tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <linux/bitops.h>
  12 #include <api/fs/debugfs.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <sys/resource.h>
  17 #include "asm/bug.h"
  18 #include "callchain.h"
  19 #include "cgroup.h"
  20 #include "evsel.h"
  21 #include "evlist.h"
  22 #include "util.h"
  23 #include "cpumap.h"
  24 #include "thread_map.h"
  25 #include "target.h"
  26 #include "perf_regs.h"
  27 #include "debug.h"
  28 #include "trace-event.h"
  29 #include "stat.h"
  30
  31 static struct {
  32         bool sample_id_all;
  33         bool exclude_guest;
  34         bool mmap2;
  35         bool cloexec;
  36         bool clockid;
  37         bool clockid_wrong;
  38 } perf_missing_features;
  39
  40 static clockid_t clockid;
  41
  42 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  43 {
  44         return 0;
  45 }
  46
  47 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  48 {
  49 }
  50
  51 static struct {
  52         size_t  size;
  53         int     (*init)(struct perf_evsel *evsel);
  54         void    (*fini)(struct perf_evsel *evsel);
  55 } perf_evsel__object = {
  56         .size = sizeof(struct perf_evsel),
  57         .init = perf_evsel__no_extra_init,
  58         .fini = perf_evsel__no_extra_fini,
  59 };
  60
  61 int perf_evsel__object_config(size_t object_size,
  62                               int (*init)(struct perf_evsel *evsel),
  63                               void (*fini)(struct perf_evsel *evsel))
  64 {
  65
  66         if (object_size == 0)
  67                 goto set_methods;
  68
  69         if (perf_evsel__object.size > object_size)
  70                 return -EINVAL;
  71
  72         perf_evsel__object.size = object_size;
  73
  74 set_methods:
  75         if (init != NULL)
  76                 perf_evsel__object.init = init;
  77
  78         if (fini != NULL)
  79                 perf_evsel__object.fini = fini;
  80
  81         return 0;
  82 }
  83
  84 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  85
  86 int __perf_evsel__sample_size(u64 sample_type)
  87 {
  88         u64 mask = sample_type & PERF_SAMPLE_MASK;
  89         int size = 0;
  90         int i;
  91
  92         for (i = 0; i < 64; i++) {
  93                 if (mask & (1ULL << i))
  94                         size++;
  95         }
  96
  97         size *= sizeof(u64);
  98
  99         return size;
 100 }
 101
 102 /**
 103  * __perf_evsel__calc_id_pos - calculate id_pos.
 104  * @sample_type: sample type
 105  *
 106  * This function returns the position of the event id (PERF_SAMPLE_ID or
 107  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 108  * sample_event.
 109  */
 110 static int __perf_evsel__calc_id_pos(u64 sample_type)
 111 {
 112         int idx = 0;
 113
 114         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 115                 return 0;
 116
 117         if (!(sample_type & PERF_SAMPLE_ID))
 118                 return -1;
 119
 120         if (sample_type & PERF_SAMPLE_IP)
 121                 idx += 1;
 122
 123         if (sample_type & PERF_SAMPLE_TID)
 124                 idx += 1;
 125
 126         if (sample_type & PERF_SAMPLE_TIME)
 127                 idx += 1;
 128
 129         if (sample_type & PERF_SAMPLE_ADDR)
 130                 idx += 1;
 131
 132         return idx;
 133 }
 134
 135 /**
 136  * __perf_evsel__calc_is_pos - calculate is_pos.
 137  * @sample_type: sample type
 138  *
 139  * This function returns the position (counting backwards) of the event id
 140  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 141  * sample_id_all is used there is an id sample appended to non-sample events.
 142  */
 143 static int __perf_evsel__calc_is_pos(u64 sample_type)
 144 {
 145         int idx = 1;
 146
 147         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 148                 return 1;
 149
 150         if (!(sample_type & PERF_SAMPLE_ID))
 151                 return -1;
 152
 153         if (sample_type & PERF_SAMPLE_CPU)
 154                 idx += 1;
 155
 156         if (sample_type & PERF_SAMPLE_STREAM_ID)
 157                 idx += 1;
 158
 159         return idx;
 160 }
 161
 162 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 163 {
 164         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 165         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 166 }
 167
 168 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 169                                   enum perf_event_sample_format bit)
 170 {
 171         if (!(evsel->attr.sample_type & bit)) {
 172                 evsel->attr.sample_type |= bit;
 173                 evsel->sample_size += sizeof(u64);
 174                 perf_evsel__calc_id_pos(evsel);
 175         }
 176 }
 177
 178 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 179                                     enum perf_event_sample_format bit)
 180 {
 181         if (evsel->attr.sample_type & bit) {
 182                 evsel->attr.sample_type &= ~bit;
 183                 evsel->sample_size -= sizeof(u64);
 184                 perf_evsel__calc_id_pos(evsel);
 185         }
 186 }
 187
 188 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 189                                bool can_sample_identifier)
 190 {
 191         if (can_sample_identifier) {
 192                 perf_evsel__reset_sample_bit(evsel, ID);
 193                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 194         } else {
 195                 perf_evsel__set_sample_bit(evsel, ID);
 196         }
 197         evsel->attr.read_format |= PERF_FORMAT_ID;
 198 }
 199
 200 void perf_evsel__init(struct perf_evsel *evsel,
 201                       struct perf_event_attr *attr, int idx)
 202 {
 203         evsel->idx         = idx;
 204         evsel->tracking    = !idx;
 205         evsel->attr        = *attr;
 206         evsel->leader      = evsel;
 207         evsel->unit        = "";
 208         evsel->scale       = 1.0;
 209         INIT_LIST_HEAD(&evsel->node);
 210         INIT_LIST_HEAD(&evsel->config_terms);
 211         perf_evsel__object.init(evsel);
 212         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 213         perf_evsel__calc_id_pos(evsel);
 214         evsel->cmdline_group_boundary = false;
 215 }
 216
 217 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 218 {
 219         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 220
 221         if (evsel != NULL)
 222                 perf_evsel__init(evsel, attr, idx);
 223
 224         return evsel;
 225 }
 226
 227 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 228 {
 229         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 230
 231         if (evsel != NULL) {
 232                 struct perf_event_attr attr = {
 233                         .type          = PERF_TYPE_TRACEPOINT,
 234                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 235                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 236                 };
 237
 238                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 239                         goto out_free;
 240
 241                 evsel->tp_format = trace_event__tp_format(sys, name);
 242                 if (evsel->tp_format == NULL)
 243                         goto out_free;
 244
 245                 event_attr_init(&attr);
 246                 attr.config = evsel->tp_format->id;
 247                 attr.sample_period = 1;
 248                 perf_evsel__init(evsel, &attr, idx);
 249         }
 250
 251         return evsel;
 252
 253 out_free:
 254         zfree(&evsel->name);
 255         free(evsel);
 256         return NULL;
 257 }
 258
 259 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 260         "cycles",
 261         "instructions",
 262         "cache-references",
 263         "cache-misses",
 264         "branches",
 265         "branch-misses",
 266         "bus-cycles",
 267         "stalled-cycles-frontend",
 268         "stalled-cycles-backend",
 269         "ref-cycles",
 270 };
 271
 272 static const char *__perf_evsel__hw_name(u64 config)
 273 {
 274         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 275                 return perf_evsel__hw_names[config];
 276
 277         return "unknown-hardware";
 278 }
 279
 280 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 281 {
 282         int colon = 0, r = 0;
 283         struct perf_event_attr *attr = &evsel->attr;
 284         bool exclude_guest_default = false;
 285
 286 #define MOD_PRINT(context, mod) do {                                    \
 287                 if (!attr->exclude_##context) {                         \
 288                         if (!colon) colon = ++r;                        \
 289                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 290                 } } while(0)
 291
 292         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 293                 MOD_PRINT(kernel, 'k');
 294                 MOD_PRINT(user, 'u');
 295                 MOD_PRINT(hv, 'h');
 296                 exclude_guest_default = true;
 297         }
 298
 299         if (attr->precise_ip) {
 300                 if (!colon)
 301                         colon = ++r;
 302                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 303                 exclude_guest_default = true;
 304         }
 305
 306         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 307                 MOD_PRINT(host, 'H');
 308                 MOD_PRINT(guest, 'G');
 309         }
 310 #undef MOD_PRINT
 311         if (colon)
 312                 bf[colon - 1] = ':';
 313         return r;
 314 }
 315
 316 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 317 {
 318         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 319         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 320 }
 321
 322 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 323         "cpu-clock",
 324         "task-clock",
 325         "page-faults",
 326         "context-switches",
 327         "cpu-migrations",
 328         "minor-faults",
 329         "major-faults",
 330         "alignment-faults",
 331         "emulation-faults",
 332         "dummy",
 333 };
 334
 335 static const char *__perf_evsel__sw_name(u64 config)
 336 {
 337         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 338                 return perf_evsel__sw_names[config];
 339         return "unknown-software";
 340 }
 341
 342 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 343 {
 344         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 345         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 346 }
 347
 348 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 349 {
 350         int r;
 351
 352         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 353
 354         if (type & HW_BREAKPOINT_R)
 355                 r += scnprintf(bf + r, size - r, "r");
 356
 357         if (type & HW_BREAKPOINT_W)
 358                 r += scnprintf(bf + r, size - r, "w");
 359
 360         if (type & HW_BREAKPOINT_X)
 361                 r += scnprintf(bf + r, size - r, "x");
 362
 363         return r;
 364 }
 365
 366 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 367 {
 368         struct perf_event_attr *attr = &evsel->attr;
 369         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 370         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 371 }
 372
 373 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 374                                 [PERF_EVSEL__MAX_ALIASES] = {
 375  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 376  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 377  { "LLC",       "L2",                                                   },
 378  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 379  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 380  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 381  { "node",                                                              },
 382 };
 383
 384 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 385                                    [PERF_EVSEL__MAX_ALIASES] = {
