1 /* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <[email protected]> 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 PERF_COUNT_HW_REF_CPU_CYCLES = 9, 58 59 PERF_COUNT_HW_MAX, /* non-ABI */ 60 }; 61 62 /* 63 * Generalized hardware cache events: 64 * 65 * { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } 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 PERF_COUNT_HW_CACHE_NODE = 6, 77 78 PERF_COUNT_HW_CACHE_MAX, /* non-ABI */ 79 }; 80 81 enum perf_hw_cache_op_id { 82 PERF_COUNT_HW_CACHE_OP_READ = 0, 83 PERF_COUNT_HW_CACHE_OP_WRITE = 1, 84 PERF_COUNT_HW_CACHE_OP_PREFETCH = 2, 85 86 PERF_COUNT_HW_CACHE_OP_MAX, /* non-ABI */ 87 }; 88 89 enum perf_hw_cache_op_result_id { 90 PERF_COUNT_HW_CACHE_RESULT_ACCESS = 0, 91 PERF_COUNT_HW_CACHE_RESULT_MISS = 1, 92 93 PERF_COUNT_HW_CACHE_RESULT_MAX, /* non-ABI */ 94 }; 95 96 /* 97 * Special "software" events provided by the kernel, even if the hardware 98 * does not support performance events. These events measure various 99 * physical and sw events of the kernel (and allow the profiling of them as 100 * well): 101 */ 102 enum perf_sw_ids { 103 PERF_COUNT_SW_CPU_CLOCK = 0, 104 PERF_COUNT_SW_TASK_CLOCK = 1, 105 PERF_COUNT_SW_PAGE_FAULTS = 2, 106 PERF_COUNT_SW_CONTEXT_SWITCHES = 3, 107 PERF_COUNT_SW_CPU_MIGRATIONS = 4, 108 PERF_COUNT_SW_PAGE_FAULTS_MIN = 5, 109 PERF_COUNT_SW_PAGE_FAULTS_MAJ = 6, 110 PERF_COUNT_SW_ALIGNMENT_FAULTS = 7, 111 PERF_COUNT_SW_EMULATION_FAULTS = 8, 112 113 PERF_COUNT_SW_MAX, /* non-ABI */ 114 }; 115 116 /* 117 * Bits that can be set in attr.sample_type to request information 118 * in the overflow packets. 119 */ 120 enum perf_event_sample_format { 121 PERF_SAMPLE_IP = 1U << 0, 122 PERF_SAMPLE_TID = 1U << 1, 123 PERF_SAMPLE_TIME = 1U << 2, 124 PERF_SAMPLE_ADDR = 1U << 3, 125 PERF_SAMPLE_READ = 1U << 4, 126 PERF_SAMPLE_CALLCHAIN = 1U << 5, 127 PERF_SAMPLE_ID = 1U << 6, 128 PERF_SAMPLE_CPU = 1U << 7, 129 PERF_SAMPLE_PERIOD = 1U << 8, 130 PERF_SAMPLE_STREAM_ID = 1U << 9, 131 PERF_SAMPLE_RAW = 1U << 10, 132 PERF_SAMPLE_BRANCH_STACK = 1U << 11, 133 PERF_SAMPLE_REGS_USER = 1U << 12, 134 PERF_SAMPLE_STACK_USER = 1U << 13, 135 136 PERF_SAMPLE_MAX = 1U << 14, /* non-ABI */ 137 }; 138 139 /* 140 * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set 141 * 142 * If the user does not pass priv level information via branch_sample_type, 143 * the kernel uses the event's priv level. Branch and event priv levels do 144 * not have to match. Branch priv level is checked for permissions. 145 * 146 * The branch types can be combined, however BRANCH_ANY covers all types 147 * of branches and therefore it supersedes all the other types. 148 */ 149 enum perf_branch_sample_type { 150 PERF_SAMPLE_BRANCH_USER = 1U << 0, /* user branches */ 151 PERF_SAMPLE_BRANCH_KERNEL = 1U << 1, /* kernel branches */ 152 PERF_SAMPLE_BRANCH_HV = 1U << 2, /* hypervisor branches */ 153 154 PERF_SAMPLE_BRANCH_ANY = 1U << 3, /* any branch types */ 155 PERF_SAMPLE_BRANCH_ANY_CALL = 1U << 4, /* any call branch */ 156 PERF_SAMPLE_BRANCH_ANY_RETURN = 1U << 5, /* any return branch */ 157 PERF_SAMPLE_BRANCH_IND_CALL = 1U << 6, /* indirect calls */ 158 159 PERF_SAMPLE_BRANCH_MAX = 1U << 7, /* non-ABI */ 160 }; 161 162 #define PERF_SAMPLE_BRANCH_PLM_ALL \ 163 (PERF_SAMPLE_BRANCH_USER|\ 164 PERF_SAMPLE_BRANCH_KERNEL|\ 165 PERF_SAMPLE_BRANCH_HV) 166 167 /* 168 * Values to determine ABI of the registers dump. 169 */ 170 enum perf_sample_regs_abi { 171 PERF_SAMPLE_REGS_ABI_NONE = 0, 172 PERF_SAMPLE_REGS_ABI_32 = 1, 173 PERF_SAMPLE_REGS_ABI_64 = 2, 174 }; 175 176 /* 177 * The format of the data returned by read() on a perf event fd, 178 * as specified by attr.read_format: 179 * 180 * struct read_format { 181 * { u64 value; 182 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 183 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 184 * { u64 id; } && PERF_FORMAT_ID 185 * } && !PERF_FORMAT_GROUP 186 * 187 * { u64 nr; 188 * { u64 time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED 189 * { u64 time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING 190 * { u64 value; 191 * { u64 id; } && PERF_FORMAT_ID 192 * } cntr[nr]; 193 * } && PERF_FORMAT_GROUP 194 * }; 195 */ 196 enum perf_event_read_format { 197 PERF_FORMAT_TOTAL_TIME_ENABLED = 1U << 0, 198 PERF_FORMAT_TOTAL_TIME_RUNNING = 1U << 1, 199 PERF_FORMAT_ID = 1U << 2, 200 PERF_FORMAT_GROUP = 1U << 3, 201 202 PERF_FORMAT_MAX = 1U << 4, /* non-ABI */ 203 }; 204 205 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */ 206 #define PERF_ATTR_SIZE_VER1 72 /* add: config2 */ 207 #define PERF_ATTR_SIZE_VER2 80 /* add: branch_sample_type */ 208 #define PERF_ATTR_SIZE_VER3 96 /* add: sample_regs_user */ 209 /* add: sample_stack_user */ 210 211 /* 212 * Hardware event_id to monitor via a performance monitoring event: 213 */ 214 struct perf_event_attr { 215 216 /* 217 * Major type: hardware/software/tracepoint/etc. 218 */ 219 __u32 type; 220 221 /* 222 * Size of the attr structure, for fwd/bwd compat. 223 */ 224 __u32 size; 225 226 /* 227 * Type specific configuration information. 228 */ 229 __u64 config; 230 231 union { 232 __u64 sample_period; 233 __u64 sample_freq; 234 }; 235 236 __u64 sample_type; 237 __u64 read_format; 238 239 __u64 disabled : 1, /* off by default */ 240 inherit : 1, /* children inherit it */ 241 pinned : 1, /* must always be on PMU */ 242 exclusive : 1, /* only group on PMU */ 243 exclude_user : 1, /* don't count user */ 244 exclude_kernel : 1, /* ditto kernel */ 245 exclude_hv : 1, /* ditto hypervisor */ 246 exclude_idle : 1, /* don't count when idle */ 247 mmap : 1, /* include mmap data */ 248 comm : 1, /* include comm data */ 249 freq : 1, /* use freq, not period */ 250 inherit_stat : 1, /* per task counts */ 251 enable_on_exec : 1, /* next exec enables */ 252 task : 1, /* trace fork/exit */ 253 watermark : 1, /* wakeup_watermark */ 254 /* 255 * precise_ip: 256 * 257 * 0 - SAMPLE_IP can have arbitrary skid 258 * 1 - SAMPLE_IP must have constant skid 259 * 2 - SAMPLE_IP requested to have 0 skid 260 * 3 - SAMPLE_IP must have 0 skid 261 * 262 * See also PERF_RECORD_MISC_EXACT_IP 263 */ 264 precise_ip : 2, /* skid constraint */ 265 mmap_data : 1, /* non-exec mmap data */ 266 sample_id_all : 1, /* sample_type all events */ 267 268 exclude_host : 1, /* don't count in host */ 269 exclude_guest : 1, /* don't count in guest */ 270 271 exclude_callchain_kernel : 1, /* exclude kernel callchains */ 272 exclude_callchain_user : 1, /* exclude user callchains */ 273 274 __reserved_1 : 41; 275 276 union { 277 __u32 wakeup_events; /* wakeup every n events */ 278 __u32 wakeup_watermark; /* bytes before wakeup */ 279 }; 280 281 __u32 bp_type; 282 union { 283 __u64 bp_addr; 284 __u64 config1; /* extension of config */ 285 }; 286 union { 287 __u64 bp_len; 288 __u64 config2; /* extension of config1 */ 289 }; 290 __u64 branch_sample_type; /* enum perf_branch_sample_type */ 291 292 /* 293 * Defines set of user regs to dump on samples. 