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         thread_map__put(evsel->threads);
1037         zfree(&evsel->group_name);
1038         zfree(&evsel->name);
1039         perf_evsel__object.fini(evsel);
1040 }
1041
1042 void perf_evsel__delete(struct perf_evsel *evsel)
1043 {
1044         perf_evsel__exit(evsel);
1045         free(evsel);
1046 }
1047
1048 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1049                                 struct perf_counts_values *count)
1050 {
1051         struct perf_counts_values tmp;
1052
1053         if (!evsel->prev_raw_counts)
1054                 return;
1055
1056         if (cpu == -1) {
1057                 tmp = evsel->prev_raw_counts->aggr;
1058                 evsel->prev_raw_counts->aggr = *count;
1059         } else {
1060                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1061                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1062         }
1063
1064         count->val = count->val - tmp.val;
1065         count->ena = count->ena - tmp.ena;
1066         count->run = count->run - tmp.run;
1067 }
1068
1069 void perf_counts_values__scale(struct perf_counts_values *count,
1070                                bool scale, s8 *pscaled)
1071 {
1072         s8 scaled = 0;
1073
1074         if (scale) {
1075                 if (count->run == 0) {
1076                         scaled = -1;
1077                         count->val = 0;
1078                 } else if (count->run < count->ena) {
1079                         scaled = 1;
1080                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1081                 }
1082         } else
1083                 count->ena = count->run = 0;
1084
1085         if (pscaled)
1086                 *pscaled = scaled;
1087 }
1088
1089 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1090                      struct perf_counts_values *count)
1091 {
1092         memset(count, 0, sizeof(*count));
1093
1094         if (FD(evsel, cpu, thread) < 0)
1095                 return -EINVAL;
1096
1097         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1098                 return -errno;
1099
1100         return 0;
1101 }
1102
1103 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1104                               int cpu, int thread, bool scale)
1105 {
1106         struct perf_counts_values count;
1107         size_t nv = scale ? 3 : 1;
1108
1109         if (FD(evsel, cpu, thread) < 0)
1110                 return -EINVAL;
1111
1112         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1113                 return -ENOMEM;
1114
1115         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1116                 return -errno;
1117
1118         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1119         perf_counts_values__scale(&count, scale, NULL);
1120         *perf_counts(evsel->counts, cpu, thread) = count;
1121         return 0;
1122 }
1123
1124 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1125 {
1126         struct perf_evsel *leader = evsel->leader;
1127         int fd;
1128
1129         if (perf_evsel__is_group_leader(evsel))
1130                 return -1;
1131
1132         /*
1133          * Leader must be already processed/open,
1134          * if not it's a bug.
1135          */
1136         BUG_ON(!leader->fd);
1137
1138         fd = FD(leader, cpu, thread);
1139         BUG_ON(fd == -1);
1140
1141         return fd;
1142 }
1143
1144 struct bit_names {
1145         int bit;
1146         const char *name;
1147 };
1148
1149 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1150 {
1151         bool first_bit = true;
1152         int i = 0;
1153
1154         do {
1155                 if (value & bits[i].bit) {
1156                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1157                         first_bit = false;
1158                 }
1159         } while (bits[++i].name != NULL);
1160 }
1161
1162 static void __p_sample_type(char *buf, size_t size, u64 value)
1163 {
1164 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1165         struct bit_names bits[] = {
1166                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1167                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1168                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1169                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1170                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1171                 { .name = NULL, }
1172         };
1173 #undef bit_name
1174         __p_bits(buf, size, value, bits);
1175 }
1176
1177 static void __p_read_format(char *buf, size_t size, u64 value)
1178 {
1179 #define bit_name(n) { PERF_FORMAT_##n, #n }
1180         struct bit_names bits[] = {
1181                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1182                 bit_name(ID), bit_name(GROUP),
1183                 { .name = NULL, }
1184         };
1185 #undef bit_name
1186         __p_bits(buf, size, value, bits);
1187 }
1188
1189 #define BUF_SIZE                1024
1190
1191 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1192 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1193 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1194 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1195 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1196
1197 #define PRINT_ATTRn(_n, _f, _p)                         \
1198 do {                                                    \
1199         if (attr->_f) {                                 \
1200                 _p(attr->_f);                           \
1201                 ret += attr__fprintf(fp, _n, buf, priv);\
1202         }                                               \
1203 } while (0)
1204
1205 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1206
1207 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1208                              attr__fprintf_f attr__fprintf, void *priv)
1209 {
1210         char buf[BUF_SIZE];
1211         int ret = 0;
1212
1213         PRINT_ATTRf(type, p_unsigned);
1214         PRINT_ATTRf(size, p_unsigned);
1215         PRINT_ATTRf(config, p_hex);
