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         /*
 299          * Now let the usual logic to set up the perf_event_attr defaults
 300          * to kick in when we return and before perf_evsel__open() is called.
 301          */
 302 new_event:
 303         evsel = perf_evsel__new(&attr);
 304         if (evsel == NULL)
 305                 goto out;
 306
 307         evsel->precise_max = true;
 308
 309         /* use asprintf() because free(evsel) assumes name is allocated */
 310         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 311                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 312                      attr.exclude_kernel ? "u" : "",
 313                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 314                 goto error_free;
 315 out:
 316         return evsel;
 317 error_free:
 318         perf_evsel__delete(evsel);
 319         evsel = NULL;
 320         goto out;
 321 }
 322
 323 /*
 324  * Returns pointer with encoded error via <linux/err.h> interface.
 325  */
 326 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 327 {
 328         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 329         int err = -ENOMEM;
 330
 331         if (evsel == NULL) {
 332                 goto out_err;
 333         } else {
 334                 struct perf_event_attr attr = {
 335                         .type          = PERF_TYPE_TRACEPOINT,
 336                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 337                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 338                 };
 339
 340                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 341                         goto out_free;
 342
 343                 evsel->tp_format = trace_event__tp_format(sys, name);
 344                 if (IS_ERR(evsel->tp_format)) {
 345                         err = PTR_ERR(evsel->tp_format);
 346                         goto out_free;
 347                 }
 348
 349                 event_attr_init(&attr);
 350                 attr.config = evsel->tp_format->id;
 351                 attr.sample_period = 1;
 352                 perf_evsel__init(evsel, &attr, idx);
 353         }
 354
 355         return evsel;
 356
 357 out_free:
 358         zfree(&evsel->name);
 359         free(evsel);
 360 out_err:
 361         return ERR_PTR(err);
 362 }
 363
 364 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 365         "cycles",
 366         "instructions",
 367         "cache-references",
 368         "cache-misses",
 369         "branches",
 370         "branch-misses",
 371         "bus-cycles",
 372         "stalled-cycles-frontend",
 373         "stalled-cycles-backend",
 374         "ref-cycles",
 375 };
 376
 377 static const char *__perf_evsel__hw_name(u64 config)
 378 {
 379         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 380                 return perf_evsel__hw_names[config];
 381
 382         return "unknown-hardware";
 383 }
 384
 385 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 386 {
 387         int colon = 0, r = 0;
 388         struct perf_event_attr *attr = &evsel->attr;
 389         bool exclude_guest_default = false;
 390
 391 #define MOD_PRINT(context, mod) do {                                    \
 392                 if (!attr->exclude_##context) {                         \
 393                         if (!colon) colon = ++r;                        \
 394                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 395                 } } while(0)
 396
 397         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 398                 MOD_PRINT(kernel, 'k');
 399                 MOD_PRINT(user, 'u');
 400                 MOD_PRINT(hv, 'h');
 401                 exclude_guest_default = true;
 402         }
 403
 404         if (attr->precise_ip) {
 405                 if (!colon)
 406                         colon = ++r;
 407                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 412                 MOD_PRINT(host, 'H');
 413                 MOD_PRINT(guest, 'G');
 414         }
 415 #undef MOD_PRINT
 416         if (colon)
 417                 bf[colon - 1] = ':';
 418         return r;
 419 }
 420
 421 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 422 {
 423         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 424         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 425 }
 426
 427 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 428         "cpu-clock",
 429         "task-clock",
 430         "page-faults",
 431         "context-switches",
 432         "cpu-migrations",
 433         "minor-faults",
 434         "major-faults",
 435         "alignment-faults",
 436         "emulation-faults",
 437         "dummy",
 438 };
 439
 440 static const char *__perf_evsel__sw_name(u64 config)
 441 {
 442         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 443                 return perf_evsel__sw_names[config];
 444         return "unknown-software";
 445 }
 446
 447 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 448 {
 449         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 450         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 451 }
 452
 453 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 454 {
 455         int r;
 456
 457         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 458
 459         if (type & HW_BREAKPOINT_R)
 460                 r += scnprintf(bf + r, size - r, "r");
 461
 462         if (type & HW_BREAKPOINT_W)
 463                 r += scnprintf(bf + r, size - r, "w");
 464
 465         if (type & HW_BREAKPOINT_X)
 466                 r += scnprintf(bf + r, size - r, "x");
 467
 468         return r;
 469 }
 470
 471 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 472 {
 473         struct perf_event_attr *attr = &evsel->attr;
 474         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 475         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 476 }
 477
 478 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 479                                 [PERF_EVSEL__MAX_ALIASES] = {
 480  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 481  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 482  { "LLC",       "L2",                                                   },
 483  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 484  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 485  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 486  { "node",                                                              },
 487 };
 488
 489 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 490                                    [PERF_EVSEL__MAX_ALIASES] = {
 491  { "load",      "loads",        "read",                                 },
 492  { "store",     "stores",       "write",                                },
 493  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 494 };
 495
 496 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 497                                        [PERF_EVSEL__MAX_ALIASES] = {
 498  { "refs",      "Reference",    "ops",          "access",               },
 499  { "misses",    "miss",                                                 },
 500 };
 501
 502 #define C(x)            PERF_COUNT_HW_CACHE_##x
 503 #define CACHE_READ      (1 << C(OP_READ))
 504 #define CACHE_WRITE     (1 << C(OP_WRITE))
 505 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 506 #define COP(x)          (1 << x)
 507
 508 /*
 509  * cache operartion stat
 510  * L1I : Read and prefetch only
 511  * ITLB and BPU : Read-only
 512  */
 513 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 514  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 515  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 516  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 517  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 518  [C(ITLB)]      = (CACHE_READ),
 519  [C(BPU)]       = (CACHE_READ),
 520  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 521 };
 522
 523 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 524 {
 525         if (perf_evsel__hw_cache_stat[type] & COP(op))
 526                 return true;    /* valid */
 527         else
 528                 return false;   /* invalid */
 529 }
 530
 531 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 532                                             char *bf, size_t size)
 533 {
 534         if (result) {
 535                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 536                                  perf_evsel__hw_cache_op[op][0],
 537                                  perf_evsel__hw_cache_result[result][0]);
 538         }
 539
 540         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 541                          perf_evsel__hw_cache_op[op][1]);
 542 }
 543
 544 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 545 {
 546         u8 op, result, type = (config >>  0) & 0xff;
 547         const char *err = "unknown-ext-hardware-cache-type";
 548
 549         if (type >= PERF_COUNT_HW_CACHE_MAX)
 550                 goto out_err;
 551
 552         op = (config >>  8) & 0xff;
 553         err = "unknown-ext-hardware-cache-op";
 554         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 555                 goto out_err;
 556
 557         result = (config >> 16) & 0xff;
 558         err = "unknown-ext-hardware-cache-result";
 559         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 560                 goto out_err;
 561
 562         err = "invalid-cache";
 563         if (!perf_evsel__is_cache_op_valid(type, op))
 564                 goto out_err;
 565
 566         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 567 out_err:
 568         return scnprintf(bf, size, "%s", err);
 569 }
 570
 571 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 572 {
 573         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 574         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 575 }
 576
 577 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 578 {
 579         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 580         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 581 }
 582
 583 static int perf_evsel__tool_name(char *bf, size_t size)
 584 {
 585         int ret = scnprintf(bf, size, "duration_time");
 586         return ret;
 587 }
 588
 589 const char *perf_evsel__name(struct perf_evsel *evsel)
 590 {
 591         char bf[128];
 592
 593         if (evsel->name)
 594                 return evsel->name;
 595
 596         switch (evsel->attr.type) {
 597         case PERF_TYPE_RAW:
 598                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 599                 break;
 600
 601         case PERF_TYPE_HARDWARE:
 602                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 603                 break;
 604
 605         case PERF_TYPE_HW_CACHE:
 606                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 607                 break;
 608
 609         case PERF_TYPE_SOFTWARE:
 610                 if (evsel->tool_event)
 611                         perf_evsel__tool_name(bf, sizeof(bf));
 612                 else
 613                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 614                 break;
 615
 616         case PERF_TYPE_TRACEPOINT:
 617                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 618                 break;
 619
 620         case PERF_TYPE_BREAKPOINT:
 621                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 622                 break;
 623
 624         default:
 625                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 626                           evsel->attr.type);
 627                 break;
 628         }
 629
 630         evsel->name = strdup(bf);
 631
 632         return evsel->name ?: "unknown";
 633 }
 634
 635 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 636 {
 637         return evsel->group_name ?: "anon group";
 638 }
 639
 640 /*
 641  * Returns the group details for the specified leader,
 642  * with following rules.
