include/linux/perf_counter.h

   1 /*
   2  *  Performance counters:
   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_COUNTER_H
  15 #define _LINUX_PERF_COUNTER_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 counter event types, used by the
  40  * attr.event_id parameter of the sys_perf_counter_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 counters:
  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" counters provided by the kernel, even if the hardware
  93  * does not support performance counters. These counters 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_counter_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 counter 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_counter_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 to monitor via a performance monitoring counter:
 162  */
 163 struct perf_counter_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
 203                                 __reserved_1   : 50;
 204
 205         __u32                   wakeup_events;  /* wakeup every n events */
 206         __u32                   __reserved_2;
 207
 208         __u64                   __reserved_3;
 209 };
 210
 211 /*
 212  * Ioctls that can be done on a perf counter fd:
 213  */
 214 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 215 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 216 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
 217 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
 218 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
 219 #define PERF_COUNTER_IOC_SET_OUTPUT     _IO ('$', 5)
 220
 221 enum perf_counter_ioc_flags {
 222         PERF_IOC_FLAG_GROUP             = 1U << 0,
 223 };
 224
 225 /*
 226  * Structure of the page that can be mapped via mmap
 227  */
 228 struct perf_counter_mmap_page {
 229         __u32   version;                /* version number of this structure */
 230         __u32   compat_version;         /* lowest version this is compat with */
 231
 232         /*
 233          * Bits needed to read the hw counters in user-space.
 234          *
 235          *   u32 seq;
 236          *   s64 count;
 237          *
 238          *   do {
 239          *     seq = pc->lock;
 240          *
 241          *     barrier()
 242          *     if (pc->index) {
 243          *       count = pmc_read(pc->index - 1);
 244          *       count += pc->offset;
 245          *     } else
 246          *       goto regular_read;
 247          *
 248          *     barrier();
 249          *   } while (pc->lock != seq);
 250          *
 251          * NOTE: for obvious reason this only works on self-monitoring
 252          *       processes.
 253          */
 254         __u32   lock;                   /* seqlock for synchronization */
 255         __u32   index;                  /* hardware counter identifier */
 256         __s64   offset;                 /* add to hardware counter value */
 257         __u64   time_enabled;           /* time counter active */
 258         __u64   time_running;           /* time counter on cpu */
 259
 260                 /*
 261                  * Hole for extension of the self monitor capabilities
 262                  */
 263
 264         __u64   __reserved[123];        /* align to 1k */
 265
 266         /*
 267          * Control data for the mmap() data buffer.
 268          *
 269          * User-space reading the @data_head value should issue an rmb(), on
 270          * SMP capable platforms, after reading this value -- see
 271          * perf_counter_wakeup().
 272          *
 273          * When the mapping is PROT_WRITE the @data_tail value should be
 274          * written by userspace to reflect the last read data. In this case
 275          * the kernel will not over-write unread data.
