1 /* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <tglx (at) linutronix.de> 5 * Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar 6 * Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra 7 * 8 * Data type definitions, declarations, prototypes. 9 * 10 * Started by: Thomas Gleixner and Ingo Molnar 11 * 12 * For licencing details see kernel-base/COPYING 13 */ 14 #ifndef _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 PERF_TYPE_BREAKPOINT = 5, 35 36 PERF_TYPE_MAX, /* non-ABI */ 37 }; 38 39 /* 40 * Generalized performance event event_id types, used by the 41 * attr.event_id parameter of the sys_perf_event_open() 42 * syscall: 43 */ 44 enum perf_hw_id { 45 /* 46 * Common hardware events, generalized by the kernel: 47 */ 48 PERF_COUNT_HW_CPU_CYCLES = 0, 49 PERF_COUNT_HW_INSTRUCTIONS = 1, 50 PERF_COUNT_HW_CACHE_REFERENCES = 2, 51 PERF_COUNT_HW_CACHE_MISSES = 3, 52 PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, 53 PERF_COUNT_HW_BRANCH_MISSES = 5, 54 PERF_COUNT_HW_BUS_CYCLES = 6, 55 PERF_COUNT_HW_STALLED_CYCLES_FRONTEND = 7, 56 PERF_COUNT_HW_STALLED_CYCLES_BACKEND = 8, 57 58 PERF_COUNT_HW_MAX, /* non-ABI */ 59 }; 60 61 /* 62 * Generalized hardware cache events: 63 * 64 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x 65 * { read, write, prefetch } x 66 * { accesses, misses } 67 */ 68 enum perf_hw_cache_id { 69 PERF_COUNT_HW_CACHE_L1D = 0, 70 PERF_COUNT_HW_CACHE_L1I = 1, 71 PERF_COUNT_HW_CACHE_LL = 2, 72 PERF_COUNT_HW_CACHE_DTLB = 3, 73 PERF_COUNT_HW_CACHE_ITLB = 4, 74 PERF_COUNT_HW_CACHE_BPU = 5, 75 PERF_COUNT_HW_CACHE_NODE = 6, 76 77 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 78 }; 79 80 enum perf_hw_cache_op_id { 81 PERF_COUNT_HW_CACHE_OP_READ = 0, 82 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 83 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 84 85 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 86 }; 87 88 enum perf_hw_cache_op_result_id { 89 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 90 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 91 92 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 93 }; 94 95 /* 96 * Special "software" events provided by the kernel, even if the hardware 97 * does not support performance events. These events measure various 98 * physical and sw events of the kernel (and allow the profiling of them as 99 * well): 100 */ 101 enum perf_sw_ids { 102 PERF_COUNT_SW_CPU_CLOCK = 0, 103 PERF_COUNT_SW_TASK_CLOCK = 1, 104 PERF_COUNT_SW_PAGE_FAULTS = 2, 105 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 106 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 107 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 108 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 109 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7, 110 PERF_COUNT_SW_EMULATION_FAULTS = 8, 111 112 PERF_COUNT_SW_MAX, /* non-ABI */ 113 }; 114 115 /* 116 * Bits that can be set in attr.sample_type to request information 117 * in the overflow packets. 118 */ 119 enum perf_event_sample_format { 120 PERF_SAMPLE_IP = 1U << 0, 121 PERF_SAMPLE_TID = 1U << 1, 122 PERF_SAMPLE_TIME = 1U << 2, 123 PERF_SAMPLE_ADDR = 1U << 3, 124 PERF_SAMPLE_READ = 1U << 4, 125 PERF_SAMPLE_CALLCHAIN = 1U << 5, 126 PERF_SAMPLE_ID = 1U << 6, 127 PERF_SAMPLE_CPU = 1U << 7, 128 PERF_SAMPLE_PERIOD = 1U << 8, 129 PERF_SAMPLE_STREAM_ID = 1U << 9, 130 PERF_SAMPLE_RAW = 1U << 10, 131 132 PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */ 133 }; 134 135 /* 136 * The format of the data returned by read() on a perf event fd, 137 * as specified by attr.read_format: 138 * 139 * struct read_format { 140 * { u64 value; 141 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 142 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 143 * { u64 id; } && PERF_FORMAT_ID 144 * } && !PERF_FORMAT_GROUP 145 * 146 * { u64 nr; 147 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 148 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 149 * { u64 value; 150 * { u64 id; } && PERF_FORMAT_ID 151 * } cntr[nr]; 152 * } && PERF_FORMAT_GROUP 153 * }; 154 */ 155 enum perf_event_read_format { 156 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 157 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 158 PERF_FORMAT_ID = 1U << 2, 159 PERF_FORMAT_GROUP = 1U << 3, 160 161 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 162 }; 163 164 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 165 166 /* 167 * Hardware event_id to monitor via a performance monitoring event: 168 */ 169 struct perf_event_attr { 170 171 /* 172 * Major type: hardware/software/tracepoint/etc. 173 */ 174 __u32 type; 175 176 /* 177 * Size of the attr structure, for fwd/bwd compat. 