1 /* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <[email protected]> 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 */ 28 enum 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 */ 44 enum 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 */ 69 enum 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 81 enum 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 89 enum 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 */ 102 enum 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 */ 122 enum 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 */ 156 enum 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_NO_FLAGS_SHIFT = 14, /* no flags */ 175 PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT = 15, /* no cycles */ 176 177 PERF_SAMPLE_BRANCH_MAX_SHIFT /* non-ABI */ 178 }; 179 180 enum perf_branch_sample_type { 181 PERF_SAMPLE_BRANCH_USER = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT, 182 PERF_SAMPLE_BRANCH_KERNEL = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT, 183 PERF_SAMPLE_BRANCH_HV = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT, 184 185 PERF_SAMPLE_BRANCH_ANY = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT, 186 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT, 187 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT, 188 PERF_SAMPLE_BRANCH_IND_CALL = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT, 189 PERF_SAMPLE_BRANCH_ABORT_TX = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT, 190 PERF_SAMPLE_BRANCH_IN_TX = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT, 191 PERF_SAMPLE_BRANCH_NO_TX = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT, 192 PERF_SAMPLE_BRANCH_COND = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT, 193 194 PERF_SAMPLE_BRANCH_CALL_STACK = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT, 195 PERF_SAMPLE_BRANCH_IND_JUMP = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT, 196 PERF_SAMPLE_BRANCH_CALL = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT, 197 198 PERF_SAMPLE_BRANCH_NO_FLAGS = 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT, 199 PERF_SAMPLE_BRANCH_NO_CYCLES = 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT, 200 201 PERF_SAMPLE_BRANCH_MAX = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT, 202 }; 203 204 #define PERF_SAMPLE_BRANCH_PLM_ALL \ 205 (PERF_SAMPLE_BRANCH_USER|\ 206 PERF_SAMPLE_BRANCH_KERNEL|\ 207 PERF_SAMPLE_BRANCH_HV) 208 209 /* 210 * Values to determine ABI of the registers dump. 211 */ 212 enum perf_sample_regs_abi { 213 PERF_SAMPLE_REGS_ABI_NONE = 0, 214 PERF_SAMPLE_REGS_ABI_32 = 1, 215 PERF_SAMPLE_REGS_ABI_64 = 2, 216 }; 217 218 /* 219 * Values for the memory transaction event qualifier, mostly for 220 * abort events. Multiple bits can be set. 221 */ 222 enum { 223 PERF_TXN_ELISION = (1 << 0), /* From elision */ 224 PERF_TXN_TRANSACTION = (1 << 1), /* From transaction */ 225 PERF_TXN_SYNC = (1 << 2), /* Instruction is related */ 226 PERF_TXN_ASYNC = (1 << 3), /* Instruction not related */ 227 PERF_TXN_RETRY = (1 << 4), /* Retry possible */ 228 PERF_TXN_CONFLICT = (1 << 5), /* Conflict abort */ 229 PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */ 230 PERF_TXN_CAPACITY_READ = (1 << 7), /* Capacity read abort */ 231 232 PERF_TXN_MAX = (1 << 8), /* non-ABI */ 233 234 /* bits 32..63 are reserved for the abort code */ 235 236 PERF_TXN_ABORT_MASK = (0xffffffffULL << 32), 237 PERF_TXN_ABORT_SHIFT = 32, 238 }; 239 240 /* 241 * The format of the data returned by read() on a perf event fd, 242 * as specified by attr.read_format: 243 * 244 * struct read_format { 