tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->evlist      = NULL;
 236         evsel->bpf_fd      = -1;
 237         INIT_LIST_HEAD(&evsel->node);
 238         INIT_LIST_HEAD(&evsel->config_terms);
 239         perf_evsel__object.init(evsel);
 240         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 241         perf_evsel__calc_id_pos(evsel);
 242         evsel->cmdline_group_boundary = false;
 243         evsel->metric_expr   = NULL;
 244         evsel->metric_name   = NULL;
 245         evsel->metric_events = NULL;
 246         evsel->collect_stat  = false;
 247         evsel->pmu_name      = NULL;
 248 }
 249
 250 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 251 {
 252         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 253
 254         if (!evsel)
 255                 return NULL;
 256         perf_evsel__init(evsel, attr, idx);
 257
 258         if (perf_evsel__is_bpf_output(evsel)) {
 259                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 260                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 261                 evsel->attr.sample_period = 1;
 262         }
 263
 264         if (perf_evsel__is_clock(evsel)) {
 265                 /*
 266                  * The evsel->unit points to static alias->unit
 267                  * so it's ok to use static string in here.
 268                  */
 269                 static const char *unit = "msec";
 270
 271                 evsel->unit = unit;
 272                 evsel->scale = 1e-6;
 273         }
 274
 275         return evsel;
 276 }
 277
 278 static bool perf_event_can_profile_kernel(void)
 279 {
 280         return geteuid() == 0 || perf_event_paranoid() == -1;
 281 }
 282
 283 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 284 {
 285         struct perf_event_attr attr = {
 286                 .type   = PERF_TYPE_HARDWARE,
 287                 .config = PERF_COUNT_HW_CPU_CYCLES,
 288                 .exclude_kernel = !perf_event_can_profile_kernel(),
 289         };
 290         struct perf_evsel *evsel;
 291
 292         event_attr_init(&attr);
 293
 294         if (!precise)
 295                 goto new_event;
 296         /*
 297          * Unnamed union member, not supported as struct member named
 298          * initializer in older compilers such as gcc 4.4.7
 299          *
 300          * Just for probing the precise_ip:
 301          */
 302         attr.sample_period = 1;
 303
 304         perf_event_attr__set_max_precise_ip(&attr);
 305         /*
 306          * Now let the usual logic to set up the perf_event_attr defaults
 307          * to kick in when we return and before perf_evsel__open() is called.
 308          */
 309         attr.sample_period = 0;
 310 new_event:
 311         evsel = perf_evsel__new(&attr);
 312         if (evsel == NULL)
 313                 goto out;
 314
 315         /* use asprintf() because free(evsel) assumes name is allocated */
 316         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 317                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 318                      attr.exclude_kernel ? "u" : "",
 319                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 320                 goto error_free;
 321 out:
 322         return evsel;
 323 error_free:
 324         perf_evsel__delete(evsel);
 325         evsel = NULL;
 326         goto out;
 327 }
 328
 329 /*
 330  * Returns pointer with encoded error via <linux/err.h> interface.
 331  */
 332 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 333 {
 334         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 335         int err = -ENOMEM;
 336
 337         if (evsel == NULL) {
 338                 goto out_err;
 339         } else {
 340                 struct perf_event_attr attr = {
 341                         .type          = PERF_TYPE_TRACEPOINT,
 342                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 343                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 344                 };
 345
 346                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 347                         goto out_free;
 348
 349                 evsel->tp_format = trace_event__tp_format(sys, name);
 350                 if (IS_ERR(evsel->tp_format)) {
 351                         err = PTR_ERR(evsel->tp_format);
 352                         goto out_free;
 353                 }
 354
 355                 event_attr_init(&attr);
 356                 attr.config = evsel->tp_format->id;
 357                 attr.sample_period = 1;
 358                 perf_evsel__init(evsel, &attr, idx);
 359         }
 360
 361         return evsel;
 362
 363 out_free:
 364         zfree(&evsel->name);
 365         free(evsel);
 366 out_err:
 367         return ERR_PTR(err);
 368 }
 369
 370 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 371         "cycles",
 372         "instructions",
 373         "cache-references",
 374         "cache-misses",
 375         "branches",
 376         "branch-misses",
 377         "bus-cycles",
 378         "stalled-cycles-frontend",
 379         "stalled-cycles-backend",
 380         "ref-cycles",
 381 };
 382
 383 static const char *__perf_evsel__hw_name(u64 config)
 384 {
 385         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 386                 return perf_evsel__hw_names[config];
 387
 388         return "unknown-hardware";
 389 }
 390
 391 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 392 {
 393         int colon = 0, r = 0;
 394         struct perf_event_attr *attr = &evsel->attr;
 395         bool exclude_guest_default = false;
 396
 397 #define MOD_PRINT(context, mod) do {                                    \
 398                 if (!attr->exclude_##context) {                         \
 399                         if (!colon) colon = ++r;                        \
 400                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 401                 } } while(0)
 402
 403         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 404                 MOD_PRINT(kernel, 'k');
 405                 MOD_PRINT(user, 'u');
 406                 MOD_PRINT(hv, 'h');
 407                 exclude_guest_default = true;
 408         }
 409
 410         if (attr->precise_ip) {
 411                 if (!colon)
 412                         colon = ++r;
 413                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 414                 exclude_guest_default = true;
 415         }
 416
 417         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 418                 MOD_PRINT(host, 'H');
 419                 MOD_PRINT(guest, 'G');
 420         }
 421 #undef MOD_PRINT
 422         if (colon)
 423                 bf[colon - 1] = ':';
 424         return r;
 425 }
 426
 427 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 428 {
 429         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 430         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 431 }
 432
 433 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 434         "cpu-clock",
 435         "task-clock",
 436         "page-faults",
 437         "context-switches",
 438         "cpu-migrations",
 439         "minor-faults",
 440         "major-faults",
 441         "alignment-faults",
 442         "emulation-faults",
 443         "dummy",
 444 };
 445
 446 static const char *__perf_evsel__sw_name(u64 config)
 447 {
 448         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 449                 return perf_evsel__sw_names[config];
 450         return "unknown-software";
 451 }
 452
 453 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 454 {
 455         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 456         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 457 }
 458
 459 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 460 {
 461         int r;
 462
 463         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 464
 465         if (type & HW_BREAKPOINT_R)
 466                 r += scnprintf(bf + r, size - r, "r");
 467
 468         if (type & HW_BREAKPOINT_W)
 469                 r += scnprintf(bf + r, size - r, "w");
 470
 471         if (type & HW_BREAKPOINT_X)
 472                 r += scnprintf(bf + r, size - r, "x");
 473
 474         return r;
 475 }
 476
 477 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 478 {
 479         struct perf_event_attr *attr = &evsel->attr;
 480         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 481         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 482 }
 483
 484 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 485                                 [PERF_EVSEL__MAX_ALIASES] = {
 486  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 487  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 488  { "LLC",       "L2",                                                   },
 489  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 490  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 491  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 492  { "node",                                                              },
 493 };
 494
 495 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 496                                    [PERF_EVSEL__MAX_ALIASES] = {
 497  { "load",      "loads",        "read",                                 },
 498  { "store",     "stores",       "write",                                },
 499  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 500 };
 501
 502 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 503                                        [PERF_EVSEL__MAX_ALIASES] = {
 504  { "refs",      "Reference",    "ops",          "access",               },
 505  { "misses",    "miss",                                                 },
 506 };
 507
 508 #define C(x)            PERF_COUNT_HW_CACHE_##x
 509 #define CACHE_READ      (1 << C(OP_READ))
 510 #define CACHE_WRITE     (1 << C(OP_WRITE))
 511 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 512 #define COP(x)          (1 << x)
 513
 514 /*
 515  * cache operartion stat
 516  * L1I : Read and prefetch only
 517  * ITLB and BPU : Read-only
 518  */
 519 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 520  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 521  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 522  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 523  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524  [C(ITLB)]      = (CACHE_READ),
 525  [C(BPU)]       = (CACHE_READ),
 526  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 527 };
 528
 529 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 530 {
 531         if (perf_evsel__hw_cache_stat[type] & COP(op))
 532                 return true;    /* valid */
 533         else
 534                 return false;   /* invalid */
 535 }
 536
 537 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 538                                             char *bf, size_t size)
 539 {
 540         if (result) {
 541                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 542                                  perf_evsel__hw_cache_op[op][0],
 543                                  perf_evsel__hw_cache_result[result][0]);
 544         }
 545
 546         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 547                          perf_evsel__hw_cache_op[op][1]);
 548 }
 549
 550 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 551 {
 552         u8 op, result, type = (config >>  0) & 0xff;
 553         const char *err = "unknown-ext-hardware-cache-type";
 554
 555         if (type >= PERF_COUNT_HW_CACHE_MAX)
 556                 goto out_err;
 557
 558         op = (config >>  8) & 0xff;
 559         err = "unknown-ext-hardware-cache-op";
 560         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 561                 goto out_err;
 562
 563         result = (config >> 16) & 0xff;
 564         err = "unknown-ext-hardware-cache-result";
 565         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 566                 goto out_err;
 567
 568         err = "invalid-cache";
 569         if (!perf_evsel__is_cache_op_valid(type, op))
 570                 goto out_err;
 571
 572         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 573 out_err:
 574         return scnprintf(bf, size, "%s", err);
 575 }
 576
 577 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 578 {
 579         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 580         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 581 }
 582
 583 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 584 {
 585         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 586         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 587 }
 588
 589 const char *perf_evsel__name(struct perf_evsel *evsel)
 590 {
 591         char bf[128];
 592
 593         if (!evsel)
 594                 goto out_unknown;
 595
 596         if (evsel->name)
 597                 return evsel->name;
 598
 599         switch (evsel->attr.type) {
 600         case PERF_TYPE_RAW:
 601                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 602                 break;
 603
 604         case PERF_TYPE_HARDWARE:
 605                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 606                 break;
 607
 608         case PERF_TYPE_HW_CACHE:
 609                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 610                 break;
 611
 612         case PERF_TYPE_SOFTWARE:
 613                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 614                 break;
 615
 616         case PERF_TYPE_TRACEPOINT:
 617                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 618                 break;
 619
 620         case PERF_TYPE_BREAKPOINT:
 621                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 622                 break;
 623
 624         default:
 625                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 626                           evsel->attr.type);
 627                 break;
 628         }
 629
 630         evsel->name = strdup(bf);
 631
 632         if (evsel->name)
 633                 return evsel->name;
 634 out_unknown:
 635         return "unknown";
 636 }
 637
 638 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 639 {
 640         return evsel->group_name ?: "anon group";
 641 }
 642
 643 /*
 644  * Returns the group details for the specified leader,
 645  * with following rules.
