include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
   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, time_mult, time_shift, idx, width;
 303          *   u64 count, enabled, running;
 304          *   u64 cyc, time_offset;
 305          *   s64 pmc = 0;
 306          *
 307          *   do {
 308          *     seq = pc->lock;
 309          *     barrier()
 310          *
 311          *     enabled = pc->time_enabled;
 312          *     running = pc->time_running;
 313          *
 314          *     if (pc->cap_usr_time && enabled != running) {
 315          *       cyc = rdtsc();
 316          *       time_offset = pc->time_offset;
 317          *       time_mult   = pc->time_mult;
 318          *       time_shift  = pc->time_shift;
 319          *     }
 320          *
 321          *     idx = pc->index;
 322          *     count = pc->offset;
 323          *     if (pc->cap_usr_rdpmc && idx) {
 324          *       width = pc->pmc_width;
 325          *       pmc = rdpmc(idx - 1);
 326          *     }
 327          *
 328          *     barrier();
 329          *   } while (pc->lock != seq);
 330          *
 331          * NOTE: for obvious reason this only works on self-monitoring
 332          *       processes.
 333          */
 334         __u32   lock;                   /* seqlock for synchronization */
 335         __u32   index;                  /* hardware event identifier */
 336         __s64   offset;                 /* add to hardware event value */
 337         __u64   time_enabled;           /* time event active */
 338         __u64   time_running;           /* time event on cpu */
 339         union {
 340                 __u64   capabilities;
 341                 __u64   cap_usr_time  : 1,
 342                         cap_usr_rdpmc : 1,
 343                         cap_____res   : 62;
 344         };
 345
 346         /*
 347          * If cap_usr_rdpmc this field provides the bit-width of the value
 348          * read using the rdpmc() or equivalent instruction. This can be used
 349          * to sign extend the result like:
 350          *
 351          *   pmc <<= 64 - width;
 352          *   pmc >>= 64 - width; // signed shift right
 353          *   count += pmc;
 354          */
 355         __u16   pmc_width;
 356
 357         /*
 358          * If cap_usr_time the below fields can be used to compute the time
 359          * delta since time_enabled (in ns) using rdtsc or similar.
 360          *
 361          *   u64 quot, rem;
 362          *   u64 delta;
 363          *
 364          *   quot = (cyc >> time_shift);
 365          *   rem = cyc & ((1 << time_shift) - 1);
 366          *   delta = time_offset + quot * time_mult +
 367          *              ((rem * time_mult) >> time_shift);
 368          *
 369          * Where time_offset,time_mult,time_shift and cyc are read in the
 370          * seqcount loop described above. This delta can then be added to
 371          * enabled and possible running (if idx), improving the scaling:
 372          *
 373          *   enabled += delta;
 374          *   if (idx)
 375          *     running += delta;
 376          *
 377          *   quot = count / running;
 378          *   rem  = count % running;
 379          *   count = quot * enabled + (rem * enabled) / running;
 380          */
 381         __u16   time_shift;
 382         __u32   time_mult;
 383         __u64   time_offset;
 384
 385                 /*
 386                  * Hole for extension of the self monitor capabilities
 387                  */
 388
 389         __u64   __reserved[120];        /* align to 1k */
 390
 391         /*
 392          * Control data for the mmap() data buffer.
 393          *
 394          * User-space reading the @data_head value should issue an rmb(), on
 395          * SMP capable platforms, after reading this value -- see
 396          * perf_event_wakeup().
 397          *
 398          * When the mapping is PROT_WRITE the @data_tail value should be
 399          * written by userspace to reflect the last read data. In this case
 400          * the kernel will not over-write unread data.
 401          */
 402         __u64   data_head;              /* head in the data section */
 403         __u64   data_tail;              /* user-space written tail */
 404 };
 405
 406 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 407 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 408 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 409 #define PERF_RECORD_MISC_USER                   (2 << 0)
 410 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 411 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 412 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 413
 414 /*
 415  * Indicates that the content of PERF_SAMPLE_IP points to
 416  * the actual instruction that triggered the event. See also
 417  * perf_event_attr::precise_ip.
