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
  30 static struct {
  31         bool sample_id_all;
  32         bool exclude_guest;
  33         bool mmap2;
  34         bool cloexec;
  35 } perf_missing_features;
  36
  37 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  38 {
  39         return 0;
  40 }
  41
  42 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  43 {
  44 }
  45
  46 static struct {
  47         size_t  size;
  48         int     (*init)(struct perf_evsel *evsel);
  49         void    (*fini)(struct perf_evsel *evsel);
  50 } perf_evsel__object = {
  51         .size = sizeof(struct perf_evsel),
  52         .init = perf_evsel__no_extra_init,
  53         .fini = perf_evsel__no_extra_fini,
  54 };
  55
  56 int perf_evsel__object_config(size_t object_size,
  57                               int (*init)(struct perf_evsel *evsel),
  58                               void (*fini)(struct perf_evsel *evsel))
  59 {
  60
  61         if (object_size == 0)
  62                 goto set_methods;
  63
  64         if (perf_evsel__object.size > object_size)
  65                 return -EINVAL;
  66
  67         perf_evsel__object.size = object_size;
  68
  69 set_methods:
  70         if (init != NULL)
  71                 perf_evsel__object.init = init;
  72
  73         if (fini != NULL)
  74                 perf_evsel__object.fini = fini;
  75
  76         return 0;
  77 }
  78
  79 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  80
  81 int __perf_evsel__sample_size(u64 sample_type)
  82 {
  83         u64 mask = sample_type & PERF_SAMPLE_MASK;
  84         int size = 0;
  85         int i;
  86
  87         for (i = 0; i < 64; i++) {
  88                 if (mask & (1ULL << i))
  89                         size++;
  90         }
  91
  92         size *= sizeof(u64);
  93
  94         return size;
  95 }
  96
  97 /**
  98  * __perf_evsel__calc_id_pos - calculate id_pos.
  99  * @sample_type: sample type
 100  *
 101  * This function returns the position of the event id (PERF_SAMPLE_ID or
 102  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 103  * sample_event.
 104  */
 105 static int __perf_evsel__calc_id_pos(u64 sample_type)
 106 {
 107         int idx = 0;
 108
 109         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 110                 return 0;
 111
 112         if (!(sample_type & PERF_SAMPLE_ID))
 113                 return -1;
 114
 115         if (sample_type & PERF_SAMPLE_IP)
 116                 idx += 1;
 117
 118         if (sample_type & PERF_SAMPLE_TID)
 119                 idx += 1;
 120
 121         if (sample_type & PERF_SAMPLE_TIME)
 122                 idx += 1;
 123
 124         if (sample_type & PERF_SAMPLE_ADDR)
 125                 idx += 1;
 126
 127         return idx;
 128 }
 129
 130 /**
 131  * __perf_evsel__calc_is_pos - calculate is_pos.
 132  * @sample_type: sample type
 133  *
 134  * This function returns the position (counting backwards) of the event id
 135  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 136  * sample_id_all is used there is an id sample appended to non-sample events.
 137  */
 138 static int __perf_evsel__calc_is_pos(u64 sample_type)
 139 {
 140         int idx = 1;
 141
 142         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 143                 return 1;
 144
 145         if (!(sample_type & PERF_SAMPLE_ID))
 146                 return -1;
 147
 148         if (sample_type & PERF_SAMPLE_CPU)
 149                 idx += 1;
 150
 151         if (sample_type & PERF_SAMPLE_STREAM_ID)
 152                 idx += 1;
 153
 154         return idx;
 155 }
 156
 157 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 158 {
 159         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 160         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 161 }
 162
 163 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 164                                   enum perf_event_sample_format bit)
 165 {
 166         if (!(evsel->attr.sample_type & bit)) {
 167                 evsel->attr.sample_type |= bit;
 168                 evsel->sample_size += sizeof(u64);
 169                 perf_evsel__calc_id_pos(evsel);
 170         }
 171 }
 172
 173 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 174                                     enum perf_event_sample_format bit)
 175 {
 176         if (evsel->attr.sample_type & bit) {
 177                 evsel->attr.sample_type &= ~bit;
 178                 evsel->sample_size -= sizeof(u64);
 179                 perf_evsel__calc_id_pos(evsel);
 180         }
 181 }
 182
 183 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 184                                bool can_sample_identifier)
 185 {
 186         if (can_sample_identifier) {
 187                 perf_evsel__reset_sample_bit(evsel, ID);
 188                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 189         } else {
 190                 perf_evsel__set_sample_bit(evsel, ID);
 191         }
 192         evsel->attr.read_format |= PERF_FORMAT_ID;
 193 }
 194
 195 void perf_evsel__init(struct perf_evsel *evsel,
 196                       struct perf_event_attr *attr, int idx)
 197 {
 198         evsel->idx         = idx;
 199         evsel->tracking    = !idx;
 200         evsel->attr        = *attr;
 201         evsel->leader      = evsel;
 202         evsel->unit        = "";
 203         evsel->scale       = 1.0;
 204         INIT_LIST_HEAD(&evsel->node);
 205         perf_evsel__object.init(evsel);
 206         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 207         perf_evsel__calc_id_pos(evsel);
 208 }
 209
 210 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 211 {
 212         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 213
 214         if (evsel != NULL)
 215                 perf_evsel__init(evsel, attr, idx);
 216
 217         return evsel;
 218 }
 219
 220 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 221 {
 222         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 223
 224         if (evsel != NULL) {
 225                 struct perf_event_attr attr = {
 226                         .type          = PERF_TYPE_TRACEPOINT,
 227                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 228                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 229                 };
 230
 231                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 232                         goto out_free;
 233
 234                 evsel->tp_format = trace_event__tp_format(sys, name);
 235                 if (evsel->tp_format == NULL)
 236                         goto out_free;
 237
 238                 event_attr_init(&attr);
 239                 attr.config = evsel->tp_format->id;
 240                 attr.sample_period = 1;
 241                 perf_evsel__init(evsel, &attr, idx);
 242         }
 243
 244         return evsel;
 245
 246 out_free:
 247         zfree(&evsel->name);
 248         free(evsel);
 249         return NULL;
 250 }
 251
 252 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 253         "cycles",
 254         "instructions",
 255         "cache-references",
 256         "cache-misses",
 257         "branches",
 258         "branch-misses",
 259         "bus-cycles",
 260         "stalled-cycles-frontend",
 261         "stalled-cycles-backend",
 262         "ref-cycles",
 263 };
 264
 265 static const char *__perf_evsel__hw_name(u64 config)
 266 {
 267         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 268                 return perf_evsel__hw_names[config];
 269
 270         return "unknown-hardware";
 271 }
 272
 273 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 274 {
 275         int colon = 0, r = 0;
 276         struct perf_event_attr *attr = &evsel->attr;
 277         bool exclude_guest_default = false;
 278
 279 #define MOD_PRINT(context, mod) do {                                    \
 280                 if (!attr->exclude_##context) {                         \
 281                         if (!colon) colon = ++r;                        \
 282                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 283                 } } while(0)
 284
 285         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 286                 MOD_PRINT(kernel, 'k');
 287                 MOD_PRINT(user, 'u');
 288                 MOD_PRINT(hv, 'h');
 289                 exclude_guest_default = true;
 290         }
 291
 292         if (attr->precise_ip) {
 293                 if (!colon)
 294                         colon = ++r;
 295                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 296                 exclude_guest_default = true;
 297         }
 298
 299         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 300                 MOD_PRINT(host, 'H');
 301                 MOD_PRINT(guest, 'G');
 302         }
 303 #undef MOD_PRINT
 304         if (colon)
 305                 bf[colon - 1] = ':';
 306         return r;
 307 }
 308
 309 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 310 {
 311         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 312         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 313 }
 314
 315 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 316         "cpu-clock",
 317         "task-clock",
 318         "page-faults",
 319         "context-switches",
 320         "cpu-migrations",
 321         "minor-faults",
 322         "major-faults",
 323         "alignment-faults",
 324         "emulation-faults",
 325         "dummy",
 326 };
 327
 328 static const char *__perf_evsel__sw_name(u64 config)
 329 {
 330         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 331                 return perf_evsel__sw_names[config];
 332         return "unknown-software";
 333 }
 334
 335 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 336 {
 337         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 338         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 339 }
 340
 341 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 342 {
 343         int r;
 344
 345         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 346
