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 <linux/bitops.h>
  12 #include <api/fs/tracing_path.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <linux/err.h>
  17 #include <sys/resource.h>
  18 #include "asm/bug.h"
  19 #include "callchain.h"
  20 #include "cgroup.h"
  21 #include "evsel.h"
  22 #include "evlist.h"
  23 #include "util.h"
  24 #include "cpumap.h"
  25 #include "thread_map.h"
  26 #include "target.h"
  27 #include "perf_regs.h"
  28 #include "debug.h"
  29 #include "trace-event.h"
  30 #include "stat.h"
  31
  32 static struct {
  33         bool sample_id_all;
  34         bool exclude_guest;
  35         bool mmap2;
  36         bool cloexec;
  37         bool clockid;
  38         bool clockid_wrong;
  39         bool lbr_flags;
  40 } perf_missing_features;
  41
  42 static clockid_t clockid;
  43
  44 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  45 {
  46         return 0;
  47 }
  48
  49 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  50 {
  51 }
  52
  53 static struct {
  54         size_t  size;
  55         int     (*init)(struct perf_evsel *evsel);
  56         void    (*fini)(struct perf_evsel *evsel);
  57 } perf_evsel__object = {
  58         .size = sizeof(struct perf_evsel),
  59         .init = perf_evsel__no_extra_init,
  60         .fini = perf_evsel__no_extra_fini,
  61 };
  62
  63 int perf_evsel__object_config(size_t object_size,
  64                               int (*init)(struct perf_evsel *evsel),
  65                               void (*fini)(struct perf_evsel *evsel))
  66 {
  67
  68         if (object_size == 0)
  69                 goto set_methods;
  70
  71         if (perf_evsel__object.size > object_size)
  72                 return -EINVAL;
  73
  74         perf_evsel__object.size = object_size;
  75
  76 set_methods:
  77         if (init != NULL)
  78                 perf_evsel__object.init = init;
  79
  80         if (fini != NULL)
  81                 perf_evsel__object.fini = fini;
  82
  83         return 0;
  84 }
  85
  86 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  87
  88 int __perf_evsel__sample_size(u64 sample_type)
  89 {
  90         u64 mask = sample_type & PERF_SAMPLE_MASK;
  91         int size = 0;
  92         int i;
  93
  94         for (i = 0; i < 64; i++) {
  95                 if (mask & (1ULL << i))
  96                         size++;
  97         }
  98
  99         size *= sizeof(u64);
 100
 101         return size;
 102 }
 103
 104 /**
 105  * __perf_evsel__calc_id_pos - calculate id_pos.
 106  * @sample_type: sample type
 107  *
 108  * This function returns the position of the event id (PERF_SAMPLE_ID or
 109  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 110  * sample_event.
 111  */
 112 static int __perf_evsel__calc_id_pos(u64 sample_type)
 113 {
 114         int idx = 0;
 115
 116         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 117                 return 0;
 118
 119         if (!(sample_type & PERF_SAMPLE_ID))
 120                 return -1;
 121
 122         if (sample_type & PERF_SAMPLE_IP)
 123                 idx += 1;
 124
 125         if (sample_type & PERF_SAMPLE_TID)
 126                 idx += 1;
 127
 128         if (sample_type & PERF_SAMPLE_TIME)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_ADDR)
 132                 idx += 1;
 133
 134         return idx;
 135 }
 136
 137 /**
 138  * __perf_evsel__calc_is_pos - calculate is_pos.
 139  * @sample_type: sample type
 140  *
 141  * This function returns the position (counting backwards) of the event id
 142  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 143  * sample_id_all is used there is an id sample appended to non-sample events.
 144  */
 145 static int __perf_evsel__calc_is_pos(u64 sample_type)
 146 {
 147         int idx = 1;
 148
 149         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 150                 return 1;
 151
 152         if (!(sample_type & PERF_SAMPLE_ID))
 153                 return -1;
 154
 155         if (sample_type & PERF_SAMPLE_CPU)
 156                 idx += 1;
 157
 158         if (sample_type & PERF_SAMPLE_STREAM_ID)
 159                 idx += 1;
 160
 161         return idx;
 162 }
 163
 164 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 165 {
 166         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 167         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 168 }
 169
 170 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 171                                   enum perf_event_sample_format bit)
 172 {
 173         if (!(evsel->attr.sample_type & bit)) {
 174                 evsel->attr.sample_type |= bit;
 175                 evsel->sample_size += sizeof(u64);
 176                 perf_evsel__calc_id_pos(evsel);
 177         }
 178 }
 179
 180 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 181                                     enum perf_event_sample_format bit)
 182 {
 183         if (evsel->attr.sample_type & bit) {
 184                 evsel->attr.sample_type &= ~bit;
 185                 evsel->sample_size -= sizeof(u64);
 186                 perf_evsel__calc_id_pos(evsel);
 187         }
 188 }
 189
 190 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 191                                bool can_sample_identifier)
 192 {
 193         if (can_sample_identifier) {
 194                 perf_evsel__reset_sample_bit(evsel, ID);
 195                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 196         } else {
 197                 perf_evsel__set_sample_bit(evsel, ID);
 198         }
 199         evsel->attr.read_format |= PERF_FORMAT_ID;
 200 }
 201
 202 void perf_evsel__init(struct perf_evsel *evsel,
 203                       struct perf_event_attr *attr, int idx)
 204 {
 205         evsel->idx         = idx;
 206         evsel->tracking    = !idx;
 207         evsel->attr        = *attr;
 208         evsel->leader      = evsel;
 209         evsel->unit        = "";
 210         evsel->scale       = 1.0;
 211         evsel->evlist      = NULL;
 212         evsel->bpf_fd      = -1;
 213         INIT_LIST_HEAD(&evsel->node);
 214         INIT_LIST_HEAD(&evsel->config_terms);
 215         perf_evsel__object.init(evsel);
 216         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 217         perf_evsel__calc_id_pos(evsel);
 218         evsel->cmdline_group_boundary = false;
 219 }
 220
 221 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 222 {
 223         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 224
 225         if (evsel != NULL)
 226                 perf_evsel__init(evsel, attr, idx);
 227
 228         if (perf_evsel__is_bpf_output(evsel)) {
 229                 evsel->attr.sample_type |= PERF_SAMPLE_RAW;
 230                 evsel->attr.sample_period = 1;
 231         }
 232
 233         return evsel;
 234 }
 235
 236 /*
 237  * Returns pointer with encoded error via <linux/err.h> interface.
