tools/include/uapi/linux/perf_event.h

   1 /* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
   2 /*
   3  * Performance events:
   4  *
   5  *    Copyright (C) 2008-2009, Thomas Gleixner <[email protected]>
   6  *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
   7  *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
   8  *
   9  * Data type definitions, declarations, prototypes.
  10  *
  11  *    Started by: Thomas Gleixner and Ingo Molnar
  12  *
  13  * For licencing details see kernel-base/COPYING
  14  */
  15 #ifndef _UAPI_LINUX_PERF_EVENT_H
  16 #define _UAPI_LINUX_PERF_EVENT_H
  17
  18 #include <linux/types.h>
  19 #include <linux/ioctl.h>
  20 #include <asm/byteorder.h>
  21
  22 /*
  23  * User-space ABI bits:
  24  */
  25
  26 /*
  27  * attr.type
  28  */
  29 enum perf_type_id {
  30         PERF_TYPE_HARDWARE                      = 0,
  31         PERF_TYPE_SOFTWARE                      = 1,
  32         PERF_TYPE_TRACEPOINT                    = 2,
  33         PERF_TYPE_HW_CACHE                      = 3,
  34         PERF_TYPE_RAW                           = 4,
  35         PERF_TYPE_BREAKPOINT                    = 5,
  36
  37         PERF_TYPE_MAX,                          /* non-ABI */
  38 };
  39
  40 /*
  41  * attr.config layout for type PERF_TYPE_HARDWARE and PERF_TYPE_HW_CACHE
  42  * PERF_TYPE_HARDWARE:                  0xEEEEEEEE000000AA
  43  *                                      AA: hardware event ID
  44  *                                      EEEEEEEE: PMU type ID
  45  * PERF_TYPE_HW_CACHE:                  0xEEEEEEEE00DDCCBB
  46  *                                      BB: hardware cache ID
  47  *                                      CC: hardware cache op ID
  48  *                                      DD: hardware cache op result ID
  49  *                                      EEEEEEEE: PMU type ID
  50  * If the PMU type ID is 0, the PERF_TYPE_RAW will be applied.
  51  */
  52 #define PERF_PMU_TYPE_SHIFT             32
  53 #define PERF_HW_EVENT_MASK              0xffffffff
  54
  55 /*
  56  * Generalized performance event event_id types, used by the
  57  * attr.event_id parameter of the sys_perf_event_open()
  58  * syscall:
  59  */
  60 enum perf_hw_id {
  61         /*
  62          * Common hardware events, generalized by the kernel:
  63          */
  64         PERF_COUNT_HW_CPU_CYCLES                = 0,
  65         PERF_COUNT_HW_INSTRUCTIONS              = 1,
  66         PERF_COUNT_HW_CACHE_REFERENCES          = 2,
  67         PERF_COUNT_HW_CACHE_MISSES              = 3,
  68         PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
  69         PERF_COUNT_HW_BRANCH_MISSES             = 5,
  70         PERF_COUNT_HW_BUS_CYCLES                = 6,
  71         PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
  72         PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
  73         PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
  74
  75         PERF_COUNT_HW_MAX,                      /* non-ABI */
  76 };
  77
  78 /*
  79  * Generalized hardware cache events:
  80  *
  81  *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
  82  *       { read, write, prefetch } x
  83  *       { accesses, misses }
  84  */
  85 enum perf_hw_cache_id {
  86         PERF_COUNT_HW_CACHE_L1D                 = 0,
  87         PERF_COUNT_HW_CACHE_L1I                 = 1,
  88         PERF_COUNT_HW_CACHE_LL                  = 2,
  89         PERF_COUNT_HW_CACHE_DTLB                = 3,
  90         PERF_COUNT_HW_CACHE_ITLB                = 4,
  91         PERF_COUNT_HW_CACHE_BPU                 = 5,
  92         PERF_COUNT_HW_CACHE_NODE                = 6,
  93
  94         PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
  95 };
  96
  97 enum perf_hw_cache_op_id {
  98         PERF_COUNT_HW_CACHE_OP_READ             = 0,
  99         PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
 100         PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
 101
 102         PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
 103 };
 104
 105 enum perf_hw_cache_op_result_id {
 106         PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
 107         PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
 108
 109         PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
 110 };
 111
 112 /*
 113  * Special "software" events provided by the kernel, even if the hardware
 114  * does not support performance events. These events measure various
 115  * physical and sw events of the kernel (and allow the profiling of them as
 116  * well):
 117  */
 118 enum perf_sw_ids {
 119         PERF_COUNT_SW_CPU_CLOCK                 = 0,
 120         PERF_COUNT_SW_TASK_CLOCK                = 1,
 121         PERF_COUNT_SW_PAGE_FAULTS               = 2,
 122         PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
 123         PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
 124         PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
 125         PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
 126         PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
 127         PERF_COUNT_SW_EMULATION_FAULTS          = 8,
 128         PERF_COUNT_SW_DUMMY                     = 9,
 129         PERF_COUNT_SW_BPF_OUTPUT                = 10,
 130         PERF_COUNT_SW_CGROUP_SWITCHES           = 11,
 131
 132         PERF_COUNT_SW_MAX,                      /* non-ABI */
 133 };
 134
 135 /*
 136  * Bits that can be set in attr.sample_type to request information
 137  * in the overflow packets.
 138  */
 139 enum perf_event_sample_format {
 140         PERF_SAMPLE_IP                          = 1U << 0,
 141         PERF_SAMPLE_TID                         = 1U << 1,
 142         PERF_SAMPLE_TIME                        = 1U << 2,
 143         PERF_SAMPLE_ADDR                        = 1U << 3,
 144         PERF_SAMPLE_READ                        = 1U << 4,
 145         PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
 146         PERF_SAMPLE_ID                          = 1U << 6,
 147         PERF_SAMPLE_CPU                         = 1U << 7,
 148         PERF_SAMPLE_PERIOD                      = 1U << 8,
 149         PERF_SAMPLE_STREAM_ID                   = 1U << 9,
 150         PERF_SAMPLE_RAW                         = 1U << 10,
 151         PERF_SAMPLE_BRANCH_STACK                = 1U << 11,
 152         PERF_SAMPLE_REGS_USER                   = 1U << 12,
 153         PERF_SAMPLE_STACK_USER                  = 1U << 13,
 154         PERF_SAMPLE_WEIGHT                      = 1U << 14,
 155         PERF_SAMPLE_DATA_SRC                    = 1U << 15,
 156         PERF_SAMPLE_IDENTIFIER                  = 1U << 16,
 157         PERF_SAMPLE_TRANSACTION                 = 1U << 17,
 158         PERF_SAMPLE_REGS_INTR                   = 1U << 18,
 159         PERF_SAMPLE_PHYS_ADDR                   = 1U << 19,
 160         PERF_SAMPLE_AUX                         = 1U << 20,
 161         PERF_SAMPLE_CGROUP                      = 1U << 21,
 162         PERF_SAMPLE_DATA_PAGE_SIZE              = 1U << 22,
 163         PERF_SAMPLE_CODE_PAGE_SIZE              = 1U << 23,
 164         PERF_SAMPLE_WEIGHT_STRUCT               = 1U << 24,
 165
 166         PERF_SAMPLE_MAX = 1U << 25,             /* non-ABI */
 167
 168         __PERF_SAMPLE_CALLCHAIN_EARLY           = 1ULL << 63, /* non-ABI; internal use */
 169 };
 170
 171 #define PERF_SAMPLE_WEIGHT_TYPE (PERF_SAMPLE_WEIGHT | PERF_SAMPLE_WEIGHT_STRUCT)
 172 /*
 173  * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
 174  *
 175  * If the user does not pass priv level information via branch_sample_type,
 176  * the kernel uses the event's priv level. Branch and event priv levels do
 177  * not have to match. Branch priv level is checked for permissions.
