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 "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "../perf-sys.h"
  49 #include "util/parse-branch-options.h"
  50 #include <internal/xyarray.h>
  51 #include <internal/lib.h>
  52
  53 #include <linux/ctype.h>
  54
  55 struct perf_missing_features perf_missing_features;
  56
  57 static clockid_t clockid;
  58
  59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  60 {
  61         return 0;
  62 }
  63
  64 void __weak test_attr__ready(void) { }
  65
  66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  67 {
  68 }
  69
  70 static struct {
  71         size_t  size;
  72         int     (*init)(struct evsel *evsel);
  73         void    (*fini)(struct evsel *evsel);
  74 } perf_evsel__object = {
  75         .size = sizeof(struct evsel),
  76         .init = evsel__no_extra_init,
  77         .fini = evsel__no_extra_fini,
  78 };
  79
  80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
  81                          void (*fini)(struct evsel *evsel))
  82 {
  83
  84         if (object_size == 0)
  85                 goto set_methods;
  86
  87         if (perf_evsel__object.size > object_size)
  88                 return -EINVAL;
  89
  90         perf_evsel__object.size = object_size;
  91
  92 set_methods:
  93         if (init != NULL)
  94                 perf_evsel__object.init = init;
  95
  96         if (fini != NULL)
  97                 perf_evsel__object.fini = fini;
  98
  99         return 0;
 100 }
 101
 102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 103
 104 int __evsel__sample_size(u64 sample_type)
 105 {
 106         u64 mask = sample_type & PERF_SAMPLE_MASK;
 107         int size = 0;
 108         int i;
 109
 110         for (i = 0; i < 64; i++) {
 111                 if (mask & (1ULL << i))
 112                         size++;
 113         }
 114
 115         size *= sizeof(u64);
 116
 117         return size;
 118 }
 119
 120 /**
 121  * __perf_evsel__calc_id_pos - calculate id_pos.
 122  * @sample_type: sample type
 123  *
 124  * This function returns the position of the event id (PERF_SAMPLE_ID or
 125  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 126  * perf_record_sample.
 127  */
 128 static int __perf_evsel__calc_id_pos(u64 sample_type)
 129 {
 130         int idx = 0;
 131
 132         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 133                 return 0;
 134
 135         if (!(sample_type & PERF_SAMPLE_ID))
 136                 return -1;
 137
 138         if (sample_type & PERF_SAMPLE_IP)
 139                 idx += 1;
 140
 141         if (sample_type & PERF_SAMPLE_TID)
 142                 idx += 1;
 143
 144         if (sample_type & PERF_SAMPLE_TIME)
 145                 idx += 1;
 146
 147         if (sample_type & PERF_SAMPLE_ADDR)
 148                 idx += 1;
 149
 150         return idx;
 151 }
 152
 153 /**
 154  * __perf_evsel__calc_is_pos - calculate is_pos.
 155  * @sample_type: sample type
 156  *
 157  * This function returns the position (counting backwards) of the event id
 158  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 159  * sample_id_all is used there is an id sample appended to non-sample events.
 160  */
 161 static int __perf_evsel__calc_is_pos(u64 sample_type)
 162 {
 163         int idx = 1;
 164
 165         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 166                 return 1;
 167
 168         if (!(sample_type & PERF_SAMPLE_ID))
 169                 return -1;
 170
 171         if (sample_type & PERF_SAMPLE_CPU)
 172                 idx += 1;
 173
 174         if (sample_type & PERF_SAMPLE_STREAM_ID)
 175                 idx += 1;
 176
 177         return idx;
 178 }
 179
 180 void evsel__calc_id_pos(struct evsel *evsel)
 181 {
 182         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 183         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 184 }
 185
 186 void __evsel__set_sample_bit(struct evsel *evsel,
 187                                   enum perf_event_sample_format bit)
 188 {
 189         if (!(evsel->core.attr.sample_type & bit)) {
 190                 evsel->core.attr.sample_type |= bit;
 191                 evsel->sample_size += sizeof(u64);
 192                 evsel__calc_id_pos(evsel);
 193         }
 194 }
 195
 196 void __evsel__reset_sample_bit(struct evsel *evsel,
 197                                     enum perf_event_sample_format bit)
 198 {
 199         if (evsel->core.attr.sample_type & bit) {
 200                 evsel->core.attr.sample_type &= ~bit;
 201                 evsel->sample_size -= sizeof(u64);
 202                 evsel__calc_id_pos(evsel);
 203         }
 204 }
 205
 206 void evsel__set_sample_id(struct evsel *evsel,
 207                                bool can_sample_identifier)
 208 {
 209         if (can_sample_identifier) {
 210                 evsel__reset_sample_bit(evsel, ID);
 211                 evsel__set_sample_bit(evsel, IDENTIFIER);
 212         } else {
 213                 evsel__set_sample_bit(evsel, ID);
 214         }
 215         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 216 }
 217
 218 /**
 219  * evsel__is_function_event - Return whether given evsel is a function
 220  * trace event
 221  *
 222  * @evsel - evsel selector to be tested
 223  *
 224  * Return %true if event is function trace event
 225  */
 226 bool evsel__is_function_event(struct evsel *evsel)
 227 {
 228 #define FUNCTION_EVENT "ftrace:function"
 229
 230         return evsel->name &&
 231                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 232
 233 #undef FUNCTION_EVENT
 234 }
 235
 236 void evsel__init(struct evsel *evsel,
 237                  struct perf_event_attr *attr, int idx)
 238 {
 239         perf_evsel__init(&evsel->core, attr);
 240         evsel->idx         = idx;
 241         evsel->tracking    = !idx;
 242         evsel->leader      = evsel;
 243         evsel->unit        = "";
 244         evsel->scale       = 1.0;
 245         evsel->max_events  = ULONG_MAX;
 246         evsel->evlist      = NULL;
 247         evsel->bpf_obj     = NULL;
 248         evsel->bpf_fd      = -1;
 249         INIT_LIST_HEAD(&evsel->config_terms);
 250         perf_evsel__object.init(evsel);
 251         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 252         evsel__calc_id_pos(evsel);
 253         evsel->cmdline_group_boundary = false;
 254         evsel->metric_expr   = NULL;
 255         evsel->metric_name   = NULL;
 256         evsel->metric_events = NULL;
 257         evsel->per_pkg_mask  = NULL;
 258         evsel->collect_stat  = false;
 259         evsel->pmu_name      = NULL;
 260 }
 261
 262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 263 {
 264         struct evsel *evsel = zalloc(perf_evsel__object.size);
 265
 266         if (!evsel)
 267                 return NULL;
 268         evsel__init(evsel, attr, idx);
 269
 270         if (evsel__is_bpf_output(evsel)) {
 271                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 272                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 273                 evsel->core.attr.sample_period = 1;
 274         }
 275
 276         if (evsel__is_clock(evsel)) {
 277                 /*
 278                  * The evsel->unit points to static alias->unit
 279                  * so it's ok to use static string in here.
 280                  */
 281                 static const char *unit = "msec";
 282
 283                 evsel->unit = unit;
 284                 evsel->scale = 1e-6;
 285         }
 286
 287         return evsel;
 288 }
 289
 290 static bool perf_event_can_profile_kernel(void)
 291 {
 292         return perf_event_paranoid_check(1);
 293 }
 294
 295 struct evsel *evsel__new_cycles(bool precise)
 296 {
 297         struct perf_event_attr attr = {
 298                 .type   = PERF_TYPE_HARDWARE,
 299                 .config = PERF_COUNT_HW_CPU_CYCLES,
 300                 .exclude_kernel = !perf_event_can_profile_kernel(),
 301         };
 302         struct evsel *evsel;
 303
 304         event_attr_init(&attr);
 305
 306         if (!precise)
 307                 goto new_event;
 308
 309         /*
 310          * Now let the usual logic to set up the perf_event_attr defaults
 311          * to kick in when we return and before perf_evsel__open() is called.
 312          */
 313 new_event:
 314         evsel = evsel__new(&attr);
 315         if (evsel == NULL)
 316                 goto out;
 317
 318         evsel->precise_max = true;
 319
 320         /* use asprintf() because free(evsel) assumes name is allocated */
 321         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 322                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 323                      attr.exclude_kernel ? "u" : "",
 324                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 325                 goto error_free;
 326 out:
 327         return evsel;
 328 error_free:
 329         evsel__delete(evsel);
 330         evsel = NULL;
 331         goto out;
 332 }
 333
 334 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 335 {
 336         struct evsel_config_term *pos, *tmp;
 337
 338         list_for_each_entry(pos, &src->config_terms, list) {
 339                 tmp = malloc(sizeof(*tmp));
 340                 if (tmp == NULL)
 341                         return -ENOMEM;
 342
 343                 *tmp = *pos;
 344                 if (tmp->free_str) {
 345                         tmp->val.str = strdup(pos->val.str);
 346                         if (tmp->val.str == NULL) {
 347                                 free(tmp);
 348                                 return -ENOMEM;
 349                         }
 350                 }
 351                 list_add_tail(&tmp->list, &dst->config_terms);
 352         }
 353         return 0;
 354 }
 355
 356 /**
 357  * evsel__clone - create a new evsel copied from @orig
 358  * @orig: original evsel
 359  *
 360  * The assumption is that @orig is not configured nor opened yet.
 361  * So we only care about the attributes that can be set while it's parsed.
