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         INIT_LIST_HEAD(&evsel->config_terms);
 211         perf_evsel__object.init(evsel);
 212         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 213         perf_evsel__calc_id_pos(evsel);
 214         evsel->cmdline_group_boundary = false;
 215 }
 216
 217 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 218 {
 219         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 220
 221         if (evsel != NULL)
 222                 perf_evsel__init(evsel, attr, idx);
 223
 224         return evsel;
 225 }
 226
 227 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 228 {
 229         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 230
 231         if (evsel != NULL) {
 232                 struct perf_event_attr attr = {
 233                         .type          = PERF_TYPE_TRACEPOINT,
 234                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 235                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 236                 };
 237
 238                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 239                         goto out_free;
 240
 241                 evsel->tp_format = trace_event__tp_format(sys, name);
 242                 if (evsel->tp_format == NULL)
 243                         goto out_free;
 244
 245                 event_attr_init(&attr);
 246                 attr.config = evsel->tp_format->id;
 247                 attr.sample_period = 1;
 248                 perf_evsel__init(evsel, &attr, idx);
 249         }
 250
 251         return evsel;
 252
 253 out_free:
 254         zfree(&evsel->name);
 255         free(evsel);
 256         return NULL;
 257 }
 258
 259 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 260         "cycles",
 261         "instructions",
 262         "cache-references",
 263         "cache-misses",
 264         "branches",
 265         "branch-misses",
 266         "bus-cycles",
 267         "stalled-cycles-frontend",
 268         "stalled-cycles-backend",
 269         "ref-cycles",
 270 };
 271
 272 static const char *__perf_evsel__hw_name(u64 config)
 273 {
 274         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 275                 return perf_evsel__hw_names[config];
 276
 277         return "unknown-hardware";
 278 }
 279
 280 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 281 {
 282         int colon = 0, r = 0;
 283         struct perf_event_attr *attr = &evsel->attr;
 284         bool exclude_guest_default = false;
 285
 286 #define MOD_PRINT(context, mod) do {                                    \
 287                 if (!attr->exclude_##context) {                         \
 288                         if (!colon) colon = ++r;                        \
 289                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 290                 } } while(0)
 291
 292         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 293                 MOD_PRINT(kernel, 'k');
 294                 MOD_PRINT(user, 'u');
 295                 MOD_PRINT(hv, 'h');
 296                 exclude_guest_default = true;
 297         }
 298
 299         if (attr->precise_ip) {
 300                 if (!colon)
 301                         colon = ++r;
 302                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 303                 exclude_guest_default = true;
 304         }
 305
 306         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 307                 MOD_PRINT(host, 'H');
 308                 MOD_PRINT(guest, 'G');
 309         }
 310 #undef MOD_PRINT
 311         if (colon)
 312                 bf[colon - 1] = ':';
 313         return r;
 314 }
 315
 316 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 317 {
 318         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 319         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 320 }
 321
 322 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 323         "cpu-clock",
 324         "task-clock",
 325         "page-faults",
 326         "context-switches",
 327         "cpu-migrations",
 328         "minor-faults",
 329         "major-faults",
 330         "alignment-faults",
 331         "emulation-faults",
 332         "dummy",
 333 };
 334
 335 static const char *__perf_evsel__sw_name(u64 config)
 336 {
 337         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 338                 return perf_evsel__sw_names[config];
 339         return "unknown-software";
 340 }
 341
 342 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 343 {
 344         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 345         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 346 }
 347
 348 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 349 {
 350         int r;
 351
 352         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 353
 354         if (type & HW_BREAKPOINT_R)
 355                 r += scnprintf(bf + r, size - r, "r");
 356
 357         if (type & HW_BREAKPOINT_W)
 358                 r += scnprintf(bf + r, size - r, "w");
 359
 360         if (type & HW_BREAKPOINT_X)
 361                 r += scnprintf(bf + r, size - r, "x");
 362
 363         return r;
 364 }
 365
 366 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 367 {
 368         struct perf_event_attr *attr = &evsel->attr;
 369         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 370         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 371 }
 372
 373 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 374                                 [PERF_EVSEL__MAX_ALIASES] = {
 375  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 376  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 377  { "LLC",       "L2",                                                   },
 378  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 379  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 380  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 381  { "node",                                                              },
 382 };
 383
 384 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 385                                    [PERF_EVSEL__MAX_ALIASES] = {
 386  { "load",      "loads",        "read",                                 },
 387  { "store",     "stores",       "write",                                },
 388  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 389 };
 390
 391 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 392                                        [PERF_EVSEL__MAX_ALIASES] = {
 393  { "refs",      "Reference",    "ops",          "access",               },
 394  { "misses",    "miss",                                                 },
 395 };
 396
 397 #define C(x)            PERF_COUNT_HW_CACHE_##x
 398 #define CACHE_READ      (1 << C(OP_READ))
 399 #define CACHE_WRITE     (1 << C(OP_WRITE))
 400 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 401 #define COP(x)          (1 << x)
 402
 403 /*
 404  * cache operartion stat
 405  * L1I : Read and prefetch only
 406  * ITLB and BPU : Read-only
 407  */
 408 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 409  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 410  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 411  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 412  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 413  [C(ITLB)]      = (CACHE_READ),
 414  [C(BPU)]       = (CACHE_READ),
 415  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 416 };
 417
 418 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 419 {
 420         if (perf_evsel__hw_cache_stat[type] & COP(op))
 421                 return true;    /* valid */
 422         else
 423                 return false;   /* invalid */
 424 }
 425
 426 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 427                                             char *bf, size_t size)
 428 {
 429         if (result) {
 430                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 431                                  perf_evsel__hw_cache_op[op][0],
 432                                  perf_evsel__hw_cache_result[result][0]);
 433         }
 434
 435         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 436                          perf_evsel__hw_cache_op[op][1]);
 437 }
 438
 439 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 440 {
 441         u8 op, result, type = (config >>  0) & 0xff;
 442         const char *err = "unknown-ext-hardware-cache-type";
 443
 444         if (type > PERF_COUNT_HW_CACHE_MAX)
 445                 goto out_err;
 446
 447         op = (config >>  8) & 0xff;
 448         err = "unknown-ext-hardware-cache-op";
 449         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 450                 goto out_err;
 451
 452         result = (config >> 16) & 0xff;
 453         err = "unknown-ext-hardware-cache-result";
 454         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 455                 goto out_err;
 456
 457         err = "invalid-cache";
 458         if (!perf_evsel__is_cache_op_valid(type, op))
 459                 goto out_err;
 460
 461         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 462 out_err:
 463         return scnprintf(bf, size, "%s", err);
 464 }
 465
 466 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 467 {
 468         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 469         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 470 }
 471
 472 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 473 {
 474         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 475         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 476 }
 477
 478 const char *perf_evsel__name(struct perf_evsel *evsel)
 479 {
 480         char bf[128];
 481
 482         if (evsel->name)
 483                 return evsel->name;
 484
