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         evsel->evlist      = NULL;
 210         INIT_LIST_HEAD(&evsel->node);
 211         INIT_LIST_HEAD(&evsel->config_terms);
 212         perf_evsel__object.init(evsel);
 213         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 214         perf_evsel__calc_id_pos(evsel);
 215         evsel->cmdline_group_boundary = false;
 216 }
 217
 218 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 219 {
 220         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 221
 222         if (evsel != NULL)
 223                 perf_evsel__init(evsel, attr, idx);
 224
 225         return evsel;
 226 }
 227
 228 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 229 {
 230         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 231
 232         if (evsel != NULL) {
 233                 struct perf_event_attr attr = {
 234                         .type          = PERF_TYPE_TRACEPOINT,
 235                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 236                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 237                 };
 238
 239                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 240                         goto out_free;
 241
 242                 evsel->tp_format = trace_event__tp_format(sys, name);
 243                 if (evsel->tp_format == NULL)
 244                         goto out_free;
 245
 246                 event_attr_init(&attr);
 247                 attr.config = evsel->tp_format->id;
 248                 attr.sample_period = 1;
 249                 perf_evsel__init(evsel, &attr, idx);
 250         }
 251
 252         return evsel;
 253
 254 out_free:
 255         zfree(&evsel->name);
 256         free(evsel);
 257         return NULL;
 258 }
 259
 260 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 261         "cycles",
 262         "instructions",
 263         "cache-references",
 264         "cache-misses",
 265         "branches",
 266         "branch-misses",
 267         "bus-cycles",
 268         "stalled-cycles-frontend",
 269         "stalled-cycles-backend",
 270         "ref-cycles",
 271 };
 272
 273 static const char *__perf_evsel__hw_name(u64 config)
 274 {
 275         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 276                 return perf_evsel__hw_names[config];
 277
 278         return "unknown-hardware";
 279 }
 280
 281 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 282 {
 283         int colon = 0, r = 0;
 284         struct perf_event_attr *attr = &evsel->attr;
 285         bool exclude_guest_default = false;
 286
 287 #define MOD_PRINT(context, mod) do {                                    \
 288                 if (!attr->exclude_##context) {                         \
 289                         if (!colon) colon = ++r;                        \
 290                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 291                 } } while(0)
 292
 293         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 294                 MOD_PRINT(kernel, 'k');
 295                 MOD_PRINT(user, 'u');
 296                 MOD_PRINT(hv, 'h');
 297                 exclude_guest_default = true;
 298         }
 299
 300         if (attr->precise_ip) {
 301                 if (!colon)
 302                         colon = ++r;
 303                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 304                 exclude_guest_default = true;
 305         }
 306
 307         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 308                 MOD_PRINT(host, 'H');
 309                 MOD_PRINT(guest, 'G');
 310         }
 311 #undef MOD_PRINT
 312         if (colon)
 313                 bf[colon - 1] = ':';
 314         return r;
 315 }
 316
 317 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 318 {
 319         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 320         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 321 }
 322
 323 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 324         "cpu-clock",
 325         "task-clock",
 326         "page-faults",
 327         "context-switches",
 328         "cpu-migrations",
 329         "minor-faults",
 330         "major-faults",
 331         "alignment-faults",
 332         "emulation-faults",
 333         "dummy",
 334 };
 335
 336 static const char *__perf_evsel__sw_name(u64 config)
 337 {
 338         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 339                 return perf_evsel__sw_names[config];
 340         return "unknown-software";
 341 }
 342
 343 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 344 {
 345         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 346         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 347 }
 348
 349 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 350 {
 351         int r;
 352
 353         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 354
 355         if (type & HW_BREAKPOINT_R)
 356                 r += scnprintf(bf + r, size - r, "r");
 357
 358         if (type & HW_BREAKPOINT_W)
 359                 r += scnprintf(bf + r, size - r, "w");
 360
 361         if (type & HW_BREAKPOINT_X)
 362                 r += scnprintf(bf + r, size - r, "x");
 363
 364         return r;
 365 }
 366
 367 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 368 {
 369         struct perf_event_attr *attr = &evsel->attr;
 370         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 371         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 372 }
 373
 374 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 375                                 [PERF_EVSEL__MAX_ALIASES] = {
 376  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 377  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 378  { "LLC",       "L2",                                                   },
 379  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 380  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 381  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 382  { "node",                                                              },
 383 };
 384
 385 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 386                                    [PERF_EVSEL__MAX_ALIASES] = {
 387  { "load",      "loads",        "read",                                 },
 388  { "store",     "stores",       "write",                                },
 389  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 390 };
 391
 392 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 393                                        [PERF_EVSEL__MAX_ALIASES] = {
 394  { "refs",      "Reference",    "ops",          "access",               },
 395  { "misses",    "miss",                                                 },
 396 };
 397
 398 #define C(x)            PERF_COUNT_HW_CACHE_##x
 399 #define CACHE_READ      (1 << C(OP_READ))
 400 #define CACHE_WRITE     (1 << C(OP_WRITE))
 401 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 402 #define COP(x)          (1 << x)
 403
 404 /*
 405  * cache operartion stat
 406  * L1I : Read and prefetch only
 407  * ITLB and BPU : Read-only
 408  */
 409 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 410  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 411  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 412  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 413  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 414  [C(ITLB)]      = (CACHE_READ),
 415  [C(BPU)]       = (CACHE_READ),
 416  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 417 };
 418
 419 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 420 {
 421         if (perf_evsel__hw_cache_stat[type] & COP(op))
 422                 return true;    /* valid */
 423         else
 424                 return false;   /* invalid */
 425 }
 426
 427 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 428                                             char *bf, size_t size)
 429 {
 430         if (result) {
 431                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 432                                  perf_evsel__hw_cache_op[op][0],
 433                                  perf_evsel__hw_cache_result[result][0]);
 434         }
 435
 436         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 437                          perf_evsel__hw_cache_op[op][1]);
 438 }
 439
 440 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 441 {
 442         u8 op, result, type = (config >>  0) & 0xff;
 443         const char *err = "unknown-ext-hardware-cache-type";
 444
 445         if (type > PERF_COUNT_HW_CACHE_MAX)
 446                 goto out_err;
 447
 448         op = (config >>  8) & 0xff;
 449         err = "unknown-ext-hardware-cache-op";
 450         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 451                 goto out_err;
 452
 453         result = (config >> 16) & 0xff;
 454         err = "unknown-ext-hardware-cache-result";
 455         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 456                 goto out_err;
 457
 458         err = "invalid-cache";
 459         if (!perf_evsel__is_cache_op_valid(type, op))
 460                 goto out_err;
 461
 462         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 463 out_err:
 464         return scnprintf(bf, size, "%s", err);
 465 }
 466
 467 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 468 {
 469         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 470         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 471 }
 472
 473 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 474 {
 475         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 476         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 477 }
 478
 479 const char *perf_evsel__name(struct perf_evsel *evsel)
 480 {
 481         char bf[128];
 482
 483         if (evsel->name)
 484                 return evsel->name;
 485
 486         switch (evsel->attr.type) {
 487         case PERF_TYPE_RAW:
 488                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 489                 break;
 490
 491         case PERF_TYPE_HARDWARE:
 492                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 493                 break;
 494
 495         case PERF_TYPE_HW_CACHE:
 496                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 497                 break;
 498
 499         case PERF_TYPE_SOFTWARE:
 500                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 501                 break;
 502
 503         case PERF_TYPE_TRACEPOINT:
 504                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 505                 break;
 506
 507         case PERF_TYPE_BREAKPOINT:
 508                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 509                 break;