 643  *
 644  *  For record -e '{cycles,instructions}'
 645  *    'anon group { cycles:u, instructions:u }'
 646  *
 647  *  For record -e 'cycles,instructions' and report --group
 648  *    'cycles:u, instructions:u'
 649  */
 650 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 651 {
 652         int ret = 0;
 653         struct perf_evsel *pos;
 654         const char *group_name = perf_evsel__group_name(evsel);
 655
 656         if (!evsel->forced_leader)
 657                 ret = scnprintf(buf, size, "%s { ", group_name);
 658
 659         ret += scnprintf(buf + ret, size - ret, "%s",
 660                          perf_evsel__name(evsel));
 661
 662         for_each_group_member(pos, evsel)
 663                 ret += scnprintf(buf + ret, size - ret, ", %s",
 664                                  perf_evsel__name(pos));
 665
 666         if (!evsel->forced_leader)
 667                 ret += scnprintf(buf + ret, size - ret, " }");
 668
 669         return ret;
 670 }
 671
 672 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 673                                            struct record_opts *opts,
 674                                            struct callchain_param *param)
 675 {
 676         bool function = perf_evsel__is_function_event(evsel);
 677         struct perf_event_attr *attr = &evsel->attr;
 678
 679         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 680
 681         attr->sample_max_stack = param->max_stack;
 682
 683         if (param->record_mode == CALLCHAIN_LBR) {
 684                 if (!opts->branch_stack) {
 685                         if (attr->exclude_user) {
 686                                 pr_warning("LBR callstack option is only available "
 687                                            "to get user callchain information. "
 688                                            "Falling back to framepointers.\n");
 689                         } else {
 690                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 691                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 692                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 693                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 694                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 695                         }
 696                 } else
 697                          pr_warning("Cannot use LBR callstack with branch stack. "
 698                                     "Falling back to framepointers.\n");
 699         }
 700
 701         if (param->record_mode == CALLCHAIN_DWARF) {
 702                 if (!function) {
 703                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 704                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 705                         attr->sample_regs_user |= PERF_REGS_MASK;
 706                         attr->sample_stack_user = param->dump_size;
 707                         attr->exclude_callchain_user = 1;
 708                 } else {
 709                         pr_info("Cannot use DWARF unwind for function trace event,"
 710                                 " falling back to framepointers.\n");
 711                 }
 712         }
 713
 714         if (function) {
 715                 pr_info("Disabling user space callchains for function trace event.\n");
 716                 attr->exclude_callchain_user = 1;
 717         }
 718 }
 719
 720 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 721                                   struct record_opts *opts,
 722                                   struct callchain_param *param)
 723 {
 724         if (param->enabled)
 725                 return __perf_evsel__config_callchain(evsel, opts, param);
 726 }
 727
 728 static void
 729 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 730                             struct callchain_param *param)
 731 {
 732         struct perf_event_attr *attr = &evsel->attr;
 733
 734         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 735         if (param->record_mode == CALLCHAIN_LBR) {
 736                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 737                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 738                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 739         }
 740         if (param->record_mode == CALLCHAIN_DWARF) {
 741                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 742                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 743         }
 744 }
 745
 746 static void apply_config_terms(struct perf_evsel *evsel,
 747                                struct record_opts *opts, bool track)
 748 {
 749         struct perf_evsel_config_term *term;
 750         struct list_head *config_terms = &evsel->config_terms;
 751         struct perf_event_attr *attr = &evsel->attr;
 752         /* callgraph default */
 753         struct callchain_param param = {
 754                 .record_mode = callchain_param.record_mode,
 755         };
 756         u32 dump_size = 0;
 757         int max_stack = 0;
 758         const char *callgraph_buf = NULL;
 759
 760         list_for_each_entry(term, config_terms, list) {
 761                 switch (term->type) {
 762                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 763                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 764                                 attr->sample_period = term->val.period;
 765                                 attr->freq = 0;
 766                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 767                         }
 768                         break;
 769                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 770                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 771                                 attr->sample_freq = term->val.freq;
 772                                 attr->freq = 1;
 773                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 774                         }
 775                         break;
 776                 case PERF_EVSEL__CONFIG_TERM_TIME:
 777                         if (term->val.time)
 778                                 perf_evsel__set_sample_bit(evsel, TIME);
 779                         else
 780                                 perf_evsel__reset_sample_bit(evsel, TIME);
 781                         break;
 782                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 783                         callgraph_buf = term->val.callgraph;
 784                         break;
 785                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 786                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 787                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 788                                 parse_branch_str(term->val.branch,
 789                                                  &attr->branch_sample_type);
 790                         } else
 791                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 794                         dump_size = term->val.stack_user;
 795                         break;
 796                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 797                         max_stack = term->val.max_stack;
 798                         break;
 799                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 800                         evsel->max_events = term->val.max_events;
 801                         break;
 802                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 803                         /*
 804                          * attr->inherit should has already been set by
 805                          * perf_evsel__config. If user explicitly set
 806                          * inherit using config terms, override global
 807                          * opt->no_inherit setting.
 808                          */
 809                         attr->inherit = term->val.inherit ? 1 : 0;
 810                         break;
 811                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 812                         attr->write_backward = term->val.overwrite ? 1 : 0;
 813                         break;
 814                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 815                         break;
 816                 default:
 817                         break;
 818                 }
 819         }
 820
 821         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 822         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 823                 bool sample_address = false;
 824
 825                 if (max_stack) {
 826                         param.max_stack = max_stack;
 827                         if (callgraph_buf == NULL)
 828                                 callgraph_buf = "fp";
 829                 }
 830
 831                 /* parse callgraph parameters */
 832                 if (callgraph_buf != NULL) {
 833                         if (!strcmp(callgraph_buf, "no")) {
 834                                 param.enabled = false;
 835                                 param.record_mode = CALLCHAIN_NONE;
 836                         } else {
 837                                 param.enabled = true;
 838                                 if (parse_callchain_record(callgraph_buf, &param)) {
 839                                         pr_err("per-event callgraph setting for %s failed. "
 840                                                "Apply callgraph global setting for it\n",
 841                                                evsel->name);
 842                                         return;
 843                                 }
 844                                 if (param.record_mode == CALLCHAIN_DWARF)
 845                                         sample_address = true;
 846                         }
 847                 }
 848                 if (dump_size > 0) {
 849                         dump_size = round_up(dump_size, sizeof(u64));
 850                         param.dump_size = dump_size;
 851                 }
 852
 853                 /* If global callgraph set, clear it */
 854                 if (callchain_param.enabled)
 855                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 856
 857                 /* set perf-event callgraph */
 858                 if (param.enabled) {
 859                         if (sample_address) {
 860                                 perf_evsel__set_sample_bit(evsel, ADDR);
 861                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 862                                 evsel->attr.mmap_data = track;
 863                         }
 864                         perf_evsel__config_callchain(evsel, opts, &param);
 865                 }
 866         }
 867 }
 868
 869 static bool is_dummy_event(struct perf_evsel *evsel)
 870 {
 871         return (evsel->attr.type == PERF_TYPE_SOFTWARE) &&
 872                (evsel->attr.config == PERF_COUNT_SW_DUMMY);
 873 }
 874
 875 /*
 876  * The enable_on_exec/disabled value strategy:
 877  *
 878  *  1) For any type of traced program:
 879  *    - all independent events and group leaders are disabled
 880  *    - all group members are enabled
 881  *
 882  *     Group members are ruled by group leaders. They need to
 883  *     be enabled, because the group scheduling relies on that.
 884  *
 885  *  2) For traced programs executed by perf:
 886  *     - all independent events and group leaders have
 887  *       enable_on_exec set
 888  *     - we don't specifically enable or disable any event during
 889  *       the record command
 890  *
 891  *     Independent events and group leaders are initially disabled
 892  *     and get enabled by exec. Group members are ruled by group
 893  *     leaders as stated in 1).
 894  *
 895  *  3) For traced programs attached by perf (pid/tid):
 896  *     - we specifically enable or disable all events during
 897  *       the record command
 898  *
 899  *     When attaching events to already running traced we
 900  *     enable/disable events specifically, as there's no
 901  *     initial traced exec call.
 902  */
 903 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 904                         struct callchain_param *callchain)
 905 {
 906         struct perf_evsel *leader = evsel->leader;
 907         struct perf_event_attr *attr = &evsel->attr;
 908         int track = evsel->tracking;
 909         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 910
 911         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 912         attr->inherit       = !opts->no_inherit;
 913         attr->write_backward = opts->overwrite ? 1 : 0;
 914
 915         perf_evsel__set_sample_bit(evsel, IP);
 916         perf_evsel__set_sample_bit(evsel, TID);
 917
 918         if (evsel->sample_read) {
 919                 perf_evsel__set_sample_bit(evsel, READ);
 920
 921                 /*
 922                  * We need ID even in case of single event, because
 923                  * PERF_SAMPLE_READ process ID specific data.
 924                  */
 925                 perf_evsel__set_sample_id(evsel, false);
 926
 927                 /*
 928                  * Apply group format only if we belong to group
 929                  * with more than one members.
 930                  */
 931                 if (leader->nr_members > 1) {
 932                         attr->read_format |= PERF_FORMAT_GROUP;
 933                         attr->inherit = 0;
 934                 }
 935         }
 936
 937         /*
 938          * We default some events to have a default interval. But keep
 939          * it a weak assumption overridable by the user.
 940          */
 941         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 942                                      opts->user_interval != ULLONG_MAX)) {
 943                 if (opts->freq) {
 944                         perf_evsel__set_sample_bit(evsel, PERIOD);
 945                         attr->freq              = 1;
 946                         attr->sample_freq       = opts->freq;
 947                 } else {
 948                         attr->sample_period = opts->default_interval;
 949                 }
 950         }
 951
 952         /*
 953          * Disable sampling for all group members other
 954          * than leader in case leader 'leads' the sampling.
 955          */
 956         if ((leader != evsel) && leader->sample_read) {
 957                 attr->freq           = 0;
 958                 attr->sample_freq    = 0;
 959                 attr->sample_period  = 0;
 960                 attr->write_backward = 0;
 961
 962                 /*
 963                  * We don't get sample for slave events, we make them
 964                  * when delivering group leader sample. Set the slave
 965                  * event to follow the master sample_type to ease up
 966                  * report.
 967                  */
 968                 attr->sample_type = leader->attr.sample_type;
 969         }
 970
 971         if (opts->no_samples)
 972                 attr->sample_freq = 0;
 973
 974         if (opts->inherit_stat) {
 975                 evsel->attr.read_format |=
 976                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 977                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 978                         PERF_FORMAT_ID;
 979                 attr->inherit_stat = 1;
 980         }
 981
 982         if (opts->sample_address) {
 983                 perf_evsel__set_sample_bit(evsel, ADDR);
 984                 attr->mmap_data = track;
 985         }
 986
 987         /*
 988          * We don't allow user space callchains for  function trace
 989          * event, due to issues with page faults while tracing page
 990          * fault handler and its overall trickiness nature.
 991          */
 992         if (perf_evsel__is_function_event(evsel))
 993                 evsel->attr.exclude_callchain_user = 1;
 994
 995         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 996                 perf_evsel__config_callchain(evsel, opts, callchain);
 997
 998         if (opts->sample_intr_regs) {
 999                 attr->sample_regs_intr = opts->sample_intr_regs;
1000                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1001         }
1002
1003         if (opts->sample_user_regs) {
1004                 attr->sample_regs_user |= opts->sample_user_regs;
1005                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1006         }
1007
1008         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1009                 perf_evsel__set_sample_bit(evsel, CPU);
1010
1011         /*
1012          * When the user explicitly disabled time don't force it here.