 485         switch (evsel->attr.type) {
 486         case PERF_TYPE_RAW:
 487                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 488                 break;
 489
 490         case PERF_TYPE_HARDWARE:
 491                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 492                 break;
 493
 494         case PERF_TYPE_HW_CACHE:
 495                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 496                 break;
 497
 498         case PERF_TYPE_SOFTWARE:
 499                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 500                 break;
 501
 502         case PERF_TYPE_TRACEPOINT:
 503                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 504                 break;
 505
 506         case PERF_TYPE_BREAKPOINT:
 507                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 508                 break;
 509
 510         default:
 511                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 512                           evsel->attr.type);
 513                 break;
 514         }
 515
 516         evsel->name = strdup(bf);
 517
 518         return evsel->name ?: "unknown";
 519 }
 520
 521 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 522 {
 523         return evsel->group_name ?: "anon group";
 524 }
 525
 526 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 527 {
 528         int ret;
 529         struct perf_evsel *pos;
 530         const char *group_name = perf_evsel__group_name(evsel);
 531
 532         ret = scnprintf(buf, size, "%s", group_name);
 533
 534         ret += scnprintf(buf + ret, size - ret, " { %s",
 535                          perf_evsel__name(evsel));
 536
 537         for_each_group_member(pos, evsel)
 538                 ret += scnprintf(buf + ret, size - ret, ", %s",
 539                                  perf_evsel__name(pos));
 540
 541         ret += scnprintf(buf + ret, size - ret, " }");
 542
 543         return ret;
 544 }
 545
 546 static void
 547 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 548                              struct record_opts *opts)
 549 {
 550         bool function = perf_evsel__is_function_event(evsel);
 551         struct perf_event_attr *attr = &evsel->attr;
 552
 553         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 554
 555         if (callchain_param.record_mode == CALLCHAIN_LBR) {
 556                 if (!opts->branch_stack) {
 557                         if (attr->exclude_user) {
 558                                 pr_warning("LBR callstack option is only available "
 559                                            "to get user callchain information. "
 560                                            "Falling back to framepointers.\n");
 561                         } else {
 562                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 563                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 564                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 565                         }
 566                 } else
 567                          pr_warning("Cannot use LBR callstack with branch stack. "
 568                                     "Falling back to framepointers.\n");
 569         }
 570
 571         if (callchain_param.record_mode == CALLCHAIN_DWARF) {
 572                 if (!function) {
 573                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 574                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 575                         attr->sample_regs_user = PERF_REGS_MASK;
 576                         attr->sample_stack_user = callchain_param.dump_size;
 577                         attr->exclude_callchain_user = 1;
 578                 } else {
 579                         pr_info("Cannot use DWARF unwind for function trace event,"
 580                                 " falling back to framepointers.\n");
 581                 }
 582         }
 583
 584         if (function) {
 585                 pr_info("Disabling user space callchains for function trace event.\n");
 586                 attr->exclude_callchain_user = 1;
 587         }
 588 }
 589
 590 static void apply_config_terms(struct perf_evsel *evsel)
 591 {
 592         struct perf_evsel_config_term *term;
 593         struct list_head *config_terms = &evsel->config_terms;
 594         struct perf_event_attr *attr = &evsel->attr;
 595
 596         list_for_each_entry(term, config_terms, list) {
 597                 switch (term->type) {
 598                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 599                         attr->sample_period = term->val.period;
 600                         break;
 601                 case PERF_EVSEL__CONFIG_TERM_TIME:
 602                         if (term->val.time)
 603                                 perf_evsel__set_sample_bit(evsel, TIME);
 604                         else
 605                                 perf_evsel__reset_sample_bit(evsel, TIME);
 606                         break;
 607                 default:
 608                         break;
 609                 }
 610         }
 611 }
 612
 613 /*
 614  * The enable_on_exec/disabled value strategy:
 615  *
 616  *  1) For any type of traced program:
 617  *    - all independent events and group leaders are disabled
 618  *    - all group members are enabled
 619  *
 620  *     Group members are ruled by group leaders. They need to
 621  *     be enabled, because the group scheduling relies on that.
 622  *
 623  *  2) For traced programs executed by perf:
 624  *     - all independent events and group leaders have
 625  *       enable_on_exec set
 626  *     - we don't specifically enable or disable any event during
 627  *       the record command
 628  *
 629  *     Independent events and group leaders are initially disabled
 630  *     and get enabled by exec. Group members are ruled by group
 631  *     leaders as stated in 1).
 632  *
 633  *  3) For traced programs attached by perf (pid/tid):
 634  *     - we specifically enable or disable all events during
 635  *       the record command
 636  *
 637  *     When attaching events to already running traced we
 638  *     enable/disable events specifically, as there's no
 639  *     initial traced exec call.
 640  */
 641 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 642 {
 643         struct perf_evsel *leader = evsel->leader;
 644         struct perf_event_attr *attr = &evsel->attr;
 645         int track = evsel->tracking;
 646         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 647
 648         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 649         attr->inherit       = !opts->no_inherit;
 650
 651         perf_evsel__set_sample_bit(evsel, IP);
 652         perf_evsel__set_sample_bit(evsel, TID);
 653
 654         if (evsel->sample_read) {
 655                 perf_evsel__set_sample_bit(evsel, READ);
 656
 657                 /*
 658                  * We need ID even in case of single event, because
 659                  * PERF_SAMPLE_READ process ID specific data.
 660                  */
 661                 perf_evsel__set_sample_id(evsel, false);
 662
 663                 /*
 664                  * Apply group format only if we belong to group
 665                  * with more than one members.
 666                  */
 667                 if (leader->nr_members > 1) {
 668                         attr->read_format |= PERF_FORMAT_GROUP;
 669                         attr->inherit = 0;
 670                 }
 671         }
 672
 673         /*
 674          * We default some events to have a default interval. But keep
 675          * it a weak assumption overridable by the user.
 676          */
 677         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 678                                      opts->user_interval != ULLONG_MAX)) {
 679                 if (opts->freq) {
 680                         perf_evsel__set_sample_bit(evsel, PERIOD);
 681                         attr->freq              = 1;
 682                         attr->sample_freq       = opts->freq;
 683                 } else {
 684                         attr->sample_period = opts->default_interval;
 685                 }
 686         }
 687
 688         /*
 689          * Disable sampling for all group members other
 690          * than leader in case leader 'leads' the sampling.
 691          */
 692         if ((leader != evsel) && leader->sample_read) {
 693                 attr->sample_freq   = 0;
 694                 attr->sample_period = 0;
 695         }
 696
 697         if (opts->no_samples)
 698                 attr->sample_freq = 0;
 699
 700         if (opts->inherit_stat)
 701                 attr->inherit_stat = 1;
 702
 703         if (opts->sample_address) {
 704                 perf_evsel__set_sample_bit(evsel, ADDR);
 705                 attr->mmap_data = track;
 706         }
 707
 708         /*
 709          * We don't allow user space callchains for  function trace
 710          * event, due to issues with page faults while tracing page
 711          * fault handler and its overall trickiness nature.
 712          */
 713         if (perf_evsel__is_function_event(evsel))
 714                 evsel->attr.exclude_callchain_user = 1;
 715
 716         if (callchain_param.enabled && !evsel->no_aux_samples)
 717                 perf_evsel__config_callgraph(evsel, opts);
 718
 719         if (opts->sample_intr_regs) {
 720                 attr->sample_regs_intr = PERF_REGS_MASK;
 721                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 722         }
 723
 724         if (target__has_cpu(&opts->target))
 725                 perf_evsel__set_sample_bit(evsel, CPU);
 726
 727         if (opts->period)
 728                 perf_evsel__set_sample_bit(evsel, PERIOD);
 729
 730         /*
 731          * When the user explicitely disabled time don't force it here.
