1/* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de> 5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar 6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra 7 * 8 * Data type definitions, declarations, prototypes. 9 * 10 * Started by: Thomas Gleixner and Ingo Molnar 11 * 12 * For licencing details see kernel-base/COPYING 13 */ 14#ifndef _UAPI_LINUX_PERF_EVENT_H 15#define _UAPI_LINUX_PERF_EVENT_H 16 17#include <linux/types.h> 18#include <linux/ioctl.h> 19#include <asm/byteorder.h> 20 21/* 22 * User-space ABI bits: 23 */ 24 25/* 26 * attr.type 27 */ 28enum perf_type_id { 29 PERF_TYPE_HARDWARE = 0, 30 PERF_TYPE_SOFTWARE = 1, 31 PERF_TYPE_TRACEPOINT = 2, 32 PERF_TYPE_HW_CACHE = 3, 33 PERF_TYPE_RAW = 4, 34 PERF_TYPE_BREAKPOINT = 5, 35 36 PERF_TYPE_MAX, /* non-ABI */ 37}; 38 39/* 40 * Generalized performance event event_id types, used by the 41 * attr.event_id parameter of the sys_perf_event_open() 42 * syscall: 43 */ 44enum perf_hw_id { 45 /* 46 * Common hardware events, generalized by the kernel: 47 */ 48 PERF_COUNT_HW_CPU_CYCLES = 0, 49 PERF_COUNT_HW_INSTRUCTIONS = 1, 50 PERF_COUNT_HW_CACHE_REFERENCES = 2, 51 PERF_COUNT_HW_CACHE_MISSES = 3, 52 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, 53 PERF_COUNT_HW_BRANCH_MISSES = 5, 54 PERF_COUNT_HW_BUS_CYCLES = 6, 55 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND = 7, 56 PERF_COUNT_HW_STALLED_CYCLES_BACKEND = 8, 57 PERF_COUNT_HW_REF_CPU_CYCLES = 9, 58 59 PERF_COUNT_HW_MAX, /* non-ABI */ 60}; 61 62/* 63 * Generalized hardware cache events: 64 * 65 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x 66 * { read, write, prefetch } x 67 * { accesses, misses } 68 */ 69enum perf_hw_cache_id { 70 PERF_COUNT_HW_CACHE_L1D = 0, 71 PERF_COUNT_HW_CACHE_L1I = 1, 72 PERF_COUNT_HW_CACHE_LL = 2, 73 PERF_COUNT_HW_CACHE_DTLB = 3, 74 PERF_COUNT_HW_CACHE_ITLB = 4, 75 PERF_COUNT_HW_CACHE_BPU = 5, 76 PERF_COUNT_HW_CACHE_NODE = 6, 77 78 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 79}; 80 81enum perf_hw_cache_op_id { 82 PERF_COUNT_HW_CACHE_OP_READ = 0, 83 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 84 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 85 86 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 87}; 88 89enum perf_hw_cache_op_result_id { 90 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 91 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 92 93 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 94}; 95 96/* 97 * Special "software" events provided by the kernel, even if the hardware 98 * does not support performance events. These events measure various 99 * physical and sw events of the kernel (and allow the profiling of them as 100 * well): 101 */ 102enum perf_sw_ids { 103 PERF_COUNT_SW_CPU_CLOCK = 0, 104 PERF_COUNT_SW_TASK_CLOCK = 1, 105 PERF_COUNT_SW_PAGE_FAULTS = 2, 106 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 107 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 108 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 109 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 110 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7, 111 PERF_COUNT_SW_EMULATION_FAULTS = 8, 112 PERF_COUNT_SW_DUMMY = 9, 113 PERF_COUNT_SW_BPF_OUTPUT = 10, 114 115 PERF_COUNT_SW_MAX, /* non-ABI */ 116}; 117 118/* 119 * Bits that can be set in attr.sample_type to request information 120 * in the overflow packets. 