 347         if (type & HW_BREAKPOINT_R)
 348                 r += scnprintf(bf + r, size - r, "r");
 349
 350         if (type & HW_BREAKPOINT_W)
 351                 r += scnprintf(bf + r, size - r, "w");
 352
 353         if (type & HW_BREAKPOINT_X)
 354                 r += scnprintf(bf + r, size - r, "x");
 355
 356         return r;
 357 }
 358
 359 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 360 {
 361         struct perf_event_attr *attr = &evsel->attr;
 362         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 363         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 364 }
 365
 366 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 367                                 [PERF_EVSEL__MAX_ALIASES] = {
 368  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 369  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 370  { "LLC",       "L2",                                                   },
 371  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 372  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 373  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 374  { "node",                                                              },
 375 };
 376
 377 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 378                                    [PERF_EVSEL__MAX_ALIASES] = {
 379  { "load",      "loads",        "read",                                 },
 380  { "store",     "stores",       "write",                                },
 381  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 382 };
 383
 384 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 385                                        [PERF_EVSEL__MAX_ALIASES] = {
 386  { "refs",      "Reference",    "ops",          "access",               },
 387  { "misses",    "miss",                                                 },
 388 };
 389
 390 #define C(x)            PERF_COUNT_HW_CACHE_##x
 391 #define CACHE_READ      (1 << C(OP_READ))
 392 #define CACHE_WRITE     (1 << C(OP_WRITE))
 393 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 394 #define COP(x)          (1 << x)
 395
 396 /*
 397  * cache operartion stat
 398  * L1I : Read and prefetch only
 399  * ITLB and BPU : Read-only
 400  */
 401 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 402  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 403  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 404  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 405  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 406  [C(ITLB)]      = (CACHE_READ),
 407  [C(BPU)]       = (CACHE_READ),
 408  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 409 };
 410
 411 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 412 {
 413         if (perf_evsel__hw_cache_stat[type] & COP(op))
 414                 return true;    /* valid */
 415         else
 416                 return false;   /* invalid */
 417 }
 418
 419 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 420                                             char *bf, size_t size)
 421 {
 422         if (result) {
 423                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 424                                  perf_evsel__hw_cache_op[op][0],
 425                                  perf_evsel__hw_cache_result[result][0]);
 426         }
 427
 428         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 429                          perf_evsel__hw_cache_op[op][1]);
 430 }
 431
 432 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 433 {
 434         u8 op, result, type = (config >>  0) & 0xff;
 435         const char *err = "unknown-ext-hardware-cache-type";
 436
 437         if (type > PERF_COUNT_HW_CACHE_MAX)
 438                 goto out_err;
 439
 440         op = (config >>  8) & 0xff;
 441         err = "unknown-ext-hardware-cache-op";
 442         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 443                 goto out_err;
 444
 445         result = (config >> 16) & 0xff;
 446         err = "unknown-ext-hardware-cache-result";
 447         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 448                 goto out_err;
 449
 450         err = "invalid-cache";
 451         if (!perf_evsel__is_cache_op_valid(type, op))
 452                 goto out_err;
 453
 454         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 455 out_err:
 456         return scnprintf(bf, size, "%s", err);
 457 }
 458
 459 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 460 {
 461         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 462         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 463 }
 464
 465 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 466 {
 467         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 468         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 469 }
 470
 471 const char *perf_evsel__name(struct perf_evsel *evsel)
 472 {
 473         char bf[128];
 474
 475         if (evsel->name)
 476                 return evsel->name;
 477
 478         switch (evsel->attr.type) {
 479         case PERF_TYPE_RAW:
 480                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 481                 break;
 482
 483         case PERF_TYPE_HARDWARE:
 484                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 485                 break;
 486
 487         case PERF_TYPE_HW_CACHE:
 488                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 489                 break;
 490
 491         case PERF_TYPE_SOFTWARE:
 492                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 493                 break;
 494
 495         case PERF_TYPE_TRACEPOINT:
 496                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 497                 break;
 498
 499         case PERF_TYPE_BREAKPOINT:
 500                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 501                 break;
 502
 503         default:
 504                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 505                           evsel->attr.type);
 506                 break;
 507         }
 508
 509         evsel->name = strdup(bf);
 510
 511         return evsel->name ?: "unknown";
 512 }
 513
 514 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 515 {
 516         return evsel->group_name ?: "anon group";
 517 }
 518
 519 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 520 {
 521         int ret;
 522         struct perf_evsel *pos;
 523         const char *group_name = perf_evsel__group_name(evsel);
 524
 525         ret = scnprintf(buf, size, "%s", group_name);
 526
 527         ret += scnprintf(buf + ret, size - ret, " { %s",
 528                          perf_evsel__name(evsel));
 529
 530         for_each_group_member(pos, evsel)
 531                 ret += scnprintf(buf + ret, size - ret, ", %s",
 532                                  perf_evsel__name(pos));
 533
 534         ret += scnprintf(buf + ret, size - ret, " }");
 535
 536         return ret;
 537 }
 538
 539 static void
 540 perf_evsel__config_callgraph(struct perf_evsel *evsel)
 541 {
 542         bool function = perf_evsel__is_function_event(evsel);
 543         struct perf_event_attr *attr = &evsel->attr;
 544
 545         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 546
 547         if (callchain_param.record_mode == CALLCHAIN_DWARF) {
 548                 if (!function) {
 549                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 550                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 551                         attr->sample_regs_user = PERF_REGS_MASK;
 552                         attr->sample_stack_user = callchain_param.dump_size;
 553                         attr->exclude_callchain_user = 1;
 554                 } else {
 555                         pr_info("Cannot use DWARF unwind for function trace event,"
 556                                 " falling back to framepointers.\n");
 557                 }
 558         }
 559
 560         if (function) {
 561                 pr_info("Disabling user space callchains for function trace event.\n");
 562                 attr->exclude_callchain_user = 1;
 563         }
 564 }
 565
 566 /*
 567  * The enable_on_exec/disabled value strategy:
 568  *
 569  *  1) For any type of traced program:
 570  *    - all independent events and group leaders are disabled
 571  *    - all group members are enabled
 572  *
 573  *     Group members are ruled by group leaders. They need to
 574  *     be enabled, because the group scheduling relies on that.
 575  *
 576  *  2) For traced programs executed by perf:
 577  *     - all independent events and group leaders have
 578  *       enable_on_exec set
 579  *     - we don't specifically enable or disable any event during
 580  *       the record command
 581  *
 582  *     Independent events and group leaders are initially disabled
 583  *     and get enabled by exec. Group members are ruled by group
 584  *     leaders as stated in 1).
 585  *
 586  *  3) For traced programs attached by perf (pid/tid):
 587  *     - we specifically enable or disable all events during
 588  *       the record command
 589  *
 590  *     When attaching events to already running traced we
 591  *     enable/disable events specifically, as there's no
 592  *     initial traced exec call.
 593  */
 594 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 595 {
 596         struct perf_evsel *leader = evsel->leader;
 597         struct perf_event_attr *attr = &evsel->attr;
 598         int track = evsel->tracking;
 599         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 600
 601         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 602         attr->inherit       = !opts->no_inherit;
 603
 604         perf_evsel__set_sample_bit(evsel, IP);
 605         perf_evsel__set_sample_bit(evsel, TID);
 606
 607         if (evsel->sample_read) {
 608                 perf_evsel__set_sample_bit(evsel, READ);
 609
 610                 /*
 611                  * We need ID even in case of single event, because
 612                  * PERF_SAMPLE_READ process ID specific data.
