include/linux/perf_event.h

   1 /*
   2  * Performance events:
   3  *
   4  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
   5  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   7  *
   8  * Data type definitions, declarations, prototypes.
   9  *
  10  *    Started by: Thomas Gleixner and Ingo Molnar
  11  *
  12  * For licencing details see kernel-base/COPYING
  13  */
  14 #ifndef _LINUX_PERF_EVENT_H
  15 #define _LINUX_PERF_EVENT_H
  16
  17 #include <linux/types.h>
  18 #include <linux/ioctl.h>
  19 #include <asm/byteorder.h>
  20
  21 /*
  22  * User-space ABI bits:
  23  */
  24
  25 /*
  26  * attr.type
  27  */
  28 enum perf_type_id {
  29         PERF_TYPE_HARDWARE                      = 0,
  30         PERF_TYPE_SOFTWARE                      = 1,
  31         PERF_TYPE_TRACEPOINT                    = 2,
  32         PERF_TYPE_HW_CACHE                      = 3,
  33         PERF_TYPE_RAW                           = 4,
  34         PERF_TYPE_BREAKPOINT                    = 5,
  35
  36         PERF_TYPE_MAX,                          /* non-ABI */
  37 };
  38
  39 /*
  40  * Generalized performance event event_id types, used by the
  41  * attr.event_id parameter of the sys_perf_event_open()
  42  * syscall:
  43  */
  44 enum perf_hw_id {
  45         /*
  46          * Common hardware events, generalized by the kernel:
  47          */
  48         PERF_COUNT_HW_CPU_CYCLES                = 0,
  49         PERF_COUNT_HW_INSTRUCTIONS              = 1,
  50         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
  51         PERF_COUNT_HW_CACHE_MISSES              = 3,
  52         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
  53         PERF_COUNT_HW_BRANCH_MISSES             = 5,
  54         PERF_COUNT_HW_BUS_CYCLES                = 6,
  55         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
  56         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
  57
  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 rmb(), on
 304          * SMP capable platforms, after reading this value -- see
 305          * perf_event_wakeup().
 306          *
 307          * When the mapping is PROT_WRITE the @data_tail value should be
 308          * written by userspace to reflect the last read data. In this case
 309          * the kernel will not over-write unread data.
 310          */
 311         __u64   data_head;              /* head in the data section */
 312         __u64   data_tail;              /* user-space written tail */
 313 };
 314
 315 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 316 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 317 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 318 #define PERF_RECORD_MISC_USER                   (2 << 0)
 319 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 320 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 321 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 322
 323 /*
 324  * Indicates that the content of PERF_SAMPLE_IP points to
 325  * the actual instruction that triggered the event. See also
 326  * perf_event_attr::precise_ip.
 327  */
 328 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 329 /*
 330  * Reserve the last bit to indicate some extended misc field
 331  */
 332 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 333
 334 struct perf_event_header {
 335         __u32   type;
 336         __u16   misc;
 337         __u16   size;
 338 };
 339
 340 enum perf_event_type {
 341
 342         /*
 343          * If perf_event_attr.sample_id_all is set then all event types will
 344          * have the sample_type selected fields related to where/when
 345          * (identity) an event took place (TID, TIME, ID, CPU, STREAM_ID)
 346          * described in PERF_RECORD_SAMPLE below, it will be stashed just after
 347          * the perf_event_header and the fields already present for the existing
 348          * fields, i.e. at the end of the payload. That way a newer perf.data
 349          * file will be supported by older perf tools, with these new optional
 350          * fields being ignored.
