1/* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de> 5 * Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar 6 * Copyright (C) 2008-2009, 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 _LINUX_PERF_EVENT_H 15#define _LINUX_PERF_EVENT_H 16 17#include <linux/types.h> 18#include <linux/ioctl.h> 19#include <asm/byteorder.h> 20 21/* 22 * User-space ABI bits: 23 */ 24 25/* 26 * attr.type 27 */ 28enum perf_type_id { 29 PERF_TYPE_HARDWARE = 0, 30 PERF_TYPE_SOFTWARE = 1, 31 PERF_TYPE_TRACEPOINT = 2, 32 PERF_TYPE_HW_CACHE = 3, 33 PERF_TYPE_RAW = 4, 34 PERF_TYPE_BREAKPOINT = 5, 35 36 PERF_TYPE_MAX, /* non-ABI */ 37}; 38 39/* 40 * Generalized performance event event_id types, used by the 41 * attr.event_id parameter of the sys_perf_event_open() 42 * syscall: 43 */ 44enum perf_hw_id { 45 /* 46 * Common hardware events, generalized by the kernel: 47 */ 48 PERF_COUNT_HW_CPU_CYCLES = 0, 49 PERF_COUNT_HW_INSTRUCTIONS = 1, 50 PERF_COUNT_HW_CACHE_REFERENCES = 2, 51 PERF_COUNT_HW_CACHE_MISSES = 3, 52 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, 53 PERF_COUNT_HW_BRANCH_MISSES = 5, 54 PERF_COUNT_HW_BUS_CYCLES = 6, 55 56 PERF_COUNT_HW_MAX, /* non-ABI */ 57}; 58 59/* 60 * Generalized hardware cache events: 61 * 62 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x 63 * { read, write, prefetch } x 64 * { accesses, misses } 65 */ 66enum perf_hw_cache_id { 67 PERF_COUNT_HW_CACHE_L1D = 0, 68 PERF_COUNT_HW_CACHE_L1I = 1, 69 PERF_COUNT_HW_CACHE_LL = 2, 70 PERF_COUNT_HW_CACHE_DTLB = 3, 71 PERF_COUNT_HW_CACHE_ITLB = 4, 72 PERF_COUNT_HW_CACHE_BPU = 5, 73 74 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 75}; 76 77enum perf_hw_cache_op_id { 78 PERF_COUNT_HW_CACHE_OP_READ = 0, 79 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 80 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 81 82 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 83}; 84 85enum perf_hw_cache_op_result_id { 86 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 87 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 88 89 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 90}; 91 92/* 93 * Special "software" events provided by the kernel, even if the hardware 94 * does not support performance events. These events measure various 95 * physical and sw events of the kernel (and allow the profiling of them as 96 * well): 97 */ 98enum perf_sw_ids { 99 PERF_COUNT_SW_CPU_CLOCK = 0, 100 PERF_COUNT_SW_TASK_CLOCK = 1, 101 PERF_COUNT_SW_PAGE_FAULTS = 2, 102 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 103 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 104 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 105 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 106 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7, 107 PERF_COUNT_SW_EMULATION_FAULTS = 8, 108 109 PERF_COUNT_SW_MAX, /* non-ABI */ 110}; 111 112/* 113 * Bits that can be set in attr.sample_type to request information 114 * in the overflow packets. 115 */ 116enum perf_event_sample_format { 117 PERF_SAMPLE_IP = 1U << 0, 118 PERF_SAMPLE_TID = 1U << 1, 119 PERF_SAMPLE_TIME = 1U << 2, 120 PERF_SAMPLE_ADDR = 1U << 3, 121 PERF_SAMPLE_READ = 1U << 4, 122 PERF_SAMPLE_CALLCHAIN = 1U << 5, 123 PERF_SAMPLE_ID = 1U << 6, 124 PERF_SAMPLE_CPU = 1U << 7, 125 PERF_SAMPLE_PERIOD = 1U << 8, 126 PERF_SAMPLE_STREAM_ID = 1U << 9, 127 PERF_SAMPLE_RAW = 1U << 10, 128 129 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */ 130}; 131 132/* 133 * The format of the data returned by read() on a perf event fd, 134 * as specified by attr.read_format: 135 * 136 * struct read_format { 137 * { u64 value; 138 * { u64 time_enabled; } && PERF_FORMAT_ENABLED 139 * { u64 time_running; } && PERF_FORMAT_RUNNING 140 * { u64 id; } && PERF_FORMAT_ID 141 * } && !PERF_FORMAT_GROUP 142 * 143 * { u64 nr; 144 * { u64 time_enabled; } && PERF_FORMAT_ENABLED 145 * { u64 time_running; } && PERF_FORMAT_RUNNING 146 * { u64 value; 147 * { u64 id; } && PERF_FORMAT_ID 148 * } cntr[nr]; 149 * } && PERF_FORMAT_GROUP 