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                                 /*
 207                                  * precise_ip:
 208                                  *
 209                                  *  0 - SAMPLE_IP can have arbitrary skid
 210                                  *  1 - SAMPLE_IP must have constant skid
 211                                  *  2 - SAMPLE_IP requested to have 0 skid
 212                                  *  3 - SAMPLE_IP must have 0 skid
 213                                  *
 214                                  *  See also PERF_RECORD_MISC_EXACT_IP
 215                                  */
 216                                 precise_ip     :  2, /* skid constraint       */
 217                                 mmap_data      :  1, /* non-exec mmap data    */
 218                                 sample_id_all  :  1, /* sample_type all events */
 219
 220                                 __reserved_1   : 45;
 221
 222         union {
 223                 __u32           wakeup_events;    /* wakeup every n events */
 224                 __u32           wakeup_watermark; /* bytes before wakeup   */
 225         };
 226
 227         __u32                   bp_type;
 228         __u64                   bp_addr;
 229         __u64                   bp_len;
 230 };
 231
 232 /*
 233  * Ioctls that can be done on a perf event fd:
 234  */
 235 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
 236 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
 237 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
 238 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
 239 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
 240 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
 241 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 242
 243 enum perf_event_ioc_flags {
 244         PERF_IOC_FLAG_GROUP             = 1U << 0,
 245 };
 246
 247 /*
 248  * Structure of the page that can be mapped via mmap
 249  */
 250 struct perf_event_mmap_page {
 251         __u32   version;                /* version number of this structure */
 252         __u32   compat_version;         /* lowest version this is compat with */
 253
 254         /*
 255          * Bits needed to read the hw events in user-space.
 256          *
 257          *   u32 seq;
 258          *   s64 count;
 259          *
 260          *   do {
 261          *     seq = pc->lock;
 262          *
 263          *     barrier()
 264          *     if (pc->index) {
 265          *       count = pmc_read(pc->index - 1);
 266          *       count += pc->offset;
 267          *     } else
 268          *       goto regular_read;
 269          *
 270          *     barrier();
 271          *   } while (pc->lock != seq);
 272          *
 273          * NOTE: for obvious reason this only works on self-monitoring
 274          *       processes.
 275          */
 276         __u32   lock;                   /* seqlock for synchronization */
 277         __u32   index;                  /* hardware event identifier */
 278         __s64   offset;                 /* add to hardware event value */
 279         __u64   time_enabled;           /* time event active */
 280         __u64   time_running;           /* time event on cpu */
 281
 282                 /*
 283                  * Hole for extension of the self monitor capabilities
 284                  */
 285
 286         __u64   __reserved[123];        /* align to 1k */
 287
 288         /*
 289          * Control data for the mmap() data buffer.
 290          *
 291          * User-space reading the @data_head value should issue an rmb(), on
 292          * SMP capable platforms, after reading this value -- see
 293          * perf_event_wakeup().
 294          *
 295          * When the mapping is PROT_WRITE the @data_tail value should be
 296          * written by userspace to reflect the last read data. In this case
 297          * the kernel will not over-write unread data.
 298          */
 299         __u64   data_head;              /* head in the data section */
 300         __u64   data_tail;              /* user-space written tail */
 301 };
 302
 303 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 304 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 305 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 306 #define PERF_RECORD_MISC_USER                   (2 << 0)
 307 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 308 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 309 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 310
 311 /*
 312  * Indicates that the content of PERF_SAMPLE_IP points to
 313  * the actual instruction that triggered the event. See also
 314  * perf_event_attr::precise_ip.
 315  */
 316 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 317 /*
 318  * Reserve the last bit to indicate some extended misc field
 319  */
 320 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 321
 322 struct perf_event_header {
 323         __u32   type;
 324         __u16   misc;
 325         __u16   size;
 326 };
 327
 328 enum perf_event_type {
 329
 330         /*
 331          * If perf_event_attr.sample_id_all is set then all event types will
 332          * have the sample_type selected fields related to where/when
 333          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
 334          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
 335          * the perf_event_header and the fields already present for the existing
 336          * fields, i.e. at the end of the payload. That way a newer perf.data
 337          * file will be supported by older perf tools, with these new optional
 338          * fields being ignored.