 276          */
 277         __u64   data_head;              /* head in the data section */
 278         __u64   data_tail;              /* user-space written tail */
 279 };
 280
 281 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
 282 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
 283 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
 284 #define PERF_EVENT_MISC_USER                    (2 << 0)
 285 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
 286
 287 struct perf_event_header {
 288         __u32   type;
 289         __u16   misc;
 290         __u16   size;
 291 };
 292
 293 enum perf_event_type {
 294
 295         /*
 296          * The MMAP events record the PROT_EXEC mappings so that we can
 297          * correlate userspace IPs to code. They have the following structure:
 298          *
 299          * struct {
 300          *      struct perf_event_header        header;
 301          *
 302          *      u32                             pid, tid;
 303          *      u64                             addr;
 304          *      u64                             len;
 305          *      u64                             pgoff;
 306          *      char                            filename[];
 307          * };
 308          */
 309         PERF_EVENT_MMAP                 = 1,
 310
 311         /*
 312          * struct {
 313          *      struct perf_event_header        header;
 314          *      u64                             id;
 315          *      u64                             lost;
 316          * };
 317          */
 318         PERF_EVENT_LOST                 = 2,
 319
 320         /*
 321          * struct {
 322          *      struct perf_event_header        header;
 323          *
 324          *      u32                             pid, tid;
 325          *      char                            comm[];
 326          * };
 327          */
 328         PERF_EVENT_COMM                 = 3,
 329
 330         /*
 331          * struct {
 332          *      struct perf_event_header        header;
 333          *      u32                             pid, ppid;
 334          *      u32                             tid, ptid;
 335          * };
 336          */
 337         PERF_EVENT_EXIT                 = 4,
 338
 339         /*
 340          * struct {
 341          *      struct perf_event_header        header;
 342          *      u64                             time;
 343          *      u64                             id;
 344          *      u64                             stream_id;
 345          * };
 346          */
 347         PERF_EVENT_THROTTLE             = 5,
 348         PERF_EVENT_UNTHROTTLE           = 6,
 349
 350         /*
 351          * struct {
 352          *      struct perf_event_header        header;
 353          *      u32                             pid, ppid;
 354          *      u32                             tid, ptid;
 355          * };
 356          */
 357         PERF_EVENT_FORK                 = 7,
 358
 359         /*
 360          * struct {
 361          *      struct perf_event_header        header;
 362          *      u32                             pid, tid;
 363          *
 364          *      struct read_format              values;
 365          * };
 366          */
 367         PERF_EVENT_READ                 = 8,
 368
 369         /*
 370          * struct {
 371          *      struct perf_event_header        header;
 372          *
 373          *      { u64                   ip;       } && PERF_SAMPLE_IP
 374          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 375          *      { u64                   time;     } && PERF_SAMPLE_TIME
 376          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 377          *      { u64                   id;       } && PERF_SAMPLE_ID
 378          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 379          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 380          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 381          *
 382          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 383          *
 384          *      { u64                   nr,
 385          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 386          *
 387          *      #
 388          *      # The RAW record below is opaque data wrt the ABI
 389          *      #
 390          *      # That is, the ABI doesn't make any promises wrt to
 391          *      # the stability of its content, it may vary depending
 392          *      # on event, hardware, kernel version and phase of
 393          *      # the moon.
 394          *      #
 395          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 396          *      #
 397          *
 398          *      { u32                   size;
 399          *        char                  data[size];}&& PERF_SAMPLE_RAW
 400          * };
 401          */
 402         PERF_EVENT_SAMPLE               = 9,
 403
 404         PERF_EVENT_MAX,                 /* non-ABI */
 405 };
 406
 407 enum perf_callchain_context {
 408         PERF_CONTEXT_HV                 = (__u64)-32,
 409         PERF_CONTEXT_KERNEL             = (__u64)-128,
 410         PERF_CONTEXT_USER               = (__u64)-512,
 411
 412         PERF_CONTEXT_GUEST              = (__u64)-2048,
 413         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 414         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 415
 416         PERF_CONTEXT_MAX                = (__u64)-4095,
 417 };
 418
 419 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 420 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 421