 386  { "load",      "loads",        "read",                                 },
 387  { "store",     "stores",       "write",                                },
 388  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 389 };
 390
 391 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 392                                        [PERF_EVSEL__MAX_ALIASES] = {
 393  { "refs",      "Reference",    "ops",          "access",               },
 394  { "misses",    "miss",                                                 },
 395 };
 396
 397 #define C(x)            PERF_COUNT_HW_CACHE_##x
 398 #define CACHE_READ      (1 << C(OP_READ))
 399 #define CACHE_WRITE     (1 << C(OP_WRITE))
 400 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 401 #define COP(x)          (1 << x)
 402
 403 /*
 404  * cache operartion stat
 405  * L1I : Read and prefetch only
 406  * ITLB and BPU : Read-only
 407  */
 408 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 409  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 410  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 411  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 412  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 413  [C(ITLB)]      = (CACHE_READ),
 414  [C(BPU)]       = (CACHE_READ),
 415  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 416 };
 417
 418 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 419 {
 420         if (perf_evsel__hw_cache_stat[type] & COP(op))
 421                 return true;    /* valid */
 422         else
 423                 return false;   /* invalid */
 424 }
 425
 426 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 427                                             char *bf, size_t size)
 428 {
 429         if (result) {
 430                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 431                                  perf_evsel__hw_cache_op[op][0],
 432                                  perf_evsel__hw_cache_result[result][0]);
 433         }
 434
 435         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 436                          perf_evsel__hw_cache_op[op][1]);
 437 }
 438
 439 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 440 {
 441         u8 op, result, type = (config >>  0) & 0xff;
 442         const char *err = "unknown-ext-hardware-cache-type";
 443
 444         if (type > PERF_COUNT_HW_CACHE_MAX)
 445                 goto out_err;
 446
 447         op = (config >>  8) & 0xff;
 448         err = "unknown-ext-hardware-cache-op";
 449         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 450                 goto out_err;
 451
 452         result = (config >> 16) & 0xff;
 453         err = "unknown-ext-hardware-cache-result";
 454         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 455                 goto out_err;
 456
 457         err = "invalid-cache";
 458         if (!perf_evsel__is_cache_op_valid(type, op))
 459                 goto out_err;
 460
 461         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 462 out_err:
 463         return scnprintf(bf, size, "%s", err);
 464 }
 465
 466 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 467 {
 468         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 469         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 470 }
 471
 472 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 473 {
 474         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 475         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 476 }
 477
 478 const char *perf_evsel__name(struct perf_evsel *evsel)
 479 {
 480         char bf[128];
 481
 482         if (evsel->name)
 483                 return evsel->name;
 484
 485         switch (evsel->attr.type) {
 486         case PERF_TYPE_RAW:
 487                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 488                 break;
 489
 490         case PERF_TYPE_HARDWARE:
 491                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 492                 break;
 493
 494         case PERF_TYPE_HW_CACHE:
 495                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 496                 break;
 497
 498         case PERF_TYPE_SOFTWARE:
 499                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 500                 break;
 501
 502         case PERF_TYPE_TRACEPOINT:
 503                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 504                 break;
 505
 506         case PERF_TYPE_BREAKPOINT:
 507                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 508                 break;
 509
 510         default:
 511                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 512                           evsel->attr.type);
 513                 break;
 514         }
 515
 516         evsel->name = strdup(bf);
 517
 518         return evsel->name ?: "unknown";
 519 }
 520
 521 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 522 {
 523         return evsel->group_name ?: "anon group";
 524 }
 525
 526 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 527 {
 528         int ret;
 529         struct perf_evsel *pos;
 530         const char *group_name = perf_evsel__group_name(evsel);
 531
 532         ret = scnprintf(buf, size, "%s", group_name);
 533
 534         ret += scnprintf(buf + ret, size - ret, " { %s",
 535                          perf_evsel__name(evsel));
 536
 537         for_each_group_member(pos, evsel)
 538                 ret += scnprintf(buf + ret, size - ret, ", %s",
 539                                  perf_evsel__name(pos));
 540
 541         ret += scnprintf(buf + ret, size - ret, " }");
 542
 543         return ret;
 544 }
 545
 546 static void
 547 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 548                              struct record_opts *opts,
 549                              struct callchain_param *param)
 550 {
 551         bool function = perf_evsel__is_function_event(evsel);
 552         struct perf_event_attr *attr = &evsel->attr;
 553
 554         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 555
 556         if (param->record_mode == CALLCHAIN_LBR) {
 557                 if (!opts->branch_stack) {
 558                         if (attr->exclude_user) {
 559                                 pr_warning("LBR callstack option is only available "
 560                                            "to get user callchain information. "
 561                                            "Falling back to framepointers.\n");
 562                         } else {
 563                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 564                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 565                                                         PERF_SAMPLE_BRANCH_CALL_STACK;
 566                         }
 567                 } else
 568                          pr_warning("Cannot use LBR callstack with branch stack. "
 569                                     "Falling back to framepointers.\n");
 570         }
 571
 572         if (param->record_mode == CALLCHAIN_DWARF) {
 573                 if (!function) {
 574                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 575                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 576                         attr->sample_regs_user = PERF_REGS_MASK;
 577                         attr->sample_stack_user = param->dump_size;
 578                         attr->exclude_callchain_user = 1;
 579                 } else {
 580                         pr_info("Cannot use DWARF unwind for function trace event,"
 581                                 " falling back to framepointers.\n");
 582                 }
 583         }
 584
 585         if (function) {
 586                 pr_info("Disabling user space callchains for function trace event.\n");
 587                 attr->exclude_callchain_user = 1;
 588         }
 589 }
 590
 591 static void
 592 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 593                             struct callchain_param *param)
 594 {
 595         struct perf_event_attr *attr = &evsel->attr;
 596
 597         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 598         if (param->record_mode == CALLCHAIN_LBR) {
 599                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 600                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 601                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 602         }
 603         if (param->record_mode == CALLCHAIN_DWARF) {
 604                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 605                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 606         }
 607 }
 608
 609 static void apply_config_terms(struct perf_evsel *evsel,
 610                                struct record_opts *opts)
 611 {
 612         struct perf_evsel_config_term *term;
 613         struct list_head *config_terms = &evsel->config_terms;
 614         struct perf_event_attr *attr = &evsel->attr;
 615         struct callchain_param param;
 616         u32 dump_size = 0;
 617         char *callgraph_buf = NULL;
 618
 619         /* callgraph default */
 620         param.record_mode = callchain_param.record_mode;
 621
 622         list_for_each_entry(term, config_terms, list) {
 623                 switch (term->type) {
 624                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 625                         attr->sample_period = term->val.period;
 626                         attr->freq = 0;
 627                         break;
 628                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 629                         attr->sample_freq = term->val.freq;
 630                         attr->freq = 1;
 631                         break;
 632                 case PERF_EVSEL__CONFIG_TERM_TIME:
 633                         if (term->val.time)
 634                                 perf_evsel__set_sample_bit(evsel, TIME);
 635                         else
 636                                 perf_evsel__reset_sample_bit(evsel, TIME);
 637                         break;
 638                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 639                         callgraph_buf = term->val.callgraph;
 640                         break;
 641                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 642                         dump_size = term->val.stack_user;
 643                         break;
 644                 default:
 645                         break;
 646                 }
 647         }
 648
 649         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 650         if ((callgraph_buf != NULL) || (dump_size > 0)) {
 651
 652                 /* parse callgraph parameters */
 653                 if (callgraph_buf != NULL) {
 654                         param.enabled = true;
 655                         if (parse_callchain_record(callgraph_buf, &param)) {
 656                                 pr_err("per-event callgraph setting for %s failed. "
 657                                        "Apply callgraph global setting for it\n",
 658                                        evsel->name);
 659                                 return;
 660                         }
 661                 }
 662                 if (dump_size > 0) {
 663                         dump_size = round_up(dump_size, sizeof(u64));
 664                         param.dump_size = dump_size;
 665                 }
 666
 667                 /* If global callgraph set, clear it */
 668                 if (callchain_param.enabled)
 669                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 670
 671                 /* set perf-event callgraph */
 672                 if (param.enabled)
 673                         perf_evsel__config_callgraph(evsel, opts, &param);
 674         }
 675 }
 676
 677 /*
 678  * The enable_on_exec/disabled value strategy:
 679  *
 680  *  1) For any type of traced program:
 681  *    - all independent events and group leaders are disabled
 682  *    - all group members are enabled
 683  *
 684  *     Group members are ruled by group leaders. They need to
 685  *     be enabled, because the group scheduling relies on that.
