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
 219                                 __reserved_1   : 46;
 220
 221         union {
 222                 __u32           wakeup_events;    /* wakeup every n events */
 223                 __u32           wakeup_watermark; /* bytes before wakeup   */
 224         };
 225
 226         __u32                   bp_type;
 227         __u64                   bp_addr;
 228         __u64                   bp_len;
 229 };
 230
 231 /*
 232  * Ioctls that can be done on a perf event fd:
 233  */
 234 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
 235 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
 236 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
 237 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
 238 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
 239 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
 240 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 241
 242 enum perf_event_ioc_flags {
 243         PERF_IOC_FLAG_GROUP             = 1U << 0,
 244 };
 245
 246 /*
 247  * Structure of the page that can be mapped via mmap
 248  */
 249 struct perf_event_mmap_page {
 250         __u32   version;                /* version number of this structure */
 251         __u32   compat_version;         /* lowest version this is compat with */
 252
 253         /*
 254          * Bits needed to read the hw events in user-space.
 255          *
 256          *   u32 seq;
 257          *   s64 count;
 258          *
 259          *   do {
 260          *     seq = pc->lock;
 261          *
 262          *     barrier()
 263          *     if (pc->index) {
 264          *       count = pmc_read(pc->index - 1);
 265          *       count += pc->offset;
 266          *     } else
 267          *       goto regular_read;
 268          *
 269          *     barrier();
 270          *   } while (pc->lock != seq);
 271          *
 272          * NOTE: for obvious reason this only works on self-monitoring
 273          *       processes.
 274          */
 275         __u32   lock;                   /* seqlock for synchronization */
 276         __u32   index;                  /* hardware event identifier */
 277         __s64   offset;                 /* add to hardware event value */
 278         __u64   time_enabled;           /* time event active */
 279         __u64   time_running;           /* time event on cpu */
 280
 281                 /*
 282                  * Hole for extension of the self monitor capabilities
 283                  */
 284
 285         __u64   __reserved[123];        /* align to 1k */
 286
 287         /*
 288          * Control data for the mmap() data buffer.
 289          *
 290          * User-space reading the @data_head value should issue an rmb(), on
 291          * SMP capable platforms, after reading this value -- see
 292          * perf_event_wakeup().
 293          *
 294          * When the mapping is PROT_WRITE the @data_tail value should be
 295          * written by userspace to reflect the last read data. In this case
 296          * the kernel will not over-write unread data.
 297          */
 298         __u64   data_head;              /* head in the data section */
 299         __u64   data_tail;              /* user-space written tail */
 300 };
 301
 302 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 303 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 304 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 305 #define PERF_RECORD_MISC_USER                   (2 << 0)
 306 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 307 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 308 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 309
 310 /*
 311  * Indicates that the content of PERF_SAMPLE_IP points to
 312  * the actual instruction that triggered the event. See also
 313  * perf_event_attr::precise_ip.
 314  */
 315 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 316 /*
 317  * Reserve the last bit to indicate some extended misc field
 318  */
 319 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 320
 321 struct perf_event_header {
 322         __u32   type;
 323         __u16   misc;
 324         __u16   size;
 325 };
 326
 327 enum perf_event_type {
 328
 329         /*
 330          * The MMAP events record the PROT_EXEC mappings so that we can
 331          * correlate userspace IPs to code. They have the following structure:
 332          *
 333          * struct {
 334          *      struct perf_event_header        header;
 335          *
 336          *      u32                             pid, tid;
 337          *      u64                             addr;
 338          *      u64                             len;
 339          *      u64                             pgoff;
 340          *      char                            filename[];
 341          * };
 342          */
 343         PERF_RECORD_MMAP                        = 1,
 344
 345         /*
 346          * struct {
 347          *      struct perf_event_header        header;
 348          *      u64                             id;
 349          *      u64                             lost;
 350          * };
 351          */
 352         PERF_RECORD_LOST                        = 2,
 353
 354         /*
 355          * struct {
