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         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                                 precise        :  1, /* OoO invariant counter */
 207
 208                                 __reserved_1   : 48;
 209
 210         union {
 211                 __u32           wakeup_events;    /* wakeup every n events */
 212                 __u32           wakeup_watermark; /* bytes before wakeup   */
 213         };
 214
 215         __u32                   bp_type;
 216         __u64                   bp_addr;
 217         __u64                   bp_len;
 218 };
 219
 220 /*
 221  * Ioctls that can be done on a perf event fd:
 222  */
 223 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
 224 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
 225 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
 226 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
 227 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
 228 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
 229 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 230
 231 enum perf_event_ioc_flags {
 232         PERF_IOC_FLAG_GROUP             = 1U << 0,
 233 };
 234
 235 /*
 236  * Structure of the page that can be mapped via mmap
 237  */
 238 struct perf_event_mmap_page {
 239         __u32   version;                /* version number of this structure */
 240         __u32   compat_version;         /* lowest version this is compat with */
 241
 242         /*
 243          * Bits needed to read the hw events in user-space.
 244          *
 245          *   u32 seq;
 246          *   s64 count;
 247          *
 248          *   do {
 249          *     seq = pc->lock;
 250          *
 251          *     barrier()
 252          *     if (pc->index) {
 253          *       count = pmc_read(pc->index - 1);
 254          *       count += pc->offset;
 255          *     } else
 256          *       goto regular_read;
 257          *
 258          *     barrier();
 259          *   } while (pc->lock != seq);
 260          *
 261          * NOTE: for obvious reason this only works on self-monitoring
 262          *       processes.
 263          */
 264         __u32   lock;                   /* seqlock for synchronization */
 265         __u32   index;                  /* hardware event identifier */
 266         __s64   offset;                 /* add to hardware event value */
 267         __u64   time_enabled;           /* time event active */
 268         __u64   time_running;           /* time event on cpu */
 269
 270                 /*
 271                  * Hole for extension of the self monitor capabilities
 272                  */
 273
 274         __u64   __reserved[123];        /* align to 1k */
 275
 276         /*
 277          * Control data for the mmap() data buffer.
 278          *
 279          * User-space reading the @data_head value should issue an rmb(), on
 280          * SMP capable platforms, after reading this value -- see
 281          * perf_event_wakeup().
 282          *
 283          * When the mapping is PROT_WRITE the @data_tail value should be
 284          * written by userspace to reflect the last read data. In this case
 285          * the kernel will not over-write unread data.
 286          */
 287         __u64   data_head;              /* head in the data section */
 288         __u64   data_tail;              /* user-space written tail */
 289 };
 290
 291 #define PERF_RECORD_MISC_CPUMODE_MASK           (3 << 0)
 292 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 293 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 294 #define PERF_RECORD_MISC_USER                   (2 << 0)
 295 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 296
 297 #define PERF_RECORD_MISC_EXACT                  (1 << 14)
 298 /*
 299  * Reserve the last bit to indicate some extended misc field
 300  */
 301 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 302
 303 struct perf_event_header {
 304         __u32   type;
 305         __u16   misc;
 306         __u16   size;
 307 };
 308
 309 enum perf_event_type {
 310
 311         /*
 312          * The MMAP events record the PROT_EXEC mappings so that we can
 313          * correlate userspace IPs to code. They have the following structure:
 314          *
 315          * struct {
 316          *      struct perf_event_header        header;
 317          *
 318          *      u32                             pid, tid;
 319          *      u64                             addr;
 320          *      u64                             len;
 321          *      u64                             pgoff;
 322          *      char                            filename[];
 323          * };
 324          */
 325         PERF_RECORD_MMAP                        = 1,
 326
 327         /*
 328          * struct {
 329          *      struct perf_event_header        header;
 330          *      u64                             id;
 331          *      u64                             lost;
 332          * };
 333          */
 334         PERF_RECORD_LOST                        = 2,
 335
 336         /*
 337          * struct {
 338          *      struct perf_event_header        header;
 339          *
 340          *      u32                             pid, tid;
 341          *      char                            comm[];
 342          * };
 343          */
 344         PERF_RECORD_COMM                        = 3,
 345
 346         /*
 347          * struct {
 348          *      struct perf_event_header        header;