 351          *
 352          * The MMAP events record the PROT_EXEC mappings so that we can
 353          * correlate userspace IPs to code. They have the following structure:
 354          *
 355          * struct {
 356          *      struct perf_event_header        header;
 357          *
 358          *      u32                             pid, tid;
 359          *      u64                             addr;
 360          *      u64                             len;
 361          *      u64                             pgoff;
 362          *      char                            filename[];
 363          * };
 364          */
 365         PERF_RECORD_MMAP                        = 1,
 366
 367         /*
 368          * struct {
 369          *      struct perf_event_header        header;
 370          *      u64                             id;
 371          *      u64                             lost;
 372          * };
 373          */
 374         PERF_RECORD_LOST                        = 2,
 375
 376         /*
 377          * struct {
 378          *      struct perf_event_header        header;
 379          *
 380          *      u32                             pid, tid;
 381          *      char                            comm[];
 382          * };
 383          */
 384         PERF_RECORD_COMM                        = 3,
 385
 386         /*
 387          * struct {
 388          *      struct perf_event_header        header;
 389          *      u32                             pid, ppid;
 390          *      u32                             tid, ptid;
 391          *      u64                             time;
 392          * };
 393          */
 394         PERF_RECORD_EXIT                        = 4,
 395
 396         /*
 397          * struct {
 398          *      struct perf_event_header        header;
 399          *      u64                             time;
 400          *      u64                             id;
 401          *      u64                             stream_id;
 402          * };
 403          */
 404         PERF_RECORD_THROTTLE                    = 5,
 405         PERF_RECORD_UNTHROTTLE                  = 6,
 406
 407         /*
 408          * struct {
 409          *      struct perf_event_header        header;
 410          *      u32                             pid, ppid;
 411          *      u32                             tid, ptid;
 412          *      u64                             time;
 413          * };
 414          */
 415         PERF_RECORD_FORK                        = 7,
 416
 417         /*
 418          * struct {
 419          *      struct perf_event_header        header;
 420          *      u32                             pid, tid;
 421          *
 422          *      struct read_format              values;
 423          * };
 424          */
 425         PERF_RECORD_READ                        = 8,
 426
 427         /*
 428          * struct {
 429          *      struct perf_event_header        header;
 430          *
 431          *      { u64                   ip;       } && PERF_SAMPLE_IP
 432          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 433          *      { u64                   time;     } && PERF_SAMPLE_TIME
 434          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 435          *      { u64                   id;       } && PERF_SAMPLE_ID
 436          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 437          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 438          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 439          *
 440          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 441          *
 442          *      { u64                   nr,
 443          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 444          *
 445          *      #
 446          *      # The RAW record below is opaque data wrt the ABI
 447          *      #
 448          *      # That is, the ABI doesn't make any promises wrt to
 449          *      # the stability of its content, it may vary depending
 450          *      # on event, hardware, kernel version and phase of
 451          *      # the moon.
 452          *      #
 453          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 454          *      #
 455          *
 456          *      { u32                   size;
 457          *        char                  data[size];}&& PERF_SAMPLE_RAW
 458          * };
 459          */
 460         PERF_RECORD_SAMPLE                      = 9,
 461
 462         PERF_RECORD_MAX,                        /* non-ABI */
 463 };
 464
 465 enum perf_callchain_context {
 466         PERF_CONTEXT_HV                 = (__u64)-32,
 467         PERF_CONTEXT_KERNEL             = (__u64)-128,
 468         PERF_CONTEXT_USER               = (__u64)-512,
 469
 470         PERF_CONTEXT_GUEST              = (__u64)-2048,
 471         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 472         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 473
 474         PERF_CONTEXT_MAX                = (__u64)-4095,
 475 };
 476
 477 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 478 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 479 #define PERF_FLAG_PID_CGROUP            (1U << 2) /* pid=cgroup id, per-cpu mode only */
 480
 481 #ifdef __KERNEL__
 482 /*
 483  * Kernel-internal data types and definitions:
 484  */
 485
 486 #ifdef CONFIG_PERF_EVENTS
 487 # include <linux/cgroup.h>
 488 # include <asm/perf_event.h>
 489 # include <asm/local64.h>
 490 #endif
 491
 492 struct perf_guest_info_callbacks {
 493         int                             (*is_in_guest)(void);
