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 <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->max_events  = ULONG_MAX;
 236         evsel->evlist      = NULL;
 237         evsel->bpf_fd      = -1;
 238         INIT_LIST_HEAD(&evsel->node);
 239         INIT_LIST_HEAD(&evsel->config_terms);
 240         perf_evsel__object.init(evsel);
 241         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 242         perf_evsel__calc_id_pos(evsel);
 243         evsel->cmdline_group_boundary = false;
 244         evsel->metric_expr   = NULL;
 245         evsel->metric_name   = NULL;
 246         evsel->metric_events = NULL;
 247         evsel->collect_stat  = false;
 248         evsel->pmu_name      = NULL;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (!evsel)
 256                 return NULL;
 257         perf_evsel__init(evsel, attr, idx);
 258
 259         if (perf_evsel__is_bpf_output(evsel)) {
 260                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 261                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 262                 evsel->attr.sample_period = 1;
 263         }
 264
 265         if (perf_evsel__is_clock(evsel)) {
 266                 /*
 267                  * The evsel->unit points to static alias->unit
 268                  * so it's ok to use static string in here.
 269                  */
 270                 static const char *unit = "msec";
 271
 272                 evsel->unit = unit;
 273                 evsel->scale = 1e-6;
 274         }
 275
 276         return evsel;
 277 }
 278
 279 static bool perf_event_can_profile_kernel(void)
 280 {
 281         return geteuid() == 0 || perf_event_paranoid() == -1;
 282 }
 283
 284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 285 {
 286         struct perf_event_attr attr = {
 287                 .type   = PERF_TYPE_HARDWARE,
 288                 .config = PERF_COUNT_HW_CPU_CYCLES,
 289                 .exclude_kernel = !perf_event_can_profile_kernel(),
 290         };
 291         struct perf_evsel *evsel;
 292
 293         event_attr_init(&attr);
 294
 295         if (!precise)
 296                 goto new_event;
 297         /*
 298          * Unnamed union member, not supported as struct member named
 299          * initializer in older compilers such as gcc 4.4.7
 300          *
 301          * Just for probing the precise_ip:
 302          */
 303         attr.sample_period = 1;
 304
 305         perf_event_attr__set_max_precise_ip(&attr);
 306         /*
 307          * Now let the usual logic to set up the perf_event_attr defaults
 308          * to kick in when we return and before perf_evsel__open() is called.
 309          */
 310         attr.sample_period = 0;
 311 new_event:
 312         evsel = perf_evsel__new(&attr);
 313         if (evsel == NULL)
 314                 goto out;
 315
 316         /* use asprintf() because free(evsel) assumes name is allocated */
 317         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 318                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 319                      attr.exclude_kernel ? "u" : "",
 320                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 321                 goto error_free;
 322 out:
 323         return evsel;
 324 error_free:
 325         perf_evsel__delete(evsel);
 326         evsel = NULL;
 327         goto out;
 328 }
 329
 330 /*
 331  * Returns pointer with encoded error via <linux/err.h> interface.
 332  */
 333 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 334 {
 335         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 336         int err = -ENOMEM;
 337
 338         if (evsel == NULL) {
 339                 goto out_err;
 340         } else {
 341                 struct perf_event_attr attr = {
 342                         .type          = PERF_TYPE_TRACEPOINT,
 343                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 344                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 345                 };
 346
 347                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 348                         goto out_free;
 349
 350                 evsel->tp_format = trace_event__tp_format(sys, name);
 351                 if (IS_ERR(evsel->tp_format)) {
 352                         err = PTR_ERR(evsel->tp_format);
 353                         goto out_free;
 354                 }
 355
 356                 event_attr_init(&attr);
 357                 attr.config = evsel->tp_format->id;
 358                 attr.sample_period = 1;
 359                 perf_evsel__init(evsel, &attr, idx);
 360         }
 361
 362         return evsel;
 363
 364 out_free:
 365         zfree(&evsel->name);
 366         free(evsel);
 367 out_err:
 368         return ERR_PTR(err);
 369 }
 370
 371 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 372         "cycles",
 373         "instructions",
 374         "cache-references",
 375         "cache-misses",
 376         "branches",
 377         "branch-misses",
 378         "bus-cycles",
 379         "stalled-cycles-frontend",
 380         "stalled-cycles-backend",
 381         "ref-cycles",
 382 };
 383
 384 static const char *__perf_evsel__hw_name(u64 config)
 385 {
 386         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 387                 return perf_evsel__hw_names[config];
 388
 389         return "unknown-hardware";
 390 }
 391
 392 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 393 {
 394         int colon = 0, r = 0;
 395         struct perf_event_attr *attr = &evsel->attr;
 396         bool exclude_guest_default = false;
 397
 398 #define MOD_PRINT(context, mod) do {                                    \
 399                 if (!attr->exclude_##context) {                         \
 400                         if (!colon) colon = ++r;                        \
 401                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 402                 } } while(0)
 403
 404         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 405                 MOD_PRINT(kernel, 'k');
 406                 MOD_PRINT(user, 'u');
 407                 MOD_PRINT(hv, 'h');
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->precise_ip) {
 412                 if (!colon)
 413                         colon = ++r;
 414                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 415                 exclude_guest_default = true;
 416         }
 417
 418         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 419                 MOD_PRINT(host, 'H');
 420                 MOD_PRINT(guest, 'G');
 421         }
 422 #undef MOD_PRINT
 423         if (colon)
 424                 bf[colon - 1] = ':';
 425         return r;
 426 }
 427
 428 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 429 {
 430         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 431         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 432 }
 433
 434 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 435         "cpu-clock",
 436         "task-clock",
 437         "page-faults",
 438         "context-switches",
 439         "cpu-migrations",
 440         "minor-faults",
 441         "major-faults",
 442         "alignment-faults",
 443         "emulation-faults",
 444         "dummy",
 445 };
 446
 447 static const char *__perf_evsel__sw_name(u64 config)
 448 {
 449         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 450                 return perf_evsel__sw_names[config];
 451         return "unknown-software";
 452 }
 453
 454 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 455 {
 456         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 457         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 458 }
 459
 460 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 461 {
 462         int r;
 463
 464         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 465
 466         if (type & HW_BREAKPOINT_R)
 467                 r += scnprintf(bf + r, size - r, "r");
 468
 469         if (type & HW_BREAKPOINT_W)
 470                 r += scnprintf(bf + r, size - r, "w");
 471
 472         if (type & HW_BREAKPOINT_X)
 473                 r += scnprintf(bf + r, size - r, "x");
 474
 475         return r;
 476 }
 477
 478 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 479 {
 480         struct perf_event_attr *attr = &evsel->attr;
 481         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 482         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 483 }
 484
 485 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 486                                 [PERF_EVSEL__MAX_ALIASES] = {
 487  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 488  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 489  { "LLC",       "L2",                                                   },
 490  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 491  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 492  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 493  { "node",                                                              },
 494 };
 495
 496 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 497                                    [PERF_EVSEL__MAX_ALIASES] = {
 498  { "load",      "loads",        "read",                                 },
 499  { "store",     "stores",       "write",                                },
 500  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 501 };
 502
 503 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 504                                        [PERF_EVSEL__MAX_ALIASES] = {
 505  { "refs",      "Reference",    "ops",          "access",               },
 506  { "misses",    "miss",                                                 },
 507 };
 508
 509 #define C(x)            PERF_COUNT_HW_CACHE_##x
 510 #define CACHE_READ      (1 << C(OP_READ))
 511 #define CACHE_WRITE     (1 << C(OP_WRITE))
 512 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 513 #define COP(x)          (1 << x)
 514
 515 /*
 516  * cache operartion stat
 517  * L1I : Read and prefetch only
 518  * ITLB and BPU : Read-only
 519  */
 520 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 521  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 522  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 523  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(ITLB)]      = (CACHE_READ),
 526  [C(BPU)]       = (CACHE_READ),
 527  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 528 };
 529
 530 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 531 {
 532         if (perf_evsel__hw_cache_stat[type] & COP(op))
 533                 return true;    /* valid */
 534         else
 535                 return false;   /* invalid */
 536 }
 537
 538 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 539                                             char *bf, size_t size)
 540 {
 541         if (result) {
 542                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 543                                  perf_evsel__hw_cache_op[op][0],
 544                                  perf_evsel__hw_cache_result[result][0]);
 545         }
 546
 547         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 548                          perf_evsel__hw_cache_op[op][1]);
 549 }
 550
 551 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 552 {
 553         u8 op, result, type = (config >>  0) & 0xff;
 554         const char *err = "unknown-ext-hardware-cache-type";
 555
 556         if (type >= PERF_COUNT_HW_CACHE_MAX)
 557                 goto out_err;
 558
 559         op = (config >>  8) & 0xff;
 560         err = "unknown-ext-hardware-cache-op";
 561         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 562                 goto out_err;
 563
 564         result = (config >> 16) & 0xff;
 565         err = "unknown-ext-hardware-cache-result";
 566         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 567                 goto out_err;
 568
 569         err = "invalid-cache";
 570         if (!perf_evsel__is_cache_op_valid(type, op))
 571                 goto out_err;
 572
 573         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 574 out_err:
 575         return scnprintf(bf, size, "%s", err);
 576 }
 577
 578 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 579 {
 580         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 581         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 582 }
 583
 584 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 585 {
 586         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 587         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 588 }
 589
 590 const char *perf_evsel__name(struct perf_evsel *evsel)
 591 {
 592         char bf[128];
 593
 594         if (evsel->name)
 595                 return evsel->name;
 596
 597         switch (evsel->attr.type) {
 598         case PERF_TYPE_RAW:
 599                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 600                 break;
 601
 602         case PERF_TYPE_HARDWARE:
 603                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 604                 break;
 605
 606         case PERF_TYPE_HW_CACHE:
 607                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 608                 break;
 609
 610         case PERF_TYPE_SOFTWARE:
 611                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 612                 break;
 613
 614         case PERF_TYPE_TRACEPOINT:
 615                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 616                 break;
 617
 618         case PERF_TYPE_BREAKPOINT:
 619                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 620                 break;
 621
 622         default:
 623                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 624                           evsel->attr.type);
 625                 break;
 626         }
 627
 628         evsel->name = strdup(bf);
 629
 630         return evsel->name ?: "unknown";
 631 }
 632
 633 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 634 {
 635         return evsel->group_name ?: "anon group";
 636 }
 637
 638 /*
 639  * Returns the group details for the specified leader,
 640  * with following rules.