 178  *
 179  * The branch types can be combined, however BRANCH_ANY covers all types
 180  * of branches and therefore it supersedes all the other types.
 181  */
 182 enum perf_branch_sample_type_shift {
 183         PERF_SAMPLE_BRANCH_USER_SHIFT           = 0, /* user branches */
 184         PERF_SAMPLE_BRANCH_KERNEL_SHIFT         = 1, /* kernel branches */
 185         PERF_SAMPLE_BRANCH_HV_SHIFT             = 2, /* hypervisor branches */
 186
 187         PERF_SAMPLE_BRANCH_ANY_SHIFT            = 3, /* any branch types */
 188         PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT       = 4, /* any call branch */
 189         PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT     = 5, /* any return branch */
 190         PERF_SAMPLE_BRANCH_IND_CALL_SHIFT       = 6, /* indirect calls */
 191         PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT       = 7, /* transaction aborts */
 192         PERF_SAMPLE_BRANCH_IN_TX_SHIFT          = 8, /* in transaction */
 193         PERF_SAMPLE_BRANCH_NO_TX_SHIFT          = 9, /* not in transaction */
 194         PERF_SAMPLE_BRANCH_COND_SHIFT           = 10, /* conditional branches */
 195
 196         PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT     = 11, /* call/ret stack */
 197         PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT       = 12, /* indirect jumps */
 198         PERF_SAMPLE_BRANCH_CALL_SHIFT           = 13, /* direct call */
 199
 200         PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT       = 14, /* no flags */
 201         PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT      = 15, /* no cycles */
 202
 203         PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT      = 16, /* save branch type */
 204
 205         PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT       = 17, /* save low level index of raw branch records */
 206
 207         PERF_SAMPLE_BRANCH_MAX_SHIFT            /* non-ABI */
 208 };
 209
 210 enum perf_branch_sample_type {
 211         PERF_SAMPLE_BRANCH_USER         = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
 212         PERF_SAMPLE_BRANCH_KERNEL       = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
 213         PERF_SAMPLE_BRANCH_HV           = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
 214
 215         PERF_SAMPLE_BRANCH_ANY          = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
 216         PERF_SAMPLE_BRANCH_ANY_CALL     = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
 217         PERF_SAMPLE_BRANCH_ANY_RETURN   = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
 218         PERF_SAMPLE_BRANCH_IND_CALL     = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
 219         PERF_SAMPLE_BRANCH_ABORT_TX     = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
 220         PERF_SAMPLE_BRANCH_IN_TX        = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
 221         PERF_SAMPLE_BRANCH_NO_TX        = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
 222         PERF_SAMPLE_BRANCH_COND         = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
 223
 224         PERF_SAMPLE_BRANCH_CALL_STACK   = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
 225         PERF_SAMPLE_BRANCH_IND_JUMP     = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
 226         PERF_SAMPLE_BRANCH_CALL         = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
 227
 228         PERF_SAMPLE_BRANCH_NO_FLAGS     = 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
 229         PERF_SAMPLE_BRANCH_NO_CYCLES    = 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
 230
 231         PERF_SAMPLE_BRANCH_TYPE_SAVE    =
 232                 1U << PERF_SAMPLE_BRANCH_TYPE_SAVE_SHIFT,
 233
 234         PERF_SAMPLE_BRANCH_HW_INDEX     = 1U << PERF_SAMPLE_BRANCH_HW_INDEX_SHIFT,
 235
 236         PERF_SAMPLE_BRANCH_MAX          = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
 237 };
 238
 239 /*
 240  * Common flow change classification
 241  */
 242 enum {
 243         PERF_BR_UNKNOWN         = 0,    /* unknown */
 244         PERF_BR_COND            = 1,    /* conditional */
 245         PERF_BR_UNCOND          = 2,    /* unconditional  */
 246         PERF_BR_IND             = 3,    /* indirect */
 247         PERF_BR_CALL            = 4,    /* function call */
 248         PERF_BR_IND_CALL        = 5,    /* indirect function call */
 249         PERF_BR_RET             = 6,    /* function return */
 250         PERF_BR_SYSCALL         = 7,    /* syscall */
 251         PERF_BR_SYSRET          = 8,    /* syscall return */
 252         PERF_BR_COND_CALL       = 9,    /* conditional function call */
 253         PERF_BR_COND_RET        = 10,   /* conditional function return */
 254         PERF_BR_MAX,
 255 };
 256
 257 #define PERF_SAMPLE_BRANCH_PLM_ALL \
 258         (PERF_SAMPLE_BRANCH_USER|\
 259          PERF_SAMPLE_BRANCH_KERNEL|\
 260          PERF_SAMPLE_BRANCH_HV)
 261
 262 /*
 263  * Values to determine ABI of the registers dump.
 264  */
 265 enum perf_sample_regs_abi {
 266         PERF_SAMPLE_REGS_ABI_NONE       = 0,
 267         PERF_SAMPLE_REGS_ABI_32         = 1,
 268         PERF_SAMPLE_REGS_ABI_64         = 2,
 269 };
 270
 271 /*
 272  * Values for the memory transaction event qualifier, mostly for
 273  * abort events. Multiple bits can be set.
 274  */
 275 enum {
 276         PERF_TXN_ELISION        = (1 << 0), /* From elision */
 277         PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
 278         PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
 279         PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
 280         PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
 281         PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
 282         PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
 283         PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
 284
 285         PERF_TXN_MAX            = (1 << 8), /* non-ABI */
 286
 287         /* bits 32..63 are reserved for the abort code */
 288
 289         PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
 290         PERF_TXN_ABORT_SHIFT = 32,
 291 };
 292
 293 /*
 294  * The format of the data returned by read() on a perf event fd,
 295  * as specified by attr.read_format:
 296  *
 297  * struct read_format {
 298  *      { u64           value;
 299  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 300  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 301  *        { u64         id;           } && PERF_FORMAT_ID
 302  *      } && !PERF_FORMAT_GROUP
 303  *
 304  *      { u64           nr;
 305  *        { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
 306  *        { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
 307  *        { u64         value;
 308  *          { u64       id;           } && PERF_FORMAT_ID
 309  *        }             cntr[nr];
 310  *      } && PERF_FORMAT_GROUP
 311  * };
 312  */
 313 enum perf_event_read_format {
 314         PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
 315         PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
 316         PERF_FORMAT_ID                          = 1U << 2,
 317         PERF_FORMAT_GROUP                       = 1U << 3,
 318
 319         PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
 320 };
 321
 322 #define PERF_ATTR_SIZE_VER0     64      /* sizeof first published struct */
 323 #define PERF_ATTR_SIZE_VER1     72      /* add: config2 */
 324 #define PERF_ATTR_SIZE_VER2     80      /* add: branch_sample_type */
 325 #define PERF_ATTR_SIZE_VER3     96      /* add: sample_regs_user */
 326                                         /* add: sample_stack_user */
 327 #define PERF_ATTR_SIZE_VER4     104     /* add: sample_regs_intr */
 328 #define PERF_ATTR_SIZE_VER5     112     /* add: aux_watermark */
 329 #define PERF_ATTR_SIZE_VER6     120     /* add: aux_sample_size */
 330 #define PERF_ATTR_SIZE_VER7     128     /* add: sig_data */
 331
 332 /*
 333  * Hardware event_id to monitor via a performance monitoring event:
 334  *
 335  * @sample_max_stack: Max number of frame pointers in a callchain,
 336  *                    should be < /proc/sys/kernel/perf_event_max_stack
 337  */
 338 struct perf_event_attr {
 339
 340         /*
 341          * Major type: hardware/software/tracepoint/etc.
