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__append_filter(struct perf_evsel *evsel,
 840                               const char *op, const char *filter)
 841 {
 842         char *new_filter;
 843
 844         if (evsel->filter == NULL)
 845                 return perf_evsel__set_filter(evsel, filter);
 846
 847         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 848                 free(evsel->filter);
 849                 evsel->filter = new_filter;
 850                 return 0;
 851         }
 852
 853         return -1;
 854 }
 855
 856 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 857 {
 858         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 859                                      PERF_EVENT_IOC_ENABLE,
 860                                      0);
 861 }
 862
 863 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 864 {
 865         if (ncpus == 0 || nthreads == 0)
 866                 return 0;
 867
 868         if (evsel->system_wide)
 869                 nthreads = 1;
 870
 871         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 872         if (evsel->sample_id == NULL)
 873                 return -ENOMEM;
 874
 875         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 876         if (evsel->id == NULL) {
 877                 xyarray__delete(evsel->sample_id);
 878                 evsel->sample_id = NULL;
 879                 return -ENOMEM;
 880         }
 881
 882         return 0;
 883 }
 884
 885 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 886 {
 887         xyarray__delete(evsel->fd);
 888         evsel->fd = NULL;
 889 }
 890
 891 static void perf_evsel__free_id(struct perf_evsel *evsel)
 892 {
 893         xyarray__delete(evsel->sample_id);
 894         evsel->sample_id = NULL;
 895         zfree(&evsel->id);
 896 }
 897
 898 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 899 {
 900         int cpu, thread;
 901
 902         if (evsel->system_wide)
 903                 nthreads = 1;
 904
 905         for (cpu = 0; cpu < ncpus; cpu++)
 906                 for (thread = 0; thread < nthreads; ++thread) {
 907                         close(FD(evsel, cpu, thread));
 908                         FD(evsel, cpu, thread) = -1;
 909                 }
 910 }
 911
 912 void perf_evsel__exit(struct perf_evsel *evsel)
 913 {
 914         assert(list_empty(&evsel->node));
 915         perf_evsel__free_fd(evsel);
 916         perf_evsel__free_id(evsel);
 917         close_cgroup(evsel->cgrp);
 918         cpu_map__put(evsel->cpus);
 919         thread_map__put(evsel->threads);
 920         zfree(&evsel->group_name);
 921         zfree(&evsel->name);
 922         perf_evsel__object.fini(evsel);
 923 }
 924
 925 void perf_evsel__delete(struct perf_evsel *evsel)
 926 {
 927         perf_evsel__exit(evsel);
 928         free(evsel);
 929 }
 930
 931 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
 932                                 struct perf_counts_values *count)
 933 {
 934         struct perf_counts_values tmp;
 935
 936         if (!evsel->prev_raw_counts)
 937                 return;
 938
 939         if (cpu == -1) {
 940                 tmp = evsel->prev_raw_counts->aggr;
 941                 evsel->prev_raw_counts->aggr = *count;
 942         } else {
 943                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
 944                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
 945         }
 946
 947         count->val = count->val - tmp.val;
 948         count->ena = count->ena - tmp.ena;
 949         count->run = count->run - tmp.run;
 950 }
 951
 952 void perf_counts_values__scale(struct perf_counts_values *count,
 953                                bool scale, s8 *pscaled)
 954 {
 955         s8 scaled = 0;
 956
 957         if (scale) {
 958                 if (count->run == 0) {
 959                         scaled = -1;
 960                         count->val = 0;
 961                 } else if (count->run < count->ena) {
 962                         scaled = 1;
 963                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
 964                 }
 965         } else
 966                 count->ena = count->run = 0;
 967
 968         if (pscaled)
 969                 *pscaled = scaled;
 970 }
 971
 972 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
 973                      struct perf_counts_values *count)
 974 {
 975         memset(count, 0, sizeof(*count));
 976
 977         if (FD(evsel, cpu, thread) < 0)
 978                 return -EINVAL;
 979
 980         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
 981                 return -errno;
 982
 983         return 0;
 984 }
 985
 986 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 987                               int cpu, int thread, bool scale)
 988 {
 989         struct perf_counts_values count;
 990         size_t nv = scale ? 3 : 1;
 991
 992         if (FD(evsel, cpu, thread) < 0)
 993                 return -EINVAL;
 994
 995         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
 996                 return -ENOMEM;
 997
 998         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 999                 return -errno;
1000
1001         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1002         perf_counts_values__scale(&count, scale, NULL);
1003         *perf_counts(evsel->counts, cpu, thread) = count;
1004         return 0;
1005 }
1006
1007 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1008 {
1009         struct perf_evsel *leader = evsel->leader;
1010         int fd;
1011
1012         if (perf_evsel__is_group_leader(evsel))
1013                 return -1;
1014
1015         /*
1016          * Leader must be already processed/open,
1017          * if not it's a bug.