 510
 511         default:
 512                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 513                           evsel->attr.type);
 514                 break;
 515         }
 516
 517         evsel->name = strdup(bf);
 518
 519         return evsel->name ?: "unknown";
 520 }
 521
 522 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 523 {
 524         return evsel->group_name ?: "anon group";
 525 }
 526
 527 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 528 {
 529         int ret;
 530         struct perf_evsel *pos;
 531         const char *group_name = perf_evsel__group_name(evsel);
 532
 533         ret = scnprintf(buf, size, "%s", group_name);
 534
 535         ret += scnprintf(buf + ret, size - ret, " { %s",
 536                          perf_evsel__name(evsel));
 537
 538         for_each_group_member(pos, evsel)
 539                 ret += scnprintf(buf + ret, size - ret, ", %s",
 540                                  perf_evsel__name(pos));
 541
 542         ret += scnprintf(buf + ret, size - ret, " }");
 543
 544         return ret;
 545 }
 546
 547 static void
 548 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 549                              struct record_opts *opts,
 550                              struct callchain_param *param)
 551 {
 552         bool function = perf_evsel__is_function_event(evsel);
 553         struct perf_event_attr *attr = &evsel->attr;
 554
 555         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 556
 557         if (param->record_mode == CALLCHAIN_LBR) {
 558                 if (!opts->branch_stack) {
 559                         if (attr->exclude_user) {
 560                                 pr_warning("LBR callstack option is only available "
 561                                            "to get user callchain information. "
 562                                            "Falling back to framepointers.\n");
 563                         } else {
 564                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 565                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 566                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 567                         }
 568                 } else
 569                          pr_warning("Cannot use LBR callstack with branch stack. "
 570                                     "Falling back to framepointers.\n");
 571         }
 572
 573         if (param->record_mode == CALLCHAIN_DWARF) {
 574                 if (!function) {
 575                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 576                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 577                         attr->sample_regs_user = PERF_REGS_MASK;
 578                         attr->sample_stack_user = param->dump_size;
 579                         attr->exclude_callchain_user = 1;
 580                 } else {
 581                         pr_info("Cannot use DWARF unwind for function trace event,"
 582                                 " falling back to framepointers.\n");
 583                 }
 584         }
 585
 586         if (function) {
 587                 pr_info("Disabling user space callchains for function trace event.\n");
 588                 attr->exclude_callchain_user = 1;
 589         }
 590 }
 591
 592 static void
 593 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 594                             struct callchain_param *param)
 595 {
 596         struct perf_event_attr *attr = &evsel->attr;
 597
 598         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 599         if (param->record_mode == CALLCHAIN_LBR) {
 600                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 601                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 602                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 603         }
 604         if (param->record_mode == CALLCHAIN_DWARF) {
 605                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 606                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 607         }
 608 }
 609
 610 static void apply_config_terms(struct perf_evsel *evsel,
 611                                struct record_opts *opts)
 612 {
 613         struct perf_evsel_config_term *term;
 614         struct list_head *config_terms = &evsel->config_terms;
 615         struct perf_event_attr *attr = &evsel->attr;
 616         struct callchain_param param;
 617         u32 dump_size = 0;
 618         char *callgraph_buf = NULL;
 619
 620         /* callgraph default */
 621         param.record_mode = callchain_param.record_mode;
 622
 623         list_for_each_entry(term, config_terms, list) {
 624                 switch (term->type) {
 625                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 626                         attr->sample_period = term->val.period;
 627                         attr->freq = 0;
 628                         break;
 629                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 630                         attr->sample_freq = term->val.freq;
 631                         attr->freq = 1;
 632                         break;
 633                 case PERF_EVSEL__CONFIG_TERM_TIME:
 634                         if (term->val.time)
 635                                 perf_evsel__set_sample_bit(evsel, TIME);
 636                         else
 637                                 perf_evsel__reset_sample_bit(evsel, TIME);
 638                         break;
 639                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 640                         callgraph_buf = term->val.callgraph;
 641                         break;
 642                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 643                         dump_size = term->val.stack_user;
 644                         break;
 645                 default:
 646                         break;
 647                 }
 648         }
 649
 650         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 651         if ((callgraph_buf != NULL) || (dump_size > 0)) {
 652
 653                 /* parse callgraph parameters */
 654                 if (callgraph_buf != NULL) {
 655                         if (!strcmp(callgraph_buf, "no")) {
 656                                 param.enabled = false;
 657                                 param.record_mode = CALLCHAIN_NONE;
 658                         } else {
 659                                 param.enabled = true;
 660                                 if (parse_callchain_record(callgraph_buf, &param)) {
 661                                         pr_err("per-event callgraph setting for %s failed. "
 662                                                "Apply callgraph global setting for it\n",
 663                                                evsel->name);
 664                                         return;
 665                                 }
 666                         }
 667                 }
 668                 if (dump_size > 0) {
 669                         dump_size = round_up(dump_size, sizeof(u64));
 670                         param.dump_size = dump_size;
 671                 }
 672
 673                 /* If global callgraph set, clear it */
 674                 if (callchain_param.enabled)
 675                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 676
 677                 /* set perf-event callgraph */
 678                 if (param.enabled)
 679                         perf_evsel__config_callgraph(evsel, opts, &param);
 680         }
 681 }
 682
 683 /*
 684  * The enable_on_exec/disabled value strategy:
 685  *
 686  *  1) For any type of traced program:
 687  *    - all independent events and group leaders are disabled
 688  *    - all group members are enabled
 689  *
 690  *     Group members are ruled by group leaders. They need to
 691  *     be enabled, because the group scheduling relies on that.
 692  *
 693  *  2) For traced programs executed by perf:
 694  *     - all independent events and group leaders have
 695  *       enable_on_exec set
 696  *     - we don't specifically enable or disable any event during
 697  *       the record command
 698  *
 699  *     Independent events and group leaders are initially disabled
 700  *     and get enabled by exec. Group members are ruled by group
 701  *     leaders as stated in 1).
 702  *
 703  *  3) For traced programs attached by perf (pid/tid):
 704  *     - we specifically enable or disable all events during
 705  *       the record command
 706  *
 707  *     When attaching events to already running traced we
 708  *     enable/disable events specifically, as there's no
 709  *     initial traced exec call.
 710  */
 711 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 712 {
 713         struct perf_evsel *leader = evsel->leader;
 714         struct perf_event_attr *attr = &evsel->attr;
 715         int track = evsel->tracking;
 716         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 717
 718         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 719         attr->inherit       = !opts->no_inherit;
 720
 721         perf_evsel__set_sample_bit(evsel, IP);
 722         perf_evsel__set_sample_bit(evsel, TID);
 723
 724         if (evsel->sample_read) {
 725                 perf_evsel__set_sample_bit(evsel, READ);
 726
 727                 /*
 728                  * We need ID even in case of single event, because
 729                  * PERF_SAMPLE_READ process ID specific data.
 730                  */
 731                 perf_evsel__set_sample_id(evsel, false);
 732
 733                 /*
 734                  * Apply group format only if we belong to group
 735                  * with more than one members.
 736                  */
 737                 if (leader->nr_members > 1) {
 738                         attr->read_format |= PERF_FORMAT_GROUP;
 739                         attr->inherit = 0;
 740                 }
 741         }
 742
 743         /*
 744          * We default some events to have a default interval. But keep
 745          * it a weak assumption overridable by the user.
 746          */
 747         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 748                                      opts->user_interval != ULLONG_MAX)) {
 749                 if (opts->freq) {
 750                         perf_evsel__set_sample_bit(evsel, PERIOD);
 751                         attr->freq              = 1;
 752                         attr->sample_freq       = opts->freq;
 753                 } else {
 754                         attr->sample_period = opts->default_interval;
 755                 }
 756         }
 757
 758         /*
 759          * Disable sampling for all group members other
 760          * than leader in case leader 'leads' the sampling.