 362  */
 363 struct evsel *evsel__clone(struct evsel *orig)
 364 {
 365         struct evsel *evsel;
 366
 367         BUG_ON(orig->core.fd);
 368         BUG_ON(orig->counts);
 369         BUG_ON(orig->priv);
 370         BUG_ON(orig->per_pkg_mask);
 371
 372         /* cannot handle BPF objects for now */
 373         if (orig->bpf_obj)
 374                 return NULL;
 375
 376         evsel = evsel__new(&orig->core.attr);
 377         if (evsel == NULL)
 378                 return NULL;
 379
 380         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 381         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 382         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 383         evsel->core.nr_members = orig->core.nr_members;
 384         evsel->core.system_wide = orig->core.system_wide;
 385
 386         if (orig->name) {
 387                 evsel->name = strdup(orig->name);
 388                 if (evsel->name == NULL)
 389                         goto out_err;
 390         }
 391         if (orig->group_name) {
 392                 evsel->group_name = strdup(orig->group_name);
 393                 if (evsel->group_name == NULL)
 394                         goto out_err;
 395         }
 396         if (orig->pmu_name) {
 397                 evsel->pmu_name = strdup(orig->pmu_name);
 398                 if (evsel->pmu_name == NULL)
 399                         goto out_err;
 400         }
 401         if (orig->filter) {
 402                 evsel->filter = strdup(orig->filter);
 403                 if (evsel->filter == NULL)
 404                         goto out_err;
 405         }
 406         evsel->cgrp = cgroup__get(orig->cgrp);
 407         evsel->tp_format = orig->tp_format;
 408         evsel->handler = orig->handler;
 409         evsel->leader = orig->leader;
 410
 411         evsel->max_events = orig->max_events;
 412         evsel->tool_event = orig->tool_event;
 413         evsel->unit = orig->unit;
 414         evsel->scale = orig->scale;
 415         evsel->snapshot = orig->snapshot;
 416         evsel->per_pkg = orig->per_pkg;
 417         evsel->percore = orig->percore;
 418         evsel->precise_max = orig->precise_max;
 419         evsel->use_uncore_alias = orig->use_uncore_alias;
 420         evsel->is_libpfm_event = orig->is_libpfm_event;
 421
 422         evsel->exclude_GH = orig->exclude_GH;
 423         evsel->sample_read = orig->sample_read;
 424         evsel->auto_merge_stats = orig->auto_merge_stats;
 425         evsel->collect_stat = orig->collect_stat;
 426         evsel->weak_group = orig->weak_group;
 427
 428         if (evsel__copy_config_terms(evsel, orig) < 0)
 429                 goto out_err;
 430
 431         return evsel;
 432
 433 out_err:
 434         evsel__delete(evsel);
 435         return NULL;
 436 }
 437
 438 /*
 439  * Returns pointer with encoded error via <linux/err.h> interface.
 440  */
 441 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 442 {
 443         struct evsel *evsel = zalloc(perf_evsel__object.size);
 444         int err = -ENOMEM;
 445
 446         if (evsel == NULL) {
 447                 goto out_err;
 448         } else {
 449                 struct perf_event_attr attr = {
 450                         .type          = PERF_TYPE_TRACEPOINT,
 451                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 452                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 453                 };
 454
 455                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 456                         goto out_free;
 457
 458                 evsel->tp_format = trace_event__tp_format(sys, name);
 459                 if (IS_ERR(evsel->tp_format)) {
 460                         err = PTR_ERR(evsel->tp_format);
 461                         goto out_free;
 462                 }
 463
 464                 event_attr_init(&attr);
 465                 attr.config = evsel->tp_format->id;
 466                 attr.sample_period = 1;
 467                 evsel__init(evsel, &attr, idx);
 468         }
 469
 470         return evsel;
 471
 472 out_free:
 473         zfree(&evsel->name);
 474         free(evsel);
 475 out_err:
 476         return ERR_PTR(err);
 477 }
 478
 479 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
 480         "cycles",
 481         "instructions",
 482         "cache-references",
 483         "cache-misses",
 484         "branches",
 485         "branch-misses",
 486         "bus-cycles",
 487         "stalled-cycles-frontend",
 488         "stalled-cycles-backend",
 489         "ref-cycles",
 490 };
 491
 492 static const char *__evsel__hw_name(u64 config)
 493 {
 494         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 495                 return evsel__hw_names[config];
 496
 497         return "unknown-hardware";
 498 }
 499
 500 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 501 {
 502         int colon = 0, r = 0;
 503         struct perf_event_attr *attr = &evsel->core.attr;
 504         bool exclude_guest_default = false;
 505
 506 #define MOD_PRINT(context, mod) do {                                    \
 507                 if (!attr->exclude_##context) {                         \
 508                         if (!colon) colon = ++r;                        \
 509                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 510                 } } while(0)
 511
 512         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 513                 MOD_PRINT(kernel, 'k');
 514                 MOD_PRINT(user, 'u');
 515                 MOD_PRINT(hv, 'h');
 516                 exclude_guest_default = true;
 517         }
 518
 519         if (attr->precise_ip) {
 520                 if (!colon)
 521                         colon = ++r;
 522                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 523                 exclude_guest_default = true;
 524         }
 525
 526         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 527                 MOD_PRINT(host, 'H');
 528                 MOD_PRINT(guest, 'G');
 529         }
 530 #undef MOD_PRINT
 531         if (colon)
 532                 bf[colon - 1] = ':';
 533         return r;
 534 }
 535
 536 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 537 {
 538         int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 539         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 540 }
 541
 542 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
 543         "cpu-clock",
 544         "task-clock",
 545         "page-faults",
 546         "context-switches",
 547         "cpu-migrations",
 548         "minor-faults",
 549         "major-faults",
 550         "alignment-faults",
 551         "emulation-faults",
 552         "dummy",
 553 };
 554
 555 static const char *__evsel__sw_name(u64 config)
 556 {
 557         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 558                 return evsel__sw_names[config];
 559         return "unknown-software";
 560 }
 561
 562 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 563 {
 564         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 565         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 566 }
 567
 568 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 569 {
 570         int r;
 571
 572         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 573
 574         if (type & HW_BREAKPOINT_R)
 575                 r += scnprintf(bf + r, size - r, "r");
 576
 577         if (type & HW_BREAKPOINT_W)
 578                 r += scnprintf(bf + r, size - r, "w");
 579
 580         if (type & HW_BREAKPOINT_X)
 581                 r += scnprintf(bf + r, size - r, "x");
 582
 583         return r;
 584 }
 585
 586 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 587 {
 588         struct perf_event_attr *attr = &evsel->core.attr;
 589         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 590         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 591 }
 592
 593 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 594  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 595  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 596  { "LLC",       "L2",                                                   },
 597  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 598  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 599  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 600  { "node",                                                              },
 601 };
 602
 603 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 604  { "load",      "loads",        "read",                                 },
 605  { "store",     "stores",       "write",                                },
 606  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 607 };
 608
 609 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 610  { "refs",      "Reference",    "ops",          "access",               },
 611  { "misses",    "miss",                                                 },
 612 };
 613
 614 #define C(x)            PERF_COUNT_HW_CACHE_##x
 615 #define CACHE_READ      (1 << C(OP_READ))
 616 #define CACHE_WRITE     (1 << C(OP_WRITE))
 617 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 618 #define COP(x)          (1 << x)
 619
 620 /*
 621  * cache operartion stat
 622  * L1I : Read and prefetch only
 623  * ITLB and BPU : Read-only
 624  */
 625 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
 626  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 627  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 628  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 629  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 630  [C(ITLB)]      = (CACHE_READ),
 631  [C(BPU)]       = (CACHE_READ),
 632  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 633 };
 634
 635 bool evsel__is_cache_op_valid(u8 type, u8 op)
 636 {
 637         if (evsel__hw_cache_stat[type] & COP(op))
 638                 return true;    /* valid */
 639         else
 640                 return false;   /* invalid */
 641 }
 642
 643 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 644 {
 645         if (result) {
 646                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 647                                  evsel__hw_cache_op[op][0],
 648                                  evsel__hw_cache_result[result][0]);
 649         }
 650
 651         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 652                          evsel__hw_cache_op[op][1]);
 653 }
 654
 655 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 656 {
 657         u8 op, result, type = (config >>  0) & 0xff;
 658         const char *err = "unknown-ext-hardware-cache-type";
 659
 660         if (type >= PERF_COUNT_HW_CACHE_MAX)
 661                 goto out_err;
 662
 663         op = (config >>  8) & 0xff;
 664         err = "unknown-ext-hardware-cache-op";
 665         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 666                 goto out_err;
 667
 668         result = (config >> 16) & 0xff;
 669         err = "unknown-ext-hardware-cache-result";
 670         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 671                 goto out_err;
 672
 673         err = "invalid-cache";
 674         if (!evsel__is_cache_op_valid(type, op))
 675                 goto out_err;
 676
 677         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 678 out_err:
 679         return scnprintf(bf, size, "%s", err);
 680 }
 681
 682 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 683 {
 684         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 685         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 686 }
 687
 688 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 689 {
 690         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 691         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 692 }
 693
 694 static int evsel__tool_name(char *bf, size_t size)
 695 {
 696         int ret = scnprintf(bf, size, "duration_time");
 697         return ret;
 698 }
 699
 700 const char *evsel__name(struct evsel *evsel)
 701 {
 702         char bf[128];
 703
 704         if (!evsel)
 705                 goto out_unknown;
 706
 707         if (evsel->name)
 708                 return evsel->name;
 709
 710         switch (evsel->core.attr.type) {
 711         case PERF_TYPE_RAW:
 712                 evsel__raw_name(evsel, bf, sizeof(bf));
 713                 break;
 714
 715         case PERF_TYPE_HARDWARE:
 716                 evsel__hw_name(evsel, bf, sizeof(bf));
 717                 break;
 718
 719         case PERF_TYPE_HW_CACHE:
 720                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 721                 break;
 722
 723         case PERF_TYPE_SOFTWARE:
 724                 if (evsel->tool_event)
 725                         evsel__tool_name(bf, sizeof(bf));
 726                 else
 727                         evsel__sw_name(evsel, bf, sizeof(bf));
 728                 break;
 729
 730         case PERF_TYPE_TRACEPOINT:
 731                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 732                 break;
 733
 734         case PERF_TYPE_BREAKPOINT:
 735                 evsel__bp_name(evsel, bf, sizeof(bf));
 736                 break;
 737
 738         default:
 739                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 740                           evsel->core.attr.type);
 741                 break;
 742         }
 743
 744         evsel->name = strdup(bf);
 745
 746         if (evsel->name)
 747                 return evsel->name;
 748 out_unknown:
 749         return "unknown";
 750 }
 751
 752 const char *evsel__group_name(struct evsel *evsel)
 753 {
 754         return evsel->group_name ?: "anon group";
 755 }
 756
 757 /*
 758  * Returns the group details for the specified leader,
 759  * with following rules.