 418  */
 419 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 420 /*
 421  * Reserve the last bit to indicate some extended misc field
 422  */
 423 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 424
 425 struct perf_event_header {
 426         __u32   type;
 427         __u16   misc;
 428         __u16   size;
 429 };
 430
 431 enum perf_event_type {
 432
 433         /*
 434          * If perf_event_attr.sample_id_all is set then all event types will
 435          * have the sample_type selected fields related to where/when
 436          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
 437          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
 438          * the perf_event_header and the fields already present for the existing
 439          * fields, i.e. at the end of the payload. That way a newer perf.data
 440          * file will be supported by older perf tools, with these new optional
 441          * fields being ignored.
 442          *
 443          * The MMAP events record the PROT_EXEC mappings so that we can
 444          * correlate userspace IPs to code. They have the following structure:
 445          *
 446          * struct {
 447          *      struct perf_event_header        header;
 448          *
 449          *      u32                             pid, tid;
 450          *      u64                             addr;
 451          *      u64                             len;
 452          *      u64                             pgoff;
 453          *      char                            filename[];
 454          * };
 455          */
 456         PERF_RECORD_MMAP                        = 1,
 457
 458         /*
 459          * struct {
 460          *      struct perf_event_header        header;
 461          *      u64                             id;
 462          *      u64                             lost;
 463          * };
 464          */
 465         PERF_RECORD_LOST                        = 2,
 466
 467         /*
 468          * struct {
 469          *      struct perf_event_header        header;
 470          *
 471          *      u32                             pid, tid;
 472          *      char                            comm[];
 473          * };
 474          */
 475         PERF_RECORD_COMM                        = 3,
 476
 477         /*
 478          * struct {
 479          *      struct perf_event_header        header;
 480          *      u32                             pid, ppid;
 481          *      u32                             tid, ptid;
 482          *      u64                             time;
 483          * };
 484          */
 485         PERF_RECORD_EXIT                        = 4,
 486
 487         /*
 488          * struct {
 489          *      struct perf_event_header        header;
 490          *      u64                             time;
 491          *      u64                             id;
 492          *      u64                             stream_id;
 493          * };
 494          */
 495         PERF_RECORD_THROTTLE                    = 5,
 496         PERF_RECORD_UNTHROTTLE                  = 6,
 497
 498         /*
 499          * struct {
 500          *      struct perf_event_header        header;
 501          *      u32                             pid, ppid;
 502          *      u32                             tid, ptid;
 503          *      u64                             time;
 504          * };
 505          */
 506         PERF_RECORD_FORK                        = 7,
 507
 508         /*
 509          * struct {
 510          *      struct perf_event_header        header;
 511          *      u32                             pid, tid;
 512          *
 513          *      struct read_format              values;
 514          * };
 515          */
 516         PERF_RECORD_READ                        = 8,
 517
 518         /*
 519          * struct {
 520          *      struct perf_event_header        header;
 521          *
 522          *      { u64                   ip;       } && PERF_SAMPLE_IP
 523          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 524          *      { u64                   time;     } && PERF_SAMPLE_TIME
 525          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 526          *      { u64                   id;       } && PERF_SAMPLE_ID
 527          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 528          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 529          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 530          *
 531          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 532          *
 533          *      { u64                   nr,
 534          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 535          *
 536          *      #
 537          *      # The RAW record below is opaque data wrt the ABI
 538          *      #
 539          *      # That is, the ABI doesn't make any promises wrt to
 540          *      # the stability of its content, it may vary depending
 541          *      # on event, hardware, kernel version and phase of
 542          *      # the moon.