 613                  */
 614                 perf_evsel__set_sample_id(evsel, false);
 615
 616                 /*
 617                  * Apply group format only if we belong to group
 618                  * with more than one members.
 619                  */
 620                 if (leader->nr_members > 1) {
 621                         attr->read_format |= PERF_FORMAT_GROUP;
 622                         attr->inherit = 0;
 623                 }
 624         }
 625
 626         /*
 627          * We default some events to have a default interval. But keep
 628          * it a weak assumption overridable by the user.
 629          */
 630         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 631                                      opts->user_interval != ULLONG_MAX)) {
 632                 if (opts->freq) {
 633                         perf_evsel__set_sample_bit(evsel, PERIOD);
 634                         attr->freq              = 1;
 635                         attr->sample_freq       = opts->freq;
 636                 } else {
 637                         attr->sample_period = opts->default_interval;
 638                 }
 639         }
 640
 641         /*
 642          * Disable sampling for all group members other
 643          * than leader in case leader 'leads' the sampling.
 644          */
 645         if ((leader != evsel) && leader->sample_read) {
 646                 attr->sample_freq   = 0;
 647                 attr->sample_period = 0;
 648         }
 649
 650         if (opts->no_samples)
 651                 attr->sample_freq = 0;
 652
 653         if (opts->inherit_stat)
 654                 attr->inherit_stat = 1;
 655
 656         if (opts->sample_address) {
 657                 perf_evsel__set_sample_bit(evsel, ADDR);
 658                 attr->mmap_data = track;
 659         }
 660
 661         /*
 662          * We don't allow user space callchains for  function trace
 663          * event, due to issues with page faults while tracing page
 664          * fault handler and its overall trickiness nature.
 665          */
 666         if (perf_evsel__is_function_event(evsel))
 667                 evsel->attr.exclude_callchain_user = 1;
 668
 669         if (callchain_param.enabled && !evsel->no_aux_samples)
 670                 perf_evsel__config_callgraph(evsel);
 671
 672         if (opts->sample_intr_regs) {
 673                 attr->sample_regs_intr = PERF_REGS_MASK;
 674                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 675         }
 676
 677         if (target__has_cpu(&opts->target))
 678                 perf_evsel__set_sample_bit(evsel, CPU);
 679
 680         if (opts->period)
 681                 perf_evsel__set_sample_bit(evsel, PERIOD);
 682
 683         /*
 684          * When the user explicitely disabled time don't force it here.
 685          */
 686         if (opts->sample_time &&
 687             (!perf_missing_features.sample_id_all &&
 688             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu)))
 689                 perf_evsel__set_sample_bit(evsel, TIME);
 690
 691         if (opts->raw_samples && !evsel->no_aux_samples) {
 692                 perf_evsel__set_sample_bit(evsel, TIME);
 693                 perf_evsel__set_sample_bit(evsel, RAW);
 694                 perf_evsel__set_sample_bit(evsel, CPU);
 695         }
 696
 697         if (opts->sample_address)
 698                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 699
 700         if (opts->no_buffering) {
 701                 attr->watermark = 0;
 702                 attr->wakeup_events = 1;
 703         }
 704         if (opts->branch_stack && !evsel->no_aux_samples) {
 705                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 706                 attr->branch_sample_type = opts->branch_stack;
 707         }
 708
 709         if (opts->sample_weight)
 710                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 711
 712         attr->mmap  = track;
 713         attr->mmap2 = track && !perf_missing_features.mmap2;
 714         attr->comm  = track;
 715
 716         if (opts->sample_transaction)
 717                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 718
 719         /*
 720          * XXX see the function comment above
 721          *
 722          * Disabling only independent events or group leaders,
 723          * keeping group members enabled.
 724          */
 725         if (perf_evsel__is_group_leader(evsel))
 726                 attr->disabled = 1;
 727
 728         /*
 729          * Setting enable_on_exec for independent events and
 730          * group leaders for traced executed by perf.
 731          */
 732         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 733                 !opts->initial_delay)
 734                 attr->enable_on_exec = 1;
 735
 736         if (evsel->immediate) {
 737                 attr->disabled = 0;
 738                 attr->enable_on_exec = 0;
 739         }
 740 }
 741
 742 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 743 {
 744         int cpu, thread;
 745
 746         if (evsel->system_wide)
 747                 nthreads = 1;
 748
 749         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 750
 751         if (evsel->fd) {
 752                 for (cpu = 0; cpu < ncpus; cpu++) {
 753                         for (thread = 0; thread < nthreads; thread++) {
 754                                 FD(evsel, cpu, thread) = -1;
 755                         }
 756                 }
 757         }
 758
 759         return evsel->fd != NULL ? 0 : -ENOMEM;
 760 }
 761
 762 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 763                           int ioc,  void *arg)
 764 {
 765         int cpu, thread;
 766
 767         if (evsel->system_wide)
 768                 nthreads = 1;
 769
 770         for (cpu = 0; cpu < ncpus; cpu++) {
 771                 for (thread = 0; thread < nthreads; thread++) {
 772                         int fd = FD(evsel, cpu, thread),
 773                             err = ioctl(fd, ioc, arg);
 774
 775                         if (err)
 776                                 return err;
 777                 }
 778         }
 779
 780         return 0;
 781 }
 782
 783 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 784                            const char *filter)
 785 {
 786         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 787                                      PERF_EVENT_IOC_SET_FILTER,
 788                                      (void *)filter);
 789 }
 790
 791 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 792 {
 793         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 794                                      PERF_EVENT_IOC_ENABLE,
 795                                      0);
 796 }
 797
 798 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 799 {
 800         if (evsel->system_wide)
 801                 nthreads = 1;
 802
 803         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 804         if (evsel->sample_id == NULL)
 805                 return -ENOMEM;
 806
 807         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 808         if (evsel->id == NULL) {
 809                 xyarray__delete(evsel->sample_id);
 810                 evsel->sample_id = NULL;
 811                 return -ENOMEM;
 812         }
 813
 814         return 0;
 815 }
 816
 817 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
 818 {
 819         memset(evsel->counts, 0, (sizeof(*evsel->counts) +
 820                                  (ncpus * sizeof(struct perf_counts_values))));
 821 }
 822
 823 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
 824 {
 825         evsel->counts = zalloc((sizeof(*evsel->counts) +
 826                                 (ncpus * sizeof(struct perf_counts_values))));
 827         return evsel->counts != NULL ? 0 : -ENOMEM;
 828 }
 829
 830 static void perf_evsel__free_fd(struct perf_evsel *evsel)
 831 {
 832         xyarray__delete(evsel->fd);
 833         evsel->fd = NULL;
 834 }
 835
 836 static void perf_evsel__free_id(struct perf_evsel *evsel)
 837 {
 838         xyarray__delete(evsel->sample_id);
 839         evsel->sample_id = NULL;
 840         zfree(&evsel->id);
 841 }
 842
 843 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 844 {
 845         int cpu, thread;
 846
 847         if (evsel->system_wide)
 848                 nthreads = 1;
 849
 850         for (cpu = 0; cpu < ncpus; cpu++)
 851                 for (thread = 0; thread < nthreads; ++thread) {
 852                         close(FD(evsel, cpu, thread));
 853                         FD(evsel, cpu, thread) = -1;
 854                 }
 855 }
 856
 857 void perf_evsel__free_counts(struct perf_evsel *evsel)
 858 {
 859         zfree(&evsel->counts);
 860 }
 861
 862 void perf_evsel__exit(struct perf_evsel *evsel)
 863 {
 864         assert(list_empty(&evsel->node));
 865         perf_evsel__free_fd(evsel);
 866         perf_evsel__free_id(evsel);
 867         close_cgroup(evsel->cgrp);
 868         zfree(&evsel->group_name);
 869         zfree(&evsel->name);
 870         perf_evsel__object.fini(evsel);
 871 }
 872
 873 void perf_evsel__delete(struct perf_evsel *evsel)
 874 {
 875         perf_evsel__exit(evsel);
 876         free(evsel);
 877 }
 878
 879 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu,
 880                                 struct perf_counts_values *count)