121 */ 122enum perf_event_sample_format { 123 PERF_SAMPLE_IP = 1U << 0, 124 PERF_SAMPLE_TID = 1U << 1, 125 PERF_SAMPLE_TIME = 1U << 2, 126 PERF_SAMPLE_ADDR = 1U << 3, 127 PERF_SAMPLE_READ = 1U << 4, 128 PERF_SAMPLE_CALLCHAIN = 1U << 5, 129 PERF_SAMPLE_ID = 1U << 6, 130 PERF_SAMPLE_CPU = 1U << 7, 131 PERF_SAMPLE_PERIOD = 1U << 8, 132 PERF_SAMPLE_STREAM_ID = 1U << 9, 133 PERF_SAMPLE_RAW = 1U << 10, 134 PERF_SAMPLE_BRANCH_STACK = 1U << 11, 135 PERF_SAMPLE_REGS_USER = 1U << 12, 136 PERF_SAMPLE_STACK_USER = 1U << 13, 137 PERF_SAMPLE_WEIGHT = 1U << 14, 138 PERF_SAMPLE_DATA_SRC = 1U << 15, 139 PERF_SAMPLE_IDENTIFIER = 1U << 16, 140 PERF_SAMPLE_TRANSACTION = 1U << 17, 141 PERF_SAMPLE_REGS_INTR = 1U << 18, 142 143 PERF_SAMPLE_MAX = 1U << 19, /* non-ABI */ 144}; 145 146/* 147 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set 148 * 149 * If the user does not pass priv level information via branch_sample_type, 150 * the kernel uses the event's priv level. Branch and event priv levels do 151 * not have to match. Branch priv level is checked for permissions. 152 * 153 * The branch types can be combined, however BRANCH_ANY covers all types 154 * of branches and therefore it supersedes all the other types. 155 */ 156enum perf_branch_sample_type_shift { 157 PERF_SAMPLE_BRANCH_USER_SHIFT = 0, /* user branches */ 158 PERF_SAMPLE_BRANCH_KERNEL_SHIFT = 1, /* kernel branches */ 159 PERF_SAMPLE_BRANCH_HV_SHIFT = 2, /* hypervisor branches */ 160 161 PERF_SAMPLE_BRANCH_ANY_SHIFT = 3, /* any branch types */ 162 PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT = 4, /* any call branch */ 163 PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT = 5, /* any return branch */ 164 PERF_SAMPLE_BRANCH_IND_CALL_SHIFT = 6, /* indirect calls */ 165 PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT = 7, /* transaction aborts */ 166 PERF_SAMPLE_BRANCH_IN_TX_SHIFT = 8, /* in transaction */ 167 PERF_SAMPLE_BRANCH_NO_TX_SHIFT = 9, /* not in transaction */ 168 PERF_SAMPLE_BRANCH_COND_SHIFT = 10, /* conditional branches */ 169 170 PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT = 11, /* call/ret stack */ 171 PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT = 12, /* indirect jumps */ 172 PERF_SAMPLE_BRANCH_CALL_SHIFT = 13, /* direct call */ 173 174 PERF_SAMPLE_BRANCH_MAX_SHIFT /* non-ABI */ 175}; 176 177enum perf_branch_sample_type { 178 PERF_SAMPLE_BRANCH_USER = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT, 179 PERF_SAMPLE_BRANCH_KERNEL = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT, 180 PERF_SAMPLE_BRANCH_HV = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT, 181 182 PERF_SAMPLE_BRANCH_ANY = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT, 183 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT, 184 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT, 185 PERF_SAMPLE_BRANCH_IND_CALL = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT, 186 PERF_SAMPLE_BRANCH_ABORT_TX = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT, 187 PERF_SAMPLE_BRANCH_IN_TX = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT, 188 PERF_SAMPLE_BRANCH_NO_TX = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT, 189 PERF_SAMPLE_BRANCH_COND = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT, 190 191 PERF_SAMPLE_BRANCH_CALL_STACK = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT, 192 PERF_SAMPLE_BRANCH_IND_JUMP = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT, 193 PERF_SAMPLE_BRANCH_CALL = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT, 194 195 PERF_SAMPLE_BRANCH_MAX = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT, 196}; 197 198#define PERF_SAMPLE_BRANCH_PLM_ALL \ 199 (PERF_SAMPLE_BRANCH_USER|\ 200 PERF_SAMPLE_BRANCH_KERNEL|\ 201 PERF_SAMPLE_BRANCH_HV) 202 203/* 204 * Values to determine ABI of the registers dump. 205 */ 206enum perf_sample_regs_abi { 207 PERF_SAMPLE_REGS_ABI_NONE = 0, 208 PERF_SAMPLE_REGS_ABI_32 = 1, 209 PERF_SAMPLE_REGS_ABI_64 = 2, 210}; 211 212/* 213 * Values for the memory transaction event qualifier, mostly for 214 * abort events. Multiple bits can be set. 215 */ 216enum { 217 PERF_TXN_ELISION = (1 << 0), /* From elision */ 218 PERF_TXN_TRANSACTION = (1 << 1), /* From transaction */ 219 PERF_TXN_SYNC = (1 << 2), /* Instruction is related */ 220 PERF_TXN_ASYNC = (1 << 3), /* Instruction not related */ 221 PERF_TXN_RETRY = (1 << 4), /* Retry possible */ 222 PERF_TXN_CONFLICT = (1 << 5), /* Conflict abort */ 223 PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */ 224 PERF_TXN_CAPACITY_READ = (1 << 7), /* Capacity read abort */ 225 226 PERF_TXN_MAX = (1 << 8), /* non-ABI */ 227 228 /* bits 32..63 are reserved for the abort code */ 229 230 PERF_TXN_ABORT_MASK = (0xffffffffULL << 32), 231 PERF_TXN_ABORT_SHIFT = 32, 232}; 233 234/* 235 * The format of the data returned by read() on a perf event fd, 236 * as specified by attr.read_format: 237 * 238 * struct read_format { 239 * { u64 value; 240 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 241 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 242 * { u64 id; } && PERF_FORMAT_ID 243 * } && !PERF_FORMAT_GROUP 244 * 245 * { u64 nr; 246 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 247 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 248 * { u64 value; 249 * { u64 id; } && PERF_FORMAT_ID 250 * } cntr[nr]; 251 * } && PERF_FORMAT_GROUP 252 * }; 253 */ 254enum perf_event_read_format { 255 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 256 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 257 PERF_FORMAT_ID = 1U << 2, 258 PERF_FORMAT_GROUP = 1U << 3, 259 260 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 261}; 262 263#define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 264#define PERF_ATTR_SIZE_VER1 72 /* add: config2 */ 265#define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */ 266#define PERF_ATTR_SIZE_VER3 96 /* add: sample_regs_user */ 267 /* add: sample_stack_user */ 268#define PERF_ATTR_SIZE_VER4 104 /* add: sample_regs_intr */ 269#define PERF_ATTR_SIZE_VER5 112 /* add: aux_watermark */ 270 271/* 272 * Hardware event_id to monitor via a performance monitoring event: 273 */ 274struct perf_event_attr { 275 276 /* 277 * Major type: hardware/software/tracepoint/etc. 278 */ 279 __u32 type; 280 281 /* 282 * Size of the attr structure, for fwd/bwd compat. 283 */ 284 __u32 size; 285 286 /* 287 * Type specific configuration information. 288 */ 289 __u64 config; 290 291 union { 292 __u64 sample_period; 293 __u64 sample_freq; 294 }; 295 296 __u64 sample_type; 297 __u64 read_format; 298 299 __u64 disabled : 1, /* off by default */ 300 inherit : 1, /* children inherit it */ 301 pinned : 1, /* must always be on PMU */ 302 exclusive : 1, /* only group on PMU */ 303 exclude_user : 1, /* don't count user */ 304 exclude_kernel : 1, /* ditto kernel */ 305 exclude_hv : 1, /* ditto hypervisor */ 306 exclude_idle : 1, /* don't count when idle */ 307 mmap : 1, /* include mmap data */ 308 comm : 1, /* include comm data */ 309 freq : 1, /* use freq, not period */ 310 inherit_stat : 1, /* per task counts */ 311 enable_on_exec : 1, /* next exec enables */ 312 task : 1, /* trace fork/exit */ 313 watermark : 1, /* wakeup_watermark */ 314 /* 315 * precise_ip: 316 * 317 * 0 - SAMPLE_IP can have arbitrary skid 318 * 1 - SAMPLE_IP must have constant skid 319 * 2 - SAMPLE_IP requested to have 0 skid 320 * 3 - SAMPLE_IP must have 0 skid 321 * 322 * See also PERF_RECORD_MISC_EXACT_IP 323 */ 324 precise_ip : 2, /* skid constraint */ 325 mmap_data : 1, /* non-exec mmap data */ 326 sample_id_all : 1, /* sample_type all events */ 327 328 exclude_host : 1, /* don't count in host */ 329 exclude_guest : 1, /* don't count in guest */ 330 331 exclude_callchain_kernel : 1, /* exclude kernel callchains */ 332 exclude_callchain_user : 1, /* exclude user callchains */ 333 mmap2 : 1, /* include