp_get_event_constraints; 4654 x86_pmu.cpu_events = glm_events_attrs; 4655 /* Goldmont Plus has 4-wide pipeline */ 4656 event_attr_td_total_slots_scale_glm.event_str = "4"; 4657 extra_attr = slm_format_attr; 4658 pr_cont("Goldmont plus events, "); 4659 name = "goldmont_plus"; 4660 break; 4661 4662 case INTEL_FAM6_ATOM_TREMONT_X: 4663 x86_pmu.late_ack = true; 4664 memcpy(hw_cache_event_ids, glp_hw_cache_event_ids, 4665 sizeof(hw_cache_event_ids)); 4666 memcpy(hw_cache_extra_regs, tnt_hw_cache_extra_regs, 4667 sizeof(hw_cache_extra_regs)); 4668 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 4669 4670 intel_pmu_lbr_init_skl(); 4671 4672 x86_pmu.event_constraints = intel_slm_event_constraints; 4673 x86_pmu.extra_regs = intel_tnt_extra_regs; 4674 /* 4675 * It's recommended to use CPU_CLK_UNHALTED.CORE_P + NPEBS 4676 * for precise cycles. 4677 */ 4678 x86_pmu.pebs_aliases = NULL; 4679 x86_pmu.pebs_prec_dist = true; 4680 x86_pmu.lbr_pt_coexist = true; 4681 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4682 x86_pmu.get_event_constraints = tnt_get_event_constraints; 4683 extra_attr = slm_format_attr; 4684 pr_cont("Tremont events, "); 4685 name = "Tremont"; 4686 break; 4687 4688 case INTEL_FAM6_WESTMERE: 4689 case INTEL_FAM6_WESTMERE_EP: 4690 case INTEL_FAM6_WESTMERE_EX: 4691 memcpy(hw_cache_event_ids, westmere_hw_cache_event_ids, 4692 sizeof(hw_cache_event_ids)); 4693 memcpy(hw_cache_extra_regs, nehalem_hw_cache_extra_regs, 4694 sizeof(hw_cache_extra_regs)); 4695 4696 intel_pmu_lbr_init_nhm(); 4697 4698 x86_pmu.event_constraints = intel_westmere_event_constraints; 4699 x86_pmu.enable_all = intel_pmu_nhm_enable_all; 4700 x86_pmu.pebs_constraints = intel_westmere_pebs_event_constraints; 4701 x86_pmu.extra_regs = intel_westmere_extra_regs; 4702 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4703 4704 mem_attr = nhm_mem_events_attrs; 4705 4706 /* UOPS_ISSUED.STALLED_CYCLES */ 4707 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4708 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4709 /* UOPS_EXECUTED.CORE_ACTIVE_CYCLES,c=1,i=1 */ 4710 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4711 X86_CONFIG(.event=0xb1, .umask=0x3f, .inv=1, .cmask=1); 4712 4713 intel_pmu_pebs_data_source_nhm(); 4714 extra_attr = nhm_format_attr; 4715 pr_cont("Westmere events, "); 4716 name = "westmere"; 4717 break; 4718 4719 case INTEL_FAM6_SANDYBRIDGE: 4720 case INTEL_FAM6_SANDYBRIDGE_X: 4721 x86_add_quirk(intel_sandybridge_quirk); 4722 x86_add_quirk(intel_ht_bug); 4723 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 4724 sizeof(hw_cache_event_ids)); 4725 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 4726 sizeof(hw_cache_extra_regs)); 4727 4728 intel_pmu_lbr_init_snb(); 4729 4730 x86_pmu.event_constraints = intel_snb_event_constraints; 4731 x86_pmu.pebs_constraints = intel_snb_pebs_event_constraints; 4732 x86_pmu.pebs_aliases = intel_pebs_aliases_snb; 4733 if (boot_cpu_data.x86_model == INTEL_FAM6_SANDYBRIDGE_X) 4734 x86_pmu.extra_regs = intel_snbep_extra_regs; 4735 else 4736 x86_pmu.extra_regs = intel_snb_extra_regs; 4737 4738 4739 /* all extra regs are per-cpu when HT is on */ 4740 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4741 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4742 4743 x86_pmu.cpu_events = snb_events_attrs; 4744 mem_attr = snb_mem_events_attrs; 4745 4746 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 4747 