include/linux/perf_event.h

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