tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <[email protected]>
   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 <linux/bitops.h>
  12 #include <api/fs/debugfs.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <sys/resource.h>
  17 #include "asm/bug.h"
  18 #include "callchain.h"
  19 #include "cgroup.h"
  20 #include "evsel.h"
  21 #include "evlist.h"
  22 #include "util.h"
  23 #include "cpumap.h"
  24 #include "thread_map.h"
  25 #include "target.h"
  26 #include "perf_regs.h"
  27 #include "debug.h"
  28 #include "trace-event.h"
  29 #include "stat.h"
  30
  31 static struct {
  32         bool sample_id_all;
  33         bool exclude_guest;
  34         bool mmap2;
  35         bool cloexec;
  36         bool clockid;
  37         bool clockid_wrong;
  38 } perf_missing_features;
  39
  40 static clockid_t clockid;
  41
  42 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  43 {
  44         return 0;
  45 }
  46
  47 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  48 {
  49 }
  50
  51 static struct {
  52         size_t  size;
  53         int     (*init)(struct perf_evsel *evsel);
  54         void    (*fini)(struct perf_evsel *evsel);
  55 } perf_evsel__object = {
  56         .size = sizeof(struct perf_evsel),
  57         .init = perf_evsel__no_extra_init,
  58         .fini = perf_evsel__no_extra_fini,
  59 };
  60
  61 int perf_evsel__object_config(size_t object_size,
  62                               int (*init)(struct perf_evsel *evsel),
  63                               void (*fini)(struct perf_evsel *evsel))
  64 {
  65
  66         if (object_size == 0)
  67                 goto set_methods;
  68
  69         if (perf_evsel__object.size > object_size)
  70                 return -EINVAL;
  71
  72         perf_evsel__object.size = object_size;
  73
  74 set_methods:
  75         if (init != NULL)
  76                 perf_evsel__object.init = init;
  77
  78         if (fini != NULL)
  79                 perf_evsel__object.fini = fini;
  80
  81         return 0;
  82 }
  83
  84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  85
  86 int __perf_evsel__sample_size(u64 sample_type)
  87 {
  88         u64 mask = sample_type & PERF_SAMPLE_MASK;
  89         int size = 0;
  90         int i;
  91
  92         for (i = 0; i < 64; i++) {
  93                 if (mask & (1ULL << i))
  94                         size++;
  95         }
  96
  97         size *= sizeof(u64);
  98
  99         return size;
 100 }
 101
 102 /**
 103  * __perf_evsel__calc_id_pos - calculate id_pos.
 104  * @sample_type: sample type
 105  *
 106  * This function returns the position of the event id (PERF_SAMPLE_ID or
 107  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 108  * sample_event.
 109  */
 110 static int __perf_evsel__calc_id_pos(u64 sample_type)
 111 {
 112         int idx = 0;
 113
 114         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 115                 return 0;
 116
 117         if (!(sample_type & PERF_SAMPLE_ID))
 118                 return -1;
 119
 120         if (sample_type & PERF_SAMPLE_IP)
 121                 idx += 1;
 122
 123         if (sample_type & PERF_SAMPLE_TID)
 124                 idx += 1;
 125
 126         if (sample_type & PERF_SAMPLE_TIME)
 127                 idx += 1;
 128
 129         if (sample_type & PERF_SAMPLE_ADDR)
 130                 idx += 1;
 131
 132         return idx;
 133 }
 134
 135 /**
 136  * __perf_evsel__calc_is_pos - calculate is_pos.
 137  * @sample_type: sample type
 138  *
 139  * This function returns the position (counting backwards) of the event id
 140  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 141  * sample_id_all is used there is an id sample appended to non-sample events.
 142  */
 143 static int __perf_evsel__calc_is_pos(u64 sample_type)
 144 {
 145         int idx = 1;
 146
 147         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 148                 return 1;
 149
 150         if (!(sample_type & PERF_SAMPLE_ID))
 151                 return -1;
 152
 153         if (sample_type & PERF_SAMPLE_CPU)
 154                 idx += 1;
 155
 156         if (sample_type & PERF_SAMPLE_STREAM_ID)
 157                 idx += 1;
 158
 159         return idx;
 160 }
 161
 162 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 163 {
 164         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 165         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 166 }
 167
 168 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 169                                   enum perf_event_sample_format bit)
 170 {
 171         if (!(evsel->attr.sample_type & bit)) {
 172                 evsel->attr.sample_type |= bit;
 173                 evsel->sample_size += sizeof(u64);
 174                 perf_evsel__calc_id_pos(evsel);
 175         }
 176 }
 177
 178 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 179                                     enum perf_event_sample_format bit)
 180 {
 181         if (evsel->attr.sample_type & bit) {
 182                 evsel->attr.sample_type &= ~bit;
 183                 evsel->sample_size -= sizeof(u64);
 184                 perf_evsel__calc_id_pos(evsel);
 185         }
 186 }
 187
 188 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 189                                bool can_sample_identifier)
 190 {
 191         if (can_sample_identifier) {
 192                 perf_evsel__reset_sample_bit(evsel, ID);
 193                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 194         } else {
 195                 perf_evsel__set_sample_bit(evsel, ID);
 196         }
 197         evsel->attr.read_format |= PERF_FORMAT_ID;
 198 }
 199
 200 void perf_evsel__init(struct perf_evsel *evsel,
 201                       struct perf_event_attr *attr, int idx)
 202 {
 203         evsel->idx         = idx;
 204         evsel->tracking    = !idx;
 205         evsel->attr        = *attr;
 206         evsel->leader      = evsel;
 207         evsel->unit        = "";
 208         evsel->scale       = 1.0;
 209         INIT_LIST_HEAD(&evsel->node);
 210         perf_evsel__object.init(evsel);
 211         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 212         perf_evsel__calc_id_pos(evsel);
 213 }
 214
 215 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 216 {
 217         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 218
 219         if (evsel != NULL)
 220                 perf_evsel__init(evsel, attr, idx);
 221
 222         return evsel;
 223 }
 224
 225 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 226 {
 227         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 228
 229         if (evsel != NULL) {
 230                 struct perf_event_attr attr = {
 231                         .type          = PERF_TYPE_TRACEPOINT,
 232                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 233                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 234                 };
 235
 236                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 237                         goto out_free;
 238
 239                 evsel->tp_format = trace_event__tp_format(sys, name);
 240                 if (evsel->tp_format == NULL)
 241                         goto out_free;
 242
 243                 event_attr_init(&attr);
 244                 attr.config = evsel->tp_format->id;
 245                 attr.sample_period = 1;
 246                 perf_evsel__init(evsel, &attr, idx);
 247         }
 248
 249         return evsel;
 250
 251 out_free:
 252         zfree(&evsel->name);
 253         free(evsel);
 254         return NULL;
 255 }
 256
 257 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 258         "cycles",
 259         "instructions",
 260         "cache-references",
 261         "cache-misses",
 262         "branches",
 263         "branch-misses",
 264         "bus-cycles",
 265         "stalled-cycles-frontend",
 266         "stalled-cycles-backend",
 267         "ref-cycles",
 268 };
 269
 270 static const char *__perf_evsel__hw_name(u64 config)
 271 {
 272         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 273                 return perf_evsel__hw_names[config];
 274
 275         return "unknown-hardware";
 276 }
 277
 278 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 279 {
 280         int colon = 0, r = 0;
 281         struct perf_event_attr *attr = &evsel->attr;
 282         bool exclude_guest_default = false;
 283
 284 #define MOD_PRINT(context, mod) do {                                    \
 285                 if (!attr->exclude_##context) {                         \
 286                         if (!colon) colon = ++r;                        \
 287                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 288                 } } while(0)
 289
 290         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 291                 MOD_PRINT(kernel, 'k');
 292                 MOD_PRINT(user, 'u');
 293                 MOD_PRINT(hv, 'h');
 294                 exclude_guest_default = true;
 295         }
 296
 297         if (attr->precise_ip) {
 298                 if (!colon)
 299                         colon = ++r;
 300                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 301                 exclude_guest_default = true;
 302         }
 303
 304         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 305                 MOD_PRINT(host, 'H');
 306                 MOD_PRINT(guest, 'G');
 307         }
 308 #undef MOD_PRINT
 309         if (colon)
 310                 bf[colon - 1] = ':';
 311         return r;
 312 }
 313
 314 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 315 {
 316         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 317         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 318 }
 319
 320 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 321         "cpu-clock",
 322         "task-clock",
 323         "page-faults",
 324         "context-switches",
 325         "cpu-migrations",
 326         "minor-faults",
 327         "major-faults",
 328         "alignment-faults",
 329         "emulation-faults",
 330         "dummy",
 331 };
 332
 333 static const char *__perf_evsel__sw_name(u64 config)
 334 {
 335         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 336                 return perf_evsel__sw_names[config];