 732          */
 733         if (opts->sample_time &&
 734             (!perf_missing_features.sample_id_all &&
 735             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 736              opts->sample_time_set)))
 737                 perf_evsel__set_sample_bit(evsel, TIME);
 738
 739         if (opts->raw_samples && !evsel->no_aux_samples) {
 740                 perf_evsel__set_sample_bit(evsel, TIME);
 741                 perf_evsel__set_sample_bit(evsel, RAW);
 742                 perf_evsel__set_sample_bit(evsel, CPU);
 743         }
 744
 745         if (opts->sample_address)
 746                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 747
 748         if (opts->no_buffering) {
 749                 attr->watermark = 0;
 750                 attr->wakeup_events = 1;
 751         }
 752         if (opts->branch_stack && !evsel->no_aux_samples) {
 753                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 754                 attr->branch_sample_type = opts->branch_stack;
 755         }
 756
 757         if (opts->sample_weight)
 758                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 759
 760         attr->task  = track;
 761         attr->mmap  = track;
 762         attr->mmap2 = track && !perf_missing_features.mmap2;
 763         attr->comm  = track;
 764
 765         if (opts->record_switch_events)
 766                 attr->context_switch = track;
 767
 768         if (opts->sample_transaction)
 769                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 770
 771         if (opts->running_time) {
 772                 evsel->attr.read_format |=
 773                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 774                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 775         }
 776
 777         /*
 778          * XXX see the function comment above
 779          *
 780          * Disabling only independent events or group leaders,
 781          * keeping group members enabled.
 782          */
 783         if (perf_evsel__is_group_leader(evsel))
 784                 attr->disabled = 1;
 785
 786         /*
 787          * Setting enable_on_exec for independent events and
 788          * group leaders for traced executed by perf.
 789          */
 790         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 791                 !opts->initial_delay)
 792                 attr->enable_on_exec = 1;
 793
 794         if (evsel->immediate) {
 795                 attr->disabled = 0;
 796                 attr->enable_on_exec = 0;
 797         }
 798
 799         clockid = opts->clockid;
 800         if (opts->use_clockid) {
 801                 attr->use_clockid = 1;
 802                 attr->clockid = opts->clockid;
 803         }
 804
 805         /*
 806          * Apply event specific term settings,
 807          * it overloads any global configuration.
 808          */
 809         apply_config_terms(evsel);
 810 }
 811
 812 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 813 {
 814         int cpu, thread;
 815
 816         if (evsel->system_wide)
 817                 nthreads = 1;
 818
 819         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 820
 821         if (evsel->fd) {
 822                 for (cpu = 0; cpu < ncpus; cpu++) {
 823                         for (thread = 0; thread < nthreads; thread++) {
 824                                 FD(evsel, cpu, thread) = -1;
 825                         }
 826                 }
 827         }
 828
 829         return evsel->fd != NULL ? 0 : -ENOMEM;
 830 }
 831
 832 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 833                           int ioc,  void *arg)
 834 {
 835         int cpu, thread;
 836
 837         if (evsel->system_wide)
 838                 nthreads = 1;
 839
 840         for (cpu = 0; cpu < ncpus; cpu++) {
 841                 for (thread = 0; thread < nthreads; thread++) {
 842                         int fd = FD(evsel, cpu, thread),
 843                             err = ioctl(fd, ioc, arg);
 844
 845                         if (err)
 846                                 return err;
 847                 }
 848         }
 849
 850         return 0;
 851 }
 852
 853 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 854                              const char *filter)
 855 {
 856         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 857                                      PERF_EVENT_IOC_SET_FILTER,
 858                                      (void *)filter);
 859 }
 860
 861 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 862 {
 863         char *new_filter = strdup(filter);
 864
 865         if (new_filter != NULL) {
 866                 free(evsel->filter);
 867                 evsel->filter = new_filter;
 868                 return 0;
 869         }
 870
 871         return -1;
 872 }
 873
 874 int perf_evsel__append_filter(struct perf_evsel *evsel,
 875                               const char *op, const char *filter)
 876 {
 877         char *new_filter;
 878
 879         if (evsel->filter == NULL)
 880                 return perf_evsel__set_filter(evsel, filter);
 881
 882         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 883                 free(evsel->filter);
 884                 evsel->filter = new_filter;
 885                 return 0;
 886         }
 887
 888         return -1;
 889 }
 890
 891 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 892 {
 893         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 894                                      PERF_EVENT_IOC_ENABLE,
 895                                      0);
 896 }
 897
 898 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 899 {
 900         if (ncpus == 0 || nthreads == 0)
 901                 return 0;
 902
 903         if (evsel->system_wide)
 904                 nthreads = 1;
 905
 906         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 907         if (evsel->sample_id == NULL)
 908                 return -ENOMEM;
 909
 910         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 911         if (evsel->id == NULL) {
 912                 xyarray__delete(evsel->sample_id);
 913                 evsel->sample_id = NULL;
 914                 return -ENOMEM;
 915         }
 916
 917         return 0;
 918 }
 919
 920 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 921 {
 922         xyarray__delete(evsel->fd);
 923         evsel->fd = NULL;
 924 }
 925
 926 static void perf_evsel__free_id(struct perf_evsel *evsel)
 927 {
 928         xyarray__delete(evsel->sample_id);
 929         evsel->sample_id = NULL;
 930         zfree(&evsel->id);
 931 }
 932
 933 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
 934 {
 935         struct perf_evsel_config_term *term, *h;
 936
 937         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
 938                 list_del(&term->list);
 939                 free(term);
 940         }
 941 }
 942
 943 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 944 {
 945         int cpu, thread;
 946
 947         if (evsel->system_wide)
 948                 nthreads = 1;
 949
 950         for (cpu = 0; cpu < ncpus; cpu++)
 951                 for (thread = 0; thread < nthreads; ++thread) {
 952                         close(FD(evsel, cpu, thread));
 953                         FD(evsel, cpu, thread) = -1;
 954                 }
 955 }
 956
 957 void perf_evsel__exit(struct perf_evsel *evsel)
 958 {
 959         assert(list_empty(&evsel->node));
 960         perf_evsel__free_fd(evsel);
 961         perf_evsel__free_id(evsel);
 962         perf_evsel__free_config_terms(evsel);
 963         close_cgroup(evsel->cgrp);
 964         cpu_map__put(evsel->cpus);
 965         thread_map__put(evsel->threads);
 966         zfree(&evsel->group_name);
 967         zfree(&evsel->name);
 968         perf_evsel__object.fini(evsel);
 969 }
 970
 971 void perf_evsel__delete(struct perf_evsel *evsel)
 972 {
 973         perf_evsel__exit(evsel);
 974         free(evsel);
 975 }
 976
 977 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
 978                                 struct perf_counts_values *count)
 979 {
 980         struct perf_counts_values tmp;
 981
 982         if (!evsel->prev_raw_counts)
 983                 return;
 984
 985         if (cpu == -1) {
 986                 tmp = evsel->prev_raw_counts->aggr;
 987                 evsel->prev_raw_counts->aggr = *count;
 988         } else {
 989                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
 990                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
 991         }
 992
 993         count->val = count->val - tmp.val;
 994         count->ena = count->ena - tmp.ena;
 995         count->run = count->run - tmp.run;
 996 }
 997
 998 void perf_counts_values__scale(struct perf_counts_values *count,
 999                                bool scale, s8 *pscaled)
1000 {
1001         s8 scaled = 0;
1002
1003         if (scale) {
1004                 if (count->run == 0) {
1005                         scaled = -1;
1006                         count->val = 0;
1007                 } else if (count->run < count->ena) {
1008                         scaled = 1;
1009                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1010                 }
1011         } else
1012                 count->ena = count->run = 0;
1013
1014         if (pscaled)
1015                 *pscaled = scaled;
1016 }
1017
1018 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1019                      struct perf_counts_values *count)
1020 {
1021         memset(count, 0, sizeof(*count));
1022
1023         if (FD(evsel, cpu, thread) < 0)
1024                 return -EINVAL;
1025
1026         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1027                 return -errno;
1028
1029         return 0;
1030 }
1031
1032 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1033                               int cpu, int thread, bool scale)
1034 {
1035         struct perf_counts_values count;
1036         size_t nv = scale ? 3 : 1;
1037
1038         if (FD(evsel, cpu, thread) < 0)
1039                 return -EINVAL;
1040
1041         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1042                 return -ENOMEM;
1043
1044         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1045                 return -errno;
1046
1047         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1048         perf_counts_values__scale(&count, scale, NULL);
1049         *perf_counts(evsel->counts, cpu, thread) = count;
1050         return 0;
1051 }
1052
1053 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1054 {
1055         struct perf_evsel *leader = evsel->leader;
1056         int fd;
1057
1058         if (perf_evsel__is_group_leader(evsel))
1059                 return -1;
1060
1061         /*
1062          * Leader must be already processed/open,
1063          * if not it's a bug.