1018          */
1019         BUG_ON(!leader->fd);
1020
1021         fd = FD(leader, cpu, thread);
1022         BUG_ON(fd == -1);
1023
1024         return fd;
1025 }
1026
1027 struct bit_names {
1028         int bit;
1029         const char *name;
1030 };
1031
1032 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1033 {
1034         bool first_bit = true;
1035         int i = 0;
1036
1037         do {
1038                 if (value & bits[i].bit) {
1039                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1040                         first_bit = false;
1041                 }
1042         } while (bits[++i].name != NULL);
1043 }
1044
1045 static void __p_sample_type(char *buf, size_t size, u64 value)
1046 {
1047 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1048         struct bit_names bits[] = {
1049                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1050                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1051                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1052                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1053                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1054                 { .name = NULL, }
1055         };
1056 #undef bit_name
1057         __p_bits(buf, size, value, bits);
1058 }
1059
1060 static void __p_read_format(char *buf, size_t size, u64 value)
1061 {
1062 #define bit_name(n) { PERF_FORMAT_##n, #n }
1063         struct bit_names bits[] = {
1064                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1065                 bit_name(ID), bit_name(GROUP),
1066                 { .name = NULL, }
1067         };
1068 #undef bit_name
1069         __p_bits(buf, size, value, bits);
1070 }
1071
1072 #define BUF_SIZE                1024
1073
1074 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1075 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1076 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1077 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1078 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1079
1080 #define PRINT_ATTRn(_n, _f, _p)                         \
1081 do {                                                    \
1082         if (attr->_f) {                                 \
1083                 _p(attr->_f);                           \
1084                 ret += attr__fprintf(fp, _n, buf, priv);\
1085         }                                               \
1086 } while (0)
1087
1088 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1089
1090 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1091                              attr__fprintf_f attr__fprintf, void *priv)
1092 {
1093         char buf[BUF_SIZE];
1094         int ret = 0;
1095
1096         PRINT_ATTRf(type, p_unsigned);
1097         PRINT_ATTRf(size, p_unsigned);
1098         PRINT_ATTRf(config, p_hex);
1099         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1100         PRINT_ATTRf(sample_type, p_sample_type);
1101         PRINT_ATTRf(read_format, p_read_format);
1102
1103         PRINT_ATTRf(disabled, p_unsigned);
1104         PRINT_ATTRf(inherit, p_unsigned);
1105         PRINT_ATTRf(pinned, p_unsigned);
1106         PRINT_ATTRf(exclusive, p_unsigned);
1107         PRINT_ATTRf(exclude_user, p_unsigned);
1108         PRINT_ATTRf(exclude_kernel, p_unsigned);
1109         PRINT_ATTRf(exclude_hv, p_unsigned);
1110         PRINT_ATTRf(exclude_idle, p_unsigned);
1111         PRINT_ATTRf(mmap, p_unsigned);
1112         PRINT_ATTRf(comm, p_unsigned);
1113         PRINT_ATTRf(freq, p_unsigned);
1114         PRINT_ATTRf(inherit_stat, p_unsigned);
1115         PRINT_ATTRf(enable_on_exec, p_unsigned);
1116         PRINT_ATTRf(task, p_unsigned);
1117         PRINT_ATTRf(watermark, p_unsigned);
1118         PRINT_ATTRf(precise_ip, p_unsigned);
1119         PRINT_ATTRf(mmap_data, p_unsigned);
1120         PRINT_ATTRf(sample_id_all, p_unsigned);
1121         PRINT_ATTRf(exclude_host, p_unsigned);
1122         PRINT_ATTRf(exclude_guest, p_unsigned);
1123         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1124         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1125         PRINT_ATTRf(mmap2, p_unsigned);
1126         PRINT_ATTRf(comm_exec, p_unsigned);
1127         PRINT_ATTRf(use_clockid, p_unsigned);
1128
1129         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1130         PRINT_ATTRf(bp_type, p_unsigned);
1131         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1132         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1133         PRINT_ATTRf(sample_regs_user, p_hex);
1134         PRINT_ATTRf(sample_stack_user, p_unsigned);
1135         PRINT_ATTRf(clockid, p_signed);
1136         PRINT_ATTRf(sample_regs_intr, p_hex);
1137         PRINT_ATTRf(aux_watermark, p_unsigned);
1138
1139         return ret;
1140 }
1141
1142 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1143                                 void *priv __attribute__((unused)))
1144 {
1145         return fprintf(fp, "  %-32s %s\n", name, val);
1146 }
1147
1148 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1149                               struct thread_map *threads)
1150 {
1151         int cpu, thread, nthreads;
1152         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1153         int pid = -1, err;
1154         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1155
1156         if (evsel->system_wide)
1157                 nthreads = 1;
1158         else
1159                 nthreads = threads->nr;
1160
1161         if (evsel->fd == NULL &&
1162             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1163                 return -ENOMEM;
1164
1165         if (evsel->cgrp) {
1166                 flags |= PERF_FLAG_PID_CGROUP;
1167                 pid = evsel->cgrp->fd;
1168         }
1169
1170 fallback_missing_features:
1171         if (perf_missing_features.clockid_wrong)
1172                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1173         if (perf_missing_features.clockid) {
1174                 evsel->attr.use_clockid = 0;
1175                 evsel->attr.clockid = 0;
1176         }
1177         if (perf_missing_features.cloexec)
1178                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1179         if (perf_missing_features.mmap2)
1180                 evsel->attr.mmap2 = 0;
1181         if (perf_missing_features.exclude_guest)
1182                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1183 retry_sample_id:
1184         if (perf_missing_features.sample_id_all)
1185                 evsel->attr.sample_id_all = 0;
1186
1187         if (verbose >= 2) {
1188                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1189                 fprintf(stderr, "perf_event_attr:\n");
1190                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1191                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1192         }
1193
1194         for (cpu = 0; cpu < cpus->nr; cpu++) {
1195
1196                 for (thread = 0; thread < nthreads; thread++) {
1197                         int group_fd;
1198
1199                         if (!evsel->cgrp && !evsel->system_wide)
1200                                 pid = thread_map__pid(threads, thread);
1201
1202                         group_fd = get_group_fd(evsel, cpu, thread);
1203 retry_open:
1204                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1205                                   pid, cpus->map[cpu], group_fd, flags);
1206
1207                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1208                                                                      pid,
1209                                                                      cpus->map[cpu],
1210                                                                      group_fd, flags);
1211                         if (FD(evsel, cpu, thread) < 0) {
1212                                 err = -errno;
1213                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1214                                           err);
1215                                 goto try_fallback;
1216                         }
1217                         set_rlimit = NO_CHANGE;
1218
1219                         /*
1220                          * If we succeeded but had to kill clockid, fail and
1221                          * have perf_evsel__open_strerror() print us a nice
1222                          * error.
