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 "bpf_counter.h"
  29 #include "callchain.h"
  30 #include "cgroup.h"
  31 #include "counts.h"
  32 #include "event.h"
  33 #include "evsel.h"
  34 #include "util/env.h"
  35 #include "util/evsel_config.h"
  36 #include "util/evsel_fprintf.h"
  37 #include "evlist.h"
  38 #include <perf/cpumap.h>
  39 #include "thread_map.h"
  40 #include "target.h"
  41 #include "perf_regs.h"
  42 #include "record.h"
  43 #include "debug.h"
  44 #include "trace-event.h"
  45 #include "stat.h"
  46 #include "string2.h"
  47 #include "memswap.h"
  48 #include "util.h"
  49 #include "hashmap.h"
  50 #include "pmu-hybrid.h"
  51 #include "../perf-sys.h"
  52 #include "util/parse-branch-options.h"
  53 #include <internal/xyarray.h>
  54 #include <internal/lib.h>
  55
  56 #include <linux/ctype.h>
  57
  58 struct perf_missing_features perf_missing_features;
  59
  60 static clockid_t clockid;
  61
  62 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  63 {
  64         return 0;
  65 }
  66
  67 void __weak test_attr__ready(void) { }
  68
  69 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  70 {
  71 }
  72
  73 static struct {
  74         size_t  size;
  75         int     (*init)(struct evsel *evsel);
  76         void    (*fini)(struct evsel *evsel);
  77 } perf_evsel__object = {
  78         .size = sizeof(struct evsel),
  79         .init = evsel__no_extra_init,
  80         .fini = evsel__no_extra_fini,
  81 };
  82
  83 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
  84                          void (*fini)(struct evsel *evsel))
  85 {
  86
  87         if (object_size == 0)
  88                 goto set_methods;
  89
  90         if (perf_evsel__object.size > object_size)
  91                 return -EINVAL;
  92
  93         perf_evsel__object.size = object_size;
  94
  95 set_methods:
  96         if (init != NULL)
  97                 perf_evsel__object.init = init;
  98
  99         if (fini != NULL)
 100                 perf_evsel__object.fini = fini;
 101
 102         return 0;
 103 }
 104
 105 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 106
 107 int __evsel__sample_size(u64 sample_type)
 108 {
 109         u64 mask = sample_type & PERF_SAMPLE_MASK;
 110         int size = 0;
 111         int i;
 112
 113         for (i = 0; i < 64; i++) {
 114                 if (mask & (1ULL << i))
 115                         size++;
 116         }
 117
 118         size *= sizeof(u64);
 119
 120         return size;
 121 }
 122
 123 /**
 124  * __perf_evsel__calc_id_pos - calculate id_pos.
 125  * @sample_type: sample type
 126  *
 127  * This function returns the position of the event id (PERF_SAMPLE_ID or
 128  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 129  * perf_record_sample.
 130  */
 131 static int __perf_evsel__calc_id_pos(u64 sample_type)
 132 {
 133         int idx = 0;
 134
 135         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 136                 return 0;
 137
 138         if (!(sample_type & PERF_SAMPLE_ID))
 139                 return -1;
 140
 141         if (sample_type & PERF_SAMPLE_IP)
 142                 idx += 1;
 143
 144         if (sample_type & PERF_SAMPLE_TID)
 145                 idx += 1;
 146
 147         if (sample_type & PERF_SAMPLE_TIME)
 148                 idx += 1;
 149
 150         if (sample_type & PERF_SAMPLE_ADDR)
 151                 idx += 1;
 152
 153         return idx;
 154 }
 155
 156 /**
 157  * __perf_evsel__calc_is_pos - calculate is_pos.
 158  * @sample_type: sample type
 159  *
 160  * This function returns the position (counting backwards) of the event id
 161  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 162  * sample_id_all is used there is an id sample appended to non-sample events.