 641  *
 642  *  For record -e '{cycles,instructions}'
 643  *    'anon group { cycles:u, instructions:u }'
 644  *
 645  *  For record -e 'cycles,instructions' and report --group
 646  *    'cycles:u, instructions:u'
 647  */
 648 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 649 {
 650         int ret = 0;
 651         struct perf_evsel *pos;
 652         const char *group_name = perf_evsel__group_name(evsel);
 653
 654         if (!evsel->forced_leader)
 655                 ret = scnprintf(buf, size, "%s { ", group_name);
 656
 657         ret += scnprintf(buf + ret, size - ret, "%s",
 658                          perf_evsel__name(evsel));
 659
 660         for_each_group_member(pos, evsel)
 661                 ret += scnprintf(buf + ret, size - ret, ", %s",
 662                                  perf_evsel__name(pos));
 663
 664         if (!evsel->forced_leader)
 665                 ret += scnprintf(buf + ret, size - ret, " }");
 666
 667         return ret;
 668 }
 669
 670 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 671                                            struct record_opts *opts,
 672                                            struct callchain_param *param)
 673 {
 674         bool function = perf_evsel__is_function_event(evsel);
 675         struct perf_event_attr *attr = &evsel->attr;
 676
 677         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 678
 679         attr->sample_max_stack = param->max_stack;
 680
 681         if (param->record_mode == CALLCHAIN_LBR) {
 682                 if (!opts->branch_stack) {
 683                         if (attr->exclude_user) {
 684                                 pr_warning("LBR callstack option is only available "
 685                                            "to get user callchain information. "
 686                                            "Falling back to framepointers.\n");
 687                         } else {
 688                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 689                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 690                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 691                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 692                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 693                         }
 694                 } else
 695                          pr_warning("Cannot use LBR callstack with branch stack. "
 696                                     "Falling back to framepointers.\n");
 697         }
 698
 699         if (param->record_mode == CALLCHAIN_DWARF) {
 700                 if (!function) {
 701                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 702                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 703                         attr->sample_regs_user |= PERF_REGS_MASK;
 704                         attr->sample_stack_user = param->dump_size;
 705                         attr->exclude_callchain_user = 1;
 706                 } else {
 707                         pr_info("Cannot use DWARF unwind for function trace event,"
 708                                 " falling back to framepointers.\n");
 709                 }
 710         }
 711
 712         if (function) {
 713                 pr_info("Disabling user space callchains for function trace event.\n");
 714                 attr->exclude_callchain_user = 1;
 715         }
 716 }
 717
 718 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 719                                   struct record_opts *opts,
 720                                   struct callchain_param *param)
 721 {
 722         if (param->enabled)
 723                 return __perf_evsel__config_callchain(evsel, opts, param);
 724 }
 725
 726 static void
 727 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 728                             struct callchain_param *param)
 729 {
 730         struct perf_event_attr *attr = &evsel->attr;
 731
 732         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 733         if (param->record_mode == CALLCHAIN_LBR) {
 734                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 735                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 736                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 737         }
 738         if (param->record_mode == CALLCHAIN_DWARF) {
 739                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 740                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 741         }
 742 }
 743
 744 static void apply_config_terms(struct perf_evsel *evsel,
 745                                struct record_opts *opts, bool track)
 746 {
 747         struct perf_evsel_config_term *term;
 748         struct list_head *config_terms = &evsel->config_terms;
 749         struct perf_event_attr *attr = &evsel->attr;
 750         /* callgraph default */
 751         struct callchain_param param = {
 752                 .record_mode = callchain_param.record_mode,
 753         };
 754         u32 dump_size = 0;
 755         int max_stack = 0;
 756         const char *callgraph_buf = NULL;
 757
 758         list_for_each_entry(term, config_terms, list) {
 759                 switch (term->type) {
 760                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 761                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 762                                 attr->sample_period = term->val.period;
 763                                 attr->freq = 0;
 764                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 765                         }
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 768                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 769                                 attr->sample_freq = term->val.freq;
 770                                 attr->freq = 1;
 771                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 772                         }
 773                         break;
 774                 case PERF_EVSEL__CONFIG_TERM_TIME:
 775                         if (term->val.time)
 776                                 perf_evsel__set_sample_bit(evsel, TIME);
 777                         else
 778                                 perf_evsel__reset_sample_bit(evsel, TIME);
 779                         break;
 780                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 781                         callgraph_buf = term->val.callgraph;
 782                         break;
 783                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 784                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 785                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 786                                 parse_branch_str(term->val.branch,
 787                                                  &attr->branch_sample_type);
 788                         } else
 789                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 790                         break;
 791                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 792                         dump_size = term->val.stack_user;
 793                         break;
 794                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 795                         max_stack = term->val.max_stack;
 796                         break;
 797                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 798                         evsel->max_events = term->val.max_events;
 799                         break;
 800                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 801                         /*
 802                          * attr->inherit should has already been set by
 803                          * perf_evsel__config. If user explicitly set
 804                          * inherit using config terms, override global
 805                          * opt->no_inherit setting.
 806                          */
 807                         attr->inherit = term->val.inherit ? 1 : 0;
 808                         break;
 809                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 810                         attr->write_backward = term->val.overwrite ? 1 : 0;
 811                         break;
 812                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 813                         break;
 814                 default:
 815                         break;
 816                 }
 817         }
 818
 819         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 820         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 821                 bool sample_address = false;
 822
 823                 if (max_stack) {
 824                         param.max_stack = max_stack;
 825                         if (callgraph_buf == NULL)
 826                                 callgraph_buf = "fp";
 827                 }
 828
 829                 /* parse callgraph parameters */
 830                 if (callgraph_buf != NULL) {
 831                         if (!strcmp(callgraph_buf, "no")) {
 832                                 param.enabled = false;
 833                                 param.record_mode = CALLCHAIN_NONE;
 834                         } else {
 835                                 param.enabled = true;
 836                                 if (parse_callchain_record(callgraph_buf, &param)) {
 837                                         pr_err("per-event callgraph setting for %s failed. "
 838                                                "Apply callgraph global setting for it\n",
 839                                                evsel->name);
 840                                         return;
 841                                 }
 842                                 if (param.record_mode == CALLCHAIN_DWARF)
 843                                         sample_address = true;
 844                         }
 845                 }
 846                 if (dump_size > 0) {
 847                         dump_size = round_up(dump_size, sizeof(u64));
 848                         param.dump_size = dump_size;
 849                 }
 850
 851                 /* If global callgraph set, clear it */
 852                 if (callchain_param.enabled)
 853                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 854
 855                 /* set perf-event callgraph */
 856                 if (param.enabled) {
 857                         if (sample_address) {
 858                                 perf_evsel__set_sample_bit(evsel, ADDR);
 859                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 860                                 evsel->attr.mmap_data = track;
 861                         }
 862                         perf_evsel__config_callchain(evsel, opts, &param);
 863                 }
 864         }
 865 }
 866
 867 static bool is_dummy_event(struct perf_evsel *evsel)
 868 {
 869         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 870                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 871 }
 872
 873 /*
 874  * The enable_on_exec/disabled value strategy:
 875  *
 876  *  1) For any type of traced program:
 877  *    - all independent events and group leaders are disabled
 878  *    - all group members are enabled
 879  *
 880  *     Group members are ruled by group leaders. They need to
 881  *     be enabled, because the group scheduling relies on that.
 882  *
 883  *  2) For traced programs executed by perf:
 884  *     - all independent events and group leaders have
 885  *       enable_on_exec set
 886  *     - we don't specifically enable or disable any event during
 887  *       the record command
 888  *
 889  *     Independent events and group leaders are initially disabled
 890  *     and get enabled by exec. Group members are ruled by group
 891  *     leaders as stated in 1).
 892  *
 893  *  3) For traced programs attached by perf (pid/tid):
 894  *     - we specifically enable or disable all events during
 895  *       the record command
 896  *
 897  *     When attaching events to already running traced we
 898  *     enable/disable events specifically, as there's no
 899  *     initial traced exec call.
 900  */
 901 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 902                         struct callchain_param *callchain)
 903 {
 904         struct perf_evsel *leader = evsel->leader;
 905         struct perf_event_attr *attr = &evsel->attr;
 906         int track = evsel->tracking;
 907         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 908
 909         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 910         attr->inherit       = !opts->no_inherit;
 911         attr->write_backward = opts->overwrite ? 1 : 0;
 912
 913         perf_evsel__set_sample_bit(evsel, IP);
 914         perf_evsel__set_sample_bit(evsel, TID);
 915
 916         if (evsel->sample_read) {
 917                 perf_evsel__set_sample_bit(evsel, READ);
 918
 919                 /*
 920                  * We need ID even in case of single event, because
 921                  * PERF_SAMPLE_READ process ID specific data.
 922                  */
 923                 perf_evsel__set_sample_id(evsel, false);
 924
 925                 /*
 926                  * Apply group format only if we belong to group
 927                  * with more than one members.
 928                  */
 929                 if (leader->nr_members > 1) {
 930                         attr->read_format |= PERF_FORMAT_GROUP;
 931                         attr->inherit = 0;
 932                 }
 933         }
 934
 935         /*
 936          * We default some events to have a default interval. But keep
 937          * it a weak assumption overridable by the user.
 938          */
 939         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 940                                      opts->user_interval != ULLONG_MAX)) {
 941                 if (opts->freq) {
 942                         perf_evsel__set_sample_bit(evsel, PERIOD);
 943                         attr->freq              = 1;
 944                         attr->sample_freq       = opts->freq;
 945                 } else {
 946                         attr->sample_period = opts->default_interval;
 947                 }
 948         }
 949
 950         /*
 951          * Disable sampling for all group members other
 952          * than leader in case leader 'leads' the sampling.
