include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, 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         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
  56         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
  57         PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
  58
  59         PERF_COUNT_HW_MAX,                      /* non-ABI */
  60 };
  61
  62 /*
  63  * Generalized hardware cache events:
  64  *
  65  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
  66  *       { read, write, prefetch } x
  67  *       { accesses, misses }
  68  */
  69 enum perf_hw_cache_id {
  70         PERF_COUNT_HW_CACHE_L1D                 = 0,
  71         PERF_COUNT_HW_CACHE_L1I                 = 1,
  72         PERF_COUNT_HW_CACHE_LL                  = 2,
  73         PERF_COUNT_HW_CACHE_DTLB                = 3,
  74         PERF_COUNT_HW_CACHE_ITLB                = 4,
  75         PERF_COUNT_HW_CACHE_BPU                 = 5,
  76         PERF_COUNT_HW_CACHE_NODE                = 6,
  77
  78         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
  79 };
  80
  81 enum perf_hw_cache_op_id {
  82         PERF_COUNT_HW_CACHE_OP_READ             = 0,
  83         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
  84         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
  85
  86         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
  87 };
  88
  89 enum perf_hw_cache_op_result_id {
  90         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
  91         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
  92
  93         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
  94 };
  95
  96 /*
  97  * Special "software" events provided by the kernel, even if the hardware
  98  * does not support performance events. These events measure various
  99  * physical and sw events of the kernel (and allow the profiling of them as
 100  * well):
 101  */
 102 enum perf_sw_ids {
 103         PERF_COUNT_SW_CPU_CLOCK                 = 0,
 104         PERF_COUNT_SW_TASK_CLOCK                = 1,
 105         PERF_COUNT_SW_PAGE_FAULTS               = 2,
 106         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
 107         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
 108         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
 109         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
 110         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
 111         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
 112
 113         PERF_COUNT_SW_MAX,                      /* non-ABI */
 114 };
 115
 116 /*
 117  * Bits that can be set in attr.sample_type to request information
 118  * in the overflow packets.
 119  */
 120 enum perf_event_sample_format {
 121         PERF_SAMPLE_IP                          = 1U << 0,
 122         PERF_SAMPLE_TID                         = 1U << 1,
 123         PERF_SAMPLE_TIME                        = 1U << 2,
 124         PERF_SAMPLE_ADDR                        = 1U << 3,
 125         PERF_SAMPLE_READ                        = 1U << 4,
 126         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
 127         PERF_SAMPLE_ID                          = 1U << 6,
 128         PERF_SAMPLE_CPU                         = 1U << 7,
 129         PERF_SAMPLE_PERIOD                      = 1U << 8,
 130         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
 131         PERF_SAMPLE_RAW                         = 1U << 10,
 132         PERF_SAMPLE_BRANCH_STACK                = 1U << 11,
 133
 134         PERF_SAMPLE_MAX = 1U << 12,             /* non-ABI */
 135 };
 136
 137 /*
 138  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
 139  *
 140  * If the user does not pass priv level information via branch_sample_type,
 141  * the kernel uses the event's priv level. Branch and event priv levels do
 142  * not have to match. Branch priv level is checked for permissions.
 143  *
 144  * The branch types can be combined, however BRANCH_ANY covers all types
 145  * of branches and therefore it supersedes all the other types.
 146  */
 147 enum perf_branch_sample_type {
 148         PERF_SAMPLE_BRANCH_USER         = 1U << 0, /* user branches */
 149         PERF_SAMPLE_BRANCH_KERNEL       = 1U << 1, /* kernel branches */
 150         PERF_SAMPLE_BRANCH_HV           = 1U << 2, /* hypervisor branches */
 151
 152         PERF_SAMPLE_BRANCH_ANY          = 1U << 3, /* any branch types */
 153         PERF_SAMPLE_BRANCH_ANY_CALL     = 1U << 4, /* any call branch */
 154         PERF_SAMPLE_BRANCH_ANY_RETURN   = 1U << 5, /* any return branch */
 155         PERF_SAMPLE_BRANCH_IND_CALL     = 1U << 6, /* indirect calls */
 156
 157         PERF_SAMPLE_BRANCH_MAX          = 1U << 7, /* non-ABI */
 158 };
 159
 160 #define PERF_SAMPLE_BRANCH_PLM_ALL \
 161         (PERF_SAMPLE_BRANCH_USER|\
 162          PERF_SAMPLE_BRANCH_KERNEL|\
 163          PERF_SAMPLE_BRANCH_HV)
 164
 165 /*
 166  * The format of the data returned by read() on a perf event fd,
 167  * as specified by attr.read_format:
 168  *
 169  * struct read_format {
 170  *      { u64           value;
 171  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 172  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 173  *        { u64         id;           } && PERF_FORMAT_ID
 174  *      } && !PERF_FORMAT_GROUP
 175  *
 176  *      { u64           nr;
 177  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 178  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 179  *        { u64         value;
 180  *          { u64       id;           } && PERF_FORMAT_ID
 181  *        }             cntr[nr];
 182  *      } && PERF_FORMAT_GROUP
 183  * };
 184  */
 185 enum perf_event_read_format {
 186         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
 187         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
 188         PERF_FORMAT_ID                          = 1U << 2,
 189         PERF_FORMAT_GROUP                       = 1U << 3,
 190
 191         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
 192 };
 193
 194 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
 195 #define PERF_ATTR_SIZE_VER1     72      /* add: config2 */
 196 #define PERF_ATTR_SIZE_VER2     80      /* add: branch_sample_type */
 197
 198 /*
 199  * Hardware event_id to monitor via a performance monitoring event:
 200  */
 201 struct perf_event_attr {
 202
 203         /*
 204          * Major type: hardware/software/tracepoint/etc.
