tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->max_events  = ULONG_MAX;
 236         evsel->evlist      = NULL;
 237         evsel->bpf_fd      = -1;
 238         INIT_LIST_HEAD(&evsel->node);
 239         INIT_LIST_HEAD(&evsel->config_terms);
 240         perf_evsel__object.init(evsel);
 241         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 242         perf_evsel__calc_id_pos(evsel);
 243         evsel->cmdline_group_boundary = false;
 244         evsel->metric_expr   = NULL;
 245         evsel->metric_name   = NULL;
 246         evsel->metric_events = NULL;
 247         evsel->collect_stat  = false;
 248         evsel->pmu_name      = NULL;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (!evsel)
 256                 return NULL;
 257         perf_evsel__init(evsel, attr, idx);
 258
 259         if (perf_evsel__is_bpf_output(evsel)) {
 260                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 261                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 262                 evsel->attr.sample_period = 1;
 263         }
 264
 265         if (perf_evsel__is_clock(evsel)) {
 266                 /*
 267                  * The evsel->unit points to static alias->unit
 268                  * so it's ok to use static string in here.
 269                  */
 270                 static const char *unit = "msec";
 271
 272                 evsel->unit = unit;
 273                 evsel->scale = 1e-6;
 274         }
 275
 276         return evsel;
 277 }
 278
 279 static bool perf_event_can_profile_kernel(void)
 280 {
 281         return geteuid() == 0 || perf_event_paranoid() == -1;
 282 }
 283
 284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 285 {
 286         struct perf_event_attr attr = {
 287                 .type   = PERF_TYPE_HARDWARE,
 288                 .config = PERF_COUNT_HW_CPU_CYCLES,
 289                 .exclude_kernel = !perf_event_can_profile_kernel(),
 290         };
 291         struct perf_evsel *evsel;
 292
 293         event_attr_init(&attr);
 294
 295         if (!precise)
 296                 goto new_event;
 297         /*
 298          * Unnamed union member, not supported as struct member named
 299          * initializer in older compilers such as gcc 4.4.7
 300          *
 301          * Just for probing the precise_ip:
 302          */
 303         attr.sample_period = 1;
 304
 305         perf_event_attr__set_max_precise_ip(&attr);
 306         /*
 307          * Now let the usual logic to set up the perf_event_attr defaults
 308          * to kick in when we return and before perf_evsel__open() is called.
 309          */
 310         attr.sample_period = 0;
 311 new_event:
 312         evsel = perf_evsel__new(&attr);
 313         if (evsel == NULL)
 314                 goto out;
 315
 316         /* use asprintf() because free(evsel) assumes name is allocated */
 317         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 318                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 319                      attr.exclude_kernel ? "u" : "",
 320                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 321                 goto error_free;
 322 out:
 323         return evsel;
 324 error_free:
 325         perf_evsel__delete(evsel);
 326         evsel = NULL;
 327         goto out;
 328 }
 329
 330 /*
 331  * Returns pointer with encoded error via <linux/err.h> interface.
 332  */
 333 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 334 {
 335         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 336         int err = -ENOMEM;
 337
 338         if (evsel == NULL) {
 339                 goto out_err;
 340         } else {
 341                 struct perf_event_attr attr = {
 342                         .type          = PERF_TYPE_TRACEPOINT,
 343                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 344                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 345                 };
 346
 347                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 348                         goto out_free;
 349
 350                 evsel->tp_format = trace_event__tp_format(sys, name);
 351                 if (IS_ERR(evsel->tp_format)) {
 352                         err = PTR_ERR(evsel->tp_format);
 353                         goto out_free;
 354                 }
 355
 356                 event_attr_init(&attr);
 357                 attr.config = evsel->tp_format->id;
 358                 attr.sample_period = 1;
 359                 perf_evsel__init(evsel, &attr, idx);
 360         }
 361
 362         return evsel;
 363
 364 out_free:
 365         zfree(&evsel->name);
 366         free(evsel);
 367 out_err:
 368         return ERR_PTR(err);
 369 }
 370
 371 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 372         "cycles",
 373         "instructions",
 374         "cache-references",
 375         "cache-misses",
 376         "branches",
 377         "branch-misses",
 378         "bus-cycles",
 379         "stalled-cycles-frontend",
 380         "stalled-cycles-backend",
 381         "ref-cycles",
 382 };
 383
 384 static const char *__perf_evsel__hw_name(u64 config)
 385 {
 386         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 387                 return perf_evsel__hw_names[config];
 388
 389         return "unknown-hardware";
 390 }
 391
 392 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 393 {
 394         int colon = 0, r = 0;
 395         struct perf_event_attr *attr = &evsel->attr;
 396         bool exclude_guest_default = false;
 397
 398 #define MOD_PRINT(context, mod) do {                                    \
 399                 if (!attr->exclude_##context) {                         \
 400                         if (!colon) colon = ++r;                        \
 401                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 402                 } } while(0)
 403
 404         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 405                 MOD_PRINT(kernel, 'k');
 406                 MOD_PRINT(user, 'u');
 407                 MOD_PRINT(hv, 'h');
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->precise_ip) {
 412                 if (!colon)
 413                         colon = ++r;
 414                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 415                 exclude_guest_default = true;
 416         }
 417
 418         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 419                 MOD_PRINT(host, 'H');
 420                 MOD_PRINT(guest, 'G');
 421         }
 422 #undef MOD_PRINT
 423         if (colon)
 424                 bf[colon - 1] = ':';
 425         return r;
 426 }
 427
 428 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 429 {
 430         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 431         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 432 }
 433
 434 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 435         "cpu-clock",
 436         "task-clock",
 437         "page-faults",
 438         "context-switches",
 439         "cpu-migrations",
 440         "minor-faults",
 441         "major-faults",
 442         "alignment-faults",
 443         "emulation-faults",
 444         "dummy",
 445 };
 446
 447 static const char *__perf_evsel__sw_name(u64 config)
 448 {
 449         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 450                 return perf_evsel__sw_names[config];
 451         return "unknown-software";
 452 }
 453
 454 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 455 {
 456         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 457         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 458 }
 459
 460 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 461 {
 462         int r;
 463
 464         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 465
 466         if (type & HW_BREAKPOINT_R)
 467                 r += scnprintf(bf + r, size - r, "r");
 468
 469         if (type & HW_BREAKPOINT_W)
 470                 r += scnprintf(bf + r, size - r, "w");
 471
 472         if (type & HW_BREAKPOINT_X)
 473                 r += scnprintf(bf + r, size - r, "x");
 474
 475         return r;
 476 }
 477
 478 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 479 {
 480         struct perf_event_attr *attr = &evsel->attr;
 481         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 482         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 483 }
 484
 485 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 486                                 [PERF_EVSEL__MAX_ALIASES] = {
 487  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 488  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 489  { "LLC",       "L2",                                                   },
 490  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 491  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 492  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 493  { "node",                                                              },
 494 };
 495
 496 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 497                                    [PERF_EVSEL__MAX_ALIASES] = {
 498  { "load",      "loads",        "read",                                 },
 499  { "store",     "stores",       "write",                                },
 500  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 501 };
 502
 503 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 504                                        [PERF_EVSEL__MAX_ALIASES] = {
 505  { "refs",      "Reference",    "ops",          "access",               },
 506  { "misses",    "miss",                                                 },
 507 };
 508
 509 #define C(x)            PERF_COUNT_HW_CACHE_##x
 510 #define CACHE_READ      (1 << C(OP_READ))
 511 #define CACHE_WRITE     (1 << C(OP_WRITE))
 512 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 513 #define COP(x)          (1 << x)
 514
 515 /*
 516  * cache operartion stat
 517  * L1I : Read and prefetch only
 518  * ITLB and BPU : Read-only
 519  */
 520 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 521  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 522  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 523  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(ITLB)]      = (CACHE_READ),
 526  [C(BPU)]       = (CACHE_READ),
 527  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 528 };
 529
 530 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 531 {
 532         if (perf_evsel__hw_cache_stat[type] & COP(op))
 533                 return true;    /* valid */
 534         else
 535                 return false;   /* invalid */
 536 }
 537
 538 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 539                                             char *bf, size_t size)
 540 {
 541         if (result) {
 542                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 543                                  perf_evsel__hw_cache_op[op][0],
 544                                  perf_evsel__hw_cache_result[result][0]);
 545         }
 546
 547         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 548                          perf_evsel__hw_cache_op[op][1]);
 549 }
 550
 551 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 552 {
 553         u8 op, result, type = (config >>  0) & 0xff;
 554         const char *err = "unknown-ext-hardware-cache-type";
 555
 556         if (type >= PERF_COUNT_HW_CACHE_MAX)
 557                 goto out_err;
 558
 559         op = (config >>  8) & 0xff;
 560         err = "unknown-ext-hardware-cache-op";
 561         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 562                 goto out_err;
 563
 564         result = (config >> 16) & 0xff;
 565         err = "unknown-ext-hardware-cache-result";
 566         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 567                 goto out_err;
 568
 569         err = "invalid-cache";
 570         if (!perf_evsel__is_cache_op_valid(type, op))
 571                 goto out_err;
 572
 573         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 574 out_err:
 575         return scnprintf(bf, size, "%s", err);
 576 }
 577
 578 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 579 {
 580         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 581         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 582 }
 583
 584 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 585 {
 586         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 587         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 588 }
 589
 590 const char *perf_evsel__name(struct perf_evsel *evsel)
 591 {
 592         char bf[128];
 593
 594         if (evsel->name)
 595                 return evsel->name;
 596
 597         switch (evsel->attr.type) {
 598         case PERF_TYPE_RAW:
 599                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 600                 break;
 601
 602         case PERF_TYPE_HARDWARE:
 603                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 604                 break;
 605
 606         case PERF_TYPE_HW_CACHE:
 607                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 608                 break;
 609
 610         case PERF_TYPE_SOFTWARE:
 611                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 612                 break;
 613
 614         case PERF_TYPE_TRACEPOINT:
 615                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 616                 break;
 617
 618         case PERF_TYPE_BREAKPOINT:
 619                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 620                 break;
 621
 622         default:
 623                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 624                           evsel->attr.type);
 625                 break;
 626         }
 627
 628         evsel->name = strdup(bf);
 629
 630         return evsel->name ?: "unknown";
 631 }
 632
 633 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 634 {
 635         return evsel->group_name ?: "anon group";
 636 }
 637
 638 /*
 639  * Returns the group details for the specified leader,
 640  * with following rules.