178 */ 179 __u32 size; 180 181 /* 182 * Type specific configuration information. 183 */ 184 __u64 config; 185 186 union { 187 __u64 sample_period; 188 __u64 sample_freq; 189 }; 190 191 __u64 sample_type; 192 __u64 read_format; 193 194 __u64 disabled : 1, /* off by default */ 195 inherit : 1, /* children inherit it */ 196 pinned : 1, /* must always be on PMU */ 197 exclusive : 1, /* only group on PMU */ 198 exclude_user : 1, /* don't count user */ 199 exclude_kernel : 1, /* ditto kernel */ 200 exclude_hv : 1, /* ditto hypervisor */ 201 exclude_idle : 1, /* don't count when idle */ 202 mmap : 1, /* include mmap data */ 203 comm : 1, /* include comm data */ 204 freq : 1, /* use freq, not period */ 205 inherit_stat : 1, /* per task counts */ 206 enable_on_exec : 1, /* next exec enables */ 207 task : 1, /* trace fork/exit */ 208 watermark : 1, /* wakeup_watermark */ 209 /* 210 * precise_ip: 211 * 212 * 0 - SAMPLE_IP can have arbitrary skid 213 * 1 - SAMPLE_IP must have constant skid 214 * 2 - SAMPLE_IP requested to have 0 skid 215 * 3 - SAMPLE_IP must have 0 skid 216 * 217 * See also PERF_RECORD_MISC_EXACT_IP 218 */ 219 precise_ip : 2, /* skid constraint */ 220 mmap_data : 1, /* non-exec mmap data */ 221 sample_id_all : 1, /* sample_type all events */ 222 223 exclude_host : 1, /* don't count in host */ 224 exclude_guest : 1, /* don't count in guest */ 225 226 __reserved_1 : 43; 227 228 union { 229 __u32 wakeup_events; /* wakeup every n events */ 230 __u32 wakeup_watermark; /* bytes before wakeup */ 231 }; 232 233 __u32 bp_type; 234 union { 235 __u64 bp_addr; 236 __u64 config1; /* extension of config */ 237 }; 238 union { 239 __u64 bp_len; 240 __u64 config2; /* extension of config1 */ 241 }; 242 }; 243 244 /* 245 * Ioctls that can be done on a perf event fd: 246 */ 247 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 248 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 249 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 250 #define PERF_EVENT_IOC_RESET _IO ('$', 3) 251 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 252 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 253 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 254 255 enum perf_event_ioc_flags { 256 PERF_IOC_FLAG_GROUP = 1U << 0, 257 }; 258 259 /* 260 * Structure of the page that can be mapped via mmap 261 */ 262 struct perf_event_mmap_page { 263 __u32 version; /* version number of this structure */ 264 __u32 compat_version; /* lowest version this is compat with */ 265 266 /* 267 * Bits needed to read the hw events in user-space. 268 * 269 * u32 seq; 270 * s64 count; 271 * 272 * do { 273 * seq = pc->lock; 274 * 275 * barrier() 276 * if (pc->index) { 277 * count = pmc_read(pc->index - 1); 278 * count += pc->offset; 279 * } else 280 * goto regular_read; 281 * 282 * barrier(); 283 * } while (pc->lock != seq); 284 * 285 * NOTE: for obvious reason this only works on self-monitoring 286 * processes. 287 */ 288 __u32 lock; /* seqlock for synchronization */ 289 __u32 index; /* hardware event identifier */ 290 __s64 offset; /* add to hardware event value */ 291 __u64 time_enabled; /* time event active */ 292 __u64 time_running; /* time event on cpu */ 293 294 /* 295 * Hole for extension of the self monitor capabilities 296 */ 297 298 __u64 __reserved[123]; /* align to 1k */ 299 300 /* 301 * Control data for the mmap() data buffer. 302 * 303 * User-space reading the @data_head value should issue an rmb(), on 304 * SMP capable platforms, after reading this value -- see 305 * perf_event_wakeup(). 