tools/perf/util/evsel.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9
  10 #include <byteswap.h>
  11 #include <linux/bitops.h>
  12 #include <api/fs/tracing_path.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <linux/err.h>
  17 #include <sys/resource.h>
  18 #include "asm/bug.h"
  19 #include "callchain.h"
  20 #include "cgroup.h"
  21 #include "evsel.h"
  22 #include "evlist.h"
  23 #include "util.h"
  24 #include "cpumap.h"
  25 #include "thread_map.h"
  26 #include "target.h"
  27 #include "perf_regs.h"
  28 #include "debug.h"
  29 #include "trace-event.h"
  30 #include "stat.h"
  31
  32 static struct {
  33         bool sample_id_all;
  34         bool exclude_guest;
  35         bool mmap2;
  36         bool cloexec;
  37         bool clockid;
  38         bool clockid_wrong;
  39         bool lbr_flags;
  40 } perf_missing_features;
  41
  42 static clockid_t clockid;
  43
  44 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  45 {
  46         return 0;
  47 }
  48
  49 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  50 {
  51 }
  52
  53 static struct {
  54         size_t  size;
  55         int     (*init)(struct perf_evsel *evsel);
  56         void    (*fini)(struct perf_evsel *evsel);
  57 } perf_evsel__object = {
  58         .size = sizeof(struct perf_evsel),
  59         .init = perf_evsel__no_extra_init,
  60         .fini = perf_evsel__no_extra_fini,
  61 };
  62
  63 int perf_evsel__object_config(size_t object_size,
  64                               int (*init)(struct perf_evsel *evsel),
  65                               void (*fini)(struct perf_evsel *evsel))
  66 {
  67
  68         if (object_size == 0)
  69                 goto set_methods;
  70
  71         if (perf_evsel__object.size > object_size)
  72                 return -EINVAL;
  73
  74         perf_evsel__object.size = object_size;
  75
  76 set_methods:
  77         if (init != NULL)
  78                 perf_evsel__object.init = init;
  79
  80         if (fini != NULL)
  81                 perf_evsel__object.fini = fini;
  82
  83         return 0;
  84 }
  85
  86 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  87
  88 int __perf_evsel__sample_size(u64 sample_type)
  89 {
  90         u64 mask = sample_type & PERF_SAMPLE_MASK;
  91         int size = 0;
  92         int i;
  93
  94         for (i = 0; i < 64; i++) {
  95                 if (mask & (1ULL << i))
  96                         size++;
  97         }
  98
  99         size *= sizeof(u64);
 100
 101         return size;
 102 }
 103
 104 /**
 105  * __perf_evsel__calc_id_pos - calculate id_pos.
 106  * @sample_type: sample type
 107  *
 108  * This function returns the position of the event id (PERF_SAMPLE_ID or
 109  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 110  * sample_event.
 111  */
 112 static int __perf_evsel__calc_id_pos(u64 sample_type)
 113 {
 114         int idx = 0;
 115
 116         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 117                 return 0;
 118
 119         if (!(sample_type & PERF_SAMPLE_ID))
 120                 return -1;
 121
 122         if (sample_type & PERF_SAMPLE_IP)
 123                 idx += 1;
 124
 125         if (sample_type & PERF_SAMPLE_TID)
 126                 idx += 1;
 127
 128         if (sample_type & PERF_SAMPLE_TIME)
 129                 idx += 1;
 130
 131         if (sample_type & PERF_SAMPLE_ADDR)
 132                 idx += 1;
 133
 134         return idx;
 135 }
 136
 137 /**
 138  * __perf_evsel__calc_is_pos - calculate is_pos.
 139  * @sample_type: sample type
 140  *
 141  * This function returns the position (counting backwards) of the event id
 142  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 143  * sample_id_all is used there is an id sample appended to non-sample events.
 144  */
 145 static int __perf_evsel__calc_is_pos(u64 sample_type)
 146 {
 147         int idx = 1;
 148
 149         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 150                 return 1;
 151
 152         if (!(sample_type & PERF_SAMPLE_ID))
 153                 return -1;
 154
 155         if (sample_type & PERF_SAMPLE_CPU)
 156                 idx += 1;
 157
 158         if (sample_type & PERF_SAMPLE_STREAM_ID)
 159                 idx += 1;
 160
 161         return idx;
 162 }
 163
 164 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 165 {
 166         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 167         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 168 }
 169
 170 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 171                                   enum perf_event_sample_format bit)
 172 {
 173         if (!(evsel->attr.sample_type & bit)) {
 174                 evsel->attr.sample_type |= bit;
 175                 evsel->sample_size += sizeof(u64);
 176                 perf_evsel__calc_id_pos(evsel);
 177         }
 178 }
 179
 180 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 181                                     enum perf_event_sample_format bit)
 182 {
 183         if (evsel->attr.sample_type & bit) {
 184                 evsel->attr.sample_type &= ~bit;
 185                 evsel->sample_size -= sizeof(u64);
 186                 perf_evsel__calc_id_pos(evsel);
 187         }
 188 }
 189
 190 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 191                                bool can_sample_identifier)
 192 {
 193         if (can_sample_identifier) {
 194                 perf_evsel__reset_sample_bit(evsel, ID);
 195                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 196         } else {
 197                 perf_evsel__set_sample_bit(evsel, ID);
 198         }
 199         evsel->attr.read_format |= PERF_FORMAT_ID;
 200 }
 201
 202 void perf_evsel__init(struct perf_evsel *evsel,
 203                       struct perf_event_attr *attr, int idx)
 204 {
 205         evsel->idx         = idx;
 206         evsel->tracking    = !idx;
 207         evsel->attr        = *attr;
 208         evsel->leader      = evsel;
 209         evsel->unit        = "";
 210         evsel->scale       = 1.0;
 211         evsel->evlist      = NULL;
 212         evsel->bpf_fd      = -1;
 213         INIT_LIST_HEAD(&evsel->node);
 214         INIT_LIST_HEAD(&evsel->config_terms);
 215         perf_evsel__object.init(evsel);
 216         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 217         perf_evsel__calc_id_pos(evsel);
 218         evsel->cmdline_group_boundary = false;
 219 }
 220
 221 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 222 {
 223         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 224
 225         if (evsel != NULL)
 226                 perf_evsel__init(evsel, attr, idx);
 227
 228         if (perf_evsel__is_bpf_output(evsel)) {
 229                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 230                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 231                 evsel->attr.sample_period = 1;
 232         }
 233
 234         return evsel;
 235 }
 236
 237 /*
 238  * Returns pointer with encoded error via <linux/err.h> interface.