 686  *
 687  *  2) For traced programs executed by perf:
 688  *     - all independent events and group leaders have
 689  *       enable_on_exec set
 690  *     - we don't specifically enable or disable any event during
 691  *       the record command
 692  *
 693  *     Independent events and group leaders are initially disabled
 694  *     and get enabled by exec. Group members are ruled by group
 695  *     leaders as stated in 1).
 696  *
 697  *  3) For traced programs attached by perf (pid/tid):
 698  *     - we specifically enable or disable all events during
 699  *       the record command
 700  *
 701  *     When attaching events to already running traced we
 702  *     enable/disable events specifically, as there's no
 703  *     initial traced exec call.
 704  */
 705 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 706 {
 707         struct perf_evsel *leader = evsel->leader;
 708         struct perf_event_attr *attr = &evsel->attr;
 709         int track = evsel->tracking;
 710         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 711
 712         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 713         attr->inherit       = !opts->no_inherit;
 714
 715         perf_evsel__set_sample_bit(evsel, IP);
 716         perf_evsel__set_sample_bit(evsel, TID);
 717
 718         if (evsel->sample_read) {
 719                 perf_evsel__set_sample_bit(evsel, READ);
 720
 721                 /*
 722                  * We need ID even in case of single event, because
 723                  * PERF_SAMPLE_READ process ID specific data.
 724                  */
 725                 perf_evsel__set_sample_id(evsel, false);
 726
 727                 /*
 728                  * Apply group format only if we belong to group
 729                  * with more than one members.
 730                  */
 731                 if (leader->nr_members > 1) {
 732                         attr->read_format |= PERF_FORMAT_GROUP;
 733                         attr->inherit = 0;
 734                 }
 735         }
 736
 737         /*
 738          * We default some events to have a default interval. But keep
 739          * it a weak assumption overridable by the user.
 740          */
 741         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 742                                      opts->user_interval != ULLONG_MAX)) {
 743                 if (opts->freq) {
 744                         perf_evsel__set_sample_bit(evsel, PERIOD);
 745                         attr->freq              = 1;
 746                         attr->sample_freq       = opts->freq;
 747                 } else {
 748                         attr->sample_period = opts->default_interval;
 749                 }
 750         }
 751
 752         /*
 753          * Disable sampling for all group members other
 754          * than leader in case leader 'leads' the sampling.
 755          */
 756         if ((leader != evsel) && leader->sample_read) {
 757                 attr->sample_freq   = 0;
 758                 attr->sample_period = 0;
 759         }
 760
 761         if (opts->no_samples)
 762                 attr->sample_freq = 0;
 763
 764         if (opts->inherit_stat)
 765                 attr->inherit_stat = 1;
 766
 767         if (opts->sample_address) {
 768                 perf_evsel__set_sample_bit(evsel, ADDR);
 769                 attr->mmap_data = track;
 770         }
 771
 772         /*
 773          * We don't allow user space callchains for  function trace
 774          * event, due to issues with page faults while tracing page
 775          * fault handler and its overall trickiness nature.
 776          */
 777         if (perf_evsel__is_function_event(evsel))
 778                 evsel->attr.exclude_callchain_user = 1;
 779
 780         if (callchain_param.enabled && !evsel->no_aux_samples)
 781                 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
 782
 783         if (opts->sample_intr_regs) {
 784                 attr->sample_regs_intr = PERF_REGS_MASK;
 785                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 786         }
 787
 788         if (target__has_cpu(&opts->target))
 789                 perf_evsel__set_sample_bit(evsel, CPU);
 790
 791         if (opts->period)
 792                 perf_evsel__set_sample_bit(evsel, PERIOD);
 793
 794         /*
 795          * When the user explicitely disabled time don't force it here.
 796          */
 797         if (opts->sample_time &&
 798             (!perf_missing_features.sample_id_all &&
 799             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 800              opts->sample_time_set)))
 801                 perf_evsel__set_sample_bit(evsel, TIME);
 802
 803         if (opts->raw_samples && !evsel->no_aux_samples) {
 804                 perf_evsel__set_sample_bit(evsel, TIME);
 805                 perf_evsel__set_sample_bit(evsel, RAW);
 806                 perf_evsel__set_sample_bit(evsel, CPU);
 807         }
 808
 809         if (opts->sample_address)
 810                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 811
 812         if (opts->no_buffering) {
 813                 attr->watermark = 0;
 814                 attr->wakeup_events = 1;
 815         }
 816         if (opts->branch_stack && !evsel->no_aux_samples) {
 817                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 818                 attr->branch_sample_type = opts->branch_stack;
 819         }
 820
 821         if (opts->sample_weight)
 822                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 823
 824         attr->task  = track;
 825         attr->mmap  = track;
 826         attr->mmap2 = track && !perf_missing_features.mmap2;
 827         attr->comm  = track;
 828
 829         if (opts->record_switch_events)
 830                 attr->context_switch = track;
 831
 832         if (opts->sample_transaction)
 833                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 834
 835         if (opts->running_time) {
 836                 evsel->attr.read_format |=
 837                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 838                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 839         }
 840
 841         /*
 842          * XXX see the function comment above
 843          *
 844          * Disabling only independent events or group leaders,
 845          * keeping group members enabled.
 846          */
 847         if (perf_evsel__is_group_leader(evsel))
 848                 attr->disabled = 1;
 849
 850         /*
 851          * Setting enable_on_exec for independent events and
 852          * group leaders for traced executed by perf.
 853          */
 854         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 855                 !opts->initial_delay)
 856                 attr->enable_on_exec = 1;
 857
 858         if (evsel->immediate) {
 859                 attr->disabled = 0;
 860                 attr->enable_on_exec = 0;
 861         }
 862
 863         clockid = opts->clockid;
 864         if (opts->use_clockid) {
 865                 attr->use_clockid = 1;
 866                 attr->clockid = opts->clockid;
 867         }
 868
 869         /*
 870          * Apply event specific term settings,
 871          * it overloads any global configuration.