 337         return "unknown-software";
 338 }
 339
 340 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 341 {
 342         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 343         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 344 }
 345
 346 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 347 {
 348         int r;
 349
 350         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 351
 352         if (type & HW_BREAKPOINT_R)
 353                 r += scnprintf(bf + r, size - r, "r");
 354
 355         if (type & HW_BREAKPOINT_W)
 356                 r += scnprintf(bf + r, size - r, "w");
 357
 358         if (type & HW_BREAKPOINT_X)
 359                 r += scnprintf(bf + r, size - r, "x");
 360
 361         return r;
 362 }
 363
 364 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 365 {
 366         struct perf_event_attr *attr = &evsel->attr;
 367         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 368         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 369 }
 370
 371 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 372                                 [PERF_EVSEL__MAX_ALIASES] = {
 373  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 374  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 375  { "LLC",       "L2",                                                   },
 376  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 377  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 378  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 379  { "node",                                                              },
 380 };
 381
 382 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 383                                    [PERF_EVSEL__MAX_ALIASES] = {
 384  { "load",      "loads",        "read",                                 },
 385  { "store",     "stores",       "write",                                },
 386  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 387 };
 388
 389 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 390                                        [PERF_EVSEL__MAX_ALIASES] = {
 391  { "refs",      "Reference",    "ops",          "access",               },
 392  { "misses",    "miss",                                                 },
 393 };
 394
 395 #define C(x)            PERF_COUNT_HW_CACHE_##x
 396 #define CACHE_READ      (1 << C(OP_READ))
 397 #define CACHE_WRITE     (1 << C(OP_WRITE))
 398 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 399 #define COP(x)          (1 << x)
 400
 401 /*
 402  * cache operartion stat
 403  * L1I : Read and prefetch only
 404  * ITLB and BPU : Read-only
 405  */
 406 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 407  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 408  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 409  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 410  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 411  [C(ITLB)]      = (CACHE_READ),
 412  [C(BPU)]       = (CACHE_READ),
 413  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 414 };
 415
 416 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 417 {
 418         if (perf_evsel__hw_cache_stat[type] & COP(op))
 419                 return true;    /* valid */
 420         else
 421                 return false;   /* invalid */
 422 }
 423
 424 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 425                                             char *bf, size_t size)
 426 {
 427         if (result) {
 428                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 429                                  perf_evsel__hw_cache_op[op][0],
 430                                  perf_evsel__hw_cache_result[result][0]);
 431         }
 432
 433         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 434                          perf_evsel__hw_cache_op[op][1]);
 435 }
 436
 437 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 438 {
 439         u8 op, result, type = (config >>  0) & 0xff;
 440         const char *err = "unknown-ext-hardware-cache-type";
 441
 442         if (type > PERF_COUNT_HW_CACHE_MAX)
 443                 goto out_err;
 444
 445         op = (config >>  8) & 0xff;
 446         err = "unknown-ext-hardware-cache-op";
 447         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 448                 goto out_err;
 449
 450         result = (config >> 16) & 0xff;
 451         err = "unknown-ext-hardware-cache-result";
 452         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 453                 goto out_err;
 454
 455         err = "invalid-cache";
 456         if (!perf_evsel__is_cache_op_valid(type, op))
 457                 goto out_err;
 458
 459         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 460 out_err:
 461         return scnprintf(bf, size, "%s", err);
 462 }
 463
 464 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 465 {
 466         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 467         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 468 }
 469
 470 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 471 {
 472         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 473         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 474 }
 475
 476 const char *perf_evsel__name(struct perf_evsel *evsel)
 477 {
 478         char bf[128];
 479
 480         if (evsel->name)
 481                 return evsel->name;
 482
 483         switch (evsel->attr.type) {
 484         case PERF_TYPE_RAW:
 485                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 486                 break;
 487
 488         case PERF_TYPE_HARDWARE:
 489                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 490                 break;
 491
 492         case PERF_TYPE_HW_CACHE:
 493                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 494                 break;
 495
 496         case PERF_TYPE_SOFTWARE:
 497                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 498                 break;
 499
 500         case PERF_TYPE_TRACEPOINT:
 501                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 502                 break;
 503
 504         case PERF_TYPE_BREAKPOINT:
 505                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 506                 break;
 507
 508         default:
 509                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 510                           evsel->attr.type);
 511                 break;
 512         }
 513
 514         evsel->name = strdup(bf);
 515
 516         return evsel->name ?: "unknown";
 517 }
 518
 519 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 520 {
 521         return evsel->group_name ?: "anon group";
 522 }
 523
 524 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 525 {
 526         int ret;
 527         struct perf_evsel *pos;
 528         const char *group_name = perf_evsel__group_name(evsel);
 529
 530         ret = scnprintf(buf, size, "%s", group_name);
 531
 532         ret += scnprintf(buf + ret, size - ret, " { %s",
 533                          perf_evsel__name(evsel));
 534
 535         for_each_group_member(pos, evsel)
 536                 ret += scnprintf(buf + ret, size - ret, ", %s",
 537                                  perf_evsel__name(pos));
 538
 539         ret += scnprintf(buf + ret, size - ret, " }");
 540
 541         return ret;
 542 }
 543
 544 static void
 545 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 546                              struct record_opts *opts)
 547 {
 548         bool function = perf_evsel__is_function_event(evsel);
 549         struct perf_event_attr *attr = &evsel->attr;
 550
 551         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 552
 553         if (callchain_param.record_mode == CALLCHAIN_LBR) {
 554                 if (!opts->branch_stack) {
 555                         if (attr->exclude_user) {
 556                                 pr_warning("LBR callstack option is only available "
 557                                            "to get user callchain information. "
 558                                            "Falling back to framepointers.\n");
 559                         } else {
 560                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 561                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 562                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 563                         }
 564                 } else
 565                          pr_warning("Cannot use LBR callstack with branch stack. "
 566                                     "Falling back to framepointers.\n");
 567         }
 568
 569         if (callchain_param.record_mode == CALLCHAIN_DWARF) {
 570                 if (!function) {
 571                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 572                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 573                         attr->sample_regs_user = PERF_REGS_MASK;
 574                         attr->sample_stack_user = callchain_param.dump_size;
 575                         attr->exclude_callchain_user = 1;
 576                 } else {
 577                         pr_info("Cannot use DWARF unwind for function trace event,"
 578                                 " falling back to framepointers.\n");
 579                 }
 580         }
 581
 582         if (function) {
 583                 pr_info("Disabling user space callchains for function trace event.\n");
 584                 attr->exclude_callchain_user = 1;
 585         }
 586 }
 587
 588 /*
 589  * The enable_on_exec/disabled value strategy:
 590  *
 591  *  1) For any type of traced program:
 592  *    - all independent events and group leaders are disabled
 593  *    - all group members are enabled
 594  *
 595  *     Group members are ruled by group leaders. They need to
 596  *     be enabled, because the group scheduling relies on that.
 597  *
 598  *  2) For traced programs executed by perf:
 599  *     - all independent events and group leaders have
 600  *       enable_on_exec set
 601  *     - we don't specifically enable or disable any event during
 602  *       the record command
 603  *
 604  *     Independent events and group leaders are initially disabled
 605  *     and get enabled by exec. Group members are ruled by group
 606  *     leaders as stated in 1).
