tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "memswap.h"
  40 #include "util/parse-branch-options.h"
  41
  42 #include "sane_ctype.h"
  43
  44 struct perf_missing_features perf_missing_features;
  45
  46 static clockid_t clockid;
  47
  48 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  49 {
  50         return 0;
  51 }
  52
  53 void __weak test_attr__ready(void) { }
  54
  55 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  56 {
  57 }
  58
  59 static struct {
  60         size_t  size;
  61         int     (*init)(struct perf_evsel *evsel);
  62         void    (*fini)(struct perf_evsel *evsel);
  63 } perf_evsel__object = {
  64         .size = sizeof(struct perf_evsel),
  65         .init = perf_evsel__no_extra_init,
  66         .fini = perf_evsel__no_extra_fini,
  67 };
  68
  69 int perf_evsel__object_config(size_t object_size,
  70                               int (*init)(struct perf_evsel *evsel),
  71                               void (*fini)(struct perf_evsel *evsel))
  72 {
  73
  74         if (object_size == 0)
  75                 goto set_methods;
  76
  77         if (perf_evsel__object.size > object_size)
  78                 return -EINVAL;
  79
  80         perf_evsel__object.size = object_size;
  81
  82 set_methods:
  83         if (init != NULL)
  84                 perf_evsel__object.init = init;
  85
  86         if (fini != NULL)
  87                 perf_evsel__object.fini = fini;
  88
  89         return 0;
  90 }
  91
  92 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  93
  94 int __perf_evsel__sample_size(u64 sample_type)
  95 {
  96         u64 mask = sample_type & PERF_SAMPLE_MASK;
  97         int size = 0;
  98         int i;
  99
 100         for (i = 0; i < 64; i++) {
 101                 if (mask & (1ULL << i))
 102                         size++;
 103         }
 104
 105         size *= sizeof(u64);
 106
 107         return size;
 108 }
 109
 110 /**
 111  * __perf_evsel__calc_id_pos - calculate id_pos.
 112  * @sample_type: sample type
 113  *
 114  * This function returns the position of the event id (PERF_SAMPLE_ID or
 115  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 116  * sample_event.
 117  */
 118 static int __perf_evsel__calc_id_pos(u64 sample_type)
 119 {
 120         int idx = 0;
 121
 122         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 123                 return 0;
 124
 125         if (!(sample_type & PERF_SAMPLE_ID))
 126                 return -1;
 127
 128         if (sample_type & PERF_SAMPLE_IP)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_TID)
 132                 idx += 1;
 133
 134         if (sample_type & PERF_SAMPLE_TIME)
 135                 idx += 1;
 136
 137         if (sample_type & PERF_SAMPLE_ADDR)
 138                 idx += 1;
 139
 140         return idx;
 141 }
 142
 143 /**
 144  * __perf_evsel__calc_is_pos - calculate is_pos.
 145  * @sample_type: sample type
 146  *
 147  * This function returns the position (counting backwards) of the event id
 148  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 149  * sample_id_all is used there is an id sample appended to non-sample events.
 150  */
 151 static int __perf_evsel__calc_is_pos(u64 sample_type)
 152 {
 153         int idx = 1;
 154
 155         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 156                 return 1;
 157
 158         if (!(sample_type & PERF_SAMPLE_ID))
 159                 return -1;
 160
 161         if (sample_type & PERF_SAMPLE_CPU)
 162                 idx += 1;
 163
 164         if (sample_type & PERF_SAMPLE_STREAM_ID)
 165                 idx += 1;
 166
 167         return idx;
 168 }
 169
 170 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 171 {
 172         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 173         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 174 }
 175
 176 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 177                                   enum perf_event_sample_format bit)
 178 {
 179         if (!(evsel->attr.sample_type & bit)) {
 180                 evsel->attr.sample_type |= bit;
 181                 evsel->sample_size += sizeof(u64);
 182                 perf_evsel__calc_id_pos(evsel);
 183         }
 184 }
 185
 186 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 187                                     enum perf_event_sample_format bit)
 188 {
 189         if (evsel->attr.sample_type & bit) {
 190                 evsel->attr.sample_type &= ~bit;
 191                 evsel->sample_size -= sizeof(u64);
 192                 perf_evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 197                                bool can_sample_identifier)
 198 {
 199         if (can_sample_identifier) {
 200                 perf_evsel__reset_sample_bit(evsel, ID);
 201                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 202         } else {
 203                 perf_evsel__set_sample_bit(evsel, ID);
 204         }
 205         evsel->attr.read_format |= PERF_FORMAT_ID;
 206 }
 207
 208 /**
 209  * perf_evsel__is_function_event - Return whether given evsel is a function
 210  * trace event
 211  *
 212  * @evsel - evsel selector to be tested
 213  *
 214  * Return %true if event is function trace event
 215  */
 216 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 217 {
 218 #define FUNCTION_EVENT "ftrace:function"
 219
 220         return evsel->name &&
 221                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 222
 223 #undef FUNCTION_EVENT
 224 }
 225
 226 void perf_evsel__init(struct perf_evsel *evsel,
 227                       struct perf_event_attr *attr, int idx)
 228 {
 229         evsel->idx         = idx;
 230         evsel->tracking    = !idx;
 231         evsel->attr        = *attr;
 232         evsel->leader      = evsel;
 233         evsel->unit        = "";
 234         evsel->scale       = 1.0;
 235         evsel->max_events  = ULONG_MAX;
 236         evsel->evlist      = NULL;
 237         evsel->bpf_fd      = -1;
 238         INIT_LIST_HEAD(&evsel->node);
 239         INIT_LIST_HEAD(&evsel->config_terms);
 240         perf_evsel__object.init(evsel);
 241         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 242         perf_evsel__calc_id_pos(evsel);
 243         evsel->cmdline_group_boundary = false;
 244         evsel->metric_expr   = NULL;
 245         evsel->metric_name   = NULL;
 246         evsel->metric_events = NULL;
 247         evsel->collect_stat  = false;
 248         evsel->pmu_name      = NULL;
 249 }
 250
 251 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 252 {
 253         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 254
 255         if (!evsel)
 256                 return NULL;
 257         perf_evsel__init(evsel, attr, idx);
 258
 259         if (perf_evsel__is_bpf_output(evsel)) {
 260                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 261                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 262                 evsel->attr.sample_period = 1;
 263         }
 264
 265         if (perf_evsel__is_clock(evsel)) {
 266                 /*
 267                  * The evsel->unit points to static alias->unit
 268                  * so it's ok to use static string in here.
 269                  */
 270                 static const char *unit = "msec";
 271
 272                 evsel->unit = unit;
 273                 evsel->scale = 1e-6;
 274         }
 275
 276         return evsel;
 277 }
 278
 279 static bool perf_event_can_profile_kernel(void)
 280 {
 281         return geteuid() == 0 || perf_event_paranoid() == -1;
 282 }
 283
 284 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 285 {
 286         struct perf_event_attr attr = {
 287                 .type   = PERF_TYPE_HARDWARE,
 288                 .config = PERF_COUNT_HW_CPU_CYCLES,
 289                 .exclude_kernel = !perf_event_can_profile_kernel(),
 290         };
 291         struct perf_evsel *evsel;
 292
 293         event_attr_init(&attr);
 294
 295         if (!precise)
 296                 goto new_event;
 297         /*
 298          * Unnamed union member, not supported as struct member named
 299          * initializer in older compilers such as gcc 4.4.7
 300          *
 301          * Just for probing the precise_ip:
 302          */
 303         attr.sample_period = 1;
 304
 305         perf_event_attr__set_max_precise_ip(&attr);
 306         /*
 307          * Now let the usual logic to set up the perf_event_attr defaults
 308          * to kick in when we return and before perf_evsel__open() is called.
 309          */
 310         attr.sample_period = 0;
 311 new_event:
 312         evsel = perf_evsel__new(&attr);
 313         if (evsel == NULL)
 314                 goto out;
 315
 316         /* use asprintf() because free(evsel) assumes name is allocated */
 317         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 318                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 319                      attr.exclude_kernel ? "u" : "",
 320                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 321                 goto error_free;
 322 out:
 323         return evsel;
 324 error_free:
 325         perf_evsel__delete(evsel);
 326         evsel = NULL;
 327         goto out;
 328 }
 329
 330 /*
 331  * Returns pointer with encoded error via <linux/err.h> interface.
