tools/perf/util/event.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef __PERF_RECORD_H
   3 #define __PERF_RECORD_H
   4 /*
   5  * The linux/stddef.h isn't need here, but is needed for __always_inline used
   6  * in files included from uapi/linux/perf_event.h such as
   7  * /usr/include/linux/swab.h and /usr/include/linux/byteorder/little_endian.h,
   8  * detected in at least musl libc, used in Alpine Linux. -acme
   9  */
  10 #include <stdio.h>
  11 #include <linux/stddef.h>
  12 #include <perf/event.h>
  13 #include <linux/types.h>
  14
  15 #include "perf_regs.h"
  16
  17 struct dso;
  18 struct machine;
  19 struct perf_event_attr;
  20
  21 #ifdef __LP64__
  22 /*
  23  * /usr/include/inttypes.h uses just 'lu' for PRIu64, but we end up defining
  24  * __u64 as long long unsigned int, and then -Werror=format= kicks in and
  25  * complains of the mismatched types, so use these two special extra PRI
  26  * macros to overcome that.
  27  */
  28 #define PRI_lu64 "l" PRIu64
  29 #define PRI_lx64 "l" PRIx64
  30 #define PRI_ld64 "l" PRId64
  31 #else
  32 #define PRI_lu64 PRIu64
  33 #define PRI_lx64 PRIx64
  34 #define PRI_ld64 PRId64
  35 #endif
  36
  37 #define PERF_SAMPLE_MASK                                \
  38         (PERF_SAMPLE_IP | PERF_SAMPLE_TID |             \
  39          PERF_SAMPLE_TIME | PERF_SAMPLE_ADDR |          \
  40         PERF_SAMPLE_ID | PERF_SAMPLE_STREAM_ID |        \
  41          PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD |         \
  42          PERF_SAMPLE_IDENTIFIER)
  43
  44 /* perf sample has 16 bits size limit */
  45 #define PERF_SAMPLE_MAX_SIZE (1 << 16)
  46
  47 struct regs_dump {
  48         u64 abi;
  49         u64 mask;
  50         u64 *regs;
  51
  52         /* Cached values/mask filled by first register access. */
  53         u64 cache_regs[PERF_REGS_MAX];
  54         u64 cache_mask;
  55 };
  56
  57 struct stack_dump {
  58         u16 offset;
  59         u64 size;
  60         char *data;
  61 };
  62
  63 struct sample_read_value {
  64         u64 value;
  65         u64 id;
  66 };
  67
  68 struct sample_read {
  69         u64 time_enabled;
  70         u64 time_running;
  71         union {
  72                 struct {
  73                         u64 nr;
  74                         struct sample_read_value *values;
  75                 } group;
  76                 struct sample_read_value one;
  77         };
  78 };
  79
  80 struct ip_callchain {
  81         u64 nr;
  82         u64 ips[];
  83 };
  84
  85 struct branch_stack;
  86
  87 enum {
  88         PERF_IP_FLAG_BRANCH             = 1ULL << 0,
  89         PERF_IP_FLAG_CALL               = 1ULL << 1,
  90         PERF_IP_FLAG_RETURN             = 1ULL << 2,
  91         PERF_IP_FLAG_CONDITIONAL        = 1ULL << 3,
  92         PERF_IP_FLAG_SYSCALLRET         = 1ULL << 4,
  93         PERF_IP_FLAG_ASYNC              = 1ULL << 5,
  94         PERF_IP_FLAG_INTERRUPT          = 1ULL << 6,
  95         PERF_IP_FLAG_TX_ABORT           = 1ULL << 7,
  96         PERF_IP_FLAG_TRACE_BEGIN        = 1ULL << 8,
  97         PERF_IP_FLAG_TRACE_END          = 1ULL << 9,
