include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <linux/types.h>
  18 #include <linux/ioctl.h>
  19 #include <asm/byteorder.h>
  20
  21 /*
  22  * User-space ABI bits:
  23  */
  24
  25 /*
  26  * attr.type
  27  */
  28 enum perf_type_id {
  29         PERF_TYPE_HARDWARE                      = 0,
  30         PERF_TYPE_SOFTWARE                      = 1,
  31         PERF_TYPE_TRACEPOINT                    = 2,
  32         PERF_TYPE_HW_CACHE                      = 3,
  33         PERF_TYPE_RAW                           = 4,
  34         PERF_TYPE_BREAKPOINT                    = 5,
  35
  36         PERF_TYPE_MAX,                          /* non-ABI */
  37 };
  38
  39 /*
  40  * Generalized performance event event_id types, used by the
  41  * attr.event_id parameter of the sys_perf_event_open()
  42  * syscall:
  43  */
  44 enum perf_hw_id {
  45         /*
  46          * Common hardware events, generalized by the kernel:
  47          */
  48         PERF_COUNT_HW_CPU_CYCLES                = 0,
  49         PERF_COUNT_HW_INSTRUCTIONS              = 1,
  50         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
  51         PERF_COUNT_HW_CACHE_MISSES              = 3,
  52         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
  53         PERF_COUNT_HW_BRANCH_MISSES             = 5,
  54         PERF_COUNT_HW_BUS_CYCLES                = 6,
  55         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
  56         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
  57         PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
  58
  59         PERF_COUNT_HW_MAX,                      /* non-ABI */
  60 };
  61
  62 /*
  63  * Generalized hardware cache events:
  64  *
  65  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
  66  *       { read, write, prefetch } x
  67  *       { accesses, misses }
  68  */
  69 enum perf_hw_cache_id {
  70         PERF_COUNT_HW_CACHE_L1D                 = 0,
  71         PERF_COUNT_HW_CACHE_L1I                 = 1,
  72         PERF_COUNT_HW_CACHE_LL                  = 2,
  73         PERF_COUNT_HW_CACHE_DTLB                = 3,
  74         PERF_COUNT_HW_CACHE_ITLB                = 4,
  75         PERF_COUNT_HW_CACHE_BPU                 = 5,
  76         PERF_COUNT_HW_CACHE_NODE                = 6,
  77
  78         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
  79 };
  80
  81 enum perf_hw_cache_op_id {
  82         PERF_COUNT_HW_CACHE_OP_READ             = 0,
  83         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
  84         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
  85
  86         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
  87 };
  88
  89 enum perf_hw_cache_op_result_id {
  90         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
  91         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
  92
  93         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
  94 };
  95
  96 /*
  97  * Special "software" events provided by the kernel, even if the hardware
  98  * does not support performance events. These events measure various
  99  * physical and sw events of the kernel (and allow the profiling of them as
 100  * well):
 101  */
 102 enum perf_sw_ids {
 103         PERF_COUNT_SW_CPU_CLOCK                 = 0,
 104         PERF_COUNT_SW_TASK_CLOCK                = 1,
 105         PERF_COUNT_SW_PAGE_FAULTS               = 2,
 106         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
 107         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
 108         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
 109         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
 110         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
 111         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
 112
 113         PERF_COUNT_SW_MAX,                      /* non-ABI */
 114 };
 115
 116 /*
 117  * Bits that can be set in attr.sample_type to request information
 118  * in the overflow packets.
 119  */
 120 enum perf_event_sample_format {
 121         PERF_SAMPLE_IP                          = 1U << 0,
 122         PERF_SAMPLE_TID                         = 1U << 1,
 123         PERF_SAMPLE_TIME                        = 1U << 2,
 124         PERF_SAMPLE_ADDR                        = 1U << 3,
 125         PERF_SAMPLE_READ                        = 1U << 4,
 126         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
 127         PERF_SAMPLE_ID                          = 1U << 6,
 128         PERF_SAMPLE_CPU                         = 1U << 7,
 129         PERF_SAMPLE_PERIOD                      = 1U << 8,
 130         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
 131         PERF_SAMPLE_RAW                         = 1U << 10,
 132         PERF_SAMPLE_BRANCH_STACK                = 1U << 11,
 133
 134         PERF_SAMPLE_MAX = 1U << 12,             /* non-ABI */
 135 };
 136
 137 /*
 138  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
 139  *
 140  * If the user does not pass priv level information via branch_sample_type,
 141  * the kernel uses the event's priv level. Branch and event priv levels do
 142  * not have to match. Branch priv level is checked for permissions.
 143  *
 144  * The branch types can be combined, however BRANCH_ANY covers all types
 145  * of branches and therefore it supersedes all the other types.