 641  *
 642  *  For record -e '{cycles,instructions}'
 643  *    'anon group { cycles:u, instructions:u }'
 644  *
 645  *  For record -e 'cycles,instructions' and report --group
 646  *    'cycles:u, instructions:u'
 647  */
 648 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 649 {
 650         int ret = 0;
 651         struct perf_evsel *pos;
 652         const char *group_name = perf_evsel__group_name(evsel);
 653
 654         if (!evsel->forced_leader)
 655                 ret = scnprintf(buf, size, "%s { ", group_name);
 656
 657         ret += scnprintf(buf + ret, size - ret, "%s",
 658                          perf_evsel__name(evsel));
 659
 660         for_each_group_member(pos, evsel)
 661                 ret += scnprintf(buf + ret, size - ret, ", %s",
 662                                  perf_evsel__name(pos));
 663
 664         if (!evsel->forced_leader)
 665                 ret += scnprintf(buf + ret, size - ret, " }");
 666
 667         return ret;
 668 }
 669
 670 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 671                                            struct record_opts *opts,
 672                                            struct callchain_param *param)
 673 {
 674         bool function = perf_evsel__is_function_event(evsel);
 675         struct perf_event_attr *attr = &evsel->attr;
 676
 677         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 678
 679         attr->sample_max_stack = param->max_stack;
 680
 681         if (param->record_mode == CALLCHAIN_LBR) {
 682                 if (!opts->branch_stack) {
 683                         if (attr->exclude_user) {
 684                                 pr_warning("LBR callstack option is only available "
 685                                            "to get user callchain information. "
 686                                            "Falling back to framepointers.\n");
 687                         } else {
 688                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 689                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 690                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 691                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 692                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 693                         }
 694                 } else
 695                          pr_warning("Cannot use LBR callstack with branch stack. "
 696                                     "Falling back to framepointers.\n");
 697         }
 698
 699         if (param->record_mode == CALLCHAIN_DWARF) {
 700                 if (!function) {
 701                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 702                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 703                         attr->sample_regs_user |= PERF_REGS_MASK;
 704                         attr->sample_stack_user = param->dump_size;
 705                         attr->exclude_callchain_user = 1;
 706                 } else {
 707                         pr_info("Cannot use DWARF unwind for function trace event,"
 708                                 " falling back to framepointers.\n");
 709                 }
 710         }
 711
 712         if (function) {
 713                 pr_info("Disabling user space callchains for function trace event.\n");
 714                 attr->exclude_callchain_user = 1;
 715         }
 716 }
 717
 718 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 719                                   struct record_opts *opts,
 720                                   struct callchain_param *param)
 721 {
 722         if (param->enabled)
 723                 return __perf_evsel__config_callchain(evsel, opts, param);
 724 }
 725
 726 static void
 727 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 728                             struct callchain_param *param)
 729 {
 730         struct perf_event_attr *attr = &evsel->attr;
 731
 732         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 733         if (param->record_mode == CALLCHAIN_LBR) {
 734                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 735                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 736                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 737         }
 738         if (param->record_mode == CALLCHAIN_DWARF) {
 739                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 740                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 741         }
 742 }
 743
 744 static void apply_config_terms(struct perf_evsel *evsel,
 745                                struct record_opts *opts, bool track)
 746 {
 747         struct perf_evsel_config_term *term;
 748         struct list_head *config_terms = &evsel->config_terms;
 749         struct perf_event_attr *attr = &evsel->attr;
 750         /* callgraph default */
 751         struct callchain_param param = {
 752                 .record_mode = callchain_param.record_mode,
 753         };
 754         u32 dump_size = 0;
 755         int max_stack = 0;
 756         const char *callgraph_buf = NULL;
 757
 758         list_for_each_entry(term, config_terms, list) {
 759                 switch (term->type) {
 760                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 761                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 762                                 attr->sample_period = term->val.period;
 763                                 attr->freq = 0;
 764                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 765                         }
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 768                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 769                                 attr->sample_freq = term->val.freq;
 770                                 attr->freq = 1;
 771                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 772                         }
 773                         break;
 774                 case PERF_EVSEL__CONFIG_TERM_TIME:
 775                         if (term->val.time)
 776                                 perf_evsel__set_sample_bit(evsel, TIME);
 777                         else
 778                                 perf_evsel__reset_sample_bit(evsel, TIME);
 779                         break;
 780                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 781                         callgraph_buf = term->val.callgraph;
 782                         break;
 783                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 784                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 785                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 786                                 parse_branch_str(term->val.branch,
 787                                                  &attr->branch_sample_type);
 788                         } else
 789                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 790                         break;
 791                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 792                         dump_size = term->val.stack_user;
 793                         break;
 794                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 795                         max_stack = term->val.max_stack;
 796                         break;
 797                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 798                         evsel->max_events = term->val.max_events;
 799                         break;
 800                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 801                         /*
 802                          * attr->inherit should has already been set by
 803                          * perf_evsel__config. If user explicitly set
 804                          * inherit using config terms, override global
 805                          * opt->no_inherit setting.
 806                          */
 807                         attr->inherit = term->val.inherit ? 1 : 0;
 808                         break;
 809                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 810                         attr->write_backward = term->val.overwrite ? 1 : 0;
 811                         break;
 812                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 813                         break;
 814                 default:
 815                         break;
 816                 }
 817         }
 818
 819         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 820         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 821                 bool sample_address = false;
 822
 823                 if (max_stack) {
 824                         param.max_stack = max_stack;
 825                         if (callgraph_buf == NULL)
 826                                 callgraph_buf = "fp";
 827                 }
 828
 829                 /* parse callgraph parameters */
 830                 if (callgraph_buf != NULL) {
 831                         if (!strcmp(callgraph_buf, "no")) {
 832                                 param.enabled = false;
 833                                 param.record_mode = CALLCHAIN_NONE;
 834                         } else {
 835                                 param.enabled = true;
 836                                 if (parse_callchain_record(callgraph_buf, &param)) {
 837                                         pr_err("per-event callgraph setting for %s failed. "
 838                                                "Apply callgraph global setting for it\n",
 839                                                evsel->name);
 840                                         return;
 841                                 }
 842                                 if (param.record_mode == CALLCHAIN_DWARF)
 843                                         sample_address = true;
 844                         }
 845                 }
 846                 if (dump_size > 0) {
 847                         dump_size = round_up(dump_size, sizeof(u64));
 848                         param.dump_size = dump_size;
 849                 }
 850
 851                 /* If global callgraph set, clear it */
 852                 if (callchain_param.enabled)
 853                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 854
 855                 /* set perf-event callgraph */
 856                 if (param.enabled) {
 857                         if (sample_address) {
 858                                 perf_evsel__set_sample_bit(evsel, ADDR);
 859                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 860                                 evsel->attr.mmap_data = track;
 861                         }
 862                         perf_evsel__config_callchain(evsel, opts, &param);
 863                 }
 864         }
 865 }
 866
 867 static bool is_dummy_event(struct perf_evsel *evsel)
 868 {
 869         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 870                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 871 }
 872
 873 /*
 874  * The enable_on_exec/disabled value strategy:
 875  *
 876  *  1) For any type of traced program:
 877  *    - all independent events and group leaders are disabled
 878  *    - all group members are enabled
 879  *
 880  *     Group members are ruled by group leaders. They need to
 881  *     be enabled, because the group scheduling relies on that.
 882  *
 883  *  2) For traced programs executed by perf:
 884  *     - all independent events and group leaders have
 885  *       enable_on_exec set
 886  *     - we don't specifically enable or disable any event during
 887  *       the record command
 888  *
 889  *     Independent events and group leaders are initially disabled
 890  *     and get enabled by exec. Group members are ruled by group
 891  *     leaders as stated in 1).
 892  *
 893  *  3) For traced programs attached by perf (pid/tid):
 894  *     - we specifically enable or disable all events during
 895  *       the record command
 896  *
 897  *     When attaching events to already running traced we
 898  *     enable/disable events specifically, as there's no
 899  *     initial traced exec call.
 900  */
 901 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 902                         struct callchain_param *callchain)
 903 {
 904         struct perf_evsel *leader = evsel->leader;
 905         struct perf_event_attr *attr = &evsel->attr;
 906         int track = evsel->tracking;
 907         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 908
 909         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 910         attr->inherit       = !opts->no_inherit;
 911         attr->write_backward = opts->overwrite ? 1 : 0;
 912
 913         perf_evsel__set_sample_bit(evsel, IP);
 914         perf_evsel__set_sample_bit(evsel, TID);
 915
 916         if (evsel->sample_read) {
 917                 perf_evsel__set_sample_bit(evsel, READ);
 918
 919                 /*
 920                  * We need ID even in case of single event, because
 921                  * PERF_SAMPLE_READ process ID specific data.
 922                  */
 923                 perf_evsel__set_sample_id(evsel, false);
 924
 925                 /*
 926                  * Apply group format only if we belong to group
 927                  * with more than one members.
 928                  */
 929                 if (leader->nr_members > 1) {
 930                         attr->read_format |= PERF_FORMAT_GROUP;
 931                         attr->inherit = 0;
 932                 }
 933         }
 934
 935         /*
 936          * We default some events to have a default interval. But keep
 937          * it a weak assumption overridable by the user.
 938          */
 939         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 940                                      opts->user_interval != ULLONG_MAX)) {
 941                 if (opts->freq) {
 942                         perf_evsel__set_sample_bit(evsel, PERIOD);
 943                         attr->freq              = 1;
 944                         attr->sample_freq       = opts->freq;
 945                 } else {
 946                         attr->sample_period = opts->default_interval;
 947                 }
 948         }
 949
 950         /*
 951          * Disable sampling for all group members other
 952          * than leader in case leader 'leads' the sampling.
