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 <lk/debugfs.h>
  13 #include <traceevent/event-parse.h>
  14 #include <linux/hw_breakpoint.h>
  15 #include <linux/perf_event.h>
  16 #include <sys/resource.h>
  17 #include "asm/bug.h"
  18 #include "evsel.h"
  19 #include "evlist.h"
  20 #include "util.h"
  21 #include "cpumap.h"
  22 #include "thread_map.h"
  23 #include "target.h"
  24 #include "perf_regs.h"
  25 #include "debug.h"
  26
  27 static struct {
  28         bool sample_id_all;
  29         bool exclude_guest;
  30 } perf_missing_features;
  31
  32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  33
  34 static int __perf_evsel__sample_size(u64 sample_type)
  35 {
  36         u64 mask = sample_type & PERF_SAMPLE_MASK;
  37         int size = 0;
  38         int i;
  39
  40         for (i = 0; i < 64; i++) {
  41                 if (mask & (1ULL << i))
  42                         size++;
  43         }
  44
  45         size *= sizeof(u64);
  46
  47         return size;
  48 }
  49
  50 void hists__init(struct hists *hists)
  51 {
  52         memset(hists, 0, sizeof(*hists));
  53         hists->entries_in_array[0] = hists->entries_in_array[1] = RB_ROOT;
  54         hists->entries_in = &hists->entries_in_array[0];
  55         hists->entries_collapsed = RB_ROOT;
  56         hists->entries = RB_ROOT;
  57         pthread_mutex_init(&hists->lock, NULL);
  58 }
  59
  60 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
  61                                   enum perf_event_sample_format bit)
  62 {
  63         if (!(evsel->attr.sample_type & bit)) {
  64                 evsel->attr.sample_type |= bit;
  65                 evsel->sample_size += sizeof(u64);
  66         }
  67 }
  68
  69 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
  70                                     enum perf_event_sample_format bit)
  71 {
  72         if (evsel->attr.sample_type & bit) {
  73                 evsel->attr.sample_type &= ~bit;
  74                 evsel->sample_size -= sizeof(u64);
  75         }
  76 }
  77
  78 void perf_evsel__set_sample_id(struct perf_evsel *evsel)
  79 {
  80         perf_evsel__set_sample_bit(evsel, ID);
  81         evsel->attr.read_format |= PERF_FORMAT_ID;
  82 }
  83
  84 void perf_evsel__init(struct perf_evsel *evsel,
  85                       struct perf_event_attr *attr, int idx)
  86 {
  87         evsel->idx         = idx;
  88         evsel->attr        = *attr;
  89         evsel->leader      = evsel;
  90         INIT_LIST_HEAD(&evsel->node);
  91         hists__init(&evsel->hists);
  92         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
  93 }
  94
  95 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
  96 {
  97         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
  98
  99         if (evsel != NULL)
 100                 perf_evsel__init(evsel, attr, idx);
 101
 102         return evsel;
 103 }
 104
 105 struct event_format *event_format__new(const char *sys, const char *name)
 106 {
 107         int fd, n;
 108         char *filename;
 109         void *bf = NULL, *nbf;
 110         size_t size = 0, alloc_size = 0;
 111         struct event_format *format = NULL;
 112
 113         if (asprintf(&filename, "%s/%s/%s/format", tracing_events_path, sys, name) < 0)
 114                 goto out;
 115
 116         fd = open(filename, O_RDONLY);
 117         if (fd < 0)
 118                 goto out_free_filename;
 119
 120         do {
 121                 if (size == alloc_size) {
 122                         alloc_size += BUFSIZ;
 123                         nbf = realloc(bf, alloc_size);
 124                         if (nbf == NULL)
 125                                 goto out_free_bf;
 126                         bf = nbf;
 127                 }
 128
 129                 n = read(fd, bf + size, alloc_size - size);
 130                 if (n < 0)
 131                         goto out_free_bf;
 132                 size += n;
 133         } while (n > 0);
 134
 135         pevent_parse_format(&format, bf, size, sys);
 136
 137 out_free_bf:
 138         free(bf);
 139         close(fd);
 140 out_free_filename:
 141         free(filename);
 142 out:
 143         return format;
 144 }
 145
 146 struct perf_evsel *perf_evsel__newtp(const char *sys, const char *name, int idx)
 147 {
 148         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
 149
 150         if (evsel != NULL) {
 151                 struct perf_event_attr attr = {
 152                         .type          = PERF_TYPE_TRACEPOINT,
 153                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 154                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 155                 };
 156
 157                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 158                         goto out_free;
 159
 160                 evsel->tp_format = event_format__new(sys, name);
 161                 if (evsel->tp_format == NULL)
 162                         goto out_free;
 163
 164                 event_attr_init(&attr);
 165                 attr.config = evsel->tp_format->id;
 166                 attr.sample_period = 1;
 167                 perf_evsel__init(evsel, &attr, idx);
 168         }
 169
 170         return evsel;
 171
 172 out_free:
 173         free(evsel->name);
 174         free(evsel);
 175         return NULL;
 176 }
 177
 178 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 179         "cycles",
 180         "instructions",
 181         "cache-references",
 182         "cache-misses",
 183         "branches",
 184         "branch-misses",
 185         "bus-cycles",
 186         "stalled-cycles-frontend",
 187         "stalled-cycles-backend",
 188         "ref-cycles",
 189 };
 190
 191 static const char *__perf_evsel__hw_name(u64 config)
 192 {
 193         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 194                 return perf_evsel__hw_names[config];
 195
 196         return "unknown-hardware";
 197 }
 198
 199 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 200 {
 201         int colon = 0, r = 0;
 202         struct perf_event_attr *attr = &evsel->attr;
 203         bool exclude_guest_default = false;
 204
 205 #define MOD_PRINT(context, mod) do {                                    \
 206                 if (!attr->exclude_##context) {                         \
 207                         if (!colon) colon = ++r;                        \
 208                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 209                 } } while(0)
 210
 211         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 212                 MOD_PRINT(kernel, 'k');
 213                 MOD_PRINT(user, 'u');
 214                 MOD_PRINT(hv, 'h');
 215                 exclude_guest_default = true;
 216         }
 217
 218         if (attr->precise_ip) {
 219                 if (!colon)
 220                         colon = ++r;
 221                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 222                 exclude_guest_default = true;
 223         }
 224
 225         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 226                 MOD_PRINT(host, 'H');
 227                 MOD_PRINT(guest, 'G');
 228         }
 229 #undef MOD_PRINT
 230         if (colon)
 231                 bf[colon - 1] = ':';
 232         return r;
 233 }
 234
 235 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 236 {
 237         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 238         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 239 }
 240
 241 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 242         "cpu-clock",
 243         "task-clock",
 244         "page-faults",
 245         "context-switches",
 246         "cpu-migrations",
 247         "minor-faults",
 248         "major-faults",
 249         "alignment-faults",
 250         "emulation-faults",
 251 };
 252
 253 static const char *__perf_evsel__sw_name(u64 config)
 254 {
 255         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 256                 return perf_evsel__sw_names[config];
 257         return "unknown-software";
 258 }
 259
 260 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 261 {
 262         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 263         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 264 }
