include/linux/perf_counter.h

   1 /*
   2  *  NOTE: this file will be removed in a future kernel release, it is
   3  *  provided as a courtesy copy of user-space code that relies on the
   4  *  old (pre-rename) symbols and constants.
   5  *
   6  *  Performance events:
   7  *
   8  *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
   9  *    Copyright (C) 2008-2009, Red Hat, Inc., Ingo Molnar
  10  *    Copyright (C) 2008-2009, Red Hat, Inc., Peter Zijlstra
  11  *
  12  *  Data type definitions, declarations, prototypes.
  13  *
  14  *    Started by: Thomas Gleixner and Ingo Molnar
  15  *
  16  *  For licencing details see kernel-base/COPYING
  17  */
  18 #ifndef _LINUX_PERF_COUNTER_H
  19 #define _LINUX_PERF_COUNTER_H
  20
  21 #include <linux/types.h>
  22 #include <linux/ioctl.h>
  23 #include <asm/byteorder.h>
  24
  25 /*
  26  * User-space ABI bits:
  27  */
  28
  29 /*
  30  * attr.type
  31  */
  32 enum perf_type_id {
  33         PERF_TYPE_HARDWARE                      = 0,
  34         PERF_TYPE_SOFTWARE                      = 1,
  35         PERF_TYPE_TRACEPOINT                    = 2,
  36         PERF_TYPE_HW_CACHE                      = 3,
  37         PERF_TYPE_RAW                           = 4,
  38
  39         PERF_TYPE_MAX,                          /* non-ABI */
  40 };
  41
  42 /*
  43  * Generalized performance counter event types, used by the
  44  * attr.event_id parameter of the sys_perf_counter_open()
  45  * syscall:
  46  */
  47 enum perf_hw_id {
  48         /*
  49          * Common hardware events, generalized by the kernel:
  50          */
  51         PERF_COUNT_HW_CPU_CYCLES                = 0,
  52         PERF_COUNT_HW_INSTRUCTIONS              = 1,
  53         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
  54         PERF_COUNT_HW_CACHE_MISSES              = 3,
  55         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
  56         PERF_COUNT_HW_BRANCH_MISSES             = 5,
  57         PERF_COUNT_HW_BUS_CYCLES                = 6,
  58
  59         PERF_COUNT_HW_MAX,                      /* non-ABI */
  60 };
  61
  62 /*
  63  * Generalized hardware cache counters:
  64  *
  65  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU } x
  66  *       { read, write, prefetch } x
  67  *       { accesses, misses }
  68  */
  69 enum perf_hw_cache_id {
  70         PERF_COUNT_HW_CACHE_L1D                 = 0,
  71         PERF_COUNT_HW_CACHE_L1I                 = 1,
  72         PERF_COUNT_HW_CACHE_LL                  = 2,
  73         PERF_COUNT_HW_CACHE_DTLB                = 3,
  74         PERF_COUNT_HW_CACHE_ITLB                = 4,
  75         PERF_COUNT_HW_CACHE_BPU                 = 5,
  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" counters provided by the kernel, even if the hardware
  97  * does not support performance counters. These counters 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_counter_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 counter fd,
 137  * as specified by attr.read_format:
 138  *
 139  * struct read_format {
 140  *      { u64           value;
 141  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
 142  *        { u64         time_running; } && PERF_FORMAT_RUNNING
 143  *        { u64         id;           } && PERF_FORMAT_ID
 144  *      } && !PERF_FORMAT_GROUP
 145  *
 146  *      { u64           nr;
 147  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
 148  *        { u64         time_running; } && PERF_FORMAT_RUNNING
 149  *        { u64         value;
 150  *          { u64       id;           } && PERF_FORMAT_ID
 151  *        }             cntr[nr];
 152  *      } && PERF_FORMAT_GROUP
 153  * };
 154  */
 155 enum perf_counter_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 to monitor via a performance monitoring counter:
 168  */
 169 struct perf_counter_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                                 __reserved_1   : 49;
 211
 212         union {
 213                 __u32           wakeup_events;    /* wakeup every n events */
 214                 __u32           wakeup_watermark; /* bytes before wakeup   */
 215         };
 216         __u32                   __reserved_2;
 217
 218         __u64                   __reserved_3;
 219 };
 220
 221 /*
 222  * Ioctls that can be done on a perf counter fd:
 223  */
 224 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 225 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 226 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
 227 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
 228 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
 229 #define PERF_COUNTER_IOC_SET_OUTPUT     _IO ('$', 5)
 230 #define PERF_COUNTER_IOC_SET_FILTER     _IOW('$', 6, char *)
 231
 232 enum perf_counter_ioc_flags {
 233         PERF_IOC_FLAG_GROUP             = 1U << 0,
 234 };
 235
 236 /*
 237  * Structure of the page that can be mapped via mmap
 238  */
 239 struct perf_counter_mmap_page {
 240         __u32   version;                /* version number of this structure */
 241         __u32   compat_version;         /* lowest version this is compat with */
 242
 243         /*
 244          * Bits needed to read the hw counters in user-space.
 245          *
 246          *   u32 seq;
 247          *   s64 count;
 248          *
 249          *   do {
 250          *     seq = pc->lock;
 251          *
 252          *     barrier()
 253          *     if (pc->index) {
 254          *       count = pmc_read(pc->index - 1);
 255          *       count += pc->offset;
 256          *     } else
 257          *       goto regular_read;
 258          *
 259          *     barrier();
 260          *   } while (pc->lock != seq);
 261          *
 262          * NOTE: for obvious reason this only works on self-monitoring
 263          *       processes.
