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 "evsel.h"
  19 #include "evlist.h"
  20 #include "util.h"
  21 #include "cpumap.h"
  22 #include "thread_map.h"
  23 #include "target.h"
  24 #include "perf_regs.h"
  25 #include "debug.h"
  26 #include "trace-event.h"
  27
  28 static struct {
  29         bool sample_id_all;
  30         bool exclude_guest;
  31         bool mmap2;
  32         bool cloexec;
  33 } perf_missing_features;
  34
  35 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  36
  37 int __perf_evsel__sample_size(u64 sample_type)
  38 {
  39         u64 mask = sample_type & PERF_SAMPLE_MASK;
  40         int size = 0;
  41         int i;
  42
  43         for (i = 0; i < 64; i++) {
  44                 if (mask & (1ULL << i))
  45                         size++;
  46         }
  47
  48         size *= sizeof(u64);
  49
  50         return size;
  51 }
  52
  53 /**
  54  * __perf_evsel__calc_id_pos - calculate id_pos.
  55  * @sample_type: sample type
  56  *
  57  * This function returns the position of the event id (PERF_SAMPLE_ID or
  58  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
  59  * sample_event.
  60  */
  61 static int __perf_evsel__calc_id_pos(u64 sample_type)
  62 {
  63         int idx = 0;
  64
  65         if (sample_type & PERF_SAMPLE_IDENTIFIER)
  66                 return 0;
  67
  68         if (!(sample_type & PERF_SAMPLE_ID))
  69                 return -1;
  70
  71         if (sample_type & PERF_SAMPLE_IP)
  72                 idx += 1;
  73
  74         if (sample_type & PERF_SAMPLE_TID)
  75                 idx += 1;
  76
  77         if (sample_type & PERF_SAMPLE_TIME)
  78                 idx += 1;
  79
  80         if (sample_type & PERF_SAMPLE_ADDR)
  81                 idx += 1;
  82
  83         return idx;
  84 }
  85
  86 /**
  87  * __perf_evsel__calc_is_pos - calculate is_pos.
  88  * @sample_type: sample type
  89  *
  90  * This function returns the position (counting backwards) of the event id
  91  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
  92  * sample_id_all is used there is an id sample appended to non-sample events.
  93  */
  94 static int __perf_evsel__calc_is_pos(u64 sample_type)
  95 {
  96         int idx = 1;
  97
  98         if (sample_type & PERF_SAMPLE_IDENTIFIER)
  99                 return 1;
 100
 101         if (!(sample_type & PERF_SAMPLE_ID))
 102                 return -1;
 103
 104         if (sample_type & PERF_SAMPLE_CPU)
 105                 idx += 1;
 106
 107         if (sample_type & PERF_SAMPLE_STREAM_ID)
 108                 idx += 1;
 109
 110         return idx;
 111 }
 112
 113 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 114 {
 115         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 116         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 117 }
 118
 119 void hists__init(struct hists *hists)
 120 {
 121         memset(hists, 0, sizeof(*hists));
 122         hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
 123         hists->entries_in = &hists->entries_in_array[0];
 124         hists->entries_collapsed = RB_ROOT;
 125         hists->entries = RB_ROOT;
 126         pthread_mutex_init(&hists->lock, NULL);
 127 }
 128
 129 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 130                                   enum perf_event_sample_format bit)
 131 {
 132         if (!(evsel->attr.sample_type & bit)) {
 133                 evsel->attr.sample_type |= bit;
 134                 evsel->sample_size += sizeof(u64);
 135                 perf_evsel__calc_id_pos(evsel);
 136         }
 137 }
 138
 139 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 140                                     enum perf_event_sample_format bit)
 141 {
 142         if (evsel->attr.sample_type & bit) {
 143                 evsel->attr.sample_type &= ~bit;
 144                 evsel->sample_size -= sizeof(u64);
 145                 perf_evsel__calc_id_pos(evsel);
 146         }
 147 }
 148
 149 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 150                                bool can_sample_identifier)
 151 {
 152         if (can_sample_identifier) {
 153                 perf_evsel__reset_sample_bit(evsel, ID);
 154                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 155         } else {
 156                 perf_evsel__set_sample_bit(evsel, ID);
 157         }
 158         evsel->attr.read_format |= PERF_FORMAT_ID;
 159 }
 160
 161 void perf_evsel__init(struct perf_evsel *evsel,
 162                       struct perf_event_attr *attr, int idx)
 163 {
 164         evsel->idx         = idx;
 165         evsel->tracking    = !idx;
 166         evsel->attr        = *attr;
 167         evsel->leader      = evsel;
 168         evsel->unit        = "";
 169         evsel->scale       = 1.0;
 170         INIT_LIST_HEAD(&evsel->node);
 171         hists__init(&evsel->hists);
 172         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 173         perf_evsel__calc_id_pos(evsel);
 174 }
 175
 176 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 177 {
 178         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
 179
 180         if (evsel != NULL)
 181                 perf_evsel__init(evsel, attr, idx);
 182
 183         return evsel;
 184 }
 185
 186 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 187 {
 188         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
 189
 190         if (evsel != NULL) {
 191                 struct perf_event_attr attr = {
 192                         .type          = PERF_TYPE_TRACEPOINT,
 193                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 194                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 195                 };
 196
 197                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 198                         goto out_free;
 199
 200                 evsel->tp_format = trace_event__tp_format(sys, name);
 201                 if (evsel->tp_format == NULL)
 202                         goto out_free;
 203
 204                 event_attr_init(&attr);
 205                 attr.config = evsel->tp_format->id;
 206                 attr.sample_period = 1;
 207                 perf_evsel__init(evsel, &attr, idx);
 208         }
 209
 210         return evsel;
 211
 212 out_free:
 213         zfree(&evsel->name);
 214         free(evsel);
 215         return NULL;
 216 }
 217
 218 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 219         "cycles",
 220         "instructions",
 221         "cache-references",
 222         "cache-misses",
 223         "branches",
 224         "branch-misses",
 225         "bus-cycles",
 226         "stalled-cycles-frontend",
 227         "stalled-cycles-backend",
 228         "ref-cycles",
 229 };
 230
 231 static const char *__perf_evsel__hw_name(u64 config)
 232 {
 233         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 234                 return perf_evsel__hw_names[config];
 235
 236         return "unknown-hardware";
 237 }
 238
 239 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 240 {
 241         int colon = 0, r = 0;
 242         struct perf_event_attr *attr = &evsel->attr;
 243         bool exclude_guest_default = false;
 244
 245 #define MOD_PRINT(context, mod) do {                                    \
 246                 if (!attr->exclude_##context) {                         \
 247                         if (!colon) colon = ++r;                        \
 248                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 249                 } } while(0)
 250
 251         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 252                 MOD_PRINT(kernel, 'k');
 253                 MOD_PRINT(user, 'u');
 254                 MOD_PRINT(hv, 'h');
 255                 exclude_guest_default = true;
 256         }
 257
 258         if (attr->precise_ip) {
 259                 if (!colon)
 260                         colon = ++r;
 261                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 262                 exclude_guest_default = true;
 263         }
 264
 265         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 266                 MOD_PRINT(host, 'H');
 267                 MOD_PRINT(guest, 'G');
 268         }
 269 #undef MOD_PRINT
 270         if (colon)
 271                 bf[colon - 1] = ':';
 272         return r;
 273 }
 274
 275 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 276 {
 277         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 278         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 279 }
 280
 281 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 282         "cpu-clock",
 283         "task-clock",
 284         "page-faults",
 285         "context-switches",
 286         "cpu-migrations",
 287         "minor-faults",
 288         "major-faults",
 289         "alignment-faults",
 290         "emulation-faults",
 291         "dummy",
 292 };
 293
 294 static const char *__perf_evsel__sw_name(u64 config)
 295 {
 296         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 297                 return perf_evsel__sw_names[config];
 298         return "unknown-software";
 299 }
 300
 301 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 302 {
 303         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 304         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 305 }
 306
 307 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 308 {
 309         int r;
 310
 311         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 312
 313         if (type & HW_BREAKPOINT_R)
 314                 r += scnprintf(bf + r, size - r, "r");
 315
 316         if (type & HW_BREAKPOINT_W)
 317                 r += scnprintf(bf + r, size - r, "w");
 318
 319         if (type & HW_BREAKPOINT_X)
 320                 r += scnprintf(bf + r, size - r, "x");