 761          */
 762         if ((leader != evsel) && leader->sample_read) {
 763                 attr->sample_freq   = 0;
 764                 attr->sample_period = 0;
 765         }
 766
 767         if (opts->no_samples)
 768                 attr->sample_freq = 0;
 769
 770         if (opts->inherit_stat)
 771                 attr->inherit_stat = 1;
 772
 773         if (opts->sample_address) {
 774                 perf_evsel__set_sample_bit(evsel, ADDR);
 775                 attr->mmap_data = track;
 776         }
 777
 778         /*
 779          * We don't allow user space callchains for  function trace
 780          * event, due to issues with page faults while tracing page
 781          * fault handler and its overall trickiness nature.
 782          */
 783         if (perf_evsel__is_function_event(evsel))
 784                 evsel->attr.exclude_callchain_user = 1;
 785
 786         if (callchain_param.enabled && !evsel->no_aux_samples)
 787                 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
 788
 789         if (opts->sample_intr_regs) {
 790                 attr->sample_regs_intr = opts->sample_intr_regs;
 791                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 792         }
 793
 794         if (target__has_cpu(&opts->target))
 795                 perf_evsel__set_sample_bit(evsel, CPU);
 796
 797         if (opts->period)
 798                 perf_evsel__set_sample_bit(evsel, PERIOD);
 799
 800         /*
 801          * When the user explicitely disabled time don't force it here.
 802          */
 803         if (opts->sample_time &&
 804             (!perf_missing_features.sample_id_all &&
 805             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 806              opts->sample_time_set)))
 807                 perf_evsel__set_sample_bit(evsel, TIME);
 808
 809         if (opts->raw_samples && !evsel->no_aux_samples) {
 810                 perf_evsel__set_sample_bit(evsel, TIME);
 811                 perf_evsel__set_sample_bit(evsel, RAW);
 812                 perf_evsel__set_sample_bit(evsel, CPU);
 813         }
 814
 815         if (opts->sample_address)
 816                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 817
 818         if (opts->no_buffering) {
 819                 attr->watermark = 0;
 820                 attr->wakeup_events = 1;
 821         }
 822         if (opts->branch_stack && !evsel->no_aux_samples) {
 823                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 824                 attr->branch_sample_type = opts->branch_stack;
 825         }
 826
 827         if (opts->sample_weight)
 828                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 829
 830         attr->task  = track;
 831         attr->mmap  = track;
 832         attr->mmap2 = track && !perf_missing_features.mmap2;
 833         attr->comm  = track;
 834
 835         if (opts->record_switch_events)
 836                 attr->context_switch = track;
 837
 838         if (opts->sample_transaction)
 839                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 840
 841         if (opts->running_time) {
 842                 evsel->attr.read_format |=
 843                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 844                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 845         }
 846
 847         /*
 848          * XXX see the function comment above
 849          *
 850          * Disabling only independent events or group leaders,
 851          * keeping group members enabled.
 852          */
 853         if (perf_evsel__is_group_leader(evsel))
 854                 attr->disabled = 1;
 855
 856         /*
 857          * Setting enable_on_exec for independent events and
 858          * group leaders for traced executed by perf.
 859          */
 860         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 861                 !opts->initial_delay)
 862                 attr->enable_on_exec = 1;
 863
 864         if (evsel->immediate) {
 865                 attr->disabled = 0;
 866                 attr->enable_on_exec = 0;
 867         }
 868
 869         clockid = opts->clockid;
 870         if (opts->use_clockid) {
 871                 attr->use_clockid = 1;
 872                 attr->clockid = opts->clockid;
 873         }
 874
 875         /*
 876          * Apply event specific term settings,
 877          * it overloads any global configuration.
 878          */
 879         apply_config_terms(evsel, opts);
 880 }
 881
 882 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 883 {
 884         int cpu, thread;
 885
 886         if (evsel->system_wide)
 887                 nthreads = 1;
 888
 889         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 890
 891         if (evsel->fd) {
 892                 for (cpu = 0; cpu < ncpus; cpu++) {
 893                         for (thread = 0; thread < nthreads; thread++) {
 894                                 FD(evsel, cpu, thread) = -1;
 895                         }
 896                 }
 897         }
 898
 899         return evsel->fd != NULL ? 0 : -ENOMEM;
 900 }
 901
 902 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 903                           int ioc,  void *arg)
 904 {
 905         int cpu, thread;
 906
 907         if (evsel->system_wide)
 908                 nthreads = 1;
 909
 910         for (cpu = 0; cpu < ncpus; cpu++) {
 911                 for (thread = 0; thread < nthreads; thread++) {
 912                         int fd = FD(evsel, cpu, thread),
 913                             err = ioctl(fd, ioc, arg);
 914
 915                         if (err)
 916                                 return err;
 917                 }
 918         }
 919
 920         return 0;
 921 }
 922
 923 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 924                              const char *filter)
 925 {
 926         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 927                                      PERF_EVENT_IOC_SET_FILTER,
 928                                      (void *)filter);
 929 }
 930
 931 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 932 {
 933         char *new_filter = strdup(filter);
 934
 935         if (new_filter != NULL) {
 936                 free(evsel->filter);
 937                 evsel->filter = new_filter;
 938                 return 0;
 939         }
 940
 941         return -1;
 942 }
 943
 944 int perf_evsel__append_filter(struct perf_evsel *evsel,
 945                               const char *op, const char *filter)
 946 {
 947         char *new_filter;
 948
 949         if (evsel->filter == NULL)
 950                 return perf_evsel__set_filter(evsel, filter);
 951
 952         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 953                 free(evsel->filter);
 954                 evsel->filter = new_filter;
 955                 return 0;
 956         }
 957
 958         return -1;
 959 }
 960
 961 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 962 {
 963         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 964                                      PERF_EVENT_IOC_ENABLE,
 965                                      0);
 966 }
 967
 968 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 969 {
 970         if (ncpus == 0 || nthreads == 0)
 971                 return 0;
 972
 973         if (evsel->system_wide)
 974                 nthreads = 1;
 975
 976         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 977         if (evsel->sample_id == NULL)
 978                 return -ENOMEM;
 979
 980         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 981         if (evsel->id == NULL) {
 982                 xyarray__delete(evsel->sample_id);
 983                 evsel->sample_id = NULL;
 984                 return -ENOMEM;
 985         }
 986
 987         return 0;
 988 }
 989
 990 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 991 {
 992         xyarray__delete(evsel->fd);
 993         evsel->fd = NULL;
 994 }
 995
 996 static void perf_evsel__free_id(struct perf_evsel *evsel)
 997 {
 998         xyarray__delete(evsel->sample_id);
 999         evsel->sample_id = NULL;
1000         zfree(&evsel->id);
1001 }
1002
1003 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1004 {
1005         struct perf_evsel_config_term *term, *h;
1006
1007         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1008                 list_del(&term->list);
1009                 free(term);
1010         }
1011 }
1012
1013 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1014 {
1015         int cpu, thread;
1016
1017         if (evsel->system_wide)
1018                 nthreads = 1;
1019
1020         for (cpu = 0; cpu < ncpus; cpu++)
1021                 for (thread = 0; thread < nthreads; ++thread) {
1022                         close(FD(evsel, cpu, thread));
1023                         FD(evsel, cpu, thread) = -1;
1024                 }
1025 }
1026
1027 void perf_evsel__exit(struct perf_evsel *evsel)
1028 {
1029         assert(list_empty(&evsel->node));
1030         assert(evsel->evlist == NULL);
1031         perf_evsel__free_fd(evsel);
1032         perf_evsel__free_id(evsel);
1033         perf_evsel__free_config_terms(evsel);
1034         close_cgroup(evsel->cgrp);
1035         cpu_map__put(evsel->cpus);
1036         cpu_map__put(evsel->own_cpus);
1037         thread_map__put(evsel->threads);
1038         zfree(&evsel->group_name);
1039         zfree(&evsel->name);
1040         perf_evsel__object.fini(evsel);
1041 }
1042
1043 void perf_evsel__delete(struct perf_evsel *evsel)
1044 {
1045         perf_evsel__exit(evsel);
1046         free(evsel);
1047 }
1048
1049 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1050                                 struct perf_counts_values *count)
1051 {
1052         struct perf_counts_values tmp;
1053
1054         if (!evsel->prev_raw_counts)
1055                 return;
1056
1057         if (cpu == -1) {
1058                 tmp = evsel->prev_raw_counts->aggr;
1059                 evsel->prev_raw_counts->aggr = *count;
1060         } else {
1061                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1062                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1063         }
1064
1065         count->val = count->val - tmp.val;
1066         count->ena = count->ena - tmp.ena;
1067         count->run = count->run - tmp.run;
1068 }
1069
1070 void perf_counts_values__scale(struct perf_counts_values *count,
1071                                bool scale, s8 *pscaled)
1072 {
1073         s8 scaled = 0;
1074
1075         if (scale) {
1076                 if (count->run == 0) {
1077                         scaled = -1;
1078                         count->val = 0;
1079                 } else if (count->run < count->ena) {
1080                         scaled = 1;
1081                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1082                 }
1083         } else
1084                 count->ena = count->run = 0;
1085
1086         if (pscaled)
1087                 *pscaled = scaled;
1088 }
1089
1090 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1091                      struct perf_counts_values *count)
1092 {
1093         memset(count, 0, sizeof(*count));
1094
1095         if (FD(evsel, cpu, thread) < 0)
1096                 return -EINVAL;
1097
1098         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1099                 return -errno;
1100
1101         return 0;
1102 }
1103
1104 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1105                               int cpu, int thread, bool scale)
1106 {
1107         struct perf_counts_values count;
1108         size_t nv = scale ? 3 : 1;
1109
1110         if (FD(evsel, cpu, thread) < 0)
1111                 return -EINVAL;
1112
1113         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1114                 return -ENOMEM;
1115
1116         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1117                 return -errno;
1118
1119         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1120         perf_counts_values__scale(&count, scale, NULL);
1121         *perf_counts(evsel->counts, cpu, thread) = count;
1122         return 0;
1123 }
1124
1125 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1126 {
1127         struct perf_evsel *leader = evsel->leader;
1128         int fd;
1129
1130         if (perf_evsel__is_group_leader(evsel))
1131                 return -1;
1132
1133         /*
1134          * Leader must be already processed/open,
1135          * if not it's a bug.