1223                          */
1224                         if (perf_missing_features.clockid ||
1225                             perf_missing_features.clockid_wrong) {
1226                                 err = -EINVAL;
1227                                 goto out_close;
1228                         }
1229                 }
1230         }
1231
1232         return 0;
1233
1234 try_fallback:
1235         /*
1236          * perf stat needs between 5 and 22 fds per CPU. When we run out
1237          * of them try to increase the limits.
1238          */
1239         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1240                 struct rlimit l;
1241                 int old_errno = errno;
1242
1243                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1244                         if (set_rlimit == NO_CHANGE)
1245                                 l.rlim_cur = l.rlim_max;
1246                         else {
1247                                 l.rlim_cur = l.rlim_max + 1000;
1248                                 l.rlim_max = l.rlim_cur;
1249                         }
1250                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1251                                 set_rlimit++;
1252                                 errno = old_errno;
1253                                 goto retry_open;
1254                         }
1255                 }
1256                 errno = old_errno;
1257         }
1258
1259         if (err != -EINVAL || cpu > 0 || thread > 0)
1260                 goto out_close;
1261
1262         /*
1263          * Must probe features in the order they were added to the
1264          * perf_event_attr interface.
1265          */
1266         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1267                 perf_missing_features.clockid_wrong = true;
1268                 goto fallback_missing_features;
1269         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1270                 perf_missing_features.clockid = true;
1271                 goto fallback_missing_features;
1272         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1273                 perf_missing_features.cloexec = true;
1274                 goto fallback_missing_features;
1275         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1276                 perf_missing_features.mmap2 = true;
1277                 goto fallback_missing_features;
1278         } else if (!perf_missing_features.exclude_guest &&
1279                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1280                 perf_missing_features.exclude_guest = true;
1281                 goto fallback_missing_features;
1282         } else if (!perf_missing_features.sample_id_all) {
1283                 perf_missing_features.sample_id_all = true;
1284                 goto retry_sample_id;
1285         }
1286
1287 out_close:
1288         do {
1289                 while (--thread >= 0) {
1290                         close(FD(evsel, cpu, thread));
1291                         FD(evsel, cpu, thread) = -1;
1292                 }
1293                 thread = nthreads;
1294         } while (--cpu >= 0);
1295         return err;
1296 }
1297
1298 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1299 {
1300         if (evsel->fd == NULL)
1301                 return;
1302
1303         perf_evsel__close_fd(evsel, ncpus, nthreads);
1304         perf_evsel__free_fd(evsel);
1305 }
1306
1307 static struct {
1308         struct cpu_map map;
1309         int cpus[1];
1310 } empty_cpu_map = {
1311         .map.nr = 1,
1312         .cpus   = { -1, },
1313 };
1314
1315 static struct {
1316         struct thread_map map;
1317         int threads[1];
1318 } empty_thread_map = {
1319         .map.nr  = 1,
1320         .threads = { -1, },
1321 };
1322
1323 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1324                      struct thread_map *threads)
1325 {
1326         if (cpus == NULL) {
1327                 /* Work around old compiler warnings about strict aliasing */
1328                 cpus = &empty_cpu_map.map;
1329         }
1330
1331         if (threads == NULL)
1332                 threads = &empty_thread_map.map;
1333
1334         return __perf_evsel__open(evsel, cpus, threads);
1335 }
1336
1337 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1338                              struct cpu_map *cpus)
1339 {
1340         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1341 }
1342
1343 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1344                                 struct thread_map *threads)
1345 {
1346         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1347 }
1348
1349 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1350                                        const union perf_event *event,
1351                                        struct perf_sample *sample)
1352 {
1353         u64 type = evsel->attr.sample_type;
1354         const u64 *array = event->sample.array;
1355         bool swapped = evsel->needs_swap;
1356         union u64_swap u;
1357
1358         array += ((event->header.size -
1359                    sizeof(event->header)) / sizeof(u64)) - 1;
1360
1361         if (type & PERF_SAMPLE_IDENTIFIER) {
1362                 sample->id = *array;
1363                 array--;
1364         }
1365
1366         if (type & PERF_SAMPLE_CPU) {
1367                 u.val64 = *array;
1368                 if (swapped) {
1369                         /* undo swap of u64, then swap on individual u32s */