 339          *
 340          * The MMAP events record the PROT_EXEC mappings so that we can
 341          * correlate userspace IPs to code. They have the following structure:
 342          *
 343          * struct {
 344          *      struct perf_event_header        header;
 345          *
 346          *      u32                             pid, tid;
 347          *      u64                             addr;
 348          *      u64                             len;
 349          *      u64                             pgoff;
 350          *      char                            filename[];
 351          * };
 352          */
 353         PERF_RECORD_MMAP                        = 1,
 354
 355         /*
 356          * struct {
 357          *      struct perf_event_header        header;
 358          *      u64                             id;
 359          *      u64                             lost;
 360          * };
 361          */
 362         PERF_RECORD_LOST                        = 2,
 363
 364         /*
 365          * struct {
 366          *      struct perf_event_header        header;
 367          *
 368          *      u32                             pid, tid;
 369          *      char                            comm[];
 370          * };
 371          */
 372         PERF_RECORD_COMM                        = 3,
 373
 374         /*
 375          * struct {
 376          *      struct perf_event_header        header;
 377          *      u32                             pid, ppid;
 378          *      u32                             tid, ptid;
 379          *      u64                             time;
 380          * };
 381          */
 382         PERF_RECORD_EXIT                        = 4,
 383
 384         /*
 385          * struct {
 386          *      struct perf_event_header        header;
 387          *      u64                             time;
 388          *      u64                             id;
 389          *      u64                             stream_id;
 390          * };
 391          */
 392         PERF_RECORD_THROTTLE                    = 5,
 393         PERF_RECORD_UNTHROTTLE                  = 6,
 394
 395         /*
 396          * struct {
 397          *      struct perf_event_header        header;
 398          *      u32                             pid, ppid;
 399          *      u32                             tid, ptid;
 400          *      u64                             time;
 401          * };
 402          */
 403         PERF_RECORD_FORK                        = 7,
 404
 405         /*
 406          * struct {
 407          *      struct perf_event_header        header;
 408          *      u32                             pid, tid;
 409          *
 410          *      struct read_format              values;
 411          * };
 412          */
 413         PERF_RECORD_READ                        = 8,
 414
 415         /*
 416          * struct {
 417          *      struct perf_event_header        header;
 418          *
 419          *      { u64                   ip;       } && PERF_SAMPLE_IP
 420          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 421          *      { u64                   time;     } && PERF_SAMPLE_TIME
 422          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 423          *      { u64                   id;       } && PERF_SAMPLE_ID
 424          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 425          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 426          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 427          *
 428          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 429          *
 430          *      { u64                   nr,
 431          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 432          *
 433          *      #
 434          *      # The RAW record below is opaque data wrt the ABI
 435          *      #
 436          *      # That is, the ABI doesn't make any promises wrt to
 437          *      # the stability of its content, it may vary depending
 438          *      # on event, hardware, kernel version and phase of
 439          *      # the moon.