 146  */
 147 enum perf_branch_sample_type {
 148         PERF_SAMPLE_BRANCH_USER         = 1U << 0, /* user branches */
 149         PERF_SAMPLE_BRANCH_KERNEL       = 1U << 1, /* kernel branches */
 150         PERF_SAMPLE_BRANCH_HV           = 1U << 2, /* hypervisor branches */
 151
 152         PERF_SAMPLE_BRANCH_ANY          = 1U << 3, /* any branch types */
 153         PERF_SAMPLE_BRANCH_ANY_CALL     = 1U << 4, /* any call branch */
 154         PERF_SAMPLE_BRANCH_ANY_RETURN   = 1U << 5, /* any return branch */
 155         PERF_SAMPLE_BRANCH_IND_CALL     = 1U << 6, /* indirect calls */
 156
 157         PERF_SAMPLE_BRANCH_MAX          = 1U << 7, /* non-ABI */
 158 };
 159
 160 #define PERF_SAMPLE_BRANCH_PLM_ALL \
 161         (PERF_SAMPLE_BRANCH_USER|\
 162          PERF_SAMPLE_BRANCH_KERNEL|\
 163          PERF_SAMPLE_BRANCH_HV)
 164
 165 /*
 166  * The format of the data returned by read() on a perf event fd,
 167  * as specified by attr.read_format:
 168  *
 169  * struct read_format {
 170  *      { u64           value;
 171  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 172  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 173  *        { u64         id;           } && PERF_FORMAT_ID
 174  *      } && !PERF_FORMAT_GROUP
 175  *
 176  *      { u64           nr;
 177  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 178  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 179  *        { u64         value;
 180  *          { u64       id;           } && PERF_FORMAT_ID
 181  *        }             cntr[nr];
 182  *      } && PERF_FORMAT_GROUP
 183  * };
 184  */
 185 enum perf_event_read_format {
 186         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
 187         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
 188         PERF_FORMAT_ID                          = 1U << 2,
 189         PERF_FORMAT_GROUP                       = 1U << 3,
 190
 191         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
 192 };
 193
 194 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
 195 #define PERF_ATTR_SIZE_VER1     72      /* add: config2 */
 196 #define PERF_ATTR_SIZE_VER2     80      /* add: branch_sample_type */
 197
 198 /*
 199  * Hardware event_id to monitor via a performance monitoring event:
 200  */
 201 struct perf_event_attr {
 202
 203         /*
 204          * Major type: hardware/software/tracepoint/etc.
 205          */
 206         __u32                   type;
 207
 208         /*
 209          * Size of the attr structure, for fwd/bwd compat.
 210          */
 211         __u32                   size;
 212
 213         /*
 214          * Type specific configuration information.
 215          */
 216         __u64                   config;
 217
 218         union {
 219                 __u64           sample_period;
 220                 __u64           sample_freq;
 221         };
 222
 223         __u64                   sample_type;
 224         __u64                   read_format;
 225
 226         __u64                   disabled       :  1, /* off by default        */
 227                                 inherit        :  1, /* children inherit it   */
 228                                 pinned         :  1, /* must always be on PMU */
 229                                 exclusive      :  1, /* only group on PMU     */
 230                                 exclude_user   :  1, /* don't count user      */
 231                                 exclude_kernel :  1, /* ditto kernel          */
 232                                 exclude_hv     :  1, /* ditto hypervisor      */
 233                                 exclude_idle   :  1, /* don't count when idle */
 234                                 mmap           :  1, /* include mmap data     */
 235                                 comm           :  1, /* include comm data     */
 236                                 freq           :  1, /* use freq, not period  */
 237                                 inherit_stat   :  1, /* per task counts       */
 238                                 enable_on_exec :  1, /* next exec enables     */
 239                                 task           :  1, /* trace fork/exit       */
 240                                 watermark      :  1, /* wakeup_watermark      */
 241                                 /*
 242                                  * precise_ip:
 243                                  *
 244                                  *  0 - SAMPLE_IP can have arbitrary skid
 245                                  *  1 - SAMPLE_IP must have constant skid
 246                                  *  2 - SAMPLE_IP requested to have 0 skid
 247                                  *  3 - SAMPLE_IP must have 0 skid
 248                                  *
 249                                  *  See also PERF_RECORD_MISC_EXACT_IP
 250                                  */
 251                                 precise_ip     :  2, /* skid constraint       */
 252                                 mmap_data      :  1, /* non-exec mmap data    */
 253                                 sample_id_all  :  1, /* sample_type all events */
 254
 255                                 exclude_host   :  1, /* don't count in host   */
 256                                 exclude_guest  :  1, /* don't count in guest  */
 257
 258                                 __reserved_1   : 43;
 259
 260         union {
 261                 __u32           wakeup_events;    /* wakeup every n events */
 262                 __u32           wakeup_watermark; /* bytes before wakeup   */
 263         };
 264
 265         __u32                   bp_type;
 266         union {
 267                 __u64           bp_addr;
 268                 __u64           config1; /* extension of config */
 269         };
 270         union {
 271                 __u64           bp_len;
 272                 __u64           config2; /* extension of config1 */
 273         };
 274         __u64   branch_sample_type; /* enum branch_sample_type */
 275 };
 276
 277 #define perf_flags(attr)        (*(&(attr)->read_format + 1))
 278
 279 /*
 280  * Ioctls that can be done on a perf event fd:
 281  */
 282 #define PERF_EVENT_IOC_ENABLE           _IO ('$', 0)
 283 #define PERF_EVENT_IOC_DISABLE          _IO ('$', 1)
 284 #define PERF_EVENT_IOC_REFRESH          _IO ('$', 2)
 285 #define PERF_EVENT_IOC_RESET            _IO ('$', 3)
 286 #define PERF_EVENT_IOC_PERIOD           _IOW('$', 4, __u64)
 287 #define PERF_EVENT_IOC_SET_OUTPUT       _IO ('$', 5)
 288 #define PERF_EVENT_IOC_SET_FILTER       _IOW('$', 6, char *)
 289
 290 enum perf_event_ioc_flags {
 291         PERF_IOC_FLAG_GROUP             = 1U << 0,
 292 };
 293
 294 /*
 295  * Structure of the page that can be mapped via mmap
 296  */
 297 struct perf_event_mmap_page {
 298         __u32   version;                /* version number of this structure */
 299         __u32   compat_version;         /* lowest version this is compat with */
 300
 301         /*
 302          * Bits needed to read the hw events in user-space.
