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                                 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 counter fd:
 217  */
 218 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 219 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 220 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
 221 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
 222 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
 223 #define PERF_COUNTER_IOC_SET_OUTPUT     _IO ('$', 5)
 224
 225 enum perf_counter_ioc_flags {
 226         PERF_IOC_FLAG_GROUP             = 1U << 0,
 227 };
 228
 229 /*
 230  * Structure of the page that can be mapped via mmap
 231  */
 232 struct perf_counter_mmap_page {
 233         __u32   version;                /* version number of this structure */
 234         __u32   compat_version;         /* lowest version this is compat with */
 235
 236         /*
 237          * Bits needed to read the hw counters in user-space.
 238          *
 239          *   u32 seq;
 240          *   s64 count;
 241          *
 242          *   do {
 243          *     seq = pc->lock;
 244          *
 245          *     barrier()
 246          *     if (pc->index) {
 247          *       count = pmc_read(pc->index - 1);
 248          *       count += pc->offset;
 249          *     } else
 250          *       goto regular_read;
 251          *
 252          *     barrier();
 253          *   } while (pc->lock != seq);
 254          *
 255          * NOTE: for obvious reason this only works on self-monitoring
 256          *       processes.
 257          */
 258         __u32   lock;                   /* seqlock for synchronization */
 259         __u32   index;                  /* hardware counter identifier */
 260         __s64   offset;                 /* add to hardware counter value */
 261         __u64   time_enabled;           /* time counter active */
 262         __u64   time_running;           /* time counter on cpu */
 263
 264                 /*
 265                  * Hole for extension of the self monitor capabilities
 266                  */
 267
 268         __u64   __reserved[123];        /* align to 1k */
 269
 270         /*
 271          * Control data for the mmap() data buffer.
 272          *
 273          * User-space reading the @data_head value should issue an rmb(), on
 274          * SMP capable platforms, after reading this value -- see
 275          * perf_counter_wakeup().
 276          *
 277          * When the mapping is PROT_WRITE the @data_tail value should be
 278          * written by userspace to reflect the last read data. In this case
 279          * the kernel will not over-write unread data.
 280          */
 281         __u64   data_head;              /* head in the data section */
 282         __u64   data_tail;              /* user-space written tail */
 283 };
 284
 285 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
 286 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
 287 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
 288 #define PERF_EVENT_MISC_USER                    (2 << 0)
 289 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
 290
 291 struct perf_event_header {
 292         __u32   type;
 293         __u16   misc;
 294         __u16   size;
 295 };
 296
 297 enum perf_event_type {
 298
 299         /*
 300          * The MMAP events record the PROT_EXEC mappings so that we can
 301          * correlate userspace IPs to code. They have the following structure:
 302          *
 303          * struct {
 304          *      struct perf_event_header        header;
 305          *
 306          *      u32                             pid, tid;
 307          *      u64                             addr;
 308          *      u64                             len;
 309          *      u64                             pgoff;
 310          *      char                            filename[];
 311          * };
 312          */
 313         PERF_EVENT_MMAP                 = 1,
 314
 315         /*
 316          * struct {
 317          *      struct perf_event_header        header;
 318          *      u64                             id;
 319          *      u64                             lost;
 320          * };
 321          */
 322         PERF_EVENT_LOST                 = 2,
 323
 324         /*
 325          * struct {
 326          *      struct perf_event_header        header;
 327          *
 328          *      u32                             pid, tid;
 329          *      char                            comm[];
 330          * };
 331          */
 332         PERF_EVENT_COMM                 = 3,
 333
 334         /*
 335          * struct {
 336          *      struct perf_event_header        header;
 337          *      u32                             pid, ppid;
 338          *      u32                             tid, ptid;
 339          * };
 340          */
 341         PERF_EVENT_EXIT                 = 4,
 342
 343         /*
 344          * struct {
 345          *      struct perf_event_header        header;
 346          *      u64                             time;
 347          *      u64                             id;
 348          *      u64                             stream_id;
 349          * };
 350          */
 351         PERF_EVENT_THROTTLE             = 5,
 352         PERF_EVENT_UNTHROTTLE           = 6,
 353
 354         /*
 355          * struct {
 356          *      struct perf_event_header        header;
 357          *      u32                             pid, ppid;
 358          *      u32                             tid, ptid;
 359          * };
 360          */
 361         PERF_EVENT_FORK                 = 7,
 362
 363         /*
 364          * struct {
 365          *      struct perf_event_header        header;
 366          *      u32                             pid, tid;
 367          *
 368          *      struct read_format              values;
 369          * };
 370          */
 371         PERF_EVENT_READ                 = 8,
 372
 373         /*
 374          * struct {
 375          *      struct perf_event_header        header;
 376          *
 377          *      { u64                   ip;       } && PERF_SAMPLE_IP
 378          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 379          *      { u64                   time;     } && PERF_SAMPLE_TIME
 380          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 381          *      { u64                   id;       } && PERF_SAMPLE_ID
 382          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 383          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 384          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 385          *
 386          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 387          *
 388          *      { u64                   nr,
 389          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 390          *
 391          *      #
 392          *      # The RAW record below is opaque data wrt the ABI
 393          *      #
 394          *      # That is, the ABI doesn't make any promises wrt to
 395          *      # the stability of its content, it may vary depending
 396          *      # on event, hardware, kernel version and phase of
 397          *      # the moon.