1013          */
1014         if (opts->sample_time &&
1015             (!perf_missing_features.sample_id_all &&
1016             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1017              opts->sample_time_set)))
1018                 perf_evsel__set_sample_bit(evsel, TIME);
1019
1020         if (opts->raw_samples && !evsel->no_aux_samples) {
1021                 perf_evsel__set_sample_bit(evsel, TIME);
1022                 perf_evsel__set_sample_bit(evsel, RAW);
1023                 perf_evsel__set_sample_bit(evsel, CPU);
1024         }
1025
1026         if (opts->sample_address)
1027                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1028
1029         if (opts->sample_phys_addr)
1030                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1031
1032         if (opts->no_buffering) {
1033                 attr->watermark = 0;
1034                 attr->wakeup_events = 1;
1035         }
1036         if (opts->branch_stack && !evsel->no_aux_samples) {
1037                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1038                 attr->branch_sample_type = opts->branch_stack;
1039         }
1040
1041         if (opts->sample_weight)
1042                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1043
1044         attr->task  = track;
1045         attr->mmap  = track;
1046         attr->mmap2 = track && !perf_missing_features.mmap2;
1047         attr->comm  = track;
1048         attr->ksymbol = track && !perf_missing_features.ksymbol;
1049         attr->bpf_event = track && !opts->no_bpf_event &&
1050                 !perf_missing_features.bpf_event;
1051
1052         if (opts->record_namespaces)
1053                 attr->namespaces  = track;
1054
1055         if (opts->record_switch_events)
1056                 attr->context_switch = track;
1057
1058         if (opts->sample_transaction)
1059                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1060
1061         if (opts->running_time) {
1062                 evsel->attr.read_format |=
1063                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1064                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1065         }
1066
1067         /*
1068          * XXX see the function comment above
1069          *
1070          * Disabling only independent events or group leaders,
1071          * keeping group members enabled.
1072          */
1073         if (perf_evsel__is_group_leader(evsel))
1074                 attr->disabled = 1;
1075
1076         /*
1077          * Setting enable_on_exec for independent events and
1078          * group leaders for traced executed by perf.
1079          */
1080         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1081                 !opts->initial_delay)
1082                 attr->enable_on_exec = 1;
1083
1084         if (evsel->immediate) {
1085                 attr->disabled = 0;
1086                 attr->enable_on_exec = 0;
1087         }
1088
1089         clockid = opts->clockid;
1090         if (opts->use_clockid) {
1091                 attr->use_clockid = 1;
1092                 attr->clockid = opts->clockid;
1093         }
1094
1095         if (evsel->precise_max)
1096                 attr->precise_ip = 3;
1097
1098         if (opts->all_user) {
1099                 attr->exclude_kernel = 1;
1100                 attr->exclude_user   = 0;
1101         }
1102
1103         if (opts->all_kernel) {
1104                 attr->exclude_kernel = 0;
1105                 attr->exclude_user   = 1;
1106         }
1107
1108         if (evsel->own_cpus || evsel->unit)
1109                 evsel->attr.read_format |= PERF_FORMAT_ID;
1110
1111         /*
1112          * Apply event specific term settings,
1113          * it overloads any global configuration.
1114          */
1115         apply_config_terms(evsel, opts, track);
1116
1117         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1118
1119         /* The --period option takes the precedence. */
1120         if (opts->period_set) {
1121                 if (opts->period)
1122                         perf_evsel__set_sample_bit(evsel, PERIOD);
1123                 else
1124                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1125         }
1126
1127         /*
1128          * For initial_delay, a dummy event is added implicitly.
1129          * The software event will trigger -EOPNOTSUPP error out,
1130          * if BRANCH_STACK bit is set.
1131          */
1132         if (opts->initial_delay && is_dummy_event(evsel))
1133                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1134 }
1135
1136 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1137 {
1138         if (evsel->system_wide)
1139                 nthreads = 1;
1140
1141         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1142
1143         if (evsel->fd) {
1144                 int cpu, thread;
1145                 for (cpu = 0; cpu < ncpus; cpu++) {
1146                         for (thread = 0; thread < nthreads; thread++) {
1147                                 FD(evsel, cpu, thread) = -1;
1148                         }
1149                 }
1150         }
1151
1152         return evsel->fd != NULL ? 0 : -ENOMEM;
1153 }
1154
1155 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1156                           int ioc,  void *arg)
1157 {
1158         int cpu, thread;
1159
1160         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1161                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1162                         int fd = FD(evsel, cpu, thread),
1163                             err = ioctl(fd, ioc, arg);
1164
1165                         if (err)
1166                                 return err;
1167                 }
1168         }
1169
1170         return 0;
1171 }
1172
1173 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1174 {
1175         return perf_evsel__run_ioctl(evsel,
1176                                      PERF_EVENT_IOC_SET_FILTER,
1177                                      (void *)filter);
1178 }
1179
1180 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1181 {
1182         char *new_filter = strdup(filter);
1183
1184         if (new_filter != NULL) {
1185                 free(evsel->filter);
1186                 evsel->filter = new_filter;
1187                 return 0;
1188         }
1189
1190         return -1;
1191 }
1192
1193 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1194                                      const char *fmt, const char *filter)
1195 {
1196         char *new_filter;
1197
1198         if (evsel->filter == NULL)
1199                 return perf_evsel__set_filter(evsel, filter);
1200
1201         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1202                 free(evsel->filter);
1203                 evsel->filter = new_filter;
1204                 return 0;
1205         }
1206
1207         return -1;
1208 }
1209
1210 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1211 {
1212         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1213 }
1214
1215 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1216 {
1217         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1218 }
1219
1220 int perf_evsel__enable(struct perf_evsel *evsel)
1221 {
1222         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_ENABLE, 0);
1223
1224         if (!err)
1225                 evsel->disabled = false;
1226
1227         return err;
1228 }
1229
1230 int perf_evsel__disable(struct perf_evsel *evsel)
1231 {
1232         int err = perf_evsel__run_ioctl(evsel, PERF_EVENT_IOC_DISABLE, 0);
1233         /*
1234          * We mark it disabled here so that tools that disable a event can
1235          * ignore events after they disable it. I.e. the ring buffer may have
1236          * already a few more events queued up before the kernel got the stop
1237          * request.
1238          */
1239         if (!err)
1240                 evsel->disabled = true;
1241
1242         return err;
1243 }
1244
1245 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1246 {
1247         if (ncpus == 0 || nthreads == 0)
1248                 return 0;
1249
1250         if (evsel->system_wide)
1251                 nthreads = 1;
1252
1253         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1254         if (evsel->sample_id == NULL)
1255                 return -ENOMEM;
1256
1257         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1258         if (evsel->id == NULL) {
1259                 xyarray__delete(evsel->sample_id);
1260                 evsel->sample_id = NULL;
1261                 return -ENOMEM;
1262         }
1263
1264         return 0;
1265 }
1266
1267 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1268 {
1269         xyarray__delete(evsel->fd);
1270         evsel->fd = NULL;
1271 }
1272
1273 static void perf_evsel__free_id(struct perf_evsel *evsel)
1274 {
1275         xyarray__delete(evsel->sample_id);
1276         evsel->sample_id = NULL;
1277         zfree(&evsel->id);
1278 }
1279
1280 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1281 {
1282         struct perf_evsel_config_term *term, *h;
1283
1284         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1285                 list_del(&term->list);
1286                 free(term);
1287         }
1288 }
1289
1290 void perf_evsel__close_fd(struct perf_evsel *evsel)
1291 {
1292         int cpu, thread;
1293
1294         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1295                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1296                         close(FD(evsel, cpu, thread));
1297                         FD(evsel, cpu, thread) = -1;
1298                 }
1299 }
1300
1301 void perf_evsel__exit(struct perf_evsel *evsel)
1302 {
1303         assert(list_empty(&evsel->node));
1304         assert(evsel->evlist == NULL);
1305         perf_evsel__free_counts(evsel);
1306         perf_evsel__free_fd(evsel);
1307         perf_evsel__free_id(evsel);
1308         perf_evsel__free_config_terms(evsel);
1309         cgroup__put(evsel->cgrp);
1310         cpu_map__put(evsel->cpus);
1311         cpu_map__put(evsel->own_cpus);
1312         thread_map__put(evsel->threads);
1313         zfree(&evsel->group_name);
1314         zfree(&evsel->name);
1315         perf_evsel__object.fini(evsel);
1316 }
1317
1318 void perf_evsel__delete(struct perf_evsel *evsel)
1319 {
1320         perf_evsel__exit(evsel);
1321         free(evsel);
1322 }
1323
1324 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1325                                 struct perf_counts_values *count)
1326 {
1327         struct perf_counts_values tmp;
1328
1329         if (!evsel->prev_raw_counts)
1330                 return;
1331
1332         if (cpu == -1) {
1333                 tmp = evsel->prev_raw_counts->aggr;
1334                 evsel->prev_raw_counts->aggr = *count;
1335         } else {
1336                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1337                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1338         }
1339
1340         count->val = count->val - tmp.val;
1341         count->ena = count->ena - tmp.ena;
1342         count->run = count->run - tmp.run;
1343 }
1344
1345 void perf_counts_values__scale(struct perf_counts_values *count,
1346                                bool scale, s8 *pscaled)
1347 {
1348         s8 scaled = 0;
1349
1350         if (scale) {
1351                 if (count->run == 0) {
1352                         scaled = -1;
1353                         count->val = 0;
1354                 } else if (count->run < count->ena) {
1355                         scaled = 1;
1356                         count->val = (u64)((double) count->val * count->ena / count->run);
1357                 }
1358         }
1359
1360         if (pscaled)
1361                 *pscaled = scaled;
1362 }
1363
1364 static int perf_evsel__read_size(struct perf_evsel *evsel)
1365 {
1366         u64 read_format = evsel->attr.read_format;
1367         int entry = sizeof(u64); /* value */
1368         int size = 0;
1369         int nr = 1;
1370
1371         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1372                 size += sizeof(u64);
1373
1374         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1375                 size += sizeof(u64);
1376
1377         if (read_format & PERF_FORMAT_ID)
1378                 entry += sizeof(u64);
1379
1380         if (read_format & PERF_FORMAT_GROUP) {
1381                 nr = evsel->nr_members;
1382                 size += sizeof(u64);
1383         }
1384
1385         size += entry * nr;
1386         return size;
1387 }
1388
1389 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1390                      struct perf_counts_values *count)
1391 {
1392         size_t size = perf_evsel__read_size(evsel);
1393
1394         memset(count, 0, sizeof(*count));
1395
1396         if (FD(evsel, cpu, thread) < 0)