  98         PERF_IP_FLAG_IN_TX              = 1ULL << 10,
  99         PERF_IP_FLAG_VMENTRY            = 1ULL << 11,
 100         PERF_IP_FLAG_VMEXIT             = 1ULL << 12,
 101 };
 102
 103 #define PERF_IP_FLAG_CHARS "bcrosyiABEx"
 104
 105 #define PERF_BRANCH_MASK                (\
 106         PERF_IP_FLAG_BRANCH             |\
 107         PERF_IP_FLAG_CALL               |\
 108         PERF_IP_FLAG_RETURN             |\
 109         PERF_IP_FLAG_CONDITIONAL        |\
 110         PERF_IP_FLAG_SYSCALLRET         |\
 111         PERF_IP_FLAG_ASYNC              |\
 112         PERF_IP_FLAG_INTERRUPT          |\
 113         PERF_IP_FLAG_TX_ABORT           |\
 114         PERF_IP_FLAG_TRACE_BEGIN        |\
 115         PERF_IP_FLAG_TRACE_END          |\
 116         PERF_IP_FLAG_VMENTRY            |\
 117         PERF_IP_FLAG_VMEXIT)
 118
 119 #define MAX_INSN 16
 120
 121 struct aux_sample {
 122         u64 size;
 123         void *data;
 124 };
 125
 126 struct perf_sample {
 127         u64 ip;
 128         u32 pid, tid;
 129         u64 time;
 130         u64 addr;
 131         u64 id;
 132         u64 stream_id;
 133         u64 period;
 134         u64 weight;
 135         u64 transaction;
 136         u64 insn_cnt;
 137         u64 cyc_cnt;
 138         u32 cpu;
 139         u32 raw_size;
 140         u64 data_src;
 141         u64 phys_addr;
 142         u64 data_page_size;
 143         u64 code_page_size;
 144         u64 cgroup;
 145         u32 flags;
 146         u16 insn_len;
 147         u8  cpumode;
 148         u16 misc;
 149         u16 ins_lat;
 150         u16 p_stage_cyc;
 151         bool no_hw_idx;         /* No hw_idx collected in branch_stack */
 152         char insn[MAX_INSN];
 153         void *raw_data;
 154         struct ip_callchain *callchain;
 155         struct branch_stack *branch_stack;
 156         struct regs_dump  user_regs;
 157         struct regs_dump  intr_regs;
 158         struct stack_dump user_stack;
 159         struct sample_read read;
 160         struct aux_sample aux_sample;
 161 };
 162
 163 #define PERF_MEM_DATA_SRC_NONE \
 164         (PERF_MEM_S(OP, NA) |\
 165          PERF_MEM_S(LVL, NA) |\
 166          PERF_MEM_S(SNOOP, NA) |\
 167          PERF_MEM_S(LOCK, NA) |\
 168          PERF_MEM_S(TLB, NA))
 169
 170 /* Attribute type for custom synthesized events */
 171 #define PERF_TYPE_SYNTH         (INT_MAX + 1U)
 172
 173 /* Attribute config for custom synthesized events */
 174 enum perf_synth_id {
 175         PERF_SYNTH_INTEL_PTWRITE,
 176         PERF_SYNTH_INTEL_MWAIT,
 177         PERF_SYNTH_INTEL_PWRE,
 178         PERF_SYNTH_INTEL_EXSTOP,
 179         PERF_SYNTH_INTEL_PWRX,
 180         PERF_SYNTH_INTEL_CBR,
 181         PERF_SYNTH_INTEL_PSB,
 182 };
 183
 184 /*
 185  * Raw data formats for synthesized events. Note that 4 bytes of padding are
 186  * present to match the 'size' member of PERF_SAMPLE_RAW data which is always
 187  * 8-byte aligned. That means we must dereference raw_data with an offset of 4.
 188  * Refer perf_sample__synth_ptr() and perf_synth__raw_data().  It also means the
 189  * structure sizes are 4 bytes bigger than the raw_size, refer
 190  * perf_synth__raw_size().
