Ingo Molnar | 57c0c15 | 2009-09-21 12:20:38 +0200 | [diff] [blame] | 1 | /* |
| 2 | * NOTE: this file will be removed in a future kernel release, it is |
| 3 | * provided as a courtesy copy of user-space code that relies on the |
| 4 | * old (pre-rename) symbols and constants. |
| 5 | * |
| 6 | * Performance events: |
| 7 | * |
| 8 | * Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de> |
| 9 | * Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar |
| 10 | * Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra |
| 11 | * |
| 12 | * Data type definitions, declarations, prototypes. |
| 13 | * |
| 14 | * Started by: Thomas Gleixner and Ingo Molnar |
| 15 | * |
| 16 | * For licencing details see kernel-base/COPYING |
| 17 | */ |
| 18 | #ifndef _LINUX_PERF_COUNTER_H |
| 19 | #define _LINUX_PERF_COUNTER_H |
| 20 | |
| 21 | #include <linux/types.h> |
| 22 | #include <linux/ioctl.h> |
| 23 | #include <asm/byteorder.h> |
| 24 | |
| 25 | /* |
| 26 | * User-space ABI bits: |
| 27 | */ |
| 28 | |
| 29 | /* |
| 30 | * attr.type |
| 31 | */ |
| 32 | enum perf_type_id { |
| 33 | PERF_TYPE_HARDWARE = 0, |
| 34 | PERF_TYPE_SOFTWARE = 1, |
| 35 | PERF_TYPE_TRACEPOINT = 2, |
| 36 | PERF_TYPE_HW_CACHE = 3, |
| 37 | PERF_TYPE_RAW = 4, |
| 38 | |
| 39 | PERF_TYPE_MAX, /* non-ABI */ |
| 40 | }; |
| 41 | |
| 42 | /* |
| 43 | * Generalized performance counter event types, used by the |
| 44 | * attr.event_id parameter of the sys_perf_counter_open() |
| 45 | * syscall: |
| 46 | */ |
| 47 | enum perf_hw_id { |
| 48 | /* |
| 49 | * Common hardware events, generalized by the kernel: |
| 50 | */ |
| 51 | PERF_COUNT_HW_CPU_CYCLES = 0, |
| 52 | PERF_COUNT_HW_INSTRUCTIONS = 1, |
| 53 | PERF_COUNT_HW_CACHE_REFERENCES = 2, |
| 54 | PERF_COUNT_HW_CACHE_MISSES = 3, |
| 55 | PERF_COUNT_HW_BRANCH_INSTRUCTIONS = 4, |
| 56 | PERF_COUNT_HW_BRANCH_MISSES = 5, |
| 57 | PERF_COUNT_HW_BUS_CYCLES = 6, |
| 58 | |
| 59 | PERF_COUNT_HW_MAX, /* non-ABI */ |
| 60 | }; |
| 61 | |
| 62 | /* |
| 63 | * Generalized hardware cache counters: |
| 64 | * |
| 65 | * { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x |
| 66 | * { read, write, prefetch } x |
| 67 | * { accesses, misses } |
| 68 | */ |
| 69 | enum perf_hw_cache_id { |
| 70 | PERF_COUNT_HW_CACHE_L1D = 0, |
| 71 | PERF_COUNT_HW_CACHE_L1I = 1, |
| 72 | PERF_COUNT_HW_CACHE_LL = 2, |
| 73 | PERF_COUNT_HW_CACHE_DTLB = 3, |
| 74 | PERF_COUNT_HW_CACHE_ITLB = 4, |
| 75 | PERF_COUNT_HW_CACHE_BPU = 5, |
| 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" counters provided by the kernel, even if the hardware |
| 97 | * does not support performance counters. These counters 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 | |
| 110 | PERF_COUNT_SW_MAX, /* non-ABI */ |
| 111 | }; |
| 112 | |
| 113 | /* |
| 114 | * Bits that can be set in attr.sample_type to request information |
| 115 | * in the overflow packets. |
| 116 | */ |
| 117 | enum perf_counter_sample_format { |
| 118 | PERF_SAMPLE_IP = 1U << 0, |
| 119 | PERF_SAMPLE_TID = 1U << 1, |
| 120 | PERF_SAMPLE_TIME = 1U << 2, |
| 121 | PERF_SAMPLE_ADDR = 1U << 3, |
| 122 | PERF_SAMPLE_READ = 1U << 4, |
| 123 | PERF_SAMPLE_CALLCHAIN = 1U << 5, |
| 124 | PERF_SAMPLE_ID = 1U << 6, |
| 125 | PERF_SAMPLE_CPU = 1U << 7, |
| 126 | PERF_SAMPLE_PERIOD = 1U << 8, |
| 127 | PERF_SAMPLE_STREAM_ID = 1U << 9, |