 543          *      #
 544          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 545          *      #
 546          *
 547          *      { u32                   size;
 548          *        char                  data[size];}&& PERF_SAMPLE_RAW
 549          *
 550          *      { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
 551          * };
 552          */
 553         PERF_RECORD_SAMPLE                      = 9,
 554
 555         PERF_RECORD_MAX,                        /* non-ABI */
 556 };
 557
 558 #define PERF_MAX_STACK_DEPTH            127
 559
 560 enum perf_callchain_context {
 561         PERF_CONTEXT_HV                 = (__u64)-32,
 562         PERF_CONTEXT_KERNEL             = (__u64)-128,
 563         PERF_CONTEXT_USER               = (__u64)-512,
 564
 565         PERF_CONTEXT_GUEST              = (__u64)-2048,
 566         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 567         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 568
 569         PERF_CONTEXT_MAX                = (__u64)-4095,
 570 };
 571
 572 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 573 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 574 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
 575
 576 #ifdef __KERNEL__
 577 /*
 578  * Kernel-internal data types and definitions:
 579  */
 580
 581 #ifdef CONFIG_PERF_EVENTS
 582 # include <linux/cgroup.h>
 583 # include <asm/perf_event.h>
 584 # include <asm/local64.h>
 585 #endif
 586
 587 struct perf_guest_info_callbacks {
 588         int                             (*is_in_guest)(void);
 589         int                             (*is_user_mode)(void);
 590         unsigned long                   (*get_guest_ip)(void);
 591 };
 592
 593 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 594 #include <asm/hw_breakpoint.h>
 595 #endif
 596
 597 #include <linux/list.h>
 598 #include <linux/mutex.h>
 599 #include <linux/rculist.h>
 600 #include <linux/rcupdate.h>
 601 #include <linux/spinlock.h>
 602 #include <linux/hrtimer.h>
 603 #include <linux/fs.h>
 604 #include <linux/pid_namespace.h>
 605 #include <linux/workqueue.h>
 606 #include <linux/ftrace.h>
 607 #include <linux/cpu.h>
 608 #include <linux/irq_work.h>
 609 #include <linux/static_key.h>
 610 #include <linux/atomic.h>
 611 #include <linux/sysfs.h>
 612 #include <asm/local.h>
 613
 614 struct perf_callchain_entry {
 615         __u64                           nr;
 616         __u64                           ip[PERF_MAX_STACK_DEPTH];
 617 };
 618
 619 struct perf_raw_record {
 620         u32                             size;
 621         void                            *data;
 622 };
 623
 624 /*
 625  * single taken branch record layout:
 626  *
 627  *      from: source instruction (may not always be a branch insn)
 628  *        to: branch target
 629  *   mispred: branch target was mispredicted
 630  * predicted: branch target was predicted
 631  *
 632  * support for mispred, predicted is optional. In case it
 633  * is not supported mispred = predicted = 0.
 634  */
 635 struct perf_branch_entry {
 636         __u64   from;
 637         __u64   to;
 638         __u64   mispred:1,  /* target mispredicted */
 639                 predicted:1,/* target predicted */
 640                 reserved:62;
 641 };
 642
 643 /*
 644  * branch stack layout:
 645  *  nr: number of taken branches stored in entries[]
 646  *
 647  * Note that nr can vary from sample to sample
 648  * branches (to, from) are stored from most recent
 649  * to least recent, i.e., entries[0] contains the most
 650  * recent branch.
 651  */
 652 struct perf_branch_stack {
 653         __u64                           nr;
 654         struct perf_branch_entry        entries[0];
 655 };
 656
 657 struct task_struct;
 658
 659 /*
 660  * extra PMU register associated with an event
 661  */
 662 struct hw_perf_event_extra {
 663         u64             config; /* register value */
 664         unsigned int    reg;    /* register address or index */
 665         int             alloc;  /* extra register already allocated */
 666         int             idx;    /* index in shared_regs->regs[] */
 667 };
 668
 669 /**
 670  * struct hw_perf_event - performance event hardware details:
 671  */
 672 struct hw_perf_event {
 673 #ifdef CONFIG_PERF_EVENTS
 674         union {
 675                 struct { /* hardware */
 676                         u64             config;
 677                         u64             last_tag;
 678                         unsigned long   config_base;
 679                         unsigned long   event_base;
 680                         int             event_base_rdpmc;
 681                         int             idx;
 682                         int             last_cpu;
 683
 684                         struct hw_perf_event_extra extra_reg;
 685                         struct hw_perf_event_extra branch_reg;
 686                 };
 687                 struct { /* software */
 688                         struct hrtimer  hrtimer;
 689                 };
 690 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 691                 struct { /* breakpoint */
 692                         struct arch_hw_breakpoint       info;
 693                         struct list_head                bp_list;
 694                         /*
 695                          * Crufty hack to avoid the chicken and egg
 696                          * problem hw_breakpoint has with context
 697                          * creation and event initalization.