 342          */
 343         __u32                   type;
 344
 345         /*
 346          * Size of the attr structure, for fwd/bwd compat.
 347          */
 348         __u32                   size;
 349
 350         /*
 351          * Type specific configuration information.
 352          */
 353         __u64                   config;
 354
 355         union {
 356                 __u64           sample_period;
 357                 __u64           sample_freq;
 358         };
 359
 360         __u64                   sample_type;
 361         __u64                   read_format;
 362
 363         __u64                   disabled       :  1, /* off by default        */
 364                                 inherit        :  1, /* children inherit it   */
 365                                 pinned         :  1, /* must always be on PMU */
 366                                 exclusive      :  1, /* only group on PMU     */
 367                                 exclude_user   :  1, /* don't count user      */
 368                                 exclude_kernel :  1, /* ditto kernel          */
 369                                 exclude_hv     :  1, /* ditto hypervisor      */
 370                                 exclude_idle   :  1, /* don't count when idle */
 371                                 mmap           :  1, /* include mmap data     */
 372                                 comm           :  1, /* include comm data     */
 373                                 freq           :  1, /* use freq, not period  */
 374                                 inherit_stat   :  1, /* per task counts       */
 375                                 enable_on_exec :  1, /* next exec enables     */
 376                                 task           :  1, /* trace fork/exit       */
 377                                 watermark      :  1, /* wakeup_watermark      */
 378                                 /*
 379                                  * precise_ip:
 380                                  *
 381                                  *  0 - SAMPLE_IP can have arbitrary skid
 382                                  *  1 - SAMPLE_IP must have constant skid
 383                                  *  2 - SAMPLE_IP requested to have 0 skid
 384                                  *  3 - SAMPLE_IP must have 0 skid
 385                                  *
 386                                  *  See also PERF_RECORD_MISC_EXACT_IP
 387                                  */
 388                                 precise_ip     :  2, /* skid constraint       */
 389                                 mmap_data      :  1, /* non-exec mmap data    */
 390                                 sample_id_all  :  1, /* sample_type all events */
 391
 392                                 exclude_host   :  1, /* don't count in host   */
 393                                 exclude_guest  :  1, /* don't count in guest  */
 394
 395                                 exclude_callchain_kernel : 1, /* exclude kernel callchains */
 396                                 exclude_callchain_user   : 1, /* exclude user callchains */
 397                                 mmap2          :  1, /* include mmap with inode data     */
 398                                 comm_exec      :  1, /* flag comm events that are due to an exec */
 399                                 use_clockid    :  1, /* use @clockid for time fields */
 400                                 context_switch :  1, /* context switch data */
 401                                 write_backward :  1, /* Write ring buffer from end to beginning */
 402                                 namespaces     :  1, /* include namespaces data */
 403                                 ksymbol        :  1, /* include ksymbol events */
 404                                 bpf_event      :  1, /* include bpf events */
 405                                 aux_output     :  1, /* generate AUX records instead of events */
 406                                 cgroup         :  1, /* include cgroup events */
 407                                 text_poke      :  1, /* include text poke events */
 408                                 build_id       :  1, /* use build id in mmap2 events */
 409                                 inherit_thread :  1, /* children only inherit if cloned with CLONE_THREAD */
 410                                 remove_on_exec :  1, /* event is removed from task on exec */
 411                                 sigtrap        :  1, /* send synchronous SIGTRAP on event */
 412                                 __reserved_1   : 26;
 413
 414         union {
 415                 __u32           wakeup_events;    /* wakeup every n events */
 416                 __u32           wakeup_watermark; /* bytes before wakeup   */
 417         };
 418
 419         __u32                   bp_type;
 420         union {
 421                 __u64           bp_addr;
 422                 __u64           kprobe_func; /* for perf_kprobe */
 423                 __u64           uprobe_path; /* for perf_uprobe */
 424                 __u64           config1; /* extension of config */
 425         };
 426         union {
 427                 __u64           bp_len;
 428                 __u64           kprobe_addr; /* when kprobe_func == NULL */
 429                 __u64           probe_offset; /* for perf_[k,u]probe */
 430                 __u64           config2; /* extension of config1 */
 431         };
 432         __u64   branch_sample_type; /* enum perf_branch_sample_type */
 433
 434         /*
 435          * Defines set of user regs to dump on samples.
 436          * See asm/perf_regs.h for details.
 437          */
 438         __u64   sample_regs_user;
 439
 440         /*
 441          * Defines size of the user stack to dump on samples.
 442          */
 443         __u32   sample_stack_user;
 444
 445         __s32   clockid;
 446         /*
 447          * Defines set of regs to dump for each sample
 448          * state captured on:
 449          *  - precise = 0: PMU interrupt
 450          *  - precise > 0: sampled instruction
 451          *
 452          * See asm/perf_regs.h for details.
 453          */
 454         __u64   sample_regs_intr;
 455
 456         /*
 457          * Wakeup watermark for AUX area
 458          */
 459         __u32   aux_watermark;
 460         __u16   sample_max_stack;
 461         __u16   __reserved_2;
 462         __u32   aux_sample_size;
 463         __u32   __reserved_3;
 464
 465         /*
 466          * User provided data if sigtrap=1, passed back to user via
 467          * siginfo_t::si_perf_data, e.g. to permit user to identify the event.
 468          */
 469         __u64   sig_data;
 470 };
 471
 472 /*
 473  * Structure used by below PERF_EVENT_IOC_QUERY_BPF command
 474  * to query bpf programs attached to the same perf tracepoint
 475  * as the given perf event.