 953          */
 954         if ((leader != evsel) && leader->sample_read) {
 955                 attr->freq           = 0;
 956                 attr->sample_freq    = 0;
 957                 attr->sample_period  = 0;
 958                 attr->write_backward = 0;
 959         }
 960
 961         if (opts->no_samples)
 962                 attr->sample_freq = 0;
 963
 964         if (opts->inherit_stat) {
 965                 evsel->attr.read_format |=
 966                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 967                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 968                         PERF_FORMAT_ID;
 969                 attr->inherit_stat = 1;
 970         }
 971
 972         if (opts->sample_address) {
 973                 perf_evsel__set_sample_bit(evsel, ADDR);
 974                 attr->mmap_data = track;
 975         }
 976
 977         /*
 978          * We don't allow user space callchains for  function trace
 979          * event, due to issues with page faults while tracing page
 980          * fault handler and its overall trickiness nature.
 981          */
 982         if (perf_evsel__is_function_event(evsel))
 983                 evsel->attr.exclude_callchain_user = 1;
 984
 985         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 986                 perf_evsel__config_callchain(evsel, opts, callchain);
 987
 988         if (opts->sample_intr_regs) {
 989                 attr->sample_regs_intr = opts->sample_intr_regs;
 990                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 991         }
 992
 993         if (opts->sample_user_regs) {
 994                 attr->sample_regs_user |= opts->sample_user_regs;
 995                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 996         }
 997
 998         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 999                 perf_evsel__set_sample_bit(evsel, CPU);
1000
1001         /*
1002          * When the user explicitly disabled time don't force it here.
1003          */
1004         if (opts->sample_time &&
1005             (!perf_missing_features.sample_id_all &&
1006             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1007              opts->sample_time_set)))
1008                 perf_evsel__set_sample_bit(evsel, TIME);
1009
1010         if (opts->raw_samples && !evsel->no_aux_samples) {
1011                 perf_evsel__set_sample_bit(evsel, TIME);
1012                 perf_evsel__set_sample_bit(evsel, RAW);
1013                 perf_evsel__set_sample_bit(evsel, CPU);
1014         }
1015
1016         if (opts->sample_address)
1017                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1018
1019         if (opts->sample_phys_addr)
1020                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1021
1022         if (opts->no_buffering) {
1023                 attr->watermark = 0;
1024                 attr->wakeup_events = 1;
1025         }
1026         if (opts->branch_stack && !evsel->no_aux_samples) {
1027                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1028                 attr->branch_sample_type = opts->branch_stack;
1029         }
1030
1031         if (opts->sample_weight)
1032                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1033
1034         attr->task  = track;
1035         attr->mmap  = track;
1036         attr->mmap2 = track && !perf_missing_features.mmap2;
1037         attr->comm  = track;
1038
1039         if (opts->record_namespaces)
1040                 attr->namespaces  = track;
1041
1042         if (opts->record_switch_events)
1043                 attr->context_switch = track;
1044
1045         if (opts->sample_transaction)
1046                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1047
1048         if (opts->running_time) {
1049                 evsel->attr.read_format |=
1050                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1051                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1052         }
1053
1054         /*
1055          * XXX see the function comment above
1056          *
1057          * Disabling only independent events or group leaders,
1058          * keeping group members enabled.
1059          */
1060         if (perf_evsel__is_group_leader(evsel))
1061                 attr->disabled = 1;
1062
1063         /*
1064          * Setting enable_on_exec for independent events and
1065          * group leaders for traced executed by perf.
1066          */
1067         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1068                 !opts->initial_delay)
1069                 attr->enable_on_exec = 1;
1070
1071         if (evsel->immediate) {
1072                 attr->disabled = 0;
1073                 attr->enable_on_exec = 0;
1074         }
1075
1076         clockid = opts->clockid;
1077         if (opts->use_clockid) {
1078                 attr->use_clockid = 1;
1079                 attr->clockid = opts->clockid;
1080         }
1081
1082         if (evsel->precise_max)
1083                 perf_event_attr__set_max_precise_ip(attr);
1084
1085         if (opts->all_user) {
1086                 attr->exclude_kernel = 1;
1087                 attr->exclude_user   = 0;
1088         }
1089
1090         if (opts->all_kernel) {
1091                 attr->exclude_kernel = 0;
1092                 attr->exclude_user   = 1;
1093         }
1094
1095         if (evsel->own_cpus || evsel->unit)
1096                 evsel->attr.read_format |= PERF_FORMAT_ID;
1097
1098         /*
1099          * Apply event specific term settings,
1100          * it overloads any global configuration.
1101          */
1102         apply_config_terms(evsel, opts, track);
1103
1104         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1105
1106         /* The --period option takes the precedence. */
1107         if (opts->period_set) {
1108                 if (opts->period)
1109                         perf_evsel__set_sample_bit(evsel, PERIOD);
1110                 else
1111                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1112         }
1113
1114         /*
1115          * For initial_delay, a dummy event is added implicitly.
1116          * The software event will trigger -EOPNOTSUPP error out,
1117          * if BRANCH_STACK bit is set.
1118          */
1119         if (opts->initial_delay && is_dummy_event(evsel))
1120                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1121 }
1122
1123 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1124 {
1125         if (evsel->system_wide)
1126                 nthreads = 1;
1127
1128         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1129
1130         if (evsel->fd) {
1131                 int cpu, thread;
1132                 for (cpu = 0; cpu < ncpus; cpu++) {
1133                         for (thread = 0; thread < nthreads; thread++) {
1134                                 FD(evsel, cpu, thread) = -1;
1135                         }
1136                 }
1137         }
1138
1139         return evsel->fd != NULL ? 0 : -ENOMEM;
1140 }
1141
1142 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1143                           int ioc,  void *arg)
1144 {
1145         int cpu, thread;
1146
1147         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1148                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1149                         int fd = FD(evsel, cpu, thread),
1150                             err = ioctl(fd, ioc, arg);
1151
1152                         if (err)
1153                                 return err;
1154                 }
1155         }
1156
1157         return 0;
1158 }
1159
1160 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1161 {
1162         return perf_evsel__run_ioctl(evsel,
1163                                      PERF_EVENT_IOC_SET_FILTER,
1164                                      (void *)filter);
1165 }
1166
1167 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1168 {
1169         char *new_filter = strdup(filter);
1170
1171         if (new_filter != NULL) {
1172                 free(evsel->filter);
1173                 evsel->filter = new_filter;
1174                 return 0;
1175         }
1176
1177         return -1;
1178 }
1179
1180 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1181                                      const char *fmt, const char *filter)
1182 {
1183         char *new_filter;
1184
1185         if (evsel->filter == NULL)
1186                 return perf_evsel__set_filter(evsel, filter);
1187
1188         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1189                 free(evsel->filter);
1190                 evsel->filter = new_filter;
1191                 return 0;
1192         }
1193
1194         return -1;
1195 }
1196
1197 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1198 {
1199         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1200 }
1201
1202 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1203 {
1204         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1205 }
1206
1207 int perf_evsel__enable(struct perf_evsel *evsel)
1208 {
1209         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1210
1211         if (!err)
1212                 evsel->disabled = false;
1213
1214         return err;
1215 }
1216
1217 int perf_evsel__disable(struct perf_evsel *evsel)
1218 {
1219         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1220         /*
1221          * We mark it disabled here so that tools that disable a event can
1222          * ignore events after they disable it. I.e. the ring buffer may have
1223          * already a few more events queued up before the kernel got the stop
1224          * request.
