include/linux/perf_event.h

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