 646  *
 647  *  For record -e '{cycles,instructions}'
 648  *    'anon group { cycles:u, instructions:u }'
 649  *
 650  *  For record -e 'cycles,instructions' and report --group
 651  *    'cycles:u, instructions:u'
 652  */
 653 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 654 {
 655         int ret = 0;
 656         struct perf_evsel *pos;
 657         const char *group_name = perf_evsel__group_name(evsel);
 658
 659         if (!evsel->forced_leader)
 660                 ret = scnprintf(buf, size, "%s { ", group_name);
 661
 662         ret += scnprintf(buf + ret, size - ret, "%s",
 663                          perf_evsel__name(evsel));
 664
 665         for_each_group_member(pos, evsel)
 666                 ret += scnprintf(buf + ret, size - ret, ", %s",
 667                                  perf_evsel__name(pos));
 668
 669         if (!evsel->forced_leader)
 670                 ret += scnprintf(buf + ret, size - ret, " }");
 671
 672         return ret;
 673 }
 674
 675 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 676                                            struct record_opts *opts,
 677                                            struct callchain_param *param)
 678 {
 679         bool function = perf_evsel__is_function_event(evsel);
 680         struct perf_event_attr *attr = &evsel->attr;
 681
 682         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 683
 684         attr->sample_max_stack = param->max_stack;
 685
 686         if (param->record_mode == CALLCHAIN_LBR) {
 687                 if (!opts->branch_stack) {
 688                         if (attr->exclude_user) {
 689                                 pr_warning("LBR callstack option is only available "
 690                                            "to get user callchain information. "
 691                                            "Falling back to framepointers.\n");
 692                         } else {
 693                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 694                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 695                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 696                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 697                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 698                         }
 699                 } else
 700                          pr_warning("Cannot use LBR callstack with branch stack. "
 701                                     "Falling back to framepointers.\n");
 702         }
 703
 704         if (param->record_mode == CALLCHAIN_DWARF) {
 705                 if (!function) {
 706                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 707                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 708                         attr->sample_regs_user |= PERF_REGS_MASK;
 709                         attr->sample_stack_user = param->dump_size;
 710                         attr->exclude_callchain_user = 1;
 711                 } else {
 712                         pr_info("Cannot use DWARF unwind for function trace event,"
 713                                 " falling back to framepointers.\n");
 714                 }
 715         }
 716
 717         if (function) {
 718                 pr_info("Disabling user space callchains for function trace event.\n");
 719                 attr->exclude_callchain_user = 1;
 720         }
 721 }
 722
 723 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 724                                   struct record_opts *opts,
 725                                   struct callchain_param *param)
 726 {
 727         if (param->enabled)
 728                 return __perf_evsel__config_callchain(evsel, opts, param);
 729 }
 730
 731 static void
 732 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 733                             struct callchain_param *param)
 734 {
 735         struct perf_event_attr *attr = &evsel->attr;
 736
 737         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 738         if (param->record_mode == CALLCHAIN_LBR) {
 739                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 740                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 741                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 742         }
 743         if (param->record_mode == CALLCHAIN_DWARF) {
 744                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 745                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 746         }
 747 }
 748
 749 static void apply_config_terms(struct perf_evsel *evsel,
 750                                struct record_opts *opts, bool track)
 751 {
 752         struct perf_evsel_config_term *term;
 753         struct list_head *config_terms = &evsel->config_terms;
 754         struct perf_event_attr *attr = &evsel->attr;
 755         /* callgraph default */
 756         struct callchain_param param = {
 757                 .record_mode = callchain_param.record_mode,
 758         };
 759         u32 dump_size = 0;
 760         int max_stack = 0;
 761         const char *callgraph_buf = NULL;
 762
 763         list_for_each_entry(term, config_terms, list) {
 764                 switch (term->type) {
 765                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 766                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 767                                 attr->sample_period = term->val.period;
 768                                 attr->freq = 0;
 769                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 770                         }
 771                         break;
 772                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 773                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 774                                 attr->sample_freq = term->val.freq;
 775                                 attr->freq = 1;
 776                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 777                         }
 778                         break;
 779                 case PERF_EVSEL__CONFIG_TERM_TIME:
 780                         if (term->val.time)
 781                                 perf_evsel__set_sample_bit(evsel, TIME);
 782                         else
 783                                 perf_evsel__reset_sample_bit(evsel, TIME);
 784                         break;
 785                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 786                         callgraph_buf = term->val.callgraph;
 787                         break;
 788                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 789                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 790                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 791                                 parse_branch_str(term->val.branch,
 792                                                  &attr->branch_sample_type);
 793                         } else
 794                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 795                         break;
 796                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 797                         dump_size = term->val.stack_user;
 798                         break;
 799                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 800                         max_stack = term->val.max_stack;
 801                         break;
 802                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 803                         /*
 804                          * attr->inherit should has already been set by
 805                          * perf_evsel__config. If user explicitly set
 806                          * inherit using config terms, override global
 807                          * opt->no_inherit setting.
 808                          */
 809                         attr->inherit = term->val.inherit ? 1 : 0;
 810                         break;
 811                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 812                         attr->write_backward = term->val.overwrite ? 1 : 0;
 813                         break;
 814                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 815                         break;
 816                 default:
 817                         break;
 818                 }
 819         }
 820
 821         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 822         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 823                 bool sample_address = false;
 824
 825                 if (max_stack) {
 826                         param.max_stack = max_stack;
 827                         if (callgraph_buf == NULL)
 828                                 callgraph_buf = "fp";
 829                 }
 830
 831                 /* parse callgraph parameters */
 832                 if (callgraph_buf != NULL) {
 833                         if (!strcmp(callgraph_buf, "no")) {
 834                                 param.enabled = false;
 835                                 param.record_mode = CALLCHAIN_NONE;
 836                         } else {
 837                                 param.enabled = true;
 838                                 if (parse_callchain_record(callgraph_buf, &param)) {
 839                                         pr_err("per-event callgraph setting for %s failed. "
 840                                                "Apply callgraph global setting for it\n",
 841                                                evsel->name);
 842                                         return;
 843                                 }
 844                                 if (param.record_mode == CALLCHAIN_DWARF)
 845                                         sample_address = true;
 846                         }
 847                 }
 848                 if (dump_size > 0) {
 849                         dump_size = round_up(dump_size, sizeof(u64));
 850                         param.dump_size = dump_size;
 851                 }
 852
 853                 /* If global callgraph set, clear it */
 854                 if (callchain_param.enabled)
 855                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 856
 857                 /* set perf-event callgraph */
 858                 if (param.enabled) {
 859                         if (sample_address) {
 860                                 perf_evsel__set_sample_bit(evsel, ADDR);
 861                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 862                                 evsel->attr.mmap_data = track;
 863                         }
 864                         perf_evsel__config_callchain(evsel, opts, &param);
 865                 }
 866         }
 867 }
 868
 869 static bool is_dummy_event(struct perf_evsel *evsel)
 870 {
 871         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 872                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 873 }
 874
 875 /*
 876  * The enable_on_exec/disabled value strategy:
 877  *
 878  *  1) For any type of traced program:
 879  *    - all independent events and group leaders are disabled
 880  *    - all group members are enabled
 881  *
 882  *     Group members are ruled by group leaders. They need to
 883  *     be enabled, because the group scheduling relies on that.
 884  *
 885  *  2) For traced programs executed by perf:
 886  *     - all independent events and group leaders have
 887  *       enable_on_exec set
 888  *     - we don't specifically enable or disable any event during
 889  *       the record command
 890  *
 891  *     Independent events and group leaders are initially disabled
 892  *     and get enabled by exec. Group members are ruled by group
 893  *     leaders as stated in 1).
 894  *
 895  *  3) For traced programs attached by perf (pid/tid):
 896  *     - we specifically enable or disable all events during
 897  *       the record command
 898  *
 899  *     When attaching events to already running traced we
 900  *     enable/disable events specifically, as there's no
 901  *     initial traced exec call.
 902  */
 903 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 904                         struct callchain_param *callchain)
 905 {
 906         struct perf_evsel *leader = evsel->leader;
 907         struct perf_event_attr *attr = &evsel->attr;
 908         int track = evsel->tracking;
 909         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 910
 911         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 912         attr->inherit       = !opts->no_inherit;
 913         attr->write_backward = opts->overwrite ? 1 : 0;
 914
 915         perf_evsel__set_sample_bit(evsel, IP);
 916         perf_evsel__set_sample_bit(evsel, TID);
 917
 918         if (evsel->sample_read) {
 919                 perf_evsel__set_sample_bit(evsel, READ);
 920
 921                 /*
 922                  * We need ID even in case of single event, because
 923                  * PERF_SAMPLE_READ process ID specific data.
 924                  */
 925                 perf_evsel__set_sample_id(evsel, false);
 926
 927                 /*
 928                  * Apply group format only if we belong to group
 929                  * with more than one members.
 930                  */
 931                 if (leader->nr_members > 1) {
 932                         attr->read_format |= PERF_FORMAT_GROUP;
 933                         attr->inherit = 0;
 934                 }
 935         }
 936
 937         /*
 938          * We default some events to have a default interval. But keep
 939          * it a weak assumption overridable by the user.