 760  *
 761  *  For record -e '{cycles,instructions}'
 762  *    'anon group { cycles:u, instructions:u }'
 763  *
 764  *  For record -e 'cycles,instructions' and report --group
 765  *    'cycles:u, instructions:u'
 766  */
 767 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 768 {
 769         int ret = 0;
 770         struct evsel *pos;
 771         const char *group_name = evsel__group_name(evsel);
 772
 773         if (!evsel->forced_leader)
 774                 ret = scnprintf(buf, size, "%s { ", group_name);
 775
 776         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 777
 778         for_each_group_member(pos, evsel)
 779                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 780
 781         if (!evsel->forced_leader)
 782                 ret += scnprintf(buf + ret, size - ret, " }");
 783
 784         return ret;
 785 }
 786
 787 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 788                                       struct callchain_param *param)
 789 {
 790         bool function = evsel__is_function_event(evsel);
 791         struct perf_event_attr *attr = &evsel->core.attr;
 792
 793         evsel__set_sample_bit(evsel, CALLCHAIN);
 794
 795         attr->sample_max_stack = param->max_stack;
 796
 797         if (opts->kernel_callchains)
 798                 attr->exclude_callchain_user = 1;
 799         if (opts->user_callchains)
 800                 attr->exclude_callchain_kernel = 1;
 801         if (param->record_mode == CALLCHAIN_LBR) {
 802                 if (!opts->branch_stack) {
 803                         if (attr->exclude_user) {
 804                                 pr_warning("LBR callstack option is only available "
 805                                            "to get user callchain information. "
 806                                            "Falling back to framepointers.\n");
 807                         } else {
 808                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 809                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 810                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 811                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 812                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 813                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 814                         }
 815                 } else
 816                          pr_warning("Cannot use LBR callstack with branch stack. "
 817                                     "Falling back to framepointers.\n");
 818         }
 819
 820         if (param->record_mode == CALLCHAIN_DWARF) {
 821                 if (!function) {
 822                         evsel__set_sample_bit(evsel, REGS_USER);
 823                         evsel__set_sample_bit(evsel, STACK_USER);
 824                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 825                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 826                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 827                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 828                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 829                         } else {
 830                                 attr->sample_regs_user |= PERF_REGS_MASK;
 831                         }
 832                         attr->sample_stack_user = param->dump_size;
 833                         attr->exclude_callchain_user = 1;
 834                 } else {
 835                         pr_info("Cannot use DWARF unwind for function trace event,"
 836                                 " falling back to framepointers.\n");
 837                 }
 838         }
 839
 840         if (function) {
 841                 pr_info("Disabling user space callchains for function trace event.\n");
 842                 attr->exclude_callchain_user = 1;
 843         }
 844 }
 845
 846 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 847                              struct callchain_param *param)
 848 {
 849         if (param->enabled)
 850                 return __evsel__config_callchain(evsel, opts, param);
 851 }
 852
 853 static void
 854 perf_evsel__reset_callgraph(struct evsel *evsel,
 855                             struct callchain_param *param)
 856 {
 857         struct perf_event_attr *attr = &evsel->core.attr;
 858
 859         evsel__reset_sample_bit(evsel, CALLCHAIN);
 860         if (param->record_mode == CALLCHAIN_LBR) {
 861                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 862                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 863                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 864                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 865         }
 866         if (param->record_mode == CALLCHAIN_DWARF) {
 867                 evsel__reset_sample_bit(evsel, REGS_USER);
 868                 evsel__reset_sample_bit(evsel, STACK_USER);
 869         }
 870 }
 871
 872 static void evsel__apply_config_terms(struct evsel *evsel,
 873                                       struct record_opts *opts, bool track)
 874 {
 875         struct evsel_config_term *term;
 876         struct list_head *config_terms = &evsel->config_terms;
 877         struct perf_event_attr *attr = &evsel->core.attr;
 878         /* callgraph default */
 879         struct callchain_param param = {
 880                 .record_mode = callchain_param.record_mode,
 881         };
 882         u32 dump_size = 0;
 883         int max_stack = 0;
 884         const char *callgraph_buf = NULL;
 885
 886         list_for_each_entry(term, config_terms, list) {
 887                 switch (term->type) {
 888                 case EVSEL__CONFIG_TERM_PERIOD:
 889                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 890                                 attr->sample_period = term->val.period;
 891                                 attr->freq = 0;
 892                                 evsel__reset_sample_bit(evsel, PERIOD);
 893                         }
 894                         break;
 895                 case EVSEL__CONFIG_TERM_FREQ:
 896                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 897                                 attr->sample_freq = term->val.freq;
 898                                 attr->freq = 1;
 899                                 evsel__set_sample_bit(evsel, PERIOD);
 900                         }
 901                         break;
 902                 case EVSEL__CONFIG_TERM_TIME:
 903                         if (term->val.time)
 904                                 evsel__set_sample_bit(evsel, TIME);
 905                         else
 906                                 evsel__reset_sample_bit(evsel, TIME);
 907                         break;
 908                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 909                         callgraph_buf = term->val.str;
 910                         break;
 911                 case EVSEL__CONFIG_TERM_BRANCH:
 912                         if (term->val.str && strcmp(term->val.str, "no")) {
 913                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 914                                 parse_branch_str(term->val.str,
 915                                                  &attr->branch_sample_type);
 916                         } else
 917                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 918                         break;
 919                 case EVSEL__CONFIG_TERM_STACK_USER:
 920                         dump_size = term->val.stack_user;
 921                         break;
 922                 case EVSEL__CONFIG_TERM_MAX_STACK:
 923                         max_stack = term->val.max_stack;
 924                         break;
 925                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 926                         evsel->max_events = term->val.max_events;
 927                         break;
 928                 case EVSEL__CONFIG_TERM_INHERIT:
 929                         /*
 930                          * attr->inherit should has already been set by
 931                          * evsel__config. If user explicitly set
 932                          * inherit using config terms, override global
 933                          * opt->no_inherit setting.
 934                          */
 935                         attr->inherit = term->val.inherit ? 1 : 0;
 936                         break;
 937                 case EVSEL__CONFIG_TERM_OVERWRITE:
 938                         attr->write_backward = term->val.overwrite ? 1 : 0;
 939                         break;
 940                 case EVSEL__CONFIG_TERM_DRV_CFG:
 941                         break;
 942                 case EVSEL__CONFIG_TERM_PERCORE:
 943                         break;
 944                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
 945                         attr->aux_output = term->val.aux_output ? 1 : 0;
 946                         break;
 947                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 948                         /* Already applied by auxtrace */
 949                         break;
 950                 case EVSEL__CONFIG_TERM_CFG_CHG:
 951                         break;
 952                 default:
 953                         break;
 954                 }
 955         }
 956
 957         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 958         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 959                 bool sample_address = false;
 960
 961                 if (max_stack) {
 962                         param.max_stack = max_stack;
 963                         if (callgraph_buf == NULL)
 964                                 callgraph_buf = "fp";
 965                 }
 966
 967                 /* parse callgraph parameters */
 968                 if (callgraph_buf != NULL) {
 969                         if (!strcmp(callgraph_buf, "no")) {
 970                                 param.enabled = false;
 971                                 param.record_mode = CALLCHAIN_NONE;
 972                         } else {
 973                                 param.enabled = true;
 974                                 if (parse_callchain_record(callgraph_buf, &param)) {
 975                                         pr_err("per-event callgraph setting for %s failed. "
 976                                                "Apply callgraph global setting for it\n",
 977                                                evsel->name);
 978                                         return;
 979                                 }
 980                                 if (param.record_mode == CALLCHAIN_DWARF)
 981                                         sample_address = true;
 982                         }
 983                 }
 984                 if (dump_size > 0) {
 985                         dump_size = round_up(dump_size, sizeof(u64));
 986                         param.dump_size = dump_size;
 987                 }
 988
 989                 /* If global callgraph set, clear it */
 990                 if (callchain_param.enabled)
 991                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 992
 993                 /* set perf-event callgraph */
 994                 if (param.enabled) {
 995                         if (sample_address) {
 996                                 evsel__set_sample_bit(evsel, ADDR);
 997                                 evsel__set_sample_bit(evsel, DATA_SRC);
 998                                 evsel->core.attr.mmap_data = track;
 999                         }
1000                         evsel__config_callchain(evsel, opts, &param);
1001                 }
1002         }
1003 }
1004
1005 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1006 {
1007         struct evsel_config_term *term, *found_term = NULL;
1008
1009         list_for_each_entry(term, &evsel->config_terms, list) {
1010                 if (term->type == type)
1011                         found_term = term;
1012         }
1013
1014         return found_term;
1015 }
1016
1017 /*
1018  * The enable_on_exec/disabled value strategy:
1019  *
1020  *  1) For any type of traced program:
1021  *    - all independent events and group leaders are disabled
1022  *    - all group members are enabled
1023  *
1024  *     Group members are ruled by group leaders. They need to
1025  *     be enabled, because the group scheduling relies on that.
1026  *
1027  *  2) For traced programs executed by perf:
1028  *     - all independent events and group leaders have
1029  *       enable_on_exec set
1030  *     - we don't specifically enable or disable any event during
1031  *       the record command
1032  *
1033  *     Independent events and group leaders are initially disabled
1034  *     and get enabled by exec. Group members are ruled by group
1035  *     leaders as stated in 1).
1036  *
1037  *  3) For traced programs attached by perf (pid/tid):
1038  *     - we specifically enable or disable all events during
1039  *       the record command
1040  *
1041  *     When attaching events to already running traced we
1042  *     enable/disable events specifically, as there's no
1043  *     initial traced exec call.