 422 #ifdef __KERNEL__
 423 /*
 424  * Kernel-internal data types and definitions:
 425  */
 426
 427 #ifdef CONFIG_PERF_COUNTERS
 428 # include <asm/perf_counter.h>
 429 #endif
 430
 431 #include <linux/list.h>
 432 #include <linux/mutex.h>
 433 #include <linux/rculist.h>
 434 #include <linux/rcupdate.h>
 435 #include <linux/spinlock.h>
 436 #include <linux/hrtimer.h>
 437 #include <linux/fs.h>
 438 #include <linux/pid_namespace.h>
 439 #include <asm/atomic.h>
 440
 441 #define PERF_MAX_STACK_DEPTH            255
 442
 443 struct perf_callchain_entry {
 444         __u64                           nr;
 445         __u64                           ip[PERF_MAX_STACK_DEPTH];
 446 };
 447
 448 struct perf_raw_record {
 449         u32                             size;
 450         void                            *data;
 451 };
 452
 453 struct task_struct;
 454
 455 /**
 456  * struct hw_perf_counter - performance counter hardware details:
 457  */
 458 struct hw_perf_counter {
 459 #ifdef CONFIG_PERF_COUNTERS
 460         union {
 461                 struct { /* hardware */
 462                         u64             config;
 463                         unsigned long   config_base;
 464                         unsigned long   counter_base;
 465                         int             idx;
 466                 };
 467                 union { /* software */
 468                         atomic64_t      count;
 469                         struct hrtimer  hrtimer;
 470                 };
 471         };
 472         atomic64_t                      prev_count;
 473         u64                             sample_period;
 474         u64                             last_period;
 475         atomic64_t                      period_left;
 476         u64                             interrupts;
 477
 478         u64                             freq_count;
 479         u64                             freq_interrupts;
 480         u64                             freq_stamp;
 481 #endif
 482 };
 483
 484 struct perf_counter;
 485
 486 /**
 487  * struct pmu - generic performance monitoring unit
 488  */
 489 struct pmu {
 490         int (*enable)                   (struct perf_counter *counter);
 491         void (*disable)                 (struct perf_counter *counter);
 492         void (*read)                    (struct perf_counter *counter);
 493         void (*unthrottle)              (struct perf_counter *counter);
 494 };
 495
 496 /**
 497  * enum perf_counter_active_state - the states of a counter
 498  */
 499 enum perf_counter_active_state {
 500         PERF_COUNTER_STATE_ERROR        = -2,
 501         PERF_COUNTER_STATE_OFF          = -1,
 502         PERF_COUNTER_STATE_INACTIVE     =  0,
 503         PERF_COUNTER_STATE_ACTIVE       =  1,
 504 };
 505
 506 struct file;
 507
 508 struct perf_mmap_data {
 509         struct rcu_head                 rcu_head;
 510         int                             nr_pages;       /* nr of data pages  */
 511         int                             writable;       /* are we writable   */
 512         int                             nr_locked;      /* nr pages mlocked  */
 513
 514         atomic_t                        poll;           /* POLL_ for wakeups */
 515         atomic_t                        events;         /* event limit       */
 516
 517         atomic_long_t                   head;           /* write position    */
 518         atomic_long_t                   done_head;      /* completed head    */
 519
 520         atomic_t                        lock;           /* concurrent writes */
 521         atomic_t                        wakeup;         /* needs a wakeup    */
 522         atomic_t                        lost;           /* nr records lost   */
 523
 524         struct perf_counter_mmap_page   *user_page;
 525         void                            *data_pages[0];
 526 };
 527
 528 struct perf_pending_entry {
 529         struct perf_pending_entry *next;
 530         void (*func)(struct perf_pending_entry *);
 531 };
 532
 533 /**
 534  * struct perf_counter - performance counter kernel representation:
 535  */
 536 struct perf_counter {
 537 #ifdef CONFIG_PERF_COUNTERS
 538         struct list_head                list_entry;
 539         struct list_head                event_entry;
 540         struct list_head                sibling_list;
 541         int                             nr_siblings;
 542         struct perf_counter             *group_leader;
 543         struct perf_counter             *output;
 544         const struct pmu                *pmu;
 545
 546         enum perf_counter_active_state  state;
 547         atomic64_t                      count;
 548
 549         /*
 550          * These are the total time in nanoseconds that the counter
 551          * has been enabled (i.e. eligible to run, and the task has
 552          * been scheduled in, if this is a per-task counter)
 553          * and running (scheduled onto the CPU), respectively.
 554          *
 555          * They are computed from tstamp_enabled, tstamp_running and
 556          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 557          */
 558         u64                             total_time_enabled;
 559         u64                             total_time_running;
 560
 561         /*
 562          * These are timestamps used for computing total_time_enabled
 563          * and total_time_running when the counter is in INACTIVE or
 564          * ACTIVE state, measured in nanoseconds from an arbitrary point
 565          * in time.
