perf_event.h (43707B) - Raw
1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */ 2 /* 3 * Performance events: 4 * 5 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de> 6 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar 7 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra 8 * 9 * Data type definitions, declarations, prototypes. 10 * 11 * Started by: Thomas Gleixner and Ingo Molnar 12 * 13 * For licencing details see kernel-base/COPYING 14 */ 15 #ifndef _LINUX_PERF_EVENT_H 16 #define _LINUX_PERF_EVENT_H 17 18 #include <linux/types.h> 19 #include <linux/ioctl.h> 20 #include <asm/byteorder.h> 21 22 /* 23 * User-space ABI bits: 24 */ 25 26 /* 27 * attr.type 28 */ 29 enum perf_type_id { 30 PERF_TYPE_HARDWARE = 0, 31 PERF_TYPE_SOFTWARE = 1, 32 PERF_TYPE_TRACEPOINT = 2, 33 PERF_TYPE_HW_CACHE = 3, 34 PERF_TYPE_RAW = 4, 35 PERF_TYPE_BREAKPOINT = 5, 36 37 PERF_TYPE_MAX, /* non-ABI */ 38 }; 39 40 /* 41 * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE 42 * PERF_TYPE_HARDWARE: 0xEEEEEEEE000000AA 43 * AA: hardware event ID 44 * EEEEEEEE: PMU type ID 45 * PERF_TYPE_HW_CACHE: 0xEEEEEEEE00DDCCBB 46 * BB: hardware cache ID 47 * CC: hardware cache op ID 48 * DD: hardware cache op result ID 49 * EEEEEEEE: PMU type ID 50 * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied. 51 */ 52 #define PERF_PMU_TYPE_SHIFT 32 53 #define PERF_HW_EVENT_MASK 0xffffffff 54 55 /* 56 * Generalized performance event event_id types, used by the 57 * attr.event_id parameter of the sys_perf_event_open() 58 * syscall: 59 */ 60 enum perf_hw_id { 61 /* 62 * Common hardware events, generalized by the kernel: 63 */ 64 PERF_COUNT_HW_CPU_CYCLES = 0, 65 PERF_COUNT_HW_INSTRUCTIONS = 1, 66 PERF_COUNT_HW_CACHE_REFERENCES = 2, 67 PERF_COUNT_HW_CACHE_MISSES = 3, 68 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, 69 PERF_COUNT_HW_BRANCH_MISSES = 5, 70 PERF_COUNT_HW_BUS_CYCLES = 6, 71 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND = 7, 72 PERF_COUNT_HW_STALLED_CYCLES_BACKEND = 8, 73 PERF_COUNT_HW_REF_CPU_CYCLES = 9, 74 75 PERF_COUNT_HW_MAX, /* non-ABI */ 76 }; 77 78 /* 79 * Generalized hardware cache events: 80 * 81 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x 82 * { read, write, prefetch } x 83 * { accesses, misses } 84 */ 85 enum perf_hw_cache_id { 86 PERF_COUNT_HW_CACHE_L1D = 0, 87 PERF_COUNT_HW_CACHE_L1I = 1, 88 PERF_COUNT_HW_CACHE_LL = 2, 89 PERF_COUNT_HW_CACHE_DTLB = 3, 90 PERF_COUNT_HW_CACHE_ITLB = 4, 91 PERF_COUNT_HW_CACHE_BPU = 5, 92 PERF_COUNT_HW_CACHE_NODE = 6, 93 94 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 95 }; 96 97 enum perf_hw_cache_op_id { 98 PERF_COUNT_HW_CACHE_OP_READ = 0, 99 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 100 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 101 102 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 103 }; 104 105 enum perf_hw_cache_op_result_id { 106 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 107 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 108 109 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 110 }; 111 112 /* 113 * Special "software" events provided by the kernel, even if the hardware 114 * does not support performance events. These events measure various 115 * physical and sw events of the kernel (and allow the profiling of them as 116 * well): 117 */ 118 enum perf_sw_ids { 119 PERF_COUNT_SW_CPU_CLOCK = 0, 120 PERF_COUNT_SW_TASK_CLOCK = 1, 121 PERF_COUNT_SW_PAGE_FAULTS = 2, 122 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 123 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 124 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 125 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 126 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7, 127 PERF_COUNT_SW_EMULATION_FAULTS = 8, 128 PERF_COUNT_SW_DUMMY = 9, 129 PERF_COUNT_SW_BPF_OUTPUT = 10, 130 PERF_COUNT_SW_CGROUP_SWITCHES = 11, 131 132 PERF_COUNT_SW_MAX, /* non-ABI */ 133 }; 134 135 /* 136 * Bits that can be set in attr.sample_type to request information 137 * in the overflow packets. 138 */ 139 enum perf_event_sample_format { 140 PERF_SAMPLE_IP = 1U << 0, 141 PERF_SAMPLE_TID = 1U << 1, 142 PERF_SAMPLE_TIME = 1U << 2, 143 PERF_SAMPLE_ADDR = 1U << 3, 144 PERF_SAMPLE_READ = 1U << 4, 145 PERF_SAMPLE_CALLCHAIN = 1U << 5, 146 PERF_SAMPLE_ID = 1U << 6, 147 PERF_SAMPLE_CPU = 1U << 7, 148 PERF_SAMPLE_PERIOD = 1U << 8, 149 PERF_SAMPLE_STREAM_ID = 1U << 9, 150 PERF_SAMPLE_RAW = 1U << 10, 151 PERF_SAMPLE_BRANCH_STACK = 1U << 11, 152 PERF_SAMPLE_REGS_USER = 1U << 12, 153 PERF_SAMPLE_STACK_USER = 1U << 13, 154 PERF_SAMPLE_WEIGHT = 1U << 14, 155 PERF_SAMPLE_DATA_SRC = 1U << 15, 156 PERF_SAMPLE_IDENTIFIER = 1U << 16, 157 PERF_SAMPLE_TRANSACTION = 1U << 17, 158 PERF_SAMPLE_REGS_INTR = 1U << 18, 159 PERF_SAMPLE_PHYS_ADDR = 1U << 19, 160 PERF_SAMPLE_AUX = 1U << 20, 161 PERF_SAMPLE_CGROUP = 1U << 21, 162 PERF_SAMPLE_DATA_PAGE_SIZE = 1U << 22, 163 PERF_SAMPLE_CODE_PAGE_SIZE = 1U << 23, 164 PERF_SAMPLE_WEIGHT_STRUCT = 1U << 24, 165 166 PERF_SAMPLE_MAX = 1U << 25, /* non-ABI */ 167 }; 168 169 #define PERF_SAMPLE_WEIGHT_TYPE (PERF_SAMPLE_WEIGHT | PERF_SAMPLE_WEIGHT_STRUCT) 170 /* 171 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set 172 * 173 * If the user does not pass priv level information via branch_sample_type, 174 * the kernel uses the event's priv level. Branch and event priv levels do 175 * not have to match. Branch priv level is checked for permissions. 176 * 177 * The branch types can be combined, however BRANCH_ANY covers all types 178 * of branches and therefore it supersedes all the other types. 