| 128 | PERF_SAMPLE_RAW = 1U << 10, |
| 129 | |
| 130 | PERF_SAMPLE_MAX = 1U << 11, /* non-ABI */ |
| 131 | }; |
| 132 | |
| 133 | /* |
| 134 | * The format of the data returned by read() on a perf counter fd, |
| 135 | * as specified by attr.read_format: |
| 136 | * |
| 137 | * struct read_format { |
| 138 | * { u64 value; |
| 139 | * { u64 time_enabled; } && PERF_FORMAT_ENABLED |
| 140 | * { u64 time_running; } && PERF_FORMAT_RUNNING |
| 141 | * { u64 id; } && PERF_FORMAT_ID |
| 142 | * } && !PERF_FORMAT_GROUP |
| 143 | * |
| 144 | * { u64 nr; |
| 145 | * { u64 time_enabled; } && PERF_FORMAT_ENABLED |
| 146 | * { u64 time_running; } && PERF_FORMAT_RUNNING |
| 147 | * { u64 value; |
| 148 | * { u64 id; } && PERF_FORMAT_ID |
| 149 | * } cntr[nr]; |
| 150 | * } && PERF_FORMAT_GROUP |
| 151 | * }; |
| 152 | */ |
| 153 | enum perf_counter_read_format { |
| 154 | PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, |
| 155 | PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, |
| 156 | PERF_FORMAT_ID = 1U << 2, |
| 157 | PERF_FORMAT_GROUP = 1U << 3, |
| 158 | |
| 159 | PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ |
| 160 | }; |
| 161 | |
| 162 | #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ |
| 163 | |
| 164 | /* |
| 165 | * Hardware event to monitor via a performance monitoring counter: |
| 166 | */ |
| 167 | struct perf_counter_attr { |
| 168 | |
| 169 | /* |
| 170 | * Major type: hardware/software/tracepoint/etc. |
| 171 | */ |
| 172 | __u32 type; |
| 173 | |
| 174 | /* |
| 175 | * Size of the attr structure, for fwd/bwd compat. |
| 176 | */ |
| 177 | __u32 size; |
| 178 | |
| 179 | /* |
| 180 | * Type specific configuration information. |
| 181 | */ |
| 182 | __u64 config; |
| 183 | |
| 184 | union { |
| 185 | __u64 sample_period; |
| 186 | __u64 sample_freq; |
| 187 | }; |
| 188 | |
| 189 | __u64 sample_type; |
| 190 | __u64 read_format; |
| 191 | |
| 192 | __u64 disabled : 1, /* off by default */ |
| 193 | inherit : 1, /* children inherit it */ |
| 194 | pinned : 1, /* must always be on PMU */ |
| 195 | exclusive : 1, /* only group on PMU */ |
| 196 | exclude_user : 1, /* don't count user */ |
| 197 | exclude_kernel : 1, /* ditto kernel */ |
| 198 | exclude_hv : 1, /* ditto hypervisor */ |
| 199 | exclude_idle : 1, /* don't count when idle */ |
| 200 | mmap : 1, /* include mmap data */ |
| 201 | comm : 1, /* include comm data */ |
| 202 | freq : 1, /* use freq, not period */ |
| 203 | inherit_stat : 1, /* per task counts */ |
| 204 | enable_on_exec : 1, /* next exec enables */ |
| 205 | task : 1, /* trace fork/exit */ |
| 206 | watermark : 1, /* wakeup_watermark */ |
| 207 | |
| 208 | __reserved_1 : 49; |
| 209 | |
| 210 | union { |
| 211 | __u32 wakeup_events; /* wakeup every n events */ |
| 212 | __u32 wakeup_watermark; /* bytes before wakeup */ |
| 213 | }; |
| 214 | __u32 __reserved_2; |
| 215 | |
| 216 | __u64 __reserved_3; |
| 217 | }; |
| 218 | |
| 219 | /* |
| 220 | * Ioctls that can be done on a perf counter fd: |
| 221 | */ |
| 222 | #define PERF_COUNTER_IOC_ENABLE _IO ('$', 0) |
| 223 | #define PERF_COUNTER_IOC_DISABLE _IO ('$', 1) |
| 224 | #define PERF_COUNTER_IOC_REFRESH _IO ('$', 2) |
| 225 | #define PERF_COUNTER_IOC_RESET _IO ('$', 3) |
| 226 | #define PERF_COUNTER_IOC_PERIOD _IOW('$', 4, u64) |
