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         evsel->cmdline_group_boundary = false;
 214 }
 215
 216 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 217 {
 218         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 219
 220         if (evsel != NULL)
 221                 perf_evsel__init(evsel, attr, idx);
 222
 223         return evsel;
 224 }
 225
 226 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 227 {
 228         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 229
 230         if (evsel != NULL) {
 231                 struct perf_event_attr attr = {
 232                         .type          = PERF_TYPE_TRACEPOINT,
 233                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 234                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 235                 };
 236
 237                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 238                         goto out_free;
 239
 240                 evsel->tp_format = trace_event__tp_format(sys, name);
 241                 if (evsel->tp_format == NULL)
 242                         goto out_free;
 243
 244                 event_attr_init(&attr);
 245                 attr.config = evsel->tp_format->id;
 246                 attr.sample_period = 1;
 247                 perf_evsel__init(evsel, &attr, idx);
 248         }
 249
 250         return evsel;
 251
 252 out_free:
 253         zfree(&evsel->name);
 254         free(evsel);
 255         return NULL;
 256 }
 257
 258 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 259         "cycles",
 260         "instructions",
 261         "cache-references",
 262         "cache-misses",
 263         "branches",
 264         "branch-misses",
 265         "bus-cycles",
 266         "stalled-cycles-frontend",
 267         "stalled-cycles-backend",
 268         "ref-cycles",
 269 };
 270
 271 static const char *__perf_evsel__hw_name(u64 config)
 272 {
 273         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 274                 return perf_evsel__hw_names[config];
 275
 276         return "unknown-hardware";
 277 }
 278
 279 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 280 {
 281         int colon = 0, r = 0;
 282         struct perf_event_attr *attr = &evsel->attr;
 283         bool exclude_guest_default = false;
 284
 285 #define MOD_PRINT(context, mod) do {                                    \
 286                 if (!attr->exclude_##context) {                         \
 287                         if (!colon) colon = ++r;                        \
 288                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 289                 } } while(0)
 290
 291         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 292                 MOD_PRINT(kernel, 'k');
 293                 MOD_PRINT(user, 'u');
 294                 MOD_PRINT(hv, 'h');
 295                 exclude_guest_default = true;
 296         }
 297
 298         if (attr->precise_ip) {
 299                 if (!colon)
 300                         colon = ++r;
 301                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 302                 exclude_guest_default = true;
 303         }
 304
 305         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 306                 MOD_PRINT(host, 'H');
 307                 MOD_PRINT(guest, 'G');
 308         }
 309 #undef MOD_PRINT
 310         if (colon)
 311                 bf[colon - 1] = ':';
 312         return r;
 313 }
 314
 315 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 316 {
 317         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 318         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 319 }
 320
 321 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 322         "cpu-clock",
 323         "task-clock",
 324         "page-faults",
 325         "context-switches",
 326         "cpu-migrations",
 327         "minor-faults",
 328         "major-faults",
 329         "alignment-faults",
 330         "emulation-faults",
 331         "dummy",
 332 };
 333
 334 static const char *__perf_evsel__sw_name(u64 config)
 335 {
 336         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 337                 return perf_evsel__sw_names[config];
 338         return "unknown-software";
 339 }
 340
 341 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 342 {
 343         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 344         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 345 }
 346
 347 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 348 {
 349         int r;
 350
 351         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 352
 353         if (type & HW_BREAKPOINT_R)
 354                 r += scnprintf(bf + r, size - r, "r");
 355
 356         if (type & HW_BREAKPOINT_W)
 357                 r += scnprintf(bf + r, size - r, "w");
 358
 359         if (type & HW_BREAKPOINT_X)
 360                 r += scnprintf(bf + r, size - r, "x");
 361
 362         return r;
 363 }
 364
 365 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 366 {
 367         struct perf_event_attr *attr = &evsel->attr;
 368         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 369         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 370 }
 371
 372 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 373                                 [PERF_EVSEL__MAX_ALIASES] = {
 374  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 375  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 376  { "LLC",       "L2",                                                   },
 377  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 378  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 379  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 380  { "node",                                                              },
 381 };
 382
 383 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 384                                    [PERF_EVSEL__MAX_ALIASES] = {
 385  { "load",      "loads",        "read",                                 },
 386  { "store",     "stores",       "write",                                },
 387  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 388 };
 389
 390 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 391                                        [PERF_EVSEL__MAX_ALIASES] = {
 392  { "refs",      "Reference",    "ops",          "access",               },
 393  { "misses",    "miss",                                                 },
 394 };
 395
 396 #define C(x)            PERF_COUNT_HW_CACHE_##x
 397 #define CACHE_READ      (1 << C(OP_READ))
 398 #define CACHE_WRITE     (1 << C(OP_WRITE))
 399 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 400 #define COP(x)          (1 << x)
 401
 402 /*
 403  * cache operartion stat
 404  * L1I : Read and prefetch only
 405  * ITLB and BPU : Read-only
 406  */
 407 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 408  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 409  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 410  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 411  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 412  [C(ITLB)]      = (CACHE_READ),
 413  [C(BPU)]       = (CACHE_READ),
 414  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 415 };
 416
 417 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 418 {
 419         if (perf_evsel__hw_cache_stat[type] & COP(op))
 420                 return true;    /* valid */
 421         else
 422                 return false;   /* invalid */
 423 }
 424
 425 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 426                                             char *bf, size_t size)
 427 {
 428         if (result) {
 429                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 430                                  perf_evsel__hw_cache_op[op][0],
 431                                  perf_evsel__hw_cache_result[result][0]);
 432         }
 433
 434         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 435                          perf_evsel__hw_cache_op[op][1]);
 436 }
 437
 438 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 439 {
 440         u8 op, result, type = (config >>  0) & 0xff;
 441         const char *err = "unknown-ext-hardware-cache-type";
 442
 443         if (type > PERF_COUNT_HW_CACHE_MAX)
 444                 goto out_err;
 445
 446         op = (config >>  8) & 0xff;
 447         err = "unknown-ext-hardware-cache-op";
 448         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 449                 goto out_err;
 450
 451         result = (config >> 16) & 0xff;
 452         err = "unknown-ext-hardware-cache-result";
 453         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 454                 goto out_err;
 455
 456         err = "invalid-cache";
 457         if (!perf_evsel__is_cache_op_valid(type, op))
 458                 goto out_err;
 459
 460         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 461 out_err:
 462         return scnprintf(bf, size, "%s", err);
 463 }
 464
 465 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 466 {
 467         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 468         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 469 }
 470
 471 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 472 {
 473         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 474         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 475 }
 476
 477 const char *perf_evsel__name(struct perf_evsel *evsel)
 478 {
 479         char bf[128];
 480
 481         if (evsel->name)
 482                 return evsel->name;
 483
 484         switch (evsel->attr.type) {
 485         case PERF_TYPE_RAW:
 486                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 487                 break;
 488
 489         case PERF_TYPE_HARDWARE:
 490                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 491                 break;
 492
 493         case PERF_TYPE_HW_CACHE:
 494                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 495                 break;
 496
 497         case PERF_TYPE_SOFTWARE:
 498                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 499                 break;