 356          *      struct perf_event_header        header;
 357          *
 358          *      u32                             pid, tid;
 359          *      char                            comm[];
 360          * };
 361          */
 362         PERF_RECORD_COMM                        = 3,
 363
 364         /*
 365          * struct {
 366          *      struct perf_event_header        header;
 367          *      u32                             pid, ppid;
 368          *      u32                             tid, ptid;
 369          *      u64                             time;
 370          * };
 371          */
 372         PERF_RECORD_EXIT                        = 4,
 373
 374         /*
 375          * struct {
 376          *      struct perf_event_header        header;
 377          *      u64                             time;
 378          *      u64                             id;
 379          *      u64                             stream_id;
 380          * };
 381          */
 382         PERF_RECORD_THROTTLE                    = 5,
 383         PERF_RECORD_UNTHROTTLE                  = 6,
 384
 385         /*
 386          * struct {
 387          *      struct perf_event_header        header;
 388          *      u32                             pid, ppid;
 389          *      u32                             tid, ptid;
 390          *      u64                             time;
 391          * };
 392          */
 393         PERF_RECORD_FORK                        = 7,
 394
 395         /*
 396          * struct {
 397          *      struct perf_event_header        header;
 398          *      u32                             pid, tid;
 399          *
 400          *      struct read_format              values;
 401          * };
 402          */
 403         PERF_RECORD_READ                        = 8,
 404
 405         /*
 406          * struct {
 407          *      struct perf_event_header        header;
 408          *
 409          *      { u64                   ip;       } && PERF_SAMPLE_IP
 410          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 411          *      { u64                   time;     } && PERF_SAMPLE_TIME
 412          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 413          *      { u64                   id;       } && PERF_SAMPLE_ID
 414          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 415          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 416          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 417          *
 418          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 419          *
 420          *      { u64                   nr,
 421          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 422          *
 423          *      #
 424          *      # The RAW record below is opaque data wrt the ABI
 425          *      #
 426          *      # That is, the ABI doesn't make any promises wrt to
 427          *      # the stability of its content, it may vary depending
 428          *      # on event, hardware, kernel version and phase of
 429          *      # the moon.
 430          *      #
 431          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 432          *      #
 433          *
 434          *      { u32                   size;
 435          *        char                  data[size];}&& PERF_SAMPLE_RAW
 436          * };
 437          */
 438         PERF_RECORD_SAMPLE                      = 9,
 439
 440         PERF_RECORD_MAX,                        /* non-ABI */
 441 };
 442
 443 enum perf_callchain_context {
 444         PERF_CONTEXT_HV                 = (__u64)-32,
 445         PERF_CONTEXT_KERNEL             = (__u64)-128,
 446         PERF_CONTEXT_USER               = (__u64)-512,
 447
 448         PERF_CONTEXT_GUEST              = (__u64)-2048,
 449         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 450         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 451
 452         PERF_CONTEXT_MAX                = (__u64)-4095,
 453 };
 454
 455 #define PERF_FLAG_FD_NO_GROUP   (1U << 0)
 456 #define PERF_FLAG_FD_OUTPUT     (1U << 1)
 457
 458 #ifdef __KERNEL__
 459 /*
 460  * Kernel-internal data types and definitions:
 461  */
 462
 463 #ifdef CONFIG_PERF_EVENTS
 464 # include <asm/perf_event.h>
 465 # include <asm/local64.h>
 466 #endif
 467
 468 struct perf_guest_info_callbacks {
 469         int (*is_in_guest) (void);
 470         int (*is_user_mode) (void);
 471         unsigned long (*get_guest_ip) (void);
 472 };
 473
 474 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 475 #include <asm/hw_breakpoint.h>
 476 #endif
 477
 478 #include <linux/list.h>
 479 #include <linux/mutex.h>
 480 #include <linux/rculist.h>
 481 #include <linux/rcupdate.h>
 482 #include <linux/spinlock.h>
 483 #include <linux/hrtimer.h>
 484 #include <linux/fs.h>
 485 #include <linux/pid_namespace.h>
 486 #include <linux/workqueue.h>
 487 #include <linux/ftrace.h>
 488 #include <linux/cpu.h>
 489 #include <linux/irq_work.h>
 490 #include <asm/atomic.h>
 491 #include <asm/local.h>
 492
 493 #define PERF_MAX_STACK_DEPTH            255
 494
 495 struct perf_callchain_entry {
 496         __u64                           nr;
 497         __u64                           ip[PERF_MAX_STACK_DEPTH];
 498 };
 499
 500 struct perf_raw_record {