 321
 322         return r;
 323 }
 324
 325 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 326 {
 327         struct perf_event_attr *attr = &evsel->attr;
 328         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 329         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 330 }
 331
 332 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 333                                 [PERF_EVSEL__MAX_ALIASES] = {
 334  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 335  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 336  { "LLC",       "L2",                                                   },
 337  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 338  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 339  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 340  { "node",                                                              },
 341 };
 342
 343 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 344                                    [PERF_EVSEL__MAX_ALIASES] = {
 345  { "load",      "loads",        "read",                                 },
 346  { "store",     "stores",       "write",                                },
 347  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 348 };
 349
 350 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 351                                        [PERF_EVSEL__MAX_ALIASES] = {
 352  { "refs",      "Reference",    "ops",          "access",               },
 353  { "misses",    "miss",                                                 },
 354 };
 355
 356 #define C(x)            PERF_COUNT_HW_CACHE_##x
 357 #define CACHE_READ      (1 << C(OP_READ))
 358 #define CACHE_WRITE     (1 << C(OP_WRITE))
 359 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 360 #define COP(x)          (1 << x)
 361
 362 /*
 363  * cache operartion stat
 364  * L1I : Read and prefetch only
 365  * ITLB and BPU : Read-only
 366  */
 367 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 368  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 369  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 370  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 371  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 372  [C(ITLB)]      = (CACHE_READ),
 373  [C(BPU)]       = (CACHE_READ),
 374  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 375 };
 376
 377 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 378 {
 379         if (perf_evsel__hw_cache_stat[type] & COP(op))
 380                 return true;    /* valid */
 381         else
 382                 return false;   /* invalid */
 383 }
 384
 385 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 386                                             char *bf, size_t size)
 387 {
 388         if (result) {
 389                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 390                                  perf_evsel__hw_cache_op[op][0],
 391                                  perf_evsel__hw_cache_result[result][0]);
 392         }
 393
 394         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 395                          perf_evsel__hw_cache_op[op][1]);
 396 }
 397
 398 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 399 {
 400         u8 op, result, type = (config >>  0) & 0xff;
 401         const char *err = "unknown-ext-hardware-cache-type";
 402
 403         if (type > PERF_COUNT_HW_CACHE_MAX)
 404                 goto out_err;
 405
 406         op = (config >>  8) & 0xff;
 407         err = "unknown-ext-hardware-cache-op";
 408         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 409                 goto out_err;
 410
 411         result = (config >> 16) & 0xff;
 412         err = "unknown-ext-hardware-cache-result";
 413         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 414                 goto out_err;
 415
 416         err = "invalid-cache";
 417         if (!perf_evsel__is_cache_op_valid(type, op))
 418                 goto out_err;
 419
 420         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 421 out_err:
 422         return scnprintf(bf, size, "%s", err);
 423 }
 424
 425 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 426 {
 427         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 428         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 429 }
 430
 431 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 432 {
 433         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 434         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 435 }
 436
 437 const char *perf_evsel__name(struct perf_evsel *evsel)
 438 {
 439         char bf[128];
 440
 441         if (evsel->name)
 442                 return evsel->name;
 443
 444         switch (evsel->attr.type) {
 445         case PERF_TYPE_RAW:
 446                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 447                 break;
 448
 449         case PERF_TYPE_HARDWARE:
 450                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 451                 break;
 452
 453         case PERF_TYPE_HW_CACHE:
 454                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 455                 break;
 456
 457         case PERF_TYPE_SOFTWARE:
 458                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 459                 break;
 460
 461         case PERF_TYPE_TRACEPOINT:
 462                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 463                 break;
 464
 465         case PERF_TYPE_BREAKPOINT:
 466                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 467                 break;
 468
 469         default:
 470                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 471                           evsel->attr.type);
 472                 break;
 473         }
 474
 475         evsel->name = strdup(bf);
 476
 477         return evsel->name ?: "unknown";
 478 }
 479
 480 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 481 {
 482         return evsel->group_name ?: "anon group";
 483 }
 484
 485 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 486 {
 487         int ret;
 488         struct perf_evsel *pos;
 489         const char *group_name = perf_evsel__group_name(evsel);
 490
 491         ret = scnprintf(buf, size, "%s", group_name);
 492
 493         ret += scnprintf(buf + ret, size - ret, " { %s",
 494                          perf_evsel__name(evsel));
 495
 496         for_each_group_member(pos, evsel)
 497                 ret += scnprintf(buf + ret, size - ret, ", %s",
 498                                  perf_evsel__name(pos));
 499
 500         ret += scnprintf(buf + ret, size - ret, " }");
 501
 502         return ret;
 503 }
 504
 505 static void
 506 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 507                              struct record_opts *opts)
 508 {
 509         bool function = perf_evsel__is_function_event(evsel);
 510         struct perf_event_attr *attr = &evsel->attr;
 511
 512         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 513
 514         if (opts->call_graph == CALLCHAIN_DWARF) {
 515                 if (!function) {
 516                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 517                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 518                         attr->sample_regs_user = PERF_REGS_MASK;
 519                         attr->sample_stack_user = opts->stack_dump_size;
 520                         attr->exclude_callchain_user = 1;
 521                 } else {
 522                         pr_info("Cannot use DWARF unwind for function trace event,"
 523                                 " falling back to framepointers.\n");
 524                 }
 525         }
 526
 527         if (function) {
 528                 pr_info("Disabling user space callchains for function trace event.\n");
 529                 attr->exclude_callchain_user = 1;
 530         }
 531 }
 532
 533 /*
 534  * The enable_on_exec/disabled value strategy:
 535  *
 536  *  1) For any type of traced program:
 537  *    - all independent events and group leaders are disabled
 538  *    - all group members are enabled
 539  *
 540  *     Group members are ruled by group leaders. They need to
 541  *     be enabled, because the group scheduling relies on that.
 542  *
 543  *  2) For traced programs executed by perf:
 544  *     - all independent events and group leaders have
 545  *       enable_on_exec set
 546  *     - we don't specifically enable or disable any event during
 547  *       the record command
 548  *
 549  *     Independent events and group leaders are initially disabled
 550  *     and get enabled by exec. Group members are ruled by group
 551  *     leaders as stated in 1).
 552  *
 553  *  3) For traced programs attached by perf (pid/tid):
 554  *     - we specifically enable or disable all events during
 555  *       the record command
 556  *
 557  *     When attaching events to already running traced we
 558  *     enable/disable events specifically, as there's no
 559  *     initial traced exec call.
 560  */
 561 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 562 {
 563         struct perf_evsel *leader = evsel->leader;
 564         struct perf_event_attr *attr = &evsel->attr;
 565         int track = evsel->tracking;
 566         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 567
 568         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 569         attr->inherit       = !opts->no_inherit;
 570
 571         perf_evsel__set_sample_bit(evsel, IP);
 572         perf_evsel__set_sample_bit(evsel, TID);
 573
 574         if (evsel->sample_read) {
 575                 perf_evsel__set_sample_bit(evsel, READ);
 576
 577                 /*
 578                  * We need ID even in case of single event, because
 579                  * PERF_SAMPLE_READ process ID specific data.