 500
 501         case PERF_TYPE_TRACEPOINT:
 502                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 503                 break;
 504
 505         case PERF_TYPE_BREAKPOINT:
 506                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 507                 break;
 508
 509         default:
 510                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 511                           evsel->attr.type);
 512                 break;
 513         }
 514
 515         evsel->name = strdup(bf);
 516
 517         return evsel->name ?: "unknown";
 518 }
 519
 520 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 521 {
 522         return evsel->group_name ?: "anon group";
 523 }
 524
 525 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 526 {
 527         int ret;
 528         struct perf_evsel *pos;
 529         const char *group_name = perf_evsel__group_name(evsel);
 530
 531         ret = scnprintf(buf, size, "%s", group_name);
 532
 533         ret += scnprintf(buf + ret, size - ret, " { %s",
 534                          perf_evsel__name(evsel));
 535
 536         for_each_group_member(pos, evsel)
 537                 ret += scnprintf(buf + ret, size - ret, ", %s",
 538                                  perf_evsel__name(pos));
 539
 540         ret += scnprintf(buf + ret, size - ret, " }");
 541
 542         return ret;
 543 }
 544
 545 static void
 546 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 547                              struct record_opts *opts)
 548 {
 549         bool function = perf_evsel__is_function_event(evsel);
 550         struct perf_event_attr *attr = &evsel->attr;
 551
 552         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 553
 554         if (callchain_param.record_mode == CALLCHAIN_LBR) {
 555                 if (!opts->branch_stack) {
 556                         if (attr->exclude_user) {
 557                                 pr_warning("LBR callstack option is only available "
 558                                            "to get user callchain information. "
 559                                            "Falling back to framepointers.\n");
 560                         } else {
 561                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 562                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 563                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 564                         }
 565                 } else
 566                          pr_warning("Cannot use LBR callstack with branch stack. "
 567                                     "Falling back to framepointers.\n");
 568         }
 569
 570         if (callchain_param.record_mode == CALLCHAIN_DWARF) {
 571                 if (!function) {
 572                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 573                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 574                         attr->sample_regs_user = PERF_REGS_MASK;
 575                         attr->sample_stack_user = callchain_param.dump_size;
 576                         attr->exclude_callchain_user = 1;
 577                 } else {
 578                         pr_info("Cannot use DWARF unwind for function trace event,"
 579                                 " falling back to framepointers.\n");
 580                 }
 581         }
 582
 583         if (function) {
 584                 pr_info("Disabling user space callchains for function trace event.\n");
 585                 attr->exclude_callchain_user = 1;
 586         }
 587 }
 588
 589 /*
 590  * The enable_on_exec/disabled value strategy:
 591  *
 592  *  1) For any type of traced program:
 593  *    - all independent events and group leaders are disabled
 594  *    - all group members are enabled
 595  *
 596  *     Group members are ruled by group leaders. They need to
 597  *     be enabled, because the group scheduling relies on that.
 598  *
 599  *  2) For traced programs executed by perf:
 600  *     - all independent events and group leaders have
 601  *       enable_on_exec set
 602  *     - we don't specifically enable or disable any event during
 603  *       the record command
 604  *
 605  *     Independent events and group leaders are initially disabled
 606  *     and get enabled by exec. Group members are ruled by group
 607  *     leaders as stated in 1).
 608  *
 609  *  3) For traced programs attached by perf (pid/tid):
 610  *     - we specifically enable or disable all events during
 611  *       the record command
 612  *
 613  *     When attaching events to already running traced we
 614  *     enable/disable events specifically, as there's no
 615  *     initial traced exec call.
 616  */
 617 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 618 {
 619         struct perf_evsel *leader = evsel->leader;
 620         struct perf_event_attr *attr = &evsel->attr;
 621         int track = evsel->tracking;
 622         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 623
 624         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 625         attr->inherit       = !opts->no_inherit;
 626
 627         perf_evsel__set_sample_bit(evsel, IP);
 628         perf_evsel__set_sample_bit(evsel, TID);
 629
 630         if (evsel->sample_read) {
 631                 perf_evsel__set_sample_bit(evsel, READ);
 632
 633                 /*
 634                  * We need ID even in case of single event, because
 635                  * PERF_SAMPLE_READ process ID specific data.
 636                  */
 637                 perf_evsel__set_sample_id(evsel, false);
 638
 639                 /*
 640                  * Apply group format only if we belong to group
 641                  * with more than one members.
 642                  */
 643                 if (leader->nr_members > 1) {
 644                         attr->read_format |= PERF_FORMAT_GROUP;
 645                         attr->inherit = 0;
 646                 }
 647         }
 648
 649         /*
 650          * We default some events to have a default interval. But keep
 651          * it a weak assumption overridable by the user.
 652          */
 653         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 654                                      opts->user_interval != ULLONG_MAX)) {
 655                 if (opts->freq) {
 656                         perf_evsel__set_sample_bit(evsel, PERIOD);
 657                         attr->freq              = 1;
 658                         attr->sample_freq       = opts->freq;
 659                 } else {
 660                         attr->sample_period = opts->default_interval;
 661                 }
 662         }
 663
 664         /*
 665          * Disable sampling for all group members other
 666          * than leader in case leader 'leads' the sampling.
 667          */
 668         if ((leader != evsel) && leader->sample_read) {
 669                 attr->sample_freq   = 0;
 670                 attr->sample_period = 0;
 671         }
 672
 673         if (opts->no_samples)
 674                 attr->sample_freq = 0;
 675
 676         if (opts->inherit_stat)
 677                 attr->inherit_stat = 1;
 678
 679         if (opts->sample_address) {
 680                 perf_evsel__set_sample_bit(evsel, ADDR);
 681                 attr->mmap_data = track;
 682         }
 683
 684         /*
 685          * We don't allow user space callchains for  function trace
 686          * event, due to issues with page faults while tracing page
 687          * fault handler and its overall trickiness nature.
 688          */
 689         if (perf_evsel__is_function_event(evsel))
 690                 evsel->attr.exclude_callchain_user = 1;
 691
 692         if (callchain_param.enabled && !evsel->no_aux_samples)
 693                 perf_evsel__config_callgraph(evsel, opts);
 694
 695         if (opts->sample_intr_regs) {
 696                 attr->sample_regs_intr = PERF_REGS_MASK;
 697                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 698         }
 699
 700         if (target__has_cpu(&opts->target))
 701                 perf_evsel__set_sample_bit(evsel, CPU);
 702
 703         if (opts->period)
 704                 perf_evsel__set_sample_bit(evsel, PERIOD);
 705
 706         /*
 707          * When the user explicitely disabled time don't force it here.
 708          */
 709         if (opts->sample_time &&
 710             (!perf_missing_features.sample_id_all &&
 711             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 712              opts->sample_time_set)))
 713                 perf_evsel__set_sample_bit(evsel, TIME);
 714
 715         if (opts->raw_samples && !evsel->no_aux_samples) {
 716                 perf_evsel__set_sample_bit(evsel, TIME);
 717                 perf_evsel__set_sample_bit(evsel, RAW);
 718                 perf_evsel__set_sample_bit(evsel, CPU);
 719         }
 720
 721         if (opts->sample_address)
 722                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 723
 724         if (opts->no_buffering) {
 725                 attr->watermark = 0;
 726                 attr->wakeup_events = 1;
 727         }
 728         if (opts->branch_stack && !evsel->no_aux_samples) {
 729                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 730                 attr->branch_sample_type = opts->branch_stack;
 731         }
 732
 733         if (opts->sample_weight)
 734                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 735
 736         attr->task  = track;
 737         attr->mmap  = track;
 738         attr->mmap2 = track && !perf_missing_features.mmap2;
 739         attr->comm  = track;
 740
 741         if (opts->sample_transaction)
 742                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 743
 744         if (opts->running_time) {
 745                 evsel->attr.read_format |=
 746                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 747                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 748         }
 749
 750         /*
 751          * XXX see the function comment above
 752          *
 753          * Disabling only independent events or group leaders,
 754          * keeping group members enabled.