 494         int                             (*is_user_mode)(void);
 495         unsigned long                   (*get_guest_ip)(void);
 496 };
 497
 498 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 499 #include <asm/hw_breakpoint.h>
 500 #endif
 501
 502 #include <linux/list.h>
 503 #include <linux/mutex.h>
 504 #include <linux/rculist.h>
 505 #include <linux/rcupdate.h>
 506 #include <linux/spinlock.h>
 507 #include <linux/hrtimer.h>
 508 #include <linux/fs.h>
 509 #include <linux/pid_namespace.h>
 510 #include <linux/workqueue.h>
 511 #include <linux/ftrace.h>
 512 #include <linux/cpu.h>
 513 #include <linux/irq_work.h>
 514 #include <linux/jump_label.h>
 515 #include <linux/atomic.h>
 516 #include <asm/local.h>
 517
 518 #define PERF_MAX_STACK_DEPTH            255
 519
 520 struct perf_callchain_entry {
 521         __u64                           nr;
 522         __u64                           ip[PERF_MAX_STACK_DEPTH];
 523 };
 524
 525 struct perf_raw_record {
 526         u32                             size;
 527         void                            *data;
 528 };
 529
 530 struct perf_branch_entry {
 531         __u64                           from;
 532         __u64                           to;
 533         __u64                           flags;
 534 };
 535
 536 struct perf_branch_stack {
 537         __u64                           nr;
 538         struct perf_branch_entry        entries[0];
 539 };
 540
 541 struct task_struct;
 542
 543 /*
 544  * extra PMU register associated with an event
 545  */
 546 struct hw_perf_event_extra {
 547         u64             config; /* register value */
 548         unsigned int    reg;    /* register address or index */
 549         int             alloc;  /* extra register already allocated */
 550         int             idx;    /* index in shared_regs->regs[] */
 551 };
 552
 553 /**
 554  * struct hw_perf_event - performance event hardware details:
 555  */
 556 struct hw_perf_event {
 557 #ifdef CONFIG_PERF_EVENTS
 558         union {
 559                 struct { /* hardware */
 560                         u64             config;
 561                         u64             last_tag;
 562                         unsigned long   config_base;
 563                         unsigned long   event_base;
 564                         int             idx;
 565                         int             last_cpu;
 566                         struct hw_perf_event_extra extra_reg;
 567                 };
 568                 struct { /* software */
 569                         struct hrtimer  hrtimer;
 570                 };
 571 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 572                 struct { /* breakpoint */
 573                         struct arch_hw_breakpoint       info;
 574                         struct list_head                bp_list;
 575                         /*
 576                          * Crufty hack to avoid the chicken and egg
 577                          * problem hw_breakpoint has with context
 578                          * creation and event initalization.
 579                          */
 580                         struct task_struct              *bp_target;
 581                 };
 582 #endif
 583         };
 584         int                             state;
 585         local64_t                       prev_count;
 586         u64                             sample_period;
 587         u64                             last_period;
 588         local64_t                       period_left;
 589         u64                             interrupts;
 590
 591         u64                             freq_time_stamp;
 592         u64                             freq_count_stamp;
 593 #endif
 594 };
 595
 596 /*
 597  * hw_perf_event::state flags
 598  */
 599 #define PERF_HES_STOPPED        0x01 /* the counter is stopped */
 600 #define PERF_HES_UPTODATE       0x02 /* event->count up-to-date */
 601 #define PERF_HES_ARCH           0x04
 602
 603 struct perf_event;
 604
 605 /*
 606  * Common implementation detail of pmu::{start,commit,cancel}_txn
 607  */
 608 #define PERF_EVENT_TXN 0x1
 609
 610 /**
 611  * struct pmu - generic performance monitoring unit
 612  */
 613 struct pmu {
 614         struct list_head                entry;
 615
 616         struct device                   *dev;
 617         char                            *name;
 618         int                             type;
 619
 620         int * __percpu                  pmu_disable_count;
 621         struct perf_cpu_context * __percpu pmu_cpu_context;
 622         int                             task_ctx_nr;
 623
 624         /*
 625          * Fully disable/enable this PMU, can be used to protect from the PMI
 626          * as well as for lazy/batch writing of the MSRs.
 627          */
 628         void (*pmu_enable)              (struct pmu *pmu); /* optional */
 629         void (*pmu_disable)             (struct pmu *pmu); /* optional */
 630
 631         /*
 632          * Try and initialize the event for this PMU.
 633          * Should return -ENOENT when the @event doesn't match this PMU.
 634          */
 635         int (*event_init)               (struct perf_event *event);
 636
 637 #define PERF_EF_START   0x01            /* start the counter when adding    */
 638 #define PERF_EF_RELOAD  0x02            /* reload the counter when starting */
 639 #define PERF_EF_UPDATE  0x04            /* update the counter when stopping */
 640
 641         /*
 642          * Adds/Removes a counter to/from the PMU, can be done inside
 643          * a transaction, see the ->*_txn() methods.