 580                  */
 581                 perf_evsel__set_sample_id(evsel, false);
 582
 583                 /*
 584                  * Apply group format only if we belong to group
 585                  * with more than one members.
 586                  */
 587                 if (leader->nr_members > 1) {
 588                         attr->read_format |= PERF_FORMAT_GROUP;
 589                         attr->inherit = 0;
 590                 }
 591         }
 592
 593         /*
 594          * We default some events to have a default interval. But keep
 595          * it a weak assumption overridable by the user.
 596          */
 597         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 598                                      opts->user_interval != ULLONG_MAX)) {
 599                 if (opts->freq) {
 600                         perf_evsel__set_sample_bit(evsel, PERIOD);
 601                         attr->freq              = 1;
 602                         attr->sample_freq       = opts->freq;
 603                 } else {
 604                         attr->sample_period = opts->default_interval;
 605                 }
 606         }
 607
 608         /*
 609          * Disable sampling for all group members other
 610          * than leader in case leader 'leads' the sampling.
 611          */
 612         if ((leader != evsel) && leader->sample_read) {
 613                 attr->sample_freq   = 0;
 614                 attr->sample_period = 0;
 615         }
 616
 617         if (opts->no_samples)
 618                 attr->sample_freq = 0;
 619
 620         if (opts->inherit_stat)
 621                 attr->inherit_stat = 1;
 622
 623         if (opts->sample_address) {
 624                 perf_evsel__set_sample_bit(evsel, ADDR);
 625                 attr->mmap_data = track;
 626         }
 627
 628         if (opts->call_graph_enabled && !evsel->no_aux_samples)
 629                 perf_evsel__config_callgraph(evsel, opts);
 630
 631         if (target__has_cpu(&opts->target))
 632                 perf_evsel__set_sample_bit(evsel, CPU);
 633
 634         if (opts->period)
 635                 perf_evsel__set_sample_bit(evsel, PERIOD);
 636
 637         /*
 638          * When the user explicitely disabled time don't force it here.
 639          */
 640         if (opts->sample_time &&
 641             (!perf_missing_features.sample_id_all &&
 642             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
 643                 perf_evsel__set_sample_bit(evsel, TIME);
 644
 645         if (opts->raw_samples && !evsel->no_aux_samples) {
 646                 perf_evsel__set_sample_bit(evsel, TIME);
 647                 perf_evsel__set_sample_bit(evsel, RAW);
 648                 perf_evsel__set_sample_bit(evsel, CPU);
 649         }
 650
 651         if (opts->sample_address)
 652                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 653
 654         if (opts->no_buffering) {
 655                 attr->watermark = 0;
 656                 attr->wakeup_events = 1;
 657         }
 658         if (opts->branch_stack && !evsel->no_aux_samples) {
 659                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 660                 attr->branch_sample_type = opts->branch_stack;
 661         }
 662
 663         if (opts->sample_weight)
 664                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 665
 666         attr->mmap  = track;
 667         attr->mmap2 = track && !perf_missing_features.mmap2;
 668         attr->comm  = track;
 669
 670         if (opts->sample_transaction)
 671                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 672
 673         /*
 674          * XXX see the function comment above
 675          *
 676          * Disabling only independent events or group leaders,
 677          * keeping group members enabled.
 678          */
 679         if (perf_evsel__is_group_leader(evsel))
 680                 attr->disabled = 1;
 681
 682         /*
 683          * Setting enable_on_exec for independent events and
 684          * group leaders for traced executed by perf.
 685          */
 686         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 687                 !opts->initial_delay)
 688                 attr->enable_on_exec = 1;
 689
 690         if (evsel->immediate) {
 691                 attr->disabled = 0;
 692                 attr->enable_on_exec = 0;
 693         }
 694 }
 695
 696 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 697 {
 698         int cpu, thread;
 699
 700         if (evsel->system_wide)
 701                 nthreads = 1;
 702
 703         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 704
 705         if (evsel->fd) {
 706                 for (cpu = 0; cpu < ncpus; cpu++) {
 707                         for (thread = 0; thread < nthreads; thread++) {
 708                                 FD(evsel, cpu, thread) = -1;
 709                         }
 710                 }
 711         }
 712
 713         return evsel->fd != NULL ? 0 : -ENOMEM;
 714 }
 715
 716 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 717                           int ioc,  void *arg)
 718 {
 719         int cpu, thread;
 720
 721         if (evsel->system_wide)
 722                 nthreads = 1;
 723
 724         for (cpu = 0; cpu < ncpus; cpu++) {
 725                 for (thread = 0; thread < nthreads; thread++) {
 726                         int fd = FD(evsel, cpu, thread),
 727                             err = ioctl(fd, ioc, arg);
 728
 729                         if (err)
 730                                 return err;
 731                 }
 732         }
 733
 734         return 0;
 735 }
 736
 737 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 738                            const char *filter)
 739 {
 740         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 741                                      PERF_EVENT_IOC_SET_FILTER,
 742                                      (void *)filter);
 743 }
 744
 745 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 746 {
 747         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 748                                      PERF_EVENT_IOC_ENABLE,
 749                                      0);
 750 }
 751
 752 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 753 {
 754         if (evsel->system_wide)
 755                 nthreads = 1;
 756
 757         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 758         if (evsel->sample_id == NULL)
 759                 return -ENOMEM;
 760
 761         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 762         if (evsel->id == NULL) {
 763                 xyarray__delete(evsel->sample_id);
 764                 evsel->sample_id = NULL;
 765                 return -ENOMEM;
 766         }
 767
 768         return 0;
 769 }
 770
 771 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
 772 {
 773         memset(evsel->counts, 0, (sizeof(*evsel->counts) +
 774                                  (ncpus * sizeof(struct perf_counts_values))));
 775 }
 776
 777 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
 778 {
 779         evsel->counts = zalloc((sizeof(*evsel->counts) +
 780                                 (ncpus * sizeof(struct perf_counts_values))));
 781         return evsel->counts != NULL ? 0 : -ENOMEM;
 782 }
 783
 784 void perf_evsel__free_fd(struct perf_evsel *evsel)
 785 {
 786         xyarray__delete(evsel->fd);
 787         evsel->fd = NULL;
 788 }
 789
 790 void perf_evsel__free_id(struct perf_evsel *evsel)
 791 {
 792         xyarray__delete(evsel->sample_id);
 793         evsel->sample_id = NULL;
 794         zfree(&evsel->id);
 795 }
 796
 797 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 798 {
 799         int cpu, thread;
 800
 801         if (evsel->system_wide)
 802                 nthreads = 1;
 803
 804         for (cpu = 0; cpu < ncpus; cpu++)
 805                 for (thread = 0; thread < nthreads; ++thread) {
 806                         close(FD(evsel, cpu, thread));
 807                         FD(evsel, cpu, thread) = -1;
 808                 }
 809 }
 810
 811 void perf_evsel__free_counts(struct perf_evsel *evsel)
 812 {
 813         zfree(&evsel->counts);
 814 }
 815
 816 void perf_evsel__exit(struct perf_evsel *evsel)
 817 {
 818         assert(list_empty(&evsel->node));
 819         perf_evsel__free_fd(evsel);
 820         perf_evsel__free_id(evsel);
 821 }
 822
 823 void perf_evsel__delete(struct perf_evsel *evsel)
 824 {
 825         perf_evsel__exit(evsel);
 826         close_cgroup(evsel->cgrp);
 827         zfree(&evsel->group_name);
 828         if (evsel->tp_format)
 829                 pevent_free_format(evsel->tp_format);
 830         zfree(&evsel->name);
 831         free(evsel);
 832 }
 833
 834 static inline void compute_deltas(struct perf_evsel *evsel,
 835                                   int cpu,
 836                                   struct perf_counts_values *count)
 837 {
 838         struct perf_counts_values tmp;
 839
 840         if (!evsel->prev_raw_counts)
 841                 return;
 842
 843         if (cpu == -1) {