179 */ 180 enum perf_branch_sample_type_shift { 181 PERF_SAMPLE_BRANCH_USER_SHIFT = 0, /* user branches */ 182 PERF_SAMPLE_BRANCH_KERNEL_SHIFT = 1, /* kernel branches */ 183 PERF_SAMPLE_BRANCH_HV_SHIFT = 2, /* hypervisor branches */ 184 185 PERF_SAMPLE_BRANCH_ANY_SHIFT = 3, /* any branch types */ 186 PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT = 4, /* any call branch */ 187 PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT = 5, /* any return branch */ 188 PERF_SAMPLE_BRANCH_IND_CALL_SHIFT = 6, /* indirect calls */ 189 PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT = 7, /* transaction aborts */ 190 PERF_SAMPLE_BRANCH_IN_TX_SHIFT = 8, /* in transaction */ 191 PERF_SAMPLE_BRANCH_NO_TX_SHIFT = 9, /* not in transaction */ 192 PERF_SAMPLE_BRANCH_COND_SHIFT = 10, /* conditional branches */ 193 194 PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT = 11, /* call/ret stack */ 195 PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT = 12, /* indirect jumps */ 196 PERF_SAMPLE_BRANCH_CALL_SHIFT = 13, /* direct call */ 197 198 PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT = 14, /* no flags */ 199 PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT = 15, /* no cycles */ 200 201 PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT = 16, /* save branch type */ 202 203 PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT = 17, /* save low level index of raw branch records */ 204 205 PERF_SAMPLE_BRANCH_PRIV_SAVE_SHIFT = 18, /* save privilege mode */ 206 207 PERF_SAMPLE_BRANCH_COUNTERS_SHIFT = 19, /* save occurrences of events on a branch */ 208 209 PERF_SAMPLE_BRANCH_MAX_SHIFT /* non-ABI */ 210 }; 211 212 enum perf_branch_sample_type { 213 PERF_SAMPLE_BRANCH_USER = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT, 214 PERF_SAMPLE_BRANCH_KERNEL = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT, 215 PERF_SAMPLE_BRANCH_HV = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT, 216 217 PERF_SAMPLE_BRANCH_ANY = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT, 218 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT, 219 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT, 220 PERF_SAMPLE_BRANCH_IND_CALL = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT, 221 PERF_SAMPLE_BRANCH_ABORT_TX = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT, 222 PERF_SAMPLE_BRANCH_IN_TX = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT, 223 PERF_SAMPLE_BRANCH_NO_TX = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT, 224 PERF_SAMPLE_BRANCH_COND = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT, 225 226 PERF_SAMPLE_BRANCH_CALL_STACK = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT, 227 PERF_SAMPLE_BRANCH_IND_JUMP = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT, 228 PERF_SAMPLE_BRANCH_CALL = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT, 229 230 PERF_SAMPLE_BRANCH_NO_FLAGS = 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT, 231 PERF_SAMPLE_BRANCH_NO_CYCLES = 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT, 232 233 PERF_SAMPLE_BRANCH_TYPE_SAVE = 234 1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT, 235 236 PERF_SAMPLE_BRANCH_HW_INDEX = 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT, 237 238 PERF_SAMPLE_BRANCH_PRIV_SAVE = 1U << PERF_SAMPLE_BRANCH_PRIV_SAVE_SHIFT, 239 240 PERF_SAMPLE_BRANCH_COUNTERS = 1U << PERF_SAMPLE_BRANCH_COUNTERS_SHIFT, 241 242 PERF_SAMPLE_BRANCH_MAX = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT, 243 }; 244 245 /* 246 * Common flow change classification 247 */ 248 enum { 249 PERF_BR_UNKNOWN = 0, /* unknown */ 250 PERF_BR_COND = 1, /* conditional */ 251 PERF_BR_UNCOND = 2, /* unconditional */ 252 PERF_BR_IND = 3, /* indirect */ 253 PERF_BR_CALL = 4, /* function call */ 254 PERF_BR_IND_CALL = 5, /* indirect function call */ 255 PERF_BR_RET = 6, /* function return */ 256 PERF_BR_SYSCALL = 7, /* syscall */ 257 PERF_BR_SYSRET = 8, /* syscall return */ 258 PERF_BR_COND_CALL = 9, /* conditional function call */ 259 PERF_BR_COND_RET = 10, /* conditional function return */ 260 PERF_BR_ERET = 11, /* exception return */ 261 PERF_BR_IRQ = 12, /* irq */ 262 PERF_BR_SERROR = 13, /* system error */ 263 PERF_BR_NO_TX = 14, /* not in transaction */ 264 PERF_BR_EXTEND_ABI = 15, /* extend ABI */ 265 PERF_BR_MAX, 266 }; 267 268 /* 269 * Common branch speculation outcome classification 270 */ 271 enum { 272 PERF_BR_SPEC_NA = 0, /* Not available */ 273 PERF_BR_SPEC_WRONG_PATH = 1, /* Speculative but on wrong path */ 274 PERF_BR_NON_SPEC_CORRECT_PATH = 2, /* Non-speculative but on correct path */ 275 PERF_BR_SPEC_CORRECT_PATH = 3, /* Speculative and on correct path */ 276 PERF_BR_SPEC_MAX, 277 }; 278 279 enum { 280 PERF_BR_NEW_FAULT_ALGN = 0, /* Alignment fault */ 281 PERF_BR_NEW_FAULT_DATA = 1, /* Data fault */ 282 PERF_BR_NEW_FAULT_INST = 2, /* Inst fault */ 283 PERF_BR_NEW_ARCH_1 = 3, /* Architecture specific */ 284 PERF_BR_NEW_ARCH_2 = 4, /* Architecture specific */ 285 PERF_BR_NEW_ARCH_3 = 5, /* Architecture specific */ 286 PERF_BR_NEW_ARCH_4 = 6, /* Architecture specific */ 287 PERF_BR_NEW_ARCH_5 = 7, /* Architecture specific */ 288 PERF_BR_NEW_MAX, 289 }; 290 291 enum { 292 PERF_BR_PRIV_UNKNOWN = 0, 293 PERF_BR_PRIV_USER = 1, 294 PERF_BR_PRIV_KERNEL = 2, 295 PERF_BR_PRIV_HV = 3, 296 }; 297 298 #define PERF_BR_ARM64_FIQ PERF_BR_NEW_ARCH_1 299 #define PERF_BR_ARM64_DEBUG_HALT PERF_BR_NEW_ARCH_2 300 #define PERF_BR_ARM64_DEBUG_EXIT PERF_BR_NEW_ARCH_3 301 #define PERF_BR_ARM64_DEBUG_INST PERF_BR_NEW_ARCH_4 302 #define PERF_BR_ARM64_DEBUG_DATA PERF_BR_NEW_ARCH_5 303 304 #define PERF_SAMPLE_BRANCH_PLM_ALL \ 305 (PERF_SAMPLE_BRANCH_USER|\ 306 PERF_SAMPLE_BRANCH_KERNEL|\ 307 PERF_SAMPLE_BRANCH_HV) 308 309 /* 310 * Values to determine ABI of the registers dump. 311 */ 312 enum perf_sample_regs_abi { 313 PERF_SAMPLE_REGS_ABI_NONE = 0, 314 PERF_SAMPLE_REGS_ABI_32 = 1, 315 PERF_SAMPLE_REGS_ABI_64 = 2, 316 }; 317 318 /* 319 * Values for the memory transaction event qualifier, mostly for 320 * abort events. Multiple bits can be set. 321 */ 322 enum { 323 PERF_TXN_ELISION = (1 << 0), /* From elision */ 324 PERF_TXN_TRANSACTION = (1 << 1), /* From transaction */ 325 PERF_TXN_SYNC = (1 << 