1044  */
1045 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1046                    struct callchain_param *callchain)
1047 {
1048         struct evsel *leader = evsel->leader;
1049         struct perf_event_attr *attr = &evsel->core.attr;
1050         int track = evsel->tracking;
1051         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1052
1053         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1054         attr->inherit       = !opts->no_inherit;
1055         attr->write_backward = opts->overwrite ? 1 : 0;
1056
1057         evsel__set_sample_bit(evsel, IP);
1058         evsel__set_sample_bit(evsel, TID);
1059
1060         if (evsel->sample_read) {
1061                 evsel__set_sample_bit(evsel, READ);
1062
1063                 /*
1064                  * We need ID even in case of single event, because
1065                  * PERF_SAMPLE_READ process ID specific data.
1066                  */
1067                 evsel__set_sample_id(evsel, false);
1068
1069                 /*
1070                  * Apply group format only if we belong to group
1071                  * with more than one members.
1072                  */
1073                 if (leader->core.nr_members > 1) {
1074                         attr->read_format |= PERF_FORMAT_GROUP;
1075                         attr->inherit = 0;
1076                 }
1077         }
1078
1079         /*
1080          * We default some events to have a default interval. But keep
1081          * it a weak assumption overridable by the user.
1082          */
1083         if (!attr->sample_period) {
1084                 if (opts->freq) {
1085                         attr->freq              = 1;
1086                         attr->sample_freq       = opts->freq;
1087                 } else {
1088                         attr->sample_period = opts->default_interval;
1089                 }
1090         }
1091         /*
1092          * If attr->freq was set (here or earlier), ask for period
1093          * to be sampled.
1094          */
1095         if (attr->freq)
1096                 evsel__set_sample_bit(evsel, PERIOD);
1097
1098         if (opts->no_samples)
1099                 attr->sample_freq = 0;
1100
1101         if (opts->inherit_stat) {
1102                 evsel->core.attr.read_format |=
1103                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1104                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1105                         PERF_FORMAT_ID;
1106                 attr->inherit_stat = 1;
1107         }
1108
1109         if (opts->sample_address) {
1110                 evsel__set_sample_bit(evsel, ADDR);
1111                 attr->mmap_data = track;
1112         }
1113
1114         /*
1115          * We don't allow user space callchains for  function trace
1116          * event, due to issues with page faults while tracing page
1117          * fault handler and its overall trickiness nature.
1118          */
1119         if (evsel__is_function_event(evsel))
1120                 evsel->core.attr.exclude_callchain_user = 1;
1121
1122         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1123                 evsel__config_callchain(evsel, opts, callchain);
1124
1125         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1126             !evsel__is_dummy_event(evsel)) {
1127                 attr->sample_regs_intr = opts->sample_intr_regs;
1128                 evsel__set_sample_bit(evsel, REGS_INTR);
1129         }
1130
1131         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1132             !evsel__is_dummy_event(evsel)) {
1133                 attr->sample_regs_user |= opts->sample_user_regs;
1134                 evsel__set_sample_bit(evsel, REGS_USER);
1135         }
1136
1137         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1138                 evsel__set_sample_bit(evsel, CPU);
1139
1140         /*
1141          * When the user explicitly disabled time don't force it here.
1142          */
1143         if (opts->sample_time &&
1144             (!perf_missing_features.sample_id_all &&
1145             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1146              opts->sample_time_set)))
1147                 evsel__set_sample_bit(evsel, TIME);
1148
1149         if (opts->raw_samples && !evsel->no_aux_samples) {
1150                 evsel__set_sample_bit(evsel, TIME);
1151                 evsel__set_sample_bit(evsel, RAW);
1152                 evsel__set_sample_bit(evsel, CPU);
1153         }
1154
1155         if (opts->sample_address)
1156                 evsel__set_sample_bit(evsel, DATA_SRC);
1157
1158         if (opts->sample_phys_addr)
1159                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1160
1161         if (opts->no_buffering) {
1162                 attr->watermark = 0;
1163                 attr->wakeup_events = 1;
1164         }
1165         if (opts->branch_stack && !evsel->no_aux_samples) {
1166                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1167                 attr->branch_sample_type = opts->branch_stack;
1168         }
1169
1170         if (opts->sample_weight)
1171                 evsel__set_sample_bit(evsel, WEIGHT);
1172
1173         attr->task  = track;
1174         attr->mmap  = track;
1175         attr->mmap2 = track && !perf_missing_features.mmap2;
1176         attr->comm  = track;
1177         /*
1178          * ksymbol is tracked separately with text poke because it needs to be
1179          * system wide and enabled immediately.
1180          */
1181         if (!opts->text_poke)
1182                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1183         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1184
1185         if (opts->record_namespaces)
1186                 attr->namespaces  = track;
1187
1188         if (opts->record_cgroup) {
1189                 attr->cgroup = track && !perf_missing_features.cgroup;
1190                 evsel__set_sample_bit(evsel, CGROUP);
1191         }
1192
1193         if (opts->record_switch_events)
1194                 attr->context_switch = track;
1195
1196         if (opts->sample_transaction)
1197                 evsel__set_sample_bit(evsel, TRANSACTION);
1198
1199         if (opts->running_time) {
1200                 evsel->core.attr.read_format |=
1201                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1202                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1203         }
1204
1205         /*
1206          * XXX see the function comment above
1207          *
1208          * Disabling only independent events or group leaders,
1209          * keeping group members enabled.
1210          */
1211         if (evsel__is_group_leader(evsel))
1212                 attr->disabled = 1;
1213
1214         /*
1215          * Setting enable_on_exec for independent events and
1216          * group leaders for traced executed by perf.
1217          */
1218         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1219             !opts->initial_delay)
1220                 attr->enable_on_exec = 1;
1221
1222         if (evsel->immediate) {
1223                 attr->disabled = 0;
1224                 attr->enable_on_exec = 0;
1225         }
1226
1227         clockid = opts->clockid;
1228         if (opts->use_clockid) {
1229                 attr->use_clockid = 1;
1230                 attr->clockid = opts->clockid;
1231         }
1232
1233         if (evsel->precise_max)
1234                 attr->precise_ip = 3;
1235
1236         if (opts->all_user) {
1237                 attr->exclude_kernel = 1;
1238                 attr->exclude_user   = 0;
1239         }
1240
1241         if (opts->all_kernel) {
1242                 attr->exclude_kernel = 0;
1243                 attr->exclude_user   = 1;
1244         }
1245
1246         if (evsel->core.own_cpus || evsel->unit)
1247                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1248
1249         /*
1250          * Apply event specific term settings,
1251          * it overloads any global configuration.
1252          */
1253         evsel__apply_config_terms(evsel, opts, track);
1254
1255         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1256
1257         /* The --period option takes the precedence. */
1258         if (opts->period_set) {
1259                 if (opts->period)
1260                         evsel__set_sample_bit(evsel, PERIOD);
1261                 else
1262                         evsel__reset_sample_bit(evsel, PERIOD);
1263         }
1264
1265         /*
1266          * A dummy event never triggers any actual counter and therefore
1267          * cannot be used with branch_stack.
1268          *
1269          * For initial_delay, a dummy event is added implicitly.
1270          * The software event will trigger -EOPNOTSUPP error out,
1271          * if BRANCH_STACK bit is set.
1272          */
1273         if (evsel__is_dummy_event(evsel))
1274                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1275 }
1276
1277 int evsel__set_filter(struct evsel *evsel, const char *filter)
1278 {
1279         char *new_filter = strdup(filter);
1280
1281         if (new_filter != NULL) {
1282                 free(evsel->filter);
1283                 evsel->filter = new_filter;
1284                 return 0;
1285         }
1286
1287         return -1;
1288 }
1289
1290 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1291 {
1292         char *new_filter;
1293
1294         if (evsel->filter == NULL)
1295                 return evsel__set_filter(evsel, filter);
1296
1297         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1298                 free(evsel->filter);
1299                 evsel->filter = new_filter;
1300                 return 0;
1301         }
1302
1303         return -1;
1304 }
1305
1306 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1307 {
1308         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1309 }
1310
1311 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1312 {
1313         return evsel__append_filter(evsel, "%s,%s", filter);
1314 }
1315
1316 /* Caller has to clear disabled after going through all CPUs. */
1317 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1318 {
1319         return perf_evsel__enable_cpu(&evsel->core, cpu);
1320 }
1321
1322 int evsel__enable(struct evsel *evsel)
1323 {
1324         int err = perf_evsel__enable(&evsel->core);
1325
1326         if (!err)
1327                 evsel->disabled = false;
1328         return err;
1329 }
1330
1331 /* Caller has to set disabled after going through all CPUs. */
1332 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1333 {
1334         return perf_evsel__disable_cpu(&evsel->core, cpu);
1335 }
1336
1337 int evsel__disable(struct evsel *evsel)
1338 {
1339         int err = perf_evsel__disable(&evsel->core);
1340         /*
1341          * We mark it disabled here so that tools that disable a event can
1342          * ignore events after they disable it. I.e. the ring buffer may have
1343          * already a few more events queued up before the kernel got the stop
1344          * request.