 698                          */
 699                         struct task_struct              *bp_target;
 700                 };
 701 #endif
 702         };
 703         int                             state;
 704         local64_t                       prev_count;
 705         u64                             sample_period;
 706         u64                             last_period;
 707         local64_t                       period_left;
 708         u64                             interrupts_seq;
 709         u64                             interrupts;
 710
 711         u64                             freq_time_stamp;
 712         u64                             freq_count_stamp;
 713 #endif
 714 };
 715
 716 /*
 717  * hw_perf_event::state flags
 718  */
 719 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 720 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 721 #define PERF_HES_ARCH           0x04
 722
 723 struct perf_event;
 724
 725 /*
 726  * Common implementation detail of pmu::{start,commit,cancel}_txn
 727  */
 728 #define PERF_EVENT_TXN 0x1
 729
 730 /**
 731  * struct pmu - generic performance monitoring unit
 732  */
 733 struct pmu {
 734         struct list_head                entry;
 735
 736         struct device                   *dev;
 737         const struct attribute_group    **attr_groups;
 738         char                            *name;
 739         int                             type;
 740
 741         int * __percpu                  pmu_disable_count;
 742         struct perf_cpu_context * __percpu pmu_cpu_context;
 743         int                             task_ctx_nr;
 744
 745         /*
 746          * Fully disable/enable this PMU, can be used to protect from the PMI
 747          * as well as for lazy/batch writing of the MSRs.
 748          */
 749         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 750         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 751
 752         /*
 753          * Try and initialize the event for this PMU.
 754          * Should return -ENOENT when the @event doesn't match this PMU.
 755          */
 756         int (*event_init)               (struct perf_event *event);
 757
 758 #define PERF_EF_START   0x01            /* start the counter when adding    */
 759 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 760 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 761
 762         /*
 763          * Adds/Removes a counter to/from the PMU, can be done inside
 764          * a transaction, see the ->*_txn() methods.
 765          */
 766         int  (*add)                     (struct perf_event *event, int flags);
 767         void (*del)                     (struct perf_event *event, int flags);
 768
 769         /*
 770          * Starts/Stops a counter present on the PMU. The PMI handler
 771          * should stop the counter when perf_event_overflow() returns
 772          * !0. ->start() will be used to continue.
 773          */
 774         void (*start)                   (struct perf_event *event, int flags);
 775         void (*stop)                    (struct perf_event *event, int flags);
 776
 777         /*
 778          * Updates the counter value of the event.
 779          */
 780         void (*read)                    (struct perf_event *event);
 781
 782         /*
 783          * Group events scheduling is treated as a transaction, add
 784          * group events as a whole and perform one schedulability test.
 785          * If the test fails, roll back the whole group
 786          *
 787          * Start the transaction, after this ->add() doesn't need to
 788          * do schedulability tests.
 789          */
 790         void (*start_txn)               (struct pmu *pmu); /* optional */
 791         /*
 792          * If ->start_txn() disabled the ->add() schedulability test
 793          * then ->commit_txn() is required to perform one. On success
 794          * the transaction is closed. On error the transaction is kept
 795          * open until ->cancel_txn() is called.
 796          */
 797         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 798         /*
 799          * Will cancel the transaction, assumes ->del() is called
 800          * for each successful ->add() during the transaction.
 801          */
 802         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 803
 804         /*
 805          * Will return the value for perf_event_mmap_page::index for this event,
 806          * if no implementation is provided it will default to: event->hw.idx + 1.
 807          */
 808         int (*event_idx)                (struct perf_event *event); /*optional */
 809
 810         /*
 811          * flush branch stack on context-switches (needed in cpu-wide mode)
 812          */
 813         void (*flush_branch_stack)      (void);
 814 };
 815
 816 /**
 817  * enum perf_event_active_state - the states of a event
 818  */
 819 enum perf_event_active_state {
 820         PERF_EVENT_STATE_ERROR          = -2,
 821         PERF_EVENT_STATE_OFF            = -1,
 822         PERF_EVENT_STATE_INACTIVE       =  0,
 823         PERF_EVENT_STATE_ACTIVE         =  1,
 824 };
 825
 826 struct file;
 827 struct perf_sample_data;
 828
 829 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 830                                         struct perf_sample_data *,
 831                                         struct pt_regs *regs);
 832
 833 enum perf_group_flag {
 834         PERF_GROUP_SOFTWARE             = 0x1,
 835 };
 836
 837 #define SWEVENT_HLIST_BITS              8
 838 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 839
 840 struct swevent_hlist {
 841         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 842         struct rcu_head                 rcu_head;
 843 };
 844
 845 #define PERF_ATTACH_CONTEXT     0x01
 846 #define PERF_ATTACH_GROUP       0x02
 847 #define PERF_ATTACH_TASK        0x04
 848
 849 #ifdef CONFIG_CGROUP_PERF
 850 /*
 851  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 852  * This is a per-cpu dynamically allocated data structure.