 755          */
 756         if (perf_evsel__is_group_leader(evsel))
 757                 attr->disabled = 1;
 758
 759         /*
 760          * Setting enable_on_exec for independent events and
 761          * group leaders for traced executed by perf.
 762          */
 763         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 764                 !opts->initial_delay)
 765                 attr->enable_on_exec = 1;
 766
 767         if (evsel->immediate) {
 768                 attr->disabled = 0;
 769                 attr->enable_on_exec = 0;
 770         }
 771
 772         clockid = opts->clockid;
 773         if (opts->use_clockid) {
 774                 attr->use_clockid = 1;
 775                 attr->clockid = opts->clockid;
 776         }
 777 }
 778
 779 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 780 {
 781         int cpu, thread;
 782
 783         if (evsel->system_wide)
 784                 nthreads = 1;
 785
 786         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 787
 788         if (evsel->fd) {
 789                 for (cpu = 0; cpu < ncpus; cpu++) {
 790                         for (thread = 0; thread < nthreads; thread++) {
 791                                 FD(evsel, cpu, thread) = -1;
 792                         }
 793                 }
 794         }
 795
 796         return evsel->fd != NULL ? 0 : -ENOMEM;
 797 }
 798
 799 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 800                           int ioc,  void *arg)
 801 {
 802         int cpu, thread;
 803
 804         if (evsel->system_wide)
 805                 nthreads = 1;
 806
 807         for (cpu = 0; cpu < ncpus; cpu++) {
 808                 for (thread = 0; thread < nthreads; thread++) {
 809                         int fd = FD(evsel, cpu, thread),
 810                             err = ioctl(fd, ioc, arg);
 811
 812                         if (err)
 813                                 return err;
 814                 }
 815         }
 816
 817         return 0;
 818 }
 819
 820 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 821                              const char *filter)
 822 {
 823         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 824                                      PERF_EVENT_IOC_SET_FILTER,
 825                                      (void *)filter);
 826 }
 827
 828 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 829 {
 830         char *new_filter = strdup(filter);
 831
 832         if (new_filter != NULL) {
 833                 free(evsel->filter);
 834                 evsel->filter = new_filter;
 835                 return 0;
 836         }
 837
 838         return -1;
 839 }
 840
 841 int perf_evsel__append_filter(struct perf_evsel *evsel,
 842                               const char *op, const char *filter)
 843 {
 844         char *new_filter;
 845
 846         if (evsel->filter == NULL)
 847                 return perf_evsel__set_filter(evsel, filter);
 848
 849         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 850                 free(evsel->filter);
 851                 evsel->filter = new_filter;
 852                 return 0;
 853         }
 854
 855         return -1;
 856 }
 857
 858 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 859 {
 860         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 861                                      PERF_EVENT_IOC_ENABLE,
 862                                      0);
 863 }
 864
 865 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 866 {
 867         if (ncpus == 0 || nthreads == 0)
 868                 return 0;
 869
 870         if (evsel->system_wide)
 871                 nthreads = 1;
 872
 873         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 874         if (evsel->sample_id == NULL)
 875                 return -ENOMEM;
 876
 877         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 878         if (evsel->id == NULL) {
 879                 xyarray__delete(evsel->sample_id);
 880                 evsel->sample_id = NULL;
 881                 return -ENOMEM;
 882         }
 883
 884         return 0;
 885 }
 886
 887 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 888 {
 889         xyarray__delete(evsel->fd);
 890         evsel->fd = NULL;
 891 }
 892
 893 static void perf_evsel__free_id(struct perf_evsel *evsel)
 894 {
 895         xyarray__delete(evsel->sample_id);
 896         evsel->sample_id = NULL;
 897         zfree(&evsel->id);
 898 }
 899
 900 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 901 {
 902         int cpu, thread;
 903
 904         if (evsel->system_wide)
 905                 nthreads = 1;
 906
 907         for (cpu = 0; cpu < ncpus; cpu++)
 908                 for (thread = 0; thread < nthreads; ++thread) {
 909                         close(FD(evsel, cpu, thread));
 910                         FD(evsel, cpu, thread) = -1;
 911                 }
 912 }
 913
 914 void perf_evsel__exit(struct perf_evsel *evsel)
 915 {
 916         assert(list_empty(&evsel->node));
 917         perf_evsel__free_fd(evsel);
 918         perf_evsel__free_id(evsel);
 919         close_cgroup(evsel->cgrp);
 920         cpu_map__put(evsel->cpus);
 921         thread_map__put(evsel->threads);
 922         zfree(&evsel->group_name);
 923         zfree(&evsel->name);
 924         perf_evsel__object.fini(evsel);
 925 }
 926
 927 void perf_evsel__delete(struct perf_evsel *evsel)
 928 {
 929         perf_evsel__exit(evsel);
 930         free(evsel);
 931 }
 932
 933 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
 934                                 struct perf_counts_values *count)
 935 {
 936         struct perf_counts_values tmp;
 937
 938         if (!evsel->prev_raw_counts)
 939                 return;
 940
 941         if (cpu == -1) {
 942                 tmp = evsel->prev_raw_counts->aggr;
 943                 evsel->prev_raw_counts->aggr = *count;
 944         } else {
 945                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
 946                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
 947         }
 948
 949         count->val = count->val - tmp.val;
 950         count->ena = count->ena - tmp.ena;
 951         count->run = count->run - tmp.run;
 952 }
 953
 954 void perf_counts_values__scale(struct perf_counts_values *count,
 955                                bool scale, s8 *pscaled)
 956 {
 957         s8 scaled = 0;
 958
 959         if (scale) {
 960                 if (count->run == 0) {
 961                         scaled = -1;
 962                         count->val = 0;
 963                 } else if (count->run < count->ena) {
 964                         scaled = 1;
 965                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
 966                 }
 967         } else
 968                 count->ena = count->run = 0;
 969
 970         if (pscaled)
 971                 *pscaled = scaled;
 972 }
 973
 974 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
 975                      struct perf_counts_values *count)
 976 {
 977         memset(count, 0, sizeof(*count));
 978
 979         if (FD(evsel, cpu, thread) < 0)
 980                 return -EINVAL;
 981
 982         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
 983                 return -errno;
 984
 985         return 0;
 986 }
 987
 988 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 989                               int cpu, int thread, bool scale)
 990 {
 991         struct perf_counts_values count;
 992         size_t nv = scale ? 3 : 1;
 993
 994         if (FD(evsel, cpu, thread) < 0)
 995                 return -EINVAL;
 996
 997         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
 998                 return -ENOMEM;
 999
1000         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1001                 return -errno;
1002
1003         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1004         perf_counts_values__scale(&count, scale, NULL);
1005         *perf_counts(evsel->counts, cpu, thread) = count;
1006         return 0;
1007 }
1008
1009 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1010 {
1011         struct perf_evsel *leader = evsel->leader;
1012         int fd;
1013
1014         if (perf_evsel__is_group_leader(evsel))
1015                 return -1;
1016
1017         /*
1018          * Leader must be already processed/open,
1019          * if not it's a bug.