1136          */
1137         BUG_ON(!leader->fd);
1138
1139         fd = FD(leader, cpu, thread);
1140         BUG_ON(fd == -1);
1141
1142         return fd;
1143 }
1144
1145 struct bit_names {
1146         int bit;
1147         const char *name;
1148 };
1149
1150 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1151 {
1152         bool first_bit = true;
1153         int i = 0;
1154
1155         do {
1156                 if (value & bits[i].bit) {
1157                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1158                         first_bit = false;
1159                 }
1160         } while (bits[++i].name != NULL);
1161 }
1162
1163 static void __p_sample_type(char *buf, size_t size, u64 value)
1164 {
1165 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1166         struct bit_names bits[] = {
1167                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1168                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1169                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1170                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1171                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1172                 { .name = NULL, }
1173         };
1174 #undef bit_name
1175         __p_bits(buf, size, value, bits);
1176 }
1177
1178 static void __p_read_format(char *buf, size_t size, u64 value)
1179 {
1180 #define bit_name(n) { PERF_FORMAT_##n, #n }
1181         struct bit_names bits[] = {
1182                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1183                 bit_name(ID), bit_name(GROUP),
1184                 { .name = NULL, }
1185         };
1186 #undef bit_name
1187         __p_bits(buf, size, value, bits);
1188 }
1189
1190 #define BUF_SIZE                1024
1191
1192 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1193 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1194 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1195 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1196 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1197
1198 #define PRINT_ATTRn(_n, _f, _p)                         \
1199 do {                                                    \
1200         if (attr->_f) {                                 \
1201                 _p(attr->_f);                           \
1202                 ret += attr__fprintf(fp, _n, buf, priv);\
1203         }                                               \
1204 } while (0)
1205
1206 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1207
1208 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1209                              attr__fprintf_f attr__fprintf, void *priv)
1210 {
1211         char buf[BUF_SIZE];
1212         int ret = 0;
1213
1214         PRINT_ATTRf(type, p_unsigned);
1215         PRINT_ATTRf(size, p_unsigned);
1216         PRINT_ATTRf(config, p_hex);
1217         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1218         PRINT_ATTRf(sample_type, p_sample_type);
1219         PRINT_ATTRf(read_format, p_read_format);
1220
1221         PRINT_ATTRf(disabled, p_unsigned);
1222         PRINT_ATTRf(inherit, p_unsigned);
1223         PRINT_ATTRf(pinned, p_unsigned);
1224         PRINT_ATTRf(exclusive, p_unsigned);
1225         PRINT_ATTRf(exclude_user, p_unsigned);
1226         PRINT_ATTRf(exclude_kernel, p_unsigned);
1227         PRINT_ATTRf(exclude_hv, p_unsigned);
1228         PRINT_ATTRf(exclude_idle, p_unsigned);
1229         PRINT_ATTRf(mmap, p_unsigned);
1230         PRINT_ATTRf(comm, p_unsigned);
1231         PRINT_ATTRf(freq, p_unsigned);
1232         PRINT_ATTRf(inherit_stat, p_unsigned);
1233         PRINT_ATTRf(enable_on_exec, p_unsigned);
1234         PRINT_ATTRf(task, p_unsigned);
1235         PRINT_ATTRf(watermark, p_unsigned);
1236         PRINT_ATTRf(precise_ip, p_unsigned);
1237         PRINT_ATTRf(mmap_data, p_unsigned);
1238         PRINT_ATTRf(sample_id_all, p_unsigned);
1239         PRINT_ATTRf(exclude_host, p_unsigned);
1240         PRINT_ATTRf(exclude_guest, p_unsigned);
1241         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1242         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1243         PRINT_ATTRf(mmap2, p_unsigned);
1244         PRINT_ATTRf(comm_exec, p_unsigned);
1245         PRINT_ATTRf(use_clockid, p_unsigned);
1246         PRINT_ATTRf(context_switch, p_unsigned);
1247
1248         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1249         PRINT_ATTRf(bp_type, p_unsigned);
1250         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1251         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1252         PRINT_ATTRf(sample_regs_user, p_hex);
1253         PRINT_ATTRf(sample_stack_user, p_unsigned);
1254         PRINT_ATTRf(clockid, p_signed);
1255         PRINT_ATTRf(sample_regs_intr, p_hex);
1256         PRINT_ATTRf(aux_watermark, p_unsigned);
1257
1258         return ret;
1259 }
1260
1261 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1262                                 void *priv __attribute__((unused)))
1263 {
1264         return fprintf(fp, "  %-32s %s\n", name, val);
1265 }
1266
1267 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1268                               struct thread_map *threads)
1269 {
1270         int cpu, thread, nthreads;
1271         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1272         int pid = -1, err;
1273         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1274
1275         if (evsel->system_wide)
1276                 nthreads = 1;
1277         else
1278                 nthreads = threads->nr;
1279
1280         if (evsel->fd == NULL &&
1281             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1282                 return -ENOMEM;
1283
1284         if (evsel->cgrp) {
1285                 flags |= PERF_FLAG_PID_CGROUP;
1286                 pid = evsel->cgrp->fd;
1287         }
1288
1289 fallback_missing_features:
1290         if (perf_missing_features.clockid_wrong)
1291                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1292         if (perf_missing_features.clockid) {
1293                 evsel->attr.use_clockid = 0;
1294                 evsel->attr.clockid = 0;
1295         }
1296         if (perf_missing_features.cloexec)
1297                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1298         if (perf_missing_features.mmap2)
1299                 evsel->attr.mmap2 = 0;
1300         if (perf_missing_features.exclude_guest)
1301                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1302 retry_sample_id:
1303         if (perf_missing_features.sample_id_all)
1304                 evsel->attr.sample_id_all = 0;
1305
1306         if (verbose >= 2) {
1307                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1308                 fprintf(stderr, "perf_event_attr:\n");
1309                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1310                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1311         }
1312
1313         for (cpu = 0; cpu < cpus->nr; cpu++) {
1314
1315                 for (thread = 0; thread < nthreads; thread++) {
1316                         int group_fd;
1317
1318                         if (!evsel->cgrp && !evsel->system_wide)
1319                                 pid = thread_map__pid(threads, thread);
1320
1321                         group_fd = get_group_fd(evsel, cpu, thread);
1322 retry_open:
1323                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1324                                   pid, cpus->map[cpu], group_fd, flags);
1325
1326                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1327                                                                      pid,
1328                                                                      cpus->map[cpu],
1329                                                                      group_fd, flags);
1330                         if (FD(evsel, cpu, thread) < 0) {
1331                                 err = -errno;
1332                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1333                                           err);
1334                                 goto try_fallback;
1335                         }
1336                         set_rlimit = NO_CHANGE;
1337
1338                         /*
1339                          * If we succeeded but had to kill clockid, fail and
1340                          * have perf_evsel__open_strerror() print us a nice
1341                          * error.
