SAFE public projects git trees. - safe/jmp/linux-2.6/blob - include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@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 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 225
 226 enum perf_event_ioc_flags {
 227         PERF_IOC_FLAG_GROUP             = 1U << 0,
 228 };
 229
 230 /*
 231  * Structure of the page that can be mapped via mmap
 232  */
 233 struct perf_event_mmap_page {
 234         __u32   version;                /* version number of this structure */
 235         __u32   compat_version;         /* lowest version this is compat with */
 236
 237         /*
 238          * Bits needed to read the hw events in user-space.
 239          *
 240          *   u32 seq;
 241          *   s64 count;
 242          *
 243          *   do {
 244          *     seq = pc->lock;
 245          *
 246          *     barrier()
 247          *     if (pc->index) {
 248          *       count = pmc_read(pc->index - 1);
 249          *       count += pc->offset;
 250          *     } else
 251          *       goto regular_read;
 252          *
 253          *     barrier();
 254          *   } while (pc->lock != seq);
 255          *
 256          * NOTE: for obvious reason this only works on self-monitoring
 257          *       processes.
 258          */
 259         __u32   lock;                   /* seqlock for synchronization */
 260         __u32   index;                  /* hardware event identifier */
 261         __s64   offset;                 /* add to hardware event value */
 262         __u64   time_enabled;           /* time event active */
 263         __u64   time_running;           /* time event on cpu */
 264
 265                 /*
 266                  * Hole for extension of the self monitor capabilities
 267                  */
 268
 269         __u64   __reserved[123];        /* align to 1k */
 270
 271         /*
 272          * Control data for the mmap() data buffer.
 273          *
 274          * User-space reading the @data_head value should issue an rmb(), on
 275          * SMP capable platforms, after reading this value -- see
 276          * perf_event_wakeup().
 277          *
 278          * When the mapping is PROT_WRITE the @data_tail value should be
 279          * written by userspace to reflect the last read data. In this case
 280          * the kernel will not over-write unread data.
 281          */
 282         __u64   data_head;              /* head in the data section */
 283         __u64   data_tail;              /* user-space written tail */
 284 };
 285
 286 #define PERF_RECORD_MISC_CPUMODE_MASK           (3 << 0)
 287 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN                (0 << 0)
 288 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 289 #define PERF_RECORD_MISC_USER                   (2 << 0)
 290 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 291
 292 struct perf_event_header {
 293         __u32   type;
 294         __u16   misc;
 295         __u16   size;
 296 };
 297
 298 enum perf_event_type {
 299
 300         /*
 301          * The MMAP events record the PROT_EXEC mappings so that we can
 302          * correlate userspace IPs to code. They have the following structure:
 303          *
 304          * struct {
 305          *      struct perf_event_header        header;
 306          *
 307          *      u32                             pid, tid;
 308          *      u64                             addr;
 309          *      u64                             len;
 310          *      u64                             pgoff;
 311          *      char                            filename[];
 312          * };
 313          */
 314         PERF_RECORD_MMAP                        = 1,
 315
 316         /*
 317          * struct {
 318          *      struct perf_event_header        header;
 319          *      u64                             id;
 320          *      u64                             lost;
 321          * };
 322          */
 323         PERF_RECORD_LOST                        = 2,
 324
 325         /*
 326          * struct {
 327          *      struct perf_event_header        header;
 328          *
 329          *      u32                             pid, tid;
 330          *      char                            comm[];
 331          * };
 332          */
 333         PERF_RECORD_COMM                        = 3,
 334
 335         /*
 336          * struct {
 337          *      struct perf_event_header        header;
 338          *      u32                             pid, ppid;
 339          *      u32                             tid, ptid;
 340          *      u64                             time;
 341          * };
 342          */
 343         PERF_RECORD_EXIT                        = 4,
 344
 345         /*
 346          * struct {
 347          *      struct perf_event_header        header;
 348          *      u64                             time;
 349          *      u64                             id;
 350          *      u64                             stream_id;
 351          * };
 352          */
 353         PERF_RECORD_THROTTLE            = 5,
 354         PERF_RECORD_UNTHROTTLE          = 6,
 355
 356         /*
 357          * struct {
 358          *      struct perf_event_header        header;
 359          *      u32                             pid, ppid;
 360          *      u32                             tid, ptid;
 361          *      u64                             time;
 362          * };
 363          */
 364         PERF_RECORD_FORK                        = 7,
 365
 366         /*
 367          * struct {
 368          *      struct perf_event_header        header;
 369          *      u32                             pid, tid;
 370          *
 371          *      struct read_format              values;
 372          * };
 373          */
 374         PERF_RECORD_READ                        = 8,
 375
 376         /*
 377          * struct {
 378          *      struct perf_event_header        header;
 379          *
 380          *      { u64                   ip;       } && PERF_SAMPLE_IP
 381          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 382          *      { u64                   time;     } && PERF_SAMPLE_TIME
 383          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 384          *      { u64                   id;       } && PERF_SAMPLE_ID
 385          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 386          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 387          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 388          *
 389          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 390          *
 391          *      { u64                   nr,
 392          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 393          *
 394          *      #
 395          *      # The RAW record below is opaque data wrt the ABI
 396          *      #
 397          *      # That is, the ABI doesn't make any promises wrt to
 398          *      # the stability of its content, it may vary depending
 399          *      # on event, hardware, kernel version and phase of
 400          *      # the moon.