 205          */
 206         __u32                   type;
 207
 208         /*
 209          * Size of the attr structure, for fwd/bwd compat.
 210          */
 211         __u32                   size;
 212
 213         /*
 214          * Type specific configuration information.
 215          */
 216         __u64                   config;
 217
 218         union {
 219                 __u64           sample_period;
 220                 __u64           sample_freq;
 221         };
 222
 223         __u64                   sample_type;
 224         __u64                   read_format;
 225
 226         __u64                   disabled       :  1, /* off by default        */
 227                                 inherit        :  1, /* children inherit it   */
 228                                 pinned         :  1, /* must always be on PMU */
 229                                 exclusive      :  1, /* only group on PMU     */
 230                                 exclude_user   :  1, /* don't count user      */
 231                                 exclude_kernel :  1, /* ditto kernel          */
 232                                 exclude_hv     :  1, /* ditto hypervisor      */
 233                                 exclude_idle   :  1, /* don't count when idle */
 234                                 mmap           :  1, /* include mmap data     */
 235                                 comm           :  1, /* include comm data     */
 236                                 freq           :  1, /* use freq, not period  */
 237                                 inherit_stat   :  1, /* per task counts       */
 238                                 enable_on_exec :  1, /* next exec enables     */
 239                                 task           :  1, /* trace fork/exit       */
 240                                 watermark      :  1, /* wakeup_watermark      */
 241                                 /*
 242                                  * precise_ip:
 243                                  *
 244                                  *  0 - SAMPLE_IP can have arbitrary skid
 245                                  *  1 - SAMPLE_IP must have constant skid
 246                                  *  2 - SAMPLE_IP requested to have 0 skid
 247                                  *  3 - SAMPLE_IP must have 0 skid
 248                                  *
 249                                  *  See also PERF_RECORD_MISC_EXACT_IP
 250                                  */
 251                                 precise_ip     :  2, /* skid constraint       */
 252                                 mmap_data      :  1, /* non-exec mmap data    */
 253                                 sample_id_all  :  1, /* sample_type all events */
 254
 255                                 exclude_host   :  1, /* don't count in host   */
 256                                 exclude_guest  :  1, /* don't count in guest  */
 257
 258                                 __reserved_1   : 43;
 259
 260         union {
 261                 __u32           wakeup_events;    /* wakeup every n events */
 262                 __u32           wakeup_watermark; /* bytes before wakeup   */
 263         };
 264
 265         __u32                   bp_type;
 266         union {
 267                 __u64           bp_addr;
 268                 __u64           config1; /* extension of config */
 269         };
 270         union {
 271                 __u64           bp_len;
 272                 __u64           config2; /* extension of config1 */
 273         };
 274         __u64   branch_sample_type; /* enum branch_sample_type */
 275 };
 276
 277 /*
 278  * Ioctls that can be done on a perf event fd:
 279  */
 280 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
 281 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
 282 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
 283 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
 284 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
 285 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
 286 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 287
 288 enum perf_event_ioc_flags {
 289         PERF_IOC_FLAG_GROUP             = 1U << 0,
 290 };
 291
 292 /*
 293  * Structure of the page that can be mapped via mmap
 294  */
 295 struct perf_event_mmap_page {
 296         __u32   version;                /* version number of this structure */
 297         __u32   compat_version;         /* lowest version this is compat with */
 298
 299         /*
 300          * Bits needed to read the hw events in user-space.
 301          *
 302          *   u32 seq;
 303          *   s64 count;
 304          *
 305          *   do {
 306          *     seq = pc->lock;
 307          *
 308          *     barrier()
 309          *     if (pc->index) {
 310          *       count = pmc_read(pc->index - 1);
 311          *       count += pc->offset;
 312          *     } else
 313          *       goto regular_read;
 314          *
 315          *     barrier();
 316          *   } while (pc->lock != seq);
 317          *
 318          * NOTE: for obvious reason this only works on self-monitoring
 319          *       processes.
 320          */
 321         __u32   lock;                   /* seqlock for synchronization */
 322         __u32   index;                  /* hardware event identifier */
 323         __s64   offset;                 /* add to hardware event value */
 324         __u64   time_enabled;           /* time event active */
 325         __u64   time_running;           /* time event on cpu */
 326         __u32   time_mult, time_shift;
 327         __u64   time_offset;
 328
 329                 /*
 330                  * Hole for extension of the self monitor capabilities
 331                  */
 332
 333         __u64   __reserved[121];        /* align to 1k */
 334
 335         /*
 336          * Control data for the mmap() data buffer.