 265
 266 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 267 {
 268         int r;
 269
 270         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 271
 272         if (type & HW_BREAKPOINT_R)
 273                 r += scnprintf(bf + r, size - r, "r");
 274
 275         if (type & HW_BREAKPOINT_W)
 276                 r += scnprintf(bf + r, size - r, "w");
 277
 278         if (type & HW_BREAKPOINT_X)
 279                 r += scnprintf(bf + r, size - r, "x");
 280
 281         return r;
 282 }
 283
 284 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 285 {
 286         struct perf_event_attr *attr = &evsel->attr;
 287         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 288         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 289 }
 290
 291 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 292                                 [PERF_EVSEL__MAX_ALIASES] = {
 293  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 294  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 295  { "LLC",       "L2",                                                   },
 296  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 297  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 298  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 299  { "node",                                                              },
 300 };
 301
 302 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 303                                    [PERF_EVSEL__MAX_ALIASES] = {
 304  { "load",      "loads",        "read",                                 },
 305  { "store",     "stores",       "write",                                },
 306  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 307 };
 308
 309 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 310                                        [PERF_EVSEL__MAX_ALIASES] = {
 311  { "refs",      "Reference",    "ops",          "access",               },
 312  { "misses",    "miss",                                                 },
 313 };
 314
 315 #define C(x)            PERF_COUNT_HW_CACHE_##x
 316 #define CACHE_READ      (1 << C(OP_READ))
 317 #define CACHE_WRITE     (1 << C(OP_WRITE))
 318 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 319 #define COP(x)          (1 << x)
 320
 321 /*
 322  * cache operartion stat
 323  * L1I : Read and prefetch only
 324  * ITLB and BPU : Read-only
 325  */
 326 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 327  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 328  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 329  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 330  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 331  [C(ITLB)]      = (CACHE_READ),
 332  [C(BPU)]       = (CACHE_READ),
 333  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 334 };
 335
 336 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 337 {
 338         if (perf_evsel__hw_cache_stat[type] & COP(op))
 339                 return true;    /* valid */
 340         else
 341                 return false;   /* invalid */
 342 }
 343
 344 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 345                                             char *bf, size_t size)
 346 {
 347         if (result) {
 348                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 349                                  perf_evsel__hw_cache_op[op][0],
 350                                  perf_evsel__hw_cache_result[result][0]);
 351         }
 352
 353         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 354                          perf_evsel__hw_cache_op[op][1]);
 355 }
 356
 357 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 358 {
 359         u8 op, result, type = (config >>  0) & 0xff;
 360         const char *err = "unknown-ext-hardware-cache-type";
 361
 362         if (type > PERF_COUNT_HW_CACHE_MAX)
 363                 goto out_err;
 364
 365         op = (config >>  8) & 0xff;
 366         err = "unknown-ext-hardware-cache-op";
 367         if (op > PERF_COUNT_HW_CACHE_OP_MAX)
 368                 goto out_err;
 369
 370         result = (config >> 16) & 0xff;
 371         err = "unknown-ext-hardware-cache-result";
 372         if (result > PERF_COUNT_HW_CACHE_RESULT_MAX)
 373                 goto out_err;
 374
 375         err = "invalid-cache";
 376         if (!perf_evsel__is_cache_op_valid(type, op))
 377                 goto out_err;
 378
 379         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 380 out_err:
 381         return scnprintf(bf, size, "%s", err);
 382 }
 383
 384 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 385 {
 386         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 387         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 388 }
 389
 390 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 391 {
 392         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 393         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 394 }
 395
 396 const char *perf_evsel__name(struct perf_evsel *evsel)
 397 {
 398         char bf[128];
 399
 400         if (evsel->name)
 401                 return evsel->name;
 402
 403         switch (evsel->attr.type) {
 404         case PERF_TYPE_RAW:
 405                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 406                 break;
 407
 408         case PERF_TYPE_HARDWARE:
 409                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 410                 break;
 411
 412         case PERF_TYPE_HW_CACHE:
 413                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 414                 break;
 415
 416         case PERF_TYPE_SOFTWARE:
 417                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 418                 break;
 419
 420         case PERF_TYPE_TRACEPOINT:
 421                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 422                 break;
 423
 424         case PERF_TYPE_BREAKPOINT:
 425                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 426                 break;
 427
 428         default:
 429                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 430                           evsel->attr.type);
 431                 break;
 432         }
 433
 434         evsel->name = strdup(bf);
 435
 436         return evsel->name ?: "unknown";
 437 }
 438
 439 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 440 {
 441         return evsel->group_name ?: "anon group";
 442 }
 443
 444 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 445 {
 446         int ret;
 447         struct perf_evsel *pos;
 448         const char *group_name = perf_evsel__group_name(evsel);
 449
 450         ret = scnprintf(buf, size, "%s", group_name);
 451
 452         ret += scnprintf(buf + ret, size - ret, " { %s",
 453                          perf_evsel__name(evsel));
 454
 455         for_each_group_member(pos, evsel)
 456                 ret += scnprintf(buf + ret, size - ret, ", %s",
 457                                  perf_evsel__name(pos));
 458
 459         ret += scnprintf(buf + ret, size - ret, " }");
 460
 461         return ret;
 462 }
 463
 464 /*
 465  * The enable_on_exec/disabled value strategy:
 466  *
 467  *  1) For any type of traced program:
 468  *    - all independent events and group leaders are disabled
 469  *    - all group members are enabled
 470  *
 471  *     Group members are ruled by group leaders. They need to
 472  *     be enabled, because the group scheduling relies on that.
 473  *
 474  *  2) For traced programs executed by perf:
 475  *     - all independent events and group leaders have
 476  *       enable_on_exec set
 477  *     - we don't specifically enable or disable any event during
 478  *       the record command
 479  *
 480  *     Independent events and group leaders are initially disabled
 481  *     and get enabled by exec. Group members are ruled by group
 482  *     leaders as stated in 1).
 483  *
 484  *  3) For traced programs attached by perf (pid/tid):
 485  *     - we specifically enable or disable all events during
 486  *       the record command
 487  *
 488  *     When attaching events to already running traced we
 489  *     enable/disable events specifically, as there's no
 490  *     initial traced exec call.