 501         u32                             size;
 502         void                            *data;
 503 };
 504
 505 struct perf_branch_entry {
 506         __u64                           from;
 507         __u64                           to;
 508         __u64                           flags;
 509 };
 510
 511 struct perf_branch_stack {
 512         __u64                           nr;
 513         struct perf_branch_entry        entries[0];
 514 };
 515
 516 struct task_struct;
 517
 518 /**
 519  * struct hw_perf_event - performance event hardware details:
 520  */
 521 struct hw_perf_event {
 522 #ifdef CONFIG_PERF_EVENTS
 523         union {
 524                 struct { /* hardware */
 525                         u64             config;
 526                         u64             last_tag;
 527                         unsigned long   config_base;
 528                         unsigned long   event_base;
 529                         int             idx;
 530                         int             last_cpu;
 531                 };
 532                 struct { /* software */
 533                         struct hrtimer  hrtimer;
 534                 };
 535 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 536                 struct { /* breakpoint */
 537                         struct arch_hw_breakpoint       info;
 538                         struct list_head                bp_list;
 539                 };
 540 #endif
 541         };
 542         int                             state;
 543         local64_t                       prev_count;
 544         u64                             sample_period;
 545         u64                             last_period;
 546         local64_t                       period_left;
 547         u64                             interrupts;
 548
 549         u64                             freq_time_stamp;
 550         u64                             freq_count_stamp;
 551 #endif
 552 };
 553
 554 /*
 555  * hw_perf_event::state flags
 556  */
 557 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 558 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 559 #define PERF_HES_ARCH           0x04
 560
 561 struct perf_event;
 562
 563 /*
 564  * Common implementation detail of pmu::{start,commit,cancel}_txn
 565  */
 566 #define PERF_EVENT_TXN 0x1
 567
 568 /**
 569  * struct pmu - generic performance monitoring unit
 570  */
 571 struct pmu {
 572         struct list_head                entry;
 573
 574         int * __percpu                  pmu_disable_count;
 575         struct perf_cpu_context * __percpu pmu_cpu_context;
 576         int                             task_ctx_nr;
 577
 578         /*
 579          * Fully disable/enable this PMU, can be used to protect from the PMI
 580          * as well as for lazy/batch writing of the MSRs.
 581          */
 582         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 583         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 584
 585         /*
 586          * Try and initialize the event for this PMU.
 587          * Should return -ENOENT when the @event doesn't match this PMU.
 588          */
 589         int (*event_init)               (struct perf_event *event);
 590
 591 #define PERF_EF_START   0x01            /* start the counter when adding    */
 592 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 593 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 594
 595         /*
 596          * Adds/Removes a counter to/from the PMU, can be done inside
 597          * a transaction, see the ->*_txn() methods.
 598          */
 599         int  (*add)                     (struct perf_event *event, int flags);
 600         void (*del)                     (struct perf_event *event, int flags);
 601
 602         /*
 603          * Starts/Stops a counter present on the PMU. The PMI handler
 604          * should stop the counter when perf_event_overflow() returns
 605          * !0. ->start() will be used to continue.
 606          */
 607         void (*start)                   (struct perf_event *event, int flags);
 608         void (*stop)                    (struct perf_event *event, int flags);
 609
 610         /*
 611          * Updates the counter value of the event.
 612          */
 613         void (*read)                    (struct perf_event *event);
 614
 615         /*
 616          * Group events scheduling is treated as a transaction, add
 617          * group events as a whole and perform one schedulability test.
 618          * If the test fails, roll back the whole group
 619          *
 620          * Start the transaction, after this ->add() doesn't need to
 621          * do schedulability tests.
 622          */
 623         void (*start_txn)       (struct pmu *pmu); /* optional */
 624         /*
 625          * If ->start_txn() disabled the ->add() schedulability test
 626          * then ->commit_txn() is required to perform one. On success
 627          * the transaction is closed. On error the transaction is kept
 628          * open until ->cancel_txn() is called.
 629          */
 630         int  (*commit_txn)      (struct pmu *pmu); /* optional */
 631         /*
 632          * Will cancel the transaction, assumes ->del() is called
 633          * for each successfull ->add() during the transaction.