 332  */
 333 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 334 {
 335         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 336         int err = -ENOMEM;
 337
 338         if (evsel == NULL) {
 339                 goto out_err;
 340         } else {
 341                 struct perf_event_attr attr = {
 342                         .type          = PERF_TYPE_TRACEPOINT,
 343                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 344                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 345                 };
 346
 347                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 348                         goto out_free;
 349
 350                 evsel->tp_format = trace_event__tp_format(sys, name);
 351                 if (IS_ERR(evsel->tp_format)) {
 352                         err = PTR_ERR(evsel->tp_format);
 353                         goto out_free;
 354                 }
 355
 356                 event_attr_init(&attr);
 357                 attr.config = evsel->tp_format->id;
 358                 attr.sample_period = 1;
 359                 perf_evsel__init(evsel, &attr, idx);
 360         }
 361
 362         return evsel;
 363
 364 out_free:
 365         zfree(&evsel->name);
 366         free(evsel);
 367 out_err:
 368         return ERR_PTR(err);
 369 }
 370
 371 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 372         "cycles",
 373         "instructions",
 374         "cache-references",
 375         "cache-misses",
 376         "branches",
 377         "branch-misses",
 378         "bus-cycles",
 379         "stalled-cycles-frontend",
 380         "stalled-cycles-backend",
 381         "ref-cycles",
 382 };
 383
 384 static const char *__perf_evsel__hw_name(u64 config)
 385 {
 386         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 387                 return perf_evsel__hw_names[config];
 388
 389         return "unknown-hardware";
 390 }
 391
 392 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 393 {
 394         int colon = 0, r = 0;
 395         struct perf_event_attr *attr = &evsel->attr;
 396         bool exclude_guest_default = false;
 397
 398 #define MOD_PRINT(context, mod) do {                                    \
 399                 if (!attr->exclude_##context) {                         \
 400                         if (!colon) colon = ++r;                        \
 401                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 402                 } } while(0)
 403
 404         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 405                 MOD_PRINT(kernel, 'k');
 406                 MOD_PRINT(user, 'u');
 407                 MOD_PRINT(hv, 'h');
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->precise_ip) {
 412                 if (!colon)
 413                         colon = ++r;
 414                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 415                 exclude_guest_default = true;
 416         }
 417
 418         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 419                 MOD_PRINT(host, 'H');
 420                 MOD_PRINT(guest, 'G');
 421         }
 422 #undef MOD_PRINT
 423         if (colon)
 424                 bf[colon - 1] = ':';
 425         return r;
 426 }
 427
 428 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 429 {
 430         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 431         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 432 }
 433
 434 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 435         "cpu-clock",
 436         "task-clock",
 437         "page-faults",
 438         "context-switches",
 439         "cpu-migrations",
 440         "minor-faults",
 441         "major-faults",
 442         "alignment-faults",
 443         "emulation-faults",
 444         "dummy",
 445 };
 446
 447 static const char *__perf_evsel__sw_name(u64 config)
 448 {
 449         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 450                 return perf_evsel__sw_names[config];
 451         return "unknown-software";
 452 }
 453
 454 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 455 {
 456         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 457         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 458 }
 459
 460 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 461 {
 462         int r;
 463
 464         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 465
 466         if (type & HW_BREAKPOINT_R)
 467                 r += scnprintf(bf + r, size - r, "r");
 468
 469         if (type & HW_BREAKPOINT_W)
 470                 r += scnprintf(bf + r, size - r, "w");
 471
 472         if (type & HW_BREAKPOINT_X)
 473                 r += scnprintf(bf + r, size - r, "x");
 474
 475         return r;
 476 }
 477
 478 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 479 {
 480         struct perf_event_attr *attr = &evsel->attr;
 481         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 482         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 483 }
 484
 485 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 486                                 [PERF_EVSEL__MAX_ALIASES] = {
 487  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 488  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 489  { "LLC",       "L2",                                                   },
 490  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 491  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 492  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 493  { "node",                                                              },
 494 };
 495
 496 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 497                                    [PERF_EVSEL__MAX_ALIASES] = {
 498  { "load",      "loads",        "read",                                 },
 499  { "store",     "stores",       "write",                                },
 500  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 501 };
 502
 503 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 504                                        [PERF_EVSEL__MAX_ALIASES] = {
 505  { "refs",      "Reference",    "ops",          "access",               },
 506  { "misses",    "miss",                                                 },
 507 };
 508
 509 #define C(x)            PERF_COUNT_HW_CACHE_##x
 510 #define CACHE_READ      (1 << C(OP_READ))
 511 #define CACHE_WRITE     (1 << C(OP_WRITE))
 512 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 513 #define COP(x)          (1 << x)
 514
 515 /*
 516  * cache operartion stat
 517  * L1I : Read and prefetch only
 518  * ITLB and BPU : Read-only
 519  */
 520 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 521  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 522  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 523  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(ITLB)]      = (CACHE_READ),
 526  [C(BPU)]       = (CACHE_READ),
 527  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 528 };
 529
 530 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 531 {
 532         if (perf_evsel__hw_cache_stat[type] & COP(op))
 533                 return true;    /* valid */
 534         else
 535                 return false;   /* invalid */
 536 }
 537
 538 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 539                                             char *bf, size_t size)
 540 {
 541         if (result) {
 542                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 543                                  perf_evsel__hw_cache_op[op][0],
 544                                  perf_evsel__hw_cache_result[result][0]);
 545         }
 546
 547         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 548                          perf_evsel__hw_cache_op[op][1]);
 549 }
 550
 551 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 552 {
 553         u8 op, result, type = (config >>  0) & 0xff;
 554         const char *err = "unknown-ext-hardware-cache-type";
 555
 556         if (type >= PERF_COUNT_HW_CACHE_MAX)
 557                 goto out_err;
 558
 559         op = (config >>  8) & 0xff;
 560         err = "unknown-ext-hardware-cache-op";
 561         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 562                 goto out_err;
 563
 564         result = (config >> 16) & 0xff;
 565         err = "unknown-ext-hardware-cache-result";
 566         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 567                 goto out_err;
 568
 569         err = "invalid-cache";
 570         if (!perf_evsel__is_cache_op_valid(type, op))
 571                 goto out_err;
 572
 573         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 574 out_err:
 575         return scnprintf(bf, size, "%s", err);
 576 }
 577
 578 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 579 {
 580         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 581         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 582 }
 583
 584 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 585 {
 586         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 587         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 588 }
 589
 590 const char *perf_evsel__name(struct perf_evsel *evsel)
 591 {
 592         char bf[128];
 593
 594         if (evsel->name)
 595                 return evsel->name;
 596
 597         switch (evsel->attr.type) {
 598         case PERF_TYPE_RAW:
 599                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 600                 break;
 601
 602         case PERF_TYPE_HARDWARE:
 603                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 604                 break;
 605
 606         case PERF_TYPE_HW_CACHE:
 607                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 608                 break;
 609
 610         case PERF_TYPE_SOFTWARE:
 611                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 612                 break;
 613
 614         case PERF_TYPE_TRACEPOINT:
 615                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 616                 break;
 617
 618         case PERF_TYPE_BREAKPOINT:
 619                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 620                 break;
 621
 622         default:
 623                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 624                           evsel->attr.type);
 625                 break;
 626         }
 627
 628         evsel->name = strdup(bf);
 629
 630         return evsel->name ?: "unknown";
 631 }
 632
 633 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 634 {
 635         return evsel->group_name ?: "anon group";
 636 }
 637
 638 /*
 639  * Returns the group details for the specified leader,
 640  * with following rules.
 641  *
 642  *  For record -e '{cycles,instructions}'
 643  *    'anon group { cycles:u, instructions:u }'
 644  *
 645  *  For record -e 'cycles,instructions' and report --group
 646  *    'cycles:u, instructions:u'
 647  */
 648 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 649 {
 650         int ret = 0;
 651         struct perf_evsel *pos;
 652         const char *group_name = perf_evsel__group_name(evsel);
 653
 654         if (!evsel->forced_leader)
 655                 ret = scnprintf(buf, size, "%s { ", group_name);
 656
 657         ret += scnprintf(buf + ret, size - ret, "%s",
 658                          perf_evsel__name(evsel));
 659
 660         for_each_group_member(pos, evsel)
 661                 ret += scnprintf(buf + ret, size - ret, ", %s",
 662                                  perf_evsel__name(pos));
 663
 664         if (!evsel->forced_leader)
 665                 ret += scnprintf(buf + ret, size - ret, " }");
 666
 667         return ret;
 668 }
 669
 670 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 671                                            struct record_opts *opts,
 672                                            struct callchain_param *param)
 673 {
 674         bool function = perf_evsel__is_function_event(evsel);
 675         struct perf_event_attr *attr = &evsel->attr;
 676
 677         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 678
 679         attr->sample_max_stack = param->max_stack;
 680
 681         if (param->record_mode == CALLCHAIN_LBR) {
 682                 if (!opts->branch_stack) {
 683                         if (attr->exclude_user) {
 684                                 pr_warning("LBR callstack option is only available "
 685                                            "to get user callchain information. "
 686                                            "Falling back to framepointers.\n");
 687                         } else {
 688                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 689                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 690                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 691                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 692                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 693                         }
 694                 } else
 695                          pr_warning("Cannot use LBR callstack with branch stack. "
 696                                     "Falling back to framepointers.\n");
 697         }
 698
 699         if (param->record_mode == CALLCHAIN_DWARF) {
 700                 if (!function) {
 701                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 702                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 703                         attr->sample_regs_user |= PERF_REGS_MASK;
 704                         attr->sample_stack_user = param->dump_size;
 705                         attr->exclude_callchain_user = 1;
 706                 } else {
 707                         pr_info("Cannot use DWARF unwind for function trace event,"
 708                                 " falling back to framepointers.\n");
 709                 }
 710         }
 711
 712         if (function) {
 713                 pr_info("Disabling user space callchains for function trace event.\n");
 714                 attr->exclude_callchain_user = 1;
 715         }
 716 }
 717
 718 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 719                                   struct record_opts *opts,
 720                                   struct callchain_param *param)
 721 {
 722         if (param->enabled)
 723                 return __perf_evsel__config_callchain(evsel, opts, param);
 724 }
 725
 726 static void
 727 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 728                             struct callchain_param *param)
 729 {
 730         struct perf_event_attr *attr = &evsel->attr;
 731
 732         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 733         if (param->record_mode == CALLCHAIN_LBR) {
 734                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 735                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 736                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 737         }
 738         if (param->record_mode == CALLCHAIN_DWARF) {
 739                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 740                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 741         }
 742 }
 743
 744 static void apply_config_terms(struct perf_evsel *evsel,
 745                                struct record_opts *opts, bool track)
 746 {
 747         struct perf_evsel_config_term *term;
 748         struct list_head *config_terms = &evsel->config_terms;
 749         struct perf_event_attr *attr = &evsel->attr;
 750         /* callgraph default */
 751         struct callchain_param param = {
 752                 .record_mode = callchain_param.record_mode,
 753         };
 754         u32 dump_size = 0;
 755         int max_stack = 0;
 756         const char *callgraph_buf = NULL;
 757
 758         list_for_each_entry(term, config_terms, list) {
 759                 switch (term->type) {
 760                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 761                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 762                                 attr->sample_period = term->val.period;
 763                                 attr->freq = 0;
 764                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 765                         }
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 768                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 769                                 attr->sample_freq = term->val.freq;
 770                                 attr->freq = 1;
 771                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 772                         }
 773                         break;
 774                 case PERF_EVSEL__CONFIG_TERM_TIME:
 775                         if (term->val.time)
 776                                 perf_evsel__set_sample_bit(evsel, TIME);
 777                         else
 778                                 perf_evsel__reset_sample_bit(evsel, TIME);
 779                         break;
 780                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 781                         callgraph_buf = term->val.callgraph;
 782                         break;
 783                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 784                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 785                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 786                                 parse_branch_str(term->val.branch,
 787                                                  &attr->branch_sample_type);
 788                         } else
 789                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 790                         break;
 791                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 792                         dump_size = term->val.stack_user;
 793                         break;
 794                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 795                         max_stack = term->val.max_stack;
 796                         break;
 797                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 798                         evsel->max_events = term->val.max_events;
 799                         break;
 800                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 801                         /*
 802                          * attr->inherit should has already been set by
 803                          * perf_evsel__config. If user explicitly set
 804                          * inherit using config terms, override global
 805                          * opt->no_inherit setting.
