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