| 227 | #define PERF_COUNTER_IOC_SET_OUTPUT _IO ('$', 5) |
| 228 | |
| 229 | enum perf_counter_ioc_flags { |
| 230 | PERF_IOC_FLAG_GROUP = 1U << 0, |
| 231 | }; |
| 232 | |
| 233 | /* |
| 234 | * Structure of the page that can be mapped via mmap |
| 235 | */ |
| 236 | struct perf_counter_mmap_page { |
| 237 | __u32 version; /* version number of this structure */ |
| 238 | __u32 compat_version; /* lowest version this is compat with */ |
| 239 | |
| 240 | /* |
| 241 | * Bits needed to read the hw counters in user-space. |
| 242 | * |
| 243 | * u32 seq; |
| 244 | * s64 count; |
| 245 | * |
| 246 | * do { |
| 247 | * seq = pc->lock; |
| 248 | * |
| 249 | * barrier() |
| 250 | * if (pc->index) { |
| 251 | * count = pmc_read(pc->index - 1); |
| 252 | * count += pc->offset; |
| 253 | * } else |
| 254 | * goto regular_read; |
| 255 | * |
| 256 | * barrier(); |
| 257 | * } while (pc->lock != seq); |
| 258 | * |
| 259 | * NOTE: for obvious reason this only works on self-monitoring |
| 260 | * processes. |
| 261 | */ |
| 262 | __u32 lock; /* seqlock for synchronization */ |
| 263 | __u32 index; /* hardware counter identifier */ |
| 264 | __s64 offset; /* add to hardware counter value */ |
| 265 | __u64 time_enabled; /* time counter active */ |
| 266 | __u64 time_running; /* time counter on cpu */ |
| 267 | |
| 268 | /* |
| 269 | * Hole for extension of the self monitor capabilities |
| 270 | */ |
| 271 | |
| 272 | __u64 __reserved[123]; /* align to 1k */ |
| 273 | |
| 274 | /* |
| 275 | * Control data for the mmap() data buffer. |
| 276 | * |
| 277 | * User-space reading the @data_head value should issue an rmb(), on |
| 278 | * SMP capable platforms, after reading this value -- see |
| 279 | * perf_counter_wakeup(). |
| 280 | * |
| 281 | * When the mapping is PROT_WRITE the @data_tail value should be |
| 282 | * written by userspace to reflect the last read data. In this case |
| 283 | * the kernel will not over-write unread data. |
| 284 | */ |
| 285 | __u64 data_head; /* head in the data section */ |
| 286 | __u64 data_tail; /* user-space written tail */ |
| 287 | }; |
| 288 | |
| 289 | #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0) |
| 290 | #define PERF_EVENT_MISC_CPUMODE_UNKNOWN (0 << 0) |
| 291 | #define PERF_EVENT_MISC_KERNEL (1 << 0) |
| 292 | #define PERF_EVENT_MISC_USER (2 << 0) |
| 293 | #define PERF_EVENT_MISC_HYPERVISOR (3 << 0) |
| 294 | |
| 295 | struct perf_event_header { |
| 296 | __u32 type; |
| 297 | __u16 misc; |
| 298 | __u16 size; |
| 299 | }; |
| 300 | |
| 301 | enum perf_event_type { |
| 302 | |
| 303 | /* |
| 304 | * The MMAP events record the PROT_EXEC mappings so that we can |
| 305 | * correlate userspace IPs to code. They have the following structure: |
| 306 | * |
| 307 | * struct { |
| 308 | * struct perf_event_header header; |
| 309 | * |
| 310 | * u32 pid, tid; |
| 311 | * u64 addr; |
| 312 | * u64 len; |
| 313 | * u64 pgoff; |
| 314 | * char filename[]; |
| 315 | * }; |
| 316 | */ |
| 317 | PERF_EVENT_MMAP = 1, |
| 318 | |
| 319 | /* |
| 320 | * struct { |
| 321 | * struct perf_event_header header; |
| 322 | * u64 id; |
| 323 | * u64 lost; |
| 324 | * }; |
| 325 | */ |
| 326 | PERF_EVENT_LOST = 2, |
| 327 | |
| 328 | /* |
| 329 | * struct { |
| 330 | * struct perf_event_header header; |
| 331 | * |
| 332 | * u32 pid, tid; |
| 333 | * char comm[]; |