 163  */
 164 static int __perf_evsel__calc_is_pos(u64 sample_type)
 165 {
 166         int idx = 1;
 167
 168         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 169                 return 1;
 170
 171         if (!(sample_type & PERF_SAMPLE_ID))
 172                 return -1;
 173
 174         if (sample_type & PERF_SAMPLE_CPU)
 175                 idx += 1;
 176
 177         if (sample_type & PERF_SAMPLE_STREAM_ID)
 178                 idx += 1;
 179
 180         return idx;
 181 }
 182
 183 void evsel__calc_id_pos(struct evsel *evsel)
 184 {
 185         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 186         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 187 }
 188
 189 void __evsel__set_sample_bit(struct evsel *evsel,
 190                                   enum perf_event_sample_format bit)
 191 {
 192         if (!(evsel->core.attr.sample_type & bit)) {
 193                 evsel->core.attr.sample_type |= bit;
 194                 evsel->sample_size += sizeof(u64);
 195                 evsel__calc_id_pos(evsel);
 196         }
 197 }
 198
 199 void __evsel__reset_sample_bit(struct evsel *evsel,
 200                                     enum perf_event_sample_format bit)
 201 {
 202         if (evsel->core.attr.sample_type & bit) {
 203                 evsel->core.attr.sample_type &= ~bit;
 204                 evsel->sample_size -= sizeof(u64);
 205                 evsel__calc_id_pos(evsel);
 206         }
 207 }
 208
 209 void evsel__set_sample_id(struct evsel *evsel,
 210                                bool can_sample_identifier)
 211 {
 212         if (can_sample_identifier) {
 213                 evsel__reset_sample_bit(evsel, ID);
 214                 evsel__set_sample_bit(evsel, IDENTIFIER);
 215         } else {
 216                 evsel__set_sample_bit(evsel, ID);
 217         }
 218         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 219 }
 220
 221 /**
 222  * evsel__is_function_event - Return whether given evsel is a function
 223  * trace event
 224  *
 225  * @evsel - evsel selector to be tested
 226  *
 227  * Return %true if event is function trace event
 228  */
 229 bool evsel__is_function_event(struct evsel *evsel)
 230 {
 231 #define FUNCTION_EVENT "ftrace:function"
 232
 233         return evsel->name &&
 234                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 235
 236 #undef FUNCTION_EVENT
 237 }
 238
 239 void evsel__init(struct evsel *evsel,
 240                  struct perf_event_attr *attr, int idx)
 241 {
 242         perf_evsel__init(&evsel->core, attr, idx);
 243         evsel->tracking    = !idx;
 244         evsel->leader      = evsel;
 245         evsel->unit        = "";
 246         evsel->scale       = 1.0;
 247         evsel->max_events  = ULONG_MAX;
 248         evsel->evlist      = NULL;
 249         evsel->bpf_obj     = NULL;
 250         evsel->bpf_fd      = -1;
 251         INIT_LIST_HEAD(&evsel->config_terms);
 252         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 253         perf_evsel__object.init(evsel);
 254         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 255         evsel__calc_id_pos(evsel);
 256         evsel->cmdline_group_boundary = false;
 257         evsel->metric_expr   = NULL;
 258         evsel->metric_name   = NULL;
 259         evsel->metric_events = NULL;
 260         evsel->per_pkg_mask  = NULL;
 261         evsel->collect_stat  = false;
 262         evsel->pmu_name      = NULL;
 263 }
 264
 265 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 266 {
 267         struct evsel *evsel = zalloc(perf_evsel__object.size);
 268
 269         if (!evsel)
 270                 return NULL;
 271         evsel__init(evsel, attr, idx);
 272
 273         if (evsel__is_bpf_output(evsel)) {
 274                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 275                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 276                 evsel->core.attr.sample_period = 1;
 277         }
 278
 279         if (evsel__is_clock(evsel)) {
 280                 /*
 281                  * The evsel->unit points to static alias->unit
 282                  * so it's ok to use static string in here.
 283                  */
 284                 static const char *unit = "msec";
 285
 286                 evsel->unit = unit;
 287                 evsel->scale = 1e-6;
 288         }
 289
 290         return evsel;
 291 }
 292
 293 static bool perf_event_can_profile_kernel(void)
 294 {
 295         return perf_event_paranoid_check(1);
 296 }
 297
 298 struct evsel *evsel__new_cycles(bool precise, __u32 type, __u64 config)
 299 {
 300         struct perf_event_attr attr = {
 301                 .type   = type,
 302                 .config = config,
 303                 .exclude_kernel = !perf_event_can_profile_kernel(),
 304         };
 305         struct evsel *evsel;
 306
 307         event_attr_init(&attr);
 308
 309         if (!precise)
 310                 goto new_event;
 311
 312         /*
 313          * Now let the usual logic to set up the perf_event_attr defaults
 314          * to kick in when we return and before perf_evsel__open() is called.