 872          */
 873         apply_config_terms(evsel, opts);
 874 }
 875
 876 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 877 {
 878         int cpu, thread;
 879
 880         if (evsel->system_wide)
 881                 nthreads = 1;
 882
 883         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 884
 885         if (evsel->fd) {
 886                 for (cpu = 0; cpu < ncpus; cpu++) {
 887                         for (thread = 0; thread < nthreads; thread++) {
 888                                 FD(evsel, cpu, thread) = -1;
 889                         }
 890                 }
 891         }
 892
 893         return evsel->fd != NULL ? 0 : -ENOMEM;
 894 }
 895
 896 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 897                           int ioc,  void *arg)
 898 {
 899         int cpu, thread;
 900
 901         if (evsel->system_wide)
 902                 nthreads = 1;
 903
 904         for (cpu = 0; cpu < ncpus; cpu++) {
 905                 for (thread = 0; thread < nthreads; thread++) {
 906                         int fd = FD(evsel, cpu, thread),
 907                             err = ioctl(fd, ioc, arg);
 908
 909                         if (err)
 910                                 return err;
 911                 }
 912         }
 913
 914         return 0;
 915 }
 916
 917 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 918                              const char *filter)
 919 {
 920         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 921                                      PERF_EVENT_IOC_SET_FILTER,
 922                                      (void *)filter);
 923 }
 924
 925 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 926 {
 927         char *new_filter = strdup(filter);
 928
 929         if (new_filter != NULL) {
 930                 free(evsel->filter);
 931                 evsel->filter = new_filter;
 932                 return 0;
 933         }
 934
 935         return -1;
 936 }
 937
 938 int perf_evsel__append_filter(struct perf_evsel *evsel,
 939                               const char *op, const char *filter)
 940 {
 941         char *new_filter;
 942
 943         if (evsel->filter == NULL)
 944                 return perf_evsel__set_filter(evsel, filter);
 945
 946         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 947                 free(evsel->filter);
 948                 evsel->filter = new_filter;
 949                 return 0;
 950         }
 951
 952         return -1;
 953 }
 954
 955 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 956 {
 957         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 958                                      PERF_EVENT_IOC_ENABLE,
 959                                      0);
 960 }
 961
 962 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 963 {
 964         if (ncpus == 0 || nthreads == 0)
 965                 return 0;
 966
 967         if (evsel->system_wide)
 968                 nthreads = 1;
 969
 970         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 971         if (evsel->sample_id == NULL)
 972                 return -ENOMEM;
 973
 974         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 975         if (evsel->id == NULL) {
 976                 xyarray__delete(evsel->sample_id);
 977                 evsel->sample_id = NULL;
 978                 return -ENOMEM;
 979         }
 980
 981         return 0;
 982 }
 983
 984 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 985 {
 986         xyarray__delete(evsel->fd);
 987         evsel->fd = NULL;
 988 }
 989
 990 static void perf_evsel__free_id(struct perf_evsel *evsel)
 991 {
 992         xyarray__delete(evsel->sample_id);
 993         evsel->sample_id = NULL;
 994         zfree(&evsel->id);
 995 }
 996
 997 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
 998 {
 999         struct perf_evsel_config_term *term, *h;
1000
1001         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1002                 list_del(&term->list);
1003                 free(term);
1004         }
1005 }
1006
1007 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1008 {
1009         int cpu, thread;
1010
1011         if (evsel->system_wide)
1012                 nthreads = 1;
1013
1014         for (cpu = 0; cpu < ncpus; cpu++)
1015                 for (thread = 0; thread < nthreads; ++thread) {
1016                         close(FD(evsel, cpu, thread));
1017                         FD(evsel, cpu, thread) = -1;
1018                 }
1019 }
1020
1021 void perf_evsel__exit(struct perf_evsel *evsel)
1022 {
1023         assert(list_empty(&evsel->node));
1024         perf_evsel__free_fd(evsel);
1025         perf_evsel__free_id(evsel);
1026         perf_evsel__free_config_terms(evsel);
1027         close_cgroup(evsel->cgrp);
1028         cpu_map__put(evsel->cpus);
1029         thread_map__put(evsel->threads);
1030         zfree(&evsel->group_name);
1031         zfree(&evsel->name);
1032         perf_evsel__object.fini(evsel);
1033 }
1034
1035 void perf_evsel__delete(struct perf_evsel *evsel)
1036 {
1037         perf_evsel__exit(evsel);
1038         free(evsel);
1039 }
1040
1041 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1042                                 struct perf_counts_values *count)
1043 {
1044         struct perf_counts_values tmp;
1045
1046         if (!evsel->prev_raw_counts)
1047                 return;
1048
1049         if (cpu == -1) {
1050                 tmp = evsel->prev_raw_counts->aggr;
1051                 evsel->prev_raw_counts->aggr = *count;
1052         } else {
1053                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1054                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1055         }
1056
1057         count->val = count->val - tmp.val;
1058         count->ena = count->ena - tmp.ena;
1059         count->run = count->run - tmp.run;
1060 }
1061
1062 void perf_counts_values__scale(struct perf_counts_values *count,
1063                                bool scale, s8 *pscaled)
1064 {
1065         s8 scaled = 0;
1066
1067         if (scale) {
1068                 if (count->run == 0) {
1069                         scaled = -1;
1070                         count->val = 0;
1071                 } else if (count->run < count->ena) {
1072                         scaled = 1;
1073                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1074                 }
1075         } else
1076                 count->ena = count->run = 0;
1077
1078         if (pscaled)
1079                 *pscaled = scaled;
1080 }
1081
1082 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1083                      struct perf_counts_values *count)
1084 {
1085         memset(count, 0, sizeof(*count));
1086
1087         if (FD(evsel, cpu, thread) < 0)
1088                 return -EINVAL;
1089
1090         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1091                 return -errno;
1092
1093         return 0;
1094 }
1095
1096 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1097                               int cpu, int thread, bool scale)
1098 {
1099         struct perf_counts_values count;
1100         size_t nv = scale ? 3 : 1;
1101
1102         if (FD(evsel, cpu, thread) < 0)
1103                 return -EINVAL;
1104
1105         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1106                 return -ENOMEM;
1107
1108         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1109                 return -errno;
1110
1111         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1112         perf_counts_values__scale(&count, scale, NULL);
1113         *perf_counts(evsel->counts, cpu, thread) = count;
1114         return 0;
1115 }
1116
1117 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1118 {
1119         struct perf_evsel *leader = evsel->leader;
1120         int fd;
1121
1122         if (perf_evsel__is_group_leader(evsel))
1123                 return -1;
1124
1125         /*
1126          * Leader must be already processed/open,
1127          * if not it's a bug.
1128          */
1129         BUG_ON(!leader->fd);
1130
1131         fd = FD(leader, cpu, thread);
1132         BUG_ON(fd == -1);
1133
1134         return fd;
1135 }
1136
1137 struct bit_names {
1138         int bit;
1139         const char *name;
1140 };
1141
1142 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1143 {
1144         bool first_bit = true;
1145         int i = 0;
1146
1147         do {
1148                 if (value & bits[i].bit) {
1149                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1150                         first_bit = false;
1151                 }
1152         } while (bits[++i].name != NULL);
1153 }
1154
1155 static void __p_sample_type(char *buf, size_t size, u64 value)
1156 {
1157 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1158         struct bit_names bits[] = {
1159                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1160                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1161                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1162                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1163                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1164                 { .name = NULL, }
1165         };
1166 #undef bit_name
1167         __p_bits(buf, size, value, bits);
1168 }
1169
1170 static void __p_read_format(char *buf, size_t size, u64 value)
1171 {
1172 #define bit_name(n) { PERF_FORMAT_##n, #n }
1173         struct bit_names bits[] = {
1174                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1175                 bit_name(ID), bit_name(GROUP),
1176                 { .name = NULL, }
1177         };
1178 #undef bit_name
1179         __p_bits(buf, size, value, bits);
1180 }
1181
1182 #define BUF_SIZE                1024
1183
1184 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1185 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1186 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1187 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1188 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1189
1190 #define PRINT_ATTRn(_n, _f, _p)                         \
1191 do {                                                    \
1192         if (attr->_f) {                                 \
1193                 _p(attr->_f);                           \
1194                 ret += attr__fprintf(fp, _n, buf, priv);\
1195         }                                               \
1196 } while (0)
1197
1198 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1199
1200 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1201                              attr__fprintf_f attr__fprintf, void *priv)
1202 {
1203         char buf[BUF_SIZE];
1204         int ret = 0;
1205
1206         PRINT_ATTRf(type, p_unsigned);
1207         PRINT_ATTRf(size, p_unsigned);
1208         PRINT_ATTRf(config, p_hex);
1209         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1210         PRINT_ATTRf(sample_type, p_sample_type);
1211         PRINT_ATTRf(read_format, p_read_format);
1212
1213         PRINT_ATTRf(disabled, p_unsigned);
1214         PRINT_ATTRf(inherit, p_unsigned);
1215         PRINT_ATTRf(pinned, p_unsigned);
1216         PRINT_ATTRf(exclusive, p_unsigned);
1217         PRINT_ATTRf(exclude_user, p_unsigned);
1218         PRINT_ATTRf(exclude_kernel, p_unsigned);
1219         PRINT_ATTRf(exclude_hv, p_unsigned);
1220         PRINT_ATTRf(exclude_idle, p_unsigned);
1221         PRINT_ATTRf(mmap, p_unsigned);
1222         PRINT_ATTRf(comm, p_unsigned);
1223         PRINT_ATTRf(freq, p_unsigned);
1224         PRINT_ATTRf(inherit_stat, p_unsigned);
1225         PRINT_ATTRf(enable_on_exec, p_unsigned);
1226         PRINT_ATTRf(task, p_unsigned);
1227         PRINT_ATTRf(watermark, p_unsigned);
1228         PRINT_ATTRf(precise_ip, p_unsigned);
1229         PRINT_ATTRf(mmap_data, p_unsigned);
1230         PRINT_ATTRf(sample_id_all, p_unsigned);
1231         PRINT_ATTRf(exclude_host, p_unsigned);
1232         PRINT_ATTRf(exclude_guest, p_unsigned);
1233         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1234         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1235         PRINT_ATTRf(mmap2, p_unsigned);