1342                          */
1343                         if (perf_missing_features.clockid ||
1344                             perf_missing_features.clockid_wrong) {
1345                                 err = -EINVAL;
1346                                 goto out_close;
1347                         }
1348                 }
1349         }
1350
1351         return 0;
1352
1353 try_fallback:
1354         /*
1355          * perf stat needs between 5 and 22 fds per CPU. When we run out
1356          * of them try to increase the limits.
1357          */
1358         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1359                 struct rlimit l;
1360                 int old_errno = errno;
1361
1362                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1363                         if (set_rlimit == NO_CHANGE)
1364                                 l.rlim_cur = l.rlim_max;
1365                         else {
1366                                 l.rlim_cur = l.rlim_max + 1000;
1367                                 l.rlim_max = l.rlim_cur;
1368                         }
1369                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1370                                 set_rlimit++;
1371                                 errno = old_errno;
1372                                 goto retry_open;
1373                         }
1374                 }
1375                 errno = old_errno;
1376         }
1377
1378         if (err != -EINVAL || cpu > 0 || thread > 0)
1379                 goto out_close;
1380
1381         /*
1382          * Must probe features in the order they were added to the
1383          * perf_event_attr interface.
1384          */
1385         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1386                 perf_missing_features.clockid_wrong = true;
1387                 goto fallback_missing_features;
1388         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1389                 perf_missing_features.clockid = true;
1390                 goto fallback_missing_features;
1391         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1392                 perf_missing_features.cloexec = true;
1393                 goto fallback_missing_features;
1394         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1395                 perf_missing_features.mmap2 = true;
1396                 goto fallback_missing_features;
1397         } else if (!perf_missing_features.exclude_guest &&
1398                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1399                 perf_missing_features.exclude_guest = true;
1400                 goto fallback_missing_features;
1401         } else if (!perf_missing_features.sample_id_all) {
1402                 perf_missing_features.sample_id_all = true;
1403                 goto retry_sample_id;
1404         }
1405
1406 out_close:
1407         do {
1408                 while (--thread >= 0) {
1409                         close(FD(evsel, cpu, thread));
1410                         FD(evsel, cpu, thread) = -1;
1411                 }
1412                 thread = nthreads;
1413         } while (--cpu >= 0);
1414         return err;
1415 }
1416
1417 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1418 {
1419         if (evsel->fd == NULL)
1420                 return;
1421
1422         perf_evsel__close_fd(evsel, ncpus, nthreads);
1423         perf_evsel__free_fd(evsel);
1424 }
1425
1426 static struct {
1427         struct cpu_map map;
1428         int cpus[1];
1429 } empty_cpu_map = {
1430         .map.nr = 1,
1431         .cpus   = { -1, },
1432 };
1433
1434 static struct {
1435         struct thread_map map;
1436         int threads[1];
1437 } empty_thread_map = {
1438         .map.nr  = 1,
1439         .threads = { -1, },
1440 };
1441
1442 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1443                      struct thread_map *threads)
1444 {
1445         if (cpus == NULL) {
1446                 /* Work around old compiler warnings about strict aliasing */
1447                 cpus = &empty_cpu_map.map;
1448         }
1449
1450         if (threads == NULL)
1451                 threads = &empty_thread_map.map;
1452
1453         return __perf_evsel__open(evsel, cpus, threads);
1454 }
1455
1456 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1457                              struct cpu_map *cpus)
1458 {
1459         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1460 }
1461
1462 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1463                                 struct thread_map *threads)
1464 {
1465         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1466 }
1467
1468 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1469                                        const union perf_event *event,
1470                                        struct perf_sample *sample)
1471 {
1472         u64 type = evsel->attr.sample_type;
1473         const u64 *array = event->sample.array;
1474         bool swapped = evsel->needs_swap;
1475         union u64_swap u;
1476
1477         array += ((event->header.size -
1478                    sizeof(event->header)) / sizeof(u64)) - 1;
1479
1480         if (type & PERF_SAMPLE_IDENTIFIER) {
1481                 sample->id = *array;
1482                 array--;
1483         }
1484
1485         if (type & PERF_SAMPLE_CPU) {
1486                 u.val64 = *array;
1487                 if (swapped) {
1488                         /* undo swap of u64, then swap on individual u32s */
1489                         u.val64 = bswap_64(u.val64);
1490                         u.val32[0] = bswap_32(u.val32[0]);
1491                 }
1492
1493                 sample->cpu = u.val32[0];
1494                 array--;
1495         }
1496
1497         if (type & PERF_SAMPLE_STREAM_ID) {
1498                 sample->stream_id = *array;
1499                 array--;
1500         }
1501
1502         if (type & PERF_SAMPLE_ID) {
1503                 sample->id = *array;
1504                 array--;
1505         }
1506
1507         if (type & PERF_SAMPLE_TIME) {
1508                 sample->time = *array;
1509                 array--;
1510         }
1511
1512         if (type & PERF_SAMPLE_TID) {
1513                 u.val64 = *array;
1514                 if (swapped) {
1515                         /* undo swap of u64, then swap on individual u32s */
1516                         u.val64 = bswap_64(u.val64);
1517                         u.val32[0] = bswap_32(u.val32[0]);
1518                         u.val32[1] = bswap_32(u.val32[1]);
1519                 }
1520
1521                 sample->pid = u.val32[0];
1522                 sample->tid = u.val32[1];
1523                 array--;
1524         }
1525
1526         return 0;
1527 }
1528
1529 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1530                             u64 size)
1531 {
1532         return size > max_size || offset + size > endp;
1533 }
1534
1535 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1536         do {                                                            \
1537                 if (overflow(endp, (max_size), (offset), (size)))       \
1538                         return -EFAULT;                                 \
1539         } while (0)
1540
1541 #define OVERFLOW_CHECK_u64(offset) \
1542         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1543
1544 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1545                              struct perf_sample *data)
1546 {
1547         u64 type = evsel->attr.sample_type;
1548         bool swapped = evsel->needs_swap;
1549         const u64 *array;
1550         u16 max_size = event->header.size;
1551         const void *endp = (void *)event + max_size;
1552         u64 sz;
1553
1554         /*
1555          * used for cross-endian analysis. See git commit 65014ab3
1556          * for why this goofiness is needed.
1557          */
1558         union u64_swap u;
1559
1560         memset(data, 0, sizeof(*data));
1561         data->cpu = data->pid = data->tid = -1;
1562         data->stream_id = data->id = data->time = -1ULL;
1563         data->period = evsel->attr.sample_period;
1564         data->weight = 0;
1565
1566         if (event->header.type != PERF_RECORD_SAMPLE) {
1567                 if (!evsel->attr.sample_id_all)
1568                         return 0;
1569                 return perf_evsel__parse_id_sample(evsel, event, data);
1570         }
1571
1572         array = event->sample.array;
1573
1574         /*
1575          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1576          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1577          * check the format does not go past the end of the event.