 440          *      #
 441          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 442          *      #
 443          *
 444          *      { u32                   size;
 445          *        char                  data[size];}&& PERF_SAMPLE_RAW
 446          * };
 447          */
 448         PERF_RECORD_SAMPLE                      = 9,
 449
 450         PERF_RECORD_MAX,                        /* non-ABI */
 451 };
 452
 453 enum perf_callchain_context {
 454         PERF_CONTEXT_HV                 = (__u64)-32,
 455         PERF_CONTEXT_KERNEL             = (__u64)-128,
 456         PERF_CONTEXT_USER               = (__u64)-512,
 457
 458         PERF_CONTEXT_GUEST              = (__u64)-2048,
 459         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 460         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 461
 462         PERF_CONTEXT_MAX                = (__u64)-4095,
 463 };
 464
 465 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 466 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 467 #define PERF_FLAG_PID_CGROUP    (1U << 2) /* pid=cgroup id, per-cpu mode only */
 468
 469 #ifdef __KERNEL__
 470 /*
 471  * Kernel-internal data types and definitions:
 472  */
 473
 474 #ifdef CONFIG_PERF_EVENTS
 475 # include <linux/cgroup.h>
 476 # include <asm/perf_event.h>
 477 # include <asm/local64.h>
 478 #endif
 479
 480 struct perf_guest_info_callbacks {
 481         int (*is_in_guest) (void);
 482         int (*is_user_mode) (void);
 483         unsigned long (*get_guest_ip) (void);
 484 };
 485
 486 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 487 #include <asm/hw_breakpoint.h>
 488 #endif
 489
 490 #include <linux/list.h>
 491 #include <linux/mutex.h>
 492 #include <linux/rculist.h>
 493 #include <linux/rcupdate.h>
 494 #include <linux/spinlock.h>
 495 #include <linux/hrtimer.h>
 496 #include <linux/fs.h>
 497 #include <linux/pid_namespace.h>
 498 #include <linux/workqueue.h>
 499 #include <linux/ftrace.h>
 500 #include <linux/cpu.h>
 501 #include <linux/irq_work.h>
 502 #include <linux/jump_label_ref.h>
 503 #include <asm/atomic.h>
 504 #include <asm/local.h>
 505
 506 #define PERF_MAX_STACK_DEPTH            255
 507
 508 struct perf_callchain_entry {
 509         __u64                           nr;
 510         __u64                           ip[PERF_MAX_STACK_DEPTH];
 511 };
 512
 513 struct perf_raw_record {
 514         u32                             size;
 515         void                            *data;
 516 };
 517
 518 struct perf_branch_entry {
 519         __u64                           from;
 520         __u64                           to;
 521         __u64                           flags;
 522 };
 523
 524 struct perf_branch_stack {
 525         __u64                           nr;
 526         struct perf_branch_entry        entries[0];
 527 };
 528
 529 struct task_struct;
 530
 531 /**
 532  * struct hw_perf_event - performance event hardware details:
 533  */
 534 struct hw_perf_event {
 535 #ifdef CONFIG_PERF_EVENTS
 536         union {
 537                 struct { /* hardware */
 538                         u64             config;
 539                         u64             last_tag;
 540                         unsigned long   config_base;
 541                         unsigned long   event_base;
 542                         int             idx;
 543                         int             last_cpu;
 544                 };
 545                 struct { /* software */
 546                         struct hrtimer  hrtimer;
 547                 };
 548 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 549                 struct { /* breakpoint */
 550                         struct arch_hw_breakpoint       info;
 551                         struct list_head                bp_list;
 552                         /*
 553                          * Crufty hack to avoid the chicken and egg
 554                          * problem hw_breakpoint has with context
 555                          * creation and event initalization.
 556                          */
 557                         struct task_struct              *bp_target;
 558                 };
 559 #endif
 560         };
 561         int                             state;
 562         local64_t                       prev_count;
 563         u64                             sample_period;
 564         u64                             last_period;
 565         local64_t                       period_left;
 566         u64                             interrupts;
 567
 568         u64                             freq_time_stamp;
 569         u64                             freq_count_stamp;
 570 #endif
 571 };
 572
 573 /*
 574  * hw_perf_event::state flags
 575  */
 576 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 577 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 578 #define PERF_HES_ARCH           0x04
 579
 580 struct perf_event;
 581
 582 /*
 583  * Common implementation detail of pmu::{start,commit,cancel}_txn
 584  */
 585 #define PERF_EVENT_TXN 0x1
 586
 587 /**
 588  * struct pmu - generic performance monitoring unit
 589  */
 590 struct pmu {
 591         struct list_head                entry;
 592
 593         struct device                   *dev;
 594         char                            *name;
 595         int                             type;
 596
 597         int * __percpu                  pmu_disable_count;
 598         struct perf_cpu_context * __percpu pmu_cpu_context;
 599         int                             task_ctx_nr;
 600
 601         /*
 602          * Fully disable/enable this PMU, can be used to protect from the PMI
 603          * as well as for lazy/batch writing of the MSRs.
 604          */
 605         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 606         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 607
 608         /*
 609          * Try and initialize the event for this PMU.