1236         PRINT_ATTRf(comm_exec, p_unsigned);
1237         PRINT_ATTRf(use_clockid, p_unsigned);
1238         PRINT_ATTRf(context_switch, p_unsigned);
1239
1240         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1241         PRINT_ATTRf(bp_type, p_unsigned);
1242         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1243         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1244         PRINT_ATTRf(sample_regs_user, p_hex);
1245         PRINT_ATTRf(sample_stack_user, p_unsigned);
1246         PRINT_ATTRf(clockid, p_signed);
1247         PRINT_ATTRf(sample_regs_intr, p_hex);
1248         PRINT_ATTRf(aux_watermark, p_unsigned);
1249
1250         return ret;
1251 }
1252
1253 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1254                                 void *priv __attribute__((unused)))
1255 {
1256         return fprintf(fp, "  %-32s %s\n", name, val);
1257 }
1258
1259 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1260                               struct thread_map *threads)
1261 {
1262         int cpu, thread, nthreads;
1263         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1264         int pid = -1, err;
1265         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1266
1267         if (evsel->system_wide)
1268                 nthreads = 1;
1269         else
1270                 nthreads = threads->nr;
1271
1272         if (evsel->fd == NULL &&
1273             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1274                 return -ENOMEM;
1275
1276         if (evsel->cgrp) {
1277                 flags |= PERF_FLAG_PID_CGROUP;
1278                 pid = evsel->cgrp->fd;
1279         }
1280
1281 fallback_missing_features:
1282         if (perf_missing_features.clockid_wrong)
1283                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1284         if (perf_missing_features.clockid) {
1285                 evsel->attr.use_clockid = 0;
1286                 evsel->attr.clockid = 0;
1287         }
1288         if (perf_missing_features.cloexec)
1289                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1290         if (perf_missing_features.mmap2)
1291                 evsel->attr.mmap2 = 0;
1292         if (perf_missing_features.exclude_guest)
1293                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1294 retry_sample_id:
1295         if (perf_missing_features.sample_id_all)
1296                 evsel->attr.sample_id_all = 0;
1297
1298         if (verbose >= 2) {
1299                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1300                 fprintf(stderr, "perf_event_attr:\n");
1301                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1302                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1303         }
1304
1305         for (cpu = 0; cpu < cpus->nr; cpu++) {
1306
1307                 for (thread = 0; thread < nthreads; thread++) {
1308                         int group_fd;
1309
1310                         if (!evsel->cgrp && !evsel->system_wide)
1311                                 pid = thread_map__pid(threads, thread);
1312
1313                         group_fd = get_group_fd(evsel, cpu, thread);
1314 retry_open:
1315                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1316                                   pid, cpus->map[cpu], group_fd, flags);
1317
1318                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1319                                                                      pid,
1320                                                                      cpus->map[cpu],
1321                                                                      group_fd, flags);
1322                         if (FD(evsel, cpu, thread) < 0) {
1323                                 err = -errno;
1324                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1325                                           err);
1326                                 goto try_fallback;
1327                         }
1328                         set_rlimit = NO_CHANGE;
1329
1330                         /*
1331                          * If we succeeded but had to kill clockid, fail and
1332                          * have perf_evsel__open_strerror() print us a nice
1333                          * error.
1334                          */
1335                         if (perf_missing_features.clockid ||
1336                             perf_missing_features.clockid_wrong) {
1337                                 err = -EINVAL;
1338                                 goto out_close;
1339                         }
1340                 }
1341         }
1342
1343         return 0;
1344
1345 try_fallback:
1346         /*
1347          * perf stat needs between 5 and 22 fds per CPU. When we run out
1348          * of them try to increase the limits.
1349          */
1350         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1351                 struct rlimit l;
1352                 int old_errno = errno;
1353
1354                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1355                         if (set_rlimit == NO_CHANGE)
1356                                 l.rlim_cur = l.rlim_max;
1357                         else {
1358                                 l.rlim_cur = l.rlim_max + 1000;
1359                                 l.rlim_max = l.rlim_cur;
1360                         }
1361                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1362                                 set_rlimit++;
1363                                 errno = old_errno;
1364                                 goto retry_open;
1365                         }
1366                 }
1367                 errno = old_errno;
1368         }
1369
1370         if (err != -EINVAL || cpu > 0 || thread > 0)
1371                 goto out_close;
1372
1373         /*
1374          * Must probe features in the order they were added to the
1375          * perf_event_attr interface.
1376          */
1377         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1378                 perf_missing_features.clockid_wrong = true;
1379                 goto fallback_missing_features;
1380         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1381                 perf_missing_features.clockid = true;
1382                 goto fallback_missing_features;
1383         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1384                 perf_missing_features.cloexec = true;
1385                 goto fallback_missing_features;
1386         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1387                 perf_missing_features.mmap2 = true;
1388                 goto fallback_missing_features;
1389         } else if (!perf_missing_features.exclude_guest &&
1390                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1391                 perf_missing_features.exclude_guest = true;
1392                 goto fallback_missing_features;
1393         } else if (!perf_missing_features.sample_id_all) {
1394                 perf_missing_features.sample_id_all = true;
1395                 goto retry_sample_id;
1396         }
1397
1398 out_close:
1399         do {
1400                 while (--thread >= 0) {
1401                         close(FD(evsel, cpu, thread));
1402                         FD(evsel, cpu, thread) = -1;
1403                 }
1404                 thread = nthreads;
1405         } while (--cpu >= 0);
1406         return err;
1407 }
1408
1409 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1410 {
1411         if (evsel->fd == NULL)
1412                 return;
1413
1414         perf_evsel__close_fd(evsel, ncpus, nthreads);
1415         perf_evsel__free_fd(evsel);
1416 }
1417
1418 static struct {
1419         struct cpu_map map;
1420         int cpus[1];
1421 } empty_cpu_map = {
1422         .map.nr = 1,
1423         .cpus   = { -1, },
1424 };
1425
1426 static struct {
1427         struct thread_map map;
1428         int threads[1];
1429 } empty_thread_map = {
1430         .map.nr  = 1,
1431         .threads = { -1, },
1432 };
1433
1434 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1435                      struct thread_map *threads)
1436 {
1437         if (cpus == NULL) {
1438                 /* Work around old compiler warnings about strict aliasing */
1439                 cpus = &empty_cpu_map.map;
1440         }
1441
1442         if (threads == NULL)
1443                 threads = &empty_thread_map.map;
1444
1445         return __perf_evsel__open(evsel, cpus, threads);
1446 }
1447
1448 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1449                              struct cpu_map *cpus)
1450 {
1451         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1452 }
1453
1454 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1455                                 struct thread_map *threads)
1456 {
1457         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1458 }
1459
1460 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1461                                        const union perf_event *event,
1462                                        struct perf_sample *sample)
1463 {
1464         u64 type = evsel->attr.sample_type;
1465         const u64 *array = event->sample.array;
1466         bool swapped = evsel->needs_swap;
1467         union u64_swap u;
1468
1469         array += ((event->header.size -
1470                    sizeof(event->header)) / sizeof(u64)) - 1;
1471
1472         if (type & PERF_SAMPLE_IDENTIFIER) {
1473                 sample->id = *array;
1474                 array--;
1475         }
1476
1477         if (type & PERF_SAMPLE_CPU) {
1478                 u.val64 = *array;
1479                 if (swapped) {
1480                         /* undo swap of u64, then swap on individual u32s */
1481                         u.val64 = bswap_64(u.val64);
1482                         u.val32[0] = bswap_32(u.val32[0]);
1483                 }
1484
1485                 sample->cpu = u.val32[0];
1486                 array--;
1487         }
1488
1489         if (type & PERF_SAMPLE_STREAM_ID) {
1490                 sample->stream_id = *array;
1491                 array--;
1492         }
1493
1494         if (type & PERF_SAMPLE_ID) {
1495                 sample->id = *array;
1496                 array--;
1497         }
1498
1499         if (type & PERF_SAMPLE_TIME) {
1500                 sample->time = *array;
1501                 array--;
1502         }
1503
1504         if (type & PERF_SAMPLE_TID) {
1505                 u.val64 = *array;
1506                 if (swapped) {
1507                         /* undo swap of u64, then swap on individual u32s */
1508                         u.val64 = bswap_64(u.val64);
1509                         u.val32[0] = bswap_32(u.val32[0]);
1510                         u.val32[1] = bswap_32(u.val32[1]);
1511                 }
1512
1513                 sample->pid = u.val32[0];
1514                 sample->tid = u.val32[1];
1515                 array--;
1516         }
1517
1518         return 0;
1519 }
1520
1521 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1522                             u64 size)
1523 {
1524         return size > max_size || offset + size > endp;
1525 }
1526
1527 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1528         do {                                                            \
1529                 if (overflow(endp, (max_size), (offset), (size)))       \
1530                         return -EFAULT;                                 \
1531         } while (0)
1532
1533 #define OVERFLOW_CHECK_u64(offset) \
1534         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1535
1536 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1537                              struct perf_sample *data)
1538 {
1539         u64 type = evsel->attr.sample_type;
1540         bool swapped = evsel->needs_swap;
1541         const u64 *array;
1542         u16 max_size = event->header.size;
1543         const void *endp = (void *)event + max_size;
1544         u64 sz;
1545
1546         /*
1547          * used for cross-endian analysis. See git commit 65014ab3
1548          * for why this goofiness is needed.
