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