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 <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__apply_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__set_filter(struct perf_evsel *evsel, const char *filter)
 827 {
 828         char *new_filter = strdup(filter);
 829
 830         if (new_filter != NULL) {
 831                 free(evsel->filter);
 832                 evsel->filter = new_filter;
 833                 return 0;
 834         }
 835
 836         return -1;
 837 }
 838
 839 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 840 {
 841         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 842                                      PERF_EVENT_IOC_ENABLE,
 843                                      0);
 844 }
 845
 846 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 847 {
 848         if (ncpus == 0 || nthreads == 0)
 849                 return 0;
 850
 851         if (evsel->system_wide)
 852                 nthreads = 1;
 853
 854         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 855         if (evsel->sample_id == NULL)
 856                 return -ENOMEM;
 857
 858         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 859         if (evsel->id == NULL) {
 860                 xyarray__delete(evsel->sample_id);
 861                 evsel->sample_id = NULL;
 862                 return -ENOMEM;
 863         }
 864
 865         return 0;
 866 }
 867
 868 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 869 {
 870         xyarray__delete(evsel->fd);
 871         evsel->fd = NULL;
 872 }
 873
 874 static void perf_evsel__free_id(struct perf_evsel *evsel)
 875 {
 876         xyarray__delete(evsel->sample_id);
 877         evsel->sample_id = NULL;
 878         zfree(&evsel->id);
 879 }
 880
 881 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 882 {
 883         int cpu, thread;
 884
 885         if (evsel->system_wide)
 886                 nthreads = 1;
 887
 888         for (cpu = 0; cpu < ncpus; cpu++)
 889                 for (thread = 0; thread < nthreads; ++thread) {
 890                         close(FD(evsel, cpu, thread));
 891                         FD(evsel, cpu, thread) = -1;
 892                 }
 893 }
 894
 895 void perf_evsel__exit(struct perf_evsel *evsel)
 896 {
 897         assert(list_empty(&evsel->node));
 898         perf_evsel__free_fd(evsel);
 899         perf_evsel__free_id(evsel);
 900         close_cgroup(evsel->cgrp);
 901         cpu_map__put(evsel->cpus);
 902         thread_map__put(evsel->threads);
 903         zfree(&evsel->group_name);
 904         zfree(&evsel->name);
 905         perf_evsel__object.fini(evsel);
 906 }
 907
 908 void perf_evsel__delete(struct perf_evsel *evsel)
 909 {
 910         perf_evsel__exit(evsel);
 911         free(evsel);
 912 }
 913
 914 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
 915                                 struct perf_counts_values *count)
 916 {
 917         struct perf_counts_values tmp;
 918
 919         if (!evsel->prev_raw_counts)
 920                 return;
 921
 922         if (cpu == -1) {
 923                 tmp = evsel->prev_raw_counts->aggr;
 924                 evsel->prev_raw_counts->aggr = *count;
 925         } else {
 926                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
 927                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
 928         }
 929
 930         count->val = count->val - tmp.val;
 931         count->ena = count->ena - tmp.ena;
 932         count->run = count->run - tmp.run;
 933 }
 934
 935 void perf_counts_values__scale(struct perf_counts_values *count,
 936                                bool scale, s8 *pscaled)
 937 {
 938         s8 scaled = 0;
 939
 940         if (scale) {
 941                 if (count->run == 0) {
 942                         scaled = -1;
 943                         count->val = 0;
 944                 } else if (count->run < count->ena) {
 945                         scaled = 1;
 946                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
 947                 }
 948         } else
 949                 count->ena = count->run = 0;
 950
 951         if (pscaled)
 952                 *pscaled = scaled;
 953 }
 954
 955 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
 956                      struct perf_counts_values *count)
 957 {
 958         memset(count, 0, sizeof(*count));
 959
 960         if (FD(evsel, cpu, thread) < 0)
 961                 return -EINVAL;
 962
 963         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
 964                 return -errno;
 965
 966         return 0;
 967 }
 968
 969 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 970                               int cpu, int thread, bool scale)
 971 {
 972         struct perf_counts_values count;
 973         size_t nv = scale ? 3 : 1;
 974
 975         if (FD(evsel, cpu, thread) < 0)
 976                 return -EINVAL;
 977
 978         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
 979                 return -ENOMEM;
 980
 981         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 982                 return -errno;
 983
 984         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
 985         perf_counts_values__scale(&count, scale, NULL);
 986         *perf_counts(evsel->counts, cpu, thread) = count;
 987         return 0;
 988 }
 989
 990 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 991 {
 992         struct perf_evsel *leader = evsel->leader;
 993         int fd;
 994
 995         if (perf_evsel__is_group_leader(evsel))
 996                 return -1;
 997
 998         /*
 999          * Leader must be already processed/open,
1000          * if not it's a bug.