1397                 return -EINVAL;
1398
1399         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1400                 return -errno;
1401
1402         return 0;
1403 }
1404
1405 static int
1406 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1407 {
1408         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1409
1410         return perf_evsel__read(evsel, cpu, thread, count);
1411 }
1412
1413 static void
1414 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1415                       u64 val, u64 ena, u64 run)
1416 {
1417         struct perf_counts_values *count;
1418
1419         count = perf_counts(counter->counts, cpu, thread);
1420
1421         count->val    = val;
1422         count->ena    = ena;
1423         count->run    = run;
1424         count->loaded = true;
1425 }
1426
1427 static int
1428 perf_evsel__process_group_data(struct perf_evsel *leader,
1429                                int cpu, int thread, u64 *data)
1430 {
1431         u64 read_format = leader->attr.read_format;
1432         struct sample_read_value *v;
1433         u64 nr, ena = 0, run = 0, i;
1434
1435         nr = *data++;
1436
1437         if (nr != (u64) leader->nr_members)
1438                 return -EINVAL;
1439
1440         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1441                 ena = *data++;
1442
1443         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1444                 run = *data++;
1445
1446         v = (struct sample_read_value *) data;
1447
1448         perf_evsel__set_count(leader, cpu, thread,
1449                               v[0].value, ena, run);
1450
1451         for (i = 1; i < nr; i++) {
1452                 struct perf_evsel *counter;
1453
1454                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1455                 if (!counter)
1456                         return -EINVAL;
1457
1458                 perf_evsel__set_count(counter, cpu, thread,
1459                                       v[i].value, ena, run);
1460         }
1461
1462         return 0;
1463 }
1464
1465 static int
1466 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1467 {
1468         struct perf_stat_evsel *ps = leader->stats;
1469         u64 read_format = leader->attr.read_format;
1470         int size = perf_evsel__read_size(leader);
1471         u64 *data = ps->group_data;
1472
1473         if (!(read_format & PERF_FORMAT_ID))
1474                 return -EINVAL;
1475
1476         if (!perf_evsel__is_group_leader(leader))
1477                 return -EINVAL;
1478
1479         if (!data) {
1480                 data = zalloc(size);
1481                 if (!data)
1482                         return -ENOMEM;
1483
1484                 ps->group_data = data;
1485         }
1486
1487         if (FD(leader, cpu, thread) < 0)
1488                 return -EINVAL;
1489
1490         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1491                 return -errno;
1492
1493         return perf_evsel__process_group_data(leader, cpu, thread, data);
1494 }
1495
1496 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1497 {
1498         u64 read_format = evsel->attr.read_format;
1499
1500         if (read_format & PERF_FORMAT_GROUP)
1501                 return perf_evsel__read_group(evsel, cpu, thread);
1502         else
1503                 return perf_evsel__read_one(evsel, cpu, thread);
1504 }
1505
1506 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1507                               int cpu, int thread, bool scale)
1508 {
1509         struct perf_counts_values count;
1510         size_t nv = scale ? 3 : 1;
1511
1512         if (FD(evsel, cpu, thread) < 0)
1513                 return -EINVAL;
1514
1515         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1516                 return -ENOMEM;
1517
1518         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1519                 return -errno;
1520
1521         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1522         perf_counts_values__scale(&count, scale, NULL);
1523         *perf_counts(evsel->counts, cpu, thread) = count;
1524         return 0;
1525 }
1526
1527 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1528 {
1529         struct perf_evsel *leader = evsel->leader;
1530         int fd;
1531
1532         if (perf_evsel__is_group_leader(evsel))
1533                 return -1;
1534
1535         /*
1536          * Leader must be already processed/open,
1537          * if not it's a bug.
1538          */
1539         BUG_ON(!leader->fd);
1540
1541         fd = FD(leader, cpu, thread);
1542         BUG_ON(fd == -1);
1543
1544         return fd;
1545 }
1546
1547 struct bit_names {
1548         int bit;
1549         const char *name;
1550 };
1551
1552 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1553 {
1554         bool first_bit = true;
1555         int i = 0;
1556
1557         do {
1558                 if (value & bits[i].bit) {
1559                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1560                         first_bit = false;
1561                 }
1562         } while (bits[++i].name != NULL);
1563 }
1564
1565 static void __p_sample_type(char *buf, size_t size, u64 value)
1566 {
1567 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1568         struct bit_names bits[] = {
1569                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1570                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1571                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1572                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1573                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1574                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1575                 { .name = NULL, }
1576         };
1577 #undef bit_name
1578         __p_bits(buf, size, value, bits);
1579 }
1580
1581 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1582 {
1583 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1584         struct bit_names bits[] = {
1585                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1586                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1587                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1588                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1589                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1590                 { .name = NULL, }
1591         };
1592 #undef bit_name
1593         __p_bits(buf, size, value, bits);
1594 }
1595
1596 static void __p_read_format(char *buf, size_t size, u64 value)
1597 {
1598 #define bit_name(n) { PERF_FORMAT_##n, #n }
1599         struct bit_names bits[] = {
1600                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1601                 bit_name(ID), bit_name(GROUP),
1602                 { .name = NULL, }
1603         };
1604 #undef bit_name
1605         __p_bits(buf, size, value, bits);
1606 }
1607
1608 #define BUF_SIZE                1024
1609
1610 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1611 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1612 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1613 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1614 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1615 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1616
1617 #define PRINT_ATTRn(_n, _f, _p)                         \
1618 do {                                                    \
1619         if (attr->_f) {                                 \
1620                 _p(attr->_f);                           \
1621                 ret += attr__fprintf(fp, _n, buf, priv);\
1622         }                                               \
1623 } while (0)
1624
1625 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1626
1627 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1628                              attr__fprintf_f attr__fprintf, void *priv)
1629 {
1630         char buf[BUF_SIZE];
1631         int ret = 0;
1632
1633         PRINT_ATTRf(type, p_unsigned);
1634         PRINT_ATTRf(size, p_unsigned);
1635         PRINT_ATTRf(config, p_hex);
1636         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1637         PRINT_ATTRf(sample_type, p_sample_type);
1638         PRINT_ATTRf(read_format, p_read_format);
1639
1640         PRINT_ATTRf(disabled, p_unsigned);
1641         PRINT_ATTRf(inherit, p_unsigned);
1642         PRINT_ATTRf(pinned, p_unsigned);
1643         PRINT_ATTRf(exclusive, p_unsigned);
1644         PRINT_ATTRf(exclude_user, p_unsigned);
1645         PRINT_ATTRf(exclude_kernel, p_unsigned);
1646         PRINT_ATTRf(exclude_hv, p_unsigned);
1647         PRINT_ATTRf(exclude_idle, p_unsigned);
1648         PRINT_ATTRf(mmap, p_unsigned);
1649         PRINT_ATTRf(comm, p_unsigned);
1650         PRINT_ATTRf(freq, p_unsigned);
1651         PRINT_ATTRf(inherit_stat, p_unsigned);
1652         PRINT_ATTRf(enable_on_exec, p_unsigned);
1653         PRINT_ATTRf(task, p_unsigned);
1654         PRINT_ATTRf(watermark, p_unsigned);
1655         PRINT_ATTRf(precise_ip, p_unsigned);
1656         PRINT_ATTRf(mmap_data, p_unsigned);
1657         PRINT_ATTRf(sample_id_all, p_unsigned);
1658         PRINT_ATTRf(exclude_host, p_unsigned);
1659         PRINT_ATTRf(exclude_guest, p_unsigned);
1660         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1661         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1662         PRINT_ATTRf(mmap2, p_unsigned);
1663         PRINT_ATTRf(comm_exec, p_unsigned);
1664         PRINT_ATTRf(use_clockid, p_unsigned);
1665         PRINT_ATTRf(context_switch, p_unsigned);
1666         PRINT_ATTRf(write_backward, p_unsigned);
1667         PRINT_ATTRf(namespaces, p_unsigned);
1668         PRINT_ATTRf(ksymbol, p_unsigned);
1669         PRINT_ATTRf(bpf_event, p_unsigned);
1670
1671         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1672         PRINT_ATTRf(bp_type, p_unsigned);
1673         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1674         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1675         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1676         PRINT_ATTRf(sample_regs_user, p_hex);
1677         PRINT_ATTRf(sample_stack_user, p_unsigned);
1678         PRINT_ATTRf(clockid, p_signed);
1679         PRINT_ATTRf(sample_regs_intr, p_hex);
1680         PRINT_ATTRf(aux_watermark, p_unsigned);
1681         PRINT_ATTRf(sample_max_stack, p_unsigned);
1682
1683         return ret;
1684 }
1685
1686 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1687                                 void *priv __maybe_unused)
1688 {
1689         return fprintf(fp, "  %-32s %s\n", name, val);
1690 }
1691
1692 static void perf_evsel__remove_fd(struct perf_evsel *pos,
1693                                   int nr_cpus, int nr_threads,
1694                                   int thread_idx)
1695 {
1696         for (int cpu = 0; cpu < nr_cpus; cpu++)
1697                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1698                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1699 }
1700
1701 static int update_fds(struct perf_evsel *evsel,
1702                       int nr_cpus, int cpu_idx,
1703                       int nr_threads, int thread_idx)
1704 {
1705         struct perf_evsel *pos;
1706
1707         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1708                 return -EINVAL;
1709
1710         evlist__for_each_entry(evsel->evlist, pos) {
1711                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1712
1713                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1714
1715                 /*
1716                  * Since fds for next evsel has not been created,
1717                  * there is no need to iterate whole event list.
1718                  */
1719                 if (pos == evsel)
1720                         break;
1721         }
1722         return 0;
1723 }
1724
1725 static bool ignore_missing_thread(struct perf_evsel *evsel,
1726                                   int nr_cpus, int cpu,
1727                                   struct thread_map *threads,
1728                                   int thread, int err)
1729 {
1730         pid_t ignore_pid = thread_map__pid(threads, thread);
1731
1732         if (!evsel->ignore_missing_thread)
1733                 return false;
1734
1735         /* The system wide setup does not work with threads. */
1736         if (evsel->system_wide)
1737                 return false;
1738
1739         /* The -ESRCH is perf event syscall errno for pid's not found. */
1740         if (err != -ESRCH)
1741                 return false;
1742
1743         /* If there's only one thread, let it fail. */
1744         if (threads->nr == 1)
1745                 return false;
1746
1747         /*
1748          * We should remove fd for missing_thread first
1749          * because thread_map__remove() will decrease threads->nr.