 303          *
 304          *   u32 seq, time_mult, time_shift, idx, width;
 305          *   u64 count, enabled, running;
 306          *   u64 cyc, time_offset;
 307          *   s64 pmc = 0;
 308          *
 309          *   do {
 310          *     seq = pc->lock;
 311          *     barrier()
 312          *
 313          *     enabled = pc->time_enabled;
 314          *     running = pc->time_running;
 315          *
 316          *     if (pc->cap_usr_time && enabled != running) {
 317          *       cyc = rdtsc();
 318          *       time_offset = pc->time_offset;
 319          *       time_mult   = pc->time_mult;
 320          *       time_shift  = pc->time_shift;
 321          *     }
 322          *
 323          *     idx = pc->index;
 324          *     count = pc->offset;
 325          *     if (pc->cap_usr_rdpmc && idx) {
 326          *       width = pc->pmc_width;
 327          *       pmc = rdpmc(idx - 1);
 328          *     }
 329          *
 330          *     barrier();
 331          *   } while (pc->lock != seq);
 332          *
 333          * NOTE: for obvious reason this only works on self-monitoring
 334          *       processes.
 335          */
 336         __u32   lock;                   /* seqlock for synchronization */
 337         __u32   index;                  /* hardware event identifier */
 338         __s64   offset;                 /* add to hardware event value */
 339         __u64   time_enabled;           /* time event active */
 340         __u64   time_running;           /* time event on cpu */
 341         union {
 342                 __u64   capabilities;
 343                 __u64   cap_usr_time  : 1,
 344                         cap_usr_rdpmc : 1,
 345                         cap_____res   : 62;
 346         };
 347
 348         /*
 349          * If cap_usr_rdpmc this field provides the bit-width of the value
 350          * read using the rdpmc() or equivalent instruction. This can be used
 351          * to sign extend the result like:
 352          *
 353          *   pmc <<= 64 - width;
 354          *   pmc >>= 64 - width; // signed shift right
 355          *   count += pmc;
 356          */
 357         __u16   pmc_width;
 358
 359         /*
 360          * If cap_usr_time the below fields can be used to compute the time
 361          * delta since time_enabled (in ns) using rdtsc or similar.
 362          *
 363          *   u64 quot, rem;
 364          *   u64 delta;
 365          *
 366          *   quot = (cyc >> time_shift);
 367          *   rem = cyc & ((1 << time_shift) - 1);
 368          *   delta = time_offset + quot * time_mult +
 369          *              ((rem * time_mult) >> time_shift);
 370          *
 371          * Where time_offset,time_mult,time_shift and cyc are read in the
 372          * seqcount loop described above. This delta can then be added to
 373          * enabled and possible running (if idx), improving the scaling:
 374          *
 375          *   enabled += delta;
 376          *   if (idx)
 377          *     running += delta;
 378          *
 379          *   quot = count / running;
 380          *   rem  = count % running;
 381          *   count = quot * enabled + (rem * enabled) / running;
 382          */
 383         __u16   time_shift;
 384         __u32   time_mult;
 385         __u64   time_offset;
 386
 387                 /*
 388                  * Hole for extension of the self monitor capabilities
 389                  */
 390
 391         __u64   __reserved[120];        /* align to 1k */
 392
 393         /*
 394          * Control data for the mmap() data buffer.
 395          *
 396          * User-space reading the @data_head value should issue an rmb(), on
 397          * SMP capable platforms, after reading this value -- see
 398          * perf_event_wakeup().
 399          *
 400          * When the mapping is PROT_WRITE the @data_tail value should be
 401          * written by userspace to reflect the last read data. In this case
 402          * the kernel will not over-write unread data.
 403          */
 404         __u64   data_head;              /* head in the data section */
 405         __u64   data_tail;              /* user-space written tail */
 406 };
 407
 408 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 409 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 410 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 411 #define PERF_RECORD_MISC_USER                   (2 << 0)
 412 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 413 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 414 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 415
 416 /*
 417  * Indicates that the content of PERF_SAMPLE_IP points to
 418  * the actual instruction that triggered the event. See also
 419  * perf_event_attr::precise_ip.
 420  */
 421 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 422 /*
 423  * Reserve the last bit to indicate some extended misc field
 424  */
 425 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 426
 427 struct perf_event_header {
 428         __u32   type;
 429         __u16   misc;
 430         __u16   size;
 431 };
 432
 433 enum perf_event_type {
 434
 435         /*
 436          * If perf_event_attr.sample_id_all is set then all event types will
 437          * have the sample_type selected fields related to where/when
 438          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
 439          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
 440          * the perf_event_header and the fields already present for the existing
 441          * fields, i.e. at the end of the payload. That way a newer perf.data
 442          * file will be supported by older perf tools, with these new optional
 443          * fields being ignored.
 444          *
 445          * The MMAP events record the PROT_EXEC mappings so that we can
 446          * correlate userspace IPs to code. They have the following structure:
 447          *
 448          * struct {
 449          *      struct perf_event_header        header;
 450          *
 451          *      u32                             pid, tid;
 452          *      u64                             addr;
 453          *      u64                             len;
 454          *      u64                             pgoff;
 455          *      char                            filename[];
 456          * };
 457          */
 458         PERF_RECORD_MMAP                        = 1,
 459
 460         /*
 461          * struct {
 462          *      struct perf_event_header        header;
 463          *      u64                             id;
 464          *      u64                             lost;
 465          * };
 466          */
 467         PERF_RECORD_LOST                        = 2,
 468
 469         /*
 470          * struct {
 471          *      struct perf_event_header        header;
 472          *
 473          *      u32                             pid, tid;
 474          *      char                            comm[];
 475          * };
 476          */
 477         PERF_RECORD_COMM                        = 3,
 478
 479         /*
 480          * struct {
 481          *      struct perf_event_header        header;
 482          *      u32                             pid, ppid;
 483          *      u32                             tid, ptid;
 484          *      u64                             time;
 485          * };
 486          */
 487         PERF_RECORD_EXIT                        = 4,
 488
 489         /*
 490          * struct {
 491          *      struct perf_event_header        header;
 492          *      u64                             time;
 493          *      u64                             id;
 494          *      u64                             stream_id;
 495          * };
 496          */
 497         PERF_RECORD_THROTTLE                    = 5,
 498         PERF_RECORD_UNTHROTTLE                  = 6,
 499
 500         /*
 501          * struct {
 502          *      struct perf_event_header        header;
 503          *      u32                             pid, ppid;
 504          *      u32                             tid, ptid;
 505          *      u64                             time;
 506          * };
 507          */
 508         PERF_RECORD_FORK                        = 7,
 509
 510         /*
 511          * struct {
 512          *      struct perf_event_header        header;
 513          *      u32                             pid, tid;
 514          *
 515          *      struct read_format              values;
 516          * };
 517          */
 518         PERF_RECORD_READ                        = 8,
 519
 520         /*
 521          * struct {
 522          *      struct perf_event_header        header;
 523          *
 524          *      { u64                   ip;       } && PERF_SAMPLE_IP
 525          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 526          *      { u64                   time;     } && PERF_SAMPLE_TIME
 527          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 528          *      { u64                   id;       } && PERF_SAMPLE_ID
 529          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 530          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 531          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 532          *
 533          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 534          *
 535          *      { u64                   nr,
 536          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 537          *
 538          *      #
 539          *      # The RAW record below is opaque data wrt the ABI
 540          *      #
 541          *      # That is, the ABI doesn't make any promises wrt to
 542          *      # the stability of its content, it may vary depending
 543          *      # on event, hardware, kernel version and phase of
 544          *      # the moon.
