tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/compiler.h>
  19 #include <linux/err.h>
  20 #include <linux/zalloc.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include <stdlib.h>
  26 #include <perf/evsel.h>
  27 #include "asm/bug.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "evlist.h"
  34 #include "cpumap.h"
  35 #include "thread_map.h"
  36 #include "target.h"
  37 #include "perf_regs.h"
  38 #include "record.h"
  39 #include "debug.h"
  40 #include "trace-event.h"
  41 #include "stat.h"
  42 #include "string2.h"
  43 #include "memswap.h"
  44 #include "util.h"
  45 #include "../perf-sys.h"
  46 #include "util/parse-branch-options.h"
  47 #include <internal/xyarray.h>
  48
  49 #include <linux/ctype.h>
  50
  51 struct perf_missing_features perf_missing_features;
  52
  53 static clockid_t clockid;
  54
  55 static int perf_evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  56 {
  57         return 0;
  58 }
  59
  60 void __weak test_attr__ready(void) { }
  61
  62 static void perf_evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  63 {
  64 }
  65
  66 static struct {
  67         size_t  size;
  68         int     (*init)(struct evsel *evsel);
  69         void    (*fini)(struct evsel *evsel);
  70 } perf_evsel__object = {
  71         .size = sizeof(struct evsel),
  72         .init = perf_evsel__no_extra_init,
  73         .fini = perf_evsel__no_extra_fini,
  74 };
  75
  76 int perf_evsel__object_config(size_t object_size,
  77                               int (*init)(struct evsel *evsel),
  78                               void (*fini)(struct evsel *evsel))
  79 {
  80
  81         if (object_size == 0)
  82                 goto set_methods;
  83
  84         if (perf_evsel__object.size > object_size)
  85                 return -EINVAL;
  86
  87         perf_evsel__object.size = object_size;
  88
  89 set_methods:
  90         if (init != NULL)
  91                 perf_evsel__object.init = init;
  92
  93         if (fini != NULL)
  94                 perf_evsel__object.fini = fini;
  95
  96         return 0;
  97 }
  98
  99 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 100
 101 int __perf_evsel__sample_size(u64 sample_type)
 102 {
 103         u64 mask = sample_type & PERF_SAMPLE_MASK;
 104         int size = 0;
 105         int i;
 106
 107         for (i = 0; i < 64; i++) {
 108                 if (mask & (1ULL << i))
 109                         size++;
 110         }
 111
 112         size *= sizeof(u64);
 113
 114         return size;
 115 }
 116
 117 /**
 118  * __perf_evsel__calc_id_pos - calculate id_pos.
 119  * @sample_type: sample type
 120  *
 121  * This function returns the position of the event id (PERF_SAMPLE_ID or
 122  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 123  * perf_record_sample.
 124  */
 125 static int __perf_evsel__calc_id_pos(u64 sample_type)
 126 {
 127         int idx = 0;
 128
 129         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 130                 return 0;
 131
 132         if (!(sample_type & PERF_SAMPLE_ID))
 133                 return -1;
 134
 135         if (sample_type & PERF_SAMPLE_IP)
 136                 idx += 1;
 137
 138         if (sample_type & PERF_SAMPLE_TID)
 139                 idx += 1;
 140
 141         if (sample_type & PERF_SAMPLE_TIME)
 142                 idx += 1;
 143
 144         if (sample_type & PERF_SAMPLE_ADDR)
 145                 idx += 1;
 146
 147         return idx;
 148 }
 149
 150 /**
 151  * __perf_evsel__calc_is_pos - calculate is_pos.
 152  * @sample_type: sample type
 153  *
 154  * This function returns the position (counting backwards) of the event id
 155  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 156  * sample_id_all is used there is an id sample appended to non-sample events.
 157  */
 158 static int __perf_evsel__calc_is_pos(u64 sample_type)
 159 {
 160         int idx = 1;
 161
 162         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 163                 return 1;
 164
 165         if (!(sample_type & PERF_SAMPLE_ID))
 166                 return -1;
 167
 168         if (sample_type & PERF_SAMPLE_CPU)
 169                 idx += 1;
 170
 171         if (sample_type & PERF_SAMPLE_STREAM_ID)
 172                 idx += 1;
 173
 174         return idx;
 175 }
 176
 177 void perf_evsel__calc_id_pos(struct evsel *evsel)
 178 {
 179         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 180         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 181 }
 182
 183 void __perf_evsel__set_sample_bit(struct evsel *evsel,
 184                                   enum perf_event_sample_format bit)
 185 {
 186         if (!(evsel->core.attr.sample_type & bit)) {
 187                 evsel->core.attr.sample_type |= bit;
 188                 evsel->sample_size += sizeof(u64);
 189                 perf_evsel__calc_id_pos(evsel);
 190         }
 191 }
 192
 193 void __perf_evsel__reset_sample_bit(struct evsel *evsel,
 194                                     enum perf_event_sample_format bit)
 195 {
 196         if (evsel->core.attr.sample_type & bit) {
 197                 evsel->core.attr.sample_type &= ~bit;
 198                 evsel->sample_size -= sizeof(u64);
 199                 perf_evsel__calc_id_pos(evsel);
 200         }
 201 }
 202
 203 void perf_evsel__set_sample_id(struct evsel *evsel,
 204                                bool can_sample_identifier)
 205 {
 206         if (can_sample_identifier) {
 207                 perf_evsel__reset_sample_bit(evsel, ID);
 208                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 209         } else {
 210                 perf_evsel__set_sample_bit(evsel, ID);
 211         }
 212         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 213 }
 214
 215 /**
 216  * perf_evsel__is_function_event - Return whether given evsel is a function
 217  * trace event
 218  *
 219  * @evsel - evsel selector to be tested
 220  *
 221  * Return %true if event is function trace event
 222  */
 223 bool perf_evsel__is_function_event(struct evsel *evsel)
 224 {
 225 #define FUNCTION_EVENT "ftrace:function"
 226
 227         return evsel->name &&
 228                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 229
 230 #undef FUNCTION_EVENT
 231 }
 232
 233 void evsel__init(struct evsel *evsel,
 234                  struct perf_event_attr *attr, int idx)
 235 {
 236         perf_evsel__init(&evsel->core, attr);
 237         evsel->idx         = idx;
 238         evsel->tracking    = !idx;
 239         evsel->leader      = evsel;
 240         evsel->unit        = "";
 241         evsel->scale       = 1.0;
 242         evsel->max_events  = ULONG_MAX;
 243         evsel->evlist      = NULL;
 244         evsel->bpf_obj     = NULL;
 245         evsel->bpf_fd      = -1;
 246         INIT_LIST_HEAD(&evsel->config_terms);
 247         perf_evsel__object.init(evsel);
 248         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 249         perf_evsel__calc_id_pos(evsel);
 250         evsel->cmdline_group_boundary = false;
 251         evsel->metric_expr   = NULL;
 252         evsel->metric_name   = NULL;
 253         evsel->metric_events = NULL;
 254         evsel->collect_stat  = false;
 255         evsel->pmu_name      = NULL;
 256 }
 257
 258 struct evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 259 {
 260         struct evsel *evsel = zalloc(perf_evsel__object.size);
 261
 262         if (!evsel)
 263                 return NULL;
 264         evsel__init(evsel, attr, idx);
 265
 266         if (perf_evsel__is_bpf_output(evsel)) {
 267                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 268                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 269                 evsel->core.attr.sample_period = 1;
 270         }
 271
 272         if (perf_evsel__is_clock(evsel)) {
 273                 /*
 274                  * The evsel->unit points to static alias->unit
 275                  * so it's ok to use static string in here.
 276                  */
 277                 static const char *unit = "msec";
 278
 279                 evsel->unit = unit;
 280                 evsel->scale = 1e-6;
 281         }
 282
 283         return evsel;
 284 }
 285
 286 static bool perf_event_can_profile_kernel(void)
 287 {
 288         return perf_event_paranoid_check(1);
 289 }
 290
 291 struct evsel *perf_evsel__new_cycles(bool precise)
 292 {
 293         struct perf_event_attr attr = {
 294                 .type   = PERF_TYPE_HARDWARE,
 295                 .config = PERF_COUNT_HW_CPU_CYCLES,
 296                 .exclude_kernel = !perf_event_can_profile_kernel(),
 297         };
 298         struct evsel *evsel;
 299
 300         event_attr_init(&attr);
 301
 302         if (!precise)
 303                 goto new_event;
 304
 305         /*
 306          * Now let the usual logic to set up the perf_event_attr defaults
 307          * to kick in when we return and before perf_evsel__open() is called.
 308          */
 309 new_event:
 310         evsel = evsel__new(&attr);
 311         if (evsel == NULL)
 312                 goto out;
 313
 314         evsel->precise_max = true;
 315
 316         /* use asprintf() because free(evsel) assumes name is allocated */
 317         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 318                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 319                      attr.exclude_kernel ? "u" : "",
 320                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 321                 goto error_free;
 322 out:
 323         return evsel;
 324 error_free:
 325         evsel__delete(evsel);
 326         evsel = NULL;
 327         goto out;
 328 }
 329
 330 /*
 331  * Returns pointer with encoded error via <linux/err.h> interface.