 315          */
 316 new_event:
 317         evsel = evsel__new(&attr);
 318         if (evsel == NULL)
 319                 goto out;
 320
 321         evsel->precise_max = true;
 322
 323         /* use asprintf() because free(evsel) assumes name is allocated */
 324         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 325                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 326                      attr.exclude_kernel ? "u" : "",
 327                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 328                 goto error_free;
 329 out:
 330         return evsel;
 331 error_free:
 332         evsel__delete(evsel);
 333         evsel = NULL;
 334         goto out;
 335 }
 336
 337 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 338 {
 339         struct evsel_config_term *pos, *tmp;
 340
 341         list_for_each_entry(pos, &src->config_terms, list) {
 342                 tmp = malloc(sizeof(*tmp));
 343                 if (tmp == NULL)
 344                         return -ENOMEM;
 345
 346                 *tmp = *pos;
 347                 if (tmp->free_str) {
 348                         tmp->val.str = strdup(pos->val.str);
 349                         if (tmp->val.str == NULL) {
 350                                 free(tmp);
 351                                 return -ENOMEM;
 352                         }
 353                 }
 354                 list_add_tail(&tmp->list, &dst->config_terms);
 355         }
 356         return 0;
 357 }
 358
 359 /**
 360  * evsel__clone - create a new evsel copied from @orig
 361  * @orig: original evsel
 362  *
 363  * The assumption is that @orig is not configured nor opened yet.
 364  * So we only care about the attributes that can be set while it's parsed.
 365  */
 366 struct evsel *evsel__clone(struct evsel *orig)
 367 {
 368         struct evsel *evsel;
 369
 370         BUG_ON(orig->core.fd);
 371         BUG_ON(orig->counts);
 372         BUG_ON(orig->priv);
 373         BUG_ON(orig->per_pkg_mask);
 374
 375         /* cannot handle BPF objects for now */
 376         if (orig->bpf_obj)
 377                 return NULL;
 378
 379         evsel = evsel__new(&orig->core.attr);
 380         if (evsel == NULL)
 381                 return NULL;
 382
 383         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 384         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 385         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 386         evsel->core.nr_members = orig->core.nr_members;
 387         evsel->core.system_wide = orig->core.system_wide;
 388
 389         if (orig->name) {
 390                 evsel->name = strdup(orig->name);
 391                 if (evsel->name == NULL)
 392                         goto out_err;
 393         }
 394         if (orig->group_name) {
 395                 evsel->group_name = strdup(orig->group_name);
 396                 if (evsel->group_name == NULL)
 397                         goto out_err;
 398         }
 399         if (orig->pmu_name) {
 400                 evsel->pmu_name = strdup(orig->pmu_name);
 401                 if (evsel->pmu_name == NULL)
 402                         goto out_err;
 403         }
 404         if (orig->filter) {
 405                 evsel->filter = strdup(orig->filter);
 406                 if (evsel->filter == NULL)
 407                         goto out_err;
 408         }
 409         evsel->cgrp = cgroup__get(orig->cgrp);
 410         evsel->tp_format = orig->tp_format;
 411         evsel->handler = orig->handler;
 412         evsel->leader = orig->leader;
 413
 414         evsel->max_events = orig->max_events;
 415         evsel->tool_event = orig->tool_event;
 416         evsel->unit = orig->unit;
 417         evsel->scale = orig->scale;
 418         evsel->snapshot = orig->snapshot;
 419         evsel->per_pkg = orig->per_pkg;
 420         evsel->percore = orig->percore;
 421         evsel->precise_max = orig->precise_max;
 422         evsel->use_uncore_alias = orig->use_uncore_alias;
 423         evsel->is_libpfm_event = orig->is_libpfm_event;
 424
 425         evsel->exclude_GH = orig->exclude_GH;
 426         evsel->sample_read = orig->sample_read;
 427         evsel->auto_merge_stats = orig->auto_merge_stats;
 428         evsel->collect_stat = orig->collect_stat;
 429         evsel->weak_group = orig->weak_group;
 430         evsel->use_config_name = orig->use_config_name;
 431
 432         if (evsel__copy_config_terms(evsel, orig) < 0)
 433                 goto out_err;
 434
 435         return evsel;
 436
 437 out_err:
 438         evsel__delete(evsel);
 439         return NULL;
 440 }
 441
 442 /*
 443  * Returns pointer with encoded error via <linux/err.h> interface.