 545          *      #
 546          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 547          *      #
 548          *
 549          *      { u32                   size;
 550          *        char                  data[size];}&& PERF_SAMPLE_RAW
 551          *
 552          *      { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
 553          * };
 554          */
 555         PERF_RECORD_SAMPLE                      = 9,
 556
 557         PERF_RECORD_MAX,                        /* non-ABI */
 558 };
 559
 560 #define PERF_MAX_STACK_DEPTH            127
 561
 562 enum perf_callchain_context {
 563         PERF_CONTEXT_HV                 = (__u64)-32,
 564         PERF_CONTEXT_KERNEL             = (__u64)-128,
 565         PERF_CONTEXT_USER               = (__u64)-512,
 566
 567         PERF_CONTEXT_GUEST              = (__u64)-2048,
 568         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 569         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 570
 571         PERF_CONTEXT_MAX                = (__u64)-4095,
 572 };
 573
 574 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 575 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 576 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
 577
 578 #ifdef __KERNEL__
 579 /*
 580  * Kernel-internal data types and definitions:
 581  */
 582
 583 #ifdef CONFIG_PERF_EVENTS
 584 # include <linux/cgroup.h>
 585 # include <asm/perf_event.h>
 586 # include <asm/local64.h>
 587 #endif
 588
 589 struct perf_guest_info_callbacks {
 590         int                             (*is_in_guest)(void);
 591         int                             (*is_user_mode)(void);
 592         unsigned long                   (*get_guest_ip)(void);
 593 };
 594
 595 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 596 #include <asm/hw_breakpoint.h>
 597 #endif
 598
 599 #include <linux/list.h>
 600 #include <linux/mutex.h>
 601 #include <linux/rculist.h>
 602 #include <linux/rcupdate.h>
 603 #include <linux/spinlock.h>
 604 #include <linux/hrtimer.h>
 605 #include <linux/fs.h>
 606 #include <linux/pid_namespace.h>
 607 #include <linux/workqueue.h>
 608 #include <linux/ftrace.h>
 609 #include <linux/cpu.h>
 610 #include <linux/irq_work.h>
 611 #include <linux/static_key.h>
 612 #include <linux/atomic.h>
 613 #include <linux/sysfs.h>
 614 #include <asm/local.h>
 615
 616 struct perf_callchain_entry {
 617         __u64                           nr;
 618         __u64                           ip[PERF_MAX_STACK_DEPTH];
 619 };
 620
 621 struct perf_raw_record {
 622         u32                             size;
 623         void                            *data;
 624 };
 625
 626 /*
 627  * single taken branch record layout:
 628  *
 629  *      from: source instruction (may not always be a branch insn)
 630  *        to: branch target
 631  *   mispred: branch target was mispredicted
 632  * predicted: branch target was predicted
 633  *
 634  * support for mispred, predicted is optional. In case it
 635  * is not supported mispred = predicted = 0.
 636  */
 637 struct perf_branch_entry {
 638         __u64   from;
 639         __u64   to;
 640         __u64   mispred:1,  /* target mispredicted */
 641                 predicted:1,/* target predicted */
 642                 reserved:62;
 643 };
 644
 645 /*
 646  * branch stack layout:
 647  *  nr: number of taken branches stored in entries[]
 648  *
 649  * Note that nr can vary from sample to sample
 650  * branches (to, from) are stored from most recent
 651  * to least recent, i.e., entries[0] contains the most
 652  * recent branch.
 653  */
 654 struct perf_branch_stack {
 655         __u64                           nr;
 656         struct perf_branch_entry        entries[0];
 657 };
 658
 659 struct task_struct;
 660
 661 /*
 662  * extra PMU register associated with an event
 663  */
 664 struct hw_perf_event_extra {
 665         u64             config; /* register value */
 666         unsigned int    reg;    /* register address or index */
 667         int             alloc;  /* extra register already allocated */
 668         int             idx;    /* index in shared_regs->regs[] */
 669 };
 670
 671 /**
 672  * struct hw_perf_event - performance event hardware details:
 673  */
 674 struct hw_perf_event {
 675 #ifdef CONFIG_PERF_EVENTS
 676         union {
 677                 struct { /* hardware */
 678                         u64             config;
 679                         u64             last_tag;
 680                         unsigned long   config_base;
 681                         unsigned long   event_base;
 682                         int             event_base_rdpmc;
 683                         int             idx;
 684                         int             last_cpu;
 685
 686                         struct hw_perf_event_extra extra_reg;
 687                         struct hw_perf_event_extra branch_reg;
 688                 };
 689                 struct { /* software */
 690                         struct hrtimer  hrtimer;
 691                 };
 692 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 693                 struct { /* breakpoint */
 694                         struct arch_hw_breakpoint       info;
 695                         struct list_head                bp_list;
 696                         /*
 697                          * Crufty hack to avoid the chicken and egg
 698                          * problem hw_breakpoint has with context
 699                          * creation and event initalization.