 644          */
 645         int  (*add)                     (struct perf_event *event, int flags);
 646         void (*del)                     (struct perf_event *event, int flags);
 647
 648         /*
 649          * Starts/Stops a counter present on the PMU. The PMI handler
 650          * should stop the counter when perf_event_overflow() returns
 651          * !0. ->start() will be used to continue.
 652          */
 653         void (*start)                   (struct perf_event *event, int flags);
 654         void (*stop)                    (struct perf_event *event, int flags);
 655
 656         /*
 657          * Updates the counter value of the event.
 658          */
 659         void (*read)                    (struct perf_event *event);
 660
 661         /*
 662          * Group events scheduling is treated as a transaction, add
 663          * group events as a whole and perform one schedulability test.
 664          * If the test fails, roll back the whole group
 665          *
 666          * Start the transaction, after this ->add() doesn't need to
 667          * do schedulability tests.
 668          */
 669         void (*start_txn)               (struct pmu *pmu); /* optional */
 670         /*
 671          * If ->start_txn() disabled the ->add() schedulability test
 672          * then ->commit_txn() is required to perform one. On success
 673          * the transaction is closed. On error the transaction is kept
 674          * open until ->cancel_txn() is called.
 675          */
 676         int  (*commit_txn)              (struct pmu *pmu); /* optional */
 677         /*
 678          * Will cancel the transaction, assumes ->del() is called
 679          * for each successful ->add() during the transaction.
 680          */
 681         void (*cancel_txn)              (struct pmu *pmu); /* optional */
 682 };
 683
 684 /**
 685  * enum perf_event_active_state - the states of a event
 686  */
 687 enum perf_event_active_state {
 688         PERF_EVENT_STATE_ERROR          = -2,
 689         PERF_EVENT_STATE_OFF            = -1,
 690         PERF_EVENT_STATE_INACTIVE       =  0,
 691         PERF_EVENT_STATE_ACTIVE         =  1,
 692 };
 693
 694 struct file;
 695 struct perf_sample_data;
 696
 697 typedef void (*perf_overflow_handler_t)(struct perf_event *,
 698                                         struct perf_sample_data *,
 699                                         struct pt_regs *regs);
 700
 701 enum perf_group_flag {
 702         PERF_GROUP_SOFTWARE             = 0x1,
 703 };
 704
 705 #define SWEVENT_HLIST_BITS              8
 706 #define SWEVENT_HLIST_SIZE              (1 << SWEVENT_HLIST_BITS)
 707
 708 struct swevent_hlist {
 709         struct hlist_head               heads[SWEVENT_HLIST_SIZE];
 710         struct rcu_head                 rcu_head;
 711 };
 712
 713 #define PERF_ATTACH_CONTEXT     0x01
 714 #define PERF_ATTACH_GROUP       0x02
 715 #define PERF_ATTACH_TASK        0x04
 716
 717 #ifdef CONFIG_CGROUP_PERF
 718 /*
 719  * perf_cgroup_info keeps track of time_enabled for a cgroup.
 720  * This is a per-cpu dynamically allocated data structure.
 721  */
 722 struct perf_cgroup_info {
 723         u64                             time;
 724         u64                             timestamp;
 725 };
 726
 727 struct perf_cgroup {
 728         struct                          cgroup_subsys_state css;
 729         struct                          perf_cgroup_info *info; /* timing info, one per cpu */
 730 };
 731 #endif
 732
 733 struct ring_buffer;
 734
 735 /**
 736  * struct perf_event - performance event kernel representation:
 737  */
 738 struct perf_event {
 739 #ifdef CONFIG_PERF_EVENTS
 740         struct list_head                group_entry;
 741         struct list_head                event_entry;
 742         struct list_head                sibling_list;
 743         struct hlist_node               hlist_entry;
 744         int                             nr_siblings;
 745         int                             group_flags;
 746         struct perf_event               *group_leader;
 747         struct pmu                      *pmu;
 748
 749         enum perf_event_active_state    state;
 750         unsigned int                    attach_state;
 751         local64_t                       count;
 752         atomic64_t                      child_count;
 753
 754         /*
 755          * These are the total time in nanoseconds that the event
 756          * has been enabled (i.e. eligible to run, and the task has
 757          * been scheduled in, if this is a per-task event)
 758          * and running (scheduled onto the CPU), respectively.
