1 /* 2 * Performance events x86 architecture header 3 * 4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de> 5 * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar 6 * Copyright (C) 2009 Jaswinder Singh Rajput 7 * Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter 8 * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra 9 * Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com> 10 * Copyright (C) 2009 Google, Inc., Stephane Eranian 11 * 12 * For licencing details see kernel-base/COPYING 13 */ 14 15 #include <linux/perf_event.h> 16 17 #include <asm/intel_ds.h> 18 19 /* To enable MSR tracing please use the generic trace points. */ 20 21 /* 22 * | NHM/WSM | SNB | 23 * register ------------------------------- 24 * | HT | no HT | HT | no HT | 25 *----------------------------------------- 26 * offcore | core | core | cpu | core | 27 * lbr_sel | core | core | cpu | core | 28 * ld_lat | cpu | core | cpu | core | 29 *----------------------------------------- 30 * 31 * Given that there is a small number of shared regs, 32 * we can pre-allocate their slot in the per-cpu 33 * per-core reg tables. 34 */ 35 enum extra_reg_type { 36 EXTRA_REG_NONE = -1, /* not used */ 37 38 EXTRA_REG_RSP_0 = 0, /* offcore_response_0 */ 39 EXTRA_REG_RSP_1 = 1, /* offcore_response_1 */ 40 EXTRA_REG_LBR = 2, /* lbr_select */ 41 EXTRA_REG_LDLAT = 3, /* ld_lat_threshold */ 42 EXTRA_REG_FE = 4, /* fe_* */ 43 44 EXTRA_REG_MAX /* number of entries needed */ 45 }; 46 47 struct event_constraint { 48 union { 49 unsigned long idxmsk[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; 50 u64 idxmsk64; 51 }; 52 u64 code; 53 u64 cmask; 54 int weight; 55 int overlap; 56 int flags; 57 }; 58 /* 59 * struct hw_perf_event.flags flags 60 */ 61 #define PERF_X86_EVENT_PEBS_LDLAT 0x0001 /* ld+ldlat data address sampling */ 62 #define PERF_X86_EVENT_PEBS_ST 0x0002 /* st data address sampling */ 63 #define PERF_X86_EVENT_PEBS_ST_HSW 0x0004 /* haswell style datala, store */ 64 #define PERF_X86_EVENT_COMMITTED 0x0008 /* event passed commit_txn */ 65 #define PERF_X86_EVENT_PEBS_LD_HSW 0x0010 /* haswell style datala, load */ 66 #define PERF_X86_EVENT_PEBS_NA_HSW 0x0020 /* haswell style datala, unknown */ 67 #define PERF_X86_EVENT_EXCL 0x0040 /* HT exclusivity on counter */ 68 #define PERF_X86_EVENT_DYNAMIC 0x0080 /* dynamic alloc'd constraint */ 69 #define PERF_X86_EVENT_RDPMC_ALLOWED 0x0100 /* grant rdpmc permission */ 70 #define PERF_X86_EVENT_EXCL_ACCT 0x0200 /* accounted EXCL event */ 71 #define PERF_X86_EVENT_AUTO_RELOAD 0x0400 /* use PEBS auto-reload */ 72 #define PERF_X86_EVENT_LARGE_PEBS 0x0800 /* use large PEBS */ 73 74 75 struct amd_nb { 76 int nb_id; /* NorthBridge id */ 77 int refcnt; /* reference count */ 78 struct perf_event *owners[X86_PMC_IDX_MAX]; 79 struct event_constraint event_constraints[X86_PMC_IDX_MAX]; 80 }; 81 82 #define PEBS_COUNTER_MASK ((1ULL << MAX_PEBS_EVENTS) - 1) 83 84 /* 85 * Flags PEBS can handle without an PMI. 86 * 87 * TID can only be handled by flushing at context switch. 88 * REGS_USER can be handled for events limited to ring 3. 89 * 90 */ 91 #define LARGE_PEBS_FLAGS \ 92 (PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_ADDR | \ 93 PERF_SAMPLE_ID | PERF_SAMPLE_CPU | PERF_SAMPLE_STREAM_ID | \ 94 PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_IDENTIFIER | \ 95 PERF_SAMPLE_TRANSACTION | PERF_SAMPLE_PHYS_ADDR | \ 96 PERF_SAMPLE_REGS_INTR | PERF_SAMPLE_REGS_USER | \ 97 PERF_SAMPLE_PERIOD) 98 99 #define PEBS_GP_REGS \ 100 ((1ULL << PERF_REG_X86_AX) | \ 101 (1ULL << PERF_REG_X86_BX) | \ 102 (1ULL << PERF_REG_X86_CX) | \ 103 (1ULL << PERF_REG_X86_DX) | \ 104 (1ULL << PERF_REG_X86_DI) | \ 105 (1ULL << PERF_REG_X86_SI) | \ 106 (1ULL << PERF_REG_X86_SP) | \ 107 (1ULL << PERF_REG_X86_BP) | \ 108 (1ULL << PERF_REG_X86_IP) | \ 109 (1ULL << PERF_REG_X86_FLAGS) | \ 110 (1ULL << PERF_REG_X86_R8) | \ 111 (1ULL << PERF_REG_X86_R9) | \ 112 (1ULL << PERF_REG_X86_R10) | \ 113 (1ULL << PERF_REG_X86_R11) | \ 114 (1ULL << PERF_REG_X86_R12) | \ 115 (1ULL << PERF_REG_X86_R13) | \ 116 (1ULL << PERF_REG_X86_R14) | \ 117 (1ULL << PERF_REG_X86_R15)) 118 119 /* 120 * Per register state. 