245 * { u64 value; 246 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 247 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 248 * { u64 id; } && PERF_FORMAT_ID 249 * } && !PERF_FORMAT_GROUP 250 * 251 * { u64 nr; 252 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 253 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 254 * { u64 value; 255 * { u64 id; } && PERF_FORMAT_ID 256 * } cntr[nr]; 257 * } && PERF_FORMAT_GROUP 258 * }; 259 */ 260 enum perf_event_read_format { 261 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 262 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 263 PERF_FORMAT_ID = 1U << 2, 264 PERF_FORMAT_GROUP = 1U << 3, 265 266 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 267 }; 268 269 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 270 #define PERF_ATTR_SIZE_VER1 72 /* add: config2 */ 271 #define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */ 272 #define PERF_ATTR_SIZE_VER3 96 /* add: sample_regs_user */ 273 /* add: sample_stack_user */ 274 #define PERF_ATTR_SIZE_VER4 104 /* add: sample_regs_intr */ 275 #define PERF_ATTR_SIZE_VER5 112 /* add: aux_watermark */ 276 277 /* 278 * Hardware event_id to monitor via a performance monitoring event: 279 */ 280 struct perf_event_attr { 281 282 /* 283 * Major type: hardware/software/tracepoint/etc. 284 */ 285 __u32 type; 286 287 /* 288 * Size of the attr structure, for fwd/bwd compat. 289 */ 290 __u32 size; 291 292 /* 293 * Type specific configuration information. 294 */ 295 __u64 config; 296 297 union { 298 __u64 sample_period; 299 __u64 sample_freq; 300 }; 301 302 __u64 sample_type; 303 __u64 read_format; 304 305 __u64 disabled : 1, /* off by default */ 306 inherit : 1, /* children inherit it */ 307 pinned : 1, /* must always be on PMU */ 308 exclusive : 1, /* only group on PMU */ 309 exclude_user : 1, /* don't count user */ 310 exclude_kernel : 1, /* ditto kernel */ 311 exclude_hv : 1, /* ditto hypervisor */ 312 exclude_idle : 1, /* don't count when idle */ 313 mmap : 1, /* include mmap data */ 314 comm : 1, /* include comm data */ 315 freq : 1, /* use freq, not period */ 316 inherit_stat : 1, /* per task counts */ 317 enable_on_exec : 1, /* next exec enables */ 318 task : 1, /* trace fork/exit */ 319 watermark : 1, /* wakeup_watermark */ 320 /* 321 * precise_ip: 322 * 323 * 0 - SAMPLE_IP can have arbitrary skid 324 * 1 - SAMPLE_IP must have constant skid 325 * 2 - SAMPLE_IP requested to have 0 skid 326 * 3 - SAMPLE_IP must have 0 skid 327 * 328 * See also PERF_RECORD_MISC_EXACT_IP 329 */ 330 precise_ip : 2, /* skid constraint */ 331 mmap_data : 1, /* non-exec mmap data */ 332 sample_id_all : 1, /* sample_type all events */ 333 334 exclude_host : 1, /* don't count in host */ 335 exclude_guest : 1, /* don't count in guest */ 336 337 exclude_callchain_kernel : 1, /* exclude kernel callchains */ 338 exclude_callchain_user : 1, /* exclude user callchains */ 339 mmap2 : 1, /* include mmap with inode data */ 340 comm_exec : 1, /* flag comm events that are due to an exec */ 341 use_clockid : 1, /* use @clockid for time fields */ 342 context_switch : 1, /* context switch data */ 343 __reserved_1 : 37; 344 345 union { 346 __u32 wakeup_events; /* wakeup every n events */ 347 __u32 wakeup_watermark; /* bytes before wakeup */ 348 }; 349 350 __u32 bp_type; 351 union { 352 __u64 bp_addr; 353 __u64 config1; /* extension of config */ 354 }; 355 union { 356 __u64 bp_len; 357 __u64 config2; /* extension of config1 */ 358 }; 359 __u64 branch_sample_type; /* enum perf_branch_sample_type */ 360 361 /* 362 * Defines set of user regs to dump on samples. 