 332  */
 333 struct evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 334 {
 335         struct evsel *evsel = zalloc(perf_evsel__object.size);
 336         int err = -ENOMEM;
 337
 338         if (evsel == NULL) {
 339                 goto out_err;
 340         } else {
 341                 struct perf_event_attr attr = {
 342                         .type          = PERF_TYPE_TRACEPOINT,
 343                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 344                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 345                 };
 346
 347                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 348                         goto out_free;
 349
 350                 evsel->tp_format = trace_event__tp_format(sys, name);
 351                 if (IS_ERR(evsel->tp_format)) {
 352                         err = PTR_ERR(evsel->tp_format);
 353                         goto out_free;
 354                 }
 355
 356                 event_attr_init(&attr);
 357                 attr.config = evsel->tp_format->id;
 358                 attr.sample_period = 1;
 359                 evsel__init(evsel, &attr, idx);
 360         }
 361
 362         return evsel;
 363
 364 out_free:
 365         zfree(&evsel->name);
 366         free(evsel);
 367 out_err:
 368         return ERR_PTR(err);
 369 }
 370
 371 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 372         "cycles",
 373         "instructions",
 374         "cache-references",
 375         "cache-misses",
 376         "branches",
 377         "branch-misses",
 378         "bus-cycles",
 379         "stalled-cycles-frontend",
 380         "stalled-cycles-backend",
 381         "ref-cycles",
 382 };
 383
 384 static const char *__perf_evsel__hw_name(u64 config)
 385 {
 386         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 387                 return perf_evsel__hw_names[config];
 388
 389         return "unknown-hardware";
 390 }
 391
 392 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 393 {
 394         int colon = 0, r = 0;
 395         struct perf_event_attr *attr = &evsel->core.attr;
 396         bool exclude_guest_default = false;
 397
 398 #define MOD_PRINT(context, mod) do {                                    \
 399                 if (!attr->exclude_##context) {                         \
 400                         if (!colon) colon = ++r;                        \
 401                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 402                 } } while(0)
 403
 404         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 405                 MOD_PRINT(kernel, 'k');
 406                 MOD_PRINT(user, 'u');
 407                 MOD_PRINT(hv, 'h');
 408                 exclude_guest_default = true;
 409         }
 410
 411         if (attr->precise_ip) {
 412                 if (!colon)
 413                         colon = ++r;
 414                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 415                 exclude_guest_default = true;
 416         }
 417
 418         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 419                 MOD_PRINT(host, 'H');
 420                 MOD_PRINT(guest, 'G');
 421         }
 422 #undef MOD_PRINT
 423         if (colon)
 424                 bf[colon - 1] = ':';
 425         return r;
 426 }
 427
 428 static int perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 429 {
 430         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config));
 431         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 432 }
 433
 434 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 435         "cpu-clock",
 436         "task-clock",
 437         "page-faults",
 438         "context-switches",
 439         "cpu-migrations",
 440         "minor-faults",
 441         "major-faults",
 442         "alignment-faults",
 443         "emulation-faults",
 444         "dummy",
 445 };
 446
 447 static const char *__perf_evsel__sw_name(u64 config)
 448 {
 449         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 450                 return perf_evsel__sw_names[config];
 451         return "unknown-software";
 452 }
 453
 454 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 455 {
 456         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config));
 457         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 458 }
 459
 460 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 461 {
 462         int r;
 463
 464         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 465
 466         if (type & HW_BREAKPOINT_R)
 467                 r += scnprintf(bf + r, size - r, "r");
 468
 469         if (type & HW_BREAKPOINT_W)
 470                 r += scnprintf(bf + r, size - r, "w");
 471
 472         if (type & HW_BREAKPOINT_X)
 473                 r += scnprintf(bf + r, size - r, "x");
 474
 475         return r;
 476 }
 477
 478 static int perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 479 {
 480         struct perf_event_attr *attr = &evsel->core.attr;
 481         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 482         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 483 }
 484
 485 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 486                                 [PERF_EVSEL__MAX_ALIASES] = {
 487  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 488  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 489  { "LLC",       "L2",                                                   },
 490  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 491  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 492  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 493  { "node",                                                              },
 494 };
 495
 496 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 497                                    [PERF_EVSEL__MAX_ALIASES] = {
 498  { "load",      "loads",        "read",                                 },
 499  { "store",     "stores",       "write",                                },
 500  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 501 };
 502
 503 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 504                                        [PERF_EVSEL__MAX_ALIASES] = {
 505  { "refs",      "Reference",    "ops",          "access",               },
 506  { "misses",    "miss",                                                 },
 507 };
 508
 509 #define C(x)            PERF_COUNT_HW_CACHE_##x
 510 #define CACHE_READ      (1 << C(OP_READ))
 511 #define CACHE_WRITE     (1 << C(OP_WRITE))
 512 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 513 #define COP(x)          (1 << x)
 514
 515 /*
 516  * cache operartion stat
 517  * L1I : Read and prefetch only
 518  * ITLB and BPU : Read-only
 519  */
 520 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 521  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 522  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 523  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 524  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 525  [C(ITLB)]      = (CACHE_READ),
 526  [C(BPU)]       = (CACHE_READ),
 527  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 528 };
 529
 530 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 531 {
 532         if (perf_evsel__hw_cache_stat[type] & COP(op))
 533                 return true;    /* valid */
 534         else
 535                 return false;   /* invalid */
 536 }
 537
 538 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 539                                             char *bf, size_t size)
 540 {
 541         if (result) {
 542                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 543                                  perf_evsel__hw_cache_op[op][0],
 544                                  perf_evsel__hw_cache_result[result][0]);
 545         }
 546
 547         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 548                          perf_evsel__hw_cache_op[op][1]);
 549 }
 550
 551 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 552 {
 553         u8 op, result, type = (config >>  0) & 0xff;
 554         const char *err = "unknown-ext-hardware-cache-type";
 555
 556         if (type >= PERF_COUNT_HW_CACHE_MAX)
 557                 goto out_err;
 558
 559         op = (config >>  8) & 0xff;
 560         err = "unknown-ext-hardware-cache-op";
 561         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 562                 goto out_err;
 563
 564         result = (config >> 16) & 0xff;
 565         err = "unknown-ext-hardware-cache-result";
 566         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 567                 goto out_err;
 568
 569         err = "invalid-cache";
 570         if (!perf_evsel__is_cache_op_valid(type, op))
 571                 goto out_err;
 572
 573         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 574 out_err:
 575         return scnprintf(bf, size, "%s", err);
 576 }
 577
 578 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 579 {
 580         int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 581         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 582 }
 583
 584 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 585 {
 586         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 587         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 588 }
 589
 590 static int perf_evsel__tool_name(char *bf, size_t size)
 591 {
 592         int ret = scnprintf(bf, size, "duration_time");
 593         return ret;
 594 }
 595
 596 const char *perf_evsel__name(struct evsel *evsel)
 597 {
 598         char bf[128];
 599
 600         if (!evsel)
 601                 goto out_unknown;
 602
 603         if (evsel->name)
 604                 return evsel->name;
 605
 606         switch (evsel->core.attr.type) {
 607         case PERF_TYPE_RAW:
 608                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 609                 break;
 610
 611         case PERF_TYPE_HARDWARE:
 612                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 613                 break;
 614
 615         case PERF_TYPE_HW_CACHE:
 616                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 617                 break;
 618
 619         case PERF_TYPE_SOFTWARE:
 620                 if (evsel->tool_event)
 621                         perf_evsel__tool_name(bf, sizeof(bf));
 622                 else
 623                         perf_evsel__sw_name(evsel, bf, sizeof(bf));
 624                 break;
 625
 626         case PERF_TYPE_TRACEPOINT:
 627                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 628                 break;
 629
 630         case PERF_TYPE_BREAKPOINT:
 631                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 632                 break;
 633
 634         default:
 635                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 636                           evsel->core.attr.type);
 637                 break;
 638         }
 639
 640         evsel->name = strdup(bf);
 641
 642         if (evsel->name)
 643                 return evsel->name;
 644 out_unknown:
 645         return "unknown";
 646 }
 647
 648 const char *perf_evsel__group_name(struct evsel *evsel)
 649 {
 650         return evsel->group_name ?: "anon group";
 651 }
 652
 653 /*
 654  * Returns the group details for the specified leader,
 655  * with following rules.
 656  *
 657  *  For record -e '{cycles,instructions}'
 658  *    'anon group { cycles:u, instructions:u }'
 659  *
 660  *  For record -e 'cycles,instructions' and report --group
 661  *    'cycles:u, instructions:u'
 662  */
 663 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 664 {
 665         int ret = 0;
 666         struct evsel *pos;
 667         const char *group_name = perf_evsel__group_name(evsel);
 668
 669         if (!evsel->forced_leader)
 670                 ret = scnprintf(buf, size, "%s { ", group_name);
 671
 672         ret += scnprintf(buf + ret, size - ret, "%s",
 673                          perf_evsel__name(evsel));
 674
 675         for_each_group_member(pos, evsel)
 676                 ret += scnprintf(buf + ret, size - ret, ", %s",
 677                                  perf_evsel__name(pos));
 678
 679         if (!evsel->forced_leader)
 680                 ret += scnprintf(buf + ret, size - ret, " }");
 681
 682         return ret;
 683 }
 684
 685 static void __perf_evsel__config_callchain(struct evsel *evsel,
 686                                            struct record_opts *opts,
 687                                            struct callchain_param *param)
 688 {
 689         bool function = perf_evsel__is_function_event(evsel);
 690         struct perf_event_attr *attr = &evsel->core.attr;
 691
 692         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 693
 694         attr->sample_max_stack = param->max_stack;
 695
 696         if (opts->kernel_callchains)
 697                 attr->exclude_callchain_user = 1;
 698         if (opts->user_callchains)
 699                 attr->exclude_callchain_kernel = 1;
 700         if (param->record_mode == CALLCHAIN_LBR) {
 701                 if (!opts->branch_stack) {
 702                         if (attr->exclude_user) {
 703                                 pr_warning("LBR callstack option is only available "
 704                                            "to get user callchain information. "
 705                                            "Falling back to framepointers.\n");
 706                         } else {
 707                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 708                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 709                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 710                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 711                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 712                         }
 713                 } else
 714                          pr_warning("Cannot use LBR callstack with branch stack. "
 715                                     "Falling back to framepointers.\n");
 716         }
 717
 718         if (param->record_mode == CALLCHAIN_DWARF) {
 719                 if (!function) {
 720                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 721                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 722                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 723                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 724                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 725                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 726                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 727                         } else {
 728                                 attr->sample_regs_user |= PERF_REGS_MASK;
 729                         }
 730                         attr->sample_stack_user = param->dump_size;
 731                         attr->exclude_callchain_user = 1;
 732                 } else {
 733                         pr_info("Cannot use DWARF unwind for function trace event,"
 734                                 " falling back to framepointers.\n");
 735                 }
 736         }
 737
 738         if (function) {
 739                 pr_info("Disabling user space callchains for function trace event.\n");
 740                 attr->exclude_callchain_user = 1;
 741         }
 742 }
 743
 744 void perf_evsel__config_callchain(struct evsel *evsel,
 745                                   struct record_opts *opts,
 746                                   struct callchain_param *param)
 747 {
 748         if (param->enabled)
 749                 return __perf_evsel__config_callchain(evsel, opts, param);
 750 }
 751
 752 static void
 753 perf_evsel__reset_callgraph(struct evsel *evsel,
 754                             struct callchain_param *param)
 755 {
 756         struct perf_event_attr *attr = &evsel->core.attr;
 757
 758         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 759         if (param->record_mode == CALLCHAIN_LBR) {
 760                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 761                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 762                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 763         }
 764         if (param->record_mode == CALLCHAIN_DWARF) {
 765                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 766                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 767         }
 768 }
 769
 770 static void apply_config_terms(struct evsel *evsel,
 771                                struct record_opts *opts, bool track)
 772 {
 773         struct perf_evsel_config_term *term;
 774         struct list_head *config_terms = &evsel->config_terms;
 775         struct perf_event_attr *attr = &evsel->core.attr;
 776         /* callgraph default */
 777         struct callchain_param param = {
 778                 .record_mode = callchain_param.record_mode,
 779         };
 780         u32 dump_size = 0;
 781         int max_stack = 0;
 782         const char *callgraph_buf = NULL;
 783
 784         list_for_each_entry(term, config_terms, list) {
 785                 switch (term->type) {
 786                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 787                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 788                                 attr->sample_period = term->val.period;
 789                                 attr->freq = 0;
 790                                 perf_evsel__reset_sample_bit(evsel, PERIOD);
 791                         }
 792                         break;
 793                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 794                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 795                                 attr->sample_freq = term->val.freq;
 796                                 attr->freq = 1;
 797                                 perf_evsel__set_sample_bit(evsel, PERIOD);
 798                         }
 799                         break;
 800                 case PERF_EVSEL__CONFIG_TERM_TIME:
 801                         if (term->val.time)
 802                                 perf_evsel__set_sample_bit(evsel, TIME);
 803                         else
 804                                 perf_evsel__reset_sample_bit(evsel, TIME);
 805                         break;
 806                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 807                         callgraph_buf = term->val.callgraph;
 808                         break;
 809                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 810                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 811                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 812                                 parse_branch_str(term->val.branch,
 813                                                  &attr->branch_sample_type);
 814                         } else
 815                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 816                         break;
 817                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 818                         dump_size = term->val.stack_user;
 819                         break;
 820                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 821                         max_stack = term->val.max_stack;
 822                         break;
 823                 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS:
 824                         evsel->max_events = term->val.max_events;
 825                         break;
 826                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 827                         /*
 828                          * attr->inherit should has already been set by
 829                          * perf_evsel__config. If user explicitly set
 830                          * inherit using config terms, override global
 831                          * opt->no_inherit setting.