 337          *
 338          * User-space reading the @data_head value should issue an rmb(), on
 339          * SMP capable platforms, after reading this value -- see
 340          * perf_event_wakeup().
 341          *
 342          * When the mapping is PROT_WRITE the @data_tail value should be
 343          * written by userspace to reflect the last read data. In this case
 344          * the kernel will not over-write unread data.
 345          */
 346         __u64   data_head;              /* head in the data section */
 347         __u64   data_tail;              /* user-space written tail */
 348 };
 349
 350 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 351 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 352 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 353 #define PERF_RECORD_MISC_USER                   (2 << 0)
 354 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 355 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 356 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 357
 358 /*
 359  * Indicates that the content of PERF_SAMPLE_IP points to
 360  * the actual instruction that triggered the event. See also
 361  * perf_event_attr::precise_ip.
 362  */
 363 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 364 /*
 365  * Reserve the last bit to indicate some extended misc field
 366  */
 367 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 368
 369 struct perf_event_header {
 370         __u32   type;
 371         __u16   misc;
 372         __u16   size;
 373 };
 374
 375 enum perf_event_type {
 376
 377         /*
 378          * If perf_event_attr.sample_id_all is set then all event types will
 379          * have the sample_type selected fields related to where/when
 380          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
 381          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
 382          * the perf_event_header and the fields already present for the existing
 383          * fields, i.e. at the end of the payload. That way a newer perf.data
 384          * file will be supported by older perf tools, with these new optional
 385          * fields being ignored.
 386          *
 387          * The MMAP events record the PROT_EXEC mappings so that we can
 388          * correlate userspace IPs to code. They have the following structure:
 389          *
 390          * struct {
 391          *      struct perf_event_header        header;
 392          *
 393          *      u32                             pid, tid;
 394          *      u64                             addr;
 395          *      u64                             len;
 396          *      u64                             pgoff;
 397          *      char                            filename[];
 398          * };
 399          */
 400         PERF_RECORD_MMAP                        = 1,
 401
 402         /*
 403          * struct {
 404          *      struct perf_event_header        header;
 405          *      u64                             id;
 406          *      u64                             lost;
 407          * };
 408          */
 409         PERF_RECORD_LOST                        = 2,
 410
 411         /*
 412          * struct {
 413          *      struct perf_event_header        header;
 414          *
 415          *      u32                             pid, tid;
 416          *      char                            comm[];
 417          * };
 418          */
 419         PERF_RECORD_COMM                        = 3,
 420
 421         /*
 422          * struct {
 423          *      struct perf_event_header        header;
 424          *      u32                             pid, ppid;
 425          *      u32                             tid, ptid;
 426          *      u64                             time;
 427          * };
 428          */
 429         PERF_RECORD_EXIT                        = 4,
 430
 431         /*
 432          * struct {
 433          *      struct perf_event_header        header;
 434          *      u64                             time;
 435          *      u64                             id;
 436          *      u64                             stream_id;
 437          * };
 438          */
 439         PERF_RECORD_THROTTLE                    = 5,
 440         PERF_RECORD_UNTHROTTLE                  = 6,
 441
 442         /*
 443          * struct {
 444          *      struct perf_event_header        header;
 445          *      u32                             pid, ppid;
 446          *      u32                             tid, ptid;
 447          *      u64                             time;
 448          * };
 449          */
 450         PERF_RECORD_FORK                        = 7,
 451
 452         /*
 453          * struct {
 454          *      struct perf_event_header        header;
 455          *      u32                             pid, tid;
 456          *
 457          *      struct read_format              values;
 458          * };
 459          */
 460         PERF_RECORD_READ                        = 8,
 461
 462         /*
 463          * struct {
 464          *      struct perf_event_header        header;
 465          *
 466          *      { u64                   ip;       } && PERF_SAMPLE_IP
 467          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 468          *      { u64                   time;     } && PERF_SAMPLE_TIME
 469          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 470          *      { u64                   id;       } && PERF_SAMPLE_ID
 471          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 472          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 473          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 474          *
 475          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 476          *
 477          *      { u64                   nr,
 478          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 479          *
 480          *      #
 481          *      # The RAW record below is opaque data wrt the ABI
 482          *      #
 483          *      # That is, the ABI doesn't make any promises wrt to
 484          *      # the stability of its content, it may vary depending
 485          *      # on event, hardware, kernel version and phase of
 486          *      # the moon.