1345          */
1346         if (!err)
1347                 evsel->disabled = true;
1348
1349         return err;
1350 }
1351
1352 static void evsel__free_config_terms(struct evsel *evsel)
1353 {
1354         struct evsel_config_term *term, *h;
1355
1356         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1357                 list_del_init(&term->list);
1358                 if (term->free_str)
1359                         zfree(&term->val.str);
1360                 free(term);
1361         }
1362 }
1363
1364 void evsel__exit(struct evsel *evsel)
1365 {
1366         assert(list_empty(&evsel->core.node));
1367         assert(evsel->evlist == NULL);
1368         evsel__free_counts(evsel);
1369         perf_evsel__free_fd(&evsel->core);
1370         perf_evsel__free_id(&evsel->core);
1371         evsel__free_config_terms(evsel);
1372         cgroup__put(evsel->cgrp);
1373         perf_cpu_map__put(evsel->core.cpus);
1374         perf_cpu_map__put(evsel->core.own_cpus);
1375         perf_thread_map__put(evsel->core.threads);
1376         zfree(&evsel->group_name);
1377         zfree(&evsel->name);
1378         zfree(&evsel->pmu_name);
1379         zfree(&evsel->per_pkg_mask);
1380         zfree(&evsel->metric_events);
1381         perf_evsel__object.fini(evsel);
1382 }
1383
1384 void evsel__delete(struct evsel *evsel)
1385 {
1386         evsel__exit(evsel);
1387         free(evsel);
1388 }
1389
1390 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1391                            struct perf_counts_values *count)
1392 {
1393         struct perf_counts_values tmp;
1394
1395         if (!evsel->prev_raw_counts)
1396                 return;
1397
1398         if (cpu == -1) {
1399                 tmp = evsel->prev_raw_counts->aggr;
1400                 evsel->prev_raw_counts->aggr = *count;
1401         } else {
1402                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1403                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1404         }
1405
1406         count->val = count->val - tmp.val;
1407         count->ena = count->ena - tmp.ena;
1408         count->run = count->run - tmp.run;
1409 }
1410
1411 void perf_counts_values__scale(struct perf_counts_values *count,
1412                                bool scale, s8 *pscaled)
1413 {
1414         s8 scaled = 0;
1415
1416         if (scale) {
1417                 if (count->run == 0) {
1418                         scaled = -1;
1419                         count->val = 0;
1420                 } else if (count->run < count->ena) {
1421                         scaled = 1;
1422                         count->val = (u64)((double) count->val * count->ena / count->run);
1423                 }
1424         }
1425
1426         if (pscaled)
1427                 *pscaled = scaled;
1428 }
1429
1430 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1431 {
1432         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1433
1434         return perf_evsel__read(&evsel->core, cpu, thread, count);
1435 }
1436
1437 static void
1438 perf_evsel__set_count(struct evsel *counter, int cpu, int thread,
1439                       u64 val, u64 ena, u64 run)
1440 {
1441         struct perf_counts_values *count;
1442
1443         count = perf_counts(counter->counts, cpu, thread);
1444
1445         count->val    = val;
1446         count->ena    = ena;
1447         count->run    = run;
1448
1449         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1450 }
1451
1452 static int
1453 perf_evsel__process_group_data(struct evsel *leader,
1454                                int cpu, int thread, u64 *data)
1455 {
1456         u64 read_format = leader->core.attr.read_format;
1457         struct sample_read_value *v;
1458         u64 nr, ena = 0, run = 0, i;
1459
1460         nr = *data++;
1461
1462         if (nr != (u64) leader->core.nr_members)
1463                 return -EINVAL;
1464
1465         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1466                 ena = *data++;
1467
1468         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1469                 run = *data++;
1470
1471         v = (struct sample_read_value *) data;
1472
1473         perf_evsel__set_count(leader, cpu, thread,
1474                               v[0].value, ena, run);
1475
1476         for (i = 1; i < nr; i++) {
1477                 struct evsel *counter;
1478
1479                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1480                 if (!counter)
1481                         return -EINVAL;
1482
1483                 perf_evsel__set_count(counter, cpu, thread,
1484                                       v[i].value, ena, run);
1485         }
1486
1487         return 0;
1488 }
1489
1490 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1491 {
1492         struct perf_stat_evsel *ps = leader->stats;
1493         u64 read_format = leader->core.attr.read_format;
1494         int size = perf_evsel__read_size(&leader->core);
1495         u64 *data = ps->group_data;
1496
1497         if (!(read_format & PERF_FORMAT_ID))
1498                 return -EINVAL;
1499
1500         if (!evsel__is_group_leader(leader))
1501                 return -EINVAL;
1502
1503         if (!data) {
1504                 data = zalloc(size);
1505                 if (!data)
1506                         return -ENOMEM;
1507
1508                 ps->group_data = data;
1509         }
1510
1511         if (FD(leader, cpu, thread) < 0)
1512                 return -EINVAL;
1513
1514         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1515                 return -errno;
1516
1517         return perf_evsel__process_group_data(leader, cpu, thread, data);
1518 }
1519
1520 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1521 {
1522         u64 read_format = evsel->core.attr.read_format;
1523
1524         if (read_format & PERF_FORMAT_GROUP)
1525                 return evsel__read_group(evsel, cpu, thread);
1526
1527         return evsel__read_one(evsel, cpu, thread);
1528 }
1529
1530 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1531 {
1532         struct perf_counts_values count;
1533         size_t nv = scale ? 3 : 1;
1534
1535         if (FD(evsel, cpu, thread) < 0)
1536                 return -EINVAL;
1537
1538         if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1539                 return -ENOMEM;
1540
1541         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1542                 return -errno;
1543
1544         evsel__compute_deltas(evsel, cpu, thread, &count);
1545         perf_counts_values__scale(&count, scale, NULL);
1546         *perf_counts(evsel->counts, cpu, thread) = count;
1547         return 0;
1548 }
1549
1550 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1551 {
1552         struct evsel *leader = evsel->leader;
1553         int fd;
1554
1555         if (evsel__is_group_leader(evsel))
1556                 return -1;
1557
1558         /*
1559          * Leader must be already processed/open,
1560          * if not it's a bug.
1561          */
1562         BUG_ON(!leader->core.fd);
1563
1564         fd = FD(leader, cpu, thread);
1565         BUG_ON(fd == -1);
1566
1567         return fd;
1568 }
1569
1570 static void perf_evsel__remove_fd(struct evsel *pos,
1571                                   int nr_cpus, int nr_threads,
1572                                   int thread_idx)
1573 {
1574         for (int cpu = 0; cpu < nr_cpus; cpu++)
1575                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1576                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1577 }
1578
1579 static int update_fds(struct evsel *evsel,
1580                       int nr_cpus, int cpu_idx,
1581                       int nr_threads, int thread_idx)
1582 {
1583         struct evsel *pos;
1584
1585         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1586                 return -EINVAL;
1587
1588         evlist__for_each_entry(evsel->evlist, pos) {
1589                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1590
1591                 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1592
1593                 /*
1594                  * Since fds for next evsel has not been created,
1595                  * there is no need to iterate whole event list.
1596                  */
1597                 if (pos == evsel)
1598                         break;
1599         }
1600         return 0;
1601 }
1602
1603 static bool ignore_missing_thread(struct evsel *evsel,
1604                                   int nr_cpus, int cpu,
1605                                   struct perf_thread_map *threads,
1606                                   int thread, int err)
1607 {
1608         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1609
1610         if (!evsel->ignore_missing_thread)
1611                 return false;
1612
1613         /* The system wide setup does not work with threads. */
1614         if (evsel->core.system_wide)
1615                 return false;
1616
1617         /* The -ESRCH is perf event syscall errno for pid's not found. */
1618         if (err != -ESRCH)
1619                 return false;
1620
1621         /* If there's only one thread, let it fail. */
1622         if (threads->nr == 1)
1623                 return false;
1624
1625         /*
1626          * We should remove fd for missing_thread first
1627          * because thread_map__remove() will decrease threads->nr.
1628          */
1629         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1630                 return false;
1631
1632         if (thread_map__remove(threads, thread))
1633                 return false;
1634
1635         pr_warning("WARNING: Ignored open failure for pid %d\n",
1636                    ignore_pid);
1637         return true;
1638 }
1639
1640 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1641                                 void *priv __maybe_unused)
1642 {
1643         return fprintf(fp, "  %-32s %s\n", name, val);
1644 }
1645
1646 static void display_attr(struct perf_event_attr *attr)
1647 {
1648         if (verbose >= 2 || debug_peo_args) {
1649                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1650                 fprintf(stderr, "perf_event_attr:\n");
1651                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1652                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1653         }
1654 }
1655
1656 static int perf_event_open(struct evsel *evsel,
1657                            pid_t pid, int cpu, int group_fd,
1658                            unsigned long flags)
1659 {
1660         int precise_ip = evsel->core.attr.precise_ip;
1661         int fd;
1662
1663         while (1) {
1664                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1665                           pid, cpu, group_fd, flags);
1666
1667                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1668                 if (fd >= 0)
1669                         break;
1670
1671                 /* Do not try less precise if not requested. */
1672                 if (!evsel->precise_max)
1673                         break;
1674
1675                 /*
1676                  * We tried all the precise_ip values, and it's
1677                  * still failing, so leave it to standard fallback.
1678                  */
1679                 if (!evsel->core.attr.precise_ip) {
1680                         evsel->core.attr.precise_ip = precise_ip;
1681                         break;
1682                 }
1683
1684                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1685                 evsel->core.attr.precise_ip--;
1686                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1687                 display_attr(&evsel->core.attr);
1688         }
1689
1690         return fd;
1691 }
1692
1693 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1694                 struct perf_thread_map *threads,
1695                 int start_cpu, int end_cpu)
1696 {
1697         int cpu, thread, nthreads;
1698         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1699         int pid = -1, err, old_errno;
1700         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1701
1702         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1703             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1704                 return -EINVAL;
1705
1706         if (cpus == NULL) {
1707                 static struct perf_cpu_map *empty_cpu_map;
1708
1709                 if (empty_cpu_map == NULL) {
1710                         empty_cpu_map = perf_cpu_map__dummy_new();
1711                         if (empty_cpu_map == NULL)
1712                                 return -ENOMEM;
1713                 }
1714
1715                 cpus = empty_cpu_map;
1716         }
1717
1718         if (threads == NULL) {
1719                 static struct perf_thread_map *empty_thread_map;
1720
1721                 if (empty_thread_map == NULL) {
1722                         empty_thread_map = thread_map__new_by_tid(-1);
1723                         if (empty_thread_map == NULL)
1724                                 return -ENOMEM;
1725                 }
1726
1727                 threads = empty_thread_map;
1728         }
1729
1730         if (evsel->core.system_wide)
1731                 nthreads = 1;
1732         else
1733                 nthreads = threads->nr;
1734
1735         if (evsel->core.fd == NULL &&
1736             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1737                 return -ENOMEM;
1738
1739         if (evsel->cgrp) {
1740                 flags |= PERF_FLAG_PID_CGROUP;
1741                 pid = evsel->cgrp->fd;
1742         }
1743
1744 fallback_missing_features:
1745         if (perf_missing_features.clockid_wrong)
1746                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1747         if (perf_missing_features.clockid) {
1748                 evsel->core.attr.use_clockid = 0;
1749                 evsel->core.attr.clockid = 0;
1750         }
1751         if (perf_missing_features.cloexec)
1752                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1753         if (perf_missing_features.mmap2)
1754                 evsel->core.attr.mmap2 = 0;
1755         if (perf_missing_features.exclude_guest)
1756                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1757         if (perf_missing_features.lbr_flags)
1758                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1759                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1760         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1761                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1762         if (perf_missing_features.ksymbol)
1763                 evsel->core.attr.ksymbol = 0;
1764         if (perf_missing_features.bpf)
1765                 evsel->core.attr.bpf_event = 0;
1766         if (perf_missing_features.branch_hw_idx)
1767                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1768 retry_sample_id:
1769         if (perf_missing_features.sample_id_all)
1770                 evsel->core.attr.sample_id_all = 0;
1771
1772         display_attr(&evsel->core.attr);
1773
1774         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1775
1776                 for (thread = 0; thread < nthreads; thread++) {
1777                         int fd, group_fd;
1778
1779                         if (!evsel->cgrp && !evsel->core.system_wide)
1780                                 pid = perf_thread_map__pid(threads, thread);
1781
1782                         group_fd = get_group_fd(evsel, cpu, thread);
1783 retry_open:
1784                         test_attr__ready();
1785
1786                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1787                                              group_fd, flags);
1788
1789                         FD(evsel, cpu, thread) = fd;
1790
1791                         if (unlikely(test_attr__enabled)) {
1792                                 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1793                                                 fd, group_fd, flags);
1794                         }
1795
1796                         if (fd < 0) {
1797                                 err = -errno;
1798
1799                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1800                                         /*
1801                                          * We just removed 1 thread, so take a step
1802                                          * back on thread index and lower the upper
1803                                          * nthreads limit.