 832                          */
 833                         attr->inherit = term->val.inherit ? 1 : 0;
 834                         break;
 835                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 836                         attr->write_backward = term->val.overwrite ? 1 : 0;
 837                         break;
 838                 case PERF_EVSEL__CONFIG_TERM_DRV_CFG:
 839                         break;
 840                 case PERF_EVSEL__CONFIG_TERM_PERCORE:
 841                         break;
 842                 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT:
 843                         attr->aux_output = term->val.aux_output ? 1 : 0;
 844                         break;
 845                 default:
 846                         break;
 847                 }
 848         }
 849
 850         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 851         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 852                 bool sample_address = false;
 853
 854                 if (max_stack) {
 855                         param.max_stack = max_stack;
 856                         if (callgraph_buf == NULL)
 857                                 callgraph_buf = "fp";
 858                 }
 859
 860                 /* parse callgraph parameters */
 861                 if (callgraph_buf != NULL) {
 862                         if (!strcmp(callgraph_buf, "no")) {
 863                                 param.enabled = false;
 864                                 param.record_mode = CALLCHAIN_NONE;
 865                         } else {
 866                                 param.enabled = true;
 867                                 if (parse_callchain_record(callgraph_buf, &param)) {
 868                                         pr_err("per-event callgraph setting for %s failed. "
 869                                                "Apply callgraph global setting for it\n",
 870                                                evsel->name);
 871                                         return;
 872                                 }
 873                                 if (param.record_mode == CALLCHAIN_DWARF)
 874                                         sample_address = true;
 875                         }
 876                 }
 877                 if (dump_size > 0) {
 878                         dump_size = round_up(dump_size, sizeof(u64));
 879                         param.dump_size = dump_size;
 880                 }
 881
 882                 /* If global callgraph set, clear it */
 883                 if (callchain_param.enabled)
 884                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 885
 886                 /* set perf-event callgraph */
 887                 if (param.enabled) {
 888                         if (sample_address) {
 889                                 perf_evsel__set_sample_bit(evsel, ADDR);
 890                                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 891                                 evsel->core.attr.mmap_data = track;
 892                         }
 893                         perf_evsel__config_callchain(evsel, opts, &param);
 894                 }
 895         }
 896 }
 897
 898 static bool is_dummy_event(struct evsel *evsel)
 899 {
 900         return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) &&
 901                (evsel->core.attr.config == PERF_COUNT_SW_DUMMY);
 902 }
 903
 904 /*
 905  * The enable_on_exec/disabled value strategy:
 906  *
 907  *  1) For any type of traced program:
 908  *    - all independent events and group leaders are disabled
 909  *    - all group members are enabled
 910  *
 911  *     Group members are ruled by group leaders. They need to
 912  *     be enabled, because the group scheduling relies on that.
 913  *
 914  *  2) For traced programs executed by perf:
 915  *     - all independent events and group leaders have
 916  *       enable_on_exec set
 917  *     - we don't specifically enable or disable any event during
 918  *       the record command
 919  *
 920  *     Independent events and group leaders are initially disabled
 921  *     and get enabled by exec. Group members are ruled by group
 922  *     leaders as stated in 1).
 923  *
 924  *  3) For traced programs attached by perf (pid/tid):
 925  *     - we specifically enable or disable all events during
 926  *       the record command
 927  *
 928  *     When attaching events to already running traced we
 929  *     enable/disable events specifically, as there's no
 930  *     initial traced exec call.
 931  */
 932 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts,
 933                         struct callchain_param *callchain)
 934 {
 935         struct evsel *leader = evsel->leader;
 936         struct perf_event_attr *attr = &evsel->core.attr;
 937         int track = evsel->tracking;
 938         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 939
 940         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 941         attr->inherit       = !opts->no_inherit;
 942         attr->write_backward = opts->overwrite ? 1 : 0;
 943
 944         perf_evsel__set_sample_bit(evsel, IP);
 945         perf_evsel__set_sample_bit(evsel, TID);
 946
 947         if (evsel->sample_read) {
 948                 perf_evsel__set_sample_bit(evsel, READ);
 949
 950                 /*
 951                  * We need ID even in case of single event, because
 952                  * PERF_SAMPLE_READ process ID specific data.
 953                  */
 954                 perf_evsel__set_sample_id(evsel, false);
 955
 956                 /*
 957                  * Apply group format only if we belong to group
 958                  * with more than one members.
 959                  */
 960                 if (leader->core.nr_members > 1) {
 961                         attr->read_format |= PERF_FORMAT_GROUP;
 962                         attr->inherit = 0;
 963                 }
 964         }
 965
 966         /*
 967          * We default some events to have a default interval. But keep
 968          * it a weak assumption overridable by the user.
 969          */
 970         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 971                                      opts->user_interval != ULLONG_MAX)) {
 972                 if (opts->freq) {
 973                         perf_evsel__set_sample_bit(evsel, PERIOD);
 974                         attr->freq              = 1;
 975                         attr->sample_freq       = opts->freq;
 976                 } else {
 977                         attr->sample_period = opts->default_interval;
 978                 }
 979         }
 980
 981         /*
 982          * Disable sampling for all group members other
 983          * than leader in case leader 'leads' the sampling.
 984          */
 985         if ((leader != evsel) && leader->sample_read) {
 986                 attr->freq           = 0;
 987                 attr->sample_freq    = 0;
 988                 attr->sample_period  = 0;
 989                 attr->write_backward = 0;
 990
 991                 /*
 992                  * We don't get sample for slave events, we make them
 993                  * when delivering group leader sample. Set the slave
 994                  * event to follow the master sample_type to ease up
 995                  * report.
 996                  */
 997                 attr->sample_type = leader->core.attr.sample_type;
 998         }
 999
1000         if (opts->no_samples)
1001                 attr->sample_freq = 0;
1002
1003         if (opts->inherit_stat) {
1004                 evsel->core.attr.read_format |=
1005                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1006                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1007                         PERF_FORMAT_ID;
1008                 attr->inherit_stat = 1;
1009         }
1010
1011         if (opts->sample_address) {
1012                 perf_evsel__set_sample_bit(evsel, ADDR);
1013                 attr->mmap_data = track;
1014         }
1015
1016         /*
1017          * We don't allow user space callchains for  function trace
1018          * event, due to issues with page faults while tracing page
1019          * fault handler and its overall trickiness nature.
1020          */
1021         if (perf_evsel__is_function_event(evsel))
1022                 evsel->core.attr.exclude_callchain_user = 1;
1023
1024         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1025                 perf_evsel__config_callchain(evsel, opts, callchain);
1026
1027         if (opts->sample_intr_regs) {
1028                 attr->sample_regs_intr = opts->sample_intr_regs;
1029                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
1030         }
1031
1032         if (opts->sample_user_regs) {
1033                 attr->sample_regs_user |= opts->sample_user_regs;
1034                 perf_evsel__set_sample_bit(evsel, REGS_USER);
1035         }
1036
1037         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1038                 perf_evsel__set_sample_bit(evsel, CPU);
1039
1040         /*
1041          * When the user explicitly disabled time don't force it here.
1042          */
1043         if (opts->sample_time &&
1044             (!perf_missing_features.sample_id_all &&
1045             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1046              opts->sample_time_set)))
1047                 perf_evsel__set_sample_bit(evsel, TIME);
1048
1049         if (opts->raw_samples && !evsel->no_aux_samples) {
1050                 perf_evsel__set_sample_bit(evsel, TIME);
1051                 perf_evsel__set_sample_bit(evsel, RAW);
1052                 perf_evsel__set_sample_bit(evsel, CPU);
1053         }
1054
1055         if (opts->sample_address)
1056                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
1057
1058         if (opts->sample_phys_addr)
1059                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
1060
1061         if (opts->no_buffering) {
1062                 attr->watermark = 0;
1063                 attr->wakeup_events = 1;
1064         }
1065         if (opts->branch_stack && !evsel->no_aux_samples) {
1066                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
1067                 attr->branch_sample_type = opts->branch_stack;
1068         }
1069
1070         if (opts->sample_weight)
1071                 perf_evsel__set_sample_bit(evsel, WEIGHT);
1072
1073         attr->task  = track;
1074         attr->mmap  = track;
1075         attr->mmap2 = track && !perf_missing_features.mmap2;
1076         attr->comm  = track;
1077         attr->ksymbol = track && !perf_missing_features.ksymbol;
1078         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1079
1080         if (opts->record_namespaces)
1081                 attr->namespaces  = track;
1082
1083         if (opts->record_switch_events)
1084                 attr->context_switch = track;
1085
1086         if (opts->sample_transaction)
1087                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1088
1089         if (opts->running_time) {
1090                 evsel->core.attr.read_format |=
1091                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1092                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1093         }
1094
1095         /*
1096          * XXX see the function comment above
1097          *
1098          * Disabling only independent events or group leaders,
1099          * keeping group members enabled.
1100          */
1101         if (perf_evsel__is_group_leader(evsel))
1102                 attr->disabled = 1;
1103
1104         /*
1105          * Setting enable_on_exec for independent events and
1106          * group leaders for traced executed by perf.
1107          */
1108         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1109                 !opts->initial_delay)
1110                 attr->enable_on_exec = 1;
1111
1112         if (evsel->immediate) {
1113                 attr->disabled = 0;
1114                 attr->enable_on_exec = 0;
1115         }
1116
1117         clockid = opts->clockid;
1118         if (opts->use_clockid) {
1119                 attr->use_clockid = 1;
1120                 attr->clockid = opts->clockid;
1121         }
1122
1123         if (evsel->precise_max)
1124                 attr->precise_ip = 3;
1125
1126         if (opts->all_user) {
1127                 attr->exclude_kernel = 1;
1128                 attr->exclude_user   = 0;
1129         }
1130
1131         if (opts->all_kernel) {
1132                 attr->exclude_kernel = 0;
1133                 attr->exclude_user   = 1;
1134         }
1135
1136         if (evsel->core.own_cpus || evsel->unit)
1137                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1138
1139         /*
1140          * Apply event specific term settings,
1141          * it overloads any global configuration.
1142          */
1143         apply_config_terms(evsel, opts, track);
1144
1145         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1146
1147         /* The --period option takes the precedence. */
1148         if (opts->period_set) {
1149                 if (opts->period)
1150                         perf_evsel__set_sample_bit(evsel, PERIOD);
1151                 else
1152                         perf_evsel__reset_sample_bit(evsel, PERIOD);
1153         }
1154
1155         /*
1156          * For initial_delay, a dummy event is added implicitly.