 953          */
 954         if ((leader != evsel) && leader->sample_read) {
 955                 attr->freq           = 0;
 956                 attr->sample_freq    = 0;
 957                 attr->sample_period  = 0;
 958                 attr->write_backward = 0;
 959                 attr->sample_id_all  = 0;
 960         }
 961
 962         if (opts->no_samples)
 963                 attr->sample_freq = 0;
 964
 965         if (opts->inherit_stat) {
 966                 evsel->attr.read_format |=
 967                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 968                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 969                         PERF_FORMAT_ID;
 970                 attr->inherit_stat = 1;
 971         }
 972
 973         if (opts->sample_address) {
 974                 perf_evsel__set_sample_bit(evsel, ADDR);
 975                 attr->mmap_data = track;
 976         }
 977
 978         /*
 979          * We don't allow user space callchains for  function trace
 980          * event, due to issues with page faults while tracing page
 981          * fault handler and its overall trickiness nature.
 982          */
 983         if (perf_evsel__is_function_event(evsel))
 984                 evsel->attr.exclude_callchain_user = 1;
 985
 986         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 987                 perf_evsel__config_callchain(evsel, opts, callchain);
 988
 989         if (opts->sample_intr_regs) {
 990                 attr->sample_regs_intr = opts->sample_intr_regs;
 991                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 992         }
 993
 994         if (opts->sample_user_regs) {
 995                 attr->sample_regs_user |= opts->sample_user_regs;
 996                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 997         }
 998
 999         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1000                 perf_evsel__set_sample_bit(evsel, CPU);
1001
1002         /*
1003          * When the user explicitly disabled time don't force it here.
1004          */
1005         if (opts->sample_time &&
1006             (!perf_missing_features.sample_id_all &&
1007             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1008              opts->sample_time_set)))
1009                 perf_evsel__set_sample_bit(evsel, TIME);
1010
1011         if (opts->raw_samples && !evsel->no_aux_samples) {
1012                 perf_evsel__set_sample_bit(evsel, TIME);
1013                 perf_evsel__set_sample_bit(evsel, RAW);
1014                 perf_evsel__set_sample_bit(evsel, CPU);
1015         }
1016
1017         if (opts->sample_address)
1018                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1019
1020         if (opts->sample_phys_addr)
1021                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1022
1023         if (opts->no_buffering) {
1024                 attr->watermark = 0;
1025                 attr->wakeup_events = 1;
1026         }
1027         if (opts->branch_stack && !evsel->no_aux_samples) {
1028                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1029                 attr->branch_sample_type = opts->branch_stack;
1030         }
1031
1032         if (opts->sample_weight)
1033                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1034
1035         attr->task  = track;
1036         attr->mmap  = track;
1037         attr->mmap2 = track && !perf_missing_features.mmap2;
1038         attr->comm  = track;
1039
1040         if (opts->record_namespaces)
1041                 attr->namespaces  = track;
1042
1043         if (opts->record_switch_events)
1044                 attr->context_switch = track;
1045
1046         if (opts->sample_transaction)
1047                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1048
1049         if (opts->running_time) {
1050                 evsel->attr.read_format |=
1051                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1052                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1053         }
1054
1055         /*
1056          * XXX see the function comment above
1057          *
1058          * Disabling only independent events or group leaders,
1059          * keeping group members enabled.
1060          */
1061         if (perf_evsel__is_group_leader(evsel))
1062                 attr->disabled = 1;
1063
1064         /*
1065          * Setting enable_on_exec for independent events and
1066          * group leaders for traced executed by perf.
1067          */
1068         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1069                 !opts->initial_delay)
1070                 attr->enable_on_exec = 1;
1071
1072         if (evsel->immediate) {
1073                 attr->disabled = 0;
1074                 attr->enable_on_exec = 0;
1075         }
1076
1077         clockid = opts->clockid;
1078         if (opts->use_clockid) {
1079                 attr->use_clockid = 1;
1080                 attr->clockid = opts->clockid;
1081         }
1082
1083         if (evsel->precise_max)
1084                 perf_event_attr__set_max_precise_ip(attr);
1085
1086         if (opts->all_user) {
1087                 attr->exclude_kernel = 1;
1088                 attr->exclude_user   = 0;
1089         }
1090
1091         if (opts->all_kernel) {
1092                 attr->exclude_kernel = 0;
1093                 attr->exclude_user   = 1;
1094         }
1095
1096         if (evsel->own_cpus)
1097                 evsel->attr.read_format |= PERF_FORMAT_ID;
1098
1099         /*
1100          * Apply event specific term settings,
1101          * it overloads any global configuration.
1102          */
1103         apply_config_terms(evsel, opts, track);
1104
1105         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1106
1107         /* The --period option takes the precedence. */
1108         if (opts->period_set) {
1109                 if (opts->period)
1110                         perf_evsel__set_sample_bit(evsel, PERIOD);
1111                 else
1112                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1113         }
1114
1115         /*
1116          * For initial_delay, a dummy event is added implicitly.
1117          * The software event will trigger -EOPNOTSUPP error out,
1118          * if BRANCH_STACK bit is set.
1119          */
1120         if (opts->initial_delay && is_dummy_event(evsel))
1121                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1122 }
1123
1124 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1125 {
1126         if (evsel->system_wide)
1127                 nthreads = 1;
1128
1129         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1130
1131         if (evsel->fd) {
1132                 int cpu, thread;
1133                 for (cpu = 0; cpu < ncpus; cpu++) {
1134                         for (thread = 0; thread < nthreads; thread++) {
1135                                 FD(evsel, cpu, thread) = -1;
1136                         }
1137                 }
1138         }
1139
1140         return evsel->fd != NULL ? 0 : -ENOMEM;
1141 }
1142
1143 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1144                           int ioc,  void *arg)
1145 {
1146         int cpu, thread;
1147
1148         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1149                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1150                         int fd = FD(evsel, cpu, thread),
1151                             err = ioctl(fd, ioc, arg);
1152
1153                         if (err)
1154                                 return err;
1155                 }
1156         }
1157
1158         return 0;
1159 }
1160
1161 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1162 {
1163         return perf_evsel__run_ioctl(evsel,
1164                                      PERF_EVENT_IOC_SET_FILTER,
1165                                      (void *)filter);
1166 }
1167
1168 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1169 {
1170         char *new_filter = strdup(filter);
1171
1172         if (new_filter != NULL) {
1173                 free(evsel->filter);
1174                 evsel->filter = new_filter;
1175                 return 0;
1176         }
1177
1178         return -1;
1179 }
1180
1181 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1182                                      const char *fmt, const char *filter)
1183 {
1184         char *new_filter;
1185
1186         if (evsel->filter == NULL)
1187                 return perf_evsel__set_filter(evsel, filter);
1188
1189         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1190                 free(evsel->filter);
1191                 evsel->filter = new_filter;
1192                 return 0;
1193         }
1194
1195         return -1;
1196 }
1197
1198 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1199 {
1200         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1201 }
1202
1203 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1204 {
1205         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1206 }
1207
1208 int perf_evsel__enable(struct perf_evsel *evsel)
1209 {
1210         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1211
1212         if (!err)
1213                 evsel->disabled = false;
1214
1215         return err;
1216 }
1217
1218 int perf_evsel__disable(struct perf_evsel *evsel)
1219 {
1220         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1221         /*
1222          * We mark it disabled here so that tools that disable a event can
1223          * ignore events after they disable it. I.e. the ring buffer may have
1224          * already a few more events queued up before the kernel got the stop
1225          * request.