1001          */
1002         BUG_ON(!leader->fd);
1003
1004         fd = FD(leader, cpu, thread);
1005         BUG_ON(fd == -1);
1006
1007         return fd;
1008 }
1009
1010 struct bit_names {
1011         int bit;
1012         const char *name;
1013 };
1014
1015 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1016 {
1017         bool first_bit = true;
1018         int i = 0;
1019
1020         do {
1021                 if (value & bits[i].bit) {
1022                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1023                         first_bit = false;
1024                 }
1025         } while (bits[++i].name != NULL);
1026 }
1027
1028 static void __p_sample_type(char *buf, size_t size, u64 value)
1029 {
1030 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1031         struct bit_names bits[] = {
1032                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1033                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1034                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1035                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1036                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1037                 { .name = NULL, }
1038         };
1039 #undef bit_name
1040         __p_bits(buf, size, value, bits);
1041 }
1042
1043 static void __p_read_format(char *buf, size_t size, u64 value)
1044 {
1045 #define bit_name(n) { PERF_FORMAT_##n, #n }
1046         struct bit_names bits[] = {
1047                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1048                 bit_name(ID), bit_name(GROUP),
1049                 { .name = NULL, }
1050         };
1051 #undef bit_name
1052         __p_bits(buf, size, value, bits);
1053 }
1054
1055 #define BUF_SIZE                1024
1056
1057 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1058 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1059 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1060 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1061 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1062
1063 #define PRINT_ATTRn(_n, _f, _p)                         \
1064 do {                                                    \
1065         if (attr->_f) {                                 \
1066                 _p(attr->_f);                           \
1067                 ret += attr__fprintf(fp, _n, buf, priv);\
1068         }                                               \
1069 } while (0)
1070
1071 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1072
1073 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1074                              attr__fprintf_f attr__fprintf, void *priv)
1075 {
1076         char buf[BUF_SIZE];
1077         int ret = 0;
1078
1079         PRINT_ATTRf(type, p_unsigned);
1080         PRINT_ATTRf(size, p_unsigned);
1081         PRINT_ATTRf(config, p_hex);
1082         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1083         PRINT_ATTRf(sample_type, p_sample_type);
1084         PRINT_ATTRf(read_format, p_read_format);
1085
1086         PRINT_ATTRf(disabled, p_unsigned);
1087         PRINT_ATTRf(inherit, p_unsigned);
1088         PRINT_ATTRf(pinned, p_unsigned);
1089         PRINT_ATTRf(exclusive, p_unsigned);
1090         PRINT_ATTRf(exclude_user, p_unsigned);
1091         PRINT_ATTRf(exclude_kernel, p_unsigned);
1092         PRINT_ATTRf(exclude_hv, p_unsigned);
1093         PRINT_ATTRf(exclude_idle, p_unsigned);
1094         PRINT_ATTRf(mmap, p_unsigned);
1095         PRINT_ATTRf(comm, p_unsigned);
1096         PRINT_ATTRf(freq, p_unsigned);
1097         PRINT_ATTRf(inherit_stat, p_unsigned);
1098         PRINT_ATTRf(enable_on_exec, p_unsigned);
1099         PRINT_ATTRf(task, p_unsigned);
1100         PRINT_ATTRf(watermark, p_unsigned);
1101         PRINT_ATTRf(precise_ip, p_unsigned);
1102         PRINT_ATTRf(mmap_data, p_unsigned);
1103         PRINT_ATTRf(sample_id_all, p_unsigned);
1104         PRINT_ATTRf(exclude_host, p_unsigned);
1105         PRINT_ATTRf(exclude_guest, p_unsigned);
1106         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1107         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1108         PRINT_ATTRf(mmap2, p_unsigned);
1109         PRINT_ATTRf(comm_exec, p_unsigned);
1110         PRINT_ATTRf(use_clockid, p_unsigned);
1111
1112         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1113         PRINT_ATTRf(bp_type, p_unsigned);
1114         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1115         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1116         PRINT_ATTRf(sample_regs_user, p_hex);
1117         PRINT_ATTRf(sample_stack_user, p_unsigned);
1118         PRINT_ATTRf(clockid, p_signed);
1119         PRINT_ATTRf(sample_regs_intr, p_hex);
1120         PRINT_ATTRf(aux_watermark, p_unsigned);
1121
1122         return ret;
1123 }
1124
1125 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1126                                 void *priv __attribute__((unused)))
1127 {
1128         return fprintf(fp, "  %-32s %s\n", name, val);
1129 }
1130
1131 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1132                               struct thread_map *threads)
1133 {
1134         int cpu, thread, nthreads;
1135         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1136         int pid = -1, err;
1137         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1138
1139         if (evsel->system_wide)
1140                 nthreads = 1;
1141         else
1142                 nthreads = threads->nr;
1143
1144         if (evsel->fd == NULL &&
1145             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1146                 return -ENOMEM;
1147
1148         if (evsel->cgrp) {
1149                 flags |= PERF_FLAG_PID_CGROUP;
1150                 pid = evsel->cgrp->fd;
1151         }
1152
1153 fallback_missing_features:
1154         if (perf_missing_features.clockid_wrong)
1155                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1156         if (perf_missing_features.clockid) {
1157                 evsel->attr.use_clockid = 0;
1158                 evsel->attr.clockid = 0;
1159         }
1160         if (perf_missing_features.cloexec)
1161                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1162         if (perf_missing_features.mmap2)
1163                 evsel->attr.mmap2 = 0;
1164         if (perf_missing_features.exclude_guest)
1165                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1166 retry_sample_id:
1167         if (perf_missing_features.sample_id_all)
1168                 evsel->attr.sample_id_all = 0;
1169
1170         if (verbose >= 2) {
1171                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1172                 fprintf(stderr, "perf_event_attr:\n");
1173                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1174                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1175         }
1176
1177         for (cpu = 0; cpu < cpus->nr; cpu++) {
1178
1179                 for (thread = 0; thread < nthreads; thread++) {
1180                         int group_fd;
1181
1182                         if (!evsel->cgrp && !evsel->system_wide)
1183                                 pid = thread_map__pid(threads, thread);
1184
1185                         group_fd = get_group_fd(evsel, cpu, thread);
1186 retry_open:
1187                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1188                                   pid, cpus->map[cpu], group_fd, flags);
1189
1190                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1191                                                                      pid,
1192                                                                      cpus->map[cpu],
1193                                                                      group_fd, flags);
1194                         if (FD(evsel, cpu, thread) < 0) {
1195                                 err = -errno;
1196                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1197                                           err);
1198                                 goto try_fallback;
1199                         }
1200                         set_rlimit = NO_CHANGE;
1201
1202                         /*
1203                          * If we succeeded but had to kill clockid, fail and
1204                          * have perf_evsel__open_strerror() print us a nice
1205                          * error.