1020          */
1021         BUG_ON(!leader->fd);
1022
1023         fd = FD(leader, cpu, thread);
1024         BUG_ON(fd == -1);
1025
1026         return fd;
1027 }
1028
1029 struct bit_names {
1030         int bit;
1031         const char *name;
1032 };
1033
1034 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1035 {
1036         bool first_bit = true;
1037         int i = 0;
1038
1039         do {
1040                 if (value & bits[i].bit) {
1041                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1042                         first_bit = false;
1043                 }
1044         } while (bits[++i].name != NULL);
1045 }
1046
1047 static void __p_sample_type(char *buf, size_t size, u64 value)
1048 {
1049 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1050         struct bit_names bits[] = {
1051                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1052                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1053                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1054                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1055                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1056                 { .name = NULL, }
1057         };
1058 #undef bit_name
1059         __p_bits(buf, size, value, bits);
1060 }
1061
1062 static void __p_read_format(char *buf, size_t size, u64 value)
1063 {
1064 #define bit_name(n) { PERF_FORMAT_##n, #n }
1065         struct bit_names bits[] = {
1066                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1067                 bit_name(ID), bit_name(GROUP),
1068                 { .name = NULL, }
1069         };
1070 #undef bit_name
1071         __p_bits(buf, size, value, bits);
1072 }
1073
1074 #define BUF_SIZE                1024
1075
1076 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1077 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1078 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1079 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1080 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1081
1082 #define PRINT_ATTRn(_n, _f, _p)                         \
1083 do {                                                    \
1084         if (attr->_f) {                                 \
1085                 _p(attr->_f);                           \
1086                 ret += attr__fprintf(fp, _n, buf, priv);\
1087         }                                               \
1088 } while (0)
1089
1090 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1091
1092 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1093                              attr__fprintf_f attr__fprintf, void *priv)
1094 {
1095         char buf[BUF_SIZE];
1096         int ret = 0;
1097
1098         PRINT_ATTRf(type, p_unsigned);
1099         PRINT_ATTRf(size, p_unsigned);
1100         PRINT_ATTRf(config, p_hex);
1101         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1102         PRINT_ATTRf(sample_type, p_sample_type);
1103         PRINT_ATTRf(read_format, p_read_format);
1104
1105         PRINT_ATTRf(disabled, p_unsigned);
1106         PRINT_ATTRf(inherit, p_unsigned);
1107         PRINT_ATTRf(pinned, p_unsigned);
1108         PRINT_ATTRf(exclusive, p_unsigned);
1109         PRINT_ATTRf(exclude_user, p_unsigned);
1110         PRINT_ATTRf(exclude_kernel, p_unsigned);
1111         PRINT_ATTRf(exclude_hv, p_unsigned);
1112         PRINT_ATTRf(exclude_idle, p_unsigned);
1113         PRINT_ATTRf(mmap, p_unsigned);
1114         PRINT_ATTRf(comm, p_unsigned);
1115         PRINT_ATTRf(freq, p_unsigned);
1116         PRINT_ATTRf(inherit_stat, p_unsigned);
1117         PRINT_ATTRf(enable_on_exec, p_unsigned);
1118         PRINT_ATTRf(task, p_unsigned);
1119         PRINT_ATTRf(watermark, p_unsigned);
1120         PRINT_ATTRf(precise_ip, p_unsigned);
1121         PRINT_ATTRf(mmap_data, p_unsigned);
1122         PRINT_ATTRf(sample_id_all, p_unsigned);
1123         PRINT_ATTRf(exclude_host, p_unsigned);
1124         PRINT_ATTRf(exclude_guest, p_unsigned);
1125         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1126         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1127         PRINT_ATTRf(mmap2, p_unsigned);
1128         PRINT_ATTRf(comm_exec, p_unsigned);
1129         PRINT_ATTRf(use_clockid, p_unsigned);
1130
1131         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1132         PRINT_ATTRf(bp_type, p_unsigned);
1133         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1134         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1135         PRINT_ATTRf(sample_regs_user, p_hex);
1136         PRINT_ATTRf(sample_stack_user, p_unsigned);
1137         PRINT_ATTRf(clockid, p_signed);
1138         PRINT_ATTRf(sample_regs_intr, p_hex);
1139         PRINT_ATTRf(aux_watermark, p_unsigned);
1140
1141         return ret;
1142 }
1143
1144 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1145                                 void *priv __attribute__((unused)))
1146 {
1147         return fprintf(fp, "  %-32s %s\n", name, val);
1148 }
1149
1150 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1151                               struct thread_map *threads)
1152 {
1153         int cpu, thread, nthreads;
1154         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1155         int pid = -1, err;
1156         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1157
1158         if (evsel->system_wide)
1159                 nthreads = 1;
1160         else
1161                 nthreads = threads->nr;
1162
1163         if (evsel->fd == NULL &&
1164             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1165                 return -ENOMEM;
1166
1167         if (evsel->cgrp) {
1168                 flags |= PERF_FLAG_PID_CGROUP;
1169                 pid = evsel->cgrp->fd;
1170         }
1171
1172 fallback_missing_features:
1173         if (perf_missing_features.clockid_wrong)
1174                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1175         if (perf_missing_features.clockid) {
1176                 evsel->attr.use_clockid = 0;
1177                 evsel->attr.clockid = 0;
1178         }
1179         if (perf_missing_features.cloexec)
1180                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1181         if (perf_missing_features.mmap2)
1182                 evsel->attr.mmap2 = 0;
1183         if (perf_missing_features.exclude_guest)
1184                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1185 retry_sample_id:
1186         if (perf_missing_features.sample_id_all)
1187                 evsel->attr.sample_id_all = 0;
1188
1189         if (verbose >= 2) {
1190                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1191                 fprintf(stderr, "perf_event_attr:\n");
1192                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1193                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1194         }
1195
1196         for (cpu = 0; cpu < cpus->nr; cpu++) {
1197
1198                 for (thread = 0; thread < nthreads; thread++) {
1199                         int group_fd;
1200
1201                         if (!evsel->cgrp && !evsel->system_wide)
1202                                 pid = thread_map__pid(threads, thread);
1203
1204                         group_fd = get_group_fd(evsel, cpu, thread);
1205 retry_open:
1206                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1207                                   pid, cpus->map[cpu], group_fd, flags);
1208
1209                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1210                                                                      pid,
1211                                                                      cpus->map[cpu],
1212                                                                      group_fd, flags);
1213                         if (FD(evsel, cpu, thread) < 0) {
1214                                 err = -errno;
1215                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1216                                           err);
1217                                 goto try_fallback;
1218                         }
1219                         set_rlimit = NO_CHANGE;
1220
1221                         /*
1222                          * If we succeeded but had to kill clockid, fail and
1223                          * have perf_evsel__open_strerror() print us a nice
1224                          * error.
1225                          */
1226                         if (perf_missing_features.clockid ||
1227                             perf_missing_features.clockid_wrong) {
1228                                 err = -EINVAL;
1229                                 goto out_close;
1230                         }
1231                 }
1232         }
1233
1234         return 0;
1235
1236 try_fallback:
1237         /*
1238          * perf stat needs between 5 and 22 fds per CPU. When we run out
1239          * of them try to increase the limits.