1750          */
1751         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1752                 return false;
1753
1754         if (thread_map__remove(threads, thread))
1755                 return false;
1756
1757         pr_warning("WARNING: Ignored open failure for pid %d\n",
1758                    ignore_pid);
1759         return true;
1760 }
1761
1762 static void display_attr(struct perf_event_attr *attr)
1763 {
1764         if (verbose >= 2) {
1765                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1766                 fprintf(stderr, "perf_event_attr:\n");
1767                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1768                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1769         }
1770 }
1771
1772 static int perf_event_open(struct perf_evsel *evsel,
1773                            pid_t pid, int cpu, int group_fd,
1774                            unsigned long flags)
1775 {
1776         int precise_ip = evsel->attr.precise_ip;
1777         int fd;
1778
1779         while (1) {
1780                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1781                           pid, cpu, group_fd, flags);
1782
1783                 fd = sys_perf_event_open(&evsel->attr, pid, cpu, group_fd, flags);
1784                 if (fd >= 0)
1785                         break;
1786
1787                 /*
1788                  * Do quick precise_ip fallback if:
1789                  *  - there is precise_ip set in perf_event_attr
1790                  *  - maximum precise is requested
1791                  *  - sys_perf_event_open failed with ENOTSUP error,
1792                  *    which is associated with wrong precise_ip
1793                  */
1794                 if (!precise_ip || !evsel->precise_max || (errno != ENOTSUP))
1795                         break;
1796
1797                 /*
1798                  * We tried all the precise_ip values, and it's
1799                  * still failing, so leave it to standard fallback.
1800                  */
1801                 if (!evsel->attr.precise_ip) {
1802                         evsel->attr.precise_ip = precise_ip;
1803                         break;
1804                 }
1805
1806                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1807                 evsel->attr.precise_ip--;
1808                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->attr.precise_ip);
1809                 display_attr(&evsel->attr);
1810         }
1811
1812         return fd;
1813 }
1814
1815 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1816                      struct thread_map *threads)
1817 {
1818         int cpu, thread, nthreads;
1819         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1820         int pid = -1, err;
1821         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1822
1823         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1824                 return -EINVAL;
1825
1826         if (cpus == NULL) {
1827                 static struct cpu_map *empty_cpu_map;
1828
1829                 if (empty_cpu_map == NULL) {
1830                         empty_cpu_map = cpu_map__dummy_new();
1831                         if (empty_cpu_map == NULL)
1832                                 return -ENOMEM;
1833                 }
1834
1835                 cpus = empty_cpu_map;
1836         }
1837
1838         if (threads == NULL) {
1839                 static struct thread_map *empty_thread_map;
1840
1841                 if (empty_thread_map == NULL) {
1842                         empty_thread_map = thread_map__new_by_tid(-1);
1843                         if (empty_thread_map == NULL)
1844                                 return -ENOMEM;
1845                 }
1846
1847                 threads = empty_thread_map;
1848         }
1849
1850         if (evsel->system_wide)
1851                 nthreads = 1;
1852         else
1853                 nthreads = threads->nr;
1854
1855         if (evsel->fd == NULL &&
1856             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1857                 return -ENOMEM;
1858
1859         if (evsel->cgrp) {
1860                 flags |= PERF_FLAG_PID_CGROUP;
1861                 pid = evsel->cgrp->fd;
1862         }
1863
1864 fallback_missing_features:
1865         if (perf_missing_features.clockid_wrong)
1866                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1867         if (perf_missing_features.clockid) {
1868                 evsel->attr.use_clockid = 0;
1869                 evsel->attr.clockid = 0;
1870         }
1871         if (perf_missing_features.cloexec)
1872                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1873         if (perf_missing_features.mmap2)
1874                 evsel->attr.mmap2 = 0;
1875         if (perf_missing_features.exclude_guest)
1876                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1877         if (perf_missing_features.lbr_flags)
1878                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1879                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1880         if (perf_missing_features.group_read && evsel->attr.inherit)
1881                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1882         if (perf_missing_features.ksymbol)
1883                 evsel->attr.ksymbol = 0;
1884         if (perf_missing_features.bpf_event)
1885                 evsel->attr.bpf_event = 0;
1886 retry_sample_id:
1887         if (perf_missing_features.sample_id_all)
1888                 evsel->attr.sample_id_all = 0;
1889
1890         display_attr(&evsel->attr);
1891
1892         for (cpu = 0; cpu < cpus->nr; cpu++) {
1893
1894                 for (thread = 0; thread < nthreads; thread++) {
1895                         int fd, group_fd;
1896
1897                         if (!evsel->cgrp && !evsel->system_wide)
1898                                 pid = thread_map__pid(threads, thread);
1899
1900                         group_fd = get_group_fd(evsel, cpu, thread);
1901 retry_open:
1902                         test_attr__ready();
1903
1904                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1905                                              group_fd, flags);
1906
1907                         FD(evsel, cpu, thread) = fd;
1908
1909                         if (fd < 0) {
1910                                 err = -errno;
1911
1912                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1913                                         /*
1914                                          * We just removed 1 thread, so take a step
1915                                          * back on thread index and lower the upper
1916                                          * nthreads limit.
1917                                          */
1918                                         nthreads--;
1919                                         thread--;
1920
1921                                         /* ... and pretend like nothing have happened. */
1922                                         err = 0;
1923                                         continue;
1924                                 }
1925
1926                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1927                                           err);
1928                                 goto try_fallback;
1929                         }
1930
1931                         pr_debug2(" = %d\n", fd);
1932
1933                         if (evsel->bpf_fd >= 0) {
1934                                 int evt_fd = fd;
1935                                 int bpf_fd = evsel->bpf_fd;
1936
1937                                 err = ioctl(evt_fd,
1938                                             PERF_EVENT_IOC_SET_BPF,
1939                                             bpf_fd);
1940                                 if (err && errno != EEXIST) {
1941                                         pr_err("failed to attach bpf fd %d: %s\n",
1942                                                bpf_fd, strerror(errno));
1943                                         err = -EINVAL;
1944                                         goto out_close;
1945                                 }
1946                         }
1947
1948                         set_rlimit = NO_CHANGE;
1949
1950                         /*
1951                          * If we succeeded but had to kill clockid, fail and
1952                          * have perf_evsel__open_strerror() print us a nice
1953                          * error.
1954                          */
1955                         if (perf_missing_features.clockid ||
1956                             perf_missing_features.clockid_wrong) {
1957                                 err = -EINVAL;
1958                                 goto out_close;
1959                         }
1960                 }
1961         }
1962
1963         return 0;
1964
1965 try_fallback:
1966         /*
1967          * perf stat needs between 5 and 22 fds per CPU. When we run out
1968          * of them try to increase the limits.
1969          */
1970         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1971                 struct rlimit l;
1972                 int old_errno = errno;
1973
1974                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1975                         if (set_rlimit == NO_CHANGE)
1976                                 l.rlim_cur = l.rlim_max;
1977                         else {
1978                                 l.rlim_cur = l.rlim_max + 1000;
1979                                 l.rlim_max = l.rlim_cur;
1980                         }
1981                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1982                                 set_rlimit++;
1983                                 errno = old_errno;
1984                                 goto retry_open;
1985                         }
1986                 }
1987                 errno = old_errno;
1988         }
1989
1990         if (err != -EINVAL || cpu > 0 || thread > 0)
1991                 goto out_close;
1992
1993         /*
1994          * Must probe features in the order they were added to the
1995          * perf_event_attr interface.