 881 {
 882         struct perf_counts_values tmp;
 883
 884         if (!evsel->prev_raw_counts)
 885                 return;
 886
 887         if (cpu == -1) {
 888                 tmp = evsel->prev_raw_counts->aggr;
 889                 evsel->prev_raw_counts->aggr = *count;
 890         } else {
 891                 tmp = evsel->prev_raw_counts->cpu[cpu];
 892                 evsel->prev_raw_counts->cpu[cpu] = *count;
 893         }
 894
 895         count->val = count->val - tmp.val;
 896         count->ena = count->ena - tmp.ena;
 897         count->run = count->run - tmp.run;
 898 }
 899
 900 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 901                               int cpu, int thread, bool scale)
 902 {
 903         struct perf_counts_values count;
 904         size_t nv = scale ? 3 : 1;
 905
 906         if (FD(evsel, cpu, thread) < 0)
 907                 return -EINVAL;
 908
 909         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 910                 return -ENOMEM;
 911
 912         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 913                 return -errno;
 914
 915         perf_evsel__compute_deltas(evsel, cpu, &count);
 916
 917         if (scale) {
 918                 if (count.run == 0)
 919                         count.val = 0;
 920                 else if (count.run < count.ena)
 921                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
 922         } else
 923                 count.ena = count.run = 0;
 924
 925         evsel->counts->cpu[cpu] = count;
 926         return 0;
 927 }
 928
 929 int __perf_evsel__read(struct perf_evsel *evsel,
 930                        int ncpus, int nthreads, bool scale)
 931 {
 932         size_t nv = scale ? 3 : 1;
 933         int cpu, thread;
 934         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
 935
 936         if (evsel->system_wide)
 937                 nthreads = 1;
 938
 939         aggr->val = aggr->ena = aggr->run = 0;
 940
 941         for (cpu = 0; cpu < ncpus; cpu++) {
 942                 for (thread = 0; thread < nthreads; thread++) {
 943                         if (FD(evsel, cpu, thread) < 0)
 944                                 continue;
 945
 946                         if (readn(FD(evsel, cpu, thread),
 947                                   &count, nv * sizeof(u64)) < 0)
 948                                 return -errno;
 949
 950                         aggr->val += count.val;
 951                         if (scale) {
 952                                 aggr->ena += count.ena;
 953                                 aggr->run += count.run;
 954                         }
 955                 }
 956         }
 957
 958         perf_evsel__compute_deltas(evsel, -1, aggr);
 959
 960         evsel->counts->scaled = 0;
 961         if (scale) {
 962                 if (aggr->run == 0) {
 963                         evsel->counts->scaled = -1;
 964                         aggr->val = 0;
 965                         return 0;
 966                 }
 967
 968                 if (aggr->run < aggr->ena) {
 969                         evsel->counts->scaled = 1;
 970                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 971                 }
 972         } else
 973                 aggr->ena = aggr->run = 0;
 974
 975         return 0;
 976 }
 977
 978 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 979 {
 980         struct perf_evsel *leader = evsel->leader;
 981         int fd;
 982
 983         if (perf_evsel__is_group_leader(evsel))
 984                 return -1;
 985
 986         /*
 987          * Leader must be already processed/open,
 988          * if not it's a bug.
 989          */
 990         BUG_ON(!leader->fd);
 991
 992         fd = FD(leader, cpu, thread);
 993         BUG_ON(fd == -1);
 994
 995         return fd;
 996 }
 997
 998 #define __PRINT_ATTR(fmt, cast, field)  \
 999         fprintf(fp, "  %-19s "fmt"\n", #field, cast attr->field)
1000
1001 #define PRINT_ATTR_U32(field)  __PRINT_ATTR("%u" , , field)
1002 #define PRINT_ATTR_X32(field)  __PRINT_ATTR("%#x", , field)
1003 #define PRINT_ATTR_U64(field)  __PRINT_ATTR("%" PRIu64, (uint64_t), field)
1004 #define PRINT_ATTR_X64(field)  __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
1005
1006 #define PRINT_ATTR2N(name1, field1, name2, field2)      \
1007         fprintf(fp, "  %-19s %u    %-19s %u\n",         \
1008         name1, attr->field1, name2, attr->field2)
1009
1010 #define PRINT_ATTR2(field1, field2) \
1011         PRINT_ATTR2N(#field1, field1, #field2, field2)
1012
1013 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
1014 {
1015         size_t ret = 0;
1016
1017         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1018         ret += fprintf(fp, "perf_event_attr:\n");
1019
1020         ret += PRINT_ATTR_U32(type);
1021         ret += PRINT_ATTR_U32(size);
1022         ret += PRINT_ATTR_X64(config);
1023         ret += PRINT_ATTR_U64(sample_period);
1024         ret += PRINT_ATTR_U64(sample_freq);
1025         ret += PRINT_ATTR_X64(sample_type);
1026         ret += PRINT_ATTR_X64(read_format);
1027
1028         ret += PRINT_ATTR2(disabled, inherit);
1029         ret += PRINT_ATTR2(pinned, exclusive);
1030         ret += PRINT_ATTR2(exclude_user, exclude_kernel);
1031         ret += PRINT_ATTR2(exclude_hv, exclude_idle);
1032         ret += PRINT_ATTR2(mmap, comm);
1033         ret += PRINT_ATTR2(mmap2, comm_exec);
1034         ret += PRINT_ATTR2(freq, inherit_stat);
1035         ret += PRINT_ATTR2(enable_on_exec, task);
1036         ret += PRINT_ATTR2(watermark, precise_ip);
1037         ret += PRINT_ATTR2(mmap_data, sample_id_all);
1038         ret += PRINT_ATTR2(exclude_host, exclude_guest);
1039         ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
1040                             "excl.callchain_user", exclude_callchain_user);
1041
1042         ret += PRINT_ATTR_U32(wakeup_events);
1043         ret += PRINT_ATTR_U32(wakeup_watermark);
1044         ret += PRINT_ATTR_X32(bp_type);
1045         ret += PRINT_ATTR_X64(bp_addr);
1046         ret += PRINT_ATTR_X64(config1);
1047         ret += PRINT_ATTR_U64(bp_len);
1048         ret += PRINT_ATTR_X64(config2);
1049         ret += PRINT_ATTR_X64(branch_sample_type);
1050         ret += PRINT_ATTR_X64(sample_regs_user);
1051         ret += PRINT_ATTR_U32(sample_stack_user);
1052         ret += PRINT_ATTR_X64(sample_regs_intr);
1053
1054         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
1055
1056         return ret;
1057 }
1058
1059 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1060                               struct thread_map *threads)
1061 {
1062         int cpu, thread, nthreads;
1063         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1064         int pid = -1, err;
1065         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1066
1067         if (evsel->system_wide)
1068                 nthreads = 1;
1069         else
1070                 nthreads = threads->nr;
1071
1072         if (evsel->fd == NULL &&
1073             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1074                 return -ENOMEM;
1075
1076         if (evsel->cgrp) {
1077                 flags |= PERF_FLAG_PID_CGROUP;
1078                 pid = evsel->cgrp->fd;
1079         }
1080
1081 fallback_missing_features:
1082         if (perf_missing_features.cloexec)
1083                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1084         if (perf_missing_features.mmap2)
1085                 evsel->attr.mmap2 = 0;
1086         if (perf_missing_features.exclude_guest)
1087                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1088 retry_sample_id:
1089         if (perf_missing_features.sample_id_all)
1090                 evsel->attr.sample_id_all = 0;
1091
1092         if (verbose >= 2)
1093                 perf_event_attr__fprintf(&evsel->attr, stderr);
1094
1095         for (cpu = 0; cpu < cpus->nr; cpu++) {
1096
1097                 for (thread = 0; thread < nthreads; thread++) {
1098                         int group_fd;
1099
1100                         if (!evsel->cgrp && !evsel->system_wide)
1101                                 pid = threads->map[thread];
1102
1103                         group_fd = get_group_fd(evsel, cpu, thread);
1104 retry_open:
1105                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1106                                   pid, cpus->map[cpu], group_fd, flags);
1107
1108                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1109                                                                      pid,
1110                                                                      cpus->map[cpu],
1111                                                                      group_fd, flags);
1112                         if (FD(evsel, cpu, thread) < 0) {
1113                                 err = -errno;
1114                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1115                                           err);
1116                                 goto try_fallback;
1117                         }
1118                         set_rlimit = NO_CHANGE;
1119                 }
1120         }
1121
1122         return 0;
1123
1124 try_fallback:
1125         /*
1126          * perf stat needs between 5 and 22 fds per CPU. When we run out
1127          * of them try to increase the limits.