 700                          */
 701                         struct task_struct              *bp_target;
 702                 };
 703 #endif
 704         };
 705         int                             state;
 706         local64_t                       prev_count;
 707         u64                             sample_period;
 708         u64                             last_period;
 709         local64_t                       period_left;
 710         u64                             interrupts_seq;
 711         u64                             interrupts;
 712
 713         u64                             freq_time_stamp;
 714         u64                             freq_count_stamp;
 715 #endif
 716 };
 717
 718 /*
 719  * hw_perf_event::state flags
 720  */
 721 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 722 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 723 #define PERF_HES_ARCH           0x04
 724
 725 struct perf_event;
 726
 727 /*
 728  * Common implementation detail of pmu::{start,commit,cancel}_txn
 729  */
 730 #define PERF_EVENT_TXN 0x1
 731
 732 /**
 733  * struct pmu - generic performance monitoring unit
 734  */
 735 struct pmu {
 736         struct list_head                entry;
 737
 738         struct device                   *dev;
 739         const struct attribute_group    **attr_groups;
 740         char                            *name;
 741         int                             type;
 742
 743         int * __percpu                  pmu_disable_count;
 744         struct perf_cpu_context * __percpu pmu_cpu_context;
 745         int                             task_ctx_nr;
 746
 747         /*
 748          * Fully disable/enable this PMU, can be used to protect from the PMI
 749          * as well as for lazy/batch writing of the MSRs.
 750          */
 751         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 752         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 753
 754         /*
 755          * Try and initialize the event for this PMU.
 756          * Should return -ENOENT when the @event doesn't match this PMU.
 757          */
 758         int (*event_init)               (struct perf_event *event);
 759
 760 #define PERF_EF_START   0x01            /* start the counter when adding    */
 761 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 762 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 763
 764         /*
 765          * Adds/Removes a counter to/from the PMU, can be done inside
 766          * a transaction, see the ->*_txn() methods.
 767          */
 768         int  (*add)                     (struct perf_event *event, int flags);
 769         void (*del)                     (struct perf_event *event, int flags);
 770
 771         /*
 772          * Starts/Stops a counter present on the PMU. The PMI handler
 773          * should stop the counter when perf_event_overflow() returns
 774          * !0. ->start() will be used to continue.
 775          */
 776         void (*start)                   (struct perf_event *event, int flags);
 777         void (*stop)                    (struct perf_event *event, int flags);
 778
 779         /*
 780          * Updates the counter value of the event.
 781          */
 782         void (*read)                    (struct perf_event *event);
 783
 784         /*
 785          * Group events scheduling is treated as a transaction, add
 786          * group events as a whole and perform one schedulability test.
 787          * If the test fails, roll back the whole group
 788          *
 789          * Start the transaction, after this ->add() doesn't need to
 790          * do schedulability tests.
 791          */
 792         void (*start_txn)               (struct pmu *pmu); /* optional */
 793         /*
 794          * If ->start_txn() disabled the ->add() schedulability test
 795          * then ->commit_txn() is required to perform one. On success
 796          * the transaction is closed. On error the transaction is kept
 797          * open until ->cancel_txn() is called.
 798          */
 799         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 800         /*
 801          * Will cancel the transaction, assumes ->del() is called
 802          * for each successful ->add() during the transaction.
 803          */
 804         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 805
 806         /*
 807          * Will return the value for perf_event_mmap_page::index for this event,
 808          * if no implementation is provided it will default to: event->hw.idx + 1.
 809          */
 810         int (*event_idx)                (struct perf_event *event); /*optional */
 811
 812         /*
 813          * flush branch stack on context-switches (needed in cpu-wide mode)
 814          */
 815         void (*flush_branch_stack)      (void);
 816 };
 817
 818 /**
 819  * enum perf_event_active_state - the states of a event
 820  */
 821 enum perf_event_active_state {
 822         PERF_EVENT_STATE_ERROR          = -2,
 823         PERF_EVENT_STATE_OFF            = -1,
 824         PERF_EVENT_STATE_INACTIVE       =  0,
 825         PERF_EVENT_STATE_ACTIVE         =  1,
 826 };
 827
 828 struct file;
 829 struct perf_sample_data;
 830
 831 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 832                                         struct perf_sample_data *,
 833                                         struct pt_regs *regs);
 834
 835 enum perf_group_flag {
 836         PERF_GROUP_SOFTWARE             = 0x1,
 837 };
 838
 839 #define SWEVENT_HLIST_BITS              8
 840 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 841
 842 struct swevent_hlist {
 843         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 844         struct rcu_head                 rcu_head;
 845 };
 846
 847 #define PERF_ATTACH_CONTEXT     0x01
 848 #define PERF_ATTACH_GROUP       0x02
 849 #define PERF_ATTACH_TASK        0x04
 850
 851 #ifdef CONFIG_CGROUP_PERF
 852 /*
 853  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 854  * This is a per-cpu dynamically allocated data structure.