 853  */
 854 struct perf_cgroup_info {
 855         u64                             time;
 856         u64                             timestamp;
 857 };
 858
 859 struct perf_cgroup {
 860         struct                          cgroup_subsys_state css;
 861         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
 862 };
 863 #endif
 864
 865 struct ring_buffer;
 866
 867 /**
 868  * struct perf_event - performance event kernel representation:
 869  */
 870 struct perf_event {
 871 #ifdef CONFIG_PERF_EVENTS
 872         struct list_head                group_entry;
 873         struct list_head                event_entry;
 874         struct list_head                sibling_list;
 875         struct hlist_node               hlist_entry;
 876         int                             nr_siblings;
 877         int                             group_flags;
 878         struct perf_event               *group_leader;
 879         struct pmu                      *pmu;
 880
 881         enum perf_event_active_state    state;
 882         unsigned int                    attach_state;
 883         local64_t                       count;
 884         atomic64_t                      child_count;
 885
 886         /*
 887          * These are the total time in nanoseconds that the event
 888          * has been enabled (i.e. eligible to run, and the task has
 889          * been scheduled in, if this is a per-task event)
 890          * and running (scheduled onto the CPU), respectively.
 891          *
 892          * They are computed from tstamp_enabled, tstamp_running and
 893          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 894          */
 895         u64                             total_time_enabled;
 896         u64                             total_time_running;
 897
 898         /*
 899          * These are timestamps used for computing total_time_enabled
 900          * and total_time_running when the event is in INACTIVE or
 901          * ACTIVE state, measured in nanoseconds from an arbitrary point
 902          * in time.
 903          * tstamp_enabled: the notional time when the event was enabled
 904          * tstamp_running: the notional time when the event was scheduled on
 905          * tstamp_stopped: in INACTIVE state, the notional time when the
 906          *      event was scheduled off.
 907          */
 908         u64                             tstamp_enabled;
 909         u64                             tstamp_running;
 910         u64                             tstamp_stopped;
 911
 912         /*
 913          * timestamp shadows the actual context timing but it can
 914          * be safely used in NMI interrupt context. It reflects the
 915          * context time as it was when the event was last scheduled in.
 916          *
 917          * ctx_time already accounts for ctx->timestamp. Therefore to
 918          * compute ctx_time for a sample, simply add perf_clock().
 919          */
 920         u64                             shadow_ctx_time;
 921
 922         struct perf_event_attr          attr;
 923         u16                             header_size;
 924         u16                             id_header_size;
 925         u16                             read_size;
 926         struct hw_perf_event            hw;
 927
 928         struct perf_event_context       *ctx;
 929         struct file                     *filp;
 930
 931         /*
 932          * These accumulate total time (in nanoseconds) that children
 933          * events have been enabled and running, respectively.
 934          */
 935         atomic64_t                      child_total_time_enabled;
 936         atomic64_t                      child_total_time_running;
 937
 938         /*
 939          * Protect attach/detach and child_list:
 940          */
 941         struct mutex                    child_mutex;
 942         struct list_head                child_list;
 943         struct perf_event               *parent;
 944
 945         int                             oncpu;
 946         int                             cpu;
 947
 948         struct list_head                owner_entry;
 949         struct task_struct              *owner;
 950
 951         /* mmap bits */
 952         struct mutex                    mmap_mutex;
 953         atomic_t                        mmap_count;
 954         int                             mmap_locked;
 955         struct user_struct              *mmap_user;
 956         struct ring_buffer              *rb;
 957         struct list_head                rb_entry;
 958
 959         /* poll related */
 960         wait_queue_head_t               waitq;
 961         struct fasync_struct            *fasync;
 962
 963         /* delayed work for NMIs and such */
 964         int                             pending_wakeup;
 965         int                             pending_kill;
 966         int                             pending_disable;
 967         struct irq_work                 pending;
 968
 969         atomic_t                        event_limit;
 970
 971         void (*destroy)(struct perf_event *);
 972         struct rcu_head                 rcu_head;
 973
 974         struct pid_namespace            *ns;
 975         u64                             id;
 976
 977         perf_overflow_handler_t         overflow_handler;
 978         void                            *overflow_handler_context;
 979
 980 #ifdef CONFIG_EVENT_TRACING
 981         struct ftrace_event_call        *tp_event;
 982         struct event_filter             *filter;
 983 #ifdef CONFIG_FUNCTION_TRACER
 984         struct ftrace_ops               ftrace_ops;
 985 #endif
 986 #endif
 987
 988 #ifdef CONFIG_CGROUP_PERF
 989         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 990         int                             cgrp_defer_enabled;
 991 #endif
 992
 993 #endif /* CONFIG_PERF_EVENTS */
 994 };
 995
 996 enum perf_event_context_type {
 997         task_context,
 998         cpu_context,
 999 };
1000
1001 /**
1002  * struct perf_event_context - event context structure
1003  *
1004  * Used as a container for task events and CPU events as well:
1005  */
1006 struct perf_event_context {
1007         struct pmu                      *pmu;
1008         enum perf_event_context_type    type;
1009         /*
1010          * Protect the states of the events in the list,
1011          * nr_active, and the list:
1012          */
1013         raw_spinlock_t                  lock;
1014         /*
1015          * Protect the list of events.  Locking either mutex or lock
1016          * is sufficient to ensure the list doesn't change; to change
1017          * the list you need to lock both the mutex and the spinlock.