mmap with inode data */ 334 comm_exec : 1, /* flag comm events that are due to an exec */ 335 use_clockid : 1, /* use @clockid for time fields */ 336 context_switch : 1, /* context switch data */ 337 __reserved_1 : 37; 338 339 union { 340 __u32 wakeup_events; /* wakeup every n events */ 341 __u32 wakeup_watermark; /* bytes before wakeup */ 342 }; 343 344 __u32 bp_type; 345 union { 346 __u64 bp_addr; 347 __u64 config1; /* extension of config */ 348 }; 349 union { 350 __u64 bp_len; 351 __u64 config2; /* extension of config1 */ 352 }; 353 __u64 branch_sample_type; /* enum perf_branch_sample_type */ 354 355 /* 356 * Defines set of user regs to dump on samples. 357 * See asm/perf_regs.h for details. 358 */ 359 __u64 sample_regs_user; 360 361 /* 362 * Defines size of the user stack to dump on samples. 363 */ 364 __u32 sample_stack_user; 365 366 __s32 clockid; 367 /* 368 * Defines set of regs to dump for each sample 369 * state captured on: 370 * - precise = 0: PMU interrupt 371 * - precise > 0: sampled instruction 372 * 373 * See asm/perf_regs.h for details. 374 */ 375 __u64 sample_regs_intr; 376 377 /* 378 * Wakeup watermark for AUX area 379 */ 380 __u32 aux_watermark; 381 __u32 __reserved_2; /* align to __u64 */ 382}; 383 384#define perf_flags(attr) (*(&(attr)->read_format + 1)) 385 386/* 387 * Ioctls that can be done on a perf event fd: 388 */ 389#define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 390#define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 391#define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 392#define PERF_EVENT_IOC_RESET _IO ('$', 3) 393#define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 394#define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 395#define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 396#define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *) 397#define PERF_EVENT_IOC_SET_BPF _IOW('$', 8, __u32) 398 399enum perf_event_ioc_flags { 400 PERF_IOC_FLAG_GROUP = 1U << 0, 401}; 402 403/* 404 * Structure of the page that can be mapped via mmap 405 */ 406struct perf_event_mmap_page { 407 __u32 version; /* version number of this structure */ 408 __u32 compat_version; /* lowest version this is compat with */ 409 410 /* 411 * Bits needed to read the hw events in user-space. 412 * 413 * u32 seq, time_mult, time_shift, index, width; 414 * u64 count, enabled, running; 415 * u64 cyc, time_offset; 416 * s64 pmc = 0; 417 * 418 * do { 419 * seq = pc->lock; 420 * barrier() 421 * 422 * enabled = pc->time_enabled; 423 * running = pc->time_running; 424 * 425 * if (pc->cap_usr_time && enabled != running) { 426 * cyc = rdtsc(); 427 * time_offset = pc->time_offset; 428 * time_mult = pc->time_mult; 429 * time_shift = pc->time_shift; 430 * } 431 * 432 * index = pc->index; 433 * count = pc->offset; 434 * if (pc->cap_user_rdpmc && index) { 435 * width = pc->pmc_width; 436 * pmc = rdpmc(index - 1); 437 * } 438 * 439 * barrier(); 440 * } while (pc->lock != seq); 441 * 442 * NOTE: for obvious reason this only works on self-monitoring 443 * processes. 444 */ 445 __u32 lock; /* seqlock for synchronization */ 446 __u32 index; /* hardware event identifier */ 447 __s64 offset; /* add to hardware event value */ 448 __u64 time_enabled; /* time event active */ 449 __u64 time_running; /* time event on cpu */ 450 union { 451 __u64 capabilities; 452 struct { 453 __u64 cap_bit0 : 1, /* Always 0, deprecated, see commit 860f085b74e9 */ 454 cap_bit0_is_deprecated : 1, /* Always 1, signals that bit 0 is zero */ 455 456 cap_user_rdpmc : 1, /* The RDPMC instruction can be used to read counts */ 457 cap_user_time : 1, /* The time_* fields are used */ 458 cap_user_time_zero : 1, /* The time_zero field is used */ 459 cap_____res : 59; 460 }; 461 }; 462 463 /* 464 * If cap_user_rdpmc this field provides the bit-width of the value 465 * read using the rdpmc() or equivalent instruction. This can be used 466 * to sign extend the result like: 467 * 468 * pmc <<= 64 - width; 469 * pmc >>= 64 - width; // signed shift right 470 * count += pmc; 471 */ 472 __u16 pmc_width; 473 474 /* 475 * If cap_usr_time the below fields can be used to compute the time 476 * delta since time_enabled (in ns) using rdtsc or similar. 