 610          * Should return -ENOENT when the @event doesn't match this PMU.
 611          */
 612         int (*event_init)               (struct perf_event *event);
 613
 614 #define PERF_EF_START   0x01            /* start the counter when adding    */
 615 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 616 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 617
 618         /*
 619          * Adds/Removes a counter to/from the PMU, can be done inside
 620          * a transaction, see the ->*_txn() methods.
 621          */
 622         int  (*add)                     (struct perf_event *event, int flags);
 623         void (*del)                     (struct perf_event *event, int flags);
 624
 625         /*
 626          * Starts/Stops a counter present on the PMU. The PMI handler
 627          * should stop the counter when perf_event_overflow() returns
 628          * !0. ->start() will be used to continue.
 629          */
 630         void (*start)                   (struct perf_event *event, int flags);
 631         void (*stop)                    (struct perf_event *event, int flags);
 632
 633         /*
 634          * Updates the counter value of the event.
 635          */
 636         void (*read)                    (struct perf_event *event);
 637
 638         /*
 639          * Group events scheduling is treated as a transaction, add
 640          * group events as a whole and perform one schedulability test.
 641          * If the test fails, roll back the whole group
 642          *
 643          * Start the transaction, after this ->add() doesn't need to
 644          * do schedulability tests.
 645          */
 646         void (*start_txn)       (struct pmu *pmu); /* optional */
 647         /*
 648          * If ->start_txn() disabled the ->add() schedulability test
 649          * then ->commit_txn() is required to perform one. On success
 650          * the transaction is closed. On error the transaction is kept
 651          * open until ->cancel_txn() is called.
 652          */
 653         int  (*commit_txn)      (struct pmu *pmu); /* optional */
 654         /*
 655          * Will cancel the transaction, assumes ->del() is called
 656          * for each successfull ->add() during the transaction.
 657          */
 658         void (*cancel_txn)      (struct pmu *pmu); /* optional */
 659 };
 660
 661 /**
 662  * enum perf_event_active_state - the states of a event
 663  */
 664 enum perf_event_active_state {
 665         PERF_EVENT_STATE_ERROR          = -2,
 666         PERF_EVENT_STATE_OFF            = -1,
 667         PERF_EVENT_STATE_INACTIVE       =  0,
 668         PERF_EVENT_STATE_ACTIVE         =  1,
 669 };
 670
 671 struct file;
 672
 673 #define PERF_BUFFER_WRITABLE            0x01
 674
 675 struct perf_buffer {
 676         atomic_t                        refcount;
 677         struct rcu_head                 rcu_head;
 678 #ifdef CONFIG_PERF_USE_VMALLOC
 679         struct work_struct              work;
 680         int                             page_order;     /* allocation order  */
 681 #endif
 682         int                             nr_pages;       /* nr of data pages  */
 683         int                             writable;       /* are we writable   */
 684
 685         atomic_t                        poll;           /* POLL_ for wakeups */
 686
 687         local_t                         head;           /* write position    */
 688         local_t                         nest;           /* nested writers    */
 689         local_t                         events;         /* event limit       */
 690         local_t                         wakeup;         /* wakeup stamp      */
 691         local_t                         lost;           /* nr records lost   */
 692
 693         long                            watermark;      /* wakeup watermark  */
 694
 695         struct perf_event_mmap_page     *user_page;
 696         void                            *data_pages[0];
 697 };
 698
 699 struct perf_sample_data;
 700
 701 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
 702                                         struct perf_sample_data *,
 703                                         struct pt_regs *regs);
 704
 705 enum perf_group_flag {
 706         PERF_GROUP_SOFTWARE = 0x1,
 707 };
 708
 709 #define SWEVENT_HLIST_BITS      8
 710 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
 711
 712 struct swevent_hlist {
 713         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
 714         struct rcu_head         rcu_head;
 715 };
 716
 717 #define PERF_ATTACH_CONTEXT     0x01
 718 #define PERF_ATTACH_GROUP       0x02
 719 #define PERF_ATTACH_TASK        0x04
 720
 721 #ifdef CONFIG_CGROUP_PERF
 722 /*
 723  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 724  * This is a per-cpu dynamically allocated data structure.