1018          */
1019         struct mutex                    mutex;
1020
1021         struct list_head                pinned_groups;
1022         struct list_head                flexible_groups;
1023         struct list_head                event_list;
1024         int                             nr_events;
1025         int                             nr_active;
1026         int                             is_active;
1027         int                             nr_stat;
1028         int                             nr_freq;
1029         int                             rotate_disable;
1030         atomic_t                        refcount;
1031         struct task_struct              *task;
1032
1033         /*
1034          * Context clock, runs when context enabled.
1035          */
1036         u64                             time;
1037         u64                             timestamp;
1038
1039         /*
1040          * These fields let us detect when two contexts have both
1041          * been cloned (inherited) from a common ancestor.
1042          */
1043         struct perf_event_context       *parent_ctx;
1044         u64                             parent_gen;
1045         u64                             generation;
1046         int                             pin_count;
1047         int                             nr_cgroups;      /* cgroup evts */
1048         int                             nr_branch_stack; /* branch_stack evt */
1049         struct rcu_head                 rcu_head;
1050 };
1051
1052 /*
1053  * Number of contexts where an event can trigger:
1054  *      task, softirq, hardirq, nmi.
1055  */
1056 #define PERF_NR_CONTEXTS        4
1057
1058 /**
1059  * struct perf_event_cpu_context - per cpu event context structure
1060  */
1061 struct perf_cpu_context {
1062         struct perf_event_context       ctx;
1063         struct perf_event_context       *task_ctx;
1064         int                             active_oncpu;
1065         int                             exclusive;
1066         struct list_head                rotation_list;
1067         int                             jiffies_interval;
1068         struct pmu                      *active_pmu;
1069         struct perf_cgroup              *cgrp;
1070 };
1071
1072 struct perf_output_handle {
1073         struct perf_event               *event;
1074         struct ring_buffer              *rb;
1075         unsigned long                   wakeup;
1076         unsigned long                   size;
1077         void                            *addr;
1078         int                             page;
1079 };
1080
1081 #ifdef CONFIG_PERF_EVENTS
1082
1083 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1084 extern void perf_pmu_unregister(struct pmu *pmu);
1085
1086 extern int perf_num_counters(void);
1087 extern const char *perf_pmu_name(void);
1088 extern void __perf_event_task_sched_in(struct task_struct *prev,
1089                                        struct task_struct *task);
1090 extern void __perf_event_task_sched_out(struct task_struct *prev,
1091                                         struct task_struct *next);
1092 extern int perf_event_init_task(struct task_struct *child);
1093 extern void perf_event_exit_task(struct task_struct *child);
1094 extern void perf_event_free_task(struct task_struct *task);
1095 extern void perf_event_delayed_put(struct task_struct *task);
1096 extern void perf_event_print_debug(void);
1097 extern void perf_pmu_disable(struct pmu *pmu);
1098 extern void perf_pmu_enable(struct pmu *pmu);
1099 extern int perf_event_task_disable(void);
1100 extern int perf_event_task_enable(void);
1101 extern int perf_event_refresh(struct perf_event *event, int refresh);
1102 extern void perf_event_update_userpage(struct perf_event *event);
1103 extern int perf_event_release_kernel(struct perf_event *event);
1104 extern struct perf_event *
1105 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1106                                 int cpu,
1107                                 struct task_struct *task,
1108                                 perf_overflow_handler_t callback,
1109                                 void *context);
1110 extern void perf_pmu_migrate_context(struct pmu *pmu,
1111                                 int src_cpu, int dst_cpu);
1112 extern u64 perf_event_read_value(struct perf_event *event,
1113                                  u64 *enabled, u64 *running);
1114
1115
1116 struct perf_sample_data {
1117         u64                             type;
1118
1119         u64                             ip;
1120         struct {
1121                 u32     pid;
1122                 u32     tid;
1123         }                               tid_entry;
1124         u64                             time;
1125         u64                             addr;
1126         u64                             id;
1127         u64                             stream_id;
1128         struct {
1129                 u32     cpu;
1130                 u32     reserved;
1131         }                               cpu_entry;