 238  */
 239 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 240 {
 241         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 242         int err = -ENOMEM;
 243
 244         if (evsel == NULL) {
 245                 goto out_err;
 246         } else {
 247                 struct perf_event_attr attr = {
 248                         .type          = PERF_TYPE_TRACEPOINT,
 249                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 250                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 251                 };
 252
 253                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 254                         goto out_free;
 255
 256                 evsel->tp_format = trace_event__tp_format(sys, name);
 257                 if (IS_ERR(evsel->tp_format)) {
 258                         err = PTR_ERR(evsel->tp_format);
 259                         goto out_free;
 260                 }
 261
 262                 event_attr_init(&attr);
 263                 attr.config = evsel->tp_format->id;
 264                 attr.sample_period = 1;
 265                 perf_evsel__init(evsel, &attr, idx);
 266         }
 267
 268         return evsel;
 269
 270 out_free:
 271         zfree(&evsel->name);
 272         free(evsel);
 273 out_err:
 274         return ERR_PTR(err);
 275 }
 276
 277 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 278         "cycles",
 279         "instructions",
 280         "cache-references",
 281         "cache-misses",
 282         "branches",
 283         "branch-misses",
 284         "bus-cycles",
 285         "stalled-cycles-frontend",
 286         "stalled-cycles-backend",
 287         "ref-cycles",
 288 };
 289
 290 static const char *__perf_evsel__hw_name(u64 config)
 291 {
 292         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 293                 return perf_evsel__hw_names[config];
 294
 295         return "unknown-hardware";
 296 }
 297
 298 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 299 {
 300         int colon = 0, r = 0;
 301         struct perf_event_attr *attr = &evsel->attr;
 302         bool exclude_guest_default = false;
 303
 304 #define MOD_PRINT(context, mod) do {                                    \
 305                 if (!attr->exclude_##context) {                         \
 306                         if (!colon) colon = ++r;                        \
 307                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 308                 } } while(0)
 309
 310         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 311                 MOD_PRINT(kernel, 'k');
 312                 MOD_PRINT(user, 'u');
 313                 MOD_PRINT(hv, 'h');
 314                 exclude_guest_default = true;
 315         }
 316
 317         if (attr->precise_ip) {
 318                 if (!colon)
 319                         colon = ++r;
 320                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 321                 exclude_guest_default = true;
 322         }
 323
 324         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 325                 MOD_PRINT(host, 'H');
 326                 MOD_PRINT(guest, 'G');
 327         }
 328 #undef MOD_PRINT
 329         if (colon)
 330                 bf[colon - 1] = ':';
 331         return r;
 332 }
 333
 334 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 335 {
 336         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 337         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 338 }
 339
 340 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 341         "cpu-clock",
 342         "task-clock",
 343         "page-faults",
 344         "context-switches",
 345         "cpu-migrations",
 346         "minor-faults",
 347         "major-faults",
 348         "alignment-faults",
 349         "emulation-faults",
 350         "dummy",
 351 };
 352
 353 static const char *__perf_evsel__sw_name(u64 config)
 354 {
 355         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 356                 return perf_evsel__sw_names[config];
 357         return "unknown-software";
 358 }
 359
 360 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 361 {
 362         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 363         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 364 }
 365
 366 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 367 {
 368         int r;
 369
 370         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 371
 372         if (type & HW_BREAKPOINT_R)
 373                 r += scnprintf(bf + r, size - r, "r");
 374
 375         if (type & HW_BREAKPOINT_W)
 376                 r += scnprintf(bf + r, size - r, "w");
 377
 378         if (type & HW_BREAKPOINT_X)
 379                 r += scnprintf(bf + r, size - r, "x");
 380
 381         return r;
 382 }
 383
 384 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 385 {
 386         struct perf_event_attr *attr = &evsel->attr;
 387         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 388         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 389 }
 390
 391 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 392                                 [PERF_EVSEL__MAX_ALIASES] = {
 393  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 394  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 395  { "LLC",       "L2",                                                   },
 396  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 397  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 398  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 399  { "node",                                                              },
 400 };
 401
 402 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 403                                    [PERF_EVSEL__MAX_ALIASES] = {
 404  { "load",      "loads",        "read",                                 },
 405  { "store",     "stores",       "write",                                },
 406  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 407 };
 408
 409 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 410                                        [PERF_EVSEL__MAX_ALIASES] = {
 411  { "refs",      "Reference",    "ops",          "access",               },
 412  { "misses",    "miss",                                                 },
 413 };
 414
 415 #define C(x)            PERF_COUNT_HW_CACHE_##x
 416 #define CACHE_READ      (1 << C(OP_READ))
 417 #define CACHE_WRITE     (1 << C(OP_WRITE))
 418 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 419 #define COP(x)          (1 << x)
 420
 421 /*
 422  * cache operartion stat
 423  * L1I : Read and prefetch only
 424  * ITLB and BPU : Read-only
 425  */
 426 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 427  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 428  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 429  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 430  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 431  [C(ITLB)]      = (CACHE_READ),
 432  [C(BPU)]       = (CACHE_READ),
 433  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 434 };
 435
 436 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 437 {
 438         if (perf_evsel__hw_cache_stat[type] & COP(op))
 439                 return true;    /* valid */
 440         else
 441                 return false;   /* invalid */
 442 }
 443
 444 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 445                                             char *bf, size_t size)
 446 {
 447         if (result) {
 448                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 449                                  perf_evsel__hw_cache_op[op][0],
 450                                  perf_evsel__hw_cache_result[result][0]);
 451         }
 452
 453         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 454                          perf_evsel__hw_cache_op[op][1]);
 455 }
 456
 457 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 458 {
 459         u8 op, result, type = (config >>  0) & 0xff;
 460         const char *err = "unknown-ext-hardware-cache-type";
 461
 462         if (type > PERF_COUNT_HW_CACHE_MAX)
 463                 goto out_err;
 464
 465         op = (config >>  8) & 0xff;
 466         err = "unknown-ext-hardware-cache-op";
 467         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 468                 goto out_err;
 469
 470         result = (config >> 16) & 0xff;
 471         err = "unknown-ext-hardware-cache-result";
 472         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 473                 goto out_err;
 474
 475         err = "invalid-cache";
 476         if (!perf_evsel__is_cache_op_valid(type, op))
 477                 goto out_err;
 478
 479         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 480 out_err:
 481         return scnprintf(bf, size, "%s", err);
 482 }
 483
 484 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 485 {
 486         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 487         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 488 }
 489
 490 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 491 {
 492         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 493         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 494 }
 495
 496 const char *perf_evsel__name(struct perf_evsel *evsel)
 497 {
 498         char bf[128];
 499
 500         if (evsel->name)
 501                 return evsel->name;
 502
 503         switch (evsel->attr.type) {
 504         case PERF_TYPE_RAW:
 505                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 506                 break;
 507
 508         case PERF_TYPE_HARDWARE:
 509                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 510                 break;
 511
 512         case PERF_TYPE_HW_CACHE:
 513                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 514                 break;
 515
 516         case PERF_TYPE_SOFTWARE:
 517                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 518                 break;
 519
 520         case PERF_TYPE_TRACEPOINT:
 521                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 522                 break;
 523
 524         case PERF_TYPE_BREAKPOINT:
 525                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 526                 break;
 527
 528         default:
 529                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 530                           evsel->attr.type);
 531                 break;
 532         }
 533
 534         evsel->name = strdup(bf);
 535
 536         return evsel->name ?: "unknown";
 537 }
 538
 539 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 540 {
 541         return evsel->group_name ?: "anon group";
 542 }
 543
 544 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 545 {
 546         int ret;
 547         struct perf_evsel *pos;
 548         const char *group_name = perf_evsel__group_name(evsel);
 549
 550         ret = scnprintf(buf, size, "%s", group_name);
 551
 552         ret += scnprintf(buf + ret, size - ret, " { %s",
 553                          perf_evsel__name(evsel));
 554
 555         for_each_group_member(pos, evsel)
 556                 ret += scnprintf(buf + ret, size - ret, ", %s",
 557                                  perf_evsel__name(pos));
 558
 559         ret += scnprintf(buf + ret, size - ret, " }");
 560
 561         return ret;
 562 }
 563
 564 static void
 565 perf_evsel__config_callgraph(struct perf_evsel *evsel,
 566                              struct record_opts *opts,
 567                              struct callchain_param *param)
 568 {
 569         bool function = perf_evsel__is_function_event(evsel);
 570         struct perf_event_attr *attr = &evsel->attr;
 571
 572         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 573
 574         if (param->record_mode == CALLCHAIN_LBR) {
 575                 if (!opts->branch_stack) {
 576                         if (attr->exclude_user) {
 577                                 pr_warning("LBR callstack option is only available "
 578                                            "to get user callchain information. "
 579                                            "Falling back to framepointers.\n");
 580                         } else {
 581                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 582                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 583                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 584                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 585                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 586                         }
 587                 } else
 588                          pr_warning("Cannot use LBR callstack with branch stack. "
 589                                     "Falling back to framepointers.\n");
 590         }
 591
 592         if (param->record_mode == CALLCHAIN_DWARF) {
 593                 if (!function) {
 594                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 595                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 596                         attr->sample_regs_user = PERF_REGS_MASK;
 597                         attr->sample_stack_user = param->dump_size;
 598                         attr->exclude_callchain_user = 1;
 599                 } else {
 600                         pr_info("Cannot use DWARF unwind for function trace event,"
 601                                 " falling back to framepointers.\n");
 602                 }
 603         }
 604
 605         if (function) {
 606                 pr_info("Disabling user space callchains for function trace event.\n");
 607                 attr->exclude_callchain_user = 1;
 608         }
 609 }
 610
 611 static void
 612 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 613                             struct callchain_param *param)
 614 {
 615         struct perf_event_attr *attr = &evsel->attr;
 616
 617         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 618         if (param->record_mode == CALLCHAIN_LBR) {
 619                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 620                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 621                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 622         }
 623         if (param->record_mode == CALLCHAIN_DWARF) {
 624                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 625                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 626         }
 627 }
 628
 629 static void apply_config_terms(struct perf_evsel *evsel,
 630                                struct record_opts *opts)
 631 {
 632         struct perf_evsel_config_term *term;
 633         struct list_head *config_terms = &evsel->config_terms;
 634         struct perf_event_attr *attr = &evsel->attr;
 635         struct callchain_param param;
 636         u32 dump_size = 0;
 637         char *callgraph_buf = NULL;
 638
 639         /* callgraph default */
 640         param.record_mode = callchain_param.record_mode;
 641
 642         list_for_each_entry(term, config_terms, list) {
 643                 switch (term->type) {
 644                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 645                         attr->sample_period = term->val.period;
 646                         attr->freq = 0;
 647                         break;
 648                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 649                         attr->sample_freq = term->val.freq;
 650                         attr->freq = 1;
 651                         break;
 652                 case PERF_EVSEL__CONFIG_TERM_TIME:
 653                         if (term->val.time)
 654                                 perf_evsel__set_sample_bit(evsel, TIME);
 655                         else
 656                                 perf_evsel__reset_sample_bit(evsel, TIME);
 657                         break;
 658                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 659                         callgraph_buf = term->val.callgraph;
 660                         break;
 661                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 662                         dump_size = term->val.stack_user;
 663                         break;
 664                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 665                         /*
 666                          * attr->inherit should has already been set by
 667                          * perf_evsel__config. If user explicitly set
 668                          * inherit using config terms, override global
 669                          * opt->no_inherit setting.