 191  */
 192
 193 struct perf_synth_intel_ptwrite {
 194         u32 padding;
 195         union {
 196                 struct {
 197                         u32     ip              :  1,
 198                                 reserved        : 31;
 199                 };
 200                 u32     flags;
 201         };
 202         u64     payload;
 203 };
 204
 205 struct perf_synth_intel_mwait {
 206         u32 padding;
 207         u32 reserved;
 208         union {
 209                 struct {
 210                         u64     hints           :  8,
 211                                 reserved1       : 24,
 212                                 extensions      :  2,
 213                                 reserved2       : 30;
 214                 };
 215                 u64     payload;
 216         };
 217 };
 218
 219 struct perf_synth_intel_pwre {
 220         u32 padding;
 221         u32 reserved;
 222         union {
 223                 struct {
 224                         u64     reserved1       :  7,
 225                                 hw              :  1,
 226                                 subcstate       :  4,
 227                                 cstate          :  4,
 228                                 reserved2       : 48;
 229                 };
 230                 u64     payload;
 231         };
 232 };
 233
 234 struct perf_synth_intel_exstop {
 235         u32 padding;
 236         union {
 237                 struct {
 238                         u32     ip              :  1,
 239                                 reserved        : 31;
 240                 };
 241                 u32     flags;
 242         };
 243 };
 244
 245 struct perf_synth_intel_pwrx {
 246         u32 padding;
 247         u32 reserved;
 248         union {
 249                 struct {
 250                         u64     deepest_cstate  :  4,
 251                                 last_cstate     :  4,
 252                                 wake_reason     :  4,
 253                                 reserved1       : 52;
 254                 };
 255                 u64     payload;
 256         };
 257 };
 258
 259 struct perf_synth_intel_cbr {
 260         u32 padding;
 261         union {
 262                 struct {
 263                         u32     cbr             :  8,
 264                                 reserved1       :  8,
 265                                 max_nonturbo    :  8,
 266                                 reserved2       :  8;
 267                 };
 268                 u32     flags;
 269         };
 270         u32 freq;
 271         u32 reserved3;
 272 };
 273
 274 struct perf_synth_intel_psb {
 275         u32 padding;
 276         u32 reserved;
 277         u64 offset;
 278 };
 279
 280 /*
 281  * raw_data is always 4 bytes from an 8-byte boundary, so subtract 4 to get
 282  * 8-byte alignment.
 283  */
 284 static inline void *perf_sample__synth_ptr(struct perf_sample *sample)
 285 {
 286         return sample->raw_data - 4;
 287 }
 288
 289 static inline void *perf_synth__raw_data(void *p)
 290 {
 291         return p + 4;
 292 }
 293
 294 #define perf_synth__raw_size(d) (sizeof(d) - 4)
 295
 296 #define perf_sample__bad_synth_size(s, d) ((s)->raw_size < sizeof(d) - 4)
 297
 298 enum {
 299         PERF_STAT_ROUND_TYPE__INTERVAL  = 0,
 300         PERF_STAT_ROUND_TYPE__FINAL     = 1,
 301 };
 302
 303 void perf_event__print_totals(void);
 304
 305 struct perf_cpu_map;
 306 struct perf_record_stat_config;
 307 struct perf_stat_config;
 308 struct perf_tool;
 309
 310 void perf_event__read_stat_config(struct perf_stat_config *config,
 311                                   struct perf_record_stat_config *event);
 312
 313 int perf_event__process_comm(struct perf_tool *tool,
 314                              union perf_event *event,
 315                              struct perf_sample *sample,
 316                              struct machine *machine);
 317 int perf_event__process_lost(struct perf_tool *tool,
 318                              union perf_event *event,
 319                              struct perf_sample *sample,
 320                              struct machine *machine);
 321 int perf_event__process_lost_samples(struct perf_tool *tool,
 322                                      union perf_event *event,
 323                                      struct perf_sample *sample,
 324                                      struct machine *machine);
 325 int perf_event__process_aux(struct perf_tool *tool,
 326                             union perf_event *event,
 327                             struct perf_sample *sample,
 328                             struct machine *machine);
 329 int perf_event__process_itrace_start(struct perf_tool *tool,
 330                                      union perf_event *event,
 331                                      struct perf_sample *sample,
 332                                      struct machine *machine);
 333 int perf_event__process_switch(struct perf_tool *tool,
 334                                union perf_event *event,
 335                                struct perf_sample *sample,