 855  */
 856 struct perf_cgroup_info {
 857         u64                             time;
 858         u64                             timestamp;
 859 };
 860
 861 struct perf_cgroup {
 862         struct                          cgroup_subsys_state css;
 863         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
 864 };
 865 #endif
 866
 867 struct ring_buffer;
 868
 869 /**
 870  * struct perf_event - performance event kernel representation:
 871  */
 872 struct perf_event {
 873 #ifdef CONFIG_PERF_EVENTS
 874         struct list_head                group_entry;
 875         struct list_head                event_entry;
 876         struct list_head                sibling_list;
 877         struct hlist_node               hlist_entry;
 878         int                             nr_siblings;
 879         int                             group_flags;
 880         struct perf_event               *group_leader;
 881         struct pmu                      *pmu;
 882
 883         enum perf_event_active_state    state;
 884         unsigned int                    attach_state;
 885         local64_t                       count;
 886         atomic64_t                      child_count;
 887
 888         /*
 889          * These are the total time in nanoseconds that the event
 890          * has been enabled (i.e. eligible to run, and the task has
 891          * been scheduled in, if this is a per-task event)
 892          * and running (scheduled onto the CPU), respectively.
 893          *
 894          * They are computed from tstamp_enabled, tstamp_running and
 895          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 896          */
 897         u64                             total_time_enabled;
 898         u64                             total_time_running;
 899
 900         /*
 901          * These are timestamps used for computing total_time_enabled
 902          * and total_time_running when the event is in INACTIVE or
 903          * ACTIVE state, measured in nanoseconds from an arbitrary point
 904          * in time.
 905          * tstamp_enabled: the notional time when the event was enabled
 906          * tstamp_running: the notional time when the event was scheduled on
 907          * tstamp_stopped: in INACTIVE state, the notional time when the
 908          *      event was scheduled off.
 909          */
 910         u64                             tstamp_enabled;
 911         u64                             tstamp_running;
 912         u64                             tstamp_stopped;
 913
 914         /*
 915          * timestamp shadows the actual context timing but it can
 916          * be safely used in NMI interrupt context. It reflects the
 917          * context time as it was when the event was last scheduled in.
 918          *
 919          * ctx_time already accounts for ctx->timestamp. Therefore to
 920          * compute ctx_time for a sample, simply add perf_clock().
 921          */
 922         u64                             shadow_ctx_time;
 923
 924         struct perf_event_attr          attr;
 925         u16                             header_size;
 926         u16                             id_header_size;
 927         u16                             read_size;
 928         struct hw_perf_event            hw;
 929
 930         struct perf_event_context       *ctx;
 931         atomic_long_t                   refcount;
 932
 933         /*
 934          * These accumulate total time (in nanoseconds) that children
 935          * events have been enabled and running, respectively.
 936          */
 937         atomic64_t                      child_total_time_enabled;
 938         atomic64_t                      child_total_time_running;
 939
 940         /*
 941          * Protect attach/detach and child_list:
 942          */
 943         struct mutex                    child_mutex;
 944         struct list_head                child_list;
 945         struct perf_event               *parent;
 946
 947         int                             oncpu;
 948         int                             cpu;
 949
 950         struct list_head                owner_entry;
 951         struct task_struct              *owner;
 952
 953         /* mmap bits */
 954         struct mutex                    mmap_mutex;
 955         atomic_t                        mmap_count;
 956         int                             mmap_locked;
 957         struct user_struct              *mmap_user;
 958         struct ring_buffer              *rb;
 959         struct list_head                rb_entry;
 960
 961         /* poll related */
 962         wait_queue_head_t               waitq;
 963         struct fasync_struct            *fasync;
 964
 965         /* delayed work for NMIs and such */
 966         int                             pending_wakeup;
 967         int                             pending_kill;
 968         int                             pending_disable;
 969         struct irq_work                 pending;
 970
 971         atomic_t                        event_limit;
 972
 973         void (*destroy)(struct perf_event *);
 974         struct rcu_head                 rcu_head;
 975
 976         struct pid_namespace            *ns;
 977         u64                             id;
 978
 979         perf_overflow_handler_t         overflow_handler;
 980         void                            *overflow_handler_context;
 981
 982 #ifdef CONFIG_EVENT_TRACING
 983         struct ftrace_event_call        *tp_event;
 984         struct event_filter             *filter;
 985 #ifdef CONFIG_FUNCTION_TRACER
 986         struct ftrace_ops               ftrace_ops;
 987 #endif
 988 #endif
 989
 990 #ifdef CONFIG_CGROUP_PERF
 991         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 992         int                             cgrp_defer_enabled;
 993 #endif
 994
 995 #endif /* CONFIG_PERF_EVENTS */
 996 };
 997
 998 enum perf_event_context_type {
 999         task_context,
1000         cpu_context,
1001 };
1002
1003 /**
1004  * struct perf_event_context - event context structure
1005  *
1006  * Used as a container for task events and CPU events as well:
1007  */
1008 struct perf_event_context {
1009         struct pmu                      *pmu;
1010         enum perf_event_context_type    type;
1011         /*
1012          * Protect the states of the events in the list,
1013          * nr_active, and the list:
1014          */
1015         raw_spinlock_t                  lock;
1016         /*
1017          * Protect the list of events.  Locking either mutex or lock
1018          * is sufficient to ensure the list doesn't change; to change
1019          * the list you need to lock both the mutex and the spinlock.
1020          */
1021         struct mutex                    mutex;
1022
1023         struct list_head                pinned_groups;
1024         struct list_head                flexible_groups;
1025         struct list_head                event_list;
1026         int                             nr_events;
1027         int                             nr_active;
1028         int                             is_active;
1029         int                             nr_stat;
1030         int                             nr_freq;
1031         int                             rotate_disable;
1032         atomic_t                        refcount;
1033         struct task_struct              *task;
1034
1035         /*
1036          * Context clock, runs when context enabled.
1037          */
1038         u64                             time;
1039         u64                             timestamp;
1040
1041         /*
1042          * These fields let us detect when two contexts have both
1043          * been cloned (inherited) from a common ancestor.