 487          *      #
 488          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 489          *      #
 490          *
 491          *      { u32                   size;
 492          *        char                  data[size];}&& PERF_SAMPLE_RAW
 493          *
 494          *      { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
 495          * };
 496          */
 497         PERF_RECORD_SAMPLE                      = 9,
 498
 499         PERF_RECORD_MAX,                        /* non-ABI */
 500 };
 501
 502 enum perf_callchain_context {
 503         PERF_CONTEXT_HV                 = (__u64)-32,
 504         PERF_CONTEXT_KERNEL             = (__u64)-128,
 505         PERF_CONTEXT_USER               = (__u64)-512,
 506
 507         PERF_CONTEXT_GUEST              = (__u64)-2048,
 508         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 509         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 510
 511         PERF_CONTEXT_MAX                = (__u64)-4095,
 512 };
 513
 514 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 515 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 516 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
 517
 518 #ifdef __KERNEL__
 519 /*
 520  * Kernel-internal data types and definitions:
 521  */
 522
 523 #ifdef CONFIG_PERF_EVENTS
 524 # include <linux/cgroup.h>
 525 # include <asm/perf_event.h>
 526 # include <asm/local64.h>
 527 #endif
 528
 529 struct perf_guest_info_callbacks {
 530         int                             (*is_in_guest)(void);
 531         int                             (*is_user_mode)(void);
 532         unsigned long                   (*get_guest_ip)(void);
 533 };
 534
 535 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 536 #include <asm/hw_breakpoint.h>
 537 #endif
 538
 539 #include <linux/list.h>
 540 #include <linux/mutex.h>
 541 #include <linux/rculist.h>
 542 #include <linux/rcupdate.h>
 543 #include <linux/spinlock.h>
 544 #include <linux/hrtimer.h>
 545 #include <linux/fs.h>
 546 #include <linux/pid_namespace.h>
 547 #include <linux/workqueue.h>
 548 #include <linux/ftrace.h>
 549 #include <linux/cpu.h>
 550 #include <linux/irq_work.h>
 551 #include <linux/static_key.h>
 552 #include <linux/atomic.h>
 553 #include <asm/local.h>
 554
 555 #define PERF_MAX_STACK_DEPTH            255
 556
 557 struct perf_callchain_entry {
 558         __u64                           nr;
 559         __u64                           ip[PERF_MAX_STACK_DEPTH];
 560 };
 561
 562 struct perf_raw_record {
 563         u32                             size;
 564         void                            *data;
 565 };
 566
 567 /*
 568  * single taken branch record layout:
 569  *
 570  *      from: source instruction (may not always be a branch insn)
 571  *        to: branch target
 572  *   mispred: branch target was mispredicted
 573  * predicted: branch target was predicted
 574  *
 575  * support for mispred, predicted is optional. In case it
 576  * is not supported mispred = predicted = 0.
 577  */
 578 struct perf_branch_entry {
 579         __u64   from;
 580         __u64   to;
 581         __u64   mispred:1,  /* target mispredicted */
 582                 predicted:1,/* target predicted */
 583                 reserved:62;
 584 };
 585
 586 /*
 587  * branch stack layout:
 588  *  nr: number of taken branches stored in entries[]
 589  *
 590  * Note that nr can vary from sample to sample
 591  * branches (to, from) are stored from most recent
 592  * to least recent, i.e., entries[0] contains the most
 593  * recent branch.
 594  */
 595 struct perf_branch_stack {
 596         __u64                           nr;
 597         struct perf_branch_entry        entries[0];
 598 };
 599
 600 struct task_struct;
 601
 602 /*
 603  * extra PMU register associated with an event
 604  */
 605 struct hw_perf_event_extra {
 606         u64             config; /* register value */
 607         unsigned int    reg;    /* register address or index */
 608         int             alloc;  /* extra register already allocated */
 609         int             idx;    /* index in shared_regs->regs[] */
 610 };
 611
 612 /**
 613  * struct hw_perf_event - performance event hardware details:
 614  */
 615 struct hw_perf_event {
 616 #ifdef CONFIG_PERF_EVENTS
 617         union {
 618                 struct { /* hardware */
 619                         u64             config;
 620                         u64             last_tag;
 621                         unsigned long   config_base;
 622                         unsigned long   event_base;
 623                         int             idx;
 624                         int             last_cpu;
 625
 626                         struct hw_perf_event_extra extra_reg;
 627                         struct hw_perf_event_extra branch_reg;
 628                 };
 629                 struct { /* software */
 630                         struct hrtimer  hrtimer;
 631                 };
 632 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 633                 struct { /* breakpoint */
 634                         struct arch_hw_breakpoint       info;
 635                         struct list_head                bp_list;
 636                         /*
 637                          * Crufty hack to avoid the chicken and egg
 638                          * problem hw_breakpoint has with context
 639                          * creation and event initalization.
 640                          */
 641                         struct task_struct              *bp_target;
 642                 };
 643 #endif
 644         };
 645         int                             state;
 646         local64_t                       prev_count;
 647         u64                             sample_period;
 648         u64                             last_period;
 649         local64_t                       period_left;
 650         u64                             interrupts_seq;
 651         u64                             interrupts;
 652
 653         u64                             freq_time_stamp;
 654         u64                             freq_count_stamp;
 655 #endif
 656 };
 657
 658 /*
 659  * hw_perf_event::state flags
 660  */
 661 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 662 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 663 #define PERF_HES_ARCH           0x04
 664
 665 struct perf_event;
 666
 667 /*
 668  * Common implementation detail of pmu::{start,commit,cancel}_txn
 669  */
 670 #define PERF_EVENT_TXN 0x1
 671
 672 /**
 673  * struct pmu - generic performance monitoring unit
 674  */
 675 struct pmu {
 676         struct list_head                entry;
 677
 678         struct device                   *dev;
 679         const struct attribute_group    **attr_groups;
 680         char                            *name;
 681         int                             type;
 682
 683         int * __percpu                  pmu_disable_count;
 684         struct perf_cpu_context * __percpu pmu_cpu_context;
 685         int                             task_ctx_nr;
 686
 687         /*
 688          * Fully disable/enable this PMU, can be used to protect from the PMI
 689          * as well as for lazy/batch writing of the MSRs.