363 * See asm/perf_regs.h for details. 364 */ 365 __u64 sample_regs_user; 366 367 /* 368 * Defines size of the user stack to dump on samples. 369 */ 370 __u32 sample_stack_user; 371 372 __s32 clockid; 373 /* 374 * Defines set of regs to dump for each sample 375 * state captured on: 376 * - precise = 0: PMU interrupt 377 * - precise > 0: sampled instruction 378 * 379 * See asm/perf_regs.h for details. 380 */ 381 __u64 sample_regs_intr; 382 383 /* 384 * Wakeup watermark for AUX area 385 */ 386 __u32 aux_watermark; 387 __u32 __reserved_2; /* align to __u64 */ 388 }; 389 390 #define perf_flags(attr) (*(&(attr)->read_format + 1)) 391 392 /* 393 * Ioctls that can be done on a perf event fd: 394 */ 395 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 396 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 397 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 398 #define PERF_EVENT_IOC_RESET _IO ('$', 3) 399 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 400 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 401 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 402 #define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *) 403 #define PERF_EVENT_IOC_SET_BPF _IOW('$', 8, __u32) 404 405 enum perf_event_ioc_flags { 406 PERF_IOC_FLAG_GROUP = 1U << 0, 407 }; 408 409 /* 410 * Structure of the page that can be mapped via mmap 411 */ 412 struct perf_event_mmap_page { 413 __u32 version; /* version number of this structure */ 414 __u32 compat_version; /* lowest version this is compat with */ 415 416 /* 417 * Bits needed to read the hw events in user-space. 418 * 419 * u32 seq, time_mult, time_shift, index, width; 420 * u64 count, enabled, running; 421 * u64 cyc, time_offset; 422 * s64 pmc = 0; 423 * 424 * do { 425 * seq = pc->lock; 426 * barrier() 427 * 428 * enabled = pc->time_enabled; 429 * running = pc->time_running; 430 * 431 * if (pc->cap_usr_time && enabled != running) { 432 * cyc = rdtsc(); 433 * time_offset = pc->time_offset; 434 * time_mult = pc->time_mult; 435 * time_shift = pc->time_shift; 436 * } 437 * 438 * index = pc->index; 439 * count = pc->offset; 440 * if (pc->cap_user_rdpmc && index) { 441 * width = pc->pmc_width; 442 * pmc = rdpmc(index - 1); 443 * } 444 * 445 * barrier(); 446 * } while (pc->lock != seq); 447 * 448 * NOTE: for obvious reason this only works on self-monitoring 449 * processes. 450 */ 451 __u32 lock; /* seqlock for synchronization */ 452 __u32 index; /* hardware event identifier */ 453 __s64 offset; /* add to hardware event value */ 454 __u64 time_enabled; /* time event active */ 455 __u64 time_running; /* time event on cpu */ 456 union { 457 __u64 capabilities; 458 struct { 459 __u64 cap_bit0 : 1, /* Always 0, deprecated, see commit 860f085b74e9 */ 460 cap_bit0_is_deprecated : 1, /* Always 1, signals that bit 0 is zero */ 461 462 cap_user_rdpmc : 1, /* The RDPMC instruction can be used to read counts */ 463 cap_user_time : 1, /* The time_* fields are used */ 464 cap_user_time_zero : 1, /* The time_zero field is used */ 465 cap_____res : 59; 466 }; 467 }; 468 469 /* 470 * If cap_user_rdpmc this field provides the bit-width of the value 471 * read using the rdpmc() or equivalent instruction. This can be used 472 * to sign extend the result like: 473 * 474 * pmc <<= 64 - width; 475 * pmc >>= 64 - width; // signed shift right 476 * count += pmc; 477 */ 478 __u16 pmc_width; 479 480 /* 481 * If cap_usr_time the below fields can be used to compute the time 482 * delta since time_enabled (in ns) using rdtsc or similar. 