1216         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1217         PRINT_ATTRf(sample_type, p_sample_type);
1218         PRINT_ATTRf(read_format, p_read_format);
1219
1220         PRINT_ATTRf(disabled, p_unsigned);
1221         PRINT_ATTRf(inherit, p_unsigned);
1222         PRINT_ATTRf(pinned, p_unsigned);
1223         PRINT_ATTRf(exclusive, p_unsigned);
1224         PRINT_ATTRf(exclude_user, p_unsigned);
1225         PRINT_ATTRf(exclude_kernel, p_unsigned);
1226         PRINT_ATTRf(exclude_hv, p_unsigned);
1227         PRINT_ATTRf(exclude_idle, p_unsigned);
1228         PRINT_ATTRf(mmap, p_unsigned);
1229         PRINT_ATTRf(comm, p_unsigned);
1230         PRINT_ATTRf(freq, p_unsigned);
1231         PRINT_ATTRf(inherit_stat, p_unsigned);
1232         PRINT_ATTRf(enable_on_exec, p_unsigned);
1233         PRINT_ATTRf(task, p_unsigned);
1234         PRINT_ATTRf(watermark, p_unsigned);
1235         PRINT_ATTRf(precise_ip, p_unsigned);
1236         PRINT_ATTRf(mmap_data, p_unsigned);
1237         PRINT_ATTRf(sample_id_all, p_unsigned);
1238         PRINT_ATTRf(exclude_host, p_unsigned);
1239         PRINT_ATTRf(exclude_guest, p_unsigned);
1240         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1241         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1242         PRINT_ATTRf(mmap2, p_unsigned);
1243         PRINT_ATTRf(comm_exec, p_unsigned);
1244         PRINT_ATTRf(use_clockid, p_unsigned);
1245         PRINT_ATTRf(context_switch, p_unsigned);
1246
1247         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1248         PRINT_ATTRf(bp_type, p_unsigned);
1249         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1250         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1251         PRINT_ATTRf(sample_regs_user, p_hex);
1252         PRINT_ATTRf(sample_stack_user, p_unsigned);
1253         PRINT_ATTRf(clockid, p_signed);
1254         PRINT_ATTRf(sample_regs_intr, p_hex);
1255         PRINT_ATTRf(aux_watermark, p_unsigned);
1256
1257         return ret;
1258 }
1259
1260 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1261                                 void *priv __attribute__((unused)))
1262 {
1263         return fprintf(fp, "  %-32s %s\n", name, val);
1264 }
1265
1266 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1267                               struct thread_map *threads)
1268 {
1269         int cpu, thread, nthreads;
1270         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1271         int pid = -1, err;
1272         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1273
1274         if (evsel->system_wide)
1275                 nthreads = 1;
1276         else
1277                 nthreads = threads->nr;
1278
1279         if (evsel->fd == NULL &&
1280             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1281                 return -ENOMEM;
1282
1283         if (evsel->cgrp) {
1284                 flags |= PERF_FLAG_PID_CGROUP;
1285                 pid = evsel->cgrp->fd;
1286         }
1287
1288 fallback_missing_features:
1289         if (perf_missing_features.clockid_wrong)
1290                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1291         if (perf_missing_features.clockid) {
1292                 evsel->attr.use_clockid = 0;
1293                 evsel->attr.clockid = 0;
1294         }
1295         if (perf_missing_features.cloexec)
1296                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1297         if (perf_missing_features.mmap2)
1298                 evsel->attr.mmap2 = 0;
1299         if (perf_missing_features.exclude_guest)
1300                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1301 retry_sample_id:
1302         if (perf_missing_features.sample_id_all)
1303                 evsel->attr.sample_id_all = 0;
1304
1305         if (verbose >= 2) {
1306                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1307                 fprintf(stderr, "perf_event_attr:\n");
1308                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1309                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1310         }
1311
1312         for (cpu = 0; cpu < cpus->nr; cpu++) {
1313
1314                 for (thread = 0; thread < nthreads; thread++) {
1315                         int group_fd;
1316
1317                         if (!evsel->cgrp && !evsel->system_wide)
1318                                 pid = thread_map__pid(threads, thread);
1319
1320                         group_fd = get_group_fd(evsel, cpu, thread);
1321 retry_open:
1322                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1323                                   pid, cpus->map[cpu], group_fd, flags);
1324
1325                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1326                                                                      pid,
1327                                                                      cpus->map[cpu],
1328                                                                      group_fd, flags);
1329                         if (FD(evsel, cpu, thread) < 0) {
1330                                 err = -errno;
1331                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1332                                           err);
1333                                 goto try_fallback;
1334                         }
1335                         set_rlimit = NO_CHANGE;
1336
1337                         /*
1338                          * If we succeeded but had to kill clockid, fail and
1339                          * have perf_evsel__open_strerror() print us a nice
1340                          * error.
1341                          */
1342                         if (perf_missing_features.clockid ||
1343                             perf_missing_features.clockid_wrong) {
1344                                 err = -EINVAL;
1345                                 goto out_close;
1346                         }
1347                 }
1348         }
1349
1350         return 0;
1351
1352 try_fallback:
1353         /*
1354          * perf stat needs between 5 and 22 fds per CPU. When we run out
1355          * of them try to increase the limits.