 566          * tstamp_enabled: the notional time when the counter was enabled
 567          * tstamp_running: the notional time when the counter was scheduled on
 568          * tstamp_stopped: in INACTIVE state, the notional time when the
 569          *      counter was scheduled off.
 570          */
 571         u64                             tstamp_enabled;
 572         u64                             tstamp_running;
 573         u64                             tstamp_stopped;
 574
 575         struct perf_counter_attr        attr;
 576         struct hw_perf_counter          hw;
 577
 578         struct perf_counter_context     *ctx;
 579         struct file                     *filp;
 580
 581         /*
 582          * These accumulate total time (in nanoseconds) that children
 583          * counters have been enabled and running, respectively.
 584          */
 585         atomic64_t                      child_total_time_enabled;
 586         atomic64_t                      child_total_time_running;
 587
 588         /*
 589          * Protect attach/detach and child_list:
 590          */
 591         struct mutex                    child_mutex;
 592         struct list_head                child_list;
 593         struct perf_counter             *parent;
 594
 595         int                             oncpu;
 596         int                             cpu;
 597
 598         struct list_head                owner_entry;
 599         struct task_struct              *owner;
 600
 601         /* mmap bits */
 602         struct mutex                    mmap_mutex;
 603         atomic_t                        mmap_count;
 604         struct perf_mmap_data           *data;
 605
 606         /* poll related */
 607         wait_queue_head_t               waitq;
 608         struct fasync_struct            *fasync;
 609
 610         /* delayed work for NMIs and such */
 611         int                             pending_wakeup;
 612         int                             pending_kill;
 613         int                             pending_disable;
 614         struct perf_pending_entry       pending;
 615
 616         atomic_t                        event_limit;
 617
 618         void (*destroy)(struct perf_counter *);
 619         struct rcu_head                 rcu_head;
 620
 621         struct pid_namespace            *ns;
 622         u64                             id;
 623 #endif
 624 };
 625
 626 /**
 627  * struct perf_counter_context - counter context structure
 628  *
 629  * Used as a container for task counters and CPU counters as well:
 630  */
 631 struct perf_counter_context {
 632         /*
 633          * Protect the states of the counters in the list,
 634          * nr_active, and the list:
 635          */
 636         spinlock_t                      lock;
 637         /*
 638          * Protect the list of counters.  Locking either mutex or lock
 639          * is sufficient to ensure the list doesn't change; to change
 640          * the list you need to lock both the mutex and the spinlock.
 641          */
 642         struct mutex                    mutex;
 643
 644         struct list_head                counter_list;
 645         struct list_head                event_list;
 646         int                             nr_counters;
 647         int                             nr_active;
 648         int                             is_active;
 649         int                             nr_stat;
 650         atomic_t                        refcount;
 651         struct task_struct              *task;
 652
 653         /*
 654          * Context clock, runs when context enabled.
 655          */
 656         u64                             time;
 657         u64                             timestamp;
 658
 659         /*
 660          * These fields let us detect when two contexts have both
 661          * been cloned (inherited) from a common ancestor.