 806                          */
 807                         attr->inherit = term->val.inherit ? 1 : 0;
 808                         break;
 809                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 810                         attr->write_backward = term->val.overwrite ? 1 : 0;
 811                         break;
 812                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 813                         break;
 814                 default:
 815                         break;
 816                 }
 817         }
 818
 819         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 820         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 821                 bool sample_address = false;
 822
 823                 if (max_stack) {
 824                         param.max_stack = max_stack;
 825                         if (callgraph_buf == NULL)
 826                                 callgraph_buf = "fp";
 827                 }
 828
 829                 /* parse callgraph parameters */
 830                 if (callgraph_buf != NULL) {
 831                         if (!strcmp(callgraph_buf, "no")) {
 832                                 param.enabled = false;
 833                                 param.record_mode = CALLCHAIN_NONE;
 834                         } else {
 835                                 param.enabled = true;
 836                                 if (parse_callchain_record(callgraph_buf, &param)) {
 837                                         pr_err("per-event callgraph setting for %s failed. "
 838                                                "Apply callgraph global setting for it\n",
 839                                                evsel->name);
 840                                         return;
 841                                 }
 842                                 if (param.record_mode == CALLCHAIN_DWARF)
 843                                         sample_address = true;
 844                         }
 845                 }
 846                 if (dump_size > 0) {
 847                         dump_size = round_up(dump_size, sizeof(u64));
 848                         param.dump_size = dump_size;
 849                 }
 850
 851                 /* If global callgraph set, clear it */
 852                 if (callchain_param.enabled)
 853                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 854
 855                 /* set perf-event callgraph */
 856                 if (param.enabled) {
 857                         if (sample_address) {
 858                                 perf_evsel__set_sample_bit(evsel, ADDR);
 859                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 860                                 evsel->attr.mmap_data = track;
 861                         }
 862                         perf_evsel__config_callchain(evsel, opts, &param);
 863                 }
 864         }
 865 }
 866
 867 static bool is_dummy_event(struct perf_evsel *evsel)
 868 {
 869         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 870                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 871 }
 872
 873 /*
 874  * The enable_on_exec/disabled value strategy:
 875  *
 876  *  1) For any type of traced program:
 877  *    - all independent events and group leaders are disabled
 878  *    - all group members are enabled
 879  *
 880  *     Group members are ruled by group leaders. They need to
 881  *     be enabled, because the group scheduling relies on that.
 882  *
 883  *  2) For traced programs executed by perf:
 884  *     - all independent events and group leaders have
 885  *       enable_on_exec set
 886  *     - we don't specifically enable or disable any event during
 887  *       the record command
 888  *
 889  *     Independent events and group leaders are initially disabled
 890  *     and get enabled by exec. Group members are ruled by group
 891  *     leaders as stated in 1).
 892  *
 893  *  3) For traced programs attached by perf (pid/tid):
 894  *     - we specifically enable or disable all events during
 895  *       the record command
 896  *
 897  *     When attaching events to already running traced we
 898  *     enable/disable events specifically, as there's no
 899  *     initial traced exec call.
 900  */
 901 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 902                         struct callchain_param *callchain)
 903 {
 904         struct perf_evsel *leader = evsel->leader;
 905         struct perf_event_attr *attr = &evsel->attr;
 906         int track = evsel->tracking;
 907         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 908
 909         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 910         attr->inherit       = !opts->no_inherit;
 911         attr->write_backward = opts->overwrite ? 1 : 0;
 912
 913         perf_evsel__set_sample_bit(evsel, IP);
 914         perf_evsel__set_sample_bit(evsel, TID);
 915
 916         if (evsel->sample_read) {
 917                 perf_evsel__set_sample_bit(evsel, READ);
 918
 919                 /*
 920                  * We need ID even in case of single event, because
 921                  * PERF_SAMPLE_READ process ID specific data.
 922                  */
 923                 perf_evsel__set_sample_id(evsel, false);
 924
 925                 /*
 926                  * Apply group format only if we belong to group
 927                  * with more than one members.
 928                  */
 929                 if (leader->nr_members > 1) {
 930                         attr->read_format |= PERF_FORMAT_GROUP;
 931                         attr->inherit = 0;
 932                 }
 933         }
 934
 935         /*
 936          * We default some events to have a default interval. But keep
 937          * it a weak assumption overridable by the user.
 938          */
 939         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 940                                      opts->user_interval != ULLONG_MAX)) {
 941                 if (opts->freq) {
 942                         perf_evsel__set_sample_bit(evsel, PERIOD);
 943                         attr->freq              = 1;
 944                         attr->sample_freq       = opts->freq;
 945                 } else {
 946                         attr->sample_period = opts->default_interval;
 947                 }
 948         }
 949
 950         /*
 951          * Disable sampling for all group members other
 952          * than leader in case leader 'leads' the sampling.
 953          */
 954         if ((leader != evsel) && leader->sample_read) {
 955                 attr->freq           = 0;
 956                 attr->sample_freq    = 0;
 957                 attr->sample_period  = 0;
 958                 attr->write_backward = 0;
 959
 960                 /*
 961                  * We don't get sample for slave events, we make them
 962                  * when delivering group leader sample. Set the slave
 963                  * event to follow the master sample_type to ease up
 964                  * report.
 965                  */
 966                 attr->sample_type = leader->attr.sample_type;
 967         }
 968
 969         if (opts->no_samples)
 970                 attr->sample_freq = 0;
 971
 972         if (opts->inherit_stat) {
 973                 evsel->attr.read_format |=
 974                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 975                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 976                         PERF_FORMAT_ID;
 977                 attr->inherit_stat = 1;
 978         }
 979
 980         if (opts->sample_address) {
 981                 perf_evsel__set_sample_bit(evsel, ADDR);
 982                 attr->mmap_data = track;
 983         }
 984
 985         /*
 986          * We don't allow user space callchains for  function trace
 987          * event, due to issues with page faults while tracing page
 988          * fault handler and its overall trickiness nature.
 989          */
 990         if (perf_evsel__is_function_event(evsel))
 991                 evsel->attr.exclude_callchain_user = 1;
 992
 993         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 994                 perf_evsel__config_callchain(evsel, opts, callchain);
 995
 996         if (opts->sample_intr_regs) {
 997                 attr->sample_regs_intr = opts->sample_intr_regs;
 998                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 999         }
1000
1001         if (opts->sample_user_regs) {
1002                 attr->sample_regs_user |= opts->sample_user_regs;
1003                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1004         }
1005
1006         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1007                 perf_evsel__set_sample_bit(evsel, CPU);
1008
1009         /*
1010          * When the user explicitly disabled time don't force it here.
1011          */
1012         if (opts->sample_time &&
1013             (!perf_missing_features.sample_id_all &&
1014             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1015              opts->sample_time_set)))
1016                 perf_evsel__set_sample_bit(evsel, TIME);
1017
1018         if (opts->raw_samples && !evsel->no_aux_samples) {
1019                 perf_evsel__set_sample_bit(evsel, TIME);
1020                 perf_evsel__set_sample_bit(evsel, RAW);
1021                 perf_evsel__set_sample_bit(evsel, CPU);
1022         }
1023
1024         if (opts->sample_address)
1025                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1026
1027         if (opts->sample_phys_addr)
1028                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1029
1030         if (opts->no_buffering) {
1031                 attr->watermark = 0;
1032                 attr->wakeup_events = 1;
1033         }
1034         if (opts->branch_stack && !evsel->no_aux_samples) {
1035                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1036                 attr->branch_sample_type = opts->branch_stack;
1037         }
1038
1039         if (opts->sample_weight)
1040                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1041
1042         attr->task  = track;
1043         attr->mmap  = track;
1044         attr->mmap2 = track && !perf_missing_features.mmap2;
1045         attr->comm  = track;
1046         attr->ksymbol = track && !perf_missing_features.ksymbol;
1047         attr->bpf_event = track && opts->bpf_event &&
1048                 !perf_missing_features.bpf_event;
1049
1050         if (opts->record_namespaces)
1051                 attr->namespaces  = track;
1052
1053         if (opts->record_switch_events)
1054                 attr->context_switch = track;
1055
1056         if (opts->sample_transaction)
1057                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1058
1059         if (opts->running_time) {
1060                 evsel->attr.read_format |=
1061                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1062                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1063         }
1064
1065         /*
1066          * XXX see the function comment above
1067          *
1068          * Disabling only independent events or group leaders,
1069          * keeping group members enabled.
1070          */
1071         if (perf_evsel__is_group_leader(evsel))
1072                 attr->disabled = 1;
1073
1074         /*
1075          * Setting enable_on_exec for independent events and
1076          * group leaders for traced executed by perf.
1077          */
1078         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1079                 !opts->initial_delay)
1080                 attr->enable_on_exec = 1;
1081
1082         if (evsel->immediate) {
1083                 attr->disabled = 0;
1084                 attr->enable_on_exec = 0;
1085         }
1086
1087         clockid = opts->clockid;
1088         if (opts->use_clockid) {
1089                 attr->use_clockid = 1;
1090                 attr->clockid = opts->clockid;
1091         }
1092
1093         if (evsel->precise_max)
1094                 perf_event_attr__set_max_precise_ip(attr);
1095
1096         if (opts->all_user) {
1097                 attr->exclude_kernel = 1;
1098                 attr->exclude_user   = 0;
1099         }
1100
1101         if (opts->all_kernel) {
1102                 attr->exclude_kernel = 0;
1103                 attr->exclude_user   = 1;
1104         }
1105
1106         if (evsel->own_cpus || evsel->unit)
1107                 evsel->attr.read_format |= PERF_FORMAT_ID;
1108
1109         /*
1110          * Apply event specific term settings,
1111          * it overloads any global configuration.
1112          */
1113         apply_config_terms(evsel, opts, track);
1114
1115         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1116
1117         /* The --period option takes the precedence. */
1118         if (opts->period_set) {
1119                 if (opts->period)
1120                         perf_evsel__set_sample_bit(evsel, PERIOD);
1121                 else
1122                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1123         }
1124
1125         /*
1126          * For initial_delay, a dummy event is added implicitly.
1127          * The software event will trigger -EOPNOTSUPP error out,
1128          * if BRANCH_STACK bit is set.
1129          */
1130         if (opts->initial_delay && is_dummy_event(evsel))
1131                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1132 }
1133
1134 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1135 {
1136         if (evsel->system_wide)
1137                 nthreads = 1;
1138
1139         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1140
1141         if (evsel->fd) {
1142                 int cpu, thread;
1143                 for (cpu = 0; cpu < ncpus; cpu++) {
1144                         for (thread = 0; thread < nthreads; thread++) {
1145                                 FD(evsel, cpu, thread) = -1;
1146                         }
1147                 }
1148         }
1149
1150         return evsel->fd != NULL ? 0 : -ENOMEM;
1151 }
1152
1153 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1154                           int ioc,  void *arg)
1155 {
1156         int cpu, thread;
1157
1158         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1159                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1160                         int fd = FD(evsel, cpu, thread),
1161                             err = ioctl(fd, ioc, arg);
1162
1163                         if (err)
1164                                 return err;
1165                 }
1166         }
1167
1168         return 0;
1169 }
1170
1171 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1172 {
1173         return perf_evsel__run_ioctl(evsel,
1174                                      PERF_EVENT_IOC_SET_FILTER,
1175                                      (void *)filter);
1176 }
1177
1178 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1179 {
1180         char *new_filter = strdup(filter);
1181
1182         if (new_filter != NULL) {
1183                 free(evsel->filter);
1184                 evsel->filter = new_filter;
1185                 return 0;
1186         }
1187
1188         return -1;
1189 }
1190
1191 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1192                                      const char *fmt, const char *filter)
1193 {
1194         char *new_filter;
1195
1196         if (evsel->filter == NULL)
1197                 return perf_evsel__set_filter(evsel, filter);
1198
1199         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1200                 free(evsel->filter);
1201                 evsel->filter = new_filter;
1202                 return 0;
1203         }
1204
1205         return -1;
1206 }
1207
1208 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1209 {
1210         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1211 }
1212
1213 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1214 {
1215         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1216 }
1217
1218 int perf_evsel__enable(struct perf_evsel *evsel)
1219 {
1220         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1221
1222         if (!err)
1223                 evsel->disabled = false;
1224
1225         return err;
1226 }
1227
1228 int perf_evsel__disable(struct perf_evsel *evsel)
1229 {
1230         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1231         /*
1232          * We mark it disabled here so that tools that disable a event can
1233          * ignore events after they disable it. I.e. the ring buffer may have
1234          * already a few more events queued up before the kernel got the stop
1235          * request.