1356          */
1357         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1358                 struct rlimit l;
1359                 int old_errno = errno;
1360
1361                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1362                         if (set_rlimit == NO_CHANGE)
1363                                 l.rlim_cur = l.rlim_max;
1364                         else {
1365                                 l.rlim_cur = l.rlim_max + 1000;
1366                                 l.rlim_max = l.rlim_cur;
1367                         }
1368                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1369                                 set_rlimit++;
1370                                 errno = old_errno;
1371                                 goto retry_open;
1372                         }
1373                 }
1374                 errno = old_errno;
1375         }
1376
1377         if (err != -EINVAL || cpu > 0 || thread > 0)
1378                 goto out_close;
1379
1380         /*
1381          * Must probe features in the order they were added to the
1382          * perf_event_attr interface.
1383          */
1384         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1385                 perf_missing_features.clockid_wrong = true;
1386                 goto fallback_missing_features;
1387         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1388                 perf_missing_features.clockid = true;
1389                 goto fallback_missing_features;
1390         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1391                 perf_missing_features.cloexec = true;
1392                 goto fallback_missing_features;
1393         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1394                 perf_missing_features.mmap2 = true;
1395                 goto fallback_missing_features;
1396         } else if (!perf_missing_features.exclude_guest &&
1397                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1398                 perf_missing_features.exclude_guest = true;
1399                 goto fallback_missing_features;
1400         } else if (!perf_missing_features.sample_id_all) {
1401                 perf_missing_features.sample_id_all = true;
1402                 goto retry_sample_id;
1403         }
1404
1405 out_close:
1406         do {
1407                 while (--thread >= 0) {
1408                         close(FD(evsel, cpu, thread));
1409                         FD(evsel, cpu, thread) = -1;
1410                 }
1411                 thread = nthreads;
1412         } while (--cpu >= 0);
1413         return err;
1414 }
1415
1416 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1417 {
1418         if (evsel->fd == NULL)
1419                 return;
1420
1421         perf_evsel__close_fd(evsel, ncpus, nthreads);
1422         perf_evsel__free_fd(evsel);
1423 }
1424
1425 static struct {
1426         struct cpu_map map;
1427         int cpus[1];
1428 } empty_cpu_map = {
1429         .map.nr = 1,
1430         .cpus   = { -1, },
1431 };
1432
1433 static struct {
1434         struct thread_map map;
1435         int threads[1];
1436 } empty_thread_map = {
1437         .map.nr  = 1,
1438         .threads = { -1, },
1439 };
1440
1441 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1442                      struct thread_map *threads)
1443 {
1444         if (cpus == NULL) {
1445                 /* Work around old compiler warnings about strict aliasing */
1446                 cpus = &empty_cpu_map.map;
1447         }
1448
1449         if (threads == NULL)
1450                 threads = &empty_thread_map.map;
1451
1452         return __perf_evsel__open(evsel, cpus, threads);
1453 }
1454
1455 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1456                              struct cpu_map *cpus)
1457 {
1458         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1459 }
1460
1461 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1462                                 struct thread_map *threads)
1463 {
1464         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1465 }
1466
1467 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1468                                        const union perf_event *event,
1469                                        struct perf_sample *sample)
1470 {
1471         u64 type = evsel->attr.sample_type;
1472         const u64 *array = event->sample.array;
1473         bool swapped = evsel->needs_swap;
1474         union u64_swap u;
1475
1476         array += ((event->header.size -
1477                    sizeof(event->header)) / sizeof(u64)) - 1;
1478
1479         if (type & PERF_SAMPLE_IDENTIFIER) {
1480                 sample->id = *array;
1481                 array--;
1482         }
1483
1484         if (type & PERF_SAMPLE_CPU) {
1485                 u.val64 = *array;
1486                 if (swapped) {
1487                         /* undo swap of u64, then swap on individual u32s */
1488                         u.val64 = bswap_64(u.val64);
1489                         u.val32[0] = bswap_32(u.val32[0]);
1490                 }
1491
1492                 sample->cpu = u.val32[0];
1493                 array--;
1494         }
1495
1496         if (type & PERF_SAMPLE_STREAM_ID) {
1497                 sample->stream_id = *array;
1498                 array--;
1499         }
1500
1501         if (type & PERF_SAMPLE_ID) {
1502                 sample->id = *array;
1503                 array--;
1504         }
1505
1506         if (type & PERF_SAMPLE_TIME) {
1507                 sample->time = *array;
1508                 array--;
1509         }
1510
1511         if (type & PERF_SAMPLE_TID) {
1512                 u.val64 = *array;
1513                 if (swapped) {
1514                         /* undo swap of u64, then swap on individual u32s */
1515                         u.val64 = bswap_64(u.val64);
1516                         u.val32[0] = bswap_32(u.val32[0]);
1517                         u.val32[1] = bswap_32(u.val32[1]);
1518                 }
1519
1520                 sample->pid = u.val32[0];
1521                 sample->tid = u.val32[1];
1522                 array--;
1523         }
1524
1525         return 0;
1526 }
1527
1528 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1529                             u64 size)
1530 {
1531         return size > max_size || offset + size > endp;
1532 }
1533
1534 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1535         do {                                                            \
1536                 if (overflow(endp, (max_size), (offset), (size)))       \
1537                         return -EFAULT;                                 \
1538         } while (0)
1539
1540 #define OVERFLOW_CHECK_u64(offset) \
1541         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1542
1543 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1544                              struct perf_sample *data)
1545 {
1546         u64 type = evsel->attr.sample_type;
1547         bool swapped = evsel->needs_swap;
1548         const u64 *array;
1549         u16 max_size = event->header.size;
1550         const void *endp = (void *)event + max_size;
1551         u64 sz;
1552
1553         /*
1554          * used for cross-endian analysis. See git commit 65014ab3
1555          * for why this goofiness is needed.