1996          */
1997         if (!perf_missing_features.bpf_event && evsel->attr.bpf_event) {
1998                 perf_missing_features.bpf_event = true;
1999                 pr_debug2("switching off bpf_event\n");
2000                 goto fallback_missing_features;
2001         } else if (!perf_missing_features.ksymbol && evsel->attr.ksymbol) {
2002                 perf_missing_features.ksymbol = true;
2003                 pr_debug2("switching off ksymbol\n");
2004                 goto fallback_missing_features;
2005         } else if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
2006                 perf_missing_features.write_backward = true;
2007                 pr_debug2("switching off write_backward\n");
2008                 goto out_close;
2009         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
2010                 perf_missing_features.clockid_wrong = true;
2011                 pr_debug2("switching off clockid\n");
2012                 goto fallback_missing_features;
2013         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
2014                 perf_missing_features.clockid = true;
2015                 pr_debug2("switching off use_clockid\n");
2016                 goto fallback_missing_features;
2017         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
2018                 perf_missing_features.cloexec = true;
2019                 pr_debug2("switching off cloexec flag\n");
2020                 goto fallback_missing_features;
2021         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
2022                 perf_missing_features.mmap2 = true;
2023                 pr_debug2("switching off mmap2\n");
2024                 goto fallback_missing_features;
2025         } else if (!perf_missing_features.exclude_guest &&
2026                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
2027                 perf_missing_features.exclude_guest = true;
2028                 pr_debug2("switching off exclude_guest, exclude_host\n");
2029                 goto fallback_missing_features;
2030         } else if (!perf_missing_features.sample_id_all) {
2031                 perf_missing_features.sample_id_all = true;
2032                 pr_debug2("switching off sample_id_all\n");
2033                 goto retry_sample_id;
2034         } else if (!perf_missing_features.lbr_flags &&
2035                         (evsel->attr.branch_sample_type &
2036                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2037                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2038                 perf_missing_features.lbr_flags = true;
2039                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
2040                 goto fallback_missing_features;
2041         } else if (!perf_missing_features.group_read &&
2042                     evsel->attr.inherit &&
2043                    (evsel->attr.read_format & PERF_FORMAT_GROUP) &&
2044                    perf_evsel__is_group_leader(evsel)) {
2045                 perf_missing_features.group_read = true;
2046                 pr_debug2("switching off group read\n");
2047                 goto fallback_missing_features;
2048         }
2049 out_close:
2050         if (err)
2051                 threads->err_thread = thread;
2052
2053         do {
2054                 while (--thread >= 0) {
2055                         close(FD(evsel, cpu, thread));
2056                         FD(evsel, cpu, thread) = -1;
2057                 }
2058                 thread = nthreads;
2059         } while (--cpu >= 0);
2060         return err;
2061 }
2062
2063 void perf_evsel__close(struct perf_evsel *evsel)
2064 {
2065         if (evsel->fd == NULL)
2066                 return;
2067
2068         perf_evsel__close_fd(evsel);
2069         perf_evsel__free_fd(evsel);
2070 }
2071
2072 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
2073                              struct cpu_map *cpus)
2074 {
2075         return perf_evsel__open(evsel, cpus, NULL);
2076 }
2077
2078 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
2079                                 struct thread_map *threads)
2080 {
2081         return perf_evsel__open(evsel, NULL, threads);
2082 }
2083
2084 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
2085                                        const union perf_event *event,
2086                                        struct perf_sample *sample)
2087 {
2088         u64 type = evsel->attr.sample_type;
2089         const u64 *array = event->sample.array;
2090         bool swapped = evsel->needs_swap;
2091         union u64_swap u;
2092
2093         array += ((event->header.size -
2094                    sizeof(event->header)) / sizeof(u64)) - 1;
2095
2096         if (type & PERF_SAMPLE_IDENTIFIER) {
2097                 sample->id = *array;
2098                 array--;
2099         }
2100
2101         if (type & PERF_SAMPLE_CPU) {
2102                 u.val64 = *array;
2103                 if (swapped) {
2104                         /* undo swap of u64, then swap on individual u32s */
2105                         u.val64 = bswap_64(u.val64);
2106                         u.val32[0] = bswap_32(u.val32[0]);
2107                 }
2108
2109                 sample->cpu = u.val32[0];
2110                 array--;
2111         }
2112
2113         if (type & PERF_SAMPLE_STREAM_ID) {
2114                 sample->stream_id = *array;
2115                 array--;
2116         }
2117
2118         if (type & PERF_SAMPLE_ID) {
2119                 sample->id = *array;
2120                 array--;
2121         }
2122
2123         if (type & PERF_SAMPLE_TIME) {
2124                 sample->time = *array;
2125                 array--;
2126         }
2127
2128         if (type & PERF_SAMPLE_TID) {
2129                 u.val64 = *array;
2130                 if (swapped) {
2131                         /* undo swap of u64, then swap on individual u32s */
2132                         u.val64 = bswap_64(u.val64);
2133                         u.val32[0] = bswap_32(u.val32[0]);
2134                         u.val32[1] = bswap_32(u.val32[1]);
2135                 }
2136
2137                 sample->pid = u.val32[0];
2138                 sample->tid = u.val32[1];
2139                 array--;
2140         }
2141
2142         return 0;
2143 }
2144
2145 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2146                             u64 size)
2147 {
2148         return size > max_size || offset + size > endp;
2149 }
2150
2151 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2152         do {                                                            \
2153                 if (overflow(endp, (max_size), (offset), (size)))       \
2154                         return -EFAULT;                                 \
2155         } while (0)
2156
2157 #define OVERFLOW_CHECK_u64(offset) \
2158         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2159
2160 static int
2161 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2162 {
2163         /*
2164          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2165          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2166          * check the format does not go past the end of the event.
2167          */
2168         if (sample_size + sizeof(event->header) > event->header.size)
2169                 return -EFAULT;
2170
2171         return 0;
2172 }
2173
2174 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
2175                              struct perf_sample *data)
2176 {
2177         u64 type = evsel->attr.sample_type;
2178         bool swapped = evsel->needs_swap;
2179         const u64 *array;
2180         u16 max_size = event->header.size;
2181         const void *endp = (void *)event + max_size;
2182         u64 sz;
2183
2184         /*
2185          * used for cross-endian analysis. See git commit 65014ab3
2186          * for why this goofiness is needed.
2187          */
2188         union u64_swap u;
2189
2190         memset(data, 0, sizeof(*data));
2191         data->cpu = data->pid = data->tid = -1;
2192         data->stream_id = data->id = data->time = -1ULL;
2193         data->period = evsel->attr.sample_period;
2194         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2195         data->misc    = event->header.misc;
2196         data->id = -1ULL;
2197         data->data_src = PERF_MEM_DATA_SRC_NONE;
2198
2199         if (event->header.type != PERF_RECORD_SAMPLE) {
2200                 if (!evsel->attr.sample_id_all)
2201                         return 0;
2202                 return perf_evsel__parse_id_sample(evsel, event, data);
2203         }
2204
2205         array = event->sample.array;
2206
2207         if (perf_event__check_size(event, evsel->sample_size))
2208                 return -EFAULT;
2209
2210         if (type & PERF_SAMPLE_IDENTIFIER) {
2211                 data->id = *array;
2212                 array++;
2213         }
2214
2215         if (type & PERF_SAMPLE_IP) {
2216                 data->ip = *array;
2217                 array++;
2218         }
2219
2220         if (type & PERF_SAMPLE_TID) {
2221                 u.val64 = *array;
2222                 if (swapped) {
2223                         /* undo swap of u64, then swap on individual u32s */
2224                         u.val64 = bswap_64(u.val64);
2225                         u.val32[0] = bswap_32(u.val32[0]);
2226                         u.val32[1] = bswap_32(u.val32[1]);
2227                 }
2228
2229                 data->pid = u.val32[0];
2230                 data->tid = u.val32[1];
2231                 array++;
2232         }
2233
2234         if (type & PERF_SAMPLE_TIME) {
2235                 data->time = *array;
2236                 array++;
2237         }
2238
2239         if (type & PERF_SAMPLE_ADDR) {
2240                 data->addr = *array;
2241                 array++;
2242         }
2243
2244         if (type & PERF_SAMPLE_ID) {
2245                 data->id = *array;
2246                 array++;
2247         }
2248
2249         if (type & PERF_SAMPLE_STREAM_ID) {
2250                 data->stream_id = *array;
2251                 array++;
2252         }
2253
2254         if (type & PERF_SAMPLE_CPU) {
2255
2256                 u.val64 = *array;
2257                 if (swapped) {
2258                         /* undo swap of u64, then swap on individual u32s */
2259                         u.val64 = bswap_64(u.val64);
2260                         u.val32[0] = bswap_32(u.val32[0]);
2261                 }
2262
2263                 data->cpu = u.val32[0];
2264                 array++;
2265         }
2266
2267         if (type & PERF_SAMPLE_PERIOD) {
2268                 data->period = *array;
2269                 array++;
2270         }
2271
2272         if (type & PERF_SAMPLE_READ) {
2273                 u64 read_format = evsel->attr.read_format;
2274
2275                 OVERFLOW_CHECK_u64(array);
2276                 if (read_format & PERF_FORMAT_GROUP)
2277                         data->read.group.nr = *array;
2278                 else
2279                         data->read.one.value = *array;
2280
2281                 array++;
2282
2283                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2284                         OVERFLOW_CHECK_u64(array);
2285                         data->read.time_enabled = *array;
2286                         array++;
2287                 }
2288
2289                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2290                         OVERFLOW_CHECK_u64(array);
2291                         data->read.time_running = *array;
2292                         array++;
2293                 }
2294
2295                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2296                 if (read_format & PERF_FORMAT_GROUP) {
2297                         const u64 max_group_nr = UINT64_MAX /
2298                                         sizeof(struct sample_read_value);
2299
2300                         if (data->read.group.nr > max_group_nr)
2301                                 return -EFAULT;
2302                         sz = data->read.group.nr *
2303                              sizeof(struct sample_read_value);
2304                         OVERFLOW_CHECK(array, sz, max_size);
2305                         data->read.group.values =
2306                                         (struct sample_read_value *)array;
2307                         array = (void *)array + sz;
2308                 } else {
2309                         OVERFLOW_CHECK_u64(array);
2310                         data->read.one.id = *array;
2311                         array++;
2312                 }
2313         }
2314
2315         if (evsel__has_callchain(evsel)) {
2316                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2317
2318                 OVERFLOW_CHECK_u64(array);
2319                 data->callchain = (struct ip_callchain *)array++;
2320                 if (data->callchain->nr > max_callchain_nr)
2321                         return -EFAULT;
2322                 sz = data->callchain->nr * sizeof(u64);
2323                 OVERFLOW_CHECK(array, sz, max_size);
2324                 array = (void *)array + sz;
2325         }
2326
2327         if (type & PERF_SAMPLE_RAW) {
2328                 OVERFLOW_CHECK_u64(array);
2329                 u.val64 = *array;
2330
2331                 /*
2332                  * Undo swap of u64, then swap on individual u32s,
2333                  * get the size of the raw area and undo all of the
2334                  * swap. The pevent interface handles endianity by
2335                  * itself.
2336                  */
2337                 if (swapped) {
2338                         u.val64 = bswap_64(u.val64);
2339                         u.val32[0] = bswap_32(u.val32[0]);
2340                         u.val32[1] = bswap_32(u.val32[1]);
2341                 }
2342                 data->raw_size = u.val32[0];
2343
2344                 /*
2345                  * The raw data is aligned on 64bits including the
2346                  * u32 size, so it's safe to use mem_bswap_64.