 607  *
 608  *  3) For traced programs attached by perf (pid/tid):
 609  *     - we specifically enable or disable all events during
 610  *       the record command
 611  *
 612  *     When attaching events to already running traced we
 613  *     enable/disable events specifically, as there's no
 614  *     initial traced exec call.
 615  */
 616 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 617 {
 618         struct perf_evsel *leader = evsel->leader;
 619         struct perf_event_attr *attr = &evsel->attr;
 620         int track = evsel->tracking;
 621         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 622
 623         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 624         attr->inherit       = !opts->no_inherit;
 625
 626         perf_evsel__set_sample_bit(evsel, IP);
 627         perf_evsel__set_sample_bit(evsel, TID);
 628
 629         if (evsel->sample_read) {
 630                 perf_evsel__set_sample_bit(evsel, READ);
 631
 632                 /*
 633                  * We need ID even in case of single event, because
 634                  * PERF_SAMPLE_READ process ID specific data.
 635                  */
 636                 perf_evsel__set_sample_id(evsel, false);
 637
 638                 /*
 639                  * Apply group format only if we belong to group
 640                  * with more than one members.
 641                  */
 642                 if (leader->nr_members > 1) {
 643                         attr->read_format |= PERF_FORMAT_GROUP;
 644                         attr->inherit = 0;
 645                 }
 646         }
 647
 648         /*
 649          * We default some events to have a default interval. But keep
 650          * it a weak assumption overridable by the user.
 651          */
 652         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 653                                      opts->user_interval != ULLONG_MAX)) {
 654                 if (opts->freq) {
 655                         perf_evsel__set_sample_bit(evsel, PERIOD);
 656                         attr->freq              = 1;
 657                         attr->sample_freq       = opts->freq;
 658                 } else {
 659                         attr->sample_period = opts->default_interval;
 660                 }
 661         }
 662
 663         /*
 664          * Disable sampling for all group members other
 665          * than leader in case leader 'leads' the sampling.
 666          */
 667         if ((leader != evsel) && leader->sample_read) {
 668                 attr->sample_freq   = 0;
 669                 attr->sample_period = 0;
 670         }
 671
 672         if (opts->no_samples)
 673                 attr->sample_freq = 0;
 674
 675         if (opts->inherit_stat)
 676                 attr->inherit_stat = 1;
 677
 678         if (opts->sample_address) {
 679                 perf_evsel__set_sample_bit(evsel, ADDR);
 680                 attr->mmap_data = track;
 681         }
 682
 683         /*
 684          * We don't allow user space callchains for  function trace
 685          * event, due to issues with page faults while tracing page
 686          * fault handler and its overall trickiness nature.
 687          */
 688         if (perf_evsel__is_function_event(evsel))
 689                 evsel->attr.exclude_callchain_user = 1;
 690
 691         if (callchain_param.enabled && !evsel->no_aux_samples)
 692                 perf_evsel__config_callgraph(evsel, opts);
 693
 694         if (opts->sample_intr_regs) {
 695                 attr->sample_regs_intr = PERF_REGS_MASK;
 696                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 697         }
 698
 699         if (target__has_cpu(&opts->target))
 700                 perf_evsel__set_sample_bit(evsel, CPU);
 701
 702         if (opts->period)
 703                 perf_evsel__set_sample_bit(evsel, PERIOD);
 704
 705         /*
 706          * When the user explicitely disabled time don't force it here.
 707          */
 708         if (opts->sample_time &&
 709             (!perf_missing_features.sample_id_all &&
 710             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
 711                 perf_evsel__set_sample_bit(evsel, TIME);
 712
 713         if (opts->raw_samples && !evsel->no_aux_samples) {
 714                 perf_evsel__set_sample_bit(evsel, TIME);
 715                 perf_evsel__set_sample_bit(evsel, RAW);
 716                 perf_evsel__set_sample_bit(evsel, CPU);
 717         }
 718
 719         if (opts->sample_address)
 720                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 721
 722         if (opts->no_buffering) {
 723                 attr->watermark = 0;
 724                 attr->wakeup_events = 1;
 725         }
 726         if (opts->branch_stack && !evsel->no_aux_samples) {
 727                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 728                 attr->branch_sample_type = opts->branch_stack;
 729         }
 730
 731         if (opts->sample_weight)
 732                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 733
 734         attr->task  = track;
 735         attr->mmap  = track;
 736         attr->mmap2 = track && !perf_missing_features.mmap2;
 737         attr->comm  = track;
 738
 739         if (opts->sample_transaction)
 740                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 741
 742         if (opts->running_time) {
 743                 evsel->attr.read_format |=
 744                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 745                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 746         }
 747
 748         /*
 749          * XXX see the function comment above
 750          *
 751          * Disabling only independent events or group leaders,
 752          * keeping group members enabled.
 753          */
 754         if (perf_evsel__is_group_leader(evsel))
 755                 attr->disabled = 1;
 756
 757         /*
 758          * Setting enable_on_exec for independent events and
 759          * group leaders for traced executed by perf.
 760          */
 761         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 762                 !opts->initial_delay)
 763                 attr->enable_on_exec = 1;
 764
 765         if (evsel->immediate) {
 766                 attr->disabled = 0;
 767                 attr->enable_on_exec = 0;
 768         }
 769
 770         clockid = opts->clockid;
 771         if (opts->use_clockid) {
 772                 attr->use_clockid = 1;
 773                 attr->clockid = opts->clockid;
 774         }
 775 }
 776
 777 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 778 {
 779         int cpu, thread;
 780
 781         if (evsel->system_wide)
 782                 nthreads = 1;
 783
 784         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 785
 786         if (evsel->fd) {
 787                 for (cpu = 0; cpu < ncpus; cpu++) {
 788                         for (thread = 0; thread < nthreads; thread++) {
 789                                 FD(evsel, cpu, thread) = -1;
 790                         }
 791                 }
 792         }
 793
 794         return evsel->fd != NULL ? 0 : -ENOMEM;
 795 }
 796
 797 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 798                           int ioc,  void *arg)
 799 {
 800         int cpu, thread;
 801
 802         if (evsel->system_wide)
 803                 nthreads = 1;
 804
 805         for (cpu = 0; cpu < ncpus; cpu++) {
 806                 for (thread = 0; thread < nthreads; thread++) {
 807                         int fd = FD(evsel, cpu, thread),
 808                             err = ioctl(fd, ioc, arg);
 809
 810                         if (err)
 811                                 return err;
 812                 }
 813         }
 814
 815         return 0;
 816 }
 817
 818 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 819                            const char *filter)
 820 {
 821         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 822                                      PERF_EVENT_IOC_SET_FILTER,
 823                                      (void *)filter);
 824 }
 825
 826 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 827 {
 828         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 829                                      PERF_EVENT_IOC_ENABLE,
 830                                      0);
 831 }
 832
 833 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 834 {
 835         if (ncpus == 0 || nthreads == 0)
 836                 return 0;
 837
 838         if (evsel->system_wide)
 839                 nthreads = 1;
 840
 841         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 842         if (evsel->sample_id == NULL)
 843                 return -ENOMEM;
 844
 845         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 846         if (evsel->id == NULL) {
 847                 xyarray__delete(evsel->sample_id);
 848                 evsel->sample_id = NULL;
 849                 return -ENOMEM;
 850         }
 851
 852         return 0;
 853 }
 854
 855 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 856 {
 857         xyarray__delete(evsel->fd);
 858         evsel->fd = NULL;
 859 }
 860
 861 static void perf_evsel__free_id(struct perf_evsel *evsel)
 862 {
 863         xyarray__delete(evsel->sample_id);
 864         evsel->sample_id = NULL;
 865         zfree(&evsel->id);
 866 }
 867
 868 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 869 {
 870         int cpu, thread;
 871
 872         if (evsel->system_wide)
 873                 nthreads = 1;
 874
 875         for (cpu = 0; cpu < ncpus; cpu++)
 876                 for (thread = 0; thread < nthreads; ++thread) {
 877                         close(FD(evsel, cpu, thread));
 878                         FD(evsel, cpu, thread) = -1;
 879                 }
 880 }
 881
 882 void perf_evsel__exit(struct perf_evsel *evsel)
 883 {
 884         assert(list_empty(&evsel->node));
 885         perf_evsel__free_fd(evsel);
 886         perf_evsel__free_id(evsel);
 887         close_cgroup(evsel->cgrp);
 888         zfree(&evsel->group_name);
 889         zfree(&evsel->name);
 890         perf_evsel__object.fini(evsel);
 891 }
 892
 893 void perf_evsel__delete(struct perf_evsel *evsel)
 894 {
 895         perf_evsel__exit(evsel);
 896         free(evsel);
 897 }
 898
 899 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu,
 900                                 struct perf_counts_values *count)
 901 {
 902         struct perf_counts_values tmp;
 903
 904         if (!evsel->prev_raw_counts)
 905                 return;
 906
 907         if (cpu == -1) {
 908                 tmp = evsel->prev_raw_counts->aggr;
 909                 evsel->prev_raw_counts->aggr = *count;
 910         } else {
 911                 tmp = evsel->prev_raw_counts->cpu[cpu];
 912                 evsel->prev_raw_counts->cpu[cpu] = *count;