1240          */
1241         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1242                 struct rlimit l;
1243                 int old_errno = errno;
1244
1245                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1246                         if (set_rlimit == NO_CHANGE)
1247                                 l.rlim_cur = l.rlim_max;
1248                         else {
1249                                 l.rlim_cur = l.rlim_max + 1000;
1250                                 l.rlim_max = l.rlim_cur;
1251                         }
1252                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1253                                 set_rlimit++;
1254                                 errno = old_errno;
1255                                 goto retry_open;
1256                         }
1257                 }
1258                 errno = old_errno;
1259         }
1260
1261         if (err != -EINVAL || cpu > 0 || thread > 0)
1262                 goto out_close;
1263
1264         /*
1265          * Must probe features in the order they were added to the
1266          * perf_event_attr interface.
1267          */
1268         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1269                 perf_missing_features.clockid_wrong = true;
1270                 goto fallback_missing_features;
1271         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1272                 perf_missing_features.clockid = true;
1273                 goto fallback_missing_features;
1274         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1275                 perf_missing_features.cloexec = true;
1276                 goto fallback_missing_features;
1277         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1278                 perf_missing_features.mmap2 = true;
1279                 goto fallback_missing_features;
1280         } else if (!perf_missing_features.exclude_guest &&
1281                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1282                 perf_missing_features.exclude_guest = true;
1283                 goto fallback_missing_features;
1284         } else if (!perf_missing_features.sample_id_all) {
1285                 perf_missing_features.sample_id_all = true;
1286                 goto retry_sample_id;
1287         }
1288
1289 out_close:
1290         do {
1291                 while (--thread >= 0) {
1292                         close(FD(evsel, cpu, thread));
1293                         FD(evsel, cpu, thread) = -1;
1294                 }
1295                 thread = nthreads;
1296         } while (--cpu >= 0);
1297         return err;
1298 }
1299
1300 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1301 {
1302         if (evsel->fd == NULL)
1303                 return;
1304
1305         perf_evsel__close_fd(evsel, ncpus, nthreads);
1306         perf_evsel__free_fd(evsel);
1307 }
1308
1309 static struct {
1310         struct cpu_map map;
1311         int cpus[1];
1312 } empty_cpu_map = {
1313         .map.nr = 1,
1314         .cpus   = { -1, },
1315 };
1316
1317 static struct {
1318         struct thread_map map;
1319         int threads[1];
1320 } empty_thread_map = {
1321         .map.nr  = 1,
1322         .threads = { -1, },
1323 };
1324
1325 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1326                      struct thread_map *threads)
1327 {
1328         if (cpus == NULL) {
1329                 /* Work around old compiler warnings about strict aliasing */
1330                 cpus = &empty_cpu_map.map;
1331         }
1332
1333         if (threads == NULL)
1334                 threads = &empty_thread_map.map;
1335
1336         return __perf_evsel__open(evsel, cpus, threads);
1337 }
1338
1339 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1340                              struct cpu_map *cpus)
1341 {
1342         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1343 }
1344
1345 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1346                                 struct thread_map *threads)
1347 {
1348         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1349 }
1350
1351 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1352                                        const union perf_event *event,
1353                                        struct perf_sample *sample)
1354 {
1355         u64 type = evsel->attr.sample_type;
1356         const u64 *array = event->sample.array;
1357         bool swapped = evsel->needs_swap;
1358         union u64_swap u;
1359
1360         array += ((event->header.size -
1361                    sizeof(event->header)) / sizeof(u64)) - 1;
1362
1363         if (type & PERF_SAMPLE_IDENTIFIER) {
1364                 sample->id = *array;
1365                 array--;
1366         }
1367
1368         if (type & PERF_SAMPLE_CPU) {
1369                 u.val64 = *array;
1370                 if (swapped) {
1371                         /* undo swap of u64, then swap on individual u32s */
1372                         u.val64 = bswap_64(u.val64);
1373                         u.val32[0] = bswap_32(u.val32[0]);
1374                 }
1375
1376                 sample->cpu = u.val32[0];
1377                 array--;
1378         }
1379
1380         if (type & PERF_SAMPLE_STREAM_ID) {
1381                 sample->stream_id = *array;
1382                 array--;
1383         }
1384
1385         if (type & PERF_SAMPLE_ID) {
1386                 sample->id = *array;
1387                 array--;
1388         }
1389
1390         if (type & PERF_SAMPLE_TIME) {
1391                 sample->time = *array;
1392                 array--;
1393         }
1394
1395         if (type & PERF_SAMPLE_TID) {
1396                 u.val64 = *array;
1397                 if (swapped) {
1398                         /* undo swap of u64, then swap on individual u32s */
1399                         u.val64 = bswap_64(u.val64);
1400                         u.val32[0] = bswap_32(u.val32[0]);
1401                         u.val32[1] = bswap_32(u.val32[1]);
1402                 }
1403
1404                 sample->pid = u.val32[0];
1405                 sample->tid = u.val32[1];
1406                 array--;
1407         }
1408
1409         return 0;
1410 }
1411
1412 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1413                             u64 size)
1414 {
1415         return size > max_size || offset + size > endp;
1416 }
1417
1418 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1419         do {                                                            \
1420                 if (overflow(endp, (max_size), (offset), (size)))       \
1421                         return -EFAULT;                                 \
1422         } while (0)
1423
1424 #define OVERFLOW_CHECK_u64(offset) \
1425         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1426
1427 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1428                              struct perf_sample *data)
1429 {
1430         u64 type = evsel->attr.sample_type;
1431         bool swapped = evsel->needs_swap;
1432         const u64 *array;
1433         u16 max_size = event->header.size;
1434         const void *endp = (void *)event + max_size;
1435         u64 sz;
1436
1437         /*
1438          * used for cross-endian analysis. See git commit 65014ab3
1439          * for why this goofiness is needed.
1440          */
1441         union u64_swap u;
1442
1443         memset(data, 0, sizeof(*data));
1444         data->cpu = data->pid = data->tid = -1;
1445         data->stream_id = data->id = data->time = -1ULL;
1446         data->period = evsel->attr.sample_period;
1447         data->weight = 0;
1448
1449         if (event->header.type != PERF_RECORD_SAMPLE) {
1450                 if (!evsel->attr.sample_id_all)
1451                         return 0;
1452                 return perf_evsel__parse_id_sample(evsel, event, data);
1453         }
1454
1455         array = event->sample.array;
1456
1457         /*
1458          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1459          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1460          * check the format does not go past the end of the event.