 844                 tmp = evsel->prev_raw_counts->aggr;
 845                 evsel->prev_raw_counts->aggr = *count;
 846         } else {
 847                 tmp = evsel->prev_raw_counts->cpu[cpu];
 848                 evsel->prev_raw_counts->cpu[cpu] = *count;
 849         }
 850
 851         count->val = count->val - tmp.val;
 852         count->ena = count->ena - tmp.ena;
 853         count->run = count->run - tmp.run;
 854 }
 855
 856 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 857                               int cpu, int thread, bool scale)
 858 {
 859         struct perf_counts_values count;
 860         size_t nv = scale ? 3 : 1;
 861
 862         if (FD(evsel, cpu, thread) < 0)
 863                 return -EINVAL;
 864
 865         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 866                 return -ENOMEM;
 867
 868         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 869                 return -errno;
 870
 871         compute_deltas(evsel, cpu, &count);
 872
 873         if (scale) {
 874                 if (count.run == 0)
 875                         count.val = 0;
 876                 else if (count.run < count.ena)
 877                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
 878         } else
 879                 count.ena = count.run = 0;
 880
 881         evsel->counts->cpu[cpu] = count;
 882         return 0;
 883 }
 884
 885 int __perf_evsel__read(struct perf_evsel *evsel,
 886                        int ncpus, int nthreads, bool scale)
 887 {
 888         size_t nv = scale ? 3 : 1;
 889         int cpu, thread;
 890         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
 891
 892         if (evsel->system_wide)
 893                 nthreads = 1;
 894
 895         aggr->val = aggr->ena = aggr->run = 0;
 896
 897         for (cpu = 0; cpu < ncpus; cpu++) {
 898                 for (thread = 0; thread < nthreads; thread++) {
 899                         if (FD(evsel, cpu, thread) < 0)
 900                                 continue;
 901
 902                         if (readn(FD(evsel, cpu, thread),
 903                                   &count, nv * sizeof(u64)) < 0)
 904                                 return -errno;
 905
 906                         aggr->val += count.val;
 907                         if (scale) {
 908                                 aggr->ena += count.ena;
 909                                 aggr->run += count.run;
 910                         }
 911                 }
 912         }
 913
 914         compute_deltas(evsel, -1, aggr);
 915
 916         evsel->counts->scaled = 0;
 917         if (scale) {
 918                 if (aggr->run == 0) {
 919                         evsel->counts->scaled = -1;
 920                         aggr->val = 0;
 921                         return 0;
 922                 }
 923
 924                 if (aggr->run < aggr->ena) {
 925                         evsel->counts->scaled = 1;
 926                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 927                 }
 928         } else
 929                 aggr->ena = aggr->run = 0;
 930
 931         return 0;
 932 }
 933
 934 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 935 {
 936         struct perf_evsel *leader = evsel->leader;
 937         int fd;
 938
 939         if (perf_evsel__is_group_leader(evsel))
 940                 return -1;
 941
 942         /*
 943          * Leader must be already processed/open,
 944          * if not it's a bug.
 945          */
 946         BUG_ON(!leader->fd);
 947
 948         fd = FD(leader, cpu, thread);
 949         BUG_ON(fd == -1);
 950
 951         return fd;
 952 }
 953
 954 #define __PRINT_ATTR(fmt, cast, field)  \
 955         fprintf(fp, "  %-19s "fmt"\n", #field, cast attr->field)
 956
 957 #define PRINT_ATTR_U32(field)  __PRINT_ATTR("%u" , , field)
 958 #define PRINT_ATTR_X32(field)  __PRINT_ATTR("%#x", , field)
 959 #define PRINT_ATTR_U64(field)  __PRINT_ATTR("%" PRIu64, (uint64_t), field)
 960 #define PRINT_ATTR_X64(field)  __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
 961
 962 #define PRINT_ATTR2N(name1, field1, name2, field2)      \
 963         fprintf(fp, "  %-19s %u    %-19s %u\n",         \
 964         name1, attr->field1, name2, attr->field2)
 965
 966 #define PRINT_ATTR2(field1, field2) \
 967         PRINT_ATTR2N(#field1, field1, #field2, field2)
 968
 969 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
 970 {
 971         size_t ret = 0;
 972
 973         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
 974         ret += fprintf(fp, "perf_event_attr:\n");
 975
 976         ret += PRINT_ATTR_U32(type);
 977         ret += PRINT_ATTR_U32(size);
 978         ret += PRINT_ATTR_X64(config);
 979         ret += PRINT_ATTR_U64(sample_period);
 980         ret += PRINT_ATTR_U64(sample_freq);
 981         ret += PRINT_ATTR_X64(sample_type);
 982         ret += PRINT_ATTR_X64(read_format);
 983
 984         ret += PRINT_ATTR2(disabled, inherit);
 985         ret += PRINT_ATTR2(pinned, exclusive);
 986         ret += PRINT_ATTR2(exclude_user, exclude_kernel);
 987         ret += PRINT_ATTR2(exclude_hv, exclude_idle);
 988         ret += PRINT_ATTR2(mmap, comm);
 989         ret += PRINT_ATTR2(mmap2, comm_exec);
 990         ret += PRINT_ATTR2(freq, inherit_stat);
 991         ret += PRINT_ATTR2(enable_on_exec, task);
 992         ret += PRINT_ATTR2(watermark, precise_ip);
 993         ret += PRINT_ATTR2(mmap_data, sample_id_all);
 994         ret += PRINT_ATTR2(exclude_host, exclude_guest);
 995         ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
 996                             "excl.callchain_user", exclude_callchain_user);
 997
 998         ret += PRINT_ATTR_U32(wakeup_events);
 999         ret += PRINT_ATTR_U32(wakeup_watermark);
1000         ret += PRINT_ATTR_X32(bp_type);
1001         ret += PRINT_ATTR_X64(bp_addr);
1002         ret += PRINT_ATTR_X64(config1);
1003         ret += PRINT_ATTR_U64(bp_len);
1004         ret += PRINT_ATTR_X64(config2);
1005         ret += PRINT_ATTR_X64(branch_sample_type);
1006         ret += PRINT_ATTR_X64(sample_regs_user);
1007         ret += PRINT_ATTR_U32(sample_stack_user);
1008
1009         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1010
1011         return ret;
1012 }
1013
1014 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1015                               struct thread_map *threads)
1016 {
1017         int cpu, thread, nthreads;
1018         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1019         int pid = -1, err;
1020         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1021
1022         if (evsel->system_wide)
1023                 nthreads = 1;
1024         else
1025                 nthreads = threads->nr;
1026
1027         if (evsel->fd == NULL &&
1028             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1029                 return -ENOMEM;
1030
1031         if (evsel->cgrp) {
1032                 flags |= PERF_FLAG_PID_CGROUP;
1033                 pid = evsel->cgrp->fd;
1034         }
1035
1036 fallback_missing_features:
1037         if (perf_missing_features.cloexec)
1038                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1039         if (perf_missing_features.mmap2)
1040                 evsel->attr.mmap2 = 0;
1041         if (perf_missing_features.exclude_guest)
1042                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1043 retry_sample_id:
1044         if (perf_missing_features.sample_id_all)
1045                 evsel->attr.sample_id_all = 0;
1046
1047         if (verbose >= 2)
1048                 perf_event_attr__fprintf(&evsel->attr, stderr);
1049
1050         for (cpu = 0; cpu < cpus->nr; cpu++) {
1051
1052                 for (thread = 0; thread < nthreads; thread++) {
1053                         int group_fd;
1054
1055                         if (!evsel->cgrp && !evsel->system_wide)
1056                                 pid = threads->map[thread];
1057
1058                         group_fd = get_group_fd(evsel, cpu, thread);
1059 retry_open:
1060                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1061                                   pid, cpus->map[cpu], group_fd, flags);
1062
1063                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1064                                                                      pid,
1065                                                                      cpus->map[cpu],
1066                                                                      group_fd, flags);
1067                         if (FD(evsel, cpu, thread) < 0) {
1068                                 err = -errno;
1069                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1070                                           err);
1071                                 goto try_fallback;
1072                         }
1073                         set_rlimit = NO_CHANGE;
1074                 }
1075         }
1076
1077         return 0;
1078
1079 try_fallback:
1080         /*
1081          * perf stat needs between 5 and 22 fds per CPU. When we run out
1082          * of them try to increase the limits.