1132         u64                             period;
1133         struct perf_callchain_entry     *callchain;
1134         struct perf_raw_record          *raw;
1135         struct perf_branch_stack        *br_stack;
1136 };
1137
1138 static inline void perf_sample_data_init(struct perf_sample_data *data,
1139                                          u64 addr, u64 period)
1140 {
1141         /* remaining struct members initialized in perf_prepare_sample() */
1142         data->addr = addr;
1143         data->raw  = NULL;
1144         data->br_stack = NULL;
1145         data->period    = period;
1146 }
1147
1148 extern void perf_output_sample(struct perf_output_handle *handle,
1149                                struct perf_event_header *header,
1150                                struct perf_sample_data *data,
1151                                struct perf_event *event);
1152 extern void perf_prepare_sample(struct perf_event_header *header,
1153                                 struct perf_sample_data *data,
1154                                 struct perf_event *event,
1155                                 struct pt_regs *regs);
1156
1157 extern int perf_event_overflow(struct perf_event *event,
1158                                  struct perf_sample_data *data,
1159                                  struct pt_regs *regs);
1160
1161 static inline bool is_sampling_event(struct perf_event *event)
1162 {
1163         return event->attr.sample_period != 0;
1164 }
1165
1166 /*
1167  * Return 1 for a software event, 0 for a hardware event
1168  */
1169 static inline int is_software_event(struct perf_event *event)
1170 {
1171         return event->pmu->task_ctx_nr == perf_sw_context;
1172 }
1173
1174 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1175
1176 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1177
1178 #ifndef perf_arch_fetch_caller_regs
1179 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1180 #endif
1181
1182 /*
1183  * Take a snapshot of the regs. Skip ip and frame pointer to
1184  * the nth caller. We only need a few of the regs:
1185  * - ip for PERF_SAMPLE_IP
1186  * - cs for user_mode() tests
1187  * - bp for callchains
1188  * - eflags, for future purposes, just in case
1189  */
1190 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1191 {
1192         memset(regs, 0, sizeof(*regs));
1193
1194         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1195 }
1196
1197 static __always_inline void
1198 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1199 {
1200         struct pt_regs hot_regs;
1201
1202         if (static_key_false(&perf_swevent_enabled[event_id])) {
1203                 if (!regs) {
1204                         perf_fetch_caller_regs(&hot_regs);
1205                         regs = &hot_regs;
1206                 }
1207                 __perf_sw_event(event_id, nr, regs, addr);
1208         }
1209 }
1210
1211 extern struct static_key_deferred perf_sched_events;
1212
1213 static inline void perf_event_task_sched_in(struct task_struct *prev,
1214                                             struct task_struct *task)
1215 {
1216         if (static_key_false(&perf_sched_events.key))
1217                 __perf_event_task_sched_in(prev, task);
1218 }
1219
1220 static inline void perf_event_task_sched_out(struct task_struct *prev,
1221                                              struct task_struct *next)
1222 {
1223         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1224
1225         if (static_key_false(&perf_sched_events.key))
1226                 __perf_event_task_sched_out(prev, next);
1227 }
1228
1229 extern void perf_event_mmap(struct vm_area_struct *vma);
1230 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1231 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1232 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1233
1234 extern void perf_event_comm(struct task_struct *tsk);
1235 extern void perf_event_fork(struct task_struct *tsk);
1236
1237 /* Callchains */
1238 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1239
1240 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1241 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1242
1243 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1244 {
1245         if (entry->nr < PERF_MAX_STACK_DEPTH)
1246                 entry->ip[entry->nr++] = ip;
1247 }
1248
1249 extern int sysctl_perf_event_paranoid;
1250 extern int sysctl_perf_event_mlock;
1251 extern int sysctl_perf_event_sample_rate;
1252
1253 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1254                 void __user *buffer, size_t *lenp,
1255                 loff_t *ppos);
1256
1257 static inline bool perf_paranoid_tracepoint_raw(void)
1258 {
1259         return sysctl_perf_event_paranoid > -1;
1260 }
1261
1262 static inline bool perf_paranoid_cpu(void)
1263 {
1264         return sysctl_perf_event_paranoid > 0;
1265 }
1266
1267 static inline bool perf_paranoid_kernel(void)
1268 {
1269         return sysctl_perf_event_paranoid > 1;
1270 }
1271
1272 extern void perf_event_init(void);
1273 extern void perf_tp_event(u64 addr, u64 count, void *record,