 476  */
 477 struct perf_event_query_bpf {
 478         /*
 479          * The below ids array length
 480          */
 481         __u32   ids_len;
 482         /*
 483          * Set by the kernel to indicate the number of
 484          * available programs
 485          */
 486         __u32   prog_cnt;
 487         /*
 488          * User provided buffer to store program ids
 489          */
 490         __u32   ids[0];
 491 };
 492
 493 /*
 494  * Ioctls that can be done on a perf event fd:
 495  */
 496 #define PERF_EVENT_IOC_ENABLE                   _IO ('$', 0)
 497 #define PERF_EVENT_IOC_DISABLE                  _IO ('$', 1)
 498 #define PERF_EVENT_IOC_REFRESH                  _IO ('$', 2)
 499 #define PERF_EVENT_IOC_RESET                    _IO ('$', 3)
 500 #define PERF_EVENT_IOC_PERIOD                   _IOW('$', 4, __u64)
 501 #define PERF_EVENT_IOC_SET_OUTPUT               _IO ('$', 5)
 502 #define PERF_EVENT_IOC_SET_FILTER               _IOW('$', 6, char *)
 503 #define PERF_EVENT_IOC_ID                       _IOR('$', 7, __u64 *)
 504 #define PERF_EVENT_IOC_SET_BPF                  _IOW('$', 8, __u32)
 505 #define PERF_EVENT_IOC_PAUSE_OUTPUT             _IOW('$', 9, __u32)
 506 #define PERF_EVENT_IOC_QUERY_BPF                _IOWR('$', 10, struct perf_event_query_bpf *)
 507 #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES        _IOW('$', 11, struct perf_event_attr *)
 508
 509 enum perf_event_ioc_flags {
 510         PERF_IOC_FLAG_GROUP             = 1U << 0,
 511 };
 512
 513 /*
 514  * Structure of the page that can be mapped via mmap
 515  */
 516 struct perf_event_mmap_page {
 517         __u32   version;                /* version number of this structure */
 518         __u32   compat_version;         /* lowest version this is compat with */
 519
 520         /*
 521          * Bits needed to read the hw events in user-space.
 522          *
 523          *   u32 seq, time_mult, time_shift, index, width;
 524          *   u64 count, enabled, running;
 525          *   u64 cyc, time_offset;
 526          *   s64 pmc = 0;
 527          *
 528          *   do {
 529          *     seq = pc->lock;
 530          *     barrier()
 531          *
 532          *     enabled = pc->time_enabled;
 533          *     running = pc->time_running;
 534          *
 535          *     if (pc->cap_usr_time && enabled != running) {
 536          *       cyc = rdtsc();
 537          *       time_offset = pc->time_offset;
 538          *       time_mult   = pc->time_mult;
 539          *       time_shift  = pc->time_shift;
 540          *     }
 541          *
 542          *     index = pc->index;
 543          *     count = pc->offset;
 544          *     if (pc->cap_user_rdpmc && index) {
 545          *       width = pc->pmc_width;
 546          *       pmc = rdpmc(index - 1);
 547          *     }
 548          *
 549          *     barrier();
 550          *   } while (pc->lock != seq);
 551          *
 552          * NOTE: for obvious reason this only works on self-monitoring
 553          *       processes.
 554          */
 555         __u32   lock;                   /* seqlock for synchronization */
 556         __u32   index;                  /* hardware event identifier */
 557         __s64   offset;                 /* add to hardware event value */
 558         __u64   time_enabled;           /* time event active */
 559         __u64   time_running;           /* time event on cpu */
 560         union {
 561                 __u64   capabilities;
 562                 struct {
 563                         __u64   cap_bit0                : 1, /* Always 0, deprecated, see commit 860f085b74e9 */
 564                                 cap_bit0_is_deprecated  : 1, /* Always 1, signals that bit 0 is zero */
 565
 566                                 cap_user_rdpmc          : 1, /* The RDPMC instruction can be used to read counts */
 567                                 cap_user_time           : 1, /* The time_{shift,mult,offset} fields are used */
 568                                 cap_user_time_zero      : 1, /* The time_zero field is used */
 569                                 cap_user_time_short     : 1, /* the time_{cycle,mask} fields are used */
 570                                 cap_____res             : 58;
 571                 };
 572         };
 573
 574         /*
 575          * If cap_user_rdpmc this field provides the bit-width of the value
 576          * read using the rdpmc() or equivalent instruction. This can be used
 577          * to sign extend the result like:
 578          *
 579          *   pmc <<= 64 - width;
 580          *   pmc >>= 64 - width; // signed shift right
 581          *   count += pmc;
 582          */
 583         __u16   pmc_width;
 584
 585         /*
 586          * If cap_usr_time the below fields can be used to compute the time
 587          * delta since time_enabled (in ns) using rdtsc or similar.
 588          *
 589          *   u64 quot, rem;
 590          *   u64 delta;
 591          *
 592          *   quot = (cyc >> time_shift);
 593          *   rem = cyc & (((u64)1 << time_shift) - 1);
 594          *   delta = time_offset + quot * time_mult +
 595          *              ((rem * time_mult) >> time_shift);
 596          *
 597          * Where time_offset,time_mult,time_shift and cyc are read in the
 598          * seqcount loop described above. This delta can then be added to
 599          * enabled and possible running (if index), improving the scaling:
 600          *
 601          *   enabled += delta;
 602          *   if (index)
 603          *     running += delta;
 604          *
 605          *   quot = count / running;
 606          *   rem  = count % running;
 607          *   count = quot * enabled + (rem * enabled) / running;
 608          */
 609         __u16   time_shift;
 610         __u32   time_mult;
 611         __u64   time_offset;
 612         /*
 613          * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
 614          * from sample timestamps.
 615          *
 616          *   time = timestamp - time_zero;
 617          *   quot = time / time_mult;
 618          *   rem  = time % time_mult;
 619          *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
 620          *
 621          * And vice versa:
 622          *
 623          *   quot = cyc >> time_shift;
 624          *   rem  = cyc & (((u64)1 << time_shift) - 1);
 625          *   timestamp = time_zero + quot * time_mult +
 626          *               ((rem * time_mult) >> time_shift);
 627          */
 628         __u64   time_zero;
 629
 630         __u32   size;                   /* Header size up to __reserved[] fields. */
 631         __u32   __reserved_1;
 632
 633         /*
 634          * If cap_usr_time_short, the hardware clock is less than 64bit wide
 635          * and we must compute the 'cyc' value, as used by cap_usr_time, as:
 636          *
 637          *   cyc = time_cycles + ((cyc - time_cycles) & time_mask)
 638          *
 639          * NOTE: this form is explicitly chosen such that cap_usr_time_short
 640          *       is a correction on top of cap_usr_time, and code that doesn't
 641          *       know about cap_usr_time_short still works under the assumption
 642          *       the counter doesn't wrap.
 643          */
 644         __u64   time_cycles;
 645         __u64   time_mask;
 646
 647                 /*
 648                  * Hole for extension of the self monitor capabilities
 649                  */
 650
 651         __u8    __reserved[116*8];      /* align to 1k. */
 652
 653         /*
 654          * Control data for the mmap() data buffer.
 655          *
 656          * User-space reading the @data_head value should issue an smp_rmb(),
 657          * after reading this value.
 658          *
 659          * When the mapping is PROT_WRITE the @data_tail value should be
 660          * written by userspace to reflect the last read data, after issueing
 661          * an smp_mb() to separate the data read from the ->data_tail store.
 662          * In this case the kernel will not over-write unread data.
 663          *
 664          * See perf_output_put_handle() for the data ordering.
 665          *
 666          * data_{offset,size} indicate the location and size of the perf record
 667          * buffer within the mmapped area.
 668          */
 669         __u64   data_head;              /* head in the data section */
 670         __u64   data_tail;              /* user-space written tail */
 671         __u64   data_offset;            /* where the buffer starts */
 672         __u64   data_size;              /* data buffer size */
 673
 674         /*
 675          * AUX area is defined by aux_{offset,size} fields that should be set
 676          * by the userspace, so that
 677          *
 678          *   aux_offset >= data_offset + data_size
 679          *
 680          * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
 681          *
 682          * Ring buffer pointers aux_{head,tail} have the same semantics as
 683          * data_{head,tail} and same ordering rules apply.