1083          */
1084         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1085                 struct rlimit l;
1086                 int old_errno = errno;
1087
1088                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1089                         if (set_rlimit == NO_CHANGE)
1090                                 l.rlim_cur = l.rlim_max;
1091                         else {
1092                                 l.rlim_cur = l.rlim_max + 1000;
1093                                 l.rlim_max = l.rlim_cur;
1094                         }
1095                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1096                                 set_rlimit++;
1097                                 errno = old_errno;
1098                                 goto retry_open;
1099                         }
1100                 }
1101                 errno = old_errno;
1102         }
1103
1104         if (err != -EINVAL || cpu > 0 || thread > 0)
1105                 goto out_close;
1106
1107         if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1108                 perf_missing_features.cloexec = true;
1109                 goto fallback_missing_features;
1110         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1111                 perf_missing_features.mmap2 = true;
1112                 goto fallback_missing_features;
1113         } else if (!perf_missing_features.exclude_guest &&
1114                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1115                 perf_missing_features.exclude_guest = true;
1116                 goto fallback_missing_features;
1117         } else if (!perf_missing_features.sample_id_all) {
1118                 perf_missing_features.sample_id_all = true;
1119                 goto retry_sample_id;
1120         }
1121
1122 out_close:
1123         do {
1124                 while (--thread >= 0) {
1125                         close(FD(evsel, cpu, thread));
1126                         FD(evsel, cpu, thread) = -1;
1127                 }
1128                 thread = nthreads;
1129         } while (--cpu >= 0);
1130         return err;
1131 }
1132
1133 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1134 {
1135         if (evsel->fd == NULL)
1136                 return;
1137
1138         perf_evsel__close_fd(evsel, ncpus, nthreads);
1139         perf_evsel__free_fd(evsel);
1140 }
1141
1142 static struct {
1143         struct cpu_map map;
1144         int cpus[1];
1145 } empty_cpu_map = {
1146         .map.nr = 1,
1147         .cpus   = { -1, },
1148 };
1149
1150 static struct {
1151         struct thread_map map;
1152         int threads[1];
1153 } empty_thread_map = {
1154         .map.nr  = 1,
1155         .threads = { -1, },
1156 };
1157
1158 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1159                      struct thread_map *threads)
1160 {
1161         if (cpus == NULL) {
1162                 /* Work around old compiler warnings about strict aliasing */
1163                 cpus = &empty_cpu_map.map;
1164         }
1165
1166         if (threads == NULL)
1167                 threads = &empty_thread_map.map;
1168
1169         return __perf_evsel__open(evsel, cpus, threads);
1170 }
1171
1172 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1173                              struct cpu_map *cpus)
1174 {
1175         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1176 }
1177
1178 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1179                                 struct thread_map *threads)
1180 {
1181         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1182 }
1183
1184 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1185                                        const union perf_event *event,
1186                                        struct perf_sample *sample)
1187 {
1188         u64 type = evsel->attr.sample_type;
1189         const u64 *array = event->sample.array;
1190         bool swapped = evsel->needs_swap;
1191         union u64_swap u;
1192
1193         array += ((event->header.size -
1194                    sizeof(event->header)) / sizeof(u64)) - 1;
1195
1196         if (type & PERF_SAMPLE_IDENTIFIER) {
1197                 sample->id = *array;
1198                 array--;
1199         }
1200
1201         if (type & PERF_SAMPLE_CPU) {
1202                 u.val64 = *array;
1203                 if (swapped) {
1204                         /* undo swap of u64, then swap on individual u32s */
1205                         u.val64 = bswap_64(u.val64);
1206                         u.val32[0] = bswap_32(u.val32[0]);
1207                 }
1208
1209                 sample->cpu = u.val32[0];
1210                 array--;
1211         }
1212
1213         if (type & PERF_SAMPLE_STREAM_ID) {
1214                 sample->stream_id = *array;
1215                 array--;
1216         }
1217
1218         if (type & PERF_SAMPLE_ID) {
1219                 sample->id = *array;
1220                 array--;
1221         }
1222
1223         if (type & PERF_SAMPLE_TIME) {
1224                 sample->time = *array;
1225                 array--;
1226         }
1227
1228         if (type & PERF_SAMPLE_TID) {
1229                 u.val64 = *array;
1230                 if (swapped) {
1231                         /* undo swap of u64, then swap on individual u32s */
1232                         u.val64 = bswap_64(u.val64);
1233                         u.val32[0] = bswap_32(u.val32[0]);
1234                         u.val32[1] = bswap_32(u.val32[1]);
1235                 }
1236
1237                 sample->pid = u.val32[0];
1238                 sample->tid = u.val32[1];
1239                 array--;
1240         }
1241
1242         return 0;
1243 }
1244
1245 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1246                             u64 size)
1247 {
1248         return size > max_size || offset + size > endp;
1249 }
1250
1251 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1252         do {                                                            \
1253                 if (overflow(endp, (max_size), (offset), (size)))       \
1254                         return -EFAULT;                                 \
1255         } while (0)
1256
1257 #define OVERFLOW_CHECK_u64(offset) \
1258         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1259
1260 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1261                              struct perf_sample *data)
1262 {
1263         u64 type = evsel->attr.sample_type;
1264         bool swapped = evsel->needs_swap;
1265         const u64 *array;
1266         u16 max_size = event->header.size;
1267         const void *endp = (void *)event + max_size;
1268         u64 sz;
1269
1270         /*
1271          * used for cross-endian analysis. See git commit 65014ab3
1272          * for why this goofiness is needed.
1273          */
1274         union u64_swap u;
1275
1276         memset(data, 0, sizeof(*data));
1277         data->cpu = data->pid = data->tid = -1;
1278         data->stream_id = data->id = data->time = -1ULL;
1279         data->period = evsel->attr.sample_period;
1280         data->weight = 0;
1281
1282         if (event->header.type != PERF_RECORD_SAMPLE) {
1283                 if (!evsel->attr.sample_id_all)
1284                         return 0;
1285                 return perf_evsel__parse_id_sample(evsel, event, data);
1286         }
1287
1288         array = event->sample.array;
1289
1290         /*
1291          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1292          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1293          * check the format does not go past the end of the event.