 401          *      #
 402          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 403          *      #
 404          *
 405          *      { u32                   size;
 406          *        char                  data[size];}&& PERF_SAMPLE_RAW
 407          * };
 408          */
 409         PERF_RECORD_SAMPLE              = 9,
 410
 411         PERF_RECORD_MAX,                        /* non-ABI */
 412 };
 413
 414 enum perf_callchain_context {
 415         PERF_CONTEXT_HV                 = (__u64)-32,
 416         PERF_CONTEXT_KERNEL             = (__u64)-128,
 417         PERF_CONTEXT_USER               = (__u64)-512,
 418
 419         PERF_CONTEXT_GUEST              = (__u64)-2048,
 420         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 421         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 422
 423         PERF_CONTEXT_MAX                = (__u64)-4095,
 424 };
 425
 426 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 427 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 428
 429 #ifdef __KERNEL__
 430 /*
 431  * Kernel-internal data types and definitions:
 432  */
 433
 434 #ifdef CONFIG_PERF_EVENTS
 435 # include <asm/perf_event.h>
 436 #endif
 437
 438 #include <linux/list.h>
 439 #include <linux/mutex.h>
 440 #include <linux/rculist.h>
 441 #include <linux/rcupdate.h>
 442 #include <linux/spinlock.h>
 443 #include <linux/hrtimer.h>
 444 #include <linux/fs.h>
 445 #include <linux/pid_namespace.h>
 446 #include <linux/workqueue.h>
 447 #include <asm/atomic.h>
 448
 449 #define PERF_MAX_STACK_DEPTH            255
 450
 451 struct perf_callchain_entry {
 452         __u64                           nr;
 453         __u64                           ip[PERF_MAX_STACK_DEPTH];
 454 };
 455
 456 struct perf_raw_record {
 457         u32                             size;
 458         void                            *data;
 459 };
 460
 461 struct task_struct;
 462
 463 /**
 464  * struct hw_perf_event - performance event hardware details:
 465  */
 466 struct hw_perf_event {
 467 #ifdef CONFIG_PERF_EVENTS
 468         union {
 469                 struct { /* hardware */
 470                         u64             config;
 471                         unsigned long   config_base;
 472                         unsigned long   event_base;
 473                         int             idx;
 474                 };
 475                 union { /* software */
 476                         atomic64_t      count;
 477                         struct hrtimer  hrtimer;
 478                 };
 479         };
 480         atomic64_t                      prev_count;
 481         u64                             sample_period;
 482         u64                             last_period;
 483         atomic64_t                      period_left;
 484         u64                             interrupts;
 485
 486         u64                             freq_count;
 487         u64                             freq_interrupts;
 488         u64                             freq_stamp;
 489 #endif
 490 };
 491
 492 struct perf_event;
 493
 494 /**
 495  * struct pmu - generic performance monitoring unit
 496  */
 497 struct pmu {
 498         int (*enable)                   (struct perf_event *event);
 499         void (*disable)                 (struct perf_event *event);
 500         void (*read)                    (struct perf_event *event);
 501         void (*unthrottle)              (struct perf_event *event);
 502 };
 503
 504 /**
 505  * enum perf_event_active_state - the states of a event
 506  */
 507 enum perf_event_active_state {
 508         PERF_EVENT_STATE_ERROR          = -2,
 509         PERF_EVENT_STATE_OFF            = -1,
 510         PERF_EVENT_STATE_INACTIVE       =  0,
 511         PERF_EVENT_STATE_ACTIVE         =  1,
 512 };
 513
 514 struct file;
 515
 516 struct perf_mmap_data {