1461          */
1462         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1463                 return -EFAULT;
1464
1465         data->id = -1ULL;
1466         if (type & PERF_SAMPLE_IDENTIFIER) {
1467                 data->id = *array;
1468                 array++;
1469         }
1470
1471         if (type & PERF_SAMPLE_IP) {
1472                 data->ip = *array;
1473                 array++;
1474         }
1475
1476         if (type & PERF_SAMPLE_TID) {
1477                 u.val64 = *array;
1478                 if (swapped) {
1479                         /* undo swap of u64, then swap on individual u32s */
1480                         u.val64 = bswap_64(u.val64);
1481                         u.val32[0] = bswap_32(u.val32[0]);
1482                         u.val32[1] = bswap_32(u.val32[1]);
1483                 }
1484
1485                 data->pid = u.val32[0];
1486                 data->tid = u.val32[1];
1487                 array++;
1488         }
1489
1490         if (type & PERF_SAMPLE_TIME) {
1491                 data->time = *array;
1492                 array++;
1493         }
1494
1495         data->addr = 0;
1496         if (type & PERF_SAMPLE_ADDR) {
1497                 data->addr = *array;
1498                 array++;
1499         }
1500
1501         if (type & PERF_SAMPLE_ID) {
1502                 data->id = *array;
1503                 array++;
1504         }
1505
1506         if (type & PERF_SAMPLE_STREAM_ID) {
1507                 data->stream_id = *array;
1508                 array++;
1509         }
1510
1511         if (type & PERF_SAMPLE_CPU) {
1512
1513                 u.val64 = *array;
1514                 if (swapped) {
1515                         /* undo swap of u64, then swap on individual u32s */
1516                         u.val64 = bswap_64(u.val64);
1517                         u.val32[0] = bswap_32(u.val32[0]);
1518                 }
1519
1520                 data->cpu = u.val32[0];
1521                 array++;
1522         }
1523
1524         if (type & PERF_SAMPLE_PERIOD) {
1525                 data->period = *array;
1526                 array++;
1527         }
1528
1529         if (type & PERF_SAMPLE_READ) {
1530                 u64 read_format = evsel->attr.read_format;
1531
1532                 OVERFLOW_CHECK_u64(array);
1533                 if (read_format & PERF_FORMAT_GROUP)
1534                         data->read.group.nr = *array;
1535                 else
1536                         data->read.one.value = *array;
1537
1538                 array++;
1539
1540                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1541                         OVERFLOW_CHECK_u64(array);
1542                         data->read.time_enabled = *array;
1543                         array++;
1544                 }
1545
1546                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1547                         OVERFLOW_CHECK_u64(array);
1548                         data->read.time_running = *array;
1549                         array++;
1550                 }
1551
1552                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1553                 if (read_format & PERF_FORMAT_GROUP) {
1554                         const u64 max_group_nr = UINT64_MAX /
1555                                         sizeof(struct sample_read_value);
1556
1557                         if (data->read.group.nr > max_group_nr)
1558                                 return -EFAULT;
1559                         sz = data->read.group.nr *
1560                              sizeof(struct sample_read_value);
1561                         OVERFLOW_CHECK(array, sz, max_size);
1562                         data->read.group.values =
1563                                         (struct sample_read_value *)array;
1564                         array = (void *)array + sz;
1565                 } else {
1566                         OVERFLOW_CHECK_u64(array);
1567                         data->read.one.id = *array;
1568                         array++;
1569                 }
1570         }
1571
1572         if (type & PERF_SAMPLE_CALLCHAIN) {
1573                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1574
1575                 OVERFLOW_CHECK_u64(array);
1576                 data->callchain = (struct ip_callchain *)array++;
1577                 if (data->callchain->nr > max_callchain_nr)
1578                         return -EFAULT;
1579                 sz = data->callchain->nr * sizeof(u64);
1580                 OVERFLOW_CHECK(array, sz, max_size);
1581                 array = (void *)array + sz;
1582         }
1583
1584         if (type & PERF_SAMPLE_RAW) {
1585                 OVERFLOW_CHECK_u64(array);
1586                 u.val64 = *array;
1587                 if (WARN_ONCE(swapped,
1588                               "Endianness of raw data not corrected!\n")) {
1589                         /* undo swap of u64, then swap on individual u32s */
1590                         u.val64 = bswap_64(u.val64);
1591                         u.val32[0] = bswap_32(u.val32[0]);
1592                         u.val32[1] = bswap_32(u.val32[1]);
1593                 }
1594                 data->raw_size = u.val32[0];
1595                 array = (void *)array + sizeof(u32);
1596
1597                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1598                 data->raw_data = (void *)array;
1599                 array = (void *)array + data->raw_size;
1600         }
1601
1602         if (type & PERF_SAMPLE_BRANCH_STACK) {
1603                 const u64 max_branch_nr = UINT64_MAX /
1604                                           sizeof(struct branch_entry);
1605
1606                 OVERFLOW_CHECK_u64(array);
1607                 data->branch_stack = (struct branch_stack *)array++;
1608
1609                 if (data->branch_stack->nr > max_branch_nr)
1610                         return -EFAULT;
1611                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1612                 OVERFLOW_CHECK(array, sz, max_size);
1613                 array = (void *)array + sz;
1614         }
1615
1616         if (type & PERF_SAMPLE_REGS_USER) {
1617                 OVERFLOW_CHECK_u64(array);
1618                 data->user_regs.abi = *array;
1619                 array++;
1620
1621                 if (data->user_regs.abi) {
1622                         u64 mask = evsel->attr.sample_regs_user;
1623
1624                         sz = hweight_long(mask) * sizeof(u64);
1625                         OVERFLOW_CHECK(array, sz, max_size);
1626                         data->user_regs.mask = mask;
1627                         data->user_regs.regs = (u64 *)array;
1628                         array = (void *)array + sz;
1629                 }
1630         }
1631
1632         if (type & PERF_SAMPLE_STACK_USER) {
1633                 OVERFLOW_CHECK_u64(array);
1634                 sz = *array++;
1635
1636                 data->user_stack.offset = ((char *)(array - 1)
1637                                           - (char *) event);
1638
1639                 if (!sz) {
1640                         data->user_stack.size = 0;
1641                 } else {
1642                         OVERFLOW_CHECK(array, sz, max_size);
1643                         data->user_stack.data = (char *)array;
1644                         array = (void *)array + sz;
1645                         OVERFLOW_CHECK_u64(array);
1646                         data->user_stack.size = *array++;
1647                         if (WARN_ONCE(data->user_stack.size > sz,
1648                                       "user stack dump failure\n"))
1649                                 return -EFAULT;
1650                 }
1651         }
1652
1653         data->weight = 0;
1654         if (type & PERF_SAMPLE_WEIGHT) {
1655                 OVERFLOW_CHECK_u64(array);
1656                 data->weight = *array;
1657                 array++;
1658         }
1659
1660         data->data_src = PERF_MEM_DATA_SRC_NONE;
1661         if (type & PERF_SAMPLE_DATA_SRC) {
1662                 OVERFLOW_CHECK_u64(array);
1663                 data->data_src = *array;
1664                 array++;
1665         }
1666
1667         data->transaction = 0;
1668         if (type & PERF_SAMPLE_TRANSACTION) {
1669                 OVERFLOW_CHECK_u64(array);
1670                 data->transaction = *array;
1671                 array++;
1672         }
1673
1674         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1675         if (type & PERF_SAMPLE_REGS_INTR) {
1676                 OVERFLOW_CHECK_u64(array);
1677                 data->intr_regs.abi = *array;
1678                 array++;
1679
1680                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1681                         u64 mask = evsel->attr.sample_regs_intr;
1682
1683                         sz = hweight_long(mask) * sizeof(u64);
1684                         OVERFLOW_CHECK(array, sz, max_size);
1685                         data->intr_regs.mask = mask;
1686                         data->intr_regs.regs = (u64 *)array;
1687                         array = (void *)array + sz;
1688                 }
1689         }
1690
1691         return 0;
1692 }
1693
1694 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1695                                      u64 read_format)
1696 {
1697         size_t sz, result = sizeof(struct sample_event);
1698
1699         if (type & PERF_SAMPLE_IDENTIFIER)
