tools/perf/util/evsel.c

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