 725  */
 726 struct perf_cgroup_info {
 727         u64 time;
 728         u64 timestamp;
 729 };
 730
 731 struct perf_cgroup {
 732         struct cgroup_subsys_state css;
 733         struct perf_cgroup_info *info;  /* timing info, one per cpu */
 734 };
 735 #endif
 736
 737 /**
 738  * struct perf_event - performance event kernel representation:
 739  */
 740 struct perf_event {
 741 #ifdef CONFIG_PERF_EVENTS
 742         struct list_head                group_entry;
 743         struct list_head                event_entry;
 744         struct list_head                sibling_list;
 745         struct hlist_node               hlist_entry;
 746         int                             nr_siblings;
 747         int                             group_flags;
 748         struct perf_event               *group_leader;
 749         struct pmu                      *pmu;
 750
 751         enum perf_event_active_state    state;
 752         unsigned int                    attach_state;
 753         local64_t                       count;
 754         atomic64_t                      child_count;
 755
 756         /*
 757          * These are the total time in nanoseconds that the event
 758          * has been enabled (i.e. eligible to run, and the task has
 759          * been scheduled in, if this is a per-task event)
 760          * and running (scheduled onto the CPU), respectively.
 761          *
 762          * They are computed from tstamp_enabled, tstamp_running and
 763          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 764          */
 765         u64                             total_time_enabled;
 766         u64                             total_time_running;
 767
 768         /*
 769          * These are timestamps used for computing total_time_enabled
 770          * and total_time_running when the event is in INACTIVE or
 771          * ACTIVE state, measured in nanoseconds from an arbitrary point
 772          * in time.
 773          * tstamp_enabled: the notional time when the event was enabled
 774          * tstamp_running: the notional time when the event was scheduled on
 775          * tstamp_stopped: in INACTIVE state, the notional time when the
 776          *      event was scheduled off.
 777          */
 778         u64                             tstamp_enabled;
 779         u64                             tstamp_running;
 780         u64                             tstamp_stopped;
 781
 782         /*
 783          * timestamp shadows the actual context timing but it can
 784          * be safely used in NMI interrupt context. It reflects the
 785          * context time as it was when the event was last scheduled in.
 786          *
 787          * ctx_time already accounts for ctx->timestamp. Therefore to
 788          * compute ctx_time for a sample, simply add perf_clock().
 789          */
 790         u64                             shadow_ctx_time;
 791
 792         struct perf_event_attr          attr;
 793         u16                             header_size;
 794         u16                             id_header_size;
 795         u16                             read_size;
 796         struct hw_perf_event            hw;
 797
 798         struct perf_event_context       *ctx;
 799         struct file                     *filp;
 800
 801         /*
 802          * These accumulate total time (in nanoseconds) that children
 803          * events have been enabled and running, respectively.
 804          */
 805         atomic64_t                      child_total_time_enabled;
 806         atomic64_t                      child_total_time_running;
 807
 808         /*
 809          * Protect attach/detach and child_list:
 810          */
 811         struct mutex                    child_mutex;
 812         struct list_head                child_list;
 813         struct perf_event               *parent;
 814
 815         int                             oncpu;
 816         int                             cpu;
 817
 818         struct list_head                owner_entry;
 819         struct task_struct              *owner;
 820
 821         /* mmap bits */
 822         struct mutex                    mmap_mutex;
 823         atomic_t                        mmap_count;
 824         int                             mmap_locked;
 825         struct user_struct              *mmap_user;
 826         struct perf_buffer              *buffer;
 827
 828         /* poll related */
 829         wait_queue_head_t               waitq;
 830         struct fasync_struct            *fasync;
 831
 832         /* delayed work for NMIs and such */
 833         int                             pending_wakeup;
 834         int                             pending_kill;
 835         int                             pending_disable;
 836         struct irq_work                 pending;
 837
 838         atomic_t                        event_limit;
 839
 840         void (*destroy)(struct perf_event *);
 841         struct rcu_head                 rcu_head;
 842
 843         struct pid_namespace            *ns;
 844         u64                             id;
 845
 846         perf_overflow_handler_t         overflow_handler;
 847
 848 #ifdef CONFIG_EVENT_TRACING
 849         struct ftrace_event_call        *tp_event;
 850         struct event_filter             *filter;
 851 #endif
 852
 853 #ifdef CONFIG_CGROUP_PERF
 854         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 855         int                             cgrp_defer_enabled;
 856 #endif
 857
 858 #endif /* CONFIG_PERF_EVENTS */
 859 };
 860
 861 enum perf_event_context_type {
 862         task_context,
 863         cpu_context,
 864 };
 865
 866 /**
 867  * struct perf_event_context - event context structure
 868  *
 869  * Used as a container for task events and CPU events as well:
 870  */
 871 struct perf_event_context {
 872         enum perf_event_context_type    type;
 873         struct pmu                      *pmu;
 874         /*
 875          * Protect the states of the events in the list,
 876          * nr_active, and the list:
 877          */
 878         raw_spinlock_t                  lock;
 879         /*
 880          * Protect the list of events.  Locking either mutex or lock
 881          * is sufficient to ensure the list doesn't change; to change
 882          * the list you need to lock both the mutex and the spinlock.