1206                          */
1207                         if (perf_missing_features.clockid ||
1208                             perf_missing_features.clockid_wrong) {
1209                                 err = -EINVAL;
1210                                 goto out_close;
1211                         }
1212                 }
1213         }
1214
1215         return 0;
1216
1217 try_fallback:
1218         /*
1219          * perf stat needs between 5 and 22 fds per CPU. When we run out
1220          * of them try to increase the limits.
1221          */
1222         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1223                 struct rlimit l;
1224                 int old_errno = errno;
1225
1226                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1227                         if (set_rlimit == NO_CHANGE)
1228                                 l.rlim_cur = l.rlim_max;
1229                         else {
1230                                 l.rlim_cur = l.rlim_max + 1000;
1231                                 l.rlim_max = l.rlim_cur;
1232                         }
1233                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1234                                 set_rlimit++;
1235                                 errno = old_errno;
1236                                 goto retry_open;
1237                         }
1238                 }
1239                 errno = old_errno;
1240         }
1241
1242         if (err != -EINVAL || cpu > 0 || thread > 0)
1243                 goto out_close;
1244
1245         /*
1246          * Must probe features in the order they were added to the
1247          * perf_event_attr interface.
1248          */
1249         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1250                 perf_missing_features.clockid_wrong = true;
1251                 goto fallback_missing_features;
1252         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1253                 perf_missing_features.clockid = true;
1254                 goto fallback_missing_features;
1255         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1256                 perf_missing_features.cloexec = true;
1257                 goto fallback_missing_features;
1258         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1259                 perf_missing_features.mmap2 = true;
1260                 goto fallback_missing_features;
1261         } else if (!perf_missing_features.exclude_guest &&
1262                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1263                 perf_missing_features.exclude_guest = true;
1264                 goto fallback_missing_features;
1265         } else if (!perf_missing_features.sample_id_all) {
1266                 perf_missing_features.sample_id_all = true;
1267                 goto retry_sample_id;
1268         }
1269
1270 out_close:
1271         do {
1272                 while (--thread >= 0) {
1273                         close(FD(evsel, cpu, thread));
1274                         FD(evsel, cpu, thread) = -1;
1275                 }
1276                 thread = nthreads;
1277         } while (--cpu >= 0);
1278         return err;
1279 }
1280
1281 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1282 {
1283         if (evsel->fd == NULL)
1284                 return;
1285
1286         perf_evsel__close_fd(evsel, ncpus, nthreads);
1287         perf_evsel__free_fd(evsel);
1288 }
1289
1290 static struct {
1291         struct cpu_map map;
1292         int cpus[1];
1293 } empty_cpu_map = {
1294         .map.nr = 1,
1295         .cpus   = { -1, },
1296 };
1297
1298 static struct {
1299         struct thread_map map;
1300         int threads[1];
1301 } empty_thread_map = {
1302         .map.nr  = 1,
1303         .threads = { -1, },
1304 };
1305
1306 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1307                      struct thread_map *threads)
1308 {
1309         if (cpus == NULL) {
1310                 /* Work around old compiler warnings about strict aliasing */
1311                 cpus = &empty_cpu_map.map;
1312         }
1313
1314         if (threads == NULL)
1315                 threads = &empty_thread_map.map;
1316
1317         return __perf_evsel__open(evsel, cpus, threads);
1318 }
1319
1320 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1321                              struct cpu_map *cpus)
1322 {
1323         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1324 }
1325
1326 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1327                                 struct thread_map *threads)
1328 {
1329         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1330 }
1331
1332 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1333                                        const union perf_event *event,
1334                                        struct perf_sample *sample)
1335 {
1336         u64 type = evsel->attr.sample_type;
1337         const u64 *array = event->sample.array;
1338         bool swapped = evsel->needs_swap;
1339         union u64_swap u;
1340
1341         array += ((event->header.size -
1342                    sizeof(event->header)) / sizeof(u64)) - 1;
1343
1344         if (type & PERF_SAMPLE_IDENTIFIER) {
1345                 sample->id = *array;
1346                 array--;
1347         }
1348
1349         if (type & PERF_SAMPLE_CPU) {
1350                 u.val64 = *array;
1351                 if (swapped) {
1352                         /* undo swap of u64, then swap on individual u32s */