1225          */
1226         if (!err)
1227                 evsel->disabled = true;
1228
1229         return err;
1230 }
1231
1232 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1233 {
1234         if (ncpus == 0 || nthreads == 0)
1235                 return 0;
1236
1237         if (evsel->system_wide)
1238                 nthreads = 1;
1239
1240         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1241         if (evsel->sample_id == NULL)
1242                 return -ENOMEM;
1243
1244         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1245         if (evsel->id == NULL) {
1246                 xyarray__delete(evsel->sample_id);
1247                 evsel->sample_id = NULL;
1248                 return -ENOMEM;
1249         }
1250
1251         return 0;
1252 }
1253
1254 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1255 {
1256         xyarray__delete(evsel->fd);
1257         evsel->fd = NULL;
1258 }
1259
1260 static void perf_evsel__free_id(struct perf_evsel *evsel)
1261 {
1262         xyarray__delete(evsel->sample_id);
1263         evsel->sample_id = NULL;
1264         zfree(&evsel->id);
1265 }
1266
1267 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1268 {
1269         struct perf_evsel_config_term *term, *h;
1270
1271         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1272                 list_del(&term->list);
1273                 free(term);
1274         }
1275 }
1276
1277 void perf_evsel__close_fd(struct perf_evsel *evsel)
1278 {
1279         int cpu, thread;
1280
1281         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1282                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1283                         close(FD(evsel, cpu, thread));
1284                         FD(evsel, cpu, thread) = -1;
1285                 }
1286 }
1287
1288 void perf_evsel__exit(struct perf_evsel *evsel)
1289 {
1290         assert(list_empty(&evsel->node));
1291         assert(evsel->evlist == NULL);
1292         perf_evsel__free_fd(evsel);
1293         perf_evsel__free_id(evsel);
1294         perf_evsel__free_config_terms(evsel);
1295         cgroup__put(evsel->cgrp);
1296         cpu_map__put(evsel->cpus);
1297         cpu_map__put(evsel->own_cpus);
1298         thread_map__put(evsel->threads);
1299         zfree(&evsel->group_name);
1300         zfree(&evsel->name);
1301         perf_evsel__object.fini(evsel);
1302 }
1303
1304 void perf_evsel__delete(struct perf_evsel *evsel)
1305 {
1306         perf_evsel__exit(evsel);
1307         free(evsel);
1308 }
1309
1310 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1311                                 struct perf_counts_values *count)
1312 {
1313         struct perf_counts_values tmp;
1314
1315         if (!evsel->prev_raw_counts)
1316                 return;
1317
1318         if (cpu == -1) {
1319                 tmp = evsel->prev_raw_counts->aggr;
1320                 evsel->prev_raw_counts->aggr = *count;
1321         } else {
1322                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1323                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1324         }
1325
1326         count->val = count->val - tmp.val;
1327         count->ena = count->ena - tmp.ena;
1328         count->run = count->run - tmp.run;
1329 }
1330
1331 void perf_counts_values__scale(struct perf_counts_values *count,
1332                                bool scale, s8 *pscaled)
1333 {
1334         s8 scaled = 0;
1335
1336         if (scale) {
1337                 if (count->run == 0) {
1338                         scaled = -1;
1339                         count->val = 0;
1340                 } else if (count->run < count->ena) {
1341                         scaled = 1;
1342                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1343                 }
1344         } else
1345                 count->ena = count->run = 0;
1346
1347         if (pscaled)
1348                 *pscaled = scaled;
1349 }
1350
1351 static int perf_evsel__read_size(struct perf_evsel *evsel)
1352 {
1353         u64 read_format = evsel->attr.read_format;
1354         int entry = sizeof(u64); /* value */
1355         int size = 0;
1356         int nr = 1;
1357
1358         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1359                 size += sizeof(u64);
1360
1361         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1362                 size += sizeof(u64);
1363
1364         if (read_format & PERF_FORMAT_ID)
1365                 entry += sizeof(u64);
1366
1367         if (read_format & PERF_FORMAT_GROUP) {
1368                 nr = evsel->nr_members;
1369                 size += sizeof(u64);
1370         }
1371
1372         size += entry * nr;
1373         return size;
1374 }
1375
1376 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1377                      struct perf_counts_values *count)
1378 {
1379         size_t size = perf_evsel__read_size(evsel);
1380
1381         memset(count, 0, sizeof(*count));
1382
1383         if (FD(evsel, cpu, thread) < 0)
1384                 return -EINVAL;
1385
1386         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1387                 return -errno;
1388
1389         return 0;
1390 }
1391
1392 static int
1393 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1394 {
1395         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1396
1397         return perf_evsel__read(evsel, cpu, thread, count);
1398 }
1399
1400 static void
1401 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1402                       u64 val, u64 ena, u64 run)
1403 {
1404         struct perf_counts_values *count;
1405
1406         count = perf_counts(counter->counts, cpu, thread);
1407
1408         count->val    = val;
1409         count->ena    = ena;
1410         count->run    = run;
1411         count->loaded = true;
1412 }
1413
1414 static int
1415 perf_evsel__process_group_data(struct perf_evsel *leader,
1416                                int cpu, int thread, u64 *data)
1417 {
1418         u64 read_format = leader->attr.read_format;
1419         struct sample_read_value *v;
1420         u64 nr, ena = 0, run = 0, i;
1421
1422         nr = *data++;
1423
1424         if (nr != (u64) leader->nr_members)
1425                 return -EINVAL;
1426
1427         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1428                 ena = *data++;
1429
1430         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1431                 run = *data++;
1432
1433         v = (struct sample_read_value *) data;
1434
1435         perf_evsel__set_count(leader, cpu, thread,
1436                               v[0].value, ena, run);
1437
1438         for (i = 1; i < nr; i++) {
1439                 struct perf_evsel *counter;
1440
1441                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1442                 if (!counter)
1443                         return -EINVAL;
1444
1445                 perf_evsel__set_count(counter, cpu, thread,
1446                                       v[i].value, ena, run);
1447         }
1448
1449         return 0;
1450 }
1451
1452 static int
1453 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1454 {
1455         struct perf_stat_evsel *ps = leader->stats;
1456         u64 read_format = leader->attr.read_format;
1457         int size = perf_evsel__read_size(leader);
1458         u64 *data = ps->group_data;
1459
1460         if (!(read_format & PERF_FORMAT_ID))
1461                 return -EINVAL;
1462
1463         if (!perf_evsel__is_group_leader(leader))
1464                 return -EINVAL;
1465
1466         if (!data) {
1467                 data = zalloc(size);
1468                 if (!data)
1469                         return -ENOMEM;
1470
1471                 ps->group_data = data;
1472         }
1473
1474         if (FD(leader, cpu, thread) < 0)
1475                 return -EINVAL;
1476
1477         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1478                 return -errno;
1479
1480         return perf_evsel__process_group_data(leader, cpu, thread, data);
1481 }
1482
1483 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1484 {
1485         u64 read_format = evsel->attr.read_format;
1486
1487         if (read_format & PERF_FORMAT_GROUP)
1488                 return perf_evsel__read_group(evsel, cpu, thread);
1489         else
1490                 return perf_evsel__read_one(evsel, cpu, thread);
1491 }
1492
1493 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1494                               int cpu, int thread, bool scale)
1495 {
1496         struct perf_counts_values count;
1497         size_t nv = scale ? 3 : 1;
1498
1499         if (FD(evsel, cpu, thread) < 0)
1500                 return -EINVAL;
1501
1502         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1503                 return -ENOMEM;
1504
1505         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1506                 return -errno;
1507
1508         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1509         perf_counts_values__scale(&count, scale, NULL);
1510         *perf_counts(evsel->counts, cpu, thread) = count;
1511         return 0;
1512 }
1513
1514 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1515 {
1516         struct perf_evsel *leader = evsel->leader;
1517         int fd;
1518
1519         if (perf_evsel__is_group_leader(evsel))
1520                 return -1;
1521
1522         /*
1523          * Leader must be already processed/open,
1524          * if not it's a bug.
1525          */
1526         BUG_ON(!leader->fd);
1527
1528         fd = FD(leader, cpu, thread);
1529         BUG_ON(fd == -1);
1530
1531         return fd;
1532 }
1533
1534 struct bit_names {
1535         int bit;
1536         const char *name;
1537 };
1538
1539 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1540 {
1541         bool first_bit = true;
1542         int i = 0;
1543
1544         do {
1545                 if (value & bits[i].bit) {
1546                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1547                         first_bit = false;
1548                 }
1549         } while (bits[++i].name != NULL);
1550 }
1551
1552 static void __p_sample_type(char *buf, size_t size, u64 value)
1553 {
1554 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1555         struct bit_names bits[] = {
1556                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1557                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1558                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1559                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1560                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1561                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1562                 { .name = NULL, }
1563         };
1564 #undef bit_name
1565         __p_bits(buf, size, value, bits);
1566 }
1567
1568 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1569 {
1570 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1571         struct bit_names bits[] = {
1572                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1573                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1574                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1575                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1576                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1577                 { .name = NULL, }
1578         };
1579 #undef bit_name
1580         __p_bits(buf, size, value, bits);
1581 }
1582
1583 static void __p_read_format(char *buf, size_t size, u64 value)
1584 {
1585 #define bit_name(n) { PERF_FORMAT_##n, #n }
1586         struct bit_names bits[] = {
1587                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1588                 bit_name(ID), bit_name(GROUP),
1589                 { .name = NULL, }
1590         };
1591 #undef bit_name
1592         __p_bits(buf, size, value, bits);
1593 }
1594
1595 #define BUF_SIZE                1024
1596
1597 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1598 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1599 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1600 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1601 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1602 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1603
1604 #define PRINT_ATTRn(_n, _f, _p)                         \
1605 do {                                                    \
1606         if (attr->_f) {                                 \
1607                 _p(attr->_f);                           \
1608                 ret += attr__fprintf(fp, _n, buf, priv);\
1609         }                                               \
1610 } while (0)
1611
1612 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1613
1614 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1615                              attr__fprintf_f attr__fprintf, void *priv)
1616 {
1617         char buf[BUF_SIZE];
1618         int ret = 0;
1619
1620         PRINT_ATTRf(type, p_unsigned);
1621         PRINT_ATTRf(size, p_unsigned);
1622         PRINT_ATTRf(config, p_hex);
1623         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1624         PRINT_ATTRf(sample_type, p_sample_type);
1625         PRINT_ATTRf(read_format, p_read_format);
1626
1627         PRINT_ATTRf(disabled, p_unsigned);
1628         PRINT_ATTRf(inherit, p_unsigned);
1629         PRINT_ATTRf(pinned, p_unsigned);
1630         PRINT_ATTRf(exclusive, p_unsigned);
1631         PRINT_ATTRf(exclude_user, p_unsigned);
1632         PRINT_ATTRf(exclude_kernel, p_unsigned);
1633         PRINT_ATTRf(exclude_hv, p_unsigned);
1634         PRINT_ATTRf(exclude_idle, p_unsigned);
1635         PRINT_ATTRf(mmap, p_unsigned);
1636         PRINT_ATTRf(comm, p_unsigned);
1637         PRINT_ATTRf(freq, p_unsigned);
1638         PRINT_ATTRf(inherit_stat, p_unsigned);
1639         PRINT_ATTRf(enable_on_exec, p_unsigned);
1640         PRINT_ATTRf(task, p_unsigned);
1641         PRINT_ATTRf(watermark, p_unsigned);
1642         PRINT_ATTRf(precise_ip, p_unsigned);
1643         PRINT_ATTRf(mmap_data, p_unsigned);
1644         PRINT_ATTRf(sample_id_all, p_unsigned);
1645         PRINT_ATTRf(exclude_host, p_unsigned);
1646         PRINT_ATTRf(exclude_guest, p_unsigned);
1647         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1648         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1649         PRINT_ATTRf(mmap2, p_unsigned);
1650         PRINT_ATTRf(comm_exec, p_unsigned);
1651         PRINT_ATTRf(use_clockid, p_unsigned);
1652         PRINT_ATTRf(context_switch, p_unsigned);
1653         PRINT_ATTRf(write_backward, p_unsigned);
1654         PRINT_ATTRf(namespaces, p_unsigned);
1655
1656         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1657         PRINT_ATTRf(bp_type, p_unsigned);
1658         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1659         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1660         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1661         PRINT_ATTRf(sample_regs_user, p_hex);
1662         PRINT_ATTRf(sample_stack_user, p_unsigned);
1663         PRINT_ATTRf(clockid, p_signed);
1664         PRINT_ATTRf(sample_regs_intr, p_hex);
1665         PRINT_ATTRf(aux_watermark, p_unsigned);
1666         PRINT_ATTRf(sample_max_stack, p_unsigned);
1667
1668         return ret;
1669 }
1670
1671 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1672                                 void *priv __maybe_unused)
1673 {
1674         return fprintf(fp, "  %-32s %s\n", name, val);
1675 }
1676
1677 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1678                                   int nr_cpus, int nr_threads,
1679                                   int thread_idx)
1680 {
1681         for (int cpu = 0; cpu < nr_cpus; cpu++)
1682                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1683                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1684 }
1685
1686 static int update_fds(struct perf_evsel *evsel,
1687                       int nr_cpus, int cpu_idx,
1688                       int nr_threads, int thread_idx)
1689 {
1690         struct perf_evsel *pos;
1691
1692         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1693                 return -EINVAL;
1694
1695         evlist__for_each_entry(evsel->evlist, pos) {
1696                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1697
1698                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1699
1700                 /*
1701                  * Since fds for next evsel has not been created,
1702                  * there is no need to iterate whole event list.