1549          */
1550         union u64_swap u;
1551
1552         memset(data, 0, sizeof(*data));
1553         data->cpu = data->pid = data->tid = -1;
1554         data->stream_id = data->id = data->time = -1ULL;
1555         data->period = evsel->attr.sample_period;
1556         data->weight = 0;
1557
1558         if (event->header.type != PERF_RECORD_SAMPLE) {
1559                 if (!evsel->attr.sample_id_all)
1560                         return 0;
1561                 return perf_evsel__parse_id_sample(evsel, event, data);
1562         }
1563
1564         array = event->sample.array;
1565
1566         /*
1567          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1568          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1569          * check the format does not go past the end of the event.
1570          */
1571         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1572                 return -EFAULT;
1573
1574         data->id = -1ULL;
1575         if (type & PERF_SAMPLE_IDENTIFIER) {
1576                 data->id = *array;
1577                 array++;
1578         }
1579
1580         if (type & PERF_SAMPLE_IP) {
1581                 data->ip = *array;
1582                 array++;
1583         }
1584
1585         if (type & PERF_SAMPLE_TID) {
1586                 u.val64 = *array;
1587                 if (swapped) {
1588                         /* undo swap of u64, then swap on individual u32s */
1589                         u.val64 = bswap_64(u.val64);
1590                         u.val32[0] = bswap_32(u.val32[0]);
1591                         u.val32[1] = bswap_32(u.val32[1]);
1592                 }
1593
1594                 data->pid = u.val32[0];
1595                 data->tid = u.val32[1];
1596                 array++;
1597         }
1598
1599         if (type & PERF_SAMPLE_TIME) {
1600                 data->time = *array;
1601                 array++;
1602         }
1603
1604         data->addr = 0;
1605         if (type & PERF_SAMPLE_ADDR) {
1606                 data->addr = *array;
1607                 array++;
1608         }
1609
1610         if (type & PERF_SAMPLE_ID) {
1611                 data->id = *array;
1612                 array++;
1613         }
1614
1615         if (type & PERF_SAMPLE_STREAM_ID) {
1616                 data->stream_id = *array;
1617                 array++;
1618         }
1619
1620         if (type & PERF_SAMPLE_CPU) {
1621
1622                 u.val64 = *array;
1623                 if (swapped) {
1624                         /* undo swap of u64, then swap on individual u32s */
1625                         u.val64 = bswap_64(u.val64);
1626                         u.val32[0] = bswap_32(u.val32[0]);
1627                 }
1628
1629                 data->cpu = u.val32[0];
1630                 array++;
1631         }
1632
1633         if (type & PERF_SAMPLE_PERIOD) {
1634                 data->period = *array;
1635                 array++;
1636         }
1637
1638         if (type & PERF_SAMPLE_READ) {
1639                 u64 read_format = evsel->attr.read_format;
1640
1641                 OVERFLOW_CHECK_u64(array);
1642                 if (read_format & PERF_FORMAT_GROUP)
1643                         data->read.group.nr = *array;
1644                 else
1645                         data->read.one.value = *array;
1646
1647                 array++;
1648
1649                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1650                         OVERFLOW_CHECK_u64(array);
1651                         data->read.time_enabled = *array;
1652                         array++;
1653                 }
1654
1655                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1656                         OVERFLOW_CHECK_u64(array);
1657                         data->read.time_running = *array;
1658                         array++;
1659                 }
1660
1661                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1662                 if (read_format & PERF_FORMAT_GROUP) {
1663                         const u64 max_group_nr = UINT64_MAX /
1664                                         sizeof(struct sample_read_value);
1665
1666                         if (data->read.group.nr > max_group_nr)
1667                                 return -EFAULT;
1668                         sz = data->read.group.nr *
1669                              sizeof(struct sample_read_value);
1670                         OVERFLOW_CHECK(array, sz, max_size);
1671                         data->read.group.values =
1672                                         (struct sample_read_value *)array;
1673                         array = (void *)array + sz;
1674                 } else {
1675                         OVERFLOW_CHECK_u64(array);
1676                         data->read.one.id = *array;
1677                         array++;
1678                 }
1679         }
1680
1681         if (type & PERF_SAMPLE_CALLCHAIN) {
1682                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1683
1684                 OVERFLOW_CHECK_u64(array);
1685                 data->callchain = (struct ip_callchain *)array++;
1686                 if (data->callchain->nr > max_callchain_nr)
1687                         return -EFAULT;
1688                 sz = data->callchain->nr * sizeof(u64);
1689                 OVERFLOW_CHECK(array, sz, max_size);
1690                 array = (void *)array + sz;
1691         }
1692
1693         if (type & PERF_SAMPLE_RAW) {
1694                 OVERFLOW_CHECK_u64(array);
1695                 u.val64 = *array;
1696                 if (WARN_ONCE(swapped,
1697                               "Endianness of raw data not corrected!\n")) {
1698                         /* undo swap of u64, then swap on individual u32s */
1699                         u.val64 = bswap_64(u.val64);
1700                         u.val32[0] = bswap_32(u.val32[0]);
1701                         u.val32[1] = bswap_32(u.val32[1]);
1702                 }
1703                 data->raw_size = u.val32[0];
1704                 array = (void *)array + sizeof(u32);
1705
1706                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1707                 data->raw_data = (void *)array;
1708                 array = (void *)array + data->raw_size;
1709         }
1710
1711         if (type & PERF_SAMPLE_BRANCH_STACK) {
1712                 const u64 max_branch_nr = UINT64_MAX /
1713                                           sizeof(struct branch_entry);
1714
1715                 OVERFLOW_CHECK_u64(array);
1716                 data->branch_stack = (struct branch_stack *)array++;
1717
1718                 if (data->branch_stack->nr > max_branch_nr)
1719                         return -EFAULT;
1720                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1721                 OVERFLOW_CHECK(array, sz, max_size);
1722                 array = (void *)array + sz;
1723         }
1724
1725         if (type & PERF_SAMPLE_REGS_USER) {
1726                 OVERFLOW_CHECK_u64(array);
1727                 data->user_regs.abi = *array;
1728                 array++;
1729
1730                 if (data->user_regs.abi) {
1731                         u64 mask = evsel->attr.sample_regs_user;
1732
1733                         sz = hweight_long(mask) * sizeof(u64);
1734                         OVERFLOW_CHECK(array, sz, max_size);
1735                         data->user_regs.mask = mask;
1736                         data->user_regs.regs = (u64 *)array;
1737                         array = (void *)array + sz;
1738                 }
1739         }
1740
1741         if (type & PERF_SAMPLE_STACK_USER) {
1742                 OVERFLOW_CHECK_u64(array);
1743                 sz = *array++;
1744
1745                 data->user_stack.offset = ((char *)(array - 1)
1746                                           - (char *) event);
1747
1748                 if (!sz) {
1749                         data->user_stack.size = 0;
1750                 } else {
1751                         OVERFLOW_CHECK(array, sz, max_size);
1752                         data->user_stack.data = (char *)array;
1753                         array = (void *)array + sz;
1754                         OVERFLOW_CHECK_u64(array);
1755                         data->user_stack.size = *array++;
1756                         if (WARN_ONCE(data->user_stack.size > sz,
1757                                       "user stack dump failure\n"))
1758                                 return -EFAULT;
1759                 }
1760         }
1761
1762         data->weight = 0;
1763         if (type & PERF_SAMPLE_WEIGHT) {
1764                 OVERFLOW_CHECK_u64(array);
1765                 data->weight = *array;
1766                 array++;
1767         }
1768
1769         data->data_src = PERF_MEM_DATA_SRC_NONE;
1770         if (type & PERF_SAMPLE_DATA_SRC) {
1771                 OVERFLOW_CHECK_u64(array);
1772                 data->data_src = *array;
1773                 array++;
1774         }
1775
1776         data->transaction = 0;
1777         if (type & PERF_SAMPLE_TRANSACTION) {
1778                 OVERFLOW_CHECK_u64(array);
1779                 data->transaction = *array;
1780                 array++;
1781         }
1782
1783         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1784         if (type & PERF_SAMPLE_REGS_INTR) {
1785                 OVERFLOW_CHECK_u64(array);
1786                 data->intr_regs.abi = *array;
1787                 array++;
1788
1789                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1790                         u64 mask = evsel->attr.sample_regs_intr;
1791
1792                         sz = hweight_long(mask) * sizeof(u64);
1793                         OVERFLOW_CHECK(array, sz, max_size);