1353                         u.val64 = bswap_64(u.val64);
1354                         u.val32[0] = bswap_32(u.val32[0]);
1355                 }
1356
1357                 sample->cpu = u.val32[0];
1358                 array--;
1359         }
1360
1361         if (type & PERF_SAMPLE_STREAM_ID) {
1362                 sample->stream_id = *array;
1363                 array--;
1364         }
1365
1366         if (type & PERF_SAMPLE_ID) {
1367                 sample->id = *array;
1368                 array--;
1369         }
1370
1371         if (type & PERF_SAMPLE_TIME) {
1372                 sample->time = *array;
1373                 array--;
1374         }
1375
1376         if (type & PERF_SAMPLE_TID) {
1377                 u.val64 = *array;
1378                 if (swapped) {
1379                         /* undo swap of u64, then swap on individual u32s */
1380                         u.val64 = bswap_64(u.val64);
1381                         u.val32[0] = bswap_32(u.val32[0]);
1382                         u.val32[1] = bswap_32(u.val32[1]);
1383                 }
1384
1385                 sample->pid = u.val32[0];
1386                 sample->tid = u.val32[1];
1387                 array--;
1388         }
1389
1390         return 0;
1391 }
1392
1393 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1394                             u64 size)
1395 {
1396         return size > max_size || offset + size > endp;
1397 }
1398
1399 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1400         do {                                                            \
1401                 if (overflow(endp, (max_size), (offset), (size)))       \
1402                         return -EFAULT;                                 \
1403         } while (0)
1404
1405 #define OVERFLOW_CHECK_u64(offset) \
1406         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1407
1408 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1409                              struct perf_sample *data)
1410 {
1411         u64 type = evsel->attr.sample_type;
1412         bool swapped = evsel->needs_swap;
1413         const u64 *array;
1414         u16 max_size = event->header.size;
1415         const void *endp = (void *)event + max_size;
1416         u64 sz;
1417
1418         /*
1419          * used for cross-endian analysis. See git commit 65014ab3
1420          * for why this goofiness is needed.
1421          */
1422         union u64_swap u;
1423
1424         memset(data, 0, sizeof(*data));
1425         data->cpu = data->pid = data->tid = -1;
1426         data->stream_id = data->id = data->time = -1ULL;
1427         data->period = evsel->attr.sample_period;
1428         data->weight = 0;
1429
1430         if (event->header.type != PERF_RECORD_SAMPLE) {
1431                 if (!evsel->attr.sample_id_all)
1432                         return 0;
1433                 return perf_evsel__parse_id_sample(evsel, event, data);
1434         }
1435
1436         array = event->sample.array;
1437
1438         /*
1439          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1440          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1441          * check the format does not go past the end of the event.
1442          */
1443         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1444                 return -EFAULT;
1445
1446         data->id = -1ULL;
1447         if (type & PERF_SAMPLE_IDENTIFIER) {
1448                 data->id = *array;
1449                 array++;
1450         }
1451
1452         if (type & PERF_SAMPLE_IP) {
1453                 data->ip = *array;
1454                 array++;
1455         }
1456
1457         if (type & PERF_SAMPLE_TID) {
1458                 u.val64 = *array;
1459                 if (swapped) {
1460                         /* undo swap of u64, then swap on individual u32s */
1461                         u.val64 = bswap_64(u.val64);
1462                         u.val32[0] = bswap_32(u.val32[0]);
1463                         u.val32[1] = bswap_32(u.val32[1]);
1464                 }
1465
1466                 data->pid = u.val32[0];
1467                 data->tid = u.val32[1];
1468                 array++;
1469         }
1470
1471         if (type & PERF_SAMPLE_TIME) {
1472                 data->time = *array;
1473                 array++;
1474         }
1475
1476         data->addr = 0;
1477         if (type & PERF_SAMPLE_ADDR) {
1478                 data->addr = *array;
1479                 array++;
1480         }
1481
1482         if (type & PERF_SAMPLE_ID) {
1483                 data->id = *array;
1484                 array++;
1485         }
1486
1487         if (type & PERF_SAMPLE_STREAM_ID) {
1488                 data->stream_id = *array;
1489                 array++;
1490         }
1491
1492         if (type & PERF_SAMPLE_CPU) {
1493
1494                 u.val64 = *array;
1495                 if (swapped) {
1496                         /* undo swap of u64, then swap on individual u32s */
1497                         u.val64 = bswap_64(u.val64);
1498                         u.val32[0] = bswap_32(u.val32[0]);
1499                 }
1500
1501                 data->cpu = u.val32[0];
1502                 array++;
1503         }
1504
1505         if (type & PERF_SAMPLE_PERIOD) {
1506                 data->period = *array;
1507                 array++;
1508         }
1509
1510         if (type & PERF_SAMPLE_READ) {
1511                 u64 read_format = evsel->attr.read_format;
1512
1513                 OVERFLOW_CHECK_u64(array);