 759          *
 760          * They are computed from tstamp_enabled, tstamp_running and
 761          * tstamp_stopped when the event is in INACTIVE or ACTIVE state.
 762          */
 763         u64                             total_time_enabled;
 764         u64                             total_time_running;
 765
 766         /*
 767          * These are timestamps used for computing total_time_enabled
 768          * and total_time_running when the event is in INACTIVE or
 769          * ACTIVE state, measured in nanoseconds from an arbitrary point
 770          * in time.
 771          * tstamp_enabled: the notional time when the event was enabled
 772          * tstamp_running: the notional time when the event was scheduled on
 773          * tstamp_stopped: in INACTIVE state, the notional time when the
 774          *      event was scheduled off.
 775          */
 776         u64                             tstamp_enabled;
 777         u64                             tstamp_running;
 778         u64                             tstamp_stopped;
 779
 780         /*
 781          * timestamp shadows the actual context timing but it can
 782          * be safely used in NMI interrupt context. It reflects the
 783          * context time as it was when the event was last scheduled in.
 784          *
 785          * ctx_time already accounts for ctx->timestamp. Therefore to
 786          * compute ctx_time for a sample, simply add perf_clock().
 787          */
 788         u64                             shadow_ctx_time;
 789
 790         struct perf_event_attr          attr;
 791         u16                             header_size;
 792         u16                             id_header_size;
 793         u16                             read_size;
 794         struct hw_perf_event            hw;
 795
 796         struct perf_event_context       *ctx;
 797         struct file                     *filp;
 798
 799         /*
 800          * These accumulate total time (in nanoseconds) that children
 801          * events have been enabled and running, respectively.
 802          */
 803         atomic64_t                      child_total_time_enabled;
 804         atomic64_t                      child_total_time_running;
 805
 806         /*
 807          * Protect attach/detach and child_list:
 808          */
 809         struct mutex                    child_mutex;
 810         struct list_head                child_list;
 811         struct perf_event               *parent;
 812
 813         int                             oncpu;
 814         int                             cpu;
 815
 816         struct list_head                owner_entry;
 817         struct task_struct              *owner;
 818
 819         /* mmap bits */
 820         struct mutex                    mmap_mutex;
 821         atomic_t                        mmap_count;
 822         int                             mmap_locked;
 823         struct user_struct              *mmap_user;
 824         struct ring_buffer              *rb;
 825
 826         /* poll related */
 827         wait_queue_head_t               waitq;
 828         struct fasync_struct            *fasync;
 829
 830         /* delayed work for NMIs and such */
 831         int                             pending_wakeup;
 832         int                             pending_kill;
 833         int                             pending_disable;
 834         struct irq_work                 pending;
 835
 836         atomic_t                        event_limit;
 837
 838         void (*destroy)(struct perf_event *);
 839         struct rcu_head                 rcu_head;
 840
 841         struct pid_namespace            *ns;
 842         u64                             id;
 843
 844         perf_overflow_handler_t         overflow_handler;
 845         void                            *overflow_handler_context;
 846
 847 #ifdef CONFIG_EVENT_TRACING
 848         struct ftrace_event_call        *tp_event;
 849         struct event_filter             *filter;
 850 #endif
 851
 852 #ifdef CONFIG_CGROUP_PERF
 853         struct perf_cgroup              *cgrp; /* cgroup event is attach to */
 854         int                             cgrp_defer_enabled;
 855 #endif
 856
 857 #endif /* CONFIG_PERF_EVENTS */
 858 };
 859
 860 enum perf_event_context_type {
 861         task_context,
 862         cpu_context,
 863 };
 864
 865 /**
 866  * struct perf_event_context - event context structure
 867  *
 868  * Used as a container for task events and CPU events as well:
 869  */
 870 struct perf_event_context {
 871         struct pmu                      *pmu;
 872         enum perf_event_context_type    type;
 873         /*
 874          * Protect the states of the events in the list,
 875          * nr_active, and the list:
 876          */
 877         raw_spinlock_t                  lock;
 878         /*
 879          * Protect the list of events.  Locking either mutex or lock
 880          * is sufficient to ensure the list doesn't change; to change
 881          * the list you need to lock both the mutex and the spinlock.
