mono/profiler/perf_event.h

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