 491  */
 492 void perf_evsel__config(struct perf_evsel *evsel,
 493                         struct perf_record_opts *opts)
 494 {
 495         struct perf_evsel *leader = evsel->leader;
 496         struct perf_event_attr *attr = &evsel->attr;
 497         int track = !evsel->idx; /* only the first counter needs these */
 498
 499         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 500         attr->inherit       = !opts->no_inherit;
 501
 502         perf_evsel__set_sample_bit(evsel, IP);
 503         perf_evsel__set_sample_bit(evsel, TID);
 504
 505         if (evsel->sample_read) {
 506                 perf_evsel__set_sample_bit(evsel, READ);
 507
 508                 /*
 509                  * We need ID even in case of single event, because
 510                  * PERF_SAMPLE_READ process ID specific data.
 511                  */
 512                 perf_evsel__set_sample_id(evsel);
 513
 514                 /*
 515                  * Apply group format only if we belong to group
 516                  * with more than one members.
 517                  */
 518                 if (leader->nr_members > 1) {
 519                         attr->read_format |= PERF_FORMAT_GROUP;
 520                         attr->inherit = 0;
 521                 }
 522         }
 523
 524         /*
 525          * We default some events to a 1 default interval. But keep
 526          * it a weak assumption overridable by the user.
 527          */
 528         if (!attr->sample_period || (opts->user_freq != UINT_MAX &&
 529                                      opts->user_interval != ULLONG_MAX)) {
 530                 if (opts->freq) {
 531                         perf_evsel__set_sample_bit(evsel, PERIOD);
 532                         attr->freq              = 1;
 533                         attr->sample_freq       = opts->freq;
 534                 } else {
 535                         attr->sample_period = opts->default_interval;
 536                 }
 537         }
 538
 539         /*
 540          * Disable sampling for all group members other
 541          * than leader in case leader 'leads' the sampling.
 542          */
 543         if ((leader != evsel) && leader->sample_read) {
 544                 attr->sample_freq   = 0;
 545                 attr->sample_period = 0;
 546         }
 547
 548         if (opts->no_samples)
 549                 attr->sample_freq = 0;
 550
 551         if (opts->inherit_stat)
 552                 attr->inherit_stat = 1;
 553
 554         if (opts->sample_address) {
 555                 perf_evsel__set_sample_bit(evsel, ADDR);
 556                 attr->mmap_data = track;
 557         }
 558
 559         if (opts->call_graph) {
 560                 perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 561
 562                 if (opts->call_graph == CALLCHAIN_DWARF) {
 563                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 564                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 565                         attr->sample_regs_user = PERF_REGS_MASK;
 566                         attr->sample_stack_user = opts->stack_dump_size;
 567                         attr->exclude_callchain_user = 1;
 568                 }
 569         }
 570
 571         if (perf_target__has_cpu(&opts->target))
 572                 perf_evsel__set_sample_bit(evsel, CPU);
 573
 574         if (opts->period)
 575                 perf_evsel__set_sample_bit(evsel, PERIOD);
 576
 577         if (!perf_missing_features.sample_id_all &&
 578             (opts->sample_time || !opts->no_inherit ||
 579              perf_target__has_cpu(&opts->target)))
 580                 perf_evsel__set_sample_bit(evsel, TIME);
 581
 582         if (opts->raw_samples) {
 583                 perf_evsel__set_sample_bit(evsel, TIME);
 584                 perf_evsel__set_sample_bit(evsel, RAW);
 585                 perf_evsel__set_sample_bit(evsel, CPU);
 586         }
 587
 588         if (opts->sample_address)
 589                 attr->sample_type       |= PERF_SAMPLE_DATA_SRC;
 590
 591         if (opts->no_delay) {
 592                 attr->watermark = 0;
 593                 attr->wakeup_events = 1;
 594         }
 595         if (opts->branch_stack) {
 596                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 597                 attr->branch_sample_type = opts->branch_stack;
 598         }
 599
 600         if (opts->sample_weight)
 601                 attr->sample_type       |= PERF_SAMPLE_WEIGHT;
 602
 603         attr->mmap = track;
 604         attr->comm = track;
 605
 606         /*
 607          * XXX see the function comment above
 608          *
 609          * Disabling only independent events or group leaders,
 610          * keeping group members enabled.
 611          */
 612         if (perf_evsel__is_group_leader(evsel))
 613                 attr->disabled = 1;
 614
 615         /*
 616          * Setting enable_on_exec for independent events and
 617          * group leaders for traced executed by perf.
 618          */
 619         if (perf_target__none(&opts->target) && perf_evsel__is_group_leader(evsel))
 620                 attr->enable_on_exec = 1;
 621 }
 622
 623 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 624 {
 625         int cpu, thread;
 626         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
 627
 628         if (evsel->fd) {
 629                 for (cpu = 0; cpu < ncpus; cpu++) {
 630                         for (thread = 0; thread < nthreads; thread++) {
 631                                 FD(evsel, cpu, thread) = -1;
 632                         }
 633                 }
 634         }
 635
 636         return evsel->fd != NULL ? 0 : -ENOMEM;
 637 }
 638
 639 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads,
 640                           int ioc,  void *arg)
 641 {
 642         int cpu, thread;
 643
 644         for (cpu = 0; cpu < ncpus; cpu++) {
 645                 for (thread = 0; thread < nthreads; thread++) {
 646                         int fd = FD(evsel, cpu, thread),
 647                             err = ioctl(fd, ioc, arg);
 648
 649                         if (err)
 650                                 return err;
 651                 }
 652         }
 653
 654         return 0;
 655 }
 656
 657 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads,
 658                            const char *filter)
 659 {
 660         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 661                                      PERF_EVENT_IOC_SET_FILTER,