121 */ 122 struct er_account { 123 raw_spinlock_t lock; /* per-core: protect structure */ 124 u64 config; /* extra MSR config */ 125 u64 reg; /* extra MSR number */ 126 atomic_t ref; /* reference count */ 127 }; 128 129 /* 130 * Per core/cpu state 131 * 132 * Used to coordinate shared registers between HT threads or 133 * among events on a single PMU. 134 */ 135 struct intel_shared_regs { 136 struct er_account regs[EXTRA_REG_MAX]; 137 int refcnt; /* per-core: #HT threads */ 138 unsigned core_id; /* per-core: core id */ 139 }; 140 141 enum intel_excl_state_type { 142 INTEL_EXCL_UNUSED = 0, /* counter is unused */ 143 INTEL_EXCL_SHARED = 1, /* counter can be used by both threads */ 144 INTEL_EXCL_EXCLUSIVE = 2, /* counter can be used by one thread only */ 145 }; 146 147 struct intel_excl_states { 148 enum intel_excl_state_type state[X86_PMC_IDX_MAX]; 149 bool sched_started; /* true if scheduling has started */ 150 }; 151 152 struct intel_excl_cntrs { 153 raw_spinlock_t lock; 154 155 struct intel_excl_states states[2]; 156 157 union { 158 u16 has_exclusive[2]; 159 u32 exclusive_present; 160 }; 161 162 int refcnt; /* per-core: #HT threads */ 163 unsigned core_id; /* per-core: core id */ 164 }; 165 166 struct x86_perf_task_context; 167 #define MAX_LBR_ENTRIES 32 168 169 enum { 170 X86_PERF_KFREE_SHARED = 0, 171 X86_PERF_KFREE_EXCL = 1, 172 X86_PERF_KFREE_MAX 173 }; 174 175 struct cpu_hw_events { 176 /* 177 * Generic x86 PMC bits 178 */ 179 struct perf_event *events[X86_PMC_IDX_MAX]; /* in counter order */ 180 unsigned long active_mask[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; 181 unsigned long running[BITS_TO_LONGS(X86_PMC_IDX_MAX)]; 182 int enabled; 183 184 int n_events; /* the # of events in the below arrays */ 185 int n_added; /* the # last events in the below arrays; 186 they've never been enabled yet */ 187 int n_txn; /* the # last events in the below arrays; 188 added in the current transaction */ 189 int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */ 190 u64 tags[X86_PMC_IDX_MAX]; 191 192 struct perf_event *event_list[X86_PMC_IDX_MAX]; /* in enabled order */ 193 struct event_constraint *event_constraint[X86_PMC_IDX_MAX]; 194 195 int n_excl; /* the number of exclusive events */ 196 197 unsigned int txn_flags; 198 int is_fake; 199 200 /* 201 * Intel DebugStore bits 202 */ 203 struct debug_store *ds; 204 void *ds_pebs_vaddr; 205 void *ds_bts_vaddr; 206 u64 pebs_enabled; 207 int n_pebs; 208 int n_large_pebs; 209 210 /* 211 * Intel LBR bits 212 */ 213 int lbr_users; 214 struct perf_branch_stack lbr_stack; 215 struct perf_branch_entry lbr_entries[MAX_LBR_ENTRIES]; 216 struct er_account *lbr_sel; 217 u64 br_sel; 218 struct x86_perf_task_context *last_task_ctx; 219 int last_log_id; 220 221 /* 222 * Intel host/guest exclude bits 223 */ 224 u64 intel_ctrl_guest_mask; 225 u64 intel_ctrl_host_mask; 226 struct perf_guest_switch_msr guest_switch_msrs[X86_PMC_IDX_MAX]; 227 228 /* 229 * Intel checkpoint mask 230 */ 231 u64 intel_cp_status; 232 233 /* 234 * manage shared (per-core, per-cpu) registers 235 * used on Intel NHM/WSM/SNB 236 */ 237 struct intel_shared_regs *shared_regs; 238 /* 239 * manage exclusive counter access between hyperthread 240 */ 241 struct event_constraint *constraint_list; /* in enable order */ 242 struct intel_excl_cntrs *excl_cntrs; 243 int excl_thread_id; /* 0 or 1 */ 244 245 /* 246 * SKL TSX_FORCE_ABORT shadow 247 */ 248 u64 tfa_shadow; 249 250 /* 251 * AMD specific bits 252 */ 253 struct amd_nb *amd_nb; 254 /* Inverted mask of bits to clear in the perf_ctr ctrl registers */ 255 u64 perf_ctr_virt_mask; 256 257 void *kfree_on_online[X86_PERF_KFREE_MAX]; 258 }; 259 260 #define __EVENT_CONSTRAINT(c, n, m, w, o, f) {\ 