 398          *      #
 399          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 400          *      #
 401          *
 402          *      { u32                   size;
 403          *        char                  data[size];}&& PERF_SAMPLE_RAW
 404          * };
 405          */
 406         PERF_EVENT_SAMPLE               = 9,
 407
 408         PERF_EVENT_MAX,                 /* non-ABI */
 409 };
 410
 411 enum perf_callchain_context {
 412         PERF_CONTEXT_HV                 = (__u64)-32,
 413         PERF_CONTEXT_KERNEL             = (__u64)-128,
 414         PERF_CONTEXT_USER               = (__u64)-512,
 415
 416         PERF_CONTEXT_GUEST              = (__u64)-2048,
 417         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 418         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 419
 420         PERF_CONTEXT_MAX                = (__u64)-4095,
 421 };
 422
 423 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 424 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 425
 426 #ifdef __KERNEL__
 427 /*
 428  * Kernel-internal data types and definitions:
 429  */
 430
 431 #ifdef CONFIG_PERF_COUNTERS
 432 # include <asm/perf_counter.h>
 433 #endif
 434
 435 #include <linux/list.h>
 436 #include <linux/mutex.h>
 437 #include <linux/rculist.h>
 438 #include <linux/rcupdate.h>
 439 #include <linux/spinlock.h>
 440 #include <linux/hrtimer.h>
 441 #include <linux/fs.h>
 442 #include <linux/pid_namespace.h>
 443 #include <asm/atomic.h>
 444
 445 #define PERF_MAX_STACK_DEPTH            255
 446
 447 struct perf_callchain_entry {
 448         __u64                           nr;
 449         __u64                           ip[PERF_MAX_STACK_DEPTH];
 450 };
 451
 452 struct perf_raw_record {
 453         u32                             size;
 454         void                            *data;
 455 };
 456
 457 struct task_struct;
 458
 459 /**
 460  * struct hw_perf_counter - performance counter hardware details:
 461  */
 462 struct hw_perf_counter {
 463 #ifdef CONFIG_PERF_COUNTERS
 464         union {
 465                 struct { /* hardware */
 466                         u64             config;
 467                         unsigned long   config_base;
 468                         unsigned long   counter_base;
 469                         int             idx;
 470                 };
 471                 union { /* software */
 472                         atomic64_t      count;
 473                         struct hrtimer  hrtimer;
 474                 };
 475         };
 476         atomic64_t                      prev_count;
 477         u64                             sample_period;
 478         u64                             last_period;
 479         atomic64_t                      period_left;
 480         u64                             interrupts;
 481
 482         u64                             freq_count;
 483         u64                             freq_interrupts;
 484         u64                             freq_stamp;
 485 #endif
 486 };
 487
 488 struct perf_counter;
 489
 490 /**
 491  * struct pmu - generic performance monitoring unit
 492  */
 493 struct pmu {
 494         int (*enable)                   (struct perf_counter *counter);
 495         void (*disable)                 (struct perf_counter *counter);
 496         void (*read)                    (struct perf_counter *counter);
 497         void (*unthrottle)              (struct perf_counter *counter);
 498 };
 499
 500 /**
 501  * enum perf_counter_active_state - the states of a counter
 502  */
 503 enum perf_counter_active_state {
 504         PERF_COUNTER_STATE_ERROR        = -2,
 505         PERF_COUNTER_STATE_OFF          = -1,
 506         PERF_COUNTER_STATE_INACTIVE     =  0,
 507         PERF_COUNTER_STATE_ACTIVE       =  1,
 508 };
 509
 510 struct file;
 511
 512 struct perf_mmap_data {
 513         struct rcu_head                 rcu_head;
 514         int                             nr_pages;       /* nr of data pages  */
 515         int                             writable;       /* are we writable   */
 516         int                             nr_locked;      /* nr pages mlocked  */
 517
 518         atomic_t                        poll;           /* POLL_ for wakeups */
 519         atomic_t                        events;         /* event limit       */
 520
 521         atomic_long_t                   head;           /* write position    */
 522         atomic_long_t                   done_head;      /* completed head    */
 523
 524         atomic_t                        lock;           /* concurrent writes */
 525         atomic_t                        wakeup;         /* needs a wakeup    */
 526         atomic_t                        lost;           /* nr records lost   */
 527
 528         long                            watermark;      /* wakeup watermark  */
 529
