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