 662                                      (void *)filter);
 663 }
 664
 665 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads)
 666 {
 667         return perf_evsel__run_ioctl(evsel, ncpus, nthreads,
 668                                      PERF_EVENT_IOC_ENABLE,
 669                                      0);
 670 }
 671
 672 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
 673 {
 674         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
 675         if (evsel->sample_id == NULL)
 676                 return -ENOMEM;
 677
 678         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
 679         if (evsel->id == NULL) {
 680                 xyarray__delete(evsel->sample_id);
 681                 evsel->sample_id = NULL;
 682                 return -ENOMEM;
 683         }
 684
 685         return 0;
 686 }
 687
 688 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus)
 689 {
 690         memset(evsel->counts, 0, (sizeof(*evsel->counts) +
 691                                  (ncpus * sizeof(struct perf_counts_values))));
 692 }
 693
 694 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
 695 {
 696         evsel->counts = zalloc((sizeof(*evsel->counts) +
 697                                 (ncpus * sizeof(struct perf_counts_values))));
 698         return evsel->counts != NULL ? 0 : -ENOMEM;
 699 }
 700
 701 void perf_evsel__free_fd(struct perf_evsel *evsel)
 702 {
 703         xyarray__delete(evsel->fd);
 704         evsel->fd = NULL;
 705 }
 706
 707 void perf_evsel__free_id(struct perf_evsel *evsel)
 708 {
 709         xyarray__delete(evsel->sample_id);
 710         evsel->sample_id = NULL;
 711         free(evsel->id);
 712         evsel->id = NULL;
 713 }
 714
 715 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
 716 {
 717         int cpu, thread;
 718
 719         for (cpu = 0; cpu < ncpus; cpu++)
 720                 for (thread = 0; thread < nthreads; ++thread) {
 721                         close(FD(evsel, cpu, thread));
 722                         FD(evsel, cpu, thread) = -1;
 723                 }
 724 }
 725
 726 void perf_evsel__free_counts(struct perf_evsel *evsel)
 727 {
 728         free(evsel->counts);
 729 }
 730
 731 void perf_evsel__exit(struct perf_evsel *evsel)
 732 {
 733         assert(list_empty(&evsel->node));
 734         perf_evsel__free_fd(evsel);
 735         perf_evsel__free_id(evsel);
 736 }
 737
 738 void perf_evsel__delete(struct perf_evsel *evsel)
 739 {
 740         perf_evsel__exit(evsel);
 741         close_cgroup(evsel->cgrp);
 742         free(evsel->group_name);
 743         if (evsel->tp_format)
 744                 pevent_free_format(evsel->tp_format);
 745         free(evsel->name);
 746         free(evsel);
 747 }
 748
 749 static inline void compute_deltas(struct perf_evsel *evsel,
 750                                   int cpu,
 751                                   struct perf_counts_values *count)
 752 {
 753         struct perf_counts_values tmp;
 754
 755         if (!evsel->prev_raw_counts)
 756                 return;
 757
 758         if (cpu == -1) {
 759                 tmp = evsel->prev_raw_counts->aggr;
 760                 evsel->prev_raw_counts->aggr = *count;
 761         } else {
 762                 tmp = evsel->prev_raw_counts->cpu[cpu];
 763                 evsel->prev_raw_counts->cpu[cpu] = *count;
 764         }
 765
 766         count->val = count->val - tmp.val;
 767         count->ena = count->ena - tmp.ena;
 768         count->run = count->run - tmp.run;
 769 }
 770
 771 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
 772                               int cpu, int thread, bool scale)
 773 {
 774         struct perf_counts_values count;
 775         size_t nv = scale ? 3 : 1;
 776
 777         if (FD(evsel, cpu, thread) < 0)
 778                 return -EINVAL;
 779
 780         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 781                 return -ENOMEM;
 782
 783         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 784                 return -errno;
 785
 786         compute_deltas(evsel, cpu, &count);
 787
 788         if (scale) {
 789                 if (count.run == 0)
 790                         count.val = 0;
 791                 else if (count.run < count.ena)
 792                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
 793         } else
 794                 count.ena = count.run = 0;
 795
 796         evsel->counts->cpu[cpu] = count;
 797         return 0;
 798 }
 799
 800 int __perf_evsel__read(struct perf_evsel *evsel,
 801                        int ncpus, int nthreads, bool scale)
 802 {
 803         size_t nv = scale ? 3 : 1;
 804         int cpu, thread;
 805         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
 806
 807         aggr->val = aggr->ena = aggr->run = 0;
 808
 809         for (cpu = 0; cpu < ncpus; cpu++) {
 810                 for (thread = 0; thread < nthreads; thread++) {
 811                         if (FD(evsel, cpu, thread) < 0)
 812                                 continue;
 813
 814                         if (readn(FD(evsel, cpu, thread),
 815                                   &count, nv * sizeof(u64)) < 0)
 816                                 return -errno;
 817
 818                         aggr->val += count.val;
 819                         if (scale) {
 820                                 aggr->ena += count.ena;
 821                                 aggr->run += count.run;
 822                         }
 823                 }
 824         }
 825
 826         compute_deltas(evsel, -1, aggr);
 827
 828         evsel->counts->scaled = 0;
 829         if (scale) {
 830                 if (aggr->run == 0) {
 831                         evsel->counts->scaled = -1;
 832                         aggr->val = 0;
 833                         return 0;
 834                 }
 835
 836                 if (aggr->run < aggr->ena) {
 837                         evsel->counts->scaled = 1;
 838                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 839                 }
 840         } else
 841                 aggr->ena = aggr->run = 0;
 842
 843         return 0;
 844 }
 845
 846 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
 847 {
 848         struct perf_evsel *leader = evsel->leader;
 849         int fd;
 850
 851         if (perf_evsel__is_group_leader(evsel))
 852                 return -1;
 853
 854         /*
 855          * Leader must be already processed/open,
 856          * if not it's a bug.