2), /* Instruction is related */ 326 PERF_TXN_ASYNC = (1 << 3), /* Instruction not related */ 327 PERF_TXN_RETRY = (1 << 4), /* Retry possible */ 328 PERF_TXN_CONFLICT = (1 << 5), /* Conflict abort */ 329 PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */ 330 PERF_TXN_CAPACITY_READ = (1 << 7), /* Capacity read abort */ 331 332 PERF_TXN_MAX = (1 << 8), /* non-ABI */ 333 334 /* bits 32..63 are reserved for the abort code */ 335 336 PERF_TXN_ABORT_MASK = (0xffffffffULL << 32), 337 PERF_TXN_ABORT_SHIFT = 32, 338 }; 339 340 /* 341 * The format of the data returned by read() on a perf event fd, 342 * as specified by attr.read_format: 343 * 344 * struct read_format { 345 * { u64 value; 346 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 347 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 348 * { u64 id; } && PERF_FORMAT_ID 349 * { u64 lost; } && PERF_FORMAT_LOST 350 * } && !PERF_FORMAT_GROUP 351 * 352 * { u64 nr; 353 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 354 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 355 * { u64 value; 356 * { u64 id; } && PERF_FORMAT_ID 357 * { u64 lost; } && PERF_FORMAT_LOST 358 * } cntr[nr]; 359 * } && PERF_FORMAT_GROUP 360 * }; 361 */ 362 enum perf_event_read_format { 363 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 364 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 365 PERF_FORMAT_ID = 1U << 2, 366 PERF_FORMAT_GROUP = 1U << 3, 367 PERF_FORMAT_LOST = 1U << 4, 368 369 PERF_FORMAT_MAX = 1U << 5, /* non-ABI */ 370 }; 371 372 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 373 #define PERF_ATTR_SIZE_VER1 72 /* add: config2 */ 374 #define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */ 375 #define PERF_ATTR_SIZE_VER3 96 /* add: sample_regs_user */ 376 /* add: sample_stack_user */ 377 #define PERF_ATTR_SIZE_VER4 104 /* add: sample_regs_intr */ 378 #define PERF_ATTR_SIZE_VER5 112 /* add: aux_watermark */ 379 #define PERF_ATTR_SIZE_VER6 120 /* add: aux_sample_size */ 380 #define PERF_ATTR_SIZE_VER7 128 /* add: sig_data */ 381 #define PERF_ATTR_SIZE_VER8 136 /* add: config3 */ 382 383 /* 384 * Hardware event_id to monitor via a performance monitoring event: 385 * 386 * @sample_max_stack: Max number of frame pointers in a callchain, 387 * should be < /proc/sys/kernel/perf_event_max_stack 388 */ 389 struct perf_event_attr { 390 391 /* 392 * Major type: hardware/software/tracepoint/etc. 393 */ 394 __u32 type; 395 396 /* 397 * Size of the attr structure, for fwd/bwd compat. 398 */ 399 __u32 size; 400 401 /* 402 * Type specific configuration information. 403 */ 404 __u64 config; 405 406 union { 407 __u64 sample_period; 408 __u64 sample_freq; 409 }; 410 411 __u64 sample_type; 412 __u64 read_format; 413 414 __u64 disabled : 1, /* off by default */ 415 inherit : 1, /* children inherit it */ 416 pinned : 1, /* must always be on PMU */ 417 exclusive : 1, /* only group on PMU */ 418 exclude_user : 1, /* don't count user */ 419 exclude_kernel : 1, /* ditto kernel */ 420 exclude_hv : 1, /* ditto hypervisor */ 421 exclude_idle : 1, /* don't count when idle */ 422 mmap : 1, /* include mmap data */ 423 comm : 1, /* include comm data */ 424 freq : 1, /* use freq, not period */ 425 inherit_stat : 1, /* per task counts */ 426 enable_on_exec : 1, /* next exec enables */ 427 task : 1, /* trace fork/exit */ 428 watermark : 1, /* wakeup_watermark */ 429 /* 430 * precise_ip: 431 * 432 * 0 - SAMPLE_IP can have arbitrary skid 433 * 1 - SAMPLE_IP must have constant skid 434 * 2 - SAMPLE_IP requested to have 0 skid 435 * 3 - SAMPLE_IP must have 0 skid 436 * 437 * See also PERF_RECORD_MISC_EXACT_IP 438 */ 439 precise_ip : 2, /* skid constraint */ 440 mmap_data : 1, /* non-exec mmap data */ 441 sample_id_all : 1, /* sample_type all events */ 442 443 exclude_host : 1, /* don't count in host */ 444 exclude_guest : 1, /* don't count in guest */ 445 446 exclude_callchain_kernel : 1, /* exclude kernel callchains */ 447 exclude_callchain_user : 1, /* exclude user callchains */ 448 mmap2 : 1, /* include mmap with inode data */ 449 comm_exec : 1, /* flag comm events that are due to an exec */ 450 use_clockid : 1, /* use @clockid for time fields */ 451 context_switch : 1, /* context switch data */ 452 write_backward : 1, /* Write ring buffer from end to beginning */ 453 namespaces : 1, /* include namespaces data */ 454 ksymbol : 1, /* include ksymbol events */ 455 bpf_event : 1, /* include bpf events */ 456 aux_output : 1, /* generate AUX records instead of events */ 457 cgroup : 1, /* include cgroup events */ 458 text_poke : 1, /* include text poke events */ 459 build_id : 1, /* use build id in mmap2 events */ 460 inherit_thread : 1, /* children only inherit if cloned with CLONE_THREAD */ 461 remove_on_exec : 1, /* event is removed from task on exec */ 462 sigtrap : 1, /* send synchronous SIGTRAP on event */ 463 __reserved_1 : 26; 464 465 union { 466 __u32 wakeup_events; /* wakeup every n events */ 467 __u32 wakeup_watermark; /* bytes before wakeup */ 468 }; 469 470 __u32 bp_type; 471 union { 472 __u64 bp_addr; 473 __u64 kprobe_func; /* for perf_kprobe */ 474 __u64 uprobe_path; /* for perf_uprobe */ 475 __u64 config1; /* extension of config */ 476 }; 477 union { 478 __u64 bp_len; 479 __u64 kprobe_addr; /* when kprobe_func == NULL */ 480 __u64 probe_offset; /* for perf_[k,u]probe */ 481 __u64 config2; /* extension of config1 */ 482 }; 483 __u64 branch_sample_type; /* enum perf_branch_sample_type */ 484 485 /* 486 * Defines set of user regs to dump on samples. 487 * See asm/perf_regs.h for details. 488 */ 489 __u64 sample_regs_user; 490 491 /* 492 * Defines size of the user stack to dump on samples. 493 */ 494 __u32 sample_stack_user; 495 496 __s32 clockid; 497 /* 498 * Defines set of regs to dump for each sample 499 * state captured on: 500 * - precise = 0: PMU interrupt 501 * - precise > 0: sampled instruction 502 * 503 * See asm/perf_regs.h for details. 