1703                  */
1704                 if (pos == evsel)
1705                         break;
1706         }
1707         return 0;
1708 }
1709
1710 static bool ignore_missing_thread(struct perf_evsel *evsel,
1711                                   int nr_cpus, int cpu,
1712                                   struct thread_map *threads,
1713                                   int thread, int err)
1714 {
1715         pid_t ignore_pid = thread_map__pid(threads, thread);
1716
1717         if (!evsel->ignore_missing_thread)
1718                 return false;
1719
1720         /* The system wide setup does not work with threads. */
1721         if (evsel->system_wide)
1722                 return false;
1723
1724         /* The -ESRCH is perf event syscall errno for pid's not found. */
1725         if (err != -ESRCH)
1726                 return false;
1727
1728         /* If there's only one thread, let it fail. */
1729         if (threads->nr == 1)
1730                 return false;
1731
1732         /*
1733          * We should remove fd for missing_thread first
1734          * because thread_map__remove() will decrease threads->nr.
1735          */
1736         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1737                 return false;
1738
1739         if (thread_map__remove(threads, thread))
1740                 return false;
1741
1742         pr_warning("WARNING: Ignored open failure for pid %d\n",
1743                    ignore_pid);
1744         return true;
1745 }
1746
1747 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1748                      struct thread_map *threads)
1749 {
1750         int cpu, thread, nthreads;
1751         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1752         int pid = -1, err;
1753         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1754
1755         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1756                 return -EINVAL;
1757
1758         if (cpus == NULL) {
1759                 static struct cpu_map *empty_cpu_map;
1760
1761                 if (empty_cpu_map == NULL) {
1762                         empty_cpu_map = cpu_map__dummy_new();
1763                         if (empty_cpu_map == NULL)
1764                                 return -ENOMEM;
1765                 }
1766
1767                 cpus = empty_cpu_map;
1768         }
1769
1770         if (threads == NULL) {
1771                 static struct thread_map *empty_thread_map;
1772
1773                 if (empty_thread_map == NULL) {
1774                         empty_thread_map = thread_map__new_by_tid(-1);
1775                         if (empty_thread_map == NULL)
1776                                 return -ENOMEM;
1777                 }
1778
1779                 threads = empty_thread_map;
1780         }
1781
1782         if (evsel->system_wide)
1783                 nthreads = 1;
1784         else
1785                 nthreads = threads->nr;
1786
1787         if (evsel->fd == NULL &&
1788             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1789                 return -ENOMEM;
1790
1791         if (evsel->cgrp) {
1792                 flags |= PERF_FLAG_PID_CGROUP;
1793                 pid = evsel->cgrp->fd;
1794         }
1795
1796 fallback_missing_features:
1797         if (perf_missing_features.clockid_wrong)
1798                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1799         if (perf_missing_features.clockid) {
1800                 evsel->attr.use_clockid = 0;
1801                 evsel->attr.clockid = 0;
1802         }
1803         if (perf_missing_features.cloexec)
1804                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1805         if (perf_missing_features.mmap2)
1806                 evsel->attr.mmap2 = 0;
1807         if (perf_missing_features.exclude_guest)
1808                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1809         if (perf_missing_features.lbr_flags)
1810                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1811                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1812         if (perf_missing_features.group_read && evsel->attr.inherit)
1813                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1814 retry_sample_id:
1815         if (perf_missing_features.sample_id_all)
1816                 evsel->attr.sample_id_all = 0;
1817
1818         if (verbose >= 2) {
1819                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1820                 fprintf(stderr, "perf_event_attr:\n");
1821                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1822                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1823         }
1824
1825         for (cpu = 0; cpu < cpus->nr; cpu++) {
1826
1827                 for (thread = 0; thread < nthreads; thread++) {
1828                         int fd, group_fd;
1829
1830                         if (!evsel->cgrp && !evsel->system_wide)
1831                                 pid = thread_map__pid(threads, thread);
1832
1833                         group_fd = get_group_fd(evsel, cpu, thread);
1834 retry_open:
1835                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1836                                   pid, cpus->map[cpu], group_fd, flags);
1837
1838                         test_attr__ready();
1839
1840                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1841                                                  group_fd, flags);
1842
1843                         FD(evsel, cpu, thread) = fd;
1844
1845                         if (fd < 0) {
1846                                 err = -errno;
1847
1848                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1849                                         /*
1850                                          * We just removed 1 thread, so take a step
1851                                          * back on thread index and lower the upper
1852                                          * nthreads limit.
1853                                          */
1854                                         nthreads--;
1855                                         thread--;
1856
1857                                         /* ... and pretend like nothing have happened. */
1858                                         err = 0;
1859                                         continue;
1860                                 }
1861
1862                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1863                                           err);
1864                                 goto try_fallback;
1865                         }
1866
1867                         pr_debug2(" = %d\n", fd);
1868
1869                         if (evsel->bpf_fd >= 0) {
1870                                 int evt_fd = fd;
1871                                 int bpf_fd = evsel->bpf_fd;
1872
1873                                 err = ioctl(evt_fd,
1874                                             PERF_EVENT_IOC_SET_BPF,
1875                                             bpf_fd);
1876                                 if (err && errno != EEXIST) {
1877                                         pr_err("failed to attach bpf fd %d: %s\n",
1878                                                bpf_fd, strerror(errno));
1879                                         err = -EINVAL;
1880                                         goto out_close;
1881                                 }
1882                         }
1883
1884                         set_rlimit = NO_CHANGE;
1885
1886                         /*
1887                          * If we succeeded but had to kill clockid, fail and
1888                          * have perf_evsel__open_strerror() print us a nice
1889                          * error.
1890                          */
1891                         if (perf_missing_features.clockid ||
1892                             perf_missing_features.clockid_wrong) {
1893                                 err = -EINVAL;
1894                                 goto out_close;
1895                         }
1896                 }
1897         }
1898
1899         return 0;
1900
1901 try_fallback:
1902         /*
1903          * perf stat needs between 5 and 22 fds per CPU. When we run out
1904          * of them try to increase the limits.
1905          */
1906         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1907                 struct rlimit l;
1908                 int old_errno = errno;
1909
1910                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1911                         if (set_rlimit == NO_CHANGE)
1912                                 l.rlim_cur = l.rlim_max;
1913                         else {
1914                                 l.rlim_cur = l.rlim_max + 1000;
1915                                 l.rlim_max = l.rlim_cur;
1916                         }
1917                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1918                                 set_rlimit++;
1919                                 errno = old_errno;
1920                                 goto retry_open;
1921                         }
1922                 }
1923                 errno = old_errno;
1924         }
1925
1926         if (err != -EINVAL || cpu > 0 || thread > 0)
1927                 goto out_close;
1928
1929         /*
1930          * Must probe features in the order they were added to the
1931          * perf_event_attr interface.