 940          */
 941         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 942                                      opts->user_interval != ULLONG_MAX)) {
 943                 if (opts->freq) {
 944                         perf_evsel__set_sample_bit(evsel, PERIOD);
 945                         attr->freq              = 1;
 946                         attr->sample_freq       = opts->freq;
 947                 } else {
 948                         attr->sample_period = opts->default_interval;
 949                 }
 950         }
 951
 952         /*
 953          * Disable sampling for all group members other
 954          * than leader in case leader 'leads' the sampling.
 955          */
 956         if ((leader != evsel) && leader->sample_read) {
 957                 attr->freq           = 0;
 958                 attr->sample_freq    = 0;
 959                 attr->sample_period  = 0;
 960                 attr->write_backward = 0;
 961         }
 962
 963         if (opts->no_samples)
 964                 attr->sample_freq = 0;
 965
 966         if (opts->inherit_stat) {
 967                 evsel->attr.read_format |=
 968                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 969                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 970                         PERF_FORMAT_ID;
 971                 attr->inherit_stat = 1;
 972         }
 973
 974         if (opts->sample_address) {
 975                 perf_evsel__set_sample_bit(evsel, ADDR);
 976                 attr->mmap_data = track;
 977         }
 978
 979         /*
 980          * We don't allow user space callchains for  function trace
 981          * event, due to issues with page faults while tracing page
 982          * fault handler and its overall trickiness nature.
 983          */
 984         if (perf_evsel__is_function_event(evsel))
 985                 evsel->attr.exclude_callchain_user = 1;
 986
 987         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 988                 perf_evsel__config_callchain(evsel, opts, callchain);
 989
 990         if (opts->sample_intr_regs) {
 991                 attr->sample_regs_intr = opts->sample_intr_regs;
 992                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 993         }
 994
 995         if (opts->sample_user_regs) {
 996                 attr->sample_regs_user |= opts->sample_user_regs;
 997                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 998         }
 999
1000         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1001                 perf_evsel__set_sample_bit(evsel, CPU);
1002
1003         /*
1004          * When the user explicitly disabled time don't force it here.
1005          */
1006         if (opts->sample_time &&
1007             (!perf_missing_features.sample_id_all &&
1008             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1009              opts->sample_time_set)))
1010                 perf_evsel__set_sample_bit(evsel, TIME);
1011
1012         if (opts->raw_samples && !evsel->no_aux_samples) {
1013                 perf_evsel__set_sample_bit(evsel, TIME);
1014                 perf_evsel__set_sample_bit(evsel, RAW);
1015                 perf_evsel__set_sample_bit(evsel, CPU);
1016         }
1017
1018         if (opts->sample_address)
1019                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1020
1021         if (opts->sample_phys_addr)
1022                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1023
1024         if (opts->no_buffering) {
1025                 attr->watermark = 0;
1026                 attr->wakeup_events = 1;
1027         }
1028         if (opts->branch_stack && !evsel->no_aux_samples) {
1029                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1030                 attr->branch_sample_type = opts->branch_stack;
1031         }
1032
1033         if (opts->sample_weight)
1034                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1035
1036         attr->task  = track;
1037         attr->mmap  = track;
1038         attr->mmap2 = track && !perf_missing_features.mmap2;
1039         attr->comm  = track;
1040
1041         if (opts->record_namespaces)
1042                 attr->namespaces  = track;
1043
1044         if (opts->record_switch_events)
1045                 attr->context_switch = track;
1046
1047         if (opts->sample_transaction)
1048                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1049
1050         if (opts->running_time) {
1051                 evsel->attr.read_format |=
1052                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1053                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1054         }
1055
1056         /*
1057          * XXX see the function comment above
1058          *
1059          * Disabling only independent events or group leaders,
1060          * keeping group members enabled.
1061          */
1062         if (perf_evsel__is_group_leader(evsel))
1063                 attr->disabled = 1;
1064
1065         /*
1066          * Setting enable_on_exec for independent events and
1067          * group leaders for traced executed by perf.
1068          */
1069         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1070                 !opts->initial_delay)
1071                 attr->enable_on_exec = 1;
1072
1073         if (evsel->immediate) {
1074                 attr->disabled = 0;
1075                 attr->enable_on_exec = 0;
1076         }
1077
1078         clockid = opts->clockid;
1079         if (opts->use_clockid) {
1080                 attr->use_clockid = 1;
1081                 attr->clockid = opts->clockid;
1082         }
1083
1084         if (evsel->precise_max)
1085                 perf_event_attr__set_max_precise_ip(attr);
1086
1087         if (opts->all_user) {
1088                 attr->exclude_kernel = 1;
1089                 attr->exclude_user   = 0;
1090         }
1091
1092         if (opts->all_kernel) {
1093                 attr->exclude_kernel = 0;
1094                 attr->exclude_user   = 1;
1095         }
1096
1097         if (evsel->own_cpus || evsel->unit)
1098                 evsel->attr.read_format |= PERF_FORMAT_ID;
1099
1100         /*
1101          * Apply event specific term settings,
1102          * it overloads any global configuration.
1103          */
1104         apply_config_terms(evsel, opts, track);
1105
1106         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1107
1108         /* The --period option takes the precedence. */
1109         if (opts->period_set) {
1110                 if (opts->period)
1111                         perf_evsel__set_sample_bit(evsel, PERIOD);
1112                 else
1113                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1114         }
1115
1116         /*
1117          * For initial_delay, a dummy event is added implicitly.
1118          * The software event will trigger -EOPNOTSUPP error out,
1119          * if BRANCH_STACK bit is set.
1120          */
1121         if (opts->initial_delay && is_dummy_event(evsel))
1122                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1123 }
1124
1125 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1126 {
1127         if (evsel->system_wide)
1128                 nthreads = 1;
1129
1130         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1131
1132         if (evsel->fd) {
1133                 int cpu, thread;
1134                 for (cpu = 0; cpu < ncpus; cpu++) {
1135                         for (thread = 0; thread < nthreads; thread++) {
1136                                 FD(evsel, cpu, thread) = -1;
1137                         }
1138                 }
1139         }
1140
1141         return evsel->fd != NULL ? 0 : -ENOMEM;
1142 }
1143
1144 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1145                           int ioc,  void *arg)
1146 {
1147         int cpu, thread;
1148
1149         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1150                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1151                         int fd = FD(evsel, cpu, thread),
1152                             err = ioctl(fd, ioc, arg);
1153
1154                         if (err)
1155                                 return err;
1156                 }
1157         }
1158
1159         return 0;
1160 }
1161
1162 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1163 {
1164         return perf_evsel__run_ioctl(evsel,
1165                                      PERF_EVENT_IOC_SET_FILTER,
1166                                      (void *)filter);
1167 }
1168
1169 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1170 {
1171         char *new_filter = strdup(filter);
1172
1173         if (new_filter != NULL) {
1174                 free(evsel->filter);
1175                 evsel->filter = new_filter;
1176                 return 0;
1177         }
1178
1179         return -1;
1180 }
1181
1182 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1183                                      const char *fmt, const char *filter)
1184 {
1185         char *new_filter;
1186
1187         if (evsel->filter == NULL)
1188                 return perf_evsel__set_filter(evsel, filter);
1189
1190         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1191                 free(evsel->filter);
1192                 evsel->filter = new_filter;
1193                 return 0;
1194         }
1195
1196         return -1;
1197 }
1198
1199 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1200 {
1201         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1202 }
1203
1204 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1205 {
1206         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1207 }
1208
1209 int perf_evsel__enable(struct perf_evsel *evsel)
1210 {
1211         return perf_evsel__run_ioctl(evsel,
1212                                      PERF_EVENT_IOC_ENABLE,
1213                                      0);
1214 }
1215
1216 int perf_evsel__disable(struct perf_evsel *evsel)
1217 {
1218         return perf_evsel__run_ioctl(evsel,
1219                                      PERF_EVENT_IOC_DISABLE,
1220                                      0);
1221 }
1222
1223 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1224 {
1225         if (ncpus == 0 || nthreads == 0)
1226                 return 0;
1227
1228         if (evsel->system_wide)
1229                 nthreads = 1;
1230
1231         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1232         if (evsel->sample_id == NULL)
1233                 return -ENOMEM;
1234
1235         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1236         if (evsel->id == NULL) {
1237                 xyarray__delete(evsel->sample_id);
1238                 evsel->sample_id = NULL;
1239                 return -ENOMEM;
1240         }
1241
1242         return 0;
1243 }
1244
1245 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1246 {
1247         xyarray__delete(evsel->fd);
1248         evsel->fd = NULL;
1249 }
1250
1251 static void perf_evsel__free_id(struct perf_evsel *evsel)
1252 {
1253         xyarray__delete(evsel->sample_id);
1254         evsel->sample_id = NULL;
1255         zfree(&evsel->id);
1256 }
1257
1258 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1259 {
1260         struct perf_evsel_config_term *term, *h;
1261
1262         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1263                 list_del(&term->list);
1264                 free(term);
1265         }
1266 }
1267
1268 void perf_evsel__close_fd(struct perf_evsel *evsel)
1269 {
1270         int cpu, thread;
1271
1272         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1273                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1274                         close(FD(evsel, cpu, thread));
1275                         FD(evsel, cpu, thread) = -1;
1276                 }
1277 }
1278
1279 void perf_evsel__exit(struct perf_evsel *evsel)
1280 {
1281         assert(list_empty(&evsel->node));
1282         assert(evsel->evlist == NULL);
1283         perf_evsel__free_counts(evsel);
1284         perf_evsel__free_fd(evsel);
1285         perf_evsel__free_id(evsel);
1286         perf_evsel__free_config_terms(evsel);
1287         cgroup__put(evsel->cgrp);
1288         cpu_map__put(evsel->cpus);
1289         cpu_map__put(evsel->own_cpus);
1290         thread_map__put(evsel->threads);
1291         zfree(&evsel->group_name);
1292         zfree(&evsel->name);
1293         zfree(&evsel->pmu_name);
1294         zfree(&evsel->per_pkg_mask);
1295         zfree(&evsel->metric_events);
1296         perf_evsel__object.fini(evsel);
1297 }
1298
1299 void perf_evsel__delete(struct perf_evsel *evsel)
1300 {
1301         perf_evsel__exit(evsel);
1302         free(evsel);
1303 }
1304
1305 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1306                                 struct perf_counts_values *count)
1307 {
1308         struct perf_counts_values tmp;
1309
1310         if (!evsel->prev_raw_counts)
1311                 return;
1312
1313         if (cpu == -1) {
1314                 tmp = evsel->prev_raw_counts->aggr;
1315                 evsel->prev_raw_counts->aggr = *count;
1316         } else {
1317                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1318                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1319         }
1320
1321         count->val = count->val - tmp.val;
1322         count->ena = count->ena - tmp.ena;