504 */ 505 __u64 sample_regs_intr; 506 507 /* 508 * Wakeup watermark for AUX area 509 */ 510 __u32 aux_watermark; 511 __u16 sample_max_stack; 512 __u16 __reserved_2; 513 __u32 aux_sample_size; 514 515 union { 516 __u32 aux_action; 517 struct { 518 __u32 aux_start_paused : 1, /* start AUX area tracing paused */ 519 aux_pause : 1, /* on overflow, pause AUX area tracing */ 520 aux_resume : 1, /* on overflow, resume AUX area tracing */ 521 __reserved_3 : 29; 522 }; 523 }; 524 525 /* 526 * User provided data if sigtrap=1, passed back to user via 527 * siginfo_t::si_perf_data, e.g. to permit user to identify the event. 528 * Note, siginfo_t::si_perf_data is long-sized, and sig_data will be 529 * truncated accordingly on 32 bit architectures. 530 */ 531 __u64 sig_data; 532 533 __u64 config3; /* extension of config2 */ 534 }; 535 536 /* 537 * Structure used by below PERF_EVENT_IOC_QUERY_BPF command 538 * to query bpf programs attached to the same perf tracepoint 539 * as the given perf event. 540 */ 541 struct perf_event_query_bpf { 542 /* 543 * The below ids array length 544 */ 545 __u32 ids_len; 546 /* 547 * Set by the kernel to indicate the number of 548 * available programs 549 */ 550 __u32 prog_cnt; 551 /* 552 * User provided buffer to store program ids 553 */ 554 __u32 ids[]; 555 }; 556 557 /* 558 * Ioctls that can be done on a perf event fd: 559 */ 560 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 561 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 562 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 563 #define PERF_EVENT_IOC_RESET _IO ('$', 3) 564 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 565 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 566 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 567 #define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *) 568 #define PERF_EVENT_IOC_SET_BPF _IOW('$', 8, __u32) 569 #define PERF_EVENT_IOC_PAUSE_OUTPUT _IOW('$', 9, __u32) 570 #define PERF_EVENT_IOC_QUERY_BPF _IOWR('$', 10, struct perf_event_query_bpf *) 571 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES _IOW('$', 11, struct perf_event_attr *) 572 573 enum perf_event_ioc_flags { 574 PERF_IOC_FLAG_GROUP = 1U << 0, 575 }; 576 577 /* 578 * Structure of the page that can be mapped via mmap 579 */ 580 struct perf_event_mmap_page { 581 __u32 version; /* version number of this structure */ 582 __u32 compat_version; /* lowest version this is compat with */ 583 584 /* 585 * Bits needed to read the hw events in user-space. 586 * 587 * u32 seq, time_mult, time_shift, index, width; 588 * u64 count, enabled, running; 589 * u64 cyc, time_offset; 590 * s64 pmc = 0; 591 * 592 * do { 593 * seq = pc->lock; 594 * barrier() 595 * 596 * enabled = pc->time_enabled; 597 * running = pc->time_running; 598 * 599 * if (pc->cap_usr_time && enabled != running) { 600 * cyc = rdtsc(); 601 * time_offset = pc->time_offset; 602 * time_mult = pc->time_mult; 603 * time_shift = pc->time_shift; 604 * } 605 * 606 * index = pc->index; 607 * count = pc->offset; 608 * if (pc->cap_user_rdpmc && index) { 609 * width = pc->pmc_width; 610 * pmc = rdpmc(index - 1); 611 * } 612 * 613 * barrier(); 614 * } while (pc->lock != seq); 615 * 616 * NOTE: for obvious reason this only works on self-monitoring 617 * processes. 618 */ 619 __u32 lock; /* seqlock for synchronization */ 620 __u32 index; /* hardware event identifier */ 621 __s64 offset; /* add to hardware event value */ 622 __u64 time_enabled; /* time event active */ 623 __u64 time_running; /* time event on cpu */ 624 union { 625 __u64 capabilities; 626 struct { 627 __u64 cap_bit0 : 1, /* Always 0, deprecated, see commit 860f085b74e9 */ 628 cap_bit0_is_deprecated : 1, /* Always 1, signals that bit 0 is zero */ 629 630 cap_user_rdpmc : 1, /* The RDPMC instruction can be used to read counts */ 631 cap_user_time : 1, /* The time_{shift,mult,offset} fields are used */ 632 cap_user_time_zero : 1, /* The time_zero field is used */ 633 cap_user_time_short : 1, /* the time_{cycle,mask} fields are used */ 634 cap_____res : 58; 635 }; 636 }; 637 638 /* 639 * If cap_user_rdpmc this field provides the bit-width of the value 640 * read using the rdpmc() or equivalent instruction. This can be used 641 * to sign extend the result like: 642 * 643 * pmc <<= 64 - width; 644 * pmc >>= 64 - width; // signed shift right 645 * count += pmc; 646 */ 647 __u16 pmc_width; 648 649 /* 650 * If cap_usr_time the below fields can be used to compute the time 651 * delta since time_enabled (in ns) using rdtsc or similar. 652 * 653 * u64 quot, rem; 654 * u64 delta; 655 * 656 * quot = (cyc >> time_shift); 657 * rem = cyc & (((u64)1 << time_shift) - 1); 658 * delta = time_offset + quot * time_mult + 659 * ((rem * time_mult) >> time_shift); 660 * 661 * Where time_offset,time_mult,time_shift and cyc are read in the 662 * seqcount loop described above. This delta can then be added to 663 * enabled and possible running (if index), improving the scaling: 664 * 665 * enabled += delta; 666 * if (index) 667 * running += delta; 668 * 669 * quot = count / running; 670 * rem = count % running; 671 * count = quot * enabled + (rem * enabled) / running; 672 */ 673 __u16 time_shift; 674 __u32 time_mult; 675 __u64 time_offset; 676 /* 677 * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated 678 * from sample timestamps. 