 239  */
 240 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 241 {
 242         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 243         int err = -ENOMEM;
 244
 245         if (evsel == NULL) {
 246                 goto out_err;
 247         } else {
 248                 struct perf_event_attr attr = {
 249                         .type          = PERF_TYPE_TRACEPOINT,
 250                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 251                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 252                 };
 253
 254                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 255                         goto out_free;
 256
 257                 evsel->tp_format = trace_event__tp_format(sys, name);
 258                 if (IS_ERR(evsel->tp_format)) {
 259                         err = PTR_ERR(evsel->tp_format);
 260                         goto out_free;
 261                 }
 262
 263                 event_attr_init(&attr);
 264                 attr.config = evsel->tp_format->id;
 265                 attr.sample_period = 1;
 266                 perf_evsel__init(evsel, &attr, idx);
 267         }
 268
 269         return evsel;
 270
 271 out_free:
 272         zfree(&evsel->name);
 273         free(evsel);
 274 out_err:
 275         return ERR_PTR(err);
 276 }
 277
 278 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 279         "cycles",
 280         "instructions",
 281         "cache-references",
 282         "cache-misses",
 283         "branches",
 284         "branch-misses",
 285         "bus-cycles",
 286         "stalled-cycles-frontend",
 287         "stalled-cycles-backend",
 288         "ref-cycles",
 289 };
 290
 291 static const char *__perf_evsel__hw_name(u64 config)
 292 {
 293         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 294                 return perf_evsel__hw_names[config];
 295
 296         return "unknown-hardware";
 297 }
 298
 299 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 300 {
 301         int colon = 0, r = 0;
 302         struct perf_event_attr *attr = &evsel->attr;
 303         bool exclude_guest_default = false;
 304
 305 #define MOD_PRINT(context, mod) do {                                    \
 306                 if (!attr->exclude_##context) {                         \
 307                         if (!colon) colon = ++r;                        \
 308                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 309                 } } while(0)
 310
 311         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 312                 MOD_PRINT(kernel, 'k');
 313                 MOD_PRINT(user, 'u');
 314                 MOD_PRINT(hv, 'h');
 315                 exclude_guest_default = true;
 316         }
 317
 318         if (attr->precise_ip) {
 319                 if (!colon)
 320                         colon = ++r;
 321                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 322                 exclude_guest_default = true;
 323         }
 324
 325         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 326                 MOD_PRINT(host, 'H');
 327                 MOD_PRINT(guest, 'G');
 328         }
 329 #undef MOD_PRINT
 330         if (colon)
 331                 bf[colon - 1] = ':';
 332         return r;
 333 }
 334
 335 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 336 {
 337         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 338         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 339 }
 340
 341 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 342         "cpu-clock",
 343         "task-clock",
 344         "page-faults",
 345         "context-switches",
 346         "cpu-migrations",
 347         "minor-faults",
 348         "major-faults",
 349         "alignment-faults",
 350         "emulation-faults",
 351         "dummy",
 352 };
 353
 354 static const char *__perf_evsel__sw_name(u64 config)
 355 {
 356         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 357                 return perf_evsel__sw_names[config];
 358         return "unknown-software";
 359 }
 360
 361 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 362 {
 363         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 364         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 365 }
 366
 367 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 368 {
 369         int r;
 370
 371         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 372
 373         if (type & HW_BREAKPOINT_R)
 374                 r += scnprintf(bf + r, size - r, "r");
 375
 376         if (type & HW_BREAKPOINT_W)
 377                 r += scnprintf(bf + r, size - r, "w");
 378
 379         if (type & HW_BREAKPOINT_X)
 380                 r += scnprintf(bf + r, size - r, "x");
 381
 382         return r;
 383 }
 384
 385 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 386 {
 387         struct perf_event_attr *attr = &evsel->attr;
 388         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 389         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 390 }
 391
 392 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 393                                 [PERF_EVSEL__MAX_ALIASES] = {
 394  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 395  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 396  { "LLC",       "L2",                                                   },
 397  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 398  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 399  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 400  { "node",                                                              },
 401 };
 402
 403 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 404                                    [PERF_EVSEL__MAX_ALIASES] = {
 405  { "load",      "loads",        "read",                                 },
 406  { "store",     "stores",       "write",                                },
 407  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 408 };
 409
 410 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 411                                        [PERF_EVSEL__MAX_ALIASES] = {
 412  { "refs",      "Reference",    "ops",          "access",               },
 413  { "misses",    "miss",                                                 },
 414 };
 415
 416 #define C(x)            PERF_COUNT_HW_CACHE_##x
 417 #define CACHE_READ      (1 << C(OP_READ))
 418 #define CACHE_WRITE     (1 << C(OP_WRITE))
 419 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 420 #define COP(x)          (1 << x)
 421
 422 /*
 423  * cache operartion stat
 424  * L1I : Read and prefetch only
 425  * ITLB and BPU : Read-only
 426  */
 427 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 428  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 429  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 430  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 431  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 432  [C(ITLB)]      = (CACHE_READ),
 433  [C(BPU)]       = (CACHE_READ),
 434  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 435 };
 436
 437 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 438 {
 439         if (perf_evsel__hw_cache_stat[type] & COP(op))
 440                 return true;    /* valid */
 441         else
 442                 return false;   /* invalid */
 443 }
 444
 445 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 446                                             char *bf, size_t size)
 447 {
 448         if (result) {
 449                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 450                                  perf_evsel__hw_cache_op[op][0],
 451                                  perf_evsel__hw_cache_result[result][0]);
 452         }
 453
 454         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 455                          perf_evsel__hw_cache_op[op][1]);
 456 }
 457
 458 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 459 {
 460         u8 op, result, type = (config >>  0) & 0xff;
 461         const char *err = "unknown-ext-hardware-cache-type";
 462
 463         if (type > PERF_COUNT_HW_CACHE_MAX)
 464                 goto out_err;
 465
 466         op = (config >>  8) & 0xff;
 467         err = "unknown-ext-hardware-cache-op";
 468         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 469                 goto out_err;
 470
 471         result = (config >> 16) & 0xff;
 472         err = "unknown-ext-hardware-cache-result";
 473         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 474                 goto out_err;
 475
 476         err = "invalid-cache";
 477         if (!perf_evsel__is_cache_op_valid(type, op))
 478                 goto out_err;
 479
 480         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 481 out_err:
 482         return scnprintf(bf, size, "%s", err);
 483 }
 484
 485 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 486 {
 487         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 488         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 489 }
 490
 491 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 492 {
 493         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 494         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 495 }
 496
 497 const char *perf_evsel__name(struct perf_evsel *evsel)
 498 {
 499         char bf[128];
 500
 501         if (evsel->name)
 502                 return evsel->name;
 503
 504         switch (evsel->attr.type) {
 505         case PERF_TYPE_RAW:
 506                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 507                 break;
 508
 509         case PERF_TYPE_HARDWARE:
 510                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 511                 break;
 512
 513         case PERF_TYPE_HW_CACHE:
 514                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 515                 break;
 516
 517         case PERF_TYPE_SOFTWARE:
 518                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 519                 break;
 520
 521         case PERF_TYPE_TRACEPOINT:
 522                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 523                 break;
 524
 525         case PERF_TYPE_BREAKPOINT:
 526                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 527                 break;
 528
 529         default:
 530                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 531                           evsel->attr.type);
 532                 break;
 533         }
 534
 535         evsel->name = strdup(bf);
 536
 537         return evsel->name ?: "unknown";
 538 }
 539
 540 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 541 {
 542         return evsel->group_name ?: "anon group";
 543 }
 544
 545 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 546 {
 547         int ret;
 548         struct perf_evsel *pos;
 549         const char *group_name = perf_evsel__group_name(evsel);
 550
 551         ret = scnprintf(buf, size, "%s", group_name);
 552
 553         ret += scnprintf(buf + ret, size - ret, " { %s",
 554                          perf_evsel__name(evsel));
 555
 556         for_each_group_member(pos, evsel)
 557                 ret += scnprintf(buf + ret, size - ret, ", %s",
 558                                  perf_evsel__name(pos));
 559
 560         ret += scnprintf(buf + ret, size - ret, " }");
 561
 562         return ret;
 563 }
 564
 565 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 566                                   struct record_opts *opts,
 567                                   struct callchain_param *param)
 568 {
 569         bool function = perf_evsel__is_function_event(evsel);
 570         struct perf_event_attr *attr = &evsel->attr;
 571
 572         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 573
 574         if (param->record_mode == CALLCHAIN_LBR) {
 575                 if (!opts->branch_stack) {
 576                         if (attr->exclude_user) {
 577                                 pr_warning("LBR callstack option is only available "
 578                                            "to get user callchain information. "
 579                                            "Falling back to framepointers.\n");
 580                         } else {
 581                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 582                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 583                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 584                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 585                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 586                         }
 587                 } else
 588                          pr_warning("Cannot use LBR callstack with branch stack. "
 589                                     "Falling back to framepointers.\n");
 590         }
 591
 592         if (param->record_mode == CALLCHAIN_DWARF) {
 593                 if (!function) {
 594                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 595                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 596                         attr->sample_regs_user = PERF_REGS_MASK;
 597                         attr->sample_stack_user = param->dump_size;
 598                         attr->exclude_callchain_user = 1;
 599                 } else {
 600                         pr_info("Cannot use DWARF unwind for function trace event,"
 601                                 " falling back to framepointers.\n");
 602                 }
 603         }
 604
 605         if (function) {
 606                 pr_info("Disabling user space callchains for function trace event.\n");
 607                 attr->exclude_callchain_user = 1;
 608         }
 609 }
 610
 611 static void
 612 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 613                             struct callchain_param *param)
 614 {
 615         struct perf_event_attr *attr = &evsel->attr;
 616
 617         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 618         if (param->record_mode == CALLCHAIN_LBR) {
 619                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 620                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 621                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 622         }
 623         if (param->record_mode == CALLCHAIN_DWARF) {
 624                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 625                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 626         }
 627 }
 628
 629 static void apply_config_terms(struct perf_evsel *evsel,
 630                                struct record_opts *opts)
 631 {
 632         struct perf_evsel_config_term *term;
 633         struct list_head *config_terms = &evsel->config_terms;
 634         struct perf_event_attr *attr = &evsel->attr;
 635         struct callchain_param param;
 636         u32 dump_size = 0;
 637         char *callgraph_buf = NULL;
 638
 639         /* callgraph default */
 640         param.record_mode = callchain_param.record_mode;
 641
 642         list_for_each_entry(term, config_terms, list) {
 643                 switch (term->type) {
 644                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 645                         attr->sample_period = term->val.period;
 646                         attr->freq = 0;
 647                         break;
 648                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 649                         attr->sample_freq = term->val.freq;
 650                         attr->freq = 1;
 651                         break;
 652                 case PERF_EVSEL__CONFIG_TERM_TIME:
 653                         if (term->val.time)
 654                                 perf_evsel__set_sample_bit(evsel, TIME);
 655                         else
 656                                 perf_evsel__reset_sample_bit(evsel, TIME);
 657                         break;
 658                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 659                         callgraph_buf = term->val.callgraph;
 660                         break;
 661                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 662                         dump_size = term->val.stack_user;
 663                         break;
 664                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 665                         /*
 666                          * attr->inherit should has already been set by
 667                          * perf_evsel__config. If user explicitly set
 668                          * inherit using config terms, override global
 669                          * opt->no_inherit setting.