 530         struct perf_counter_mmap_page   *user_page;
 531         void                            *data_pages[0];
 532 };
 533
 534 struct perf_pending_entry {
 535         struct perf_pending_entry *next;
 536         void (*func)(struct perf_pending_entry *);
 537 };
 538
 539 /**
 540  * struct perf_counter - performance counter kernel representation:
 541  */
 542 struct perf_counter {
 543 #ifdef CONFIG_PERF_COUNTERS
 544         struct list_head                list_entry;
 545         struct list_head                event_entry;
 546         struct list_head                sibling_list;
 547         int                             nr_siblings;
 548         struct perf_counter             *group_leader;
 549         struct perf_counter             *output;
 550         const struct pmu                *pmu;
 551
 552         enum perf_counter_active_state  state;
 553         atomic64_t                      count;
 554
 555         /*
 556          * These are the total time in nanoseconds that the counter
 557          * has been enabled (i.e. eligible to run, and the task has
 558          * been scheduled in, if this is a per-task counter)
 559          * and running (scheduled onto the CPU), respectively.
 560          *
 561          * They are computed from tstamp_enabled, tstamp_running and
 562          * tstamp_stopped when the counter is in INACTIVE or ACTIVE state.
 563          */
 564         u64                             total_time_enabled;
 565         u64                             total_time_running;
 566
 567         /*
 568          * These are timestamps used for computing total_time_enabled
 569          * and total_time_running when the counter is in INACTIVE or
 570          * ACTIVE state, measured in nanoseconds from an arbitrary point
 571          * in time.
 572          * tstamp_enabled: the notional time when the counter was enabled
 573          * tstamp_running: the notional time when the counter was scheduled on
 574          * tstamp_stopped: in INACTIVE state, the notional time when the
 575          *      counter was scheduled off.
 576          */
 577         u64                             tstamp_enabled;
 578         u64                             tstamp_running;
 579         u64                             tstamp_stopped;
 580
 581         struct perf_counter_attr        attr;
 582         struct hw_perf_counter          hw;
 583
 584         struct perf_counter_context     *ctx;
 585         struct file                     *filp;
 586
 587         /*
 588          * These accumulate total time (in nanoseconds) that children
 589          * counters have been enabled and running, respectively.
 590          */
 591         atomic64_t                      child_total_time_enabled;
 592         atomic64_t                      child_total_time_running;
 593
 594         /*
 595          * Protect attach/detach and child_list:
 596          */
 597         struct mutex                    child_mutex;
 598         struct list_head                child_list;
 599         struct perf_counter             *parent;
 600
 601         int                             oncpu;
 602         int                             cpu;
 603
 604         struct list_head                owner_entry;
 605         struct task_struct              *owner;
 606
 607         /* mmap bits */
 608         struct mutex                    mmap_mutex;
 609         atomic_t                        mmap_count;
 610         struct perf_mmap_data           *data;
 611
 612         /* poll related */
 613         wait_queue_head_t               waitq;
 614         struct fasync_struct            *fasync;
 615
 616         /* delayed work for NMIs and such */
 617         int                             pending_wakeup;
 618         int                             pending_kill;
 619         int                             pending_disable;
 620         struct perf_pending_entry       pending;
 621
 622         atomic_t                        event_limit;
 623
 624         void (*destroy)(struct perf_counter *);
 625         struct rcu_head                 rcu_head;
 626
 627         struct pid_namespace            *ns;
 628         u64                             id;
 629 #endif
 630 };
 631
 632 /**
 633  * struct perf_counter_context - counter context structure
 634  *
 635  * Used as a container for task counters and CPU counters as well:
 636  */
 637 struct perf_counter_context {
 638         /*
 639          * Protect the states of the counters in the list,
 640          * nr_active, and the list:
 641          */
 642         spinlock_t                      lock;
 643         /*
 644          * Protect the list of counters.  Locking either mutex or lock
 645          * is sufficient to ensure the list doesn't change; to change
 646          * the list you need to lock both the mutex and the spinlock.