 913         }
 914
 915         count->val = count->val - tmp.val;
 916         count->ena = count->ena - tmp.ena;
 917         count->run = count->run - tmp.run;
 918 }
 919
 920 void perf_counts_values__scale(struct perf_counts_values *count,
 921                                bool scale, s8 *pscaled)
 922 {
 923         s8 scaled = 0;
 924
 925         if (scale) {
 926                 if (count->run == 0) {
 927                         scaled = -1;
 928                         count->val = 0;
 929                 } else if (count->run < count->ena) {
 930                         scaled = 1;
 931                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
 932                 }
 933         } else
 934                 count->ena = count->run = 0;
 935
 936         if (pscaled)
 937                 *pscaled = scaled;
 938 }
 939
 940 int perf_evsel__read_cb(struct perf_evsel *evsel, int cpu, int thread,
 941                         perf_evsel__read_cb_t cb)
 942 {
 943         struct perf_counts_values count;
 944
 945         memset(&count, 0, sizeof(count));
 946
 947         if (FD(evsel, cpu, thread) < 0)
 948                 return -EINVAL;
 949
 950         if (readn(FD(evsel, cpu, thread), &count, sizeof(count)) < 0)
 951                 return -errno;
 952
 953         return cb(evsel, cpu, thread, &count);
 954 }
 955
 956 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 957                               int cpu, int thread, bool scale)
 958 {
 959         struct perf_counts_values count;
 960         size_t nv = scale ? 3 : 1;
 961
 962         if (FD(evsel, cpu, thread) < 0)
 963                 return -EINVAL;
 964
 965         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 966                 return -ENOMEM;
 967
 968         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 969                 return -errno;
 970
 971         perf_evsel__compute_deltas(evsel, cpu, &count);
 972         perf_counts_values__scale(&count, scale, NULL);
 973         evsel->counts->cpu[cpu] = count;
 974         return 0;
 975 }
 976
 977 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 978 {
 979         struct perf_evsel *leader = evsel->leader;
 980         int fd;
 981
 982         if (perf_evsel__is_group_leader(evsel))
 983                 return -1;
 984
 985         /*
 986          * Leader must be already processed/open,
 987          * if not it's a bug.
 988          */
 989         BUG_ON(!leader->fd);
 990
 991         fd = FD(leader, cpu, thread);
 992         BUG_ON(fd == -1);
 993
 994         return fd;
 995 }
 996
 997 struct bit_names {
 998         int bit;
 999         const char *name;
1000 };
1001
1002 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1003 {
1004         bool first_bit = true;
1005         int i = 0;
1006
1007         do {
1008                 if (value & bits[i].bit) {
1009                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1010                         first_bit = false;
1011                 }
1012         } while (bits[++i].name != NULL);
1013 }
1014
1015 static void __p_sample_type(char *buf, size_t size, u64 value)
1016 {
1017 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1018         struct bit_names bits[] = {
1019                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1020                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1021                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1022                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1023                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1024                 { .name = NULL, }
1025         };
1026 #undef bit_name
1027         __p_bits(buf, size, value, bits);
1028 }
1029
1030 static void __p_read_format(char *buf, size_t size, u64 value)
1031 {
1032 #define bit_name(n) { PERF_FORMAT_##n, #n }
1033         struct bit_names bits[] = {
1034                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1035                 bit_name(ID), bit_name(GROUP),
1036                 { .name = NULL, }
1037         };
1038 #undef bit_name
1039         __p_bits(buf, size, value, bits);
1040 }
1041
1042 #define BUF_SIZE                1024
1043
1044 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1045 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1046 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1047 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1048 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1049
1050 #define PRINT_ATTRn(_n, _f, _p)                         \
1051 do {                                                    \
1052         if (attr->_f) {                                 \
1053                 _p(attr->_f);                           \
1054                 ret += attr__fprintf(fp, _n, buf, priv);\
1055         }                                               \
1056 } while (0)
1057
1058 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1059
1060 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1061                              attr__fprintf_f attr__fprintf, void *priv)
1062 {
1063         char buf[BUF_SIZE];
1064         int ret = 0;
1065
1066         PRINT_ATTRf(type, p_unsigned);
1067         PRINT_ATTRf(size, p_unsigned);
1068         PRINT_ATTRf(config, p_hex);
1069         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1070         PRINT_ATTRf(sample_type, p_sample_type);
1071         PRINT_ATTRf(read_format, p_read_format);
1072
1073         PRINT_ATTRf(disabled, p_unsigned);
1074         PRINT_ATTRf(inherit, p_unsigned);
1075         PRINT_ATTRf(pinned, p_unsigned);
1076         PRINT_ATTRf(exclusive, p_unsigned);
1077         PRINT_ATTRf(exclude_user, p_unsigned);
1078         PRINT_ATTRf(exclude_kernel, p_unsigned);
1079         PRINT_ATTRf(exclude_hv, p_unsigned);
1080         PRINT_ATTRf(exclude_idle, p_unsigned);
1081         PRINT_ATTRf(mmap, p_unsigned);
1082         PRINT_ATTRf(comm, p_unsigned);
1083         PRINT_ATTRf(freq, p_unsigned);
1084         PRINT_ATTRf(inherit_stat, p_unsigned);
1085         PRINT_ATTRf(enable_on_exec, p_unsigned);
1086         PRINT_ATTRf(task, p_unsigned);
1087         PRINT_ATTRf(watermark, p_unsigned);
1088         PRINT_ATTRf(precise_ip, p_unsigned);
1089         PRINT_ATTRf(mmap_data, p_unsigned);
1090         PRINT_ATTRf(sample_id_all, p_unsigned);
1091         PRINT_ATTRf(exclude_host, p_unsigned);
1092         PRINT_ATTRf(exclude_guest, p_unsigned);
1093         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1094         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1095         PRINT_ATTRf(mmap2, p_unsigned);
1096         PRINT_ATTRf(comm_exec, p_unsigned);
1097         PRINT_ATTRf(use_clockid, p_unsigned);
1098
1099         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1100         PRINT_ATTRf(bp_type, p_unsigned);
1101         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1102         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1103         PRINT_ATTRf(sample_regs_user, p_hex);
1104         PRINT_ATTRf(sample_stack_user, p_unsigned);
1105         PRINT_ATTRf(clockid, p_signed);
1106         PRINT_ATTRf(sample_regs_intr, p_hex);
1107         PRINT_ATTRf(aux_watermark, p_unsigned);
1108
1109         return ret;
1110 }
1111
1112 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1113                                 void *priv __attribute__((unused)))
1114 {
1115         return fprintf(fp, "  %-32s %s\n", name, val);
1116 }
1117
1118 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1119                               struct thread_map *threads)
1120 {
1121         int cpu, thread, nthreads;
1122         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1123         int pid = -1, err;
1124         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1125
1126         if (evsel->system_wide)
1127                 nthreads = 1;
1128         else
1129                 nthreads = threads->nr;
1130
1131         if (evsel->fd == NULL &&
1132             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1133                 return -ENOMEM;
1134
1135         if (evsel->cgrp) {
1136                 flags |= PERF_FLAG_PID_CGROUP;
1137                 pid = evsel->cgrp->fd;
1138         }
1139
1140 fallback_missing_features:
1141         if (perf_missing_features.clockid_wrong)
1142                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1143         if (perf_missing_features.clockid) {
1144                 evsel->attr.use_clockid = 0;
1145                 evsel->attr.clockid = 0;
1146         }
1147         if (perf_missing_features.cloexec)
1148                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1149         if (perf_missing_features.mmap2)
1150                 evsel->attr.mmap2 = 0;
1151         if (perf_missing_features.exclude_guest)
1152                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1153 retry_sample_id:
1154         if (perf_missing_features.sample_id_all)
1155                 evsel->attr.sample_id_all = 0;
1156
1157         if (verbose >= 2) {
1158                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1159                 fprintf(stderr, "perf_event_attr:\n");
1160                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1161                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1162         }
1163
1164         for (cpu = 0; cpu < cpus->nr; cpu++) {
1165
1166                 for (thread = 0; thread < nthreads; thread++) {
1167                         int group_fd;
1168
1169                         if (!evsel->cgrp && !evsel->system_wide)
1170                                 pid = threads->map[thread];
1171
1172                         group_fd = get_group_fd(evsel, cpu, thread);
1173 retry_open:
1174                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1175                                   pid, cpus->map[cpu], group_fd, flags);
1176
1177                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1178                                                                      pid,
1179                                                                      cpus->map[cpu],
1180                                                                      group_fd, flags);
1181                         if (FD(evsel, cpu, thread) < 0) {
1182                                 err = -errno;
1183                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1184                                           err);
1185                                 goto try_fallback;
1186                         }
1187                         set_rlimit = NO_CHANGE;
1188
1189                         /*
1190                          * If we succeeded but had to kill clockid, fail and
1191                          * have perf_evsel__open_strerror() print us a nice
1192                          * error.