261 { .idxmsk64 = (n) }, \ 262 .code = (c), \ 263 .cmask = (m), \ 264 .weight = (w), \ 265 .overlap = (o), \ 266 .flags = f, \ 267 } 268 269 #define EVENT_CONSTRAINT(c, n, m) \ 270 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 0, 0) 271 272 #define INTEL_EXCLEVT_CONSTRAINT(c, n) \ 273 __EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT, HWEIGHT(n),\ 274 0, PERF_X86_EVENT_EXCL) 275 276 /* 277 * The overlap flag marks event constraints with overlapping counter 278 * masks. This is the case if the counter mask of such an event is not 279 * a subset of any other counter mask of a constraint with an equal or 280 * higher weight, e.g.: 281 * 282 * c_overlaps = EVENT_CONSTRAINT_OVERLAP(0, 0x09, 0); 283 * c_another1 = EVENT_CONSTRAINT(0, 0x07, 0); 284 * c_another2 = EVENT_CONSTRAINT(0, 0x38, 0); 285 * 286 * The event scheduler may not select the correct counter in the first 287 * cycle because it needs to know which subsequent events will be 288 * scheduled. It may fail to schedule the events then. So we set the 289 * overlap flag for such constraints to give the scheduler a hint which 290 * events to select for counter rescheduling. 291 * 292 * Care must be taken as the rescheduling algorithm is O(n!) which 293 * will increase scheduling cycles for an over-committed system 294 * dramatically. The number of such EVENT_CONSTRAINT_OVERLAP() macros 295 * and its counter masks must be kept at a minimum. 296 */ 297 #define EVENT_CONSTRAINT_OVERLAP(c, n, m) \ 298 __EVENT_CONSTRAINT(c, n, m, HWEIGHT(n), 1, 0) 299 300 /* 301 * Constraint on the Event code. 302 */ 303 #define INTEL_EVENT_CONSTRAINT(c, n) \ 304 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT) 305 306 /* 307 * Constraint on the Event code + UMask + fixed-mask 308 * 309 * filter mask to validate fixed counter events. 310 * the following filters disqualify for fixed counters: 311 * - inv 312 * - edge 313 * - cnt-mask 314 * - in_tx 315 * - in_tx_checkpointed 316 * The other filters are supported by fixed counters. 317 * The any-thread option is supported starting with v3. 318 */ 319 #define FIXED_EVENT_FLAGS (X86_RAW_EVENT_MASK|HSW_IN_TX|HSW_IN_TX_CHECKPOINTED) 320 #define FIXED_EVENT_CONSTRAINT(c, n) \ 321 EVENT_CONSTRAINT(c, (1ULL << (32+n)), FIXED_EVENT_FLAGS) 322 323 /* 324 * Constraint on the Event code + UMask 325 */ 326 #define INTEL_UEVENT_CONSTRAINT(c, n) \ 327 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK) 328 329 /* Constraint on specific umask bit only + event */ 330 #define INTEL_UBIT_EVENT_CONSTRAINT(c, n) \ 331 EVENT_CONSTRAINT(c, n, ARCH_PERFMON_EVENTSEL_EVENT|(c)) 332 333 /* Like UEVENT_CONSTRAINT, but match flags too */ 334 #define INTEL_FLAGS_UEVENT_CONSTRAINT(c, n) \ 335 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS) 336 337 #define INTEL_EXCLUEVT_CONSTRAINT(c, n) \ 338 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK, \ 339 HWEIGHT(n), 0, PERF_X86_EVENT_EXCL) 340 341 #define INTEL_PLD_CONSTRAINT(c, n) \ 342 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 343 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LDLAT) 344 345 #define INTEL_PST_CONSTRAINT(c, n) \ 346 __EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 347 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST) 348 349 /* Event constraint, but match on all event flags too. */ 350 #define INTEL_FLAGS_EVENT_CONSTRAINT(c, n) \ 351 EVENT_CONSTRAINT(c, n, INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS) 352 353 /* Check only flags, but allow all event/umask */ 354 #define INTEL_ALL_EVENT_CONSTRAINT(code, n) \ 355 EVENT_CONSTRAINT(code, n, X86_ALL_EVENT_FLAGS) 356 357 /* Check flags and event code, and set the HSW store flag */ 358 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_ST(code, n) \ 359 __EVENT_CONSTRAINT(code, n, \ 360 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \ 361 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW) 362 363 /* Check