1514                 if (read_format & PERF_FORMAT_GROUP)
1515                         data->read.group.nr = *array;
1516                 else
1517                         data->read.one.value = *array;
1518
1519                 array++;
1520
1521                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1522                         OVERFLOW_CHECK_u64(array);
1523                         data->read.time_enabled = *array;
1524                         array++;
1525                 }
1526
1527                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1528                         OVERFLOW_CHECK_u64(array);
1529                         data->read.time_running = *array;
1530                         array++;
1531                 }
1532
1533                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1534                 if (read_format & PERF_FORMAT_GROUP) {
1535                         const u64 max_group_nr = UINT64_MAX /
1536                                         sizeof(struct sample_read_value);
1537
1538                         if (data->read.group.nr > max_group_nr)
1539                                 return -EFAULT;
1540                         sz = data->read.group.nr *
1541                              sizeof(struct sample_read_value);
1542                         OVERFLOW_CHECK(array, sz, max_size);
1543                         data->read.group.values =
1544                                         (struct sample_read_value *)array;
1545                         array = (void *)array + sz;
1546                 } else {
1547                         OVERFLOW_CHECK_u64(array);
1548                         data->read.one.id = *array;
1549                         array++;
1550                 }
1551         }
1552
1553         if (type & PERF_SAMPLE_CALLCHAIN) {
1554                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1555
1556                 OVERFLOW_CHECK_u64(array);
1557                 data->callchain = (struct ip_callchain *)array++;
1558                 if (data->callchain->nr > max_callchain_nr)
1559                         return -EFAULT;
1560                 sz = data->callchain->nr * sizeof(u64);
1561                 OVERFLOW_CHECK(array, sz, max_size);
1562                 array = (void *)array + sz;
1563         }
1564
1565         if (type & PERF_SAMPLE_RAW) {
1566                 OVERFLOW_CHECK_u64(array);
1567                 u.val64 = *array;
1568                 if (WARN_ONCE(swapped,
1569                               "Endianness of raw data not corrected!\n")) {
1570                         /* undo swap of u64, then swap on individual u32s */
1571                         u.val64 = bswap_64(u.val64);
1572                         u.val32[0] = bswap_32(u.val32[0]);
1573                         u.val32[1] = bswap_32(u.val32[1]);
1574                 }
1575                 data->raw_size = u.val32[0];
1576                 array = (void *)array + sizeof(u32);
1577
1578                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1579                 data->raw_data = (void *)array;
1580                 array = (void *)array + data->raw_size;
1581         }
1582
1583         if (type & PERF_SAMPLE_BRANCH_STACK) {
1584                 const u64 max_branch_nr = UINT64_MAX /
1585                                           sizeof(struct branch_entry);
1586
1587                 OVERFLOW_CHECK_u64(array);
1588                 data->branch_stack = (struct branch_stack *)array++;
1589
1590                 if (data->branch_stack->nr > max_branch_nr)
1591                         return -EFAULT;
1592                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1593                 OVERFLOW_CHECK(array, sz, max_size);
1594                 array = (void *)array + sz;
1595         }
1596
1597         if (type & PERF_SAMPLE_REGS_USER) {
1598                 OVERFLOW_CHECK_u64(array);
1599                 data->user_regs.abi = *array;
1600                 array++;
1601
1602                 if (data->user_regs.abi) {
1603                         u64 mask = evsel->attr.sample_regs_user;
1604
1605                         sz = hweight_long(mask) * sizeof(u64);
1606                         OVERFLOW_CHECK(array, sz, max_size);
1607                         data->user_regs.mask = mask;
1608                         data->user_regs.regs = (u64 *)array;
1609                         array = (void *)array + sz;
1610                 }
1611         }
1612
1613         if (type & PERF_SAMPLE_STACK_USER) {
1614                 OVERFLOW_CHECK_u64(array);
1615                 sz = *array++;
1616
1617                 data->user_stack.offset = ((char *)(array - 1)
1618                                           - (char *) event);
1619
1620                 if (!sz) {
1621                         data->user_stack.size = 0;
1622                 } else {
1623                         OVERFLOW_CHECK(array, sz, max_size);
1624                         data->user_stack.data = (char *)array;
1625                         array = (void *)array + sz;
1626                         OVERFLOW_CHECK_u64(array);
1627                         data->user_stack.size = *array++;
1628                         if (WARN_ONCE(data->user_stack.size > sz,
1629                                       "user stack dump failure\n"))
1630                                 return -EFAULT;
1631                 }
1632         }
1633
1634         data->weight = 0;
1635         if (type & PERF_SAMPLE_WEIGHT) {
1636                 OVERFLOW_CHECK_u64(array);