679 * 680 * time = timestamp - time_zero; 681 * quot = time / time_mult; 682 * rem = time % time_mult; 683 * cyc = (quot << time_shift) + (rem << time_shift) / time_mult; 684 * 685 * And vice versa: 686 * 687 * quot = cyc >> time_shift; 688 * rem = cyc & (((u64)1 << time_shift) - 1); 689 * timestamp = time_zero + quot * time_mult + 690 * ((rem * time_mult) >> time_shift); 691 */ 692 __u64 time_zero; 693 694 __u32 size; /* Header size up to __reserved[] fields. */ 695 __u32 __reserved_1; 696 697 /* 698 * If cap_usr_time_short, the hardware clock is less than 64bit wide 699 * and we must compute the 'cyc' value, as used by cap_usr_time, as: 700 * 701 * cyc = time_cycles + ((cyc - time_cycles) & time_mask) 702 * 703 * NOTE: this form is explicitly chosen such that cap_usr_time_short 704 * is a correction on top of cap_usr_time, and code that doesn't 705 * know about cap_usr_time_short still works under the assumption 706 * the counter doesn't wrap. 707 */ 708 __u64 time_cycles; 709 __u64 time_mask; 710 711 /* 712 * Hole for extension of the self monitor capabilities 713 */ 714 715 __u8 __reserved[116*8]; /* align to 1k. */ 716 717 /* 718 * Control data for the mmap() data buffer. 719 * 720 * User-space reading the @data_head value should issue an smp_rmb(), 721 * after reading this value. 722 * 723 * When the mapping is PROT_WRITE the @data_tail value should be 724 * written by userspace to reflect the last read data, after issueing 725 * an smp_mb() to separate the data read from the ->data_tail store. 726 * In this case the kernel will not over-write unread data. 727 * 728 * See perf_output_put_handle() for the data ordering. 729 * 730 * data_{offset,size} indicate the location and size of the perf record 731 * buffer within the mmapped area. 732 */ 733 __u64 data_head; /* head in the data section */ 734 __u64 data_tail; /* user-space written tail */ 735 __u64 data_offset; /* where the buffer starts */ 736 __u64 data_size; /* data buffer size */ 737 738 /* 739 * AUX area is defined by aux_{offset,size} fields that should be set 740 * by the userspace, so that 741 * 742 * aux_offset >= data_offset + data_size 743 * 744 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size. 745 * 746 * Ring buffer pointers aux_{head,tail} have the same semantics as 747 * data_{head,tail} and same ordering rules apply. 748 */ 749 __u64 aux_head; 750 __u64 aux_tail; 751 __u64 aux_offset; 752 __u64 aux_size; 753 }; 754 755 /* 756 * The current state of perf_event_header::misc bits usage: 757 * ('|' used bit, '-' unused bit) 758 * 759 * 012 CDEF 760 * |||---------|||| 761 * 762 * Where: 763 * 0-2 CPUMODE_MASK 764 * 765 * C PROC_MAP_PARSE_TIMEOUT 766 * D MMAP_DATA / COMM_EXEC / FORK_EXEC / SWITCH_OUT 767 * E MMAP_BUILD_ID / EXACT_IP / SCHED_OUT_PREEMPT 768 * F (reserved) 769 */ 770 771 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 772 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 773 #define PERF_RECORD_MISC_KERNEL (1 << 0) 774 #define PERF_RECORD_MISC_USER (2 << 0) 775 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 776 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 777 #define PERF_RECORD_MISC_GUEST_USER (5 << 0) 778 779 /* 780 * Indicates that /proc/PID/maps parsing are truncated by time out. 781 */ 782 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12) 783 /* 784 * Following PERF_RECORD_MISC_* are used on different 785 * events, so can reuse the same bit position: 786 * 787 * PERF_RECORD_MISC_MMAP_DATA - PERF_RECORD_MMAP* events 788 * PERF_RECORD_MISC_COMM_EXEC - PERF_RECORD_COMM event 789 * PERF_RECORD_MISC_FORK_EXEC - PERF_RECORD_FORK event (perf internal) 790 * PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events 791 */ 792 #define PERF_RECORD_MISC_MMAP_DATA (1 << 13) 793 #define PERF_RECORD_MISC_COMM_EXEC (1 << 13) 794 #define PERF_RECORD_MISC_FORK_EXEC (1 << 13) 795 #define PERF_RECORD_MISC_SWITCH_OUT (1 << 13) 796 /* 797 * These PERF_RECORD_MISC_* flags below are safely reused 798 * for the following events: 799 * 800 * PERF_RECORD_MISC_EXACT_IP - PERF_RECORD_SAMPLE of precise events 801 * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events 802 * PERF_RECORD_MISC_MMAP_BUILD_ID - PERF_RECORD_MMAP2 event 803 * 804 * 805 * PERF_RECORD_MISC_EXACT_IP: 806 * Indicates that the content of PERF_SAMPLE_IP points to 807 * the actual instruction that triggered the event. See also 808 * perf_event_attr::precise_ip. 809 * 810 * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT: 811 * Indicates that thread was preempted in TASK_RUNNING state. 812 * 813 * PERF_RECORD_MISC_MMAP_BUILD_ID: 814 * Indicates that mmap2 event carries build id data. 815 */ 816 #define PERF_RECORD_MISC_EXACT_IP (1 << 14) 817 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT (1 << 14) 818 #define PERF_RECORD_MISC_MMAP_BUILD_ID (1 << 14) 819 /* 820 * Reserve the last bit to indicate some extended misc field 821 */ 822 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 823 824 struct perf_event_header { 825 __u32 type; 826 __u16 misc; 827 __u16 size; 828 }; 829 830 struct perf_ns_link_info { 831 __u64 dev; 832 __u64 ino; 833 }; 834 835 enum { 836 NET_NS_INDEX = 0, 837 UTS_NS_INDEX = 1, 838 IPC_NS_INDEX = 2, 839 PID_NS_INDEX = 3, 840 USER_NS_INDEX = 4, 841 MNT_NS_INDEX = 5, 842 CGROUP_NS_INDEX = 6, 843 844 NR_NAMESPACES, /* number of available namespaces */ 845 }; 846 847 enum perf_event_type { 848 849 /* 850 * If perf_event_attr.sample_id_all is set then all event types will 851 * have the sample_type selected fields related to where/when 852 * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU, 853 * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed 854 * just after the perf_event_header and the fields already present for 855 * the existing fields, i.e. at the end of the payload. That way a newer 856 * perf.data file will be supported by older perf tools, with these new 857 * optional fields being ignored. 