 670                          */
 671                         attr->inherit = term->val.inherit ? 1 : 0;
 672                         break;
 673                 default:
 674                         break;
 675                 }
 676         }
 677
 678         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 679         if ((callgraph_buf != NULL) || (dump_size > 0)) {
 680
 681                 /* parse callgraph parameters */
 682                 if (callgraph_buf != NULL) {
 683                         if (!strcmp(callgraph_buf, "no")) {
 684                                 param.enabled = false;
 685                                 param.record_mode = CALLCHAIN_NONE;
 686                         } else {
 687                                 param.enabled = true;
 688                                 if (parse_callchain_record(callgraph_buf, &param)) {
 689                                         pr_err("per-event callgraph setting for %s failed. "
 690                                                "Apply callgraph global setting for it\n",
 691                                                evsel->name);
 692                                         return;
 693                                 }
 694                         }
 695                 }
 696                 if (dump_size > 0) {
 697                         dump_size = round_up(dump_size, sizeof(u64));
 698                         param.dump_size = dump_size;
 699                 }
 700
 701                 /* If global callgraph set, clear it */
 702                 if (callchain_param.enabled)
 703                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 704
 705                 /* set perf-event callgraph */
 706                 if (param.enabled)
 707                         perf_evsel__config_callgraph(evsel, opts, &param);
 708         }
 709 }
 710
 711 /*
 712  * The enable_on_exec/disabled value strategy:
 713  *
 714  *  1) For any type of traced program:
 715  *    - all independent events and group leaders are disabled
 716  *    - all group members are enabled
 717  *
 718  *     Group members are ruled by group leaders. They need to
 719  *     be enabled, because the group scheduling relies on that.
 720  *
 721  *  2) For traced programs executed by perf:
 722  *     - all independent events and group leaders have
 723  *       enable_on_exec set
 724  *     - we don't specifically enable or disable any event during
 725  *       the record command
 726  *
 727  *     Independent events and group leaders are initially disabled
 728  *     and get enabled by exec. Group members are ruled by group
 729  *     leaders as stated in 1).
 730  *
 731  *  3) For traced programs attached by perf (pid/tid):
 732  *     - we specifically enable or disable all events during
 733  *       the record command
 734  *
 735  *     When attaching events to already running traced we
 736  *     enable/disable events specifically, as there's no
 737  *     initial traced exec call.
 738  */
 739 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts)
 740 {
 741         struct perf_evsel *leader = evsel->leader;
 742         struct perf_event_attr *attr = &evsel->attr;
 743         int track = evsel->tracking;
 744         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 745
 746         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 747         attr->inherit       = !opts->no_inherit;
 748
 749         perf_evsel__set_sample_bit(evsel, IP);
 750         perf_evsel__set_sample_bit(evsel, TID);
 751
 752         if (evsel->sample_read) {
 753                 perf_evsel__set_sample_bit(evsel, READ);
 754
 755                 /*
 756                  * We need ID even in case of single event, because
 757                  * PERF_SAMPLE_READ process ID specific data.
 758                  */
 759                 perf_evsel__set_sample_id(evsel, false);
 760
 761                 /*
 762                  * Apply group format only if we belong to group
 763                  * with more than one members.
 764                  */
 765                 if (leader->nr_members > 1) {
 766                         attr->read_format |= PERF_FORMAT_GROUP;
 767                         attr->inherit = 0;
 768                 }
 769         }
 770
 771         /*
 772          * We default some events to have a default interval. But keep
 773          * it a weak assumption overridable by the user.
 774          */
 775         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 776                                      opts->user_interval != ULLONG_MAX)) {
 777                 if (opts->freq) {
 778                         perf_evsel__set_sample_bit(evsel, PERIOD);
 779                         attr->freq              = 1;
 780                         attr->sample_freq       = opts->freq;
 781                 } else {
 782                         attr->sample_period = opts->default_interval;
 783                 }
 784         }
 785
 786         /*
 787          * Disable sampling for all group members other
 788          * than leader in case leader 'leads' the sampling.
 789          */
 790         if ((leader != evsel) && leader->sample_read) {
 791                 attr->sample_freq   = 0;
 792                 attr->sample_period = 0;
 793         }
 794
 795         if (opts->no_samples)
 796                 attr->sample_freq = 0;
 797
 798         if (opts->inherit_stat)
 799                 attr->inherit_stat = 1;
 800
 801         if (opts->sample_address) {
 802                 perf_evsel__set_sample_bit(evsel, ADDR);
 803                 attr->mmap_data = track;
 804         }
 805
 806         /*
 807          * We don't allow user space callchains for  function trace
 808          * event, due to issues with page faults while tracing page
 809          * fault handler and its overall trickiness nature.
 810          */
 811         if (perf_evsel__is_function_event(evsel))
 812                 evsel->attr.exclude_callchain_user = 1;
 813
 814         if (callchain_param.enabled && !evsel->no_aux_samples)
 815                 perf_evsel__config_callgraph(evsel, opts, &callchain_param);
 816
 817         if (opts->sample_intr_regs) {
 818                 attr->sample_regs_intr = opts->sample_intr_regs;
 819                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 820         }
 821
 822         if (target__has_cpu(&opts->target))
 823                 perf_evsel__set_sample_bit(evsel, CPU);
 824
 825         if (opts->period)
 826                 perf_evsel__set_sample_bit(evsel, PERIOD);
 827
 828         /*
 829          * When the user explicitely disabled time don't force it here.
 830          */
 831         if (opts->sample_time &&
 832             (!perf_missing_features.sample_id_all &&
 833             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 834              opts->sample_time_set)))
 835                 perf_evsel__set_sample_bit(evsel, TIME);
 836
 837         if (opts->raw_samples && !evsel->no_aux_samples) {
 838                 perf_evsel__set_sample_bit(evsel, TIME);
 839                 perf_evsel__set_sample_bit(evsel, RAW);
 840                 perf_evsel__set_sample_bit(evsel, CPU);
 841         }
 842
 843         if (opts->sample_address)
 844                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 845
 846         if (opts->no_buffering) {
 847                 attr->watermark = 0;
 848                 attr->wakeup_events = 1;
 849         }
 850         if (opts->branch_stack && !evsel->no_aux_samples) {
 851                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 852                 attr->branch_sample_type = opts->branch_stack;
 853         }
 854
 855         if (opts->sample_weight)
 856                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 857
 858         attr->task  = track;
 859         attr->mmap  = track;
 860         attr->mmap2 = track && !perf_missing_features.mmap2;
 861         attr->comm  = track;
 862
 863         if (opts->record_switch_events)
 864                 attr->context_switch = track;
 865
 866         if (opts->sample_transaction)
 867                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 868
 869         if (opts->running_time) {
 870                 evsel->attr.read_format |=
 871                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 872                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 873         }
 874
 875         /*
 876          * XXX see the function comment above
 877          *
 878          * Disabling only independent events or group leaders,
 879          * keeping group members enabled.
 880          */
 881         if (perf_evsel__is_group_leader(evsel))
 882                 attr->disabled = 1;
 883
 884         /*
 885          * Setting enable_on_exec for independent events and
 886          * group leaders for traced executed by perf.
 887          */
 888         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 889                 !opts->initial_delay)
 890                 attr->enable_on_exec = 1;
 891
 892         if (evsel->immediate) {
 893                 attr->disabled = 0;
 894                 attr->enable_on_exec = 0;
 895         }
 896
 897         clockid = opts->clockid;
 898         if (opts->use_clockid) {
 899                 attr->use_clockid = 1;
 900                 attr->clockid = opts->clockid;
 901         }
 902
 903         if (evsel->precise_max)
 904                 perf_event_attr__set_max_precise_ip(attr);
 905
 906         if (opts->all_user) {
 907                 attr->exclude_kernel = 1;
 908                 attr->exclude_user   = 0;
 909         }
 910
 911         if (opts->all_kernel) {
 912                 attr->exclude_kernel = 0;
 913                 attr->exclude_user   = 1;
 914         }
 915
 916         /*
 917          * Apply event specific term settings,
 918          * it overloads any global configuration.
