1 /* 2 * Performance events: 3 * 4 * Copyright (C) 2008-2009, Thomas Gleixner <[email protected]> 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 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 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 */ 66 enum 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 77 enum 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 85 enum 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 */ 98 enum 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 */ 116 enum 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 */ 152 enum 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 */ 166 struct 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 *) 241 242 enum 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 */ 249 struct 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) 320 321 struct perf_event_header { 322 __u32 type; 323 __u16 misc; 324 __u16 size; 325 }; 326 327 enum 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 }; 442 443 enum 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 #ifdef __KERNEL__ 459 /* 460 * Kernel-internal data types and definitions: 461 */ 462 463 #ifdef CONFIG_PERF_EVENTS 464 # include <asm/perf_event.h> 465 # include <asm/local64.h> 466 #endif 467 468 struct perf_guest_info_callbacks { 469 int (*is_in_guest) (void); 470 int (*is_user_mode) (void); 471 unsigned long (*get_guest_ip) (void); 472 }; 473 474 #ifdef CONFIG_HAVE_HW_BREAKPOINT 475 #include <asm/hw_breakpoint.h> 476 #endif 477 478 #include <linux/list.h> 479 #include <linux/mutex.h> 480 #include <linux/rculist.h> 481 #include <linux/rcupdate.h> 482 #include <linux/spinlock.h> 483 #include <linux/hrtimer.h> 484 #include <linux/fs.h> 485 #include <linux/pid_namespace.h> 486 #include <linux/workqueue.h> 487 #include <linux/ftrace.h> 488 #include <linux/cpu.h> 489 #include <asm/atomic.h> 490 #include <asm/local.h> 491 492 #define PERF_MAX_STACK_DEPTH 255 493 494 struct perf_callchain_entry { 495 __u64 nr; 496 __u64 ip[PERF_MAX_STACK_DEPTH]; 497 }; 498 499 struct perf_raw_record { 500 u32 size; 501 void *data; 502 }; 503 504 struct perf_branch_entry { 505 __u64 from; 506 __u64 to; 507 __u64 flags; 508 }; 509 510 struct perf_branch_stack { 511 __u64 nr; 512 struct perf_branch_entry entries[0]; 513 }; 514 515 struct task_struct; 516 517 /** 518 * struct hw_perf_event - performance event hardware details: 519 */ 520 struct hw_perf_event { 521 #ifdef CONFIG_PERF_EVENTS 522 union { 523 struct { /* hardware */ 524 u64 config; 525 u64 last_tag; 526 unsigned long config_base; 527 unsigned long event_base; 528 int idx; 529 int last_cpu; 530 }; 531 struct { /* software */ 532 s64 remaining; 533 struct hrtimer hrtimer; 534 }; 535 #ifdef CONFIG_HAVE_HW_BREAKPOINT 536 struct { /* breakpoint */ 537 struct arch_hw_breakpoint info; 538 struct list_head bp_list; 539 }; 540 #endif 541 }; 542 local64_t prev_count; 543 u64 sample_period; 544 u64 last_period; 545 local64_t period_left; 546 u64 interrupts; 547 548 u64 freq_time_stamp; 549 u64 freq_count_stamp; 550 #endif 551 }; 552 553 struct perf_event; 554 555 /* 556 * Common implementation detail of pmu::{start,commit,cancel}_txn 557 */ 558 #define PERF_EVENT_TXN 0x1 559 560 /** 561 * struct pmu - generic performance monitoring unit 562 */ 563 struct pmu { 564 int (*enable) (struct perf_event *event); 565 void (*disable) (struct perf_event *event); 566 int (*start) (struct perf_event *event); 567 void (*stop) (struct perf_event *event); 568 void (*read) (struct perf_event *event); 569 void (*unthrottle) (struct perf_event *event); 570 571 /* 572 * Group events scheduling is treated as a transaction, add group 573 * events as a whole and perform one schedulability test. If the test 574 * fails, roll back the whole group 575 */ 576 577 /* 578 * Start the transaction, after this ->enable() doesn't need 579 * to do schedulability tests. 