1637                 data->weight = *array;
1638                 array++;
1639         }
1640
1641         data->data_src = PERF_MEM_DATA_SRC_NONE;
1642         if (type & PERF_SAMPLE_DATA_SRC) {
1643                 OVERFLOW_CHECK_u64(array);
1644                 data->data_src = *array;
1645                 array++;
1646         }
1647
1648         data->transaction = 0;
1649         if (type & PERF_SAMPLE_TRANSACTION) {
1650                 OVERFLOW_CHECK_u64(array);
1651                 data->transaction = *array;
1652                 array++;
1653         }
1654
1655         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1656         if (type & PERF_SAMPLE_REGS_INTR) {
1657                 OVERFLOW_CHECK_u64(array);
1658                 data->intr_regs.abi = *array;
1659                 array++;
1660
1661                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1662                         u64 mask = evsel->attr.sample_regs_intr;
1663
1664                         sz = hweight_long(mask) * sizeof(u64);
1665                         OVERFLOW_CHECK(array, sz, max_size);
1666                         data->intr_regs.mask = mask;
1667                         data->intr_regs.regs = (u64 *)array;
1668                         array = (void *)array + sz;
1669                 }
1670         }
1671
1672         return 0;
1673 }
1674
1675 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1676                                      u64 read_format)
1677 {
1678         size_t sz, result = sizeof(struct sample_event);
1679
1680         if (type & PERF_SAMPLE_IDENTIFIER)
1681                 result += sizeof(u64);
1682
1683         if (type & PERF_SAMPLE_IP)
1684                 result += sizeof(u64);
1685
1686         if (type & PERF_SAMPLE_TID)
1687                 result += sizeof(u64);
1688
1689         if (type & PERF_SAMPLE_TIME)
1690                 result += sizeof(u64);
1691
1692         if (type & PERF_SAMPLE_ADDR)
1693                 result += sizeof(u64);
1694
1695         if (type & PERF_SAMPLE_ID)
1696                 result += sizeof(u64);
1697
1698         if (type & PERF_SAMPLE_STREAM_ID)
1699                 result += sizeof(u64);
1700
1701         if (type & PERF_SAMPLE_CPU)
1702                 result += sizeof(u64);
1703
1704         if (type & PERF_SAMPLE_PERIOD)
1705                 result += sizeof(u64);
1706
1707         if (type & PERF_SAMPLE_READ) {
1708                 result += sizeof(u64);
1709                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1710                         result += sizeof(u64);
1711                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1712                         result += sizeof(u64);
1713                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1714                 if (read_format & PERF_FORMAT_GROUP) {
1715                         sz = sample->read.group.nr *
1716                              sizeof(struct sample_read_value);
1717                         result += sz;
1718                 } else {
1719                         result += sizeof(u64);
1720                 }
1721         }
1722
1723         if (type & PERF_SAMPLE_CALLCHAIN) {
1724                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1725                 result += sz;
1726         }
1727
1728         if (type & PERF_SAMPLE_RAW) {
1729                 result += sizeof(u32);
1730                 result += sample->raw_size;
1731         }
1732
1733         if (type & PERF_SAMPLE_BRANCH_STACK) {
1734                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1735                 sz += sizeof(u64);
1736                 result += sz;
1737         }
1738
1739         if (type & PERF_SAMPLE_REGS_USER) {
1740                 if (sample->user_regs.abi) {
1741                         result += sizeof(u64);
1742                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1743                         result += sz;
1744                 } else {
1745                         result += sizeof(u64);
1746                 }
1747         }
1748
1749         if (type & PERF_SAMPLE_STACK_USER) {
1750                 sz = sample->user_stack.size;
1751                 result += sizeof(u64);
1752                 if (sz) {
1753                         result += sz;
1754                         result += sizeof(u64);
1755                 }
1756         }
1757
1758         if (type & PERF_SAMPLE_WEIGHT)
1759                 result += sizeof(u64);
1760
1761         if (type & PERF_SAMPLE_DATA_SRC)
1762                 result += sizeof(u64);
1763
1764         if (type & PERF_SAMPLE_TRANSACTION)
1765                 result += sizeof(u64);
1766
1767         if (type & PERF_SAMPLE_REGS_INTR) {
1768                 if (sample->intr_regs.abi) {
1769                         result += sizeof(u64);
1770                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1771                         result += sz;
1772                 } else {
1773                         result += sizeof(u64);
1774                 }
1775         }
1776
1777         return result;
1778 }
1779
1780 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1781                                   u64 read_format,
1782                                   const struct perf_sample *sample,
1783                                   bool swapped)
1784 {
1785         u64 *array;
1786         size_t sz;
1787         /*
1788          * used for cross-endian analysis. See git commit 65014ab3
1789          * for why this goofiness is needed.