294 * See asm/perf_regs.h for details. 295 */ 296 __u64 sample_regs_user; 297 298 /* 299 * Defines size of the user stack to dump on samples. 300 */ 301 __u32 sample_stack_user; 302 303 /* Align to u64. */ 304 __u32 __reserved_2; 305 }; 306 307 #define perf_flags(attr) (*(&(attr)->read_format + 1)) 308 309 /* 310 * Ioctls that can be done on a perf event fd: 311 */ 312 #define PERF_EVENT_IOC_ENABLE _IO ('$', 0) 313 #define PERF_EVENT_IOC_DISABLE _IO ('$', 1) 314 #define PERF_EVENT_IOC_REFRESH _IO ('$', 2) 315 #define PERF_EVENT_IOC_RESET _IO ('$', 3) 316 #define PERF_EVENT_IOC_PERIOD _IOW('$', 4, __u64) 317 #define PERF_EVENT_IOC_SET_OUTPUT _IO ('$', 5) 318 #define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *) 319 320 enum perf_event_ioc_flags { 321 PERF_IOC_FLAG_GROUP = 1U << 0, 322 }; 323 324 /* 325 * Structure of the page that can be mapped via mmap 326 */ 327 struct perf_event_mmap_page { 328 __u32 version; /* version number of this structure */ 329 __u32 compat_version; /* lowest version this is compat with */ 330 331 /* 332 * Bits needed to read the hw events in user-space. 333 * 334 * u32 seq, time_mult, time_shift, idx, width; 335 * u64 count, enabled, running; 336 * u64 cyc, time_offset; 337 * s64 pmc = 0; 338 * 339 * do { 340 * seq = pc->lock; 341 * barrier() 342 * 343 * enabled = pc->time_enabled; 344 * running = pc->time_running; 345 * 346 * if (pc->cap_usr_time && enabled != running) { 347 * cyc = rdtsc(); 348 * time_offset = pc->time_offset; 349 * time_mult = pc->time_mult; 350 * time_shift = pc->time_shift; 351 * } 352 * 353 * idx = pc->index; 354 * count = pc->offset; 355 * if (pc->cap_usr_rdpmc && idx) { 356 * width = pc->pmc_width; 357 * pmc = rdpmc(idx - 1); 358 * } 359 * 360 * barrier(); 361 * } while (pc->lock != seq); 362 * 363 * NOTE: for obvious reason this only works on self-monitoring 364 * processes. 365 */ 366 __u32 lock; /* seqlock for synchronization */ 367 __u32 index; /* hardware event identifier */ 368 __s64 offset; /* add to hardware event value */ 369 __u64 time_enabled; /* time event active */ 370 __u64 time_running; /* time event on cpu */ 371 union { 372 __u64 capabilities; 373 __u64 cap_usr_time : 1, 374 cap_usr_rdpmc : 1, 375 cap_____res : 62; 376 }; 377 378 /* 379 * If cap_usr_rdpmc this field provides the bit-width of the value 380 * read using the rdpmc() or equivalent instruction. This can be used 381 * to sign extend the result like: 382 * 383 * pmc <<= 64 - width; 384 * pmc >>= 64 - width; // signed shift right 385 * count += pmc; 386 */ 387 __u16 pmc_width; 388 389 /* 390 * If cap_usr_time the below fields can be used to compute the time 391 * delta since time_enabled (in ns) using rdtsc or similar. 