 349          *      u32                             pid, ppid;
 350          *      u32                             tid, ptid;
 351          *      u64                             time;
 352          * };
 353          */
 354         PERF_RECORD_EXIT                        = 4,
 355
 356         /*
 357          * struct {
 358          *      struct perf_event_header        header;
 359          *      u64                             time;
 360          *      u64                             id;
 361          *      u64                             stream_id;
 362          * };
 363          */
 364         PERF_RECORD_THROTTLE                    = 5,
 365         PERF_RECORD_UNTHROTTLE                  = 6,
 366
 367         /*
 368          * struct {
 369          *      struct perf_event_header        header;
 370          *      u32                             pid, ppid;
 371          *      u32                             tid, ptid;
 372          *      u64                             time;
 373          * };
 374          */
 375         PERF_RECORD_FORK                        = 7,
 376
 377         /*
 378          * struct {
 379          *      struct perf_event_header        header;
 380          *      u32                             pid, tid;
 381          *
 382          *      struct read_format              values;
 383          * };
 384          */
 385         PERF_RECORD_READ                        = 8,
 386
 387         /*
 388          * struct {
 389          *      struct perf_event_header        header;
 390          *
 391          *      { u64                   ip;       } && PERF_SAMPLE_IP
 392          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 393          *      { u64                   time;     } && PERF_SAMPLE_TIME
 394          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 395          *      { u64                   id;       } && PERF_SAMPLE_ID
 396          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 397          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 398          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 399          *
 400          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 401          *
 402          *      { u64                   nr,
 403          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 404          *
 405          *      #
 406          *      # The RAW record below is opaque data wrt the ABI
 407          *      #
 408          *      # That is, the ABI doesn't make any promises wrt to
 409          *      # the stability of its content, it may vary depending
 410          *      # on event, hardware, kernel version and phase of
 411          *      # the moon.
 412          *      #
 413          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 414          *      #
 415          *
 416          *      { u32                   size;
 417          *        char                  data[size];}&& PERF_SAMPLE_RAW
 418          * };
 419          */
 420         PERF_RECORD_SAMPLE                      = 9,
 421
 422         PERF_RECORD_MAX,                        /* non-ABI */
 423 };
 424
 425 enum perf_callchain_context {
 426         PERF_CONTEXT_HV                 = (__u64)-32,
 427         PERF_CONTEXT_KERNEL             = (__u64)-128,
 428         PERF_CONTEXT_USER               = (__u64)-512,
 429
 430         PERF_CONTEXT_GUEST              = (__u64)-2048,
 431         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 432         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 433
 434         PERF_CONTEXT_MAX                = (__u64)-4095,
 435 };
 436
 437 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 438 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 439
 440 #ifdef __KERNEL__
 441 /*
 442  * Kernel-internal data types and definitions:
 443  */
 444
 445 #ifdef CONFIG_PERF_EVENTS
 446 # include <asm/perf_event.h>
 447 #endif
 448
 449 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 450 #include <asm/hw_breakpoint.h>
 451 #endif
 452
 453 #include <linux/list.h>
 454 #include <linux/mutex.h>
 455 #include <linux/rculist.h>
 456 #include <linux/rcupdate.h>
 457 #include <linux/spinlock.h>
 458 #include <linux/hrtimer.h>
 459 #include <linux/fs.h>
 460 #include <linux/pid_namespace.h>
 461 #include <linux/workqueue.h>
 462 #include <linux/ftrace.h>
 463 #include <linux/cpu.h>
 464 #include <asm/atomic.h>
 465
 466 #define PERF_MAX_STACK_DEPTH            255
 467
 468 struct perf_callchain_entry {
 469         __u64                           nr;
 470         __u64                           ip[PERF_MAX_STACK_DEPTH];
 471 };
 472
 473 struct perf_raw_record {
 474         u32                             size;
 475         void                            *data;
 476 };
 477
 478 struct perf_branch_entry {
 479         __u64                           from;
 480         __u64                           to;
 481         __u64                           flags;
 482 };
 483
 484 struct perf_branch_stack {
 485         __u64                           nr;
 486         struct perf_branch_entry        entries[0];
 487 };
 488
 489 struct task_struct;
 490
 491 /**
 492  * struct hw_perf_event - performance event hardware details:
 493  */
 494 struct hw_perf_event {
 495 #ifdef CONFIG_PERF_EVENTS
 496         union {
 497                 struct { /* hardware */
 498                         u64             config;
 499                         u64             last_tag;
 500                         unsigned long   config_base;