1064          */
1065         BUG_ON(!leader->fd);
1066
1067         fd = FD(leader, cpu, thread);
1068         BUG_ON(fd == -1);
1069
1070         return fd;
1071 }
1072
1073 struct bit_names {
1074         int bit;
1075         const char *name;
1076 };
1077
1078 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1079 {
1080         bool first_bit = true;
1081         int i = 0;
1082
1083         do {
1084                 if (value & bits[i].bit) {
1085                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1086                         first_bit = false;
1087                 }
1088         } while (bits[++i].name != NULL);
1089 }
1090
1091 static void __p_sample_type(char *buf, size_t size, u64 value)
1092 {
1093 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1094         struct bit_names bits[] = {
1095                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1096                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1097                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1098                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1099                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1100                 { .name = NULL, }
1101         };
1102 #undef bit_name
1103         __p_bits(buf, size, value, bits);
1104 }
1105
1106 static void __p_read_format(char *buf, size_t size, u64 value)
1107 {
1108 #define bit_name(n) { PERF_FORMAT_##n, #n }
1109         struct bit_names bits[] = {
1110                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1111                 bit_name(ID), bit_name(GROUP),
1112                 { .name = NULL, }
1113         };
1114 #undef bit_name
1115         __p_bits(buf, size, value, bits);
1116 }
1117
1118 #define BUF_SIZE                1024
1119
1120 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1121 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1122 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1123 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1124 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1125
1126 #define PRINT_ATTRn(_n, _f, _p)                         \
1127 do {                                                    \
1128         if (attr->_f) {                                 \
1129                 _p(attr->_f);                           \
1130                 ret += attr__fprintf(fp, _n, buf, priv);\
1131         }                                               \
1132 } while (0)
1133
1134 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1135
1136 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1137                              attr__fprintf_f attr__fprintf, void *priv)
1138 {
1139         char buf[BUF_SIZE];
1140         int ret = 0;
1141
1142         PRINT_ATTRf(type, p_unsigned);
1143         PRINT_ATTRf(size, p_unsigned);
1144         PRINT_ATTRf(config, p_hex);
1145         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1146         PRINT_ATTRf(sample_type, p_sample_type);
1147         PRINT_ATTRf(read_format, p_read_format);
1148
1149         PRINT_ATTRf(disabled, p_unsigned);
1150         PRINT_ATTRf(inherit, p_unsigned);
1151         PRINT_ATTRf(pinned, p_unsigned);
1152         PRINT_ATTRf(exclusive, p_unsigned);
1153         PRINT_ATTRf(exclude_user, p_unsigned);
1154         PRINT_ATTRf(exclude_kernel, p_unsigned);
1155         PRINT_ATTRf(exclude_hv, p_unsigned);
1156         PRINT_ATTRf(exclude_idle, p_unsigned);
1157         PRINT_ATTRf(mmap, p_unsigned);
1158         PRINT_ATTRf(comm, p_unsigned);
1159         PRINT_ATTRf(freq, p_unsigned);
1160         PRINT_ATTRf(inherit_stat, p_unsigned);
1161         PRINT_ATTRf(enable_on_exec, p_unsigned);
1162         PRINT_ATTRf(task, p_unsigned);
1163         PRINT_ATTRf(watermark, p_unsigned);
1164         PRINT_ATTRf(precise_ip, p_unsigned);
1165         PRINT_ATTRf(mmap_data, p_unsigned);
1166         PRINT_ATTRf(sample_id_all, p_unsigned);
1167         PRINT_ATTRf(exclude_host, p_unsigned);
1168         PRINT_ATTRf(exclude_guest, p_unsigned);
1169         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1170         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1171         PRINT_ATTRf(mmap2, p_unsigned);
1172         PRINT_ATTRf(comm_exec, p_unsigned);
1173         PRINT_ATTRf(use_clockid, p_unsigned);
1174         PRINT_ATTRf(context_switch, p_unsigned);
1175
1176         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1177         PRINT_ATTRf(bp_type, p_unsigned);
1178         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1179         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1180         PRINT_ATTRf(sample_regs_user, p_hex);
1181         PRINT_ATTRf(sample_stack_user, p_unsigned);
1182         PRINT_ATTRf(clockid, p_signed);
1183         PRINT_ATTRf(sample_regs_intr, p_hex);
1184         PRINT_ATTRf(aux_watermark, p_unsigned);
1185
1186         return ret;
1187 }
1188
1189 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1190                                 void *priv __attribute__((unused)))
1191 {
1192         return fprintf(fp, "  %-32s %s\n", name, val);
1193 }
1194
1195 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1196                               struct thread_map *threads)
1197 {
1198         int cpu, thread, nthreads;
1199         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1200         int pid = -1, err;
1201         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1202
1203         if (evsel->system_wide)
1204                 nthreads = 1;
1205         else
1206                 nthreads = threads->nr;
1207
1208         if (evsel->fd == NULL &&
1209             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1210                 return -ENOMEM;
1211
1212         if (evsel->cgrp) {
1213                 flags |= PERF_FLAG_PID_CGROUP;
1214                 pid = evsel->cgrp->fd;
1215         }
1216
1217 fallback_missing_features:
1218         if (perf_missing_features.clockid_wrong)
1219                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1220         if (perf_missing_features.clockid) {
1221                 evsel->attr.use_clockid = 0;
1222                 evsel->attr.clockid = 0;
1223         }
1224         if (perf_missing_features.cloexec)
1225                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1226         if (perf_missing_features.mmap2)
1227                 evsel->attr.mmap2 = 0;
1228         if (perf_missing_features.exclude_guest)
1229                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1230 retry_sample_id:
1231         if (perf_missing_features.sample_id_all)
1232                 evsel->attr.sample_id_all = 0;
1233
1234         if (verbose >= 2) {
1235                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1236                 fprintf(stderr, "perf_event_attr:\n");
1237                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1238                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1239         }
1240
1241         for (cpu = 0; cpu < cpus->nr; cpu++) {
1242
1243                 for (thread = 0; thread < nthreads; thread++) {
1244                         int group_fd;
1245
1246                         if (!evsel->cgrp && !evsel->system_wide)
1247                                 pid = thread_map__pid(threads, thread);
1248
1249                         group_fd = get_group_fd(evsel, cpu, thread);
1250 retry_open:
1251                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1252                                   pid, cpus->map[cpu], group_fd, flags);
1253
1254                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1255                                                                      pid,
1256                                                                      cpus->map[cpu],
1257                                                                      group_fd, flags);
1258                         if (FD(evsel, cpu, thread) < 0) {
1259                                 err = -errno;
1260                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1261                                           err);
1262                                 goto try_fallback;
1263                         }
1264                         set_rlimit = NO_CHANGE;
1265
1266                         /*
1267                          * If we succeeded but had to kill clockid, fail and
1268                          * have perf_evsel__open_strerror() print us a nice
1269                          * error.