1128          */
1129         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1130                 struct rlimit l;
1131                 int old_errno = errno;
1132
1133                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1134                         if (set_rlimit == NO_CHANGE)
1135                                 l.rlim_cur = l.rlim_max;
1136                         else {
1137                                 l.rlim_cur = l.rlim_max + 1000;
1138                                 l.rlim_max = l.rlim_cur;
1139                         }
1140                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1141                                 set_rlimit++;
1142                                 errno = old_errno;
1143                                 goto retry_open;
1144                         }
1145                 }
1146                 errno = old_errno;
1147         }
1148
1149         if (err != -EINVAL || cpu > 0 || thread > 0)
1150                 goto out_close;
1151
1152         if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1153                 perf_missing_features.cloexec = true;
1154                 goto fallback_missing_features;
1155         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1156                 perf_missing_features.mmap2 = true;
1157                 goto fallback_missing_features;
1158         } else if (!perf_missing_features.exclude_guest &&
1159                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1160                 perf_missing_features.exclude_guest = true;
1161                 goto fallback_missing_features;
1162         } else if (!perf_missing_features.sample_id_all) {
1163                 perf_missing_features.sample_id_all = true;
1164                 goto retry_sample_id;
1165         }
1166
1167 out_close:
1168         do {
1169                 while (--thread >= 0) {
1170                         close(FD(evsel, cpu, thread));
1171                         FD(evsel, cpu, thread) = -1;
1172                 }
1173                 thread = nthreads;
1174         } while (--cpu >= 0);
1175         return err;
1176 }
1177
1178 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1179 {
1180         if (evsel->fd == NULL)
1181                 return;
1182
1183         perf_evsel__close_fd(evsel, ncpus, nthreads);
1184         perf_evsel__free_fd(evsel);
1185 }
1186
1187 static struct {
1188         struct cpu_map map;
1189         int cpus[1];
1190 } empty_cpu_map = {
1191         .map.nr = 1,
1192         .cpus   = { -1, },
1193 };
1194
1195 static struct {
1196         struct thread_map map;
1197         int threads[1];
1198 } empty_thread_map = {
1199         .map.nr  = 1,
1200         .threads = { -1, },
1201 };
1202
1203 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1204                      struct thread_map *threads)
1205 {
1206         if (cpus == NULL) {
1207                 /* Work around old compiler warnings about strict aliasing */
1208                 cpus = &empty_cpu_map.map;
1209         }
1210
1211         if (threads == NULL)
1212                 threads = &empty_thread_map.map;
1213
1214         return __perf_evsel__open(evsel, cpus, threads);
1215 }
1216
1217 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1218                              struct cpu_map *cpus)
1219 {
1220         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1221 }
1222
1223 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1224                                 struct thread_map *threads)
1225 {
1226         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1227 }
1228
1229 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1230                                        const union perf_event *event,
1231                                        struct perf_sample *sample)
1232 {
1233         u64 type = evsel->attr.sample_type;
1234         const u64 *array = event->sample.array;
1235         bool swapped = evsel->needs_swap;
1236         union u64_swap u;
1237
1238         array += ((event->header.size -
1239                    sizeof(event->header)) / sizeof(u64)) - 1;
1240
1241         if (type & PERF_SAMPLE_IDENTIFIER) {
1242                 sample->id = *array;
1243                 array--;
1244         }
1245
1246         if (type & PERF_SAMPLE_CPU) {
1247                 u.val64 = *array;
1248                 if (swapped) {
1249                         /* undo swap of u64, then swap on individual u32s */
1250                         u.val64 = bswap_64(u.val64);
1251                         u.val32[0] = bswap_32(u.val32[0]);
1252                 }
1253
1254                 sample->cpu = u.val32[0];
1255                 array--;
1256         }
1257
1258         if (type & PERF_SAMPLE_STREAM_ID) {
1259                 sample->stream_id = *array;
1260                 array--;
1261         }
1262
1263         if (type & PERF_SAMPLE_ID) {
1264                 sample->id = *array;
1265                 array--;
1266         }
1267
1268         if (type & PERF_SAMPLE_TIME) {
1269                 sample->time = *array;
1270                 array--;
1271         }
1272
1273         if (type & PERF_SAMPLE_TID) {
1274                 u.val64 = *array;
1275                 if (swapped) {
1276                         /* undo swap of u64, then swap on individual u32s */
1277                         u.val64 = bswap_64(u.val64);
1278                         u.val32[0] = bswap_32(u.val32[0]);
1279                         u.val32[1] = bswap_32(u.val32[1]);
1280                 }
1281
1282                 sample->pid = u.val32[0];
1283                 sample->tid = u.val32[1];
1284                 array--;
1285         }
1286
1287         return 0;
1288 }
1289
1290 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1291                             u64 size)
1292 {
1293         return size > max_size || offset + size > endp;
1294 }
1295
1296 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1297         do {                                                            \
1298                 if (overflow(endp, (max_size), (offset), (size)))       \
1299                         return -EFAULT;                                 \
1300         } while (0)
1301
1302 #define OVERFLOW_CHECK_u64(offset) \
1303         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1304
1305 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1306                              struct perf_sample *data)
1307 {
1308         u64 type = evsel->attr.sample_type;
1309         bool swapped = evsel->needs_swap;
1310         const u64 *array;
1311         u16 max_size = event->header.size;
1312         const void *endp = (void *)event + max_size;
1313         u64 sz;
1314
1315         /*
1316          * used for cross-endian analysis. See git commit 65014ab3
1317          * for why this goofiness is needed.
1318          */
1319         union u64_swap u;
1320
1321         memset(data, 0, sizeof(*data));
1322         data->cpu = data->pid = data->tid = -1;
1323         data->stream_id = data->id = data->time = -1ULL;
1324         data->period = evsel->attr.sample_period;
1325         data->weight = 0;
1326
1327         if (event->header.type != PERF_RECORD_SAMPLE) {
1328                 if (!evsel->attr.sample_id_all)
1329                         return 0;
1330                 return perf_evsel__parse_id_sample(evsel, event, data);
1331         }
1332
1333         array = event->sample.array;
1334
1335         /*
1336          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1337          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1338          * check the format does not go past the end of the event.