1226          */
1227         if (!err)
1228                 evsel->disabled = true;
1229
1230         return err;
1231 }
1232
1233 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1234 {
1235         if (ncpus == 0 || nthreads == 0)
1236                 return 0;
1237
1238         if (evsel->system_wide)
1239                 nthreads = 1;
1240
1241         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1242         if (evsel->sample_id == NULL)
1243                 return -ENOMEM;
1244
1245         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1246         if (evsel->id == NULL) {
1247                 xyarray__delete(evsel->sample_id);
1248                 evsel->sample_id = NULL;
1249                 return -ENOMEM;
1250         }
1251
1252         return 0;
1253 }
1254
1255 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1256 {
1257         xyarray__delete(evsel->fd);
1258         evsel->fd = NULL;
1259 }
1260
1261 static void perf_evsel__free_id(struct perf_evsel *evsel)
1262 {
1263         xyarray__delete(evsel->sample_id);
1264         evsel->sample_id = NULL;
1265         zfree(&evsel->id);
1266 }
1267
1268 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1269 {
1270         struct perf_evsel_config_term *term, *h;
1271
1272         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1273                 list_del(&term->list);
1274                 free(term);
1275         }
1276 }
1277
1278 void perf_evsel__close_fd(struct perf_evsel *evsel)
1279 {
1280         int cpu, thread;
1281
1282         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1283                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1284                         close(FD(evsel, cpu, thread));
1285                         FD(evsel, cpu, thread) = -1;
1286                 }
1287 }
1288
1289 void perf_evsel__exit(struct perf_evsel *evsel)
1290 {
1291         assert(list_empty(&evsel->node));
1292         assert(evsel->evlist == NULL);
1293         perf_evsel__free_fd(evsel);
1294         perf_evsel__free_id(evsel);
1295         perf_evsel__free_config_terms(evsel);
1296         cgroup__put(evsel->cgrp);
1297         cpu_map__put(evsel->cpus);
1298         cpu_map__put(evsel->own_cpus);
1299         thread_map__put(evsel->threads);
1300         zfree(&evsel->group_name);
1301         zfree(&evsel->name);
1302         perf_evsel__object.fini(evsel);
1303 }
1304
1305 void perf_evsel__delete(struct perf_evsel *evsel)
1306 {
1307         perf_evsel__exit(evsel);
1308         free(evsel);
1309 }
1310
1311 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1312                                 struct perf_counts_values *count)
1313 {
1314         struct perf_counts_values tmp;
1315
1316         if (!evsel->prev_raw_counts)
1317                 return;
1318
1319         if (cpu == -1) {
1320                 tmp = evsel->prev_raw_counts->aggr;
1321                 evsel->prev_raw_counts->aggr = *count;
1322         } else {
1323                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1324                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1325         }
1326
1327         count->val = count->val - tmp.val;
1328         count->ena = count->ena - tmp.ena;
1329         count->run = count->run - tmp.run;
1330 }
1331
1332 void perf_counts_values__scale(struct perf_counts_values *count,
1333                                bool scale, s8 *pscaled)
1334 {
1335         s8 scaled = 0;
1336
1337         if (scale) {
1338                 if (count->run == 0) {
1339                         scaled = -1;
1340                         count->val = 0;
1341                 } else if (count->run < count->ena) {
1342                         scaled = 1;
1343                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1344                 }
1345         } else
1346                 count->ena = count->run = 0;
1347
1348         if (pscaled)
1349                 *pscaled = scaled;
1350 }
1351
1352 static int perf_evsel__read_size(struct perf_evsel *evsel)
1353 {
1354         u64 read_format = evsel->attr.read_format;
1355         int entry = sizeof(u64); /* value */
1356         int size = 0;
1357         int nr = 1;
1358
1359         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1360                 size += sizeof(u64);
1361
1362         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1363                 size += sizeof(u64);
1364
1365         if (read_format & PERF_FORMAT_ID)
1366                 entry += sizeof(u64);
1367
1368         if (read_format & PERF_FORMAT_GROUP) {
1369                 nr = evsel->nr_members;
1370                 size += sizeof(u64);
1371         }
1372
1373         size += entry * nr;
1374         return size;
1375 }
1376
1377 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1378                      struct perf_counts_values *count)
1379 {
1380         size_t size = perf_evsel__read_size(evsel);
1381
1382         memset(count, 0, sizeof(*count));
1383
1384         if (FD(evsel, cpu, thread) < 0)
1385                 return -EINVAL;
1386
1387         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1388                 return -errno;
1389
1390         return 0;
1391 }
1392
1393 static int
1394 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1395 {
1396         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1397
1398         return perf_evsel__read(evsel, cpu, thread, count);
1399 }
1400
1401 static void
1402 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1403                       u64 val, u64 ena, u64 run)
1404 {
1405         struct perf_counts_values *count;
1406
1407         count = perf_counts(counter->counts, cpu, thread);
1408
1409         count->val    = val;
1410         count->ena    = ena;
1411         count->run    = run;
1412         count->loaded = true;
1413 }
1414
1415 static int
1416 perf_evsel__process_group_data(struct perf_evsel *leader,
1417                                int cpu, int thread, u64 *data)
1418 {
1419         u64 read_format = leader->attr.read_format;
1420         struct sample_read_value *v;
1421         u64 nr, ena = 0, run = 0, i;
1422
1423         nr = *data++;
1424
1425         if (nr != (u64) leader->nr_members)
1426                 return -EINVAL;
1427
1428         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1429                 ena = *data++;
1430
1431         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1432                 run = *data++;
1433
1434         v = (struct sample_read_value *) data;
1435
1436         perf_evsel__set_count(leader, cpu, thread,
1437                               v[0].value, ena, run);
1438
1439         for (i = 1; i < nr; i++) {
1440                 struct perf_evsel *counter;
1441
1442                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1443                 if (!counter)
1444                         return -EINVAL;
1445
1446                 perf_evsel__set_count(counter, cpu, thread,
1447                                       v[i].value, ena, run);
1448         }
1449
1450         return 0;
1451 }
1452
1453 static int
1454 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1455 {
1456         struct perf_stat_evsel *ps = leader->stats;
1457         u64 read_format = leader->attr.read_format;
1458         int size = perf_evsel__read_size(leader);
1459         u64 *data = ps->group_data;
1460
1461         if (!(read_format & PERF_FORMAT_ID))
1462                 return -EINVAL;
1463
1464         if (!perf_evsel__is_group_leader(leader))
1465                 return -EINVAL;
1466
1467         if (!data) {
1468                 data = zalloc(size);
1469                 if (!data)
1470                         return -ENOMEM;
1471
1472                 ps->group_data = data;
1473         }
1474
1475         if (FD(leader, cpu, thread) < 0)
1476                 return -EINVAL;
1477
1478         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1479                 return -errno;
1480
1481         return perf_evsel__process_group_data(leader, cpu, thread, data);
1482 }
1483
1484 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1485 {
1486         u64 read_format = evsel->attr.read_format;
1487
1488         if (read_format & PERF_FORMAT_GROUP)
1489                 return perf_evsel__read_group(evsel, cpu, thread);
1490         else
1491                 return perf_evsel__read_one(evsel, cpu, thread);
1492 }
1493
1494 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1495                               int cpu, int thread, bool scale)
1496 {
1497         struct perf_counts_values count;
1498         size_t nv = scale ? 3 : 1;
1499
1500         if (FD(evsel, cpu, thread) < 0)
1501                 return -EINVAL;
1502
1503         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1504                 return -ENOMEM;
1505
1506         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1507                 return -errno;
1508
1509         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1510         perf_counts_values__scale(&count, scale, NULL);
1511         *perf_counts(evsel->counts, cpu, thread) = count;
1512         return 0;
1513 }
1514
1515 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1516 {
1517         struct perf_evsel *leader = evsel->leader;
1518         int fd;
1519
1520         if (perf_evsel__is_group_leader(evsel))
1521                 return -1;
1522
1523         /*
1524          * Leader must be already processed/open,
1525          * if not it's a bug.
1526          */
1527         BUG_ON(!leader->fd);
1528
1529         fd = FD(leader, cpu, thread);
1530         BUG_ON(fd == -1);
1531
1532         return fd;
1533 }
1534
1535 struct bit_names {
1536         int bit;
1537         const char *name;
1538 };
1539
1540 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1541 {
1542         bool first_bit = true;
1543         int i = 0;
1544
1545         do {
1546                 if (value & bits[i].bit) {
1547                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1548                         first_bit = false;
1549                 }
1550         } while (bits[++i].name != NULL);
1551 }
1552
1553 static void __p_sample_type(char *buf, size_t size, u64 value)
1554 {
1555 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1556         struct bit_names bits[] = {
1557                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1558                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1559                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1560                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1561                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1562                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1563                 { .name = NULL, }
1564         };
1565 #undef bit_name
1566         __p_bits(buf, size, value, bits);
1567 }
1568
1569 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1570 {
1571 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1572         struct bit_names bits[] = {
1573                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1574                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1575                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1576                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1577                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1578                 { .name = NULL, }
1579         };
1580 #undef bit_name
1581         __p_bits(buf, size, value, bits);
1582 }
1583
1584 static void __p_read_format(char *buf, size_t size, u64 value)
1585 {
1586 #define bit_name(n) { PERF_FORMAT_##n, #n }
1587         struct bit_names bits[] = {
1588                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1589                 bit_name(ID), bit_name(GROUP),
1590                 { .name = NULL, }
1591         };
1592 #undef bit_name
1593         __p_bits(buf, size, value, bits);
1594 }
1595
1596 #define BUF_SIZE                1024
1597
1598 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1599 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1600 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1601 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1602 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1603 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1604
1605 #define PRINT_ATTRn(_n, _f, _p)                         \
1606 do {                                                    \
1607         if (attr->_f) {                                 \
1608                 _p(attr->_f);                           \
1609                 ret += attr__fprintf(fp, _n, buf, priv);\
1610         }                                               \
1611 } while (0)
1612
1613 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1614
1615 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1616                              attr__fprintf_f attr__fprintf, void *priv)
1617 {
1618         char buf[BUF_SIZE];
1619         int ret = 0;
1620
1621         PRINT_ATTRf(type, p_unsigned);
1622         PRINT_ATTRf(size, p_unsigned);
1623         PRINT_ATTRf(config, p_hex);
1624         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1625         PRINT_ATTRf(sample_type, p_sample_type);
1626         PRINT_ATTRf(read_format, p_read_format);
1627
1628         PRINT_ATTRf(disabled, p_unsigned);
1629         PRINT_ATTRf(inherit, p_unsigned);
1630         PRINT_ATTRf(pinned, p_unsigned);
1631         PRINT_ATTRf(exclusive, p_unsigned);
1632         PRINT_ATTRf(exclude_user, p_unsigned);
1633         PRINT_ATTRf(exclude_kernel, p_unsigned);
1634         PRINT_ATTRf(exclude_hv, p_unsigned);
1635         PRINT_ATTRf(exclude_idle, p_unsigned);
1636         PRINT_ATTRf(mmap, p_unsigned);
1637         PRINT_ATTRf(comm, p_unsigned);
1638         PRINT_ATTRf(freq, p_unsigned);
1639         PRINT_ATTRf(inherit_stat, p_unsigned);
1640         PRINT_ATTRf(enable_on_exec, p_unsigned);
1641         PRINT_ATTRf(task, p_unsigned);
1642         PRINT_ATTRf(watermark, p_unsigned);
1643         PRINT_ATTRf(precise_ip, p_unsigned);
1644         PRINT_ATTRf(mmap_data, p_unsigned);
1645         PRINT_ATTRf(sample_id_all, p_unsigned);
1646         PRINT_ATTRf(exclude_host, p_unsigned);
1647         PRINT_ATTRf(exclude_guest, p_unsigned);
1648         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1649         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1650         PRINT_ATTRf(mmap2, p_unsigned);
1651         PRINT_ATTRf(comm_exec, p_unsigned);
1652         PRINT_ATTRf(use_clockid, p_unsigned);
1653         PRINT_ATTRf(context_switch, p_unsigned);
1654         PRINT_ATTRf(write_backward, p_unsigned);
1655         PRINT_ATTRf(namespaces, p_unsigned);
1656
1657         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1658         PRINT_ATTRf(bp_type, p_unsigned);
1659         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1660         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1661         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1662         PRINT_ATTRf(sample_regs_user, p_hex);
1663         PRINT_ATTRf(sample_stack_user, p_unsigned);
1664         PRINT_ATTRf(clockid, p_signed);
1665         PRINT_ATTRf(sample_regs_intr, p_hex);
1666         PRINT_ATTRf(aux_watermark, p_unsigned);
1667         PRINT_ATTRf(sample_max_stack, p_unsigned);
1668
1669         return ret;
1670 }
1671
1672 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1673                                 void *priv __maybe_unused)
1674 {
1675         return fprintf(fp, "  %-32s %s\n", name, val);
1676 }
1677
1678 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1679                                   int nr_cpus, int nr_threads,
1680                                   int thread_idx)
1681 {
1682         for (int cpu = 0; cpu < nr_cpus; cpu++)
1683                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1684                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1685 }
1686
1687 static int update_fds(struct perf_evsel *evsel,
1688                       int nr_cpus, int cpu_idx,
1689                       int nr_threads, int thread_idx)
1690 {
1691         struct perf_evsel *pos;
1692
1693         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1694                 return -EINVAL;
1695
1696         evlist__for_each_entry(evsel->evlist, pos) {
1697                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1698
1699                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1700
1701                 /*
1702                  * Since fds for next evsel has not been created,
1703                  * there is no need to iterate whole event list.