flags and event code, and set the HSW load flag */ 364 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_LD(code, n) \ 365 __EVENT_CONSTRAINT(code, n, \ 366 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \ 367 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LD_HSW) 368 369 #define INTEL_FLAGS_EVENT_CONSTRAINT_DATALA_XLD(code, n) \ 370 __EVENT_CONSTRAINT(code, n, \ 371 ARCH_PERFMON_EVENTSEL_EVENT|X86_ALL_EVENT_FLAGS, \ 372 HWEIGHT(n), 0, \ 373 PERF_X86_EVENT_PEBS_LD_HSW|PERF_X86_EVENT_EXCL) 374 375 /* Check flags and event code/umask, and set the HSW store flag */ 376 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_ST(code, n) \ 377 __EVENT_CONSTRAINT(code, n, \ 378 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 379 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_ST_HSW) 380 381 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XST(code, n) \ 382 __EVENT_CONSTRAINT(code, n, \ 383 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 384 HWEIGHT(n), 0, \ 385 PERF_X86_EVENT_PEBS_ST_HSW|PERF_X86_EVENT_EXCL) 386 387 /* Check flags and event code/umask, and set the HSW load flag */ 388 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_LD(code, n) \ 389 __EVENT_CONSTRAINT(code, n, \ 390 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 391 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_LD_HSW) 392 393 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_XLD(code, n) \ 394 __EVENT_CONSTRAINT(code, n, \ 395 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 396 HWEIGHT(n), 0, \ 397 PERF_X86_EVENT_PEBS_LD_HSW|PERF_X86_EVENT_EXCL) 398 399 /* Check flags and event code/umask, and set the HSW N/A flag */ 400 #define INTEL_FLAGS_UEVENT_CONSTRAINT_DATALA_NA(code, n) \ 401 __EVENT_CONSTRAINT(code, n, \ 402 INTEL_ARCH_EVENT_MASK|X86_ALL_EVENT_FLAGS, \ 403 HWEIGHT(n), 0, PERF_X86_EVENT_PEBS_NA_HSW) 404 405 406 /* 407 * We define the end marker as having a weight of -1 408 * to enable blacklisting of events using a counter bitmask 409 * of zero and thus a weight of zero. 410 * The end marker has a weight that cannot possibly be 411 * obtained from counting the bits in the bitmask. 412 */ 413 #define EVENT_CONSTRAINT_END { .weight = -1 } 414 415 /* 416 * Check for end marker with weight == -1 417 */ 418 #define for_each_event_constraint(e, c) \ 419 for ((e) = (c); (e)->weight != -1; (e)++) 420 421 /* 422 * Extra registers for specific events. 423 * 424 * Some events need large masks and require external MSRs. 425 * Those extra MSRs end up being shared for all events on 426 * a PMU and sometimes between PMU of sibling HT threads. 427 * In either case, the kernel needs to handle conflicting 428 * accesses to those extra, shared, regs. The data structure 429 * to manage those registers is stored in cpu_hw_event. 430 */ 431 struct extra_reg { 432 unsigned int event; 433 unsigned int msr; 434 u64 config_mask; 435 u64 valid_mask; 436 int idx; /* per_xxx->regs[] reg index */ 437 bool extra_msr_access; 438 }; 439 440 #define EVENT_EXTRA_REG(e, ms, m, vm, i) { \ 441 .event = (e), \ 442 .msr = (ms), \ 443 .config_mask = (m), \ 444 .valid_mask = (vm), \ 445 .idx = EXTRA_REG_##i, \ 446 .extra_msr_access = true, \ 447 } 448 449 #define INTEL_EVENT_EXTRA_REG(event, msr, vm, idx) \ 450 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT, vm, idx) 451 452 #define INTEL_UEVENT_EXTRA_REG(event, msr, vm, idx) \ 453 EVENT_EXTRA_REG(event, msr, ARCH_PERFMON_EVENTSEL_EVENT | \ 454 ARCH_PERFMON_EVENTSEL_UMASK, vm, idx) 455 456 #define INTEL_UEVENT_PEBS_LDLAT_EXTRA_REG(c) \ 457 INTEL_UEVENT_EXTRA_REG(c, \ 458 MSR_PEBS_LD_LAT_THRESHOLD, \ 459 0xffff, \ 460 LDLAT) 461 462 #define EVENT_EXTRA_END EVENT_EXTRA_REG(0, 0, 0, 0, RSP_0) 463 464 union perf_capabilities { 465 struct { 466 u64 lbr_format:6; 467 u64 pebs_trap:1; 468 u64 pebs_arch_reg:1; 469 u64 pebs_format:4; 470 u64 smm_freeze:1; 471 /* 472 * PMU supports separate counter range for