1339          */
1340         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1341                 return -EFAULT;
1342
1343         data->id = -1ULL;
1344         if (type & PERF_SAMPLE_IDENTIFIER) {
1345                 data->id = *array;
1346                 array++;
1347         }
1348
1349         if (type & PERF_SAMPLE_IP) {
1350                 data->ip = *array;
1351                 array++;
1352         }
1353
1354         if (type & PERF_SAMPLE_TID) {
1355                 u.val64 = *array;
1356                 if (swapped) {
1357                         /* undo swap of u64, then swap on individual u32s */
1358                         u.val64 = bswap_64(u.val64);
1359                         u.val32[0] = bswap_32(u.val32[0]);
1360                         u.val32[1] = bswap_32(u.val32[1]);
1361                 }
1362
1363                 data->pid = u.val32[0];
1364                 data->tid = u.val32[1];
1365                 array++;
1366         }
1367
1368         if (type & PERF_SAMPLE_TIME) {
1369                 data->time = *array;
1370                 array++;
1371         }
1372
1373         data->addr = 0;
1374         if (type & PERF_SAMPLE_ADDR) {
1375                 data->addr = *array;
1376                 array++;
1377         }
1378
1379         if (type & PERF_SAMPLE_ID) {
1380                 data->id = *array;
1381                 array++;
1382         }
1383
1384         if (type & PERF_SAMPLE_STREAM_ID) {
1385                 data->stream_id = *array;
1386                 array++;
1387         }
1388
1389         if (type & PERF_SAMPLE_CPU) {
1390
1391                 u.val64 = *array;
1392                 if (swapped) {
1393                         /* undo swap of u64, then swap on individual u32s */
1394                         u.val64 = bswap_64(u.val64);
1395                         u.val32[0] = bswap_32(u.val32[0]);
1396                 }
1397
1398                 data->cpu = u.val32[0];
1399                 array++;
1400         }
1401
1402         if (type & PERF_SAMPLE_PERIOD) {
1403                 data->period = *array;
1404                 array++;
1405         }
1406
1407         if (type & PERF_SAMPLE_READ) {
1408                 u64 read_format = evsel->attr.read_format;
1409
1410                 OVERFLOW_CHECK_u64(array);
1411                 if (read_format & PERF_FORMAT_GROUP)
1412                         data->read.group.nr = *array;
1413                 else
1414                         data->read.one.value = *array;
1415
1416                 array++;
1417
1418                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1419                         OVERFLOW_CHECK_u64(array);
1420                         data->read.time_enabled = *array;
1421                         array++;
1422                 }
1423
1424                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1425                         OVERFLOW_CHECK_u64(array);
1426                         data->read.time_running = *array;
1427                         array++;
1428                 }
1429
1430                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1431                 if (read_format & PERF_FORMAT_GROUP) {
1432                         const u64 max_group_nr = UINT64_MAX /
1433                                         sizeof(struct sample_read_value);
1434
1435                         if (data->read.group.nr > max_group_nr)
1436                                 return -EFAULT;
1437                         sz = data->read.group.nr *
1438                              sizeof(struct sample_read_value);
1439                         OVERFLOW_CHECK(array, sz, max_size);
1440                         data->read.group.values =
1441                                         (struct sample_read_value *)array;
1442                         array = (void *)array + sz;
1443                 } else {
1444                         OVERFLOW_CHECK_u64(array);
1445                         data->read.one.id = *array;
1446                         array++;
1447                 }
1448         }
1449
1450         if (type & PERF_SAMPLE_CALLCHAIN) {
1451                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1452
1453                 OVERFLOW_CHECK_u64(array);
1454                 data->callchain = (struct ip_callchain *)array++;
1455                 if (data->callchain->nr > max_callchain_nr)
1456                         return -EFAULT;
1457                 sz = data->callchain->nr * sizeof(u64);
1458                 OVERFLOW_CHECK(array, sz, max_size);
1459                 array = (void *)array + sz;
1460         }
1461
1462         if (type & PERF_SAMPLE_RAW) {
1463                 OVERFLOW_CHECK_u64(array);
1464                 u.val64 = *array;
1465                 if (WARN_ONCE(swapped,
1466                               "Endianness of raw data not corrected!\n")) {
1467                         /* undo swap of u64, then swap on individual u32s */
1468                         u.val64 = bswap_64(u.val64);
1469                         u.val32[0] = bswap_32(u.val32[0]);
1470                         u.val32[1] = bswap_32(u.val32[1]);
1471                 }
1472                 data->raw_size = u.val32[0];
1473                 array = (void *)array + sizeof(u32);
1474
1475                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1476                 data->raw_data = (void *)array;
1477                 array = (void *)array + data->raw_size;
1478         }
1479
1480         if (type & PERF_SAMPLE_BRANCH_STACK) {
1481                 const u64 max_branch_nr = UINT64_MAX /
1482                                           sizeof(struct branch_entry);
1483
1484                 OVERFLOW_CHECK_u64(array);
1485                 data->branch_stack = (struct branch_stack *)array++;
1486
1487                 if (data->branch_stack->nr > max_branch_nr)
1488                         return -EFAULT;
1489                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1490                 OVERFLOW_CHECK(array, sz, max_size);
1491                 array = (void *)array + sz;
1492         }
1493
1494         if (type & PERF_SAMPLE_REGS_USER) {
1495                 OVERFLOW_CHECK_u64(array);
1496                 data->user_regs.abi = *array;
1497                 array++;
1498
1499                 if (data->user_regs.abi) {
1500                         u64 mask = evsel->attr.sample_regs_user;
1501
1502                         sz = hweight_long(mask) * sizeof(u64);
1503                         OVERFLOW_CHECK(array, sz, max_size);
1504                         data->user_regs.mask = mask;
1505                         data->user_regs.regs = (u64 *)array;
1506                         array = (void *)array + sz;
1507                 }
1508         }
1509
1510         if (type & PERF_SAMPLE_STACK_USER) {
1511                 OVERFLOW_CHECK_u64(array);
1512                 sz = *array++;
1513
1514                 data->user_stack.offset = ((char *)(array - 1)
1515                                           - (char *) event);
1516
1517                 if (!sz) {
1518                         data->user_stack.size = 0;
1519                 } else {
1520                         OVERFLOW_CHECK(array, sz, max_size);
1521                         data->user_stack.data = (char *)array;
1522                         array = (void *)array + sz;
1523                         OVERFLOW_CHECK_u64(array);
1524                         data->user_stack.size = *array++;
1525                         if (WARN_ONCE(data->user_stack.size > sz,
1526                                       "user stack dump failure\n"))
1527                                 return -EFAULT;
1528                 }
1529         }
1530
1531         data->weight = 0;
1532         if (type & PERF_SAMPLE_WEIGHT) {
1533                 OVERFLOW_CHECK_u64(array);
1534                 data->weight = *array;
1535                 array++;
1536         }
1537
1538         data->data_src = PERF_MEM_DATA_SRC_NONE;
1539         if (type & PERF_SAMPLE_DATA_SRC) {
1540                 OVERFLOW_CHECK_u64(array);
1541                 data->data_src = *array;
1542                 array++;
1543         }
1544
1545         data->transaction = 0;
1546         if (type & PERF_SAMPLE_TRANSACTION) {
1547                 OVERFLOW_CHECK_u64(array);
1548                 data->transaction = *array;
1549                 array++;
1550         }
1551
1552         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1553         if (type & PERF_SAMPLE_REGS_INTR) {
1554                 OVERFLOW_CHECK_u64(array);
1555                 data->intr_regs.abi = *array;
1556                 array++;
1557
1558                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1559                         u64 mask = evsel->attr.sample_regs_intr;
1560
1561                         sz = hweight_long(mask) * sizeof(u64);
1562                         OVERFLOW_CHECK(array, sz, max_size);
1563                         data->intr_regs.mask = mask;
1564                         data->intr_regs.regs = (u64 *)array;
1565                         array = (void *)array + sz;
1566                 }
1567         }
1568
1569         return 0;
1570 }
1571
1572 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1573                                      u64 read_format)
1574 {
1575         size_t sz, result = sizeof(struct sample_event);
1576
1577         if (type & PERF_SAMPLE_IDENTIFIER)
1578                 result += sizeof(u64);
1579
1580         if (type & PERF_SAMPLE_IP)
1581                 result += sizeof(u64);
1582
1583         if (type & PERF_SAMPLE_TID)
1584                 result += sizeof(u64);
1585
1586         if (type & PERF_SAMPLE_TIME)
1587                 result += sizeof(u64);
1588
1589         if (type & PERF_SAMPLE_ADDR)
1590                 result += sizeof(u64);
1591
1592         if (type & PERF_SAMPLE_ID)
1593                 result += sizeof(u64);
1594
1595         if (type & PERF_SAMPLE_STREAM_ID)
1596                 result += sizeof(u64);
1597
1598         if (type & PERF_SAMPLE_CPU)
1599                 result += sizeof(u64);
1600
1601         if (type & PERF_SAMPLE_PERIOD)
1602                 result += sizeof(u64);
1603
1604         if (type & PERF_SAMPLE_READ) {
1605                 result += sizeof(u64);
1606                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1607                         result += sizeof(u64);
1608                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1609                         result += sizeof(u64);
1610                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1611                 if (read_format & PERF_FORMAT_GROUP) {
1612                         sz = sample->read.group.nr *
1613                              sizeof(struct sample_read_value);
1614                         result += sz;
1615                 } else {
1616                         result += sizeof(u64);
1617                 }
1618         }
1619
1620         if (type & PERF_SAMPLE_CALLCHAIN) {
1621                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1622                 result += sz;
1623         }
1624
1625         if (type & PERF_SAMPLE_RAW) {
1626                 result += sizeof(u32);
1627                 result += sample->raw_size;
1628         }
1629
1630         if (type & PERF_SAMPLE_BRANCH_STACK) {
1631                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1632                 sz += sizeof(u64);
1633                 result += sz;
1634         }
1635
1636         if (type & PERF_SAMPLE_REGS_USER) {
1637                 if (sample->user_regs.abi) {
1638                         result += sizeof(u64);
1639                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1640                         result += sz;
1641                 } else {
1642                         result += sizeof(u64);
1643                 }
1644         }
1645
1646         if (type & PERF_SAMPLE_STACK_USER) {
1647                 sz = sample->user_stack.size;
1648                 result += sizeof(u64);
1649                 if (sz) {
1650                         result += sz;
1651                         result += sizeof(u64);
1652                 }
1653         }
1654
1655         if (type & PERF_SAMPLE_WEIGHT)
1656                 result += sizeof(u64);
1657
1658         if (type & PERF_SAMPLE_DATA_SRC)
1659                 result += sizeof(u64);
1660
1661         if (type & PERF_SAMPLE_TRANSACTION)
1662                 result += sizeof(u64);
1663
1664         if (type & PERF_SAMPLE_REGS_INTR) {
1665                 if (sample->intr_regs.abi) {
1666                         result += sizeof(u64);
1667                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1668                         result += sz;
1669                 } else {
1670                         result += sizeof(u64);
1671                 }
1672         }
1673
1674         return result;
1675 }
1676
1677 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1678                                   u64 read_format,
1679                                   const struct perf_sample *sample,
1680                                   bool swapped)
1681 {
1682         u64 *array;
1683         size_t sz;
1684         /*
1685          * used for cross-endian analysis. See git commit 65014ab3
1686          * for why this goofiness is needed.