858 * 859 * struct sample_id { 860 * { u32 pid, tid; } && PERF_SAMPLE_TID 861 * { u64 time; } && PERF_SAMPLE_TIME 862 * { u64 id; } && PERF_SAMPLE_ID 863 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 864 * { u32 cpu, res; } && PERF_SAMPLE_CPU 865 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 866 * } && perf_event_attr::sample_id_all 867 * 868 * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. The 869 * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed 870 * relative to header.size. 871 */ 872 873 /* 874 * The MMAP events record the PROT_EXEC mappings so that we can 875 * correlate userspace IPs to code. They have the following structure: 876 * 877 * struct { 878 * struct perf_event_header header; 879 * 880 * u32 pid, tid; 881 * u64 addr; 882 * u64 len; 883 * u64 pgoff; 884 * char filename[]; 885 * struct sample_id sample_id; 886 * }; 887 */ 888 PERF_RECORD_MMAP = 1, 889 890 /* 891 * struct { 892 * struct perf_event_header header; 893 * u64 id; 894 * u64 lost; 895 * struct sample_id sample_id; 896 * }; 897 */ 898 PERF_RECORD_LOST = 2, 899 900 /* 901 * struct { 902 * struct perf_event_header header; 903 * 904 * u32 pid, tid; 905 * char comm[]; 906 * struct sample_id sample_id; 907 * }; 908 */ 909 PERF_RECORD_COMM = 3, 910 911 /* 912 * struct { 913 * struct perf_event_header header; 914 * u32 pid, ppid; 915 * u32 tid, ptid; 916 * u64 time; 917 * struct sample_id sample_id; 918 * }; 919 */ 920 PERF_RECORD_EXIT = 4, 921 922 /* 923 * struct { 924 * struct perf_event_header header; 925 * u64 time; 926 * u64 id; 927 * u64 stream_id; 928 * struct sample_id sample_id; 929 * }; 930 */ 931 PERF_RECORD_THROTTLE = 5, 932 PERF_RECORD_UNTHROTTLE = 6, 933 934 /* 935 * struct { 936 * struct perf_event_header header; 937 * u32 pid, ppid; 938 * u32 tid, ptid; 939 * u64 time; 940 * struct sample_id sample_id; 941 * }; 942 */ 943 PERF_RECORD_FORK = 7, 944 945 /* 946 * struct { 947 * struct perf_event_header header; 948 * u32 pid, tid; 949 * 950 * struct read_format values; 951 * struct sample_id sample_id; 952 * }; 953 */ 954 PERF_RECORD_READ = 8, 955 956 /* 957 * struct { 958 * struct perf_event_header header; 959 * 960 * # 961 * # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID. 962 * # The advantage of PERF_SAMPLE_IDENTIFIER is that its position 963 * # is fixed relative to header. 964 * # 965 * 966 * { u64 id; } && PERF_SAMPLE_IDENTIFIER 967 * { u64 ip; } && PERF_SAMPLE_IP 968 * { u32 pid, tid; } && PERF_SAMPLE_TID 969 * { u64 time; } && PERF_SAMPLE_TIME 970 * { u64 addr; } && PERF_SAMPLE_ADDR 971 * { u64 id; } && PERF_SAMPLE_ID 972 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 973 * { u32 cpu, res; } && PERF_SAMPLE_CPU 974 * { u64 period; } && PERF_SAMPLE_PERIOD 975 * 976 * { struct read_format values; } && PERF_SAMPLE_READ 977 * 978 * { u64 nr, 979 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 980 * 981 * # 982 * # The RAW record below is opaque data wrt the ABI 983 * # 984 * # That is, the ABI doesn't make any promises wrt to 985 * # the stability of its content, it may vary depending 986 * # on event, hardware, kernel version and phase of 987 * # the moon. 988 * # 989 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 990 * # 991 * 992 * { u32 size; 993 * char data[size];}&& PERF_SAMPLE_RAW 994 * 995 * { u64 nr; 996 * { u64 hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX 997 * { u64 from, to, flags } lbr[nr]; 998 * # 999 * # The format of the counters is decided by the 1000 * # "branch_counter_nr" and "branch_counter_width", 1001 * # which are defined in the ABI. 1002 * # 1003 * { u64 counters; } cntr[nr] && PERF_SAMPLE_BRANCH_COUNTERS 1004 * } && PERF_SAMPLE_BRANCH_STACK 1005 * 1006 * { u64 abi; # enum perf_sample_regs_abi 1007 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER 1008 * 1009 * { u64 size; 1010 * char data[size]; 1011 * u64 dyn_size; } && PERF_SAMPLE_STACK_USER 1012 * 1013 * { union perf_sample_weight 1014 * { 1015 * u64 full; && PERF_SAMPLE_WEIGHT 1016 * #if defined(__LITTLE_ENDIAN_BITFIELD) 1017 * struct { 1018 * u32 var1_dw; 1019 * u16 var2_w; 1020 * u16 var3_w; 1021 * } && PERF_SAMPLE_WEIGHT_STRUCT 1022 * #elif defined(__BIG_ENDIAN_BITFIELD) 1023 * struct { 1024 * u16 var3_w; 1025 * u16 var2_w; 1026 * u32 var1_dw; 1027 * } && PERF_SAMPLE_WEIGHT_STRUCT 1028 * #endif 1029 * } 1030 * } 1031 * { u64 data_src; } && PERF_SAMPLE_DATA_SRC 1032 * { u64 transaction; } && PERF_SAMPLE_TRANSACTION 1033 * { u64 abi; # enum perf_sample_regs_abi 1034 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR 1035 * { u64 phys_addr;} && PERF_SAMPLE_PHYS_ADDR 1036 * { u64 size; 1037 * char data[size]; } && PERF_SAMPLE_AUX 1038 * { u64 data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE 1039 * { u64 code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE 1040 * }; 1041 */ 1042 PERF_RECORD_SAMPLE = 9, 1043 1044 /* 1045 * The MMAP2 records are an augmented version of MMAP, they add 1046 * maj, min, ino numbers to be used to uniquely identify each mapping 1047 * 1048 * struct { 1049 * struct perf_event_header header; 1050 * 1051 * u32 pid, tid; 1052 * u64 addr; 1053 * u64 len; 1054 * u64 pgoff; 1055 * union { 1056 * struct { 1057 * u32 maj; 1058 * u32 min; 1059 * u64 ino; 1060 * u64 ino_generation; 1061 * }; 1062 * struct { 1063 * u8 build_id_size; 1064 * u8 __reserved_1; 1065 * u16 __reserved_2; 1066 * u8 build_id[20]; 1067 * }; 1068 * }; 1069 * u32 prot, flags; 1070 * char filename[]; 1071 * struct sample_id sample_id; 1072 * }; 1073 */ 1074 PERF_RECORD_MMAP2 = 10, 1075 1076 /* 1077 * Records that new data landed in the AUX buffer part. 1078 * 1079 * struct { 1080 * struct perf_event_header header; 1081 * 1082 * u64 aux_offset; 1083 * u64 aux_size; 1084 * u64 flags; 1085 * struct sample_id sample_id; 1086 * }; 1087 */ 1088 PERF_RECORD_AUX = 11, 1089 1090 /* 1091 * Indicates that instruction trace has started 1092 * 1093 * struct { 1094 * struct perf_event_header header; 1095 * u32 pid; 1096 * u32 tid; 1097 * struct sample_id sample_id; 1098 * }; 1099 */ 1100 PERF_RECORD_ITRACE_START = 12, 1101 1102 /* 1103 * Records the dropped/lost sample number. 