1236          */
1237         if (!err)
1238                 evsel->disabled = true;
1239
1240         return err;
1241 }
1242
1243 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1244 {
1245         if (ncpus == 0 || nthreads == 0)
1246                 return 0;
1247
1248         if (evsel->system_wide)
1249                 nthreads = 1;
1250
1251         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1252         if (evsel->sample_id == NULL)
1253                 return -ENOMEM;
1254
1255         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1256         if (evsel->id == NULL) {
1257                 xyarray__delete(evsel->sample_id);
1258                 evsel->sample_id = NULL;
1259                 return -ENOMEM;
1260         }
1261
1262         return 0;
1263 }
1264
1265 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1266 {
1267         xyarray__delete(evsel->fd);
1268         evsel->fd = NULL;
1269 }
1270
1271 static void perf_evsel__free_id(struct perf_evsel *evsel)
1272 {
1273         xyarray__delete(evsel->sample_id);
1274         evsel->sample_id = NULL;
1275         zfree(&evsel->id);
1276 }
1277
1278 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1279 {
1280         struct perf_evsel_config_term *term, *h;
1281
1282         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1283                 list_del(&term->list);
1284                 free(term);
1285         }
1286 }
1287
1288 void perf_evsel__close_fd(struct perf_evsel *evsel)
1289 {
1290         int cpu, thread;
1291
1292         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1293                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1294                         close(FD(evsel, cpu, thread));
1295                         FD(evsel, cpu, thread) = -1;
1296                 }
1297 }
1298
1299 void perf_evsel__exit(struct perf_evsel *evsel)
1300 {
1301         assert(list_empty(&evsel->node));
1302         assert(evsel->evlist == NULL);
1303         perf_evsel__free_fd(evsel);
1304         perf_evsel__free_id(evsel);
1305         perf_evsel__free_config_terms(evsel);
1306         cgroup__put(evsel->cgrp);
1307         cpu_map__put(evsel->cpus);
1308         cpu_map__put(evsel->own_cpus);
1309         thread_map__put(evsel->threads);
1310         zfree(&evsel->group_name);
1311         zfree(&evsel->name);
1312         perf_evsel__object.fini(evsel);
1313 }
1314
1315 void perf_evsel__delete(struct perf_evsel *evsel)
1316 {
1317         perf_evsel__exit(evsel);
1318         free(evsel);
1319 }
1320
1321 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1322                                 struct perf_counts_values *count)
1323 {
1324         struct perf_counts_values tmp;
1325
1326         if (!evsel->prev_raw_counts)
1327                 return;
1328
1329         if (cpu == -1) {
1330                 tmp = evsel->prev_raw_counts->aggr;
1331                 evsel->prev_raw_counts->aggr = *count;
1332         } else {
1333                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1334                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1335         }
1336
1337         count->val = count->val - tmp.val;
1338         count->ena = count->ena - tmp.ena;
1339         count->run = count->run - tmp.run;
1340 }
1341
1342 void perf_counts_values__scale(struct perf_counts_values *count,
1343                                bool scale, s8 *pscaled)
1344 {
1345         s8 scaled = 0;
1346
1347         if (scale) {
1348                 if (count->run == 0) {
1349                         scaled = -1;
1350                         count->val = 0;
1351                 } else if (count->run < count->ena) {
1352                         scaled = 1;
1353                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1354                 }
1355         } else
1356                 count->ena = count->run = 0;
1357
1358         if (pscaled)
1359                 *pscaled = scaled;
1360 }
1361
1362 static int perf_evsel__read_size(struct perf_evsel *evsel)
1363 {
1364         u64 read_format = evsel->attr.read_format;
1365         int entry = sizeof(u64); /* value */
1366         int size = 0;
1367         int nr = 1;
1368
1369         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1370                 size += sizeof(u64);
1371
1372         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1373                 size += sizeof(u64);
1374
1375         if (read_format & PERF_FORMAT_ID)
1376                 entry += sizeof(u64);
1377
1378         if (read_format & PERF_FORMAT_GROUP) {
1379                 nr = evsel->nr_members;
1380                 size += sizeof(u64);
1381         }
1382
1383         size += entry * nr;
1384         return size;
1385 }
1386
1387 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1388                      struct perf_counts_values *count)
1389 {
1390         size_t size = perf_evsel__read_size(evsel);
1391
1392         memset(count, 0, sizeof(*count));
1393
1394         if (FD(evsel, cpu, thread) < 0)
1395                 return -EINVAL;
1396
1397         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1398                 return -errno;
1399
1400         return 0;
1401 }
1402
1403 static int
1404 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1405 {
1406         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1407
1408         return perf_evsel__read(evsel, cpu, thread, count);
1409 }
1410
1411 static void
1412 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1413                       u64 val, u64 ena, u64 run)
1414 {
1415         struct perf_counts_values *count;
1416
1417         count = perf_counts(counter->counts, cpu, thread);
1418
1419         count->val    = val;
1420         count->ena    = ena;
1421         count->run    = run;
1422         count->loaded = true;
1423 }
1424
1425 static int
1426 perf_evsel__process_group_data(struct perf_evsel *leader,
1427                                int cpu, int thread, u64 *data)
1428 {
1429         u64 read_format = leader->attr.read_format;
1430         struct sample_read_value *v;
1431         u64 nr, ena = 0, run = 0, i;
1432
1433         nr = *data++;
1434
1435         if (nr != (u64) leader->nr_members)
1436                 return -EINVAL;
1437
1438         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1439                 ena = *data++;
1440
1441         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1442                 run = *data++;
1443
1444         v = (struct sample_read_value *) data;
1445
1446         perf_evsel__set_count(leader, cpu, thread,
1447                               v[0].value, ena, run);
1448
1449         for (i = 1; i < nr; i++) {
1450                 struct perf_evsel *counter;
1451
1452                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1453                 if (!counter)
1454                         return -EINVAL;
1455
1456                 perf_evsel__set_count(counter, cpu, thread,
1457                                       v[i].value, ena, run);
1458         }
1459
1460         return 0;
1461 }
1462
1463 static int
1464 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1465 {
1466         struct perf_stat_evsel *ps = leader->stats;
1467         u64 read_format = leader->attr.read_format;
1468         int size = perf_evsel__read_size(leader);
1469         u64 *data = ps->group_data;
1470
1471         if (!(read_format & PERF_FORMAT_ID))
1472                 return -EINVAL;
1473
1474         if (!perf_evsel__is_group_leader(leader))
1475                 return -EINVAL;
1476
1477         if (!data) {
1478                 data = zalloc(size);
1479                 if (!data)
1480                         return -ENOMEM;
1481
1482                 ps->group_data = data;
1483         }
1484
1485         if (FD(leader, cpu, thread) < 0)
1486                 return -EINVAL;
1487
1488         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1489                 return -errno;
1490
1491         return perf_evsel__process_group_data(leader, cpu, thread, data);
1492 }
1493
1494 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1495 {
1496         u64 read_format = evsel->attr.read_format;
1497
1498         if (read_format & PERF_FORMAT_GROUP)
1499                 return perf_evsel__read_group(evsel, cpu, thread);
1500         else
1501                 return perf_evsel__read_one(evsel, cpu, thread);
1502 }
1503
1504 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1505                               int cpu, int thread, bool scale)
1506 {
1507         struct perf_counts_values count;
1508         size_t nv = scale ? 3 : 1;
1509
1510         if (FD(evsel, cpu, thread) < 0)
1511                 return -EINVAL;
1512
1513         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1514                 return -ENOMEM;
1515
1516         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1517                 return -errno;
1518
1519         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1520         perf_counts_values__scale(&count, scale, NULL);
1521         *perf_counts(evsel->counts, cpu, thread) = count;
1522         return 0;
1523 }
1524
1525 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1526 {
1527         struct perf_evsel *leader = evsel->leader;
1528         int fd;
1529
1530         if (perf_evsel__is_group_leader(evsel))
1531                 return -1;
1532
1533         /*
1534          * Leader must be already processed/open,
1535          * if not it's a bug.