2347                  */
2348                 if (swapped)
2349                         mem_bswap_64((void *) array, data->raw_size);
2350
2351                 array = (void *)array + sizeof(u32);
2352
2353                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2354                 data->raw_data = (void *)array;
2355                 array = (void *)array + data->raw_size;
2356         }
2357
2358         if (type & PERF_SAMPLE_BRANCH_STACK) {
2359                 const u64 max_branch_nr = UINT64_MAX /
2360                                           sizeof(struct branch_entry);
2361
2362                 OVERFLOW_CHECK_u64(array);
2363                 data->branch_stack = (struct branch_stack *)array++;
2364
2365                 if (data->branch_stack->nr > max_branch_nr)
2366                         return -EFAULT;
2367                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2368                 OVERFLOW_CHECK(array, sz, max_size);
2369                 array = (void *)array + sz;
2370         }
2371
2372         if (type & PERF_SAMPLE_REGS_USER) {
2373                 OVERFLOW_CHECK_u64(array);
2374                 data->user_regs.abi = *array;
2375                 array++;
2376
2377                 if (data->user_regs.abi) {
2378                         u64 mask = evsel->attr.sample_regs_user;
2379
2380                         sz = hweight_long(mask) * sizeof(u64);
2381                         OVERFLOW_CHECK(array, sz, max_size);
2382                         data->user_regs.mask = mask;
2383                         data->user_regs.regs = (u64 *)array;
2384                         array = (void *)array + sz;
2385                 }
2386         }
2387
2388         if (type & PERF_SAMPLE_STACK_USER) {
2389                 OVERFLOW_CHECK_u64(array);
2390                 sz = *array++;
2391
2392                 data->user_stack.offset = ((char *)(array - 1)
2393                                           - (char *) event);
2394
2395                 if (!sz) {
2396                         data->user_stack.size = 0;
2397                 } else {
2398                         OVERFLOW_CHECK(array, sz, max_size);
2399                         data->user_stack.data = (char *)array;
2400                         array = (void *)array + sz;
2401                         OVERFLOW_CHECK_u64(array);
2402                         data->user_stack.size = *array++;
2403                         if (WARN_ONCE(data->user_stack.size > sz,
2404                                       "user stack dump failure\n"))
2405                                 return -EFAULT;
2406                 }
2407         }
2408
2409         if (type & PERF_SAMPLE_WEIGHT) {
2410                 OVERFLOW_CHECK_u64(array);
2411                 data->weight = *array;
2412                 array++;
2413         }
2414
2415         if (type & PERF_SAMPLE_DATA_SRC) {
2416                 OVERFLOW_CHECK_u64(array);
2417                 data->data_src = *array;
2418                 array++;
2419         }
2420
2421         if (type & PERF_SAMPLE_TRANSACTION) {
2422                 OVERFLOW_CHECK_u64(array);
2423                 data->transaction = *array;
2424                 array++;
2425         }
2426
2427         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2428         if (type & PERF_SAMPLE_REGS_INTR) {
2429                 OVERFLOW_CHECK_u64(array);
2430                 data->intr_regs.abi = *array;
2431                 array++;
2432
2433                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2434                         u64 mask = evsel->attr.sample_regs_intr;
2435
2436                         sz = hweight_long(mask) * sizeof(u64);
2437                         OVERFLOW_CHECK(array, sz, max_size);
2438                         data->intr_regs.mask = mask;
2439                         data->intr_regs.regs = (u64 *)array;
2440                         array = (void *)array + sz;
2441                 }
2442         }
2443
2444         data->phys_addr = 0;
2445         if (type & PERF_SAMPLE_PHYS_ADDR) {
2446                 data->phys_addr = *array;
2447                 array++;
2448         }
2449
2450         return 0;
2451 }
2452
2453 int perf_evsel__parse_sample_timestamp(struct perf_evsel *evsel,
2454                                        union perf_event *event,
2455                                        u64 *timestamp)
2456 {
2457         u64 type = evsel->attr.sample_type;
2458         const u64 *array;
2459
2460         if (!(type & PERF_SAMPLE_TIME))
2461                 return -1;
2462
2463         if (event->header.type != PERF_RECORD_SAMPLE) {
2464                 struct perf_sample data = {
2465                         .time = -1ULL,
2466                 };
2467
2468                 if (!evsel->attr.sample_id_all)
2469                         return -1;
2470                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2471                         return -1;
2472
2473                 *timestamp = data.time;
2474                 return 0;
2475         }
2476
2477         array = event->sample.array;
2478
2479         if (perf_event__check_size(event, evsel->sample_size))
2480                 return -EFAULT;
2481
2482         if (type & PERF_SAMPLE_IDENTIFIER)
2483                 array++;
2484
2485         if (type & PERF_SAMPLE_IP)
2486                 array++;
2487
2488         if (type & PERF_SAMPLE_TID)
2489                 array++;
2490
2491         if (type & PERF_SAMPLE_TIME)
2492                 *timestamp = *array;
2493
2494         return 0;
2495 }
2496
2497 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2498                                      u64 read_format)
2499 {
2500         size_t sz, result = sizeof(struct sample_event);
2501
2502         if (type & PERF_SAMPLE_IDENTIFIER)
2503                 result += sizeof(u64);
2504
2505         if (type & PERF_SAMPLE_IP)
2506                 result += sizeof(u64);
2507
2508         if (type & PERF_SAMPLE_TID)
2509                 result += sizeof(u64);
2510
2511         if (type & PERF_SAMPLE_TIME)
2512                 result += sizeof(u64);
2513
2514         if (type & PERF_SAMPLE_ADDR)
2515                 result += sizeof(u64);
2516
2517         if (type & PERF_SAMPLE_ID)
2518                 result += sizeof(u64);
2519
2520         if (type & PERF_SAMPLE_STREAM_ID)
2521                 result += sizeof(u64);
2522
2523         if (type & PERF_SAMPLE_CPU)
2524                 result += sizeof(u64);
2525
2526         if (type & PERF_SAMPLE_PERIOD)
2527                 result += sizeof(u64);
2528
2529         if (type & PERF_SAMPLE_READ) {
2530                 result += sizeof(u64);
2531                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2532                         result += sizeof(u64);
2533                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2534                         result += sizeof(u64);
2535                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2536                 if (read_format & PERF_FORMAT_GROUP) {
2537                         sz = sample->read.group.nr *
2538                              sizeof(struct sample_read_value);
2539                         result += sz;
2540                 } else {
2541                         result += sizeof(u64);
2542                 }
2543         }
2544
2545         if (type & PERF_SAMPLE_CALLCHAIN) {
2546                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2547                 result += sz;
2548         }
2549
2550         if (type & PERF_SAMPLE_RAW) {
2551                 result += sizeof(u32);
2552                 result += sample->raw_size;
2553         }
2554
2555         if (type & PERF_SAMPLE_BRANCH_STACK) {
2556                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2557                 sz += sizeof(u64);
2558                 result += sz;
2559         }
2560
2561         if (type & PERF_SAMPLE_REGS_USER) {
2562                 if (sample->user_regs.abi) {
2563                         result += sizeof(u64);
2564                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2565                         result += sz;
2566                 } else {
2567                         result += sizeof(u64);
2568                 }
2569         }
2570
2571         if (type & PERF_SAMPLE_STACK_USER) {
2572                 sz = sample->user_stack.size;
2573                 result += sizeof(u64);
2574                 if (sz) {
2575                         result += sz;
2576                         result += sizeof(u64);
2577                 }
2578         }
2579
2580         if (type & PERF_SAMPLE_WEIGHT)
2581                 result += sizeof(u64);
2582
2583         if (type & PERF_SAMPLE_DATA_SRC)
2584                 result += sizeof(u64);
2585
2586         if (type & PERF_SAMPLE_TRANSACTION)
2587                 result += sizeof(u64);
2588
2589         if (type & PERF_SAMPLE_REGS_INTR) {
2590                 if (sample->intr_regs.abi) {
2591                         result += sizeof(u64);
2592                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2593                         result += sz;
2594                 } else {
2595                         result += sizeof(u64);
2596                 }
2597         }
2598
2599         if (type & PERF_SAMPLE_PHYS_ADDR)
2600                 result += sizeof(u64);
2601
2602         return result;
2603 }
2604
2605 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2606                                   u64 read_format,
2607                                   const struct perf_sample *sample)
2608 {
2609         u64 *array;
2610         size_t sz;
2611         /*
2612          * used for cross-endian analysis. See git commit 65014ab3
2613          * for why this goofiness is needed.