1932          */
1933         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1934                 perf_missing_features.write_backward = true;
1935                 pr_debug2("switching off write_backward\n");
1936                 goto out_close;
1937         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1938                 perf_missing_features.clockid_wrong = true;
1939                 pr_debug2("switching off clockid\n");
1940                 goto fallback_missing_features;
1941         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1942                 perf_missing_features.clockid = true;
1943                 pr_debug2("switching off use_clockid\n");
1944                 goto fallback_missing_features;
1945         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1946                 perf_missing_features.cloexec = true;
1947                 pr_debug2("switching off cloexec flag\n");
1948                 goto fallback_missing_features;
1949         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1950                 perf_missing_features.mmap2 = true;
1951                 pr_debug2("switching off mmap2\n");
1952                 goto fallback_missing_features;
1953         } else if (!perf_missing_features.exclude_guest &&
1954                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1955                 perf_missing_features.exclude_guest = true;
1956                 pr_debug2("switching off exclude_guest, exclude_host\n");
1957                 goto fallback_missing_features;
1958         } else if (!perf_missing_features.sample_id_all) {
1959                 perf_missing_features.sample_id_all = true;
1960                 pr_debug2("switching off sample_id_all\n");
1961                 goto retry_sample_id;
1962         } else if (!perf_missing_features.lbr_flags &&
1963                         (evsel->attr.branch_sample_type &
1964                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1965                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1966                 perf_missing_features.lbr_flags = true;
1967                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1968                 goto fallback_missing_features;
1969         } else if (!perf_missing_features.group_read &&
1970                     evsel->attr.inherit &&
1971                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1972                    perf_evsel__is_group_leader(evsel)) {
1973                 perf_missing_features.group_read = true;
1974                 pr_debug2("switching off group read\n");
1975                 goto fallback_missing_features;
1976         }
1977 out_close:
1978         if (err)
1979                 threads->err_thread = thread;
1980
1981         do {
1982                 while (--thread >= 0) {
1983                         close(FD(evsel, cpu, thread));
1984                         FD(evsel, cpu, thread) = -1;
1985                 }
1986                 thread = nthreads;
1987         } while (--cpu >= 0);
1988         return err;
1989 }
1990
1991 void perf_evsel__close(struct perf_evsel *evsel)
1992 {
1993         if (evsel->fd == NULL)
1994                 return;
1995
1996         perf_evsel__close_fd(evsel);
1997         perf_evsel__free_fd(evsel);
1998 }
1999
2000 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2001                              struct cpu_map *cpus)
2002 {
2003         return perf_evsel__open(evsel, cpus, NULL);
2004 }
2005
2006 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2007                                 struct thread_map *threads)
2008 {
2009         return perf_evsel__open(evsel, NULL, threads);
2010 }
2011
2012 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2013                                        const union perf_event *event,
2014                                        struct perf_sample *sample)
2015 {
2016         u64 type = evsel->attr.sample_type;
2017         const u64 *array = event->sample.array;
2018         bool swapped = evsel->needs_swap;
2019         union u64_swap u;
2020
2021         array += ((event->header.size -
2022                    sizeof(event->header)) / sizeof(u64)) - 1;
2023
2024         if (type & PERF_SAMPLE_IDENTIFIER) {
2025                 sample->id = *array;
2026                 array--;
2027         }
2028
2029         if (type & PERF_SAMPLE_CPU) {
2030                 u.val64 = *array;
2031                 if (swapped) {
2032                         /* undo swap of u64, then swap on individual u32s */
2033                         u.val64 = bswap_64(u.val64);
2034                         u.val32[0] = bswap_32(u.val32[0]);
2035                 }
2036
2037                 sample->cpu = u.val32[0];
2038                 array--;
2039         }
2040
2041         if (type & PERF_SAMPLE_STREAM_ID) {
2042                 sample->stream_id = *array;
2043                 array--;
2044         }
2045
2046         if (type & PERF_SAMPLE_ID) {
2047                 sample->id = *array;
2048                 array--;
2049         }
2050
2051         if (type & PERF_SAMPLE_TIME) {
2052                 sample->time = *array;
2053                 array--;
2054         }
2055
2056         if (type & PERF_SAMPLE_TID) {
2057                 u.val64 = *array;
2058                 if (swapped) {
2059                         /* undo swap of u64, then swap on individual u32s */
2060                         u.val64 = bswap_64(u.val64);
2061                         u.val32[0] = bswap_32(u.val32[0]);
2062                         u.val32[1] = bswap_32(u.val32[1]);
2063                 }
2064
2065                 sample->pid = u.val32[0];
2066                 sample->tid = u.val32[1];
2067                 array--;
2068         }
2069
2070         return 0;
2071 }
2072
2073 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2074                             u64 size)
2075 {
2076         return size > max_size || offset + size > endp;
2077 }
2078
2079 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2080         do {                                                            \
2081                 if (overflow(endp, (max_size), (offset), (size)))       \
2082                         return -EFAULT;                                 \
2083         } while (0)
2084
2085 #define OVERFLOW_CHECK_u64(offset) \
2086         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2087
2088 static int
2089 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2090 {
2091         /*
2092          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2093          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2094          * check the format does not go past the end of the event.
2095          */
2096         if (sample_size + sizeof(event->header) > event->header.size)
2097                 return -EFAULT;
2098
2099         return 0;
2100 }
2101
2102 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2103                              struct perf_sample *data)
2104 {
2105         u64 type = evsel->attr.sample_type;
2106         bool swapped = evsel->needs_swap;
2107         const u64 *array;
2108         u16 max_size = event->header.size;
2109         const void *endp = (void *)event + max_size;
2110         u64 sz;
2111
2112         /*
2113          * used for cross-endian analysis. See git commit 65014ab3
2114          * for why this goofiness is needed.
2115          */
2116         union u64_swap u;
2117
2118         memset(data, 0, sizeof(*data));
2119         data->cpu = data->pid = data->tid = -1;
2120         data->stream_id = data->id = data->time = -1ULL;
2121         data->period = evsel->attr.sample_period;
2122         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2123         data->misc    = event->header.misc;
2124         data->id = -1ULL;
2125         data->data_src = PERF_MEM_DATA_SRC_NONE;
2126
2127         if (event->header.type != PERF_RECORD_SAMPLE) {
2128                 if (!evsel->attr.sample_id_all)
2129                         return 0;
2130                 return perf_evsel__parse_id_sample(evsel, event, data);
2131         }
2132
2133         array = event->sample.array;
2134
2135         if (perf_event__check_size(event, evsel->sample_size))
2136                 return -EFAULT;
2137
2138         if (type & PERF_SAMPLE_IDENTIFIER) {
2139                 data->id = *array;
2140                 array++;
2141         }
2142
2143         if (type & PERF_SAMPLE_IP) {
2144                 data->ip = *array;
2145                 array++;
2146         }
2147
2148         if (type & PERF_SAMPLE_TID) {
2149                 u.val64 = *array;
2150                 if (swapped) {
2151                         /* undo swap of u64, then swap on individual u32s */
2152                         u.val64 = bswap_64(u.val64);
2153                         u.val32[0] = bswap_32(u.val32[0]);
2154                         u.val32[1] = bswap_32(u.val32[1]);
2155                 }
2156
2157                 data->pid = u.val32[0];
2158                 data->tid = u.val32[1];
2159                 array++;
2160         }
2161
2162         if (type & PERF_SAMPLE_TIME) {
2163                 data->time = *array;
2164                 array++;
2165         }
2166
2167         if (type & PERF_SAMPLE_ADDR) {
2168                 data->addr = *array;
2169                 array++;
2170         }
2171
2172         if (type & PERF_SAMPLE_ID) {
2173                 data->id = *array;
2174                 array++;
2175         }
2176
2177         if (type & PERF_SAMPLE_STREAM_ID) {
2178                 data->stream_id = *array;
2179                 array++;
2180         }
2181
2182         if (type & PERF_SAMPLE_CPU) {
2183
2184                 u.val64 = *array;
2185                 if (swapped) {
2186                         /* undo swap of u64, then swap on individual u32s */
2187                         u.val64 = bswap_64(u.val64);
2188                         u.val32[0] = bswap_32(u.val32[0]);
2189                 }
2190
2191                 data->cpu = u.val32[0];
2192                 array++;
2193         }
2194
2195         if (type & PERF_SAMPLE_PERIOD) {
2196                 data->period = *array;
2197                 array++;
2198         }
2199
2200         if (type & PERF_SAMPLE_READ) {
2201                 u64 read_format = evsel->attr.read_format;
2202
2203                 OVERFLOW_CHECK_u64(array);
2204                 if (read_format & PERF_FORMAT_GROUP)
2205                         data->read.group.nr = *array;
2206                 else
2207                         data->read.one.value = *array;
2208
2209                 array++;
2210
2211                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2212                         OVERFLOW_CHECK_u64(array);
2213                         data->read.time_enabled = *array;
2214                         array++;
2215                 }
2216
2217                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2218                         OVERFLOW_CHECK_u64(array);
2219                         data->read.time_running = *array;
2220                         array++;
2221                 }
2222
2223                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2224                 if (read_format & PERF_FORMAT_GROUP) {
2225                         const u64 max_group_nr = UINT64_MAX /
2226                                         sizeof(struct sample_read_value);
2227
2228                         if (data->read.group.nr > max_group_nr)
2229                                 return -EFAULT;
2230                         sz = data->read.group.nr *
2231                              sizeof(struct sample_read_value);
2232                         OVERFLOW_CHECK(array, sz, max_size);
2233                         data->read.group.values =
2234                                         (struct sample_read_value *)array;
2235                         array = (void *)array + sz;
2236                 } else {
2237                         OVERFLOW_CHECK_u64(array);
2238                         data->read.one.id = *array;
2239                         array++;
2240                 }
2241         }
2242
2243         if (evsel__has_callchain(evsel)) {
2244                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2245
2246                 OVERFLOW_CHECK_u64(array);
2247                 data->callchain = (struct ip_callchain *)array++;
2248                 if (data->callchain->nr > max_callchain_nr)
2249                         return -EFAULT;
2250                 sz = data->callchain->nr * sizeof(u64);
2251                 OVERFLOW_CHECK(array, sz, max_size);
2252                 array = (void *)array + sz;
2253         }
2254
2255         if (type & PERF_SAMPLE_RAW) {
2256                 OVERFLOW_CHECK_u64(array);
2257                 u.val64 = *array;
2258
2259                 /*
2260                  * Undo swap of u64, then swap on individual u32s,
2261                  * get the size of the raw area and undo all of the
2262                  * swap. The pevent interface handles endianity by
2263                  * itself.
2264                  */
2265                 if (swapped) {
2266                         u.val64 = bswap_64(u.val64);
2267                         u.val32[0] = bswap_32(u.val32[0]);
2268                         u.val32[1] = bswap_32(u.val32[1]);
2269                 }
2270                 data->raw_size = u.val32[0];
2271
2272                 /*
2273                  * The raw data is aligned on 64bits including the
2274                  * u32 size, so it's safe to use mem_bswap_64.