477 * 478 * u64 quot, rem; 479 * u64 delta; 480 * 481 * quot = (cyc >> time_shift); 482 * rem = cyc & (((u64)1 << time_shift) - 1); 483 * delta = time_offset + quot * time_mult + 484 * ((rem * time_mult) >> time_shift); 485 * 486 * Where time_offset,time_mult,time_shift and cyc are read in the 487 * seqcount loop described above. This delta can then be added to 488 * enabled and possible running (if index), improving the scaling: 489 * 490 * enabled += delta; 491 * if (index) 492 * running += delta; 493 * 494 * quot = count / running; 495 * rem = count % running; 496 * count = quot * enabled + (rem * enabled) / running; 497 */ 498 __u16 time_shift; 499 __u32 time_mult; 500 __u64 time_offset; 501 /* 502 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated 503 * from sample timestamps. 504 * 505 * time = timestamp - time_zero; 506 * quot = time / time_mult; 507 * rem = time % time_mult; 508 * cyc = (quot << time_shift) + (rem << time_shift) / time_mult; 509 * 510 * And vice versa: 511 * 512 * quot = cyc >> time_shift; 513 * rem = cyc & (((u64)1 << time_shift) - 1); 514 * timestamp = time_zero + quot * time_mult + 515 * ((rem * time_mult) >> time_shift); 516 */ 517 __u64 time_zero; 518 __u32 size; /* Header size up to __reserved[] fields. */ 519 520 /* 521 * Hole for extension of the self monitor capabilities 522 */ 523 524 __u8 __reserved[118*8+4]; /* align to 1k. */ 525 526 /* 527 * Control data for the mmap() data buffer. 528 * 529 * User-space reading the @data_head value should issue an smp_rmb(), 530 * after reading this value. 531 * 532 * When the mapping is PROT_WRITE the @data_tail value should be 533 * written by userspace to reflect the last read data, after issueing 534 * an smp_mb() to separate the data read from the ->data_tail store. 535 * In this case the kernel will not over-write unread data. 536 * 537 * See perf_output_put_handle() for the data ordering. 538 * 539 * data_{offset,size} indicate the location and size of the perf record 540 * buffer within the mmapped area. 541 */ 542 __u64 data_head; /* head in the data section */ 543 __u64 data_tail; /* user-space written tail */ 544 __u64 data_offset; /* where the buffer starts */ 545 __u64 data_size; /* data buffer size */ 546 547 /* 548 * AUX area is defined by aux_{offset,size} fields that should be set 549 * by the userspace, so that 550 * 551 * aux_offset >= data_offset + data_size 552 * 553 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size. 554 * 555 * Ring buffer pointers aux_{head,tail} have the same semantics as 556 * data_{head,tail} and same ordering rules apply. 557 */ 558 __u64 aux_head; 559 __u64 aux_tail; 560 __u64 aux_offset; 561 __u64 aux_size; 562}; 563 564#define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 565#define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 566#define PERF_RECORD_MISC_KERNEL (1 << 0) 567#define PERF_RECORD_MISC_USER (2 << 0) 568#define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 569#define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 570#define PERF_RECORD_MISC_GUEST_USER (5 << 0) 571 572/* 573 * Indicates that /proc/PID/maps parsing are truncated by time out. 574 */ 575#define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12) 576/* 577 * PERF_RECORD_MISC_MMAP_DATA and PERF_RECORD_MISC_COMM_EXEC are used on 578 * different events so can reuse the same bit position. 