1687          */
1688         union u64_swap u;
1689
1690         array = event->sample.array;
1691
1692         if (type & PERF_SAMPLE_IDENTIFIER) {
1693                 *array = sample->id;
1694                 array++;
1695         }
1696
1697         if (type & PERF_SAMPLE_IP) {
1698                 *array = sample->ip;
1699                 array++;
1700         }
1701
1702         if (type & PERF_SAMPLE_TID) {
1703                 u.val32[0] = sample->pid;
1704                 u.val32[1] = sample->tid;
1705                 if (swapped) {
1706                         /*
1707                          * Inverse of what is done in perf_evsel__parse_sample
1708                          */
1709                         u.val32[0] = bswap_32(u.val32[0]);
1710                         u.val32[1] = bswap_32(u.val32[1]);
1711                         u.val64 = bswap_64(u.val64);
1712                 }
1713
1714                 *array = u.val64;
1715                 array++;
1716         }
1717
1718         if (type & PERF_SAMPLE_TIME) {
1719                 *array = sample->time;
1720                 array++;
1721         }
1722
1723         if (type & PERF_SAMPLE_ADDR) {
1724                 *array = sample->addr;
1725                 array++;
1726         }
1727
1728         if (type & PERF_SAMPLE_ID) {
1729                 *array = sample->id;
1730                 array++;
1731         }
1732
1733         if (type & PERF_SAMPLE_STREAM_ID) {
1734                 *array = sample->stream_id;
1735                 array++;
1736         }
1737
1738         if (type & PERF_SAMPLE_CPU) {
1739                 u.val32[0] = sample->cpu;
1740                 if (swapped) {
1741                         /*
1742                          * Inverse of what is done in perf_evsel__parse_sample
1743                          */
1744                         u.val32[0] = bswap_32(u.val32[0]);
1745                         u.val64 = bswap_64(u.val64);
1746                 }
1747                 *array = u.val64;
1748                 array++;
1749         }
1750
1751         if (type & PERF_SAMPLE_PERIOD) {
1752                 *array = sample->period;
1753                 array++;
1754         }
1755
1756         if (type & PERF_SAMPLE_READ) {
1757                 if (read_format & PERF_FORMAT_GROUP)
1758                         *array = sample->read.group.nr;
1759                 else
1760                         *array = sample->read.one.value;
1761                 array++;
1762
1763                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1764                         *array = sample->read.time_enabled;
1765                         array++;
1766                 }
1767
1768                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1769                         *array = sample->read.time_running;
1770                         array++;
1771                 }
1772
1773                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1774                 if (read_format & PERF_FORMAT_GROUP) {
1775                         sz = sample->read.group.nr *
1776                              sizeof(struct sample_read_value);
1777                         memcpy(array, sample->read.group.values, sz);
1778                         array = (void *)array + sz;
1779                 } else {
1780                         *array = sample->read.one.id;
1781                         array++;
1782                 }
1783         }
1784
1785         if (type & PERF_SAMPLE_CALLCHAIN) {
1786                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1787                 memcpy(array, sample->callchain, sz);
1788                 array = (void *)array + sz;
1789         }
1790
1791         if (type & PERF_SAMPLE_RAW) {
1792                 u.val32[0] = sample->raw_size;
1793                 if (WARN_ONCE(swapped,
1794                               "Endianness of raw data not corrected!\n")) {
1795                         /*
1796                          * Inverse of what is done in perf_evsel__parse_sample
1797                          */
1798                         u.val32[0] = bswap_32(u.val32[0]);
1799                         u.val32[1] = bswap_32(u.val32[1]);
1800                         u.val64 = bswap_64(u.val64);
1801                 }
1802                 *array = u.val64;
1803                 array = (void *)array + sizeof(u32);
1804
1805                 memcpy(array, sample->raw_data, sample->raw_size);
1806                 array = (void *)array + sample->raw_size;
1807         }
1808
1809         if (type & PERF_SAMPLE_BRANCH_STACK) {
1810                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1811                 sz += sizeof(u64);
1812                 memcpy(array, sample->branch_stack, sz);
1813                 array = (void *)array + sz;
1814         }
1815
1816         if (type & PERF_SAMPLE_REGS_USER) {
1817                 if (sample->user_regs.abi) {
1818                         *array++ = sample->user_regs.abi;
1819                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1820                         memcpy(array, sample->user_regs.regs, sz);
1821                         array = (void *)array + sz;
1822                 } else {
1823                         *array++ = 0;
1824                 }
1825         }
1826
1827         if (type & PERF_SAMPLE_STACK_USER) {
1828                 sz = sample->user_stack.size;
1829                 *array++ = sz;
1830                 if (sz) {
1831                         memcpy(array, sample->user_stack.data, sz);
1832                         array = (void *)array + sz;
1833                         *array++ = sz;
1834                 }
1835         }
1836
1837         if (type & PERF_SAMPLE_WEIGHT) {
1838                 *array = sample->weight;
1839                 array++;
1840         }
1841
1842         if (type & PERF_SAMPLE_DATA_SRC) {
1843                 *array = sample->data_src;
1844                 array++;
1845         }
1846
1847         if (type & PERF_SAMPLE_TRANSACTION) {
1848                 *array = sample->transaction;
1849                 array++;
1850         }
1851
1852         if (type & PERF_SAMPLE_REGS_INTR) {
1853                 if (sample->intr_regs.abi) {
1854                         *array++ = sample->intr_regs.abi;
1855                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1856                         memcpy(array, sample->intr_regs.regs, sz);
1857                         array = (void *)array + sz;
1858                 } else {
1859                         *array++ = 0;
1860                 }
1861         }
1862
1863         return 0;
1864 }
1865
1866 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1867 {
1868         return pevent_find_field(evsel->tp_format, name);
1869 }
1870
1871 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1872                          const char *name)
1873 {
1874         struct format_field *field = perf_evsel__field(evsel, name);
1875         int offset;
1876
1877         if (!field)
1878                 return NULL;
1879
1880         offset = field->offset;
1881
1882         if (field->flags & FIELD_IS_DYNAMIC) {
1883                 offset = *(int *)(sample->raw_data + field->offset);
1884                 offset &= 0xffff;
1885         }
1886
1887         return sample->raw_data + offset;
1888 }
1889
1890 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1891                        const char *name)
1892 {
1893         struct format_field *field = perf_evsel__field(evsel, name);
1894         void *ptr;
1895         u64 value;
1896
1897         if (!field)
1898                 return 0;
1899
1900         ptr = sample->raw_data + field->offset;
1901
1902         switch (field->size) {
1903         case 1:
1904                 return *(u8 *)ptr;
1905         case 2:
1906                 value = *(u16 *)ptr;
1907                 break;
1908         case 4:
1909                 value = *(u32 *)ptr;
1910                 break;
1911         case 8:
1912                 value = *(u64 *)ptr;
1913                 break;
1914         default:
1915                 return 0;
1916         }
1917
1918         if (!evsel->needs_swap)
1919                 return value;
1920
1921         switch (field->size) {
1922         case 2:
1923                 return bswap_16(value);
1924         case 4:
1925                 return bswap_32(value);
1926         case 8:
1927                 return bswap_64(value);
1928         default:
1929                 return 0;
1930         }
1931
1932         return 0;
1933 }
1934
1935 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1936 {
1937         va_list args;
1938         int ret = 0;
1939
1940         if (!*first) {
1941                 ret += fprintf(fp, ",");
1942         } else {
1943                 ret += fprintf(fp, ":");
1944                 *first = false;