writing 473 * values > 32bit. 474 */ 475 u64 full_width_write:1; 476 }; 477 u64 capabilities; 478 }; 479 480 struct x86_pmu_quirk { 481 struct x86_pmu_quirk *next; 482 void (*func)(void); 483 }; 484 485 union x86_pmu_config { 486 struct { 487 u64 event:8, 488 umask:8, 489 usr:1, 490 os:1, 491 edge:1, 492 pc:1, 493 interrupt:1, 494 __reserved1:1, 495 en:1, 496 inv:1, 497 cmask:8, 498 event2:4, 499 __reserved2:4, 500 go:1, 501 ho:1; 502 } bits; 503 u64 value; 504 }; 505 506 #define X86_CONFIG(args...) ((union x86_pmu_config){.bits = {args}}).value 507 508 enum { 509 x86_lbr_exclusive_lbr, 510 x86_lbr_exclusive_bts, 511 x86_lbr_exclusive_pt, 512 x86_lbr_exclusive_max, 513 }; 514 515 /* 516 * struct x86_pmu - generic x86 pmu 517 */ 518 struct x86_pmu { 519 /* 520 * Generic x86 PMC bits 521 */ 522 const char *name; 523 int version; 524 int (*handle_irq)(struct pt_regs *); 525 void (*disable_all)(void); 526 void (*enable_all)(int added); 527 void (*enable)(struct perf_event *); 528 void (*disable)(struct perf_event *); 529 void (*add)(struct perf_event *); 530 void (*del)(struct perf_event *); 531 void (*read)(struct perf_event *event); 532 int (*hw_config)(struct perf_event *event); 533 int (*schedule_events)(struct cpu_hw_events *cpuc, int n, int *assign); 534 unsigned eventsel; 535 unsigned perfctr; 536 int (*addr_offset)(int index, bool eventsel); 537 int (*rdpmc_index)(int index); 538 u64 (*event_map)(int); 539 int max_events; 540 int num_counters; 541 int num_counters_fixed; 542 int cntval_bits; 543 u64 cntval_mask; 544 union { 545 unsigned long events_maskl; 546 unsigned long events_mask[BITS_TO_LONGS(ARCH_PERFMON_EVENTS_COUNT)]; 547 }; 548 int events_mask_len; 549 int apic; 550 u64 max_period; 551 struct event_constraint * 552 (*get_event_constraints)(struct cpu_hw_events *cpuc, 553 int idx, 554 struct perf_event *event); 555 556 void (*put_event_constraints)(struct cpu_hw_events *cpuc, 557 struct perf_event *event); 558 559 void (*start_scheduling)(struct cpu_hw_events *cpuc); 560 561 void (*commit_scheduling)(struct cpu_hw_events *cpuc, int idx, int cntr); 562 563 void (*stop_scheduling)(struct cpu_hw_events *cpuc); 564 565 struct event_constraint *event_constraints; 566 struct x86_pmu_quirk *quirks; 567 int perfctr_second_write; 568 u64 (*limit_period)(struct perf_event *event, u64 l); 569 570 /* PMI handler bits */ 571 unsigned int late_ack :1, 572 counter_freezing :1; 573 /* 574 * sysfs attrs 575 */ 576 int attr_rdpmc_broken; 577 int attr_rdpmc; 578 struct attribute **format_attrs; 579 struct attribute **event_attrs; 580 struct attribute **caps_attrs; 581 582 ssize_t (*events_sysfs_show)(char *page, u64 config); 583 struct attribute **cpu_events; 584 585 unsigned long attr_freeze_on_smi; 586 struct attribute **attrs; 587 588 /* 589 * CPU Hotplug hooks 590 */ 591 int (*cpu_prepare)(int cpu); 592 void (*cpu_starting)(int cpu); 593 void (*cpu_dying)(int cpu); 594 void (*cpu_dead)(int cpu); 595 596 void (*check_microcode)(void); 597 void (*sched_task)(struct perf_event_context *ctx, 598 bool sched_in); 599 600 /* 601 * Intel Arch Perfmon v2+ 602 */ 603 u64 intel_ctrl; 604 union perf_capabilities intel_cap; 605 606 /* 607 * Intel DebugStore bits 608 */ 609 unsigned int bts :1, 610 bts_active :1, 611 pebs :1, 612 pebs_active :1, 613 pebs_broken :1, 614 pebs_prec_dist :1, 615 pebs_no_tlb :1, 616 pebs_no_isolation :1; 617 int pebs_record_size; 618 int pebs_buffer_size; 619 void (*drain_pebs)(struct pt_regs *regs); 620 struct event_constraint *pebs_constraints; 621 void (*pebs_aliases)(struct perf_event *event); 622 int max_pebs_events; 623 unsigned long large_pebs_flags; 624 625 /* 626 * Intel LBR 627 */ 628 unsigned long lbr_tos, lbr_from, lbr_to; /* MSR base regs */ 629 int lbr_nr; /* hardware stack size */ 630 u64 lbr_sel_mask; /* LBR_SELECT valid bits */ 631 const int *lbr_sel_map; /* lbr_select mappings */ 632 bool lbr_double_abort; /* duplicated lbr aborts */ 633 bool lbr_pt_coexist; /* (LBR|BTS) may coexist with PT */ 634 635 /* 636 * Intel PT/LBR/BTS are exclusive 637 */ 638 atomic_t lbr_exclusive[x86_lbr_exclusive_max]; 639 640 /* 641 * AMD bits 642 */ 643 unsigned int amd_nb_constraints : 1; 644 645 /* 646 * Extra registers for events 647 */ 648 struct extra_reg *extra_regs; 649 unsigned int flags; 650 651 /* 652 * Intel host/guest support (KVM) 653 */ 654 struct perf_guest_switch_msr *(*guest_get_msrs)(int *nr); 655 656 /* 657 * Check period value for PERF_EVENT_IOC_PERIOD ioctl. 