1536          */
1537         BUG_ON(!leader->fd);
1538
1539         fd = FD(leader, cpu, thread);
1540         BUG_ON(fd == -1);
1541
1542         return fd;
1543 }
1544
1545 struct bit_names {
1546         int bit;
1547         const char *name;
1548 };
1549
1550 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1551 {
1552         bool first_bit = true;
1553         int i = 0;
1554
1555         do {
1556                 if (value & bits[i].bit) {
1557                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1558                         first_bit = false;
1559                 }
1560         } while (bits[++i].name != NULL);
1561 }
1562
1563 static void __p_sample_type(char *buf, size_t size, u64 value)
1564 {
1565 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1566         struct bit_names bits[] = {
1567                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1568                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1569                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1570                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1571                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1572                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1573                 { .name = NULL, }
1574         };
1575 #undef bit_name
1576         __p_bits(buf, size, value, bits);
1577 }
1578
1579 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1580 {
1581 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1582         struct bit_names bits[] = {
1583                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1584                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1585                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1586                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1587                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1588                 { .name = NULL, }
1589         };
1590 #undef bit_name
1591         __p_bits(buf, size, value, bits);
1592 }
1593
1594 static void __p_read_format(char *buf, size_t size, u64 value)
1595 {
1596 #define bit_name(n) { PERF_FORMAT_##n, #n }
1597         struct bit_names bits[] = {
1598                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1599                 bit_name(ID), bit_name(GROUP),
1600                 { .name = NULL, }
1601         };
1602 #undef bit_name
1603         __p_bits(buf, size, value, bits);
1604 }
1605
1606 #define BUF_SIZE                1024
1607
1608 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1609 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1610 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1611 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1612 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1613 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1614
1615 #define PRINT_ATTRn(_n, _f, _p)                         \
1616 do {                                                    \
1617         if (attr->_f) {                                 \
1618                 _p(attr->_f);                           \
1619                 ret += attr__fprintf(fp, _n, buf, priv);\
1620         }                                               \
1621 } while (0)
1622
1623 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1624
1625 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1626                              attr__fprintf_f attr__fprintf, void *priv)
1627 {
1628         char buf[BUF_SIZE];
1629         int ret = 0;
1630
1631         PRINT_ATTRf(type, p_unsigned);
1632         PRINT_ATTRf(size, p_unsigned);
1633         PRINT_ATTRf(config, p_hex);
1634         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1635         PRINT_ATTRf(sample_type, p_sample_type);
1636         PRINT_ATTRf(read_format, p_read_format);
1637
1638         PRINT_ATTRf(disabled, p_unsigned);
1639         PRINT_ATTRf(inherit, p_unsigned);
1640         PRINT_ATTRf(pinned, p_unsigned);
1641         PRINT_ATTRf(exclusive, p_unsigned);
1642         PRINT_ATTRf(exclude_user, p_unsigned);
1643         PRINT_ATTRf(exclude_kernel, p_unsigned);
1644         PRINT_ATTRf(exclude_hv, p_unsigned);
1645         PRINT_ATTRf(exclude_idle, p_unsigned);
1646         PRINT_ATTRf(mmap, p_unsigned);
1647         PRINT_ATTRf(comm, p_unsigned);
1648         PRINT_ATTRf(freq, p_unsigned);
1649         PRINT_ATTRf(inherit_stat, p_unsigned);
1650         PRINT_ATTRf(enable_on_exec, p_unsigned);
1651         PRINT_ATTRf(task, p_unsigned);
1652         PRINT_ATTRf(watermark, p_unsigned);
1653         PRINT_ATTRf(precise_ip, p_unsigned);
1654         PRINT_ATTRf(mmap_data, p_unsigned);
1655         PRINT_ATTRf(sample_id_all, p_unsigned);
1656         PRINT_ATTRf(exclude_host, p_unsigned);
1657         PRINT_ATTRf(exclude_guest, p_unsigned);
1658         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1659         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1660         PRINT_ATTRf(mmap2, p_unsigned);
1661         PRINT_ATTRf(comm_exec, p_unsigned);
1662         PRINT_ATTRf(use_clockid, p_unsigned);
1663         PRINT_ATTRf(context_switch, p_unsigned);
1664         PRINT_ATTRf(write_backward, p_unsigned);
1665         PRINT_ATTRf(namespaces, p_unsigned);
1666         PRINT_ATTRf(ksymbol, p_unsigned);
1667         PRINT_ATTRf(bpf_event, p_unsigned);
1668
1669         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1670         PRINT_ATTRf(bp_type, p_unsigned);
1671         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1672         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1673         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1674         PRINT_ATTRf(sample_regs_user, p_hex);
1675         PRINT_ATTRf(sample_stack_user, p_unsigned);
1676         PRINT_ATTRf(clockid, p_signed);
1677         PRINT_ATTRf(sample_regs_intr, p_hex);
1678         PRINT_ATTRf(aux_watermark, p_unsigned);
1679         PRINT_ATTRf(sample_max_stack, p_unsigned);
1680
1681         return ret;
1682 }
1683
1684 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1685                                 void *priv __maybe_unused)
1686 {
1687         return fprintf(fp, "  %-32s %s\n", name, val);
1688 }
1689
1690 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1691                                   int nr_cpus, int nr_threads,
1692                                   int thread_idx)
1693 {
1694         for (int cpu = 0; cpu < nr_cpus; cpu++)
1695                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1696                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1697 }
1698
1699 static int update_fds(struct perf_evsel *evsel,
1700                       int nr_cpus, int cpu_idx,
1701                       int nr_threads, int thread_idx)
1702 {
1703         struct perf_evsel *pos;
1704
1705         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1706                 return -EINVAL;
1707
1708         evlist__for_each_entry(evsel->evlist, pos) {
1709                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1710
1711                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1712
1713                 /*
1714                  * Since fds for next evsel has not been created,
1715                  * there is no need to iterate whole event list.
1716                  */
1717                 if (pos == evsel)
1718                         break;
1719         }
1720         return 0;
1721 }
1722
1723 static bool ignore_missing_thread(struct perf_evsel *evsel,
1724                                   int nr_cpus, int cpu,
1725                                   struct thread_map *threads,
1726                                   int thread, int err)
1727 {
1728         pid_t ignore_pid = thread_map__pid(threads, thread);
1729
1730         if (!evsel->ignore_missing_thread)
1731                 return false;
1732
1733         /* The system wide setup does not work with threads. */
1734         if (evsel->system_wide)
1735                 return false;
1736
1737         /* The -ESRCH is perf event syscall errno for pid's not found. */
1738         if (err != -ESRCH)
1739                 return false;
1740
1741         /* If there's only one thread, let it fail. */
1742         if (threads->nr == 1)
1743                 return false;
1744
1745         /*
1746          * We should remove fd for missing_thread first
1747          * because thread_map__remove() will decrease threads->nr.
1748          */
1749         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1750                 return false;
1751
1752         if (thread_map__remove(threads, thread))
1753                 return false;
1754
1755         pr_warning("WARNING: Ignored open failure for pid %d\n",
1756                    ignore_pid);
1757         return true;
1758 }
1759
1760 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1761                      struct thread_map *threads)
1762 {
1763         int cpu, thread, nthreads;
1764         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1765         int pid = -1, err;
1766         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1767
1768         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1769                 return -EINVAL;
1770
1771         if (cpus == NULL) {
1772                 static struct cpu_map *empty_cpu_map;
1773
1774                 if (empty_cpu_map == NULL) {
1775                         empty_cpu_map = cpu_map__dummy_new();
1776                         if (empty_cpu_map == NULL)
1777                                 return -ENOMEM;
1778                 }
1779
1780                 cpus = empty_cpu_map;
1781         }
1782
1783         if (threads == NULL) {
1784                 static struct thread_map *empty_thread_map;
1785
1786                 if (empty_thread_map == NULL) {
1787                         empty_thread_map = thread_map__new_by_tid(-1);
1788                         if (empty_thread_map == NULL)
1789                                 return -ENOMEM;
1790                 }
1791
1792                 threads = empty_thread_map;
1793         }
1794
1795         if (evsel->system_wide)
1796                 nthreads = 1;
1797         else
1798                 nthreads = threads->nr;
1799
1800         if (evsel->fd == NULL &&
1801             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1802                 return -ENOMEM;
1803
1804         if (evsel->cgrp) {
1805                 flags |= PERF_FLAG_PID_CGROUP;
1806                 pid = evsel->cgrp->fd;
1807         }
1808
1809 fallback_missing_features:
1810         if (perf_missing_features.clockid_wrong)
1811                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1812         if (perf_missing_features.clockid) {
1813                 evsel->attr.use_clockid = 0;
1814                 evsel->attr.clockid = 0;
1815         }
1816         if (perf_missing_features.cloexec)
1817                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1818         if (perf_missing_features.mmap2)
1819                 evsel->attr.mmap2 = 0;
1820         if (perf_missing_features.exclude_guest)
1821                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1822         if (perf_missing_features.lbr_flags)
1823                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1824                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1825         if (perf_missing_features.group_read && evsel->attr.inherit)
1826                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1827         if (perf_missing_features.ksymbol)
1828                 evsel->attr.ksymbol = 0;
1829         if (perf_missing_features.bpf_event)
1830                 evsel->attr.bpf_event = 0;
1831 retry_sample_id:
1832         if (perf_missing_features.sample_id_all)
1833                 evsel->attr.sample_id_all = 0;
1834
1835         if (verbose >= 2) {
1836                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1837                 fprintf(stderr, "perf_event_attr:\n");
1838                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1839                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1840         }
1841
1842         for (cpu = 0; cpu < cpus->nr; cpu++) {
1843
1844                 for (thread = 0; thread < nthreads; thread++) {
1845                         int fd, group_fd;
1846
1847                         if (!evsel->cgrp && !evsel->system_wide)
1848                                 pid = thread_map__pid(threads, thread);
1849
1850                         group_fd = get_group_fd(evsel, cpu, thread);
1851 retry_open:
1852                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1853                                   pid, cpus->map[cpu], group_fd, flags);
1854
1855                         test_attr__ready();
1856
1857                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1858                                                  group_fd, flags);
1859
1860                         FD(evsel, cpu, thread) = fd;
1861
1862                         if (fd < 0) {
1863                                 err = -errno;
1864
1865                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1866                                         /*
1867                                          * We just removed 1 thread, so take a step
1868                                          * back on thread index and lower the upper
1869                                          * nthreads limit.
1870                                          */
1871                                         nthreads--;
1872                                         thread--;
1873
1874                                         /* ... and pretend like nothing have happened. */
1875                                         err = 0;
1876                                         continue;
1877                                 }
1878
1879                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1880                                           err);
1881                                 goto try_fallback;
1882                         }
1883
1884                         pr_debug2(" = %d\n", fd);
1885
1886                         if (evsel->bpf_fd >= 0) {
1887                                 int evt_fd = fd;
1888                                 int bpf_fd = evsel->bpf_fd;
1889
1890                                 err = ioctl(evt_fd,
1891                                             PERF_EVENT_IOC_SET_BPF,
1892                                             bpf_fd);
1893                                 if (err && errno != EEXIST) {
1894                                         pr_err("failed to attach bpf fd %d: %s\n",
1895                                                bpf_fd, strerror(errno));
1896                                         err = -EINVAL;
1897                                         goto out_close;
1898                                 }
1899                         }
1900
1901                         set_rlimit = NO_CHANGE;
1902
1903                         /*
1904                          * If we succeeded but had to kill clockid, fail and
1905                          * have perf_evsel__open_strerror() print us a nice
1906                          * error.
1907                          */
1908                         if (perf_missing_features.clockid ||
1909                             perf_missing_features.clockid_wrong) {
1910                                 err = -EINVAL;
1911                                 goto out_close;
1912                         }
1913                 }
1914         }
1915
1916         return 0;
1917
1918 try_fallback:
1919         /*
1920          * perf stat needs between 5 and 22 fds per CPU. When we run out
1921          * of them try to increase the limits.
1922          */
1923         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1924                 struct rlimit l;
1925                 int old_errno = errno;
1926
1927                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1928                         if (set_rlimit == NO_CHANGE)
1929                                 l.rlim_cur = l.rlim_max;
1930                         else {
1931                                 l.rlim_cur = l.rlim_max + 1000;
1932                                 l.rlim_max = l.rlim_cur;
1933                         }
1934                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1935                                 set_rlimit++;
1936                                 errno = old_errno;
1937                                 goto retry_open;
1938                         }
1939                 }
1940                 errno = old_errno;
1941         }
1942
1943         if (err != -EINVAL || cpu > 0 || thread > 0)
1944                 goto out_close;
1945
1946         /*
1947          * Must probe features in the order they were added to the
1948          * perf_event_attr interface.