1370                         u.val64 = bswap_64(u.val64);
1371                         u.val32[0] = bswap_32(u.val32[0]);
1372                 }
1373
1374                 sample->cpu = u.val32[0];
1375                 array--;
1376         }
1377
1378         if (type & PERF_SAMPLE_STREAM_ID) {
1379                 sample->stream_id = *array;
1380                 array--;
1381         }
1382
1383         if (type & PERF_SAMPLE_ID) {
1384                 sample->id = *array;
1385                 array--;
1386         }
1387
1388         if (type & PERF_SAMPLE_TIME) {
1389                 sample->time = *array;
1390                 array--;
1391         }
1392
1393         if (type & PERF_SAMPLE_TID) {
1394                 u.val64 = *array;
1395                 if (swapped) {
1396                         /* undo swap of u64, then swap on individual u32s */
1397                         u.val64 = bswap_64(u.val64);
1398                         u.val32[0] = bswap_32(u.val32[0]);
1399                         u.val32[1] = bswap_32(u.val32[1]);
1400                 }
1401
1402                 sample->pid = u.val32[0];
1403                 sample->tid = u.val32[1];
1404                 array--;
1405         }
1406
1407         return 0;
1408 }
1409
1410 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1411                             u64 size)
1412 {
1413         return size > max_size || offset + size > endp;
1414 }
1415
1416 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1417         do {                                                            \
1418                 if (overflow(endp, (max_size), (offset), (size)))       \
1419                         return -EFAULT;                                 \
1420         } while (0)
1421
1422 #define OVERFLOW_CHECK_u64(offset) \
1423         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1424
1425 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1426                              struct perf_sample *data)
1427 {
1428         u64 type = evsel->attr.sample_type;
1429         bool swapped = evsel->needs_swap;
1430         const u64 *array;
1431         u16 max_size = event->header.size;
1432         const void *endp = (void *)event + max_size;
1433         u64 sz;
1434
1435         /*
1436          * used for cross-endian analysis. See git commit 65014ab3
1437          * for why this goofiness is needed.
1438          */
1439         union u64_swap u;
1440
1441         memset(data, 0, sizeof(*data));
1442         data->cpu = data->pid = data->tid = -1;
1443         data->stream_id = data->id = data->time = -1ULL;
1444         data->period = evsel->attr.sample_period;
1445         data->weight = 0;
1446
1447         if (event->header.type != PERF_RECORD_SAMPLE) {
1448                 if (!evsel->attr.sample_id_all)
1449                         return 0;
1450                 return perf_evsel__parse_id_sample(evsel, event, data);
1451         }
1452
1453         array = event->sample.array;
1454
1455         /*
1456          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1457          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1458          * check the format does not go past the end of the event.
1459          */
1460         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1461                 return -EFAULT;
1462
1463         data->id = -1ULL;
1464         if (type & PERF_SAMPLE_IDENTIFIER) {
1465                 data->id = *array;
1466                 array++;
1467         }
1468
1469         if (type & PERF_SAMPLE_IP) {
1470                 data->ip = *array;
1471                 array++;
1472         }
1473
1474         if (type & PERF_SAMPLE_TID) {
1475                 u.val64 = *array;
1476                 if (swapped) {
1477                         /* undo swap of u64, then swap on individual u32s */
1478                         u.val64 = bswap_64(u.val64);
1479                         u.val32[0] = bswap_32(u.val32[0]);
1480                         u.val32[1] = bswap_32(u.val32[1]);
1481                 }
1482
1483                 data->pid = u.val32[0];
1484                 data->tid = u.val32[1];
1485                 array++;
1486         }
1487
1488         if (type & PERF_SAMPLE_TIME) {
1489                 data->time = *array;
1490                 array++;
1491         }
1492
1493         data->addr = 0;
1494         if (type & PERF_SAMPLE_ADDR) {
1495                 data->addr = *array;
1496                 array++;
1497         }
1498
1499         if (type & PERF_SAMPLE_ID) {
1500                 data->id = *array;
1501                 array++;
1502         }
1503
1504         if (type & PERF_SAMPLE_STREAM_ID) {
1505                 data->stream_id = *array;
1506                 array++;
1507         }
1508
1509         if (type & PERF_SAMPLE_CPU) {
1510
1511                 u.val64 = *array;
1512                 if (swapped) {
1513                         /* undo swap of u64, then swap on individual u32s */
1514                         u.val64 = bswap_64(u.val64);
1515                         u.val32[0] = bswap_32(u.val32[0]);
1516                 }
1517
1518                 data->cpu = u.val32[0];
1519                 array++;
1520         }
1521
1522         if (type & PERF_SAMPLE_PERIOD) {
1523                 data->period = *array;
1524                 array++;
1525         }
1526
1527         if (type & PERF_SAMPLE_READ) {
1528                 u64 read_format = evsel->attr.read_format;
1529
1530                 OVERFLOW_CHECK_u64(array);
