SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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
 110         PERF_COUNT_SW_MAX,                      /* non-ABI */
 111 };
 112
 113 /*
 114  * Bits that can be set in attr.sample_type to request information
 115  * in the overflow packets.
 116  */
 117 enum perf_counter_sample_format {
 118         PERF_SAMPLE_IP                          = 1U << 0,
 119         PERF_SAMPLE_TID                         = 1U << 1,
 120         PERF_SAMPLE_TIME                        = 1U << 2,
 121         PERF_SAMPLE_ADDR                        = 1U << 3,
 122         PERF_SAMPLE_READ                        = 1U << 4,
 123         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
 124         PERF_SAMPLE_ID                          = 1U << 6,
 125         PERF_SAMPLE_CPU                         = 1U << 7,
 126         PERF_SAMPLE_PERIOD                      = 1U << 8,
 127         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
 128         PERF_SAMPLE_RAW                         = 1U << 10,
 129
 130         PERF_SAMPLE_MAX = 1U << 11,             /* non-ABI */
 131 };
 132
 133 /*
 134  * The format of the data returned by read() on a perf counter fd,
 135  * as specified by attr.read_format:
 136  *
 137  * struct read_format {
 138  *      { u64           value;
 139  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
 140  *        { u64         time_running; } && PERF_FORMAT_RUNNING
 141  *        { u64         id;           } && PERF_FORMAT_ID
 142  *      } && !PERF_FORMAT_GROUP
 143  *
 144  *      { u64           nr;
 145  *        { u64         time_enabled; } && PERF_FORMAT_ENABLED
 146  *        { u64         time_running; } && PERF_FORMAT_RUNNING
 147  *        { u64         value;
 148  *          { u64       id;           } && PERF_FORMAT_ID
 149  *        }             cntr[nr];
 150  *      } && PERF_FORMAT_GROUP
 151  * };
 152  */
 153 enum perf_counter_read_format {
 154         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
 155         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
 156         PERF_FORMAT_ID                          = 1U << 2,
 157         PERF_FORMAT_GROUP                       = 1U << 3,
 158
 159         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
 160 };
 161
 162 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
 163
 164 /*
 165  * Hardware event to monitor via a performance monitoring counter:
 166  */
 167 struct perf_counter_attr {
 168
 169         /*
 170          * Major type: hardware/software/tracepoint/etc.
 171          */
 172         __u32                   type;
 173
 174         /*
 175          * Size of the attr structure, for fwd/bwd compat.
 176          */
 177         __u32                   size;
 178
 179         /*
 180          * Type specific configuration information.
 181          */
 182         __u64                   config;
 183
 184         union {
 185                 __u64           sample_period;
 186                 __u64           sample_freq;
 187         };
 188
 189         __u64                   sample_type;
 190         __u64                   read_format;
 191
 192         __u64                   disabled       :  1, /* off by default        */
 193                                 inherit        :  1, /* children inherit it   */
 194                                 pinned         :  1, /* must always be on PMU */
 195                                 exclusive      :  1, /* only group on PMU     */
 196                                 exclude_user   :  1, /* don't count user      */
 197                                 exclude_kernel :  1, /* ditto kernel          */
 198                                 exclude_hv     :  1, /* ditto hypervisor      */
 199                                 exclude_idle   :  1, /* don't count when idle */
 200                                 mmap           :  1, /* include mmap data     */
 201                                 comm           :  1, /* include comm data     */
 202                                 freq           :  1, /* use freq, not period  */
 203                                 inherit_stat   :  1, /* per task counts       */
 204                                 enable_on_exec :  1, /* next exec enables     */
 205                                 task           :  1, /* trace fork/exit       */
 206                                 watermark      :  1, /* wakeup_watermark      */
 207
 208                                 __reserved_1   : 49;
 209
 210         union {