1949          */
1950         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
1951                 perf_missing_features.bpf_event = true;
1952                 pr_debug2("switching off bpf_event\n");
1953                 goto fallback_missing_features;
1954         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
1955                 perf_missing_features.ksymbol = true;
1956                 pr_debug2("switching off ksymbol\n");
1957                 goto fallback_missing_features;
1958         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1959                 perf_missing_features.write_backward = true;
1960                 pr_debug2("switching off write_backward\n");
1961                 goto out_close;
1962         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1963                 perf_missing_features.clockid_wrong = true;
1964                 pr_debug2("switching off clockid\n");
1965                 goto fallback_missing_features;
1966         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1967                 perf_missing_features.clockid = true;
1968                 pr_debug2("switching off use_clockid\n");
1969                 goto fallback_missing_features;
1970         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1971                 perf_missing_features.cloexec = true;
1972                 pr_debug2("switching off cloexec flag\n");
1973                 goto fallback_missing_features;
1974         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1975                 perf_missing_features.mmap2 = true;
1976                 pr_debug2("switching off mmap2\n");
1977                 goto fallback_missing_features;
1978         } else if (!perf_missing_features.exclude_guest &&
1979                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1980                 perf_missing_features.exclude_guest = true;
1981                 pr_debug2("switching off exclude_guest, exclude_host\n");
1982                 goto fallback_missing_features;
1983         } else if (!perf_missing_features.sample_id_all) {
1984                 perf_missing_features.sample_id_all = true;
1985                 pr_debug2("switching off sample_id_all\n");
1986                 goto retry_sample_id;
1987         } else if (!perf_missing_features.lbr_flags &&
1988                         (evsel->attr.branch_sample_type &
1989                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1990                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1991                 perf_missing_features.lbr_flags = true;
1992                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1993                 goto fallback_missing_features;
1994         } else if (!perf_missing_features.group_read &&
1995                     evsel->attr.inherit &&
1996                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
1997                    perf_evsel__is_group_leader(evsel)) {
1998                 perf_missing_features.group_read = true;
1999                 pr_debug2("switching off group read\n");
2000                 goto fallback_missing_features;
2001         }
2002 out_close:
2003         if (err)
2004                 threads->err_thread = thread;
2005
2006         do {
2007                 while (--thread >= 0) {
2008                         close(FD(evsel, cpu, thread));
2009                         FD(evsel, cpu, thread) = -1;
2010                 }
2011                 thread = nthreads;
2012         } while (--cpu >= 0);
2013         return err;
2014 }
2015
2016 void perf_evsel__close(struct perf_evsel *evsel)
2017 {
2018         if (evsel->fd == NULL)
2019                 return;
2020
2021         perf_evsel__close_fd(evsel);
2022         perf_evsel__free_fd(evsel);
2023 }
2024
2025 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2026                              struct cpu_map *cpus)
2027 {
2028         return perf_evsel__open(evsel, cpus, NULL);
2029 }
2030
2031 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2032                                 struct thread_map *threads)
2033 {
2034         return perf_evsel__open(evsel, NULL, threads);
2035 }
2036
2037 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2038                                        const union perf_event *event,
2039                                        struct perf_sample *sample)
2040 {
2041         u64 type = evsel->attr.sample_type;
2042         const u64 *array = event->sample.array;
2043         bool swapped = evsel->needs_swap;
2044         union u64_swap u;
2045
2046         array += ((event->header.size -
2047                    sizeof(event->header)) / sizeof(u64)) - 1;
2048
2049         if (type & PERF_SAMPLE_IDENTIFIER) {
2050                 sample->id = *array;
2051                 array--;
2052         }
2053
2054         if (type & PERF_SAMPLE_CPU) {
2055                 u.val64 = *array;
2056                 if (swapped) {
2057                         /* undo swap of u64, then swap on individual u32s */
2058                         u.val64 = bswap_64(u.val64);
2059                         u.val32[0] = bswap_32(u.val32[0]);
2060                 }
2061
2062                 sample->cpu = u.val32[0];
2063                 array--;
2064         }
2065
2066         if (type & PERF_SAMPLE_STREAM_ID) {
2067                 sample->stream_id = *array;
2068                 array--;
2069         }
2070
2071         if (type & PERF_SAMPLE_ID) {
2072                 sample->id = *array;
2073                 array--;
2074         }
2075
2076         if (type & PERF_SAMPLE_TIME) {
2077                 sample->time = *array;
2078                 array--;
2079         }
2080
2081         if (type & PERF_SAMPLE_TID) {
2082                 u.val64 = *array;
2083                 if (swapped) {
2084                         /* undo swap of u64, then swap on individual u32s */
2085                         u.val64 = bswap_64(u.val64);
2086                         u.val32[0] = bswap_32(u.val32[0]);
2087                         u.val32[1] = bswap_32(u.val32[1]);
2088                 }
2089
2090                 sample->pid = u.val32[0];
2091                 sample->tid = u.val32[1];
2092                 array--;
2093         }
2094
2095         return 0;
2096 }
2097
2098 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2099                             u64 size)
2100 {
2101         return size > max_size || offset + size > endp;
2102 }
2103
2104 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2105         do {                                                            \
2106                 if (overflow(endp, (max_size), (offset), (size)))       \
2107                         return -EFAULT;                                 \
2108         } while (0)
2109
2110 #define OVERFLOW_CHECK_u64(offset) \
2111         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2112
2113 static int
2114 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2115 {
2116         /*
2117          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2118          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2119          * check the format does not go past the end of the event.
2120          */
2121         if (sample_size + sizeof(event->header) > event->header.size)
2122                 return -EFAULT;
2123
2124         return 0;
2125 }
2126
2127 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2128                              struct perf_sample *data)
2129 {
2130         u64 type = evsel->attr.sample_type;
2131         bool swapped = evsel->needs_swap;
2132         const u64 *array;
2133         u16 max_size = event->header.size;
2134         const void *endp = (void *)event + max_size;
2135         u64 sz;
2136
2137         /*
2138          * used for cross-endian analysis. See git commit 65014ab3
2139          * for why this goofiness is needed.
2140          */
2141         union u64_swap u;
2142
2143         memset(data, 0, sizeof(*data));
2144         data->cpu = data->pid = data->tid = -1;
2145         data->stream_id = data->id = data->time = -1ULL;
2146         data->period = evsel->attr.sample_period;
2147         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2148         data->misc    = event->header.misc;
2149         data->id = -1ULL;
2150         data->data_src = PERF_MEM_DATA_SRC_NONE;
2151
2152         if (event->header.type != PERF_RECORD_SAMPLE) {
2153                 if (!evsel->attr.sample_id_all)
2154                         return 0;
2155                 return perf_evsel__parse_id_sample(evsel, event, data);
2156         }
2157
2158         array = event->sample.array;
2159
2160         if (perf_event__check_size(event, evsel->sample_size))
2161                 return -EFAULT;
2162
2163         if (type & PERF_SAMPLE_IDENTIFIER) {
2164                 data->id = *array;
2165                 array++;
2166         }
2167
2168         if (type & PERF_SAMPLE_IP) {
2169                 data->ip = *array;
2170                 array++;
2171         }
2172
2173         if (type & PERF_SAMPLE_TID) {
2174                 u.val64 = *array;
2175                 if (swapped) {
2176                         /* undo swap of u64, then swap on individual u32s */
2177                         u.val64 = bswap_64(u.val64);
2178                         u.val32[0] = bswap_32(u.val32[0]);
2179                         u.val32[1] = bswap_32(u.val32[1]);
2180                 }
2181
2182                 data->pid = u.val32[0];
2183                 data->tid = u.val32[1];
2184                 array++;
2185         }
2186
2187         if (type & PERF_SAMPLE_TIME) {
2188                 data->time = *array;
2189                 array++;
2190         }
2191
2192         if (type & PERF_SAMPLE_ADDR) {
2193                 data->addr = *array;
2194                 array++;
2195         }
2196
2197         if (type & PERF_SAMPLE_ID) {
2198                 data->id = *array;
2199                 array++;
2200         }
2201
2202         if (type & PERF_SAMPLE_STREAM_ID) {
2203                 data->stream_id = *array;
2204                 array++;
2205         }
2206
2207         if (type & PERF_SAMPLE_CPU) {
2208
2209                 u.val64 = *array;
2210                 if (swapped) {
2211                         /* undo swap of u64, then swap on individual u32s */
2212                         u.val64 = bswap_64(u.val64);
2213                         u.val32[0] = bswap_32(u.val32[0]);
2214                 }
2215
2216                 data->cpu = u.val32[0];
2217                 array++;
2218         }
2219
2220         if (type & PERF_SAMPLE_PERIOD) {
2221                 data->period = *array;
2222                 array++;
2223         }
2224
2225         if (type & PERF_SAMPLE_READ) {
2226                 u64 read_format = evsel->attr.read_format;
2227
2228                 OVERFLOW_CHECK_u64(array);
2229                 if (read_format & PERF_FORMAT_GROUP)
2230                         data->read.group.nr = *array;
2231                 else
2232                         data->read.one.value = *array;
2233
2234                 array++;
2235
2236                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2237                         OVERFLOW_CHECK_u64(array);
2238                         data->read.time_enabled = *array;
2239                         array++;
2240                 }
2241
2242                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2243                         OVERFLOW_CHECK_u64(array);
2244                         data->read.time_running = *array;
2245                         array++;
2246                 }
2247
2248                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2249                 if (read_format & PERF_FORMAT_GROUP) {
2250                         const u64 max_group_nr = UINT64_MAX /
2251                                         sizeof(struct sample_read_value);
2252
2253                         if (data->read.group.nr > max_group_nr)
2254                                 return -EFAULT;
2255                         sz = data->read.group.nr *
2256                              sizeof(struct sample_read_value);
2257                         OVERFLOW_CHECK(array, sz, max_size);
2258                         data->read.group.values =
2259                                         (struct sample_read_value *)array;
2260                         array = (void *)array + sz;
2261                 } else {
2262                         OVERFLOW_CHECK_u64(array);
2263                         data->read.one.id = *array;
2264                         array++;
2265                 }
2266         }
2267
2268         if (evsel__has_callchain(evsel)) {
2269                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2270
2271                 OVERFLOW_CHECK_u64(array);
2272                 data->callchain = (struct ip_callchain *)array++;
2273                 if (data->callchain->nr > max_callchain_nr)
2274                         return -EFAULT;
2275                 sz = data->callchain->nr * sizeof(u64);
2276                 OVERFLOW_CHECK(array, sz, max_size);
2277                 array = (void *)array + sz;
2278         }
2279
2280         if (type & PERF_SAMPLE_RAW) {
2281                 OVERFLOW_CHECK_u64(array);
2282                 u.val64 = *array;
2283
2284                 /*
2285                  * Undo swap of u64, then swap on individual u32s,
2286                  * get the size of the raw area and undo all of the
2287                  * swap. The pevent interface handles endianity by
2288                  * itself.