 857          */
 858         BUG_ON(!leader->fd);
 859
 860         fd = FD(leader, cpu, thread);
 861         BUG_ON(fd == -1);
 862
 863         return fd;
 864 }
 865
 866 #define __PRINT_ATTR(fmt, cast, field)  \
 867         fprintf(fp, "  %-19s "fmt"\n", #field, cast attr->field)
 868
 869 #define PRINT_ATTR_U32(field)  __PRINT_ATTR("%u" , , field)
 870 #define PRINT_ATTR_X32(field)  __PRINT_ATTR("%#x", , field)
 871 #define PRINT_ATTR_U64(field)  __PRINT_ATTR("%" PRIu64, (uint64_t), field)
 872 #define PRINT_ATTR_X64(field)  __PRINT_ATTR("%#"PRIx64, (uint64_t), field)
 873
 874 #define PRINT_ATTR2N(name1, field1, name2, field2)      \
 875         fprintf(fp, "  %-19s %u    %-19s %u\n",         \
 876         name1, attr->field1, name2, attr->field2)
 877
 878 #define PRINT_ATTR2(field1, field2) \
 879         PRINT_ATTR2N(#field1, field1, #field2, field2)
 880
 881 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp)
 882 {
 883         size_t ret = 0;
 884
 885         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
 886         ret += fprintf(fp, "perf_event_attr:\n");
 887
 888         ret += PRINT_ATTR_U32(type);
 889         ret += PRINT_ATTR_U32(size);
 890         ret += PRINT_ATTR_X64(config);
 891         ret += PRINT_ATTR_U64(sample_period);
 892         ret += PRINT_ATTR_U64(sample_freq);
 893         ret += PRINT_ATTR_X64(sample_type);
 894         ret += PRINT_ATTR_X64(read_format);
 895
 896         ret += PRINT_ATTR2(disabled, inherit);
 897         ret += PRINT_ATTR2(pinned, exclusive);
 898         ret += PRINT_ATTR2(exclude_user, exclude_kernel);
 899         ret += PRINT_ATTR2(exclude_hv, exclude_idle);
 900         ret += PRINT_ATTR2(mmap, comm);
 901         ret += PRINT_ATTR2(freq, inherit_stat);
 902         ret += PRINT_ATTR2(enable_on_exec, task);
 903         ret += PRINT_ATTR2(watermark, precise_ip);
 904         ret += PRINT_ATTR2(mmap_data, sample_id_all);
 905         ret += PRINT_ATTR2(exclude_host, exclude_guest);
 906         ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel,
 907                             "excl.callchain_user", exclude_callchain_user);
 908
 909         ret += PRINT_ATTR_U32(wakeup_events);
 910         ret += PRINT_ATTR_U32(wakeup_watermark);
 911         ret += PRINT_ATTR_X32(bp_type);
 912         ret += PRINT_ATTR_X64(bp_addr);
 913         ret += PRINT_ATTR_X64(config1);
 914         ret += PRINT_ATTR_U64(bp_len);
 915         ret += PRINT_ATTR_X64(config2);
 916         ret += PRINT_ATTR_X64(branch_sample_type);
 917         ret += PRINT_ATTR_X64(sample_regs_user);
 918         ret += PRINT_ATTR_U32(sample_stack_user);
 919
 920         ret += fprintf(fp, "%.60s\n", graph_dotted_line);
 921
 922         return ret;
 923 }
 924
 925 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 926                               struct thread_map *threads)
 927 {
 928         int cpu, thread;
 929         unsigned long flags = 0;
 930         int pid = -1, err;
 931         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
 932
 933         if (evsel->fd == NULL &&
 934             perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
 935                 return -ENOMEM;
 936
 937         if (evsel->cgrp) {
 938                 flags = PERF_FLAG_PID_CGROUP;
 939                 pid = evsel->cgrp->fd;
 940         }
 941
 942 fallback_missing_features:
 943         if (perf_missing_features.exclude_guest)
 944                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
 945 retry_sample_id:
 946         if (perf_missing_features.sample_id_all)
 947                 evsel->attr.sample_id_all = 0;
 948
 949         if (verbose >= 2)
 950                 perf_event_attr__fprintf(&evsel->attr, stderr);
 951
 952         for (cpu = 0; cpu < cpus->nr; cpu++) {
 953
 954                 for (thread = 0; thread < threads->nr; thread++) {
 955                         int group_fd;
 956
 957                         if (!evsel->cgrp)
 958                                 pid = threads->map[thread];
 959
 960                         group_fd = get_group_fd(evsel, cpu, thread);
 961 retry_open:
 962                         pr_debug2("perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx\n",
 963                                   pid, cpus->map[cpu], group_fd, flags);
 964
 965                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
 966                                                                      pid,
 967                                                                      cpus->map[cpu],
 968                                                                      group_fd, flags);
 969                         if (FD(evsel, cpu, thread) < 0) {
 970                                 err = -errno;
 971                                 goto try_fallback;
 972                         }
 973                         set_rlimit = NO_CHANGE;
 974                 }
 975         }
 976
 977         return 0;
 978
 979 try_fallback:
 980         /*
 981          * perf stat needs between 5 and 22 fds per CPU. When we run out
 982          * of them try to increase the limits.
 983          */
 984         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
 985                 struct rlimit l;
 986                 int old_errno = errno;
 987
 988                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
 989                         if (set_rlimit == NO_CHANGE)
 990                                 l.rlim_cur = l.rlim_max;
 991                         else {
 992                                 l.rlim_cur = l.rlim_max + 1000;
 993                                 l.rlim_max = l.rlim_cur;
 994                         }
 995                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
 996                                 set_rlimit++;
 997                                 errno = old_errno;
 998                                 goto retry_open;
 999                         }
1000                 }
1001                 errno = old_errno;
1002         }
1003
1004         if (err != -EINVAL || cpu > 0 || thread > 0)
1005                 goto out_close;
1006
1007         if (!perf_missing_features.exclude_guest &&
1008             (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1009                 perf_missing_features.exclude_guest = true;
1010                 goto fallback_missing_features;
1011         } else if (!perf_missing_features.sample_id_all) {
1012                 perf_missing_features.sample_id_all = true;
1013                 goto retry_sample_id;
1014         }
1015
1016 out_close:
1017         do {
1018                 while (--thread >= 0) {
1019                         close(FD(evsel, cpu, thread));
1020                         FD(evsel, cpu, thread) = -1;
1021                 }
1022                 thread = threads->nr;
1023         } while (--cpu >= 0);
1024         return err;
1025 }
1026
1027 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
1028 {
1029         if (evsel->fd == NULL)
1030                 return;
1031
1032         perf_evsel__close_fd(evsel, ncpus, nthreads);
1033         perf_evsel__free_fd(evsel);
1034         evsel->fd = NULL;
1035 }
1036
1037 static struct {
1038         struct cpu_map map;
1039         int cpus[1];
1040 } empty_cpu_map = {
1041         .map.nr = 1,
1042         .cpus   = { -1, },
1043 };
1044
1045 static struct {
1046         struct thread_map map;
1047         int threads[1];
1048 } empty_thread_map = {
1049         .map.nr  = 1,
1050         .threads = { -1, },
1051 };
1052
1053 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1054                      struct thread_map *threads)
1055 {
1056         if (cpus == NULL) {
1057                 /* Work around old compiler warnings about strict aliasing */
1058                 cpus = &empty_cpu_map.map;
1059         }
1060
1061         if (threads == NULL)
1062                 threads = &empty_thread_map.map;
1063
1064         return __perf_evsel__open(evsel, cpus, threads);
1065 }
1066
1067 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1068                              struct cpu_map *cpus)
1069 {
1070         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map);
1071 }
1072
1073 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1074                                 struct thread_map *threads)
1075 {
1076         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads);
1077 }
1078
1079 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1080                                        const union perf_event *event,
1081                                        struct perf_sample *sample)
1082 {
1083         u64 type = evsel->attr.sample_type;
1084         const u64 *array = event->sample.array;
1085         bool swapped = evsel->needs_swap;
1086         union u64_swap u;
1087
1088         array += ((event->header.size -
1089                    sizeof(event->header)) / sizeof(u64)) - 1;
1090
1091         if (type & PERF_SAMPLE_CPU) {
1092                 u.val64 = *array;
1093                 if (swapped) {
1094                         /* undo swap of u64, then swap on individual u32s */
1095                         u.val64 = bswap_64(u.val64);
1096                         u.val32[0] = bswap_32(u.val32[0]);
1097                 }
1098
1099                 sample->cpu = u.val32[0];
1100                 array--;
1101         }
1102
1103         if (type & PERF_SAMPLE_STREAM_ID) {
1104                 sample->stream_id = *array;
1105                 array--;
1106         }
1107
1108         if (type & PERF_SAMPLE_ID) {
1109                 sample->id = *array;
1110                 array--;
1111         }
1112
1113         if (type & PERF_SAMPLE_TIME) {
1114                 sample->time = *array;
1115                 array--;
1116         }
1117
1118         if (type & PERF_SAMPLE_TID) {
1119                 u.val64 = *array;
1120                 if (swapped) {
1121                         /* undo swap of u64, then swap on individual u32s */
1122                         u.val64 = bswap_64(u.val64);
1123                         u.val32[0] = bswap_32(u.val32[0]);
1124                         u.val32[1] = bswap_32(u.val32[1]);
1125                 }
1126
1127                 sample->pid = u.val32[0];
1128                 sample->tid = u.val32[1];
1129         }
1130
1131         return 0;
1132 }
1133
1134 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1135                             u64 size)
1136 {
1137         return size > max_size || offset + size > endp;
1138 }
1139
1140 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1141         do {                                                            \
1142                 if (overflow(endp, (max_size), (offset), (size)))       \
1143                         return -EFAULT;                                 \
1144         } while (0)
1145
1146 #define OVERFLOW_CHECK_u64(offset) \
1147         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1148
1149 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1150                              struct perf_sample *data)
1151 {
1152         u64 type = evsel->attr.sample_type;
1153         bool swapped = evsel->needs_swap;
1154         const u64 *array;
1155         u16 max_size = event->header.size;
1156         const void *endp = (void *)event + max_size;
1157         u64 sz;
1158
1159         /*
1160          * used for cross-endian analysis. See git commit 65014ab3
1161          * for why this goofiness is needed.