1323         count->run = count->run - tmp.run;
1324 }
1325
1326 void perf_counts_values__scale(struct perf_counts_values *count,
1327                                bool scale, s8 *pscaled)
1328 {
1329         s8 scaled = 0;
1330
1331         if (scale) {
1332                 if (count->run == 0) {
1333                         scaled = -1;
1334                         count->val = 0;
1335                 } else if (count->run < count->ena) {
1336                         scaled = 1;
1337                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1338                 }
1339         } else
1340                 count->ena = count->run = 0;
1341
1342         if (pscaled)
1343                 *pscaled = scaled;
1344 }
1345
1346 static int perf_evsel__read_size(struct perf_evsel *evsel)
1347 {
1348         u64 read_format = evsel->attr.read_format;
1349         int entry = sizeof(u64); /* value */
1350         int size = 0;
1351         int nr = 1;
1352
1353         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1354                 size += sizeof(u64);
1355
1356         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1357                 size += sizeof(u64);
1358
1359         if (read_format & PERF_FORMAT_ID)
1360                 entry += sizeof(u64);
1361
1362         if (read_format & PERF_FORMAT_GROUP) {
1363                 nr = evsel->nr_members;
1364                 size += sizeof(u64);
1365         }
1366
1367         size += entry * nr;
1368         return size;
1369 }
1370
1371 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1372                      struct perf_counts_values *count)
1373 {
1374         size_t size = perf_evsel__read_size(evsel);
1375
1376         memset(count, 0, sizeof(*count));
1377
1378         if (FD(evsel, cpu, thread) < 0)
1379                 return -EINVAL;
1380
1381         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1382                 return -errno;
1383
1384         return 0;
1385 }
1386
1387 static int
1388 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1389 {
1390         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1391
1392         return perf_evsel__read(evsel, cpu, thread, count);
1393 }
1394
1395 static void
1396 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1397                       u64 val, u64 ena, u64 run)
1398 {
1399         struct perf_counts_values *count;
1400
1401         count = perf_counts(counter->counts, cpu, thread);
1402
1403         count->val    = val;
1404         count->ena    = ena;
1405         count->run    = run;
1406         count->loaded = true;
1407 }
1408
1409 static int
1410 perf_evsel__process_group_data(struct perf_evsel *leader,
1411                                int cpu, int thread, u64 *data)
1412 {
1413         u64 read_format = leader->attr.read_format;
1414         struct sample_read_value *v;
1415         u64 nr, ena = 0, run = 0, i;
1416
1417         nr = *data++;
1418
1419         if (nr != (u64) leader->nr_members)
1420                 return -EINVAL;
1421
1422         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1423                 ena = *data++;
1424
1425         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1426                 run = *data++;
1427
1428         v = (struct sample_read_value *) data;
1429
1430         perf_evsel__set_count(leader, cpu, thread,
1431                               v[0].value, ena, run);
1432
1433         for (i = 1; i < nr; i++) {
1434                 struct perf_evsel *counter;
1435
1436                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1437                 if (!counter)
1438                         return -EINVAL;
1439
1440                 perf_evsel__set_count(counter, cpu, thread,
1441                                       v[i].value, ena, run);
1442         }
1443
1444         return 0;
1445 }
1446
1447 static int
1448 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1449 {
1450         struct perf_stat_evsel *ps = leader->stats;
1451         u64 read_format = leader->attr.read_format;
1452         int size = perf_evsel__read_size(leader);
1453         u64 *data = ps->group_data;
1454
1455         if (!(read_format & PERF_FORMAT_ID))
1456                 return -EINVAL;
1457
1458         if (!perf_evsel__is_group_leader(leader))
1459                 return -EINVAL;
1460
1461         if (!data) {
1462                 data = zalloc(size);
1463                 if (!data)
1464                         return -ENOMEM;
1465
1466                 ps->group_data = data;
1467         }
1468
1469         if (FD(leader, cpu, thread) < 0)
1470                 return -EINVAL;
1471
1472         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1473                 return -errno;
1474
1475         return perf_evsel__process_group_data(leader, cpu, thread, data);
1476 }
1477
1478 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1479 {
1480         u64 read_format = evsel->attr.read_format;
1481
1482         if (read_format & PERF_FORMAT_GROUP)
1483                 return perf_evsel__read_group(evsel, cpu, thread);
1484         else
1485                 return perf_evsel__read_one(evsel, cpu, thread);
1486 }
1487
1488 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1489                               int cpu, int thread, bool scale)
1490 {
1491         struct perf_counts_values count;
1492         size_t nv = scale ? 3 : 1;
1493
1494         if (FD(evsel, cpu, thread) < 0)
1495                 return -EINVAL;
1496
1497         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1498                 return -ENOMEM;
1499
1500         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1501                 return -errno;
1502
1503         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1504         perf_counts_values__scale(&count, scale, NULL);
1505         *perf_counts(evsel->counts, cpu, thread) = count;
1506         return 0;
1507 }
1508
1509 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1510 {
1511         struct perf_evsel *leader = evsel->leader;
1512         int fd;
1513
1514         if (perf_evsel__is_group_leader(evsel))
1515                 return -1;
1516
1517         /*
1518          * Leader must be already processed/open,
1519          * if not it's a bug.
1520          */
1521         BUG_ON(!leader->fd);
1522
1523         fd = FD(leader, cpu, thread);
1524         BUG_ON(fd == -1);
1525
1526         return fd;
1527 }
1528
1529 struct bit_names {
1530         int bit;
1531         const char *name;
1532 };
1533
1534 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1535 {
1536         bool first_bit = true;
1537         int i = 0;
1538
1539         do {
1540                 if (value & bits[i].bit) {
1541                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1542                         first_bit = false;
1543                 }
1544         } while (bits[++i].name != NULL);
1545 }
1546
1547 static void __p_sample_type(char *buf, size_t size, u64 value)
1548 {
1549 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1550         struct bit_names bits[] = {
1551                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1552                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1553                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1554                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1555                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1556                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1557                 { .name = NULL, }
1558         };
1559 #undef bit_name
1560         __p_bits(buf, size, value, bits);
1561 }
1562
1563 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1564 {
1565 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1566         struct bit_names bits[] = {
1567                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1568                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1569                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1570                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1571                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1572                 { .name = NULL, }
1573         };
1574 #undef bit_name
1575         __p_bits(buf, size, value, bits);
1576 }
1577
1578 static void __p_read_format(char *buf, size_t size, u64 value)
1579 {
1580 #define bit_name(n) { PERF_FORMAT_##n, #n }
1581         struct bit_names bits[] = {
1582                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1583                 bit_name(ID), bit_name(GROUP),
1584                 { .name = NULL, }
1585         };
1586 #undef bit_name
1587         __p_bits(buf, size, value, bits);
1588 }
1589
1590 #define BUF_SIZE                1024
1591
1592 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1593 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1594 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1595 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1596 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1597 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1598
1599 #define PRINT_ATTRn(_n, _f, _p)                         \
1600 do {                                                    \
1601         if (attr->_f) {                                 \
1602                 _p(attr->_f);                           \
1603                 ret += attr__fprintf(fp, _n, buf, priv);\
1604         }                                               \
1605 } while (0)
1606
1607 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1608
1609 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1610                              attr__fprintf_f attr__fprintf, void *priv)
1611 {
1612         char buf[BUF_SIZE];
1613         int ret = 0;
1614
1615         PRINT_ATTRf(type, p_unsigned);
1616         PRINT_ATTRf(size, p_unsigned);
1617         PRINT_ATTRf(config, p_hex);
1618         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1619         PRINT_ATTRf(sample_type, p_sample_type);
1620         PRINT_ATTRf(read_format, p_read_format);
1621
1622         PRINT_ATTRf(disabled, p_unsigned);
1623         PRINT_ATTRf(inherit, p_unsigned);
1624         PRINT_ATTRf(pinned, p_unsigned);
1625         PRINT_ATTRf(exclusive, p_unsigned);
1626         PRINT_ATTRf(exclude_user, p_unsigned);
1627         PRINT_ATTRf(exclude_kernel, p_unsigned);
1628         PRINT_ATTRf(exclude_hv, p_unsigned);
1629         PRINT_ATTRf(exclude_idle, p_unsigned);
1630         PRINT_ATTRf(mmap, p_unsigned);
1631         PRINT_ATTRf(comm, p_unsigned);
1632         PRINT_ATTRf(freq, p_unsigned);
1633         PRINT_ATTRf(inherit_stat, p_unsigned);
1634         PRINT_ATTRf(enable_on_exec, p_unsigned);
1635         PRINT_ATTRf(task, p_unsigned);
1636         PRINT_ATTRf(watermark, p_unsigned);
1637         PRINT_ATTRf(precise_ip, p_unsigned);
1638         PRINT_ATTRf(mmap_data, p_unsigned);
1639         PRINT_ATTRf(sample_id_all, p_unsigned);
1640         PRINT_ATTRf(exclude_host, p_unsigned);
1641         PRINT_ATTRf(exclude_guest, p_unsigned);
1642         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1643         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1644         PRINT_ATTRf(mmap2, p_unsigned);
1645         PRINT_ATTRf(comm_exec, p_unsigned);
1646         PRINT_ATTRf(use_clockid, p_unsigned);
1647         PRINT_ATTRf(context_switch, p_unsigned);
1648         PRINT_ATTRf(write_backward, p_unsigned);
1649         PRINT_ATTRf(namespaces, p_unsigned);
1650
1651         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1652         PRINT_ATTRf(bp_type, p_unsigned);
1653         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1654         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1655         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1656         PRINT_ATTRf(sample_regs_user, p_hex);
1657         PRINT_ATTRf(sample_stack_user, p_unsigned);
1658         PRINT_ATTRf(clockid, p_signed);
1659         PRINT_ATTRf(sample_regs_intr, p_hex);
1660         PRINT_ATTRf(aux_watermark, p_unsigned);
1661         PRINT_ATTRf(sample_max_stack, p_unsigned);
1662
1663         return ret;
1664 }
1665
1666 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1667                                 void *priv __maybe_unused)
1668 {
1669         return fprintf(fp, "  %-32s %s\n", name, val);
1670 }
1671
1672 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1673                                   int nr_cpus, int nr_threads,
1674                                   int thread_idx)
1675 {
1676         for (int cpu = 0; cpu < nr_cpus; cpu++)
1677                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1678                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1679 }
1680
1681 static int update_fds(struct perf_evsel *evsel,
1682                       int nr_cpus, int cpu_idx,
1683                       int nr_threads, int thread_idx)
1684 {
1685         struct perf_evsel *pos;
1686
1687         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1688                 return -EINVAL;
1689
1690         evlist__for_each_entry(evsel->evlist, pos) {
1691                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1692
1693                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1694
1695                 /*
1696                  * Since fds for next evsel has not been created,
1697                  * there is no need to iterate whole event list.