 647          */
 648         struct mutex                    mutex;
 649
 650         struct list_head                counter_list;
 651         struct list_head                event_list;
 652         int                             nr_counters;
 653         int                             nr_active;
 654         int                             is_active;
 655         int                             nr_stat;
 656         atomic_t                        refcount;
 657         struct task_struct              *task;
 658
 659         /*
 660          * Context clock, runs when context enabled.
 661          */
 662         u64                             time;
 663         u64                             timestamp;
 664
 665         /*
 666          * These fields let us detect when two contexts have both
 667          * been cloned (inherited) from a common ancestor.
 668          */
 669         struct perf_counter_context     *parent_ctx;
 670         u64                             parent_gen;
 671         u64                             generation;
 672         int                             pin_count;
 673         struct rcu_head                 rcu_head;
 674 };
 675
 676 /**
 677  * struct perf_counter_cpu_context - per cpu counter context structure
 678  */
 679 struct perf_cpu_context {
 680         struct perf_counter_context     ctx;
 681         struct perf_counter_context     *task_ctx;
 682         int                             active_oncpu;
 683         int                             max_pertask;
 684         int                             exclusive;
 685
 686         /*
 687          * Recursion avoidance:
 688          *
 689          * task, softirq, irq, nmi context
 690          */
 691         int                             recursion[4];
 692 };
 693
 694 struct perf_output_handle {
 695         struct perf_counter     *counter;
 696         struct perf_mmap_data   *data;
 697         unsigned long           head;
 698         unsigned long           offset;
 699         int                     nmi;
 700         int                     sample;
 701         int                     locked;
 702         unsigned long           flags;
 703 };
 704
 705 #ifdef CONFIG_PERF_COUNTERS
 706
 707 /*
 708  * Set by architecture code:
 709  */
 710 extern int perf_max_counters;
 711
 712 extern const struct pmu *hw_perf_counter_init(struct perf_counter *counter);
 713
 714 extern void perf_counter_task_sched_in(struct task_struct *task, int cpu);
 715 extern void perf_counter_task_sched_out(struct task_struct *task,
 716                                         struct task_struct *next, int cpu);
 717 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 718 extern int perf_counter_init_task(struct task_struct *child);
 719 extern void perf_counter_exit_task(struct task_struct *child);
 720 extern void perf_counter_free_task(struct task_struct *task);
 721 extern void set_perf_counter_pending(void);
 722 extern void perf_counter_do_pending(void);
 723 extern void perf_counter_print_debug(void);
 724 extern void __perf_disable(void);
 725 extern bool __perf_enable(void);
 726 extern void perf_disable(void);
 727 extern void perf_enable(void);
 728 extern int perf_counter_task_disable(void);
 729 extern int perf_counter_task_enable(void);
 730 extern int hw_perf_group_sched_in(struct perf_counter *group_leader,
 731                struct perf_cpu_context *cpuctx,
 732                struct perf_counter_context *ctx, int cpu);
 733 extern void perf_counter_update_userpage(struct perf_counter *counter);
 734
 735 struct perf_sample_data {
 736         u64                             type;
 737
 738         u64                             ip;
 739         struct {
 740                 u32     pid;
 741                 u32     tid;
 742         }                               tid_entry;
 743         u64                             time;
 744         u64                             addr;
 745         u64                             id;
 746         u64                             stream_id;
 747         struct {
 748                 u32     cpu;
 749                 u32     reserved;
 750         }                               cpu_entry;
 751         u64                             period;
 752         struct perf_callchain_entry     *callchain;
 753         struct perf_raw_record          *raw;
 754 };
 755
 756 extern void perf_output_sample(struct perf_output_handle *handle,
 757                                struct perf_event_header *header,
 758                                struct perf_sample_data *data,
 759                                struct perf_counter *counter);
 760 extern void perf_prepare_sample(struct perf_event_header *header,
 761                                 struct perf_sample_data *data,
 762                                 struct perf_counter *counter,
 763                                 struct pt_regs *regs);
 764
 765 extern int perf_counter_overflow(struct perf_counter *counter, int nmi,