 882          */
 883         struct mutex                    mutex;
 884
 885         struct list_head                pinned_groups;
 886         struct list_head                flexible_groups;
 887         struct list_head                event_list;
 888         int                             nr_events;
 889         int                             nr_active;
 890         int                             is_active;
 891         int                             nr_stat;
 892         int                             rotate_disable;
 893         atomic_t                        refcount;
 894         struct task_struct              *task;
 895
 896         /*
 897          * Context clock, runs when context enabled.
 898          */
 899         u64                             time;
 900         u64                             timestamp;
 901
 902         /*
 903          * These fields let us detect when two contexts have both
 904          * been cloned (inherited) from a common ancestor.
 905          */
 906         struct perf_event_context       *parent_ctx;
 907         u64                             parent_gen;
 908         u64                             generation;
 909         int                             pin_count;
 910         int                             nr_cgroups; /* cgroup events present */
 911         struct rcu_head                 rcu_head;
 912 };
 913
 914 /*
 915  * Number of contexts where an event can trigger:
 916  *      task, softirq, hardirq, nmi.
 917  */
 918 #define PERF_NR_CONTEXTS        4
 919
 920 /**
 921  * struct perf_event_cpu_context - per cpu event context structure
 922  */
 923 struct perf_cpu_context {
 924         struct perf_event_context       ctx;
 925         struct perf_event_context       *task_ctx;
 926         int                             active_oncpu;
 927         int                             exclusive;
 928         struct list_head                rotation_list;
 929         int                             jiffies_interval;
 930         struct pmu                      *active_pmu;
 931         struct perf_cgroup              *cgrp;
 932 };
 933
 934 struct perf_output_handle {
 935         struct perf_event               *event;
 936         struct ring_buffer              *rb;
 937         unsigned long                   wakeup;
 938         unsigned long                   size;
 939         void                            *addr;
 940         int                             page;
 941 };
 942
 943 #ifdef CONFIG_PERF_EVENTS
 944
 945 extern int perf_pmu_register(struct pmu *pmu, char *name, int type);
 946 extern void perf_pmu_unregister(struct pmu *pmu);
 947
 948 extern int perf_num_counters(void);
 949 extern const char *perf_pmu_name(void);
 950 extern void __perf_event_task_sched_in(struct task_struct *prev,
 951                                        struct task_struct *task);
 952 extern void __perf_event_task_sched_out(struct task_struct *prev,
 953                                         struct task_struct *next);
 954 extern int perf_event_init_task(struct task_struct *child);
 955 extern void perf_event_exit_task(struct task_struct *child);
 956 extern void perf_event_free_task(struct task_struct *task);
 957 extern void perf_event_delayed_put(struct task_struct *task);
 958 extern void perf_event_print_debug(void);
 959 extern void perf_pmu_disable(struct pmu *pmu);
 960 extern void perf_pmu_enable(struct pmu *pmu);
 961 extern int perf_event_task_disable(void);
 962 extern int perf_event_task_enable(void);
 963 extern int perf_event_refresh(struct perf_event *event, int refresh);
 964 extern void perf_event_update_userpage(struct perf_event *event);
 965 extern int perf_event_release_kernel(struct perf_event *event);
 966 extern struct perf_event *
 967 perf_event_create_kernel_counter(struct perf_event_attr *attr,
 968                                 int cpu,
 969                                 struct task_struct *task,
 970                                 perf_overflow_handler_t callback,
 971                                 void *context);
 972 extern u64 perf_event_read_value(struct perf_event *event,
 973                                  u64 *enabled, u64 *running);
 974
 975 struct perf_sample_data {
 976         u64                             type;
 977
 978         u64                             ip;
 979         struct {
 980                 u32     pid;
 981                 u32     tid;
 982         }                               tid_entry;
 983         u64                             time;
 984         u64                             addr;
 985         u64                             id;
 986         u64                             stream_id;
 987         struct {
 988                 u32     cpu;
 989                 u32     reserved;