 670                          */
 671                         attr->inherit = term->val.inherit ? 1 : 0;
 672                         break;
 673                 default:
 674                         break;
 675                 }
 676         }
 677
 678         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 679         if ((callgraph_buf != NULL) || (dump_size > 0)) {
 680
 681                 /* parse callgraph parameters */
 682                 if (callgraph_buf != NULL) {
 683                         if (!strcmp(callgraph_buf, "no")) {
 684                                 param.enabled = false;
 685                                 param.record_mode = CALLCHAIN_NONE;
 686                         } else {
 687                                 param.enabled = true;
 688                                 if (parse_callchain_record(callgraph_buf, &param)) {
 689                                         pr_err("per-event callgraph setting for %s failed. "
 690                                                "Apply callgraph global setting for it\n",
 691                                                evsel->name);
 692                                         return;
 693                                 }
 694                         }
 695                 }
 696                 if (dump_size > 0) {
 697                         dump_size = round_up(dump_size, sizeof(u64));
 698                         param.dump_size = dump_size;
 699                 }
 700
 701                 /* If global callgraph set, clear it */
 702                 if (callchain_param.enabled)
 703                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 704
 705                 /* set perf-event callgraph */
 706                 if (param.enabled)
 707                         perf_evsel__config_callchain(evsel, opts, &param);
 708         }
 709 }
 710
 711 /*
 712  * The enable_on_exec/disabled value strategy:
 713  *
 714  *  1) For any type of traced program:
 715  *    - all independent events and group leaders are disabled
 716  *    - all group members are enabled
 717  *
 718  *     Group members are ruled by group leaders. They need to
 719  *     be enabled, because the group scheduling relies on that.
 720  *
 721  *  2) For traced programs executed by perf:
 722  *     - all independent events and group leaders have
 723  *       enable_on_exec set
 724  *     - we don't specifically enable or disable any event during
 725  *       the record command
 726  *
 727  *     Independent events and group leaders are initially disabled
 728  *     and get enabled by exec. Group members are ruled by group
 729  *     leaders as stated in 1).
 730  *
 731  *  3) For traced programs attached by perf (pid/tid):
 732  *     - we specifically enable or disable all events during
 733  *       the record command
 734  *
 735  *     When attaching events to already running traced we
 736  *     enable/disable events specifically, as there's no
 737  *     initial traced exec call.
 738  */
 739 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 740                         struct callchain_param *callchain)
 741 {
 742         struct perf_evsel *leader = evsel->leader;
 743         struct perf_event_attr *attr = &evsel->attr;
 744         int track = evsel->tracking;
 745         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 746
 747         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 748         attr->inherit       = !opts->no_inherit;
 749
 750         perf_evsel__set_sample_bit(evsel, IP);
 751         perf_evsel__set_sample_bit(evsel, TID);
 752
 753         if (evsel->sample_read) {
 754                 perf_evsel__set_sample_bit(evsel, READ);
 755
 756                 /*
 757                  * We need ID even in case of single event, because
 758                  * PERF_SAMPLE_READ process ID specific data.
 759                  */
 760                 perf_evsel__set_sample_id(evsel, false);
 761
 762                 /*
 763                  * Apply group format only if we belong to group
 764                  * with more than one members.
 765                  */
 766                 if (leader->nr_members > 1) {
 767                         attr->read_format |= PERF_FORMAT_GROUP;
 768                         attr->inherit = 0;
 769                 }
 770         }
 771
 772         /*
 773          * We default some events to have a default interval. But keep
 774          * it a weak assumption overridable by the user.
 775          */
 776         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 777                                      opts->user_interval != ULLONG_MAX)) {
 778                 if (opts->freq) {
 779                         perf_evsel__set_sample_bit(evsel, PERIOD);
 780                         attr->freq              = 1;
 781                         attr->sample_freq       = opts->freq;
 782                 } else {
 783                         attr->sample_period = opts->default_interval;
 784                 }
 785         }
 786
 787         /*
 788          * Disable sampling for all group members other
 789          * than leader in case leader 'leads' the sampling.
 790          */
 791         if ((leader != evsel) && leader->sample_read) {
 792                 attr->sample_freq   = 0;
 793                 attr->sample_period = 0;
 794         }
 795
 796         if (opts->no_samples)
 797                 attr->sample_freq = 0;
 798
 799         if (opts->inherit_stat)
 800                 attr->inherit_stat = 1;
 801
 802         if (opts->sample_address) {
 803                 perf_evsel__set_sample_bit(evsel, ADDR);
 804                 attr->mmap_data = track;
 805         }
 806
 807         /*
 808          * We don't allow user space callchains for  function trace
 809          * event, due to issues with page faults while tracing page
 810          * fault handler and its overall trickiness nature.
 811          */
 812         if (perf_evsel__is_function_event(evsel))
 813                 evsel->attr.exclude_callchain_user = 1;
 814
 815         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 816                 perf_evsel__config_callchain(evsel, opts, callchain);
 817
 818         if (opts->sample_intr_regs) {
 819                 attr->sample_regs_intr = opts->sample_intr_regs;
 820                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 821         }
 822
 823         if (target__has_cpu(&opts->target))
 824                 perf_evsel__set_sample_bit(evsel, CPU);
 825
 826         if (opts->period)
 827                 perf_evsel__set_sample_bit(evsel, PERIOD);
 828
 829         /*
 830          * When the user explicitly disabled time don't force it here.
 831          */
 832         if (opts->sample_time &&
 833             (!perf_missing_features.sample_id_all &&
 834             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 835              opts->sample_time_set)))
 836                 perf_evsel__set_sample_bit(evsel, TIME);
 837
 838         if (opts->raw_samples && !evsel->no_aux_samples) {
 839                 perf_evsel__set_sample_bit(evsel, TIME);
 840                 perf_evsel__set_sample_bit(evsel, RAW);
 841                 perf_evsel__set_sample_bit(evsel, CPU);
 842         }
 843
 844         if (opts->sample_address)
 845                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 846
 847         if (opts->no_buffering) {
 848                 attr->watermark = 0;
 849                 attr->wakeup_events = 1;
 850         }
 851         if (opts->branch_stack && !evsel->no_aux_samples) {
 852                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 853                 attr->branch_sample_type = opts->branch_stack;
 854         }
 855
 856         if (opts->sample_weight)
 857                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 858
 859         attr->task  = track;
 860         attr->mmap  = track;
 861         attr->mmap2 = track && !perf_missing_features.mmap2;
 862         attr->comm  = track;
 863
 864         if (opts->record_switch_events)
 865                 attr->context_switch = track;
 866
 867         if (opts->sample_transaction)
 868                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
 869
 870         if (opts->running_time) {
 871                 evsel->attr.read_format |=
 872                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 873                         PERF_FORMAT_TOTAL_TIME_RUNNING;
 874         }
 875
 876         /*
 877          * XXX see the function comment above
 878          *
 879          * Disabling only independent events or group leaders,
 880          * keeping group members enabled.
 881          */
 882         if (perf_evsel__is_group_leader(evsel))
 883                 attr->disabled = 1;
 884
 885         /*
 886          * Setting enable_on_exec for independent events and
 887          * group leaders for traced executed by perf.
 888          */
 889         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
 890                 !opts->initial_delay)
 891                 attr->enable_on_exec = 1;
 892
 893         if (evsel->immediate) {
 894                 attr->disabled = 0;
 895                 attr->enable_on_exec = 0;
 896         }
 897
 898         clockid = opts->clockid;
 899         if (opts->use_clockid) {
 900                 attr->use_clockid = 1;
 901                 attr->clockid = opts->clockid;
 902         }
 903
 904         if (evsel->precise_max)
 905                 perf_event_attr__set_max_precise_ip(attr);
 906
 907         if (opts->all_user) {
 908                 attr->exclude_kernel = 1;
 909                 attr->exclude_user   = 0;
 910         }
 911
 912         if (opts->all_kernel) {
 913                 attr->exclude_kernel = 0;
 914                 attr->exclude_user   = 1;
 915         }
 916
 917         /*
 918          * Apply event specific term settings,
 919          * it overloads any global configuration.