 919          */
 920         apply_config_terms(evsel, opts);
 921 }
 922
 923 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 924 {
 925         int cpu, thread;
 926
 927         if (evsel->system_wide)
 928                 nthreads = 1;
 929
 930         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 931
 932         if (evsel->fd) {
 933                 for (cpu = 0; cpu < ncpus; cpu++) {
 934                         for (thread = 0; thread < nthreads; thread++) {
 935                                 FD(evsel, cpu, thread) = -1;
 936                         }
 937                 }
 938         }
 939
 940         return evsel->fd != NULL ? 0 : -ENOMEM;
 941 }
 942
 943 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 944                           int ioc,  void *arg)
 945 {
 946         int cpu, thread;
 947
 948         if (evsel->system_wide)
 949                 nthreads = 1;
 950
 951         for (cpu = 0; cpu < ncpus; cpu++) {
 952                 for (thread = 0; thread < nthreads; thread++) {
 953                         int fd = FD(evsel, cpu, thread),
 954                             err = ioctl(fd, ioc, arg);
 955
 956                         if (err)
 957                                 return err;
 958                 }
 959         }
 960
 961         return 0;
 962 }
 963
 964 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 965                              const char *filter)
 966 {
 967         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 968                                      PERF_EVENT_IOC_SET_FILTER,
 969                                      (void *)filter);
 970 }
 971
 972 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 973 {
 974         char *new_filter = strdup(filter);
 975
 976         if (new_filter != NULL) {
 977                 free(evsel->filter);
 978                 evsel->filter = new_filter;
 979                 return 0;
 980         }
 981
 982         return -1;
 983 }
 984
 985 int perf_evsel__append_filter(struct perf_evsel *evsel,
 986                               const char *op, const char *filter)
 987 {
 988         char *new_filter;
 989
 990         if (evsel->filter == NULL)
 991                 return perf_evsel__set_filter(evsel, filter);
 992
 993         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 994                 free(evsel->filter);
 995                 evsel->filter = new_filter;
 996                 return 0;
 997         }
 998
 999         return -1;
1000 }
1001
1002 int perf_evsel__enable(struct perf_evsel *evsel)
1003 {
1004         int nthreads = thread_map__nr(evsel->threads);
1005         int ncpus = cpu_map__nr(evsel->cpus);
1006
1007         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1008                                      PERF_EVENT_IOC_ENABLE,
1009                                      0);
1010 }
1011
1012 int perf_evsel__disable(struct perf_evsel *evsel)
1013 {
1014         int nthreads = thread_map__nr(evsel->threads);
1015         int ncpus = cpu_map__nr(evsel->cpus);
1016
1017         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1018                                      PERF_EVENT_IOC_DISABLE,
1019                                      0);
1020 }
1021
1022 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1023 {
1024         if (ncpus == 0 || nthreads == 0)
1025                 return 0;
1026
1027         if (evsel->system_wide)
1028                 nthreads = 1;
1029
1030         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1031         if (evsel->sample_id == NULL)
1032                 return -ENOMEM;
1033
1034         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1035         if (evsel->id == NULL) {
1036                 xyarray__delete(evsel->sample_id);
1037                 evsel->sample_id = NULL;
1038                 return -ENOMEM;
1039         }
1040
1041         return 0;
1042 }
1043
1044 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1045 {
1046         xyarray__delete(evsel->fd);
1047         evsel->fd = NULL;
1048 }
1049
1050 static void perf_evsel__free_id(struct perf_evsel *evsel)
1051 {
1052         xyarray__delete(evsel->sample_id);
1053         evsel->sample_id = NULL;
1054         zfree(&evsel->id);
1055 }
1056
1057 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1058 {
1059         struct perf_evsel_config_term *term, *h;
1060
1061         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1062                 list_del(&term->list);
1063                 free(term);
1064         }
1065 }
1066
1067 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1068 {
1069         int cpu, thread;
1070
1071         if (evsel->system_wide)
1072                 nthreads = 1;
1073
1074         for (cpu = 0; cpu < ncpus; cpu++)
1075                 for (thread = 0; thread < nthreads; ++thread) {
1076                         close(FD(evsel, cpu, thread));
1077                         FD(evsel, cpu, thread) = -1;
1078                 }
1079 }
1080
1081 void perf_evsel__exit(struct perf_evsel *evsel)
1082 {
1083         assert(list_empty(&evsel->node));
1084         assert(evsel->evlist == NULL);
1085         perf_evsel__free_fd(evsel);
1086         perf_evsel__free_id(evsel);
1087         perf_evsel__free_config_terms(evsel);
1088         close_cgroup(evsel->cgrp);
1089         cpu_map__put(evsel->cpus);
1090         cpu_map__put(evsel->own_cpus);
1091         thread_map__put(evsel->threads);
1092         zfree(&evsel->group_name);
1093         zfree(&evsel->name);
1094         perf_evsel__object.fini(evsel);
1095 }
1096
1097 void perf_evsel__delete(struct perf_evsel *evsel)
1098 {
1099         perf_evsel__exit(evsel);
1100         free(evsel);
1101 }
1102
1103 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1104                                 struct perf_counts_values *count)
1105 {
1106         struct perf_counts_values tmp;
1107
1108         if (!evsel->prev_raw_counts)
1109                 return;
1110
1111         if (cpu == -1) {
1112                 tmp = evsel->prev_raw_counts->aggr;
1113                 evsel->prev_raw_counts->aggr = *count;
1114         } else {
1115                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1116                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1117         }
1118
1119         count->val = count->val - tmp.val;
1120         count->ena = count->ena - tmp.ena;
1121         count->run = count->run - tmp.run;
1122 }
1123
1124 void perf_counts_values__scale(struct perf_counts_values *count,
1125                                bool scale, s8 *pscaled)
1126 {
1127         s8 scaled = 0;
1128
1129         if (scale) {
1130                 if (count->run == 0) {
1131                         scaled = -1;
1132                         count->val = 0;
1133                 } else if (count->run < count->ena) {
1134                         scaled = 1;
1135                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1136                 }
1137         } else
1138                 count->ena = count->run = 0;
1139
1140         if (pscaled)
1141                 *pscaled = scaled;
1142 }
1143
1144 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1145                      struct perf_counts_values *count)
1146 {
1147         memset(count, 0, sizeof(*count));
1148
1149         if (FD(evsel, cpu, thread) < 0)
1150                 return -EINVAL;
1151
1152         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1153                 return -errno;
1154
1155         return 0;
1156 }
1157
1158 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1159                               int cpu, int thread, bool scale)
1160 {
1161         struct perf_counts_values count;
1162         size_t nv = scale ? 3 : 1;
1163
1164         if (FD(evsel, cpu, thread) < 0)
1165                 return -EINVAL;
1166
1167         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1168                 return -ENOMEM;
1169
1170         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1171                 return -errno;
1172
1173         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1174         perf_counts_values__scale(&count, scale, NULL);
1175         *perf_counts(evsel->counts, cpu, thread) = count;
1176         return 0;
1177 }
1178
1179 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1180 {
1181         struct perf_evsel *leader = evsel->leader;
1182         int fd;
1183
1184         if (perf_evsel__is_group_leader(evsel))
1185                 return -1;
1186
1187         /*
1188          * Leader must be already processed/open,
1189          * if not it's a bug.
1190          */
1191         BUG_ON(!leader->fd);
1192
1193         fd = FD(leader, cpu, thread);
1194         BUG_ON(fd == -1);
1195
1196         return fd;
1197 }
1198
1199 struct bit_names {
1200         int bit;
1201         const char *name;
1202 };
1203
1204 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1205 {
1206         bool first_bit = true;
1207         int i = 0;
1208
1209         do {
1210                 if (value & bits[i].bit) {
1211                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1212                         first_bit = false;
1213                 }
1214         } while (bits[++i].name != NULL);
1215 }
1216
1217 static void __p_sample_type(char *buf, size_t size, u64 value)
1218 {
1219 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1220         struct bit_names bits[] = {
1221                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1222                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1223                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1224                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1225                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1226                 bit_name(WEIGHT),
1227                 { .name = NULL, }
1228         };
1229 #undef bit_name
1230         __p_bits(buf, size, value, bits);
1231 }
1232
1233 static void __p_read_format(char *buf, size_t size, u64 value)
1234 {
1235 #define bit_name(n) { PERF_FORMAT_##n, #n }
1236         struct bit_names bits[] = {
1237                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1238                 bit_name(ID), bit_name(GROUP),
1239                 { .name = NULL, }
1240         };
1241 #undef bit_name
1242         __p_bits(buf, size, value, bits);
1243 }
1244
1245 #define BUF_SIZE                1024
1246
1247 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1248 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1249 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1250 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1251 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1252
1253 #define PRINT_ATTRn(_n, _f, _p)                         \
1254 do {                                                    \
1255         if (attr->_f) {                                 \
1256                 _p(attr->_f);                           \
1257                 ret += attr__fprintf(fp, _n, buf, priv);\
1258         }                                               \
1259 } while (0)
1260
1261 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1262
1263 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1264                              attr__fprintf_f attr__fprintf, void *priv)
1265 {
1266         char buf[BUF_SIZE];
1267         int ret = 0;
1268
1269         PRINT_ATTRf(type, p_unsigned);
1270         PRINT_ATTRf(size, p_unsigned);
1271         PRINT_ATTRf(config, p_hex);
1272         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1273         PRINT_ATTRf(sample_type, p_sample_type);
1274         PRINT_ATTRf(read_format, p_read_format);
1275
1276         PRINT_ATTRf(disabled, p_unsigned);
1277         PRINT_ATTRf(inherit, p_unsigned);
1278         PRINT_ATTRf(pinned, p_unsigned);
1279         PRINT_ATTRf(exclusive, p_unsigned);
1280         PRINT_ATTRf(exclude_user, p_unsigned);
1281         PRINT_ATTRf(exclude_kernel, p_unsigned);
1282         PRINT_ATTRf(exclude_hv, p_unsigned);
1283         PRINT_ATTRf(exclude_idle, p_unsigned);
1284         PRINT_ATTRf(mmap, p_unsigned);
1285         PRINT_ATTRf(comm, p_unsigned);
1286         PRINT_ATTRf(freq, p_unsigned);
1287         PRINT_ATTRf(inherit_stat, p_unsigned);
1288         PRINT_ATTRf(enable_on_exec, p_unsigned);
1289         PRINT_ATTRf(task, p_unsigned);
1290         PRINT_ATTRf(watermark, p_unsigned);
1291         PRINT_ATTRf(precise_ip, p_unsigned);
1292         PRINT_ATTRf(mmap_data, p_unsigned);
1293         PRINT_ATTRf(sample_id_all, p_unsigned);
1294         PRINT_ATTRf(exclude_host, p_unsigned);
1295         PRINT_ATTRf(exclude_guest, p_unsigned);
1296         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1297         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1298         PRINT_ATTRf(mmap2, p_unsigned);
1299         PRINT_ATTRf(comm_exec, p_unsigned);
1300         PRINT_ATTRf(use_clockid, p_unsigned);
1301         PRINT_ATTRf(context_switch, p_unsigned);
1302
1303         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1304         PRINT_ATTRf(bp_type, p_unsigned);
1305         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1306         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1307         PRINT_ATTRf(branch_sample_type, p_unsigned);