1193                          */
1194                         if (perf_missing_features.clockid ||
1195                             perf_missing_features.clockid_wrong) {
1196                                 err = -EINVAL;
1197                                 goto out_close;
1198                         }
1199                 }
1200         }
1201
1202         return 0;
1203
1204 try_fallback:
1205         /*
1206          * perf stat needs between 5 and 22 fds per CPU. When we run out
1207          * of them try to increase the limits.
1208          */
1209         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1210                 struct rlimit l;
1211                 int old_errno = errno;
1212
1213                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1214                         if (set_rlimit == NO_CHANGE)
1215                                 l.rlim_cur = l.rlim_max;
1216                         else {
1217                                 l.rlim_cur = l.rlim_max + 1000;
1218                                 l.rlim_max = l.rlim_cur;
1219                         }
1220                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1221                                 set_rlimit++;
1222                                 errno = old_errno;
1223                                 goto retry_open;
1224                         }
1225                 }
1226                 errno = old_errno;
1227         }
1228
1229         if (err != -EINVAL || cpu > 0 || thread > 0)
1230                 goto out_close;
1231
1232         /*
1233          * Must probe features in the order they were added to the
1234          * perf_event_attr interface.
1235          */
1236         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1237                 perf_missing_features.clockid_wrong = true;
1238                 goto fallback_missing_features;
1239         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1240                 perf_missing_features.clockid = true;
1241                 goto fallback_missing_features;
1242         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1243                 perf_missing_features.cloexec = true;
1244                 goto fallback_missing_features;
1245         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1246                 perf_missing_features.mmap2 = true;
1247                 goto fallback_missing_features;
1248         } else if (!perf_missing_features.exclude_guest &&
1249                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1250                 perf_missing_features.exclude_guest = true;
1251                 goto fallback_missing_features;
1252         } else if (!perf_missing_features.sample_id_all) {
1253                 perf_missing_features.sample_id_all = true;
1254                 goto retry_sample_id;
1255         }
1256
1257 out_close:
1258         do {
1259                 while (--thread >= 0) {
1260                         close(FD(evsel, cpu, thread));
1261                         FD(evsel, cpu, thread) = -1;
1262                 }
1263                 thread = nthreads;
1264         } while (--cpu >= 0);
1265         return err;
1266 }
1267
1268 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1269 {
1270         if (evsel->fd == NULL)
1271                 return;
1272
1273         perf_evsel__close_fd(evsel, ncpus, nthreads);
1274         perf_evsel__free_fd(evsel);
1275 }
1276
1277 static struct {
1278         struct cpu_map map;
1279         int cpus[1];
1280 } empty_cpu_map = {
1281         .map.nr = 1,
1282         .cpus   = { -1, },
1283 };
1284
1285 static struct {
1286         struct thread_map map;
1287         int threads[1];
1288 } empty_thread_map = {
1289         .map.nr  = 1,
1290         .threads = { -1, },
1291 };
1292
1293 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1294                      struct thread_map *threads)
1295 {
1296         if (cpus == NULL) {
1297                 /* Work around old compiler warnings about strict aliasing */
1298                 cpus = &empty_cpu_map.map;
1299         }
1300
1301         if (threads == NULL)
1302                 threads = &empty_thread_map.map;
1303
1304         return __perf_evsel__open(evsel, cpus, threads);
1305 }
1306
1307 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1308                              struct cpu_map *cpus)
1309 {
1310         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1311 }
1312
1313 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1314                                 struct thread_map *threads)
1315 {
1316         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1317 }
1318
1319 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1320                                        const union perf_event *event,
1321                                        struct perf_sample *sample)
1322 {
1323         u64 type = evsel->attr.sample_type;
1324         const u64 *array = event->sample.array;
1325         bool swapped = evsel->needs_swap;
1326         union u64_swap u;
1327
1328         array += ((event->header.size -
1329                    sizeof(event->header)) / sizeof(u64)) - 1;
1330
1331         if (type & PERF_SAMPLE_IDENTIFIER) {
1332                 sample->id = *array;
1333                 array--;
1334         }
1335
1336         if (type & PERF_SAMPLE_CPU) {
1337                 u.val64 = *array;
1338                 if (swapped) {
1339                         /* undo swap of u64, then swap on individual u32s */
1340                         u.val64 = bswap_64(u.val64);
1341                         u.val32[0] = bswap_32(u.val32[0]);
1342                 }
1343
1344                 sample->cpu = u.val32[0];
1345                 array--;
1346         }
1347
1348         if (type & PERF_SAMPLE_STREAM_ID) {
1349                 sample->stream_id = *array;
1350                 array--;
1351         }
1352
1353         if (type & PERF_SAMPLE_ID) {
1354                 sample->id = *array;
1355                 array--;
1356         }
1357
1358         if (type & PERF_SAMPLE_TIME) {
1359                 sample->time = *array;
1360                 array--;
1361         }
1362
1363         if (type & PERF_SAMPLE_TID) {
1364                 u.val64 = *array;
1365                 if (swapped) {
1366                         /* undo swap of u64, then swap on individual u32s */
1367                         u.val64 = bswap_64(u.val64);
1368                         u.val32[0] = bswap_32(u.val32[0]);
1369                         u.val32[1] = bswap_32(u.val32[1]);
1370                 }
1371
1372                 sample->pid = u.val32[0];
1373                 sample->tid = u.val32[1];
1374                 array--;
1375         }
1376
1377         return 0;
1378 }
1379
1380 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1381                             u64 size)
1382 {
1383         return size > max_size || offset + size > endp;
1384 }
1385
1386 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1387         do {                                                            \
1388                 if (overflow(endp, (max_size), (offset), (size)))       \
1389                         return -EFAULT;                                 \
1390         } while (0)
1391
1392 #define OVERFLOW_CHECK_u64(offset) \
1393         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1394
1395 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1396                              struct perf_sample *data)
1397 {
1398         u64 type = evsel->attr.sample_type;
1399         bool swapped = evsel->needs_swap;
1400         const u64 *array;
1401         u16 max_size = event->header.size;
1402         const void *endp = (void *)event + max_size;
1403         u64 sz;
1404
1405         /*
1406          * used for cross-endian analysis. See git commit 65014ab3
1407          * for why this goofiness is needed.