 211                 __u32           wakeup_events;    /* wakeup every n events */
 212                 __u32           wakeup_watermark; /* bytes before wakeup   */
 213         };
 214         __u32                   __reserved_2;
 215
 216         __u64                   __reserved_3;
 217 };
 218
 219 /*
 220  * Ioctls that can be done on a perf counter fd:
 221  */
 222 #define PERF_COUNTER_IOC_ENABLE         _IO ('$', 0)
 223 #define PERF_COUNTER_IOC_DISABLE        _IO ('$', 1)
 224 #define PERF_COUNTER_IOC_REFRESH        _IO ('$', 2)
 225 #define PERF_COUNTER_IOC_RESET          _IO ('$', 3)
 226 #define PERF_COUNTER_IOC_PERIOD         _IOW('$', 4, u64)
 227 #define PERF_COUNTER_IOC_SET_OUTPUT     _IO ('$', 5)
 228 #define PERF_COUNTER_IOC_SET_FILTER     _IOW('$', 6, char *)
 229
 230 enum perf_counter_ioc_flags {
 231         PERF_IOC_FLAG_GROUP             = 1U << 0,
 232 };
 233
 234 /*
 235  * Structure of the page that can be mapped via mmap
 236  */
 237 struct perf_counter_mmap_page {
 238         __u32   version;                /* version number of this structure */
 239         __u32   compat_version;         /* lowest version this is compat with */
 240
 241         /*
 242          * Bits needed to read the hw counters in user-space.
 243          *
 244          *   u32 seq;
 245          *   s64 count;
 246          *
 247          *   do {
 248          *     seq = pc->lock;
 249          *
 250          *     barrier()
 251          *     if (pc->index) {
 252          *       count = pmc_read(pc->index - 1);
 253          *       count += pc->offset;
 254          *     } else
 255          *       goto regular_read;
 256          *
 257          *     barrier();
 258          *   } while (pc->lock != seq);
 259          *
 260          * NOTE: for obvious reason this only works on self-monitoring
 261          *       processes.
 262          */
 263         __u32   lock;                   /* seqlock for synchronization */
 264         __u32   index;                  /* hardware counter identifier */
 265         __s64   offset;                 /* add to hardware counter value */
 266         __u64   time_enabled;           /* time counter active */
 267         __u64   time_running;           /* time counter on cpu */
 268
 269                 /*
 270                  * Hole for extension of the self monitor capabilities
 271                  */
 272
 273         __u64   __reserved[123];        /* align to 1k */
 274
 275         /*
 276          * Control data for the mmap() data buffer.
 277          *
 278          * User-space reading the @data_head value should issue an rmb(), on
 279          * SMP capable platforms, after reading this value -- see
 280          * perf_counter_wakeup().
 281          *
 282          * When the mapping is PROT_WRITE the @data_tail value should be
 283          * written by userspace to reflect the last read data. In this case
 284          * the kernel will not over-write unread data.
 285          */
 286         __u64   data_head;              /* head in the data section */
 287         __u64   data_tail;              /* user-space written tail */
 288 };
 289
 290 #define PERF_EVENT_MISC_CPUMODE_MASK            (3 << 0)
 291 #define PERF_EVENT_MISC_CPUMODE_UNKNOWN         (0 << 0)
 292 #define PERF_EVENT_MISC_KERNEL                  (1 << 0)
 293 #define PERF_EVENT_MISC_USER                    (2 << 0)
 294 #define PERF_EVENT_MISC_HYPERVISOR              (3 << 0)
 295
 296 struct perf_event_header {
 297         __u32   type;
 298         __u16   misc;
 299         __u16   size;
 300 };
 301
 302 enum perf_event_type {
 303
 304         /*
 305          * The MMAP events record the PROT_EXEC mappings so that we can
 306          * correlate userspace IPs to code. They have the following structure:
 307          *
 308          * struct {
 309          *      struct perf_event_header        header;
 310          *
 311          *      u32                             pid, tid;
 312          *      u64                             addr;
 313          *      u64                             len;
 314          *      u64                             pgoff;
 315          *      char                            filename[];
 316          * };
 317          */
 318         PERF_EVENT_MMAP                 = 1,
 319
 320         /*
 321          * struct {
 322          *      struct perf_event_header        header;
 323          *      u64                             id;
 324          *      u64                             lost;
 325          * };