1531                 if (read_format & PERF_FORMAT_GROUP)
1532                         data->read.group.nr = *array;
1533                 else
1534                         data->read.one.value = *array;
1535
1536                 array++;
1537
1538                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1539                         OVERFLOW_CHECK_u64(array);
1540                         data->read.time_enabled = *array;
1541                         array++;
1542                 }
1543
1544                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1545                         OVERFLOW_CHECK_u64(array);
1546                         data->read.time_running = *array;
1547                         array++;
1548                 }
1549
1550                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1551                 if (read_format & PERF_FORMAT_GROUP) {
1552                         const u64 max_group_nr = UINT64_MAX /
1553                                         sizeof(struct sample_read_value);
1554
1555                         if (data->read.group.nr > max_group_nr)
1556                                 return -EFAULT;
1557                         sz = data->read.group.nr *
1558                              sizeof(struct sample_read_value);
1559                         OVERFLOW_CHECK(array, sz, max_size);
1560                         data->read.group.values =
1561                                         (struct sample_read_value *)array;
1562                         array = (void *)array + sz;
1563                 } else {
1564                         OVERFLOW_CHECK_u64(array);
1565                         data->read.one.id = *array;
1566                         array++;
1567                 }
1568         }
1569
1570         if (type & PERF_SAMPLE_CALLCHAIN) {
1571                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1572
1573                 OVERFLOW_CHECK_u64(array);
1574                 data->callchain = (struct ip_callchain *)array++;
1575                 if (data->callchain->nr > max_callchain_nr)
1576                         return -EFAULT;
1577                 sz = data->callchain->nr * sizeof(u64);
1578                 OVERFLOW_CHECK(array, sz, max_size);
1579                 array = (void *)array + sz;
1580         }
1581
1582         if (type & PERF_SAMPLE_RAW) {
1583                 OVERFLOW_CHECK_u64(array);
1584                 u.val64 = *array;
1585                 if (WARN_ONCE(swapped,
1586                               "Endianness of raw data not corrected!\n")) {
1587                         /* undo swap of u64, then swap on individual u32s */
1588                         u.val64 = bswap_64(u.val64);
1589                         u.val32[0] = bswap_32(u.val32[0]);
1590                         u.val32[1] = bswap_32(u.val32[1]);
1591                 }
1592                 data->raw_size = u.val32[0];
1593                 array = (void *)array + sizeof(u32);
1594
1595                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1596                 data->raw_data = (void *)array;
1597                 array = (void *)array + data->raw_size;
1598         }
1599
1600         if (type & PERF_SAMPLE_BRANCH_STACK) {
1601                 const u64 max_branch_nr = UINT64_MAX /
1602                                           sizeof(struct branch_entry);
1603
1604                 OVERFLOW_CHECK_u64(array);
1605                 data->branch_stack = (struct branch_stack *)array++;
1606
1607                 if (data->branch_stack->nr > max_branch_nr)
1608                         return -EFAULT;
1609                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1610                 OVERFLOW_CHECK(array, sz, max_size);
1611                 array = (void *)array + sz;
1612         }
1613
1614         if (type & PERF_SAMPLE_REGS_USER) {
1615                 OVERFLOW_CHECK_u64(array);
1616                 data->user_regs.abi = *array;
1617                 array++;
1618
1619                 if (data->user_regs.abi) {
1620                         u64 mask = evsel->attr.sample_regs_user;
1621
1622                         sz = hweight_long(mask) * sizeof(u64);
1623                         OVERFLOW_CHECK(array, sz, max_size);
1624                         data->user_regs.mask = mask;
1625                         data->user_regs.regs = (u64 *)array;
1626                         array = (void *)array + sz;
1627                 }
1628         }
1629
1630         if (type & PERF_SAMPLE_STACK_USER) {
1631                 OVERFLOW_CHECK_u64(array);
1632                 sz = *array++;
1633
1634                 data->user_stack.offset = ((char *)(array - 1)
1635                                           - (char *) event);
1636
1637                 if (!sz) {
1638                         data->user_stack.size = 0;
1639                 } else {
1640                         OVERFLOW_CHECK(array, sz, max_size);
1641                         data->user_stack.data = (char *)array;
1642                         array = (void *)array + sz;
1643                         OVERFLOW_CHECK_u64(array);
1644                         data->user_stack.size = *array++;
1645                         if (WARN_ONCE(data->user_stack.size > sz,
1646                                       "user stack dump failure\n"))
1647                                 return -EFAULT;
1648                 }
1649         }
1650
1651         data->weight = 0;
1652         if (type & PERF_SAMPLE_WEIGHT) {
1653                 OVERFLOW_CHECK_u64(array);