 634          */
 635         void (*cancel_txn)      (struct pmu *pmu); /* optional */
 636 };
 637
 638 /**
 639  * enum perf_event_active_state - the states of a event
 640  */
 641 enum perf_event_active_state {
 642         PERF_EVENT_STATE_ERROR          = -2,
 643         PERF_EVENT_STATE_OFF            = -1,
 644         PERF_EVENT_STATE_INACTIVE       =  0,
 645         PERF_EVENT_STATE_ACTIVE         =  1,
 646 };
 647
 648 struct file;
 649
 650 #define PERF_BUFFER_WRITABLE            0x01
 651
 652 struct perf_buffer {
 653         atomic_t                        refcount;
 654         struct rcu_head                 rcu_head;
 655 #ifdef CONFIG_PERF_USE_VMALLOC
 656         struct work_struct              work;
 657         int                             page_order;     /* allocation order  */
 658 #endif
 659         int                             nr_pages;       /* nr of data pages  */
 660         int                             writable;       /* are we writable   */
 661
 662         atomic_t                        poll;           /* POLL_ for wakeups */
 663
 664         local_t                         head;           /* write position    */
 665         local_t                         nest;           /* nested writers    */
 666         local_t                         events;         /* event limit       */
 667         local_t                         wakeup;         /* wakeup stamp      */
 668         local_t                         lost;           /* nr records lost   */
 669
 670         long                            watermark;      /* wakeup watermark  */
 671
 672         struct perf_event_mmap_page     *user_page;
 673         void                            *data_pages[0];
 674 };
 675
 676 struct perf_sample_data;
 677
 678 typedef void (*perf_overflow_handler_t)(struct perf_event *, int,
 679                                         struct perf_sample_data *,
 680                                         struct pt_regs *regs);
 681
 682 enum perf_group_flag {
 683         PERF_GROUP_SOFTWARE = 0x1,
 684 };
 685
 686 #define SWEVENT_HLIST_BITS      8
 687 #define SWEVENT_HLIST_SIZE      (1 << SWEVENT_HLIST_BITS)
 688
 689 struct swevent_hlist {
 690         struct hlist_head       heads[SWEVENT_HLIST_SIZE];
 691         struct rcu_head         rcu_head;
 692 };
 693
 694 #define PERF_ATTACH_CONTEXT     0x01
 695 #define PERF_ATTACH_GROUP       0x02
 696
 697 /**
 698  * struct perf_event - performance event kernel representation:
 699  */
 700 struct perf_event {
 701 #ifdef CONFIG_PERF_EVENTS
 702         struct list_head                group_entry;
 703         struct list_head                event_entry;
 704         struct list_head                sibling_list;
 705         struct hlist_node               hlist_entry;
 706         int                             nr_siblings;
 707         int                             group_flags;
 708         struct perf_event               *group_leader;
 709         struct pmu                      *pmu;
 710
 711         enum perf_event_active_state    state;
 712         unsigned int                    attach_state;
 713         local64_t                       count;
 714         atomic64_t                      child_count;
 715
 716         /*
 717          * These are the total time in nanoseconds that the event
 718          * has been enabled (i.e. eligible to run, and the task has
 719          * been scheduled in, if this is a per-task event)
 720          * and running (scheduled onto the CPU), respectively.
 721          *
 722          * They are computed from tstamp_enabled, tstamp_running and
 723          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 724          */
 725         u64                             total_time_enabled;
 726         u64                             total_time_running;
 727
 728         /*
 729          * These are timestamps used for computing total_time_enabled
 730          * and total_time_running when the event is in INACTIVE or
 731          * ACTIVE state, measured in nanoseconds from an arbitrary point
 732          * in time.
 733          * tstamp_enabled: the notional time when the event was enabled
 734          * tstamp_running: the notional time when the event was scheduled on
 735          * tstamp_stopped: in INACTIVE state, the notional time when the
 736          *      event was scheduled off.
 737          */
 738         u64                             tstamp_enabled;
 739         u64                             tstamp_running;
 740         u64                             tstamp_stopped;
 741
 742         struct perf_event_attr          attr;
 743         struct hw_perf_event            hw;
 744
 745         struct perf_event_context       *ctx;
 746         struct file                     *filp;
 747
 748         /*
 749          * These accumulate total time (in nanoseconds) that children
 750          * events have been enabled and running, respectively.