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4748 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4749 /* UOPS_DISPATCHED.THREAD,c=1,i=1 to count stall cycles*/ 4750 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_BACKEND] = 4751 X86_CONFIG(.event=0xb1, .umask=0x01, .inv=1, .cmask=1); 4752 4753 extra_attr = nhm_format_attr; 4754 4755 pr_cont("SandyBridge events, "); 4756 name = "sandybridge"; 4757 break; 4758 4759 case INTEL_FAM6_IVYBRIDGE: 4760 case INTEL_FAM6_IVYBRIDGE_X: 4761 x86_add_quirk(intel_ht_bug); 4762 memcpy(hw_cache_event_ids, snb_hw_cache_event_ids, 4763 sizeof(hw_cache_event_ids)); 4764 /* dTLB-load-misses on IVB is different than SNB */ 4765 hw_cache_event_ids[C(DTLB)][C(OP_READ)][C(RESULT_MISS)] = 0x8108; /* DTLB_LOAD_MISSES.DEMAND_LD_MISS_CAUSES_A_WALK */ 4766 4767 memcpy(hw_cache_extra_regs, snb_hw_cache_extra_regs, 4768 sizeof(hw_cache_extra_regs)); 4769 4770 intel_pmu_lbr_init_snb(); 4771 4772 x86_pmu.event_constraints = intel_ivb_event_constraints; 4773 x86_pmu.pebs_constraints = intel_ivb_pebs_event_constraints; 4774 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4775 x86_pmu.pebs_prec_dist = true; 4776 if (boot_cpu_data.x86_model == INTEL_FAM6_IVYBRIDGE_X) 4777 x86_pmu.extra_regs = intel_snbep_extra_regs; 4778 else 4779 x86_pmu.extra_regs = intel_snb_extra_regs; 4780 /* all extra regs are per-cpu when HT is on */ 4781 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4782 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4783 4784 x86_pmu.cpu_events = snb_events_attrs; 4785 mem_attr = snb_mem_events_attrs; 4786 4787 /* UOPS_ISSUED.ANY,c=1,i=1 to count stall cycles */ 4788 intel_perfmon_event_map[PERF_COUNT_HW_STALLED_CYCLES_FRONTEND] = 4789 X86_CONFIG(.event=0x0e, .umask=0x01, .inv=1, .cmask=1); 4790 4791 extra_attr = nhm_format_attr; 4792 4793 pr_cont("IvyBridge events, "); 4794 name = "ivybridge"; 4795 break; 4796 4797 4798 case INTEL_FAM6_HASWELL_CORE: 4799 case INTEL_FAM6_HASWELL_X: 4800 case INTEL_FAM6_HASWELL_ULT: 4801 case INTEL_FAM6_HASWELL_GT3E: 4802 x86_add_quirk(intel_ht_bug); 4803 x86_add_quirk(intel_pebs_isolation_quirk); 4804 x86_pmu.late_ack = true; 4805 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4806 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4807 4808 intel_pmu_lbr_init_hsw(); 4809 4810 x86_pmu.event_constraints = intel_hsw_event_constraints; 4811 x86_pmu.pebs_constraints = intel_hsw_pebs_event_constraints; 4812 x86_pmu.extra_regs = intel_snbep_extra_regs; 4813 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4814 x86_pmu.pebs_prec_dist = true; 4815 /* all extra regs are per-cpu when HT is on */ 4816 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4817 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4818 4819 x86_pmu.hw_config = hsw_hw_config; 4820 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4821 x86_pmu.cpu_events = hsw_events_attrs; 4822 x86_pmu.lbr_double_abort = true; 4823 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4824 hsw_format_attr : nhm_format_attr; 4825 mem_attr = hsw_mem_events_attrs; 4826 tsx_attr = hsw_tsx_events_attrs; 4827 pr_cont("Haswell events, "); 4828 name = "haswell"; 4829 break; 4830 4831 case INTEL_FAM6_BROADWELL_CORE: 4832 case INTEL_FAM6_BROADWELL_XEON_D: 4833 case INTEL_FAM6_BROADWELL_GT3E: 4834 case INTEL_FAM6_BROADWELL_X: 4835 x86_add_quirk(intel_pebs_isolation_quirk); 4836 x86_pmu.late_ack = true; 4837 memcpy(hw_cache_event_ids, hsw_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4838 