 501                         unsigned long   event_base;
 502                         int             idx;
 503                         int             last_cpu;
 504                 };
 505                 struct { /* software */
 506                         s64             remaining;
 507                         struct hrtimer  hrtimer;
 508                 };
 509 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 510                 /* breakpoint */
 511                 struct arch_hw_breakpoint       info;
 512 #endif
 513         };
 514         atomic64_t                      prev_count;
 515         u64                             sample_period;
 516         u64                             last_period;
 517         atomic64_t                      period_left;
 518         u64                             interrupts;
 519
 520         u64                             freq_time_stamp;
 521         u64                             freq_count_stamp;
 522 #endif
 523 };
 524
 525 struct perf_event;
 526
 527 /**
 528  * struct pmu - generic performance monitoring unit
 529  */
 530 struct pmu {
 531         int (*enable)                   (struct perf_event *event);
 532         void (*disable)                 (struct perf_event *event);
 533         int (*start)                    (struct perf_event *event);
 534         void (*stop)                    (struct perf_event *event);
 535         void (*read)                    (struct perf_event *event);
 536         void (*unthrottle)              (struct perf_event *event);
 537 };
 538
 539 /**
 540  * enum perf_event_active_state - the states of a event
 541  */
 542 enum perf_event_active_state {
 543         PERF_EVENT_STATE_ERROR          = -2,
 544         PERF_EVENT_STATE_OFF            = -1,
 545         PERF_EVENT_STATE_INACTIVE       =  0,
 546         PERF_EVENT_STATE_ACTIVE         =  1,
 547 };
 548
 549 struct file;
 550
 551 struct perf_mmap_data {
 552         struct rcu_head                 rcu_head;
 553 #ifdef CONFIG_PERF_USE_VMALLOC
 554         struct work_struct              work;
 555 #endif
 556         int                             data_order;
 557         int                             nr_pages;       /* nr of data pages  */
 558         int                             writable;       /* are we writable   */
 559         int                             nr_locked;      /* nr pages mlocked  */
 560
 561         atomic_t                        poll;           /* POLL_ for wakeups */
 562         atomic_t                        events;         /* event_id limit       */
 563
 564         atomic_long_t                   head;           /* write position    */
 565         atomic_long_t                   done_head;      /* completed head    */
 566
 567         atomic_t                        lock;           /* concurrent writes */
 568         atomic_t                        wakeup;         /* needs a wakeup    */
 569         atomic_t                        lost;           /* nr records lost   */
 570
 571         long                            watermark;      /* wakeup watermark  */
 572
 573         struct perf_event_mmap_page     *user_page;
 574         void                            *data_pages[0];
 575 };
 576
 577 struct perf_pending_entry {
 578         struct perf_pending_entry *next;
 579         void (*func)(struct perf_pending_entry *);
 580 };
 581
 582 struct perf_sample_data;
 583
 584 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
 585                                         struct perf_sample_data *,
 586                                         struct pt_regs *regs);
 587
 588 enum perf_group_flag {
 589         PERF_GROUP_SOFTWARE = 0x1,
 590 };
 591
 592 /**
 593  * struct perf_event - performance event kernel representation:
 594  */
 595 struct perf_event {
 596 #ifdef CONFIG_PERF_EVENTS
 597         struct list_head                group_entry;
 598         struct list_head                event_entry;
 599         struct list_head                sibling_list;
 600         int                             nr_siblings;
 601         int                             group_flags;
 602         struct perf_event               *group_leader;
 603         struct perf_event               *output;
 604         const struct pmu                *pmu;
 605
 606         enum perf_event_active_state    state;
 607         atomic64_t                      count;
 608
 609         /*
 610          * These are the total time in nanoseconds that the event
 611          * has been enabled (i.e. eligible to run, and the task has
 612          * been scheduled in, if this is a per-task event)
 613          * and running (scheduled onto the CPU), respectively.
 614          *
 615          * They are computed from tstamp_enabled, tstamp_running and
 616          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 617          */
 618         u64                             total_time_enabled;
 619         u64                             total_time_running;
 620
 621         /*
 622          * These are timestamps used for computing total_time_enabled
 623          * and total_time_running when the event is in INACTIVE or
 624          * ACTIVE state, measured in nanoseconds from an arbitrary point
 625          * in time.
 626          * tstamp_enabled: the notional time when the event was enabled
 627          * tstamp_running: the notional time when the event was scheduled on
 628          * tstamp_stopped: in INACTIVE state, the notional time when the
 629          *      event was scheduled off.