2275                  */
2276                 if (swapped)
2277                         mem_bswap_64((void *) array, data->raw_size);
2278
2279                 array = (void *)array + sizeof(u32);
2280
2281                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2282                 data->raw_data = (void *)array;
2283                 array = (void *)array + data->raw_size;
2284         }
2285
2286         if (type & PERF_SAMPLE_BRANCH_STACK) {
2287                 const u64 max_branch_nr = UINT64_MAX /
2288                                           sizeof(struct branch_entry);
2289
2290                 OVERFLOW_CHECK_u64(array);
2291                 data->branch_stack = (struct branch_stack *)array++;
2292
2293                 if (data->branch_stack->nr > max_branch_nr)
2294                         return -EFAULT;
2295                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2296                 OVERFLOW_CHECK(array, sz, max_size);
2297                 array = (void *)array + sz;
2298         }
2299
2300         if (type & PERF_SAMPLE_REGS_USER) {
2301                 OVERFLOW_CHECK_u64(array);
2302                 data->user_regs.abi = *array;
2303                 array++;
2304
2305                 if (data->user_regs.abi) {
2306                         u64 mask = evsel->attr.sample_regs_user;
2307
2308                         sz = hweight_long(mask) * sizeof(u64);
2309                         OVERFLOW_CHECK(array, sz, max_size);
2310                         data->user_regs.mask = mask;
2311                         data->user_regs.regs = (u64 *)array;
2312                         array = (void *)array + sz;
2313                 }
2314         }
2315
2316         if (type & PERF_SAMPLE_STACK_USER) {
2317                 OVERFLOW_CHECK_u64(array);
2318                 sz = *array++;
2319
2320                 data->user_stack.offset = ((char *)(array - 1)
2321                                           - (char *) event);
2322
2323                 if (!sz) {
2324                         data->user_stack.size = 0;
2325                 } else {
2326                         OVERFLOW_CHECK(array, sz, max_size);
2327                         data->user_stack.data = (char *)array;
2328                         array = (void *)array + sz;
2329                         OVERFLOW_CHECK_u64(array);
2330                         data->user_stack.size = *array++;
2331                         if (WARN_ONCE(data->user_stack.size > sz,
2332                                       "user stack dump failure\n"))
2333                                 return -EFAULT;
2334                 }
2335         }
2336
2337         if (type & PERF_SAMPLE_WEIGHT) {
2338                 OVERFLOW_CHECK_u64(array);
2339                 data->weight = *array;
2340                 array++;
2341         }
2342
2343         if (type & PERF_SAMPLE_DATA_SRC) {
2344                 OVERFLOW_CHECK_u64(array);
2345                 data->data_src = *array;
2346                 array++;
2347         }
2348
2349         if (type & PERF_SAMPLE_TRANSACTION) {
2350                 OVERFLOW_CHECK_u64(array);
2351                 data->transaction = *array;
2352                 array++;
2353         }
2354
2355         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2356         if (type & PERF_SAMPLE_REGS_INTR) {
2357                 OVERFLOW_CHECK_u64(array);
2358                 data->intr_regs.abi = *array;
2359                 array++;
2360
2361                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2362                         u64 mask = evsel->attr.sample_regs_intr;
2363
2364                         sz = hweight_long(mask) * sizeof(u64);
2365                         OVERFLOW_CHECK(array, sz, max_size);
2366                         data->intr_regs.mask = mask;
2367                         data->intr_regs.regs = (u64 *)array;
2368                         array = (void *)array + sz;
2369                 }
2370         }
2371
2372         data->phys_addr = 0;
2373         if (type & PERF_SAMPLE_PHYS_ADDR) {
2374                 data->phys_addr = *array;
2375                 array++;
2376         }
2377
2378         return 0;
2379 }
2380
2381 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2382                                        union perf_event *event,
2383                                        u64 *timestamp)
2384 {
2385         u64 type = evsel->attr.sample_type;
2386         const u64 *array;
2387
2388         if (!(type & PERF_SAMPLE_TIME))
2389                 return -1;
2390
2391         if (event->header.type != PERF_RECORD_SAMPLE) {
2392                 struct perf_sample data = {
2393                         .time = -1ULL,
2394                 };
2395
2396                 if (!evsel->attr.sample_id_all)
2397                         return -1;
2398                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2399                         return -1;
2400
2401                 *timestamp = data.time;
2402                 return 0;
2403         }
2404
2405         array = event->sample.array;
2406
2407         if (perf_event__check_size(event, evsel->sample_size))
2408                 return -EFAULT;
2409
2410         if (type & PERF_SAMPLE_IDENTIFIER)
2411                 array++;
2412
2413         if (type & PERF_SAMPLE_IP)
2414                 array++;
2415
2416         if (type & PERF_SAMPLE_TID)
2417                 array++;
2418
2419         if (type & PERF_SAMPLE_TIME)
2420                 *timestamp = *array;
2421
2422         return 0;
2423 }
2424
2425 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2426                                      u64 read_format)
2427 {
2428         size_t sz, result = sizeof(struct sample_event);
2429
2430         if (type & PERF_SAMPLE_IDENTIFIER)
2431                 result += sizeof(u64);
2432
2433         if (type & PERF_SAMPLE_IP)
2434                 result += sizeof(u64);
2435
2436         if (type & PERF_SAMPLE_TID)
2437                 result += sizeof(u64);
2438
2439         if (type & PERF_SAMPLE_TIME)
2440                 result += sizeof(u64);
2441
2442         if (type & PERF_SAMPLE_ADDR)
2443                 result += sizeof(u64);
2444
2445         if (type & PERF_SAMPLE_ID)
2446                 result += sizeof(u64);
2447
2448         if (type & PERF_SAMPLE_STREAM_ID)
2449                 result += sizeof(u64);
2450
2451         if (type & PERF_SAMPLE_CPU)
2452                 result += sizeof(u64);
2453
2454         if (type & PERF_SAMPLE_PERIOD)
2455                 result += sizeof(u64);
2456
2457         if (type & PERF_SAMPLE_READ) {
2458                 result += sizeof(u64);
2459                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2460                         result += sizeof(u64);
2461                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2462                         result += sizeof(u64);
2463                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2464                 if (read_format & PERF_FORMAT_GROUP) {
2465                         sz = sample->read.group.nr *
2466                              sizeof(struct sample_read_value);
2467                         result += sz;
2468                 } else {
2469                         result += sizeof(u64);
2470                 }
2471         }
2472
2473         if (type & PERF_SAMPLE_CALLCHAIN) {
2474                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2475                 result += sz;
2476         }
2477
2478         if (type & PERF_SAMPLE_RAW) {
2479                 result += sizeof(u32);
2480                 result += sample->raw_size;
2481         }
2482
2483         if (type & PERF_SAMPLE_BRANCH_STACK) {
2484                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2485                 sz += sizeof(u64);
2486                 result += sz;
2487         }
2488
2489         if (type & PERF_SAMPLE_REGS_USER) {
2490                 if (sample->user_regs.abi) {
2491                         result += sizeof(u64);
2492                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2493                         result += sz;
2494                 } else {
2495                         result += sizeof(u64);
2496                 }
2497         }
2498
2499         if (type & PERF_SAMPLE_STACK_USER) {
2500                 sz = sample->user_stack.size;
2501                 result += sizeof(u64);
2502                 if (sz) {
2503                         result += sz;
2504                         result += sizeof(u64);
2505                 }
2506         }
2507
2508         if (type & PERF_SAMPLE_WEIGHT)
2509                 result += sizeof(u64);
2510
2511         if (type & PERF_SAMPLE_DATA_SRC)
2512                 result += sizeof(u64);
2513
2514         if (type & PERF_SAMPLE_TRANSACTION)
2515                 result += sizeof(u64);
2516
2517         if (type & PERF_SAMPLE_REGS_INTR) {
2518                 if (sample->intr_regs.abi) {
2519                         result += sizeof(u64);
2520                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2521                         result += sz;
2522                 } else {
2523                         result += sizeof(u64);
2524                 }
2525         }
2526
2527         if (type & PERF_SAMPLE_PHYS_ADDR)
2528                 result += sizeof(u64);
2529
2530         return result;
2531 }
2532
2533 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2534                                   u64 read_format,
2535                                   const struct perf_sample *sample)
2536 {
2537         u64 *array;
2538         size_t sz;
2539         /*
2540          * used for cross-endian analysis. See git commit 65014ab3
2541          * for why this goofiness is needed.
