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