579 * Ditto PERF_RECORD_MISC_SWITCH_OUT. 580 */ 581#define PERF_RECORD_MISC_MMAP_DATA (1 << 13) 582#define PERF_RECORD_MISC_COMM_EXEC (1 << 13) 583#define PERF_RECORD_MISC_SWITCH_OUT (1 << 13) 584/* 585 * Indicates that the content of PERF_SAMPLE_IP points to 586 * the actual instruction that triggered the event. See also 587 * perf_event_attr::precise_ip. 588 */ 589#define PERF_RECORD_MISC_EXACT_IP (1 << 14) 590/* 591 * Reserve the last bit to indicate some extended misc field 592 */ 593#define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 594 595struct perf_event_header { 596 __u32 type; 597 __u16 misc; 598 __u16 size; 599}; 600 601enum perf_event_type { 602 603 /* 604 * If perf_event_attr.sample_id_all is set then all event types will 605 * have the sample_type selected fields related to where/when 606 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU, 607 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed 608 * just after the perf_event_header and the fields already present for 609 * the existing fields, i.e. at the end of the payload. That way a newer 610 * perf.data file will be supported by older perf tools, with these new 611 * optional fields being ignored. 612 * 613 * struct sample_id { 614 * { u32 pid, tid; } && PERF_SAMPLE_TID 615 * { u64 time; } && PERF_SAMPLE_TIME 616 * { u64 id; } && PERF_SAMPLE_ID 617 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 618 * { u32 cpu, res; } && PERF_SAMPLE_CPU 619 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 620 * } && perf_event_attr::sample_id_all 621 * 622 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. The 623 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed 624 * relative to header.size. 625 */ 626 627 /* 628 * The MMAP events record the PROT_EXEC mappings so that we can 629 * correlate userspace IPs to code. They have the following structure: 630 * 631 * struct { 632 * struct perf_event_header header; 633 * 634 * u32 pid, tid; 635 * u64 addr; 636 * u64 len; 637 * u64 pgoff; 638 * char filename[]; 639 * struct sample_id sample_id; 640 * }; 641 */ 642 PERF_RECORD_MMAP = 1, 643 644 /* 645 * struct { 646 * struct perf_event_header header; 647 * u64 id; 648 * u64 lost; 649 * struct sample_id sample_id; 650 * }; 651 */ 652 PERF_RECORD_LOST = 2, 653 654 /* 655 * struct { 656 * struct perf_event_header header; 657 * 658 * u32 pid, tid; 659 * char comm[]; 660 * struct sample_id sample_id; 661 * }; 662 */ 663 PERF_RECORD_COMM = 3, 664 665 /* 666 * struct { 667 * struct perf_event_header header; 668 * u32 pid, ppid; 669 * u32 tid, ptid; 670 * u64 time; 671 * struct sample_id sample_id; 672 * }; 673 */ 674 PERF_RECORD_EXIT = 4, 675 676 /* 677 * struct { 678 * struct perf_event_header header; 679 * u64 time; 680 * u64 id; 681 * u64 stream_id; 682 * struct sample_id sample_id; 683 * }; 684 */ 685 PERF_RECORD_THROTTLE = 5, 686 PERF_RECORD_UNTHROTTLE = 6, 687 688 /* 689 * struct { 690 * struct perf_event_header header; 691 * u32 pid, ppid; 692 * u32 tid, ptid; 693 * u64 time; 694 * struct sample_id sample_id; 695 * }; 696 */ 697 PERF_RECORD_FORK = 7, 698 699 /* 700 * struct { 701 * struct perf_event_header header; 702 * u32 pid, tid; 703 * 704 * struct read_format values; 705 * struct sample_id sample_id; 706 * }; 707 */ 708 PERF_RECORD_READ = 8, 709 710 /* 711 * struct { 712 * struct perf_event_header header; 713 * 714 * # 715 * # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. 716 * # The advantage of PERF_SAMPLE_IDENTIFIER is that its position 717 * # is fixed relative to header. 