2614          */
2615         union u64_swap u;
2616
2617         array = event->sample.array;
2618
2619         if (type & PERF_SAMPLE_IDENTIFIER) {
2620                 *array = sample->id;
2621                 array++;
2622         }
2623
2624         if (type & PERF_SAMPLE_IP) {
2625                 *array = sample->ip;
2626                 array++;
2627         }
2628
2629         if (type & PERF_SAMPLE_TID) {
2630                 u.val32[0] = sample->pid;
2631                 u.val32[1] = sample->tid;
2632                 *array = u.val64;
2633                 array++;
2634         }
2635
2636         if (type & PERF_SAMPLE_TIME) {
2637                 *array = sample->time;
2638                 array++;
2639         }
2640
2641         if (type & PERF_SAMPLE_ADDR) {
2642                 *array = sample->addr;
2643                 array++;
2644         }
2645
2646         if (type & PERF_SAMPLE_ID) {
2647                 *array = sample->id;
2648                 array++;
2649         }
2650
2651         if (type & PERF_SAMPLE_STREAM_ID) {
2652                 *array = sample->stream_id;
2653                 array++;
2654         }
2655
2656         if (type & PERF_SAMPLE_CPU) {
2657                 u.val32[0] = sample->cpu;
2658                 u.val32[1] = 0;
2659                 *array = u.val64;
2660                 array++;
2661         }
2662
2663         if (type & PERF_SAMPLE_PERIOD) {
2664                 *array = sample->period;
2665                 array++;
2666         }
2667
2668         if (type & PERF_SAMPLE_READ) {
2669                 if (read_format & PERF_FORMAT_GROUP)
2670                         *array = sample->read.group.nr;
2671                 else
2672                         *array = sample->read.one.value;
2673                 array++;
2674
2675                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2676                         *array = sample->read.time_enabled;
2677                         array++;
2678                 }
2679
2680                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2681                         *array = sample->read.time_running;
2682                         array++;
2683                 }
2684
2685                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2686                 if (read_format & PERF_FORMAT_GROUP) {
2687                         sz = sample->read.group.nr *
2688                              sizeof(struct sample_read_value);
2689                         memcpy(array, sample->read.group.values, sz);
2690                         array = (void *)array + sz;
2691                 } else {
2692                         *array = sample->read.one.id;
2693                         array++;
2694                 }
2695         }
2696
2697         if (type & PERF_SAMPLE_CALLCHAIN) {
2698                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2699                 memcpy(array, sample->callchain, sz);
2700                 array = (void *)array + sz;
2701         }
2702
2703         if (type & PERF_SAMPLE_RAW) {
2704                 u.val32[0] = sample->raw_size;
2705                 *array = u.val64;
2706                 array = (void *)array + sizeof(u32);
2707
2708                 memcpy(array, sample->raw_data, sample->raw_size);
2709                 array = (void *)array + sample->raw_size;
2710         }
2711
2712         if (type & PERF_SAMPLE_BRANCH_STACK) {
2713                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2714                 sz += sizeof(u64);
2715                 memcpy(array, sample->branch_stack, sz);
2716                 array = (void *)array + sz;
2717         }
2718
2719         if (type & PERF_SAMPLE_REGS_USER) {
2720                 if (sample->user_regs.abi) {
2721                         *array++ = sample->user_regs.abi;
2722                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2723                         memcpy(array, sample->user_regs.regs, sz);
2724                         array = (void *)array + sz;
2725                 } else {
2726                         *array++ = 0;
2727                 }
2728         }
2729
2730         if (type & PERF_SAMPLE_STACK_USER) {
2731                 sz = sample->user_stack.size;
2732                 *array++ = sz;
2733                 if (sz) {
2734                         memcpy(array, sample->user_stack.data, sz);
2735                         array = (void *)array + sz;
2736                         *array++ = sz;
2737                 }
2738         }
2739
2740         if (type & PERF_SAMPLE_WEIGHT) {
2741                 *array = sample->weight;
2742                 array++;
2743         }
2744
2745         if (type & PERF_SAMPLE_DATA_SRC) {
2746                 *array = sample->data_src;
2747                 array++;
2748         }
2749
2750         if (type & PERF_SAMPLE_TRANSACTION) {
2751                 *array = sample->transaction;
2752                 array++;
2753         }
2754
2755         if (type & PERF_SAMPLE_REGS_INTR) {
2756                 if (sample->intr_regs.abi) {
2757                         *array++ = sample->intr_regs.abi;
2758                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2759                         memcpy(array, sample->intr_regs.regs, sz);
2760                         array = (void *)array + sz;
2761                 } else {
2762                         *array++ = 0;
2763                 }
2764         }
2765
2766         if (type & PERF_SAMPLE_PHYS_ADDR) {
2767                 *array = sample->phys_addr;
2768                 array++;
2769         }
2770
2771         return 0;
2772 }
2773
2774 struct tep_format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2775 {
2776         return tep_find_field(evsel->tp_format, name);
2777 }
2778
2779 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2780                          const char *name)
2781 {
2782         struct tep_format_field *field = perf_evsel__field(evsel, name);
2783         int offset;
2784
2785         if (!field)
2786                 return NULL;
2787
2788         offset = field->offset;
2789
2790         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2791                 offset = *(int *)(sample->raw_data + field->offset);
2792                 offset &= 0xffff;
2793         }
2794
2795         return sample->raw_data + offset;
2796 }
2797
2798 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2799                          bool needs_swap)
2800 {
2801         u64 value;
2802         void *ptr = sample->raw_data + field->offset;
2803
2804         switch (field->size) {
2805         case 1:
2806                 return *(u8 *)ptr;
2807         case 2:
2808                 value = *(u16 *)ptr;
2809                 break;
2810         case 4:
2811                 value = *(u32 *)ptr;
2812                 break;
2813         case 8:
2814                 memcpy(&value, ptr, sizeof(u64));
2815                 break;
2816         default:
2817                 return 0;
2818         }
2819
2820         if (!needs_swap)
2821                 return value;
2822
2823         switch (field->size) {
2824         case 2:
2825                 return bswap_16(value);
2826         case 4:
2827                 return bswap_32(value);
2828         case 8:
2829                 return bswap_64(value);
2830         default:
2831                 return 0;
2832         }
2833
2834         return 0;
2835 }
2836
2837 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2838                        const char *name)
2839 {
2840         struct tep_format_field *field = perf_evsel__field(evsel, name);
2841
2842         if (!field)
2843                 return 0;
2844
2845         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2846 }
2847
2848 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2849                           char *msg, size_t msgsize)
2850 {
2851         int paranoid;
2852
2853         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2854             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2855             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2856                 /*
2857                  * If it's cycles then fall back to hrtimer based
2858                  * cpu-clock-tick sw counter, which is always available even if
2859                  * no PMU support.
2860                  *
2861                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2862                  * b0a873e).
2863                  */
2864                 scnprintf(msg, msgsize, "%s",
2865 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2866
2867                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2868                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2869
2870                 zfree(&evsel->name);
2871                 return true;
2872         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2873                    (paranoid = perf_event_paranoid()) > 1) {
2874                 const char *name = perf_evsel__name(evsel);
2875                 char *new_name;
2876                 const char *sep = ":";
2877
2878                 /* Is there already the separator in the name. */
2879                 if (strchr(name, '/') ||
2880                     strchr(name, ':'))
2881                         sep = "";
2882
2883                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2884                         return false;
2885
2886                 if (evsel->name)
2887                         free(evsel->name);
2888                 evsel->name = new_name;
2889                 scnprintf(msg, msgsize,
2890 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2891                 evsel->attr.exclude_kernel = 1;
2892
2893                 return true;
2894         }
2895
2896         return false;
2897 }
2898
2899 static bool find_process(const char *name)
2900 {
2901         size_t len = strlen(name);
2902         DIR *dir;
2903         struct dirent *d;
2904         int ret = -1;
2905
2906         dir = opendir(procfs__mountpoint());
2907         if (!dir)
2908                 return false;
2909
2910         /* Walk through the directory. */
2911         while (ret && (d = readdir(dir)) != NULL) {
2912                 char path[PATH_MAX];
2913                 char *data;
2914                 size_t size;
2915
2916                 if ((d->d_type != DT_DIR) ||
2917                      !strcmp(".", d->d_name) ||
2918                      !strcmp("..", d->d_name))
2919                         continue;
2920
2921                 scnprintf(path, sizeof(path), "%s/%s/comm",
2922                           procfs__mountpoint(), d->d_name);
2923
2924                 if (filename__read_str(path, &data, &size))
2925                         continue;
2926
2927                 ret = strncmp(name, data, len);
2928                 free(data);
2929         }
2930
2931         closedir(dir);
2932         return ret ? false : true;
2933 }
2934
2935 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2936                               int err, char *msg, size_t size)
2937 {
2938         char sbuf[STRERR_BUFSIZE];
2939         int printed = 0;
2940
2941         switch (err) {
2942         case EPERM:
2943         case EACCES:
2944                 if (err == EPERM)
2945                         printed = scnprintf(msg, size,
2946                                 "No permission to enable %s event.\n\n",
2947                                 perf_evsel__name(evsel));
2948
2949                 return scnprintf(msg + printed, size - printed,
2950                  "You may not have permission to collect %sstats.\n\n"
2951                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2952                  "which controls use of the performance events system by\n"
2953                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2954                  "The current value is %d:\n\n"
2955                  "  -1: Allow use of (almost) all events by all users\n"
2956                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2957                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2958                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2959                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2960                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2961                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2962                  "      kernel.perf_event_paranoid = -1\n" ,
2963                                  target->system_wide ? "system-wide " : "",
2964                                  perf_event_paranoid());
2965         case ENOENT:
2966                 return scnprintf(msg, size, "The %s event is not supported.",
2967                                  perf_evsel__name(evsel));
2968         case EMFILE:
2969                 return scnprintf(msg, size, "%s",
2970                          "Too many events are opened.\n"
2971                          "Probably the maximum number of open file descriptors has been reached.\n"
2972                          "Hint: Try again after reducing the number of events.\n"
2973                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2974         case ENOMEM:
2975                 if (evsel__has_callchain(evsel) &&
2976                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2977                         return scnprintf(msg, size,
2978                                          "Not enough memory to setup event with callchain.\n"
2979                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2980                                          "Hint: Current value: %d", sysctl__max_stack());
2981                 break;
2982         case ENODEV:
2983                 if (target->cpu_list)
2984                         return scnprintf(msg, size, "%s",
2985          "No such device - did you specify an out-of-range profile CPU?");
2986                 break;
2987         case EOPNOTSUPP:
2988                 if (evsel->attr.sample_period != 0)
2989                         return scnprintf(msg, size,
2990         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2991                                          perf_evsel__name(evsel));
2992                 if (evsel->attr.precise_ip)
2993                         return scnprintf(msg, size, "%s",
2994         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2995 #if defined(__i386__) || defined(__x86_64__)
2996                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2997                         return scnprintf(msg, size, "%s",
2998         "No hardware sampling interrupt available.\n");
2999 #endif
3000                 break;
3001         case EBUSY:
3002                 if (find_process("oprofiled"))
3003                         return scnprintf(msg, size,
3004         "The PMU counters are busy/taken by another profiler.\n"
3005         "We found oprofile daemon running, please stop it and try again.");
3006                 break;
3007         case EINVAL:
3008                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
3009                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3010                 if (perf_missing_features.clockid)
3011                         return scnprintf(msg, size, "clockid feature not supported.");
3012                 if (perf_missing_features.clockid_wrong)
3013                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3014                 break;
3015         default:
3016                 break;
3017         }
3018
3019         return scnprintf(msg, size,
3020         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3021         "/bin/dmesg | grep -i perf may provide additional information.\n",
3022                          err, str_error_r(err, sbuf, sizeof(sbuf)),
3023                          perf_evsel__name(evsel));
3024 }
3025
3026 struct perf_env *perf_evsel__env(struct perf_evsel *evsel)
3027 {
3028         if (evsel && evsel->evlist)
3029                 return evsel->evlist->env;
3030         return NULL;
3031 }
3032
3033 static int store_evsel_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3034 {
3035         int cpu, thread;
3036
3037         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
3038                 for (thread = 0; thread < xyarray__max_y(evsel->fd);
3039                      thread++) {
3040                         int fd = FD(evsel, cpu, thread);
3041
3042                         if (perf_evlist__id_add_fd(evlist, evsel,
3043                                                    cpu, thread, fd) < 0)
3044                                 return -1;
3045                 }
3046         }
3047
3048         return 0;
3049 }
3050
3051 int perf_evsel__store_ids(struct perf_evsel *evsel, struct perf_evlist *evlist)
3052 {
3053         struct cpu_map *cpus = evsel->cpus;
3054         struct thread_map *threads = evsel->threads;
3055
3056         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
3057                 return -ENOMEM;
3058
3059         return store_evsel_ids(evsel, evlist);
3060 }