 684          */
 685         __u64   aux_head;
 686         __u64   aux_tail;
 687         __u64   aux_offset;
 688         __u64   aux_size;
 689 };
 690
 691 /*
 692  * The current state of perf_event_header::misc bits usage:
 693  * ('|' used bit, '-' unused bit)
 694  *
 695  *  012         CDEF
 696  *  |||---------||||
 697  *
 698  *  Where:
 699  *    0-2     CPUMODE_MASK
 700  *
 701  *    C       PROC_MAP_PARSE_TIMEOUT
 702  *    D       MMAP_DATA / COMM_EXEC / FORK_EXEC / SWITCH_OUT
 703  *    E       MMAP_BUILD_ID / EXACT_IP / SCHED_OUT_PREEMPT
 704  *    F       (reserved)
 705  */
 706
 707 #define PERF_RECORD_MISC_CPUMODE_MASK           (7 << 0)
 708 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN        (0 << 0)
 709 #define PERF_RECORD_MISC_KERNEL                 (1 << 0)
 710 #define PERF_RECORD_MISC_USER                   (2 << 0)
 711 #define PERF_RECORD_MISC_HYPERVISOR             (3 << 0)
 712 #define PERF_RECORD_MISC_GUEST_KERNEL           (4 << 0)
 713 #define PERF_RECORD_MISC_GUEST_USER             (5 << 0)
 714
 715 /*
 716  * Indicates that /proc/PID/maps parsing are truncated by time out.
 717  */
 718 #define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT (1 << 12)
 719 /*
 720  * Following PERF_RECORD_MISC_* are used on different
 721  * events, so can reuse the same bit position:
 722  *
 723  *   PERF_RECORD_MISC_MMAP_DATA  - PERF_RECORD_MMAP* events
 724  *   PERF_RECORD_MISC_COMM_EXEC  - PERF_RECORD_COMM event
 725  *   PERF_RECORD_MISC_FORK_EXEC  - PERF_RECORD_FORK event (perf internal)
 726  *   PERF_RECORD_MISC_SWITCH_OUT - PERF_RECORD_SWITCH* events
 727  */
 728 #define PERF_RECORD_MISC_MMAP_DATA              (1 << 13)
 729 #define PERF_RECORD_MISC_COMM_EXEC              (1 << 13)
 730 #define PERF_RECORD_MISC_FORK_EXEC              (1 << 13)
 731 #define PERF_RECORD_MISC_SWITCH_OUT             (1 << 13)
 732 /*
 733  * These PERF_RECORD_MISC_* flags below are safely reused
 734  * for the following events:
 735  *
 736  *   PERF_RECORD_MISC_EXACT_IP           - PERF_RECORD_SAMPLE of precise events
 737  *   PERF_RECORD_MISC_SWITCH_OUT_PREEMPT - PERF_RECORD_SWITCH* events
 738  *   PERF_RECORD_MISC_MMAP_BUILD_ID      - PERF_RECORD_MMAP2 event
 739  *
 740  *
 741  * PERF_RECORD_MISC_EXACT_IP:
 742  *   Indicates that the content of PERF_SAMPLE_IP points to
 743  *   the actual instruction that triggered the event. See also
 744  *   perf_event_attr::precise_ip.
 745  *
 746  * PERF_RECORD_MISC_SWITCH_OUT_PREEMPT:
 747  *   Indicates that thread was preempted in TASK_RUNNING state.
 748  *
 749  * PERF_RECORD_MISC_MMAP_BUILD_ID:
 750  *   Indicates that mmap2 event carries build id data.
 751  */
 752 #define PERF_RECORD_MISC_EXACT_IP               (1 << 14)
 753 #define PERF_RECORD_MISC_SWITCH_OUT_PREEMPT     (1 << 14)
 754 #define PERF_RECORD_MISC_MMAP_BUILD_ID          (1 << 14)
 755 /*
 756  * Reserve the last bit to indicate some extended misc field
 757  */
 758 #define PERF_RECORD_MISC_EXT_RESERVED           (1 << 15)
 759
 760 struct perf_event_header {
 761         __u32   type;
 762         __u16   misc;
 763         __u16   size;
 764 };
 765
 766 struct perf_ns_link_info {
 767         __u64   dev;
 768         __u64   ino;
 769 };
 770
 771 enum {
 772         NET_NS_INDEX            = 0,
 773         UTS_NS_INDEX            = 1,
 774         IPC_NS_INDEX            = 2,
 775         PID_NS_INDEX            = 3,
 776         USER_NS_INDEX           = 4,
 777         MNT_NS_INDEX            = 5,
 778         CGROUP_NS_INDEX         = 6,
 779
 780         NR_NAMESPACES,          /* number of available namespaces */
 781 };
 782
 783 enum perf_event_type {
 784
 785         /*
 786          * If perf_event_attr.sample_id_all is set then all event types will
 787          * have the sample_type selected fields related to where/when
 788          * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
 789          * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
 790          * just after the perf_event_header and the fields already present for
 791          * the existing fields, i.e. at the end of the payload. That way a newer
 792          * perf.data file will be supported by older perf tools, with these new
 793          * optional fields being ignored.
 794          *
 795          * struct sample_id {
 796          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 797          *      { u64                   time;     } && PERF_SAMPLE_TIME
 798          *      { u64                   id;       } && PERF_SAMPLE_ID
 799          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 800          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 801          *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
 802          * } && perf_event_attr::sample_id_all
 803          *
 804          * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
 805          * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
 806          * relative to header.size.
 807          */
 808
 809         /*
 810          * The MMAP events record the PROT_EXEC mappings so that we can
 811          * correlate userspace IPs to code. They have the following structure:
 812          *
 813          * struct {
 814          *      struct perf_event_header        header;
 815          *
 816          *      u32                             pid, tid;
 817          *      u64                             addr;
 818          *      u64                             len;
 819          *      u64                             pgoff;
 820          *      char                            filename[];
 821          *      struct sample_id                sample_id;
 822          * };
 823          */
 824         PERF_RECORD_MMAP                        = 1,
 825
 826         /*
 827          * struct {
 828          *      struct perf_event_header        header;
 829          *      u64                             id;
 830          *      u64                             lost;
 831          *      struct sample_id                sample_id;
 832          * };
 833          */
 834         PERF_RECORD_LOST                        = 2,
 835
 836         /*
 837          * struct {
 838          *      struct perf_event_header        header;
 839          *
 840          *      u32                             pid, tid;
 841          *      char                            comm[];
 842          *      struct sample_id                sample_id;
 843          * };
 844          */
 845         PERF_RECORD_COMM                        = 3,
 846
 847         /*
 848          * struct {
 849          *      struct perf_event_header        header;
 850          *      u32                             pid, ppid;
 851          *      u32                             tid, ptid;
 852          *      u64                             time;
 853          *      struct sample_id                sample_id;
 854          * };
 855          */
 856         PERF_RECORD_EXIT                        = 4,
 857
 858         /*
 859          * struct {
 860          *      struct perf_event_header        header;
 861          *      u64                             time;
 862          *      u64                             id;
 863          *      u64                             stream_id;
 864          *      struct sample_id                sample_id;
 865          * };
 866          */
 867         PERF_RECORD_THROTTLE                    = 5,
 868         PERF_RECORD_UNTHROTTLE                  = 6,
 869
 870         /*
 871          * struct {
 872          *      struct perf_event_header        header;
 873          *      u32                             pid, ppid;
 874          *      u32                             tid, ptid;
 875          *      u64                             time;
 876          *      struct sample_id                sample_id;
 877          * };
 878          */
 879         PERF_RECORD_FORK                        = 7,
 880
 881         /*
 882          * struct {
 883          *      struct perf_event_header        header;
 884          *      u32                             pid, tid;
 885          *
 886          *      struct read_format              values;
 887          *      struct sample_id                sample_id;
 888          * };
 889          */
 890         PERF_RECORD_READ                        = 8,
 891
 892         /*
 893          * struct {
 894          *      struct perf_event_header        header;
 895          *
 896          *      #
 897          *      # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
 898          *      # The advantage of PERF_SAMPLE_IDENTIFIER is that its position
 899          *      # is fixed relative to header.