392 * 393 * u64 quot, rem; 394 * u64 delta; 395 * 396 * quot = (cyc >> time_shift); 397 * rem = cyc & ((1 << time_shift) - 1); 398 * delta = time_offset + quot * time_mult + 399 * ((rem * time_mult) >> time_shift); 400 * 401 * Where time_offset,time_mult,time_shift and cyc are read in the 402 * seqcount loop described above. This delta can then be added to 403 * enabled and possible running (if idx), improving the scaling: 404 * 405 * enabled += delta; 406 * if (idx) 407 * running += delta; 408 * 409 * quot = count / running; 410 * rem = count % running; 411 * count = quot * enabled + (rem * enabled) / running; 412 */ 413 __u16 time_shift; 414 __u32 time_mult; 415 __u64 time_offset; 416 417 /* 418 * Hole for extension of the self monitor capabilities 419 */ 420 421 __u64 __reserved[120]; /* align to 1k */ 422 423 /* 424 * Control data for the mmap() data buffer. 425 * 426 * User-space reading the @data_head value should issue an rmb(), on 427 * SMP capable platforms, after reading this value -- see 428 * perf_event_wakeup(). 429 * 430 * When the mapping is PROT_WRITE the @data_tail value should be 431 * written by userspace to reflect the last read data. In this case 432 * the kernel will not over-write unread data. 433 */ 434 __u64 data_head; /* head in the data section */ 435 __u64 data_tail; /* user-space written tail */ 436 __u64 data_offset; /* where the buffer starts */ 437 __u64 data_size; /* data buffer size */ 438 439 /* 440 * AUX area is defined by aux_{offset,size} fields that should be set 441 * by the userspace, so that 442 * 443 * aux_offset >= data_offset + data_size 444 * 445 * prior to mmap()ing it. Size of the mmap()ed area should be aux_size. 446 * 447 * Ring buffer pointers aux_{head,tail} have the same semantics as 448 * data_{head,tail} and same ordering rules apply. 449 */ 450 __u64 aux_head; 451 __u64 aux_tail; 452 __u64 aux_offset; 453 __u64 aux_size; 454 }; 455 456 #define PERF_RECORD_MISC_CPUMODE_MASK (7 << 0) 457 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN (0 << 0) 458 #define PERF_RECORD_MISC_KERNEL (1 << 0) 459 #define PERF_RECORD_MISC_USER (2 << 0) 460 #define PERF_RECORD_MISC_HYPERVISOR (3 << 0) 461 #define PERF_RECORD_MISC_GUEST_KERNEL (4 << 0) 462 #define PERF_RECORD_MISC_GUEST_USER (5 << 0) 463 464 /* 465 * Indicates that the content of PERF_SAMPLE_IP points to 466 * the actual instruction that triggered the event. See also 467 * perf_event_attr::precise_ip. 468 */ 469 #define PERF_RECORD_MISC_EXACT_IP (1 << 14) 470 /* 471 * Reserve the last bit to indicate some extended misc field 472 */ 473 #define PERF_RECORD_MISC_EXT_RESERVED (1 << 15) 474 475 struct perf_event_header { 476 __u32 type; 477 __u16 misc; 478 __u16 size; 479 }; 480 481 enum perf_event_type { 482 483 /* 484 * If perf_event_attr.sample_id_all is set then all event types will 485 * have the sample_type selected fields related to where/when 486 * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID) 487 * described in PERF_RECORD_SAMPLE below, it will be stashed just after 488 * the perf_event_header and the fields already present for the existing 489 * fields, i.e. at the end of the payload. That way a newer perf.data 490 * file will be supported by older perf tools, with these new optional 491 * fields being ignored. 