1698                  */
1699                 if (pos == evsel)
1700                         break;
1701         }
1702         return 0;
1703 }
1704
1705 static bool ignore_missing_thread(struct perf_evsel *evsel,
1706                                   int nr_cpus, int cpu,
1707                                   struct thread_map *threads,
1708                                   int thread, int err)
1709 {
1710         pid_t ignore_pid = thread_map__pid(threads, thread);
1711
1712         if (!evsel->ignore_missing_thread)
1713                 return false;
1714
1715         /* The system wide setup does not work with threads. */
1716         if (evsel->system_wide)
1717                 return false;
1718
1719         /* The -ESRCH is perf event syscall errno for pid's not found. */
1720         if (err != -ESRCH)
1721                 return false;
1722
1723         /* If there's only one thread, let it fail. */
1724         if (threads->nr == 1)
1725                 return false;
1726
1727         /*
1728          * We should remove fd for missing_thread first
1729          * because thread_map__remove() will decrease threads->nr.
1730          */
1731         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1732                 return false;
1733
1734         if (thread_map__remove(threads, thread))
1735                 return false;
1736
1737         pr_warning("WARNING: Ignored open failure for pid %d\n",
1738                    ignore_pid);
1739         return true;
1740 }
1741
1742 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1743                      struct thread_map *threads)
1744 {
1745         int cpu, thread, nthreads;
1746         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1747         int pid = -1, err;
1748         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1749
1750         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1751                 return -EINVAL;
1752
1753         if (cpus == NULL) {
1754                 static struct cpu_map *empty_cpu_map;
1755
1756                 if (empty_cpu_map == NULL) {
1757                         empty_cpu_map = cpu_map__dummy_new();
1758                         if (empty_cpu_map == NULL)
1759                                 return -ENOMEM;
1760                 }
1761
1762                 cpus = empty_cpu_map;
1763         }
1764
1765         if (threads == NULL) {
1766                 static struct thread_map *empty_thread_map;
1767
1768                 if (empty_thread_map == NULL) {
1769                         empty_thread_map = thread_map__new_by_tid(-1);
1770                         if (empty_thread_map == NULL)
1771                                 return -ENOMEM;
1772                 }
1773
1774                 threads = empty_thread_map;
1775         }
1776
1777         if (evsel->system_wide)
1778                 nthreads = 1;
1779         else
1780                 nthreads = threads->nr;
1781
1782         if (evsel->fd == NULL &&
1783             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1784                 return -ENOMEM;
1785
1786         if (evsel->cgrp) {
1787                 flags |= PERF_FLAG_PID_CGROUP;
1788                 pid = evsel->cgrp->fd;
1789         }
1790
1791 fallback_missing_features:
1792         if (perf_missing_features.clockid_wrong)
1793                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1794         if (perf_missing_features.clockid) {
1795                 evsel->attr.use_clockid = 0;
1796                 evsel->attr.clockid = 0;
1797         }
1798         if (perf_missing_features.cloexec)
1799                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1800         if (perf_missing_features.mmap2)
1801                 evsel->attr.mmap2 = 0;
1802         if (perf_missing_features.exclude_guest)
1803                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1804         if (perf_missing_features.lbr_flags)
1805                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1806                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1807         if (perf_missing_features.group_read && evsel->attr.inherit)
1808                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1809 retry_sample_id:
1810         if (perf_missing_features.sample_id_all)
1811                 evsel->attr.sample_id_all = 0;
1812
1813         if (verbose >= 2) {
1814                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1815                 fprintf(stderr, "perf_event_attr:\n");
1816                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1817                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1818         }
1819
1820         for (cpu = 0; cpu < cpus->nr; cpu++) {
1821
1822                 for (thread = 0; thread < nthreads; thread++) {
1823                         int fd, group_fd;
1824
1825                         if (!evsel->cgrp && !evsel->system_wide)
1826                                 pid = thread_map__pid(threads, thread);
1827
1828                         group_fd = get_group_fd(evsel, cpu, thread);
1829 retry_open:
1830                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1831                                   pid, cpus->map[cpu], group_fd, flags);
1832
1833                         test_attr__ready();
1834
1835                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1836                                                  group_fd, flags);
1837
1838                         FD(evsel, cpu, thread) = fd;
1839
1840                         if (fd < 0) {
1841                                 err = -errno;
1842
1843                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1844                                         /*
1845                                          * We just removed 1 thread, so take a step
1846                                          * back on thread index and lower the upper
1847                                          * nthreads limit.
1848                                          */
1849                                         nthreads--;
1850                                         thread--;
1851
1852                                         /* ... and pretend like nothing have happened. */
1853                                         err = 0;
1854                                         continue;
1855                                 }
1856
1857                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1858                                           err);
1859                                 goto try_fallback;
1860                         }
1861
1862                         pr_debug2(" = %d\n", fd);
1863
1864                         if (evsel->bpf_fd >= 0) {
1865                                 int evt_fd = fd;
1866                                 int bpf_fd = evsel->bpf_fd;
1867
1868                                 err = ioctl(evt_fd,
1869                                             PERF_EVENT_IOC_SET_BPF,
1870                                             bpf_fd);
1871                                 if (err && errno != EEXIST) {
1872                                         pr_err("failed to attach bpf fd %d: %s\n",
1873                                                bpf_fd, strerror(errno));
1874                                         err = -EINVAL;
1875                                         goto out_close;
1876                                 }
1877                         }
1878
1879                         set_rlimit = NO_CHANGE;
1880
1881                         /*
1882                          * If we succeeded but had to kill clockid, fail and
1883                          * have perf_evsel__open_strerror() print us a nice
1884                          * error.
1885                          */
1886                         if (perf_missing_features.clockid ||
1887                             perf_missing_features.clockid_wrong) {
1888                                 err = -EINVAL;
1889                                 goto out_close;
1890                         }
1891                 }
1892         }
1893
1894         return 0;
1895
1896 try_fallback:
1897         /*
1898          * perf stat needs between 5 and 22 fds per CPU. When we run out
1899          * of them try to increase the limits.
1900          */
1901         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1902                 struct rlimit l;
1903                 int old_errno = errno;
1904
1905                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1906                         if (set_rlimit == NO_CHANGE)
1907                                 l.rlim_cur = l.rlim_max;
1908                         else {
1909                                 l.rlim_cur = l.rlim_max + 1000;
1910                                 l.rlim_max = l.rlim_cur;
1911                         }
1912                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1913                                 set_rlimit++;
1914                                 errno = old_errno;
1915                                 goto retry_open;
1916                         }
1917                 }
1918                 errno = old_errno;
1919         }
1920
1921         if (err != -EINVAL || cpu > 0 || thread > 0)
1922                 goto out_close;
1923
1924         /*
1925          * Must probe features in the order they were added to the
1926          * perf_event_attr interface.