 990         }                               cpu_entry;
 991         u64                             period;
 992         struct perf_callchain_entry     *callchain;
 993         struct perf_raw_record          *raw;
 994 };
 995
 996 static inline void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
 997 {
 998         data->addr = addr;
 999         data->raw  = NULL;
1000 }
1001
1002 extern void perf_output_sample(struct perf_output_handle *handle,
1003                                struct perf_event_header *header,
1004                                struct perf_sample_data *data,
1005                                struct perf_event *event);
1006 extern void perf_prepare_sample(struct perf_event_header *header,
1007                                 struct perf_sample_data *data,
1008                                 struct perf_event *event,
1009                                 struct pt_regs *regs);
1010
1011 extern int perf_event_overflow(struct perf_event *event,
1012                                  struct perf_sample_data *data,
1013                                  struct pt_regs *regs);
1014
1015 static inline bool is_sampling_event(struct perf_event *event)
1016 {
1017         return event->attr.sample_period != 0;
1018 }
1019
1020 /*
1021  * Return 1 for a software event, 0 for a hardware event
1022  */
1023 static inline int is_software_event(struct perf_event *event)
1024 {
1025         return event->pmu->task_ctx_nr == perf_sw_context;
1026 }
1027
1028 extern struct jump_label_key perf_swevent_enabled[PERF_COUNT_SW_MAX];
1029
1030 extern void __perf_sw_event(u32, u64, struct pt_regs *, u64);
1031
1032 #ifndef perf_arch_fetch_caller_regs
1033 static inline void perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip) { }
1034 #endif
1035
1036 /*
1037  * Take a snapshot of the regs. Skip ip and frame pointer to
1038  * the nth caller. We only need a few of the regs:
1039  * - ip for PERF_SAMPLE_IP
1040  * - cs for user_mode() tests
1041  * - bp for callchains
1042  * - eflags, for future purposes, just in case
1043  */
1044 static inline void perf_fetch_caller_regs(struct pt_regs *regs)
1045 {
1046         memset(regs, 0, sizeof(*regs));
1047
1048         perf_arch_fetch_caller_regs(regs, CALLER_ADDR0);
1049 }
1050
1051 static __always_inline void
1052 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)
1053 {
1054         struct pt_regs hot_regs;
1055
1056         if (static_branch(&perf_swevent_enabled[event_id])) {
1057                 if (!regs) {
1058                         perf_fetch_caller_regs(&hot_regs);
1059                         regs = &hot_regs;
1060                 }
1061                 __perf_sw_event(event_id, nr, regs, addr);
1062         }
1063 }
1064
1065 extern struct jump_label_key perf_sched_events;
1066
1067 static inline void perf_event_task_sched_in(struct task_struct *prev,
1068                                             struct task_struct *task)
1069 {
1070         if (static_branch(&perf_sched_events))
1071                 __perf_event_task_sched_in(prev, task);
1072 }
1073
1074 static inline void perf_event_task_sched_out(struct task_struct *prev,
1075                                              struct task_struct *next)
1076 {
1077         perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, NULL, 0);
1078
1079         if (static_branch(&perf_sched_events))
1080                 __perf_event_task_sched_out(prev, next);
1081 }
1082
1083 extern void perf_event_mmap(struct vm_area_struct *vma);
1084 extern struct perf_guest_info_callbacks *perf_guest_cbs;
1085 extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1086 extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
1087
1088 extern void perf_event_comm(struct task_struct *tsk);
1089 extern void perf_event_fork(struct task_struct *tsk);
1090
1091 /* Callchains */
1092 DECLARE_PER_CPU(struct perf_callchain_entry, perf_callchain_entry);
1093
1094 extern void perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs);
1095 extern void perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs);
1096
1097 static inline void perf_callchain_store(struct perf_callchain_entry *entry, u64 ip)
1098 {
1099         if (entry->nr < PERF_MAX_STACK_DEPTH)
1100                 entry->ip[entry->nr++] = ip;
1101 }
1102
1103 extern int sysctl_perf_event_paranoid;
1104 extern int sysctl_perf_event_mlock;
1105 extern int sysctl_perf_event_sample_rate;
1106
1107 extern int perf_proc_update_handler(struct ctl_table *table, int write,
1108                 void __user *buffer, size_t *lenp,
1109                 loff_t *ppos);
1110
1111 static inline bool perf_paranoid_tracepoint_raw(void)
1112 {
1113         return sysctl_perf_event_paranoid > -1;
1114 }
1115
1116 static inline bool perf_paranoid_cpu(void)
1117 {
1118         return sysctl_perf_event_paranoid > 0;