483 * 484 * u64 quot, rem; 485 * u64 delta; 486 * 487 * quot = (cyc >> time_shift); 488 * rem = cyc & (((u64)1 << time_shift) - 1); 489 * delta = time_offset + quot * time_mult + 490 * ((rem * time_mult) >> time_shift); 491 * 492 * Where time_offset,time_mult,time_shift and cyc are read in the 493 * seqcount loop described above. This delta can then be added to 494 * enabled and possible running (if index), improving the scaling: 495 * 496 * enabled += delta; 497 * if (index) 498 * running += delta; 499 * 500 * quot = count / running; 501 * rem = count % running; 502 * count = quot * enabled + (rem * enabled) / running; 503 */ 504 __u16 time_shift; 505 __u32 time_mult; 506 __u64 time_offset; 507 /* 508 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated 509 * from sample timestamps. 510 * 511 * time = timestamp - time_zero; 512 * quot = time / time_mult; 513 * rem = time % time_mult; 514 * cyc = (quot << time_shift) + (rem << time_shift) / time_mult; 515 * 516 * And vice versa: 517 * 518 * quot = cyc >> time_shift; 519 * rem = cyc & (((u64)1 << time_shift) - 1); 520 * timestamp = time_zero + quot * time_mult + 521 * ((rem * time_mult) >> time_shift); 522 */ 523 __u64 time_zero; 524 __u32 size; /* Header size up to __reserved[] fields. */ 525 526 /* 527 * Hole for extension of the self monitor capabilities 528 */ 529 530 __u8 __reserved[118*8+4]; /* align to 1k. */ 531 532 /* 533 * Control data for the mmap() data buffer. 534 * 535 * User-space reading the @data_head value should issue an smp_rmb(), 536 * after reading this value. 537 * 538 * When the mapping is PROT_WRITE the @data_tail value should be 539 * written by userspace to reflect the last read data, after issueing 540 * an smp_mb() to separate the data read from the ->data_tail store. 541 * In this case the kernel will not over-write unread data. 542 * 543 * See perf_output_put_handle() for the data ordering. 544 * 545 * data_{offset,size} indicate the location and size of the perf record 546 * buffer within the mmapped area. 547 */ 548 __u64 data_head; /* head in the data section */ 549 __u64 data_tail; /* user-space written tail */ 550 __u64 data_offset; /* where the buffer starts */ 551 __u64 data_size; /* data buffer size */ 552 553 /* 554 * AUX area is defined by aux_{offset,size} fields that should be set 555 * by the userspace, so that 556 * 557 * aux_offset >= data_offset + data_size 558 * 559 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size. 560 * 561 * Ring buffer pointers aux_{head,tail} have the same semantics as 562 * data_{head,tail} and same ordering rules apply. 563 */ 564 __u64 aux_head; 565 __u64 aux_tail; 566 __u64 aux_offset; 567 __u64 aux_size; 568 }; 569 570 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 571 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 572 #define PERF_RECORD_MISC_KERNEL (1 << 0) 573 #define PERF_RECORD_MISC_USER (2 << 0) 574 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 575 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 576 #define PERF_RECORD_MISC_GUEST_USER (5 << 0) 577 578 /* 579 * Indicates that /proc/PID/maps parsing are truncated by time out. 580 */ 581 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12) 582 /* 583 * PERF_RECORD_MISC_MMAP_DATA and PERF_RECORD_MISC_COMM_EXEC are used on 584 * different events so can reuse the same bit position. 