1308         PRINT_ATTRf(sample_regs_user, p_hex);
1309         PRINT_ATTRf(sample_stack_user, p_unsigned);
1310         PRINT_ATTRf(clockid, p_signed);
1311         PRINT_ATTRf(sample_regs_intr, p_hex);
1312         PRINT_ATTRf(aux_watermark, p_unsigned);
1313
1314         return ret;
1315 }
1316
1317 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1318                                 void *priv __attribute__((unused)))
1319 {
1320         return fprintf(fp, "  %-32s %s\n", name, val);
1321 }
1322
1323 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1324                               struct thread_map *threads)
1325 {
1326         int cpu, thread, nthreads;
1327         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1328         int pid = -1, err;
1329         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1330
1331         if (evsel->system_wide)
1332                 nthreads = 1;
1333         else
1334                 nthreads = threads->nr;
1335
1336         if (evsel->fd == NULL &&
1337             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1338                 return -ENOMEM;
1339
1340         if (evsel->cgrp) {
1341                 flags |= PERF_FLAG_PID_CGROUP;
1342                 pid = evsel->cgrp->fd;
1343         }
1344
1345 fallback_missing_features:
1346         if (perf_missing_features.clockid_wrong)
1347                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1348         if (perf_missing_features.clockid) {
1349                 evsel->attr.use_clockid = 0;
1350                 evsel->attr.clockid = 0;
1351         }
1352         if (perf_missing_features.cloexec)
1353                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1354         if (perf_missing_features.mmap2)
1355                 evsel->attr.mmap2 = 0;
1356         if (perf_missing_features.exclude_guest)
1357                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1358         if (perf_missing_features.lbr_flags)
1359                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1360                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1361 retry_sample_id:
1362         if (perf_missing_features.sample_id_all)
1363                 evsel->attr.sample_id_all = 0;
1364
1365         if (verbose >= 2) {
1366                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1367                 fprintf(stderr, "perf_event_attr:\n");
1368                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1369                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1370         }
1371
1372         for (cpu = 0; cpu < cpus->nr; cpu++) {
1373
1374                 for (thread = 0; thread < nthreads; thread++) {
1375                         int group_fd;
1376
1377                         if (!evsel->cgrp && !evsel->system_wide)
1378                                 pid = thread_map__pid(threads, thread);
1379
1380                         group_fd = get_group_fd(evsel, cpu, thread);
1381 retry_open:
1382                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1383                                   pid, cpus->map[cpu], group_fd, flags);
1384
1385                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1386                                                                      pid,
1387                                                                      cpus->map[cpu],
1388                                                                      group_fd, flags);
1389                         if (FD(evsel, cpu, thread) < 0) {
1390                                 err = -errno;
1391                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1392                                           err);
1393                                 goto try_fallback;
1394                         }
1395
1396                         if (evsel->bpf_fd >= 0) {
1397                                 int evt_fd = FD(evsel, cpu, thread);
1398                                 int bpf_fd = evsel->bpf_fd;
1399
1400                                 err = ioctl(evt_fd,
1401                                             PERF_EVENT_IOC_SET_BPF,
1402                                             bpf_fd);
1403                                 if (err && errno != EEXIST) {
1404                                         pr_err("failed to attach bpf fd %d: %s\n",
1405                                                bpf_fd, strerror(errno));
1406                                         err = -EINVAL;
1407                                         goto out_close;
1408                                 }
1409                         }
1410
1411                         set_rlimit = NO_CHANGE;
1412
1413                         /*
1414                          * If we succeeded but had to kill clockid, fail and
1415                          * have perf_evsel__open_strerror() print us a nice
1416                          * error.
1417                          */
1418                         if (perf_missing_features.clockid ||
1419                             perf_missing_features.clockid_wrong) {
1420                                 err = -EINVAL;
1421                                 goto out_close;
1422                         }
1423                 }
1424         }
1425
1426         return 0;
1427
1428 try_fallback:
1429         /*
1430          * perf stat needs between 5 and 22 fds per CPU. When we run out
1431          * of them try to increase the limits.
1432          */
1433         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1434                 struct rlimit l;
1435                 int old_errno = errno;
1436
1437                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1438                         if (set_rlimit == NO_CHANGE)
1439                                 l.rlim_cur = l.rlim_max;
1440                         else {
1441                                 l.rlim_cur = l.rlim_max + 1000;
1442                                 l.rlim_max = l.rlim_cur;
1443                         }
1444                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1445                                 set_rlimit++;
1446                                 errno = old_errno;
1447                                 goto retry_open;
1448                         }
1449                 }
1450                 errno = old_errno;
1451         }
1452
1453         if (err != -EINVAL || cpu > 0 || thread > 0)
1454                 goto out_close;
1455
1456         /*
1457          * Must probe features in the order they were added to the
1458          * perf_event_attr interface.
1459          */
1460         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1461                 perf_missing_features.clockid_wrong = true;
1462                 goto fallback_missing_features;
1463         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1464                 perf_missing_features.clockid = true;
1465                 goto fallback_missing_features;
1466         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1467                 perf_missing_features.cloexec = true;
1468                 goto fallback_missing_features;
1469         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1470                 perf_missing_features.mmap2 = true;
1471                 goto fallback_missing_features;
1472         } else if (!perf_missing_features.exclude_guest &&
1473                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1474                 perf_missing_features.exclude_guest = true;
1475                 goto fallback_missing_features;
1476         } else if (!perf_missing_features.sample_id_all) {
1477                 perf_missing_features.sample_id_all = true;
1478                 goto retry_sample_id;
1479         } else if (!perf_missing_features.lbr_flags &&
1480                         (evsel->attr.branch_sample_type &
1481                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1482                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1483                 perf_missing_features.lbr_flags = true;
1484                 goto fallback_missing_features;
1485         }
1486
1487 out_close:
1488         do {
1489                 while (--thread >= 0) {
1490                         close(FD(evsel, cpu, thread));
1491                         FD(evsel, cpu, thread) = -1;
1492                 }
1493                 thread = nthreads;
1494         } while (--cpu >= 0);
1495         return err;
1496 }
1497
1498 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1499 {
1500         if (evsel->fd == NULL)
1501                 return;
1502
1503         perf_evsel__close_fd(evsel, ncpus, nthreads);
1504         perf_evsel__free_fd(evsel);
1505 }
1506
1507 static struct {
1508         struct cpu_map map;
1509         int cpus[1];
1510 } empty_cpu_map = {
1511         .map.nr = 1,
1512         .cpus   = { -1, },
1513 };
1514
1515 static struct {
1516         struct thread_map map;
1517         int threads[1];
1518 } empty_thread_map = {
1519         .map.nr  = 1,
1520         .threads = { -1, },
1521 };
1522
1523 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1524                      struct thread_map *threads)
1525 {
1526         if (cpus == NULL) {
1527                 /* Work around old compiler warnings about strict aliasing */
1528                 cpus = &empty_cpu_map.map;
1529         }
1530
1531         if (threads == NULL)
1532                 threads = &empty_thread_map.map;
1533
1534         return __perf_evsel__open(evsel, cpus, threads);
1535 }
1536
1537 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1538                              struct cpu_map *cpus)
1539 {
1540         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1541 }
1542
1543 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1544                                 struct thread_map *threads)
1545 {
1546         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1547 }
1548
1549 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1550                                        const union perf_event *event,
1551                                        struct perf_sample *sample)
1552 {
1553         u64 type = evsel->attr.sample_type;
1554         const u64 *array = event->sample.array;
1555         bool swapped = evsel->needs_swap;
1556         union u64_swap u;
1557
1558         array += ((event->header.size -
1559                    sizeof(event->header)) / sizeof(u64)) - 1;
1560
1561         if (type & PERF_SAMPLE_IDENTIFIER) {
1562                 sample->id = *array;
1563                 array--;
1564         }
1565
1566         if (type & PERF_SAMPLE_CPU) {
1567                 u.val64 = *array;
1568                 if (swapped) {
1569                         /* undo swap of u64, then swap on individual u32s */
1570                         u.val64 = bswap_64(u.val64);
1571                         u.val32[0] = bswap_32(u.val32[0]);
1572                 }
1573
1574                 sample->cpu = u.val32[0];
1575                 array--;
1576         }
1577
1578         if (type & PERF_SAMPLE_STREAM_ID) {
1579                 sample->stream_id = *array;
1580                 array--;
1581         }
1582
1583         if (type & PERF_SAMPLE_ID) {
1584                 sample->id = *array;
1585                 array--;
1586         }
1587
1588         if (type & PERF_SAMPLE_TIME) {
1589                 sample->time = *array;
1590                 array--;
1591         }
1592
1593         if (type & PERF_SAMPLE_TID) {
1594                 u.val64 = *array;
1595                 if (swapped) {
1596                         /* undo swap of u64, then swap on individual u32s */
1597                         u.val64 = bswap_64(u.val64);
1598                         u.val32[0] = bswap_32(u.val32[0]);
1599                         u.val32[1] = bswap_32(u.val32[1]);
1600                 }
1601
1602                 sample->pid = u.val32[0];
1603                 sample->tid = u.val32[1];
1604                 array--;
1605         }
1606
1607         return 0;
1608 }
1609
1610 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1611                             u64 size)
1612 {
1613         return size > max_size || offset + size > endp;
1614 }
1615
1616 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1617         do {                                                            \
1618                 if (overflow(endp, (max_size), (offset), (size)))       \
1619                         return -EFAULT;                                 \
1620         } while (0)
1621
1622 #define OVERFLOW_CHECK_u64(offset) \
1623         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1624
1625 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1626                              struct perf_sample *data)
1627 {
1628         u64 type = evsel->attr.sample_type;
1629         bool swapped = evsel->needs_swap;
1630         const u64 *array;
1631         u16 max_size = event->header.size;
1632         const void *endp = (void *)event + max_size;
1633         u64 sz;
1634
1635         /*
1636          * used for cross-endian analysis. See git commit 65014ab3
1637          * for why this goofiness is needed.