 630          */
 631         u64                             tstamp_enabled;
 632         u64                             tstamp_running;
 633         u64                             tstamp_stopped;
 634
 635         struct perf_event_attr          attr;
 636         struct hw_perf_event            hw;
 637
 638         struct perf_event_context       *ctx;
 639         struct file                     *filp;
 640
 641         /*
 642          * These accumulate total time (in nanoseconds) that children
 643          * events have been enabled and running, respectively.
 644          */
 645         atomic64_t                      child_total_time_enabled;
 646         atomic64_t                      child_total_time_running;
 647
 648         /*
 649          * Protect attach/detach and child_list:
 650          */
 651         struct mutex                    child_mutex;
 652         struct list_head                child_list;
 653         struct perf_event               *parent;
 654
 655         int                             oncpu;
 656         int                             cpu;
 657
 658         struct list_head                owner_entry;
 659         struct task_struct              *owner;
 660
 661         /* mmap bits */
 662         struct mutex                    mmap_mutex;
 663         atomic_t                        mmap_count;
 664         struct perf_mmap_data           *data;
 665
 666         /* poll related */
 667         wait_queue_head_t               waitq;
 668         struct fasync_struct            *fasync;
 669
 670         /* delayed work for NMIs and such */
 671         int                             pending_wakeup;
 672         int                             pending_kill;
 673         int                             pending_disable;
 674         struct perf_pending_entry       pending;
 675
 676         atomic_t                        event_limit;
 677
 678         void (*destroy)(struct perf_event *);
 679         struct rcu_head                 rcu_head;
 680
 681         struct pid_namespace            *ns;
 682         u64                             id;
 683
 684         perf_overflow_handler_t         overflow_handler;
 685
 686 #ifdef CONFIG_EVENT_TRACING
 687         struct event_filter             *filter;
 688 #endif
 689
 690 #endif /* CONFIG_PERF_EVENTS */
 691 };
 692
 693 /**
 694  * struct perf_event_context - event context structure
 695  *
 696  * Used as a container for task events and CPU events as well:
 697  */
 698 struct perf_event_context {
 699         /*
 700          * Protect the states of the events in the list,
 701          * nr_active, and the list:
 702          */
 703         raw_spinlock_t                  lock;
 704         /*
 705          * Protect the list of events.  Locking either mutex or lock
 706          * is sufficient to ensure the list doesn't change; to change
 707          * the list you need to lock both the mutex and the spinlock.
 708          */
 709         struct mutex                    mutex;
 710
 711         struct list_head                pinned_groups;
 712         struct list_head                flexible_groups;
 713         struct list_head                event_list;
 714         int                             nr_events;
 715         int                             nr_active;
 716         int                             is_active;
 717         int                             nr_stat;
 718         atomic_t                        refcount;
 719         struct task_struct              *task;
 720
 721         /*
 722          * Context clock, runs when context enabled.
 723          */
 724         u64                             time;
 725         u64                             timestamp;
 726
 727         /*
 728          * These fields let us detect when two contexts have both
 729          * been cloned (inherited) from a common ancestor.