1578          */
1579         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1580                 return -EFAULT;
1581
1582         data->id = -1ULL;
1583         if (type & PERF_SAMPLE_IDENTIFIER) {
1584                 data->id = *array;
1585                 array++;
1586         }
1587
1588         if (type & PERF_SAMPLE_IP) {
1589                 data->ip = *array;
1590                 array++;
1591         }
1592
1593         if (type & PERF_SAMPLE_TID) {
1594                 u.val64 = *array;
1595                 if (swapped) {
1596                         /* undo swap of u64, then swap on individual u32s */
1597                         u.val64 = bswap_64(u.val64);
1598                         u.val32[0] = bswap_32(u.val32[0]);
1599                         u.val32[1] = bswap_32(u.val32[1]);
1600                 }
1601
1602                 data->pid = u.val32[0];
1603                 data->tid = u.val32[1];
1604                 array++;
1605         }
1606
1607         if (type & PERF_SAMPLE_TIME) {
1608                 data->time = *array;
1609                 array++;
1610         }
1611
1612         data->addr = 0;
1613         if (type & PERF_SAMPLE_ADDR) {
1614                 data->addr = *array;
1615                 array++;
1616         }
1617
1618         if (type & PERF_SAMPLE_ID) {
1619                 data->id = *array;
1620                 array++;
1621         }
1622
1623         if (type & PERF_SAMPLE_STREAM_ID) {
1624                 data->stream_id = *array;
1625                 array++;
1626         }
1627
1628         if (type & PERF_SAMPLE_CPU) {
1629
1630                 u.val64 = *array;
1631                 if (swapped) {
1632                         /* undo swap of u64, then swap on individual u32s */
1633                         u.val64 = bswap_64(u.val64);
1634                         u.val32[0] = bswap_32(u.val32[0]);
1635                 }
1636
1637                 data->cpu = u.val32[0];
1638                 array++;
1639         }
1640
1641         if (type & PERF_SAMPLE_PERIOD) {
1642                 data->period = *array;
1643                 array++;
1644         }
1645
1646         if (type & PERF_SAMPLE_READ) {
1647                 u64 read_format = evsel->attr.read_format;
1648
1649                 OVERFLOW_CHECK_u64(array);
1650                 if (read_format & PERF_FORMAT_GROUP)
1651                         data->read.group.nr = *array;
1652                 else
1653                         data->read.one.value = *array;
1654
1655                 array++;
1656
1657                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1658                         OVERFLOW_CHECK_u64(array);
1659                         data->read.time_enabled = *array;
1660                         array++;
1661                 }
1662
1663                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1664                         OVERFLOW_CHECK_u64(array);
1665                         data->read.time_running = *array;
1666                         array++;
1667                 }
1668
1669                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1670                 if (read_format & PERF_FORMAT_GROUP) {
1671                         const u64 max_group_nr = UINT64_MAX /
1672                                         sizeof(struct sample_read_value);
1673
1674                         if (data->read.group.nr > max_group_nr)
1675                                 return -EFAULT;
1676                         sz = data->read.group.nr *
1677                              sizeof(struct sample_read_value);
1678                         OVERFLOW_CHECK(array, sz, max_size);
1679                         data->read.group.values =
1680                                         (struct sample_read_value *)array;
1681                         array = (void *)array + sz;
1682                 } else {
1683                         OVERFLOW_CHECK_u64(array);
1684                         data->read.one.id = *array;
1685                         array++;
1686                 }
1687         }
1688
1689         if (type & PERF_SAMPLE_CALLCHAIN) {
1690                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1691
1692                 OVERFLOW_CHECK_u64(array);
1693                 data->callchain = (struct ip_callchain *)array++;
1694                 if (data->callchain->nr > max_callchain_nr)
1695                         return -EFAULT;
1696                 sz = data->callchain->nr * sizeof(u64);
1697                 OVERFLOW_CHECK(array, sz, max_size);
1698                 array = (void *)array + sz;
1699         }
1700
1701         if (type & PERF_SAMPLE_RAW) {
1702                 OVERFLOW_CHECK_u64(array);
1703                 u.val64 = *array;
1704                 if (WARN_ONCE(swapped,
1705                               "Endianness of raw data not corrected!\n")) {
1706                         /* undo swap of u64, then swap on individual u32s */
1707                         u.val64 = bswap_64(u.val64);
1708                         u.val32[0] = bswap_32(u.val32[0]);
1709                         u.val32[1] = bswap_32(u.val32[1]);
1710                 }
1711                 data->raw_size = u.val32[0];
1712                 array = (void *)array + sizeof(u32);
1713
1714                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1715                 data->raw_data = (void *)array;
1716                 array = (void *)array + data->raw_size;
1717         }
1718
1719         if (type & PERF_SAMPLE_BRANCH_STACK) {
1720                 const u64 max_branch_nr = UINT64_MAX /
1721                                           sizeof(struct branch_entry);
1722
1723                 OVERFLOW_CHECK_u64(array);
1724                 data->branch_stack = (struct branch_stack *)array++;
1725
1726                 if (data->branch_stack->nr > max_branch_nr)
1727                         return -EFAULT;
1728                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1729                 OVERFLOW_CHECK(array, sz, max_size);
1730                 array = (void *)array + sz;
1731         }
1732
1733         if (type & PERF_SAMPLE_REGS_USER) {
1734                 OVERFLOW_CHECK_u64(array);
1735                 data->user_regs.abi = *array;
1736                 array++;
1737
1738                 if (data->user_regs.abi) {
1739                         u64 mask = evsel->attr.sample_regs_user;
1740
1741                         sz = hweight_long(mask) * sizeof(u64);
1742                         OVERFLOW_CHECK(array, sz, max_size);
1743                         data->user_regs.mask = mask;
1744                         data->user_regs.regs = (u64 *)array;
1745                         array = (void *)array + sz;
1746                 }
1747         }
1748
1749         if (type & PERF_SAMPLE_STACK_USER) {
1750                 OVERFLOW_CHECK_u64(array);
1751                 sz = *array++;
1752
1753                 data->user_stack.offset = ((char *)(array - 1)
1754                                           - (char *) event);
1755
1756                 if (!sz) {
1757                         data->user_stack.size = 0;
1758                 } else {
1759                         OVERFLOW_CHECK(array, sz, max_size);
1760                         data->user_stack.data = (char *)array;
1761                         array = (void *)array + sz;
1762                         OVERFLOW_CHECK_u64(array);
1763                         data->user_stack.size = *array++;
1764                         if (WARN_ONCE(data->user_stack.size > sz,
1765                                       "user stack dump failure\n"))
1766                                 return -EFAULT;
1767                 }
1768         }
1769
1770         data->weight = 0;
1771         if (type & PERF_SAMPLE_WEIGHT) {
1772                 OVERFLOW_CHECK_u64(array);
1773                 data->weight = *array;
1774                 array++;
1775         }
1776
1777         data->data_src = PERF_MEM_DATA_SRC_NONE;
1778         if (type & PERF_SAMPLE_DATA_SRC) {
1779                 OVERFLOW_CHECK_u64(array);
1780                 data->data_src = *array;
1781                 array++;
1782         }
1783
1784         data->transaction = 0;
1785         if (type & PERF_SAMPLE_TRANSACTION) {
1786                 OVERFLOW_CHECK_u64(array);
1787                 data->transaction = *array;
1788                 array++;
1789         }
1790
1791         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1792         if (type & PERF_SAMPLE_REGS_INTR) {
1793                 OVERFLOW_CHECK_u64(array);
1794                 data->intr_regs.abi = *array;
1795                 array++;
1796
1797                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1798                         u64 mask = evsel->attr.sample_regs_intr;
1799
1800                         sz = hweight_long(mask) * sizeof(u64);
1801                         OVERFLOW_CHECK(array, sz, max_size);
1802                         data->intr_regs.mask = mask;
1803                         data->intr_regs.regs = (u64 *)array;
1804                         array = (void *)array + sz;
1805                 }
1806         }
1807
1808         return 0;
1809 }
1810
1811 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1812                                      u64 read_format)
1813 {
1814         size_t sz, result = sizeof(struct sample_event);
1815
1816         if (type & PERF_SAMPLE_IDENTIFIER)
1817                 result += sizeof(u64);
1818
1819         if (type & PERF_SAMPLE_IP)