580 */ 581 void (*start_txn) (const struct pmu *pmu); 582 /* 583 * If ->start_txn() disabled the ->enable() schedulability test 584 * then ->commit_txn() is required to perform one. On success 585 * the transaction is closed. On error the transaction is kept 586 * open until ->cancel_txn() is called. 587 */ 588 int (*commit_txn) (const struct pmu *pmu); 589 /* 590 * Will cancel the transaction, assumes ->disable() is called for 591 * each successfull ->enable() during the transaction. 592 */ 593 void (*cancel_txn) (const struct pmu *pmu); 594 }; 595 596 /** 597 * enum perf_event_active_state - the states of a event 598 */ 599 enum perf_event_active_state { 600 PERF_EVENT_STATE_ERROR = -2, 601 PERF_EVENT_STATE_OFF = -1, 602 PERF_EVENT_STATE_INACTIVE = 0, 603 PERF_EVENT_STATE_ACTIVE = 1, 604 }; 605 606 struct file; 607 608 #define PERF_BUFFER_WRITABLE 0x01 609 610 struct perf_buffer { 611 atomic_t refcount; 612 struct rcu_head rcu_head; 613 #ifdef CONFIG_PERF_USE_VMALLOC 614 struct work_struct work; 615 int page_order; /* allocation order */ 616 #endif 617 int nr_pages; /* nr of data pages */ 618 int writable; /* are we writable */ 619 620 atomic_t poll; /* POLL_ for wakeups */ 621 622 local_t head; /* write position */ 623 local_t nest; /* nested writers */ 624 local_t events; /* event limit */ 625 local_t wakeup; /* wakeup stamp */ 626 local_t lost; /* nr records lost */ 627 628 long watermark; /* wakeup watermark */ 629 630 struct perf_event_mmap_page *user_page; 631 void *data_pages[0]; 632 }; 633 634 struct perf_pending_entry { 635 struct perf_pending_entry *next; 636 void (*func)(struct perf_pending_entry *); 637 }; 638 639 struct perf_sample_data; 640 641 typedef void (*perf_overflow_handler_t)(struct perf_event *, int, 642 struct perf_sample_data *, 643 struct pt_regs *regs); 644 645 enum perf_group_flag { 646 PERF_GROUP_SOFTWARE = 0x1, 647 }; 648 649 #define SWEVENT_HLIST_BITS 8 650 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS) 651 652 struct swevent_hlist { 653 struct hlist_head heads[SWEVENT_HLIST_SIZE]; 654 struct rcu_head rcu_head; 655 }; 656 657 #define PERF_ATTACH_CONTEXT 0x01 658 #define PERF_ATTACH_GROUP 0x02 659 660 /** 661 * struct perf_event - performance event kernel representation: 662 */ 663 struct perf_event { 664 #ifdef CONFIG_PERF_EVENTS 665 struct list_head group_entry; 666 struct list_head event_entry; 667 struct list_head sibling_list; 668 struct hlist_node hlist_entry; 669 int nr_siblings; 670 int group_flags; 671 struct perf_event *group_leader; 672 const struct pmu *pmu; 673 674 enum perf_event_active_state state; 675 unsigned int attach_state; 676 local64_t count; 677 atomic64_t child_count; 678 679 /* 680 * These are the total time in nanoseconds that the event 681 * has been enabled (i.e. eligible to run, and the task has 682 * been scheduled in, if this is a per-task event) 683 * and running (scheduled onto the CPU), respectively. 684 * 685 * They are computed from tstamp_enabled, tstamp_running and 686 * tstamp_stopped when the event is in INACTIVE or ACTIVE state. 687 */ 688 u64 total_time_enabled; 689 u64 total_time_running; 690 691 /* 692 * These are timestamps used for computing total_time_enabled 693 * and total_time_running when the event is in INACTIVE or 694 * ACTIVE state, measured in nanoseconds from an arbitrary point 695 * in time. 696 * tstamp_enabled: the notional time when the event was enabled 697 * tstamp_running: the notional time when the event was scheduled on 698 * tstamp_stopped: in INACTIVE state, the notional time when the 699 * event was scheduled off. 700 */ 701 u64 tstamp_enabled; 702 u64 tstamp_running; 703 u64 tstamp_stopped; 704 705 struct perf_event_attr attr; 706 struct hw_perf_event hw; 707 708 struct perf_event_context *ctx; 709 struct file *filp; 710 711 /* 712 * These accumulate total time (in nanoseconds) that children 713 * events have been enabled and running, respectively. 