1704                  */
1705                 if (pos == evsel)
1706                         break;
1707         }
1708         return 0;
1709 }
1710
1711 static bool ignore_missing_thread(struct perf_evsel *evsel,
1712                                   int nr_cpus, int cpu,
1713                                   struct thread_map *threads,
1714                                   int thread, int err)
1715 {
1716         pid_t ignore_pid = thread_map__pid(threads, thread);
1717
1718         if (!evsel->ignore_missing_thread)
1719                 return false;
1720
1721         /* The system wide setup does not work with threads. */
1722         if (evsel->system_wide)
1723                 return false;
1724
1725         /* The -ESRCH is perf event syscall errno for pid's not found. */
1726         if (err != -ESRCH)
1727                 return false;
1728
1729         /* If there's only one thread, let it fail. */
1730         if (threads->nr == 1)
1731                 return false;
1732
1733         /*
1734          * We should remove fd for missing_thread first
1735          * because thread_map__remove() will decrease threads->nr.
1736          */
1737         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1738                 return false;
1739
1740         if (thread_map__remove(threads, thread))
1741                 return false;
1742
1743         pr_warning("WARNING: Ignored open failure for pid %d\n",
1744                    ignore_pid);
1745         return true;
1746 }
1747
1748 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1749                      struct thread_map *threads)
1750 {
1751         int cpu, thread, nthreads;
1752         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1753         int pid = -1, err;
1754         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1755
1756         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1757                 return -EINVAL;
1758
1759         if (cpus == NULL) {
1760                 static struct cpu_map *empty_cpu_map;
1761
1762                 if (empty_cpu_map == NULL) {
1763                         empty_cpu_map = cpu_map__dummy_new();
1764                         if (empty_cpu_map == NULL)
1765                                 return -ENOMEM;
1766                 }
1767
1768                 cpus = empty_cpu_map;
1769         }
1770
1771         if (threads == NULL) {
1772                 static struct thread_map *empty_thread_map;
1773
1774                 if (empty_thread_map == NULL) {
1775                         empty_thread_map = thread_map__new_by_tid(-1);
1776                         if (empty_thread_map == NULL)
1777                                 return -ENOMEM;
1778                 }
1779
1780                 threads = empty_thread_map;
1781         }
1782
1783         if (evsel->system_wide)
1784                 nthreads = 1;
1785         else
1786                 nthreads = threads->nr;
1787
1788         if (evsel->fd == NULL &&
1789             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1790                 return -ENOMEM;
1791
1792         if (evsel->cgrp) {
1793                 flags |= PERF_FLAG_PID_CGROUP;
1794                 pid = evsel->cgrp->fd;
1795         }
1796
1797 fallback_missing_features:
1798         if (perf_missing_features.clockid_wrong)
1799                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1800         if (perf_missing_features.clockid) {
1801                 evsel->attr.use_clockid = 0;
1802                 evsel->attr.clockid = 0;
1803         }
1804         if (perf_missing_features.cloexec)
1805                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1806         if (perf_missing_features.mmap2)
1807                 evsel->attr.mmap2 = 0;
1808         if (perf_missing_features.exclude_guest)
1809                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1810         if (perf_missing_features.lbr_flags)
1811                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1812                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1813         if (perf_missing_features.group_read && evsel->attr.inherit)
1814                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1815 retry_sample_id:
1816         if (perf_missing_features.sample_id_all)
1817                 evsel->attr.sample_id_all = 0;
1818
1819         if (verbose >= 2) {
1820                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1821                 fprintf(stderr, "perf_event_attr:\n");
1822                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1823                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1824         }
1825
1826         for (cpu = 0; cpu < cpus->nr; cpu++) {
1827
1828                 for (thread = 0; thread < nthreads; thread++) {
1829                         int fd, group_fd;
1830
1831                         if (!evsel->cgrp && !evsel->system_wide)
1832                                 pid = thread_map__pid(threads, thread);
1833
1834                         group_fd = get_group_fd(evsel, cpu, thread);
1835 retry_open:
1836                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1837                                   pid, cpus->map[cpu], group_fd, flags);
1838
1839                         test_attr__ready();
1840
1841                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1842                                                  group_fd, flags);
1843
1844                         FD(evsel, cpu, thread) = fd;
1845
1846                         if (fd < 0) {
1847                                 err = -errno;
1848
1849                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1850                                         /*
1851                                          * We just removed 1 thread, so take a step
1852                                          * back on thread index and lower the upper
1853                                          * nthreads limit.
1854                                          */
1855                                         nthreads--;
1856                                         thread--;
1857
1858                                         /* ... and pretend like nothing have happened. */
1859                                         err = 0;
1860                                         continue;
1861                                 }
1862
1863                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1864                                           err);
1865                                 goto try_fallback;
1866                         }
1867
1868                         pr_debug2(" = %d\n", fd);
1869
1870                         if (evsel->bpf_fd >= 0) {
1871                                 int evt_fd = fd;
1872                                 int bpf_fd = evsel->bpf_fd;
1873
1874                                 err = ioctl(evt_fd,
1875                                             PERF_EVENT_IOC_SET_BPF,
1876                                             bpf_fd);
1877                                 if (err && errno != EEXIST) {
1878                                         pr_err("failed to attach bpf fd %d: %s\n",
1879                                                bpf_fd, strerror(errno));
1880                                         err = -EINVAL;
1881                                         goto out_close;
1882                                 }
1883                         }
1884
1885                         set_rlimit = NO_CHANGE;
1886
1887                         /*
1888                          * If we succeeded but had to kill clockid, fail and
1889                          * have perf_evsel__open_strerror() print us a nice
1890                          * error.
1891                          */
1892                         if (perf_missing_features.clockid ||
1893                             perf_missing_features.clockid_wrong) {
1894                                 err = -EINVAL;
1895                                 goto out_close;
1896                         }
1897                 }
1898         }
1899
1900         return 0;
1901
1902 try_fallback:
1903         /*
1904          * perf stat needs between 5 and 22 fds per CPU. When we run out
1905          * of them try to increase the limits.
1906          */
1907         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1908                 struct rlimit l;
1909                 int old_errno = errno;
1910
1911                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1912                         if (set_rlimit == NO_CHANGE)
1913                                 l.rlim_cur = l.rlim_max;
1914                         else {
1915                                 l.rlim_cur = l.rlim_max + 1000;
1916                                 l.rlim_max = l.rlim_cur;
1917                         }
1918                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1919                                 set_rlimit++;
1920                                 errno = old_errno;
1921                                 goto retry_open;
1922                         }
1923                 }
1924                 errno = old_errno;
1925         }
1926
1927         if (err != -EINVAL || cpu > 0 || thread > 0)
1928                 goto out_close;
1929
1930         /*
1931          * Must probe features in the order they were added to the
1932          * perf_event_attr interface.
