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