1270                          */
1271                         if (perf_missing_features.clockid ||
1272                             perf_missing_features.clockid_wrong) {
1273                                 err = -EINVAL;
1274                                 goto out_close;
1275                         }
1276                 }
1277         }
1278
1279         return 0;
1280
1281 try_fallback:
1282         /*
1283          * perf stat needs between 5 and 22 fds per CPU. When we run out
1284          * of them try to increase the limits.
1285          */
1286         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1287                 struct rlimit l;
1288                 int old_errno = errno;
1289
1290                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1291                         if (set_rlimit == NO_CHANGE)
1292                                 l.rlim_cur = l.rlim_max;
1293                         else {
1294                                 l.rlim_cur = l.rlim_max + 1000;
1295                                 l.rlim_max = l.rlim_cur;
1296                         }
1297                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1298                                 set_rlimit++;
1299                                 errno = old_errno;
1300                                 goto retry_open;
1301                         }
1302                 }
1303                 errno = old_errno;
1304         }
1305
1306         if (err != -EINVAL || cpu > 0 || thread > 0)
1307                 goto out_close;
1308
1309         /*
1310          * Must probe features in the order they were added to the
1311          * perf_event_attr interface.
1312          */
1313         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1314                 perf_missing_features.clockid_wrong = true;
1315                 goto fallback_missing_features;
1316         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1317                 perf_missing_features.clockid = true;
1318                 goto fallback_missing_features;
1319         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1320                 perf_missing_features.cloexec = true;
1321                 goto fallback_missing_features;
1322         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1323                 perf_missing_features.mmap2 = true;
1324                 goto fallback_missing_features;
1325         } else if (!perf_missing_features.exclude_guest &&
1326                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1327                 perf_missing_features.exclude_guest = true;
1328                 goto fallback_missing_features;
1329         } else if (!perf_missing_features.sample_id_all) {
1330                 perf_missing_features.sample_id_all = true;
1331                 goto retry_sample_id;
1332         }
1333
1334 out_close:
1335         do {
1336                 while (--thread >= 0) {
1337                         close(FD(evsel, cpu, thread));
1338                         FD(evsel, cpu, thread) = -1;
1339                 }
1340                 thread = nthreads;
1341         } while (--cpu >= 0);
1342         return err;
1343 }
1344
1345 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1346 {
1347         if (evsel->fd == NULL)
1348                 return;
1349
1350         perf_evsel__close_fd(evsel, ncpus, nthreads);
1351         perf_evsel__free_fd(evsel);
1352 }
1353
1354 static struct {
1355         struct cpu_map map;
1356         int cpus[1];
1357 } empty_cpu_map = {
1358         .map.nr = 1,
1359         .cpus   = { -1, },
1360 };
1361
1362 static struct {
1363         struct thread_map map;
1364         int threads[1];
1365 } empty_thread_map = {
1366         .map.nr  = 1,
1367         .threads = { -1, },
1368 };
1369
1370 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1371                      struct thread_map *threads)
1372 {
1373         if (cpus == NULL) {
1374                 /* Work around old compiler warnings about strict aliasing */
1375                 cpus = &empty_cpu_map.map;
1376         }
1377
1378         if (threads == NULL)
1379                 threads = &empty_thread_map.map;
1380
1381         return __perf_evsel__open(evsel, cpus, threads);
1382 }
1383
1384 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1385                              struct cpu_map *cpus)
1386 {
1387         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1388 }
1389
1390 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1391                                 struct thread_map *threads)
1392 {
1393         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1394 }
1395
1396 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1397                                        const union perf_event *event,
1398                                        struct perf_sample *sample)
1399 {
1400         u64 type = evsel->attr.sample_type;
1401         const u64 *array = event->sample.array;
1402         bool swapped = evsel->needs_swap;
1403         union u64_swap u;
1404
1405         array += ((event->header.size -
1406                    sizeof(event->header)) / sizeof(u64)) - 1;
1407
1408         if (type & PERF_SAMPLE_IDENTIFIER) {
1409                 sample->id = *array;
1410                 array--;
1411         }
1412
1413         if (type & PERF_SAMPLE_CPU) {
1414                 u.val64 = *array;
1415                 if (swapped) {
1416                         /* undo swap of u64, then swap on individual u32s */
1417                         u.val64 = bswap_64(u.val64);
1418                         u.val32[0] = bswap_32(u.val32[0]);
1419                 }
1420
1421                 sample->cpu = u.val32[0];
1422                 array--;
1423         }
1424
1425         if (type & PERF_SAMPLE_STREAM_ID) {
1426                 sample->stream_id = *array;
1427                 array--;
1428         }
1429
1430         if (type & PERF_SAMPLE_ID) {
1431                 sample->id = *array;
1432                 array--;
1433         }
1434
1435         if (type & PERF_SAMPLE_TIME) {
1436                 sample->time = *array;
1437                 array--;
1438         }
1439
1440         if (type & PERF_SAMPLE_TID) {
1441                 u.val64 = *array;
1442                 if (swapped) {
1443                         /* undo swap of u64, then swap on individual u32s */
1444                         u.val64 = bswap_64(u.val64);
1445                         u.val32[0] = bswap_32(u.val32[0]);
1446                         u.val32[1] = bswap_32(u.val32[1]);
1447                 }
1448
1449                 sample->pid = u.val32[0];
1450                 sample->tid = u.val32[1];
1451                 array--;
1452         }
1453
1454         return 0;
1455 }
1456
1457 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1458                             u64 size)
1459 {
1460         return size > max_size || offset + size > endp;
1461 }
1462
1463 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1464         do {                                                            \
1465                 if (overflow(endp, (max_size), (offset), (size)))       \
1466                         return -EFAULT;                                 \
1467         } while (0)
1468
1469 #define OVERFLOW_CHECK_u64(offset) \
1470         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1471
1472 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1473                              struct perf_sample *data)
1474 {
1475         u64 type = evsel->attr.sample_type;
1476         bool swapped = evsel->needs_swap;
1477         const u64 *array;
1478         u16 max_size = event->header.size;
1479         const void *endp = (void *)event + max_size;
1480         u64 sz;
1481
1482         /*
1483          * used for cross-endian analysis. See git commit 65014ab3
1484          * for why this goofiness is needed.
1485          */
1486         union u64_swap u;
1487
1488         memset(data, 0, sizeof(*data));
1489         data->cpu = data->pid = data->tid = -1;
1490         data->stream_id = data->id = data->time = -1ULL;
1491         data->period = evsel->attr.sample_period;
1492         data->weight = 0;
1493
1494         if (event->header.type != PERF_RECORD_SAMPLE) {
1495                 if (!evsel->attr.sample_id_all)
1496                         return 0;
1497                 return perf_evsel__parse_id_sample(evsel, event, data);
1498         }
1499
1500         array = event->sample.array;
1501
1502         /*
1503          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1504          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1505          * check the format does not go past the end of the event.