1700                 result += sizeof(u64);
1701
1702         if (type & PERF_SAMPLE_IP)
1703                 result += sizeof(u64);
1704
1705         if (type & PERF_SAMPLE_TID)
1706                 result += sizeof(u64);
1707
1708         if (type & PERF_SAMPLE_TIME)
1709                 result += sizeof(u64);
1710
1711         if (type & PERF_SAMPLE_ADDR)
1712                 result += sizeof(u64);
1713
1714         if (type & PERF_SAMPLE_ID)
1715                 result += sizeof(u64);
1716
1717         if (type & PERF_SAMPLE_STREAM_ID)
1718                 result += sizeof(u64);
1719
1720         if (type & PERF_SAMPLE_CPU)
1721                 result += sizeof(u64);
1722
1723         if (type & PERF_SAMPLE_PERIOD)
1724                 result += sizeof(u64);
1725
1726         if (type & PERF_SAMPLE_READ) {
1727                 result += sizeof(u64);
1728                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1729                         result += sizeof(u64);
1730                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1731                         result += sizeof(u64);
1732                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1733                 if (read_format & PERF_FORMAT_GROUP) {
1734                         sz = sample->read.group.nr *
1735                              sizeof(struct sample_read_value);
1736                         result += sz;
1737                 } else {
1738                         result += sizeof(u64);
1739                 }
1740         }
1741
1742         if (type & PERF_SAMPLE_CALLCHAIN) {
1743                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1744                 result += sz;
1745         }
1746
1747         if (type & PERF_SAMPLE_RAW) {
1748                 result += sizeof(u32);
1749                 result += sample->raw_size;
1750         }
1751
1752         if (type & PERF_SAMPLE_BRANCH_STACK) {
1753                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1754                 sz += sizeof(u64);
1755                 result += sz;
1756         }
1757
1758         if (type & PERF_SAMPLE_REGS_USER) {
1759                 if (sample->user_regs.abi) {
1760                         result += sizeof(u64);
1761                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1762                         result += sz;
1763                 } else {
1764                         result += sizeof(u64);
1765                 }
1766         }
1767
1768         if (type & PERF_SAMPLE_STACK_USER) {
1769                 sz = sample->user_stack.size;
1770                 result += sizeof(u64);
1771                 if (sz) {
1772                         result += sz;
1773                         result += sizeof(u64);
1774                 }
1775         }
1776
1777         if (type & PERF_SAMPLE_WEIGHT)
1778                 result += sizeof(u64);
1779
1780         if (type & PERF_SAMPLE_DATA_SRC)
1781                 result += sizeof(u64);
1782
1783         if (type & PERF_SAMPLE_TRANSACTION)
1784                 result += sizeof(u64);
1785
1786         if (type & PERF_SAMPLE_REGS_INTR) {
1787                 if (sample->intr_regs.abi) {
1788                         result += sizeof(u64);
1789                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1790                         result += sz;
1791                 } else {
1792                         result += sizeof(u64);
1793                 }
1794         }
1795
1796         return result;
1797 }
1798
1799 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1800                                   u64 read_format,
1801                                   const struct perf_sample *sample,
1802                                   bool swapped)
1803 {
1804         u64 *array;
1805         size_t sz;
1806         /*
1807          * used for cross-endian analysis. See git commit 65014ab3
1808          * for why this goofiness is needed.
1809          */
1810         union u64_swap u;
1811
1812         array = event->sample.array;
1813
1814         if (type & PERF_SAMPLE_IDENTIFIER) {
1815                 *array = sample->id;
1816                 array++;
1817         }
1818
1819         if (type & PERF_SAMPLE_IP) {
1820                 *array = sample->ip;
1821                 array++;
1822         }
1823
1824         if (type & PERF_SAMPLE_TID) {
1825                 u.val32[0] = sample->pid;
1826                 u.val32[1] = sample->tid;
1827                 if (swapped) {
1828                         /*
1829                          * Inverse of what is done in perf_evsel__parse_sample
1830                          */
1831                         u.val32[0] = bswap_32(u.val32[0]);
1832                         u.val32[1] = bswap_32(u.val32[1]);
1833                         u.val64 = bswap_64(u.val64);
1834                 }
1835
1836                 *array = u.val64;
1837                 array++;
1838         }
1839
1840         if (type & PERF_SAMPLE_TIME) {
1841                 *array = sample->time;
1842                 array++;
1843         }
1844
1845         if (type & PERF_SAMPLE_ADDR) {
1846                 *array = sample->addr;
1847                 array++;
1848         }
1849
1850         if (type & PERF_SAMPLE_ID) {
1851                 *array = sample->id;
1852                 array++;
1853         }
1854
1855         if (type & PERF_SAMPLE_STREAM_ID) {
1856                 *array = sample->stream_id;
1857                 array++;
1858         }
1859
1860         if (type & PERF_SAMPLE_CPU) {
1861                 u.val32[0] = sample->cpu;
1862                 if (swapped) {
1863                         /*
1864                          * Inverse of what is done in perf_evsel__parse_sample
1865                          */
1866                         u.val32[0] = bswap_32(u.val32[0]);
1867                         u.val64 = bswap_64(u.val64);
1868                 }
1869                 *array = u.val64;
1870                 array++;
1871         }
1872
1873         if (type & PERF_SAMPLE_PERIOD) {
1874                 *array = sample->period;
1875                 array++;
1876         }
1877
1878         if (type & PERF_SAMPLE_READ) {
1879                 if (read_format & PERF_FORMAT_GROUP)
1880                         *array = sample->read.group.nr;
1881                 else
1882                         *array = sample->read.one.value;
1883                 array++;
1884
1885                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1886                         *array = sample->read.time_enabled;
1887                         array++;
1888                 }
1889
1890                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1891                         *array = sample->read.time_running;
1892                         array++;
1893                 }
1894
1895                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1896                 if (read_format & PERF_FORMAT_GROUP) {
1897                         sz = sample->read.group.nr *
1898                              sizeof(struct sample_read_value);
1899                         memcpy(array, sample->read.group.values, sz);
1900                         array = (void *)array + sz;
1901                 } else {
1902                         *array = sample->read.one.id;
1903                         array++;
1904                 }
1905         }
1906
1907         if (type & PERF_SAMPLE_CALLCHAIN) {
1908                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1909                 memcpy(array, sample->callchain, sz);
1910                 array = (void *)array + sz;
1911         }
1912
1913         if (type & PERF_SAMPLE_RAW) {
1914                 u.val32[0] = sample->raw_size;
1915                 if (WARN_ONCE(swapped,
1916                               "Endianness of raw data not corrected!\n")) {
1917                         /*
1918                          * Inverse of what is done in perf_evsel__parse_sample
1919                          */
1920                         u.val32[0] = bswap_32(u.val32[0]);
1921                         u.val32[1] = bswap_32(u.val32[1]);
1922                         u.val64 = bswap_64(u.val64);
1923                 }
1924                 *array = u.val64;
1925                 array = (void *)array + sizeof(u32);
1926
1927                 memcpy(array, sample->raw_data, sample->raw_size);
1928                 array = (void *)array + sample->raw_size;
1929         }
1930
1931         if (type & PERF_SAMPLE_BRANCH_STACK) {
1932                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1933                 sz += sizeof(u64);
1934                 memcpy(array, sample->branch_stack, sz);
1935                 array = (void *)array + sz;
1936         }
1937
1938         if (type & PERF_SAMPLE_REGS_USER) {
1939                 if (sample->user_regs.abi) {
1940                         *array++ = sample->user_regs.abi;
1941                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1942                         memcpy(array, sample->user_regs.regs, sz);
1943                         array = (void *)array + sz;
1944                 } else {
1945                         *array++ = 0;
1946                 }
1947         }
1948
1949         if (type & PERF_SAMPLE_STACK_USER) {
1950                 sz = sample->user_stack.size;
1951                 *array++ = sz;
1952                 if (sz) {
1953                         memcpy(array, sample->user_stack.data, sz);
