1 /* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx (at) 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 */ 28 enum perf_type_id { 29 PERF_TYPE_HARDWARE = 0, 30 PERF_TYPE_SOFTWARE = 1, 31 PERF_TYPE_TRACEPOINT = 2, 32 PERF_TYPE_HW_CACHE = 3, 33 PERF_TYPE_RAW = 4, 34 35 PERF_TYPE_MAX, /* non-ABI */ 36 }; 37 38 /* 39 * Generalized performance event event_id types, used by the 40 * attr.event_id parameter of the sys_perf_event_open() 41 * syscall: 42 */ 43 enum perf_hw_id { 44 /* 45 * Common hardware events, generalized by the kernel: 46 */ 47 PERF_COUNT_HW_CPU_CYCLES = 0, 48 PERF_COUNT_HW_INSTRUCTIONS = 1, 49 PERF_COUNT_HW_CACHE_REFERENCES = 2, 50 PERF_COUNT_HW_CACHE_MISSES = 3, 51 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, 52 PERF_COUNT_HW_BRANCH_MISSES = 5, 53 PERF_COUNT_HW_BUS_CYCLES = 6, 54 55 PERF_COUNT_HW_MAX, /* non-ABI */ 56 }; 57 58 /* 59 * Generalized hardware cache events: 60 * 61 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x 62 * { read, write, prefetch } x 63 * { accesses, misses } 64 */ 65 enum perf_hw_cache_id { 66 PERF_COUNT_HW_CACHE_L1D = 0, 67 PERF_COUNT_HW_CACHE_L1I = 1, 68 PERF_COUNT_HW_CACHE_LL = 2, 69 PERF_COUNT_HW_CACHE_DTLB = 3, 70 PERF_COUNT_HW_CACHE_ITLB = 4, 71 PERF_COUNT_HW_CACHE_BPU = 5, 72 73 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 74 }; 75 76 enum perf_hw_cache_op_id { 77 PERF_COUNT_HW_CACHE_OP_READ = 0, 78 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 79 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 80 81 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 82 }; 83 84 enum perf_hw_cache_op_result_id { 85 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 86 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 87 88 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 89 }; 90 91 /* 92 * Special "software" events provided by the kernel, even if the hardware 93 * does not support performance events. These events measure various 94 * physical and sw events of the kernel (and allow the profiling of them as 95 * well): 96 */ 97 enum perf_sw_ids { 98 PERF_COUNT_SW_CPU_CLOCK = 0, 99 PERF_COUNT_SW_TASK_CLOCK = 1, 100 PERF_COUNT_SW_PAGE_FAULTS = 2, 101 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 102 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 103 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 104 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 105 106 PERF_COUNT_SW_MAX, /* non-ABI */ 107 }; 108 109 /* 110 * Bits that can be set in attr.sample_type to request information 111 * in the overflow packets. 112 */ 113 enum perf_event_sample_format { 114 PERF_SAMPLE_IP = 1U << 0, 115 PERF_SAMPLE_TID = 1U << 1, 116 PERF_SAMPLE_TIME = 1U << 2, 117 PERF_SAMPLE_ADDR = 1U << 3, 118 PERF_SAMPLE_READ = 1U << 4, 119 PERF_SAMPLE_CALLCHAIN = 1U << 5, 120 PERF_SAMPLE_ID = 1U << 6, 121 PERF_SAMPLE_CPU = 1U << 7, 122 PERF_SAMPLE_PERIOD = 1U << 8, 123 PERF_SAMPLE_STREAM_ID = 1U << 9, 124 PERF_SAMPLE_RAW = 1U << 10, 125 126 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */ 127 }; 128 129 /* 130 * The format of the data returned by read() on a perf event fd, 131 * as specified by attr.read_format: 132 * 133 * struct read_format { 134 * { u64 value; 135 * { u64 time_enabled; } && PERF_FORMAT_ENABLED 136 * { u64 time_running; } && PERF_FORMAT_RUNNING 137 * { u64 id; } && PERF_FORMAT_ID 138 * } && !PERF_FORMAT_GROUP 139 * 140 * { u64 nr; 141 * { u64 time_enabled; } && PERF_FORMAT_ENABLED 142 * { u64 time_running; } && PERF_FORMAT_RUNNING 143 * { u64 value; 144 * { u64 id; } && PERF_FORMAT_ID 