1654                 data->weight = *array;
1655                 array++;
1656         }
1657
1658         data->data_src = PERF_MEM_DATA_SRC_NONE;
1659         if (type & PERF_SAMPLE_DATA_SRC) {
1660                 OVERFLOW_CHECK_u64(array);
1661                 data->data_src = *array;
1662                 array++;
1663         }
1664
1665         data->transaction = 0;
1666         if (type & PERF_SAMPLE_TRANSACTION) {
1667                 OVERFLOW_CHECK_u64(array);
1668                 data->transaction = *array;
1669                 array++;
1670         }
1671
1672         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1673         if (type & PERF_SAMPLE_REGS_INTR) {
1674                 OVERFLOW_CHECK_u64(array);
1675                 data->intr_regs.abi = *array;
1676                 array++;
1677
1678                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1679                         u64 mask = evsel->attr.sample_regs_intr;
1680
1681                         sz = hweight_long(mask) * sizeof(u64);
1682                         OVERFLOW_CHECK(array, sz, max_size);
1683                         data->intr_regs.mask = mask;
1684                         data->intr_regs.regs = (u64 *)array;
1685                         array = (void *)array + sz;
1686                 }
1687         }
1688
1689         return 0;
1690 }
1691
1692 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1693                                      u64 read_format)
1694 {
1695         size_t sz, result = sizeof(struct sample_event);
1696
1697         if (type & PERF_SAMPLE_IDENTIFIER)
1698                 result += sizeof(u64);
1699
1700         if (type & PERF_SAMPLE_IP)
1701                 result += sizeof(u64);
1702
1703         if (type & PERF_SAMPLE_TID)
1704                 result += sizeof(u64);
1705
1706         if (type & PERF_SAMPLE_TIME)
1707                 result += sizeof(u64);
1708
1709         if (type & PERF_SAMPLE_ADDR)
1710                 result += sizeof(u64);
1711
1712         if (type & PERF_SAMPLE_ID)
1713                 result += sizeof(u64);
1714
1715         if (type & PERF_SAMPLE_STREAM_ID)
1716                 result += sizeof(u64);
1717
1718         if (type & PERF_SAMPLE_CPU)
1719                 result += sizeof(u64);
1720
1721         if (type & PERF_SAMPLE_PERIOD)
1722                 result += sizeof(u64);
1723
1724         if (type & PERF_SAMPLE_READ) {
1725                 result += sizeof(u64);
1726                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1727                         result += sizeof(u64);
1728                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1729                         result += sizeof(u64);
1730                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1731                 if (read_format & PERF_FORMAT_GROUP) {
1732                         sz = sample->read.group.nr *
1733                              sizeof(struct sample_read_value);
1734                         result += sz;
1735                 } else {
1736                         result += sizeof(u64);
1737                 }
1738         }
1739
1740         if (type & PERF_SAMPLE_CALLCHAIN) {
1741                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1742                 result += sz;
1743         }
1744
1745         if (type & PERF_SAMPLE_RAW) {
1746                 result += sizeof(u32);
1747                 result += sample->raw_size;
1748         }
1749
1750         if (type & PERF_SAMPLE_BRANCH_STACK) {
1751                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1752                 sz += sizeof(u64);
1753                 result += sz;
1754         }
1755
1756         if (type & PERF_SAMPLE_REGS_USER) {
1757                 if (sample->user_regs.abi) {
1758                         result += sizeof(u64);
1759                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1760                         result += sz;
1761                 } else {
1762                         result += sizeof(u64);
1763                 }
1764         }
1765
1766         if (type & PERF_SAMPLE_STACK_USER) {
1767                 sz = sample->user_stack.size;
1768                 result += sizeof(u64);
1769                 if (sz) {
1770                         result += sz;
1771                         result += sizeof(u64);
1772                 }
1773         }
1774
1775         if (type & PERF_SAMPLE_WEIGHT)
1776                 result += sizeof(u64);
1777
1778         if (type & PERF_SAMPLE_DATA_SRC)
1779                 result += sizeof(u64);
1780
1781         if (type & PERF_SAMPLE_TRANSACTION)
1782                 result += sizeof(u64);
1783
1784         if (type & PERF_SAMPLE_REGS_INTR) {
1785                 if (sample->intr_regs.abi) {
1786                         result += sizeof(u64);
1787                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1788                         result += sz;
1789                 } else {
1790                         result += sizeof(u64);
1791                 }
1792         }
1793
1794         return result;
1795 }
1796
1797 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1798                                   u64 read_format,
1799                                   const struct perf_sample *sample,
1800                                   bool swapped)
1801 {
1802         u64 *array;
1803         size_t sz;
1804         /*
1805          * used for cross-endian analysis. See git commit 65014ab3
1806          * for why this goofiness is needed.