 517         struct rcu_head                 rcu_head;
 518 #ifdef CONFIG_PERF_USE_VMALLOC
 519         struct work_struct              work;
 520 #endif
 521         int                             data_order;
 522         int                             nr_pages;       /* nr of data pages  */
 523         int                             writable;       /* are we writable   */
 524         int                             nr_locked;      /* nr pages mlocked  */
 525
 526         atomic_t                        poll;           /* POLL_ for wakeups */
 527         atomic_t                        events;         /* event_id limit       */
 528
 529         atomic_long_t                   head;           /* write position    */
 530         atomic_long_t                   done_head;      /* completed head    */
 531
 532         atomic_t                        lock;           /* concurrent writes */
 533         atomic_t                        wakeup;         /* needs a wakeup    */
 534         atomic_t                        lost;           /* nr records lost   */
 535
 536         long                            watermark;      /* wakeup watermark  */
 537
 538         struct perf_event_mmap_page     *user_page;
 539         void                            *data_pages[0];
 540 };
 541
 542 struct perf_pending_entry {
 543         struct perf_pending_entry *next;
 544         void (*func)(struct perf_pending_entry *);
 545 };
 546
 547 typedef void (*perf_callback_t)(struct perf_event *, void *);
 548
 549 /**
 550  * struct perf_event - performance event kernel representation:
 551  */
 552 struct perf_event {
 553 #ifdef CONFIG_PERF_EVENTS
 554         struct list_head                group_entry;
 555         struct list_head                event_entry;
 556         struct list_head                sibling_list;
 557         int                             nr_siblings;
 558         struct perf_event               *group_leader;
 559         struct perf_event               *output;
 560         const struct pmu                *pmu;
 561
 562         enum perf_event_active_state    state;
 563         atomic64_t                      count;
 564
 565         /*
 566          * These are the total time in nanoseconds that the event
 567          * has been enabled (i.e. eligible to run, and the task has
 568          * been scheduled in, if this is a per-task event)
 569          * and running (scheduled onto the CPU), respectively.
 570          *
 571          * They are computed from tstamp_enabled, tstamp_running and
 572          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 573          */
 574         u64                             total_time_enabled;
 575         u64                             total_time_running;
 576
 577         /*
 578          * These are timestamps used for computing total_time_enabled
 579          * and total_time_running when the event is in INACTIVE or
 580          * ACTIVE state, measured in nanoseconds from an arbitrary point
 581          * in time.
 582          * tstamp_enabled: the notional time when the event was enabled
 583          * tstamp_running: the notional time when the event was scheduled on
 584          * tstamp_stopped: in INACTIVE state, the notional time when the
 585          *      event was scheduled off.
 586          */
 587         u64                             tstamp_enabled;
 588         u64                             tstamp_running;
 589         u64                             tstamp_stopped;
 590
 591         struct perf_event_attr  attr;
 592         struct hw_perf_event            hw;
 593
 594         struct perf_event_context       *ctx;
 595         struct file                     *filp;
 596
 597         /*
 598          * These accumulate total time (in nanoseconds) that children
 599          * events have been enabled and running, respectively.