1157          * The software event will trigger -EOPNOTSUPP error out,
1158          * if BRANCH_STACK bit is set.
1159          */
1160         if (opts->initial_delay && is_dummy_event(evsel))
1161                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
1162 }
1163
1164 int perf_evsel__set_filter(struct evsel *evsel, const char *filter)
1165 {
1166         char *new_filter = strdup(filter);
1167
1168         if (new_filter != NULL) {
1169                 free(evsel->filter);
1170                 evsel->filter = new_filter;
1171                 return 0;
1172         }
1173
1174         return -1;
1175 }
1176
1177 static int perf_evsel__append_filter(struct evsel *evsel,
1178                                      const char *fmt, const char *filter)
1179 {
1180         char *new_filter;
1181
1182         if (evsel->filter == NULL)
1183                 return perf_evsel__set_filter(evsel, filter);
1184
1185         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1186                 free(evsel->filter);
1187                 evsel->filter = new_filter;
1188                 return 0;
1189         }
1190
1191         return -1;
1192 }
1193
1194 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1195 {
1196         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1197 }
1198
1199 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1200 {
1201         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1202 }
1203
1204 int evsel__enable(struct evsel *evsel)
1205 {
1206         int err = perf_evsel__enable(&evsel->core);
1207
1208         if (!err)
1209                 evsel->disabled = false;
1210
1211         return err;
1212 }
1213
1214 int evsel__disable(struct evsel *evsel)
1215 {
1216         int err = perf_evsel__disable(&evsel->core);
1217         /*
1218          * We mark it disabled here so that tools that disable a event can
1219          * ignore events after they disable it. I.e. the ring buffer may have
1220          * already a few more events queued up before the kernel got the stop
1221          * request.
1222          */
1223         if (!err)
1224                 evsel->disabled = true;
1225
1226         return err;
1227 }
1228
1229 int perf_evsel__alloc_id(struct evsel *evsel, int ncpus, int nthreads)
1230 {
1231         if (ncpus == 0 || nthreads == 0)
1232                 return 0;
1233
1234         if (evsel->system_wide)
1235                 nthreads = 1;
1236
1237         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1238         if (evsel->sample_id == NULL)
1239                 return -ENOMEM;
1240
1241         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1242         if (evsel->id == NULL) {
1243                 xyarray__delete(evsel->sample_id);
1244                 evsel->sample_id = NULL;
1245                 return -ENOMEM;
1246         }
1247
1248         return 0;
1249 }
1250
1251 static void perf_evsel__free_id(struct evsel *evsel)
1252 {
1253         xyarray__delete(evsel->sample_id);
1254         evsel->sample_id = NULL;
1255         zfree(&evsel->id);
1256         evsel->ids = 0;
1257 }
1258
1259 static void perf_evsel__free_config_terms(struct evsel *evsel)
1260 {
1261         struct perf_evsel_config_term *term, *h;
1262
1263         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1264                 list_del_init(&term->list);
1265                 free(term);
1266         }
1267 }
1268
1269 void perf_evsel__exit(struct evsel *evsel)
1270 {
1271         assert(list_empty(&evsel->core.node));
1272         assert(evsel->evlist == NULL);
1273         perf_evsel__free_counts(evsel);
1274         perf_evsel__free_fd(&evsel->core);
1275         perf_evsel__free_id(evsel);
1276         perf_evsel__free_config_terms(evsel);
1277         cgroup__put(evsel->cgrp);
1278         perf_cpu_map__put(evsel->core.cpus);
1279         perf_cpu_map__put(evsel->core.own_cpus);
1280         perf_thread_map__put(evsel->core.threads);
1281         zfree(&evsel->group_name);
1282         zfree(&evsel->name);
1283         perf_evsel__object.fini(evsel);
1284 }
1285
1286 void evsel__delete(struct evsel *evsel)
1287 {
1288         perf_evsel__exit(evsel);
1289         free(evsel);
1290 }
1291
1292 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1293                                 struct perf_counts_values *count)
1294 {
1295         struct perf_counts_values tmp;
1296
1297         if (!evsel->prev_raw_counts)
1298                 return;
1299
1300         if (cpu == -1) {
1301                 tmp = evsel->prev_raw_counts->aggr;
1302                 evsel->prev_raw_counts->aggr = *count;
1303         } else {
1304                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1305                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1306         }
1307
1308         count->val = count->val - tmp.val;
1309         count->ena = count->ena - tmp.ena;
1310         count->run = count->run - tmp.run;
1311 }
1312
1313 void perf_counts_values__scale(struct perf_counts_values *count,
1314                                bool scale, s8 *pscaled)
1315 {
1316         s8 scaled = 0;
1317
1318         if (scale) {
1319                 if (count->run == 0) {
1320                         scaled = -1;
1321                         count->val = 0;
1322                 } else if (count->run < count->ena) {
1323                         scaled = 1;
1324                         count->val = (u64)((double) count->val * count->ena / count->run);
1325                 }
1326         }
1327
1328         if (pscaled)
1329                 *pscaled = scaled;
1330 }
1331
1332 static int
1333 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread)
1334 {
1335         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1336
1337         return perf_evsel__read(&evsel->core, cpu, thread, count);
1338 }
1339
1340 static void
1341 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1342                       u64 val, u64 ena, u64 run)
1343 {
1344         struct perf_counts_values *count;
1345
1346         count = perf_counts(counter->counts, cpu, thread);
1347
1348         count->val    = val;
1349         count->ena    = ena;
1350         count->run    = run;
1351
1352         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1353 }
1354
1355 static int
1356 perf_evsel__process_group_data(struct evsel *leader,
1357                                int cpu, int thread, u64 *data)
1358 {
1359         u64 read_format = leader->core.attr.read_format;
1360         struct sample_read_value *v;
1361         u64 nr, ena = 0, run = 0, i;
1362
1363         nr = *data++;
1364
1365         if (nr != (u64) leader->core.nr_members)
1366                 return -EINVAL;
1367
1368         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1369                 ena = *data++;
1370
1371         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1372                 run = *data++;
1373
1374         v = (struct sample_read_value *) data;
1375
1376         perf_evsel__set_count(leader, cpu, thread,
1377                               v[0].value, ena, run);
1378
1379         for (i = 1; i < nr; i++) {
1380                 struct evsel *counter;
1381
1382                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1383                 if (!counter)
1384                         return -EINVAL;
1385
1386                 perf_evsel__set_count(counter, cpu, thread,
1387                                       v[i].value, ena, run);
1388         }
1389
1390         return 0;
1391 }
1392
1393 static int
1394 perf_evsel__read_group(struct evsel *leader, int cpu, int thread)
1395 {
1396         struct perf_stat_evsel *ps = leader->stats;
1397         u64 read_format = leader->core.attr.read_format;
1398         int size = perf_evsel__read_size(&leader->core);
1399         u64 *data = ps->group_data;
1400
1401         if (!(read_format & PERF_FORMAT_ID))
1402                 return -EINVAL;
1403
1404         if (!perf_evsel__is_group_leader(leader))
1405                 return -EINVAL;
1406
1407         if (!data) {
1408                 data = zalloc(size);
1409                 if (!data)
1410                         return -ENOMEM;
1411
1412                 ps->group_data = data;
1413         }
1414
1415         if (FD(leader, cpu, thread) < 0)
1416                 return -EINVAL;
1417
1418         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1419                 return -errno;
1420
1421         return perf_evsel__process_group_data(leader, cpu, thread, data);
1422 }
1423
1424 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1425 {
1426         u64 read_format = evsel->core.attr.read_format;
1427
1428         if (read_format & PERF_FORMAT_GROUP)
1429                 return perf_evsel__read_group(evsel, cpu, thread);
1430         else
1431                 return perf_evsel__read_one(evsel, cpu, thread);
1432 }
1433
1434 int __perf_evsel__read_on_cpu(struct evsel *evsel,
1435                               int cpu, int thread, bool scale)
1436 {
1437         struct perf_counts_values count;
1438         size_t nv = scale ? 3 : 1;
1439
1440         if (FD(evsel, cpu, thread) < 0)
1441                 return -EINVAL;
1442
1443         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1444                 return -ENOMEM;
1445
1446         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1447                 return -errno;
1448
1449         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1450         perf_counts_values__scale(&count, scale, NULL);
1451         *perf_counts(evsel->counts, cpu, thread) = count;
1452         return 0;
1453 }
1454
1455 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1456 {
1457         struct evsel *leader = evsel->leader;
1458         int fd;
1459
1460         if (perf_evsel__is_group_leader(evsel))
1461                 return -1;
1462
1463         /*
1464          * Leader must be already processed/open,
1465          * if not it's a bug.