150 * }; 151 */ 152enum perf_event_read_format { 153 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 154 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 155 PERF_FORMAT_ID = 1U << 2, 156 PERF_FORMAT_GROUP = 1U << 3, 157 158 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 159}; 160 161#define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 162 163/* 164 * Hardware event_id to monitor via a performance monitoring event: 165 */ 166struct perf_event_attr { 167 168 /* 169 * Major type: hardware/software/tracepoint/etc. 170 */ 171 __u32 type; 172 173 /* 174 * Size of the attr structure, for fwd/bwd compat. 175 */ 176 __u32 size; 177 178 /* 179 * Type specific configuration information. 180 */ 181 __u64 config; 182 183 union { 184 __u64 sample_period; 185 __u64 sample_freq; 186 }; 187 188 __u64 sample_type; 189 __u64 read_format; 190 191 __u64 disabled : 1, /* off by default */ 192 inherit : 1, /* children inherit it */ 193 pinned : 1, /* must always be on PMU */ 194 exclusive : 1, /* only group on PMU */ 195 exclude_user : 1, /* don't count user */ 196 exclude_kernel : 1, /* ditto kernel */ 197 exclude_hv : 1, /* ditto hypervisor */ 198 exclude_idle : 1, /* don't count when idle */ 199 mmap : 1, /* include mmap data */ 200 comm : 1, /* include comm data */ 201 freq : 1, /* use freq, not period */ 202 inherit_stat : 1, /* per task counts */ 203 enable_on_exec : 1, /* next exec enables */ 204 task : 1, /* trace fork/exit */ 205 watermark : 1, /* wakeup_watermark */ 206 /* 207 * precise_ip: 208 * 209 * 0 - SAMPLE_IP can have arbitrary skid 210 * 1 - SAMPLE_IP must have constant skid 211 * 2 - SAMPLE_IP requested to have 0 skid 212 * 3 - SAMPLE_IP must have 0 skid 213 * 214 * See also PERF_RECORD_MISC_EXACT_IP 215 */ 216 precise_ip : 2, /* skid constraint */ 217 mmap_data : 1, /* non-exec mmap data */ 218 219 __reserved_1 : 46; 220 221 union { 222 __u32 wakeup_events; /* wakeup every n events */ 223 __u32 wakeup_watermark; /* bytes before wakeup */ 224 }; 225 226 __u32 bp_type; 227 __u64 bp_addr; 228 __u64 bp_len; 229}; 230 231/* 232 * Ioctls that can be done on a perf event fd: 233 */ 234#define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 235#define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 236#define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 237#define PERF_EVENT_IOC_RESET _IO ('$', 3) 238#define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 239#define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 240#define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 241 242enum perf_event_ioc_flags { 243 PERF_IOC_FLAG_GROUP = 1U << 0, 244}; 245 246/* 247 * Structure of the page that can be mapped via mmap 248 */ 249struct perf_event_mmap_page { 250 __u32 version; /* version number of this structure */ 251 __u32 compat_version; /* lowest version this is compat with */ 252 253 /* 254 * Bits needed to read the hw events in user-space. 255 * 256 * u32 seq; 257 * s64 count; 258 * 259 * do { 260 * seq = pc->lock; 261 * 262 * barrier() 263 * if (pc->index) { 264 * count = pmc_read(pc->index - 1); 265 * count += pc->offset; 266 * } else 267 * goto regular_read; 268 * 269 * barrier(); 270 * } while (pc->lock != seq); 271 * 272 * NOTE: for obvious reason this only works on self-monitoring 273 * processes. 274 */ 275 __u32 lock; /* seqlock for synchronization */ 276 __u32 index; /* hardware event identifier */ 277 __s64 offset; /* add to hardware event value */ 278 __u64 time_enabled; /* time event active */ 279 __u64 time_running; /* time event on cpu */ 280 281 /* 282 * Hole for extension of the self monitor capabilities 283 */ 284 285 __u64 __reserved[123]; /* align to 1k */ 286 287 /* 288 * Control data for the mmap() data buffer. 289 * 290 * User-space reading the @data_head value should issue an rmb(), on 291 * SMP capable platforms, after reading this value -- see 292 * perf_event_wakeup(). 