 730          */
 731         struct perf_event_context       *parent_ctx;
 732         u64                             parent_gen;
 733         u64                             generation;
 734         int                             pin_count;
 735         struct rcu_head                 rcu_head;
 736 };
 737
 738 /**
 739  * struct perf_event_cpu_context - per cpu event context structure
 740  */
 741 struct perf_cpu_context {
 742         struct perf_event_context       ctx;
 743         struct perf_event_context       *task_ctx;
 744         int                             active_oncpu;
 745         int                             max_pertask;
 746         int                             exclusive;
 747
 748         /*
 749          * Recursion avoidance:
 750          *
 751          * task, softirq, irq, nmi context
 752          */
 753         int                             recursion[4];
 754 };
 755
 756 struct perf_output_handle {
 757         struct perf_event               *event;
 758         struct perf_mmap_data           *data;
 759         unsigned long                   head;
 760         unsigned long                   offset;
 761         int                             nmi;
 762         int                             sample;
 763         int                             locked;
 764 };
 765
 766 #ifdef CONFIG_PERF_EVENTS
 767
 768 /*
 769  * Set by architecture code:
 770  */
 771 extern int perf_max_events;
 772
 773 extern const struct pmu *hw_perf_event_init(struct perf_event *event);
 774
 775 extern void perf_event_task_sched_in(struct task_struct *task);
 776 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
 777 extern void perf_event_task_tick(struct task_struct *task);
 778 extern int perf_event_init_task(struct task_struct *child);
 779 extern void perf_event_exit_task(struct task_struct *child);
 780 extern void perf_event_free_task(struct task_struct *task);
 781 extern void set_perf_event_pending(void);
 782 extern void perf_event_do_pending(void);
 783 extern void perf_event_print_debug(void);
 784 extern void __perf_disable(void);
 785 extern bool __perf_enable(void);
 786 extern void perf_disable(void);
 787 extern void perf_enable(void);
 788 extern int perf_event_task_disable(void);
 789 extern int perf_event_task_enable(void);
 790 extern int hw_perf_group_sched_in(struct perf_event *group_leader,
 791                struct perf_cpu_context *cpuctx,
 792                struct perf_event_context *ctx);
 793 extern void perf_event_update_userpage(struct perf_event *event);
 794 extern int perf_event_release_kernel(struct perf_event *event);
 795 extern struct perf_event *
 796 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 797                                 int cpu,
 798                                 pid_t pid,
 799                                 perf_overflow_handler_t callback);
 800 extern u64 perf_event_read_value(struct perf_event *event,
 801                                  u64 *enabled, u64 *running);
 802
 803 struct perf_sample_data {
 804         u64                             type;
 805
 806         u64                             ip;
 807         struct {
 808                 u32     pid;
 809                 u32     tid;
 810         }                               tid_entry;
 811         u64                             time;
 812         u64                             addr;
 813         u64                             id;
 814         u64                             stream_id;
 815         struct {
 816                 u32     cpu;
 817                 u32     reserved;
 818         }                               cpu_entry;
 819         u64                             period;
 820         struct perf_callchain_entry     *callchain;
 821         struct perf_raw_record          *raw;
 822 };
 823
 824 static inline
 825 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
 826 {
 827         data->addr = addr;
 828         data->raw  = NULL;
 829 }
 830
 831 extern void perf_output_sample(struct perf_output_handle *handle,
 832                                struct perf_event_header *header,
 833                                struct perf_sample_data *data,
 834                                struct perf_event *event);
 835 extern void perf_prepare_sample(struct perf_event_header *header,
 836                                 struct perf_sample_data *data,
 837                                 struct perf_event *event,
 838                                 struct pt_regs *regs);
 839
 840 extern int perf_event_overflow(struct perf_event *event, int nmi,
 841                                  struct perf_sample_data *data,
 842                                  struct pt_regs *regs);
 843
 844 /*
 845  * Return 1 for a software event, 0 for a hardware event
 846  */
 847 static inline int is_software_event(struct perf_event *event)
 848 {
 849         switch (event->attr.type) {
 850         case PERF_TYPE_SOFTWARE:
 851         case PERF_TYPE_TRACEPOINT:
 852         /* for now the breakpoint stuff also works as software event */
 853         case PERF_TYPE_BREAKPOINT:
 854                 return 1;
 855         }
 856         return 0;
 857 }
 858
 859 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
 860
 861 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
 862
 863 static inline void
 864 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 865 {
 866         if (atomic_read(&perf_swevent_enabled[event_id]))
 867                 __perf_sw_event(event_id, nr, nmi, regs, addr);
 868 }
 869
 870 extern void
 871 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
 872
 873 /*
 874  * Take a snapshot of the regs. Skip ip and frame pointer to
 875  * the nth caller. We only need a few of the regs:
 876  * - ip for PERF_SAMPLE_IP
 877  * - cs for user_mode() tests
 878  * - bp for callchains
