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