585 * Ditto PERF_RECORD_MISC_SWITCH_OUT. 586 */ 587 #define PERF_RECORD_MISC_MMAP_DATA (1 << 13) 588 #define PERF_RECORD_MISC_COMM_EXEC (1 << 13) 589 #define PERF_RECORD_MISC_SWITCH_OUT (1 << 13) 590 /* 591 * Indicates that the content of PERF_SAMPLE_IP points to 592 * the actual instruction that triggered the event. See also 593 * perf_event_attr::precise_ip. 594 */ 595 #define PERF_RECORD_MISC_EXACT_IP (1 << 14) 596 /* 597 * Reserve the last bit to indicate some extended misc field 598 */ 599 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 600 601 struct perf_event_header { 602 __u32 type; 603 __u16 misc; 604 __u16 size; 605 }; 606 607 enum perf_event_type { 608 609 /* 610 * If perf_event_attr.sample_id_all is set then all event types will 611 * have the sample_type selected fields related to where/when 612 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU, 613 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed 614 * just after the perf_event_header and the fields already present for 615 * the existing fields, i.e. at the end of the payload. That way a newer 616 * perf.data file will be supported by older perf tools, with these new 617 * optional fields being ignored. 618 * 619 * struct sample_id { 620 * { u32 pid, tid; } && PERF_SAMPLE_TID 621 * { u64 time; } && PERF_SAMPLE_TIME 622 * { u64 id; } && PERF_SAMPLE_ID 623 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 624 * { u32 cpu, res; } && PERF_SAMPLE_CPU 625 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 626 * } && perf_event_attr::sample_id_all 627 * 628 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. The 629 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed 630 * relative to header.size. 631 */ 632 633 /* 634 * The MMAP events record the PROT_EXEC mappings so that we can 635 * correlate userspace IPs to code. They have the following structure: 636 * 637 * struct { 638 * struct perf_event_header header; 639 * 640 * u32 pid, tid; 641 * u64 addr; 642 * u64 len; 643 * u64 pgoff; 644 * char filename[]; 645 * struct sample_id sample_id; 646 * }; 647 */ 648 PERF_RECORD_MMAP = 1, 649 650 /* 651 * struct { 652 * struct perf_event_header header; 653 * u64 id; 654 * u64 lost; 655 * struct sample_id sample_id; 656 * }; 657 */ 658 PERF_RECORD_LOST = 2, 659 660 /* 661 * struct { 662 * struct perf_event_header header; 663 * 664 * u32 pid, tid; 665 * char comm[]; 666 * struct sample_id sample_id; 667 * }; 668 */ 669 PERF_RECORD_COMM = 3, 670 671 /* 672 * struct { 673 * struct perf_event_header header; 674 * u32 pid, ppid; 675 * u32 tid, ptid; 676 * u64 time; 677 * struct sample_id sample_id; 678 * }; 679 */ 680 PERF_RECORD_EXIT = 4, 681 682 /* 683 * struct { 684 * struct perf_event_header header; 685 * u64 time; 686 * u64 id; 687 * u64 stream_id; 688 * struct sample_id sample_id; 689 * }; 690 */ 691 PERF_RECORD_THROTTLE = 5, 692 PERF_RECORD_UNTHROTTLE = 6, 693 694 /* 695 * struct { 696 * struct perf_event_header header; 697 * u32 pid, ppid; 698 * u32 tid, ptid; 699 * u64 time; 700 * struct sample_id sample_id; 701 * }; 702 */ 703 PERF_RECORD_FORK = 7, 704 705 /* 706 * struct { 707 * struct perf_event_header header; 708 * u32 pid, tid; 709 * 710 * struct read_format values; 711 * struct sample_id sample_id; 712 * }; 713 */ 714 PERF_RECORD_READ = 8, 715 716 /* 717 * struct { 718 * struct perf_event_header header; 719 * 720 * # 721 * # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. 722 * # The advantage of PERF_SAMPLE_IDENTIFIER is that its position 723 * # is fixed relative to header. 