 883          */
 884         struct mutex                    mutex;
 885
 886         struct list_head                pinned_groups;
 887         struct list_head                flexible_groups;
 888         struct list_head                event_list;
 889         int                             nr_events;
 890         int                             nr_active;
 891         int                             is_active;
 892         int                             nr_stat;
 893         int                             rotate_disable;
 894         atomic_t                        refcount;
 895         struct task_struct              *task;
 896
 897         /*
 898          * Context clock, runs when context enabled.
 899          */
 900         u64                             time;
 901         u64                             timestamp;
 902
 903         /*
 904          * These fields let us detect when two contexts have both
 905          * been cloned (inherited) from a common ancestor.
 906          */
 907         struct perf_event_context       *parent_ctx;
 908         u64                             parent_gen;
 909         u64                             generation;
 910         int                             pin_count;
 911         struct rcu_head                 rcu_head;
 912         int                             nr_cgroups; /* cgroup events present */
 913 };
 914
 915 /*
 916  * Number of contexts where an event can trigger:
 917  *      task, softirq, hardirq, nmi.
 918  */
 919 #define PERF_NR_CONTEXTS        4
 920
 921 /**
 922  * struct perf_event_cpu_context - per cpu event context structure
 923  */
 924 struct perf_cpu_context {
 925         struct perf_event_context       ctx;
 926         struct perf_event_context       *task_ctx;
 927         int                             active_oncpu;
 928         int                             exclusive;
 929         struct list_head                rotation_list;
 930         int                             jiffies_interval;
 931         struct pmu                      *active_pmu;
 932 #ifdef CONFIG_CGROUP_PERF
 933         struct perf_cgroup              *cgrp;
 934 #endif
 935 };
 936
 937 struct perf_output_handle {
 938         struct perf_event               *event;
 939         struct perf_buffer              *buffer;
 940         unsigned long                   wakeup;
 941         unsigned long                   size;
 942         void                            *addr;
 943         int                             page;
 944         int                             nmi;
 945         int                             sample;
 946 };
 947
 948 #ifdef CONFIG_PERF_EVENTS
 949
 950 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
 951 extern void perf_pmu_unregister(struct pmu *pmu);
 952
 953 extern int perf_num_counters(void);
 954 extern const char *perf_pmu_name(void);
 955 extern void __perf_event_task_sched_in(struct task_struct *task);
 956 extern void __perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
 957 extern int perf_event_init_task(struct task_struct *child);
 958 extern void perf_event_exit_task(struct task_struct *child);
 959 extern void perf_event_free_task(struct task_struct *task);
 960 extern void perf_event_delayed_put(struct task_struct *task);
 961 extern void perf_event_print_debug(void);
 962 extern void perf_pmu_disable(struct pmu *pmu);
 963 extern void perf_pmu_enable(struct pmu *pmu);
 964 extern int perf_event_task_disable(void);
 965 extern int perf_event_task_enable(void);
 966 extern void perf_event_update_userpage(struct perf_event *event);
 967 extern int perf_event_release_kernel(struct perf_event *event);
 968 extern struct perf_event *
 969 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 970                                 int cpu,
 971                                 struct task_struct *task,
 972                                 perf_overflow_handler_t callback);
 973 extern u64 perf_event_read_value(struct perf_event *event,
 974                                  u64 *enabled, u64 *running);
 975
 976 struct perf_sample_data {
 977         u64                             type;
 978
 979         u64                             ip;
 980         struct {
 981                 u32     pid;
 982                 u32     tid;
 983         }                               tid_entry;
 984         u64                             time;
 985         u64                             addr;
 986         u64                             id;