 662          */
 663         struct perf_counter_context     *parent_ctx;
 664         u64                             parent_gen;
 665         u64                             generation;
 666         int                             pin_count;
 667         struct rcu_head                 rcu_head;
 668 };
 669
 670 /**
 671  * struct perf_counter_cpu_context - per cpu counter context structure
 672  */
 673 struct perf_cpu_context {
 674         struct perf_counter_context     ctx;
 675         struct perf_counter_context     *task_ctx;
 676         int                             active_oncpu;
 677         int                             max_pertask;
 678         int                             exclusive;
 679
 680         /*
 681          * Recursion avoidance:
 682          *
 683          * task, softirq, irq, nmi context
 684          */
 685         int                             recursion[4];
 686 };
 687
 688 #ifdef CONFIG_PERF_COUNTERS
 689
 690 /*
 691  * Set by architecture code:
 692  */
 693 extern int perf_max_counters;
 694
 695 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
 696
 697 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 698 extern void perf_counter_task_sched_out(struct task_struct *task,
 699                                         struct task_struct *next, int cpu);
 700 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 701 extern int perf_counter_init_task(struct task_struct *child);
 702 extern void perf_counter_exit_task(struct task_struct *child);
 703 extern void perf_counter_free_task(struct task_struct *task);
 704 extern void set_perf_counter_pending(void);
 705 extern void perf_counter_do_pending(void);
 706 extern void perf_counter_print_debug(void);
 707 extern void __perf_disable(void);
 708 extern bool __perf_enable(void);
 709 extern void perf_disable(void);
 710 extern void perf_enable(void);
 711 extern int perf_counter_task_disable(void);
 712 extern int perf_counter_task_enable(void);
 713 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 714                struct perf_cpu_context *cpuctx,
 715                struct perf_counter_context *ctx, int cpu);
 716 extern void perf_counter_update_userpage(struct perf_counter *counter);
 717
 718 struct perf_sample_data {
 719         struct pt_regs                  *regs;
 720         u64                             addr;
 721         u64                             period;
 722         struct perf_raw_record          *raw;
 723 };
 724
 725 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
 726                                  struct perf_sample_data *data);
 727 extern void perf_counter_output(struct perf_counter *counter, int nmi,
 728                                 struct perf_sample_data *data);
 729
 730 /*
 731  * Return 1 for a software counter, 0 for a hardware counter
 732  */
 733 static inline int is_software_counter(struct perf_counter *counter)
 734 {
 735         return (counter->attr.type != PERF_TYPE_RAW) &&
 736                 (counter->attr.type != PERF_TYPE_HARDWARE) &&
 737                 (counter->attr.type != PERF_TYPE_HW_CACHE);
 738 }
 739
 740 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
 741
 742 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 743
 744 static inline void
 745 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 746 {
 747         if (atomic_read(&perf_swcounter_enabled[event]))
 748                 __perf_swcounter_event(event, nr, nmi, regs, addr);
 749 }
 750
 751 extern void __perf_counter_mmap(struct vm_area_struct *vma);
 752
 753 static inline void perf_counter_mmap(struct vm_area_struct *vma)
 754 {
 755         if (vma->vm_flags & VM_EXEC)
 756                 __perf_counter_mmap(vma);
 757 }
 758
 759 extern void perf_counter_comm(struct task_struct *tsk);
 760 extern void perf_counter_fork(struct task_struct *tsk);
 761
 762 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 763
 764 extern int sysctl_perf_counter_paranoid;
 765 extern int sysctl_perf_counter_mlock;
 766 extern int sysctl_perf_counter_sample_rate;
 767
 768 extern void perf_counter_init(void);
 769
 770 #ifndef perf_misc_flags
 771 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
 772                                  PERF_EVENT_MISC_KERNEL)
 773 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 774 #endif
 775
 776 #else
 777 static inline void
 778 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 779 static inline void
 780 perf_counter_task_sched_out(struct task_struct *task,
 781                             struct task_struct *next, int cpu)          { }
 782 static inline void
 783 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 784 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
 785 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 786 static inline void perf_counter_free_task(struct task_struct *task)     { }
 787 static inline void perf_counter_do_pending(void)                        { }
 788 static inline void perf_counter_print_debug(void)                       { }
 789 static inline void perf_disable(void)                                   { }
 790 static inline void perf_enable(void)                                    { }
 791 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 792 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 793
 794 static inline void
 795 perf_swcounter_event(u32 event, u64 nr, int nmi,
 796                      struct pt_regs *regs, u64 addr)                    { }
 797
 798 static inline void perf_counter_mmap(struct vm_area_struct *vma)        { }
 799 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 800 static inline void perf_counter_fork(struct task_struct *tsk)           { }
 801 static inline void perf_counter_init(void)                              { }
 802 #endif
 803
 804 #endif /* __KERNEL__ */
 805 #endif /* _LINUX_PERF_COUNTER_H */