 920          */
 921         apply_config_terms(evsel, opts);
 922 }
 923
 924 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 925 {
 926         int cpu, thread;
 927
 928         if (evsel->system_wide)
 929                 nthreads = 1;
 930
 931         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 932
 933         if (evsel->fd) {
 934                 for (cpu = 0; cpu < ncpus; cpu++) {
 935                         for (thread = 0; thread < nthreads; thread++) {
 936                                 FD(evsel, cpu, thread) = -1;
 937                         }
 938                 }
 939         }
 940
 941         return evsel->fd != NULL ? 0 : -ENOMEM;
 942 }
 943
 944 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 945                           int ioc,  void *arg)
 946 {
 947         int cpu, thread;
 948
 949         if (evsel->system_wide)
 950                 nthreads = 1;
 951
 952         for (cpu = 0; cpu < ncpus; cpu++) {
 953                 for (thread = 0; thread < nthreads; thread++) {
 954                         int fd = FD(evsel, cpu, thread),
 955                             err = ioctl(fd, ioc, arg);
 956
 957                         if (err)
 958                                 return err;
 959                 }
 960         }
 961
 962         return 0;
 963 }
 964
 965 int perf_evsel__apply_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 966                              const char *filter)
 967 {
 968         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 969                                      PERF_EVENT_IOC_SET_FILTER,
 970                                      (void *)filter);
 971 }
 972
 973 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
 974 {
 975         char *new_filter = strdup(filter);
 976
 977         if (new_filter != NULL) {
 978                 free(evsel->filter);
 979                 evsel->filter = new_filter;
 980                 return 0;
 981         }
 982
 983         return -1;
 984 }
 985
 986 int perf_evsel__append_filter(struct perf_evsel *evsel,
 987                               const char *op, const char *filter)
 988 {
 989         char *new_filter;
 990
 991         if (evsel->filter == NULL)
 992                 return perf_evsel__set_filter(evsel, filter);
 993
 994         if (asprintf(&new_filter,"(%s) %s (%s)", evsel->filter, op, filter) > 0) {
 995                 free(evsel->filter);
 996                 evsel->filter = new_filter;
 997                 return 0;
 998         }
 999
1000         return -1;
1001 }
1002
1003 int perf_evsel__enable(struct perf_evsel *evsel)
1004 {
1005         int nthreads = thread_map__nr(evsel->threads);
1006         int ncpus = cpu_map__nr(evsel->cpus);
1007
1008         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1009                                      PERF_EVENT_IOC_ENABLE,
1010                                      0);
1011 }
1012
1013 int perf_evsel__disable(struct perf_evsel *evsel)
1014 {
1015         int nthreads = thread_map__nr(evsel->threads);
1016         int ncpus = cpu_map__nr(evsel->cpus);
1017
1018         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
1019                                      PERF_EVENT_IOC_DISABLE,
1020                                      0);
1021 }
1022
1023 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1024 {
1025         if (ncpus == 0 || nthreads == 0)
1026                 return 0;
1027
1028         if (evsel->system_wide)
1029                 nthreads = 1;
1030
1031         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1032         if (evsel->sample_id == NULL)
1033                 return -ENOMEM;
1034
1035         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1036         if (evsel->id == NULL) {
1037                 xyarray__delete(evsel->sample_id);
1038                 evsel->sample_id = NULL;
1039                 return -ENOMEM;
1040         }
1041
1042         return 0;
1043 }
1044
1045 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1046 {
1047         xyarray__delete(evsel->fd);
1048         evsel->fd = NULL;
1049 }
1050
1051 static void perf_evsel__free_id(struct perf_evsel *evsel)
1052 {
1053         xyarray__delete(evsel->sample_id);
1054         evsel->sample_id = NULL;
1055         zfree(&evsel->id);
1056 }
1057
1058 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1059 {
1060         struct perf_evsel_config_term *term, *h;
1061
1062         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1063                 list_del(&term->list);
1064                 free(term);
1065         }
1066 }
1067
1068 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1069 {
1070         int cpu, thread;
1071
1072         if (evsel->system_wide)
1073                 nthreads = 1;
1074
1075         for (cpu = 0; cpu < ncpus; cpu++)
1076                 for (thread = 0; thread < nthreads; ++thread) {
1077                         close(FD(evsel, cpu, thread));
1078                         FD(evsel, cpu, thread) = -1;
1079                 }
1080 }
1081
1082 void perf_evsel__exit(struct perf_evsel *evsel)
1083 {
1084         assert(list_empty(&evsel->node));
1085         assert(evsel->evlist == NULL);
1086         perf_evsel__free_fd(evsel);
1087         perf_evsel__free_id(evsel);
1088         perf_evsel__free_config_terms(evsel);
1089         close_cgroup(evsel->cgrp);
1090         cpu_map__put(evsel->cpus);
1091         cpu_map__put(evsel->own_cpus);
1092         thread_map__put(evsel->threads);
1093         zfree(&evsel->group_name);
1094         zfree(&evsel->name);
1095         perf_evsel__object.fini(evsel);
1096 }
1097
1098 void perf_evsel__delete(struct perf_evsel *evsel)
1099 {
1100         perf_evsel__exit(evsel);
1101         free(evsel);
1102 }
1103
1104 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1105                                 struct perf_counts_values *count)
1106 {
1107         struct perf_counts_values tmp;
1108
1109         if (!evsel->prev_raw_counts)
1110                 return;
1111
1112         if (cpu == -1) {
1113                 tmp = evsel->prev_raw_counts->aggr;
1114                 evsel->prev_raw_counts->aggr = *count;
1115         } else {
1116                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1117                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1118         }
1119
1120         count->val = count->val - tmp.val;
1121         count->ena = count->ena - tmp.ena;
1122         count->run = count->run - tmp.run;
1123 }
1124
1125 void perf_counts_values__scale(struct perf_counts_values *count,
1126                                bool scale, s8 *pscaled)
1127 {
1128         s8 scaled = 0;
1129
1130         if (scale) {
1131                 if (count->run == 0) {
1132                         scaled = -1;
1133                         count->val = 0;
1134                 } else if (count->run < count->ena) {
1135                         scaled = 1;
1136                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1137                 }
1138         } else
1139                 count->ena = count->run = 0;
1140
1141         if (pscaled)
1142                 *pscaled = scaled;
1143 }
1144
1145 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1146                      struct perf_counts_values *count)
1147 {
1148         memset(count, 0, sizeof(*count));
1149
1150         if (FD(evsel, cpu, thread) < 0)
1151                 return -EINVAL;
1152
1153         if (readn(FD(evsel, cpu, thread), count, sizeof(*count)) < 0)
1154                 return -errno;
1155
1156         return 0;
1157 }
1158
1159 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1160                               int cpu, int thread, bool scale)
1161 {
1162         struct perf_counts_values count;
1163         size_t nv = scale ? 3 : 1;
1164
1165         if (FD(evsel, cpu, thread) < 0)
1166                 return -EINVAL;
1167
1168         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1169                 return -ENOMEM;
1170
1171         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
1172                 return -errno;
1173
1174         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1175         perf_counts_values__scale(&count, scale, NULL);
1176         *perf_counts(evsel->counts, cpu, thread) = count;
1177         return 0;
1178 }
1179
1180 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1181 {
1182         struct perf_evsel *leader = evsel->leader;
1183         int fd;
1184
1185         if (perf_evsel__is_group_leader(evsel))
1186                 return -1;
1187
1188         /*
1189          * Leader must be already processed/open,
1190          * if not it's a bug.