1556          */
1557         union u64_swap u;
1558
1559         memset(data, 0, sizeof(*data));
1560         data->cpu = data->pid = data->tid = -1;
1561         data->stream_id = data->id = data->time = -1ULL;
1562         data->period = evsel->attr.sample_period;
1563         data->weight = 0;
1564
1565         if (event->header.type != PERF_RECORD_SAMPLE) {
1566                 if (!evsel->attr.sample_id_all)
1567                         return 0;
1568                 return perf_evsel__parse_id_sample(evsel, event, data);
1569         }
1570
1571         array = event->sample.array;
1572
1573         /*
1574          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1575          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1576          * check the format does not go past the end of the event.
1577          */
1578         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1579                 return -EFAULT;
1580
1581         data->id = -1ULL;
1582         if (type & PERF_SAMPLE_IDENTIFIER) {
1583                 data->id = *array;
1584                 array++;
1585         }
1586
1587         if (type & PERF_SAMPLE_IP) {
1588                 data->ip = *array;
1589                 array++;
1590         }
1591
1592         if (type & PERF_SAMPLE_TID) {
1593                 u.val64 = *array;
1594                 if (swapped) {
1595                         /* undo swap of u64, then swap on individual u32s */
1596                         u.val64 = bswap_64(u.val64);
1597                         u.val32[0] = bswap_32(u.val32[0]);
1598                         u.val32[1] = bswap_32(u.val32[1]);
1599                 }
1600
1601                 data->pid = u.val32[0];
1602                 data->tid = u.val32[1];
1603                 array++;
1604         }
1605
1606         if (type & PERF_SAMPLE_TIME) {
1607                 data->time = *array;
1608                 array++;
1609         }
1610
1611         data->addr = 0;
1612         if (type & PERF_SAMPLE_ADDR) {
1613                 data->addr = *array;
1614                 array++;
1615         }
1616
1617         if (type & PERF_SAMPLE_ID) {
1618                 data->id = *array;
1619                 array++;
1620         }
1621
1622         if (type & PERF_SAMPLE_STREAM_ID) {
1623                 data->stream_id = *array;
1624                 array++;
1625         }
1626
1627         if (type & PERF_SAMPLE_CPU) {
1628
1629                 u.val64 = *array;
1630                 if (swapped) {
1631                         /* undo swap of u64, then swap on individual u32s */
1632                         u.val64 = bswap_64(u.val64);
1633                         u.val32[0] = bswap_32(u.val32[0]);
1634                 }
1635
1636                 data->cpu = u.val32[0];
1637                 array++;
1638         }
1639
1640         if (type & PERF_SAMPLE_PERIOD) {
1641                 data->period = *array;
1642                 array++;
1643         }
1644
1645         if (type & PERF_SAMPLE_READ) {
1646                 u64 read_format = evsel->attr.read_format;
1647
1648                 OVERFLOW_CHECK_u64(array);
1649                 if (read_format & PERF_FORMAT_GROUP)
1650                         data->read.group.nr = *array;
1651                 else
1652                         data->read.one.value = *array;
1653
1654                 array++;
1655
1656                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1657                         OVERFLOW_CHECK_u64(array);
1658                         data->read.time_enabled = *array;
1659                         array++;
1660                 }
1661
1662                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1663                         OVERFLOW_CHECK_u64(array);
1664                         data->read.time_running = *array;
1665                         array++;
1666                 }
1667
1668                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1669                 if (read_format & PERF_FORMAT_GROUP) {
1670                         const u64 max_group_nr = UINT64_MAX /
1671                                         sizeof(struct sample_read_value);
1672
1673                         if (data->read.group.nr > max_group_nr)
1674                                 return -EFAULT;
1675                         sz = data->read.group.nr *
1676                              sizeof(struct sample_read_value);
1677                         OVERFLOW_CHECK(array, sz, max_size);
1678                         data->read.group.values =
1679                                         (struct sample_read_value *)array;
1680                         array = (void *)array + sz;
1681                 } else {
1682                         OVERFLOW_CHECK_u64(array);
1683                         data->read.one.id = *array;
1684                         array++;
1685                 }
1686         }
1687
1688         if (type & PERF_SAMPLE_CALLCHAIN) {
1689                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1690
1691                 OVERFLOW_CHECK_u64(array);
1692                 data->callchain = (struct ip_callchain *)array++;
1693                 if (data->callchain->nr > max_callchain_nr)
1694                         return -EFAULT;
1695                 sz = data->callchain->nr * sizeof(u64);
1696                 OVERFLOW_CHECK(array, sz, max_size);
1697                 array = (void *)array + sz;
1698         }
1699
1700         if (type & PERF_SAMPLE_RAW) {
1701                 OVERFLOW_CHECK_u64(array);
1702                 u.val64 = *array;
1703                 if (WARN_ONCE(swapped,
1704                               "Endianness of raw data not corrected!\n")) {
1705                         /* undo swap of u64, then swap on individual u32s */
1706                         u.val64 = bswap_64(u.val64);
1707                         u.val32[0] = bswap_32(u.val32[0]);