 751          */
 752         atomic64_t                      child_total_time_enabled;
 753         atomic64_t                      child_total_time_running;
 754
 755         /*
 756          * Protect attach/detach and child_list:
 757          */
 758         struct mutex                    child_mutex;
 759         struct list_head                child_list;
 760         struct perf_event               *parent;
 761
 762         int                             oncpu;
 763         int                             cpu;
 764
 765         struct list_head                owner_entry;
 766         struct task_struct              *owner;
 767
 768         /* mmap bits */
 769         struct mutex                    mmap_mutex;
 770         atomic_t                        mmap_count;
 771         int                             mmap_locked;
 772         struct user_struct              *mmap_user;
 773         struct perf_buffer              *buffer;
 774
 775         /* poll related */
 776         wait_queue_head_t               waitq;
 777         struct fasync_struct            *fasync;
 778
 779         /* delayed work for NMIs and such */
 780         int                             pending_wakeup;
 781         int                             pending_kill;
 782         int                             pending_disable;
 783         struct irq_work                 pending;
 784
 785         atomic_t                        event_limit;
 786
 787         void (*destroy)(struct perf_event *);
 788         struct rcu_head                 rcu_head;
 789
 790         struct pid_namespace            *ns;
 791         u64                             id;
 792
 793         perf_overflow_handler_t         overflow_handler;
 794
 795 #ifdef CONFIG_EVENT_TRACING
 796         struct ftrace_event_call        *tp_event;
 797         struct event_filter             *filter;
 798 #endif
 799
 800 #endif /* CONFIG_PERF_EVENTS */
 801 };
 802
 803 enum perf_event_context_type {
 804         task_context,
 805         cpu_context,
 806 };
 807
 808 /**
 809  * struct perf_event_context - event context structure
 810  *
 811  * Used as a container for task events and CPU events as well:
 812  */
 813 struct perf_event_context {
 814         enum perf_event_context_type    type;
 815         struct pmu                      *pmu;
 816         /*
 817          * Protect the states of the events in the list,
 818          * nr_active, and the list:
 819          */
 820         raw_spinlock_t                  lock;
 821         /*
 822          * Protect the list of events.  Locking either mutex or lock
 823          * is sufficient to ensure the list doesn't change; to change
 824          * the list you need to lock both the mutex and the spinlock.
 825          */
 826         struct mutex                    mutex;
 827
 828         struct list_head                pinned_groups;
 829         struct list_head                flexible_groups;
 830         struct list_head                event_list;
 831         int                             nr_events;
 832         int                             nr_active;
 833         int                             is_active;
 834         int                             nr_stat;
 835         atomic_t                        refcount;
 836         struct task_struct              *task;
 837
 838         /*
 839          * Context clock, runs when context enabled.
 840          */
 841         u64                             time;
 842         u64                             timestamp;
 843
 844         /*
 845          * These fields let us detect when two contexts have both
 846          * been cloned (inherited) from a common ancestor.
 847          */
 848         struct perf_event_context       *parent_ctx;
 849         u64                             parent_gen;
 850         u64                             generation;
 851         int                             pin_count;
 852         struct rcu_head                 rcu_head;
 853 };
 854
 855 /*
 856  * Number of contexts where an event can trigger:
 857  *      task, softirq, hardirq, nmi.