1191          */
1192         BUG_ON(!leader->fd);
1193
1194         fd = FD(leader, cpu, thread);
1195         BUG_ON(fd == -1);
1196
1197         return fd;
1198 }
1199
1200 struct bit_names {
1201         int bit;
1202         const char *name;
1203 };
1204
1205 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1206 {
1207         bool first_bit = true;
1208         int i = 0;
1209
1210         do {
1211                 if (value & bits[i].bit) {
1212                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1213                         first_bit = false;
1214                 }
1215         } while (bits[++i].name != NULL);
1216 }
1217
1218 static void __p_sample_type(char *buf, size_t size, u64 value)
1219 {
1220 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1221         struct bit_names bits[] = {
1222                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1223                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1224                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1225                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1226                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1227                 bit_name(WEIGHT),
1228                 { .name = NULL, }
1229         };
1230 #undef bit_name
1231         __p_bits(buf, size, value, bits);
1232 }
1233
1234 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1235 {
1236 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1237         struct bit_names bits[] = {
1238                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1239                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1240                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1241                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1242                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1243                 { .name = NULL, }
1244         };
1245 #undef bit_name
1246         __p_bits(buf, size, value, bits);
1247 }
1248
1249 static void __p_read_format(char *buf, size_t size, u64 value)
1250 {
1251 #define bit_name(n) { PERF_FORMAT_##n, #n }
1252         struct bit_names bits[] = {
1253                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1254                 bit_name(ID), bit_name(GROUP),
1255                 { .name = NULL, }
1256         };
1257 #undef bit_name
1258         __p_bits(buf, size, value, bits);
1259 }
1260
1261 #define BUF_SIZE                1024
1262
1263 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1264 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1265 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1266 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1267 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1268 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1269
1270 #define PRINT_ATTRn(_n, _f, _p)                         \
1271 do {                                                    \
1272         if (attr->_f) {                                 \
1273                 _p(attr->_f);                           \
1274                 ret += attr__fprintf(fp, _n, buf, priv);\
1275         }                                               \
1276 } while (0)
1277
1278 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1279
1280 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1281                              attr__fprintf_f attr__fprintf, void *priv)
1282 {
1283         char buf[BUF_SIZE];
1284         int ret = 0;
1285
1286         PRINT_ATTRf(type, p_unsigned);
1287         PRINT_ATTRf(size, p_unsigned);
1288         PRINT_ATTRf(config, p_hex);
1289         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1290         PRINT_ATTRf(sample_type, p_sample_type);
1291         PRINT_ATTRf(read_format, p_read_format);
1292
1293         PRINT_ATTRf(disabled, p_unsigned);
1294         PRINT_ATTRf(inherit, p_unsigned);
1295         PRINT_ATTRf(pinned, p_unsigned);
1296         PRINT_ATTRf(exclusive, p_unsigned);
1297         PRINT_ATTRf(exclude_user, p_unsigned);
1298         PRINT_ATTRf(exclude_kernel, p_unsigned);
1299         PRINT_ATTRf(exclude_hv, p_unsigned);
1300         PRINT_ATTRf(exclude_idle, p_unsigned);
1301         PRINT_ATTRf(mmap, p_unsigned);
1302         PRINT_ATTRf(comm, p_unsigned);
1303         PRINT_ATTRf(freq, p_unsigned);
1304         PRINT_ATTRf(inherit_stat, p_unsigned);
1305         PRINT_ATTRf(enable_on_exec, p_unsigned);
1306         PRINT_ATTRf(task, p_unsigned);
1307         PRINT_ATTRf(watermark, p_unsigned);
1308         PRINT_ATTRf(precise_ip, p_unsigned);
1309         PRINT_ATTRf(mmap_data, p_unsigned);
1310         PRINT_ATTRf(sample_id_all, p_unsigned);
1311         PRINT_ATTRf(exclude_host, p_unsigned);
1312         PRINT_ATTRf(exclude_guest, p_unsigned);
1313         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1314         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1315         PRINT_ATTRf(mmap2, p_unsigned);
1316         PRINT_ATTRf(comm_exec, p_unsigned);
1317         PRINT_ATTRf(use_clockid, p_unsigned);
1318         PRINT_ATTRf(context_switch, p_unsigned);
1319         PRINT_ATTRf(write_backward, p_unsigned);
1320
1321         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1322         PRINT_ATTRf(bp_type, p_unsigned);
1323         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1324         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1325         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1326         PRINT_ATTRf(sample_regs_user, p_hex);
1327         PRINT_ATTRf(sample_stack_user, p_unsigned);
1328         PRINT_ATTRf(clockid, p_signed);
1329         PRINT_ATTRf(sample_regs_intr, p_hex);
1330         PRINT_ATTRf(aux_watermark, p_unsigned);
1331
1332         return ret;
1333 }
1334
1335 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1336                                 void *priv __attribute__((unused)))
1337 {
1338         return fprintf(fp, "  %-32s %s\n", name, val);
1339 }
1340
1341 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1342                               struct thread_map *threads)
1343 {
1344         int cpu, thread, nthreads;
1345         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1346         int pid = -1, err;
1347         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1348
1349         if (evsel->system_wide)
1350                 nthreads = 1;
1351         else
1352                 nthreads = threads->nr;
1353
1354         if (evsel->fd == NULL &&
1355             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1356                 return -ENOMEM;
1357
1358         if (evsel->cgrp) {
1359                 flags |= PERF_FLAG_PID_CGROUP;
1360                 pid = evsel->cgrp->fd;
1361         }
1362
1363 fallback_missing_features:
1364         if (perf_missing_features.clockid_wrong)
1365                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1366         if (perf_missing_features.clockid) {
1367                 evsel->attr.use_clockid = 0;
1368                 evsel->attr.clockid = 0;
1369         }
1370         if (perf_missing_features.cloexec)
1371                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1372         if (perf_missing_features.mmap2)
1373                 evsel->attr.mmap2 = 0;
1374         if (perf_missing_features.exclude_guest)
1375                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1376         if (perf_missing_features.lbr_flags)
1377                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1378                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1379 retry_sample_id:
1380         if (perf_missing_features.sample_id_all)
1381                 evsel->attr.sample_id_all = 0;
1382
1383         if (verbose >= 2) {
1384                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1385                 fprintf(stderr, "perf_event_attr:\n");
1386                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1387                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1388         }
1389
1390         for (cpu = 0; cpu < cpus->nr; cpu++) {
1391
1392                 for (thread = 0; thread < nthreads; thread++) {
1393                         int group_fd;
1394
1395                         if (!evsel->cgrp && !evsel->system_wide)
1396                                 pid = thread_map__pid(threads, thread);
1397
1398                         group_fd = get_group_fd(evsel, cpu, thread);
1399 retry_open:
1400                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
1401                                   pid, cpus->map[cpu], group_fd, flags);
1402
1403                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
1404                                                                      pid,
1405                                                                      cpus->map[cpu],
1406                                                                      group_fd, flags);
1407                         if (FD(evsel, cpu, thread) < 0) {
1408                                 err = -errno;
1409                                 pr_debug2("sys_perf_event_open failed, error %d\n",
1410                                           err);
1411                                 goto try_fallback;
1412                         }
1413
1414                         if (evsel->bpf_fd >= 0) {
1415                                 int evt_fd = FD(evsel, cpu, thread);
1416                                 int bpf_fd = evsel->bpf_fd;
1417
1418                                 err = ioctl(evt_fd,
1419                                             PERF_EVENT_IOC_SET_BPF,
1420                                             bpf_fd);
1421                                 if (err && errno != EEXIST) {
1422                                         pr_err("failed to attach bpf fd %d: %s\n",
1423                                                bpf_fd, strerror(errno));
1424                                         err = -EINVAL;
1425                                         goto out_close;
1426                                 }
1427                         }
1428
1429                         set_rlimit = NO_CHANGE;
1430
1431                         /*
1432                          * If we succeeded but had to kill clockid, fail and
1433                          * have perf_evsel__open_strerror() print us a nice
1434                          * error.
1435                          */
1436                         if (perf_missing_features.clockid ||
1437                             perf_missing_features.clockid_wrong) {
1438                                 err = -EINVAL;
1439                                 goto out_close;
1440                         }
1441                 }
1442         }
1443
1444         return 0;
1445
1446 try_fallback:
1447         /*
1448          * perf stat needs between 5 and 22 fds per CPU. When we run out
1449          * of them try to increase the limits.
1450          */
1451         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1452                 struct rlimit l;
1453                 int old_errno = errno;
1454
1455                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1456                         if (set_rlimit == NO_CHANGE)
1457                                 l.rlim_cur = l.rlim_max;
1458                         else {
1459                                 l.rlim_cur = l.rlim_max + 1000;
1460                                 l.rlim_max = l.rlim_cur;
1461                         }
1462                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1463                                 set_rlimit++;
1464                                 errno = old_errno;
1465                                 goto retry_open;
1466                         }
1467                 }
1468                 errno = old_errno;
1469         }
1470
1471         if (err != -EINVAL || cpu > 0 || thread > 0)
1472                 goto out_close;
1473
1474         /*
1475          * Must probe features in the order they were added to the
1476          * perf_event_attr interface.