1294          */
1295         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1296                 return -EFAULT;
1297
1298         data->id = -1ULL;
1299         if (type & PERF_SAMPLE_IDENTIFIER) {
1300                 data->id = *array;
1301                 array++;
1302         }
1303
1304         if (type & PERF_SAMPLE_IP) {
1305                 data->ip = *array;
1306                 array++;
1307         }
1308
1309         if (type & PERF_SAMPLE_TID) {
1310                 u.val64 = *array;
1311                 if (swapped) {
1312                         /* undo swap of u64, then swap on individual u32s */
1313                         u.val64 = bswap_64(u.val64);
1314                         u.val32[0] = bswap_32(u.val32[0]);
1315                         u.val32[1] = bswap_32(u.val32[1]);
1316                 }
1317
1318                 data->pid = u.val32[0];
1319                 data->tid = u.val32[1];
1320                 array++;
1321         }
1322
1323         if (type & PERF_SAMPLE_TIME) {
1324                 data->time = *array;
1325                 array++;
1326         }
1327
1328         data->addr = 0;
1329         if (type & PERF_SAMPLE_ADDR) {
1330                 data->addr = *array;
1331                 array++;
1332         }
1333
1334         if (type & PERF_SAMPLE_ID) {
1335                 data->id = *array;
1336                 array++;
1337         }
1338
1339         if (type & PERF_SAMPLE_STREAM_ID) {
1340                 data->stream_id = *array;
1341                 array++;
1342         }
1343
1344         if (type & PERF_SAMPLE_CPU) {
1345
1346                 u.val64 = *array;
1347                 if (swapped) {
1348                         /* undo swap of u64, then swap on individual u32s */
1349                         u.val64 = bswap_64(u.val64);
1350                         u.val32[0] = bswap_32(u.val32[0]);
1351                 }
1352
1353                 data->cpu = u.val32[0];
1354                 array++;
1355         }
1356
1357         if (type & PERF_SAMPLE_PERIOD) {
1358                 data->period = *array;
1359                 array++;
1360         }
1361
1362         if (type & PERF_SAMPLE_READ) {
1363                 u64 read_format = evsel->attr.read_format;
1364
1365                 OVERFLOW_CHECK_u64(array);
1366                 if (read_format & PERF_FORMAT_GROUP)
1367                         data->read.group.nr = *array;
1368                 else
1369                         data->read.one.value = *array;
1370
1371                 array++;
1372
1373                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1374                         OVERFLOW_CHECK_u64(array);
1375                         data->read.time_enabled = *array;
1376                         array++;
1377                 }
1378
1379                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1380                         OVERFLOW_CHECK_u64(array);
1381                         data->read.time_running = *array;
1382                         array++;
1383                 }
1384
1385                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1386                 if (read_format & PERF_FORMAT_GROUP) {
1387                         const u64 max_group_nr = UINT64_MAX /
1388                                         sizeof(struct sample_read_value);
1389
1390                         if (data->read.group.nr > max_group_nr)
1391                                 return -EFAULT;
1392                         sz = data->read.group.nr *
1393                              sizeof(struct sample_read_value);
1394                         OVERFLOW_CHECK(array, sz, max_size);
1395                         data->read.group.values =
1396                                         (struct sample_read_value *)array;
1397                         array = (void *)array + sz;
1398                 } else {
1399                         OVERFLOW_CHECK_u64(array);
1400                         data->read.one.id = *array;
1401                         array++;
1402                 }
1403         }
1404
1405         if (type & PERF_SAMPLE_CALLCHAIN) {
1406                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1407
1408                 OVERFLOW_CHECK_u64(array);
1409                 data->callchain = (struct ip_callchain *)array++;
1410                 if (data->callchain->nr > max_callchain_nr)
1411                         return -EFAULT;
1412                 sz = data->callchain->nr * sizeof(u64);
1413                 OVERFLOW_CHECK(array, sz, max_size);
1414                 array = (void *)array + sz;
1415         }
1416
1417         if (type & PERF_SAMPLE_RAW) {
1418                 OVERFLOW_CHECK_u64(array);
1419                 u.val64 = *array;
1420                 if (WARN_ONCE(swapped,
1421                               "Endianness of raw data not corrected!\n")) {
1422                         /* undo swap of u64, then swap on individual u32s */
1423                         u.val64 = bswap_64(u.val64);
1424                         u.val32[0] = bswap_32(u.val32[0]);
1425                         u.val32[1] = bswap_32(u.val32[1]);
1426                 }
1427                 data->raw_size = u.val32[0];
1428                 array = (void *)array + sizeof(u32);
1429
1430                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1431                 data->raw_data = (void *)array;
1432                 array = (void *)array + data->raw_size;
1433         }
1434
1435         if (type & PERF_SAMPLE_BRANCH_STACK) {
1436                 const u64 max_branch_nr = UINT64_MAX /
1437                                           sizeof(struct branch_entry);
1438
1439                 OVERFLOW_CHECK_u64(array);
1440                 data->branch_stack = (struct branch_stack *)array++;
1441
1442                 if (data->branch_stack->nr > max_branch_nr)
1443                         return -EFAULT;
1444                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1445                 OVERFLOW_CHECK(array, sz, max_size);
1446                 array = (void *)array + sz;
1447         }
1448
1449         if (type & PERF_SAMPLE_REGS_USER) {
1450                 OVERFLOW_CHECK_u64(array);
1451                 data->user_regs.abi = *array;
1452                 array++;
1453
1454                 if (data->user_regs.abi) {
1455                         u64 mask = evsel->attr.sample_regs_user;
1456
1457                         sz = hweight_long(mask) * sizeof(u64);
1458                         OVERFLOW_CHECK(array, sz, max_size);
1459                         data->user_regs.mask = mask;
1460                         data->user_regs.regs = (u64 *)array;
1461                         array = (void *)array + sz;
1462                 }
1463         }
1464
1465         if (type & PERF_SAMPLE_STACK_USER) {
1466                 OVERFLOW_CHECK_u64(array);
1467                 sz = *array++;
1468
1469                 data->user_stack.offset = ((char *)(array - 1)
1470                                           - (char *) event);
1471
1472                 if (!sz) {
1473                         data->user_stack.size = 0;
1474                 } else {
1475                         OVERFLOW_CHECK(array, sz, max_size);
1476                         data->user_stack.data = (char *)array;
1477                         array = (void *)array + sz;
1478                         OVERFLOW_CHECK_u64(array);
1479                         data->user_stack.size = *array++;
1480                         if (WARN_ONCE(data->user_stack.size > sz,
1481                                       "user stack dump failure\n"))
1482                                 return -EFAULT;
1483                 }
1484         }
1485
1486         data->weight = 0;
1487         if (type & PERF_SAMPLE_WEIGHT) {
1488                 OVERFLOW_CHECK_u64(array);
1489                 data->weight = *array;
1490                 array++;
1491         }
1492
1493         data->data_src = PERF_MEM_DATA_SRC_NONE;
1494         if (type & PERF_SAMPLE_DATA_SRC) {
1495                 OVERFLOW_CHECK_u64(array);
1496                 data->data_src = *array;
1497                 array++;
1498         }
1499
1500         data->transaction = 0;
1501         if (type & PERF_SAMPLE_TRANSACTION) {
1502                 OVERFLOW_CHECK_u64(array);
1503                 data->transaction = *array;
1504                 array++;
1505         }
1506
1507         return 0;
1508 }
1509
1510 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1511                                      u64 read_format)
1512 {
1513         size_t sz, result = sizeof(struct sample_event);
1514
1515         if (type & PERF_SAMPLE_IDENTIFIER)
1516                 result += sizeof(u64);
1517
1518         if (type & PERF_SAMPLE_IP)
1519                 result += sizeof(u64);
1520
1521         if (type & PERF_SAMPLE_TID)
1522                 result += sizeof(u64);
1523
1524         if (type & PERF_SAMPLE_TIME)
1525                 result += sizeof(u64);
1526
1527         if (type & PERF_SAMPLE_ADDR)
1528                 result += sizeof(u64);
1529
1530         if (type & PERF_SAMPLE_ID)
1531                 result += sizeof(u64);
1532
1533         if (type & PERF_SAMPLE_STREAM_ID)
1534                 result += sizeof(u64);
1535
1536         if (type & PERF_SAMPLE_CPU)
1537                 result += sizeof(u64);
1538
1539         if (type & PERF_SAMPLE_PERIOD)
1540                 result += sizeof(u64);
1541
1542         if (type & PERF_SAMPLE_READ) {
1543                 result += sizeof(u64);
1544                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1545                         result += sizeof(u64);
1546                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1547                         result += sizeof(u64);
1548                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1549                 if (read_format & PERF_FORMAT_GROUP) {
1550                         sz = sample->read.group.nr *
1551                              sizeof(struct sample_read_value);
1552                         result += sz;
1553                 } else {
1554                         result += sizeof(u64);
1555                 }
1556         }
1557
1558         if (type & PERF_SAMPLE_CALLCHAIN) {
1559                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1560                 result += sz;
1561         }
1562
1563         if (type & PERF_SAMPLE_RAW) {
1564                 result += sizeof(u32);
1565                 result += sample->raw_size;
1566         }
1567
1568         if (type & PERF_SAMPLE_BRANCH_STACK) {
1569                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1570                 sz += sizeof(u64);
1571                 result += sz;
1572         }
1573
1574         if (type & PERF_SAMPLE_REGS_USER) {
1575                 if (sample->user_regs.abi) {
1576                         result += sizeof(u64);
1577                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1578                         result += sz;
1579                 } else {
1580                         result += sizeof(u64);
1581                 }
1582         }
1583
1584         if (type & PERF_SAMPLE_STACK_USER) {
1585                 sz = sample->user_stack.size;
1586                 result += sizeof(u64);
1587                 if (sz) {
1588                         result += sz;
1589                         result += sizeof(u64);
1590                 }
1591         }
1592
1593         if (type & PERF_SAMPLE_WEIGHT)
1594                 result += sizeof(u64);
1595
1596         if (type & PERF_SAMPLE_DATA_SRC)
1597                 result += sizeof(u64);
1598
1599         if (type & PERF_SAMPLE_TRANSACTION)
1600                 result += sizeof(u64);
1601
1602         return result;
1603 }
1604
1605 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1606                                   u64 read_format,
1607                                   const struct perf_sample *sample,
1608                                   bool swapped)
1609 {
1610         u64 *array;
1611         size_t sz;
1612         /*
1613          * used for cross-endian analysis. See git commit 65014ab3
1614          * for why this goofiness is needed.