1708                         u.val32[1] = bswap_32(u.val32[1]);
1709                 }
1710                 data->raw_size = u.val32[0];
1711                 array = (void *)array + sizeof(u32);
1712
1713                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1714                 data->raw_data = (void *)array;
1715                 array = (void *)array + data->raw_size;
1716         }
1717
1718         if (type & PERF_SAMPLE_BRANCH_STACK) {
1719                 const u64 max_branch_nr = UINT64_MAX /
1720                                           sizeof(struct branch_entry);
1721
1722                 OVERFLOW_CHECK_u64(array);
1723                 data->branch_stack = (struct branch_stack *)array++;
1724
1725                 if (data->branch_stack->nr > max_branch_nr)
1726                         return -EFAULT;
1727                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1728                 OVERFLOW_CHECK(array, sz, max_size);
1729                 array = (void *)array + sz;
1730         }
1731
1732         if (type & PERF_SAMPLE_REGS_USER) {
1733                 OVERFLOW_CHECK_u64(array);
1734                 data->user_regs.abi = *array;
1735                 array++;
1736
1737                 if (data->user_regs.abi) {
1738                         u64 mask = evsel->attr.sample_regs_user;
1739
1740                         sz = hweight_long(mask) * sizeof(u64);
1741                         OVERFLOW_CHECK(array, sz, max_size);
1742                         data->user_regs.mask = mask;
1743                         data->user_regs.regs = (u64 *)array;
1744                         array = (void *)array + sz;
1745                 }
1746         }
1747
1748         if (type & PERF_SAMPLE_STACK_USER) {
1749                 OVERFLOW_CHECK_u64(array);
1750                 sz = *array++;
1751
1752                 data->user_stack.offset = ((char *)(array - 1)
1753                                           - (char *) event);
1754
1755                 if (!sz) {
1756                         data->user_stack.size = 0;
1757                 } else {
1758                         OVERFLOW_CHECK(array, sz, max_size);
1759                         data->user_stack.data = (char *)array;
1760                         array = (void *)array + sz;
1761                         OVERFLOW_CHECK_u64(array);
1762                         data->user_stack.size = *array++;
1763                         if (WARN_ONCE(data->user_stack.size > sz,
1764                                       "user stack dump failure\n"))
1765                                 return -EFAULT;
1766                 }
1767         }
1768
1769         data->weight = 0;
1770         if (type & PERF_SAMPLE_WEIGHT) {
1771                 OVERFLOW_CHECK_u64(array);
1772                 data->weight = *array;
1773                 array++;
1774         }
1775
1776         data->data_src = PERF_MEM_DATA_SRC_NONE;
1777         if (type & PERF_SAMPLE_DATA_SRC) {
1778                 OVERFLOW_CHECK_u64(array);
1779                 data->data_src = *array;
1780                 array++;
1781         }
1782
1783         data->transaction = 0;
1784         if (type & PERF_SAMPLE_TRANSACTION) {
1785                 OVERFLOW_CHECK_u64(array);
1786                 data->transaction = *array;
1787                 array++;
1788         }
1789
1790         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1791         if (type & PERF_SAMPLE_REGS_INTR) {
1792                 OVERFLOW_CHECK_u64(array);
1793                 data->intr_regs.abi = *array;
1794                 array++;
1795
1796                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1797                         u64 mask = evsel->attr.sample_regs_intr;
1798
1799                         sz = hweight_long(mask) * sizeof(u64);
1800                         OVERFLOW_CHECK(array, sz, max_size);
1801                         data->intr_regs.mask = mask;
1802                         data->intr_regs.regs = (u64 *)array;
1803                         array = (void *)array + sz;
1804                 }
1805         }
1806
1807         return 0;
1808 }
1809
1810 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1811                                      u64 read_format)
1812 {
1813         size_t sz, result = sizeof(struct sample_event);
1814
1815         if (type & PERF_SAMPLE_IDENTIFIER)
1816                 result += sizeof(u64);
1817
1818         if (type & PERF_SAMPLE_IP)
1819                 result += sizeof(u64);
1820
1821         if (type & PERF_SAMPLE_TID)
1822                 result += sizeof(u64);
1823
1824         if (type & PERF_SAMPLE_TIME)
1825                 result += sizeof(u64);
1826
1827         if (type & PERF_SAMPLE_ADDR)
1828                 result += sizeof(u64);
1829
1830         if (type & PERF_SAMPLE_ID)
1831                 result += sizeof(u64);
1832
1833         if (type & PERF_SAMPLE_STREAM_ID)
1834                 result += sizeof(u64);
1835
1836         if (type & PERF_SAMPLE_CPU)
1837                 result += sizeof(u64);
1838
1839         if (type & PERF_SAMPLE_PERIOD)
1840                 result += sizeof(u64);
1841
1842         if (type & PERF_SAMPLE_READ) {
1843                 result += sizeof(u64);
1844                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1845                         result += sizeof(u64);
1846                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1847                         result += sizeof(u64);
1848                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1849                 if (read_format & PERF_FORMAT_GROUP) {
1850                         sz = sample->read.group.nr *
1851                              sizeof(struct sample_read_value);
1852                         result += sz;
1853                 } else {
1854                         result += sizeof(u64);