1933          */
1934         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1935                 perf_missing_features.write_backward = true;
1936                 pr_debug2("switching off write_backward\n");
1937                 goto out_close;
1938         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1939                 perf_missing_features.clockid_wrong = true;
1940                 pr_debug2("switching off clockid\n");
1941                 goto fallback_missing_features;
1942         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1943                 perf_missing_features.clockid = true;
1944                 pr_debug2("switching off use_clockid\n");
1945                 goto fallback_missing_features;
1946         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1947                 perf_missing_features.cloexec = true;
1948                 pr_debug2("switching off cloexec flag\n");
1949                 goto fallback_missing_features;
1950         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1951                 perf_missing_features.mmap2 = true;
1952                 pr_debug2("switching off mmap2\n");
1953                 goto fallback_missing_features;
1954         } else if (!perf_missing_features.exclude_guest &&
1955                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1956                 perf_missing_features.exclude_guest = true;
1957                 pr_debug2("switching off exclude_guest, exclude_host\n");
1958                 goto fallback_missing_features;
1959         } else if (!perf_missing_features.sample_id_all) {
1960                 perf_missing_features.sample_id_all = true;
1961                 pr_debug2("switching off sample_id_all\n");
1962                 goto retry_sample_id;
1963         } else if (!perf_missing_features.lbr_flags &&
1964                         (evsel->attr.branch_sample_type &
1965                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1966                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1967                 perf_missing_features.lbr_flags = true;
1968                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1969                 goto fallback_missing_features;
1970         } else if (!perf_missing_features.group_read &&
1971                     evsel->attr.inherit &&
1972                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1973                    perf_evsel__is_group_leader(evsel)) {
1974                 perf_missing_features.group_read = true;
1975                 pr_debug2("switching off group read\n");
1976                 goto fallback_missing_features;
1977         }
1978 out_close:
1979         if (err)
1980                 threads->err_thread = thread;
1981
1982         do {
1983                 while (--thread >= 0) {
1984                         close(FD(evsel, cpu, thread));
1985                         FD(evsel, cpu, thread) = -1;
1986                 }
1987                 thread = nthreads;
1988         } while (--cpu >= 0);
1989         return err;
1990 }
1991
1992 void perf_evsel__close(struct perf_evsel *evsel)
1993 {
1994         if (evsel->fd == NULL)
1995                 return;
1996
1997         perf_evsel__close_fd(evsel);
1998         perf_evsel__free_fd(evsel);
1999 }
2000
2001 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2002                              struct cpu_map *cpus)
2003 {
2004         return perf_evsel__open(evsel, cpus, NULL);
2005 }
2006
2007 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2008                                 struct thread_map *threads)
2009 {
2010         return perf_evsel__open(evsel, NULL, threads);
2011 }
2012
2013 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2014                                        const union perf_event *event,
2015                                        struct perf_sample *sample)
2016 {
2017         u64 type = evsel->attr.sample_type;
2018         const u64 *array = event->sample.array;
2019         bool swapped = evsel->needs_swap;
2020         union u64_swap u;
2021
2022         array += ((event->header.size -
2023                    sizeof(event->header)) / sizeof(u64)) - 1;
2024
2025         if (type & PERF_SAMPLE_IDENTIFIER) {
2026                 sample->id = *array;
2027                 array--;
2028         }
2029
2030         if (type & PERF_SAMPLE_CPU) {
2031                 u.val64 = *array;
2032                 if (swapped) {
2033                         /* undo swap of u64, then swap on individual u32s */
2034                         u.val64 = bswap_64(u.val64);
2035                         u.val32[0] = bswap_32(u.val32[0]);
2036                 }
2037
2038                 sample->cpu = u.val32[0];
2039                 array--;
2040         }
2041
2042         if (type & PERF_SAMPLE_STREAM_ID) {
2043                 sample->stream_id = *array;
2044                 array--;
2045         }
2046
2047         if (type & PERF_SAMPLE_ID) {
2048                 sample->id = *array;
2049                 array--;
2050         }
2051
2052         if (type & PERF_SAMPLE_TIME) {
2053                 sample->time = *array;
2054                 array--;
2055         }
2056
2057         if (type & PERF_SAMPLE_TID) {
2058                 u.val64 = *array;
2059                 if (swapped) {
2060                         /* undo swap of u64, then swap on individual u32s */
2061                         u.val64 = bswap_64(u.val64);
2062                         u.val32[0] = bswap_32(u.val32[0]);
2063                         u.val32[1] = bswap_32(u.val32[1]);
2064                 }
2065
2066                 sample->pid = u.val32[0];
2067                 sample->tid = u.val32[1];
2068                 array--;
2069         }
2070
2071         return 0;
2072 }
2073
2074 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2075                             u64 size)
2076 {
2077         return size > max_size || offset + size > endp;
2078 }
2079
2080 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2081         do {                                                            \
2082                 if (overflow(endp, (max_size), (offset), (size)))       \
2083                         return -EFAULT;                                 \
2084         } while (0)
2085
2086 #define OVERFLOW_CHECK_u64(offset) \
2087         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2088
2089 static int
2090 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2091 {
2092         /*
2093          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2094          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2095          * check the format does not go past the end of the event.
2096          */
2097         if (sample_size + sizeof(event->header) > event->header.size)
2098                 return -EFAULT;
2099
2100         return 0;
2101 }
2102
2103 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2104                              struct perf_sample *data)
2105 {
2106         u64 type = evsel->attr.sample_type;
2107         bool swapped = evsel->needs_swap;
2108         const u64 *array;
2109         u16 max_size = event->header.size;
2110         const void *endp = (void *)event + max_size;
2111         u64 sz;
2112
2113         /*
2114          * used for cross-endian analysis. See git commit 65014ab3
2115          * for why this goofiness is needed.
2116          */
2117         union u64_swap u;
2118
2119         memset(data, 0, sizeof(*data));
2120         data->cpu = data->pid = data->tid = -1;
2121         data->stream_id = data->id = data->time = -1ULL;
2122         data->period = evsel->attr.sample_period;
2123         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2124         data->misc    = event->header.misc;
2125         data->id = -1ULL;
2126         data->data_src = PERF_MEM_DATA_SRC_NONE;
2127
2128         if (event->header.type != PERF_RECORD_SAMPLE) {
2129                 if (!evsel->attr.sample_id_all)
2130                         return 0;
2131                 return perf_evsel__parse_id_sample(evsel, event, data);
2132         }
2133
2134         array = event->sample.array;
2135
2136         if (perf_event__check_size(event, evsel->sample_size))
2137                 return -EFAULT;
2138
2139         if (type & PERF_SAMPLE_IDENTIFIER) {
2140                 data->id = *array;
2141                 array++;
2142         }
2143
2144         if (type & PERF_SAMPLE_IP) {
2145                 data->ip = *array;
2146                 array++;
2147         }
2148
2149         if (type & PERF_SAMPLE_TID) {
2150                 u.val64 = *array;
2151                 if (swapped) {
2152                         /* undo swap of u64, then swap on individual u32s */
2153                         u.val64 = bswap_64(u.val64);
2154                         u.val32[0] = bswap_32(u.val32[0]);
2155                         u.val32[1] = bswap_32(u.val32[1]);
2156                 }
2157
2158                 data->pid = u.val32[0];
2159                 data->tid = u.val32[1];
2160                 array++;
2161         }
2162
2163         if (type & PERF_SAMPLE_TIME) {
2164                 data->time = *array;
2165                 array++;
2166         }
2167
2168         if (type & PERF_SAMPLE_ADDR) {
2169                 data->addr = *array;
2170                 array++;
2171         }
2172
2173         if (type & PERF_SAMPLE_ID) {
2174                 data->id = *array;
2175                 array++;
2176         }
2177
2178         if (type & PERF_SAMPLE_STREAM_ID) {
2179                 data->stream_id = *array;
2180                 array++;
2181         }
2182
2183         if (type & PERF_SAMPLE_CPU) {
2184
2185                 u.val64 = *array;
2186                 if (swapped) {
2187                         /* undo swap of u64, then swap on individual u32s */
2188                         u.val64 = bswap_64(u.val64);
2189                         u.val32[0] = bswap_32(u.val32[0]);
2190                 }
2191
2192                 data->cpu = u.val32[0];
2193                 array++;
2194         }
2195
2196         if (type & PERF_SAMPLE_PERIOD) {
2197                 data->period = *array;
2198                 array++;
2199         }
2200
2201         if (type & PERF_SAMPLE_READ) {
2202                 u64 read_format = evsel->attr.read_format;
2203
2204                 OVERFLOW_CHECK_u64(array);
2205                 if (read_format & PERF_FORMAT_GROUP)
2206                         data->read.group.nr = *array;
2207                 else
2208                         data->read.one.value = *array;
2209
2210                 array++;
2211
2212                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2213                         OVERFLOW_CHECK_u64(array);
2214                         data->read.time_enabled = *array;
2215                         array++;
2216                 }
2217
2218                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2219                         OVERFLOW_CHECK_u64(array);
2220                         data->read.time_running = *array;
2221                         array++;
2222                 }
2223
2224                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2225                 if (read_format & PERF_FORMAT_GROUP) {
2226                         const u64 max_group_nr = UINT64_MAX /
2227                                         sizeof(struct sample_read_value);
2228
2229                         if (data->read.group.nr > max_group_nr)
2230                                 return -EFAULT;
2231                         sz = data->read.group.nr *
2232                              sizeof(struct sample_read_value);
2233                         OVERFLOW_CHECK(array, sz, max_size);
2234                         data->read.group.values =
2235                                         (struct sample_read_value *)array;
2236                         array = (void *)array + sz;
2237                 } else {
2238                         OVERFLOW_CHECK_u64(array);
2239                         data->read.one.id = *array;
2240                         array++;
2241                 }
2242         }
2243
2244         if (evsel__has_callchain(evsel)) {
2245                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2246
2247                 OVERFLOW_CHECK_u64(array);
2248                 data->callchain = (struct ip_callchain *)array++;
2249                 if (data->callchain->nr > max_callchain_nr)
2250                         return -EFAULT;
2251                 sz = data->callchain->nr * sizeof(u64);
2252                 OVERFLOW_CHECK(array, sz, max_size);
2253                 array = (void *)array + sz;
2254         }
2255
2256         if (type & PERF_SAMPLE_RAW) {
2257                 OVERFLOW_CHECK_u64(array);
2258                 u.val64 = *array;
2259
2260                 /*
2261                  * Undo swap of u64, then swap on individual u32s,
2262                  * get the size of the raw area and undo all of the
2263                  * swap. The pevent interface handles endianity by
2264                  * itself.
2265                  */
2266                 if (swapped) {
2267                         u.val64 = bswap_64(u.val64);
2268                         u.val32[0] = bswap_32(u.val32[0]);
2269                         u.val32[1] = bswap_32(u.val32[1]);
2270                 }
2271                 data->raw_size = u.val32[0];
2272
2273                 /*
2274                  * The raw data is aligned on 64bits including the
2275                  * u32 size, so it's safe to use mem_bswap_64.