1162          */
1163         union u64_swap u;
1164
1165         memset(data, 0, sizeof(*data));
1166         data->cpu = data->pid = data->tid = -1;
1167         data->stream_id = data->id = data->time = -1ULL;
1168         data->period = 1;
1169         data->weight = 0;
1170
1171         if (event->header.type != PERF_RECORD_SAMPLE) {
1172                 if (!evsel->attr.sample_id_all)
1173                         return 0;
1174                 return perf_evsel__parse_id_sample(evsel, event, data);
1175         }
1176
1177         array = event->sample.array;
1178
1179         /*
1180          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1181          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1182          * check the format does not go past the end of the event.
1183          */
1184         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1185                 return -EFAULT;
1186
1187         if (type & PERF_SAMPLE_IP) {
1188                 data->ip = *array;
1189                 array++;
1190         }
1191
1192         if (type & PERF_SAMPLE_TID) {
1193                 u.val64 = *array;
1194                 if (swapped) {
1195                         /* undo swap of u64, then swap on individual u32s */
1196                         u.val64 = bswap_64(u.val64);
1197                         u.val32[0] = bswap_32(u.val32[0]);
1198                         u.val32[1] = bswap_32(u.val32[1]);
1199                 }
1200
1201                 data->pid = u.val32[0];
1202                 data->tid = u.val32[1];
1203                 array++;
1204         }
1205
1206         if (type & PERF_SAMPLE_TIME) {
1207                 data->time = *array;
1208                 array++;
1209         }
1210
1211         data->addr = 0;
1212         if (type & PERF_SAMPLE_ADDR) {
1213                 data->addr = *array;
1214                 array++;
1215         }
1216
1217         data->id = -1ULL;
1218         if (type & PERF_SAMPLE_ID) {
1219                 data->id = *array;
1220                 array++;
1221         }
1222
1223         if (type & PERF_SAMPLE_STREAM_ID) {
1224                 data->stream_id = *array;
1225                 array++;
1226         }
1227
1228         if (type & PERF_SAMPLE_CPU) {
1229
1230                 u.val64 = *array;
1231                 if (swapped) {
1232                         /* undo swap of u64, then swap on individual u32s */
1233                         u.val64 = bswap_64(u.val64);
1234                         u.val32[0] = bswap_32(u.val32[0]);
1235                 }
1236
1237                 data->cpu = u.val32[0];
1238                 array++;
1239         }
1240
1241         if (type & PERF_SAMPLE_PERIOD) {
1242                 data->period = *array;
1243                 array++;
1244         }
1245
1246         if (type & PERF_SAMPLE_READ) {
1247                 u64 read_format = evsel->attr.read_format;
1248
1249                 OVERFLOW_CHECK_u64(array);
1250                 if (read_format & PERF_FORMAT_GROUP)
1251                         data->read.group.nr = *array;
1252                 else
1253                         data->read.one.value = *array;
1254
1255                 array++;
1256
1257                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
1258                         OVERFLOW_CHECK_u64(array);
1259                         data->read.time_enabled = *array;
1260                         array++;
1261                 }
1262
1263                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
1264                         OVERFLOW_CHECK_u64(array);
1265                         data->read.time_running = *array;
1266                         array++;
1267                 }
1268
1269                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
1270                 if (read_format & PERF_FORMAT_GROUP) {
1271                         const u64 max_group_nr = UINT64_MAX /
1272                                         sizeof(struct sample_read_value);
1273
1274                         if (data->read.group.nr > max_group_nr)
1275                                 return -EFAULT;
1276                         sz = data->read.group.nr *
1277                              sizeof(struct sample_read_value);
1278                         OVERFLOW_CHECK(array, sz, max_size);
1279                         data->read.group.values =
1280                                         (struct sample_read_value *)array;
1281                         array = (void *)array + sz;
1282                 } else {
1283                         OVERFLOW_CHECK_u64(array);
1284                         data->read.one.id = *array;
1285                         array++;
1286                 }
1287         }
1288
1289         if (type & PERF_SAMPLE_CALLCHAIN) {
1290                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
1291
1292                 OVERFLOW_CHECK_u64(array);
1293                 data->callchain = (struct ip_callchain *)array++;
1294                 if (data->callchain->nr > max_callchain_nr)
1295                         return -EFAULT;
1296                 sz = data->callchain->nr * sizeof(u64);
1297                 OVERFLOW_CHECK(array, sz, max_size);
1298                 array = (void *)array + sz;
1299         }
1300
1301         if (type & PERF_SAMPLE_RAW) {
1302                 OVERFLOW_CHECK_u64(array);
1303                 u.val64 = *array;
1304                 if (WARN_ONCE(swapped,
1305                               "Endianness of raw data not corrected!\n")) {
1306                         /* undo swap of u64, then swap on individual u32s */
1307                         u.val64 = bswap_64(u.val64);
1308                         u.val32[0] = bswap_32(u.val32[0]);
1309                         u.val32[1] = bswap_32(u.val32[1]);
1310                 }
1311                 data->raw_size = u.val32[0];
1312                 array = (void *)array + sizeof(u32);
1313
1314                 OVERFLOW_CHECK(array, data->raw_size, max_size);
1315                 data->raw_data = (void *)array;
1316                 array = (void *)array + data->raw_size;
1317         }
1318
1319         if (type & PERF_SAMPLE_BRANCH_STACK) {
1320                 const u64 max_branch_nr = UINT64_MAX /
1321                                           sizeof(struct branch_entry);
1322
1323                 OVERFLOW_CHECK_u64(array);
1324                 data->branch_stack = (struct branch_stack *)array++;
1325
1326                 if (data->branch_stack->nr > max_branch_nr)
1327                         return -EFAULT;