1790          */
1791         union u64_swap u;
1792
1793         array = event->sample.array;
1794
1795         if (type & PERF_SAMPLE_IDENTIFIER) {
1796                 *array = sample->id;
1797                 array++;
1798         }
1799
1800         if (type & PERF_SAMPLE_IP) {
1801                 *array = sample->ip;
1802                 array++;
1803         }
1804
1805         if (type & PERF_SAMPLE_TID) {
1806                 u.val32[0] = sample->pid;
1807                 u.val32[1] = sample->tid;
1808                 if (swapped) {
1809                         /*
1810                          * Inverse of what is done in perf_evsel__parse_sample
1811                          */
1812                         u.val32[0] = bswap_32(u.val32[0]);
1813                         u.val32[1] = bswap_32(u.val32[1]);
1814                         u.val64 = bswap_64(u.val64);
1815                 }
1816
1817                 *array = u.val64;
1818                 array++;
1819         }
1820
1821         if (type & PERF_SAMPLE_TIME) {
1822                 *array = sample->time;
1823                 array++;
1824         }
1825
1826         if (type & PERF_SAMPLE_ADDR) {
1827                 *array = sample->addr;
1828                 array++;
1829         }
1830
1831         if (type & PERF_SAMPLE_ID) {
1832                 *array = sample->id;
1833                 array++;
1834         }
1835
1836         if (type & PERF_SAMPLE_STREAM_ID) {
1837                 *array = sample->stream_id;
1838                 array++;
1839         }
1840
1841         if (type & PERF_SAMPLE_CPU) {
1842                 u.val32[0] = sample->cpu;
1843                 if (swapped) {
1844                         /*
1845                          * Inverse of what is done in perf_evsel__parse_sample
1846                          */
1847                         u.val32[0] = bswap_32(u.val32[0]);
1848                         u.val64 = bswap_64(u.val64);
1849                 }
1850                 *array = u.val64;
1851                 array++;
1852         }
1853
1854         if (type & PERF_SAMPLE_PERIOD) {
1855                 *array = sample->period;
1856                 array++;
1857         }
1858
1859         if (type & PERF_SAMPLE_READ) {
1860                 if (read_format & PERF_FORMAT_GROUP)
1861                         *array = sample->read.group.nr;
1862                 else
1863                         *array = sample->read.one.value;
1864                 array++;
1865
1866                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1867                         *array = sample->read.time_enabled;
1868                         array++;
1869                 }
1870
1871                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1872                         *array = sample->read.time_running;
1873                         array++;
1874                 }
1875
1876                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1877                 if (read_format & PERF_FORMAT_GROUP) {
1878                         sz = sample->read.group.nr *
1879                              sizeof(struct sample_read_value);
1880                         memcpy(array, sample->read.group.values, sz);
1881                         array = (void *)array + sz;
1882                 } else {
1883                         *array = sample->read.one.id;
1884                         array++;
1885                 }
1886         }
1887
1888         if (type & PERF_SAMPLE_CALLCHAIN) {
1889                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1890                 memcpy(array, sample->callchain, sz);
1891                 array = (void *)array + sz;
1892         }
1893
1894         if (type & PERF_SAMPLE_RAW) {
1895                 u.val32[0] = sample->raw_size;
1896                 if (WARN_ONCE(swapped,
1897                               "Endianness of raw data not corrected!\n")) {
1898                         /*
1899                          * Inverse of what is done in perf_evsel__parse_sample
1900                          */
1901                         u.val32[0] = bswap_32(u.val32[0]);
1902                         u.val32[1] = bswap_32(u.val32[1]);
1903                         u.val64 = bswap_64(u.val64);
1904                 }
1905                 *array = u.val64;
1906                 array = (void *)array + sizeof(u32);
1907
1908                 memcpy(array, sample->raw_data, sample->raw_size);
1909                 array = (void *)array + sample->raw_size;
1910         }
1911
1912         if (type & PERF_SAMPLE_BRANCH_STACK) {
1913                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1914                 sz += sizeof(u64);
1915                 memcpy(array, sample->branch_stack, sz);
1916                 array = (void *)array + sz;
1917         }
1918
1919         if (type & PERF_SAMPLE_REGS_USER) {
1920                 if (sample->user_regs.abi) {
1921                         *array++ = sample->user_regs.abi;
1922                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1923                         memcpy(array, sample->user_regs.regs, sz);
1924                         array = (void *)array + sz;
1925                 } else {
1926                         *array++ = 0;
1927                 }
1928         }
1929
1930         if (type & PERF_SAMPLE_STACK_USER) {
1931                 sz = sample->user_stack.size;
1932                 *array++ = sz;
1933                 if (sz) {
1934                         memcpy(array, sample->user_stack.data, sz);
1935                         array = (void *)array + sz;
1936                         *array++ = sz;
1937                 }
1938         }
1939
1940         if (type & PERF_SAMPLE_WEIGHT) {
1941                 *array = sample->weight;
1942                 array++;
1943         }
1944
1945         if (type & PERF_SAMPLE_DATA_SRC) {
1946                 *array = sample->data_src;
1947                 array++;
1948         }
1949
1950         if (type & PERF_SAMPLE_TRANSACTION) {
1951                 *array = sample->transaction;
1952                 array++;
1953         }
1954
1955         if (type & PERF_SAMPLE_REGS_INTR) {
1956                 if (sample->intr_regs.abi) {
1957                         *array++ = sample->intr_regs.abi;
1958                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1959                         memcpy(array, sample->intr_regs.regs, sz);
1960                         array = (void *)array + sz;
1961                 } else {
1962                         *array++ = 0;
1963                 }
1964         }
1965
1966         return 0;
1967 }
1968
1969 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1970 {
1971         return pevent_find_field(evsel->tp_format, name);
1972 }
1973
1974 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1975                          const char *name)
1976 {
1977         struct format_field *field = perf_evsel__field(evsel, name);
1978         int offset;
1979
1980         if (!field)
1981                 return NULL;
1982
1983         offset = field->offset;
1984
1985         if (field->flags & FIELD_IS_DYNAMIC) {
1986                 offset = *(int *)(sample->raw_data + field->offset);
1987                 offset &= 0xffff;
1988         }
1989
1990         return sample->raw_data + offset;
1991 }
1992
1993 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1994                        const char *name)
1995 {