 690          */
 691         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 692         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 693
 694         /*
 695          * Try and initialize the event for this PMU.
 696          * Should return -ENOENT when the @event doesn't match this PMU.
 697          */
 698         int (*event_init)               (struct perf_event *event);
 699
 700 #define PERF_EF_START   0x01            /* start the counter when adding    */
 701 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 702 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 703
 704         /*
 705          * Adds/Removes a counter to/from the PMU, can be done inside
 706          * a transaction, see the ->*_txn() methods.
 707          */
 708         int  (*add)                     (struct perf_event *event, int flags);
 709         void (*del)                     (struct perf_event *event, int flags);
 710
 711         /*
 712          * Starts/Stops a counter present on the PMU. The PMI handler
 713          * should stop the counter when perf_event_overflow() returns
 714          * !0. ->start() will be used to continue.
 715          */
 716         void (*start)                   (struct perf_event *event, int flags);
 717         void (*stop)                    (struct perf_event *event, int flags);
 718
 719         /*
 720          * Updates the counter value of the event.
 721          */
 722         void (*read)                    (struct perf_event *event);
 723
 724         /*
 725          * Group events scheduling is treated as a transaction, add
 726          * group events as a whole and perform one schedulability test.
 727          * If the test fails, roll back the whole group
 728          *
 729          * Start the transaction, after this ->add() doesn't need to
 730          * do schedulability tests.
 731          */
 732         void (*start_txn)               (struct pmu *pmu); /* optional */
 733         /*
 734          * If ->start_txn() disabled the ->add() schedulability test
 735          * then ->commit_txn() is required to perform one. On success
 736          * the transaction is closed. On error the transaction is kept
 737          * open until ->cancel_txn() is called.
 738          */
 739         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 740         /*
 741          * Will cancel the transaction, assumes ->del() is called
 742          * for each successful ->add() during the transaction.
 743          */
 744         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 745
 746         /*
 747          * Will return the value for perf_event_mmap_page::index for this event,
 748          * if no implementation is provided it will default to: event->hw.idx + 1.
 749          */
 750         int (*event_idx)                (struct perf_event *event); /*optional */
 751
 752         /*
 753          * flush branch stack on context-switches (needed in cpu-wide mode)
 754          */
 755         void (*flush_branch_stack)      (void);
 756 };
 757
 758 /**
 759  * enum perf_event_active_state - the states of a event
 760  */
 761 enum perf_event_active_state {
 762         PERF_EVENT_STATE_ERROR          = -2,
 763         PERF_EVENT_STATE_OFF            = -1,
 764         PERF_EVENT_STATE_INACTIVE       =  0,
 765         PERF_EVENT_STATE_ACTIVE         =  1,
 766 };
 767
 768 struct file;
 769 struct perf_sample_data;
 770
 771 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 772                                         struct perf_sample_data *,
 773                                         struct pt_regs *regs);
 774
 775 enum perf_group_flag {
 776         PERF_GROUP_SOFTWARE             = 0x1,
 777 };
 778
 779 #define SWEVENT_HLIST_BITS              8
 780 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 781
 782 struct swevent_hlist {
 783         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 784         struct rcu_head                 rcu_head;
 785 };
 786
 787 #define PERF_ATTACH_CONTEXT     0x01
 788 #define PERF_ATTACH_GROUP       0x02
 789 #define PERF_ATTACH_TASK        0x04
 790
 791 #ifdef CONFIG_CGROUP_PERF
 792 /*
 793  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 794  * This is a per-cpu dynamically allocated data structure.
 795  */
 796 struct perf_cgroup_info {
 797         u64                             time;
 798         u64                             timestamp;
 799 };
 800
 801 struct perf_cgroup {
 802         struct                          cgroup_subsys_state css;
 803         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
 804 };
 805 #endif
 806
 807 struct ring_buffer;
 808
 809 /**
 810  * struct perf_event - performance event kernel representation:
 811  */
 812 struct perf_event {
 813 #ifdef CONFIG_PERF_EVENTS
 814         struct list_head                group_entry;
 815         struct list_head                event_entry;
 816         struct list_head                sibling_list;
 817         struct hlist_node               hlist_entry;
 818         int                             nr_siblings;
 819         int                             group_flags;
 820         struct perf_event               *group_leader;
 821         struct pmu                      *pmu;
 822
 823         enum perf_event_active_state    state;
 824         unsigned int                    attach_state;
 825         local64_t                       count;
 826         atomic64_t                      child_count;
 827
 828         /*
 829          * These are the total time in nanoseconds that the event
 830          * has been enabled (i.e. eligible to run, and the task has
 831          * been scheduled in, if this is a per-task event)
 832          * and running (scheduled onto the CPU), respectively.