 600          */
 601         atomic64_t                      child_total_time_enabled;
 602         atomic64_t                      child_total_time_running;
 603
 604         /*
 605          * Protect attach/detach and child_list:
 606          */
 607         struct mutex                    child_mutex;
 608         struct list_head                child_list;
 609         struct perf_event               *parent;
 610
 611         int                             oncpu;
 612         int                             cpu;
 613
 614         struct list_head                owner_entry;
 615         struct task_struct              *owner;
 616
 617         /* mmap bits */
 618         struct mutex                    mmap_mutex;
 619         atomic_t                        mmap_count;
 620         struct perf_mmap_data           *data;
 621
 622         /* poll related */
 623         wait_queue_head_t               waitq;
 624         struct fasync_struct            *fasync;
 625
 626         /* delayed work for NMIs and such */
 627         int                             pending_wakeup;
 628         int                             pending_kill;
 629         int                             pending_disable;
 630         struct perf_pending_entry       pending;
 631
 632         atomic_t                        event_limit;
 633
 634         void (*destroy)(struct perf_event *);
 635         struct rcu_head                 rcu_head;
 636
 637         struct pid_namespace            *ns;
 638         u64                             id;
 639
 640 #ifdef CONFIG_EVENT_PROFILE
 641         struct event_filter             *filter;
 642 #endif
 643
 644         perf_callback_t                 callback;
 645
 646 #endif /* CONFIG_PERF_EVENTS */
 647 };
 648
 649 /**
 650  * struct perf_event_context - event context structure
 651  *
 652  * Used as a container for task events and CPU events as well:
 653  */
 654 struct perf_event_context {
 655         /*
 656          * Protect the states of the events in the list,
 657          * nr_active, and the list:
 658          */
 659         spinlock_t                      lock;
 660         /*
 661          * Protect the list of events.  Locking either mutex or lock
 662          * is sufficient to ensure the list doesn't change; to change
 663          * the list you need to lock both the mutex and the spinlock.
 664          */
 665         struct mutex                    mutex;
 666
 667         struct list_head                group_list;
 668         struct list_head                event_list;
 669         int                             nr_events;
 670         int                             nr_active;
 671         int                             is_active;
 672         int                             nr_stat;
 673         atomic_t                        refcount;
 674         struct task_struct              *task;
 675
 676         /*
 677          * Context clock, runs when context enabled.
 678          */
 679         u64                             time;
 680         u64                             timestamp;
 681
 682         /*
 683          * These fields let us detect when two contexts have both
 684          * been cloned (inherited) from a common ancestor.
 685          */
 686         struct perf_event_context       *parent_ctx;
 687         u64                             parent_gen;
 688         u64                             generation;
 689         int                             pin_count;
 690         struct rcu_head                 rcu_head;
 691 };
 692
 693 /**
 694  * struct perf_event_cpu_context - per cpu event context structure
 695  */
 696 struct perf_cpu_context {
 697         struct perf_event_context       ctx;
 698         struct perf_event_context       *task_ctx;
 699         int                             active_oncpu;
 700         int                             max_pertask;
 701         int                             exclusive;
 702
 703         /*
 704          * Recursion avoidance:
 705          *
 706          * task, softirq, irq, nmi context
 707          */
 708         int                             recursion[4];
 709 };
 710
 711 struct perf_output_handle {
 712         struct perf_event               *event;
 713         struct perf_mmap_data           *data;
 714         unsigned long                   head;
 715         unsigned long                   offset;
 716         int                             nmi;
 717         int                             sample;
 718         int                             locked;
 719         unsigned long                   flags;
 720 };
 721
 722 #ifdef CONFIG_PERF_EVENTS
 723
 724 /*
 725  * Set by architecture code:
 726  */
 727 extern int perf_max_events;
 728
 729 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
 730
 731 extern void perf_event_task_sched_in(struct task_struct *task, int cpu);
 732 extern void perf_event_task_sched_out(struct task_struct *task,
 733                                         struct task_struct *next, int cpu);
 734 extern void perf_event_task_tick(struct task_struct *task, int cpu);
 735 extern int perf_event_init_task(struct task_struct *child);
 736 extern void perf_event_exit_task(struct task_struct *child);
 737 extern void perf_event_free_task(struct task_struct *task);
 738 extern void set_perf_event_pending(void);
 739 extern void perf_event_do_pending(void);
 740 extern void perf_event_print_debug(void);
 741 extern void __perf_disable(void);
 742 extern bool __perf_enable(void);
 743 extern void perf_disable(void);
 744 extern void perf_enable(void);
 745 extern int perf_event_task_disable(void);
 746 extern int perf_event_task_enable(void);
 747 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
 748                struct perf_cpu_context *cpuctx,