1855                 }
1856         }
1857
1858         if (type & PERF_SAMPLE_CALLCHAIN) {
1859                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1860                 result += sz;
1861         }
1862
1863         if (type & PERF_SAMPLE_RAW) {
1864                 result += sizeof(u32);
1865                 result += sample->raw_size;
1866         }
1867
1868         if (type & PERF_SAMPLE_BRANCH_STACK) {
1869                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1870                 sz += sizeof(u64);
1871                 result += sz;
1872         }
1873
1874         if (type & PERF_SAMPLE_REGS_USER) {
1875                 if (sample->user_regs.abi) {
1876                         result += sizeof(u64);
1877                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1878                         result += sz;
1879                 } else {
1880                         result += sizeof(u64);
1881                 }
1882         }
1883
1884         if (type & PERF_SAMPLE_STACK_USER) {
1885                 sz = sample->user_stack.size;
1886                 result += sizeof(u64);
1887                 if (sz) {
1888                         result += sz;
1889                         result += sizeof(u64);
1890                 }
1891         }
1892
1893         if (type & PERF_SAMPLE_WEIGHT)
1894                 result += sizeof(u64);
1895
1896         if (type & PERF_SAMPLE_DATA_SRC)
1897                 result += sizeof(u64);
1898
1899         if (type & PERF_SAMPLE_TRANSACTION)
1900                 result += sizeof(u64);
1901
1902         if (type & PERF_SAMPLE_REGS_INTR) {
1903                 if (sample->intr_regs.abi) {
1904                         result += sizeof(u64);
1905                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1906                         result += sz;
1907                 } else {
1908                         result += sizeof(u64);
1909                 }
1910         }
1911
1912         return result;
1913 }
1914
1915 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1916                                   u64 read_format,
1917                                   const struct perf_sample *sample,
1918                                   bool swapped)
1919 {
1920         u64 *array;
1921         size_t sz;
1922         /*
1923          * used for cross-endian analysis. See git commit 65014ab3
1924          * for why this goofiness is needed.
1925          */
1926         union u64_swap u;
1927
1928         array = event->sample.array;
1929
1930         if (type & PERF_SAMPLE_IDENTIFIER) {
1931                 *array = sample->id;
1932                 array++;
1933         }
1934
1935         if (type & PERF_SAMPLE_IP) {
1936                 *array = sample->ip;
1937                 array++;
1938         }
1939
1940         if (type & PERF_SAMPLE_TID) {
1941                 u.val32[0] = sample->pid;
1942                 u.val32[1] = sample->tid;
1943                 if (swapped) {
1944                         /*
1945                          * Inverse of what is done in perf_evsel__parse_sample
1946                          */
1947                         u.val32[0] = bswap_32(u.val32[0]);
1948                         u.val32[1] = bswap_32(u.val32[1]);
1949                         u.val64 = bswap_64(u.val64);
1950                 }
1951
1952                 *array = u.val64;
1953                 array++;
1954         }
1955
1956         if (type & PERF_SAMPLE_TIME) {
1957                 *array = sample->time;
1958                 array++;
1959         }
1960
1961         if (type & PERF_SAMPLE_ADDR) {
1962                 *array = sample->addr;
1963                 array++;
1964         }
1965
1966         if (type & PERF_SAMPLE_ID) {
1967                 *array = sample->id;
1968                 array++;
1969         }
1970
1971         if (type & PERF_SAMPLE_STREAM_ID) {
1972                 *array = sample->stream_id;
1973                 array++;
1974         }
1975
1976         if (type & PERF_SAMPLE_CPU) {
1977                 u.val32[0] = sample->cpu;
1978                 if (swapped) {
1979                         /*
1980                          * Inverse of what is done in perf_evsel__parse_sample
1981                          */
1982                         u.val32[0] = bswap_32(u.val32[0]);
1983                         u.val64 = bswap_64(u.val64);
1984                 }
1985                 *array = u.val64;
1986                 array++;
1987         }
1988
1989         if (type & PERF_SAMPLE_PERIOD) {
1990                 *array = sample->period;
1991                 array++;
1992         }
1993
1994         if (type & PERF_SAMPLE_READ) {
1995                 if (read_format & PERF_FORMAT_GROUP)
1996                         *array = sample->read.group.nr;
1997                 else
1998                         *array = sample->read.one.value;
1999                 array++;
2000
2001                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2002                         *array = sample->read.time_enabled;
2003                         array++;
2004                 }
2005
2006                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2007                         *array = sample->read.time_running;
2008                         array++;
2009                 }
2010
2011                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2012                 if (read_format & PERF_FORMAT_GROUP) {
2013                         sz = sample->read.group.nr *
2014                              sizeof(struct sample_read_value);
2015                         memcpy(array, sample->read.group.values, sz);
2016                         array = (void *)array + sz;
2017                 } else {
2018                         *array = sample->read.one.id;
2019                         array++;