 833          *
 834          * They are computed from tstamp_enabled, tstamp_running and
 835          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 836          */
 837         u64                             total_time_enabled;
 838         u64                             total_time_running;
 839
 840         /*
 841          * These are timestamps used for computing total_time_enabled
 842          * and total_time_running when the event is in INACTIVE or
 843          * ACTIVE state, measured in nanoseconds from an arbitrary point
 844          * in time.
 845          * tstamp_enabled: the notional time when the event was enabled
 846          * tstamp_running: the notional time when the event was scheduled on
 847          * tstamp_stopped: in INACTIVE state, the notional time when the
 848          *      event was scheduled off.
 849          */
 850         u64                             tstamp_enabled;
 851         u64                             tstamp_running;
 852         u64                             tstamp_stopped;
 853
 854         /*
 855          * timestamp shadows the actual context timing but it can
 856          * be safely used in NMI interrupt context. It reflects the
 857          * context time as it was when the event was last scheduled in.
 858          *
 859          * ctx_time already accounts for ctx->timestamp. Therefore to
 860          * compute ctx_time for a sample, simply add perf_clock().
 861          */
 862         u64                             shadow_ctx_time;
 863
 864         struct perf_event_attr          attr;
 865         u16                             header_size;
 866         u16                             id_header_size;
 867         u16                             read_size;
 868         struct hw_perf_event            hw;
 869
 870         struct perf_event_context       *ctx;
 871         struct file                     *filp;
 872
 873         /*
 874          * These accumulate total time (in nanoseconds) that children
 875          * events have been enabled and running, respectively.
 876          */
 877         atomic64_t                      child_total_time_enabled;
 878         atomic64_t                      child_total_time_running;
 879
 880         /*
 881          * Protect attach/detach and child_list:
 882          */
 883         struct mutex                    child_mutex;
 884         struct list_head                child_list;
 885         struct perf_event               *parent;
 886
 887         int                             oncpu;
 888         int                             cpu;
 889
 890         struct list_head                owner_entry;
 891         struct task_struct              *owner;
 892
 893         /* mmap bits */
 894         struct mutex                    mmap_mutex;
 895         atomic_t                        mmap_count;
 896         int                             mmap_locked;
 897         struct user_struct              *mmap_user;
 898         struct ring_buffer              *rb;
 899         struct list_head                rb_entry;
 900
 901         /* poll related */
 902         wait_queue_head_t               waitq;
 903         struct fasync_struct            *fasync;
 904
 905         /* delayed work for NMIs and such */
 906         int                             pending_wakeup;
 907         int                             pending_kill;
 908         int                             pending_disable;
 909         struct irq_work                 pending;
 910
 911         atomic_t                        event_limit;
 912
 913         void (*destroy)(struct perf_event *);
 914         struct rcu_head                 rcu_head;
 915
 916         struct pid_namespace            *ns;
 917         u64                             id;
 918
 919         perf_overflow_handler_t         overflow_handler;
 920         void                            *overflow_handler_context;
 921
 922 #ifdef CONFIG_EVENT_TRACING
 923         struct ftrace_event_call        *tp_event;
 924         struct event_filter             *filter;
 925 #ifdef CONFIG_FUNCTION_TRACER
 926         struct ftrace_ops               ftrace_ops;
 927 #endif
 928 #endif
 929
 930 #ifdef CONFIG_CGROUP_PERF
 931         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 932         int                             cgrp_defer_enabled;
 933 #endif
 934
 935 #endif /* CONFIG_PERF_EVENTS */
 936 };
 937
 938 enum perf_event_context_type {
 939         task_context,
 940         cpu_context,
 941 };
 942
 943 /**
 944  * struct perf_event_context - event context structure
 945  *
 946  * Used as a container for task events and CPU events as well:
 947  */
 948 struct perf_event_context {
 949         struct pmu                      *pmu;
 950         enum perf_event_context_type    type;
 951         /*
 952          * Protect the states of the events in the list,
 953          * nr_active, and the list:
 954          */
 955         raw_spinlock_t                  lock;
 956         /*
 957          * Protect the list of events.  Locking either mutex or lock
 958          * is sufficient to ensure the list doesn't change; to change
 959          * the list you need to lock both the mutex and the spinlock.
 960          */
 961         struct mutex                    mutex;
 962
 963         struct list_head                pinned_groups;
 964         struct list_head                flexible_groups;
 965         struct list_head                event_list;
 966         int                             nr_events;
 967         int                             nr_active;
 968         int                             is_active;
 969         int                             nr_stat;
 970         int                             nr_freq;
 971         int                             rotate_disable;
 972         atomic_t                        refcount;
 973         struct task_struct              *task;
 974
 975         /*
 976          * Context clock, runs when context enabled.
 977          */
 978         u64                             time;
 979         u64                             timestamp;
 980
 981         /*
 982          * These fields let us detect when two contexts have both
 983          * been cloned (inherited) from a common ancestor.
 984          */
 985         struct perf_event_context       *parent_ctx;
 986         u64                             parent_gen;
 987         u64                             generation;
 988         int                             pin_count;
 989         int                             nr_cgroups;      /* cgroup evts */
 990         int                             nr_branch_stack; /* branch_stack evt */
 991         struct rcu_head                 rcu_head;
 992 };
 993
 994 /*
 995  * Number of contexts where an event can trigger:
 996  *      task, softirq, hardirq, nmi.