1804                                          */
1805                                         nthreads--;
1806                                         thread--;
1807
1808                                         /* ... and pretend like nothing have happened. */
1809                                         err = 0;
1810                                         continue;
1811                                 }
1812
1813                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1814                                           err);
1815                                 goto try_fallback;
1816                         }
1817
1818                         pr_debug2_peo(" = %d\n", fd);
1819
1820                         if (evsel->bpf_fd >= 0) {
1821                                 int evt_fd = fd;
1822                                 int bpf_fd = evsel->bpf_fd;
1823
1824                                 err = ioctl(evt_fd,
1825                                             PERF_EVENT_IOC_SET_BPF,
1826                                             bpf_fd);
1827                                 if (err && errno != EEXIST) {
1828                                         pr_err("failed to attach bpf fd %d: %s\n",
1829                                                bpf_fd, strerror(errno));
1830                                         err = -EINVAL;
1831                                         goto out_close;
1832                                 }
1833                         }
1834
1835                         set_rlimit = NO_CHANGE;
1836
1837                         /*
1838                          * If we succeeded but had to kill clockid, fail and
1839                          * have evsel__open_strerror() print us a nice error.
1840                          */
1841                         if (perf_missing_features.clockid ||
1842                             perf_missing_features.clockid_wrong) {
1843                                 err = -EINVAL;
1844                                 goto out_close;
1845                         }
1846                 }
1847         }
1848
1849         return 0;
1850
1851 try_fallback:
1852         /*
1853          * perf stat needs between 5 and 22 fds per CPU. When we run out
1854          * of them try to increase the limits.
1855          */
1856         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1857                 struct rlimit l;
1858
1859                 old_errno = errno;
1860                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1861                         if (set_rlimit == NO_CHANGE)
1862                                 l.rlim_cur = l.rlim_max;
1863                         else {
1864                                 l.rlim_cur = l.rlim_max + 1000;
1865                                 l.rlim_max = l.rlim_cur;
1866                         }
1867                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1868                                 set_rlimit++;
1869                                 errno = old_errno;
1870                                 goto retry_open;
1871                         }
1872                 }
1873                 errno = old_errno;
1874         }
1875
1876         if (err != -EINVAL || cpu > 0 || thread > 0)
1877                 goto out_close;
1878
1879         /*
1880          * Must probe features in the order they were added to the
1881          * perf_event_attr interface.
1882          */
1883         if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1884                 perf_missing_features.cgroup = true;
1885                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1886                 goto out_close;
1887         } else if (!perf_missing_features.branch_hw_idx &&
1888             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1889                 perf_missing_features.branch_hw_idx = true;
1890                 pr_debug2("switching off branch HW index support\n");
1891                 goto fallback_missing_features;
1892         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1893                 perf_missing_features.aux_output = true;
1894                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1895                 goto out_close;
1896         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1897                 perf_missing_features.bpf = true;
1898                 pr_debug2_peo("switching off bpf_event\n");
1899                 goto fallback_missing_features;
1900         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1901                 perf_missing_features.ksymbol = true;
1902                 pr_debug2_peo("switching off ksymbol\n");
1903                 goto fallback_missing_features;
1904         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1905                 perf_missing_features.write_backward = true;
1906                 pr_debug2_peo("switching off write_backward\n");
1907                 goto out_close;
1908         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1909                 perf_missing_features.clockid_wrong = true;
1910                 pr_debug2_peo("switching off clockid\n");
1911                 goto fallback_missing_features;
1912         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1913                 perf_missing_features.clockid = true;
1914                 pr_debug2_peo("switching off use_clockid\n");
1915                 goto fallback_missing_features;
1916         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1917                 perf_missing_features.cloexec = true;
1918                 pr_debug2_peo("switching off cloexec flag\n");
1919                 goto fallback_missing_features;
1920         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1921                 perf_missing_features.mmap2 = true;
1922                 pr_debug2_peo("switching off mmap2\n");
1923                 goto fallback_missing_features;
1924         } else if (!perf_missing_features.exclude_guest &&
1925                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1926                 perf_missing_features.exclude_guest = true;
1927                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1928                 goto fallback_missing_features;
1929         } else if (!perf_missing_features.sample_id_all) {
1930                 perf_missing_features.sample_id_all = true;
1931                 pr_debug2_peo("switching off sample_id_all\n");
1932                 goto retry_sample_id;
1933         } else if (!perf_missing_features.lbr_flags &&
1934                         (evsel->core.attr.branch_sample_type &
1935                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1936                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1937                 perf_missing_features.lbr_flags = true;
1938                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1939                 goto fallback_missing_features;
1940         } else if (!perf_missing_features.group_read &&
1941                     evsel->core.attr.inherit &&
1942                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1943                    evsel__is_group_leader(evsel)) {
1944                 perf_missing_features.group_read = true;
1945                 pr_debug2_peo("switching off group read\n");
1946                 goto fallback_missing_features;
1947         }
1948 out_close:
1949         if (err)
1950                 threads->err_thread = thread;
1951
1952         old_errno = errno;
1953         do {
1954                 while (--thread >= 0) {
1955                         if (FD(evsel, cpu, thread) >= 0)
1956                                 close(FD(evsel, cpu, thread));
1957                         FD(evsel, cpu, thread) = -1;
1958                 }
1959                 thread = nthreads;
1960         } while (--cpu >= 0);
1961         errno = old_errno;
1962         return err;
1963 }
1964
1965 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
1966                 struct perf_thread_map *threads)
1967 {
1968         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
1969 }
1970
1971 void evsel__close(struct evsel *evsel)
1972 {
1973         perf_evsel__close(&evsel->core);
1974         perf_evsel__free_id(&evsel->core);
1975 }
1976
1977 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
1978 {
1979         if (cpu == -1)
1980                 return evsel__open_cpu(evsel, cpus, NULL, 0,
1981                                         cpus ? cpus->nr : 1);
1982
1983         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
1984 }
1985
1986 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
1987 {
1988         return evsel__open(evsel, NULL, threads);
1989 }
1990
1991 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
1992                                        const union perf_event *event,
1993                                        struct perf_sample *sample)
1994 {
1995         u64 type = evsel->core.attr.sample_type;
1996         const __u64 *array = event->sample.array;
1997         bool swapped = evsel->needs_swap;
1998         union u64_swap u;
1999
2000         array += ((event->header.size -
2001                    sizeof(event->header)) / sizeof(u64)) - 1;
2002
2003         if (type & PERF_SAMPLE_IDENTIFIER) {
2004                 sample->id = *array;
2005                 array--;
2006         }
2007
2008         if (type & PERF_SAMPLE_CPU) {
2009                 u.val64 = *array;
2010                 if (swapped) {
2011                         /* undo swap of u64, then swap on individual u32s */
2012                         u.val64 = bswap_64(u.val64);
2013                         u.val32[0] = bswap_32(u.val32[0]);
2014                 }
2015
2016                 sample->cpu = u.val32[0];
2017                 array--;
2018         }
2019
2020         if (type & PERF_SAMPLE_STREAM_ID) {
2021                 sample->stream_id = *array;
2022                 array--;
2023         }
2024
2025         if (type & PERF_SAMPLE_ID) {
2026                 sample->id = *array;
2027                 array--;
2028         }
2029
2030         if (type & PERF_SAMPLE_TIME) {
2031                 sample->time = *array;
2032                 array--;
2033         }
2034
2035         if (type & PERF_SAMPLE_TID) {
2036                 u.val64 = *array;
2037                 if (swapped) {
2038                         /* undo swap of u64, then swap on individual u32s */
2039                         u.val64 = bswap_64(u.val64);
2040                         u.val32[0] = bswap_32(u.val32[0]);
2041                         u.val32[1] = bswap_32(u.val32[1]);
2042                 }
2043
2044                 sample->pid = u.val32[0];
2045                 sample->tid = u.val32[1];
2046                 array--;
2047         }
2048
2049         return 0;
2050 }
2051
2052 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2053                             u64 size)
2054 {
2055         return size > max_size || offset + size > endp;
2056 }
2057
2058 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2059         do {                                                            \
2060                 if (overflow(endp, (max_size), (offset), (size)))       \
2061                         return -EFAULT;                                 \
2062         } while (0)
2063
2064 #define OVERFLOW_CHECK_u64(offset) \
2065         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2066
2067 static int
2068 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2069 {
2070         /*
2071          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2072          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2073          * check the format does not go past the end of the event.