1820                 result += sizeof(u64);
1821
1822         if (type & PERF_SAMPLE_TID)
1823                 result += sizeof(u64);
1824
1825         if (type & PERF_SAMPLE_TIME)
1826                 result += sizeof(u64);
1827
1828         if (type & PERF_SAMPLE_ADDR)
1829                 result += sizeof(u64);
1830
1831         if (type & PERF_SAMPLE_ID)
1832                 result += sizeof(u64);
1833
1834         if (type & PERF_SAMPLE_STREAM_ID)
1835                 result += sizeof(u64);
1836
1837         if (type & PERF_SAMPLE_CPU)
1838                 result += sizeof(u64);
1839
1840         if (type & PERF_SAMPLE_PERIOD)
1841                 result += sizeof(u64);
1842
1843         if (type & PERF_SAMPLE_READ) {
1844                 result += sizeof(u64);
1845                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1846                         result += sizeof(u64);
1847                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1848                         result += sizeof(u64);
1849                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1850                 if (read_format & PERF_FORMAT_GROUP) {
1851                         sz = sample->read.group.nr *
1852                              sizeof(struct sample_read_value);
1853                         result += sz;
1854                 } else {
1855                         result += sizeof(u64);
1856                 }
1857         }
1858
1859         if (type & PERF_SAMPLE_CALLCHAIN) {
1860                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1861                 result += sz;
1862         }
1863
1864         if (type & PERF_SAMPLE_RAW) {
1865                 result += sizeof(u32);
1866                 result += sample->raw_size;
1867         }
1868
1869         if (type & PERF_SAMPLE_BRANCH_STACK) {
1870                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1871                 sz += sizeof(u64);
1872                 result += sz;
1873         }
1874
1875         if (type & PERF_SAMPLE_REGS_USER) {
1876                 if (sample->user_regs.abi) {
1877                         result += sizeof(u64);
1878                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1879                         result += sz;
1880                 } else {
1881                         result += sizeof(u64);
1882                 }
1883         }
1884
1885         if (type & PERF_SAMPLE_STACK_USER) {
1886                 sz = sample->user_stack.size;
1887                 result += sizeof(u64);
1888                 if (sz) {
1889                         result += sz;
1890                         result += sizeof(u64);
1891                 }
1892         }
1893
1894         if (type & PERF_SAMPLE_WEIGHT)
1895                 result += sizeof(u64);
1896
1897         if (type & PERF_SAMPLE_DATA_SRC)
1898                 result += sizeof(u64);
1899
1900         if (type & PERF_SAMPLE_TRANSACTION)
1901                 result += sizeof(u64);
1902
1903         if (type & PERF_SAMPLE_REGS_INTR) {
1904                 if (sample->intr_regs.abi) {
1905                         result += sizeof(u64);
1906                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1907                         result += sz;
1908                 } else {
1909                         result += sizeof(u64);
1910                 }
1911         }
1912
1913         return result;
1914 }
1915
1916 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1917                                   u64 read_format,
1918                                   const struct perf_sample *sample,
1919                                   bool swapped)
1920 {
1921         u64 *array;
1922         size_t sz;
1923         /*
1924          * used for cross-endian analysis. See git commit 65014ab3
1925          * for why this goofiness is needed.
1926          */
1927         union u64_swap u;
1928
1929         array = event->sample.array;
1930
1931         if (type & PERF_SAMPLE_IDENTIFIER) {
1932                 *array = sample->id;
1933                 array++;
1934         }
1935
1936         if (type & PERF_SAMPLE_IP) {
1937                 *array = sample->ip;
1938                 array++;
1939         }
1940
1941         if (type & PERF_SAMPLE_TID) {
1942                 u.val32[0] = sample->pid;
1943                 u.val32[1] = sample->tid;
1944                 if (swapped) {
1945                         /*
1946                          * Inverse of what is done in perf_evsel__parse_sample
1947                          */
1948                         u.val32[0] = bswap_32(u.val32[0]);
1949                         u.val32[1] = bswap_32(u.val32[1]);
1950                         u.val64 = bswap_64(u.val64);
1951                 }
1952
1953                 *array = u.val64;
1954                 array++;
1955         }
1956
1957         if (type & PERF_SAMPLE_TIME) {
1958                 *array = sample->time;
1959                 array++;
1960         }
1961
1962         if (type & PERF_SAMPLE_ADDR) {
1963                 *array = sample->addr;
1964                 array++;
1965         }
1966
1967         if (type & PERF_SAMPLE_ID) {
1968                 *array = sample->id;
1969                 array++;
1970         }
1971
1972         if (type & PERF_SAMPLE_STREAM_ID) {
1973                 *array = sample->stream_id;
1974                 array++;
1975         }
1976
1977         if (type & PERF_SAMPLE_CPU) {
1978                 u.val32[0] = sample->cpu;
1979                 if (swapped) {
1980                         /*
1981                          * Inverse of what is done in perf_evsel__parse_sample
1982                          */
1983                         u.val32[0] = bswap_32(u.val32[0]);
1984                         u.val64 = bswap_64(u.val64);
1985                 }
1986                 *array = u.val64;
1987                 array++;
1988         }
1989
1990         if (type & PERF_SAMPLE_PERIOD) {
1991                 *array = sample->period;
1992                 array++;
1993         }
1994
1995         if (type & PERF_SAMPLE_READ) {
1996                 if (read_format & PERF_FORMAT_GROUP)
1997                         *array = sample->read.group.nr;
1998                 else
1999                         *array = sample->read.one.value;
2000                 array++;
2001
2002                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2003                         *array = sample->read.time_enabled;
2004                         array++;
2005                 }
2006
2007                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2008                         *array = sample->read.time_running;
2009                         array++;
2010                 }
2011
2012                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2013                 if (read_format & PERF_FORMAT_GROUP) {
2014                         sz = sample->read.group.nr *
2015                              sizeof(struct sample_read_value);
2016                         memcpy(array, sample->read.group.values, sz);
2017                         array = (void *)array + sz;
2018                 } else {
2019                         *array = sample->read.one.id;
2020                         array++;
2021                 }
2022         }
2023
2024         if (type & PERF_SAMPLE_CALLCHAIN) {
2025                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2026                 memcpy(array, sample->callchain, sz);
2027                 array = (void *)array + sz;
2028         }
2029
2030         if (type & PERF_SAMPLE_RAW) {
2031                 u.val32[0] = sample->raw_size;
2032                 if (WARN_ONCE(swapped,
2033                               "Endianness of raw data not corrected!\n")) {
2034                         /*
2035                          * Inverse of what is done in perf_evsel__parse_sample
2036                          */
2037                         u.val32[0] = bswap_32(u.val32[0]);
2038                         u.val32[1] = bswap_32(u.val32[1]);
2039                         u.val64 = bswap_64(u.val64);
2040                 }
2041                 *array = u.val64;
2042                 array = (void *)array + sizeof(u32);
2043
2044                 memcpy(array, sample->raw_data, sample->raw_size);
2045                 array = (void *)array + sample->raw_size;
2046         }
2047
2048         if (type & PERF_SAMPLE_BRANCH_STACK) {
2049                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2050                 sz += sizeof(u64);
2051                 memcpy(array, sample->branch_stack, sz);
2052                 array = (void *)array + sz;
2053         }
2054
2055         if (type & PERF_SAMPLE_REGS_USER) {
2056                 if (sample->user_regs.abi) {
2057                         *array++ = sample->user_regs.abi;
2058                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2059                         memcpy(array, sample->user_regs.regs, sz);
2060                         array = (void *)array + sz;
2061                 } else {
2062                         *array++ = 0;
2063                 }
2064         }
2065
2066         if (type & PERF_SAMPLE_STACK_USER) {
2067                 sz = sample->user_stack.size;
2068                 *array++ = sz;
2069                 if (sz) {
2070                         memcpy(array, sample->user_stack.data, sz);
2071                         array = (void *)array + sz;
2072                         *array++ = sz;