1954                         array = (void *)array + sz;
1955                         *array++ = sz;
1956                 }
1957         }
1958
1959         if (type & PERF_SAMPLE_WEIGHT) {
1960                 *array = sample->weight;
1961                 array++;
1962         }
1963
1964         if (type & PERF_SAMPLE_DATA_SRC) {
1965                 *array = sample->data_src;
1966                 array++;
1967         }
1968
1969         if (type & PERF_SAMPLE_TRANSACTION) {
1970                 *array = sample->transaction;
1971                 array++;
1972         }
1973
1974         if (type & PERF_SAMPLE_REGS_INTR) {
1975                 if (sample->intr_regs.abi) {
1976                         *array++ = sample->intr_regs.abi;
1977                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1978                         memcpy(array, sample->intr_regs.regs, sz);
1979                         array = (void *)array + sz;
1980                 } else {
1981                         *array++ = 0;
1982                 }
1983         }
1984
1985         return 0;
1986 }
1987
1988 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1989 {
1990         return pevent_find_field(evsel->tp_format, name);
1991 }
1992
1993 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1994                          const char *name)
1995 {
1996         struct format_field *field = perf_evsel__field(evsel, name);
1997         int offset;
1998
1999         if (!field)
2000                 return NULL;
2001
2002         offset = field->offset;
2003
2004         if (field->flags & FIELD_IS_DYNAMIC) {
2005                 offset = *(int *)(sample->raw_data + field->offset);
2006                 offset &= 0xffff;
2007         }
2008
2009         return sample->raw_data + offset;
2010 }
2011
2012 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2013                        const char *name)
2014 {
2015         struct format_field *field = perf_evsel__field(evsel, name);
2016         void *ptr;
2017         u64 value;
2018
2019         if (!field)
2020                 return 0;
2021
2022         ptr = sample->raw_data + field->offset;
2023
2024         switch (field->size) {
2025         case 1:
2026                 return *(u8 *)ptr;
2027         case 2:
2028                 value = *(u16 *)ptr;
2029                 break;
2030         case 4:
2031                 value = *(u32 *)ptr;
2032                 break;
2033         case 8:
2034                 memcpy(&value, ptr, sizeof(u64));
2035                 break;
2036         default:
2037                 return 0;
2038         }
2039
2040         if (!evsel->needs_swap)
2041                 return value;
2042
2043         switch (field->size) {
2044         case 2:
2045                 return bswap_16(value);
2046         case 4:
2047                 return bswap_32(value);
2048         case 8:
2049                 return bswap_64(value);
2050         default:
2051                 return 0;
2052         }
2053
2054         return 0;
2055 }
2056
2057 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2058 {
2059         va_list args;
2060         int ret = 0;
2061
2062         if (!*first) {
2063                 ret += fprintf(fp, ",");
2064         } else {
2065                 ret += fprintf(fp, ":");
2066                 *first = false;
2067         }
2068
2069         va_start(args, fmt);
2070         ret += vfprintf(fp, fmt, args);
2071         va_end(args);
2072         return ret;
2073 }
2074
2075 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2076 {
2077         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2078 }
2079
2080 int perf_evsel__fprintf(struct perf_evsel *evsel,
2081                         struct perf_attr_details *details, FILE *fp)
2082 {
2083         bool first = true;
2084         int printed = 0;
2085
2086         if (details->event_group) {
2087                 struct perf_evsel *pos;
2088
2089                 if (!perf_evsel__is_group_leader(evsel))
2090                         return 0;
2091
2092                 if (evsel->nr_members > 1)
2093                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2094
2095                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2096                 for_each_group_member(pos, evsel)
2097                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2098
2099                 if (evsel->nr_members > 1)
2100                         printed += fprintf(fp, "}");
2101                 goto out;
2102         }
2103
2104         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2105
2106         if (details->verbose) {
2107                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2108                                                     __print_attr__fprintf, &first);
2109         } else if (details->freq) {
2110                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2111                                          (u64)evsel->attr.sample_freq);
2112         }
2113 out:
2114         fputc('\n', fp);
2115         return ++printed;
2116 }
2117
2118 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2119                           char *msg, size_t msgsize)
2120 {
2121         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2122             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2123             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2124                 /*
2125                  * If it's cycles then fall back to hrtimer based
2126                  * cpu-clock-tick sw counter, which is always available even if
2127                  * no PMU support.
2128                  *
2129                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2130                  * b0a873e).
2131                  */
2132                 scnprintf(msg, msgsize, "%s",
2133 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2134
2135                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2136                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2137
2138                 zfree(&evsel->name);
2139                 return true;
2140         }
2141
2142         return false;
2143 }
2144
2145 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2146                               int err, char *msg, size_t size)
2147 {
2148         char sbuf[STRERR_BUFSIZE];
2149
2150         switch (err) {
2151         case EPERM:
2152         case EACCES:
2153                 return scnprintf(msg, size,
2154                  "You may not have permission to collect %sstats.\n"
2155                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2156                  " -1 - Not paranoid at all\n"
2157                  "  0 - Disallow raw tracepoint access for unpriv\n"
2158                  "  1 - Disallow cpu events for unpriv\n"
2159                  "  2 - Disallow kernel profiling for unpriv",
2160                                  target->system_wide ? "system-wide " : "");
2161         case ENOENT:
2162                 return scnprintf(msg, size, "The %s event is not supported.",
2163                                  perf_evsel__name(evsel));
2164         case EMFILE:
2165                 return scnprintf(msg, size, "%s",
2166                          "Too many events are opened.\n"
2167                          "Probably the maximum number of open file descriptors has been reached.\n"
2168                          "Hint: Try again after reducing the number of events.\n"
2169                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2170         case ENODEV:
2171                 if (target->cpu_list)
2172                         return scnprintf(msg, size, "%s",
2173          "No such device - did you specify an out-of-range profile CPU?\n");
2174                 break;
2175         case EOPNOTSUPP:
2176                 if (evsel->attr.precise_ip)
2177                         return scnprintf(msg, size, "%s",
2178         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2179 #if defined(__i386__) || defined(__x86_64__)
2180                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2181                         return scnprintf(msg, size, "%s",
2182         "No hardware sampling interrupt available.\n"
2183         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2184 #endif
2185                 break;
2186         case EBUSY:
2187                 if (find_process("oprofiled"))
2188                         return scnprintf(msg, size,
2189         "The PMU counters are busy/taken by another profiler.\n"
2190         "We found oprofile daemon running, please stop it and try again.");
2191                 break;
2192         case EINVAL:
2193                 if (perf_missing_features.clockid)
2194                         return scnprintf(msg, size, "clockid feature not supported.");
2195                 if (perf_missing_features.clockid_wrong)
2196                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2197                 break;
2198         default:
2199                 break;
2200         }
2201
2202         return scnprintf(msg, size,
2203         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2204         "/bin/dmesg may provide additional information.\n"
2205         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2206                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2207                          perf_evsel__name(evsel));
2208 }