 858  */
 859 #define PERF_NR_CONTEXTS        4
 860
 861 /**
 862  * struct perf_event_cpu_context - per cpu event context structure
 863  */
 864 struct perf_cpu_context {
 865         struct perf_event_context       ctx;
 866         struct perf_event_context       *task_ctx;
 867         int                             active_oncpu;
 868         int                             exclusive;
 869         struct list_head                rotation_list;
 870         int                             jiffies_interval;
 871 };
 872
 873 struct perf_output_handle {
 874         struct perf_event               *event;
 875         struct perf_buffer              *buffer;
 876         unsigned long                   wakeup;
 877         unsigned long                   size;
 878         void                            *addr;
 879         int                             page;
 880         int                             nmi;
 881         int                             sample;
 882 };
 883
 884 #ifdef CONFIG_PERF_EVENTS
 885
 886 extern int perf_pmu_register(struct pmu *pmu);
 887 extern void perf_pmu_unregister(struct pmu *pmu);
 888
 889 extern int perf_num_counters(void);
 890 extern const char *perf_pmu_name(void);
 891 extern void perf_event_task_sched_in(struct task_struct *task);
 892 extern void perf_event_task_sched_out(struct task_struct *task, struct task_struct *next);
 893 extern int perf_event_init_task(struct task_struct *child);
 894 extern void perf_event_exit_task(struct task_struct *child);
 895 extern void perf_event_free_task(struct task_struct *task);
 896 extern void perf_event_delayed_put(struct task_struct *task);
 897 extern void perf_event_print_debug(void);
 898 extern void perf_pmu_disable(struct pmu *pmu);
 899 extern void perf_pmu_enable(struct pmu *pmu);
 900 extern int perf_event_task_disable(void);
 901 extern int perf_event_task_enable(void);
 902 extern void perf_event_update_userpage(struct perf_event *event);
 903 extern int perf_event_release_kernel(struct perf_event *event);
 904 extern struct perf_event *
 905 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 906                                 int cpu,
 907                                 struct task_struct *task,
 908                                 perf_overflow_handler_t callback);
 909 extern u64 perf_event_read_value(struct perf_event *event,
 910                                  u64 *enabled, u64 *running);
 911
 912 struct perf_sample_data {
 913         u64                             type;
 914
 915         u64                             ip;
 916         struct {
 917                 u32     pid;
 918                 u32     tid;
 919         }                               tid_entry;
 920         u64                             time;
 921         u64                             addr;
 922         u64                             id;
 923         u64                             stream_id;
 924         struct {
 925                 u32     cpu;
 926                 u32     reserved;
 927         }                               cpu_entry;
 928         u64                             period;
 929         struct perf_callchain_entry     *callchain;
 930         struct perf_raw_record          *raw;
 931 };
 932
 933 static inline
 934 void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
 935 {
 936         data->addr = addr;
 937         data->raw  = NULL;
 938 }
 939
 940 extern void perf_output_sample(struct perf_output_handle *handle,
 941                                struct perf_event_header *header,
 942                                struct perf_sample_data *data,
 943                                struct perf_event *event);
 944 extern void perf_prepare_sample(struct perf_event_header *header,
 945                                 struct perf_sample_data *data,
 946                                 struct perf_event *event,
 947                                 struct pt_regs *regs);
 948
 949 extern int perf_event_overflow(struct perf_event *event, int nmi,
 950                                  struct perf_sample_data *data,
 951                                  struct pt_regs *regs);
 952
 953 /*
 954  * Return 1 for a software event, 0 for a hardware event
 955  */
 956 static inline int is_software_event(struct perf_event *event)
 957 {
 958         return event->pmu->task_ctx_nr == perf_sw_context;
 959 }
 960
 961 extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
 962
 963 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
 964
 965 #ifndef perf_arch_fetch_caller_regs
 966 static inline void
 967 perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
 968 #endif
 969
 970 /*
 971  * Take a snapshot of the regs. Skip ip and frame pointer to
 972  * the nth caller. We only need a few of the regs:
 973  * - ip for PERF_SAMPLE_IP
 974  * - cs for user_mode() tests
 975  * - bp for callchains
 976  * - eflags, for future purposes, just in case
 977  */
 978 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
 979 {
 980         memset(regs, 0, sizeof(*regs));
 981
 982         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
 983 }
 984
 985 static inline void
 986 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 987 {
 988         if (atomic_read(&perf_swevent_enabled[event_id])) {
 989                 struct pt_regs hot_regs;
 990
 991                 if (!regs) {
 992                         perf_fetch_caller_regs(&hot_regs);
 993                         regs = &hot_regs;
 994                 }
 995                 __perf_sw_event(event_id, nr, nmi, regs, addr);
 996         }
 997 }
 998
 999 extern void perf_event_mmap(struct vm_area_struct *vma);