1119 }
1120
1121 static inline bool perf_paranoid_kernel(void)
1122 {
1123         return sysctl_perf_event_paranoid > 1;
1124 }
1125
1126 extern void perf_event_init(void);
1127 extern void perf_tp_event(u64 addr, u64 count, void *record,
1128                           int entry_size, struct pt_regs *regs,
1129                           struct hlist_head *head, int rctx);
1130 extern void perf_bp_event(struct perf_event *event, void *data);
1131
1132 #ifndef perf_misc_flags
1133 # define perf_misc_flags(regs) \
1134                 (user_mode(regs) ? PERF_RECORD_MISC_USER : PERF_RECORD_MISC_KERNEL)
1135 # define perf_instruction_pointer(regs) instruction_pointer(regs)
1136 #endif
1137
1138 extern int perf_output_begin(struct perf_output_handle *handle,
1139                              struct perf_event *event, unsigned int size);
1140 extern void perf_output_end(struct perf_output_handle *handle);
1141 extern void perf_output_copy(struct perf_output_handle *handle,
1142                              const void *buf, unsigned int len);
1143 extern int perf_swevent_get_recursion_context(void);
1144 extern void perf_swevent_put_recursion_context(int rctx);
1145 extern void perf_event_enable(struct perf_event *event);
1146 extern void perf_event_disable(struct perf_event *event);
1147 extern void perf_event_task_tick(void);
1148 #else
1149 static inline void
1150 perf_event_task_sched_in(struct task_struct *prev,
1151                          struct task_struct *task)                      { }
1152 static inline void
1153 perf_event_task_sched_out(struct task_struct *prev,
1154                           struct task_struct *next)                     { }
1155 static inline int perf_event_init_task(struct task_struct *child)       { return 0; }
1156 static inline void perf_event_exit_task(struct task_struct *child)      { }
1157 static inline void perf_event_free_task(struct task_struct *task)       { }
1158 static inline void perf_event_delayed_put(struct task_struct *task)     { }
1159 static inline void perf_event_print_debug(void)                         { }
1160 static inline int perf_event_task_disable(void)                         { return -EINVAL; }
1161 static inline int perf_event_task_enable(void)                          { return -EINVAL; }
1162 static inline int perf_event_refresh(struct perf_event *event, int refresh)
1163 {
1164         return -EINVAL;
1165 }
1166
1167 static inline void
1168 perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr)     { }
1169 static inline void
1170 perf_bp_event(struct perf_event *event, void *data)                     { }
1171
1172 static inline int perf_register_guest_info_callbacks
1173 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1174 static inline int perf_unregister_guest_info_callbacks
1175 (struct perf_guest_info_callbacks *callbacks)                           { return 0; }
1176
1177 static inline void perf_event_mmap(struct vm_area_struct *vma)          { }
1178 static inline void perf_event_comm(struct task_struct *tsk)             { }
1179 static inline void perf_event_fork(struct task_struct *tsk)             { }
1180 static inline void perf_event_init(void)                                { }
1181 static inline int  perf_swevent_get_recursion_context(void)             { return -1; }
1182 static inline void perf_swevent_put_recursion_context(int rctx)         { }
1183 static inline void perf_event_enable(struct perf_event *event)          { }
1184 static inline void perf_event_disable(struct perf_event *event)         { }
1185 static inline void perf_event_task_tick(void)                           { }
1186 #endif
1187
1188 #define perf_output_put(handle, x) perf_output_copy((handle), &(x), sizeof(x))
1189
1190 /*
1191  * This has to have a higher priority than migration_notifier in sched.c.
1192  */
1193 #define perf_cpu_notifier(fn)                                           \
1194 do {                                                                    \
1195         static struct notifier_block fn##_nb __cpuinitdata =            \
1196                 { .notifier_call = fn, .priority = CPU_PRI_PERF };      \
1197         fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,                     \
1198                 (void *)(unsigned long)smp_processor_id());             \
1199         fn(&fn##_nb, (unsigned long)CPU_STARTING,                       \
1200                 (void *)(unsigned long)smp_processor_id());             \
1201         fn(&fn##_nb, (unsigned long)CPU_ONLINE,                         \
1202                 (void *)(unsigned long)smp_processor_id());             \
1203         register_cpu_notifier(&fn##_nb);                                \
1204 } while (0)
1205
1206 #endif /* __KERNEL__ */
1207 #endif /* _LINUX_PERF_EVENT_H */