658 */ 659 int (*check_period) (struct perf_event *event, u64 period); 660 }; 661 662 struct x86_perf_task_context { 663 u64 lbr_from[MAX_LBR_ENTRIES]; 664 u64 lbr_to[MAX_LBR_ENTRIES]; 665 u64 lbr_info[MAX_LBR_ENTRIES]; 666 int tos; 667 int valid_lbrs; 668 int lbr_callstack_users; 669 int lbr_stack_state; 670 int log_id; 671 }; 672 673 #define x86_add_quirk(func_) \ 674 do { \ 675 static struct x86_pmu_quirk __quirk __initdata = { \ 676 .func = func_, \ 677 }; \ 678 __quirk.next = x86_pmu.quirks; \ 679 x86_pmu.quirks = &__quirk; \ 680 } while (0) 681 682 /* 683 * x86_pmu flags 684 */ 685 #define PMU_FL_NO_HT_SHARING 0x1 /* no hyper-threading resource sharing */ 686 #define PMU_FL_HAS_RSP_1 0x2 /* has 2 equivalent offcore_rsp regs */ 687 #define PMU_FL_EXCL_CNTRS 0x4 /* has exclusive counter requirements */ 688 #define PMU_FL_EXCL_ENABLED 0x8 /* exclusive counter active */ 689 #define PMU_FL_PEBS_ALL 0x10 /* all events are valid PEBS events */ 690 #define PMU_FL_TFA 0x20 /* deal with TSX force abort */ 691 692 #define EVENT_VAR(_id) event_attr_##_id 693 #define EVENT_PTR(_id) &event_attr_##_id.attr.attr 694 695 #define EVENT_ATTR(_name, _id) \ 696 static struct perf_pmu_events_attr EVENT_VAR(_id) = { \ 697 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \ 698 .id = PERF_COUNT_HW_##_id, \ 699 .event_str = NULL, \ 700 }; 701 702 #define EVENT_ATTR_STR(_name, v, str) \ 703 static struct perf_pmu_events_attr event_attr_##v = { \ 704 .attr = __ATTR(_name, 0444, events_sysfs_show, NULL), \ 705 .id = 0, \ 706 .event_str = str, \ 707 }; 708 709 #define EVENT_ATTR_STR_HT(_name, v, noht, ht) \ 710 static struct perf_pmu_events_ht_attr event_attr_##v = { \ 711 .attr = __ATTR(_name, 0444, events_ht_sysfs_show, NULL),\ 712 .id = 0, \ 713 .event_str_noht = noht, \ 714 .event_str_ht = ht, \ 715 } 716 717 extern struct x86_pmu x86_pmu __read_mostly; 718 719 static inline bool x86_pmu_has_lbr_callstack(void) 720 { 721 return x86_pmu.lbr_sel_map && 722 x86_pmu.lbr_sel_map[PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT] > 0; 723 } 724 725 DECLARE_PER_CPU(struct cpu_hw_events, cpu_hw_events); 726 727 int x86_perf_event_set_period(struct perf_event *event); 728 729 /* 730 * Generalized hw caching related hw_event table, filled 731 * in on a per model basis. A value of 0 means 732 * 'not supported', -1 means 'hw_event makes no sense on 733 * this CPU', any other value means the raw hw_event 734 * ID. 735 */ 736 737 #define C(x) PERF_COUNT_HW_CACHE_##x 738 739 extern u64 __read_mostly hw_cache_event_ids 740 [PERF_COUNT_HW_CACHE_MAX] 741 [PERF_COUNT_HW_CACHE_OP_MAX] 742 [PERF_COUNT_HW_CACHE_RESULT_MAX]; 743 extern u64 __read_mostly hw_cache_extra_regs 744 [PERF_COUNT_HW_CACHE_MAX] 745 [PERF_COUNT_HW_CACHE_OP_MAX] 746 [PERF_COUNT_HW_CACHE_RESULT_MAX]; 747 748 u64 x86_perf_event_update(struct perf_event *event); 749 750 static inline unsigned int x86_pmu_config_addr(int index) 751 { 752 return x86_pmu.eventsel + (x86_pmu.addr_offset ? 753 x86_pmu.addr_offset(index, true) : index); 754 } 755 756 static inline unsigned int x86_pmu_event_addr(int index) 757 { 758 return x86_pmu.perfctr + (x86_pmu.addr_offset ? 