| 334 | * }; |
| 335 | */ |
| 336 | PERF_EVENT_COMM = 3, |
| 337 | |
| 338 | /* |
| 339 | * struct { |
| 340 | * struct perf_event_header header; |
| 341 | * u32 pid, ppid; |
| 342 | * u32 tid, ptid; |
| 343 | * u64 time; |
| 344 | * }; |
| 345 | */ |
| 346 | PERF_EVENT_EXIT = 4, |
| 347 | |
| 348 | /* |
| 349 | * struct { |
| 350 | * struct perf_event_header header; |
| 351 | * u64 time; |
| 352 | * u64 id; |
| 353 | * u64 stream_id; |
| 354 | * }; |
| 355 | */ |
| 356 | PERF_EVENT_THROTTLE = 5, |
| 357 | PERF_EVENT_UNTHROTTLE = 6, |
| 358 | |
| 359 | /* |
| 360 | * struct { |
| 361 | * struct perf_event_header header; |
| 362 | * u32 pid, ppid; |
| 363 | * u32 tid, ptid; |
Anton Blanchard | a6f10a2 | 2009-09-22 22:34:24 +1000 | [diff] [blame^] | 364 | * u64 time; |
Ingo Molnar | 57c0c15 | 2009-09-21 12:20:38 +0200 | [diff] [blame] | 365 | * }; |
| 366 | */ |
| 367 | PERF_EVENT_FORK = 7, |
| 368 | |
| 369 | /* |
| 370 | * struct { |
| 371 | * struct perf_event_header header; |
| 372 | * u32 pid, tid; |
| 373 | * |
| 374 | * struct read_format values; |
| 375 | * }; |
| 376 | */ |
| 377 | PERF_EVENT_READ = 8, |
| 378 | |
| 379 | /* |
| 380 | * struct { |
| 381 | * struct perf_event_header header; |
| 382 | * |
| 383 | * { u64 ip; } && PERF_SAMPLE_IP |
| 384 | * { u32 pid, tid; } && PERF_SAMPLE_TID |
| 385 | * { u64 time; } && PERF_SAMPLE_TIME |
| 386 | * { u64 addr; } && PERF_SAMPLE_ADDR |
| 387 | * { u64 id; } && PERF_SAMPLE_ID |
| 388 | * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID |
| 389 | * { u32 cpu, res; } && PERF_SAMPLE_CPU |
| 390 | * { u64 period; } && PERF_SAMPLE_PERIOD |
| 391 | * |
| 392 | * { struct read_format values; } && PERF_SAMPLE_READ |
| 393 | * |
| 394 | * { u64 nr, |
| 395 | * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN |
| 396 | * |
| 397 | * # |
| 398 | * # The RAW record below is opaque data wrt the ABI |
| 399 | * # |
| 400 | * # That is, the ABI doesn't make any promises wrt to |
| 401 | * # the stability of its content, it may vary depending |
| 402 | * # on event, hardware, kernel version and phase of |
| 403 | * # the moon. |
| 404 | * # |
| 405 | * # In other words, PERF_SAMPLE_RAW contents are not an ABI. |
| 406 | * # |
| 407 | * |
| 408 | * { u32 size; |
| 409 | * char data[size];}&& PERF_SAMPLE_RAW |
| 410 | * }; |
| 411 | */ |
| 412 | PERF_EVENT_SAMPLE = 9, |
| 413 | |
| 414 | PERF_EVENT_MAX, /* non-ABI */ |
| 415 | }; |
| 416 | |
| 417 | enum perf_callchain_context { |
| 418 | PERF_CONTEXT_HV = (__u64)-32, |
| 419 | PERF_CONTEXT_KERNEL = (__u64)-128, |
| 420 | PERF_CONTEXT_USER = (__u64)-512, |
| 421 | |
| 422 | PERF_CONTEXT_GUEST = (__u64)-2048, |
| 423 | PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, |
| 424 | PERF_CONTEXT_GUEST_USER = (__u64)-2560, |
| 425 | |
| 426 | PERF_CONTEXT_MAX = (__u64)-4095, |
| 427 | }; |
| 428 | |
| 429 | #define PERF_FLAG_FD_NO_GROUP (1U << 0) |
| 430 | #define PERF_FLAG_FD_OUTPUT (1U << 1) |
| 431 | |
| 432 | /* |
| 433 | * In case some app still references the old symbols: |
| 434 | */ |
| 435 | |
| 436 | #define __NR_perf_counter_open __NR_perf_event_open |
| 437 | |
| 438 | #define PR_TASK_PERF_COUNTERS_DISABLE PR_TASK_PERF_EVENTS_DISABLE |
| 439 | #define PR_TASK_PERF_COUNTERS_ENABLE PR_TASK_PERF_EVENTS_ENABLE |
| 440 | |
| 441 | #endif /* _LINUX_PERF_COUNTER_H */ |