1104 * 1105 * struct { 1106 * struct perf_event_header header; 1107 * 1108 * u64 lost; 1109 * struct sample_id sample_id; 1110 * }; 1111 */ 1112 PERF_RECORD_LOST_SAMPLES = 13, 1113 1114 /* 1115 * Records a context switch in or out (flagged by 1116 * PERF_RECORD_MISC_SWITCH_OUT). See also 1117 * PERF_RECORD_SWITCH_CPU_WIDE. 1118 * 1119 * struct { 1120 * struct perf_event_header header; 1121 * struct sample_id sample_id; 1122 * }; 1123 */ 1124 PERF_RECORD_SWITCH = 14, 1125 1126 /* 1127 * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and 1128 * next_prev_tid that are the next (switching out) or previous 1129 * (switching in) pid/tid. 1130 * 1131 * struct { 1132 * struct perf_event_header header; 1133 * u32 next_prev_pid; 1134 * u32 next_prev_tid; 1135 * struct sample_id sample_id; 1136 * }; 1137 */ 1138 PERF_RECORD_SWITCH_CPU_WIDE = 15, 1139 1140 /* 1141 * struct { 1142 * struct perf_event_header header; 1143 * u32 pid; 1144 * u32 tid; 1145 * u64 nr_namespaces; 1146 * { u64 dev, inode; } [nr_namespaces]; 1147 * struct sample_id sample_id; 1148 * }; 1149 */ 1150 PERF_RECORD_NAMESPACES = 16, 1151 1152 /* 1153 * Record ksymbol register/unregister events: 1154 * 1155 * struct { 1156 * struct perf_event_header header; 1157 * u64 addr; 1158 * u32 len; 1159 * u16 ksym_type; 1160 * u16 flags; 1161 * char name[]; 1162 * struct sample_id sample_id; 1163 * }; 1164 */ 1165 PERF_RECORD_KSYMBOL = 17, 1166 1167 /* 1168 * Record bpf events: 1169 * enum perf_bpf_event_type { 1170 * PERF_BPF_EVENT_UNKNOWN = 0, 1171 * PERF_BPF_EVENT_PROG_LOAD = 1, 1172 * PERF_BPF_EVENT_PROG_UNLOAD = 2, 1173 * }; 1174 * 1175 * struct { 1176 * struct perf_event_header header; 1177 * u16 type; 1178 * u16 flags; 1179 * u32 id; 1180 * u8 tag[BPF_TAG_SIZE]; 1181 * struct sample_id sample_id; 1182 * }; 1183 */ 1184 PERF_RECORD_BPF_EVENT = 18, 1185 1186 /* 1187 * struct { 1188 * struct perf_event_header header; 1189 * u64 id; 1190 * char path[]; 1191 * struct sample_id sample_id; 1192 * }; 1193 */ 1194 PERF_RECORD_CGROUP = 19, 1195 1196 /* 1197 * Records changes to kernel text i.e. self-modified code. 'old_len' is 1198 * the number of old bytes, 'new_len' is the number of new bytes. Either 1199 * 'old_len' or 'new_len' may be zero to indicate, for example, the 1200 * addition or removal of a trampoline. 'bytes' contains the old bytes 1201 * followed immediately by the new bytes. 1202 * 1203 * struct { 1204 * struct perf_event_header header; 1205 * u64 addr; 1206 * u16 old_len; 1207 * u16 new_len; 1208 * u8 bytes[]; 1209 * struct sample_id sample_id; 1210 * }; 1211 */ 1212 PERF_RECORD_TEXT_POKE = 20, 1213 1214 /* 1215 * Data written to the AUX area by hardware due to aux_output, may need 1216 * to be matched to the event by an architecture-specific hardware ID. 1217 * This records the hardware ID, but requires sample_id to provide the 1218 * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT 1219 * records from multiple events. 1220 * 1221 * struct { 1222 * struct perf_event_header header; 1223 * u64 hw_id; 1224 * struct sample_id sample_id; 1225 * }; 1226 */ 1227 PERF_RECORD_AUX_OUTPUT_HW_ID = 21, 1228 1229 PERF_RECORD_MAX, /* non-ABI */ 1230 }; 1231 1232 enum perf_record_ksymbol_type { 1233 PERF_RECORD_KSYMBOL_TYPE_UNKNOWN = 0, 1234 PERF_RECORD_KSYMBOL_TYPE_BPF = 1, 1235 /* 1236 * Out of line code such as kprobe-replaced instructions or optimized 1237 * kprobes or ftrace trampolines. 1238 */ 1239 PERF_RECORD_KSYMBOL_TYPE_OOL = 2, 1240 PERF_RECORD_KSYMBOL_TYPE_MAX /* non-ABI */ 1241 }; 1242 1243 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER (1 << 0) 1244 1245 enum perf_bpf_event_type { 1246 PERF_BPF_EVENT_UNKNOWN = 0, 1247 PERF_BPF_EVENT_PROG_LOAD = 1, 1248 PERF_BPF_EVENT_PROG_UNLOAD = 2, 1249 PERF_BPF_EVENT_MAX, /* non-ABI */ 1250 }; 1251 1252 #define PERF_MAX_STACK_DEPTH 127 1253 #define PERF_MAX_CONTEXTS_PER_STACK 8 1254 1255 enum perf_callchain_context { 1256 PERF_CONTEXT_HV = (__u64)-32, 1257 PERF_CONTEXT_KERNEL = (__u64)-128, 1258 PERF_CONTEXT_USER = (__u64)-512, 1259 1260 PERF_CONTEXT_GUEST = (__u64)-2048, 1261 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 1262 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 1263 1264 PERF_CONTEXT_MAX = (__u64)-4095, 1265 }; 1266 1267 /** 1268 * PERF_RECORD_AUX::flags bits 1269 */ 1270 #define PERF_AUX_FLAG_TRUNCATED 0x01 /* record was truncated to fit */ 1271 #define PERF_AUX_FLAG_OVERWRITE 0x02 /* snapshot from overwrite mode */ 1272 #define PERF_AUX_FLAG_PARTIAL 0x04 /* record contains gaps */ 1273 #define PERF_AUX_FLAG_COLLISION 0x08 /* sample collided with another */ 1274 #define PERF_AUX_FLAG_PMU_FORMAT_TYPE_MASK 0xff00 /* PMU specific trace format type */ 1275 1276 /* CoreSight PMU AUX buffer formats */ 1277 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_CORESIGHT 0x0000 /* Default for backward compatibility */ 1278 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW 0x0100 /* Raw format of the source */ 1279 1280 #define PERF_FLAG_FD_NO_GROUP (1UL << 0) 1281 #define PERF_FLAG_FD_OUTPUT (1UL << 1) 1282 #define PERF_FLAG_PID_CGROUP (1UL << 2) /* pid=cgroup id, per-cpu mode only */ 1283 #define PERF_FLAG_FD_CLOEXEC (1UL << 3) /* O_CLOEXEC */ 1284 1285 #if defined(__LITTLE_ENDIAN_BITFIELD) 1286 union perf_mem_data_src { 1287 __u64 val; 1288 struct { 1289 __u64 mem_op:5, /* type of opcode */ 1290 mem_lvl:14, /* memory hierarchy level */ 1291 mem_snoop:5, /* snoop mode */ 1292 mem_lock:2, /* lock instr */ 1293 mem_dtlb:7, /* tlb access */ 1294 mem_lvl_num:4, /* memory hierarchy level number */ 1295 mem_remote:1, /* remote */ 1296 mem_snoopx:2, /* snoop mode, ext */ 1297 mem_blk:3, /* access blocked */ 1298 mem_hops:3, /* hop level */ 1299 mem_rsvd:18; 1300 }; 1301 }; 1302 #elif defined(__BIG_ENDIAN_BITFIELD) 1303 union perf_mem_data_src { 1304 __u64 val; 1305 struct { 1306 __u64 mem_rsvd:18, 1307 mem_hops:3, /* hop level */ 1308 mem_blk:3, /* access blocked */ 1309 mem_snoopx:2, /* snoop mode, ext */ 1310 mem_remote:1, /* remote */ 1311 mem_lvl_num:4, /* memory hierarchy level number */ 1312 mem_dtlb:7, /* tlb access */ 1313 mem_lock:2, /* lock instr */ 1314 mem_snoop:5, /* snoop mode */ 1315 mem_lvl:14, /* memory hierarchy level */ 1316 mem_op:5; /* type of opcode */ 1317 }; 1318 }; 1319 #else 1320 #error "Unknown endianness" 1321 #endif 1322 1323 /* type of opcode (load/store/prefetch,code) */ 1324 #define PERF_MEM_OP_NA 0x01 /* not available */ 1325 #define PERF_MEM_OP_LOAD 0x02 /* load instruction */ 1326 #define PERF_MEM_OP_STORE 0x04 /* store instruction */ 1327 #define PERF_MEM_OP_PFETCH 0x08 /* prefetch */ 1328 #define PERF_MEM_OP_EXEC 0x10 /* code (execution) */ 1329 #define PERF_MEM_OP_SHIFT 0 1330 1331 /* 1332 * PERF_MEM_LVL_* namespace being depricated to some extent in the 1333 * favour of newer composite PERF_MEM_{LVLNUM_,REMOTE_,SNOOPX_} fields. 