 900          *      #
 901          *
 902          *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
 903          *      { u64                   ip;       } && PERF_SAMPLE_IP
 904          *      { u32                   pid, tid; } && PERF_SAMPLE_TID
 905          *      { u64                   time;     } && PERF_SAMPLE_TIME
 906          *      { u64                   addr;     } && PERF_SAMPLE_ADDR
 907          *      { u64                   id;       } && PERF_SAMPLE_ID
 908          *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
 909          *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
 910          *      { u64                   period;   } && PERF_SAMPLE_PERIOD
 911          *
 912          *      { struct read_format    values;   } && PERF_SAMPLE_READ
 913          *
 914          *      { u64                   nr,
 915          *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
 916          *
 917          *      #
 918          *      # The RAW record below is opaque data wrt the ABI
 919          *      #
 920          *      # That is, the ABI doesn't make any promises wrt to
 921          *      # the stability of its content, it may vary depending
 922          *      # on event, hardware, kernel version and phase of
 923          *      # the moon.
 924          *      #
 925          *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
 926          *      #
 927          *
 928          *      { u32                   size;
 929          *        char                  data[size];}&& PERF_SAMPLE_RAW
 930          *
 931          *      { u64                   nr;
 932          *        { u64 hw_idx; } && PERF_SAMPLE_BRANCH_HW_INDEX
 933          *        { u64 from, to, flags } lbr[nr];
 934          *      } && PERF_SAMPLE_BRANCH_STACK
 935          *
 936          *      { u64                   abi; # enum perf_sample_regs_abi
 937          *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
 938          *
 939          *      { u64                   size;
 940          *        char                  data[size];
 941          *        u64                   dyn_size; } && PERF_SAMPLE_STACK_USER
 942          *
 943          *      { union perf_sample_weight
 944          *       {
 945          *              u64             full; && PERF_SAMPLE_WEIGHT
 946          *      #if defined(__LITTLE_ENDIAN_BITFIELD)
 947          *              struct {
 948          *                      u32     var1_dw;
 949          *                      u16     var2_w;
 950          *                      u16     var3_w;
 951          *              } && PERF_SAMPLE_WEIGHT_STRUCT
 952          *      #elif defined(__BIG_ENDIAN_BITFIELD)
 953          *              struct {
 954          *                      u16     var3_w;
 955          *                      u16     var2_w;
 956          *                      u32     var1_dw;
 957          *              } && PERF_SAMPLE_WEIGHT_STRUCT
 958          *      #endif
 959          *       }
 960          *      }
 961          *      { u64                   data_src; } && PERF_SAMPLE_DATA_SRC
 962          *      { u64                   transaction; } && PERF_SAMPLE_TRANSACTION
 963          *      { u64                   abi; # enum perf_sample_regs_abi
 964          *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
 965          *      { u64                   phys_addr;} && PERF_SAMPLE_PHYS_ADDR
 966          *      { u64                   size;
 967          *        char                  data[size]; } && PERF_SAMPLE_AUX
 968          *      { u64                   data_page_size;} && PERF_SAMPLE_DATA_PAGE_SIZE
 969          *      { u64                   code_page_size;} && PERF_SAMPLE_CODE_PAGE_SIZE
 970          * };
 971          */
 972         PERF_RECORD_SAMPLE                      = 9,
 973
 974         /*
 975          * The MMAP2 records are an augmented version of MMAP, they add
 976          * maj, min, ino numbers to be used to uniquely identify each mapping
 977          *
 978          * struct {
 979          *      struct perf_event_header        header;
 980          *
 981          *      u32                             pid, tid;
 982          *      u64                             addr;
 983          *      u64                             len;
 984          *      u64                             pgoff;
 985          *      union {
 986          *              struct {
 987          *                      u32             maj;
 988          *                      u32             min;
 989          *                      u64             ino;
 990          *                      u64             ino_generation;
 991          *              };
 992          *              struct {
 993          *                      u8              build_id_size;
 994          *                      u8              __reserved_1;
 995          *                      u16             __reserved_2;
 996          *                      u8              build_id[20];
 997          *              };
 998          *      };
 999          *      u32                             prot, flags;
1000          *      char                            filename[];
1001          *      struct sample_id                sample_id;
1002          * };
1003          */
1004         PERF_RECORD_MMAP2                       = 10,
1005
1006         /*
1007          * Records that new data landed in the AUX buffer part.
1008          *
1009          * struct {
1010          *      struct perf_event_header        header;
1011          *
1012          *      u64                             aux_offset;
1013          *      u64                             aux_size;
1014          *      u64                             flags;
1015          *      struct sample_id                sample_id;
1016          * };
1017          */
1018         PERF_RECORD_AUX                         = 11,
1019
1020         /*
1021          * Indicates that instruction trace has started
1022          *
1023          * struct {
1024          *      struct perf_event_header        header;
1025          *      u32                             pid;
1026          *      u32                             tid;
1027          *      struct sample_id                sample_id;
1028          * };
1029          */
1030         PERF_RECORD_ITRACE_START                = 12,
1031
1032         /*
1033          * Records the dropped/lost sample number.
1034          *
1035          * struct {
1036          *      struct perf_event_header        header;
1037          *
1038          *      u64                             lost;
1039          *      struct sample_id                sample_id;
1040          * };
1041          */
1042         PERF_RECORD_LOST_SAMPLES                = 13,
1043
1044         /*
1045          * Records a context switch in or out (flagged by
1046          * PERF_RECORD_MISC_SWITCH_OUT). See also
1047          * PERF_RECORD_SWITCH_CPU_WIDE.
1048          *
1049          * struct {
1050          *      struct perf_event_header        header;
1051          *      struct sample_id                sample_id;
1052          * };
1053          */
1054         PERF_RECORD_SWITCH                      = 14,
1055
1056         /*
1057          * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
1058          * next_prev_tid that are the next (switching out) or previous
1059          * (switching in) pid/tid.