1466          */
1467         BUG_ON(!leader->core.fd);
1468
1469         fd = FD(leader, cpu, thread);
1470         BUG_ON(fd == -1);
1471
1472         return fd;
1473 }
1474
1475 struct bit_names {
1476         int bit;
1477         const char *name;
1478 };
1479
1480 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1481 {
1482         bool first_bit = true;
1483         int i = 0;
1484
1485         do {
1486                 if (value & bits[i].bit) {
1487                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1488                         first_bit = false;
1489                 }
1490         } while (bits[++i].name != NULL);
1491 }
1492
1493 static void __p_sample_type(char *buf, size_t size, u64 value)
1494 {
1495 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1496         struct bit_names bits[] = {
1497                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1498                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1499                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1500                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1501                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1502                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1503                 { .name = NULL, }
1504         };
1505 #undef bit_name
1506         __p_bits(buf, size, value, bits);
1507 }
1508
1509 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1510 {
1511 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1512         struct bit_names bits[] = {
1513                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1514                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1515                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1516                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1517                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1518                 { .name = NULL, }
1519         };
1520 #undef bit_name
1521         __p_bits(buf, size, value, bits);
1522 }
1523
1524 static void __p_read_format(char *buf, size_t size, u64 value)
1525 {
1526 #define bit_name(n) { PERF_FORMAT_##n, #n }
1527         struct bit_names bits[] = {
1528                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1529                 bit_name(ID), bit_name(GROUP),
1530                 { .name = NULL, }
1531         };
1532 #undef bit_name
1533         __p_bits(buf, size, value, bits);
1534 }
1535
1536 #define BUF_SIZE                1024
1537
1538 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1539 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1540 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1541 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1542 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1543 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1544
1545 #define PRINT_ATTRn(_n, _f, _p)                         \
1546 do {                                                    \
1547         if (attr->_f) {                                 \
1548                 _p(attr->_f);                           \
1549                 ret += attr__fprintf(fp, _n, buf, priv);\
1550         }                                               \
1551 } while (0)
1552
1553 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1554
1555 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1556                              attr__fprintf_f attr__fprintf, void *priv)
1557 {
1558         char buf[BUF_SIZE];
1559         int ret = 0;
1560
1561         PRINT_ATTRf(type, p_unsigned);
1562         PRINT_ATTRf(size, p_unsigned);
1563         PRINT_ATTRf(config, p_hex);
1564         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1565         PRINT_ATTRf(sample_type, p_sample_type);
1566         PRINT_ATTRf(read_format, p_read_format);
1567
1568         PRINT_ATTRf(disabled, p_unsigned);
1569         PRINT_ATTRf(inherit, p_unsigned);
1570         PRINT_ATTRf(pinned, p_unsigned);
1571         PRINT_ATTRf(exclusive, p_unsigned);
1572         PRINT_ATTRf(exclude_user, p_unsigned);
1573         PRINT_ATTRf(exclude_kernel, p_unsigned);
1574         PRINT_ATTRf(exclude_hv, p_unsigned);
1575         PRINT_ATTRf(exclude_idle, p_unsigned);
1576         PRINT_ATTRf(mmap, p_unsigned);
1577         PRINT_ATTRf(comm, p_unsigned);
1578         PRINT_ATTRf(freq, p_unsigned);
1579         PRINT_ATTRf(inherit_stat, p_unsigned);
1580         PRINT_ATTRf(enable_on_exec, p_unsigned);
1581         PRINT_ATTRf(task, p_unsigned);
1582         PRINT_ATTRf(watermark, p_unsigned);
1583         PRINT_ATTRf(precise_ip, p_unsigned);
1584         PRINT_ATTRf(mmap_data, p_unsigned);
1585         PRINT_ATTRf(sample_id_all, p_unsigned);
1586         PRINT_ATTRf(exclude_host, p_unsigned);
1587         PRINT_ATTRf(exclude_guest, p_unsigned);
1588         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1589         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1590         PRINT_ATTRf(mmap2, p_unsigned);
1591         PRINT_ATTRf(comm_exec, p_unsigned);
1592         PRINT_ATTRf(use_clockid, p_unsigned);
1593         PRINT_ATTRf(context_switch, p_unsigned);
1594         PRINT_ATTRf(write_backward, p_unsigned);
1595         PRINT_ATTRf(namespaces, p_unsigned);
1596         PRINT_ATTRf(ksymbol, p_unsigned);
1597         PRINT_ATTRf(bpf_event, p_unsigned);
1598         PRINT_ATTRf(aux_output, p_unsigned);
1599
1600         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1601         PRINT_ATTRf(bp_type, p_unsigned);
1602         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1603         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1604         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1605         PRINT_ATTRf(sample_regs_user, p_hex);
1606         PRINT_ATTRf(sample_stack_user, p_unsigned);
1607         PRINT_ATTRf(clockid, p_signed);
1608         PRINT_ATTRf(sample_regs_intr, p_hex);
1609         PRINT_ATTRf(aux_watermark, p_unsigned);
1610         PRINT_ATTRf(sample_max_stack, p_unsigned);
1611
1612         return ret;
1613 }
1614
1615 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1616                                 void *priv __maybe_unused)
1617 {
1618         return fprintf(fp, "  %-32s %s\n", name, val);
1619 }
1620
1621 static void perf_evsel__remove_fd(struct evsel *pos,
1622                                   int nr_cpus, int nr_threads,
1623                                   int thread_idx)
1624 {
1625         for (int cpu = 0; cpu < nr_cpus; cpu++)
1626                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1627                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1628 }
1629
1630 static int update_fds(struct evsel *evsel,
1631                       int nr_cpus, int cpu_idx,
1632                       int nr_threads, int thread_idx)
1633 {
1634         struct evsel *pos;
1635
1636         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1637                 return -EINVAL;
1638
1639         evlist__for_each_entry(evsel->evlist, pos) {
1640                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1641
1642                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1643
1644                 /*
1645                  * Since fds for next evsel has not been created,
1646                  * there is no need to iterate whole event list.
1647                  */
1648                 if (pos == evsel)
1649                         break;
1650         }
1651         return 0;
1652 }
1653
1654 static bool ignore_missing_thread(struct evsel *evsel,
1655                                   int nr_cpus, int cpu,
1656                                   struct perf_thread_map *threads,
1657                                   int thread, int err)
1658 {
1659         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1660
1661         if (!evsel->ignore_missing_thread)
1662                 return false;
1663
1664         /* The system wide setup does not work with threads. */
1665         if (evsel->system_wide)
1666                 return false;
1667
1668         /* The -ESRCH is perf event syscall errno for pid's not found. */
1669         if (err != -ESRCH)
1670                 return false;
1671
1672         /* If there's only one thread, let it fail. */
1673         if (threads->nr == 1)
1674                 return false;
1675
1676         /*
1677          * We should remove fd for missing_thread first
1678          * because thread_map__remove() will decrease threads->nr.
1679          */
1680         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1681                 return false;
1682
1683         if (thread_map__remove(threads, thread))
1684                 return false;
1685
1686         pr_warning("WARNING: Ignored open failure for pid %d\n",
1687                    ignore_pid);
1688         return true;
1689 }
1690
1691 static void display_attr(struct perf_event_attr *attr)
1692 {
1693         if (verbose >= 2) {
1694                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1695                 fprintf(stderr, "perf_event_attr:\n");
1696                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1697                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1698         }
1699 }
1700
1701 static int perf_event_open(struct evsel *evsel,
1702                            pid_t pid, int cpu, int group_fd,
1703                            unsigned long flags)
1704 {
1705         int precise_ip = evsel->core.attr.precise_ip;
1706         int fd;
1707
1708         while (1) {
1709                 pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1710                           pid, cpu, group_fd, flags);
1711
1712                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1713                 if (fd >= 0)
1714                         break;
1715
1716                 /* Do not try less precise if not requested. */
1717                 if (!evsel->precise_max)
1718                         break;
1719
1720                 /*
1721                  * We tried all the precise_ip values, and it's
1722                  * still failing, so leave it to standard fallback.
1723                  */
1724                 if (!evsel->core.attr.precise_ip) {
1725                         evsel->core.attr.precise_ip = precise_ip;
1726                         break;
1727                 }
1728
1729                 pr_debug2("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1730                 evsel->core.attr.precise_ip--;
1731                 pr_debug2("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1732                 display_attr(&evsel->core.attr);
1733         }
1734
1735         return fd;
1736 }
1737
1738 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1739                 struct perf_thread_map *threads)
1740 {
1741         int cpu, thread, nthreads;
1742         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1743         int pid = -1, err;
1744         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1745
1746         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1747             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1748                 return -EINVAL;
1749
1750         if (cpus == NULL) {
1751                 static struct perf_cpu_map *empty_cpu_map;
1752
1753                 if (empty_cpu_map == NULL) {
1754                         empty_cpu_map = perf_cpu_map__dummy_new();
1755                         if (empty_cpu_map == NULL)
1756                                 return -ENOMEM;
1757                 }
1758
1759                 cpus = empty_cpu_map;
1760         }
1761
1762         if (threads == NULL) {
1763                 static struct perf_thread_map *empty_thread_map;
1764
1765                 if (empty_thread_map == NULL) {
1766                         empty_thread_map = thread_map__new_by_tid(-1);
1767                         if (empty_thread_map == NULL)
1768                                 return -ENOMEM;
1769                 }
1770
1771                 threads = empty_thread_map;
1772         }
1773
1774         if (evsel->system_wide)
1775                 nthreads = 1;
1776         else
1777                 nthreads = threads->nr;
1778
1779         if (evsel->core.fd == NULL &&
1780             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1781                 return -ENOMEM;
1782
1783         if (evsel->cgrp) {
1784                 flags |= PERF_FLAG_PID_CGROUP;
1785                 pid = evsel->cgrp->fd;
1786         }
1787
1788 fallback_missing_features:
1789         if (perf_missing_features.clockid_wrong)
1790                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1791         if (perf_missing_features.clockid) {
1792                 evsel->core.attr.use_clockid = 0;
1793                 evsel->core.attr.clockid = 0;
1794         }
1795         if (perf_missing_features.cloexec)
1796                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1797         if (perf_missing_features.mmap2)
1798                 evsel->core.attr.mmap2 = 0;
1799         if (perf_missing_features.exclude_guest)
1800                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1801         if (perf_missing_features.lbr_flags)
1802                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1803                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1804         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1805                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1806         if (perf_missing_features.ksymbol)
1807                 evsel->core.attr.ksymbol = 0;
1808         if (perf_missing_features.bpf)
1809                 evsel->core.attr.bpf_event = 0;
1810 retry_sample_id:
1811         if (perf_missing_features.sample_id_all)
1812                 evsel->core.attr.sample_id_all = 0;
1813
1814         display_attr(&evsel->core.attr);
1815
1816         for (cpu = 0; cpu < cpus->nr; cpu++) {
1817
1818                 for (thread = 0; thread < nthreads; thread++) {
1819                         int fd, group_fd;
1820
1821                         if (!evsel->cgrp && !evsel->system_wide)
1822                                 pid = perf_thread_map__pid(threads, thread);
1823
1824                         group_fd = get_group_fd(evsel, cpu, thread);
1825 retry_open:
1826                         test_attr__ready();
1827
1828                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1829                                              group_fd, flags);
1830
1831                         FD(evsel, cpu, thread) = fd;
1832
1833                         if (fd < 0) {
1834                                 err = -errno;
1835
1836                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1837                                         /*
1838                                          * We just removed 1 thread, so take a step
1839                                          * back on thread index and lower the upper
1840                                          * nthreads limit.
1841                                          */
1842                                         nthreads--;
1843                                         thread--;
1844
1845                                         /* ... and pretend like nothing have happened. */
1846                                         err = 0;
1847                                         continue;
1848                                 }
1849
1850                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1851                                           err);
1852                                 goto try_fallback;
1853                         }
1854
1855                         pr_debug2(" = %d\n", fd);
1856
1857                         if (evsel->bpf_fd >= 0) {
1858                                 int evt_fd = fd;
1859                                 int bpf_fd = evsel->bpf_fd;
1860
1861                                 err = ioctl(evt_fd,
1862                                             PERF_EVENT_IOC_SET_BPF,
1863                                             bpf_fd);
1864                                 if (err && errno != EEXIST) {
1865                                         pr_err("failed to attach bpf fd %d: %s\n",
1866                                                bpf_fd, strerror(errno));
1867                                         err = -EINVAL;
1868                                         goto out_close;
1869                                 }
1870                         }
1871
1872                         set_rlimit = NO_CHANGE;
1873
1874                         /*
1875                          * If we succeeded but had to kill clockid, fail and
1876                          * have perf_evsel__open_strerror() print us a nice
1877                          * error.
1878                          */
1879                         if (perf_missing_features.clockid ||
1880                             perf_missing_features.clockid_wrong) {
1881                                 err = -EINVAL;
1882                                 goto out_close;
1883                         }
1884                 }
1885         }
1886
1887         return 0;
1888
1889 try_fallback:
1890         /*
1891          * perf stat needs between 5 and 22 fds per CPU. When we run out
1892          * of them try to increase the limits.
1893          */
1894         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1895                 struct rlimit l;
1896                 int old_errno = errno;
1897
1898                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1899                         if (set_rlimit == NO_CHANGE)
1900                                 l.rlim_cur = l.rlim_max;
1901                         else {
1902                                 l.rlim_cur = l.rlim_max + 1000;
1903                                 l.rlim_max = l.rlim_cur;
1904                         }
1905                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1906                                 set_rlimit++;
1907                                 errno = old_errno;
1908                                 goto retry_open;
1909                         }
1910                 }
1911                 errno = old_errno;
1912         }
1913
1914         if (err != -EINVAL || cpu > 0 || thread > 0)
1915                 goto out_close;
1916
1917         /*
1918          * Must probe features in the order they were added to the
1919          * perf_event_attr interface.