1927          */
1928         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1929                 perf_missing_features.write_backward = true;
1930                 pr_debug2("switching off write_backward\n");
1931                 goto out_close;
1932         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1933                 perf_missing_features.clockid_wrong = true;
1934                 pr_debug2("switching off clockid\n");
1935                 goto fallback_missing_features;
1936         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1937                 perf_missing_features.clockid = true;
1938                 pr_debug2("switching off use_clockid\n");
1939                 goto fallback_missing_features;
1940         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1941                 perf_missing_features.cloexec = true;
1942                 pr_debug2("switching off cloexec flag\n");
1943                 goto fallback_missing_features;
1944         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1945                 perf_missing_features.mmap2 = true;
1946                 pr_debug2("switching off mmap2\n");
1947                 goto fallback_missing_features;
1948         } else if (!perf_missing_features.exclude_guest &&
1949                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1950                 perf_missing_features.exclude_guest = true;
1951                 pr_debug2("switching off exclude_guest, exclude_host\n");
1952                 goto fallback_missing_features;
1953         } else if (!perf_missing_features.sample_id_all) {
1954                 perf_missing_features.sample_id_all = true;
1955                 pr_debug2("switching off sample_id_all\n");
1956                 goto retry_sample_id;
1957         } else if (!perf_missing_features.lbr_flags &&
1958                         (evsel->attr.branch_sample_type &
1959                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1960                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1961                 perf_missing_features.lbr_flags = true;
1962                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1963                 goto fallback_missing_features;
1964         } else if (!perf_missing_features.group_read &&
1965                     evsel->attr.inherit &&
1966                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1967                    perf_evsel__is_group_leader(evsel)) {
1968                 perf_missing_features.group_read = true;
1969                 pr_debug2("switching off group read\n");
1970                 goto fallback_missing_features;
1971         }
1972 out_close:
1973         if (err)
1974                 threads->err_thread = thread;
1975
1976         do {
1977                 while (--thread >= 0) {
1978                         close(FD(evsel, cpu, thread));
1979                         FD(evsel, cpu, thread) = -1;
1980                 }
1981                 thread = nthreads;
1982         } while (--cpu >= 0);
1983         return err;
1984 }
1985
1986 void perf_evsel__close(struct perf_evsel *evsel)
1987 {
1988         if (evsel->fd == NULL)
1989                 return;
1990
1991         perf_evsel__close_fd(evsel);
1992         perf_evsel__free_fd(evsel);
1993 }
1994
1995 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1996                              struct cpu_map *cpus)
1997 {
1998         return perf_evsel__open(evsel, cpus, NULL);
1999 }
2000
2001 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2002                                 struct thread_map *threads)
2003 {
2004         return perf_evsel__open(evsel, NULL, threads);
2005 }
2006
2007 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2008                                        const union perf_event *event,
2009                                        struct perf_sample *sample)
2010 {
2011         u64 type = evsel->attr.sample_type;
2012         const u64 *array = event->sample.array;
2013         bool swapped = evsel->needs_swap;
2014         union u64_swap u;
2015
2016         array += ((event->header.size -
2017                    sizeof(event->header)) / sizeof(u64)) - 1;
2018
2019         if (type & PERF_SAMPLE_IDENTIFIER) {
2020                 sample->id = *array;
2021                 array--;
2022         }
2023
2024         if (type & PERF_SAMPLE_CPU) {
2025                 u.val64 = *array;
2026                 if (swapped) {
2027                         /* undo swap of u64, then swap on individual u32s */
2028                         u.val64 = bswap_64(u.val64);
2029                         u.val32[0] = bswap_32(u.val32[0]);
2030                 }
2031
2032                 sample->cpu = u.val32[0];
2033                 array--;
2034         }
2035
2036         if (type & PERF_SAMPLE_STREAM_ID) {
2037                 sample->stream_id = *array;
2038                 array--;
2039         }
2040
2041         if (type & PERF_SAMPLE_ID) {
2042                 sample->id = *array;
2043                 array--;
2044         }
2045
2046         if (type & PERF_SAMPLE_TIME) {
2047                 sample->time = *array;
2048                 array--;
2049         }
2050
2051         if (type & PERF_SAMPLE_TID) {
2052                 u.val64 = *array;
2053                 if (swapped) {
2054                         /* undo swap of u64, then swap on individual u32s */
2055                         u.val64 = bswap_64(u.val64);
2056                         u.val32[0] = bswap_32(u.val32[0]);
2057                         u.val32[1] = bswap_32(u.val32[1]);
2058                 }
2059
2060                 sample->pid = u.val32[0];
2061                 sample->tid = u.val32[1];
2062                 array--;
2063         }
2064
2065         return 0;
2066 }
2067
2068 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2069                             u64 size)
2070 {
2071         return size > max_size || offset + size > endp;
2072 }
2073
2074 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2075         do {                                                            \
2076                 if (overflow(endp, (max_size), (offset), (size)))       \
2077                         return -EFAULT;                                 \
2078         } while (0)
2079
2080 #define OVERFLOW_CHECK_u64(offset) \
2081         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2082
2083 static int
2084 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2085 {
2086         /*
2087          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2088          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2089          * check the format does not go past the end of the event.
2090          */
2091         if (sample_size + sizeof(event->header) > event->header.size)
2092                 return -EFAULT;
2093
2094         return 0;
2095 }
2096
2097 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2098                              struct perf_sample *data)
2099 {
2100         u64 type = evsel->attr.sample_type;
2101         bool swapped = evsel->needs_swap;
2102         const u64 *array;
2103         u16 max_size = event->header.size;
2104         const void *endp = (void *)event + max_size;
2105         u64 sz;
2106
2107         /*
2108          * used for cross-endian analysis. See git commit 65014ab3
2109          * for why this goofiness is needed.
2110          */
2111         union u64_swap u;
2112
2113         memset(data, 0, sizeof(*data));
2114         data->cpu = data->pid = data->tid = -1;
2115         data->stream_id = data->id = data->time = -1ULL;
2116         data->period = evsel->attr.sample_period;
2117         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2118         data->misc    = event->header.misc;
2119         data->id = -1ULL;
2120         data->data_src = PERF_MEM_DATA_SRC_NONE;
2121
2122         if (event->header.type != PERF_RECORD_SAMPLE) {
2123                 if (!evsel->attr.sample_id_all)
2124                         return 0;
2125                 return perf_evsel__parse_id_sample(evsel, event, data);
2126         }
2127
2128         array = event->sample.array;
2129
2130         if (perf_event__check_size(event, evsel->sample_size))
2131                 return -EFAULT;
2132
2133         if (type & PERF_SAMPLE_IDENTIFIER) {
2134                 data->id = *array;
2135                 array++;
2136         }
2137
2138         if (type & PERF_SAMPLE_IP) {
2139                 data->ip = *array;
2140                 array++;
2141         }
2142
2143         if (type & PERF_SAMPLE_TID) {
2144                 u.val64 = *array;
2145                 if (swapped) {
2146                         /* undo swap of u64, then swap on individual u32s */
2147                         u.val64 = bswap_64(u.val64);
2148                         u.val32[0] = bswap_32(u.val32[0]);
2149                         u.val32[1] = bswap_32(u.val32[1]);
2150                 }
2151
2152                 data->pid = u.val32[0];
2153                 data->tid = u.val32[1];
2154                 array++;
2155         }
2156
2157         if (type & PERF_SAMPLE_TIME) {
2158                 data->time = *array;
2159                 array++;
2160         }
2161
2162         if (type & PERF_SAMPLE_ADDR) {
2163                 data->addr = *array;
2164                 array++;
2165         }
2166
2167         if (type & PERF_SAMPLE_ID) {
2168                 data->id = *array;
2169                 array++;
2170         }
2171
2172         if (type & PERF_SAMPLE_STREAM_ID) {
2173                 data->stream_id = *array;
2174                 array++;
2175         }
2176
2177         if (type & PERF_SAMPLE_CPU) {
2178
2179                 u.val64 = *array;
2180                 if (swapped) {
2181                         /* undo swap of u64, then swap on individual u32s */
2182                         u.val64 = bswap_64(u.val64);
2183                         u.val32[0] = bswap_32(u.val32[0]);
2184                 }
2185
2186                 data->cpu = u.val32[0];
2187                 array++;
2188         }
2189
2190         if (type & PERF_SAMPLE_PERIOD) {
2191                 data->period = *array;
2192                 array++;
2193         }
2194
2195         if (type & PERF_SAMPLE_READ) {
2196                 u64 read_format = evsel->attr.read_format;
2197
2198                 OVERFLOW_CHECK_u64(array);
2199                 if (read_format & PERF_FORMAT_GROUP)
2200                         data->read.group.nr = *array;
2201                 else
2202                         data->read.one.value = *array;
2203
2204                 array++;
2205
2206                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2207                         OVERFLOW_CHECK_u64(array);
2208                         data->read.time_enabled = *array;
2209                         array++;
2210                 }
2211
2212                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2213                         OVERFLOW_CHECK_u64(array);
2214                         data->read.time_running = *array;
2215                         array++;
2216                 }
2217
2218                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2219                 if (read_format & PERF_FORMAT_GROUP) {
2220                         const u64 max_group_nr = UINT64_MAX /
2221                                         sizeof(struct sample_read_value);
2222
2223                         if (data->read.group.nr > max_group_nr)
2224                                 return -EFAULT;
2225                         sz = data->read.group.nr *
2226                              sizeof(struct sample_read_value);
2227                         OVERFLOW_CHECK(array, sz, max_size);
2228                         data->read.group.values =
2229                                         (struct sample_read_value *)array;
2230                         array = (void *)array + sz;
2231                 } else {
2232                         OVERFLOW_CHECK_u64(array);
2233                         data->read.one.id = *array;
2234                         array++;
2235                 }
2236         }
2237
2238         if (evsel__has_callchain(evsel)) {
2239                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2240
2241                 OVERFLOW_CHECK_u64(array);
2242                 data->callchain = (struct ip_callchain *)array++;
2243                 if (data->callchain->nr > max_callchain_nr)
2244                         return -EFAULT;
2245                 sz = data->callchain->nr * sizeof(u64);
2246                 OVERFLOW_CHECK(array, sz, max_size);
2247                 array = (void *)array + sz;
2248         }
2249
2250         if (type & PERF_SAMPLE_RAW) {
2251                 OVERFLOW_CHECK_u64(array);
2252                 u.val64 = *array;
2253
2254                 /*
2255                  * Undo swap of u64, then swap on individual u32s,
2256                  * get the size of the raw area and undo all of the
2257                  * swap. The pevent interface handles endianity by
2258                  * itself.
2259                  */
2260                 if (swapped) {
2261                         u.val64 = bswap_64(u.val64);
2262                         u.val32[0] = bswap_32(u.val32[0]);
2263                         u.val32[1] = bswap_32(u.val32[1]);
2264                 }
2265                 data->raw_size = u.val32[0];
2266
2267                 /*
2268                  * The raw data is aligned on 64bits including the
2269                  * u32 size, so it's safe to use mem_bswap_64.