1060          *
1061          * struct {
1062          *      struct perf_event_header        header;
1063          *      u32                             next_prev_pid;
1064          *      u32                             next_prev_tid;
1065          *      struct sample_id                sample_id;
1066          * };
1067          */
1068         PERF_RECORD_SWITCH_CPU_WIDE             = 15,
1069
1070         /*
1071          * struct {
1072          *      struct perf_event_header        header;
1073          *      u32                             pid;
1074          *      u32                             tid;
1075          *      u64                             nr_namespaces;
1076          *      { u64                           dev, inode; } [nr_namespaces];
1077          *      struct sample_id                sample_id;
1078          * };
1079          */
1080         PERF_RECORD_NAMESPACES                  = 16,
1081
1082         /*
1083          * Record ksymbol register/unregister events:
1084          *
1085          * struct {
1086          *      struct perf_event_header        header;
1087          *      u64                             addr;
1088          *      u32                             len;
1089          *      u16                             ksym_type;
1090          *      u16                             flags;
1091          *      char                            name[];
1092          *      struct sample_id                sample_id;
1093          * };
1094          */
1095         PERF_RECORD_KSYMBOL                     = 17,
1096
1097         /*
1098          * Record bpf events:
1099          *  enum perf_bpf_event_type {
1100          *      PERF_BPF_EVENT_UNKNOWN          = 0,
1101          *      PERF_BPF_EVENT_PROG_LOAD        = 1,
1102          *      PERF_BPF_EVENT_PROG_UNLOAD      = 2,
1103          *  };
1104          *
1105          * struct {
1106          *      struct perf_event_header        header;
1107          *      u16                             type;
1108          *      u16                             flags;
1109          *      u32                             id;
1110          *      u8                              tag[BPF_TAG_SIZE];
1111          *      struct sample_id                sample_id;
1112          * };
1113          */
1114         PERF_RECORD_BPF_EVENT                   = 18,
1115
1116         /*
1117          * struct {
1118          *      struct perf_event_header        header;
1119          *      u64                             id;
1120          *      char                            path[];
1121          *      struct sample_id                sample_id;
1122          * };
1123          */
1124         PERF_RECORD_CGROUP                      = 19,
1125
1126         /*
1127          * Records changes to kernel text i.e. self-modified code. 'old_len' is
1128          * the number of old bytes, 'new_len' is the number of new bytes. Either
1129          * 'old_len' or 'new_len' may be zero to indicate, for example, the
1130          * addition or removal of a trampoline. 'bytes' contains the old bytes
1131          * followed immediately by the new bytes.
1132          *
1133          * struct {
1134          *      struct perf_event_header        header;
1135          *      u64                             addr;
1136          *      u16                             old_len;
1137          *      u16                             new_len;
1138          *      u8                              bytes[];
1139          *      struct sample_id                sample_id;
1140          * };
1141          */
1142         PERF_RECORD_TEXT_POKE                   = 20,
1143
1144         /*
1145          * Data written to the AUX area by hardware due to aux_output, may need
1146          * to be matched to the event by an architecture-specific hardware ID.
1147          * This records the hardware ID, but requires sample_id to provide the
1148          * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT
1149          * records from multiple events.
1150          *
1151          * struct {
1152          *      struct perf_event_header        header;
1153          *      u64                             hw_id;
1154          *      struct sample_id                sample_id;
1155          * };
1156          */
1157         PERF_RECORD_AUX_OUTPUT_HW_ID            = 21,
1158
1159         PERF_RECORD_MAX,                        /* non-ABI */
1160 };
1161
1162 enum perf_record_ksymbol_type {
1163         PERF_RECORD_KSYMBOL_TYPE_UNKNOWN        = 0,
1164         PERF_RECORD_KSYMBOL_TYPE_BPF            = 1,
1165         /*
1166          * Out of line code such as kprobe-replaced instructions or optimized
1167          * kprobes or ftrace trampolines.
1168          */
1169         PERF_RECORD_KSYMBOL_TYPE_OOL            = 2,
1170         PERF_RECORD_KSYMBOL_TYPE_MAX            /* non-ABI */
1171 };
1172
1173 #define PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER    (1 << 0)
1174
1175 enum perf_bpf_event_type {
1176         PERF_BPF_EVENT_UNKNOWN          = 0,
1177         PERF_BPF_EVENT_PROG_LOAD        = 1,
1178         PERF_BPF_EVENT_PROG_UNLOAD      = 2,
1179         PERF_BPF_EVENT_MAX,             /* non-ABI */
1180 };
1181
1182 #define PERF_MAX_STACK_DEPTH            127
1183 #define PERF_MAX_CONTEXTS_PER_STACK       8
1184
1185 enum perf_callchain_context {
1186         PERF_CONTEXT_HV                 = (__u64)-32,
1187         PERF_CONTEXT_KERNEL             = (__u64)-128,
1188         PERF_CONTEXT_USER               = (__u64)-512,
1189
1190         PERF_CONTEXT_GUEST              = (__u64)-2048,
1191         PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
1192         PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
1193
1194         PERF_CONTEXT_MAX                = (__u64)-4095,
1195 };
1196
1197 /**
1198  * PERF_RECORD_AUX::flags bits
1199  */
1200 #define PERF_AUX_FLAG_TRUNCATED                 0x01    /* record was truncated to fit */
1201 #define PERF_AUX_FLAG_OVERWRITE                 0x02    /* snapshot from overwrite mode */
1202 #define PERF_AUX_FLAG_PARTIAL                   0x04    /* record contains gaps */
1203 #define PERF_AUX_FLAG_COLLISION                 0x08    /* sample collided with another */
1204 #define PERF_AUX_FLAG_PMU_FORMAT_TYPE_MASK      0xff00  /* PMU specific trace format type */
1205
1206 /* CoreSight PMU AUX buffer formats */
1207 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_CORESIGHT        0x0000 /* Default for backward compatibility */
1208 #define PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW              0x0100 /* Raw format of the source */
1209
1210 #define PERF_FLAG_FD_NO_GROUP           (1UL << 0)
1211 #define PERF_FLAG_FD_OUTPUT             (1UL << 1)
1212 #define PERF_FLAG_PID_CGROUP            (1UL << 2) /* pid=cgroup id, per-cpu mode only */
1213 #define PERF_FLAG_FD_CLOEXEC            (1UL << 3) /* O_CLOEXEC */
1214
1215 #if defined(__LITTLE_ENDIAN_BITFIELD)
1216 union perf_mem_data_src {
1217         __u64 val;
1218         struct {
1219                 __u64   mem_op:5,       /* type of opcode */
1220                         mem_lvl:14,     /* memory hierarchy level */
1221                         mem_snoop:5,    /* snoop mode */
1222                         mem_lock:2,     /* lock instr */
1223                         mem_dtlb:7,     /* tlb access */
1224                         mem_lvl_num:4,  /* memory hierarchy level number */
1225                         mem_remote:1,   /* remote */
1226                         mem_snoopx:2,   /* snoop mode, ext */
1227                         mem_blk:3,      /* access blocked */
1228                         mem_hops:3,     /* hop level */
1229                         mem_rsvd:18;
1230         };
1231 };
1232 #elif defined(__BIG_ENDIAN_BITFIELD)
1233 union perf_mem_data_src {
1234         __u64 val;
1235         struct {
1236                 __u64   mem_rsvd:18,
1237                         mem_hops:3,     /* hop level */
1238                         mem_blk:3,      /* access blocked */
1239                         mem_snoopx:2,   /* snoop mode, ext */
1240                         mem_remote:1,   /* remote */
1241                         mem_lvl_num:4,  /* memory hierarchy level number */
1242                         mem_dtlb:7,     /* tlb access */
1243                         mem_lock:2,     /* lock instr */
1244                         mem_snoop:5,    /* snoop mode */
1245                         mem_lvl:14,     /* memory hierarchy level */
1246                         mem_op:5;       /* type of opcode */
1247         };
1248 };
1249 #else
1250 #error "Unknown endianness"
1251 #endif
1252
1253 /* type of opcode (load/store/prefetch,code) */
1254 #define PERF_MEM_OP_NA          0x01 /* not available */
1255 #define PERF_MEM_OP_LOAD        0x02 /* load instruction */
1256 #define PERF_MEM_OP_STORE       0x04 /* store instruction */
1257 #define PERF_MEM_OP_PFETCH      0x08 /* prefetch */
1258 #define PERF_MEM_OP_EXEC        0x10 /* code (execution) */
1259 #define PERF_MEM_OP_SHIFT       0
1260
1261 /*
1262  * PERF_MEM_LVL_* namespace being depricated to some extent in the
1263  * favour of newer composite PERF_MEM_{LVLNUM_,REMOTE_,SNOOPX_} fields.