714 */ 715 atomic64_t child_total_time_enabled; 716 atomic64_t child_total_time_running; 717 718 /* 719 * Protect attach/detach and child_list: 720 */ 721 struct mutex child_mutex; 722 struct list_head child_list; 723 struct perf_event *parent; 724 725 int oncpu; 726 int cpu; 727 728 struct list_head owner_entry; 729 struct task_struct *owner; 730 731 /* mmap bits */ 732 struct mutex mmap_mutex; 733 atomic_t mmap_count; 734 int mmap_locked; 735 struct user_struct *mmap_user; 736 struct perf_buffer *buffer; 737 738 /* poll related */ 739 wait_queue_head_t waitq; 740 struct fasync_struct *fasync; 741 742 /* delayed work for NMIs and such */ 743 int pending_wakeup; 744 int pending_kill; 745 int pending_disable; 746 struct perf_pending_entry pending; 747 748 atomic_t event_limit; 749 750 void (*destroy)(struct perf_event *); 751 struct rcu_head rcu_head; 752 753 struct pid_namespace *ns; 754 u64 id; 755 756 perf_overflow_handler_t overflow_handler; 757 758 #ifdef CONFIG_EVENT_TRACING 759 struct ftrace_event_call *tp_event; 760 struct event_filter *filter; 761 #endif 762 763 #endif /* CONFIG_PERF_EVENTS */ 764 }; 765 766 /** 767 * struct perf_event_context - event context structure 768 * 769 * Used as a container for task events and CPU events as well: 770 */ 771 struct perf_event_context { 772 /* 773 * Protect the states of the events in the list, 774 * nr_active, and the list: 775 */ 776 raw_spinlock_t lock; 777 /* 778 * Protect the list of events. Locking either mutex or lock 779 * is sufficient to ensure the list doesn't change; to change 780 * the list you need to lock both the mutex and the spinlock. 781 */ 782 struct mutex mutex; 783 784 struct list_head pinned_groups; 785 struct list_head flexible_groups; 786 struct list_head event_list; 787 int nr_events; 788 int nr_active; 789 int is_active; 790 int nr_stat; 791 atomic_t refcount; 792 struct task_struct *task; 793 794 /* 795 * Context clock, runs when context enabled. 796 */ 797 u64 time; 798 u64 timestamp; 799 800 /* 801 * These fields let us detect when two contexts have both 802 * been cloned (inherited) from a common ancestor. 803 */ 804 struct perf_event_context *parent_ctx; 805 u64 parent_gen; 806 u64 generation; 807 int pin_count; 808 struct rcu_head rcu_head; 809 }; 810 811 /** 812 * struct perf_event_cpu_context - per cpu event context structure 813 */ 814 struct perf_cpu_context { 815 struct perf_event_context ctx; 816 struct perf_event_context *task_ctx; 817 int active_oncpu; 818 int max_pertask; 819 int exclusive; 820 struct swevent_hlist *swevent_hlist; 821 struct mutex hlist_mutex; 822 int hlist_refcount; 823 824 /* 825 * Recursion avoidance: 826 * 827 * task, softirq, irq, nmi context 828 */ 829 int recursion[4]; 830 }; 831 832 struct perf_output_handle { 833 struct perf_event *event; 834 struct perf_buffer *buffer; 835 unsigned long wakeup; 836 unsigned long size; 837 void *addr; 838 int page; 839 int nmi; 840 int sample; 841 }; 842 843 #ifdef CONFIG_PERF_EVENTS 844 845 /* 846 * Set by architecture code: 847 */ 848 extern int perf_max_events; 849 850 extern const struct pmu *hw_perf_event_init(struct perf_event *event); 851 852 extern void perf_event_task_sched_in(struct task_struct *task); 853 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next); 854 extern void perf_event_task_tick(struct task_struct *task); 855 extern int perf_event_init_task(struct task_struct *child); 856 extern void perf_event_exit_task(struct task_struct *child); 857 extern void perf_event_free_task(struct task_struct *task); 858 extern void set_perf_event_pending(void); 859 extern void perf_event_do_pending(void); 860 extern void perf_event_print_debug(void); 861 extern void __perf_disable(void); 