 749                struct perf_event_context *ctx, int cpu);
 750 extern void perf_event_update_userpage(struct perf_event *event);
 751 extern int perf_event_release_kernel(struct perf_event *event);
 752 extern struct perf_event *
 753 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 754                                 int cpu,
 755                                 pid_t pid,
 756                                 perf_callback_t callback);
 757 extern u64 perf_event_read_value(struct perf_event *event);
 758
 759 struct perf_sample_data {
 760         u64                             type;
 761
 762         u64                             ip;
 763         struct {
 764                 u32     pid;
 765                 u32     tid;
 766         }                               tid_entry;
 767         u64                             time;
 768         u64                             addr;
 769         u64                             id;
 770         u64                             stream_id;
 771         struct {
 772                 u32     cpu;
 773                 u32     reserved;
 774         }                               cpu_entry;
 775         u64                             period;
 776         struct perf_callchain_entry     *callchain;
 777         struct perf_raw_record          *raw;
 778 };
 779
 780 extern void perf_output_sample(struct perf_output_handle *handle,
 781                                struct perf_event_header *header,
 782                                struct perf_sample_data *data,
 783                                struct perf_event *event);
 784 extern void perf_prepare_sample(struct perf_event_header *header,
 785                                 struct perf_sample_data *data,
 786                                 struct perf_event *event,
 787                                 struct pt_regs *regs);
 788
 789 extern int perf_event_overflow(struct perf_event *event, int nmi,
 790                                  struct perf_sample_data *data,
 791                                  struct pt_regs *regs);
 792
 793 /*
 794  * Return 1 for a software event, 0 for a hardware event
 795  */
 796 static inline int is_software_event(struct perf_event *event)
 797 {
 798         return (event->attr.type != PERF_TYPE_RAW) &&
 799                 (event->attr.type != PERF_TYPE_HARDWARE) &&
 800                 (event->attr.type != PERF_TYPE_HW_CACHE);
 801 }
 802
 803 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
 804
 805 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
 806
 807 static inline void
 808 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 809 {
 810         if (atomic_read(&perf_swevent_enabled[event_id]))
 811                 __perf_sw_event(event_id, nr, nmi, regs, addr);
 812 }
 813
 814 extern void __perf_event_mmap(struct vm_area_struct *vma);
 815
 816 static inline void perf_event_mmap(struct vm_area_struct *vma)
 817 {
 818         if (vma->vm_flags & VM_EXEC)
 819                 __perf_event_mmap(vma);
 820 }
 821
 822 extern void perf_event_comm(struct task_struct *tsk);
 823 extern void perf_event_fork(struct task_struct *tsk);
 824
 825 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 826
 827 extern int sysctl_perf_event_paranoid;
 828 extern int sysctl_perf_event_mlock;
 829 extern int sysctl_perf_event_sample_rate;
 830
 831 extern void perf_event_init(void);
 832 extern void perf_tp_event(int event_id, u64 addr, u64 count,
 833                                  void *record, int entry_size);
 834
 835 #ifndef perf_misc_flags
 836 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
 837                                  PERF_RECORD_MISC_KERNEL)
 838 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 839 #endif
 840
 841 extern int perf_output_begin(struct perf_output_handle *handle,
 842                              struct perf_event *event, unsigned int size,
 843                              int nmi, int sample);
 844 extern void perf_output_end(struct perf_output_handle *handle);
 845 extern void perf_output_copy(struct perf_output_handle *handle,
 846                              const void *buf, unsigned int len);
 847 #else
 848 static inline void
 849 perf_event_task_sched_in(struct task_struct *task, int cpu)             { }
 850 static inline void
 851 perf_event_task_sched_out(struct task_struct *task,
 852                             struct task_struct *next, int cpu)          { }
 853 static inline void
 854 perf_event_task_tick(struct task_struct *task, int cpu)                 { }
 855 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 856 static inline void perf_event_exit_task(struct task_struct *child)      { }
 857 static inline void perf_event_free_task(struct task_struct *task)       { }
 858 static inline void perf_event_do_pending(void)                          { }
 859 static inline void perf_event_print_debug(void)                         { }
 860 static inline void perf_disable(void)                                   { }
 861 static inline void perf_enable(void)                                    { }
 862 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 863 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 864
 865 static inline void
 866 perf_sw_event(u32 event_id, u64 nr, int nmi,
 867                      struct pt_regs *regs, u64 addr)                    { }
 868
 869 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 870 static inline void perf_event_comm(struct task_struct *tsk)             { }
 871 static inline void perf_event_fork(struct task_struct *tsk)             { }
 872 static inline void perf_event_init(void)                                { }
 873
 874 #endif
 875
 876 #define perf_output_put(handle, x) \
 877         perf_output_copy((handle), &(x), sizeof(x))
 878
 879 #endif /* __KERNEL__ */
 880 #endif /* _LINUX_PERF_EVENT_H */