1506          */
1507         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1508                 return -EFAULT;
1509
1510         data->id = -1ULL;
1511         if (type & PERF_SAMPLE_IDENTIFIER) {
1512                 data->id = *array;
1513                 array++;
1514         }
1515
1516         if (type & PERF_SAMPLE_IP) {
1517                 data->ip = *array;
1518                 array++;
1519         }
1520
1521         if (type & PERF_SAMPLE_TID) {
1522                 u.val64 = *array;
1523                 if (swapped) {
1524                         /* undo swap of u64, then swap on individual u32s */
1525                         u.val64 = bswap_64(u.val64);
1526                         u.val32[0] = bswap_32(u.val32[0]);
1527                         u.val32[1] = bswap_32(u.val32[1]);
1528                 }
1529
1530                 data->pid = u.val32[0];
1531                 data->tid = u.val32[1];
1532                 array++;
1533         }
1534
1535         if (type & PERF_SAMPLE_TIME) {
1536                 data->time = *array;
1537                 array++;
1538         }
1539
1540         data->addr = 0;
1541         if (type & PERF_SAMPLE_ADDR) {
1542                 data->addr = *array;
1543                 array++;
1544         }
1545
1546         if (type & PERF_SAMPLE_ID) {
1547                 data->id = *array;
1548                 array++;
1549         }
1550
1551         if (type & PERF_SAMPLE_STREAM_ID) {
1552                 data->stream_id = *array;
1553                 array++;
1554         }
1555
1556         if (type & PERF_SAMPLE_CPU) {
1557
1558                 u.val64 = *array;
1559                 if (swapped) {
1560                         /* undo swap of u64, then swap on individual u32s */
1561                         u.val64 = bswap_64(u.val64);
1562                         u.val32[0] = bswap_32(u.val32[0]);
1563                 }
1564
1565                 data->cpu = u.val32[0];
1566                 array++;
1567         }
1568
1569         if (type & PERF_SAMPLE_PERIOD) {
1570                 data->period = *array;
1571                 array++;
1572         }
1573
1574         if (type & PERF_SAMPLE_READ) {
1575                 u64 read_format = evsel->attr.read_format;
1576
1577                 OVERFLOW_CHECK_u64(array);
1578                 if (read_format & PERF_FORMAT_GROUP)
1579                         data->read.group.nr = *array;
1580                 else
1581                         data->read.one.value = *array;
1582
1583                 array++;
1584
1585                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1586                         OVERFLOW_CHECK_u64(array);
1587                         data->read.time_enabled = *array;
1588                         array++;
1589                 }
1590
1591                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1592                         OVERFLOW_CHECK_u64(array);
1593                         data->read.time_running = *array;
1594                         array++;
1595                 }
1596
1597                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1598                 if (read_format & PERF_FORMAT_GROUP) {
1599                         const u64 max_group_nr = UINT64_MAX /
1600                                         sizeof(struct sample_read_value);
1601
1602                         if (data->read.group.nr > max_group_nr)
1603                                 return -EFAULT;
1604                         sz = data->read.group.nr *
1605                              sizeof(struct sample_read_value);
1606                         OVERFLOW_CHECK(array, sz, max_size);
1607                         data->read.group.values =
1608                                         (struct sample_read_value *)array;
1609                         array = (void *)array + sz;
1610                 } else {
1611                         OVERFLOW_CHECK_u64(array);
1612                         data->read.one.id = *array;
1613                         array++;
1614                 }
1615         }
1616
1617         if (type & PERF_SAMPLE_CALLCHAIN) {
1618                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1619
1620                 OVERFLOW_CHECK_u64(array);
1621                 data->callchain = (struct ip_callchain *)array++;
1622                 if (data->callchain->nr > max_callchain_nr)
1623                         return -EFAULT;
1624                 sz = data->callchain->nr * sizeof(u64);
1625                 OVERFLOW_CHECK(array, sz, max_size);
1626                 array = (void *)array + sz;
1627         }
1628
1629         if (type & PERF_SAMPLE_RAW) {
1630                 OVERFLOW_CHECK_u64(array);
1631                 u.val64 = *array;
1632                 if (WARN_ONCE(swapped,
1633                               "Endianness of raw data not corrected!\n")) {
1634                         /* undo swap of u64, then swap on individual u32s */
1635                         u.val64 = bswap_64(u.val64);
1636                         u.val32[0] = bswap_32(u.val32[0]);
1637                         u.val32[1] = bswap_32(u.val32[1]);
1638                 }
1639                 data->raw_size = u.val32[0];
1640                 array = (void *)array + sizeof(u32);
1641
1642                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1643                 data->raw_data = (void *)array;
1644                 array = (void *)array + data->raw_size;
1645         }
1646
1647         if (type & PERF_SAMPLE_BRANCH_STACK) {
1648                 const u64 max_branch_nr = UINT64_MAX /
1649                                           sizeof(struct branch_entry);
1650
1651                 OVERFLOW_CHECK_u64(array);
1652                 data->branch_stack = (struct branch_stack *)array++;
1653
1654                 if (data->branch_stack->nr > max_branch_nr)
1655                         return -EFAULT;
1656                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1657                 OVERFLOW_CHECK(array, sz, max_size);
1658                 array = (void *)array + sz;
1659         }
1660
1661         if (type & PERF_SAMPLE_REGS_USER) {
1662                 OVERFLOW_CHECK_u64(array);
1663                 data->user_regs.abi = *array;
1664                 array++;
1665
1666                 if (data->user_regs.abi) {
1667                         u64 mask = evsel->attr.sample_regs_user;
1668
1669                         sz = hweight_long(mask) * sizeof(u64);
1670                         OVERFLOW_CHECK(array, sz, max_size);
1671                         data->user_regs.mask = mask;
1672                         data->user_regs.regs = (u64 *)array;
1673                         array = (void *)array + sz;
1674                 }
1675         }
1676
1677         if (type & PERF_SAMPLE_STACK_USER) {
1678                 OVERFLOW_CHECK_u64(array);
1679                 sz = *array++;
1680
1681                 data->user_stack.offset = ((char *)(array - 1)
1682                                           - (char *) event);
1683
1684                 if (!sz) {
1685                         data->user_stack.size = 0;
1686                 } else {
1687                         OVERFLOW_CHECK(array, sz, max_size);
1688                         data->user_stack.data = (char *)array;
1689                         array = (void *)array + sz;
1690                         OVERFLOW_CHECK_u64(array);
1691                         data->user_stack.size = *array++;
1692                         if (WARN_ONCE(data->user_stack.size > sz,
1693                                       "user stack dump failure\n"))
1694                                 return -EFAULT;
1695                 }
1696         }
1697
1698         data->weight = 0;
1699         if (type & PERF_SAMPLE_WEIGHT) {
1700                 OVERFLOW_CHECK_u64(array);
1701                 data->weight = *array;
1702                 array++;
1703         }
1704
1705         data->data_src = PERF_MEM_DATA_SRC_NONE;
1706         if (type & PERF_SAMPLE_DATA_SRC) {
1707                 OVERFLOW_CHECK_u64(array);
1708                 data->data_src = *array;
1709                 array++;
1710         }
1711
1712         data->transaction = 0;
1713         if (type & PERF_SAMPLE_TRANSACTION) {
1714                 OVERFLOW_CHECK_u64(array);
1715                 data->transaction = *array;
1716                 array++;
1717         }
1718
1719         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1720         if (type & PERF_SAMPLE_REGS_INTR) {
1721                 OVERFLOW_CHECK_u64(array);
1722                 data->intr_regs.abi = *array;
1723                 array++;
1724
1725                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1726                         u64 mask = evsel->attr.sample_regs_intr;
1727
1728                         sz = hweight_long(mask) * sizeof(u64);
1729                         OVERFLOW_CHECK(array, sz, max_size);
1730                         data->intr_regs.mask = mask;
1731                         data->intr_regs.regs = (u64 *)array;
1732                         array = (void *)array + sz;
1733                 }
1734         }
1735
1736         return 0;
1737 }
1738
1739 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1740                                      u64 read_format)
1741 {
1742         size_t sz, result = sizeof(struct sample_event);
1743
1744         if (type & PERF_SAMPLE_IDENTIFIER)