 444  */
 445 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 446 {
 447         struct evsel *evsel = zalloc(perf_evsel__object.size);
 448         int err = -ENOMEM;
 449
 450         if (evsel == NULL) {
 451                 goto out_err;
 452         } else {
 453                 struct perf_event_attr attr = {
 454                         .type          = PERF_TYPE_TRACEPOINT,
 455                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 456                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 457                 };
 458
 459                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 460                         goto out_free;
 461
 462                 evsel->tp_format = trace_event__tp_format(sys, name);
 463                 if (IS_ERR(evsel->tp_format)) {
 464                         err = PTR_ERR(evsel->tp_format);
 465                         goto out_free;
 466                 }
 467
 468                 event_attr_init(&attr);
 469                 attr.config = evsel->tp_format->id;
 470                 attr.sample_period = 1;
 471                 evsel__init(evsel, &attr, idx);
 472         }
 473
 474         return evsel;
 475
 476 out_free:
 477         zfree(&evsel->name);
 478         free(evsel);
 479 out_err:
 480         return ERR_PTR(err);
 481 }
 482
 483 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
 484         "cycles",
 485         "instructions",
 486         "cache-references",
 487         "cache-misses",
 488         "branches",
 489         "branch-misses",
 490         "bus-cycles",
 491         "stalled-cycles-frontend",
 492         "stalled-cycles-backend",
 493         "ref-cycles",
 494 };
 495
 496 char *evsel__bpf_counter_events;
 497
 498 bool evsel__match_bpf_counter_events(const char *name)
 499 {
 500         int name_len;
 501         bool match;
 502         char *ptr;
 503
 504         if (!evsel__bpf_counter_events)
 505                 return false;
 506
 507         ptr = strstr(evsel__bpf_counter_events, name);
 508         name_len = strlen(name);
 509
 510         /* check name matches a full token in evsel__bpf_counter_events */
 511         match = (ptr != NULL) &&
 512                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 513                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 514
 515         return match;
 516 }
 517
 518 static const char *__evsel__hw_name(u64 config)
 519 {
 520         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 521                 return evsel__hw_names[config];
 522
 523         return "unknown-hardware";
 524 }
 525
 526 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 527 {
 528         int colon = 0, r = 0;
 529         struct perf_event_attr *attr = &evsel->core.attr;
 530         bool exclude_guest_default = false;
 531
 532 #define MOD_PRINT(context, mod) do {                                    \
 533                 if (!attr->exclude_##context) {                         \
 534                         if (!colon) colon = ++r;                        \
 535                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 536                 } } while(0)
 537
 538         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 539                 MOD_PRINT(kernel, 'k');
 540                 MOD_PRINT(user, 'u');
 541                 MOD_PRINT(hv, 'h');
 542                 exclude_guest_default = true;
 543         }
 544
 545         if (attr->precise_ip) {
 546                 if (!colon)
 547                         colon = ++r;
 548                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 549                 exclude_guest_default = true;
 550         }
 551
 552         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 553                 MOD_PRINT(host, 'H');
 554                 MOD_PRINT(guest, 'G');
 555         }
 556 #undef MOD_PRINT
 557         if (colon)
 558                 bf[colon - 1] = ':';
 559         return r;
 560 }
 561
 562 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 563 {
 564         int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 565         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 566 }
 567
 568 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
 569         "cpu-clock",
 570         "task-clock",
 571         "page-faults",
 572         "context-switches",
 573         "cpu-migrations",
 574         "minor-faults",
 575         "major-faults",
 576         "alignment-faults",
 577         "emulation-faults",
 578         "dummy",
 579 };
 580
 581 static const char *__evsel__sw_name(u64 config)
 582 {
 583         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 584                 return evsel__sw_names[config];
 585         return "unknown-software";
 586 }
 587
 588 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 589 {
 590         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 591         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 592 }
 593
 594 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 595 {
 596         int r;
 597
 598         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 599
 600         if (type & HW_BREAKPOINT_R)
 601                 r += scnprintf(bf + r, size - r, "r");
 602
 603         if (type & HW_BREAKPOINT_W)
 604                 r += scnprintf(bf + r, size - r, "w");
 605
 606         if (type & HW_BREAKPOINT_X)
 607                 r += scnprintf(bf + r, size - r, "x");
 608
 609         return r;
 610 }
 611
 612 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 613 {
 614         struct perf_event_attr *attr = &evsel->core.attr;
 615         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 616         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 617 }
 618
 619 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 620  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 621  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 622  { "LLC",       "L2",                                                   },
 623  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 624  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 625  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 626  { "node",                                                              },