 987         u64                             stream_id;
 988         struct {
 989                 u32     cpu;
 990                 u32     reserved;
 991         }                               cpu_entry;
 992         u64                             period;
 993         struct perf_callchain_entry     *callchain;
 994         struct perf_raw_record          *raw;
 995 };
 996
 997 static inline
 998 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
 999 {
1000         data->addr = addr;
1001         data->raw  = NULL;
1002 }
1003
1004 extern void perf_output_sample(struct perf_output_handle *handle,
1005                                struct perf_event_header *header,
1006                                struct perf_sample_data *data,
1007                                struct perf_event *event);
1008 extern void perf_prepare_sample(struct perf_event_header *header,
1009                                 struct perf_sample_data *data,
1010                                 struct perf_event *event,
1011                                 struct pt_regs *regs);
1012
1013 extern int perf_event_overflow(struct perf_event *event, int nmi,
1014                                  struct perf_sample_data *data,
1015                                  struct pt_regs *regs);
1016
1017 static inline bool is_sampling_event(struct perf_event *event)
1018 {
1019         return event->attr.sample_period != 0;
1020 }
1021
1022 /*
1023  * Return 1 for a software event, 0 for a hardware event
1024  */
1025 static inline int is_software_event(struct perf_event *event)
1026 {
1027         return event->pmu->task_ctx_nr == perf_sw_context;
1028 }
1029
1030 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
1031
1032 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
1033
1034 #ifndef perf_arch_fetch_caller_regs
1035 static inline void
1036 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1037 #endif
1038
1039 /*
1040  * Take a snapshot of the regs. Skip ip and frame pointer to
1041  * the nth caller. We only need a few of the regs:
1042  * - ip for PERF_SAMPLE_IP
1043  * - cs for user_mode() tests
1044  * - bp for callchains
1045  * - eflags, for future purposes, just in case
1046  */
1047 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1048 {
1049         memset(regs, 0, sizeof(*regs));
1050
1051         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1052 }
1053
1054 static __always_inline void
1055 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
1056 {
1057         struct pt_regs hot_regs;
1058
1059         JUMP_LABEL(&perf_swevent_enabled[event_id], have_event);
1060         return;
1061
1062 have_event:
1063         if (!regs) {
1064                 perf_fetch_caller_regs(&hot_regs);
1065                 regs = &hot_regs;
1066         }
1067         __perf_sw_event(event_id, nr, nmi, regs, addr);
1068 }
1069
1070 extern atomic_t perf_sched_events;
1071
1072 static inline void perf_event_task_sched_in(struct task_struct *task)
1073 {
1074         COND_STMT(&perf_sched_events, __perf_event_task_sched_in(task));
1075 }
1076
1077 static inline
1078 void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next)
1079 {
1080         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
1081
1082         COND_STMT(&perf_sched_events, __perf_event_task_sched_out(task, next));
1083 }
1084
1085 extern void perf_event_mmap(struct vm_area_struct *vma);
1086 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1087 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1088 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1089
1090 extern void perf_event_comm(struct task_struct *tsk);
1091 extern void perf_event_fork(struct task_struct *tsk);
1092
1093 /* Callchains */
1094 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1095
1096 extern void perf_callchain_user(struct perf_callchain_entry *entry,
1097                                 struct pt_regs *regs);
1098 extern void perf_callchain_kernel(struct perf_callchain_entry *entry,
1099                                   struct pt_regs *regs);
1100
1101
1102 static inline void
1103 perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1104 {
1105         if (entry->nr < PERF_MAX_STACK_DEPTH)
1106                 entry->ip[entry->nr++] = ip;
1107 }
1108
1109 extern int sysctl_perf_event_paranoid;
1110 extern int sysctl_perf_event_mlock;
1111 extern int sysctl_perf_event_sample_rate;
1112
1113 static inline bool perf_paranoid_tracepoint_raw(void)
1114 {
1115         return sysctl_perf_event_paranoid > -1;