1920          */
1921         if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1922                 perf_missing_features.aux_output = true;
1923                 pr_debug2("Kernel has no attr.aux_output support, bailing out\n");
1924                 goto out_close;
1925         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1926                 perf_missing_features.bpf = true;
1927                 pr_debug2("switching off bpf_event\n");
1928                 goto fallback_missing_features;
1929         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1930                 perf_missing_features.ksymbol = true;
1931                 pr_debug2("switching off ksymbol\n");
1932                 goto fallback_missing_features;
1933         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1934                 perf_missing_features.write_backward = true;
1935                 pr_debug2("switching off write_backward\n");
1936                 goto out_close;
1937         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1938                 perf_missing_features.clockid_wrong = true;
1939                 pr_debug2("switching off clockid\n");
1940                 goto fallback_missing_features;
1941         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1942                 perf_missing_features.clockid = true;
1943                 pr_debug2("switching off use_clockid\n");
1944                 goto fallback_missing_features;
1945         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1946                 perf_missing_features.cloexec = true;
1947                 pr_debug2("switching off cloexec flag\n");
1948                 goto fallback_missing_features;
1949         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1950                 perf_missing_features.mmap2 = true;
1951                 pr_debug2("switching off mmap2\n");
1952                 goto fallback_missing_features;
1953         } else if (!perf_missing_features.exclude_guest &&
1954                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1955                 perf_missing_features.exclude_guest = true;
1956                 pr_debug2("switching off exclude_guest, exclude_host\n");
1957                 goto fallback_missing_features;
1958         } else if (!perf_missing_features.sample_id_all) {
1959                 perf_missing_features.sample_id_all = true;
1960                 pr_debug2("switching off sample_id_all\n");
1961                 goto retry_sample_id;
1962         } else if (!perf_missing_features.lbr_flags &&
1963                         (evsel->core.attr.branch_sample_type &
1964                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1965                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1966                 perf_missing_features.lbr_flags = true;
1967                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1968                 goto fallback_missing_features;
1969         } else if (!perf_missing_features.group_read &&
1970                     evsel->core.attr.inherit &&
1971                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1972                    perf_evsel__is_group_leader(evsel)) {
1973                 perf_missing_features.group_read = true;
1974                 pr_debug2("switching off group read\n");
1975                 goto fallback_missing_features;
1976         }
1977 out_close:
1978         if (err)
1979                 threads->err_thread = thread;
1980
1981         do {
1982                 while (--thread >= 0) {
1983                         close(FD(evsel, cpu, thread));
1984                         FD(evsel, cpu, thread) = -1;
1985                 }
1986                 thread = nthreads;
1987         } while (--cpu >= 0);
1988         return err;
1989 }
1990
1991 void evsel__close(struct evsel *evsel)
1992 {
1993         perf_evsel__close(&evsel->core);
1994         perf_evsel__free_id(evsel);
1995 }
1996
1997 int perf_evsel__open_per_cpu(struct evsel *evsel,
1998                              struct perf_cpu_map *cpus)
1999 {
2000         return evsel__open(evsel, cpus, NULL);
2001 }
2002
2003 int perf_evsel__open_per_thread(struct evsel *evsel,
2004                                 struct perf_thread_map *threads)
2005 {
2006         return evsel__open(evsel, NULL, threads);
2007 }
2008
2009 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2010                                        const union perf_event *event,
2011                                        struct perf_sample *sample)
2012 {
2013         u64 type = evsel->core.attr.sample_type;
2014         const __u64 *array = event->sample.array;
2015         bool swapped = evsel->needs_swap;
2016         union u64_swap u;
2017
2018         array += ((event->header.size -
2019                    sizeof(event->header)) / sizeof(u64)) - 1;
2020
2021         if (type & PERF_SAMPLE_IDENTIFIER) {
2022                 sample->id = *array;
2023                 array--;
2024         }
2025
2026         if (type & PERF_SAMPLE_CPU) {
2027                 u.val64 = *array;
2028                 if (swapped) {
2029                         /* undo swap of u64, then swap on individual u32s */
2030                         u.val64 = bswap_64(u.val64);
2031                         u.val32[0] = bswap_32(u.val32[0]);
2032                 }
2033
2034                 sample->cpu = u.val32[0];
2035                 array--;
2036         }
2037
2038         if (type & PERF_SAMPLE_STREAM_ID) {
2039                 sample->stream_id = *array;
2040                 array--;
2041         }
2042
2043         if (type & PERF_SAMPLE_ID) {
2044                 sample->id = *array;
2045                 array--;
2046         }
2047
2048         if (type & PERF_SAMPLE_TIME) {
2049                 sample->time = *array;
2050                 array--;
2051         }
2052
2053         if (type & PERF_SAMPLE_TID) {
2054                 u.val64 = *array;
2055                 if (swapped) {
2056                         /* undo swap of u64, then swap on individual u32s */
2057                         u.val64 = bswap_64(u.val64);
2058                         u.val32[0] = bswap_32(u.val32[0]);
2059                         u.val32[1] = bswap_32(u.val32[1]);
2060                 }
2061
2062                 sample->pid = u.val32[0];
2063                 sample->tid = u.val32[1];
2064                 array--;
2065         }
2066
2067         return 0;
2068 }
2069
2070 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2071                             u64 size)
2072 {
2073         return size > max_size || offset + size > endp;
2074 }
2075
2076 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2077         do {                                                            \
2078                 if (overflow(endp, (max_size), (offset), (size)))       \
2079                         return -EFAULT;                                 \
2080         } while (0)
2081
2082 #define OVERFLOW_CHECK_u64(offset) \
2083         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2084
2085 static int
2086 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2087 {
2088         /*
2089          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2090          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2091          * check the format does not go past the end of the event.
2092          */
2093         if (sample_size + sizeof(event->header) > event->header.size)
2094                 return -EFAULT;
2095
2096         return 0;
2097 }
2098
2099 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2100                              struct perf_sample *data)
2101 {
2102         u64 type = evsel->core.attr.sample_type;
2103         bool swapped = evsel->needs_swap;
2104         const __u64 *array;
2105         u16 max_size = event->header.size;
2106         const void *endp = (void *)event + max_size;
2107         u64 sz;
2108
2109         /*
2110          * used for cross-endian analysis. See git commit 65014ab3
2111          * for why this goofiness is needed.
2112          */
2113         union u64_swap u;
2114
2115         memset(data, 0, sizeof(*data));
2116         data->cpu = data->pid = data->tid = -1;
2117         data->stream_id = data->id = data->time = -1ULL;
2118         data->period = evsel->core.attr.sample_period;
2119         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2120         data->misc    = event->header.misc;
2121         data->id = -1ULL;
2122         data->data_src = PERF_MEM_DATA_SRC_NONE;
2123
2124         if (event->header.type != PERF_RECORD_SAMPLE) {
2125                 if (!evsel->core.attr.sample_id_all)
2126                         return 0;
2127                 return perf_evsel__parse_id_sample(evsel, event, data);
2128         }
2129
2130         array = event->sample.array;
2131
2132         if (perf_event__check_size(event, evsel->sample_size))
2133                 return -EFAULT;
2134
2135         if (type & PERF_SAMPLE_IDENTIFIER) {
2136                 data->id = *array;
2137                 array++;
2138         }
2139
2140         if (type & PERF_SAMPLE_IP) {
2141                 data->ip = *array;
2142                 array++;
2143         }
2144
2145         if (type & PERF_SAMPLE_TID) {
2146                 u.val64 = *array;
2147                 if (swapped) {
2148                         /* undo swap of u64, then swap on individual u32s */
2149                         u.val64 = bswap_64(u.val64);
2150                         u.val32[0] = bswap_32(u.val32[0]);
2151                         u.val32[1] = bswap_32(u.val32[1]);
2152                 }
2153
2154                 data->pid = u.val32[0];
2155                 data->tid = u.val32[1];
2156                 array++;
2157         }
2158
2159         if (type & PERF_SAMPLE_TIME) {
2160                 data->time = *array;
2161                 array++;
2162         }
2163
2164         if (type & PERF_SAMPLE_ADDR) {
2165                 data->addr = *array;
2166                 array++;
2167         }
2168
2169         if (type & PERF_SAMPLE_ID) {
2170                 data->id = *array;
2171                 array++;
2172         }
2173
2174         if (type & PERF_SAMPLE_STREAM_ID) {
2175                 data->stream_id = *array;
2176                 array++;
2177         }
2178
2179         if (type & PERF_SAMPLE_CPU) {
2180
2181                 u.val64 = *array;
2182                 if (swapped) {
2183                         /* undo swap of u64, then swap on individual u32s */
2184                         u.val64 = bswap_64(u.val64);
2185                         u.val32[0] = bswap_32(u.val32[0]);
2186                 }
2187
2188                 data->cpu = u.val32[0];
2189                 array++;
2190         }
2191
2192         if (type & PERF_SAMPLE_PERIOD) {
2193                 data->period = *array;
2194                 array++;
2195         }
2196
2197         if (type & PERF_SAMPLE_READ) {
2198                 u64 read_format = evsel->core.attr.read_format;
2199
2200                 OVERFLOW_CHECK_u64(array);
2201                 if (read_format & PERF_FORMAT_GROUP)
2202                         data->read.group.nr = *array;
2203                 else
2204                         data->read.one.value = *array;
2205
2206                 array++;
2207
2208                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2209                         OVERFLOW_CHECK_u64(array);
2210                         data->read.time_enabled = *array;
2211                         array++;
2212                 }
2213
2214                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2215                         OVERFLOW_CHECK_u64(array);
2216                         data->read.time_running = *array;
2217                         array++;
2218                 }
2219
2220                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2221                 if (read_format & PERF_FORMAT_GROUP) {
2222                         const u64 max_group_nr = UINT64_MAX /
2223                                         sizeof(struct sample_read_value);
2224
2225                         if (data->read.group.nr > max_group_nr)
2226                                 return -EFAULT;
2227                         sz = data->read.group.nr *
2228                              sizeof(struct sample_read_value);
2229                         OVERFLOW_CHECK(array, sz, max_size);
2230                         data->read.group.values =
2231                                         (struct sample_read_value *)array;
2232                         array = (void *)array + sz;
2233                 } else {
2234                         OVERFLOW_CHECK_u64(array);
2235                         data->read.one.id = *array;
2236                         array++;
2237                 }
2238         }
2239
2240         if (evsel__has_callchain(evsel)) {
2241                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2242
2243                 OVERFLOW_CHECK_u64(array);
2244                 data->callchain = (struct ip_callchain *)array++;
2245                 if (data->callchain->nr > max_callchain_nr)
2246                         return -EFAULT;
2247                 sz = data->callchain->nr * sizeof(u64);
2248                 OVERFLOW_CHECK(array, sz, max_size);
2249                 array = (void *)array + sz;
2250         }
2251
2252         if (type & PERF_SAMPLE_RAW) {
2253                 OVERFLOW_CHECK_u64(array);
2254                 u.val64 = *array;
2255
2256                 /*
2257                  * Undo swap of u64, then swap on individual u32s,
2258                  * get the size of the raw area and undo all of the
2259                  * swap. The pevent interface handles endianity by
2260                  * itself.