1807          */
1808         union u64_swap u;
1809
1810         array = event->sample.array;
1811
1812         if (type & PERF_SAMPLE_IDENTIFIER) {
1813                 *array = sample->id;
1814                 array++;
1815         }
1816
1817         if (type & PERF_SAMPLE_IP) {
1818                 *array = sample->ip;
1819                 array++;
1820         }
1821
1822         if (type & PERF_SAMPLE_TID) {
1823                 u.val32[0] = sample->pid;
1824                 u.val32[1] = sample->tid;
1825                 if (swapped) {
1826                         /*
1827                          * Inverse of what is done in perf_evsel__parse_sample
1828                          */
1829                         u.val32[0] = bswap_32(u.val32[0]);
1830                         u.val32[1] = bswap_32(u.val32[1]);
1831                         u.val64 = bswap_64(u.val64);
1832                 }
1833
1834                 *array = u.val64;
1835                 array++;
1836         }
1837
1838         if (type & PERF_SAMPLE_TIME) {
1839                 *array = sample->time;
1840                 array++;
1841         }
1842
1843         if (type & PERF_SAMPLE_ADDR) {
1844                 *array = sample->addr;
1845                 array++;
1846         }
1847
1848         if (type & PERF_SAMPLE_ID) {
1849                 *array = sample->id;
1850                 array++;
1851         }
1852
1853         if (type & PERF_SAMPLE_STREAM_ID) {
1854                 *array = sample->stream_id;
1855                 array++;
1856         }
1857
1858         if (type & PERF_SAMPLE_CPU) {
1859                 u.val32[0] = sample->cpu;
1860                 if (swapped) {
1861                         /*
1862                          * Inverse of what is done in perf_evsel__parse_sample
1863                          */
1864                         u.val32[0] = bswap_32(u.val32[0]);
1865                         u.val64 = bswap_64(u.val64);
1866                 }
1867                 *array = u.val64;
1868                 array++;
1869         }
1870
1871         if (type & PERF_SAMPLE_PERIOD) {
1872                 *array = sample->period;
1873                 array++;
1874         }
1875
1876         if (type & PERF_SAMPLE_READ) {
1877                 if (read_format & PERF_FORMAT_GROUP)
1878                         *array = sample->read.group.nr;
1879                 else
1880                         *array = sample->read.one.value;
1881                 array++;
1882
1883                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1884                         *array = sample->read.time_enabled;
1885                         array++;
1886                 }
1887
1888                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1889                         *array = sample->read.time_running;
1890                         array++;
1891                 }
1892
1893                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1894                 if (read_format & PERF_FORMAT_GROUP) {
1895                         sz = sample->read.group.nr *
1896                              sizeof(struct sample_read_value);
1897                         memcpy(array, sample->read.group.values, sz);
1898                         array = (void *)array + sz;
1899                 } else {
1900                         *array = sample->read.one.id;
1901                         array++;
1902                 }
1903         }
1904
1905         if (type & PERF_SAMPLE_CALLCHAIN) {
1906                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1907                 memcpy(array, sample->callchain, sz);
1908                 array = (void *)array + sz;
1909         }
1910
1911         if (type & PERF_SAMPLE_RAW) {
1912                 u.val32[0] = sample->raw_size;
1913                 if (WARN_ONCE(swapped,
1914                               "Endianness of raw data not corrected!\n")) {
1915                         /*
1916                          * Inverse of what is done in perf_evsel__parse_sample
1917                          */
1918                         u.val32[0] = bswap_32(u.val32[0]);
1919                         u.val32[1] = bswap_32(u.val32[1]);
1920                         u.val64 = bswap_64(u.val64);
1921                 }
1922                 *array = u.val64;
1923                 array = (void *)array + sizeof(u32);
1924
1925                 memcpy(array, sample->raw_data, sample->raw_size);
1926                 array = (void *)array + sample->raw_size;
1927         }
1928
1929         if (type & PERF_SAMPLE_BRANCH_STACK) {
1930                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1931                 sz += sizeof(u64);
1932                 memcpy(array, sample->branch_stack, sz);
1933                 array = (void *)array + sz;
1934         }
1935
1936         if (type & PERF_SAMPLE_REGS_USER) {
1937                 if (sample->user_regs.abi) {
1938                         *array++ = sample->user_regs.abi;
1939                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1940                         memcpy(array, sample->user_regs.regs, sz);
1941                         array = (void *)array + sz;
1942                 } else {
1943                         *array++ = 0;
1944                 }
1945         }
1946
1947         if (type & PERF_SAMPLE_STACK_USER) {
1948                 sz = sample->user_stack.size;
1949                 *array++ = sz;
1950                 if (sz) {
1951                         memcpy(array, sample->user_stack.data, sz);
1952                         array = (void *)array + sz;
1953                         *array++ = sz;
1954                 }
1955         }
1956
1957         if (type & PERF_SAMPLE_WEIGHT) {
1958                 *array = sample->weight;
1959                 array++;
1960         }
1961
1962         if (type & PERF_SAMPLE_DATA_SRC) {
1963                 *array = sample->data_src;
1964                 array++;
1965         }
1966
1967         if (type & PERF_SAMPLE_TRANSACTION) {
1968                 *array = sample->transaction;
1969                 array++;
1970         }
1971
1972         if (type & PERF_SAMPLE_REGS_INTR) {
1973                 if (sample->intr_regs.abi) {
1974                         *array++ = sample->intr_regs.abi;
1975                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1976                         memcpy(array, sample->intr_regs.regs, sz);
1977                         array = (void *)array + sz;
1978                 } else {
1979                         *array++ = 0;
1980                 }
1981         }
1982
1983         return 0;
1984 }
1985
1986 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1987 {
1988         return pevent_find_field(evsel->tp_format, name);
1989 }
1990
1991 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1992                          const char *name)
1993 {
1994         struct format_field *field = perf_evsel__field(evsel, name);
1995         int offset;
1996
1997         if (!field)
1998                 return NULL;
1999
2000         offset = field->offset;
2001
2002         if (field->flags & FIELD_IS_DYNAMIC) {
2003                 offset = *(int *)(sample->raw_data + field->offset);
2004                 offset &= 0xffff;
2005         }
2006
2007         return sample->raw_data + offset;
2008 }
2009
2010 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2011                        const char *name)
2012 {