1477          */
1478         if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1479                 perf_missing_features.clockid_wrong = true;
1480                 goto fallback_missing_features;
1481         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1482                 perf_missing_features.clockid = true;
1483                 goto fallback_missing_features;
1484         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1485                 perf_missing_features.cloexec = true;
1486                 goto fallback_missing_features;
1487         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1488                 perf_missing_features.mmap2 = true;
1489                 goto fallback_missing_features;
1490         } else if (!perf_missing_features.exclude_guest &&
1491                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1492                 perf_missing_features.exclude_guest = true;
1493                 goto fallback_missing_features;
1494         } else if (!perf_missing_features.sample_id_all) {
1495                 perf_missing_features.sample_id_all = true;
1496                 goto retry_sample_id;
1497         } else if (!perf_missing_features.lbr_flags &&
1498                         (evsel->attr.branch_sample_type &
1499                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1500                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1501                 perf_missing_features.lbr_flags = true;
1502                 goto fallback_missing_features;
1503         }
1504
1505 out_close:
1506         do {
1507                 while (--thread >= 0) {
1508                         close(FD(evsel, cpu, thread));
1509                         FD(evsel, cpu, thread) = -1;
1510                 }
1511                 thread = nthreads;
1512         } while (--cpu >= 0);
1513         return err;
1514 }
1515
1516 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1517 {
1518         if (evsel->fd == NULL)
1519                 return;
1520
1521         perf_evsel__close_fd(evsel, ncpus, nthreads);
1522         perf_evsel__free_fd(evsel);
1523 }
1524
1525 static struct {
1526         struct cpu_map map;
1527         int cpus[1];
1528 } empty_cpu_map = {
1529         .map.nr = 1,
1530         .cpus   = { -1, },
1531 };
1532
1533 static struct {
1534         struct thread_map map;
1535         int threads[1];
1536 } empty_thread_map = {
1537         .map.nr  = 1,
1538         .threads = { -1, },
1539 };
1540
1541 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1542                      struct thread_map *threads)
1543 {
1544         if (cpus == NULL) {
1545                 /* Work around old compiler warnings about strict aliasing */
1546                 cpus = &empty_cpu_map.map;
1547         }
1548
1549         if (threads == NULL)
1550                 threads = &empty_thread_map.map;
1551
1552         return __perf_evsel__open(evsel, cpus, threads);
1553 }
1554
1555 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1556                              struct cpu_map *cpus)
1557 {
1558         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1559 }
1560
1561 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1562                                 struct thread_map *threads)
1563 {
1564         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1565 }
1566
1567 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1568                                        const union perf_event *event,
1569                                        struct perf_sample *sample)
1570 {
1571         u64 type = evsel->attr.sample_type;
1572         const u64 *array = event->sample.array;
1573         bool swapped = evsel->needs_swap;
1574         union u64_swap u;
1575
1576         array += ((event->header.size -
1577                    sizeof(event->header)) / sizeof(u64)) - 1;
1578
1579         if (type & PERF_SAMPLE_IDENTIFIER) {
1580                 sample->id = *array;
1581                 array--;
1582         }
1583
1584         if (type & PERF_SAMPLE_CPU) {
1585                 u.val64 = *array;
1586                 if (swapped) {
1587                         /* undo swap of u64, then swap on individual u32s */
1588                         u.val64 = bswap_64(u.val64);
1589                         u.val32[0] = bswap_32(u.val32[0]);
1590                 }
1591
1592                 sample->cpu = u.val32[0];
1593                 array--;
1594         }
1595
1596         if (type & PERF_SAMPLE_STREAM_ID) {
1597                 sample->stream_id = *array;
1598                 array--;
1599         }
1600
1601         if (type & PERF_SAMPLE_ID) {
1602                 sample->id = *array;
1603                 array--;
1604         }
1605
1606         if (type & PERF_SAMPLE_TIME) {
1607                 sample->time = *array;
1608                 array--;
1609         }
1610
1611         if (type & PERF_SAMPLE_TID) {
1612                 u.val64 = *array;
1613                 if (swapped) {
1614                         /* undo swap of u64, then swap on individual u32s */
1615                         u.val64 = bswap_64(u.val64);
1616                         u.val32[0] = bswap_32(u.val32[0]);
1617                         u.val32[1] = bswap_32(u.val32[1]);
1618                 }
1619
1620                 sample->pid = u.val32[0];
1621                 sample->tid = u.val32[1];
1622                 array--;
1623         }
1624
1625         return 0;
1626 }
1627
1628 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1629                             u64 size)
1630 {
1631         return size > max_size || offset + size > endp;
1632 }
1633
1634 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1635         do {                                                            \
1636                 if (overflow(endp, (max_size), (offset), (size)))       \
1637                         return -EFAULT;                                 \
1638         } while (0)
1639
1640 #define OVERFLOW_CHECK_u64(offset) \
1641         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1642
1643 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1644                              struct perf_sample *data)
1645 {
1646         u64 type = evsel->attr.sample_type;
1647         bool swapped = evsel->needs_swap;
1648         const u64 *array;
1649         u16 max_size = event->header.size;
1650         const void *endp = (void *)event + max_size;
1651         u64 sz;
1652
1653         /*
1654          * used for cross-endian analysis. See git commit 65014ab3
1655          * for why this goofiness is needed.
1656          */
1657         union u64_swap u;
1658
1659         memset(data, 0, sizeof(*data));
1660         data->cpu = data->pid = data->tid = -1;
1661         data->stream_id = data->id = data->time = -1ULL;
1662         data->period = evsel->attr.sample_period;
1663         data->weight = 0;
1664         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1665
1666         if (event->header.type != PERF_RECORD_SAMPLE) {
1667                 if (!evsel->attr.sample_id_all)
1668                         return 0;
1669                 return perf_evsel__parse_id_sample(evsel, event, data);
1670         }
1671
1672         array = event->sample.array;
1673
1674         /*
1675          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1676          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1677          * check the format does not go past the end of the event.
1678          */
1679         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1680                 return -EFAULT;
1681
1682         data->id = -1ULL;
1683         if (type & PERF_SAMPLE_IDENTIFIER) {
1684                 data->id = *array;
1685                 array++;
1686         }
1687
1688         if (type & PERF_SAMPLE_IP) {
1689                 data->ip = *array;
1690                 array++;
1691         }
1692
1693         if (type & PERF_SAMPLE_TID) {
1694                 u.val64 = *array;
1695                 if (swapped) {
1696                         /* undo swap of u64, then swap on individual u32s */
1697                         u.val64 = bswap_64(u.val64);
1698                         u.val32[0] = bswap_32(u.val32[0]);
1699                         u.val32[1] = bswap_32(u.val32[1]);
1700                 }
1701
1702                 data->pid = u.val32[0];
1703                 data->tid = u.val32[1];
1704                 array++;
1705         }
1706
1707         if (type & PERF_SAMPLE_TIME) {
1708                 data->time = *array;
1709                 array++;
1710         }
1711
1712         data->addr = 0;
1713         if (type & PERF_SAMPLE_ADDR) {
1714                 data->addr = *array;
1715                 array++;
1716         }
1717
1718         if (type & PERF_SAMPLE_ID) {
1719                 data->id = *array;
1720                 array++;
1721         }
1722
1723         if (type & PERF_SAMPLE_STREAM_ID) {
1724                 data->stream_id = *array;
1725                 array++;
1726         }
1727
1728         if (type & PERF_SAMPLE_CPU) {
1729
1730                 u.val64 = *array;
1731                 if (swapped) {
1732                         /* undo swap of u64, then swap on individual u32s */
1733                         u.val64 = bswap_64(u.val64);
1734                         u.val32[0] = bswap_32(u.val32[0]);
1735                 }
1736
1737                 data->cpu = u.val32[0];
1738                 array++;
1739         }
1740
1741         if (type & PERF_SAMPLE_PERIOD) {
1742                 data->period = *array;
1743                 array++;
1744         }
1745
1746         if (type & PERF_SAMPLE_READ) {
1747                 u64 read_format = evsel->attr.read_format;
1748
1749                 OVERFLOW_CHECK_u64(array);
1750                 if (read_format & PERF_FORMAT_GROUP)
1751                         data->read.group.nr = *array;
1752                 else
1753                         data->read.one.value = *array;
1754
1755                 array++;
1756
1757                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1758                         OVERFLOW_CHECK_u64(array);
1759                         data->read.time_enabled = *array;
1760                         array++;
1761                 }
1762
1763                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1764                         OVERFLOW_CHECK_u64(array);