862 extern bool __perf_enable(void); 863 extern void perf_disable(void); 864 extern void perf_enable(void); 865 extern int perf_event_task_disable(void); 866 extern int perf_event_task_enable(void); 867 extern void perf_event_update_userpage(struct perf_event *event); 868 extern int perf_event_release_kernel(struct perf_event *event); 869 extern struct perf_event * 870 perf_event_create_kernel_counter(struct perf_event_attr *attr, 871 int cpu, 872 pid_t pid, 873 perf_overflow_handler_t callback); 874 extern u64 perf_event_read_value(struct perf_event *event, 875 u64 *enabled, u64 *running); 876 877 struct perf_sample_data { 878 u64 type; 879 880 u64 ip; 881 struct { 882 u32 pid; 883 u32 tid; 884 } tid_entry; 885 u64 time; 886 u64 addr; 887 u64 id; 888 u64 stream_id; 889 struct { 890 u32 cpu; 891 u32 reserved; 892 } cpu_entry; 893 u64 period; 894 struct perf_callchain_entry *callchain; 895 struct perf_raw_record *raw; 896 }; 897 898 static inline 899 void perf_sample_data_init(struct perf_sample_data *data, u64 addr) 900 { 901 data->addr = addr; 902 data->raw = NULL; 903 } 904 905 extern void perf_output_sample(struct perf_output_handle *handle, 906 struct perf_event_header *header, 907 struct perf_sample_data *data, 908 struct perf_event *event); 909 extern void perf_prepare_sample(struct perf_event_header *header, 910 struct perf_sample_data *data, 911 struct perf_event *event, 912 struct pt_regs *regs); 913 914 extern int perf_event_overflow(struct perf_event *event, int nmi, 915 struct perf_sample_data *data, 916 struct pt_regs *regs); 917 918 /* 919 * Return 1 for a software event, 0 for a hardware event 920 */ 921 static inline int is_software_event(struct perf_event *event) 922 { 923 switch (event->attr.type) { 924 case PERF_TYPE_SOFTWARE: 925 case PERF_TYPE_TRACEPOINT: 926 /* for now the breakpoint stuff also works as software event */ 927 case PERF_TYPE_BREAKPOINT: 928 return 1; 929 } 930 return 0; 931 } 932 933 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX]; 934 935 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64); 936 937 #ifndef perf_arch_fetch_caller_regs 938 static inline void 939 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { } 940 #endif 941 942 /* 943 * Take a snapshot of the regs. Skip ip and frame pointer to 944 * the nth caller. We only need a few of the regs: 945 * - ip for PERF_SAMPLE_IP 946 * - cs for user_mode() tests 947 * - bp for callchains 948 * - eflags, for future purposes, just in case 949 */ 950 static inline void perf_fetch_caller_regs(struct pt_regs *regs) 951 { 952 memset(regs, 0, sizeof(*regs)); 953 954 perf_arch_fetch_caller_regs(regs, CALLER_ADDR0); 955 } 956 957 static inline void 958 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr) 959 { 960 if (atomic_read(&perf_swevent_enabled[event_id])) { 961 struct pt_regs hot_regs; 962 963 if (!regs) { 964 perf_fetch_caller_regs(&hot_regs); 965 regs = &hot_regs; 966 } 967 __perf_sw_event(event_id, nr, nmi, regs, addr); 968 } 969 } 970 971 extern void perf_event_mmap(struct vm_area_struct *vma); 972 extern struct perf_guest_info_callbacks *perf_guest_cbs; 973 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks); 974 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks); 975 976 extern void perf_event_comm(struct task_struct *tsk); 977 extern void perf_event_fork(struct task_struct *tsk); 978 979 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs); 980 981 extern int sysctl_perf_event_paranoid; 982 extern int sysctl_perf_event_mlock; 983 extern int sysctl_perf_event_sample_rate; 984 985 static inline bool perf_paranoid_tracepoint_raw(void) 986 { 987 return sysctl_perf_event_paranoid > -1; 988 } 989 990 static inline bool perf_paranoid_cpu(void) 991 { 992 return sysctl_perf_event_paranoid > 0; 993 } 994 995 static inline bool perf_paranoid_kernel(void) 