2074          */
2075         if (sample_size + sizeof(event->header) > event->header.size)
2076                 return -EFAULT;
2077
2078         return 0;
2079 }
2080
2081 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2082                         struct perf_sample *data)
2083 {
2084         u64 type = evsel->core.attr.sample_type;
2085         bool swapped = evsel->needs_swap;
2086         const __u64 *array;
2087         u16 max_size = event->header.size;
2088         const void *endp = (void *)event + max_size;
2089         u64 sz;
2090
2091         /*
2092          * used for cross-endian analysis. See git commit 65014ab3
2093          * for why this goofiness is needed.
2094          */
2095         union u64_swap u;
2096
2097         memset(data, 0, sizeof(*data));
2098         data->cpu = data->pid = data->tid = -1;
2099         data->stream_id = data->id = data->time = -1ULL;
2100         data->period = evsel->core.attr.sample_period;
2101         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2102         data->misc    = event->header.misc;
2103         data->id = -1ULL;
2104         data->data_src = PERF_MEM_DATA_SRC_NONE;
2105
2106         if (event->header.type != PERF_RECORD_SAMPLE) {
2107                 if (!evsel->core.attr.sample_id_all)
2108                         return 0;
2109                 return perf_evsel__parse_id_sample(evsel, event, data);
2110         }
2111
2112         array = event->sample.array;
2113
2114         if (perf_event__check_size(event, evsel->sample_size))
2115                 return -EFAULT;
2116
2117         if (type & PERF_SAMPLE_IDENTIFIER) {
2118                 data->id = *array;
2119                 array++;
2120         }
2121
2122         if (type & PERF_SAMPLE_IP) {
2123                 data->ip = *array;
2124                 array++;
2125         }
2126
2127         if (type & PERF_SAMPLE_TID) {
2128                 u.val64 = *array;
2129                 if (swapped) {
2130                         /* undo swap of u64, then swap on individual u32s */
2131                         u.val64 = bswap_64(u.val64);
2132                         u.val32[0] = bswap_32(u.val32[0]);
2133                         u.val32[1] = bswap_32(u.val32[1]);
2134                 }
2135
2136                 data->pid = u.val32[0];
2137                 data->tid = u.val32[1];
2138                 array++;
2139         }
2140
2141         if (type & PERF_SAMPLE_TIME) {
2142                 data->time = *array;
2143                 array++;
2144         }
2145
2146         if (type & PERF_SAMPLE_ADDR) {
2147                 data->addr = *array;
2148                 array++;
2149         }
2150
2151         if (type & PERF_SAMPLE_ID) {
2152                 data->id = *array;
2153                 array++;
2154         }
2155
2156         if (type & PERF_SAMPLE_STREAM_ID) {
2157                 data->stream_id = *array;
2158                 array++;
2159         }
2160
2161         if (type & PERF_SAMPLE_CPU) {
2162
2163                 u.val64 = *array;
2164                 if (swapped) {
2165                         /* undo swap of u64, then swap on individual u32s */
2166                         u.val64 = bswap_64(u.val64);
2167                         u.val32[0] = bswap_32(u.val32[0]);
2168                 }
2169
2170                 data->cpu = u.val32[0];
2171                 array++;
2172         }
2173
2174         if (type & PERF_SAMPLE_PERIOD) {
2175                 data->period = *array;
2176                 array++;
2177         }
2178
2179         if (type & PERF_SAMPLE_READ) {
2180                 u64 read_format = evsel->core.attr.read_format;
2181
2182                 OVERFLOW_CHECK_u64(array);
2183                 if (read_format & PERF_FORMAT_GROUP)
2184                         data->read.group.nr = *array;
2185                 else
2186                         data->read.one.value = *array;
2187
2188                 array++;
2189
2190                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2191                         OVERFLOW_CHECK_u64(array);
2192                         data->read.time_enabled = *array;
2193                         array++;
2194                 }
2195
2196                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2197                         OVERFLOW_CHECK_u64(array);
2198                         data->read.time_running = *array;
2199                         array++;
2200                 }
2201
2202                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2203                 if (read_format & PERF_FORMAT_GROUP) {
2204                         const u64 max_group_nr = UINT64_MAX /
2205                                         sizeof(struct sample_read_value);
2206
2207                         if (data->read.group.nr > max_group_nr)
2208                                 return -EFAULT;
2209                         sz = data->read.group.nr *
2210                              sizeof(struct sample_read_value);
2211                         OVERFLOW_CHECK(array, sz, max_size);
2212                         data->read.group.values =
2213                                         (struct sample_read_value *)array;
2214                         array = (void *)array + sz;
2215                 } else {
2216                         OVERFLOW_CHECK_u64(array);
2217                         data->read.one.id = *array;
2218                         array++;
2219                 }
2220         }
2221
2222         if (type & PERF_SAMPLE_CALLCHAIN) {
2223                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2224
2225                 OVERFLOW_CHECK_u64(array);
2226                 data->callchain = (struct ip_callchain *)array++;
2227                 if (data->callchain->nr > max_callchain_nr)
2228                         return -EFAULT;
2229                 sz = data->callchain->nr * sizeof(u64);
2230                 OVERFLOW_CHECK(array, sz, max_size);
2231                 array = (void *)array + sz;
2232         }
2233
2234         if (type & PERF_SAMPLE_RAW) {
2235                 OVERFLOW_CHECK_u64(array);
2236                 u.val64 = *array;
2237
2238                 /*
2239                  * Undo swap of u64, then swap on individual u32s,
2240                  * get the size of the raw area and undo all of the
2241                  * swap. The pevent interface handles endianity by
2242                  * itself.
2243                  */
2244                 if (swapped) {
2245                         u.val64 = bswap_64(u.val64);
2246                         u.val32[0] = bswap_32(u.val32[0]);
2247                         u.val32[1] = bswap_32(u.val32[1]);
2248                 }
2249                 data->raw_size = u.val32[0];
2250
2251                 /*
2252                  * The raw data is aligned on 64bits including the
2253                  * u32 size, so it's safe to use mem_bswap_64.
2254                  */
2255                 if (swapped)
2256                         mem_bswap_64((void *) array, data->raw_size);
2257
2258                 array = (void *)array + sizeof(u32);
2259
2260                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2261                 data->raw_data = (void *)array;
2262                 array = (void *)array + data->raw_size;
2263         }
2264
2265         if (type & PERF_SAMPLE_BRANCH_STACK) {
2266                 const u64 max_branch_nr = UINT64_MAX /
2267                                           sizeof(struct branch_entry);
2268
2269                 OVERFLOW_CHECK_u64(array);
2270                 data->branch_stack = (struct branch_stack *)array++;
2271
2272                 if (data->branch_stack->nr > max_branch_nr)
2273                         return -EFAULT;
2274
2275                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2276                 if (evsel__has_branch_hw_idx(evsel))
2277                         sz += sizeof(u64);
2278                 else
2279                         data->no_hw_idx = true;
2280                 OVERFLOW_CHECK(array, sz, max_size);
2281                 array = (void *)array + sz;
2282         }
2283
2284         if (type & PERF_SAMPLE_REGS_USER) {
2285                 OVERFLOW_CHECK_u64(array);
2286                 data->user_regs.abi = *array;
2287                 array++;
2288
2289                 if (data->user_regs.abi) {
2290                         u64 mask = evsel->core.attr.sample_regs_user;
2291
2292                         sz = hweight64(mask) * sizeof(u64);
2293                         OVERFLOW_CHECK(array, sz, max_size);
2294                         data->user_regs.mask = mask;
2295                         data->user_regs.regs = (u64 *)array;
2296                         array = (void *)array + sz;
2297                 }
2298         }
2299
2300         if (type & PERF_SAMPLE_STACK_USER) {
2301                 OVERFLOW_CHECK_u64(array);
2302                 sz = *array++;
2303
2304                 data->user_stack.offset = ((char *)(array - 1)
2305                                           - (char *) event);
2306
2307                 if (!sz) {
2308                         data->user_stack.size = 0;
2309                 } else {
2310                         OVERFLOW_CHECK(array, sz, max_size);
2311                         data->user_stack.data = (char *)array;
2312                         array = (void *)array + sz;
2313                         OVERFLOW_CHECK_u64(array);
2314                         data->user_stack.size = *array++;
2315                         if (WARN_ONCE(data->user_stack.size > sz,
2316                                       "user stack dump failure\n"))
2317                                 return -EFAULT;
2318                 }
2319         }
2320
2321         if (type & PERF_SAMPLE_WEIGHT) {
2322                 OVERFLOW_CHECK_u64(array);
2323                 data->weight = *array;
2324                 array++;
2325         }
2326
2327         if (type & PERF_SAMPLE_DATA_SRC) {
2328                 OVERFLOW_CHECK_u64(array);
2329                 data->data_src = *array;
2330                 array++;
2331         }
2332
2333         if (type & PERF_SAMPLE_TRANSACTION) {
2334                 OVERFLOW_CHECK_u64(array);
2335                 data->transaction = *array;
2336                 array++;
2337         }
2338
2339         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2340         if (type & PERF_SAMPLE_REGS_INTR) {
2341                 OVERFLOW_CHECK_u64(array);
2342                 data->intr_regs.abi = *array;
2343                 array++;
2344
2345                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2346                         u64 mask = evsel->core.attr.sample_regs_intr;
2347
2348                         sz = hweight64(mask) * sizeof(u64);
2349                         OVERFLOW_CHECK(array, sz, max_size);
2350                         data->intr_regs.mask = mask;
2351                         data->intr_regs.regs = (u64 *)array;
2352                         array = (void *)array + sz;
2353                 }
2354         }
2355
2356         data->phys_addr = 0;
2357         if (type & PERF_SAMPLE_PHYS_ADDR) {
2358                 data->phys_addr = *array;
2359                 array++;
2360         }
2361
2362         data->cgroup = 0;
2363         if (type & PERF_SAMPLE_CGROUP) {
2364                 data->cgroup = *array;
2365                 array++;
2366         }
2367
2368         if (type & PERF_SAMPLE_AUX) {
2369                 OVERFLOW_CHECK_u64(array);
2370                 sz = *array++;
2371
2372                 OVERFLOW_CHECK(array, sz, max_size);
2373                 /* Undo swap of data */
2374                 if (swapped)
2375                         mem_bswap_64((char *)array, sz);
2376                 data->aux_sample.size = sz;
2377                 data->aux_sample.data = (char *)array;
2378                 array = (void *)array + sz;
2379         }
2380
2381         return 0;
2382 }
2383
2384 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2385                                   u64 *timestamp)
2386 {
2387         u64 type = evsel->core.attr.sample_type;
2388         const __u64 *array;
2389
2390         if (!(type & PERF_SAMPLE_TIME))
2391                 return -1;
2392
2393         if (event->header.type != PERF_RECORD_SAMPLE) {
2394                 struct perf_sample data = {
2395                         .time = -1ULL,
2396                 };
2397
2398                 if (!evsel->core.attr.sample_id_all)
2399                         return -1;
2400                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2401                         return -1;
2402
2403                 *timestamp = data.time;
2404                 return 0;
2405         }
2406
2407         array = event->sample.array;
2408
2409         if (perf_event__check_size(event, evsel->sample_size))
2410                 return -EFAULT;