1945         }
1946
1947         va_start(args, fmt);
1948         ret += vfprintf(fp, fmt, args);
1949         va_end(args);
1950         return ret;
1951 }
1952
1953 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1954 {
1955         if (value == 0)
1956                 return 0;
1957
1958         return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1959 }
1960
1961 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1962
1963 struct bit_names {
1964         int bit;
1965         const char *name;
1966 };
1967
1968 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1969                          struct bit_names *bits, bool *first)
1970 {
1971         int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1972         bool first_bit = true;
1973
1974         do {
1975                 if (value & bits[i].bit) {
1976                         printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1977                         first_bit = false;
1978                 }
1979         } while (bits[++i].name != NULL);
1980
1981         return printed;
1982 }
1983
1984 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1985 {
1986 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1987         struct bit_names bits[] = {
1988                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1989                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1990                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1991                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1992                 bit_name(IDENTIFIER), bit_name(REGS_INTR),
1993                 { .name = NULL, }
1994         };
1995 #undef bit_name
1996         return bits__fprintf(fp, "sample_type", value, bits, first);
1997 }
1998
1999 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
2000 {
2001 #define bit_name(n) { PERF_FORMAT_##n, #n }
2002         struct bit_names bits[] = {
2003                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
2004                 bit_name(ID), bit_name(GROUP),
2005                 { .name = NULL, }
2006         };
2007 #undef bit_name
2008         return bits__fprintf(fp, "read_format", value, bits, first);
2009 }
2010
2011 int perf_evsel__fprintf(struct perf_evsel *evsel,
2012                         struct perf_attr_details *details, FILE *fp)
2013 {
2014         bool first = true;
2015         int printed = 0;
2016
2017         if (details->event_group) {
2018                 struct perf_evsel *pos;
2019
2020                 if (!perf_evsel__is_group_leader(evsel))
2021                         return 0;
2022
2023                 if (evsel->nr_members > 1)
2024                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2025
2026                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2027                 for_each_group_member(pos, evsel)
2028                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2029
2030                 if (evsel->nr_members > 1)
2031                         printed += fprintf(fp, "}");
2032                 goto out;
2033         }
2034
2035         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2036
2037         if (details->verbose || details->freq) {
2038                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
2039                                          (u64)evsel->attr.sample_freq);
2040         }
2041
2042         if (details->verbose) {
2043                 if_print(type);
2044                 if_print(config);
2045                 if_print(config1);
2046                 if_print(config2);
2047                 if_print(size);
2048                 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
2049                 if (evsel->attr.read_format)
2050                         printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
2051                 if_print(disabled);
2052                 if_print(inherit);
2053                 if_print(pinned);
2054                 if_print(exclusive);
2055                 if_print(exclude_user);
2056                 if_print(exclude_kernel);
2057                 if_print(exclude_hv);
2058                 if_print(exclude_idle);
2059                 if_print(mmap);
2060                 if_print(mmap2);
2061                 if_print(comm);
2062                 if_print(comm_exec);
2063                 if_print(freq);
2064                 if_print(inherit_stat);
2065                 if_print(enable_on_exec);
2066                 if_print(task);
2067                 if_print(watermark);
2068                 if_print(precise_ip);
2069                 if_print(mmap_data);
2070                 if_print(sample_id_all);
2071                 if_print(exclude_host);
2072                 if_print(exclude_guest);
2073                 if_print(__reserved_1);
2074                 if_print(wakeup_events);
2075                 if_print(bp_type);
2076                 if_print(branch_sample_type);
2077         }
2078 out:
2079         fputc('\n', fp);
2080         return ++printed;
2081 }
2082
2083 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2084                           char *msg, size_t msgsize)
2085 {
2086         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2087             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2088             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2089                 /*
2090                  * If it's cycles then fall back to hrtimer based
2091                  * cpu-clock-tick sw counter, which is always available even if
2092                  * no PMU support.
2093                  *
2094                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2095                  * b0a873e).
2096                  */
2097                 scnprintf(msg, msgsize, "%s",
2098 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2099
2100                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2101                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2102
2103                 zfree(&evsel->name);
2104                 return true;
2105         }
2106
2107         return false;
2108 }
2109
2110 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2111                               int err, char *msg, size_t size)
2112 {
2113         char sbuf[STRERR_BUFSIZE];
2114
2115         switch (err) {
2116         case EPERM:
2117         case EACCES:
2118                 return scnprintf(msg, size,
2119                  "You may not have permission to collect %sstats.\n"
2120                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
2121                  " -1 - Not paranoid at all\n"
2122                  "  0 - Disallow raw tracepoint access for unpriv\n"
2123                  "  1 - Disallow cpu events for unpriv\n"
2124                  "  2 - Disallow kernel profiling for unpriv",
2125                                  target->system_wide ? "system-wide " : "");
2126         case ENOENT:
2127                 return scnprintf(msg, size, "The %s event is not supported.",
2128                                  perf_evsel__name(evsel));
2129         case EMFILE:
2130                 return scnprintf(msg, size, "%s",
2131                          "Too many events are opened.\n"
2132                          "Try again after reducing the number of events.");
2133         case ENODEV:
2134                 if (target->cpu_list)
2135                         return scnprintf(msg, size, "%s",
2136          "No such device - did you specify an out-of-range profile CPU?\n");
2137                 break;
2138         case EOPNOTSUPP:
2139                 if (evsel->attr.precise_ip)
2140                         return scnprintf(msg, size, "%s",
2141         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2142 #if defined(__i386__) || defined(__x86_64__)
2143                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2144                         return scnprintf(msg, size, "%s",
2145         "No hardware sampling interrupt available.\n"
2146         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2147 #endif
2148                 break;
2149         case EBUSY:
2150                 if (find_process("oprofiled"))
2151                         return scnprintf(msg, size,
2152         "The PMU counters are busy/taken by another profiler.\n"
2153         "We found oprofile daemon running, please stop it and try again.");
2154                 break;
2155         default:
2156                 break;
2157         }
2158
2159         return scnprintf(msg, size,
2160         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2161         "/bin/dmesg may provide additional information.\n"
2162         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2163                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2164                          perf_evsel__name(evsel));
2165 }