 326          */
 327         PERF_EVENT_LOST                 = 2,
 328
 329         /*
 330          * struct {
 331          *      struct perf_event_header        header;
 332          *
 333          *      u32                             pid, tid;
 334          *      char                            comm[];
 335          * };
 336          */
 337         PERF_EVENT_COMM                 = 3,
 338
 339         /*
 340          * struct {
 341          *      struct perf_event_header        header;
 342          *      u32                             pid, ppid;
 343          *      u32                             tid, ptid;
 344          *      u64                             time;
 345          * };
 346          */
 347         PERF_EVENT_EXIT                 = 4,
 348
 349         /*
 350          * struct {
 351          *      struct perf_event_header        header;
 352          *      u64                             time;
 353          *      u64                             id;
 354          *      u64                             stream_id;
 355          * };
 356          */
 357         PERF_EVENT_THROTTLE             = 5,
 358         PERF_EVENT_UNTHROTTLE           = 6,
 359
 360         /*
 361          * struct {
 362          *      struct perf_event_header        header;
 363          *      u32                             pid, ppid;
 364          *      u32                             tid, ptid;
 365          *      u64                             time;
 366          * };
 367          */
 368         PERF_EVENT_FORK                 = 7,
 369
 370         /*
 371          * struct {
 372          *      struct perf_event_header        header;
 373          *      u32                             pid, tid;
 374          *
 375          *      struct read_format              values;
 376          * };
 377          */
 378         PERF_EVENT_READ                 = 8,
 379
 380         /*
 381          * struct {
 382          *      struct perf_event_header        header;
 383          *
 384          *      { u64                   ip;       } && PERF_SAMPLE_IP
 385          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 386          *      { u64                   time;     } && PERF_SAMPLE_TIME
 387          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 388          *      { u64                   id;       } && PERF_SAMPLE_ID
 389          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 390          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 391          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 392          *
 393          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 394          *
 395          *      { u64                   nr,
 396          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 397          *
 398          *      #
 399          *      # The RAW record below is opaque data wrt the ABI
 400          *      #
 401          *      # That is, the ABI doesn't make any promises wrt to
 402          *      # the stability of its content, it may vary depending
 403          *      # on event, hardware, kernel version and phase of
 404          *      # the moon.
 405          *      #
 406          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 407          *      #
 408          *
 409          *      { u32                   size;
 410          *        char                  data[size];}&& PERF_SAMPLE_RAW
 411          * };
 412          */
 413         PERF_EVENT_SAMPLE               = 9,
 414
 415         PERF_EVENT_MAX,                 /* non-ABI */
 416 };
 417
 418 enum perf_callchain_context {
 419         PERF_CONTEXT_HV                 = (__u64)-32,
 420         PERF_CONTEXT_KERNEL             = (__u64)-128,
 421         PERF_CONTEXT_USER               = (__u64)-512,
 422
 423         PERF_CONTEXT_GUEST              = (__u64)-2048,
 424         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
 425         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
 426
 427         PERF_CONTEXT_MAX                = (__u64)-4095,
 428 };
 429
 430 #define PERF_FLAG_FD_NO_GROUP           (1U << 0)
 431 #define PERF_FLAG_FD_OUTPUT             (1U << 1)
 432
 433 /*
 434  * In case some app still references the old symbols:
 435  */
 436
 437 #define __NR_perf_counter_open          __NR_perf_event_open
 438
 439 #define PR_TASK_PERF_COUNTERS_DISABLE   PR_TASK_PERF_EVENTS_DISABLE
 440 #define PR_TASK_PERF_COUNTERS_ENABLE    PR_TASK_PERF_EVENTS_ENABLE
 441
 442 #endif /* _LINUX_PERF_COUNTER_H */