memcpy(hw_cache_extra_regs, hsw_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4839 4840 /* L3_MISS_LOCAL_DRAM is BIT(26) in Broadwell */ 4841 hw_cache_extra_regs[C(LL)][C(OP_READ)][C(RESULT_MISS)] = HSW_DEMAND_READ | 4842 BDW_L3_MISS|HSW_SNOOP_DRAM; 4843 hw_cache_extra_regs[C(LL)][C(OP_WRITE)][C(RESULT_MISS)] = HSW_DEMAND_WRITE|BDW_L3_MISS| 4844 HSW_SNOOP_DRAM; 4845 hw_cache_extra_regs[C(NODE)][C(OP_READ)][C(RESULT_ACCESS)] = HSW_DEMAND_READ| 4846 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 4847 hw_cache_extra_regs[C(NODE)][C(OP_WRITE)][C(RESULT_ACCESS)] = HSW_DEMAND_WRITE| 4848 BDW_L3_MISS_LOCAL|HSW_SNOOP_DRAM; 4849 4850 intel_pmu_lbr_init_hsw(); 4851 4852 x86_pmu.event_constraints = intel_bdw_event_constraints; 4853 x86_pmu.pebs_constraints = intel_bdw_pebs_event_constraints; 4854 x86_pmu.extra_regs = intel_snbep_extra_regs; 4855 x86_pmu.pebs_aliases = intel_pebs_aliases_ivb; 4856 x86_pmu.pebs_prec_dist = true; 4857 /* all extra regs are per-cpu when HT is on */ 4858 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4859 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4860 4861 x86_pmu.hw_config = hsw_hw_config; 4862 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4863 x86_pmu.cpu_events = hsw_events_attrs; 4864 x86_pmu.limit_period = bdw_limit_period; 4865 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4866 hsw_format_attr : nhm_format_attr; 4867 mem_attr = hsw_mem_events_attrs; 4868 tsx_attr = hsw_tsx_events_attrs; 4869 pr_cont("Broadwell events, "); 4870 name = "broadwell"; 4871 break; 4872 4873 case INTEL_FAM6_XEON_PHI_KNL: 4874 case INTEL_FAM6_XEON_PHI_KNM: 4875 memcpy(hw_cache_event_ids, 4876 slm_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4877 memcpy(hw_cache_extra_regs, 4878 knl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4879 intel_pmu_lbr_init_knl(); 4880 4881 x86_pmu.event_constraints = intel_slm_event_constraints; 4882 x86_pmu.pebs_constraints = intel_slm_pebs_event_constraints; 4883 x86_pmu.extra_regs = intel_knl_extra_regs; 4884 4885 /* all extra regs are per-cpu when HT is on */ 4886 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4887 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4888 extra_attr = slm_format_attr; 4889 pr_cont("Knights Landing/Mill events, "); 4890 name = "knights-landing"; 4891 break; 4892 4893 case INTEL_FAM6_SKYLAKE_MOBILE: 4894 case INTEL_FAM6_SKYLAKE_DESKTOP: 4895 case INTEL_FAM6_SKYLAKE_X: 4896 case INTEL_FAM6_KABYLAKE_MOBILE: 4897 case INTEL_FAM6_KABYLAKE_DESKTOP: 4898 x86_add_quirk(intel_pebs_isolation_quirk); 4899 x86_pmu.late_ack = true; 4900 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4901 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4902 intel_pmu_lbr_init_skl(); 4903 4904 /* INT_MISC.RECOVERY_CYCLES has umask 1 in Skylake */ 4905 event_attr_td_recovery_bubbles.event_str_noht = 4906 "event=0xd,umask=0x1,cmask=1"; 4907 event_attr_td_recovery_bubbles.event_str_ht = 4908 "event=0xd,umask=0x1,cmask=1,any=1"; 4909 4910 x86_pmu.event_constraints = intel_skl_event_constraints; 4911 x86_pmu.pebs_constraints = intel_skl_pebs_event_constraints; 4912 x86_pmu.extra_regs = intel_skl_extra_regs; 4913 x86_pmu.pebs_aliases = intel_pebs_aliases_skl; 4914 x86_pmu.pebs_prec_dist = true; 4915 /* all extra regs are per-cpu when HT is on */ 4916 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4917 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4918 