1638          */
1639         union u64_swap u;
1640
1641         memset(data, 0, sizeof(*data));
1642         data->cpu = data->pid = data->tid = -1;
1643         data->stream_id = data->id = data->time = -1ULL;
1644         data->period = evsel->attr.sample_period;
1645         data->weight = 0;
1646
1647         if (event->header.type != PERF_RECORD_SAMPLE) {
1648                 if (!evsel->attr.sample_id_all)
1649                         return 0;
1650                 return perf_evsel__parse_id_sample(evsel, event, data);
1651         }
1652
1653         array = event->sample.array;
1654
1655         /*
1656          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1657          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1658          * check the format does not go past the end of the event.
1659          */
1660         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1661                 return -EFAULT;
1662
1663         data->id = -1ULL;
1664         if (type & PERF_SAMPLE_IDENTIFIER) {
1665                 data->id = *array;
1666                 array++;
1667         }
1668
1669         if (type & PERF_SAMPLE_IP) {
1670                 data->ip = *array;
1671                 array++;
1672         }
1673
1674         if (type & PERF_SAMPLE_TID) {
1675                 u.val64 = *array;
1676                 if (swapped) {
1677                         /* undo swap of u64, then swap on individual u32s */
1678                         u.val64 = bswap_64(u.val64);
1679                         u.val32[0] = bswap_32(u.val32[0]);
1680                         u.val32[1] = bswap_32(u.val32[1]);
1681                 }
1682
1683                 data->pid = u.val32[0];
1684                 data->tid = u.val32[1];
1685                 array++;
1686         }
1687
1688         if (type & PERF_SAMPLE_TIME) {
1689                 data->time = *array;
1690                 array++;
1691         }
1692
1693         data->addr = 0;
1694         if (type & PERF_SAMPLE_ADDR) {
1695                 data->addr = *array;
1696                 array++;
1697         }
1698
1699         if (type & PERF_SAMPLE_ID) {
1700                 data->id = *array;
1701                 array++;
1702         }
1703
1704         if (type & PERF_SAMPLE_STREAM_ID) {
1705                 data->stream_id = *array;
1706                 array++;
1707         }
1708
1709         if (type & PERF_SAMPLE_CPU) {
1710
1711                 u.val64 = *array;
1712                 if (swapped) {
1713                         /* undo swap of u64, then swap on individual u32s */
1714                         u.val64 = bswap_64(u.val64);
1715                         u.val32[0] = bswap_32(u.val32[0]);
1716                 }
1717
1718                 data->cpu = u.val32[0];
1719                 array++;
1720         }
1721
1722         if (type & PERF_SAMPLE_PERIOD) {
1723                 data->period = *array;
1724                 array++;
1725         }
1726
1727         if (type & PERF_SAMPLE_READ) {
1728                 u64 read_format = evsel->attr.read_format;
1729
1730                 OVERFLOW_CHECK_u64(array);
1731                 if (read_format & PERF_FORMAT_GROUP)
1732                         data->read.group.nr = *array;
1733                 else
1734                         data->read.one.value = *array;
1735
1736                 array++;
1737
1738                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1739                         OVERFLOW_CHECK_u64(array);
1740                         data->read.time_enabled = *array;
1741                         array++;
1742                 }
1743
1744                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1745                         OVERFLOW_CHECK_u64(array);
1746                         data->read.time_running = *array;
1747                         array++;
1748                 }
1749
1750                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1751                 if (read_format & PERF_FORMAT_GROUP) {
1752                         const u64 max_group_nr = UINT64_MAX /
1753                                         sizeof(struct sample_read_value);
1754
1755                         if (data->read.group.nr > max_group_nr)
1756                                 return -EFAULT;
1757                         sz = data->read.group.nr *
1758                              sizeof(struct sample_read_value);
1759                         OVERFLOW_CHECK(array, sz, max_size);
1760                         data->read.group.values =
1761                                         (struct sample_read_value *)array;
1762                         array = (void *)array + sz;
1763                 } else {
1764                         OVERFLOW_CHECK_u64(array);
1765                         data->read.one.id = *array;
1766                         array++;
1767                 }
1768         }
1769
1770         if (type & PERF_SAMPLE_CALLCHAIN) {
1771                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1772
1773                 OVERFLOW_CHECK_u64(array);
1774                 data->callchain = (struct ip_callchain *)array++;
1775                 if (data->callchain->nr > max_callchain_nr)
1776                         return -EFAULT;
1777                 sz = data->callchain->nr * sizeof(u64);
1778                 OVERFLOW_CHECK(array, sz, max_size);
1779                 array = (void *)array + sz;
1780         }
1781
1782         if (type & PERF_SAMPLE_RAW) {
1783                 OVERFLOW_CHECK_u64(array);
1784                 u.val64 = *array;
1785                 if (WARN_ONCE(swapped,
1786                               "Endianness of raw data not corrected!\n")) {
1787                         /* undo swap of u64, then swap on individual u32s */
1788                         u.val64 = bswap_64(u.val64);
1789                         u.val32[0] = bswap_32(u.val32[0]);
1790                         u.val32[1] = bswap_32(u.val32[1]);
1791                 }
1792                 data->raw_size = u.val32[0];
1793                 array = (void *)array + sizeof(u32);
1794
1795                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1796                 data->raw_data = (void *)array;
1797                 array = (void *)array + data->raw_size;
1798         }
1799
1800         if (type & PERF_SAMPLE_BRANCH_STACK) {
1801                 const u64 max_branch_nr = UINT64_MAX /
1802                                           sizeof(struct branch_entry);
1803
1804                 OVERFLOW_CHECK_u64(array);
1805                 data->branch_stack = (struct branch_stack *)array++;
1806
1807                 if (data->branch_stack->nr > max_branch_nr)
1808                         return -EFAULT;
1809                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1810                 OVERFLOW_CHECK(array, sz, max_size);
1811                 array = (void *)array + sz;
1812         }
1813
1814         if (type & PERF_SAMPLE_REGS_USER) {
1815                 OVERFLOW_CHECK_u64(array);
1816                 data->user_regs.abi = *array;
1817                 array++;
1818
1819                 if (data->user_regs.abi) {
1820                         u64 mask = evsel->attr.sample_regs_user;
1821
1822                         sz = hweight_long(mask) * sizeof(u64);
1823                         OVERFLOW_CHECK(array, sz, max_size);
1824                         data->user_regs.mask = mask;
1825                         data->user_regs.regs = (u64 *)array;
1826                         array = (void *)array + sz;
1827                 }
1828         }
1829
1830         if (type & PERF_SAMPLE_STACK_USER) {
1831                 OVERFLOW_CHECK_u64(array);
1832                 sz = *array++;
1833
1834                 data->user_stack.offset = ((char *)(array - 1)
1835                                           - (char *) event);
1836
1837                 if (!sz) {
1838                         data->user_stack.size = 0;
1839                 } else {
1840                         OVERFLOW_CHECK(array, sz, max_size);
1841                         data->user_stack.data = (char *)array;
1842                         array = (void *)array + sz;
1843                         OVERFLOW_CHECK_u64(array);
1844                         data->user_stack.size = *array++;
1845                         if (WARN_ONCE(data->user_stack.size > sz,
1846                                       "user stack dump failure\n"))
1847                                 return -EFAULT;
1848                 }
1849         }
1850
1851         data->weight = 0;
1852         if (type & PERF_SAMPLE_WEIGHT) {
1853                 OVERFLOW_CHECK_u64(array);
1854                 data->weight = *array;
1855                 array++;
1856         }
1857
1858         data->data_src = PERF_MEM_DATA_SRC_NONE;
1859         if (type & PERF_SAMPLE_DATA_SRC) {
1860                 OVERFLOW_CHECK_u64(array);
1861                 data->data_src = *array;
1862                 array++;
1863         }
1864
1865         data->transaction = 0;
1866         if (type & PERF_SAMPLE_TRANSACTION) {
1867                 OVERFLOW_CHECK_u64(array);
1868                 data->transaction = *array;
1869                 array++;
1870         }
1871
1872         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1873         if (type & PERF_SAMPLE_REGS_INTR) {
1874                 OVERFLOW_CHECK_u64(array);
1875                 data->intr_regs.abi = *array;
1876                 array++;
1877
1878                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1879                         u64 mask = evsel->attr.sample_regs_intr;
1880
1881                         sz = hweight_long(mask) * sizeof(u64);
1882                         OVERFLOW_CHECK(array, sz, max_size);
1883                         data->intr_regs.mask = mask;
1884                         data->intr_regs.regs = (u64 *)array;
1885                         array = (void *)array + sz;
1886                 }
1887         }
1888
1889         return 0;
1890 }
1891
1892 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1893                                      u64 read_format)
1894 {
1895         size_t sz, result = sizeof(struct sample_event);