1264  * Supporting this namespace inorder to not break defined ABIs.
1265  *
1266  * memory hierarchy (memory level, hit or miss)
1267  */
1268 #define PERF_MEM_LVL_NA         0x01  /* not available */
1269 #define PERF_MEM_LVL_HIT        0x02  /* hit level */
1270 #define PERF_MEM_LVL_MISS       0x04  /* miss level  */
1271 #define PERF_MEM_LVL_L1         0x08  /* L1 */
1272 #define PERF_MEM_LVL_LFB        0x10  /* Line Fill Buffer */
1273 #define PERF_MEM_LVL_L2         0x20  /* L2 */
1274 #define PERF_MEM_LVL_L3         0x40  /* L3 */
1275 #define PERF_MEM_LVL_LOC_RAM    0x80  /* Local DRAM */
1276 #define PERF_MEM_LVL_REM_RAM1   0x100 /* Remote DRAM (1 hop) */
1277 #define PERF_MEM_LVL_REM_RAM2   0x200 /* Remote DRAM (2 hops) */
1278 #define PERF_MEM_LVL_REM_CCE1   0x400 /* Remote Cache (1 hop) */
1279 #define PERF_MEM_LVL_REM_CCE2   0x800 /* Remote Cache (2 hops) */
1280 #define PERF_MEM_LVL_IO         0x1000 /* I/O memory */
1281 #define PERF_MEM_LVL_UNC        0x2000 /* Uncached memory */
1282 #define PERF_MEM_LVL_SHIFT      5
1283
1284 #define PERF_MEM_REMOTE_REMOTE  0x01  /* Remote */
1285 #define PERF_MEM_REMOTE_SHIFT   37
1286
1287 #define PERF_MEM_LVLNUM_L1      0x01 /* L1 */
1288 #define PERF_MEM_LVLNUM_L2      0x02 /* L2 */
1289 #define PERF_MEM_LVLNUM_L3      0x03 /* L3 */
1290 #define PERF_MEM_LVLNUM_L4      0x04 /* L4 */
1291 /* 5-0xa available */
1292 #define PERF_MEM_LVLNUM_ANY_CACHE 0x0b /* Any cache */
1293 #define PERF_MEM_LVLNUM_LFB     0x0c /* LFB */
1294 #define PERF_MEM_LVLNUM_RAM     0x0d /* RAM */
1295 #define PERF_MEM_LVLNUM_PMEM    0x0e /* PMEM */
1296 #define PERF_MEM_LVLNUM_NA      0x0f /* N/A */
1297
1298 #define PERF_MEM_LVLNUM_SHIFT   33
1299
1300 /* snoop mode */
1301 #define PERF_MEM_SNOOP_NA       0x01 /* not available */
1302 #define PERF_MEM_SNOOP_NONE     0x02 /* no snoop */
1303 #define PERF_MEM_SNOOP_HIT      0x04 /* snoop hit */
1304 #define PERF_MEM_SNOOP_MISS     0x08 /* snoop miss */
1305 #define PERF_MEM_SNOOP_HITM     0x10 /* snoop hit modified */
1306 #define PERF_MEM_SNOOP_SHIFT    19
1307
1308 #define PERF_MEM_SNOOPX_FWD     0x01 /* forward */
1309 /* 1 free */
1310 #define PERF_MEM_SNOOPX_SHIFT  38
1311
1312 /* locked instruction */
1313 #define PERF_MEM_LOCK_NA        0x01 /* not available */
1314 #define PERF_MEM_LOCK_LOCKED    0x02 /* locked transaction */
1315 #define PERF_MEM_LOCK_SHIFT     24
1316
1317 /* TLB access */
1318 #define PERF_MEM_TLB_NA         0x01 /* not available */
1319 #define PERF_MEM_TLB_HIT        0x02 /* hit level */
1320 #define PERF_MEM_TLB_MISS       0x04 /* miss level */
1321 #define PERF_MEM_TLB_L1         0x08 /* L1 */
1322 #define PERF_MEM_TLB_L2         0x10 /* L2 */
1323 #define PERF_MEM_TLB_WK         0x20 /* Hardware Walker*/
1324 #define PERF_MEM_TLB_OS         0x40 /* OS fault handler */
1325 #define PERF_MEM_TLB_SHIFT      26
1326
1327 /* Access blocked */
1328 #define PERF_MEM_BLK_NA         0x01 /* not available */
1329 #define PERF_MEM_BLK_DATA       0x02 /* data could not be forwarded */
1330 #define PERF_MEM_BLK_ADDR       0x04 /* address conflict */
1331 #define PERF_MEM_BLK_SHIFT      40
1332
1333 /* hop level */
1334 #define PERF_MEM_HOPS_0         0x01 /* remote core, same node */
1335 /* 2-7 available */
1336 #define PERF_MEM_HOPS_SHIFT     43
1337
1338 #define PERF_MEM_S(a, s) \
1339         (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
1340
1341 /*
1342  * single taken branch record layout:
1343  *
1344  *      from: source instruction (may not always be a branch insn)
1345  *        to: branch target
1346  *   mispred: branch target was mispredicted
1347  * predicted: branch target was predicted
1348  *
1349  * support for mispred, predicted is optional. In case it
1350  * is not supported mispred = predicted = 0.
1351  *
1352  *     in_tx: running in a hardware transaction
1353  *     abort: aborting a hardware transaction
1354  *    cycles: cycles from last branch (or 0 if not supported)
1355  *      type: branch type
1356  */
1357 struct perf_branch_entry {
1358         __u64   from;
1359         __u64   to;
1360         __u64   mispred:1,  /* target mispredicted */
1361                 predicted:1,/* target predicted */
1362                 in_tx:1,    /* in transaction */
1363                 abort:1,    /* transaction abort */
1364                 cycles:16,  /* cycle count to last branch */
1365                 type:4,     /* branch type */
1366                 reserved:40;
1367 };
1368
1369 union perf_sample_weight {
1370         __u64           full;
1371 #if defined(__LITTLE_ENDIAN_BITFIELD)
1372         struct {
1373                 __u32   var1_dw;
1374                 __u16   var2_w;
1375                 __u16   var3_w;
1376         };
1377 #elif defined(__BIG_ENDIAN_BITFIELD)
1378         struct {
1379                 __u16   var3_w;
1380                 __u16   var2_w;
1381                 __u32   var1_dw;
1382         };
1383 #else
1384 #error "Unknown endianness"
1385 #endif
1386 };
1387
1388 #endif /* _UAPI_LINUX_PERF_EVENT_H */