492 * 493 * The MMAP events record the PROT_EXEC mappings so that we can 494 * correlate userspace IPs to code. They have the following structure: 495 * 496 * struct { 497 * struct perf_event_header header; 498 * 499 * u32 pid, tid; 500 * u64 addr; 501 * u64 len; 502 * u64 pgoff; 503 * char filename[]; 504 * }; 505 */ 506 PERF_RECORD_MMAP = 1, 507 508 /* 509 * struct { 510 * struct perf_event_header header; 511 * u64 id; 512 * u64 lost; 513 * }; 514 */ 515 PERF_RECORD_LOST = 2, 516 517 /* 518 * struct { 519 * struct perf_event_header header; 520 * 521 * u32 pid, tid; 522 * char comm[]; 523 * }; 524 */ 525 PERF_RECORD_COMM = 3, 526 527 /* 528 * struct { 529 * struct perf_event_header header; 530 * u32 pid, ppid; 531 * u32 tid, ptid; 532 * u64 time; 533 * }; 534 */ 535 PERF_RECORD_EXIT = 4, 536 537 /* 538 * struct { 539 * struct perf_event_header header; 540 * u64 time; 541 * u64 id; 542 * u64 stream_id; 543 * }; 544 */ 545 PERF_RECORD_THROTTLE = 5, 546 PERF_RECORD_UNTHROTTLE = 6, 547 548 /* 549 * struct { 550 * struct perf_event_header header; 551 * u32 pid, ppid; 552 * u32 tid, ptid; 553 * u64 time; 554 * }; 555 */ 556 PERF_RECORD_FORK = 7, 557 558 /* 559 * struct { 560 * struct perf_event_header header; 561 * u32 pid, tid; 562 * 563 * struct read_format values; 564 * }; 565 */ 566 PERF_RECORD_READ = 8, 567 568 /* 569 * struct { 570 * struct perf_event_header header; 571 * 572 * { u64 ip; } && PERF_SAMPLE_IP 573 * { u32 pid, tid; } && PERF_SAMPLE_TID 574 * { u64 time; } && PERF_SAMPLE_TIME 575 * { u64 addr; } && PERF_SAMPLE_ADDR 576 * { u64 id; } && PERF_SAMPLE_ID 577 * { u64 stream_id;} && PERF_SAMPLE_STREAM_ID 578 * { u32 cpu, res; } && PERF_SAMPLE_CPU 579 * { u64 period; } && PERF_SAMPLE_PERIOD 580 * 581 * { struct read_format values; } && PERF_SAMPLE_READ 582 * 583 * { u64 nr, 584 * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN 585 * 586 * # 587 * # The RAW record below is opaque data wrt the ABI 588 * # 589 * # That is, the ABI doesn't make any promises wrt to 590 * # the stability of its content, it may vary depending 591 * # on event, hardware, kernel version and phase of 592 * # the moon. 593 * # 594 * # In other words, PERF_SAMPLE_RAW contents are not an ABI. 595 * # 596 * 597 * { u32 size; 598 * char data[size];}&& PERF_SAMPLE_RAW 599 * 600 * { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK 601 * 602 * { u64 abi; # enum perf_sample_regs_abi 603 * u64 regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER 604 * 605 * { u64 size; 606 * char data[size]; 607 * u64 dyn_size; } && PERF_SAMPLE_STACK_USER 608 * }; 609 */ 610 PERF_RECORD_SAMPLE = 9, 611 612 PERF_RECORD_MAX, /* non-ABI */ 613 }; 614 615 #define PERF_MAX_STACK_DEPTH 127 616 617 enum perf_callchain_context { 618 PERF_CONTEXT_HV = (__u64)-32, 619 PERF_CONTEXT_KERNEL = (__u64)-128, 620 PERF_CONTEXT_USER = (__u64)-512, 621 622 PERF_CONTEXT_GUEST = (__u64)-2048, 623 PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, 624 PERF_CONTEXT_GUEST_USER = (__u64)-2560, 625 626 PERF_CONTEXT_MAX = (__u64)-4095, 627 }; 628 629 #define PERF_FLAG_FD_NO_GROUP (1U << 0) 630 #define PERF_FLAG_FD_OUTPUT (1U << 1) 631 #define PERF_FLAG_PID_CGROUP (1U << 2) /* pid=cgroup id, per-cpu mode only */ 632 633 #endif /* _LINUX_PERF_EVENT_H */ 634