1615          */
1616         union u64_swap u;
1617
1618         array = event->sample.array;
1619
1620         if (type & PERF_SAMPLE_IDENTIFIER) {
1621                 *array = sample->id;
1622                 array++;
1623         }
1624
1625         if (type & PERF_SAMPLE_IP) {
1626                 *array = sample->ip;
1627                 array++;
1628         }
1629
1630         if (type & PERF_SAMPLE_TID) {
1631                 u.val32[0] = sample->pid;
1632                 u.val32[1] = sample->tid;
1633                 if (swapped) {
1634                         /*
1635                          * Inverse of what is done in perf_evsel__parse_sample
1636                          */
1637                         u.val32[0] = bswap_32(u.val32[0]);
1638                         u.val32[1] = bswap_32(u.val32[1]);
1639                         u.val64 = bswap_64(u.val64);
1640                 }
1641
1642                 *array = u.val64;
1643                 array++;
1644         }
1645
1646         if (type & PERF_SAMPLE_TIME) {
1647                 *array = sample->time;
1648                 array++;
1649         }
1650
1651         if (type & PERF_SAMPLE_ADDR) {
1652                 *array = sample->addr;
1653                 array++;
1654         }
1655
1656         if (type & PERF_SAMPLE_ID) {
1657                 *array = sample->id;
1658                 array++;
1659         }
1660
1661         if (type & PERF_SAMPLE_STREAM_ID) {
1662                 *array = sample->stream_id;
1663                 array++;
1664         }
1665
1666         if (type & PERF_SAMPLE_CPU) {
1667                 u.val32[0] = sample->cpu;
1668                 if (swapped) {
1669                         /*
1670                          * Inverse of what is done in perf_evsel__parse_sample
1671                          */
1672                         u.val32[0] = bswap_32(u.val32[0]);
1673                         u.val64 = bswap_64(u.val64);
1674                 }
1675                 *array = u.val64;
1676                 array++;
1677         }
1678
1679         if (type & PERF_SAMPLE_PERIOD) {
1680                 *array = sample->period;
1681                 array++;
1682         }
1683
1684         if (type & PERF_SAMPLE_READ) {
1685                 if (read_format & PERF_FORMAT_GROUP)
1686                         *array = sample->read.group.nr;
1687                 else
1688                         *array = sample->read.one.value;
1689                 array++;
1690
1691                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1692                         *array = sample->read.time_enabled;
1693                         array++;
1694                 }
1695
1696                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1697                         *array = sample->read.time_running;
1698                         array++;
1699                 }
1700
1701                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1702                 if (read_format & PERF_FORMAT_GROUP) {
1703                         sz = sample->read.group.nr *
1704                              sizeof(struct sample_read_value);
1705                         memcpy(array, sample->read.group.values, sz);
1706                         array = (void *)array + sz;
1707                 } else {
1708                         *array = sample->read.one.id;
1709                         array++;
1710                 }
1711         }
1712
1713         if (type & PERF_SAMPLE_CALLCHAIN) {
1714                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1715                 memcpy(array, sample->callchain, sz);
1716                 array = (void *)array + sz;
1717         }
1718
1719         if (type & PERF_SAMPLE_RAW) {
1720                 u.val32[0] = sample->raw_size;
1721                 if (WARN_ONCE(swapped,
1722                               "Endianness of raw data not corrected!\n")) {
1723                         /*
1724                          * Inverse of what is done in perf_evsel__parse_sample
1725                          */
1726                         u.val32[0] = bswap_32(u.val32[0]);
1727                         u.val32[1] = bswap_32(u.val32[1]);
1728                         u.val64 = bswap_64(u.val64);
1729                 }
1730                 *array = u.val64;
1731                 array = (void *)array + sizeof(u32);
1732
1733                 memcpy(array, sample->raw_data, sample->raw_size);
1734                 array = (void *)array + sample->raw_size;
1735         }
1736
1737         if (type & PERF_SAMPLE_BRANCH_STACK) {
1738                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1739                 sz += sizeof(u64);
1740                 memcpy(array, sample->branch_stack, sz);
1741                 array = (void *)array + sz;
1742         }
1743
1744         if (type & PERF_SAMPLE_REGS_USER) {
1745                 if (sample->user_regs.abi) {
1746                         *array++ = sample->user_regs.abi;
1747                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1748                         memcpy(array, sample->user_regs.regs, sz);
1749                         array = (void *)array + sz;
1750                 } else {
1751                         *array++ = 0;
1752                 }
1753         }
1754
1755         if (type & PERF_SAMPLE_STACK_USER) {
1756                 sz = sample->user_stack.size;
1757                 *array++ = sz;
1758                 if (sz) {
1759                         memcpy(array, sample->user_stack.data, sz);
1760                         array = (void *)array + sz;
1761                         *array++ = sz;
1762                 }
1763         }
1764
1765         if (type & PERF_SAMPLE_WEIGHT) {
1766                 *array = sample->weight;
1767                 array++;
1768         }
1769
1770         if (type & PERF_SAMPLE_DATA_SRC) {
1771                 *array = sample->data_src;
1772                 array++;
1773         }
1774
1775         if (type & PERF_SAMPLE_TRANSACTION) {
1776                 *array = sample->transaction;
1777                 array++;
1778         }
1779
1780         return 0;
1781 }
1782
1783 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1784 {
1785         return pevent_find_field(evsel->tp_format, name);
1786 }
1787
1788 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1789                          const char *name)
1790 {
1791         struct format_field *field = perf_evsel__field(evsel, name);
1792         int offset;
1793
1794         if (!field)
1795                 return NULL;
1796
1797         offset = field->offset;
1798
1799         if (field->flags & FIELD_IS_DYNAMIC) {
1800                 offset = *(int *)(sample->raw_data + field->offset);
1801                 offset &= 0xffff;
1802         }
1803
1804         return sample->raw_data + offset;
1805 }
1806
1807 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1808                        const char *name)
1809 {
1810         struct format_field *field = perf_evsel__field(evsel, name);
1811         void *ptr;
1812         u64 value;
1813
1814         if (!field)
1815                 return 0;
1816
1817         ptr = sample->raw_data + field->offset;
1818
1819         switch (field->size) {
1820         case 1:
1821                 return *(u8 *)ptr;
1822         case 2:
1823                 value = *(u16 *)ptr;
1824                 break;
1825         case 4:
1826                 value = *(u32 *)ptr;
1827                 break;
1828         case 8:
1829                 value = *(u64 *)ptr;
1830                 break;
1831         default:
1832                 return 0;
1833         }
1834
1835         if (!evsel->needs_swap)
1836                 return value;
1837
1838         switch (field->size) {
1839         case 2:
1840                 return bswap_16(value);
1841         case 4:
1842                 return bswap_32(value);
1843         case 8:
1844                 return bswap_64(value);
1845         default:
1846                 return 0;
1847         }
1848
1849         return 0;
1850 }
1851
1852 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1853 {
1854         va_list args;
1855         int ret = 0;
1856
1857         if (!*first) {
1858                 ret += fprintf(fp, ",");
1859         } else {
1860                 ret += fprintf(fp, ":");
1861                 *first = false;
1862         }
1863
1864         va_start(args, fmt);
1865         ret += vfprintf(fp, fmt, args);
1866         va_end(args);
1867         return ret;
1868 }
1869
1870 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1871 {