2542          */
2543         union u64_swap u;
2544
2545         array = event->sample.array;
2546
2547         if (type & PERF_SAMPLE_IDENTIFIER) {
2548                 *array = sample->id;
2549                 array++;
2550         }
2551
2552         if (type & PERF_SAMPLE_IP) {
2553                 *array = sample->ip;
2554                 array++;
2555         }
2556
2557         if (type & PERF_SAMPLE_TID) {
2558                 u.val32[0] = sample->pid;
2559                 u.val32[1] = sample->tid;
2560                 *array = u.val64;
2561                 array++;
2562         }
2563
2564         if (type & PERF_SAMPLE_TIME) {
2565                 *array = sample->time;
2566                 array++;
2567         }
2568
2569         if (type & PERF_SAMPLE_ADDR) {
2570                 *array = sample->addr;
2571                 array++;
2572         }
2573
2574         if (type & PERF_SAMPLE_ID) {
2575                 *array = sample->id;
2576                 array++;
2577         }
2578
2579         if (type & PERF_SAMPLE_STREAM_ID) {
2580                 *array = sample->stream_id;
2581                 array++;
2582         }
2583
2584         if (type & PERF_SAMPLE_CPU) {
2585                 u.val32[0] = sample->cpu;
2586                 u.val32[1] = 0;
2587                 *array = u.val64;
2588                 array++;
2589         }
2590
2591         if (type & PERF_SAMPLE_PERIOD) {
2592                 *array = sample->period;
2593                 array++;
2594         }
2595
2596         if (type & PERF_SAMPLE_READ) {
2597                 if (read_format & PERF_FORMAT_GROUP)
2598                         *array = sample->read.group.nr;
2599                 else
2600                         *array = sample->read.one.value;
2601                 array++;
2602
2603                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2604                         *array = sample->read.time_enabled;
2605                         array++;
2606                 }
2607
2608                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2609                         *array = sample->read.time_running;
2610                         array++;
2611                 }
2612
2613                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2614                 if (read_format & PERF_FORMAT_GROUP) {
2615                         sz = sample->read.group.nr *
2616                              sizeof(struct sample_read_value);
2617                         memcpy(array, sample->read.group.values, sz);
2618                         array = (void *)array + sz;
2619                 } else {
2620                         *array = sample->read.one.id;
2621                         array++;
2622                 }
2623         }
2624
2625         if (type & PERF_SAMPLE_CALLCHAIN) {
2626                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2627                 memcpy(array, sample->callchain, sz);
2628                 array = (void *)array + sz;
2629         }
2630
2631         if (type & PERF_SAMPLE_RAW) {
2632                 u.val32[0] = sample->raw_size;
2633                 *array = u.val64;
2634                 array = (void *)array + sizeof(u32);
2635
2636                 memcpy(array, sample->raw_data, sample->raw_size);
2637                 array = (void *)array + sample->raw_size;
2638         }
2639
2640         if (type & PERF_SAMPLE_BRANCH_STACK) {
2641                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2642                 sz += sizeof(u64);
2643                 memcpy(array, sample->branch_stack, sz);
2644                 array = (void *)array + sz;
2645         }
2646
2647         if (type & PERF_SAMPLE_REGS_USER) {
2648                 if (sample->user_regs.abi) {
2649                         *array++ = sample->user_regs.abi;
2650                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2651                         memcpy(array, sample->user_regs.regs, sz);
2652                         array = (void *)array + sz;
2653                 } else {
2654                         *array++ = 0;
2655                 }
2656         }
2657
2658         if (type & PERF_SAMPLE_STACK_USER) {
2659                 sz = sample->user_stack.size;
2660                 *array++ = sz;
2661                 if (sz) {
2662                         memcpy(array, sample->user_stack.data, sz);
2663                         array = (void *)array + sz;
2664                         *array++ = sz;
2665                 }
2666         }
2667
2668         if (type & PERF_SAMPLE_WEIGHT) {
2669                 *array = sample->weight;
2670                 array++;
2671         }
2672
2673         if (type & PERF_SAMPLE_DATA_SRC) {
2674                 *array = sample->data_src;
2675                 array++;
2676         }
2677
2678         if (type & PERF_SAMPLE_TRANSACTION) {
2679                 *array = sample->transaction;
2680                 array++;
2681         }
2682
2683         if (type & PERF_SAMPLE_REGS_INTR) {
2684                 if (sample->intr_regs.abi) {
2685                         *array++ = sample->intr_regs.abi;
2686                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2687                         memcpy(array, sample->intr_regs.regs, sz);
2688                         array = (void *)array + sz;
2689                 } else {
2690                         *array++ = 0;
2691                 }
2692         }
2693
2694         if (type & PERF_SAMPLE_PHYS_ADDR) {
2695                 *array = sample->phys_addr;
2696                 array++;
2697         }
2698
2699         return 0;
2700 }
2701
2702 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2703 {
2704         return tep_find_field(evsel->tp_format, name);
2705 }
2706
2707 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2708                          const char *name)
2709 {
2710         struct tep_format_field *field = perf_evsel__field(evsel, name);
2711         int offset;
2712
2713         if (!field)
2714                 return NULL;
2715
2716         offset = field->offset;
2717
2718         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2719                 offset = *(int *)(sample->raw_data + field->offset);
2720                 offset &= 0xffff;
2721         }
2722
2723         return sample->raw_data + offset;
2724 }
2725
2726 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2727                          bool needs_swap)
2728 {
2729         u64 value;
2730         void *ptr = sample->raw_data + field->offset;
2731
2732         switch (field->size) {
2733         case 1:
2734                 return *(u8 *)ptr;
2735         case 2:
2736                 value = *(u16 *)ptr;
2737                 break;
2738         case 4:
2739                 value = *(u32 *)ptr;
2740                 break;
2741         case 8:
2742                 memcpy(&value, ptr, sizeof(u64));
2743                 break;
2744         default:
2745                 return 0;
2746         }
2747
2748         if (!needs_swap)
2749                 return value;
2750
2751         switch (field->size) {
2752         case 2:
2753                 return bswap_16(value);
2754         case 4:
2755                 return bswap_32(value);
2756         case 8:
2757                 return bswap_64(value);
2758         default:
2759                 return 0;
2760         }
2761
2762         return 0;
2763 }
2764
2765 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2766                        const char *name)
2767 {
2768         struct tep_format_field *field = perf_evsel__field(evsel, name);
2769
2770         if (!field)
2771                 return 0;
2772
2773         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2774 }
2775
2776 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2777                           char *msg, size_t msgsize)
2778 {
2779         int paranoid;
2780
2781         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2782             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2783             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2784                 /*
2785                  * If it's cycles then fall back to hrtimer based
2786                  * cpu-clock-tick sw counter, which is always available even if
2787                  * no PMU support.
2788                  *
2789                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2790                  * b0a873e).
2791                  */
2792                 scnprintf(msg, msgsize, "%s",
2793 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2794
2795                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2796                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2797
2798                 zfree(&evsel->name);
2799                 return true;
2800         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2801                    (paranoid = perf_event_paranoid()) > 1) {
2802                 const char *name = perf_evsel__name(evsel);
2803                 char *new_name;
2804                 const char *sep = ":";
2805
2806                 /* Is there already the separator in the name. */
2807                 if (strchr(name, '/') ||
2808                     strchr(name, ':'))
2809                         sep = "";
2810
2811                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2812                         return false;
2813
2814                 if (evsel->name)
2815                         free(evsel->name);
2816                 evsel->name = new_name;
2817                 scnprintf(msg, msgsize,
2818 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2819                 evsel->attr.exclude_kernel = 1;
2820
2821                 return true;
2822         }
2823
2824         return false;
2825 }
2826
2827 static bool find_process(const char *name)
2828 {
2829         size_t len = strlen(name);
2830         DIR *dir;
2831         struct dirent *d;
2832         int ret = -1;
2833
2834         dir = opendir(procfs__mountpoint());
2835         if (!dir)
2836                 return false;
2837
2838         /* Walk through the directory. */
2839         while (ret && (d = readdir(dir)) != NULL) {
2840                 char path[PATH_MAX];
2841                 char *data;
2842                 size_t size;
2843
2844                 if ((d->d_type != DT_DIR) ||
2845                      !strcmp(".", d->d_name) ||
2846                      !strcmp("..", d->d_name))
2847                         continue;
2848
2849                 scnprintf(path, sizeof(path), "%s/%s/comm",
2850                           procfs__mountpoint(), d->d_name);
2851
2852                 if (filename__read_str(path, &data, &size))
2853                         continue;
2854
2855                 ret = strncmp(name, data, len);
2856                 free(data);
2857         }
2858
2859         closedir(dir);
2860         return ret ? false : true;
2861 }
2862
2863 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2864                               int err, char *msg, size_t size)
2865 {
2866         char sbuf[STRERR_BUFSIZE];
2867         int printed = 0;
2868
2869         switch (err) {
2870         case EPERM:
2871         case EACCES:
2872                 if (err == EPERM)
2873                         printed = scnprintf(msg, size,
2874                                 "No permission to enable %s event.\n\n",
2875                                 perf_evsel__name(evsel));
2876
2877                 return scnprintf(msg + printed, size - printed,
2878                  "You may not have permission to collect %sstats.\n\n"
2879                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2880                  "which controls use of the performance events system by\n"
2881                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2882                  "The current value is %d:\n\n"
2883                  "  -1: Allow use of (almost) all events by all users\n"
2884                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2885                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2886                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2887                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2888                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2889                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2890                  "      kernel.perf_event_paranoid = -1\n" ,
2891                                  target->system_wide ? "system-wide " : "",
2892                                  perf_event_paranoid());
2893         case ENOENT:
2894                 return scnprintf(msg, size, "The %s event is not supported.",
2895                                  perf_evsel__name(evsel));
2896         case EMFILE:
2897                 return scnprintf(msg, size, "%s",
2898                          "Too many events are opened.\n"
2899                          "Probably the maximum number of open file descriptors has been reached.\n"
2900                          "Hint: Try again after reducing the number of events.\n"
2901                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2902         case ENOMEM:
2903                 if (evsel__has_callchain(evsel) &&
2904                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2905                         return scnprintf(msg, size,
2906                                          "Not enough memory to setup event with callchain.\n"
2907                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2908                                          "Hint: Current value: %d", sysctl__max_stack());
2909                 break;
2910         case ENODEV:
2911                 if (target->cpu_list)
2912                         return scnprintf(msg, size, "%s",
2913          "No such device - did you specify an out-of-range profile CPU?");
2914                 break;
2915         case EOPNOTSUPP:
2916                 if (evsel->attr.sample_period != 0)
2917                         return scnprintf(msg, size,
2918         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2919                                          perf_evsel__name(evsel));
2920                 if (evsel->attr.precise_ip)
2921                         return scnprintf(msg, size, "%s",
2922         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2923 #if defined(__i386__) || defined(__x86_64__)
2924                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2925                         return scnprintf(msg, size, "%s",
2926         "No hardware sampling interrupt available.\n");
2927 #endif
2928                 break;
2929         case EBUSY:
2930                 if (find_process("oprofiled"))
2931                         return scnprintf(msg, size,
2932         "The PMU counters are busy/taken by another profiler.\n"
2933         "We found oprofile daemon running, please stop it and try again.");
2934                 break;
2935         case EINVAL:
2936                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2937                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2938                 if (perf_missing_features.clockid)
2939                         return scnprintf(msg, size, "clockid feature not supported.");
2940                 if (perf_missing_features.clockid_wrong)
2941                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2942                 break;
2943         default:
2944                 break;
2945         }
2946
2947         return scnprintf(msg, size,
2948         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2949         "/bin/dmesg | grep -i perf may provide additional information.\n",
2950                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2951                          perf_evsel__name(evsel));
2952 }
2953
2954 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2955 {
2956         if (evsel && evsel->evlist)
2957                 return evsel->evlist->env;
2958         return NULL;
2959 }
2960
2961 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
2962 {
2963         int cpu, thread;
2964
2965         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
2966                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
2967                      thread++) {
2968                         int fd = FD(evsel, cpu, thread);
2969
2970                         if (perf_evlist__id_add_fd(evlist, evsel,
2971                                                    cpu, thread, fd) < 0)
2972                                 return -1;
2973                 }
2974         }
2975
2976         return 0;
2977 }
2978
2979 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
2980 {
2981         struct cpu_map *cpus = evsel->cpus;
2982         struct thread_map *threads = evsel->threads;
2983
2984         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
2985                 return -ENOMEM;
2986
2987         return store_evsel_ids(evsel, evlist);
2988 }