1896
1897         if (type & PERF_SAMPLE_IDENTIFIER)
1898                 result += sizeof(u64);
1899
1900         if (type & PERF_SAMPLE_IP)
1901                 result += sizeof(u64);
1902
1903         if (type & PERF_SAMPLE_TID)
1904                 result += sizeof(u64);
1905
1906         if (type & PERF_SAMPLE_TIME)
1907                 result += sizeof(u64);
1908
1909         if (type & PERF_SAMPLE_ADDR)
1910                 result += sizeof(u64);
1911
1912         if (type & PERF_SAMPLE_ID)
1913                 result += sizeof(u64);
1914
1915         if (type & PERF_SAMPLE_STREAM_ID)
1916                 result += sizeof(u64);
1917
1918         if (type & PERF_SAMPLE_CPU)
1919                 result += sizeof(u64);
1920
1921         if (type & PERF_SAMPLE_PERIOD)
1922                 result += sizeof(u64);
1923
1924         if (type & PERF_SAMPLE_READ) {
1925                 result += sizeof(u64);
1926                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1927                         result += sizeof(u64);
1928                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1929                         result += sizeof(u64);
1930                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1931                 if (read_format & PERF_FORMAT_GROUP) {
1932                         sz = sample->read.group.nr *
1933                              sizeof(struct sample_read_value);
1934                         result += sz;
1935                 } else {
1936                         result += sizeof(u64);
1937                 }
1938         }
1939
1940         if (type & PERF_SAMPLE_CALLCHAIN) {
1941                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1942                 result += sz;
1943         }
1944
1945         if (type & PERF_SAMPLE_RAW) {
1946                 result += sizeof(u32);
1947                 result += sample->raw_size;
1948         }
1949
1950         if (type & PERF_SAMPLE_BRANCH_STACK) {
1951                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1952                 sz += sizeof(u64);
1953                 result += sz;
1954         }
1955
1956         if (type & PERF_SAMPLE_REGS_USER) {
1957                 if (sample->user_regs.abi) {
1958                         result += sizeof(u64);
1959                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1960                         result += sz;
1961                 } else {
1962                         result += sizeof(u64);
1963                 }
1964         }
1965
1966         if (type & PERF_SAMPLE_STACK_USER) {
1967                 sz = sample->user_stack.size;
1968                 result += sizeof(u64);
1969                 if (sz) {
1970                         result += sz;
1971                         result += sizeof(u64);
1972                 }
1973         }
1974
1975         if (type & PERF_SAMPLE_WEIGHT)
1976                 result += sizeof(u64);
1977
1978         if (type & PERF_SAMPLE_DATA_SRC)
1979                 result += sizeof(u64);
1980
1981         if (type & PERF_SAMPLE_TRANSACTION)
1982                 result += sizeof(u64);
1983
1984         if (type & PERF_SAMPLE_REGS_INTR) {
1985                 if (sample->intr_regs.abi) {
1986                         result += sizeof(u64);
1987                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
1988                         result += sz;
1989                 } else {
1990                         result += sizeof(u64);
1991                 }
1992         }
1993
1994         return result;
1995 }
1996
1997 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1998                                   u64 read_format,
1999                                   const struct perf_sample *sample,
2000                                   bool swapped)
2001 {
2002         u64 *array;
2003         size_t sz;
2004         /*
2005          * used for cross-endian analysis. See git commit 65014ab3
2006          * for why this goofiness is needed.
2007          */
2008         union u64_swap u;
2009
2010         array = event->sample.array;
2011
2012         if (type & PERF_SAMPLE_IDENTIFIER) {
2013                 *array = sample->id;
2014                 array++;
2015         }
2016
2017         if (type & PERF_SAMPLE_IP) {
2018                 *array = sample->ip;
2019                 array++;
2020         }
2021
2022         if (type & PERF_SAMPLE_TID) {
2023                 u.val32[0] = sample->pid;
2024                 u.val32[1] = sample->tid;
2025                 if (swapped) {
2026                         /*
2027                          * Inverse of what is done in perf_evsel__parse_sample
2028                          */
2029                         u.val32[0] = bswap_32(u.val32[0]);
2030                         u.val32[1] = bswap_32(u.val32[1]);
2031                         u.val64 = bswap_64(u.val64);
2032                 }
2033
2034                 *array = u.val64;
2035                 array++;
2036         }
2037
2038         if (type & PERF_SAMPLE_TIME) {
2039                 *array = sample->time;
2040                 array++;
2041         }
2042
2043         if (type & PERF_SAMPLE_ADDR) {
2044                 *array = sample->addr;
2045                 array++;
2046         }
2047
2048         if (type & PERF_SAMPLE_ID) {
2049                 *array = sample->id;
2050                 array++;
2051         }
2052
2053         if (type & PERF_SAMPLE_STREAM_ID) {
2054                 *array = sample->stream_id;
2055                 array++;
2056         }
2057
2058         if (type & PERF_SAMPLE_CPU) {
2059                 u.val32[0] = sample->cpu;
2060                 if (swapped) {
2061                         /*
2062                          * Inverse of what is done in perf_evsel__parse_sample
2063                          */
2064                         u.val32[0] = bswap_32(u.val32[0]);
2065                         u.val64 = bswap_64(u.val64);
2066                 }
2067                 *array = u.val64;
2068                 array++;
2069         }
2070
2071         if (type & PERF_SAMPLE_PERIOD) {
2072                 *array = sample->period;
2073                 array++;
2074         }
2075
2076         if (type & PERF_SAMPLE_READ) {
2077                 if (read_format & PERF_FORMAT_GROUP)
2078                         *array = sample->read.group.nr;
2079                 else
2080                         *array = sample->read.one.value;
2081                 array++;
2082
2083                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2084                         *array = sample->read.time_enabled;
2085                         array++;
2086                 }
2087
2088                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2089                         *array = sample->read.time_running;
2090                         array++;
2091                 }
2092
2093                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2094                 if (read_format & PERF_FORMAT_GROUP) {
2095                         sz = sample->read.group.nr *
2096                              sizeof(struct sample_read_value);
2097                         memcpy(array, sample->read.group.values, sz);
2098                         array = (void *)array + sz;
2099                 } else {
2100                         *array = sample->read.one.id;
2101                         array++;
2102                 }
2103         }
2104
2105         if (type & PERF_SAMPLE_CALLCHAIN) {
2106                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2107                 memcpy(array, sample->callchain, sz);
2108                 array = (void *)array + sz;
2109         }
2110
2111         if (type & PERF_SAMPLE_RAW) {
2112                 u.val32[0] = sample->raw_size;
2113                 if (WARN_ONCE(swapped,
2114                               "Endianness of raw data not corrected!\n")) {
2115                         /*
2116                          * Inverse of what is done in perf_evsel__parse_sample
2117                          */
2118                         u.val32[0] = bswap_32(u.val32[0]);
2119                         u.val32[1] = bswap_32(u.val32[1]);
2120                         u.val64 = bswap_64(u.val64);
2121                 }
2122                 *array = u.val64;
2123                 array = (void *)array + sizeof(u32);
2124
2125                 memcpy(array, sample->raw_data, sample->raw_size);
2126                 array = (void *)array + sample->raw_size;
2127         }
2128
2129         if (type & PERF_SAMPLE_BRANCH_STACK) {
2130                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2131                 sz += sizeof(u64);
2132                 memcpy(array, sample->branch_stack, sz);
2133                 array = (void *)array + sz;
2134         }
2135
2136         if (type & PERF_SAMPLE_REGS_USER) {
2137                 if (sample->user_regs.abi) {
2138                         *array++ = sample->user_regs.abi;
2139                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2140                         memcpy(array, sample->user_regs.regs, sz);
2141                         array = (void *)array + sz;
2142                 } else {
2143                         *array++ = 0;
2144                 }
2145         }
2146
2147         if (type & PERF_SAMPLE_STACK_USER) {
2148                 sz = sample->user_stack.size;
2149                 *array++ = sz;
2150                 if (sz) {
2151                         memcpy(array, sample->user_stack.data, sz);
2152                         array = (void *)array + sz;
2153                         *array++ = sz;
2154                 }
2155         }
2156
2157         if (type & PERF_SAMPLE_WEIGHT) {
2158                 *array = sample->weight;
2159                 array++;
2160         }
2161
2162         if (type & PERF_SAMPLE_DATA_SRC) {
2163                 *array = sample->data_src;
2164                 array++;
2165         }
2166
2167         if (type & PERF_SAMPLE_TRANSACTION) {
2168                 *array = sample->transaction;
2169                 array++;
2170         }
2171
2172         if (type & PERF_SAMPLE_REGS_INTR) {