724 * # 725 * 726 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 727 * { u64 ip; } && PERF_SAMPLE_IP 728 * { u32 pid, tid; } && PERF_SAMPLE_TID 729 * { u64 time; } && PERF_SAMPLE_TIME 730 * { u64 addr; } && PERF_SAMPLE_ADDR 731 * { u64 id; } && PERF_SAMPLE_ID 732 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 733 * { u32 cpu, res; } && PERF_SAMPLE_CPU 734 * { u64 period; } && PERF_SAMPLE_PERIOD 735 * 736 * { struct read_format values; } && PERF_SAMPLE_READ 737 * 738 * { u64 nr, 739 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 740 * 741 * # 742 * # The RAW record below is opaque data wrt the ABI 743 * # 744 * # That is, the ABI doesn't make any promises wrt to 745 * # the stability of its content, it may vary depending 746 * # on event, hardware, kernel version and phase of 747 * # the moon. 748 * # 749 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 750 * # 751 * 752 * { u32 size; 753 * char data[size];}&& PERF_SAMPLE_RAW 754 * 755 * { u64 nr; 756 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK 757 * 758 * { u64 abi; # enum perf_sample_regs_abi 759 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER 760 * 761 * { u64 size; 762 * char data[size]; 763 * u64 dyn_size; } && PERF_SAMPLE_STACK_USER 764 * 765 * { u64 weight; } && PERF_SAMPLE_WEIGHT 766 * { u64 data_src; } && PERF_SAMPLE_DATA_SRC 767 * { u64 transaction; } && PERF_SAMPLE_TRANSACTION 768 * { u64 abi; # enum perf_sample_regs_abi 769 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR 770 * }; 771 */ 772 PERF_RECORD_SAMPLE = 9, 773 774 /* 775 * The MMAP2 records are an augmented version of MMAP, they add 776 * maj, min, ino numbers to be used to uniquely identify each mapping 777 * 778 * struct { 779 * struct perf_event_header header; 780 * 781 * u32 pid, tid; 782 * u64 addr; 783 * u64 len; 784 * u64 pgoff; 785 * u32 maj; 786 * u32 min; 787 * u64 ino; 788 * u64 ino_generation; 789 * u32 prot, flags; 790 * char filename[]; 791 * struct sample_id sample_id; 792 * }; 793 */ 794 PERF_RECORD_MMAP2 = 10, 795 796 /* 797 * Records that new data landed in the AUX buffer part. 798 * 799 * struct { 800 * struct perf_event_header header; 801 * 802 * u64 aux_offset; 803 * u64 aux_size; 804 * u64 flags; 805 * struct sample_id sample_id; 806 * }; 807 */ 808 PERF_RECORD_AUX = 11, 809 810 /* 811 * Indicates that instruction trace has started 812 * 813 * struct { 814 * struct perf_event_header header; 815 * u32 pid; 816 * u32 tid; 817 * }; 818 */ 819 PERF_RECORD_ITRACE_START = 12, 820 821 /* 822 * Records the dropped/lost sample number. 823 * 824 * struct { 825 * struct perf_event_header header; 826 * 827 * u64 lost; 828 * struct sample_id sample_id; 829 * }; 830 */ 831 PERF_RECORD_LOST_SAMPLES = 13, 832 833 /* 834 * Records a context switch in or out (flagged by 835 * PERF_RECORD_MISC_SWITCH_OUT). See also 836 * PERF_RECORD_SWITCH_CPU_WIDE. 837 * 838 * struct { 839 * struct perf_event_header header; 840 * struct sample_id sample_id; 841 * }; 842 */ 843 PERF_RECORD_SWITCH = 14, 844 845 /* 846 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and 847 * next_prev_tid that are the next (switching out) or previous 848 * (switching in) pid/tid. 849 * 850 * struct { 851 * struct perf_event_header header; 852 * u32 next_prev_pid; 853 * u32 next_prev_tid; 854 * struct sample_id sample_id; 855 * }; 856 */ 857 PERF_RECORD_SWITCH_CPU_WIDE = 15, 858 859 PERF_RECORD_MAX, /* non-ABI */ 860 }; 861 862 #define PERF_MAX_STACK_DEPTH 127 863 864 enum perf_callchain_context { 865 PERF_CONTEXT_HV = (__u64)-32, 866 PERF_CONTEXT_KERNEL = (__u64)-128, 867 PERF_CONTEXT_USER = (__u64)-512, 868 869 PERF_CONTEXT_GUEST = (__u64)-2048, 870 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 871 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 872 873 PERF_CONTEXT_MAX = (__u64)-4095, 874 }; 875 876 /** 877 * PERF_RECORD_AUX::flags bits 878 */ 879 #define PERF_AUX_FLAG_TRUNCATED 0x01 /* record was truncated to fit */ 880 #define PERF_AUX_FLAG_OVERWRITE 0x02 /* snapshot from overwrite mode */ 881 882 #define