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