718 * # 719 * 720 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 721 * { u64 ip; } && PERF_SAMPLE_IP 722 * { u32 pid, tid; } && PERF_SAMPLE_TID 723 * { u64 time; } && PERF_SAMPLE_TIME 724 * { u64 addr; } && PERF_SAMPLE_ADDR 725 * { u64 id; } && PERF_SAMPLE_ID 726 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 727 * { u32 cpu, res; } && PERF_SAMPLE_CPU 728 * { u64 period; } && PERF_SAMPLE_PERIOD 729 * 730 * { struct read_format values; } && PERF_SAMPLE_READ 731 * 732 * { u64 nr, 733 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 734 * 735 * # 736 * # The RAW record below is opaque data wrt the ABI 737 * # 738 * # That is, the ABI doesn't make any promises wrt to 739 * # the stability of its content, it may vary depending 740 * # on event, hardware, kernel version and phase of 741 * # the moon. 742 * # 743 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 744 * # 745 * 746 * { u32 size; 747 * char data[size];}&& PERF_SAMPLE_RAW 748 * 749 * { u64 nr; 750 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK 751 * 752 * { u64 abi; # enum perf_sample_regs_abi 753 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER 754 * 755 * { u64 size; 756 * char data[size]; 757 * u64 dyn_size; } && PERF_SAMPLE_STACK_USER 758 * 759 * { u64 weight; } && PERF_SAMPLE_WEIGHT 760 * { u64 data_src; } && PERF_SAMPLE_DATA_SRC 761 * { u64 transaction; } && PERF_SAMPLE_TRANSACTION 762 * { u64 abi; # enum perf_sample_regs_abi 763 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR 764 * }; 765 */ 766 PERF_RECORD_SAMPLE = 9, 767 768 /* 769 * The MMAP2 records are an augmented version of MMAP, they add 770 * maj, min, ino numbers to be used to uniquely identify each mapping 771 * 772 * struct { 773 * struct perf_event_header header; 774 * 775 * u32 pid, tid; 776 * u64 addr; 777 * u64 len; 778 * u64 pgoff; 779 * u32 maj; 780 * u32 min; 781 * u64 ino; 782 * u64 ino_generation; 783 * u32 prot, flags; 784 * char filename[]; 785 * struct sample_id sample_id; 786 * }; 787 */ 788 PERF_RECORD_MMAP2 = 10, 789 790 /* 791 * Records that new data landed in the AUX buffer part. 792 * 793 * struct { 794 * struct perf_event_header header; 795 * 796 * u64 aux_offset; 797 * u64 aux_size; 798 * u64 flags; 799 * struct sample_id sample_id; 800 * }; 801 */ 802 PERF_RECORD_AUX = 11, 803 804 /* 805 * Indicates that instruction trace has started 806 * 807 * struct { 808 * struct perf_event_header header; 809 * u32 pid; 810 * u32 tid; 811 * }; 812 */ 813 PERF_RECORD_ITRACE_START = 12, 814 815 /* 816 * Records the dropped/lost sample number. 817 * 818 * struct { 819 * struct perf_event_header header; 820 * 821 * u64 lost; 822 * struct sample_id sample_id; 823 * }; 824 */ 825 PERF_RECORD_LOST_SAMPLES = 13, 826 827 /* 828 * Records a context switch in or out (flagged by 829 * PERF_RECORD_MISC_SWITCH_OUT). See also 830 * PERF_RECORD_SWITCH_CPU_WIDE. 831 * 832 * struct { 833 * struct perf_event_header header; 834 * struct sample_id sample_id; 835 * }; 836 */ 837 PERF_RECORD_SWITCH = 14, 838 839 /* 840 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and 841 * next_prev_tid that are the next (switching out) or previous 842 * (switching in) pid/tid. 843 * 844 * struct { 845 * struct perf_event_header header; 846 * u32 next_prev_pid; 847 * u32 next_prev_tid; 848 * struct sample_id sample_id; 849 * }; 850 */ 851 PERF_RECORD_SWITCH_CPU_WIDE = 15, 852 853 PERF_RECORD_MAX, /* non-ABI */ 854}; 855 856#define PERF_MAX_STACK_DEPTH 127 857 858enum perf_callchain_context { 859 PERF_CONTEXT_HV = (__u64)-32, 860 PERF_CONTEXT_KERNEL = (__u64)-128, 861 PERF_CONTEXT_USER = (__u64)-512, 862 863 PERF_CONTEXT_GUEST = (__u64)-2048, 864 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 865 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 866 867 PERF_CONTEXT_MAX = (__u64)-4095, 868}; 869 870/** 871 * PERF_RECORD_AUX::flags bits 872 */ 873#define PERF_AUX_FLAG_TRUNCATED 0x01 /* record was truncated to fit */ 874#define PERF_AUX_FLAG_OVERWRITE 0x02 /* snapshot from overwrite mode */ 875 876#define PERF_FLAG_FD_NO_GROUP (1UL << 0) 877#define PERF_FLAG_FD_OUTPUT (1UL << 1) 878#define PERF_FLAG_PID_CGROUP (1UL << 2) /* pid=cgroup id, per-cpu mode only */ 879#define PERF_FLAG_FD_CLOEXEC (1UL << 3) /* O_CLOEXEC */ 880 881union perf_mem_data_src { 882 __u64 val; 883 struct { 884 __u64 mem_op:5, /* type of opcode */ 885 mem_lvl:14, /* memory hierarchy level */ 886 mem_snoop:5, /* snoop mode */ 887 mem_lock:2, /* lock instr */ 888 mem_dtlb:7, /* tlb access */ 889 mem_rsvd:31; 890 }; 891}; 892 893/* type of opcode (load/store/prefetch,code) */ 894#define PERF_MEM_OP_NA 0x01 /* not available */ 895#define PERF_MEM_OP_LOAD 0x02 /* load instruction */ 896#define PERF_MEM_OP_STORE 0x04 /* store instruction */ 897#define PERF_MEM_OP_PFETCH 0x08 /* prefetch */ 898#define PERF_MEM_OP_EXEC 0x10 /* code (execution) */ 899#define PERF_MEM_OP_SHIFT 0 900 901/* memory hierarchy (memory level, hit or miss) */ 902#define PERF_MEM_LVL_NA 0x01 /* not available */ 903#define PERF_MEM_LVL_HIT 0x02 /* hit level */ 904#define PERF_MEM_LVL_MISS 0x04 /* miss level */ 905#define PERF_MEM_LVL_L1 0x08 /* L1 */ 906#define PERF_MEM_LVL_LFB 0x10 /* Line Fill Buffer */ 907#define PERF_MEM_LVL_L2 0x20 /* L2 */ 908#define PERF_MEM_LVL_L3 0x40 /* L3 */ 909#define PERF_MEM_LVL_LOC_RAM 0x80 /* Local DRAM */ 910#define PERF_MEM_LVL_REM_RAM1 0x100 /* Remote DRAM (1 hop) */ 911#define PERF_MEM_LVL_REM_RAM2 0x200 /* Remote DRAM (2 hops) */ 912#define PERF_MEM_LVL_REM_CCE1 0x400 /* Remote Cache (1 hop) */ 913#define PERF_MEM_LVL_REM_CCE2 0x800 /* Remote Cache (2 hops) */ 914#define PERF_MEM_LVL_IO 0x1000 /* I/O memory */ 915#define PERF_MEM_LVL_UNC 0x2000 /* Uncached memory */ 916#define PERF_MEM_LVL_SHIFT 5 917 918/* snoop mode */ 919#define PERF_MEM_SNOOP_NA 0x01 /* not available */ 920#define PERF_MEM_SNOOP_NONE 0x02 /* no snoop */ 921#define PERF_MEM_SNOOP_HIT 0x04 /* snoop hit */ 922#define PERF_MEM_SNOOP_MISS 0x08 /* snoop miss */ 923#define PERF_MEM_SNOOP_HITM 0x10 /* snoop hit modified */ 924#define PERF_MEM_SNOOP_SHIFT 19 925 926/* locked instruction */ 927#define PERF_MEM_LOCK_NA 0x01 /* not available */ 928#define PERF_MEM_LOCK_LOCKED 0x02 /* locked transaction */ 929#define PERF_MEM_LOCK_SHIFT 24 930 931/* TLB access */ 932#define PERF_MEM_TLB_NA 0x01 /* not available */ 933#define PERF_MEM_TLB_HIT 0x02 /* hit level */ 934#define PERF_MEM_TLB_MISS 0x04 /* miss level */ 935#define PERF_MEM_TLB_L1 0x08 /* L1 */ 936#define PERF_MEM_TLB_L2 0x10 /* L2 */ 937#define PERF_MEM_TLB_WK 0x20 /* Hardware Walker*/ 938#define PERF_MEM_TLB_OS 0x40 /* OS fault handler */ 939#define PERF_MEM_TLB_SHIFT 26 940 941#define PERF_MEM_S(a, s) \ 942 (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT) 943 944/* 945 * single taken branch record layout: 946 * 947 * from: source instruction (may not always be a branch insn) 948 * to: branch target 949 * mispred: branch target was mispredicted 950 * predicted: branch target was predicted 951 * 952 * support for mispred, predicted is optional. In case it 953 * is not supported mispred = predicted = 0. 954 * 955 * in_tx: running in a hardware transaction 956 * abort: aborting a hardware transaction 957 * cycles: cycles from last branch (or 0 if not supported) 958 */ 959struct perf_branch_entry { 960 __u64 from; 961 __u64 to; 962 __u64 mispred:1, /* target mispredicted */ 963 predicted:1,/* target predicted */ 964 in_tx:1, /* in transaction */ 965 abort:1, /* transaction abort */ 966 cycles:16, /* cycle count to last branch */ 967 reserved:44; 968}; 969 970#endif /* _UAPI_LINUX_PERF_EVENT_H */ 971