1794                         data->intr_regs.mask = mask;
1795                         data->intr_regs.regs = (u64 *)array;
1796                         array = (void *)array + sz;
1797                 }
1798         }
1799
1800         return 0;
1801 }
1802
1803 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1804                                      u64 read_format)
1805 {
1806         size_t sz, result = sizeof(struct sample_event);
1807
1808         if (type & PERF_SAMPLE_IDENTIFIER)
1809                 result += sizeof(u64);
1810
1811         if (type & PERF_SAMPLE_IP)
1812                 result += sizeof(u64);
1813
1814         if (type & PERF_SAMPLE_TID)
1815                 result += sizeof(u64);
1816
1817         if (type & PERF_SAMPLE_TIME)
1818                 result += sizeof(u64);
1819
1820         if (type & PERF_SAMPLE_ADDR)
1821                 result += sizeof(u64);
1822
1823         if (type & PERF_SAMPLE_ID)
1824                 result += sizeof(u64);
1825
1826         if (type & PERF_SAMPLE_STREAM_ID)
1827                 result += sizeof(u64);
1828
1829         if (type & PERF_SAMPLE_CPU)
1830                 result += sizeof(u64);
1831
1832         if (type & PERF_SAMPLE_PERIOD)
1833                 result += sizeof(u64);
1834
1835         if (type & PERF_SAMPLE_READ) {
1836                 result += sizeof(u64);
1837                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1838                         result += sizeof(u64);
1839                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1840                         result += sizeof(u64);
1841                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1842                 if (read_format & PERF_FORMAT_GROUP) {
1843                         sz = sample->read.group.nr *
1844                              sizeof(struct sample_read_value);
1845                         result += sz;
1846                 } else {
1847                         result += sizeof(u64);
1848                 }
1849         }
1850
1851         if (type & PERF_SAMPLE_CALLCHAIN) {
1852                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1853                 result += sz;
1854         }
1855
1856         if (type & PERF_SAMPLE_RAW) {
1857                 result += sizeof(u32);
1858                 result += sample->raw_size;
1859         }
1860
1861         if (type & PERF_SAMPLE_BRANCH_STACK) {
1862                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1863                 sz += sizeof(u64);
1864                 result += sz;
1865         }
1866
1867         if (type & PERF_SAMPLE_REGS_USER) {
1868                 if (sample->user_regs.abi) {
1869                         result += sizeof(u64);
1870                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1871                         result += sz;
1872                 } else {
1873                         result += sizeof(u64);
1874                 }
1875         }
1876
1877         if (type & PERF_SAMPLE_STACK_USER) {
1878                 sz = sample->user_stack.size;
1879                 result += sizeof(u64);
1880                 if (sz) {
1881                         result += sz;
1882                         result += sizeof(u64);
1883                 }
1884         }
1885
1886         if (type & PERF_SAMPLE_WEIGHT)
1887                 result += sizeof(u64);
1888
1889         if (type & PERF_SAMPLE_DATA_SRC)
1890                 result += sizeof(u64);
1891
1892         if (type & PERF_SAMPLE_TRANSACTION)
1893                 result += sizeof(u64);
1894
1895         if (type & PERF_SAMPLE_REGS_INTR) {
1896                 if (sample->intr_regs.abi) {
1897                         result += sizeof(u64);
1898                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1899                         result += sz;
1900                 } else {
1901                         result += sizeof(u64);
1902                 }
1903         }
1904
1905         return result;
1906 }
1907
1908 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1909                                   u64 read_format,
1910                                   const struct perf_sample *sample,
1911                                   bool swapped)
1912 {
1913         u64 *array;
1914         size_t sz;
1915         /*
1916          * used for cross-endian analysis. See git commit 65014ab3
1917          * for why this goofiness is needed.
1918          */
1919         union u64_swap u;
1920
1921         array = event->sample.array;
1922
1923         if (type & PERF_SAMPLE_IDENTIFIER) {
1924                 *array = sample->id;
1925                 array++;
1926         }
1927
1928         if (type & PERF_SAMPLE_IP) {
1929                 *array = sample->ip;
1930                 array++;
1931         }
1932
1933         if (type & PERF_SAMPLE_TID) {
1934                 u.val32[0] = sample->pid;
1935                 u.val32[1] = sample->tid;
1936                 if (swapped) {
1937                         /*
1938                          * Inverse of what is done in perf_evsel__parse_sample
1939                          */
1940                         u.val32[0] = bswap_32(u.val32[0]);
1941                         u.val32[1] = bswap_32(u.val32[1]);
1942                         u.val64 = bswap_64(u.val64);
1943                 }
1944
1945                 *array = u.val64;
1946                 array++;
1947         }
1948
1949         if (type & PERF_SAMPLE_TIME) {
1950                 *array = sample->time;
1951                 array++;
1952         }
1953
1954         if (type & PERF_SAMPLE_ADDR) {
1955                 *array = sample->addr;
1956                 array++;
1957         }
1958
1959         if (type & PERF_SAMPLE_ID) {
1960                 *array = sample->id;
1961                 array++;
1962         }
1963
1964         if (type & PERF_SAMPLE_STREAM_ID) {
1965                 *array = sample->stream_id;
1966                 array++;
1967         }
1968
1969         if (type & PERF_SAMPLE_CPU) {
1970                 u.val32[0] = sample->cpu;
1971                 if (swapped) {
1972                         /*
1973                          * Inverse of what is done in perf_evsel__parse_sample
1974                          */
1975                         u.val32[0] = bswap_32(u.val32[0]);
1976                         u.val64 = bswap_64(u.val64);
1977                 }
1978                 *array = u.val64;
1979                 array++;
1980         }
1981
1982         if (type & PERF_SAMPLE_PERIOD) {
1983                 *array = sample->period;
1984                 array++;
1985         }
1986
1987         if (type & PERF_SAMPLE_READ) {
1988                 if (read_format & PERF_FORMAT_GROUP)
1989                         *array = sample->read.group.nr;
1990                 else
1991                         *array = sample->read.one.value;
1992                 array++;
1993
1994                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1995                         *array = sample->read.time_enabled;
1996                         array++;
1997                 }
1998
1999                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2000                         *array = sample->read.time_running;
2001                         array++;
2002                 }
2003
2004                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2005                 if (read_format & PERF_FORMAT_GROUP) {
2006                         sz = sample->read.group.nr *
2007                              sizeof(struct sample_read_value);
2008                         memcpy(array, sample->read.group.values, sz);
2009                         array = (void *)array + sz;
2010                 } else {
2011                         *array = sample->read.one.id;
2012                         array++;
2013                 }
2014         }
2015
2016         if (type & PERF_SAMPLE_CALLCHAIN) {
2017                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2018                 memcpy(array, sample->callchain, sz);
2019                 array = (void *)array + sz;
2020         }
2021
2022         if (type & PERF_SAMPLE_RAW) {
2023                 u.val32[0] = sample->raw_size;
2024                 if (WARN_ONCE(swapped,
2025                               "Endianness of raw data not corrected!\n")) {
2026                         /*
2027                          * Inverse of what is done in perf_evsel__parse_sample
2028                          */
2029                         u.val32[0] = bswap_32(u.val32[0]);
2030                         u.val32[1] = bswap_32(u.val32[1]);
2031                         u.val64 = bswap_64(u.val64);
2032                 }
2033                 *array = u.val64;
2034                 array = (void *)array + sizeof(u32);
2035
2036                 memcpy(array, sample->raw_data, sample->raw_size);
2037                 array = (void *)array + sample->raw_size;
2038         }
2039
2040         if (type & PERF_SAMPLE_BRANCH_STACK) {
2041                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2042                 sz += sizeof(u64);
2043                 memcpy(array, sample->branch_stack, sz);
2044                 array = (void *)array + sz;
2045         }
2046
2047         if (type & PERF_SAMPLE_REGS_USER) {
2048                 if (sample->user_regs.abi) {
2049                         *array++ = sample->user_regs.abi;
2050                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2051                         memcpy(array, sample->user_regs.regs, sz);
2052                         array = (void *)array + sz;
2053                 } else {
2054                         *array++ = 0;
2055                 }
2056         }
2057
2058         if (type & PERF_SAMPLE_STACK_USER) {