996 { 997 return sysctl_perf_event_paranoid > 1; 998 } 999 1000 extern void perf_event_init(void); 1001 extern void perf_tp_event(u64 addr, u64 count, void *record, 1002 int entry_size, struct pt_regs *regs, 1003 struct hlist_head *head, int rctx); 1004 extern void perf_bp_event(struct perf_event *event, void *data); 1005 1006 #ifndef perf_misc_flags 1007 #define perf_misc_flags(regs) (user_mode(regs) ? PERF_RECORD_MISC_USER : \ 1008 PERF_RECORD_MISC_KERNEL) 1009 #define perf_instruction_pointer(regs) instruction_pointer(regs) 1010 #endif 1011 1012 extern int perf_output_begin(struct perf_output_handle *handle, 1013 struct perf_event *event, unsigned int size, 1014 int nmi, int sample); 1015 extern void perf_output_end(struct perf_output_handle *handle); 1016 extern void perf_output_copy(struct perf_output_handle *handle, 1017 const void *buf, unsigned int len); 1018 extern int perf_swevent_get_recursion_context(void); 1019 extern void perf_swevent_put_recursion_context(int rctx); 1020 extern void perf_event_enable(struct perf_event *event); 1021 extern void perf_event_disable(struct perf_event *event); 1022 #else 1023 static inline void 1024 perf_event_task_sched_in(struct task_struct *task) { } 1025 static inline void 1026 perf_event_task_sched_out(struct task_struct *task, 1027 struct task_struct *next) { } 1028 static inline void 1029 perf_event_task_tick(struct task_struct *task) { } 1030 static inline int perf_event_init_task(struct task_struct *child) { return 0; } 1031 static inline void perf_event_exit_task(struct task_struct *child) { } 1032 static inline void perf_event_free_task(struct task_struct *task) { } 1033 static inline void perf_event_do_pending(void) { } 1034 static inline void perf_event_print_debug(void) { } 1035 static inline void perf_disable(void) { } 1036 static inline void perf_enable(void) { } 1037 static inline int perf_event_task_disable(void) { return -EINVAL; } 1038 static inline int perf_event_task_enable(void) { return -EINVAL; } 1039 1040 static inline void 1041 perf_sw_event(u32 event_id, u64 nr, int nmi, 1042 struct pt_regs *regs, u64 addr) { } 1043 static inline void 1044 perf_bp_event(struct perf_event *event, void *data) { } 1045 1046 static inline int perf_register_guest_info_callbacks 1047 (struct perf_guest_info_callbacks *callbacks) { return 0; } 1048 static inline int perf_unregister_guest_info_callbacks 1049 (struct perf_guest_info_callbacks *callbacks) { return 0; } 1050 1051 static inline void perf_event_mmap(struct vm_area_struct *vma) { } 1052 static inline void perf_event_comm(struct task_struct *tsk) { } 1053 static inline void perf_event_fork(struct task_struct *tsk) { } 1054 static inline void perf_event_init(void) { } 1055 static inline int perf_swevent_get_recursion_context(void) { return -1; } 1056 static inline void perf_swevent_put_recursion_context(int rctx) { } 1057 static inline void perf_event_enable(struct perf_event *event) { } 1058 static inline void perf_event_disable(struct perf_event *event) { } 1059 #endif 1060 1061 #define perf_output_put(handle, x) \ 1062 perf_output_copy((handle), &(x), sizeof(x)) 1063 1064 /* 1065 * This has to have a higher priority than migration_notifier in sched.c. 1066 */ 1067 #define perf_cpu_notifier(fn) \ 1068 do { \ 1069 static struct notifier_block fn##_nb __cpuinitdata = \ 1070 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \ 1071 fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE, \ 1072 (void *)(unsigned long)smp_processor_id()); \ 1073 fn(&fn##_nb, (unsigned long)CPU_STARTING, \ 1074 (void *)(unsigned long)smp_processor_id()); \ 1075 fn(&fn##_nb, (unsigned long)CPU_ONLINE, \ 1076 (void *)(unsigned long)smp_processor_id()); \ 1077 register_cpu_notifier(&fn##_nb); \ 1078 } while (0) 1079 1080 #endif /* __KERNEL__ */ 1081 #endif /* _LINUX_PERF_EVENT_H */ 1082