1000 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1001 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1002 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1003
1004 extern void perf_event_comm(struct task_struct *tsk);
1005 extern void perf_event_fork(struct task_struct *tsk);
1006
1007 /* Callchains */
1008 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1009
1010 extern void perf_callchain_user(struct perf_callchain_entry *entry,
1011                                 struct pt_regs *regs);
1012 extern void perf_callchain_kernel(struct perf_callchain_entry *entry,
1013                                   struct pt_regs *regs);
1014
1015
1016 static inline void
1017 perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1018 {
1019         if (entry->nr < PERF_MAX_STACK_DEPTH)
1020                 entry->ip[entry->nr++] = ip;
1021 }
1022
1023 extern int sysctl_perf_event_paranoid;
1024 extern int sysctl_perf_event_mlock;
1025 extern int sysctl_perf_event_sample_rate;
1026
1027 static inline bool perf_paranoid_tracepoint_raw(void)
1028 {
1029         return sysctl_perf_event_paranoid > -1;
1030 }
1031
1032 static inline bool perf_paranoid_cpu(void)
1033 {
1034         return sysctl_perf_event_paranoid > 0;
1035 }
1036
1037 static inline bool perf_paranoid_kernel(void)
1038 {
1039         return sysctl_perf_event_paranoid > 1;
1040 }
1041
1042 extern void perf_event_init(void);
1043 extern void perf_tp_event(u64 addr, u64 count, void *record,
1044                           int entry_size, struct pt_regs *regs,
1045                           struct hlist_head *head, int rctx);
1046 extern void perf_bp_event(struct perf_event *event, void *data);
1047
1048 #ifndef perf_misc_flags
1049 #define perf_misc_flags(regs)   (user_mode(regs) ? PERF_RECORD_MISC_USER : \
1050                                  PERF_RECORD_MISC_KERNEL)
1051 #define perf_instruction_pointer(regs)  instruction_pointer(regs)
1052 #endif
1053
1054 extern int perf_output_begin(struct perf_output_handle *handle,
1055                              struct perf_event *event, unsigned int size,
1056                              int nmi, int sample);
1057 extern void perf_output_end(struct perf_output_handle *handle);
1058 extern void perf_output_copy(struct perf_output_handle *handle,
1059                              const void *buf, unsigned int len);
1060 extern int perf_swevent_get_recursion_context(void);
1061 extern void perf_swevent_put_recursion_context(int rctx);
1062 extern void perf_event_enable(struct perf_event *event);
1063 extern void perf_event_disable(struct perf_event *event);
1064 extern void perf_event_task_tick(void);
1065 #else
1066 static inline void
1067 perf_event_task_sched_in(struct task_struct *task)                      { }
1068 static inline void
1069 perf_event_task_sched_out(struct task_struct *task,
1070                             struct task_struct *next)                   { }
1071 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1072 static inline void perf_event_exit_task(struct task_struct *child)      { }
1073 static inline void perf_event_free_task(struct task_struct *task)       { }
1074 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1075 static inline void perf_event_print_debug(void)                         { }
1076 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1077 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1078
1079 static inline void
1080 perf_sw_event(u32 event_id, u64 nr, int nmi,
1081                      struct pt_regs *regs, u64 addr)                    { }
1082 static inline void
1083 perf_bp_event(struct perf_event *event, void *data)                     { }
1084
1085 static inline int perf_register_guest_info_callbacks
1086 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1087 static inline int perf_unregister_guest_info_callbacks
1088 (struct perf_guest_info_callbacks *callbacks) { return 0; }
1089
1090 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1091 static inline void perf_event_comm(struct task_struct *tsk)             { }
1092 static inline void perf_event_fork(struct task_struct *tsk)             { }
1093 static inline void perf_event_init(void)                                { }
1094 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1095 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1096 static inline void perf_event_enable(struct perf_event *event)          { }
1097 static inline void perf_event_disable(struct perf_event *event)         { }
1098 static inline void perf_event_task_tick(void)                           { }
1099 #endif
1100
1101 #define perf_output_put(handle, x) \
1102         perf_output_copy((handle), &(x), sizeof(x))
1103
1104 /*
1105  * This has to have a higher priority than migration_notifier in sched.c.
1106  */
1107 #define perf_cpu_notifier(fn)                                   \
1108 do {                                                            \
1109         static struct notifier_block fn##_nb __cpuinitdata =    \
1110                 { .notifier_call = fn, .priority = CPU_PRI_PERF }; \
1111         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
1112                 (void *)(unsigned long)smp_processor_id());     \
1113         fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
1114                 (void *)(unsigned long)smp_processor_id());     \
1115         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
1116                 (void *)(unsigned long)smp_processor_id());     \
1117         register_cpu_notifier(&fn##_nb);                        \
1118 } while (0)
1119
1120 #endif /* __KERNEL__ */
1121 #endif /* _LINUX_PERF_EVENT_H */