759 x86_pmu.addr_offset(index, false) : index); 760 } 761 762 static inline int x86_pmu_rdpmc_index(int index) 763 { 764 return x86_pmu.rdpmc_index ? x86_pmu.rdpmc_index(index) : index; 765 } 766 767 int x86_add_exclusive(unsigned int what); 768 769 void x86_del_exclusive(unsigned int what); 770 771 int x86_reserve_hardware(void); 772 773 void x86_release_hardware(void); 774 775 int x86_pmu_max_precise(void); 776 777 void hw_perf_lbr_event_destroy(struct perf_event *event); 778 779 int x86_setup_perfctr(struct perf_event *event); 780 781 int x86_pmu_hw_config(struct perf_event *event); 782 783 void x86_pmu_disable_all(void); 784 785 static inline void __x86_pmu_enable_event(struct hw_perf_event *hwc, 786 u64 enable_mask) 787 { 788 u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask); 789 790 if (hwc->extra_reg.reg) 791 wrmsrl(hwc->extra_reg.reg, hwc->extra_reg.config); 792 wrmsrl(hwc->config_base, (hwc->config | enable_mask) & ~disable_mask); 793 } 794 795 void x86_pmu_enable_all(int added); 796 797 int perf_assign_events(struct event_constraint **constraints, int n, 798 int wmin, int wmax, int gpmax, int *assign); 799 int x86_schedule_events(struct cpu_hw_events *cpuc, int n, int *assign); 800 801 void x86_pmu_stop(struct perf_event *event, int flags); 802 803 static inline void x86_pmu_disable_event(struct perf_event *event) 804 { 805 struct hw_perf_event *hwc = &event->hw; 806 807 wrmsrl(hwc->config_base, hwc->config); 808 } 809 810 void x86_pmu_enable_event(struct perf_event *event); 811 812 int x86_pmu_handle_irq(struct pt_regs *regs); 813 814 extern struct event_constraint emptyconstraint; 815 816 extern struct event_constraint unconstrained; 817 818 static inline bool kernel_ip(unsigned long ip) 819 { 820 #ifdef CONFIG_X86_32 821 return ip > PAGE_OFFSET; 822 #else 823 return (long)ip < 0; 824 #endif 825 } 826 827 /* 828 * Not all PMUs provide the right context information to place the reported IP 829 * into full context. Specifically segment registers are typically not 830 * supplied. 831 * 832 * Assuming the address is a linear address (it is for IBS), we fake the CS and 833 * vm86 mode using the known zero-based code segment and 'fix up' the registers 834 * to reflect this. 835 * 836 * Intel PEBS/LBR appear to typically provide the effective address, nothing 837 * much we can do about that but pray and treat it like a linear address. 838 */ 839 static inline void set_linear_ip(struct pt_regs *regs, unsigned long ip) 840 { 841 regs->cs = kernel_ip(ip) ? __KERNEL_CS : __USER_CS; 842 if (regs->flags & X86_VM_MASK) 843 regs->flags ^= (PERF_EFLAGS_VM | X86_VM_MASK); 844 regs->ip = ip; 845 } 846 847 ssize_t x86_event_sysfs_show(char *page, u64 config, u64 event); 848 ssize_t intel_event_sysfs_show(char *page, u64 config); 849 850 struct attribute **merge_attr(struct attribute **a, struct attribute **b); 851 852 ssize_t events_sysfs_show(struct device *dev, struct device_attribute *attr, 853 char *page); 854 ssize_t events_ht_sysfs_show(struct device *dev, struct device_attribute *attr, 855 char *page); 856 857 #ifdef CONFIG_CPU_SUP_AMD 858 859 int amd_pmu_init(void); 860 861 #else /* CONFIG_CPU_SUP_AMD */ 862 863 static inline int amd_pmu_init(void) 864 { 865 return 0; 866 } 867 868 #endif /* CONFIG_CPU_SUP_AMD */ 869 870 #ifdef CONFIG_CPU_SUP_INTEL 871 872 static inline bool intel_pmu_has_bts_period(struct perf_event *event, u64 period) 873 { 874 struct hw_perf_event *hwc = &event->hw; 875 unsigned int hw_event, bts_event; 876 877 if (event->attr.freq) 878 return false; 879 880 hw_event = hwc->config & INTEL_ARCH_EVENT_MASK; 881 bts_event = x86_pmu.event_map(PERF_COUNT_HW_BRANCH_INSTRUCTIONS); 882 883 return hw_event == bts_event && period == 1; 884 } 885 886 static inline bool intel_pmu_has_bts(struct perf_event *event) 887 { 888 struct hw_perf_event *hwc = &event->hw; 889 890 return intel_pmu_has_bts_period(event, hwc->sample_period); 891 } 892 893 int intel_pmu_save_and_restart(struct perf_event *event); 894 895 struct event_constraint * 896 x86_get_event_constraints(struct cpu_hw_events *cpuc, int idx, 897 struct perf_event *event); 898 899 extern