1408          */
1409         union u64_swap u;
1410
1411         memset(data, 0, sizeof(*data));
1412         data->cpu = data->pid = data->tid = -1;
1413         data->stream_id = data->id = data->time = -1ULL;
1414         data->period = evsel->attr.sample_period;
1415         data->weight = 0;
1416
1417         if (event->header.type != PERF_RECORD_SAMPLE) {
1418                 if (!evsel->attr.sample_id_all)
1419                         return 0;
1420                 return perf_evsel__parse_id_sample(evsel, event, data);
1421         }
1422
1423         array = event->sample.array;
1424
1425         /*
1426          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1427          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1428          * check the format does not go past the end of the event.
1429          */
1430         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1431                 return -EFAULT;
1432
1433         data->id = -1ULL;
1434         if (type & PERF_SAMPLE_IDENTIFIER) {
1435                 data->id = *array;
1436                 array++;
1437         }
1438
1439         if (type & PERF_SAMPLE_IP) {
1440                 data->ip = *array;
1441                 array++;
1442         }
1443
1444         if (type & PERF_SAMPLE_TID) {
1445                 u.val64 = *array;
1446                 if (swapped) {
1447                         /* undo swap of u64, then swap on individual u32s */
1448                         u.val64 = bswap_64(u.val64);
1449                         u.val32[0] = bswap_32(u.val32[0]);
1450                         u.val32[1] = bswap_32(u.val32[1]);
1451                 }
1452
1453                 data->pid = u.val32[0];
1454                 data->tid = u.val32[1];
1455                 array++;
1456         }
1457
1458         if (type & PERF_SAMPLE_TIME) {
1459                 data->time = *array;
1460                 array++;
1461         }
1462
1463         data->addr = 0;
1464         if (type & PERF_SAMPLE_ADDR) {
1465                 data->addr = *array;
1466                 array++;
1467         }
1468
1469         if (type & PERF_SAMPLE_ID) {
1470                 data->id = *array;
1471                 array++;
1472         }
1473
1474         if (type & PERF_SAMPLE_STREAM_ID) {
1475                 data->stream_id = *array;
1476                 array++;
1477         }
1478
1479         if (type & PERF_SAMPLE_CPU) {
1480
1481                 u.val64 = *array;
1482                 if (swapped) {
1483                         /* undo swap of u64, then swap on individual u32s */
1484                         u.val64 = bswap_64(u.val64);
1485                         u.val32[0] = bswap_32(u.val32[0]);
1486                 }
1487
1488                 data->cpu = u.val32[0];
1489                 array++;
1490         }
1491
1492         if (type & PERF_SAMPLE_PERIOD) {
1493                 data->period = *array;
1494                 array++;
1495         }
1496
1497         if (type & PERF_SAMPLE_READ) {
1498                 u64 read_format = evsel->attr.read_format;
1499
1500                 OVERFLOW_CHECK_u64(array);
1501                 if (read_format & PERF_FORMAT_GROUP)
1502                         data->read.group.nr = *array;
1503                 else
1504                         data->read.one.value = *array;
1505
1506                 array++;
1507
1508                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1509                         OVERFLOW_CHECK_u64(array);
1510                         data->read.time_enabled = *array;
1511                         array++;
1512                 }
1513
1514                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1515                         OVERFLOW_CHECK_u64(array);
1516                         data->read.time_running = *array;
1517                         array++;
1518                 }
1519
1520                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1521                 if (read_format & PERF_FORMAT_GROUP) {
1522                         const u64 max_group_nr = UINT64_MAX /
1523                                         sizeof(struct sample_read_value);
1524
1525                         if (data->read.group.nr > max_group_nr)
1526                                 return -EFAULT;
1527                         sz = data->read.group.nr *
1528                              sizeof(struct sample_read_value);
1529                         OVERFLOW_CHECK(array, sz, max_size);
1530                         data->read.group.values =
1531                                         (struct sample_read_value *)array;
1532                         array = (void *)array + sz;
1533                 } else {
1534                         OVERFLOW_CHECK_u64(array);
1535                         data->read.one.id = *array;
1536                         array++;
1537                 }
1538         }
1539
1540         if (type & PERF_SAMPLE_CALLCHAIN) {
1541                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1542
1543                 OVERFLOW_CHECK_u64(array);
1544                 data->callchain = (struct ip_callchain *)array++;
1545                 if (data->callchain->nr > max_callchain_nr)
1546                         return -EFAULT;
1547                 sz = data->callchain->nr * sizeof(u64);
1548                 OVERFLOW_CHECK(array, sz, max_size);
1549                 array = (void *)array + sz;
1550         }
1551
1552         if (type & PERF_SAMPLE_RAW) {
1553                 OVERFLOW_CHECK_u64(array);
1554                 u.val64 = *array;
1555                 if (WARN_ONCE(swapped,
1556                               "Endianness of raw data not corrected!\n")) {
1557                         /* undo swap of u64, then swap on individual u32s */
1558                         u.val64 = bswap_64(u.val64);
1559                         u.val32[0] = bswap_32(u.val32[0]);
1560                         u.val32[1] = bswap_32(u.val32[1]);
1561                 }
1562                 data->raw_size = u.val32[0];
1563                 array = (void *)array + sizeof(u32);
1564
1565                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1566                 data->raw_data = (void *)array;
1567                 array = (void *)array + data->raw_size;
1568         }
1569
1570         if (type & PERF_SAMPLE_BRANCH_STACK) {
1571                 const u64 max_branch_nr = UINT64_MAX /
1572                                           sizeof(struct branch_entry);
1573
1574                 OVERFLOW_CHECK_u64(array);
1575                 data->branch_stack = (struct branch_stack *)array++;
1576
1577                 if (data->branch_stack->nr > max_branch_nr)
1578                         return -EFAULT;
1579                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1580                 OVERFLOW_CHECK(array, sz, max_size);
1581                 array = (void *)array + sz;
1582         }
1583
1584         if (type & PERF_SAMPLE_REGS_USER) {
1585                 OVERFLOW_CHECK_u64(array);
1586                 data->user_regs.abi = *array;
1587                 array++;
1588
1589                 if (data->user_regs.abi) {
1590                         u64 mask = evsel->attr.sample_regs_user;
1591
1592                         sz = hweight_long(mask) * sizeof(u64);
1593                         OVERFLOW_CHECK(array, sz, max_size);
1594                         data->user_regs.mask = mask;
1595                         data->user_regs.regs = (u64 *)array;
1596                         array = (void *)array + sz;
1597                 }
1598         }
1599
1600         if (type & PERF_SAMPLE_STACK_USER) {
1601                 OVERFLOW_CHECK_u64(array);
1602                 sz = *array++;
1603
1604                 data->user_stack.offset = ((char *)(array - 1)
1605                                           - (char *) event);
1606
1607                 if (!sz) {
1608                         data->user_stack.size = 0;
1609                 } else {
1610                         OVERFLOW_CHECK(array, sz, max_size);
1611                         data->user_stack.data = (char *)array;
1612                         array = (void *)array + sz;
1613                         OVERFLOW_CHECK_u64(array);
1614                         data->user_stack.size = *array++;
1615                         if (WARN_ONCE(data->user_stack.size > sz,
1616                                       "user stack dump failure\n"))
1617                                 return -EFAULT;
1618                 }
1619         }
1620
1621         data->weight = 0;
1622         if (type & PERF_SAMPLE_WEIGHT) {
1623                 OVERFLOW_CHECK_u64(array);
1624                 data->weight = *array;
1625                 array++;
1626         }
1627
1628         data->data_src = PERF_MEM_DATA_SRC_NONE;
1629         if (type & PERF_SAMPLE_DATA_SRC) {
1630                 OVERFLOW_CHECK_u64(array);
1631                 data->data_src = *array;
1632                 array++;
1633         }
1634
1635         data->transaction = 0;
1636         if (type & PERF_SAMPLE_TRANSACTION) {
1637                 OVERFLOW_CHECK_u64(array);
1638                 data->transaction = *array;
1639                 array++;
1640         }
1641
1642         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1643         if (type & PERF_SAMPLE_REGS_INTR) {
1644                 OVERFLOW_CHECK_u64(array);
1645                 data->intr_regs.abi = *array;
1646                 array++;
1647
1648                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1649                         u64 mask = evsel->attr.sample_regs_intr;
1650
1651                         sz = hweight_long(mask) * sizeof(u64);
1652                         OVERFLOW_CHECK(array, sz, max_size);