1044          */
1045         struct perf_event_context       *parent_ctx;
1046         u64                             parent_gen;
1047         u64                             generation;
1048         int                             pin_count;
1049         int                             nr_cgroups;      /* cgroup evts */
1050         int                             nr_branch_stack; /* branch_stack evt */
1051         struct rcu_head                 rcu_head;
1052 };
1053
1054 /*
1055  * Number of contexts where an event can trigger:
1056  *      task, softirq, hardirq, nmi.
1057  */
1058 #define PERF_NR_CONTEXTS        4
1059
1060 /**
1061  * struct perf_event_cpu_context - per cpu event context structure
1062  */
1063 struct perf_cpu_context {
1064         struct perf_event_context       ctx;
1065         struct perf_event_context       *task_ctx;
1066         int                             active_oncpu;
1067         int                             exclusive;
1068         struct list_head                rotation_list;
1069         int                             jiffies_interval;
1070         struct pmu                      *active_pmu;
1071         struct perf_cgroup              *cgrp;
1072 };
1073
1074 struct perf_output_handle {
1075         struct perf_event               *event;
1076         struct ring_buffer              *rb;
1077         unsigned long                   wakeup;
1078         unsigned long                   size;
1079         void                            *addr;
1080         int                             page;
1081 };
1082
1083 #ifdef CONFIG_PERF_EVENTS
1084
1085 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
1086 extern void perf_pmu_unregister(struct pmu *pmu);
1087
1088 extern int perf_num_counters(void);
1089 extern const char *perf_pmu_name(void);
1090 extern void __perf_event_task_sched_in(struct task_struct *prev,
1091                                        struct task_struct *task);
1092 extern void __perf_event_task_sched_out(struct task_struct *prev,
1093                                         struct task_struct *next);
1094 extern int perf_event_init_task(struct task_struct *child);
1095 extern void perf_event_exit_task(struct task_struct *child);
1096 extern void perf_event_free_task(struct task_struct *task);
1097 extern void perf_event_delayed_put(struct task_struct *task);
1098 extern void perf_event_print_debug(void);
1099 extern void perf_pmu_disable(struct pmu *pmu);
1100 extern void perf_pmu_enable(struct pmu *pmu);
1101 extern int perf_event_task_disable(void);
1102 extern int perf_event_task_enable(void);
1103 extern int perf_event_refresh(struct perf_event *event, int refresh);
1104 extern void perf_event_update_userpage(struct perf_event *event);
1105 extern int perf_event_release_kernel(struct perf_event *event);
1106 extern struct perf_event *
1107 perf_event_create_kernel_counter(struct perf_event_attr *attr,
1108                                 int cpu,
1109                                 struct task_struct *task,
1110                                 perf_overflow_handler_t callback,
1111                                 void *context);
1112 extern void perf_pmu_migrate_context(struct pmu *pmu,
1113                                 int src_cpu, int dst_cpu);
1114 extern u64 perf_event_read_value(struct perf_event *event,
1115                                  u64 *enabled, u64 *running);
1116
1117
1118 struct perf_sample_data {
1119         u64                             type;
1120
1121         u64                             ip;
1122         struct {
1123                 u32     pid;
1124                 u32     tid;
1125         }                               tid_entry;
1126         u64                             time;
1127         u64                             addr;
1128         u64                             id;
1129         u64                             stream_id;
1130         struct {
1131                 u32     cpu;
1132                 u32     reserved;
1133         }                               cpu_entry;
1134         u64                             period;
1135         struct perf_callchain_entry     *callchain;
1136         struct perf_raw_record          *raw;
1137         struct perf_branch_stack        *br_stack;
1138 };
1139
1140 static inline void perf_sample_data_init(struct perf_sample_data *data,
1141                                          u64 addr, u64 period)
1142 {
1143         /* remaining struct members initialized in perf_prepare_sample() */
1144         data->addr = addr;
1145         data->raw  = NULL;
1146         data->br_stack = NULL;
1147         data->period    = period;
1148 }
1149
1150 extern void perf_output_sample(struct perf_output_handle *handle,
1151                                struct perf_event_header *header,
1152                                struct perf_sample_data *data,
1153                                struct perf_event *event);
1154 extern void perf_prepare_sample(struct perf_event_header *header,
1155                                 struct perf_sample_data *data,
1156                                 struct perf_event *event,
1157                                 struct pt_regs *regs);
1158
1159 extern int perf_event_overflow(struct perf_event *event,
1160                                  struct perf_sample_data *data,
1161                                  struct pt_regs *regs);
1162
1163 static inline bool is_sampling_event(struct perf_event *event)
1164 {
1165         return event->attr.sample_period != 0;
1166 }
1167
1168 /*
1169  * Return 1 for a software event, 0 for a hardware event
1170  */
1171 static inline int is_software_event(struct perf_event *event)
1172 {
1173         return event->pmu->task_ctx_nr == perf_sw_context;
1174 }
1175
1176 extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1177
1178 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1179
1180 #ifndef perf_arch_fetch_caller_regs
1181 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1182 #endif
1183
1184 /*
1185  * Take a snapshot of the regs. Skip ip and frame pointer to
1186  * the nth caller. We only need a few of the regs:
1187  * - ip for PERF_SAMPLE_IP
1188  * - cs for user_mode() tests
1189  * - bp for callchains
1190  * - eflags, for future purposes, just in case
1191  */
1192 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1193 {
1194         memset(regs, 0, sizeof(*regs));
1195
1196         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1197 }
1198
1199 static __always_inline void
1200 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1201 {
1202         struct pt_regs hot_regs;
1203
1204         if (static_key_false(&perf_swevent_enabled[event_id])) {
1205                 if (!regs) {
1206                         perf_fetch_caller_regs(&hot_regs);
1207                         regs = &hot_regs;
1208                 }
1209                 __perf_sw_event(event_id, nr, regs, addr);
1210         }
1211 }
1212
1213 extern struct static_key_deferred perf_sched_events;
1214
1215 static inline void perf_event_task_sched_in(struct task_struct *prev,
1216                                             struct task_struct *task)
1217 {
1218         if (static_key_false(&perf_sched_events.key))
1219                 __perf_event_task_sched_in(prev, task);