2020                 }
2021         }
2022
2023         if (type & PERF_SAMPLE_CALLCHAIN) {
2024                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2025                 memcpy(array, sample->callchain, sz);
2026                 array = (void *)array + sz;
2027         }
2028
2029         if (type & PERF_SAMPLE_RAW) {
2030                 u.val32[0] = sample->raw_size;
2031                 if (WARN_ONCE(swapped,
2032                               "Endianness of raw data not corrected!\n")) {
2033                         /*
2034                          * Inverse of what is done in perf_evsel__parse_sample
2035                          */
2036                         u.val32[0] = bswap_32(u.val32[0]);
2037                         u.val32[1] = bswap_32(u.val32[1]);
2038                         u.val64 = bswap_64(u.val64);
2039                 }
2040                 *array = u.val64;
2041                 array = (void *)array + sizeof(u32);
2042
2043                 memcpy(array, sample->raw_data, sample->raw_size);
2044                 array = (void *)array + sample->raw_size;
2045         }
2046
2047         if (type & PERF_SAMPLE_BRANCH_STACK) {
2048                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2049                 sz += sizeof(u64);
2050                 memcpy(array, sample->branch_stack, sz);
2051                 array = (void *)array + sz;
2052         }
2053
2054         if (type & PERF_SAMPLE_REGS_USER) {
2055                 if (sample->user_regs.abi) {
2056                         *array++ = sample->user_regs.abi;
2057                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2058                         memcpy(array, sample->user_regs.regs, sz);
2059                         array = (void *)array + sz;
2060                 } else {
2061                         *array++ = 0;
2062                 }
2063         }
2064
2065         if (type & PERF_SAMPLE_STACK_USER) {
2066                 sz = sample->user_stack.size;
2067                 *array++ = sz;
2068                 if (sz) {
2069                         memcpy(array, sample->user_stack.data, sz);
2070                         array = (void *)array + sz;
2071                         *array++ = sz;
2072                 }
2073         }
2074
2075         if (type & PERF_SAMPLE_WEIGHT) {
2076                 *array = sample->weight;
2077                 array++;
2078         }
2079
2080         if (type & PERF_SAMPLE_DATA_SRC) {
2081                 *array = sample->data_src;
2082                 array++;
2083         }
2084
2085         if (type & PERF_SAMPLE_TRANSACTION) {
2086                 *array = sample->transaction;
2087                 array++;
2088         }
2089
2090         if (type & PERF_SAMPLE_REGS_INTR) {
2091                 if (sample->intr_regs.abi) {
2092                         *array++ = sample->intr_regs.abi;
2093                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2094                         memcpy(array, sample->intr_regs.regs, sz);
2095                         array = (void *)array + sz;
2096                 } else {
2097                         *array++ = 0;
2098                 }
2099         }
2100
2101         return 0;
2102 }
2103
2104 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2105 {
2106         return pevent_find_field(evsel->tp_format, name);
2107 }
2108
2109 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2110                          const char *name)
2111 {
2112         struct format_field *field = perf_evsel__field(evsel, name);
2113         int offset;
2114
2115         if (!field)
2116                 return NULL;
2117
2118         offset = field->offset;
2119
2120         if (field->flags & FIELD_IS_DYNAMIC) {
2121                 offset = *(int *)(sample->raw_data + field->offset);
2122                 offset &= 0xffff;
2123         }
2124
2125         return sample->raw_data + offset;
2126 }
2127
2128 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2129                        const char *name)
2130 {
2131         struct format_field *field = perf_evsel__field(evsel, name);
2132         void *ptr;
2133         u64 value;
2134
2135         if (!field)
2136                 return 0;
2137
2138         ptr = sample->raw_data + field->offset;
2139
2140         switch (field->size) {
2141         case 1:
2142                 return *(u8 *)ptr;
2143         case 2:
2144                 value = *(u16 *)ptr;
2145                 break;
2146         case 4:
2147                 value = *(u32 *)ptr;
2148                 break;
2149         case 8:
2150                 memcpy(&value, ptr, sizeof(u64));
2151                 break;
2152         default:
2153                 return 0;
2154         }
2155
2156         if (!evsel->needs_swap)
2157                 return value;
2158
2159         switch (field->size) {
2160         case 2:
2161                 return bswap_16(value);
2162         case 4:
2163                 return bswap_32(value);
2164         case 8:
2165                 return bswap_64(value);
2166         default:
2167                 return 0;
2168         }
2169
2170         return 0;
2171 }
2172
2173 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2174 {
2175         va_list args;
2176         int ret = 0;
2177
2178         if (!*first) {
2179                 ret += fprintf(fp, ",");
2180         } else {
2181                 ret += fprintf(fp, ":");
2182                 *first = false;
2183         }
2184
2185         va_start(args, fmt);
2186         ret += vfprintf(fp, fmt, args);
2187         va_end(args);
2188         return ret;
2189 }
2190
2191 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2192 {
2193         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2194 }
2195
2196 int perf_evsel__fprintf(struct perf_evsel *evsel,
2197                         struct perf_attr_details *details, FILE *fp)