1116 }
1117
1118 static inline bool perf_paranoid_cpu(void)
1119 {
1120         return sysctl_perf_event_paranoid > 0;
1121 }
1122
1123 static inline bool perf_paranoid_kernel(void)
1124 {
1125         return sysctl_perf_event_paranoid > 1;
1126 }
1127
1128 extern void perf_event_init(void);
1129 extern void perf_tp_event(u64 addr, u64 count, void *record,
1130                           int entry_size, struct pt_regs *regs,
1131                           struct hlist_head *head, int rctx);
1132 extern void perf_bp_event(struct perf_event *event, void *data);
1133
1134 #ifndef perf_misc_flags
1135 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1136                                  PERF_RECORD_MISC_KERNEL)
1137 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
1138 #endif
1139
1140 extern int perf_output_begin(struct perf_output_handle *handle,
1141                              struct perf_event *event, unsigned int size,
1142                              int nmi, int sample);
1143 extern void perf_output_end(struct perf_output_handle *handle);
1144 extern void perf_output_copy(struct perf_output_handle *handle,
1145                              const void *buf, unsigned int len);
1146 extern int perf_swevent_get_recursion_context(void);
1147 extern void perf_swevent_put_recursion_context(int rctx);
1148 extern void perf_event_enable(struct perf_event *event);
1149 extern void perf_event_disable(struct perf_event *event);
1150 extern void perf_event_task_tick(void);
1151 #else
1152 static inline void
1153 perf_event_task_sched_in(struct task_struct *task)                      { }
1154 static inline void
1155 perf_event_task_sched_out(struct task_struct *task,
1156                             struct task_struct *next)                   { }
1157 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1158 static inline void perf_event_exit_task(struct task_struct *child)      { }
1159 static inline void perf_event_free_task(struct task_struct *task)       { }
1160 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1161 static inline void perf_event_print_debug(void)                         { }
1162 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1163 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1164
1165 static inline void
1166 perf_sw_event(u32 event_id, u64 nr, int nmi,
1167                      struct pt_regs *regs, u64 addr)                    { }
1168 static inline void
1169 perf_bp_event(struct perf_event *event, void *data)                     { }
1170
1171 static inline int perf_register_guest_info_callbacks
1172 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1173 static inline int perf_unregister_guest_info_callbacks
1174 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1175
1176 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1177 static inline void perf_event_comm(struct task_struct *tsk)             { }
1178 static inline void perf_event_fork(struct task_struct *tsk)             { }
1179 static inline void perf_event_init(void)                                { }
1180 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1181 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1182 static inline void perf_event_enable(struct perf_event *event)          { }
1183 static inline void perf_event_disable(struct perf_event *event)         { }
1184 static inline void perf_event_task_tick(void)                           { }
1185 #endif
1186
1187 #define perf_output_put(handle, x) \
1188         perf_output_copy((handle), &(x), sizeof(x))
1189
1190 /*
1191  * This has to have a higher priority than migration_notifier in sched.c.
1192  */
1193 #define perf_cpu_notifier(fn)                                   \
1194 do {                                                            \
1195         static struct notifier_block fn##_nb __cpuinitdata =    \
1196                 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1197         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1198                 (void *)(unsigned long)smp_processor_id());     \
1199         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1200                 (void *)(unsigned long)smp_processor_id());     \
1201         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1202                 (void *)(unsigned long)smp_processor_id());     \
1203         register_cpu_notifier(&fn##_nb);                        \
1204 } while (0)
1205
1206 #endif /* __KERNEL__ */
1207 #endif /* _LINUX_PERF_EVENT_H */