 264          */
 265         __u32   lock;                   /* seqlock for synchronization */
 266         __u32   index;                  /* hardware counter identifier */
 267         __s64   offset;                 /* add to hardware counter value */
 268         __u64   time_enabled;           /* time counter active */
 269         __u64   time_running;           /* time counter on cpu */
 270
 271                 /*
 272                  * Hole for extension of the self monitor capabilities
 273                  */
 274
 275         __u64   __reserved[123];        /* align to 1k */
 276
 277         /*
 278          * Control data for the mmap() data buffer.
 279          *
 280          * User-space reading the @data_head value should issue an rmb(), on
 281          * SMP capable platforms, after reading this value -- see
 282          * perf_counter_wakeup().
 283          *
 284          * When the mapping is PROT_WRITE the @data_tail value should be
 285          * written by userspace to reflect the last read data. In this case
 286          * the kernel will not over-write unread data.
 287          */
 288         __u64   data_head;              /* head in the data section */
 289         __u64   data_tail;              /* user-space written tail */
 290 };
 291
 292 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
 293 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
 294 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
 295 #define PERF_EVENT_MISC_USER                    (2 << 0)
 296 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
 297
 298 struct perf_event_header {
 299         __u32   type;
 300         __u16   misc;
 301         __u16   size;
 302 };
 303
 304 enum perf_event_type {
 305
 306         /*
 307          * The MMAP events record the PROT_EXEC mappings so that we can
 308          * correlate userspace IPs to code. They have the following structure:
 309          *
 310          * struct {
 311          *      struct perf_event_header        header;
 312          *
 313          *      u32                             pid, tid;
 314          *      u64                             addr;
 315          *      u64                             len;
 316          *      u64                             pgoff;
 317          *      char                            filename[];
 318          * };
 319          */
 320         PERF_EVENT_MMAP                 = 1,
 321
 322         /*
 323          * struct {
 324          *      struct perf_event_header        header;
 325          *      u64                             id;
 326          *      u64                             lost;
 327          * };
 328          */
 329         PERF_EVENT_LOST                 = 2,
 330
 331         /*
 332          * struct {
 333          *      struct perf_event_header        header;
 334          *
 335          *      u32                             pid, tid;
 336          *      char                            comm[];
 337          * };
 338          */
 339         PERF_EVENT_COMM                 = 3,
 340
 341         /*
 342          * struct {
 343          *      struct perf_event_header        header;
 344          *      u32                             pid, ppid;
 345          *      u32                             tid, ptid;
 346          *      u64                             time;
 347          * };
 348          */
 349         PERF_EVENT_EXIT                 = 4,
 350
 351         /*
 352          * struct {
 353          *      struct perf_event_header        header;
 354          *      u64                             time;
 355          *      u64                             id;
 356          *      u64                             stream_id;
 357          * };
 358          */
 359         PERF_EVENT_THROTTLE             = 5,
 360         PERF_EVENT_UNTHROTTLE           = 6,
 361
 362         /*
 363          * struct {
 364          *      struct perf_event_header        header;
 365          *      u32                             pid, ppid;
 366          *      u32                             tid, ptid;
 367          *      u64                             time;
 368          * };
 369          */
 370         PERF_EVENT_FORK                 = 7,
 371
 372         /*
 373          * struct {
 374          *      struct perf_event_header        header;
 375          *      u32                             pid, tid;
 376          *
 377          *      struct read_format              values;
 378          * };
 379          */
 380         PERF_EVENT_READ                 = 8,
 381
 382         /*
 383          * struct {
 384          *      struct perf_event_header        header;
 385          *
 386          *      { u64                   ip;       } && PERF_SAMPLE_IP
 387          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 388          *      { u64                   time;     } && PERF_SAMPLE_TIME
 389          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 390          *      { u64                   id;       } && PERF_SAMPLE_ID
 391          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 392          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 393          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 394          *
 395          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 396          *
 397          *      { u64                   nr,
 398          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 399          *
 400          *      #
 401          *      # The RAW record below is opaque data wrt the ABI
 402          *      #
 403          *      # That is, the ABI doesn't make any promises wrt to
 404          *      # the stability of its content, it may vary depending
 405          *      # on event, hardware, kernel version and phase of
 406          *      # the moon.
 407          *      #
 408          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 409          *      #
 410          *
 411          *      { u32                   size;
 412          *        char                  data[size];}&& PERF_SAMPLE_RAW
 413          * };
 414          */
 415         PERF_EVENT_SAMPLE               = 9,
 416
 417         PERF_EVENT_MAX,                 /* non-ABI */
 418 };
 419
 420 enum perf_callchain_context {
 421         PERF_CONTEXT_HV                 = (__u64)-32,
 422         PERF_CONTEXT_KERNEL             = (__u64)-128,
 423         PERF_CONTEXT_USER               = (__u64)-512,
 424
 425         PERF_CONTEXT_GUEST              = (__u64)-2048,
 426         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 427         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 428
 429         PERF_CONTEXT_MAX                = (__u64)-4095,
 430 };
 431
 432 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 433 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 434
 435 /*
 436  * In case some app still references the old symbols:
 437  */
 438
 439 #define __NR_perf_counter_open          __NR_perf_event_open
 440
 441 #define PR_TASK_PERF_COUNTERS_DISABLE   PR_TASK_PERF_EVENTS_DISABLE
 442 #define PR_TASK_PERF_COUNTERS_ENABLE    PR_TASK_PERF_EVENTS_ENABLE
 443
 444 #endif /* _LINUX_PERF_COUNTER_H */