2261                  */
2262                 if (swapped) {
2263                         u.val64 = bswap_64(u.val64);
2264                         u.val32[0] = bswap_32(u.val32[0]);
2265                         u.val32[1] = bswap_32(u.val32[1]);
2266                 }
2267                 data->raw_size = u.val32[0];
2268
2269                 /*
2270                  * The raw data is aligned on 64bits including the
2271                  * u32 size, so it's safe to use mem_bswap_64.
2272                  */
2273                 if (swapped)
2274                         mem_bswap_64((void *) array, data->raw_size);
2275
2276                 array = (void *)array + sizeof(u32);
2277
2278                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2279                 data->raw_data = (void *)array;
2280                 array = (void *)array + data->raw_size;
2281         }
2282
2283         if (type & PERF_SAMPLE_BRANCH_STACK) {
2284                 const u64 max_branch_nr = UINT64_MAX /
2285                                           sizeof(struct branch_entry);
2286
2287                 OVERFLOW_CHECK_u64(array);
2288                 data->branch_stack = (struct branch_stack *)array++;
2289
2290                 if (data->branch_stack->nr > max_branch_nr)
2291                         return -EFAULT;
2292                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2293                 OVERFLOW_CHECK(array, sz, max_size);
2294                 array = (void *)array + sz;
2295         }
2296
2297         if (type & PERF_SAMPLE_REGS_USER) {
2298                 OVERFLOW_CHECK_u64(array);
2299                 data->user_regs.abi = *array;
2300                 array++;
2301
2302                 if (data->user_regs.abi) {
2303                         u64 mask = evsel->core.attr.sample_regs_user;
2304
2305                         sz = hweight64(mask) * sizeof(u64);
2306                         OVERFLOW_CHECK(array, sz, max_size);
2307                         data->user_regs.mask = mask;
2308                         data->user_regs.regs = (u64 *)array;
2309                         array = (void *)array + sz;
2310                 }
2311         }
2312
2313         if (type & PERF_SAMPLE_STACK_USER) {
2314                 OVERFLOW_CHECK_u64(array);
2315                 sz = *array++;
2316
2317                 data->user_stack.offset = ((char *)(array - 1)
2318                                           - (char *) event);
2319
2320                 if (!sz) {
2321                         data->user_stack.size = 0;
2322                 } else {
2323                         OVERFLOW_CHECK(array, sz, max_size);
2324                         data->user_stack.data = (char *)array;
2325                         array = (void *)array + sz;
2326                         OVERFLOW_CHECK_u64(array);
2327                         data->user_stack.size = *array++;
2328                         if (WARN_ONCE(data->user_stack.size > sz,
2329                                       "user stack dump failure\n"))
2330                                 return -EFAULT;
2331                 }
2332         }
2333
2334         if (type & PERF_SAMPLE_WEIGHT) {
2335                 OVERFLOW_CHECK_u64(array);
2336                 data->weight = *array;
2337                 array++;
2338         }
2339
2340         if (type & PERF_SAMPLE_DATA_SRC) {
2341                 OVERFLOW_CHECK_u64(array);
2342                 data->data_src = *array;
2343                 array++;
2344         }
2345
2346         if (type & PERF_SAMPLE_TRANSACTION) {
2347                 OVERFLOW_CHECK_u64(array);
2348                 data->transaction = *array;
2349                 array++;
2350         }
2351
2352         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2353         if (type & PERF_SAMPLE_REGS_INTR) {
2354                 OVERFLOW_CHECK_u64(array);
2355                 data->intr_regs.abi = *array;
2356                 array++;
2357
2358                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2359                         u64 mask = evsel->core.attr.sample_regs_intr;
2360
2361                         sz = hweight64(mask) * sizeof(u64);
2362                         OVERFLOW_CHECK(array, sz, max_size);
2363                         data->intr_regs.mask = mask;
2364                         data->intr_regs.regs = (u64 *)array;
2365                         array = (void *)array + sz;
2366                 }
2367         }
2368
2369         data->phys_addr = 0;
2370         if (type & PERF_SAMPLE_PHYS_ADDR) {
2371                 data->phys_addr = *array;
2372                 array++;
2373         }
2374
2375         return 0;
2376 }
2377
2378 int perf_evsel__parse_sample_timestamp(struct evsel *evsel,
2379                                        union perf_event *event,
2380                                        u64 *timestamp)
2381 {
2382         u64 type = evsel->core.attr.sample_type;
2383         const __u64 *array;
2384
2385         if (!(type & PERF_SAMPLE_TIME))
2386                 return -1;
2387
2388         if (event->header.type != PERF_RECORD_SAMPLE) {
2389                 struct perf_sample data = {
2390                         .time = -1ULL,
2391                 };
2392
2393                 if (!evsel->core.attr.sample_id_all)
2394                         return -1;
2395                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2396                         return -1;
2397
2398                 *timestamp = data.time;
2399                 return 0;
2400         }
2401
2402         array = event->sample.array;
2403
2404         if (perf_event__check_size(event, evsel->sample_size))
2405                 return -EFAULT;
2406
2407         if (type & PERF_SAMPLE_IDENTIFIER)
2408                 array++;
2409
2410         if (type & PERF_SAMPLE_IP)
2411                 array++;
2412
2413         if (type & PERF_SAMPLE_TID)
2414                 array++;
2415
2416         if (type & PERF_SAMPLE_TIME)
2417                 *timestamp = *array;
2418
2419         return 0;
2420 }
2421
2422 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2423                                      u64 read_format)
2424 {
2425         size_t sz, result = sizeof(struct perf_record_sample);
2426
2427         if (type & PERF_SAMPLE_IDENTIFIER)
2428                 result += sizeof(u64);
2429
2430         if (type & PERF_SAMPLE_IP)
2431                 result += sizeof(u64);
2432
2433         if (type & PERF_SAMPLE_TID)
2434                 result += sizeof(u64);
2435
2436         if (type & PERF_SAMPLE_TIME)
2437                 result += sizeof(u64);
2438
2439         if (type & PERF_SAMPLE_ADDR)
2440                 result += sizeof(u64);
2441
2442         if (type & PERF_SAMPLE_ID)
2443                 result += sizeof(u64);
2444
2445         if (type & PERF_SAMPLE_STREAM_ID)
2446                 result += sizeof(u64);
2447
2448         if (type & PERF_SAMPLE_CPU)
2449                 result += sizeof(u64);
2450
2451         if (type & PERF_SAMPLE_PERIOD)
2452                 result += sizeof(u64);
2453
2454         if (type & PERF_SAMPLE_READ) {
2455                 result += sizeof(u64);
2456                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2457                         result += sizeof(u64);
2458                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2459                         result += sizeof(u64);
2460                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2461                 if (read_format & PERF_FORMAT_GROUP) {
2462                         sz = sample->read.group.nr *
2463                              sizeof(struct sample_read_value);
2464                         result += sz;
2465                 } else {
2466                         result += sizeof(u64);
2467                 }
2468         }
2469
2470         if (type & PERF_SAMPLE_CALLCHAIN) {
2471                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2472                 result += sz;
2473         }
2474
2475         if (type & PERF_SAMPLE_RAW) {
2476                 result += sizeof(u32);
2477                 result += sample->raw_size;
2478         }
2479
2480         if (type & PERF_SAMPLE_BRANCH_STACK) {
2481                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2482                 sz += sizeof(u64);
2483                 result += sz;
2484         }
2485
2486         if (type & PERF_SAMPLE_REGS_USER) {
2487                 if (sample->user_regs.abi) {
2488                         result += sizeof(u64);
2489                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2490                         result += sz;
2491                 } else {
2492                         result += sizeof(u64);
2493                 }
2494         }
2495
2496         if (type & PERF_SAMPLE_STACK_USER) {
2497                 sz = sample->user_stack.size;
2498                 result += sizeof(u64);
2499                 if (sz) {
2500                         result += sz;
2501                         result += sizeof(u64);
2502                 }
2503         }
2504
2505         if (type & PERF_SAMPLE_WEIGHT)
2506                 result += sizeof(u64);
2507
2508         if (type & PERF_SAMPLE_DATA_SRC)
2509                 result += sizeof(u64);
2510
2511         if (type & PERF_SAMPLE_TRANSACTION)
2512                 result += sizeof(u64);
2513
2514         if (type & PERF_SAMPLE_REGS_INTR) {
2515                 if (sample->intr_regs.abi) {
2516                         result += sizeof(u64);
2517                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2518                         result += sz;
2519                 } else {
2520                         result += sizeof(u64);
2521                 }
2522         }
2523
2524         if (type & PERF_SAMPLE_PHYS_ADDR)
2525                 result += sizeof(u64);
2526
2527         return result;
2528 }
2529
2530 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2531                                   u64 read_format,
2532                                   const struct perf_sample *sample)
2533 {
2534         __u64 *array;
2535         size_t sz;
2536         /*
2537          * used for cross-endian analysis. See git commit 65014ab3
2538          * for why this goofiness is needed.