1765                         data->read.time_running = *array;
1766                         array++;
1767                 }
1768
1769                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1770                 if (read_format & PERF_FORMAT_GROUP) {
1771                         const u64 max_group_nr = UINT64_MAX /
1772                                         sizeof(struct sample_read_value);
1773
1774                         if (data->read.group.nr > max_group_nr)
1775                                 return -EFAULT;
1776                         sz = data->read.group.nr *
1777                              sizeof(struct sample_read_value);
1778                         OVERFLOW_CHECK(array, sz, max_size);
1779                         data->read.group.values =
1780                                         (struct sample_read_value *)array;
1781                         array = (void *)array + sz;
1782                 } else {
1783                         OVERFLOW_CHECK_u64(array);
1784                         data->read.one.id = *array;
1785                         array++;
1786                 }
1787         }
1788
1789         if (type & PERF_SAMPLE_CALLCHAIN) {
1790                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1791
1792                 OVERFLOW_CHECK_u64(array);
1793                 data->callchain = (struct ip_callchain *)array++;
1794                 if (data->callchain->nr > max_callchain_nr)
1795                         return -EFAULT;
1796                 sz = data->callchain->nr * sizeof(u64);
1797                 OVERFLOW_CHECK(array, sz, max_size);
1798                 array = (void *)array + sz;
1799         }
1800
1801         if (type & PERF_SAMPLE_RAW) {
1802                 OVERFLOW_CHECK_u64(array);
1803                 u.val64 = *array;
1804                 if (WARN_ONCE(swapped,
1805                               "Endianness of raw data not corrected!\n")) {
1806                         /* undo swap of u64, then swap on individual u32s */
1807                         u.val64 = bswap_64(u.val64);
1808                         u.val32[0] = bswap_32(u.val32[0]);
1809                         u.val32[1] = bswap_32(u.val32[1]);
1810                 }
1811                 data->raw_size = u.val32[0];
1812                 array = (void *)array + sizeof(u32);
1813
1814                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1815                 data->raw_data = (void *)array;
1816                 array = (void *)array + data->raw_size;
1817         }
1818
1819         if (type & PERF_SAMPLE_BRANCH_STACK) {
1820                 const u64 max_branch_nr = UINT64_MAX /
1821                                           sizeof(struct branch_entry);
1822
1823                 OVERFLOW_CHECK_u64(array);
1824                 data->branch_stack = (struct branch_stack *)array++;
1825
1826                 if (data->branch_stack->nr > max_branch_nr)
1827                         return -EFAULT;
1828                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1829                 OVERFLOW_CHECK(array, sz, max_size);
1830                 array = (void *)array + sz;
1831         }
1832
1833         if (type & PERF_SAMPLE_REGS_USER) {
1834                 OVERFLOW_CHECK_u64(array);
1835                 data->user_regs.abi = *array;
1836                 array++;
1837
1838                 if (data->user_regs.abi) {
1839                         u64 mask = evsel->attr.sample_regs_user;
1840
1841                         sz = hweight_long(mask) * sizeof(u64);
1842                         OVERFLOW_CHECK(array, sz, max_size);
1843                         data->user_regs.mask = mask;
1844                         data->user_regs.regs = (u64 *)array;
1845                         array = (void *)array + sz;
1846                 }
1847         }
1848
1849         if (type & PERF_SAMPLE_STACK_USER) {
1850                 OVERFLOW_CHECK_u64(array);
1851                 sz = *array++;
1852
1853                 data->user_stack.offset = ((char *)(array - 1)
1854                                           - (char *) event);
1855
1856                 if (!sz) {
1857                         data->user_stack.size = 0;
1858                 } else {
1859                         OVERFLOW_CHECK(array, sz, max_size);
1860                         data->user_stack.data = (char *)array;
1861                         array = (void *)array + sz;
1862                         OVERFLOW_CHECK_u64(array);
1863                         data->user_stack.size = *array++;
1864                         if (WARN_ONCE(data->user_stack.size > sz,
1865                                       "user stack dump failure\n"))
1866                                 return -EFAULT;
1867                 }
1868         }
1869
1870         data->weight = 0;
1871         if (type & PERF_SAMPLE_WEIGHT) {
1872                 OVERFLOW_CHECK_u64(array);
1873                 data->weight = *array;
1874                 array++;
1875         }
1876
1877         data->data_src = PERF_MEM_DATA_SRC_NONE;
1878         if (type & PERF_SAMPLE_DATA_SRC) {
1879                 OVERFLOW_CHECK_u64(array);
1880                 data->data_src = *array;
1881                 array++;
1882         }
1883
1884         data->transaction = 0;
1885         if (type & PERF_SAMPLE_TRANSACTION) {
1886                 OVERFLOW_CHECK_u64(array);
1887                 data->transaction = *array;
1888                 array++;
1889         }
1890
1891         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
1892         if (type & PERF_SAMPLE_REGS_INTR) {
1893                 OVERFLOW_CHECK_u64(array);
1894                 data->intr_regs.abi = *array;
1895                 array++;
1896
1897                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
1898                         u64 mask = evsel->attr.sample_regs_intr;
1899
1900                         sz = hweight_long(mask) * sizeof(u64);
1901                         OVERFLOW_CHECK(array, sz, max_size);
1902                         data->intr_regs.mask = mask;
1903                         data->intr_regs.regs = (u64 *)array;
1904                         array = (void *)array + sz;
1905                 }
1906         }
1907
1908         return 0;
1909 }
1910
1911 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
1912                                      u64 read_format)
1913 {
1914         size_t sz, result = sizeof(struct sample_event);
1915
1916         if (type & PERF_SAMPLE_IDENTIFIER)
1917                 result += sizeof(u64);
1918
1919         if (type & PERF_SAMPLE_IP)
1920                 result += sizeof(u64);
1921
1922         if (type & PERF_SAMPLE_TID)
1923                 result += sizeof(u64);
1924
1925         if (type & PERF_SAMPLE_TIME)
1926                 result += sizeof(u64);
1927
1928         if (type & PERF_SAMPLE_ADDR)
1929                 result += sizeof(u64);
1930
1931         if (type & PERF_SAMPLE_ID)
1932                 result += sizeof(u64);
1933
1934         if (type & PERF_SAMPLE_STREAM_ID)
1935                 result += sizeof(u64);
1936
1937         if (type & PERF_SAMPLE_CPU)
1938                 result += sizeof(u64);
1939
1940         if (type & PERF_SAMPLE_PERIOD)
1941                 result += sizeof(u64);
1942
1943         if (type & PERF_SAMPLE_READ) {
1944                 result += sizeof(u64);
1945                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1946                         result += sizeof(u64);
1947                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1948                         result += sizeof(u64);
1949                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1950                 if (read_format & PERF_FORMAT_GROUP) {
1951                         sz = sample->read.group.nr *
1952                              sizeof(struct sample_read_value);
1953                         result += sz;
1954                 } else {
1955                         result += sizeof(u64);
1956                 }
1957         }
1958
1959         if (type & PERF_SAMPLE_CALLCHAIN) {
1960                 sz = (sample->callchain->nr + 1) * sizeof(u64);
1961                 result += sz;
1962         }
1963
1964         if (type & PERF_SAMPLE_RAW) {
1965                 result += sizeof(u32);
1966                 result += sample->raw_size;
1967         }
1968
1969         if (type & PERF_SAMPLE_BRANCH_STACK) {
1970                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
1971                 sz += sizeof(u64);
1972                 result += sz;
1973         }
1974
1975         if (type & PERF_SAMPLE_REGS_USER) {
1976                 if (sample->user_regs.abi) {
1977                         result += sizeof(u64);
1978                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
1979                         result += sz;
1980                 } else {
1981                         result += sizeof(u64);
1982                 }
1983         }
1984
1985         if (type & PERF_SAMPLE_STACK_USER) {
1986                 sz = sample->user_stack.size;
1987                 result += sizeof(u64);
1988                 if (sz) {
1989                         result += sz;
1990                         result += sizeof(u64);
1991                 }
1992         }
1993
1994         if (type & PERF_SAMPLE_WEIGHT)
1995                 result += sizeof(u64);
1996
1997         if (type & PERF_SAMPLE_DATA_SRC)
1998                 result += sizeof(u64);
1999
2000         if (type & PERF_SAMPLE_TRANSACTION)
2001                 result += sizeof(u64);
2002
2003         if (type & PERF_SAMPLE_REGS_INTR) {
2004                 if (sample->intr_regs.abi) {
2005                         result += sizeof(u64);
2006                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2007                         result += sz;
2008                 } else {
2009                         result += sizeof(u64);
2010                 }
2011         }
2012
2013         return result;
2014 }
2015
2016 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2017                                   u64 read_format,
2018                                   const struct perf_sample *sample,
2019                                   bool swapped)
2020 {
2021         u64 *array;
2022         size_t sz;
2023         /*
2024          * used for cross-endian analysis. See git commit 65014ab3
2025          * for why this goofiness is needed.