1745                 result += sizeof(u64);
1746
1747         if (type & PERF_SAMPLE_IP)
1748                 result += sizeof(u64);
1749
1750         if (type & PERF_SAMPLE_TID)
1751                 result += sizeof(u64);
1752
1753         if (type & PERF_SAMPLE_TIME)
1754                 result += sizeof(u64);
1755
1756         if (type & PERF_SAMPLE_ADDR)
1757                 result += sizeof(u64);
1758
1759         if (type & PERF_SAMPLE_ID)
1760                 result += sizeof(u64);
1761
1762         if (type & PERF_SAMPLE_STREAM_ID)
1763                 result += sizeof(u64);
1764
1765         if (type & PERF_SAMPLE_CPU)
1766                 result += sizeof(u64);
1767
1768         if (type & PERF_SAMPLE_PERIOD)
1769                 result += sizeof(u64);
1770
1771         if (type & PERF_SAMPLE_READ) {
1772                 result += sizeof(u64);
1773                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1774                         result += sizeof(u64);
1775                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1776                         result += sizeof(u64);
1777                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1778                 if (read_format & PERF_FORMAT_GROUP) {
1779                         sz = sample->read.group.nr *
1780                              sizeof(struct sample_read_value);
1781                         result += sz;
1782                 } else {
1783                         result += sizeof(u64);
1784                 }
1785         }
1786
1787         if (type & PERF_SAMPLE_CALLCHAIN) {
1788                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1789                 result += sz;
1790         }
1791
1792         if (type & PERF_SAMPLE_RAW) {
1793                 result += sizeof(u32);
1794                 result += sample->raw_size;
1795         }
1796
1797         if (type & PERF_SAMPLE_BRANCH_STACK) {
1798                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1799                 sz += sizeof(u64);
1800                 result += sz;
1801         }
1802
1803         if (type & PERF_SAMPLE_REGS_USER) {
1804                 if (sample->user_regs.abi) {
1805                         result += sizeof(u64);
1806                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1807                         result += sz;
1808                 } else {
1809                         result += sizeof(u64);
1810                 }
1811         }
1812
1813         if (type & PERF_SAMPLE_STACK_USER) {
1814                 sz = sample->user_stack.size;
1815                 result += sizeof(u64);
1816                 if (sz) {
1817                         result += sz;
1818                         result += sizeof(u64);
1819                 }
1820         }
1821
1822         if (type & PERF_SAMPLE_WEIGHT)
1823                 result += sizeof(u64);
1824
1825         if (type & PERF_SAMPLE_DATA_SRC)
1826                 result += sizeof(u64);
1827
1828         if (type & PERF_SAMPLE_TRANSACTION)
1829                 result += sizeof(u64);
1830
1831         if (type & PERF_SAMPLE_REGS_INTR) {
1832                 if (sample->intr_regs.abi) {
1833                         result += sizeof(u64);
1834                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1835                         result += sz;
1836                 } else {
1837                         result += sizeof(u64);
1838                 }
1839         }
1840
1841         return result;
1842 }
1843
1844 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1845                                   u64 read_format,
1846                                   const struct perf_sample *sample,
1847                                   bool swapped)
1848 {
1849         u64 *array;
1850         size_t sz;
1851         /*
1852          * used for cross-endian analysis. See git commit 65014ab3
1853          * for why this goofiness is needed.
1854          */
1855         union u64_swap u;
1856
1857         array = event->sample.array;
1858
1859         if (type & PERF_SAMPLE_IDENTIFIER) {
1860                 *array = sample->id;
1861                 array++;
1862         }
1863
1864         if (type & PERF_SAMPLE_IP) {
1865                 *array = sample->ip;
1866                 array++;
1867         }
1868
1869         if (type & PERF_SAMPLE_TID) {
1870                 u.val32[0] = sample->pid;
1871                 u.val32[1] = sample->tid;
1872                 if (swapped) {
1873                         /*
1874                          * Inverse of what is done in perf_evsel__parse_sample
1875                          */
1876                         u.val32[0] = bswap_32(u.val32[0]);
1877                         u.val32[1] = bswap_32(u.val32[1]);
1878                         u.val64 = bswap_64(u.val64);
1879                 }
1880
1881                 *array = u.val64;
1882                 array++;
1883         }
1884
1885         if (type & PERF_SAMPLE_TIME) {
1886                 *array = sample->time;
1887                 array++;
1888         }
1889
1890         if (type & PERF_SAMPLE_ADDR) {
1891                 *array = sample->addr;
1892                 array++;
1893         }
1894
1895         if (type & PERF_SAMPLE_ID) {
1896                 *array = sample->id;
1897                 array++;
1898         }
1899
1900         if (type & PERF_SAMPLE_STREAM_ID) {
1901                 *array = sample->stream_id;
1902                 array++;
1903         }
1904
1905         if (type & PERF_SAMPLE_CPU) {
1906                 u.val32[0] = sample->cpu;
1907                 if (swapped) {
1908                         /*
1909                          * Inverse of what is done in perf_evsel__parse_sample
1910                          */
1911                         u.val32[0] = bswap_32(u.val32[0]);
1912                         u.val64 = bswap_64(u.val64);
1913                 }
1914                 *array = u.val64;
1915                 array++;
1916         }
1917
1918         if (type & PERF_SAMPLE_PERIOD) {
1919                 *array = sample->period;
1920                 array++;
1921         }
1922
1923         if (type & PERF_SAMPLE_READ) {
1924                 if (read_format & PERF_FORMAT_GROUP)
1925                         *array = sample->read.group.nr;
1926                 else
1927                         *array = sample->read.one.value;
1928                 array++;
1929
1930                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1931                         *array = sample->read.time_enabled;
1932                         array++;
1933                 }
1934
1935                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1936                         *array = sample->read.time_running;
1937                         array++;
1938                 }
1939
1940                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1941                 if (read_format & PERF_FORMAT_GROUP) {
1942                         sz = sample->read.group.nr *
1943                              sizeof(struct sample_read_value);
1944                         memcpy(array, sample->read.group.values, sz);
1945                         array = (void *)array + sz;
1946                 } else {
1947                         *array = sample->read.one.id;
1948                         array++;
1949                 }
1950         }
1951
1952         if (type & PERF_SAMPLE_CALLCHAIN) {
1953                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1954                 memcpy(array, sample->callchain, sz);
1955                 array = (void *)array + sz;
1956         }
1957
1958         if (type & PERF_SAMPLE_RAW) {
1959                 u.val32[0] = sample->raw_size;
1960                 if (WARN_ONCE(swapped,
1961                               "Endianness of raw data not corrected!\n")) {
1962                         /*
1963                          * Inverse of what is done in perf_evsel__parse_sample
1964                          */
1965                         u.val32[0] = bswap_32(u.val32[0]);
1966                         u.val32[1] = bswap_32(u.val32[1]);
1967                         u.val64 = bswap_64(u.val64);
1968                 }
1969                 *array = u.val64;
1970                 array = (void *)array + sizeof(u32);
1971
1972                 memcpy(array, sample->raw_data, sample->raw_size);
1973                 array = (void *)array + sample->raw_size;
1974         }
1975
1976         if (type & PERF_SAMPLE_BRANCH_STACK) {
1977                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1978                 sz += sizeof(u64);
1979                 memcpy(array, sample->branch_stack, sz);
1980                 array = (void *)array + sz;
1981         }
1982
1983         if (type & PERF_SAMPLE_REGS_USER) {
1984                 if (sample->user_regs.abi) {
1985                         *array++ = sample->user_regs.abi;
1986                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1987                         memcpy(array, sample->user_regs.regs, sz);
1988                         array = (void *)array + sz;
1989                 } else {
1990                         *array++ = 0;
1991                 }
1992         }
1993
1994         if (type & PERF_SAMPLE_STACK_USER) {
1995                 sz = sample->user_stack.size;
1996                 *array++ = sz;
1997                 if (sz) {
1998                         memcpy(array, sample->user_stack.data, sz);