1328                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
1329                 OVERFLOW_CHECK(array, sz, max_size);
1330                 array = (void *)array + sz;
1331         }
1332
1333         if (type & PERF_SAMPLE_REGS_USER) {
1334                 u64 avail;
1335
1336                 /* First u64 tells us if we have any regs in sample. */
1337                 OVERFLOW_CHECK_u64(array);
1338                 avail = *array++;
1339
1340                 if (avail) {
1341                         u64 regs_user = evsel->attr.sample_regs_user;
1342
1343                         sz = hweight_long(regs_user) * sizeof(u64);
1344                         OVERFLOW_CHECK(array, sz, max_size);
1345                         data->user_regs.regs = (u64 *)array;
1346                         array = (void *)array + sz;
1347                 }
1348         }
1349
1350         if (type & PERF_SAMPLE_STACK_USER) {
1351                 OVERFLOW_CHECK_u64(array);
1352                 sz = *array++;
1353
1354                 data->user_stack.offset = ((char *)(array - 1)
1355                                           - (char *) event);
1356
1357                 if (!sz) {
1358                         data->user_stack.size = 0;
1359                 } else {
1360                         OVERFLOW_CHECK(array, sz, max_size);
1361                         data->user_stack.data = (char *)array;
1362                         array = (void *)array + sz;
1363                         OVERFLOW_CHECK_u64(array);
1364                         data->user_stack.size = *array++;
1365                 }
1366         }
1367
1368         data->weight = 0;
1369         if (type & PERF_SAMPLE_WEIGHT) {
1370                 OVERFLOW_CHECK_u64(array);
1371                 data->weight = *array;
1372                 array++;
1373         }
1374
1375         data->data_src = PERF_MEM_DATA_SRC_NONE;
1376         if (type & PERF_SAMPLE_DATA_SRC) {
1377                 OVERFLOW_CHECK_u64(array);
1378                 data->data_src = *array;
1379                 array++;
1380         }
1381
1382         return 0;
1383 }
1384
1385 int perf_event__synthesize_sample(union perf_event *event, u64 type,
1386                                   const struct perf_sample *sample,
1387                                   bool swapped)
1388 {
1389         u64 *array;
1390
1391         /*
1392          * used for cross-endian analysis. See git commit 65014ab3
1393          * for why this goofiness is needed.
1394          */
1395         union u64_swap u;
1396
1397         array = event->sample.array;
1398
1399         if (type & PERF_SAMPLE_IP) {
1400                 *array = sample->ip;
1401                 array++;
1402         }
1403
1404         if (type & PERF_SAMPLE_TID) {
1405                 u.val32[0] = sample->pid;
1406                 u.val32[1] = sample->tid;
1407                 if (swapped) {
1408                         /*
1409                          * Inverse of what is done in perf_evsel__parse_sample
1410                          */
1411                         u.val32[0] = bswap_32(u.val32[0]);
1412                         u.val32[1] = bswap_32(u.val32[1]);
1413                         u.val64 = bswap_64(u.val64);
1414                 }
1415
1416                 *array = u.val64;
1417                 array++;
1418         }
1419
1420         if (type & PERF_SAMPLE_TIME) {
1421                 *array = sample->time;
1422                 array++;
1423         }
1424
1425         if (type & PERF_SAMPLE_ADDR) {
1426                 *array = sample->addr;
1427                 array++;
1428         }
1429
1430         if (type & PERF_SAMPLE_ID) {
1431                 *array = sample->id;
1432                 array++;
1433         }
1434
1435         if (type & PERF_SAMPLE_STREAM_ID) {
1436                 *array = sample->stream_id;
1437                 array++;
1438         }
1439
1440         if (type & PERF_SAMPLE_CPU) {
1441                 u.val32[0] = sample->cpu;
1442                 if (swapped) {
1443                         /*
1444                          * Inverse of what is done in perf_evsel__parse_sample
1445                          */
1446                         u.val32[0] = bswap_32(u.val32[0]);
1447                         u.val64 = bswap_64(u.val64);
1448                 }
1449                 *array = u.val64;
1450                 array++;
1451         }
1452
1453         if (type & PERF_SAMPLE_PERIOD) {
1454                 *array = sample->period;
1455                 array++;
1456         }
1457
1458         return 0;
1459 }
1460
1461 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
1462 {
1463         return pevent_find_field(evsel->tp_format, name);
1464 }
1465
1466 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
1467                          const char *name)
1468 {
1469         struct format_field *field = perf_evsel__field(evsel, name);
1470         int offset;
1471
1472         if (!field)
1473                 return NULL;
1474
1475         offset = field->offset;
1476
1477         if (field->flags & FIELD_IS_DYNAMIC) {
1478                 offset = *(int *)(sample->raw_data + field->offset);
1479                 offset &= 0xffff;
1480         }
1481
1482         return sample->raw_data + offset;
1483 }
1484
1485 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
1486                        const char *name)
1487 {
1488         struct format_field *field = perf_evsel__field(evsel, name);
1489         void *ptr;
1490         u64 value;
1491
1492         if (!field)
1493                 return 0;
1494
1495         ptr = sample->raw_data + field->offset;
1496
1497         switch (field->size) {
1498         case 1:
1499                 return *(u8 *)ptr;
1500         case 2:
1501                 value = *(u16 *)ptr;
1502                 break;
1503         case 4:
1504                 value = *(u32 *)ptr;
1505                 break;
1506         case 8:
1507                 value = *(u64 *)ptr;
1508                 break;
1509         default:
1510                 return 0;
1511         }
1512
1513         if (!evsel->needs_swap)
1514                 return value;
1515
1516         switch (field->size) {
1517         case 2:
1518                 return bswap_16(value);
1519         case 4:
1520                 return bswap_32(value);
1521         case 8:
1522                 return bswap_64(value);
1523         default:
1524                 return 0;
1525         }
1526
1527         return 0;
1528 }
1529
1530 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...)