 879  * - eflags, for future purposes, just in case
 880  */
 881 static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
 882 {
 883         unsigned long ip;
 884
 885         memset(regs, 0, sizeof(*regs));
 886
 887         switch (skip) {
 888         case 1 :
 889                 ip = CALLER_ADDR0;
 890                 break;
 891         case 2 :
 892                 ip = CALLER_ADDR1;
 893                 break;
 894         case 3 :
 895                 ip = CALLER_ADDR2;
 896                 break;
 897         case 4:
 898                 ip = CALLER_ADDR3;
 899                 break;
 900         /* No need to support further for now */
 901         default:
 902                 ip = 0;
 903         }
 904
 905         return perf_arch_fetch_caller_regs(regs, ip, skip);
 906 }
 907
 908 extern void __perf_event_mmap(struct vm_area_struct *vma);
 909
 910 static inline void perf_event_mmap(struct vm_area_struct *vma)
 911 {
 912         if (vma->vm_flags & VM_EXEC)
 913                 __perf_event_mmap(vma);
 914 }
 915
 916 extern void perf_event_comm(struct task_struct *tsk);
 917 extern void perf_event_fork(struct task_struct *tsk);
 918
 919 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 920
 921 extern int sysctl_perf_event_paranoid;
 922 extern int sysctl_perf_event_mlock;
 923 extern int sysctl_perf_event_sample_rate;
 924
 925 static inline bool perf_paranoid_tracepoint_raw(void)
 926 {
 927         return sysctl_perf_event_paranoid > -1;
 928 }
 929
 930 static inline bool perf_paranoid_cpu(void)
 931 {
 932         return sysctl_perf_event_paranoid > 0;
 933 }
 934
 935 static inline bool perf_paranoid_kernel(void)
 936 {
 937         return sysctl_perf_event_paranoid > 1;
 938 }
 939
 940 extern void perf_event_init(void);
 941 extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
 942                           int entry_size, struct pt_regs *regs);
 943 extern void perf_bp_event(struct perf_event *event, void *data);
 944
 945 #ifndef perf_misc_flags
 946 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
 947                                  PERF_RECORD_MISC_KERNEL)
 948 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 949 #endif
 950
 951 extern int perf_output_begin(struct perf_output_handle *handle,
 952                              struct perf_event *event, unsigned int size,
 953                              int nmi, int sample);
 954 extern void perf_output_end(struct perf_output_handle *handle);
 955 extern void perf_output_copy(struct perf_output_handle *handle,
 956                              const void *buf, unsigned int len);
 957 extern int perf_swevent_get_recursion_context(void);
 958 extern void perf_swevent_put_recursion_context(int rctx);
 959 extern void perf_event_enable(struct perf_event *event);
 960 extern void perf_event_disable(struct perf_event *event);
 961 #else
 962 static inline void
 963 perf_event_task_sched_in(struct task_struct *task)                      { }
 964 static inline void
 965 perf_event_task_sched_out(struct task_struct *task,
 966                             struct task_struct *next)                   { }
 967 static inline void
 968 perf_event_task_tick(struct task_struct *task)                          { }
 969 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
 970 static inline void perf_event_exit_task(struct task_struct *child)      { }
 971 static inline void perf_event_free_task(struct task_struct *task)       { }
 972 static inline void perf_event_do_pending(void)                          { }
 973 static inline void perf_event_print_debug(void)                         { }
 974 static inline void perf_disable(void)                                   { }
 975 static inline void perf_enable(void)                                    { }
 976 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
 977 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
 978
 979 static inline void
 980 perf_sw_event(u32 event_id, u64 nr, int nmi,
 981                      struct pt_regs *regs, u64 addr)                    { }
 982 static inline void
 983 perf_bp_event(struct perf_event *event, void *data)                     { }
 984
 985 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
 986 static inline void perf_event_comm(struct task_struct *tsk)             { }
 987 static inline void perf_event_fork(struct task_struct *tsk)             { }
 988 static inline void perf_event_init(void)                                { }
 989 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
 990 static inline void perf_swevent_put_recursion_context(int rctx)         { }
 991 static inline void perf_event_enable(struct perf_event *event)          { }
 992 static inline void perf_event_disable(struct perf_event *event)         { }
 993 #endif
 994
 995 #define perf_output_put(handle, x) \
 996         perf_output_copy((handle), &(x), sizeof(x))
 997
 998 /*
 999  * This has to have a higher priority than migration_notifier in sched.c.
1000  */
1001 #define perf_cpu_notifier(fn)                                   \
1002 do {                                                            \
1003         static struct notifier_block fn##_nb __cpuinitdata =    \
1004                 { .notifier_call = fn, .priority = 20 };        \
1005         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1006                 (void *)(unsigned long)smp_processor_id());     \
1007         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1008                 (void *)(unsigned long)smp_processor_id());     \
1009         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1010                 (void *)(unsigned long)smp_processor_id());     \
1011         register_cpu_notifier(&fn##_nb);                        \
1012 } while (0)
1013
1014 #endif /* __KERNEL__ */
1015 #endif /* _LINUX_PERF_EVENT_H */