2059                 sz = sample->user_stack.size;
2060                 *array++ = sz;
2061                 if (sz) {
2062                         memcpy(array, sample->user_stack.data, sz);
2063                         array = (void *)array + sz;
2064                         *array++ = sz;
2065                 }
2066         }
2067
2068         if (type & PERF_SAMPLE_WEIGHT) {
2069                 *array = sample->weight;
2070                 array++;
2071         }
2072
2073         if (type & PERF_SAMPLE_DATA_SRC) {
2074                 *array = sample->data_src;
2075                 array++;
2076         }
2077
2078         if (type & PERF_SAMPLE_TRANSACTION) {
2079                 *array = sample->transaction;
2080                 array++;
2081         }
2082
2083         if (type & PERF_SAMPLE_REGS_INTR) {
2084                 if (sample->intr_regs.abi) {
2085                         *array++ = sample->intr_regs.abi;
2086                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2087                         memcpy(array, sample->intr_regs.regs, sz);
2088                         array = (void *)array + sz;
2089                 } else {
2090                         *array++ = 0;
2091                 }
2092         }
2093
2094         return 0;
2095 }
2096
2097 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2098 {
2099         return pevent_find_field(evsel->tp_format, name);
2100 }
2101
2102 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2103                          const char *name)
2104 {
2105         struct format_field *field = perf_evsel__field(evsel, name);
2106         int offset;
2107
2108         if (!field)
2109                 return NULL;
2110
2111         offset = field->offset;
2112
2113         if (field->flags & FIELD_IS_DYNAMIC) {
2114                 offset = *(int *)(sample->raw_data + field->offset);
2115                 offset &= 0xffff;
2116         }
2117
2118         return sample->raw_data + offset;
2119 }
2120
2121 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2122                        const char *name)
2123 {
2124         struct format_field *field = perf_evsel__field(evsel, name);
2125         void *ptr;
2126         u64 value;
2127
2128         if (!field)
2129                 return 0;
2130
2131         ptr = sample->raw_data + field->offset;
2132
2133         switch (field->size) {
2134         case 1:
2135                 return *(u8 *)ptr;
2136         case 2:
2137                 value = *(u16 *)ptr;
2138                 break;
2139         case 4:
2140                 value = *(u32 *)ptr;
2141                 break;
2142         case 8:
2143                 memcpy(&value, ptr, sizeof(u64));
2144                 break;
2145         default:
2146                 return 0;
2147         }
2148
2149         if (!evsel->needs_swap)
2150                 return value;
2151
2152         switch (field->size) {
2153         case 2:
2154                 return bswap_16(value);
2155         case 4:
2156                 return bswap_32(value);
2157         case 8:
2158                 return bswap_64(value);
2159         default:
2160                 return 0;
2161         }
2162
2163         return 0;
2164 }
2165
2166 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2167 {
2168         va_list args;
2169         int ret = 0;
2170
2171         if (!*first) {
2172                 ret += fprintf(fp, ",");
2173         } else {
2174                 ret += fprintf(fp, ":");
2175                 *first = false;
2176         }
2177
2178         va_start(args, fmt);
2179         ret += vfprintf(fp, fmt, args);
2180         va_end(args);
2181         return ret;
2182 }
2183
2184 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2185 {
2186         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2187 }
2188
2189 int perf_evsel__fprintf(struct perf_evsel *evsel,
2190                         struct perf_attr_details *details, FILE *fp)
2191 {
2192         bool first = true;
2193         int printed = 0;
2194
2195         if (details->event_group) {
2196                 struct perf_evsel *pos;
2197
2198                 if (!perf_evsel__is_group_leader(evsel))
2199                         return 0;
2200
2201                 if (evsel->nr_members > 1)
2202                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2203
2204                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2205                 for_each_group_member(pos, evsel)
2206                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2207
2208                 if (evsel->nr_members > 1)
2209                         printed += fprintf(fp, "}");
2210                 goto out;
2211         }
2212
2213         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2214
2215         if (details->verbose) {
2216                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2217                                                     __print_attr__fprintf, &first);
2218         } else if (details->freq) {
2219                 const char *term = "sample_freq";
2220
2221                 if (!evsel->attr.freq)
2222                         term = "sample_period";
2223
2224                 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2225                                          term, (u64)evsel->attr.sample_freq);
2226         }
2227 out:
2228         fputc('\n', fp);
2229         return ++printed;
2230 }
2231
2232 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2233                           char *msg, size_t msgsize)
2234 {
2235         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2236             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2237             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2238                 /*
2239                  * If it's cycles then fall back to hrtimer based
2240                  * cpu-clock-tick sw counter, which is always available even if
2241                  * no PMU support.
2242                  *
2243                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2244                  * b0a873e).
2245                  */
2246                 scnprintf(msg, msgsize, "%s",
2247 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2248
2249                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2250                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2251
2252                 zfree(&evsel->name);
2253                 return true;
2254         }
2255
2256         return false;
2257 }
2258
2259 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2260                               int err, char *msg, size_t size)
2261 {
2262         char sbuf[STRERR_BUFSIZE];
2263
2264         switch (err) {
2265         case EPERM:
2266         case EACCES:
2267                 return scnprintf(msg, size,
2268                  "You may not have permission to collect %sstats.\n"
2269                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2270                  " -1 - Not paranoid at all\n"
2271                  "  0 - Disallow raw tracepoint access for unpriv\n"
2272                  "  1 - Disallow cpu events for unpriv\n"
2273                  "  2 - Disallow kernel profiling for unpriv",
2274                                  target->system_wide ? "system-wide " : "");
2275         case ENOENT:
2276                 return scnprintf(msg, size, "The %s event is not supported.",
2277                                  perf_evsel__name(evsel));
2278         case EMFILE:
2279                 return scnprintf(msg, size, "%s",
2280                          "Too many events are opened.\n"
2281                          "Probably the maximum number of open file descriptors has been reached.\n"
2282                          "Hint: Try again after reducing the number of events.\n"
2283                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2284         case ENODEV:
2285                 if (target->cpu_list)
2286                         return scnprintf(msg, size, "%s",
2287          "No such device - did you specify an out-of-range profile CPU?\n");
2288                 break;
2289         case EOPNOTSUPP:
2290                 if (evsel->attr.precise_ip)
2291                         return scnprintf(msg, size, "%s",
2292         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2293 #if defined(__i386__) || defined(__x86_64__)
2294                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2295                         return scnprintf(msg, size, "%s",
2296         "No hardware sampling interrupt available.\n"
2297         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2298 #endif
2299                 break;
2300         case EBUSY:
2301                 if (find_process("oprofiled"))
2302                         return scnprintf(msg, size,
2303         "The PMU counters are busy/taken by another profiler.\n"
2304         "We found oprofile daemon running, please stop it and try again.");
2305                 break;
2306         case EINVAL:
2307                 if (perf_missing_features.clockid)
2308                         return scnprintf(msg, size, "clockid feature not supported.");
2309                 if (perf_missing_features.clockid_wrong)
2310                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2311                 break;
2312         default:
2313                 break;
2314         }
2315
2316         return scnprintf(msg, size,
2317         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2318         "/bin/dmesg may provide additional information.\n"
2319         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2320                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2321                          perf_evsel__name(evsel));
2322 }