2270                  */
2271                 if (swapped)
2272                         mem_bswap_64((void *) array, data->raw_size);
2273
2274                 array = (void *)array + sizeof(u32);
2275
2276                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2277                 data->raw_data = (void *)array;
2278                 array = (void *)array + data->raw_size;
2279         }
2280
2281         if (type & PERF_SAMPLE_BRANCH_STACK) {
2282                 const u64 max_branch_nr = UINT64_MAX /
2283                                           sizeof(struct branch_entry);
2284
2285                 OVERFLOW_CHECK_u64(array);
2286                 data->branch_stack = (struct branch_stack *)array++;
2287
2288                 if (data->branch_stack->nr > max_branch_nr)
2289                         return -EFAULT;
2290                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2291                 OVERFLOW_CHECK(array, sz, max_size);
2292                 array = (void *)array + sz;
2293         }
2294
2295         if (type & PERF_SAMPLE_REGS_USER) {
2296                 OVERFLOW_CHECK_u64(array);
2297                 data->user_regs.abi = *array;
2298                 array++;
2299
2300                 if (data->user_regs.abi) {
2301                         u64 mask = evsel->attr.sample_regs_user;
2302
2303                         sz = hweight_long(mask) * sizeof(u64);
2304                         OVERFLOW_CHECK(array, sz, max_size);
2305                         data->user_regs.mask = mask;
2306                         data->user_regs.regs = (u64 *)array;
2307                         array = (void *)array + sz;
2308                 }
2309         }
2310
2311         if (type & PERF_SAMPLE_STACK_USER) {
2312                 OVERFLOW_CHECK_u64(array);
2313                 sz = *array++;
2314
2315                 data->user_stack.offset = ((char *)(array - 1)
2316                                           - (char *) event);
2317
2318                 if (!sz) {
2319                         data->user_stack.size = 0;
2320                 } else {
2321                         OVERFLOW_CHECK(array, sz, max_size);
2322                         data->user_stack.data = (char *)array;
2323                         array = (void *)array + sz;
2324                         OVERFLOW_CHECK_u64(array);
2325                         data->user_stack.size = *array++;
2326                         if (WARN_ONCE(data->user_stack.size > sz,
2327                                       "user stack dump failure\n"))
2328                                 return -EFAULT;
2329                 }
2330         }
2331
2332         if (type & PERF_SAMPLE_WEIGHT) {
2333                 OVERFLOW_CHECK_u64(array);
2334                 data->weight = *array;
2335                 array++;
2336         }
2337
2338         if (type & PERF_SAMPLE_DATA_SRC) {
2339                 OVERFLOW_CHECK_u64(array);
2340                 data->data_src = *array;
2341                 array++;
2342         }
2343
2344         if (type & PERF_SAMPLE_TRANSACTION) {
2345                 OVERFLOW_CHECK_u64(array);
2346                 data->transaction = *array;
2347                 array++;
2348         }
2349
2350         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2351         if (type & PERF_SAMPLE_REGS_INTR) {
2352                 OVERFLOW_CHECK_u64(array);
2353                 data->intr_regs.abi = *array;
2354                 array++;
2355
2356                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2357                         u64 mask = evsel->attr.sample_regs_intr;
2358
2359                         sz = hweight_long(mask) * sizeof(u64);
2360                         OVERFLOW_CHECK(array, sz, max_size);
2361                         data->intr_regs.mask = mask;
2362                         data->intr_regs.regs = (u64 *)array;
2363                         array = (void *)array + sz;
2364                 }
2365         }
2366
2367         data->phys_addr = 0;
2368         if (type & PERF_SAMPLE_PHYS_ADDR) {
2369                 data->phys_addr = *array;
2370                 array++;
2371         }
2372
2373         return 0;
2374 }
2375
2376 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2377                                        union perf_event *event,
2378                                        u64 *timestamp)
2379 {
2380         u64 type = evsel->attr.sample_type;
2381         const u64 *array;
2382
2383         if (!(type & PERF_SAMPLE_TIME))
2384                 return -1;
2385
2386         if (event->header.type != PERF_RECORD_SAMPLE) {
2387                 struct perf_sample data = {
2388                         .time = -1ULL,
2389                 };
2390
2391                 if (!evsel->attr.sample_id_all)
2392                         return -1;
2393                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2394                         return -1;
2395
2396                 *timestamp = data.time;
2397                 return 0;
2398         }
2399
2400         array = event->sample.array;
2401
2402         if (perf_event__check_size(event, evsel->sample_size))
2403                 return -EFAULT;
2404
2405         if (type & PERF_SAMPLE_IDENTIFIER)
2406                 array++;
2407
2408         if (type & PERF_SAMPLE_IP)
2409                 array++;
2410
2411         if (type & PERF_SAMPLE_TID)
2412                 array++;
2413
2414         if (type & PERF_SAMPLE_TIME)
2415                 *timestamp = *array;
2416
2417         return 0;
2418 }
2419
2420 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2421                                      u64 read_format)
2422 {
2423         size_t sz, result = sizeof(struct sample_event);
2424
2425         if (type & PERF_SAMPLE_IDENTIFIER)
2426                 result += sizeof(u64);
2427
2428         if (type & PERF_SAMPLE_IP)
2429                 result += sizeof(u64);
2430
2431         if (type & PERF_SAMPLE_TID)
2432                 result += sizeof(u64);
2433
2434         if (type & PERF_SAMPLE_TIME)
2435                 result += sizeof(u64);
2436
2437         if (type & PERF_SAMPLE_ADDR)
2438                 result += sizeof(u64);
2439
2440         if (type & PERF_SAMPLE_ID)
2441                 result += sizeof(u64);
2442
2443         if (type & PERF_SAMPLE_STREAM_ID)
2444                 result += sizeof(u64);
2445
2446         if (type & PERF_SAMPLE_CPU)
2447                 result += sizeof(u64);
2448
2449         if (type & PERF_SAMPLE_PERIOD)
2450                 result += sizeof(u64);
2451
2452         if (type & PERF_SAMPLE_READ) {
2453                 result += sizeof(u64);
2454                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2455                         result += sizeof(u64);
2456                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2457                         result += sizeof(u64);
2458                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2459                 if (read_format & PERF_FORMAT_GROUP) {
2460                         sz = sample->read.group.nr *
2461                              sizeof(struct sample_read_value);
2462                         result += sz;
2463                 } else {
2464                         result += sizeof(u64);
2465                 }
2466         }
2467
2468         if (type & PERF_SAMPLE_CALLCHAIN) {
2469                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2470                 result += sz;
2471         }
2472
2473         if (type & PERF_SAMPLE_RAW) {
2474                 result += sizeof(u32);
2475                 result += sample->raw_size;
2476         }
2477
2478         if (type & PERF_SAMPLE_BRANCH_STACK) {
2479                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2480                 sz += sizeof(u64);
2481                 result += sz;
2482         }
2483
2484         if (type & PERF_SAMPLE_REGS_USER) {
2485                 if (sample->user_regs.abi) {
2486                         result += sizeof(u64);
2487                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2488                         result += sz;
2489                 } else {
2490                         result += sizeof(u64);
2491                 }
2492         }
2493
2494         if (type & PERF_SAMPLE_STACK_USER) {
2495                 sz = sample->user_stack.size;
2496                 result += sizeof(u64);
2497                 if (sz) {
2498                         result += sz;
2499                         result += sizeof(u64);
2500                 }
2501         }
2502
2503         if (type & PERF_SAMPLE_WEIGHT)
2504                 result += sizeof(u64);
2505
2506         if (type & PERF_SAMPLE_DATA_SRC)
2507                 result += sizeof(u64);
2508
2509         if (type & PERF_SAMPLE_TRANSACTION)
2510                 result += sizeof(u64);
2511
2512         if (type & PERF_SAMPLE_REGS_INTR) {
2513                 if (sample->intr_regs.abi) {
2514                         result += sizeof(u64);
2515                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2516                         result += sz;
2517                 } else {
2518                         result += sizeof(u64);
2519                 }
2520         }
2521
2522         if (type & PERF_SAMPLE_PHYS_ADDR)
2523                 result += sizeof(u64);
2524
2525         return result;
2526 }
2527
2528 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2529                                   u64 read_format,
2530                                   const struct perf_sample *sample)
2531 {
2532         u64 *array;
2533         size_t sz;
2534         /*
2535          * used for cross-endian analysis. See git commit 65014ab3
2536          * for why this goofiness is needed.