1999                         array = (void *)array + sz;
2000                         *array++ = sz;
2001                 }
2002         }
2003
2004         if (type & PERF_SAMPLE_WEIGHT) {
2005                 *array = sample->weight;
2006                 array++;
2007         }
2008
2009         if (type & PERF_SAMPLE_DATA_SRC) {
2010                 *array = sample->data_src;
2011                 array++;
2012         }
2013
2014         if (type & PERF_SAMPLE_TRANSACTION) {
2015                 *array = sample->transaction;
2016                 array++;
2017         }
2018
2019         if (type & PERF_SAMPLE_REGS_INTR) {
2020                 if (sample->intr_regs.abi) {
2021                         *array++ = sample->intr_regs.abi;
2022                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2023                         memcpy(array, sample->intr_regs.regs, sz);
2024                         array = (void *)array + sz;
2025                 } else {
2026                         *array++ = 0;
2027                 }
2028         }
2029
2030         return 0;
2031 }
2032
2033 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2034 {
2035         return pevent_find_field(evsel->tp_format, name);
2036 }
2037
2038 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2039                          const char *name)
2040 {
2041         struct format_field *field = perf_evsel__field(evsel, name);
2042         int offset;
2043
2044         if (!field)
2045                 return NULL;
2046
2047         offset = field->offset;
2048
2049         if (field->flags & FIELD_IS_DYNAMIC) {
2050                 offset = *(int *)(sample->raw_data + field->offset);
2051                 offset &= 0xffff;
2052         }
2053
2054         return sample->raw_data + offset;
2055 }
2056
2057 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2058                        const char *name)
2059 {
2060         struct format_field *field = perf_evsel__field(evsel, name);
2061         void *ptr;
2062         u64 value;
2063
2064         if (!field)
2065                 return 0;
2066
2067         ptr = sample->raw_data + field->offset;
2068
2069         switch (field->size) {
2070         case 1:
2071                 return *(u8 *)ptr;
2072         case 2:
2073                 value = *(u16 *)ptr;
2074                 break;
2075         case 4:
2076                 value = *(u32 *)ptr;
2077                 break;
2078         case 8:
2079                 memcpy(&value, ptr, sizeof(u64));
2080                 break;
2081         default:
2082                 return 0;
2083         }
2084
2085         if (!evsel->needs_swap)
2086                 return value;
2087
2088         switch (field->size) {
2089         case 2:
2090                 return bswap_16(value);
2091         case 4:
2092                 return bswap_32(value);
2093         case 8:
2094                 return bswap_64(value);
2095         default:
2096                 return 0;
2097         }
2098
2099         return 0;
2100 }
2101
2102 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2103 {
2104         va_list args;
2105         int ret = 0;
2106
2107         if (!*first) {
2108                 ret += fprintf(fp, ",");
2109         } else {
2110                 ret += fprintf(fp, ":");
2111                 *first = false;
2112         }
2113
2114         va_start(args, fmt);
2115         ret += vfprintf(fp, fmt, args);
2116         va_end(args);
2117         return ret;
2118 }
2119
2120 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2121 {
2122         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2123 }
2124
2125 int perf_evsel__fprintf(struct perf_evsel *evsel,
2126                         struct perf_attr_details *details, FILE *fp)
2127 {
2128         bool first = true;
2129         int printed = 0;
2130
2131         if (details->event_group) {
2132                 struct perf_evsel *pos;
2133
2134                 if (!perf_evsel__is_group_leader(evsel))
2135                         return 0;
2136
2137                 if (evsel->nr_members > 1)
2138                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2139
2140                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2141                 for_each_group_member(pos, evsel)
2142                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2143
2144                 if (evsel->nr_members > 1)
2145                         printed += fprintf(fp, "}");
2146                 goto out;
2147         }
2148
2149         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2150
2151         if (details->verbose) {
2152                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2153                                                     __print_attr__fprintf, &first);
2154         } else if (details->freq) {
2155                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2156                                          (u64)evsel->attr.sample_freq);
2157         }
2158 out:
2159         fputc('\n', fp);
2160         return ++printed;
2161 }
2162
2163 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2164                           char *msg, size_t msgsize)
2165 {
2166         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2167             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2168             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2169                 /*
2170                  * If it's cycles then fall back to hrtimer based
2171                  * cpu-clock-tick sw counter, which is always available even if
2172                  * no PMU support.
2173                  *
2174                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2175                  * b0a873e).
2176                  */
2177                 scnprintf(msg, msgsize, "%s",
2178 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2179
2180                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2181                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2182
2183                 zfree(&evsel->name);
2184                 return true;
2185         }
2186
2187         return false;
2188 }
2189
2190 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2191                               int err, char *msg, size_t size)
2192 {
2193         char sbuf[STRERR_BUFSIZE];
2194
2195         switch (err) {
2196         case EPERM:
2197         case EACCES:
2198                 return scnprintf(msg, size,
2199                  "You may not have permission to collect %sstats.\n"
2200                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2201                  " -1 - Not paranoid at all\n"
2202                  "  0 - Disallow raw tracepoint access for unpriv\n"
2203                  "  1 - Disallow cpu events for unpriv\n"
2204                  "  2 - Disallow kernel profiling for unpriv",
2205                                  target->system_wide ? "system-wide " : "");
2206         case ENOENT:
2207                 return scnprintf(msg, size, "The %s event is not supported.",
2208                                  perf_evsel__name(evsel));
2209         case EMFILE:
2210                 return scnprintf(msg, size, "%s",
2211                          "Too many events are opened.\n"
2212                          "Probably the maximum number of open file descriptors has been reached.\n"
2213                          "Hint: Try again after reducing the number of events.\n"
2214                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2215         case ENODEV:
2216                 if (target->cpu_list)
2217                         return scnprintf(msg, size, "%s",
2218          "No such device - did you specify an out-of-range profile CPU?\n");
2219                 break;
2220         case EOPNOTSUPP:
2221                 if (evsel->attr.precise_ip)
2222                         return scnprintf(msg, size, "%s",
2223         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2224 #if defined(__i386__) || defined(__x86_64__)
2225                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2226                         return scnprintf(msg, size, "%s",
2227         "No hardware sampling interrupt available.\n"
2228         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2229 #endif
2230                 break;
2231         case EBUSY:
2232                 if (find_process("oprofiled"))
2233                         return scnprintf(msg, size,
2234         "The PMU counters are busy/taken by another profiler.\n"
2235         "We found oprofile daemon running, please stop it and try again.");
2236                 break;
2237         case EINVAL:
2238                 if (perf_missing_features.clockid)
2239                         return scnprintf(msg, size, "clockid feature not supported.");
2240                 if (perf_missing_features.clockid_wrong)
2241                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2242                 break;
2243         default:
2244                 break;
2245         }
2246
2247         return scnprintf(msg, size,
2248         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2249         "/bin/dmesg may provide additional information.\n"
2250         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2251                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2252                          perf_evsel__name(evsel));
2253 }