2026          */
2027         union u64_swap u;
2028
2029         array = event->sample.array;
2030
2031         if (type & PERF_SAMPLE_IDENTIFIER) {
2032                 *array = sample->id;
2033                 array++;
2034         }
2035
2036         if (type & PERF_SAMPLE_IP) {
2037                 *array = sample->ip;
2038                 array++;
2039         }
2040
2041         if (type & PERF_SAMPLE_TID) {
2042                 u.val32[0] = sample->pid;
2043                 u.val32[1] = sample->tid;
2044                 if (swapped) {
2045                         /*
2046                          * Inverse of what is done in perf_evsel__parse_sample
2047                          */
2048                         u.val32[0] = bswap_32(u.val32[0]);
2049                         u.val32[1] = bswap_32(u.val32[1]);
2050                         u.val64 = bswap_64(u.val64);
2051                 }
2052
2053                 *array = u.val64;
2054                 array++;
2055         }
2056
2057         if (type & PERF_SAMPLE_TIME) {
2058                 *array = sample->time;
2059                 array++;
2060         }
2061
2062         if (type & PERF_SAMPLE_ADDR) {
2063                 *array = sample->addr;
2064                 array++;
2065         }
2066
2067         if (type & PERF_SAMPLE_ID) {
2068                 *array = sample->id;
2069                 array++;
2070         }
2071
2072         if (type & PERF_SAMPLE_STREAM_ID) {
2073                 *array = sample->stream_id;
2074                 array++;
2075         }
2076
2077         if (type & PERF_SAMPLE_CPU) {
2078                 u.val32[0] = sample->cpu;
2079                 if (swapped) {
2080                         /*
2081                          * Inverse of what is done in perf_evsel__parse_sample
2082                          */
2083                         u.val32[0] = bswap_32(u.val32[0]);
2084                         u.val64 = bswap_64(u.val64);
2085                 }
2086                 *array = u.val64;
2087                 array++;
2088         }
2089
2090         if (type & PERF_SAMPLE_PERIOD) {
2091                 *array = sample->period;
2092                 array++;
2093         }
2094
2095         if (type & PERF_SAMPLE_READ) {
2096                 if (read_format & PERF_FORMAT_GROUP)
2097                         *array = sample->read.group.nr;
2098                 else
2099                         *array = sample->read.one.value;
2100                 array++;
2101
2102                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2103                         *array = sample->read.time_enabled;
2104                         array++;
2105                 }
2106
2107                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2108                         *array = sample->read.time_running;
2109                         array++;
2110                 }
2111
2112                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2113                 if (read_format & PERF_FORMAT_GROUP) {
2114                         sz = sample->read.group.nr *
2115                              sizeof(struct sample_read_value);
2116                         memcpy(array, sample->read.group.values, sz);
2117                         array = (void *)array + sz;
2118                 } else {
2119                         *array = sample->read.one.id;
2120                         array++;
2121                 }
2122         }
2123
2124         if (type & PERF_SAMPLE_CALLCHAIN) {
2125                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2126                 memcpy(array, sample->callchain, sz);
2127                 array = (void *)array + sz;
2128         }
2129
2130         if (type & PERF_SAMPLE_RAW) {
2131                 u.val32[0] = sample->raw_size;
2132                 if (WARN_ONCE(swapped,
2133                               "Endianness of raw data not corrected!\n")) {
2134                         /*
2135                          * Inverse of what is done in perf_evsel__parse_sample
2136                          */
2137                         u.val32[0] = bswap_32(u.val32[0]);
2138                         u.val32[1] = bswap_32(u.val32[1]);
2139                         u.val64 = bswap_64(u.val64);
2140                 }
2141                 *array = u.val64;
2142                 array = (void *)array + sizeof(u32);
2143
2144                 memcpy(array, sample->raw_data, sample->raw_size);
2145                 array = (void *)array + sample->raw_size;
2146         }
2147
2148         if (type & PERF_SAMPLE_BRANCH_STACK) {
2149                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2150                 sz += sizeof(u64);
2151                 memcpy(array, sample->branch_stack, sz);
2152                 array = (void *)array + sz;
2153         }
2154
2155         if (type & PERF_SAMPLE_REGS_USER) {
2156                 if (sample->user_regs.abi) {
2157                         *array++ = sample->user_regs.abi;
2158                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2159                         memcpy(array, sample->user_regs.regs, sz);
2160                         array = (void *)array + sz;
2161                 } else {
2162                         *array++ = 0;
2163                 }
2164         }
2165
2166         if (type & PERF_SAMPLE_STACK_USER) {
2167                 sz = sample->user_stack.size;
2168                 *array++ = sz;
2169                 if (sz) {
2170                         memcpy(array, sample->user_stack.data, sz);
2171                         array = (void *)array + sz;
2172                         *array++ = sz;
2173                 }
2174         }
2175
2176         if (type & PERF_SAMPLE_WEIGHT) {
2177                 *array = sample->weight;
2178                 array++;
2179         }
2180
2181         if (type & PERF_SAMPLE_DATA_SRC) {
2182                 *array = sample->data_src;
2183                 array++;
2184         }
2185
2186         if (type & PERF_SAMPLE_TRANSACTION) {
2187                 *array = sample->transaction;
2188                 array++;
2189         }
2190
2191         if (type & PERF_SAMPLE_REGS_INTR) {
2192                 if (sample->intr_regs.abi) {
2193                         *array++ = sample->intr_regs.abi;
2194                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2195                         memcpy(array, sample->intr_regs.regs, sz);
2196                         array = (void *)array + sz;
2197                 } else {
2198                         *array++ = 0;
2199                 }
2200         }
2201
2202         return 0;
2203 }
2204
2205 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2206 {
2207         return pevent_find_field(evsel->tp_format, name);
2208 }
2209
2210 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2211                          const char *name)
2212 {
2213         struct format_field *field = perf_evsel__field(evsel, name);
2214         int offset;
2215
2216         if (!field)
2217                 return NULL;
2218
2219         offset = field->offset;
2220
2221         if (field->flags & FIELD_IS_DYNAMIC) {
2222                 offset = *(int *)(sample->raw_data + field->offset);
2223                 offset &= 0xffff;
2224         }
2225
2226         return sample->raw_data + offset;
2227 }
2228
2229 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2230                        const char *name)
2231 {
2232         struct format_field *field = perf_evsel__field(evsel, name);
2233         void *ptr;
2234         u64 value;
2235
2236         if (!field)
2237                 return 0;
2238
2239         ptr = sample->raw_data + field->offset;
2240
2241         switch (field->size) {
2242         case 1:
2243                 return *(u8 *)ptr;
2244         case 2:
2245                 value = *(u16 *)ptr;
2246                 break;
2247         case 4:
2248                 value = *(u32 *)ptr;
2249                 break;
2250         case 8:
2251                 memcpy(&value, ptr, sizeof(u64));
2252                 break;
2253         default:
2254                 return 0;
2255         }
2256
2257         if (!evsel->needs_swap)
2258                 return value;
2259
2260         switch (field->size) {
2261         case 2:
2262                 return bswap_16(value);
2263         case 4:
2264                 return bswap_32(value);
2265         case 8:
2266                 return bswap_64(value);
2267         default:
2268                 return 0;
2269         }
2270
2271         return 0;
2272 }
2273
2274 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2275                           char *msg, size_t msgsize)
2276 {
2277         int paranoid;
2278
2279         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2280             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2281             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2282                 /*
2283                  * If it's cycles then fall back to hrtimer based
2284                  * cpu-clock-tick sw counter, which is always available even if
2285                  * no PMU support.
2286                  *
2287                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2288                  * b0a873e).
2289                  */
2290                 scnprintf(msg, msgsize, "%s",
2291 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2292
2293                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2294                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2295
2296                 zfree(&evsel->name);
2297                 return true;
2298         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2299                    (paranoid = perf_event_paranoid()) > 1) {
2300                 const char *name = perf_evsel__name(evsel);
2301                 char *new_name;
2302
2303                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2304                         return false;
2305
2306                 if (evsel->name)
2307                         free(evsel->name);
2308                 evsel->name = new_name;
2309                 scnprintf(msg, msgsize,
2310 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2311                 evsel->attr.exclude_kernel = 1;
2312
2313                 return true;
2314         }
2315
2316         return false;
2317 }
2318
2319 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2320                               int err, char *msg, size_t size)
2321 {
2322         char sbuf[STRERR_BUFSIZE];
2323
2324         switch (err) {
2325         case EPERM:
2326         case EACCES:
2327                 return scnprintf(msg, size,
2328                  "You may not have permission to collect %sstats.\n\n"
2329                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2330                  "which controls use of the performance events system by\n"
2331                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2332                  "The current value is %d:\n\n"
2333                  "  -1: Allow use of (almost) all events by all users\n"
2334                  ">= 0: Disallow raw tracepoint access by users without CAP_IOC_LOCK\n"
2335                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2336                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN",
2337                                  target->system_wide ? "system-wide " : "",
2338                                  perf_event_paranoid());
2339         case ENOENT:
2340                 return scnprintf(msg, size, "The %s event is not supported.",
2341                                  perf_evsel__name(evsel));
2342         case EMFILE:
2343                 return scnprintf(msg, size, "%s",
2344                          "Too many events are opened.\n"
2345                          "Probably the maximum number of open file descriptors has been reached.\n"
2346                          "Hint: Try again after reducing the number of events.\n"
2347                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2348         case ENOMEM:
2349                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2350                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2351                         return scnprintf(msg, size,
2352                                          "Not enough memory to setup event with callchain.\n"
2353                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2354                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2355                 break;
2356         case ENODEV:
2357                 if (target->cpu_list)
2358                         return scnprintf(msg, size, "%s",
2359          "No such device - did you specify an out-of-range profile CPU?");
2360                 break;
2361         case EOPNOTSUPP:
2362                 if (evsel->attr.precise_ip)
2363                         return scnprintf(msg, size, "%s",
2364         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2365 #if defined(__i386__) || defined(__x86_64__)
2366                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2367                         return scnprintf(msg, size, "%s",
2368         "No hardware sampling interrupt available.\n"
2369         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2370 #endif
2371                 break;
2372         case EBUSY:
2373                 if (find_process("oprofiled"))
2374                         return scnprintf(msg, size,
2375         "The PMU counters are busy/taken by another profiler.\n"
2376         "We found oprofile daemon running, please stop it and try again.");
2377                 break;
2378         case EINVAL:
2379                 if (perf_missing_features.clockid)
2380                         return scnprintf(msg, size, "clockid feature not supported.");
2381                 if (perf_missing_features.clockid_wrong)
2382                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2383                 break;
2384         default:
2385                 break;
2386         }
2387
2388         return scnprintf(msg, size,
2389         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2390         "/bin/dmesg may provide additional information.\n"
2391         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2392                          err, strerror_r(err, sbuf, sizeof(sbuf)),
2393                          perf_evsel__name(evsel));
2394 }