2013         struct format_field *field = perf_evsel__field(evsel, name);
2014         void *ptr;
2015         u64 value;
2016
2017         if (!field)
2018                 return 0;
2019
2020         ptr = sample->raw_data + field->offset;
2021
2022         switch (field->size) {
2023         case 1:
2024                 return *(u8 *)ptr;
2025         case 2:
2026                 value = *(u16 *)ptr;
2027                 break;
2028         case 4:
2029                 value = *(u32 *)ptr;
2030                 break;
2031         case 8:
2032                 memcpy(&value, ptr, sizeof(u64));
2033                 break;
2034         default:
2035                 return 0;
2036         }
2037
2038         if (!evsel->needs_swap)
2039                 return value;
2040
2041         switch (field->size) {
2042         case 2:
2043                 return bswap_16(value);
2044         case 4:
2045                 return bswap_32(value);
2046         case 8:
2047                 return bswap_64(value);
2048         default:
2049                 return 0;
2050         }
2051
2052         return 0;
2053 }
2054
2055 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2056 {
2057         va_list args;
2058         int ret = 0;
2059
2060         if (!*first) {
2061                 ret += fprintf(fp, ",");
2062         } else {
2063                 ret += fprintf(fp, ":");
2064                 *first = false;
2065         }
2066
2067         va_start(args, fmt);
2068         ret += vfprintf(fp, fmt, args);
2069         va_end(args);
2070         return ret;
2071 }
2072
2073 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2074 {
2075         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2076 }
2077
2078 int perf_evsel__fprintf(struct perf_evsel *evsel,
2079                         struct perf_attr_details *details, FILE *fp)
2080 {
2081         bool first = true;
2082         int printed = 0;
2083
2084         if (details->event_group) {
2085                 struct perf_evsel *pos;
2086
2087                 if (!perf_evsel__is_group_leader(evsel))
2088                         return 0;
2089
2090                 if (evsel->nr_members > 1)
2091                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2092
2093                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2094                 for_each_group_member(pos, evsel)
2095                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2096
2097                 if (evsel->nr_members > 1)
2098                         printed += fprintf(fp, "}");
2099                 goto out;
2100         }
2101
2102         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2103
2104         if (details->verbose) {
2105                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2106                                                     __print_attr__fprintf, &first);
2107         } else if (details->freq) {
2108                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2109                                          (u64)evsel->attr.sample_freq);
2110         }
2111 out:
2112         fputc('\n', fp);
2113         return ++printed;
2114 }
2115
2116 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2117                           char *msg, size_t msgsize)
2118 {
2119         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2120             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2121             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2122                 /*
2123                  * If it's cycles then fall back to hrtimer based
2124                  * cpu-clock-tick sw counter, which is always available even if
2125                  * no PMU support.
2126                  *
2127                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2128                  * b0a873e).
2129                  */
2130                 scnprintf(msg, msgsize, "%s",
2131 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2132
2133                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2134                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2135
2136                 zfree(&evsel->name);
2137                 return true;
2138         }
2139
2140         return false;
2141 }
2142
2143 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2144                               int err, char *msg, size_t size)
2145 {
2146         char sbuf[STRERR_BUFSIZE];
2147
2148         switch (err) {
2149         case EPERM:
2150         case EACCES:
2151                 return scnprintf(msg, size,
2152                  "You may not have permission to collect %sstats.\n"
2153                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2154                  " -1 - Not paranoid at all\n"
2155                  "  0 - Disallow raw tracepoint access for unpriv\n"
2156                  "  1 - Disallow cpu events for unpriv\n"
2157                  "  2 - Disallow kernel profiling for unpriv",
2158                                  target->system_wide ? "system-wide " : "");
2159         case ENOENT:
2160                 return scnprintf(msg, size, "The %s event is not supported.",
2161                                  perf_evsel__name(evsel));
2162         case EMFILE:
2163                 return scnprintf(msg, size, "%s",
2164                          "Too many events are opened.\n"
2165                          "Probably the maximum number of open file descriptors has been reached.\n"
2166                          "Hint: Try again after reducing the number of events.\n"
2167                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2168         case ENODEV:
2169                 if (target->cpu_list)
2170                         return scnprintf(msg, size, "%s",
2171          "No such device - did you specify an out-of-range profile CPU?\n");
2172                 break;
2173         case EOPNOTSUPP:
2174                 if (evsel->attr.precise_ip)
2175                         return scnprintf(msg, size, "%s",
2176         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2177 #if defined(__i386__) || defined(__x86_64__)
2178                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2179                         return scnprintf(msg, size, "%s",
2180         "No hardware sampling interrupt available.\n"
2181         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2182 #endif
2183                 break;
2184         case EBUSY:
2185                 if (find_process("oprofiled"))
2186                         return scnprintf(msg, size,
2187         "The PMU counters are busy/taken by another profiler.\n"
2188         "We found oprofile daemon running, please stop it and try again.");
2189                 break;
2190         case EINVAL:
2191                 if (perf_missing_features.clockid)
2192                         return scnprintf(msg, size, "clockid feature not supported.");
2193                 if (perf_missing_features.clockid_wrong)
2194                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2195                 break;
2196         default:
2197                 break;
2198         }
2199
2200         return scnprintf(msg, size,
2201         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2202         "/bin/dmesg may provide additional information.\n"
2203         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2204                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2205                          perf_evsel__name(evsel));
2206 }