 336                                struct machine *machine);
 337 int perf_event__process_namespaces(struct perf_tool *tool,
 338                                    union perf_event *event,
 339                                    struct perf_sample *sample,
 340                                    struct machine *machine);
 341 int perf_event__process_cgroup(struct perf_tool *tool,
 342                                union perf_event *event,
 343                                struct perf_sample *sample,
 344                                struct machine *machine);
 345 int perf_event__process_mmap(struct perf_tool *tool,
 346                              union perf_event *event,
 347                              struct perf_sample *sample,
 348                              struct machine *machine);
 349 int perf_event__process_mmap2(struct perf_tool *tool,
 350                              union perf_event *event,
 351                              struct perf_sample *sample,
 352                              struct machine *machine);
 353 int perf_event__process_fork(struct perf_tool *tool,
 354                              union perf_event *event,
 355                              struct perf_sample *sample,
 356                              struct machine *machine);
 357 int perf_event__process_exit(struct perf_tool *tool,
 358                              union perf_event *event,
 359                              struct perf_sample *sample,
 360                              struct machine *machine);
 361 int perf_event__process_ksymbol(struct perf_tool *tool,
 362                                 union perf_event *event,
 363                                 struct perf_sample *sample,
 364                                 struct machine *machine);
 365 int perf_event__process_bpf(struct perf_tool *tool,
 366                             union perf_event *event,
 367                             struct perf_sample *sample,
 368                             struct machine *machine);
 369 int perf_event__process_text_poke(struct perf_tool *tool,
 370                                   union perf_event *event,
 371                                   struct perf_sample *sample,
 372                                   struct machine *machine);
 373 int perf_event__process(struct perf_tool *tool,
 374                         union perf_event *event,
 375                         struct perf_sample *sample,
 376                         struct machine *machine);
 377
 378 struct addr_location;
 379
 380 int machine__resolve(struct machine *machine, struct addr_location *al,
 381                      struct perf_sample *sample);
 382
 383 void addr_location__put(struct addr_location *al);
 384
 385 struct thread;
 386
 387 bool is_bts_event(struct perf_event_attr *attr);
 388 bool sample_addr_correlates_sym(struct perf_event_attr *attr);
 389 void thread__resolve(struct thread *thread, struct addr_location *al,
 390                      struct perf_sample *sample);
 391
 392 const char *perf_event__name(unsigned int id);
 393
 394 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp);
 395 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp);
 396 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp);
 397 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp);
 398 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp);
 399 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp);
 400 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp);
 401 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp);
 402 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp);
 403 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp);
 404 size_t perf_event__fprintf_cgroup(union perf_event *event, FILE *fp);
 405 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp);
 406 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp);
 407 size_t perf_event__fprintf_text_poke(union perf_event *event, struct machine *machine,FILE *fp);
 408 size_t perf_event__fprintf(union perf_event *event, struct machine *machine, FILE *fp);
 409
 410 int kallsyms__get_function_start(const char *kallsyms_filename,
 411                                  const char *symbol_name, u64 *addr);
 412
 413 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max);
 414 void  cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map,
 415                                u16 type, int max);
 416
 417 void event_attr_init(struct perf_event_attr *attr);
 418
 419 int perf_event_paranoid(void);
 420 bool perf_event_paranoid_check(int max_level);
 421
 422 extern int sysctl_perf_event_max_stack;
 423 extern int sysctl_perf_event_max_contexts_per_stack;
 424 extern unsigned int proc_map_timeout;
 425
 426 #define PAGE_SIZE_NAME_LEN      32
 427 char *get_page_size_name(u64 size, char *str);
 428
 429 void arch_perf_parse_sample_weight(struct perf_sample *data, const __u64 *array, u64 type);
 430 void arch_perf_synthesize_sample_weight(const struct perf_sample *data, __u64 *array, u64 type);
 431 const char *arch_perf_header_entry(const char *se_header);
 432 int arch_support_sort_key(const char *sort_key);
 433
 434 #endif /* __PERF_RECORD_H */