2276                  */
2277                 if (swapped)
2278                         mem_bswap_64((void *) array, data->raw_size);
2279
2280                 array = (void *)array + sizeof(u32);
2281
2282                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2283                 data->raw_data = (void *)array;
2284                 array = (void *)array + data->raw_size;
2285         }
2286
2287         if (type & PERF_SAMPLE_BRANCH_STACK) {
2288                 const u64 max_branch_nr = UINT64_MAX /
2289                                           sizeof(struct branch_entry);
2290
2291                 OVERFLOW_CHECK_u64(array);
2292                 data->branch_stack = (struct branch_stack *)array++;
2293
2294                 if (data->branch_stack->nr > max_branch_nr)
2295                         return -EFAULT;
2296                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2297                 OVERFLOW_CHECK(array, sz, max_size);
2298                 array = (void *)array + sz;
2299         }
2300
2301         if (type & PERF_SAMPLE_REGS_USER) {
2302                 OVERFLOW_CHECK_u64(array);
2303                 data->user_regs.abi = *array;
2304                 array++;
2305
2306                 if (data->user_regs.abi) {
2307                         u64 mask = evsel->attr.sample_regs_user;
2308
2309                         sz = hweight_long(mask) * sizeof(u64);
2310                         OVERFLOW_CHECK(array, sz, max_size);
2311                         data->user_regs.mask = mask;
2312                         data->user_regs.regs = (u64 *)array;
2313                         array = (void *)array + sz;
2314                 }
2315         }
2316
2317         if (type & PERF_SAMPLE_STACK_USER) {
2318                 OVERFLOW_CHECK_u64(array);
2319                 sz = *array++;
2320
2321                 data->user_stack.offset = ((char *)(array - 1)
2322                                           - (char *) event);
2323
2324                 if (!sz) {
2325                         data->user_stack.size = 0;
2326                 } else {
2327                         OVERFLOW_CHECK(array, sz, max_size);
2328                         data->user_stack.data = (char *)array;
2329                         array = (void *)array + sz;
2330                         OVERFLOW_CHECK_u64(array);
2331                         data->user_stack.size = *array++;
2332                         if (WARN_ONCE(data->user_stack.size > sz,
2333                                       "user stack dump failure\n"))
2334                                 return -EFAULT;
2335                 }
2336         }
2337
2338         if (type & PERF_SAMPLE_WEIGHT) {
2339                 OVERFLOW_CHECK_u64(array);
2340                 data->weight = *array;
2341                 array++;
2342         }
2343
2344         if (type & PERF_SAMPLE_DATA_SRC) {
2345                 OVERFLOW_CHECK_u64(array);
2346                 data->data_src = *array;
2347                 array++;
2348         }
2349
2350         if (type & PERF_SAMPLE_TRANSACTION) {
2351                 OVERFLOW_CHECK_u64(array);
2352                 data->transaction = *array;
2353                 array++;
2354         }
2355
2356         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2357         if (type & PERF_SAMPLE_REGS_INTR) {
2358                 OVERFLOW_CHECK_u64(array);
2359                 data->intr_regs.abi = *array;
2360                 array++;
2361
2362                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2363                         u64 mask = evsel->attr.sample_regs_intr;
2364
2365                         sz = hweight_long(mask) * sizeof(u64);
2366                         OVERFLOW_CHECK(array, sz, max_size);
2367                         data->intr_regs.mask = mask;
2368                         data->intr_regs.regs = (u64 *)array;
2369                         array = (void *)array + sz;
2370                 }
2371         }
2372
2373         data->phys_addr = 0;
2374         if (type & PERF_SAMPLE_PHYS_ADDR) {
2375                 data->phys_addr = *array;
2376                 array++;
2377         }
2378
2379         return 0;
2380 }
2381
2382 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2383                                        union perf_event *event,
2384                                        u64 *timestamp)
2385 {
2386         u64 type = evsel->attr.sample_type;
2387         const u64 *array;
2388
2389         if (!(type & PERF_SAMPLE_TIME))
2390                 return -1;
2391
2392         if (event->header.type != PERF_RECORD_SAMPLE) {
2393                 struct perf_sample data = {
2394                         .time = -1ULL,
2395                 };
2396
2397                 if (!evsel->attr.sample_id_all)
2398                         return -1;
2399                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2400                         return -1;
2401
2402                 *timestamp = data.time;
2403                 return 0;
2404         }
2405
2406         array = event->sample.array;
2407
2408         if (perf_event__check_size(event, evsel->sample_size))
2409                 return -EFAULT;
2410
2411         if (type & PERF_SAMPLE_IDENTIFIER)
2412                 array++;
2413
2414         if (type & PERF_SAMPLE_IP)
2415                 array++;
2416
2417         if (type & PERF_SAMPLE_TID)
2418                 array++;
2419
2420         if (type & PERF_SAMPLE_TIME)
2421                 *timestamp = *array;
2422
2423         return 0;
2424 }
2425
2426 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2427                                      u64 read_format)
2428 {
2429         size_t sz, result = sizeof(struct sample_event);
2430
2431         if (type & PERF_SAMPLE_IDENTIFIER)
2432                 result += sizeof(u64);
2433
2434         if (type & PERF_SAMPLE_IP)
2435                 result += sizeof(u64);
2436
2437         if (type & PERF_SAMPLE_TID)
2438                 result += sizeof(u64);
2439
2440         if (type & PERF_SAMPLE_TIME)
2441                 result += sizeof(u64);
2442
2443         if (type & PERF_SAMPLE_ADDR)
2444                 result += sizeof(u64);
2445
2446         if (type & PERF_SAMPLE_ID)
2447                 result += sizeof(u64);
2448
2449         if (type & PERF_SAMPLE_STREAM_ID)
2450                 result += sizeof(u64);
2451
2452         if (type & PERF_SAMPLE_CPU)
2453                 result += sizeof(u64);
2454
2455         if (type & PERF_SAMPLE_PERIOD)
2456                 result += sizeof(u64);
2457
2458         if (type & PERF_SAMPLE_READ) {
2459                 result += sizeof(u64);
2460                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2461                         result += sizeof(u64);
2462                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2463                         result += sizeof(u64);
2464                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2465                 if (read_format & PERF_FORMAT_GROUP) {
2466                         sz = sample->read.group.nr *
2467                              sizeof(struct sample_read_value);
2468                         result += sz;
2469                 } else {
2470                         result += sizeof(u64);
2471                 }
2472         }
2473
2474         if (type & PERF_SAMPLE_CALLCHAIN) {
2475                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2476                 result += sz;
2477         }
2478
2479         if (type & PERF_SAMPLE_RAW) {
2480                 result += sizeof(u32);
2481                 result += sample->raw_size;
2482         }
2483
2484         if (type & PERF_SAMPLE_BRANCH_STACK) {
2485                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2486                 sz += sizeof(u64);
2487                 result += sz;
2488         }
2489
2490         if (type & PERF_SAMPLE_REGS_USER) {
2491                 if (sample->user_regs.abi) {
2492                         result += sizeof(u64);
2493                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2494                         result += sz;
2495                 } else {
2496                         result += sizeof(u64);
2497                 }
2498         }
2499
2500         if (type & PERF_SAMPLE_STACK_USER) {
2501                 sz = sample->user_stack.size;
2502                 result += sizeof(u64);
2503                 if (sz) {
2504                         result += sz;
2505                         result += sizeof(u64);
2506                 }
2507         }
2508
2509         if (type & PERF_SAMPLE_WEIGHT)
2510                 result += sizeof(u64);
2511
2512         if (type & PERF_SAMPLE_DATA_SRC)
2513                 result += sizeof(u64);
2514
2515         if (type & PERF_SAMPLE_TRANSACTION)
2516                 result += sizeof(u64);
2517
2518         if (type & PERF_SAMPLE_REGS_INTR) {
2519                 if (sample->intr_regs.abi) {
2520                         result += sizeof(u64);
2521                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2522                         result += sz;
2523                 } else {
2524                         result += sizeof(u64);
2525                 }
2526         }
2527
2528         if (type & PERF_SAMPLE_PHYS_ADDR)
2529                 result += sizeof(u64);
2530
2531         return result;
2532 }
2533
2534 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2535                                   u64 read_format,
2536                                   const struct perf_sample *sample)
2537 {
2538         u64 *array;
2539         size_t sz;
2540         /*
2541          * used for cross-endian analysis. See git commit 65014ab3
2542          * for why this goofiness is needed.