1653                         data->intr_regs.mask = mask;
1654                         data->intr_regs.regs = (u64 *)array;
1655                         array = (void *)array + sz;
1656                 }
1657         }
1658
1659         return 0;
1660 }
1661
1662 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1663                                      u64 read_format)
1664 {
1665         size_t sz, result = sizeof(struct sample_event);
1666
1667         if (type & PERF_SAMPLE_IDENTIFIER)
1668                 result += sizeof(u64);
1669
1670         if (type & PERF_SAMPLE_IP)
1671                 result += sizeof(u64);
1672
1673         if (type & PERF_SAMPLE_TID)
1674                 result += sizeof(u64);
1675
1676         if (type & PERF_SAMPLE_TIME)
1677                 result += sizeof(u64);
1678
1679         if (type & PERF_SAMPLE_ADDR)
1680                 result += sizeof(u64);
1681
1682         if (type & PERF_SAMPLE_ID)
1683                 result += sizeof(u64);
1684
1685         if (type & PERF_SAMPLE_STREAM_ID)
1686                 result += sizeof(u64);
1687
1688         if (type & PERF_SAMPLE_CPU)
1689                 result += sizeof(u64);
1690
1691         if (type & PERF_SAMPLE_PERIOD)
1692                 result += sizeof(u64);
1693
1694         if (type & PERF_SAMPLE_READ) {
1695                 result += sizeof(u64);
1696                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1697                         result += sizeof(u64);
1698                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1699                         result += sizeof(u64);
1700                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1701                 if (read_format & PERF_FORMAT_GROUP) {
1702                         sz = sample->read.group.nr *
1703                              sizeof(struct sample_read_value);
1704                         result += sz;
1705                 } else {
1706                         result += sizeof(u64);
1707                 }
1708         }
1709
1710         if (type & PERF_SAMPLE_CALLCHAIN) {
1711                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1712                 result += sz;
1713         }
1714
1715         if (type & PERF_SAMPLE_RAW) {
1716                 result += sizeof(u32);
1717                 result += sample->raw_size;
1718         }
1719
1720         if (type & PERF_SAMPLE_BRANCH_STACK) {
1721                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1722                 sz += sizeof(u64);
1723                 result += sz;
1724         }
1725
1726         if (type & PERF_SAMPLE_REGS_USER) {
1727                 if (sample->user_regs.abi) {
1728                         result += sizeof(u64);
1729                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1730                         result += sz;
1731                 } else {
1732                         result += sizeof(u64);
1733                 }
1734         }
1735
1736         if (type & PERF_SAMPLE_STACK_USER) {
1737                 sz = sample->user_stack.size;
1738                 result += sizeof(u64);
1739                 if (sz) {
1740                         result += sz;
1741                         result += sizeof(u64);
1742                 }
1743         }
1744
1745         if (type & PERF_SAMPLE_WEIGHT)
1746                 result += sizeof(u64);
1747
1748         if (type & PERF_SAMPLE_DATA_SRC)
1749                 result += sizeof(u64);
1750
1751         if (type & PERF_SAMPLE_TRANSACTION)
1752                 result += sizeof(u64);
1753
1754         if (type & PERF_SAMPLE_REGS_INTR) {
1755                 if (sample->intr_regs.abi) {
1756                         result += sizeof(u64);
1757                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1758                         result += sz;
1759                 } else {
1760                         result += sizeof(u64);
1761                 }
1762         }
1763
1764         return result;
1765 }
1766
1767 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1768                                   u64 read_format,
1769                                   const struct perf_sample *sample,
1770                                   bool swapped)
1771 {
1772         u64 *array;
1773         size_t sz;
1774         /*
1775          * used for cross-endian analysis. See git commit 65014ab3
1776          * for why this goofiness is needed.
1777          */
1778         union u64_swap u;
1779
1780         array = event->sample.array;
1781
1782         if (type & PERF_SAMPLE_IDENTIFIER) {
1783                 *array = sample->id;
1784                 array++;
1785         }
1786
1787         if (type & PERF_SAMPLE_IP) {
1788                 *array = sample->ip;
1789                 array++;
1790         }
1791
1792         if (type & PERF_SAMPLE_TID) {
1793                 u.val32[0] = sample->pid;
1794                 u.val32[1] = sample->tid;
1795                 if (swapped) {
1796                         /*
1797                          * Inverse of what is done in perf_evsel__parse_sample
1798                          */
1799                         u.val32[0] = bswap_32(u.val32[0]);
1800                         u.val32[1] = bswap_32(u.val32[1]);
1801                         u.val64 = bswap_64(u.val64);
1802                 }
1803
1804                 *array = u.val64;
1805                 array++;
1806         }
1807
1808         if (type & PERF_SAMPLE_TIME) {
1809                 *array = sample->time;
1810                 array++;
1811         }
1812
1813         if (type & PERF_SAMPLE_ADDR) {
1814                 *array = sample->addr;
1815                 array++;
1816         }
1817
1818         if (type & PERF_SAMPLE_ID) {
1819                 *array = sample->id;
1820                 array++;
1821         }
1822
1823         if (type & PERF_SAMPLE_STREAM_ID) {
1824                 *array = sample->stream_id;
1825                 array++;
1826         }
1827
1828         if (type & PERF_SAMPLE_CPU) {
1829                 u.val32[0] = sample->cpu;
1830                 if (swapped) {
1831                         /*
1832                          * Inverse of what is done in perf_evsel__parse_sample
1833                          */
1834                         u.val32[0] = bswap_32(u.val32[0]);
1835                         u.val64 = bswap_64(u.val64);
1836                 }
1837                 *array = u.val64;
1838                 array++;
1839         }
1840
1841         if (type & PERF_SAMPLE_PERIOD) {
1842                 *array = sample->period;
1843                 array++;
1844         }
1845
1846         if (type & PERF_SAMPLE_READ) {
1847                 if (read_format & PERF_FORMAT_GROUP)
1848                         *array = sample->read.group.nr;
1849                 else
1850                         *array = sample->read.one.value;
1851                 array++;
1852
1853                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1854                         *array = sample->read.time_enabled;
1855                         array++;
1856                 }
1857
1858                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1859                         *array = sample->read.time_running;
1860                         array++;
1861                 }
1862
1863                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1864                 if (read_format & PERF_FORMAT_GROUP) {
1865                         sz = sample->read.group.nr *
1866                              sizeof(struct sample_read_value);
1867                         memcpy(array, sample->read.group.values, sz);
1868                         array = (void *)array + sz;
1869                 } else {
1870                         *array = sample->read.one.id;
1871                         array++;
1872                 }
1873         }
1874
1875         if (type & PERF_SAMPLE_CALLCHAIN) {
1876                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1877                 memcpy(array, sample->callchain, sz);
1878                 array = (void *)array + sz;
1879         }
1880
1881         if (type & PERF_SAMPLE_RAW) {
1882                 u.val32[0] = sample->raw_size;
1883                 if (WARN_ONCE(swapped,
1884                               "Endianness of raw data not corrected!\n")) {
1885                         /*
1886                          * Inverse of what is done in perf_evsel__parse_sample
1887                          */
1888                         u.val32[0] = bswap_32(u.val32[0]);
1889                         u.val32[1] = bswap_32(u.val32[1]);
1890                         u.val64 = bswap_64(u.val64);
1891                 }
1892                 *array = u.val64;
1893                 array = (void *)array + sizeof(u32);
1894
1895                 memcpy(array, sample->raw_data, sample->raw_size);
1896                 array = (void *)array + sample->raw_size;
1897         }
1898
1899         if (type & PERF_SAMPLE_BRANCH_STACK) {
1900                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1901                 sz += sizeof(u64);
1902                 memcpy(array, sample->branch_stack, sz);
1903                 array = (void *)array + sz;
1904         }
1905
1906         if (type & PERF_SAMPLE_REGS_USER) {
1907                 if (sample->user_regs.abi) {
1908                         *array++ = sample->user_regs.abi;
1909                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1910                         memcpy(array, sample->user_regs.regs, sz);
1911                         array = (void *)array + sz;
1912                 } else {
1913                         *array++ = 0;