1274                           int entry_size, struct pt_regs *regs,
1275                           struct hlist_head *head, int rctx);
1276 extern void perf_bp_event(struct perf_event *event, void *data);
1277
1278 #ifndef perf_misc_flags
1279 # define perf_misc_flags(regs) \
1280                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1281 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1282 #endif
1283
1284 static inline bool has_branch_stack(struct perf_event *event)
1285 {
1286         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1287 }
1288
1289 extern int perf_output_begin(struct perf_output_handle *handle,
1290                              struct perf_event *event, unsigned int size);
1291 extern void perf_output_end(struct perf_output_handle *handle);
1292 extern void perf_output_copy(struct perf_output_handle *handle,
1293                              const void *buf, unsigned int len);
1294 extern int perf_swevent_get_recursion_context(void);
1295 extern void perf_swevent_put_recursion_context(int rctx);
1296 extern void perf_event_enable(struct perf_event *event);
1297 extern void perf_event_disable(struct perf_event *event);
1298 extern void perf_event_task_tick(void);
1299 #else
1300 static inline void
1301 perf_event_task_sched_in(struct task_struct *prev,
1302                          struct task_struct *task)                      { }
1303 static inline void
1304 perf_event_task_sched_out(struct task_struct *prev,
1305                           struct task_struct *next)                     { }
1306 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1307 static inline void perf_event_exit_task(struct task_struct *child)      { }
1308 static inline void perf_event_free_task(struct task_struct *task)       { }
1309 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1310 static inline void perf_event_print_debug(void)                         { }
1311 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1312 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1313 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1314 {
1315         return -EINVAL;
1316 }
1317
1318 static inline void
1319 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1320 static inline void
1321 perf_bp_event(struct perf_event *event, void *data)                     { }
1322
1323 static inline int perf_register_guest_info_callbacks
1324 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1325 static inline int perf_unregister_guest_info_callbacks
1326 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1327
1328 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1329 static inline void perf_event_comm(struct task_struct *tsk)             { }
1330 static inline void perf_event_fork(struct task_struct *tsk)             { }
1331 static inline void perf_event_init(void)                                { }
1332 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1333 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1334 static inline void perf_event_enable(struct perf_event *event)          { }
1335 static inline void perf_event_disable(struct perf_event *event)         { }
1336 static inline void perf_event_task_tick(void)                           { }
1337 #endif
1338
1339 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1340
1341 /*
1342  * This has to have a higher priority than migration_notifier in sched.c.
1343  */
1344 #define perf_cpu_notifier(fn)                                           \
1345 do {                                                                    \
1346         static struct notifier_block fn##_nb __cpuinitdata =            \
1347                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1348         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1349                 (void *)(unsigned long)smp_processor_id());             \
1350         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1351                 (void *)(unsigned long)smp_processor_id());             \
1352         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1353                 (void *)(unsigned long)smp_processor_id());             \
1354         register_cpu_notifier(&fn##_nb);                                \
1355 } while (0)
1356
1357
1358 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1359 static ssize_t                                                          \
1360 _name##_show(struct device *dev,                                        \
1361                                struct device_attribute *attr,           \
1362                                char *page)                              \
1363 {                                                                       \
1364         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1365         return sprintf(page, _format "\n");                             \
1366 }                                                                       \
1367                                                                         \
1368 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1369
1370 #endif /* __KERNEL__ */
1371 #endif /* _LINUX_PERF_EVENT_H */