2289                  */
2290                 if (swapped) {
2291                         u.val64 = bswap_64(u.val64);
2292                         u.val32[0] = bswap_32(u.val32[0]);
2293                         u.val32[1] = bswap_32(u.val32[1]);
2294                 }
2295                 data->raw_size = u.val32[0];
2296
2297                 /*
2298                  * The raw data is aligned on 64bits including the
2299                  * u32 size, so it's safe to use mem_bswap_64.
2300                  */
2301                 if (swapped)
2302                         mem_bswap_64((void *) array, data->raw_size);
2303
2304                 array = (void *)array + sizeof(u32);
2305
2306                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2307                 data->raw_data = (void *)array;
2308                 array = (void *)array + data->raw_size;
2309         }
2310
2311         if (type & PERF_SAMPLE_BRANCH_STACK) {
2312                 const u64 max_branch_nr = UINT64_MAX /
2313                                           sizeof(struct branch_entry);
2314
2315                 OVERFLOW_CHECK_u64(array);
2316                 data->branch_stack = (struct branch_stack *)array++;
2317
2318                 if (data->branch_stack->nr > max_branch_nr)
2319                         return -EFAULT;
2320                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2321                 OVERFLOW_CHECK(array, sz, max_size);
2322                 array = (void *)array + sz;
2323         }
2324
2325         if (type & PERF_SAMPLE_REGS_USER) {
2326                 OVERFLOW_CHECK_u64(array);
2327                 data->user_regs.abi = *array;
2328                 array++;
2329
2330                 if (data->user_regs.abi) {
2331                         u64 mask = evsel->attr.sample_regs_user;
2332
2333                         sz = hweight_long(mask) * sizeof(u64);
2334                         OVERFLOW_CHECK(array, sz, max_size);
2335                         data->user_regs.mask = mask;
2336                         data->user_regs.regs = (u64 *)array;
2337                         array = (void *)array + sz;
2338                 }
2339         }
2340
2341         if (type & PERF_SAMPLE_STACK_USER) {
2342                 OVERFLOW_CHECK_u64(array);
2343                 sz = *array++;
2344
2345                 data->user_stack.offset = ((char *)(array - 1)
2346                                           - (char *) event);
2347
2348                 if (!sz) {
2349                         data->user_stack.size = 0;
2350                 } else {
2351                         OVERFLOW_CHECK(array, sz, max_size);
2352                         data->user_stack.data = (char *)array;
2353                         array = (void *)array + sz;
2354                         OVERFLOW_CHECK_u64(array);
2355                         data->user_stack.size = *array++;
2356                         if (WARN_ONCE(data->user_stack.size > sz,
2357                                       "user stack dump failure\n"))
2358                                 return -EFAULT;
2359                 }
2360         }
2361
2362         if (type & PERF_SAMPLE_WEIGHT) {
2363                 OVERFLOW_CHECK_u64(array);
2364                 data->weight = *array;
2365                 array++;
2366         }
2367
2368         if (type & PERF_SAMPLE_DATA_SRC) {
2369                 OVERFLOW_CHECK_u64(array);
2370                 data->data_src = *array;
2371                 array++;
2372         }
2373
2374         if (type & PERF_SAMPLE_TRANSACTION) {
2375                 OVERFLOW_CHECK_u64(array);
2376                 data->transaction = *array;
2377                 array++;
2378         }
2379
2380         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2381         if (type & PERF_SAMPLE_REGS_INTR) {
2382                 OVERFLOW_CHECK_u64(array);
2383                 data->intr_regs.abi = *array;
2384                 array++;
2385
2386                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2387                         u64 mask = evsel->attr.sample_regs_intr;
2388
2389                         sz = hweight_long(mask) * sizeof(u64);
2390                         OVERFLOW_CHECK(array, sz, max_size);
2391                         data->intr_regs.mask = mask;
2392                         data->intr_regs.regs = (u64 *)array;
2393                         array = (void *)array + sz;
2394                 }
2395         }
2396
2397         data->phys_addr = 0;
2398         if (type & PERF_SAMPLE_PHYS_ADDR) {
2399                 data->phys_addr = *array;
2400                 array++;
2401         }
2402
2403         return 0;
2404 }
2405
2406 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2407                                        union perf_event *event,
2408                                        u64 *timestamp)
2409 {
2410         u64 type = evsel->attr.sample_type;
2411         const u64 *array;
2412
2413         if (!(type & PERF_SAMPLE_TIME))
2414                 return -1;
2415
2416         if (event->header.type != PERF_RECORD_SAMPLE) {
2417                 struct perf_sample data = {
2418                         .time = -1ULL,
2419                 };
2420
2421                 if (!evsel->attr.sample_id_all)
2422                         return -1;
2423                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2424                         return -1;
2425
2426                 *timestamp = data.time;
2427                 return 0;
2428         }
2429
2430         array = event->sample.array;
2431
2432         if (perf_event__check_size(event, evsel->sample_size))
2433                 return -EFAULT;
2434
2435         if (type & PERF_SAMPLE_IDENTIFIER)
2436                 array++;
2437
2438         if (type & PERF_SAMPLE_IP)
2439                 array++;
2440
2441         if (type & PERF_SAMPLE_TID)
2442                 array++;
2443
2444         if (type & PERF_SAMPLE_TIME)
2445                 *timestamp = *array;
2446
2447         return 0;
2448 }
2449
2450 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2451                                      u64 read_format)
2452 {
2453         size_t sz, result = sizeof(struct sample_event);
2454
2455         if (type & PERF_SAMPLE_IDENTIFIER)
2456                 result += sizeof(u64);
2457
2458         if (type & PERF_SAMPLE_IP)
2459                 result += sizeof(u64);
2460
2461         if (type & PERF_SAMPLE_TID)
2462                 result += sizeof(u64);
2463
2464         if (type & PERF_SAMPLE_TIME)
2465                 result += sizeof(u64);
2466
2467         if (type & PERF_SAMPLE_ADDR)
2468                 result += sizeof(u64);
2469
2470         if (type & PERF_SAMPLE_ID)
2471                 result += sizeof(u64);
2472
2473         if (type & PERF_SAMPLE_STREAM_ID)
2474                 result += sizeof(u64);
2475
2476         if (type & PERF_SAMPLE_CPU)
2477                 result += sizeof(u64);
2478
2479         if (type & PERF_SAMPLE_PERIOD)
2480                 result += sizeof(u64);
2481
2482         if (type & PERF_SAMPLE_READ) {
2483                 result += sizeof(u64);
2484                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2485                         result += sizeof(u64);
2486                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2487                         result += sizeof(u64);
2488                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2489                 if (read_format & PERF_FORMAT_GROUP) {
2490                         sz = sample->read.group.nr *
2491                              sizeof(struct sample_read_value);
2492                         result += sz;
2493                 } else {
2494                         result += sizeof(u64);
2495                 }
2496         }
2497
2498         if (type & PERF_SAMPLE_CALLCHAIN) {
2499                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2500                 result += sz;
2501         }
2502
2503         if (type & PERF_SAMPLE_RAW) {
2504                 result += sizeof(u32);
2505                 result += sample->raw_size;
2506         }
2507
2508         if (type & PERF_SAMPLE_BRANCH_STACK) {
2509                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2510                 sz += sizeof(u64);
2511                 result += sz;
2512         }
2513
2514         if (type & PERF_SAMPLE_REGS_USER) {
2515                 if (sample->user_regs.abi) {
2516                         result += sizeof(u64);
2517                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2518                         result += sz;
2519                 } else {
2520                         result += sizeof(u64);
2521                 }
2522         }
2523
2524         if (type & PERF_SAMPLE_STACK_USER) {
2525                 sz = sample->user_stack.size;
2526                 result += sizeof(u64);
2527                 if (sz) {
2528                         result += sz;
2529                         result += sizeof(u64);
2530                 }
2531         }
2532
2533         if (type & PERF_SAMPLE_WEIGHT)
2534                 result += sizeof(u64);
2535
2536         if (type & PERF_SAMPLE_DATA_SRC)
2537                 result += sizeof(u64);
2538
2539         if (type & PERF_SAMPLE_TRANSACTION)
2540                 result += sizeof(u64);
2541
2542         if (type & PERF_SAMPLE_REGS_INTR) {
2543                 if (sample->intr_regs.abi) {
2544                         result += sizeof(u64);
2545                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2546                         result += sz;
2547                 } else {
2548                         result += sizeof(u64);
2549                 }
2550         }
2551
2552         if (type & PERF_SAMPLE_PHYS_ADDR)
2553                 result += sizeof(u64);
2554
2555         return result;
2556 }
2557
2558 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2559                                   u64 read_format,
2560                                   const struct perf_sample *sample)
2561 {
2562         u64 *array;
2563         size_t sz;
2564         /*
2565          * used for cross-endian analysis. See git commit 65014ab3
2566          * for why this goofiness is needed.