 997  */
 998 #define PERF_NR_CONTEXTS        4
 999
1000 /**
1001  * struct perf_event_cpu_context - per cpu event context structure
1002  */
1003 struct perf_cpu_context {
1004         struct perf_event_context       ctx;
1005         struct perf_event_context       *task_ctx;
1006         int                             active_oncpu;
1007         int                             exclusive;
1008         struct list_head                rotation_list;
1009         int                             jiffies_interval;
1010         struct pmu                      *active_pmu;
1011         struct perf_cgroup              *cgrp;
1012 };
1013
1014 struct perf_output_handle {
1015         struct perf_event               *event;
1016         struct ring_buffer              *rb;
1017         unsigned long                   wakeup;
1018         unsigned long                   size;
1019         void                            *addr;
1020         int                             page;
1021 };
1022
1023 #ifdef CONFIG_PERF_EVENTS
1024
1025 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1026 extern void perf_pmu_unregister(struct pmu *pmu);
1027
1028 extern int perf_num_counters(void);
1029 extern const char *perf_pmu_name(void);
1030 extern void __perf_event_task_sched_in(struct task_struct *prev,
1031                                        struct task_struct *task);
1032 extern void __perf_event_task_sched_out(struct task_struct *prev,
1033                                         struct task_struct *next);
1034 extern int perf_event_init_task(struct task_struct *child);
1035 extern void perf_event_exit_task(struct task_struct *child);
1036 extern void perf_event_free_task(struct task_struct *task);
1037 extern void perf_event_delayed_put(struct task_struct *task);
1038 extern void perf_event_print_debug(void);
1039 extern void perf_pmu_disable(struct pmu *pmu);
1040 extern void perf_pmu_enable(struct pmu *pmu);
1041 extern int perf_event_task_disable(void);
1042 extern int perf_event_task_enable(void);
1043 extern int perf_event_refresh(struct perf_event *event, int refresh);
1044 extern void perf_event_update_userpage(struct perf_event *event);
1045 extern int perf_event_release_kernel(struct perf_event *event);
1046 extern struct perf_event *
1047 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1048                                 int cpu,
1049                                 struct task_struct *task,
1050                                 perf_overflow_handler_t callback,
1051                                 void *context);
1052 extern u64 perf_event_read_value(struct perf_event *event,
1053                                  u64 *enabled, u64 *running);
1054
1055
1056 struct perf_sample_data {
1057         u64                             type;
1058
1059         u64                             ip;
1060         struct {
1061                 u32     pid;
1062                 u32     tid;
1063         }                               tid_entry;
1064         u64                             time;
1065         u64                             addr;
1066         u64                             id;
1067         u64                             stream_id;
1068         struct {
1069                 u32     cpu;
1070                 u32     reserved;
1071         }                               cpu_entry;
1072         u64                             period;
1073         struct perf_callchain_entry     *callchain;
1074         struct perf_raw_record          *raw;
1075         struct perf_branch_stack        *br_stack;
1076 };
1077
1078 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
1079 {
1080         data->addr = addr;
1081         data->raw  = NULL;
1082         data->br_stack = NULL;
1083 }
1084
1085 extern void perf_output_sample(struct perf_output_handle *handle,
1086                                struct perf_event_header *header,
1087                                struct perf_sample_data *data,
1088                                struct perf_event *event);
1089 extern void perf_prepare_sample(struct perf_event_header *header,
1090                                 struct perf_sample_data *data,
1091                                 struct perf_event *event,
1092                                 struct pt_regs *regs);
1093
1094 extern int perf_event_overflow(struct perf_event *event,
1095                                  struct perf_sample_data *data,
1096                                  struct pt_regs *regs);
1097
1098 static inline bool is_sampling_event(struct perf_event *event)
1099 {
1100         return event->attr.sample_period != 0;
1101 }
1102
1103 /*
1104  * Return 1 for a software event, 0 for a hardware event
1105  */
1106 static inline int is_software_event(struct perf_event *event)
1107 {
1108         return event->pmu->task_ctx_nr == perf_sw_context;
1109 }
1110
1111 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1112
1113 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1114
1115 #ifndef perf_arch_fetch_caller_regs
1116 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1117 #endif
1118
1119 /*
1120  * Take a snapshot of the regs. Skip ip and frame pointer to
1121  * the nth caller. We only need a few of the regs:
1122  * - ip for PERF_SAMPLE_IP
1123  * - cs for user_mode() tests
1124  * - bp for callchains
1125  * - eflags, for future purposes, just in case
1126  */
1127 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1128 {
1129         memset(regs, 0, sizeof(*regs));
1130
1131         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1132 }
1133
1134 static __always_inline void
1135 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1136 {
1137         struct pt_regs hot_regs;
1138
1139         if (static_key_false(&perf_swevent_enabled[event_id])) {
1140                 if (!regs) {
1141                         perf_fetch_caller_regs(&hot_regs);
1142                         regs = &hot_regs;
1143                 }
1144                 __perf_sw_event(event_id, nr, regs, addr);
1145         }
1146 }
1147
1148 extern struct static_key_deferred perf_sched_events;
1149
1150 static inline void perf_event_task_sched_in(struct task_struct *prev,
1151                                             struct task_struct *task)
1152 {
1153         if (static_key_false(&perf_sched_events.key))