1531 {
1532         va_list args;
1533         int ret = 0;
1534
1535         if (!*first) {
1536                 ret += fprintf(fp, ",");
1537         } else {
1538                 ret += fprintf(fp, ":");
1539                 *first = false;
1540         }
1541
1542         va_start(args, fmt);
1543         ret += vfprintf(fp, fmt, args);
1544         va_end(args);
1545         return ret;
1546 }
1547
1548 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value)
1549 {
1550         if (value == 0)
1551                 return 0;
1552
1553         return comma_fprintf(fp, first, " %s: %" PRIu64, field, value);
1554 }
1555
1556 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field)
1557
1558 struct bit_names {
1559         int bit;
1560         const char *name;
1561 };
1562
1563 static int bits__fprintf(FILE *fp, const char *field, u64 value,
1564                          struct bit_names *bits, bool *first)
1565 {
1566         int i = 0, printed = comma_fprintf(fp, first, " %s: ", field);
1567         bool first_bit = true;
1568
1569         do {
1570                 if (value & bits[i].bit) {
1571                         printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name);
1572                         first_bit = false;
1573                 }
1574         } while (bits[++i].name != NULL);
1575
1576         return printed;
1577 }
1578
1579 static int sample_type__fprintf(FILE *fp, bool *first, u64 value)
1580 {
1581 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1582         struct bit_names bits[] = {
1583                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1584                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1585                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1586                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1587                 { .name = NULL, }
1588         };
1589 #undef bit_name
1590         return bits__fprintf(fp, "sample_type", value, bits, first);
1591 }
1592
1593 static int read_format__fprintf(FILE *fp, bool *first, u64 value)
1594 {
1595 #define bit_name(n) { PERF_FORMAT_##n, #n }
1596         struct bit_names bits[] = {
1597                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1598                 bit_name(ID), bit_name(GROUP),
1599                 { .name = NULL, }
1600         };
1601 #undef bit_name
1602         return bits__fprintf(fp, "read_format", value, bits, first);
1603 }
1604
1605 int perf_evsel__fprintf(struct perf_evsel *evsel,
1606                         struct perf_attr_details *details, FILE *fp)
1607 {
1608         bool first = true;
1609         int printed = 0;
1610
1611         if (details->event_group) {
1612                 struct perf_evsel *pos;
1613
1614                 if (!perf_evsel__is_group_leader(evsel))
1615                         return 0;
1616
1617                 if (evsel->nr_members > 1)
1618                         printed += fprintf(fp, "%s{", evsel->group_name ?: "");
1619
1620                 printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1621                 for_each_group_member(pos, evsel)
1622                         printed += fprintf(fp, ",%s", perf_evsel__name(pos));
1623
1624                 if (evsel->nr_members > 1)
1625                         printed += fprintf(fp, "}");
1626                 goto out;
1627         }
1628
1629         printed += fprintf(fp, "%s", perf_evsel__name(evsel));
1630
1631         if (details->verbose || details->freq) {
1632                 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64,
1633                                          (u64)evsel->attr.sample_freq);
1634         }
1635
1636         if (details->verbose) {
1637                 if_print(type);
1638                 if_print(config);
1639                 if_print(config1);
1640                 if_print(config2);
1641                 if_print(size);
1642                 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type);
1643                 if (evsel->attr.read_format)
1644                         printed += read_format__fprintf(fp, &first, evsel->attr.read_format);
1645                 if_print(disabled);
1646                 if_print(inherit);
1647                 if_print(pinned);
1648                 if_print(exclusive);
1649                 if_print(exclude_user);
1650                 if_print(exclude_kernel);
1651                 if_print(exclude_hv);
1652                 if_print(exclude_idle);
1653                 if_print(mmap);
1654                 if_print(comm);
1655                 if_print(freq);
1656                 if_print(inherit_stat);
1657                 if_print(enable_on_exec);
1658                 if_print(task);
1659                 if_print(watermark);
1660                 if_print(precise_ip);
1661                 if_print(mmap_data);
1662                 if_print(sample_id_all);
1663                 if_print(exclude_host);
1664                 if_print(exclude_guest);
1665                 if_print(__reserved_1);
1666                 if_print(wakeup_events);
1667                 if_print(bp_type);
1668                 if_print(branch_sample_type);
1669         }
1670 out:
1671         fputc('\n', fp);
1672         return ++printed;
1673 }
1674
1675 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
1676                           char *msg, size_t msgsize)
1677 {
1678         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
1679             evsel->attr.type   == PERF_TYPE_HARDWARE &&
1680             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
1681                 /*
1682                  * If it's cycles then fall back to hrtimer based
1683                  * cpu-clock-tick sw counter, which is always available even if
1684                  * no PMU support.
1685                  *
1686                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
1687                  * b0a873e).
1688                  */
1689                 scnprintf(msg, msgsize, "%s",
1690 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
1691
1692                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
1693                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
1694
1695                 free(evsel->name);
1696                 evsel->name = NULL;
1697                 return true;
1698         }
1699
1700         return false;
1701 }
1702
1703 int perf_evsel__open_strerror(struct perf_evsel *evsel,
1704                               struct perf_target *target,
1705                               int err, char *msg, size_t size)
1706 {
1707         switch (err) {
1708         case EPERM:
1709         case EACCES:
1710                 return scnprintf(msg, size,
1711                  "You may not have permission to collect %sstats.\n"
1712                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n"
1713                  " -1 - Not paranoid at all\n"
1714                  "  0 - Disallow raw tracepoint access for unpriv\n"
1715                  "  1 - Disallow cpu events for unpriv\n"
1716                  "  2 - Disallow kernel profiling for unpriv",
1717                                  target->system_wide ? "system-wide " : "");
1718         case ENOENT:
1719                 return scnprintf(msg, size, "The %s event is not supported.",
1720                                  perf_evsel__name(evsel));
1721         case EMFILE:
1722                 return scnprintf(msg, size, "%s",
1723                          "Too many events are opened.\n"
1724                          "Try again after reducing the number of events.");
1725         case ENODEV:
1726                 if (target->cpu_list)
1727                         return scnprintf(msg, size, "%s",
1728          "No such device - did you specify an out-of-range profile CPU?\n");
1729                 break;
1730         case EOPNOTSUPP:
1731                 if (evsel->attr.precise_ip)
1732                         return scnprintf(msg, size, "%s",
1733         "\'precise\' request may not be supported. Try removing 'p' modifier.");
1734 #if defined(__i386__) || defined(__x86_64__)
1735                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
1736                         return scnprintf(msg, size, "%s",
1737         "No hardware sampling interrupt available.\n"
1738         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
1739 #endif
1740                 break;
1741         default:
1742                 break;
1743         }
1744
1745         return scnprintf(msg, size,
1746         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).  \n"
1747         "/bin/dmesg may provide additional information.\n"
1748         "No CONFIG_PERF_EVENTS=y kernel support configured?\n",
1749                          err, strerror(err), perf_evsel__name(evsel));
1750 }