1914                 }
1915         }
1916
1917         if (type & PERF_SAMPLE_STACK_USER) {
1918                 sz = sample->user_stack.size;
1919                 *array++ = sz;
1920                 if (sz) {
1921                         memcpy(array, sample->user_stack.data, sz);
1922                         array = (void *)array + sz;
1923                         *array++ = sz;
1924                 }
1925         }
1926
1927         if (type & PERF_SAMPLE_WEIGHT) {
1928                 *array = sample->weight;
1929                 array++;
1930         }
1931
1932         if (type & PERF_SAMPLE_DATA_SRC) {
1933                 *array = sample->data_src;
1934                 array++;
1935         }
1936
1937         if (type & PERF_SAMPLE_TRANSACTION) {
1938                 *array = sample->transaction;
1939                 array++;
1940         }
1941
1942         if (type & PERF_SAMPLE_REGS_INTR) {
1943                 if (sample->intr_regs.abi) {
1944                         *array++ = sample->intr_regs.abi;
1945                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1946                         memcpy(array, sample->intr_regs.regs, sz);
1947                         array = (void *)array + sz;
1948                 } else {
1949                         *array++ = 0;
1950                 }
1951         }
1952
1953         return 0;
1954 }
1955
1956 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1957 {
1958         return pevent_find_field(evsel->tp_format, name);
1959 }
1960
1961 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1962                          const char *name)
1963 {
1964         struct format_field *field = perf_evsel__field(evsel, name);
1965         int offset;
1966
1967         if (!field)
1968                 return NULL;
1969
1970         offset = field->offset;
1971
1972         if (field->flags & FIELD_IS_DYNAMIC) {
1973                 offset = *(int *)(sample->raw_data + field->offset);
1974                 offset &= 0xffff;
1975         }
1976
1977         return sample->raw_data + offset;
1978 }
1979
1980 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1981                        const char *name)
1982 {
1983         struct format_field *field = perf_evsel__field(evsel, name);
1984         void *ptr;
1985         u64 value;
1986
1987         if (!field)
1988                 return 0;
1989
1990         ptr = sample->raw_data + field->offset;
1991
1992         switch (field->size) {
1993         case 1:
1994                 return *(u8 *)ptr;
1995         case 2:
1996                 value = *(u16 *)ptr;
1997                 break;
1998         case 4:
1999                 value = *(u32 *)ptr;
2000                 break;
2001         case 8:
2002                 memcpy(&value, ptr, sizeof(u64));
2003                 break;
2004         default:
2005                 return 0;
2006         }
2007
2008         if (!evsel->needs_swap)
2009                 return value;
2010
2011         switch (field->size) {
2012         case 2:
2013                 return bswap_16(value);
2014         case 4:
2015                 return bswap_32(value);
2016         case 8:
2017                 return bswap_64(value);
2018         default:
2019                 return 0;
2020         }
2021
2022         return 0;
2023 }
2024
2025 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2026 {
2027         va_list args;
2028         int ret = 0;
2029
2030         if (!*first) {
2031                 ret += fprintf(fp, ",");
2032         } else {
2033                 ret += fprintf(fp, ":");
2034                 *first = false;
2035         }
2036
2037         va_start(args, fmt);
2038         ret += vfprintf(fp, fmt, args);
2039         va_end(args);
2040         return ret;
2041 }
2042
2043 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2044 {
2045         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2046 }
2047
2048 int perf_evsel__fprintf(struct perf_evsel *evsel,
2049                         struct perf_attr_details *details, FILE *fp)
2050 {
2051         bool first = true;
2052         int printed = 0;
2053
2054         if (details->event_group) {
2055                 struct perf_evsel *pos;
2056
2057                 if (!perf_evsel__is_group_leader(evsel))
2058                         return 0;
2059
2060                 if (evsel->nr_members > 1)
2061                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2062
2063                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2064                 for_each_group_member(pos, evsel)
2065                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2066
2067                 if (evsel->nr_members > 1)
2068                         printed += fprintf(fp, "}");
2069                 goto out;
2070         }
2071
2072         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2073
2074         if (details->verbose) {
2075                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2076                                                     __print_attr__fprintf, &first);
2077         } else if (details->freq) {
2078                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2079                                          (u64)evsel->attr.sample_freq);
2080         }
2081 out:
2082         fputc('\n', fp);
2083         return ++printed;
2084 }
2085
2086 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2087                           char *msg, size_t msgsize)
2088 {
2089         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2090             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2091             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2092                 /*
2093                  * If it's cycles then fall back to hrtimer based
2094                  * cpu-clock-tick sw counter, which is always available even if
2095                  * no PMU support.
2096                  *
2097                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2098                  * b0a873e).
2099                  */
2100                 scnprintf(msg, msgsize, "%s",
2101 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2102
2103                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2104                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2105
2106                 zfree(&evsel->name);
2107                 return true;
2108         }
2109
2110         return false;
2111 }
2112
2113 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2114                               int err, char *msg, size_t size)
2115 {
2116         char sbuf[STRERR_BUFSIZE];
2117
2118         switch (err) {
2119         case EPERM:
2120         case EACCES:
2121                 return scnprintf(msg, size,
2122                  "You may not have permission to collect %sstats.\n"
2123                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2124                  " -1 - Not paranoid at all\n"
2125                  "  0 - Disallow raw tracepoint access for unpriv\n"
2126                  "  1 - Disallow cpu events for unpriv\n"
2127                  "  2 - Disallow kernel profiling for unpriv",
2128                                  target->system_wide ? "system-wide " : "");
2129         case ENOENT:
2130                 return scnprintf(msg, size, "The %s event is not supported.",
2131                                  perf_evsel__name(evsel));
2132         case EMFILE:
2133                 return scnprintf(msg, size, "%s",
2134                          "Too many events are opened.\n"
2135                          "Probably the maximum number of open file descriptors has been reached.\n"
2136                          "Hint: Try again after reducing the number of events.\n"
2137                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2138         case ENODEV:
2139                 if (target->cpu_list)
2140                         return scnprintf(msg, size, "%s",
2141          "No such device - did you specify an out-of-range profile CPU?\n");
2142                 break;
2143         case EOPNOTSUPP:
2144                 if (evsel->attr.precise_ip)
2145                         return scnprintf(msg, size, "%s",
2146         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2147 #if defined(__i386__) || defined(__x86_64__)
2148                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2149                         return scnprintf(msg, size, "%s",
2150         "No hardware sampling interrupt available.\n"
2151         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2152 #endif
2153                 break;
2154         case EBUSY:
2155                 if (find_process("oprofiled"))
2156                         return scnprintf(msg, size,
2157         "The PMU counters are busy/taken by another profiler.\n"
2158         "We found oprofile daemon running, please stop it and try again.");
2159                 break;
2160         case EINVAL:
2161                 if (perf_missing_features.clockid)
2162                         return scnprintf(msg, size, "clockid feature not supported.");
2163                 if (perf_missing_features.clockid_wrong)
2164                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2165                 break;
2166         default:
2167                 break;
2168         }
2169
2170         return scnprintf(msg, size,
2171         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2172         "/bin/dmesg may provide additional information.\n"
2173         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2174                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2175                          perf_evsel__name(evsel));
2176 }