2073                 }
2074         }
2075
2076         if (type & PERF_SAMPLE_WEIGHT) {
2077                 *array = sample->weight;
2078                 array++;
2079         }
2080
2081         if (type & PERF_SAMPLE_DATA_SRC) {
2082                 *array = sample->data_src;
2083                 array++;
2084         }
2085
2086         if (type & PERF_SAMPLE_TRANSACTION) {
2087                 *array = sample->transaction;
2088                 array++;
2089         }
2090
2091         if (type & PERF_SAMPLE_REGS_INTR) {
2092                 if (sample->intr_regs.abi) {
2093                         *array++ = sample->intr_regs.abi;
2094                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2095                         memcpy(array, sample->intr_regs.regs, sz);
2096                         array = (void *)array + sz;
2097                 } else {
2098                         *array++ = 0;
2099                 }
2100         }
2101
2102         return 0;
2103 }
2104
2105 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2106 {
2107         return pevent_find_field(evsel->tp_format, name);
2108 }
2109
2110 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2111                          const char *name)
2112 {
2113         struct format_field *field = perf_evsel__field(evsel, name);
2114         int offset;
2115
2116         if (!field)
2117                 return NULL;
2118
2119         offset = field->offset;
2120
2121         if (field->flags & FIELD_IS_DYNAMIC) {
2122                 offset = *(int *)(sample->raw_data + field->offset);
2123                 offset &= 0xffff;
2124         }
2125
2126         return sample->raw_data + offset;
2127 }
2128
2129 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2130                        const char *name)
2131 {
2132         struct format_field *field = perf_evsel__field(evsel, name);
2133         void *ptr;
2134         u64 value;
2135
2136         if (!field)
2137                 return 0;
2138
2139         ptr = sample->raw_data + field->offset;
2140
2141         switch (field->size) {
2142         case 1:
2143                 return *(u8 *)ptr;
2144         case 2:
2145                 value = *(u16 *)ptr;
2146                 break;
2147         case 4:
2148                 value = *(u32 *)ptr;
2149                 break;
2150         case 8:
2151                 memcpy(&value, ptr, sizeof(u64));
2152                 break;
2153         default:
2154                 return 0;
2155         }
2156
2157         if (!evsel->needs_swap)
2158                 return value;
2159
2160         switch (field->size) {
2161         case 2:
2162                 return bswap_16(value);
2163         case 4:
2164                 return bswap_32(value);
2165         case 8:
2166                 return bswap_64(value);
2167         default:
2168                 return 0;
2169         }
2170
2171         return 0;
2172 }
2173
2174 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2175 {
2176         va_list args;
2177         int ret = 0;
2178
2179         if (!*first) {
2180                 ret += fprintf(fp, ",");
2181         } else {
2182                 ret += fprintf(fp, ":");
2183                 *first = false;
2184         }
2185
2186         va_start(args, fmt);
2187         ret += vfprintf(fp, fmt, args);
2188         va_end(args);
2189         return ret;
2190 }
2191
2192 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2193 {
2194         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2195 }
2196
2197 int perf_evsel__fprintf(struct perf_evsel *evsel,
2198                         struct perf_attr_details *details, FILE *fp)
2199 {
2200         bool first = true;
2201         int printed = 0;
2202
2203         if (details->event_group) {
2204                 struct perf_evsel *pos;
2205
2206                 if (!perf_evsel__is_group_leader(evsel))
2207                         return 0;
2208
2209                 if (evsel->nr_members > 1)
2210                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2211
2212                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2213                 for_each_group_member(pos, evsel)
2214                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2215
2216                 if (evsel->nr_members > 1)
2217                         printed += fprintf(fp, "}");
2218                 goto out;
2219         }
2220
2221         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2222
2223         if (details->verbose) {
2224                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2225                                                     __print_attr__fprintf, &first);
2226         } else if (details->freq) {
2227                 const char *term = "sample_freq";
2228
2229                 if (!evsel->attr.freq)
2230                         term = "sample_period";
2231
2232                 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2233                                          term, (u64)evsel->attr.sample_freq);
2234         }
2235 out:
2236         fputc('\n', fp);
2237         return ++printed;
2238 }
2239
2240 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2241                           char *msg, size_t msgsize)
2242 {
2243         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2244             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2245             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2246                 /*
2247                  * If it's cycles then fall back to hrtimer based
2248                  * cpu-clock-tick sw counter, which is always available even if
2249                  * no PMU support.
2250                  *
2251                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2252                  * b0a873e).
2253                  */
2254                 scnprintf(msg, msgsize, "%s",
2255 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2256
2257                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2258                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2259
2260                 zfree(&evsel->name);
2261                 return true;
2262         }
2263
2264         return false;
2265 }
2266
2267 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2268                               int err, char *msg, size_t size)
2269 {
2270         char sbuf[STRERR_BUFSIZE];
2271
2272         switch (err) {
2273         case EPERM:
2274         case EACCES:
2275                 return scnprintf(msg, size,
2276                  "You may not have permission to collect %sstats.\n"
2277                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2278                  " -1 - Not paranoid at all\n"
2279                  "  0 - Disallow raw tracepoint access for unpriv\n"
2280                  "  1 - Disallow cpu events for unpriv\n"
2281                  "  2 - Disallow kernel profiling for unpriv",
2282                                  target->system_wide ? "system-wide " : "");
2283         case ENOENT:
2284                 return scnprintf(msg, size, "The %s event is not supported.",
2285                                  perf_evsel__name(evsel));
2286         case EMFILE:
2287                 return scnprintf(msg, size, "%s",
2288                          "Too many events are opened.\n"
2289                          "Probably the maximum number of open file descriptors has been reached.\n"
2290                          "Hint: Try again after reducing the number of events.\n"
2291                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2292         case ENODEV:
2293                 if (target->cpu_list)
2294                         return scnprintf(msg, size, "%s",
2295          "No such device - did you specify an out-of-range profile CPU?\n");
2296                 break;
2297         case EOPNOTSUPP:
2298                 if (evsel->attr.precise_ip)
2299                         return scnprintf(msg, size, "%s",
2300         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2301 #if defined(__i386__) || defined(__x86_64__)
2302                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2303                         return scnprintf(msg, size, "%s",
2304         "No hardware sampling interrupt available.\n"
2305         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2306 #endif
2307                 break;
2308         case EBUSY:
2309                 if (find_process("oprofiled"))
2310                         return scnprintf(msg, size,
2311         "The PMU counters are busy/taken by another profiler.\n"
2312         "We found oprofile daemon running, please stop it and try again.");
2313                 break;
2314         case EINVAL:
2315                 if (perf_missing_features.clockid)
2316                         return scnprintf(msg, size, "clockid feature not supported.");
2317                 if (perf_missing_features.clockid_wrong)
2318                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2319                 break;
2320         default:
2321                 break;
2322         }
2323
2324         return scnprintf(msg, size,
2325         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2326         "/bin/dmesg may provide additional information.\n"
2327         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2328                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2329                          perf_evsel__name(evsel));
2330 }