int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu); 900 extern void intel_cpuc_finish(struct cpu_hw_events *cpuc); 901 902 int intel_pmu_init(void); 903 904 void init_debug_store_on_cpu(int cpu); 905 906 void fini_debug_store_on_cpu(int cpu); 907 908 void release_ds_buffers(void); 909 910 void reserve_ds_buffers(void); 911 912 extern struct event_constraint bts_constraint; 913 914 void intel_pmu_enable_bts(u64 config); 915 916 void intel_pmu_disable_bts(void); 917 918 int intel_pmu_drain_bts_buffer(void); 919 920 extern struct event_constraint intel_core2_pebs_event_constraints[]; 921 922 extern struct event_constraint intel_atom_pebs_event_constraints[]; 923 924 extern struct event_constraint intel_slm_pebs_event_constraints[]; 925 926 extern struct event_constraint intel_glm_pebs_event_constraints[]; 927 928 extern struct event_constraint intel_glp_pebs_event_constraints[]; 929 930 extern struct event_constraint intel_nehalem_pebs_event_constraints[]; 931 932 extern struct event_constraint intel_westmere_pebs_event_constraints[]; 933 934 extern struct event_constraint intel_snb_pebs_event_constraints[]; 935 936 extern struct event_constraint intel_ivb_pebs_event_constraints[]; 937 938 extern struct event_constraint intel_hsw_pebs_event_constraints[]; 939 940 extern struct event_constraint intel_bdw_pebs_event_constraints[]; 941 942 extern struct event_constraint intel_skl_pebs_event_constraints[]; 943 944 struct event_constraint *intel_pebs_constraints(struct perf_event *event); 945 946 void intel_pmu_pebs_add(struct perf_event *event); 947 948 void intel_pmu_pebs_del(struct perf_event *event); 949 950 void intel_pmu_pebs_enable(struct perf_event *event); 951 952 void intel_pmu_pebs_disable(struct perf_event *event); 953 954 void intel_pmu_pebs_enable_all(void); 955 956 void intel_pmu_pebs_disable_all(void); 957 958 void intel_pmu_pebs_sched_task(struct perf_event_context *ctx, bool sched_in); 959 960 void intel_pmu_auto_reload_read(struct perf_event *event); 961 962 void intel_ds_init(void); 963 964 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in); 965 966 u64 lbr_from_signext_quirk_wr(u64 val); 967 968 void intel_pmu_lbr_reset(void); 969 970 void intel_pmu_lbr_add(struct perf_event *event); 971 972 void intel_pmu_lbr_del(struct perf_event *event); 973 974 void intel_pmu_lbr_enable_all(bool pmi); 975 976 void intel_pmu_lbr_disable_all(void); 977 978 void intel_pmu_lbr_read(void); 979 980 void intel_pmu_lbr_init_core(void); 981 982 void intel_pmu_lbr_init_nhm(void); 983 984 void intel_pmu_lbr_init_atom(void); 985 986 void intel_pmu_lbr_init_slm(void); 987 988 void intel_pmu_lbr_init_snb(void); 989 990 void intel_pmu_lbr_init_hsw(void); 991 992 void intel_pmu_lbr_init_skl(void); 993 994 void intel_pmu_lbr_init_knl(void); 995 996 void intel_pmu_pebs_data_source_nhm(void); 997 998 void intel_pmu_pebs_data_source_skl(bool pmem); 999 1000 int intel_pmu_setup_lbr_filter(struct perf_event *event); 1001 1002 void intel_pt_interrupt(void); 1003 1004 int intel_bts_interrupt(void); 1005 1006 void intel_bts_enable_local(void); 1007 1008 void intel_bts_disable_local(void); 1009 1010 int p4_pmu_init(void); 1011 1012 int p6_pmu_init(void); 1013 1014 int knc_pmu_init(void); 1015 1016 static inline int is_ht_workaround_enabled(void) 1017 { 1018 return !!(x86_pmu.flags & PMU_FL_EXCL_ENABLED); 1019 } 1020 1021 #else /* CONFIG_CPU_SUP_INTEL */ 1022 1023 static inline void reserve_ds_buffers(void) 1024 { 1025 } 1026 1027 static inline void release_ds_buffers(void) 1028 { 1029 } 1030 1031 static inline int intel_pmu_init(void) 1032 { 1033 return 0; 1034 } 1035 1036 static inline int intel_cpuc_prepare(struct cpu_hw_events *cpuc, int cpu) 1037 { 1038 return 0; 1039 } 1040 1041 static inline void intel_cpuc_finish(struct cpu_hw_events *cpuc) 1042 { 1043 } 1044 1045 static inline int is_ht_workaround_enabled(void) 1046 { 1047 return 0; 1048 } 1049 #endif /* CONFIG_CPU_SUP_INTEL */ 1050