 766                                  struct perf_sample_data *data,
 767                                  struct pt_regs *regs);
 768
 769 /*
 770  * Return 1 for a software counter, 0 for a hardware counter
 771  */
 772 static inline int is_software_counter(struct perf_counter *counter)
 773 {
 774         return (counter->attr.type != PERF_TYPE_RAW) &&
 775                 (counter->attr.type != PERF_TYPE_HARDWARE) &&
 776                 (counter->attr.type != PERF_TYPE_HW_CACHE);
 777 }
 778
 779 extern atomic_t perf_swcounter_enabled[PERF_COUNT_SW_MAX];
 780
 781 extern void __perf_swcounter_event(u32, u64, int, struct pt_regs *, u64);
 782
 783 static inline void
 784 perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 785 {
 786         if (atomic_read(&perf_swcounter_enabled[event]))
 787                 __perf_swcounter_event(event, nr, nmi, regs, addr);
 788 }
 789
 790 extern void __perf_counter_mmap(struct vm_area_struct *vma);
 791
 792 static inline void perf_counter_mmap(struct vm_area_struct *vma)
 793 {
 794         if (vma->vm_flags & VM_EXEC)
 795                 __perf_counter_mmap(vma);
 796 }
 797
 798 extern void perf_counter_comm(struct task_struct *tsk);
 799 extern void perf_counter_fork(struct task_struct *tsk);
 800
 801 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 802
 803 extern int sysctl_perf_counter_paranoid;
 804 extern int sysctl_perf_counter_mlock;
 805 extern int sysctl_perf_counter_sample_rate;
 806
 807 extern void perf_counter_init(void);
 808 extern void perf_tpcounter_event(int event_id, u64 addr, u64 count,
 809                                  void *record, int entry_size);
 810
 811 #ifndef perf_misc_flags
 812 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_EVENT_MISC_USER : \
 813                                  PERF_EVENT_MISC_KERNEL)
 814 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
 815 #endif
 816
 817 extern int perf_output_begin(struct perf_output_handle *handle,
 818                              struct perf_counter *counter, unsigned int size,
 819                              int nmi, int sample);
 820 extern void perf_output_end(struct perf_output_handle *handle);
 821 extern void perf_output_copy(struct perf_output_handle *handle,
 822                              const void *buf, unsigned int len);
 823 #else
 824 static inline void
 825 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
 826 static inline void
 827 perf_counter_task_sched_out(struct task_struct *task,
 828                             struct task_struct *next, int cpu)          { }
 829 static inline void
 830 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 831 static inline int perf_counter_init_task(struct task_struct *child)     { return 0; }
 832 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 833 static inline void perf_counter_free_task(struct task_struct *task)     { }
 834 static inline void perf_counter_do_pending(void)                        { }
 835 static inline void perf_counter_print_debug(void)                       { }
 836 static inline void perf_disable(void)                                   { }
 837 static inline void perf_enable(void)                                    { }
 838 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 839 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 840
 841 static inline void
 842 perf_swcounter_event(u32 event, u64 nr, int nmi,
 843                      struct pt_regs *regs, u64 addr)                    { }
 844
 845 static inline void perf_counter_mmap(struct vm_area_struct *vma)        { }
 846 static inline void perf_counter_comm(struct task_struct *tsk)           { }
 847 static inline void perf_counter_fork(struct task_struct *tsk)           { }
 848 static inline void perf_counter_init(void)                              { }
 849
 850 static inline int
 851 perf_output_begin(struct perf_output_handle *handle, struct perf_counter *c,
 852                   unsigned int size, int nmi, int sample)               { }
 853 static inline void perf_output_end(struct perf_output_handle *handle)   { }
 854 static inline void
 855 perf_output_copy(struct perf_output_handle *handle,
 856                  const void *buf, unsigned int len)                     { }
 857 static inline void
 858 perf_output_sample(struct perf_output_handle *handle,
 859                    struct perf_event_header *header,
 860                    struct perf_sample_data *data,
 861                    struct perf_counter *counter)                        { }
 862 static inline void
 863 perf_prepare_sample(struct perf_event_header *header,
 864                     struct perf_sample_data *data,
 865                     struct perf_counter *counter,
 866                     struct pt_regs *regs)                               { }
 867 #endif
 868
 869 #define perf_output_put(handle, x) \
 870         perf_output_copy((handle), &(x), sizeof(x))
 871
 872 #endif /* __KERNEL__ */
 873 #endif /* _LINUX_PERF_COUNTER_H */