306 * 307 * When the mapping is PROT_WRITE the @data_tail value should be 308 * written by userspace to reflect the last read data. In this case 309 * the kernel will not over-write unread data. 310 */ 311 __u64 data_head; /* head in the data section */ 312 __u64 data_tail; /* user-space written tail */ 313 }; 314 315 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 316 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 317 #define PERF_RECORD_MISC_KERNEL (1 << 0) 318 #define PERF_RECORD_MISC_USER (2 << 0) 319 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 320 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 321 #define PERF_RECORD_MISC_GUEST_USER (5 << 0) 322 323 /* 324 * Indicates that the content of PERF_SAMPLE_IP points to 325 * the actual instruction that triggered the event. See also 326 * perf_event_attr::precise_ip. 327 */ 328 #define PERF_RECORD_MISC_EXACT_IP (1 << 14) 329 /* 330 * Reserve the last bit to indicate some extended misc field 331 */ 332 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 333 334 struct perf_event_header { 335 __u32 type; 336 __u16 misc; 337 __u16 size; 338 }; 339 340 enum perf_event_type { 341 342 /* 343 * If perf_event_attr.sample_id_all is set then all event types will 344 * have the sample_type selected fields related to where/when 345 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID) 346 * described in PERF_RECORD_SAMPLE below, it will be stashed just after 347 * the perf_event_header and the fields already present for the existing 348 * fields, i.e. at the end of the payload. That way a newer perf.data 349 * file will be supported by older perf tools, with these new optional 350 * fields being ignored. 351 * 352 * The MMAP events record the PROT_EXEC mappings so that we can 353 * correlate userspace IPs to code. They have the following structure: 354 * 355 * struct { 356 * struct perf_event_header header; 357 * 358 * u32 pid, tid; 359 * u64 addr; 360 * u64 len; 361 * u64 pgoff; 362 * char filename[]; 363 * }; 364 */ 365 PERF_RECORD_MMAP = 1, 366 367 /* 368 * struct { 369 * struct perf_event_header header; 370 * u64 id; 371 * u64 lost; 372 * }; 373 */ 374 PERF_RECORD_LOST = 2, 375 376 /* 377 * struct { 378 * struct perf_event_header header; 379 * 380 * u32 pid, tid; 381 * char comm[]; 382 * }; 383 */ 384 PERF_RECORD_COMM = 3, 385 386 /* 387 * struct { 388 * struct perf_event_header header; 389 * u32 pid, ppid; 390 * u32 tid, ptid; 391 * u64 time; 392 * }; 393 */ 394 PERF_RECORD_EXIT = 4, 395 396 /* 397 * struct { 398 * struct perf_event_header header; 399 * u64 time; 400 * u64 id; 401 * u64 stream_id; 402 * }; 403 */ 404 PERF_RECORD_THROTTLE = 5, 405 PERF_RECORD_UNTHROTTLE = 6, 406 407 /* 408 * struct { 409 * struct perf_event_header header; 410 * u32 pid, ppid; 411 * u32 tid, ptid; 412 * u64 time; 413 * }; 414 */ 415 PERF_RECORD_FORK = 7, 416 417 /* 418 * struct { 419 * struct perf_event_header header; 420 * u32 pid, tid; 421 * 422 * struct read_format values; 423 * }; 424 */ 425 PERF_RECORD_READ = 8, 426 427 /* 428 * struct { 429 * struct perf_event_header header; 430 * 431 * { u64 ip; } && PERF_SAMPLE_IP 432 * { u32 pid, tid; } && PERF_SAMPLE_TID 433 * { u64 time; } && PERF_SAMPLE_TIME 434 * { u64 addr; } && PERF_SAMPLE_ADDR 435 * { u64 id; } && PERF_SAMPLE_ID 436 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 437 * { u32 cpu, res; } && PERF_SAMPLE_CPU 438 * { u64 period; } && PERF_SAMPLE_PERIOD 439 * 440 * { struct read_format values; } && PERF_SAMPLE_READ 441 * 442 * { u64 nr, 443 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 444 * 445 * # 446 * # The RAW record below is opaque data wrt the ABI 447 * # 448 * # That is, the ABI doesn't make any promises wrt to 449 * # the stability of its content, it may vary depending 450 * # on event, hardware, kernel version and phase of 451 * # the moon. 452 * # 453 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 454 * # 455 * 456 * { u32 size; 457 * char data[size];}&& PERF_SAMPLE_RAW 458 * }; 459 */ 460 PERF_RECORD_SAMPLE = 9, 461 462 PERF_RECORD_MAX, /* non-ABI */ 463 }; 464 465 enum perf_callchain_context { 466 PERF_CONTEXT_HV = (__u64)-32, 467 PERF_CONTEXT_KERNEL = (__u64)-128, 468 PERF_CONTEXT_USER = (__u64)-512, 469 470 PERF_CONTEXT_GUEST = (__u64)-2048, 471 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 472 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 473 474 PERF_CONTEXT_MAX = (__u64)-4095, 475 }; 476 477 #define PERF_FLAG_FD_NO_GROUP (1U << 0) 478 #define PERF_FLAG_FD_OUTPUT (1U << 1) 479 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */ 480 481 #endif /* _LINUX_PERF_EVENT_H */ 482