2539          */
2540         union u64_swap u;
2541
2542         array = event->sample.array;
2543
2544         if (type & PERF_SAMPLE_IDENTIFIER) {
2545                 *array = sample->id;
2546                 array++;
2547         }
2548
2549         if (type & PERF_SAMPLE_IP) {
2550                 *array = sample->ip;
2551                 array++;
2552         }
2553
2554         if (type & PERF_SAMPLE_TID) {
2555                 u.val32[0] = sample->pid;
2556                 u.val32[1] = sample->tid;
2557                 *array = u.val64;
2558                 array++;
2559         }
2560
2561         if (type & PERF_SAMPLE_TIME) {
2562                 *array = sample->time;
2563                 array++;
2564         }
2565
2566         if (type & PERF_SAMPLE_ADDR) {
2567                 *array = sample->addr;
2568                 array++;
2569         }
2570
2571         if (type & PERF_SAMPLE_ID) {
2572                 *array = sample->id;
2573                 array++;
2574         }
2575
2576         if (type & PERF_SAMPLE_STREAM_ID) {
2577                 *array = sample->stream_id;
2578                 array++;
2579         }
2580
2581         if (type & PERF_SAMPLE_CPU) {
2582                 u.val32[0] = sample->cpu;
2583                 u.val32[1] = 0;
2584                 *array = u.val64;
2585                 array++;
2586         }
2587
2588         if (type & PERF_SAMPLE_PERIOD) {
2589                 *array = sample->period;
2590                 array++;
2591         }
2592
2593         if (type & PERF_SAMPLE_READ) {
2594                 if (read_format & PERF_FORMAT_GROUP)
2595                         *array = sample->read.group.nr;
2596                 else
2597                         *array = sample->read.one.value;
2598                 array++;
2599
2600                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2601                         *array = sample->read.time_enabled;
2602                         array++;
2603                 }
2604
2605                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2606                         *array = sample->read.time_running;
2607                         array++;
2608                 }
2609
2610                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2611                 if (read_format & PERF_FORMAT_GROUP) {
2612                         sz = sample->read.group.nr *
2613                              sizeof(struct sample_read_value);
2614                         memcpy(array, sample->read.group.values, sz);
2615                         array = (void *)array + sz;
2616                 } else {
2617                         *array = sample->read.one.id;
2618                         array++;
2619                 }
2620         }
2621
2622         if (type & PERF_SAMPLE_CALLCHAIN) {
2623                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2624                 memcpy(array, sample->callchain, sz);
2625                 array = (void *)array + sz;
2626         }
2627
2628         if (type & PERF_SAMPLE_RAW) {
2629                 u.val32[0] = sample->raw_size;
2630                 *array = u.val64;
2631                 array = (void *)array + sizeof(u32);
2632
2633                 memcpy(array, sample->raw_data, sample->raw_size);
2634                 array = (void *)array + sample->raw_size;
2635         }
2636
2637         if (type & PERF_SAMPLE_BRANCH_STACK) {
2638                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2639                 sz += sizeof(u64);
2640                 memcpy(array, sample->branch_stack, sz);
2641                 array = (void *)array + sz;
2642         }
2643
2644         if (type & PERF_SAMPLE_REGS_USER) {
2645                 if (sample->user_regs.abi) {
2646                         *array++ = sample->user_regs.abi;
2647                         sz = hweight64(sample->user_regs.mask) * sizeof(u64);
2648                         memcpy(array, sample->user_regs.regs, sz);
2649                         array = (void *)array + sz;
2650                 } else {
2651                         *array++ = 0;
2652                 }
2653         }
2654
2655         if (type & PERF_SAMPLE_STACK_USER) {
2656                 sz = sample->user_stack.size;
2657                 *array++ = sz;
2658                 if (sz) {
2659                         memcpy(array, sample->user_stack.data, sz);
2660                         array = (void *)array + sz;
2661                         *array++ = sz;
2662                 }
2663         }
2664
2665         if (type & PERF_SAMPLE_WEIGHT) {
2666                 *array = sample->weight;
2667                 array++;
2668         }
2669
2670         if (type & PERF_SAMPLE_DATA_SRC) {
2671                 *array = sample->data_src;
2672                 array++;
2673         }
2674
2675         if (type & PERF_SAMPLE_TRANSACTION) {
2676                 *array = sample->transaction;
2677                 array++;
2678         }
2679
2680         if (type & PERF_SAMPLE_REGS_INTR) {
2681                 if (sample->intr_regs.abi) {
2682                         *array++ = sample->intr_regs.abi;
2683                         sz = hweight64(sample->intr_regs.mask) * sizeof(u64);
2684                         memcpy(array, sample->intr_regs.regs, sz);
2685                         array = (void *)array + sz;
2686                 } else {
2687                         *array++ = 0;
2688                 }
2689         }
2690
2691         if (type & PERF_SAMPLE_PHYS_ADDR) {
2692                 *array = sample->phys_addr;
2693                 array++;
2694         }
2695
2696         return 0;
2697 }
2698
2699 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name)
2700 {
2701         return tep_find_field(evsel->tp_format, name);
2702 }
2703
2704 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample,
2705                          const char *name)
2706 {
2707         struct tep_format_field *field = perf_evsel__field(evsel, name);
2708         int offset;
2709
2710         if (!field)
2711                 return NULL;
2712
2713         offset = field->offset;
2714
2715         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2716                 offset = *(int *)(sample->raw_data + field->offset);
2717                 offset &= 0xffff;
2718         }
2719
2720         return sample->raw_data + offset;
2721 }
2722
2723 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2724                          bool needs_swap)
2725 {
2726         u64 value;
2727         void *ptr = sample->raw_data + field->offset;
2728
2729         switch (field->size) {
2730         case 1:
2731                 return *(u8 *)ptr;
2732         case 2:
2733                 value = *(u16 *)ptr;
2734                 break;
2735         case 4:
2736                 value = *(u32 *)ptr;
2737                 break;
2738         case 8:
2739                 memcpy(&value, ptr, sizeof(u64));
2740                 break;
2741         default:
2742                 return 0;
2743         }
2744
2745         if (!needs_swap)
2746                 return value;
2747
2748         switch (field->size) {
2749         case 2:
2750                 return bswap_16(value);
2751         case 4:
2752                 return bswap_32(value);
2753         case 8:
2754                 return bswap_64(value);
2755         default:
2756                 return 0;
2757         }
2758
2759         return 0;
2760 }
2761
2762 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample,
2763                        const char *name)
2764 {
2765         struct tep_format_field *field = perf_evsel__field(evsel, name);
2766
2767         if (!field)
2768                 return 0;
2769
2770         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2771 }
2772
2773 bool perf_evsel__fallback(struct evsel *evsel, int err,
2774                           char *msg, size_t msgsize)
2775 {
2776         int paranoid;
2777
2778         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2779             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2780             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2781                 /*
2782                  * If it's cycles then fall back to hrtimer based
2783                  * cpu-clock-tick sw counter, which is always available even if
2784                  * no PMU support.
2785                  *
2786                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2787                  * b0a873e).
2788                  */
2789                 scnprintf(msg, msgsize, "%s",
2790 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2791
2792                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2793                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2794
2795                 zfree(&evsel->name);
2796                 return true;
2797         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2798                    (paranoid = perf_event_paranoid()) > 1) {
2799                 const char *name = perf_evsel__name(evsel);
2800                 char *new_name;
2801                 const char *sep = ":";
2802
2803                 /* Is there already the separator in the name. */
2804                 if (strchr(name, '/') ||
2805                     strchr(name, ':'))
2806                         sep = "";
2807
2808                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2809                         return false;
2810
2811                 if (evsel->name)
2812                         free(evsel->name);
2813                 evsel->name = new_name;
2814                 scnprintf(msg, msgsize,
2815 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2816                 evsel->core.attr.exclude_kernel = 1;
2817
2818                 return true;
2819         }
2820
2821         return false;
2822 }
2823
2824 static bool find_process(const char *name)
2825 {
2826         size_t len = strlen(name);
2827         DIR *dir;
2828         struct dirent *d;
2829         int ret = -1;
2830
2831         dir = opendir(procfs__mountpoint());
2832         if (!dir)
2833                 return false;
2834
2835         /* Walk through the directory. */
2836         while (ret && (d = readdir(dir)) != NULL) {
2837                 char path[PATH_MAX];
2838                 char *data;
2839                 size_t size;
2840
2841                 if ((d->d_type != DT_DIR) ||
2842                      !strcmp(".", d->d_name) ||
2843                      !strcmp("..", d->d_name))
2844                         continue;
2845
2846                 scnprintf(path, sizeof(path), "%s/%s/comm",
2847                           procfs__mountpoint(), d->d_name);
2848
2849                 if (filename__read_str(path, &data, &size))
2850                         continue;
2851
2852                 ret = strncmp(name, data, len);
2853                 free(data);
2854         }
2855
2856         closedir(dir);
2857         return ret ? false : true;
2858 }
2859
2860 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target,
2861                               int err, char *msg, size_t size)
2862 {
2863         char sbuf[STRERR_BUFSIZE];
2864         int printed = 0;
2865
2866         switch (err) {
2867         case EPERM:
2868         case EACCES:
2869                 if (err == EPERM)
2870                         printed = scnprintf(msg, size,
2871                                 "No permission to enable %s event.\n\n",
2872                                 perf_evsel__name(evsel));
2873
2874                 return scnprintf(msg + printed, size - printed,
2875                  "You may not have permission to collect %sstats.\n\n"
2876                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2877                  "which controls use of the performance events system by\n"
2878                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2879                  "The current value is %d:\n\n"
2880                  "  -1: Allow use of (almost) all events by all users\n"
2881                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2882                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2883                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2884                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2885                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2886                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2887                  "      kernel.perf_event_paranoid = -1\n" ,
2888                                  target->system_wide ? "system-wide " : "",
2889                                  perf_event_paranoid());
2890         case ENOENT:
2891                 return scnprintf(msg, size, "The %s event is not supported.",
2892                                  perf_evsel__name(evsel));
2893         case EMFILE:
2894                 return scnprintf(msg, size, "%s",
2895                          "Too many events are opened.\n"
2896                          "Probably the maximum number of open file descriptors has been reached.\n"
2897                          "Hint: Try again after reducing the number of events.\n"
2898                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2899         case ENOMEM:
2900                 if (evsel__has_callchain(evsel) &&
2901                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2902                         return scnprintf(msg, size,
2903                                          "Not enough memory to setup event with callchain.\n"
2904                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2905                                          "Hint: Current value: %d", sysctl__max_stack());
2906                 break;
2907         case ENODEV:
2908                 if (target->cpu_list)
2909                         return scnprintf(msg, size, "%s",
2910          "No such device - did you specify an out-of-range profile CPU?");
2911                 break;
2912         case EOPNOTSUPP:
2913                 if (evsel->core.attr.sample_period != 0)
2914                         return scnprintf(msg, size,
2915         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2916                                          perf_evsel__name(evsel));
2917                 if (evsel->core.attr.precise_ip)
2918                         return scnprintf(msg, size, "%s",
2919         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2920 #if defined(__i386__) || defined(__x86_64__)
2921                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2922                         return scnprintf(msg, size, "%s",
2923         "No hardware sampling interrupt available.\n");
2924 #endif
2925                 break;
2926         case EBUSY:
2927                 if (find_process("oprofiled"))
2928                         return scnprintf(msg, size,
2929         "The PMU counters are busy/taken by another profiler.\n"
2930         "We found oprofile daemon running, please stop it and try again.");
2931                 break;
2932         case EINVAL:
2933                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2934                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2935                 if (perf_missing_features.clockid)
2936                         return scnprintf(msg, size, "clockid feature not supported.");
2937                 if (perf_missing_features.clockid_wrong)
2938                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2939                 if (perf_missing_features.aux_output)
2940                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2941                 break;
2942         default:
2943                 break;
2944         }
2945
2946         return scnprintf(msg, size,
2947         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2948         "/bin/dmesg | grep -i perf may provide additional information.\n",
2949                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2950                          perf_evsel__name(evsel));
2951 }
2952
2953 struct perf_env *perf_evsel__env(struct evsel *evsel)
2954 {
2955         if (evsel && evsel->evlist)
2956                 return evsel->evlist->env;
2957         return NULL;
2958 }
2959
2960 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2961 {
2962         int cpu, thread;
2963
2964         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2965                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2966                      thread++) {
2967                         int fd = FD(evsel, cpu, thread);
2968
2969                         if (perf_evlist__id_add_fd(evlist, evsel,
2970                                                    cpu, thread, fd) < 0)
2971                                 return -1;
2972                 }
2973         }
2974
2975         return 0;
2976 }
2977
2978 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2979 {
2980         struct perf_cpu_map *cpus = evsel->core.cpus;
2981         struct perf_thread_map *threads = evsel->core.threads;
2982
2983         if (perf_evsel__alloc_id(evsel, cpus->nr, threads->nr))
2984                 return -ENOMEM;
2985
2986         return store_evsel_ids(evsel, evlist);
2987 }