145 * } cntr[nr]; 146 * } && PERF_FORMAT_GROUP 147 * }; 148 */ 149 enum perf_event_read_format { 150 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 151 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 152 PERF_FORMAT_ID = 1U << 2, 153 PERF_FORMAT_GROUP = 1U << 3, 154 155 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 156 }; 157 158 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 159 160 /* 161 * Hardware event_id to monitor via a performance monitoring event: 162 */ 163 struct perf_event_attr { 164 165 /* 166 * Major type: hardware/software/tracepoint/etc. 167 */ 168 __u32 type; 169 170 /* 171 * Size of the attr structure, for fwd/bwd compat. 172 */ 173 __u32 size; 174 175 /* 176 * Type specific configuration information. 177 */ 178 __u64 config; 179 180 union { 181 __u64 sample_period; 182 __u64 sample_freq; 183 }; 184 185 __u64 sample_type; 186 __u64 read_format; 187 188 __u64 disabled : 1, /* off by default */ 189 inherit : 1, /* children inherit it */ 190 pinned : 1, /* must always be on PMU */ 191 exclusive : 1, /* only group on PMU */ 192 exclude_user : 1, /* don't count user */ 193 exclude_kernel : 1, /* ditto kernel */ 194 exclude_hv : 1, /* ditto hypervisor */ 195 exclude_idle : 1, /* don't count when idle */ 196 mmap : 1, /* include mmap data */ 197 comm : 1, /* include comm data */ 198 freq : 1, /* use freq, not period */ 199 inherit_stat : 1, /* per task counts */ 200 enable_on_exec : 1, /* next exec enables */ 201 task : 1, /* trace fork/exit */ 202 watermark : 1, /* wakeup_watermark */ 203 204 __reserved_1 : 49; 205 206 union { 207 __u32 wakeup_events; /* wakeup every n events */ 208 __u32 wakeup_watermark; /* bytes before wakeup */ 209 }; 210 __u32 __reserved_2; 211 212 __u64 __reserved_3; 213 }; 214 215 /* 216 * Ioctls that can be done on a perf event fd: 217 */ 218 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 219 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 220 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 221 #define PERF_EVENT_IOC_RESET _IO ('$', 3) 222 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 223 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 224 225 enum perf_event_ioc_flags { 226 PERF_IOC_FLAG_GROUP = 1U << 0, 227 }; 228 229 /* 230 * Structure of the page that can be mapped via mmap 231 */ 232 struct perf_event_mmap_page { 233 __u32 version; /* version number of this structure */ 234 __u32 compat_version; /* lowest version this is compat with */ 235 236 /* 237 * Bits needed to read the hw events in user-space. 238 * 239 * u32 seq; 240 * s64 count; 241 * 242 * do { 243 * seq = pc->lock; 244 * 245 * barrier() 246 * if (pc->index) { 247 * count = pmc_read(pc->index - 1); 248 * count += pc->offset; 249 * } else 250 * goto regular_read; 251 * 252 * barrier(); 253 * } while (pc->lock != seq); 254 * 255 * NOTE: for obvious reason this only works on self-monitoring 256 * processes. 257 */ 258 __u32 lock; /* seqlock for synchronization */ 259 __u32 index; /* hardware event identifier */ 260 __s64 offset; /* add to hardware event value */ 261 __u64 time_enabled; /* time event active */ 262 __u64 time_running; /* time event on cpu */ 263 264 /* 265 * Hole for extension of the self monitor capabilities 266 */ 267 268 __u64 __reserved[123]; /* align to 1k */ 269 270 /* 271 * Control data for the mmap() data buffer. 272 * 273 * User-space reading the @data_head value should issue an rmb(), on 274 * SMP capable platforms, after reading this value -- see 275 * perf_event_wakeup(). 