PERF_FLAG_FD_NO_GROUP (1UL << 0) 883 #define PERF_FLAG_FD_OUTPUT (1UL << 1) 884 #define PERF_FLAG_PID_CGROUP (1UL << 2) /* pid=cgroup id, per-cpu mode only */ 885 #define PERF_FLAG_FD_CLOEXEC (1UL << 3) /* O_CLOEXEC */ 886 887 union perf_mem_data_src { 888 __u64 val; 889 struct { 890 __u64 mem_op:5, /* type of opcode */ 891 mem_lvl:14, /* memory hierarchy level */ 892 mem_snoop:5, /* snoop mode */ 893 mem_lock:2, /* lock instr */ 894 mem_dtlb:7, /* tlb access */ 895 mem_rsvd:31; 896 }; 897 }; 898 899 /* type of opcode (load/store/prefetch,code) */ 900 #define PERF_MEM_OP_NA 0x01 /* not available */ 901 #define PERF_MEM_OP_LOAD 0x02 /* load instruction */ 902 #define PERF_MEM_OP_STORE 0x04 /* store instruction */ 903 #define PERF_MEM_OP_PFETCH 0x08 /* prefetch */ 904 #define PERF_MEM_OP_EXEC 0x10 /* code (execution) */ 905 #define PERF_MEM_OP_SHIFT 0 906 907 /* memory hierarchy (memory level, hit or miss) */ 908 #define PERF_MEM_LVL_NA 0x01 /* not available */ 909 #define PERF_MEM_LVL_HIT 0x02 /* hit level */ 910 #define PERF_MEM_LVL_MISS 0x04 /* miss level */ 911 #define PERF_MEM_LVL_L1 0x08 /* L1 */ 912 #define PERF_MEM_LVL_LFB 0x10 /* Line Fill Buffer */ 913 #define PERF_MEM_LVL_L2 0x20 /* L2 */ 914 #define PERF_MEM_LVL_L3 0x40 /* L3 */ 915 #define PERF_MEM_LVL_LOC_RAM 0x80 /* Local DRAM */ 916 #define PERF_MEM_LVL_REM_RAM1 0x100 /* Remote DRAM (1 hop) */ 917 #define PERF_MEM_LVL_REM_RAM2 0x200 /* Remote DRAM (2 hops) */ 918 #define PERF_MEM_LVL_REM_CCE1 0x400 /* Remote Cache (1 hop) */ 919 #define PERF_MEM_LVL_REM_CCE2 0x800 /* Remote Cache (2 hops) */ 920 #define PERF_MEM_LVL_IO 0x1000 /* I/O memory */ 921 #define PERF_MEM_LVL_UNC 0x2000 /* Uncached memory */ 922 #define PERF_MEM_LVL_SHIFT 5 923 924 /* snoop mode */ 925 #define PERF_MEM_SNOOP_NA 0x01 /* not available */ 926 #define PERF_MEM_SNOOP_NONE 0x02 /* no snoop */ 927 #define PERF_MEM_SNOOP_HIT 0x04 /* snoop hit */ 928 #define PERF_MEM_SNOOP_MISS 0x08 /* snoop miss */ 929 #define PERF_MEM_SNOOP_HITM 0x10 /* snoop hit modified */ 930 #define PERF_MEM_SNOOP_SHIFT 19 931 932 /* locked instruction */ 933 #define PERF_MEM_LOCK_NA 0x01 /* not available */ 934 #define PERF_MEM_LOCK_LOCKED 0x02 /* locked transaction */ 935 #define PERF_MEM_LOCK_SHIFT 24 936 937 /* TLB access */ 938 #define PERF_MEM_TLB_NA 0x01 /* not available */ 939 #define PERF_MEM_TLB_HIT 0x02 /* hit level */ 940 #define PERF_MEM_TLB_MISS 0x04 /* miss level */ 941 #define PERF_MEM_TLB_L1 0x08 /* L1 */ 942 #define PERF_MEM_TLB_L2 0x10 /* L2 */ 943 #define PERF_MEM_TLB_WK 0x20 /* Hardware Walker*/ 944 #define PERF_MEM_TLB_OS 0x40 /* OS fault handler */ 945 #define PERF_MEM_TLB_SHIFT 26 946 947 #define PERF_MEM_S(a, s) \ 948 (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT) 949 950 /* 951 * single taken branch record layout: 952 * 953 * from: source instruction (may not always be a branch insn) 954 * to: branch target 955 * mispred: branch target was mispredicted 956 * predicted: branch target was predicted 957 * 958 * support for mispred, predicted is optional. In case it 959 * is not supported mispred = predicted = 0. 960 * 961 * in_tx: running in a hardware transaction 962 * abort: aborting a hardware transaction 963 * cycles: cycles from last branch (or 0 if not supported) 964 */ 965 struct perf_branch_entry { 966 __u64 from; 967 __u64 to; 968 __u64 mispred:1, /* target mispredicted */ 969 predicted:1,/* target predicted */ 970 in_tx:1, /* in transaction */ 971 abort:1, /* transaction abort */ 972 cycles:16, /* cycle count to last branch */ 973 reserved:44; 974 }; 975 976 #endif /* _UAPI_LINUX_PERF_EVENT_H */ 977