2411
2412         if (type & PERF_SAMPLE_IDENTIFIER)
2413                 array++;
2414
2415         if (type & PERF_SAMPLE_IP)
2416                 array++;
2417
2418         if (type & PERF_SAMPLE_TID)
2419                 array++;
2420
2421         if (type & PERF_SAMPLE_TIME)
2422                 *timestamp = *array;
2423
2424         return 0;
2425 }
2426
2427 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2428 {
2429         return tep_find_field(evsel->tp_format, name);
2430 }
2431
2432 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2433 {
2434         struct tep_format_field *field = evsel__field(evsel, name);
2435         int offset;
2436
2437         if (!field)
2438                 return NULL;
2439
2440         offset = field->offset;
2441
2442         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2443                 offset = *(int *)(sample->raw_data + field->offset);
2444                 offset &= 0xffff;
2445         }
2446
2447         return sample->raw_data + offset;
2448 }
2449
2450 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2451                          bool needs_swap)
2452 {
2453         u64 value;
2454         void *ptr = sample->raw_data + field->offset;
2455
2456         switch (field->size) {
2457         case 1:
2458                 return *(u8 *)ptr;
2459         case 2:
2460                 value = *(u16 *)ptr;
2461                 break;
2462         case 4:
2463                 value = *(u32 *)ptr;
2464                 break;
2465         case 8:
2466                 memcpy(&value, ptr, sizeof(u64));
2467                 break;
2468         default:
2469                 return 0;
2470         }
2471
2472         if (!needs_swap)
2473                 return value;
2474
2475         switch (field->size) {
2476         case 2:
2477                 return bswap_16(value);
2478         case 4:
2479                 return bswap_32(value);
2480         case 8:
2481                 return bswap_64(value);
2482         default:
2483                 return 0;
2484         }
2485
2486         return 0;
2487 }
2488
2489 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2490 {
2491         struct tep_format_field *field = evsel__field(evsel, name);
2492
2493         if (!field)
2494                 return 0;
2495
2496         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2497 }
2498
2499 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2500 {
2501         int paranoid;
2502
2503         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2504             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2505             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2506                 /*
2507                  * If it's cycles then fall back to hrtimer based
2508                  * cpu-clock-tick sw counter, which is always available even if
2509                  * no PMU support.
2510                  *
2511                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2512                  * b0a873e).
2513                  */
2514                 scnprintf(msg, msgsize, "%s",
2515 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2516
2517                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2518                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2519
2520                 zfree(&evsel->name);
2521                 return true;
2522         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2523                    (paranoid = perf_event_paranoid()) > 1) {
2524                 const char *name = evsel__name(evsel);
2525                 char *new_name;
2526                 const char *sep = ":";
2527
2528                 /* If event has exclude user then don't exclude kernel. */
2529                 if (evsel->core.attr.exclude_user)
2530                         return false;
2531
2532                 /* Is there already the separator in the name. */
2533                 if (strchr(name, '/') ||
2534                     (strchr(name, ':') && !evsel->is_libpfm_event))
2535                         sep = "";
2536
2537                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2538                         return false;
2539
2540                 if (evsel->name)
2541                         free(evsel->name);
2542                 evsel->name = new_name;
2543                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2544                           "to fall back to excluding kernel and hypervisor "
2545                           " samples", paranoid);
2546                 evsel->core.attr.exclude_kernel = 1;
2547                 evsel->core.attr.exclude_hv     = 1;
2548
2549                 return true;
2550         }
2551
2552         return false;
2553 }
2554
2555 static bool find_process(const char *name)
2556 {
2557         size_t len = strlen(name);
2558         DIR *dir;
2559         struct dirent *d;
2560         int ret = -1;
2561
2562         dir = opendir(procfs__mountpoint());
2563         if (!dir)
2564                 return false;
2565
2566         /* Walk through the directory. */
2567         while (ret && (d = readdir(dir)) != NULL) {
2568                 char path[PATH_MAX];
2569                 char *data;
2570                 size_t size;
2571
2572                 if ((d->d_type != DT_DIR) ||
2573                      !strcmp(".", d->d_name) ||
2574                      !strcmp("..", d->d_name))
2575                         continue;
2576
2577                 scnprintf(path, sizeof(path), "%s/%s/comm",
2578                           procfs__mountpoint(), d->d_name);
2579
2580                 if (filename__read_str(path, &data, &size))
2581                         continue;
2582
2583                 ret = strncmp(name, data, len);
2584                 free(data);
2585         }
2586
2587         closedir(dir);
2588         return ret ? false : true;
2589 }
2590
2591 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2592                          int err, char *msg, size_t size)
2593 {
2594         char sbuf[STRERR_BUFSIZE];
2595         int printed = 0, enforced = 0;
2596
2597         switch (err) {
2598         case EPERM:
2599         case EACCES:
2600                 printed += scnprintf(msg + printed, size - printed,
2601                         "Access to performance monitoring and observability operations is limited.\n");
2602
2603                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2604                         if (enforced) {
2605                                 printed += scnprintf(msg + printed, size - printed,
2606                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2607                                         "monitoring and observability operations. Inspect system audit records for\n"
2608                                         "more perf_event access control information and adjusting the policy.\n");
2609                         }
2610                 }
2611
2612                 if (err == EPERM)
2613                         printed += scnprintf(msg, size,
2614                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2615
2616                 return scnprintf(msg + printed, size - printed,
2617                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2618                  "access to performance monitoring and observability operations for processes\n"
2619                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2620                  "More information can be found at 'Perf events and tool security' document:\n"
2621                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2622                  "perf_event_paranoid setting is %d:\n"
2623                  "  -1: Allow use of (almost) all events by all users\n"
2624                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2625                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2626                  ">= 1: Disallow CPU event access\n"
2627                  ">= 2: Disallow kernel profiling\n"
2628                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2629                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2630                  perf_event_paranoid());
2631         case ENOENT:
2632                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2633         case EMFILE:
2634                 return scnprintf(msg, size, "%s",
2635                          "Too many events are opened.\n"
2636                          "Probably the maximum number of open file descriptors has been reached.\n"
2637                          "Hint: Try again after reducing the number of events.\n"
2638                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2639         case ENOMEM:
2640                 if (evsel__has_callchain(evsel) &&
2641                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2642                         return scnprintf(msg, size,
2643                                          "Not enough memory to setup event with callchain.\n"
2644                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2645                                          "Hint: Current value: %d", sysctl__max_stack());
2646                 break;
2647         case ENODEV:
2648                 if (target->cpu_list)
2649                         return scnprintf(msg, size, "%s",
2650          "No such device - did you specify an out-of-range profile CPU?");
2651                 break;
2652         case EOPNOTSUPP:
2653                 if (evsel->core.attr.aux_output)
2654                         return scnprintf(msg, size,
2655         "%s: PMU Hardware doesn't support 'aux_output' feature",
2656                                          evsel__name(evsel));
2657                 if (evsel->core.attr.sample_period != 0)
2658                         return scnprintf(msg, size,
2659         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2660                                          evsel__name(evsel));
2661                 if (evsel->core.attr.precise_ip)
2662                         return scnprintf(msg, size, "%s",
2663         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2664 #if defined(__i386__) || defined(__x86_64__)
2665                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2666                         return scnprintf(msg, size, "%s",
2667         "No hardware sampling interrupt available.\n");
2668 #endif
2669                 break;
2670         case EBUSY:
2671                 if (find_process("oprofiled"))
2672                         return scnprintf(msg, size,
2673         "The PMU counters are busy/taken by another profiler.\n"
2674         "We found oprofile daemon running, please stop it and try again.");
2675                 break;
2676         case EINVAL:
2677                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2678                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2679                 if (perf_missing_features.clockid)
2680                         return scnprintf(msg, size, "clockid feature not supported.");
2681                 if (perf_missing_features.clockid_wrong)
2682                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2683                 if (perf_missing_features.aux_output)
2684                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2685                 break;
2686         default:
2687                 break;
2688         }
2689
2690         return scnprintf(msg, size,
2691         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2692         "/bin/dmesg | grep -i perf may provide additional information.\n",
2693                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2694 }
2695
2696 struct perf_env *evsel__env(struct evsel *evsel)
2697 {
2698         if (evsel && evsel->evlist)
2699                 return evsel->evlist->env;
2700         return &perf_env;
2701 }
2702
2703 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2704 {
2705         int cpu, thread;
2706
2707         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2708                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2709                      thread++) {
2710                         int fd = FD(evsel, cpu, thread);
2711
2712                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2713                                                    cpu, thread, fd) < 0)
2714                                 return -1;
2715                 }
2716         }
2717
2718         return 0;
2719 }
2720
2721 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2722 {
2723         struct perf_cpu_map *cpus = evsel->core.cpus;
2724         struct perf_thread_map *threads = evsel->core.threads;
2725
2726         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2727                 return -ENOMEM;
2728
2729         return store_evsel_ids(evsel, evlist);
2730 }