1154                 __perf_event_task_sched_in(prev, task);
1155 }
1156
1157 static inline void perf_event_task_sched_out(struct task_struct *prev,
1158                                              struct task_struct *next)
1159 {
1160         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1161
1162         if (static_key_false(&perf_sched_events.key))
1163                 __perf_event_task_sched_out(prev, next);
1164 }
1165
1166 extern void perf_event_mmap(struct vm_area_struct *vma);
1167 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1168 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1169 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1170
1171 extern void perf_event_comm(struct task_struct *tsk);
1172 extern void perf_event_fork(struct task_struct *tsk);
1173
1174 /* Callchains */
1175 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1176
1177 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1178 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1179
1180 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1181 {
1182         if (entry->nr < PERF_MAX_STACK_DEPTH)
1183                 entry->ip[entry->nr++] = ip;
1184 }
1185
1186 extern int sysctl_perf_event_paranoid;
1187 extern int sysctl_perf_event_mlock;
1188 extern int sysctl_perf_event_sample_rate;
1189
1190 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1191                 void __user *buffer, size_t *lenp,
1192                 loff_t *ppos);
1193
1194 static inline bool perf_paranoid_tracepoint_raw(void)
1195 {
1196         return sysctl_perf_event_paranoid > -1;
1197 }
1198
1199 static inline bool perf_paranoid_cpu(void)
1200 {
1201         return sysctl_perf_event_paranoid > 0;
1202 }
1203
1204 static inline bool perf_paranoid_kernel(void)
1205 {
1206         return sysctl_perf_event_paranoid > 1;
1207 }
1208
1209 extern void perf_event_init(void);
1210 extern void perf_tp_event(u64 addr, u64 count, void *record,
1211                           int entry_size, struct pt_regs *regs,
1212                           struct hlist_head *head, int rctx);
1213 extern void perf_bp_event(struct perf_event *event, void *data);
1214
1215 #ifndef perf_misc_flags
1216 # define perf_misc_flags(regs) \
1217                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1218 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1219 #endif
1220
1221 static inline bool has_branch_stack(struct perf_event *event)
1222 {
1223         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1224 }
1225
1226 extern int perf_output_begin(struct perf_output_handle *handle,
1227                              struct perf_event *event, unsigned int size);
1228 extern void perf_output_end(struct perf_output_handle *handle);
1229 extern void perf_output_copy(struct perf_output_handle *handle,
1230                              const void *buf, unsigned int len);
1231 extern int perf_swevent_get_recursion_context(void);
1232 extern void perf_swevent_put_recursion_context(int rctx);
1233 extern void perf_event_enable(struct perf_event *event);
1234 extern void perf_event_disable(struct perf_event *event);
1235 extern void perf_event_task_tick(void);
1236 #else
1237 static inline void
1238 perf_event_task_sched_in(struct task_struct *prev,
1239                          struct task_struct *task)                      { }
1240 static inline void
1241 perf_event_task_sched_out(struct task_struct *prev,
1242                           struct task_struct *next)                     { }
1243 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1244 static inline void perf_event_exit_task(struct task_struct *child)      { }
1245 static inline void perf_event_free_task(struct task_struct *task)       { }
1246 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1247 static inline void perf_event_print_debug(void)                         { }
1248 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1249 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1250 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1251 {
1252         return -EINVAL;
1253 }
1254
1255 static inline void
1256 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1257 static inline void
1258 perf_bp_event(struct perf_event *event, void *data)                     { }
1259
1260 static inline int perf_register_guest_info_callbacks
1261 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1262 static inline int perf_unregister_guest_info_callbacks
1263 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1264
1265 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1266 static inline void perf_event_comm(struct task_struct *tsk)             { }
1267 static inline void perf_event_fork(struct task_struct *tsk)             { }
1268 static inline void perf_event_init(void)                                { }
1269 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1270 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1271 static inline void perf_event_enable(struct perf_event *event)          { }
1272 static inline void perf_event_disable(struct perf_event *event)         { }
1273 static inline void perf_event_task_tick(void)                           { }
1274 #endif
1275
1276 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1277
1278 /*
1279  * This has to have a higher priority than migration_notifier in sched.c.
1280  */
1281 #define perf_cpu_notifier(fn)                                           \
1282 do {                                                                    \
1283         static struct notifier_block fn##_nb __cpuinitdata =            \
1284                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1285         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1286                 (void *)(unsigned long)smp_processor_id());             \
1287         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1288                 (void *)(unsigned long)smp_processor_id());             \
1289         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1290                 (void *)(unsigned long)smp_processor_id());             \
1291         register_cpu_notifier(&fn##_nb);                                \
1292 } while (0)
1293
1294 #endif /* __KERNEL__ */
1295 #endif /* _LINUX_PERF_EVENT_H */