4919 x86_pmu.hw_config = hsw_hw_config; 4920 x86_pmu.get_event_constraints = hsw_get_event_constraints; 4921 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4922 hsw_format_attr : nhm_format_attr; 4923 extra_attr = merge_attr(extra_attr, skl_format_attr); 4924 to_free = extra_attr; 4925 x86_pmu.cpu_events = hsw_events_attrs; 4926 mem_attr = hsw_mem_events_attrs; 4927 tsx_attr = hsw_tsx_events_attrs; 4928 intel_pmu_pebs_data_source_skl( 4929 boot_cpu_data.x86_model == INTEL_FAM6_SKYLAKE_X); 4930 4931 if (boot_cpu_has(X86_FEATURE_TSX_FORCE_ABORT)) { 4932 x86_pmu.flags |= PMU_FL_TFA; 4933 x86_pmu.get_event_constraints = tfa_get_event_constraints; 4934 x86_pmu.enable_all = intel_tfa_pmu_enable_all; 4935 x86_pmu.commit_scheduling = intel_tfa_commit_scheduling; 4936 intel_pmu_attrs[1] = &dev_attr_allow_tsx_force_abort.attr; 4937 } 4938 4939 pr_cont("Skylake events, "); 4940 name = "skylake"; 4941 break; 4942 4943 case INTEL_FAM6_ICELAKE_MOBILE: 4944 x86_pmu.late_ack = true; 4945 memcpy(hw_cache_event_ids, skl_hw_cache_event_ids, sizeof(hw_cache_event_ids)); 4946 memcpy(hw_cache_extra_regs, skl_hw_cache_extra_regs, sizeof(hw_cache_extra_regs)); 4947 hw_cache_event_ids[C(ITLB)][C(OP_READ)][C(RESULT_ACCESS)] = -1; 4948 intel_pmu_lbr_init_skl(); 4949 4950 x86_pmu.event_constraints = intel_icl_event_constraints; 4951 x86_pmu.pebs_constraints = intel_icl_pebs_event_constraints; 4952 x86_pmu.extra_regs = intel_icl_extra_regs; 4953 x86_pmu.pebs_aliases = NULL; 4954 x86_pmu.pebs_prec_dist = true; 4955 x86_pmu.flags |= PMU_FL_HAS_RSP_1; 4956 x86_pmu.flags |= PMU_FL_NO_HT_SHARING; 4957 4958 x86_pmu.hw_config = hsw_hw_config; 4959 x86_pmu.get_event_constraints = icl_get_event_constraints; 4960 extra_attr = boot_cpu_has(X86_FEATURE_RTM) ? 4961 hsw_format_attr : nhm_format_attr; 4962 extra_attr = merge_attr(extra_attr, skl_format_attr); 4963 x86_pmu.cpu_events = get_icl_events_attrs(); 4964 x86_pmu.rtm_abort_event = X86_CONFIG(.event=0xca, .umask=0x02); 4965 x86_pmu.lbr_pt_coexist = true; 4966 intel_pmu_pebs_data_source_skl(false); 4967 pr_cont("Icelake events, "); 4968 name = "icelake"; 4969 break; 4970 4971 default: 4972 switch (x86_pmu.version) { 4973 case 1: 4974 x86_pmu.event_constraints = intel_v1_event_constraints; 4975 pr_cont("generic architected perfmon v1, "); 4976 name = "generic_arch_v1"; 4977 break; 4978 default: 4979 /* 4980 * default constraints for v2 and up 4981 */ 4982 x86_pmu.event_constraints = intel_gen_event_constraints; 4983 pr_cont("generic architected perfmon, "); 4984 name = "generic_arch_v2+"; 4985 break; 4986 } 4987 } 4988 4989 snprintf(pmu_name_str, sizeof(pmu_name_str), "%s", name); 4990 4991 if (version >= 2 && extra_attr) { 4992 x86_pmu.format_attrs = merge_attr(intel_arch3_formats_attr, 4993 extra_attr); 4994 WARN_ON(!x86_pmu.format_attrs); 4995 } 4996 4997 x86_pmu.cpu_events = get_events_attrs(x86_pmu.cpu_events, 4998 mem_attr, tsx_attr); 4999 5000 if (x86_pmu.num_counters > INTEL_PMC_MAX_GENERIC) { 5001 WARN(1, KERN_ERR "hw perf events %d > max(%d), clipping!", 5002 x86_pmu.num_counters, INTEL_PMC_MAX_GENERIC); 5003 x86_pmu.num_counters = INTEL_PMC_MAX_GENERIC; 5004 } 5005 x86_pmu.intel_ctrl = (1ULL << x86_pmu.num_counters) - 1; 5006 5007 if (x86_pmu.num_counters_fixed > INTEL_PMC_MAX_FIXED) { 5008 WARN(1, KERN_ERR "hw perf events fixed %d > max(%d), clipping!", 5009 x86_pmu.num_counters_fixed, INTEL_PMC_MAX_FIXED); 5010 x86_pmu.num_counters_fixed = INTEL_PMC_MAX_FIXED; 5011 } 5012 5013 x86_pmu.intel_ctrl |= 5014 ((1LL << x86_pmu.num_counters_fixed)-1) << INTEL_PMC_IDX_FIXED; 5015 5016 if (x86_pmu.event_constraints) { 5017 /* 5018 * event on fixed counter2 (REF_CYCLES) only works on this 5019 * counter, so do not extend mask to generic counters 5020 */ 5021 for_each_event_constraint(c, x86_pmu.event_constraints) { 5022 if (c->cmask == FIXED_EVENT_FLAGS 5023 && c->idxmsk64 != INTEL_PMC_MSK_FIXED_REF_CYCLES) { 5024 c->idxmsk64 |= (1ULL << x86_pmu.num_counters) - 1; 5025 } 5026 c->idxmsk64 &= 5027 ~(~0ULL << (INTEL_PMC_IDX_FIXED + x86_pmu.num_counters_fixed)); 5028 c->weight = hweight64(c->idxmsk64); 5029 } 5030 } 5031 5032 /* 5033 * Access LBR MSR may cause #GP under certain circumstances. 5034 * E.g. KVM doesn't support LBR MSR 5035 * Check all LBT MSR here. 5036 * Disable LBR access if any LBR MSRs can not be accessed. 5037 */ 5038 if (x86_pmu.lbr_nr && !check_msr(x86_pmu.lbr_tos, 0x3UL)) 5039 x86_pmu.lbr_nr = 0; 5040 for (i = 0; i < x86_pmu.lbr_nr; i++) { 5041 if (!(check_msr(x86_pmu.lbr_from + i, 0xffffUL) && 5042 check_msr(x86_pmu.lbr_to + i, 0xffffUL))) 5043 x86_pmu.lbr_nr = 0; 5044 } 5045 5046 x86_pmu.caps_attrs = intel_pmu_caps_attrs; 5047 5048 if (x86_pmu.lbr_nr) { 5049 x86_pmu.caps_attrs = merge_attr(x86_pmu.caps_attrs, lbr_attrs); 5050 pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr); 5051 } 5052 5053 /* 5054 * Access extra MSR may cause #GP under certain circumstances. 5055 * E.g. KVM doesn't support offcore event 5056 * Check all extra_regs here. 5057 */ 5058 if (x86_pmu.extra_regs) { 5059 for (er = x86_pmu.extra_regs; er->msr; er++) { 5060 er->extra_msr_access = check_msr(er->msr, 0x11UL); 5061 /* Disable LBR select mapping */ 5062 if ((er->idx == EXTRA_REG_LBR) && !er->extra_msr_access) 5063 x86_pmu.lbr_sel_map = NULL; 5064 } 5065 } 5066 5067 /* Support full width counters using alternative MSR range */ 5068 if (x86_pmu.intel_cap.full_width_write) { 5069 x86_pmu.max_period = x86_pmu.cntval_mask >> 1; 5070 x86_pmu.perfctr = MSR_IA32_PMC0; 5071 pr_cont("full-width counters, "); 5072 } 5073 5074 /* 5075 * For arch perfmon 4 use counter freezing to avoid 5076 * several MSR accesses in the PMI. 5077 */ 5078 if (x86_pmu.counter_freezing) 5079 x86_pmu.handle_irq = intel_pmu_handle_irq_v4; 5080 5081 kfree(to_free); 5082 return 0; 5083 } 5084 5085 /* 5086 * HT bug: phase 2 init 5087 * Called once we have valid topology information to check 5088 * whether or not HT is enabled 5089 * If HT is off, then we disable the workaround 5090 */ 5091 static __init int fixup_ht_bug(void) 5092 { 5093 int c; 5094 /* 5095 * problem not present on this CPU model, nothing to do 5096 */ 5097 if (!(x86_pmu.flags & PMU_FL_EXCL_ENABLED)) 5098 return 0; 5099 5100 if (topology_max_smt_threads() > 1) { 5101 pr_info("PMU erratum BJ122, BV98, HSD29 worked around, HT is on\n"); 5102 return 0; 5103 } 5104 5105 cpus_read_lock(); 5106 5107 hardlockup_detector_perf_stop(); 5108 5109 x86_pmu.flags &= ~(PMU_FL_EXCL_CNTRS | PMU_FL_EXCL_ENABLED); 5110 5111 x86_pmu.start_scheduling = NULL; 5112 x86_pmu.commit_scheduling = NULL; 5113 x86_pmu.stop_scheduling = NULL; 5114 5115 hardlockup_detector_perf_restart(); 5116 5117 for_each_online_cpu(c) 5118 free_excl_cntrs(&per_cpu(cpu_hw_events, c)); 5119 5120 cpus_read_unlock(); 5121 pr_info("PMU erratum BJ122, BV98, HSD29 workaround disabled, HT off\n"); 5122 return 0; 5123 } 5124 subsys_initcall(fixup_ht_bug) 5125