1996         struct format_field *field = perf_evsel__field(evsel, name);
1997         void *ptr;
1998         u64 value;
1999
2000         if (!field)
2001                 return 0;
2002
2003         ptr = sample->raw_data + field->offset;
2004
2005         switch (field->size) {
2006         case 1:
2007                 return *(u8 *)ptr;
2008         case 2:
2009                 value = *(u16 *)ptr;
2010                 break;
2011         case 4:
2012                 value = *(u32 *)ptr;
2013                 break;
2014         case 8:
2015                 memcpy(&value, ptr, sizeof(u64));
2016                 break;
2017         default:
2018                 return 0;
2019         }
2020
2021         if (!evsel->needs_swap)
2022                 return value;
2023
2024         switch (field->size) {
2025         case 2:
2026                 return bswap_16(value);
2027         case 4:
2028                 return bswap_32(value);
2029         case 8:
2030                 return bswap_64(value);
2031         default:
2032                 return 0;
2033         }
2034
2035         return 0;
2036 }
2037
2038 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2039 {
2040         va_list args;
2041         int ret = 0;
2042
2043         if (!*first) {
2044                 ret += fprintf(fp, ",");
2045         } else {
2046                 ret += fprintf(fp, ":");
2047                 *first = false;
2048         }
2049
2050         va_start(args, fmt);
2051         ret += vfprintf(fp, fmt, args);
2052         va_end(args);
2053         return ret;
2054 }
2055
2056 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2057 {
2058         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2059 }
2060
2061 int perf_evsel__fprintf(struct perf_evsel *evsel,
2062                         struct perf_attr_details *details, FILE *fp)
2063 {
2064         bool first = true;
2065         int printed = 0;
2066
2067         if (details->event_group) {
2068                 struct perf_evsel *pos;
2069
2070                 if (!perf_evsel__is_group_leader(evsel))
2071                         return 0;
2072
2073                 if (evsel->nr_members > 1)
2074                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2075
2076                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2077                 for_each_group_member(pos, evsel)
2078                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2079
2080                 if (evsel->nr_members > 1)
2081                         printed += fprintf(fp, "}");
2082                 goto out;
2083         }
2084
2085         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2086
2087         if (details->verbose) {
2088                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2089                                                     __print_attr__fprintf, &first);
2090         } else if (details->freq) {
2091                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2092                                          (u64)evsel->attr.sample_freq);
2093         }
2094 out:
2095         fputc('\n', fp);
2096         return ++printed;
2097 }
2098
2099 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2100                           char *msg, size_t msgsize)
2101 {
2102         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2103             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2104             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2105                 /*
2106                  * If it's cycles then fall back to hrtimer based
2107                  * cpu-clock-tick sw counter, which is always available even if
2108                  * no PMU support.
2109                  *
2110                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2111                  * b0a873e).
2112                  */
2113                 scnprintf(msg, msgsize, "%s",
2114 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2115
2116                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2117                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2118
2119                 zfree(&evsel->name);
2120                 return true;
2121         }
2122
2123         return false;
2124 }
2125
2126 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2127                               int err, char *msg, size_t size)
2128 {
2129         char sbuf[STRERR_BUFSIZE];
2130
2131         switch (err) {
2132         case EPERM:
2133         case EACCES:
2134                 return scnprintf(msg, size,
2135                  "You may not have permission to collect %sstats.\n"
2136                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2137                  " -1 - Not paranoid at all\n"
2138                  "  0 - Disallow raw tracepoint access for unpriv\n"
2139                  "  1 - Disallow cpu events for unpriv\n"
2140                  "  2 - Disallow kernel profiling for unpriv",
2141                                  target->system_wide ? "system-wide " : "");
2142         case ENOENT:
2143                 return scnprintf(msg, size, "The %s event is not supported.",
2144                                  perf_evsel__name(evsel));
2145         case EMFILE:
2146                 return scnprintf(msg, size, "%s",
2147                          "Too many events are opened.\n"
2148                          "Probably the maximum number of open file descriptors has been reached.\n"
2149                          "Hint: Try again after reducing the number of events.\n"
2150                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2151         case ENODEV:
2152                 if (target->cpu_list)
2153                         return scnprintf(msg, size, "%s",
2154          "No such device - did you specify an out-of-range profile CPU?\n");
2155                 break;
2156         case EOPNOTSUPP:
2157                 if (evsel->attr.precise_ip)
2158                         return scnprintf(msg, size, "%s",
2159         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2160 #if defined(__i386__) || defined(__x86_64__)
2161                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2162                         return scnprintf(msg, size, "%s",
2163         "No hardware sampling interrupt available.\n"
2164         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2165 #endif
2166                 break;
2167         case EBUSY:
2168                 if (find_process("oprofiled"))
2169                         return scnprintf(msg, size,
2170         "The PMU counters are busy/taken by another profiler.\n"
2171         "We found oprofile daemon running, please stop it and try again.");
2172                 break;
2173         case EINVAL:
2174                 if (perf_missing_features.clockid)
2175                         return scnprintf(msg, size, "clockid feature not supported.");
2176                 if (perf_missing_features.clockid_wrong)
2177                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2178                 break;
2179         default:
2180                 break;
2181         }
2182
2183         return scnprintf(msg, size,
2184         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2185         "/bin/dmesg may provide additional information.\n"
2186         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2187                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2188                          perf_evsel__name(evsel));
2189 }