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