2537          */
2538         union u64_swap u;
2539
2540         array = event->sample.array;
2541
2542         if (type & PERF_SAMPLE_IDENTIFIER) {
2543                 *array = sample->id;
2544                 array++;
2545         }
2546
2547         if (type & PERF_SAMPLE_IP) {
2548                 *array = sample->ip;
2549                 array++;
2550         }
2551
2552         if (type & PERF_SAMPLE_TID) {
2553                 u.val32[0] = sample->pid;
2554                 u.val32[1] = sample->tid;
2555                 *array = u.val64;
2556                 array++;
2557         }
2558
2559         if (type & PERF_SAMPLE_TIME) {
2560                 *array = sample->time;
2561                 array++;
2562         }
2563
2564         if (type & PERF_SAMPLE_ADDR) {
2565                 *array = sample->addr;
2566                 array++;
2567         }
2568
2569         if (type & PERF_SAMPLE_ID) {
2570                 *array = sample->id;
2571                 array++;
2572         }
2573
2574         if (type & PERF_SAMPLE_STREAM_ID) {
2575                 *array = sample->stream_id;
2576                 array++;
2577         }
2578
2579         if (type & PERF_SAMPLE_CPU) {
2580                 u.val32[0] = sample->cpu;
2581                 u.val32[1] = 0;
2582                 *array = u.val64;
2583                 array++;
2584         }
2585
2586         if (type & PERF_SAMPLE_PERIOD) {
2587                 *array = sample->period;
2588                 array++;
2589         }
2590
2591         if (type & PERF_SAMPLE_READ) {
2592                 if (read_format & PERF_FORMAT_GROUP)
2593                         *array = sample->read.group.nr;
2594                 else
2595                         *array = sample->read.one.value;
2596                 array++;
2597
2598                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2599                         *array = sample->read.time_enabled;
2600                         array++;
2601                 }
2602
2603                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2604                         *array = sample->read.time_running;
2605                         array++;
2606                 }
2607
2608                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2609                 if (read_format & PERF_FORMAT_GROUP) {
2610                         sz = sample->read.group.nr *
2611                              sizeof(struct sample_read_value);
2612                         memcpy(array, sample->read.group.values, sz);
2613                         array = (void *)array + sz;
2614                 } else {
2615                         *array = sample->read.one.id;
2616                         array++;
2617                 }
2618         }
2619
2620         if (type & PERF_SAMPLE_CALLCHAIN) {
2621                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2622                 memcpy(array, sample->callchain, sz);
2623                 array = (void *)array + sz;
2624         }
2625
2626         if (type & PERF_SAMPLE_RAW) {
2627                 u.val32[0] = sample->raw_size;
2628                 *array = u.val64;
2629                 array = (void *)array + sizeof(u32);
2630
2631                 memcpy(array, sample->raw_data, sample->raw_size);
2632                 array = (void *)array + sample->raw_size;
2633         }
2634
2635         if (type & PERF_SAMPLE_BRANCH_STACK) {
2636                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2637                 sz += sizeof(u64);
2638                 memcpy(array, sample->branch_stack, sz);
2639                 array = (void *)array + sz;
2640         }
2641
2642         if (type & PERF_SAMPLE_REGS_USER) {
2643                 if (sample->user_regs.abi) {
2644                         *array++ = sample->user_regs.abi;
2645                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2646                         memcpy(array, sample->user_regs.regs, sz);
2647                         array = (void *)array + sz;
2648                 } else {
2649                         *array++ = 0;
2650                 }
2651         }
2652
2653         if (type & PERF_SAMPLE_STACK_USER) {
2654                 sz = sample->user_stack.size;
2655                 *array++ = sz;
2656                 if (sz) {
2657                         memcpy(array, sample->user_stack.data, sz);
2658                         array = (void *)array + sz;
2659                         *array++ = sz;
2660                 }
2661         }
2662
2663         if (type & PERF_SAMPLE_WEIGHT) {
2664                 *array = sample->weight;
2665                 array++;
2666         }
2667
2668         if (type & PERF_SAMPLE_DATA_SRC) {
2669                 *array = sample->data_src;
2670                 array++;
2671         }
2672
2673         if (type & PERF_SAMPLE_TRANSACTION) {
2674                 *array = sample->transaction;
2675                 array++;
2676         }
2677
2678         if (type & PERF_SAMPLE_REGS_INTR) {
2679                 if (sample->intr_regs.abi) {
2680                         *array++ = sample->intr_regs.abi;
2681                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2682                         memcpy(array, sample->intr_regs.regs, sz);
2683                         array = (void *)array + sz;
2684                 } else {
2685                         *array++ = 0;
2686                 }
2687         }
2688
2689         if (type & PERF_SAMPLE_PHYS_ADDR) {
2690                 *array = sample->phys_addr;
2691                 array++;
2692         }
2693
2694         return 0;
2695 }
2696
2697 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2698 {
2699         return tep_find_field(evsel->tp_format, name);
2700 }
2701
2702 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2703                          const char *name)
2704 {
2705         struct format_field *field = perf_evsel__field(evsel, name);
2706         int offset;
2707
2708         if (!field)
2709                 return NULL;
2710
2711         offset = field->offset;
2712
2713         if (field->flags & FIELD_IS_DYNAMIC) {
2714                 offset = *(int *)(sample->raw_data + field->offset);
2715                 offset &= 0xffff;
2716         }
2717
2718         return sample->raw_data + offset;
2719 }
2720
2721 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2722                          bool needs_swap)
2723 {
2724         u64 value;
2725         void *ptr = sample->raw_data + field->offset;
2726
2727         switch (field->size) {
2728         case 1:
2729                 return *(u8 *)ptr;
2730         case 2:
2731                 value = *(u16 *)ptr;
2732                 break;
2733         case 4:
2734                 value = *(u32 *)ptr;
2735                 break;
2736         case 8:
2737                 memcpy(&value, ptr, sizeof(u64));
2738                 break;
2739         default:
2740                 return 0;
2741         }
2742
2743         if (!needs_swap)
2744                 return value;
2745
2746         switch (field->size) {
2747         case 2:
2748                 return bswap_16(value);
2749         case 4:
2750                 return bswap_32(value);
2751         case 8:
2752                 return bswap_64(value);
2753         default:
2754                 return 0;
2755         }
2756
2757         return 0;
2758 }
2759
2760 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2761                        const char *name)
2762 {
2763         struct format_field *field = perf_evsel__field(evsel, name);
2764
2765         if (!field)
2766                 return 0;
2767
2768         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2769 }
2770
2771 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2772                           char *msg, size_t msgsize)
2773 {
2774         int paranoid;
2775
2776         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2777             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2778             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2779                 /*
2780                  * If it's cycles then fall back to hrtimer based
2781                  * cpu-clock-tick sw counter, which is always available even if
2782                  * no PMU support.
2783                  *
2784                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2785                  * b0a873e).
2786                  */
2787                 scnprintf(msg, msgsize, "%s",
2788 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2789
2790                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2791                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2792
2793                 zfree(&evsel->name);
2794                 return true;
2795         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2796                    (paranoid = perf_event_paranoid()) > 1) {
2797                 const char *name = perf_evsel__name(evsel);
2798                 char *new_name;
2799                 const char *sep = ":";
2800
2801                 /* Is there already the separator in the name. */
2802                 if (strchr(name, '/') ||
2803                     strchr(name, ':'))
2804                         sep = "";
2805
2806                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2807                         return false;
2808
2809                 if (evsel->name)
2810                         free(evsel->name);
2811                 evsel->name = new_name;
2812                 scnprintf(msg, msgsize,
2813 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2814                 evsel->attr.exclude_kernel = 1;
2815
2816                 return true;
2817         }
2818
2819         return false;
2820 }
2821
2822 static bool find_process(const char *name)
2823 {
2824         size_t len = strlen(name);
2825         DIR *dir;
2826         struct dirent *d;
2827         int ret = -1;
2828
2829         dir = opendir(procfs__mountpoint());
2830         if (!dir)
2831                 return false;
2832
2833         /* Walk through the directory. */
2834         while (ret && (d = readdir(dir)) != NULL) {
2835                 char path[PATH_MAX];
2836                 char *data;
2837                 size_t size;
2838
2839                 if ((d->d_type != DT_DIR) ||
2840                      !strcmp(".", d->d_name) ||
2841                      !strcmp("..", d->d_name))
2842                         continue;
2843
2844                 scnprintf(path, sizeof(path), "%s/%s/comm",
2845                           procfs__mountpoint(), d->d_name);
2846
2847                 if (filename__read_str(path, &data, &size))
2848                         continue;
2849
2850                 ret = strncmp(name, data, len);
2851                 free(data);
2852         }
2853
2854         closedir(dir);
2855         return ret ? false : true;
2856 }
2857
2858 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2859                               int err, char *msg, size_t size)
2860 {
2861         char sbuf[STRERR_BUFSIZE];
2862         int printed = 0;
2863
2864         switch (err) {
2865         case EPERM:
2866         case EACCES:
2867                 if (err == EPERM)
2868                         printed = scnprintf(msg, size,
2869                                 "No permission to enable %s event.\n\n",
2870                                 perf_evsel__name(evsel));
2871
2872                 return scnprintf(msg + printed, size - printed,
2873                  "You may not have permission to collect %sstats.\n\n"
2874                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2875                  "which controls use of the performance events system by\n"
2876                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2877                  "The current value is %d:\n\n"
2878                  "  -1: Allow use of (almost) all events by all users\n"
2879                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2880                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2881                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2882                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2883                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2884                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2885                  "      kernel.perf_event_paranoid = -1\n" ,
2886                                  target->system_wide ? "system-wide " : "",
2887                                  perf_event_paranoid());
2888         case ENOENT:
2889                 return scnprintf(msg, size, "The %s event is not supported.",
2890                                  perf_evsel__name(evsel));
2891         case EMFILE:
2892                 return scnprintf(msg, size, "%s",
2893                          "Too many events are opened.\n"
2894                          "Probably the maximum number of open file descriptors has been reached.\n"
2895                          "Hint: Try again after reducing the number of events.\n"
2896                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2897         case ENOMEM:
2898                 if (evsel__has_callchain(evsel) &&
2899                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2900                         return scnprintf(msg, size,
2901                                          "Not enough memory to setup event with callchain.\n"
2902                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2903                                          "Hint: Current value: %d", sysctl__max_stack());
2904                 break;
2905         case ENODEV:
2906                 if (target->cpu_list)
2907                         return scnprintf(msg, size, "%s",
2908          "No such device - did you specify an out-of-range profile CPU?");
2909                 break;
2910         case EOPNOTSUPP:
2911                 if (evsel->attr.sample_period != 0)
2912                         return scnprintf(msg, size,
2913         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2914                                          perf_evsel__name(evsel));
2915                 if (evsel->attr.precise_ip)
2916                         return scnprintf(msg, size, "%s",
2917         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2918 #if defined(__i386__) || defined(__x86_64__)
2919                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2920                         return scnprintf(msg, size, "%s",
2921         "No hardware sampling interrupt available.\n");
2922 #endif
2923                 break;
2924         case EBUSY:
2925                 if (find_process("oprofiled"))
2926                         return scnprintf(msg, size,
2927         "The PMU counters are busy/taken by another profiler.\n"
2928         "We found oprofile daemon running, please stop it and try again.");
2929                 break;
2930         case EINVAL:
2931                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2932                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2933                 if (perf_missing_features.clockid)
2934                         return scnprintf(msg, size, "clockid feature not supported.");
2935                 if (perf_missing_features.clockid_wrong)
2936                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2937                 break;
2938         default:
2939                 break;
2940         }
2941
2942         return scnprintf(msg, size,
2943         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2944         "/bin/dmesg | grep -i perf may provide additional information.\n",
2945                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2946                          perf_evsel__name(evsel));
2947 }
2948
2949 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2950 {
2951         if (evsel && evsel->evlist)
2952                 return evsel->evlist->env;
2953         return NULL;
2954 }