1220 }
1221
1222 static inline void perf_event_task_sched_out(struct task_struct *prev,
1223                                              struct task_struct *next)
1224 {
1225         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1226
1227         if (static_key_false(&perf_sched_events.key))
1228                 __perf_event_task_sched_out(prev, next);
1229 }
1230
1231 extern void perf_event_mmap(struct vm_area_struct *vma);
1232 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1233 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1234 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1235
1236 extern void perf_event_comm(struct task_struct *tsk);
1237 extern void perf_event_fork(struct task_struct *tsk);
1238
1239 /* Callchains */
1240 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1241
1242 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1243 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1244
1245 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1246 {
1247         if (entry->nr < PERF_MAX_STACK_DEPTH)
1248                 entry->ip[entry->nr++] = ip;
1249 }
1250
1251 extern int sysctl_perf_event_paranoid;
1252 extern int sysctl_perf_event_mlock;
1253 extern int sysctl_perf_event_sample_rate;
1254
1255 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1256                 void __user *buffer, size_t *lenp,
1257                 loff_t *ppos);
1258
1259 static inline bool perf_paranoid_tracepoint_raw(void)
1260 {
1261         return sysctl_perf_event_paranoid > -1;
1262 }
1263
1264 static inline bool perf_paranoid_cpu(void)
1265 {
1266         return sysctl_perf_event_paranoid > 0;
1267 }
1268
1269 static inline bool perf_paranoid_kernel(void)
1270 {
1271         return sysctl_perf_event_paranoid > 1;
1272 }
1273
1274 extern void perf_event_init(void);
1275 extern void perf_tp_event(u64 addr, u64 count, void *record,
1276                           int entry_size, struct pt_regs *regs,
1277                           struct hlist_head *head, int rctx,
1278                           struct task_struct *task);
1279 extern void perf_bp_event(struct perf_event *event, void *data);
1280
1281 #ifndef perf_misc_flags
1282 # define perf_misc_flags(regs) \
1283                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1284 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1285 #endif
1286
1287 static inline bool has_branch_stack(struct perf_event *event)
1288 {
1289         return event->attr.sample_type & PERF_SAMPLE_BRANCH_STACK;
1290 }
1291
1292 extern int perf_output_begin(struct perf_output_handle *handle,
1293                              struct perf_event *event, unsigned int size);
1294 extern void perf_output_end(struct perf_output_handle *handle);
1295 extern void perf_output_copy(struct perf_output_handle *handle,
1296                              const void *buf, unsigned int len);
1297 extern int perf_swevent_get_recursion_context(void);
1298 extern void perf_swevent_put_recursion_context(int rctx);
1299 extern void perf_event_enable(struct perf_event *event);
1300 extern void perf_event_disable(struct perf_event *event);
1301 extern int __perf_event_disable(void *info);
1302 extern void perf_event_task_tick(void);
1303 #else
1304 static inline void
1305 perf_event_task_sched_in(struct task_struct *prev,
1306                          struct task_struct *task)                      { }
1307 static inline void
1308 perf_event_task_sched_out(struct task_struct *prev,
1309                           struct task_struct *next)                     { }
1310 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1311 static inline void perf_event_exit_task(struct task_struct *child)      { }
1312 static inline void perf_event_free_task(struct task_struct *task)       { }
1313 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1314 static inline void perf_event_print_debug(void)                         { }
1315 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1316 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1317 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1318 {
1319         return -EINVAL;
1320 }
1321
1322 static inline void
1323 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1324 static inline void
1325 perf_bp_event(struct perf_event *event, void *data)                     { }
1326
1327 static inline int perf_register_guest_info_callbacks
1328 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1329 static inline int perf_unregister_guest_info_callbacks
1330 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1331
1332 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1333 static inline void perf_event_comm(struct task_struct *tsk)             { }
1334 static inline void perf_event_fork(struct task_struct *tsk)             { }
1335 static inline void perf_event_init(void)                                { }
1336 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1337 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1338 static inline void perf_event_enable(struct perf_event *event)          { }
1339 static inline void perf_event_disable(struct perf_event *event)         { }
1340 static inline int __perf_event_disable(void *info)                      { return -1; }
1341 static inline void perf_event_task_tick(void)                           { }
1342 #endif
1343
1344 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1345
1346 /*
1347  * This has to have a higher priority than migration_notifier in sched.c.
1348  */
1349 #define perf_cpu_notifier(fn)                                           \
1350 do {                                                                    \
1351         static struct notifier_block fn##_nb __cpuinitdata =            \
1352                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1353         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1354                 (void *)(unsigned long)smp_processor_id());             \
1355         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1356                 (void *)(unsigned long)smp_processor_id());             \
1357         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1358                 (void *)(unsigned long)smp_processor_id());             \
1359         register_cpu_notifier(&fn##_nb);                                \
1360 } while (0)
1361
1362
1363 #define PMU_FORMAT_ATTR(_name, _format)                                 \
1364 static ssize_t                                                          \
1365 _name##_show(struct device *dev,                                        \
1366                                struct device_attribute *attr,           \
1367                                char *page)                              \
1368 {                                                                       \
1369         BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);                     \
1370         return sprintf(page, _format "\n");                             \
1371 }                                                                       \
1372                                                                         \
1373 static struct device_attribute format_attr_##_name = __ATTR_RO(_name)
1374
1375 #endif /* __KERNEL__ */
1376 #endif /* _LINUX_PERF_EVENT_H */