2543          */
2544         union u64_swap u;
2545
2546         array = event->sample.array;
2547
2548         if (type & PERF_SAMPLE_IDENTIFIER) {
2549                 *array = sample->id;
2550                 array++;
2551         }
2552
2553         if (type & PERF_SAMPLE_IP) {
2554                 *array = sample->ip;
2555                 array++;
2556         }
2557
2558         if (type & PERF_SAMPLE_TID) {
2559                 u.val32[0] = sample->pid;
2560                 u.val32[1] = sample->tid;
2561                 *array = u.val64;
2562                 array++;
2563         }
2564
2565         if (type & PERF_SAMPLE_TIME) {
2566                 *array = sample->time;
2567                 array++;
2568         }
2569
2570         if (type & PERF_SAMPLE_ADDR) {
2571                 *array = sample->addr;
2572                 array++;
2573         }
2574
2575         if (type & PERF_SAMPLE_ID) {
2576                 *array = sample->id;
2577                 array++;
2578         }
2579
2580         if (type & PERF_SAMPLE_STREAM_ID) {
2581                 *array = sample->stream_id;
2582                 array++;
2583         }
2584
2585         if (type & PERF_SAMPLE_CPU) {
2586                 u.val32[0] = sample->cpu;
2587                 u.val32[1] = 0;
2588                 *array = u.val64;
2589                 array++;
2590         }
2591
2592         if (type & PERF_SAMPLE_PERIOD) {
2593                 *array = sample->period;
2594                 array++;
2595         }
2596
2597         if (type & PERF_SAMPLE_READ) {
2598                 if (read_format & PERF_FORMAT_GROUP)
2599                         *array = sample->read.group.nr;
2600                 else
2601                         *array = sample->read.one.value;
2602                 array++;
2603
2604                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2605                         *array = sample->read.time_enabled;
2606                         array++;
2607                 }
2608
2609                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2610                         *array = sample->read.time_running;
2611                         array++;
2612                 }
2613
2614                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2615                 if (read_format & PERF_FORMAT_GROUP) {
2616                         sz = sample->read.group.nr *
2617                              sizeof(struct sample_read_value);
2618                         memcpy(array, sample->read.group.values, sz);
2619                         array = (void *)array + sz;
2620                 } else {
2621                         *array = sample->read.one.id;
2622                         array++;
2623                 }
2624         }
2625
2626         if (type & PERF_SAMPLE_CALLCHAIN) {
2627                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2628                 memcpy(array, sample->callchain, sz);
2629                 array = (void *)array + sz;
2630         }
2631
2632         if (type & PERF_SAMPLE_RAW) {
2633                 u.val32[0] = sample->raw_size;
2634                 *array = u.val64;
2635                 array = (void *)array + sizeof(u32);
2636
2637                 memcpy(array, sample->raw_data, sample->raw_size);
2638                 array = (void *)array + sample->raw_size;
2639         }
2640
2641         if (type & PERF_SAMPLE_BRANCH_STACK) {
2642                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2643                 sz += sizeof(u64);
2644                 memcpy(array, sample->branch_stack, sz);
2645                 array = (void *)array + sz;
2646         }
2647
2648         if (type & PERF_SAMPLE_REGS_USER) {
2649                 if (sample->user_regs.abi) {
2650                         *array++ = sample->user_regs.abi;
2651                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2652                         memcpy(array, sample->user_regs.regs, sz);
2653                         array = (void *)array + sz;
2654                 } else {
2655                         *array++ = 0;
2656                 }
2657         }
2658
2659         if (type & PERF_SAMPLE_STACK_USER) {
2660                 sz = sample->user_stack.size;
2661                 *array++ = sz;
2662                 if (sz) {
2663                         memcpy(array, sample->user_stack.data, sz);
2664                         array = (void *)array + sz;
2665                         *array++ = sz;
2666                 }
2667         }
2668
2669         if (type & PERF_SAMPLE_WEIGHT) {
2670                 *array = sample->weight;
2671                 array++;
2672         }
2673
2674         if (type & PERF_SAMPLE_DATA_SRC) {
2675                 *array = sample->data_src;
2676                 array++;
2677         }
2678
2679         if (type & PERF_SAMPLE_TRANSACTION) {
2680                 *array = sample->transaction;
2681                 array++;
2682         }
2683
2684         if (type & PERF_SAMPLE_REGS_INTR) {
2685                 if (sample->intr_regs.abi) {
2686                         *array++ = sample->intr_regs.abi;
2687                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2688                         memcpy(array, sample->intr_regs.regs, sz);
2689                         array = (void *)array + sz;
2690                 } else {
2691                         *array++ = 0;
2692                 }
2693         }
2694
2695         if (type & PERF_SAMPLE_PHYS_ADDR) {
2696                 *array = sample->phys_addr;
2697                 array++;
2698         }
2699
2700         return 0;
2701 }
2702
2703 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2704 {
2705         return tep_find_field(evsel->tp_format, name);
2706 }
2707
2708 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2709                          const char *name)
2710 {
2711         struct tep_format_field *field = perf_evsel__field(evsel, name);
2712         int offset;
2713
2714         if (!field)
2715                 return NULL;
2716
2717         offset = field->offset;
2718
2719         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2720                 offset = *(int *)(sample->raw_data + field->offset);
2721                 offset &= 0xffff;
2722         }
2723
2724         return sample->raw_data + offset;
2725 }
2726
2727 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2728                          bool needs_swap)
2729 {
2730         u64 value;
2731         void *ptr = sample->raw_data + field->offset;
2732
2733         switch (field->size) {
2734         case 1:
2735                 return *(u8 *)ptr;
2736         case 2:
2737                 value = *(u16 *)ptr;
2738                 break;
2739         case 4:
2740                 value = *(u32 *)ptr;
2741                 break;
2742         case 8:
2743                 memcpy(&value, ptr, sizeof(u64));
2744                 break;
2745         default:
2746                 return 0;
2747         }
2748
2749         if (!needs_swap)
2750                 return value;
2751
2752         switch (field->size) {
2753         case 2:
2754                 return bswap_16(value);
2755         case 4:
2756                 return bswap_32(value);
2757         case 8:
2758                 return bswap_64(value);
2759         default:
2760                 return 0;
2761         }
2762
2763         return 0;
2764 }
2765
2766 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2767                        const char *name)
2768 {
2769         struct tep_format_field *field = perf_evsel__field(evsel, name);
2770
2771         if (!field)
2772                 return 0;
2773
2774         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2775 }
2776
2777 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2778                           char *msg, size_t msgsize)
2779 {
2780         int paranoid;
2781
2782         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2783             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2784             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2785                 /*
2786                  * If it's cycles then fall back to hrtimer based
2787                  * cpu-clock-tick sw counter, which is always available even if
2788                  * no PMU support.
2789                  *
2790                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2791                  * b0a873e).
2792                  */
2793                 scnprintf(msg, msgsize, "%s",
2794 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2795
2796                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2797                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2798
2799                 zfree(&evsel->name);
2800                 return true;
2801         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2802                    (paranoid = perf_event_paranoid()) > 1) {
2803                 const char *name = perf_evsel__name(evsel);
2804                 char *new_name;
2805                 const char *sep = ":";
2806
2807                 /* Is there already the separator in the name. */
2808                 if (strchr(name, '/') ||
2809                     strchr(name, ':'))
2810                         sep = "";
2811
2812                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2813                         return false;
2814
2815                 if (evsel->name)
2816                         free(evsel->name);
2817                 evsel->name = new_name;
2818                 scnprintf(msg, msgsize,
2819 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2820                 evsel->attr.exclude_kernel = 1;
2821
2822                 return true;
2823         }
2824
2825         return false;
2826 }
2827
2828 static bool find_process(const char *name)
2829 {
2830         size_t len = strlen(name);
2831         DIR *dir;
2832         struct dirent *d;
2833         int ret = -1;
2834
2835         dir = opendir(procfs__mountpoint());
2836         if (!dir)
2837                 return false;
2838
2839         /* Walk through the directory. */
2840         while (ret && (d = readdir(dir)) != NULL) {
2841                 char path[PATH_MAX];
2842                 char *data;
2843                 size_t size;
2844
2845                 if ((d->d_type != DT_DIR) ||
2846                      !strcmp(".", d->d_name) ||
2847                      !strcmp("..", d->d_name))
2848                         continue;
2849
2850                 scnprintf(path, sizeof(path), "%s/%s/comm",
2851                           procfs__mountpoint(), d->d_name);
2852
2853                 if (filename__read_str(path, &data, &size))
2854                         continue;
2855
2856                 ret = strncmp(name, data, len);
2857                 free(data);
2858         }
2859
2860         closedir(dir);
2861         return ret ? false : true;
2862 }
2863
2864 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2865                               int err, char *msg, size_t size)
2866 {
2867         char sbuf[STRERR_BUFSIZE];
2868         int printed = 0;
2869
2870         switch (err) {
2871         case EPERM:
2872         case EACCES:
2873                 if (err == EPERM)
2874                         printed = scnprintf(msg, size,
2875                                 "No permission to enable %s event.\n\n",
2876                                 perf_evsel__name(evsel));
2877
2878                 return scnprintf(msg + printed, size - printed,
2879                  "You may not have permission to collect %sstats.\n\n"
2880                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2881                  "which controls use of the performance events system by\n"
2882                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2883                  "The current value is %d:\n\n"
2884                  "  -1: Allow use of (almost) all events by all users\n"
2885                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2886                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2887                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2888                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2889                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2890                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2891                  "      kernel.perf_event_paranoid = -1\n" ,
2892                                  target->system_wide ? "system-wide " : "",
2893                                  perf_event_paranoid());
2894         case ENOENT:
2895                 return scnprintf(msg, size, "The %s event is not supported.",
2896                                  perf_evsel__name(evsel));
2897         case EMFILE:
2898                 return scnprintf(msg, size, "%s",
2899                          "Too many events are opened.\n"
2900                          "Probably the maximum number of open file descriptors has been reached.\n"
2901                          "Hint: Try again after reducing the number of events.\n"
2902                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2903         case ENOMEM:
2904                 if (evsel__has_callchain(evsel) &&
2905                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2906                         return scnprintf(msg, size,
2907                                          "Not enough memory to setup event with callchain.\n"
2908                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2909                                          "Hint: Current value: %d", sysctl__max_stack());
2910                 break;
2911         case ENODEV:
2912                 if (target->cpu_list)
2913                         return scnprintf(msg, size, "%s",
2914          "No such device - did you specify an out-of-range profile CPU?");
2915                 break;
2916         case EOPNOTSUPP:
2917                 if (evsel->attr.sample_period != 0)
2918                         return scnprintf(msg, size,
2919         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2920                                          perf_evsel__name(evsel));
2921                 if (evsel->attr.precise_ip)
2922                         return scnprintf(msg, size, "%s",
2923         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2924 #if defined(__i386__) || defined(__x86_64__)
2925                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2926                         return scnprintf(msg, size, "%s",
2927         "No hardware sampling interrupt available.\n");
2928 #endif
2929                 break;
2930         case EBUSY:
2931                 if (find_process("oprofiled"))
2932                         return scnprintf(msg, size,
2933         "The PMU counters are busy/taken by another profiler.\n"
2934         "We found oprofile daemon running, please stop it and try again.");
2935                 break;
2936         case EINVAL:
2937                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2938                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2939                 if (perf_missing_features.clockid)
2940                         return scnprintf(msg, size, "clockid feature not supported.");
2941                 if (perf_missing_features.clockid_wrong)
2942                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2943                 break;
2944         default:
2945                 break;
2946         }
2947
2948         return scnprintf(msg, size,
2949         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2950         "/bin/dmesg | grep -i perf may provide additional information.\n",
2951                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2952                          perf_evsel__name(evsel));
2953 }
2954
2955 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2956 {
2957         if (evsel && evsel->evlist)
2958                 return evsel->evlist->env;
2959         return NULL;
2960 }
2961
2962 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
2963 {
2964         int cpu, thread;
2965
2966         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
2967                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
2968                      thread++) {
2969                         int fd = FD(evsel, cpu, thread);
2970
2971                         if (perf_evlist__id_add_fd(evlist, evsel,
2972                                                    cpu, thread, fd) < 0)
2973                                 return -1;
2974                 }
2975         }
2976
2977         return 0;
2978 }
2979
2980 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
2981 {
2982         struct cpu_map *cpus = evsel->cpus;
2983         struct thread_map *threads = evsel->threads;
2984
2985         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
2986                 return -ENOMEM;
2987
2988         return store_evsel_ids(evsel, evlist);
2989 }