2567          */
2568         union u64_swap u;
2569
2570         array = event->sample.array;
2571
2572         if (type & PERF_SAMPLE_IDENTIFIER) {
2573                 *array = sample->id;
2574                 array++;
2575         }
2576
2577         if (type & PERF_SAMPLE_IP) {
2578                 *array = sample->ip;
2579                 array++;
2580         }
2581
2582         if (type & PERF_SAMPLE_TID) {
2583                 u.val32[0] = sample->pid;
2584                 u.val32[1] = sample->tid;
2585                 *array = u.val64;
2586                 array++;
2587         }
2588
2589         if (type & PERF_SAMPLE_TIME) {
2590                 *array = sample->time;
2591                 array++;
2592         }
2593
2594         if (type & PERF_SAMPLE_ADDR) {
2595                 *array = sample->addr;
2596                 array++;
2597         }
2598
2599         if (type & PERF_SAMPLE_ID) {
2600                 *array = sample->id;
2601                 array++;
2602         }
2603
2604         if (type & PERF_SAMPLE_STREAM_ID) {
2605                 *array = sample->stream_id;
2606                 array++;
2607         }
2608
2609         if (type & PERF_SAMPLE_CPU) {
2610                 u.val32[0] = sample->cpu;
2611                 u.val32[1] = 0;
2612                 *array = u.val64;
2613                 array++;
2614         }
2615
2616         if (type & PERF_SAMPLE_PERIOD) {
2617                 *array = sample->period;
2618                 array++;
2619         }
2620
2621         if (type & PERF_SAMPLE_READ) {
2622                 if (read_format & PERF_FORMAT_GROUP)
2623                         *array = sample->read.group.nr;
2624                 else
2625                         *array = sample->read.one.value;
2626                 array++;
2627
2628                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2629                         *array = sample->read.time_enabled;
2630                         array++;
2631                 }
2632
2633                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2634                         *array = sample->read.time_running;
2635                         array++;
2636                 }
2637
2638                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2639                 if (read_format & PERF_FORMAT_GROUP) {
2640                         sz = sample->read.group.nr *
2641                              sizeof(struct sample_read_value);
2642                         memcpy(array, sample->read.group.values, sz);
2643                         array = (void *)array + sz;
2644                 } else {
2645                         *array = sample->read.one.id;
2646                         array++;
2647                 }
2648         }
2649
2650         if (type & PERF_SAMPLE_CALLCHAIN) {
2651                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2652                 memcpy(array, sample->callchain, sz);
2653                 array = (void *)array + sz;
2654         }
2655
2656         if (type & PERF_SAMPLE_RAW) {
2657                 u.val32[0] = sample->raw_size;
2658                 *array = u.val64;
2659                 array = (void *)array + sizeof(u32);
2660
2661                 memcpy(array, sample->raw_data, sample->raw_size);
2662                 array = (void *)array + sample->raw_size;
2663         }
2664
2665         if (type & PERF_SAMPLE_BRANCH_STACK) {
2666                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2667                 sz += sizeof(u64);
2668                 memcpy(array, sample->branch_stack, sz);
2669                 array = (void *)array + sz;
2670         }
2671
2672         if (type & PERF_SAMPLE_REGS_USER) {
2673                 if (sample->user_regs.abi) {
2674                         *array++ = sample->user_regs.abi;
2675                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2676                         memcpy(array, sample->user_regs.regs, sz);
2677                         array = (void *)array + sz;
2678                 } else {
2679                         *array++ = 0;
2680                 }
2681         }
2682
2683         if (type & PERF_SAMPLE_STACK_USER) {
2684                 sz = sample->user_stack.size;
2685                 *array++ = sz;
2686                 if (sz) {
2687                         memcpy(array, sample->user_stack.data, sz);
2688                         array = (void *)array + sz;
2689                         *array++ = sz;
2690                 }
2691         }
2692
2693         if (type & PERF_SAMPLE_WEIGHT) {
2694                 *array = sample->weight;
2695                 array++;
2696         }
2697
2698         if (type & PERF_SAMPLE_DATA_SRC) {
2699                 *array = sample->data_src;
2700                 array++;
2701         }
2702
2703         if (type & PERF_SAMPLE_TRANSACTION) {
2704                 *array = sample->transaction;
2705                 array++;
2706         }
2707
2708         if (type & PERF_SAMPLE_REGS_INTR) {
2709                 if (sample->intr_regs.abi) {
2710                         *array++ = sample->intr_regs.abi;
2711                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2712                         memcpy(array, sample->intr_regs.regs, sz);
2713                         array = (void *)array + sz;
2714                 } else {
2715                         *array++ = 0;
2716                 }
2717         }
2718
2719         if (type & PERF_SAMPLE_PHYS_ADDR) {
2720                 *array = sample->phys_addr;
2721                 array++;
2722         }
2723
2724         return 0;
2725 }
2726
2727 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2728 {
2729         return tep_find_field(evsel->tp_format, name);
2730 }
2731
2732 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2733                          const char *name)
2734 {
2735         struct tep_format_field *field = perf_evsel__field(evsel, name);
2736         int offset;
2737
2738         if (!field)
2739                 return NULL;
2740
2741         offset = field->offset;
2742
2743         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2744                 offset = *(int *)(sample->raw_data + field->offset);
2745                 offset &= 0xffff;
2746         }
2747
2748         return sample->raw_data + offset;
2749 }
2750
2751 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2752                          bool needs_swap)
2753 {
2754         u64 value;
2755         void *ptr = sample->raw_data + field->offset;
2756
2757         switch (field->size) {
2758         case 1:
2759                 return *(u8 *)ptr;
2760         case 2:
2761                 value = *(u16 *)ptr;
2762                 break;
2763         case 4:
2764                 value = *(u32 *)ptr;
2765                 break;
2766         case 8:
2767                 memcpy(&value, ptr, sizeof(u64));
2768                 break;
2769         default:
2770                 return 0;
2771         }
2772
2773         if (!needs_swap)
2774                 return value;
2775
2776         switch (field->size) {
2777         case 2:
2778                 return bswap_16(value);
2779         case 4:
2780                 return bswap_32(value);
2781         case 8:
2782                 return bswap_64(value);
2783         default:
2784                 return 0;
2785         }
2786
2787         return 0;
2788 }
2789
2790 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2791                        const char *name)
2792 {
2793         struct tep_format_field *field = perf_evsel__field(evsel, name);
2794
2795         if (!field)
2796                 return 0;
2797
2798         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2799 }
2800
2801 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2802                           char *msg, size_t msgsize)
2803 {
2804         int paranoid;
2805
2806         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2807             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2808             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2809                 /*
2810                  * If it's cycles then fall back to hrtimer based
2811                  * cpu-clock-tick sw counter, which is always available even if
2812                  * no PMU support.
2813                  *
2814                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2815                  * b0a873e).
2816                  */
2817                 scnprintf(msg, msgsize, "%s",
2818 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2819
2820                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2821                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2822
2823                 zfree(&evsel->name);
2824                 return true;
2825         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2826                    (paranoid = perf_event_paranoid()) > 1) {
2827                 const char *name = perf_evsel__name(evsel);
2828                 char *new_name;
2829                 const char *sep = ":";
2830
2831                 /* Is there already the separator in the name. */
2832                 if (strchr(name, '/') ||
2833                     strchr(name, ':'))
2834                         sep = "";
2835
2836                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2837                         return false;
2838
2839                 if (evsel->name)
2840                         free(evsel->name);
2841                 evsel->name = new_name;
2842                 scnprintf(msg, msgsize,
2843 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2844                 evsel->attr.exclude_kernel = 1;
2845
2846                 return true;
2847         }
2848
2849         return false;
2850 }
2851
2852 static bool find_process(const char *name)
2853 {
2854         size_t len = strlen(name);
2855         DIR *dir;
2856         struct dirent *d;
2857         int ret = -1;
2858
2859         dir = opendir(procfs__mountpoint());
2860         if (!dir)
2861                 return false;
2862
2863         /* Walk through the directory. */
2864         while (ret && (d = readdir(dir)) != NULL) {
2865                 char path[PATH_MAX];
2866                 char *data;
2867                 size_t size;
2868
2869                 if ((d->d_type != DT_DIR) ||
2870                      !strcmp(".", d->d_name) ||
2871                      !strcmp("..", d->d_name))
2872                         continue;
2873
2874                 scnprintf(path, sizeof(path), "%s/%s/comm",
2875                           procfs__mountpoint(), d->d_name);
2876
2877                 if (filename__read_str(path, &data, &size))
2878                         continue;
2879
2880                 ret = strncmp(name, data, len);
2881                 free(data);
2882         }
2883
2884         closedir(dir);
2885         return ret ? false : true;
2886 }
2887
2888 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2889                               int err, char *msg, size_t size)
2890 {
2891         char sbuf[STRERR_BUFSIZE];
2892         int printed = 0;
2893
2894         switch (err) {
2895         case EPERM:
2896         case EACCES:
2897                 if (err == EPERM)
2898                         printed = scnprintf(msg, size,
2899                                 "No permission to enable %s event.\n\n",
2900                                 perf_evsel__name(evsel));
2901
2902                 return scnprintf(msg + printed, size - printed,
2903                  "You may not have permission to collect %sstats.\n\n"
2904                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2905                  "which controls use of the performance events system by\n"
2906                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2907                  "The current value is %d:\n\n"
2908                  "  -1: Allow use of (almost) all events by all users\n"
2909                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2910                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2911                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2912                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2913                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2914                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2915                  "      kernel.perf_event_paranoid = -1\n" ,
2916                                  target->system_wide ? "system-wide " : "",
2917                                  perf_event_paranoid());
2918         case ENOENT:
2919                 return scnprintf(msg, size, "The %s event is not supported.",
2920                                  perf_evsel__name(evsel));
2921         case EMFILE:
2922                 return scnprintf(msg, size, "%s",
2923                          "Too many events are opened.\n"
2924                          "Probably the maximum number of open file descriptors has been reached.\n"
2925                          "Hint: Try again after reducing the number of events.\n"
2926                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2927         case ENOMEM:
2928                 if (evsel__has_callchain(evsel) &&
2929                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2930                         return scnprintf(msg, size,
2931                                          "Not enough memory to setup event with callchain.\n"
2932                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2933                                          "Hint: Current value: %d", sysctl__max_stack());
2934                 break;
2935         case ENODEV:
2936                 if (target->cpu_list)
2937                         return scnprintf(msg, size, "%s",
2938          "No such device - did you specify an out-of-range profile CPU?");
2939                 break;
2940         case EOPNOTSUPP:
2941                 if (evsel->attr.sample_period != 0)
2942                         return scnprintf(msg, size,
2943         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2944                                          perf_evsel__name(evsel));
2945                 if (evsel->attr.precise_ip)
2946                         return scnprintf(msg, size, "%s",
2947         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2948 #if defined(__i386__) || defined(__x86_64__)
2949                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2950                         return scnprintf(msg, size, "%s",
2951         "No hardware sampling interrupt available.\n");
2952 #endif
2953                 break;
2954         case EBUSY:
2955                 if (find_process("oprofiled"))
2956                         return scnprintf(msg, size,
2957         "The PMU counters are busy/taken by another profiler.\n"
2958         "We found oprofile daemon running, please stop it and try again.");
2959                 break;
2960         case EINVAL:
2961                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2962                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2963                 if (perf_missing_features.clockid)
2964                         return scnprintf(msg, size, "clockid feature not supported.");
2965                 if (perf_missing_features.clockid_wrong)
2966                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2967                 break;
2968         default:
2969                 break;
2970         }
2971
2972         return scnprintf(msg, size,
2973         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2974         "/bin/dmesg | grep -i perf may provide additional information.\n",
2975                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2976                          perf_evsel__name(evsel));
2977 }
2978
2979 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
2980 {
2981         if (evsel && evsel->evlist)
2982                 return evsel->evlist->env;
2983         return NULL;
2984 }
2985
2986 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
2987 {
2988         int cpu, thread;
2989
2990         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
2991                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
2992                      thread++) {
2993                         int fd = FD(evsel, cpu, thread);
2994
2995                         if (perf_evlist__id_add_fd(evlist, evsel,
2996                                                    cpu, thread, fd) < 0)
2997                                 return -1;
2998                 }
2999         }
3000
3001         return 0;
3002 }
3003
3004 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3005 {
3006         struct cpu_map *cpus = evsel->cpus;
3007         struct thread_map *threads = evsel->threads;
3008
3009         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3010                 return -ENOMEM;
3011
3012         return store_evsel_ids(evsel, evlist);
3013 }