276 * 277 * When the mapping is PROT_WRITE the @data_tail value should be 278 * written by userspace to reflect the last read data. In this case 279 * the kernel will not over-write unread data. 280 */ 281 __u64 data_head; /* head in the data section */ 282 __u64 data_tail; /* user-space written tail */ 283 }; 284 285 #define PERF_RECORD_MISC_CPUMODE_MASK (3 << 0) 286 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 287 #define PERF_RECORD_MISC_KERNEL (1 << 0) 288 #define PERF_RECORD_MISC_USER (2 << 0) 289 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 290 291 struct perf_event_header { 292 __u32 type; 293 __u16 misc; 294 __u16 size; 295 }; 296 297 enum perf_event_type { 298 299 /* 300 * The MMAP events record the PROT_EXEC mappings so that we can 301 * correlate userspace IPs to code. They have the following structure: 302 * 303 * struct { 304 * struct perf_event_header header; 305 * 306 * u32 pid, tid; 307 * u64 addr; 308 * u64 len; 309 * u64 pgoff; 310 * char filename[]; 311 * }; 312 */ 313 PERF_RECORD_MMAP = 1, 314 315 /* 316 * struct { 317 * struct perf_event_header header; 318 * u64 id; 319 * u64 lost; 320 * }; 321 */ 322 PERF_RECORD_LOST = 2, 323 324 /* 325 * struct { 326 * struct perf_event_header header; 327 * 328 * u32 pid, tid; 329 * char comm[]; 330 * }; 331 */ 332 PERF_RECORD_COMM = 3, 333 334 /* 335 * struct { 336 * struct perf_event_header header; 337 * u32 pid, ppid; 338 * u32 tid, ptid; 339 * u64 time; 340 * }; 341 */ 342 PERF_RECORD_EXIT = 4, 343 344 /* 345 * struct { 346 * struct perf_event_header header; 347 * u64 time; 348 * u64 id; 349 * u64 stream_id; 350 * }; 351 */ 352 PERF_RECORD_THROTTLE = 5, 353 PERF_RECORD_UNTHROTTLE = 6, 354 355 /* 356 * struct { 357 * struct perf_event_header header; 358 * u32 pid, ppid; 359 * u32 tid, ptid; 360 * u64 time; 361 * }; 362 */ 363 PERF_RECORD_FORK = 7, 364 365 /* 366 * struct { 367 * struct perf_event_header header; 368 * u32 pid, tid; 369 * 370 * struct read_format values; 371 * }; 372 */ 373 PERF_RECORD_READ = 8, 374 375 /* 376 * struct { 377 * struct perf_event_header header; 378 * 379 * { u64 ip; } && PERF_SAMPLE_IP 380 * { u32 pid, tid; } && PERF_SAMPLE_TID 381 * { u64 time; } && PERF_SAMPLE_TIME 382 * { u64 addr; } && PERF_SAMPLE_ADDR 383 * { u64 id; } && PERF_SAMPLE_ID 384 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 385 * { u32 cpu, res; } && PERF_SAMPLE_CPU 386 * { u64 period; } && PERF_SAMPLE_PERIOD 387 * 388 * { struct read_format values; } && PERF_SAMPLE_READ 389 * 390 * { u64 nr, 391 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 392 * 393 * # 394 * # The RAW record below is opaque data wrt the ABI 395 * # 396 * # That is, the ABI doesn't make any promises wrt to 397 * # the stability of its content, it may vary depending 398 * # on event, hardware, kernel version and phase of 399 * # the moon. 400 * # 401 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 402 * # 403 * 404 * { u32 size; 405 * char data[size];}&& PERF_SAMPLE_RAW 406 * }; 407 */ 408 PERF_RECORD_SAMPLE = 9, 409 410 PERF_RECORD_MAX, /* non-ABI */ 411 }; 412 413 enum perf_callchain_context { 414 PERF_CONTEXT_HV = (__u64)-32, 415 PERF_CONTEXT_KERNEL = (__u64)-128, 416 PERF_CONTEXT_USER = (__u64)-512, 417 418 PERF_CONTEXT_GUEST = (__u64)-2048, 419 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 420 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 421 422 PERF_CONTEXT_MAX = (__u64)-4095, 423 }; 424 425 #define PERF_FLAG_FD_NO_GROUP (1U << 0) 426 #define PERF_FLAG_FD_OUTPUT (1U << 1) 427 428 #endif /* _LINUX_PERF_EVENT_H */ 429