2173                 if (sample->intr_regs.abi) {
2174                         *array++ = sample->intr_regs.abi;
2175                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2176                         memcpy(array, sample->intr_regs.regs, sz);
2177                         array = (void *)array + sz;
2178                 } else {
2179                         *array++ = 0;
2180                 }
2181         }
2182
2183         return 0;
2184 }
2185
2186 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2187 {
2188         return pevent_find_field(evsel->tp_format, name);
2189 }
2190
2191 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2192                          const char *name)
2193 {
2194         struct format_field *field = perf_evsel__field(evsel, name);
2195         int offset;
2196
2197         if (!field)
2198                 return NULL;
2199
2200         offset = field->offset;
2201
2202         if (field->flags & FIELD_IS_DYNAMIC) {
2203                 offset = *(int *)(sample->raw_data + field->offset);
2204                 offset &= 0xffff;
2205         }
2206
2207         return sample->raw_data + offset;
2208 }
2209
2210 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2211                        const char *name)
2212 {
2213         struct format_field *field = perf_evsel__field(evsel, name);
2214         void *ptr;
2215         u64 value;
2216
2217         if (!field)
2218                 return 0;
2219
2220         ptr = sample->raw_data + field->offset;
2221
2222         switch (field->size) {
2223         case 1:
2224                 return *(u8 *)ptr;
2225         case 2:
2226                 value = *(u16 *)ptr;
2227                 break;
2228         case 4:
2229                 value = *(u32 *)ptr;
2230                 break;
2231         case 8:
2232                 memcpy(&value, ptr, sizeof(u64));
2233                 break;
2234         default:
2235                 return 0;
2236         }
2237
2238         if (!evsel->needs_swap)
2239                 return value;
2240
2241         switch (field->size) {
2242         case 2:
2243                 return bswap_16(value);
2244         case 4:
2245                 return bswap_32(value);
2246         case 8:
2247                 return bswap_64(value);
2248         default:
2249                 return 0;
2250         }
2251
2252         return 0;
2253 }
2254
2255 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
2256 {
2257         va_list args;
2258         int ret = 0;
2259
2260         if (!*first) {
2261                 ret += fprintf(fp, ",");
2262         } else {
2263                 ret += fprintf(fp, ":");
2264                 *first = false;
2265         }
2266
2267         va_start(args, fmt);
2268         ret += vfprintf(fp, fmt, args);
2269         va_end(args);
2270         return ret;
2271 }
2272
2273 static int __print_attr__fprintf(FILE *fp, const char *name, const char *val, void *priv)
2274 {
2275         return comma_fprintf(fp, (bool *)priv, " %s: %s", name, val);
2276 }
2277
2278 int perf_evsel__fprintf(struct perf_evsel *evsel,
2279                         struct perf_attr_details *details, FILE *fp)
2280 {
2281         bool first = true;
2282         int printed = 0;
2283
2284         if (details->event_group) {
2285                 struct perf_evsel *pos;
2286
2287                 if (!perf_evsel__is_group_leader(evsel))
2288                         return 0;
2289
2290                 if (evsel->nr_members > 1)
2291                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
2292
2293                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2294                 for_each_group_member(pos, evsel)
2295                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
2296
2297                 if (evsel->nr_members > 1)
2298                         printed += fprintf(fp, "}");
2299                 goto out;
2300         }
2301
2302         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
2303
2304         if (details->verbose) {
2305                 printed += perf_event_attr__fprintf(fp, &evsel->attr,
2306                                                     __print_attr__fprintf, &first);
2307         } else if (details->freq) {
2308                 const char *term = "sample_freq";
2309
2310                 if (!evsel->attr.freq)
2311                         term = "sample_period";
2312
2313                 printed += comma_fprintf(fp, &first, " %s=%" PRIu64,
2314                                          term, (u64)evsel->attr.sample_freq);
2315         }
2316
2317         if (details->trace_fields) {
2318                 struct format_field *field;
2319
2320                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT) {
2321                         printed += comma_fprintf(fp, &first, " (not a tracepoint)");
2322                         goto out;
2323                 }
2324
2325                 field = evsel->tp_format->format.fields;
2326                 if (field == NULL) {
2327                         printed += comma_fprintf(fp, &first, " (no trace field)");
2328                         goto out;
2329                 }
2330
2331                 printed += comma_fprintf(fp, &first, " trace_fields: %s", field->name);
2332
2333                 field = field->next;
2334                 while (field) {
2335                         printed += comma_fprintf(fp, &first, "%s", field->name);
2336                         field = field->next;
2337                 }
2338         }
2339 out:
2340         fputc('\n', fp);
2341         return ++printed;
2342 }
2343
2344 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2345                           char *msg, size_t msgsize)
2346 {
2347         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2348             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2349             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2350                 /*
2351                  * If it's cycles then fall back to hrtimer based
2352                  * cpu-clock-tick sw counter, which is always available even if
2353                  * no PMU support.
2354                  *
2355                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2356                  * b0a873e).
2357                  */
2358                 scnprintf(msg, msgsize, "%s",
2359 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2360
2361                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2362                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2363
2364                 zfree(&evsel->name);
2365                 return true;
2366         }
2367
2368         return false;
2369 }
2370
2371 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2372                               int err, char *msg, size_t size)
2373 {
2374         char sbuf[STRERR_BUFSIZE];
2375
2376         switch (err) {
2377         case EPERM:
2378         case EACCES:
2379                 return scnprintf(msg, size,
2380                  "You may not have permission to collect %sstats.\n\n"
2381                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2382                  "which controls use of the performance events system by\n"
2383                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2384                  "The default value is 1:\n\n"
2385                  "  -1: Allow use of (almost) all events by all users\n"
2386                  ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2387                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2388                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2389                                  target->system_wide ? "system-wide " : "");
2390         case ENOENT:
2391                 return scnprintf(msg, size, "The %s event is not supported.",
2392                                  perf_evsel__name(evsel));
2393         case EMFILE:
2394                 return scnprintf(msg, size, "%s",
2395                          "Too many events are opened.\n"
2396                          "Probably the maximum number of open file descriptors has been reached.\n"
2397                          "Hint: Try again after reducing the number of events.\n"
2398                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2399         case ENODEV:
2400                 if (target->cpu_list)
2401                         return scnprintf(msg, size, "%s",
2402          "No such device - did you specify an out-of-range profile CPU?\n");
2403                 break;
2404         case EOPNOTSUPP:
2405                 if (evsel->attr.precise_ip)
2406                         return scnprintf(msg, size, "%s",
2407         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2408 #if defined(__i386__) || defined(__x86_64__)
2409                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2410                         return scnprintf(msg, size, "%s",
2411         "No hardware sampling interrupt available.\n"
2412         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2413 #endif
2414                 break;
2415         case EBUSY:
2416                 if (find_process("oprofiled"))
2417                         return scnprintf(msg, size,
2418         "The PMU counters are busy/taken by another profiler.\n"
2419         "We found oprofile daemon running, please stop it and try again.");
2420                 break;
2421         case EINVAL:
2422                 if (perf_missing_features.clockid)
2423                         return scnprintf(msg, size, "clockid feature not supported.");
2424                 if (perf_missing_features.clockid_wrong)
2425                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2426                 break;
2427         default:
2428                 break;
2429         }
2430
2431         return scnprintf(msg, size,
2432         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2433         "/bin/dmesg may provide additional information.\n"
2434         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
2435                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2436                          perf_evsel__name(evsel));
2437 }