293 * 294 * When the mapping is PROT_WRITE the @data_tail value should be 295 * written by userspace to reflect the last read data. In this case 296 * the kernel will not over-write unread data. 297 */ 298 __u64 data_head; /* head in the data section */ 299 __u64 data_tail; /* user-space written tail */ 300}; 301 302#define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 303#define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 304#define PERF_RECORD_MISC_KERNEL (1 << 0) 305#define PERF_RECORD_MISC_USER (2 << 0) 306#define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 307#define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 308#define PERF_RECORD_MISC_GUEST_USER (5 << 0) 309 310/* 311 * Indicates that the content of PERF_SAMPLE_IP points to 312 * the actual instruction that triggered the event. See also 313 * perf_event_attr::precise_ip. 314 */ 315#define PERF_RECORD_MISC_EXACT_IP (1 << 14) 316/* 317 * Reserve the last bit to indicate some extended misc field 318 */ 319#define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 320 321struct perf_event_header { 322 __u32 type; 323 __u16 misc; 324 __u16 size; 325}; 326 327enum perf_event_type { 328 329 /* 330 * The MMAP events record the PROT_EXEC mappings so that we can 331 * correlate userspace IPs to code. They have the following structure: 332 * 333 * struct { 334 * struct perf_event_header header; 335 * 336 * u32 pid, tid; 337 * u64 addr; 338 * u64 len; 339 * u64 pgoff; 340 * char filename[]; 341 * }; 342 */ 343 PERF_RECORD_MMAP = 1, 344 345 /* 346 * struct { 347 * struct perf_event_header header; 348 * u64 id; 349 * u64 lost; 350 * }; 351 */ 352 PERF_RECORD_LOST = 2, 353 354 /* 355 * struct { 356 * struct perf_event_header header; 357 * 358 * u32 pid, tid; 359 * char comm[]; 360 * }; 361 */ 362 PERF_RECORD_COMM = 3, 363 364 /* 365 * struct { 366 * struct perf_event_header header; 367 * u32 pid, ppid; 368 * u32 tid, ptid; 369 * u64 time; 370 * }; 371 */ 372 PERF_RECORD_EXIT = 4, 373 374 /* 375 * struct { 376 * struct perf_event_header header; 377 * u64 time; 378 * u64 id; 379 * u64 stream_id; 380 * }; 381 */ 382 PERF_RECORD_THROTTLE = 5, 383 PERF_RECORD_UNTHROTTLE = 6, 384 385 /* 386 * struct { 387 * struct perf_event_header header; 388 * u32 pid, ppid; 389 * u32 tid, ptid; 390 * u64 time; 391 * }; 392 */ 393 PERF_RECORD_FORK = 7, 394 395 /* 396 * struct { 397 * struct perf_event_header header; 398 * u32 pid, tid; 399 * 400 * struct read_format values; 401 * }; 402 */ 403 PERF_RECORD_READ = 8, 404 405 /* 406 * struct { 407 * struct perf_event_header header; 408 * 409 * { u64 ip; } && PERF_SAMPLE_IP 410 * { u32 pid, tid; } && PERF_SAMPLE_TID 411 * { u64 time; } && PERF_SAMPLE_TIME 412 * { u64 addr; } && PERF_SAMPLE_ADDR 413 * { u64 id; } && PERF_SAMPLE_ID 414 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 415 * { u32 cpu, res; } && PERF_SAMPLE_CPU 416 * { u64 period; } && PERF_SAMPLE_PERIOD 417 * 418 * { struct read_format values; } && PERF_SAMPLE_READ 419 * 420 * { u64 nr, 421 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 422 * 423 * # 424 * # The RAW record below is opaque data wrt the ABI 425 * # 426 * # That is, the ABI doesn't make any promises wrt to 427 * # the stability of its content, it may vary depending 428 * # on event, hardware, kernel version and phase of 429 * # the moon. 430 * # 431 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 432 * # 433 * 434 * { u32 size; 435 * char data[size];}&& PERF_SAMPLE_RAW 436 * }; 437 */ 438 PERF_RECORD_SAMPLE = 9, 439 440 PERF_RECORD_MAX, /* non-ABI */ 441}; 442 443enum perf_callchain_context { 444 PERF_CONTEXT_HV = (__u64)-32, 445 PERF_CONTEXT_KERNEL = (__u64)-128, 446 PERF_CONTEXT_USER = (__u64)-512, 447 448 PERF_CONTEXT_GUEST = (__u64)-2048, 449 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 450 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 451 452 PERF_CONTEXT_MAX = (__u64)-4095, 453}; 454 455#define PERF_FLAG_FD_NO_GROUP (1U << 0) 456#define PERF_FLAG_FD_OUTPUT (1U << 1) 457 458#endif /* _LINUX_PERF_EVENT_H */ 459