1872         if (value == 0)
1873                 return 0;
1874
1875         return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1876 }
1877
1878 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1879
1880 struct bit_names {
1881         int bit;
1882         const char *name;
1883 };
1884
1885 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1886                          struct bit_names *bits, bool *first)
1887 {
1888         int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1889         bool first_bit = true;
1890
1891         do {
1892                 if (value & bits[i].bit) {
1893                         printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1894                         first_bit = false;
1895                 }
1896         } while (bits[++i].name != NULL);
1897
1898         return printed;
1899 }
1900
1901 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1902 {
1903 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1904         struct bit_names bits[] = {
1905                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1906                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1907                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1908                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1909                 bit_name(IDENTIFIER),
1910                 { .name = NULL, }
1911         };
1912 #undef bit_name
1913         return bits__fprintf(fp, "sample_type", value, bits, first);
1914 }
1915
1916 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1917 {
1918 #define bit_name(n) { PERF_FORMAT_##n, #n }
1919         struct bit_names bits[] = {
1920                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1921                 bit_name(ID), bit_name(GROUP),
1922                 { .name = NULL, }
1923         };
1924 #undef bit_name
1925         return bits__fprintf(fp, "read_format", value, bits, first);
1926 }
1927
1928 int perf_evsel__fprintf(struct perf_evsel *evsel,
1929                         struct perf_attr_details *details, FILE *fp)
1930 {
1931         bool first = true;
1932         int printed = 0;
1933
1934         if (details->event_group) {
1935                 struct perf_evsel *pos;
1936
1937                 if (!perf_evsel__is_group_leader(evsel))
1938                         return 0;
1939
1940                 if (evsel->nr_members > 1)
1941                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1942
1943                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1944                 for_each_group_member(pos, evsel)
1945                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1946
1947                 if (evsel->nr_members > 1)
1948                         printed += fprintf(fp, "}");
1949                 goto out;
1950         }
1951
1952         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1953
1954         if (details->verbose || details->freq) {
1955                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1956                                          (u64)evsel->attr.sample_freq);
1957         }
1958
1959         if (details->verbose) {
1960                 if_print(type);
1961                 if_print(config);
1962                 if_print(config1);
1963                 if_print(config2);
1964                 if_print(size);
1965                 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1966                 if (evsel->attr.read_format)
1967                         printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1968                 if_print(disabled);
1969                 if_print(inherit);
1970                 if_print(pinned);
1971                 if_print(exclusive);
1972                 if_print(exclude_user);
1973                 if_print(exclude_kernel);
1974                 if_print(exclude_hv);
1975                 if_print(exclude_idle);
1976                 if_print(mmap);
1977                 if_print(mmap2);
1978                 if_print(comm);
1979                 if_print(comm_exec);
1980                 if_print(freq);
1981                 if_print(inherit_stat);
1982                 if_print(enable_on_exec);
1983                 if_print(task);
1984                 if_print(watermark);
1985                 if_print(precise_ip);
1986                 if_print(mmap_data);
1987                 if_print(sample_id_all);
1988                 if_print(exclude_host);
1989                 if_print(exclude_guest);
1990                 if_print(__reserved_1);
1991                 if_print(wakeup_events);
1992                 if_print(bp_type);
1993                 if_print(branch_sample_type);
1994         }
1995 out:
1996         fputc('\n', fp);
1997         return ++printed;
1998 }
1999
2000 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2001                           char *msg, size_t msgsize)
2002 {
2003         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2004             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2005             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2006                 /*
2007                  * If it's cycles then fall back to hrtimer based
2008                  * cpu-clock-tick sw counter, which is always available even if
2009                  * no PMU support.
2010                  *
2011                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2012                  * b0a873e).
2013                  */
2014                 scnprintf(msg, msgsize, "%s",
2015 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2016
2017                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2018                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2019
2020                 zfree(&evsel->name);
2021                 return true;
2022         }
2023
2024         return false;
2025 }
2026
2027 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2028                               int err, char *msg, size_t size)
2029 {
2030         char sbuf[STRERR_BUFSIZE];
2031
2032         switch (err) {
2033         case EPERM:
2034         case EACCES:
2035                 return scnprintf(msg, size,
2036                  "You may not have permission to collect %sstats.\n"
2037                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2038                  " -1 - Not paranoid at all\n"
2039                  "  0 - Disallow raw tracepoint access for unpriv\n"
2040                  "  1 - Disallow cpu events for unpriv\n"
2041                  "  2 - Disallow kernel profiling for unpriv",
2042                                  target->system_wide ? "system-wide " : "");
2043         case ENOENT:
2044                 return scnprintf(msg, size, "The %s event is not supported.",
2045                                  perf_evsel__name(evsel));
2046         case EMFILE:
2047                 return scnprintf(msg, size, "%s",
2048                          "Too many events are opened.\n"
2049                          "Try again after reducing the number of events.");
2050         case ENODEV:
2051                 if (target->cpu_list)
2052                         return scnprintf(msg, size, "%s",
2053          "No such device - did you specify an out-of-range profile CPU?\n");
2054                 break;
2055         case EOPNOTSUPP:
2056                 if (evsel->attr.precise_ip)
2057                         return scnprintf(msg, size, "%s",
2058         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2059 #if defined(__i386__) || defined(__x86_64__)
2060                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2061                         return scnprintf(msg, size, "%s",
2062         "No hardware sampling interrupt available.\n"
2063         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2064 #endif
2065                 break;
2066         case EBUSY:
2067                 if (find_process("oprofiled"))
2068                         return scnprintf(msg, size,
2069         "The PMU counters are busy/taken by another profiler.\n"
2070         "We found oprofile daemon running, please stop it and try again.");
2071                 break;
2072         default:
2073                 break;
2074         }
2075
2076         return scnprintf(msg, size,
2077         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2078         "/bin/dmesg may provide additional information.\n"
2079         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2080                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2081                          perf_evsel__name(evsel));
2082 }