2198 {
2199         bool first = true;
2200         int printed = 0;
2201
2202         if (details->event_group) {
2203                 struct perf_evsel *pos;
2204
2205                 if (!perf_evsel__is_group_leader(evsel))
2206                         return 0;
2207
2208                 if (evsel->nr_members > 1)
2209                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2210
2211                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2212                 for_each_group_member(pos, evsel)
2213                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2214
2215                 if (evsel->nr_members > 1)
2216                         printed += fprintf(fp, "}");
2217                 goto out;
2218         }
2219
2220         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2221
2222         if (details->verbose) {
2223                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2224                                                     __print_attr__fprintf, &first);
2225         } else if (details->freq) {
2226                 const char *term = "sample_freq";
2227
2228                 if (!evsel->attr.freq)
2229                         term = "sample_period";
2230
2231                 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2232                                          term, (u64)evsel->attr.sample_freq);
2233         }
2234 out:
2235         fputc('\n', fp);
2236         return ++printed;
2237 }
2238
2239 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2240                           char *msg, size_t msgsize)
2241 {
2242         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2243             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2244             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2245                 /*
2246                  * If it's cycles then fall back to hrtimer based
2247                  * cpu-clock-tick sw counter, which is always available even if
2248                  * no PMU support.
2249                  *
2250                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2251                  * b0a873e).
2252                  */
2253                 scnprintf(msg, msgsize, "%s",
2254 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2255
2256                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2257                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2258
2259                 zfree(&evsel->name);
2260                 return true;
2261         }
2262
2263         return false;
2264 }
2265
2266 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2267                               int err, char *msg, size_t size)
2268 {
2269         char sbuf[STRERR_BUFSIZE];
2270
2271         switch (err) {
2272         case EPERM:
2273         case EACCES:
2274                 return scnprintf(msg, size,
2275                  "You may not have permission to collect %sstats.\n"
2276                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2277                  " -1 - Not paranoid at all\n"
2278                  "  0 - Disallow raw tracepoint access for unpriv\n"
2279                  "  1 - Disallow cpu events for unpriv\n"
2280                  "  2 - Disallow kernel profiling for unpriv",
2281                                  target->system_wide ? "system-wide " : "");
2282         case ENOENT:
2283                 return scnprintf(msg, size, "The %s event is not supported.",
2284                                  perf_evsel__name(evsel));
2285         case EMFILE:
2286                 return scnprintf(msg, size, "%s",
2287                          "Too many events are opened.\n"
2288                          "Probably the maximum number of open file descriptors has been reached.\n"
2289                          "Hint: Try again after reducing the number of events.\n"
2290                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2291         case ENODEV:
2292                 if (target->cpu_list)
2293                         return scnprintf(msg, size, "%s",
2294          "No such device - did you specify an out-of-range profile CPU?\n");
2295                 break;
2296         case EOPNOTSUPP:
2297                 if (evsel->attr.precise_ip)
2298                         return scnprintf(msg, size, "%s",
2299         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2300 #if defined(__i386__) || defined(__x86_64__)
2301                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2302                         return scnprintf(msg, size, "%s",
2303         "No hardware sampling interrupt available.\n"
2304         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2305 #endif
2306                 break;
2307         case EBUSY:
2308                 if (find_process("oprofiled"))
2309                         return scnprintf(msg, size,
2310         "The PMU counters are busy/taken by another profiler.\n"
2311         "We found oprofile daemon running, please stop it and try again.");
2312                 break;
2313         case EINVAL:
2314                 if (perf_missing_features.clockid)
2315                         return scnprintf(msg, size, "clockid feature not supported.");
2316                 if (perf_missing_features.clockid_wrong)
2317                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2318                 break;
2319         default:
2320                 break;
2321         }
2322
2323         return scnprintf(msg, size,
2324         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2325         "/bin/dmesg may provide additional information.\n"
2326         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2327                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2328                          perf_evsel__name(evsel));
2329 }