1334 * Supporting this namespace inorder to not break defined ABIs. 1335 * 1336 * memory hierarchy (memory level, hit or miss) 1337 */ 1338 #define PERF_MEM_LVL_NA 0x01 /* not available */ 1339 #define PERF_MEM_LVL_HIT 0x02 /* hit level */ 1340 #define PERF_MEM_LVL_MISS 0x04 /* miss level */ 1341 #define PERF_MEM_LVL_L1 0x08 /* L1 */ 1342 #define PERF_MEM_LVL_LFB 0x10 /* Line Fill Buffer */ 1343 #define PERF_MEM_LVL_L2 0x20 /* L2 */ 1344 #define PERF_MEM_LVL_L3 0x40 /* L3 */ 1345 #define PERF_MEM_LVL_LOC_RAM 0x80 /* Local DRAM */ 1346 #define PERF_MEM_LVL_REM_RAM1 0x100 /* Remote DRAM (1 hop) */ 1347 #define PERF_MEM_LVL_REM_RAM2 0x200 /* Remote DRAM (2 hops) */ 1348 #define PERF_MEM_LVL_REM_CCE1 0x400 /* Remote Cache (1 hop) */ 1349 #define PERF_MEM_LVL_REM_CCE2 0x800 /* Remote Cache (2 hops) */ 1350 #define PERF_MEM_LVL_IO 0x1000 /* I/O memory */ 1351 #define PERF_MEM_LVL_UNC 0x2000 /* Uncached memory */ 1352 #define PERF_MEM_LVL_SHIFT 5 1353 1354 #define PERF_MEM_REMOTE_REMOTE 0x01 /* Remote */ 1355 #define PERF_MEM_REMOTE_SHIFT 37 1356 1357 #define PERF_MEM_LVLNUM_L1 0x01 /* L1 */ 1358 #define PERF_MEM_LVLNUM_L2 0x02 /* L2 */ 1359 #define PERF_MEM_LVLNUM_L3 0x03 /* L3 */ 1360 #define PERF_MEM_LVLNUM_L4 0x04 /* L4 */ 1361 #define PERF_MEM_LVLNUM_L2_MHB 0x05 /* L2 Miss Handling Buffer */ 1362 #define PERF_MEM_LVLNUM_MSC 0x06 /* Memory-side Cache */ 1363 /* 0x7 available */ 1364 #define PERF_MEM_LVLNUM_UNC 0x08 /* Uncached */ 1365 #define PERF_MEM_LVLNUM_CXL 0x09 /* CXL */ 1366 #define PERF_MEM_LVLNUM_IO 0x0a /* I/O */ 1367 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */ 1368 #define PERF_MEM_LVLNUM_LFB 0x0c /* LFB / L1 Miss Handling Buffer */ 1369 #define PERF_MEM_LVLNUM_RAM 0x0d /* RAM */ 1370 #define PERF_MEM_LVLNUM_PMEM 0x0e /* PMEM */ 1371 #define PERF_MEM_LVLNUM_NA 0x0f /* N/A */ 1372 1373 #define PERF_MEM_LVLNUM_SHIFT 33 1374 1375 /* snoop mode */ 1376 #define PERF_MEM_SNOOP_NA 0x01 /* not available */ 1377 #define PERF_MEM_SNOOP_NONE 0x02 /* no snoop */ 1378 #define PERF_MEM_SNOOP_HIT 0x04 /* snoop hit */ 1379 #define PERF_MEM_SNOOP_MISS 0x08 /* snoop miss */ 1380 #define PERF_MEM_SNOOP_HITM 0x10 /* snoop hit modified */ 1381 #define PERF_MEM_SNOOP_SHIFT 19 1382 1383 #define PERF_MEM_SNOOPX_FWD 0x01 /* forward */ 1384 #define PERF_MEM_SNOOPX_PEER 0x02 /* xfer from peer */ 1385 #define PERF_MEM_SNOOPX_SHIFT 38 1386 1387 /* locked instruction */ 1388 #define PERF_MEM_LOCK_NA 0x01 /* not available */ 1389 #define PERF_MEM_LOCK_LOCKED 0x02 /* locked transaction */ 1390 #define PERF_MEM_LOCK_SHIFT 24 1391 1392 /* TLB access */ 1393 #define PERF_MEM_TLB_NA 0x01 /* not available */ 1394 #define PERF_MEM_TLB_HIT 0x02 /* hit level */ 1395 #define PERF_MEM_TLB_MISS 0x04 /* miss level */ 1396 #define PERF_MEM_TLB_L1 0x08 /* L1 */ 1397 #define PERF_MEM_TLB_L2 0x10 /* L2 */ 1398 #define PERF_MEM_TLB_WK 0x20 /* Hardware Walker*/ 1399 #define PERF_MEM_TLB_OS 0x40 /* OS fault handler */ 1400 #define PERF_MEM_TLB_SHIFT 26 1401 1402 /* Access blocked */ 1403 #define PERF_MEM_BLK_NA 0x01 /* not available */ 1404 #define PERF_MEM_BLK_DATA 0x02 /* data could not be forwarded */ 1405 #define PERF_MEM_BLK_ADDR 0x04 /* address conflict */ 1406 #define PERF_MEM_BLK_SHIFT 40 1407 1408 /* hop level */ 1409 #define PERF_MEM_HOPS_0 0x01 /* remote core, same node */ 1410 #define PERF_MEM_HOPS_1 0x02 /* remote node, same socket */ 1411 #define PERF_MEM_HOPS_2 0x03 /* remote socket, same board */ 1412 #define PERF_MEM_HOPS_3 0x04 /* remote board */ 1413 /* 5-7 available */ 1414 #define PERF_MEM_HOPS_SHIFT 43 1415 1416 #define PERF_MEM_S(a, s) \ 1417 (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT) 1418 1419 /* 1420 * single taken branch record layout: 1421 * 1422 * from: source instruction (may not always be a branch insn) 1423 * to: branch target 1424 * mispred: branch target was mispredicted 1425 * predicted: branch target was predicted 1426 * 1427 * support for mispred, predicted is optional. In case it 1428 * is not supported mispred = predicted = 0. 1429 * 1430 * in_tx: running in a hardware transaction 1431 * abort: aborting a hardware transaction 1432 * cycles: cycles from last branch (or 0 if not supported) 1433 * type: branch type 1434 * spec: branch speculation info (or 0 if not supported) 1435 */ 1436 struct perf_branch_entry { 1437 __u64 from; 1438 __u64 to; 1439 __u64 mispred:1, /* target mispredicted */ 1440 predicted:1,/* target predicted */ 1441 in_tx:1, /* in transaction */ 1442 abort:1, /* transaction abort */ 1443 cycles:16, /* cycle count to last branch */ 1444 type:4, /* branch type */ 1445 spec:2, /* branch speculation info */ 1446 new_type:4, /* additional branch type */ 1447 priv:3, /* privilege level */ 1448 reserved:31; 1449 }; 1450 1451 /* Size of used info bits in struct perf_branch_entry */ 1452 #define PERF_BRANCH_ENTRY_INFO_BITS_MAX 33 1453 1454 union perf_sample_weight { 1455 __u64 full; 1456 #if defined(__LITTLE_ENDIAN_BITFIELD) 1457 struct { 1458 __u32 var1_dw; 1459 __u16 var2_w; 1460 __u16 var3_w; 1461 }; 1462 #elif defined(__BIG_ENDIAN_BITFIELD) 1463 struct { 1464 __u16 var3_w; 1465 __u16 var2_w; 1466 __u32 var1_dw; 1467 }; 1468 #else 1469 #error "Unknown endianness" 1470 #endif 1471 }; 1472 1473 #endif /* _LINUX_PERF_EVENT_H */