 627 };
 628
 629 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 630  { "load",      "loads",        "read",                                 },
 631  { "store",     "stores",       "write",                                },
 632  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 633 };
 634
 635 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 636  { "refs",      "Reference",    "ops",          "access",               },
 637  { "misses",    "miss",                                                 },
 638 };
 639
 640 #define C(x)            PERF_COUNT_HW_CACHE_##x
 641 #define CACHE_READ      (1 << C(OP_READ))
 642 #define CACHE_WRITE     (1 << C(OP_WRITE))
 643 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 644 #define COP(x)          (1 << x)
 645
 646 /*
 647  * cache operation stat
 648  * L1I : Read and prefetch only
 649  * ITLB and BPU : Read-only
 650  */
 651 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
 652  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 653  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 654  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 655  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 656  [C(ITLB)]      = (CACHE_READ),
 657  [C(BPU)]       = (CACHE_READ),
 658  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 659 };
 660
 661 bool evsel__is_cache_op_valid(u8 type, u8 op)
 662 {
 663         if (evsel__hw_cache_stat[type] & COP(op))
 664                 return true;    /* valid */
 665         else
 666                 return false;   /* invalid */
 667 }
 668
 669 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 670 {
 671         if (result) {
 672                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 673                                  evsel__hw_cache_op[op][0],
 674                                  evsel__hw_cache_result[result][0]);
 675         }
 676
 677         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 678                          evsel__hw_cache_op[op][1]);
 679 }
 680
 681 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 682 {
 683         u8 op, result, type = (config >>  0) & 0xff;
 684         const char *err = "unknown-ext-hardware-cache-type";
 685
 686         if (type >= PERF_COUNT_HW_CACHE_MAX)
 687                 goto out_err;
 688
 689         op = (config >>  8) & 0xff;
 690         err = "unknown-ext-hardware-cache-op";
 691         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 692                 goto out_err;
 693
 694         result = (config >> 16) & 0xff;
 695         err = "unknown-ext-hardware-cache-result";
 696         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 697                 goto out_err;
 698
 699         err = "invalid-cache";
 700         if (!evsel__is_cache_op_valid(type, op))
 701                 goto out_err;
 702
 703         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 704 out_err:
 705         return scnprintf(bf, size, "%s", err);
 706 }
 707
 708 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 709 {
 710         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 711         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 712 }
 713
 714 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 715 {
 716         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 717         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 718 }
 719
 720 static int evsel__tool_name(char *bf, size_t size)
 721 {
 722         int ret = scnprintf(bf, size, "duration_time");
 723         return ret;
 724 }
 725
 726 const char *evsel__name(struct evsel *evsel)
 727 {
 728         char bf[128];
 729
 730         if (!evsel)
 731                 goto out_unknown;
 732
 733         if (evsel->name)
 734                 return evsel->name;
 735
 736         switch (evsel->core.attr.type) {
 737         case PERF_TYPE_RAW:
 738                 evsel__raw_name(evsel, bf, sizeof(bf));
 739                 break;
 740
 741         case PERF_TYPE_HARDWARE:
 742                 evsel__hw_name(evsel, bf, sizeof(bf));
 743                 break;
 744
 745         case PERF_TYPE_HW_CACHE:
 746                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 747                 break;
 748
 749         case PERF_TYPE_SOFTWARE:
 750                 if (evsel->tool_event)
 751                         evsel__tool_name(bf, sizeof(bf));
 752                 else
 753                         evsel__sw_name(evsel, bf, sizeof(bf));
 754                 break;
 755
 756         case PERF_TYPE_TRACEPOINT:
 757                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 758                 break;
 759
 760         case PERF_TYPE_BREAKPOINT:
 761                 evsel__bp_name(evsel, bf, sizeof(bf));
 762                 break;
 763
 764         default:
 765                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 766                           evsel->core.attr.type);
 767                 break;
 768         }
 769
 770         evsel->name = strdup(bf);
 771
 772         if (evsel->name)
 773                 return evsel->name;
 774 out_unknown:
 775         return "unknown";
 776 }
 777
 778 const char *evsel__group_name(struct evsel *evsel)
 779 {
 780         return evsel->group_name ?: "anon group";
 781 }
 782
 783 /*
 784  * Returns the group details for the specified leader,
 785  * with following rules.
 786  *
 787  *  For record -e '{cycles,instructions}'
 788  *    'anon group { cycles:u, instructions:u }'
 789  *
 790  *  For record -e 'cycles,instructions' and report --group
 791  *    'cycles:u, instructions:u'
 792  */
 793 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 794 {
 795         int ret = 0;
 796         struct evsel *pos;
 797         const char *group_name = evsel__group_name(evsel);
 798
 799         if (!evsel->forced_leader)
 800                 ret = scnprintf(buf, size, "%s { ", group_name);
 801
 802         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 803
 804         for_each_group_member(pos, evsel)
 805                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 806
 807         if (!evsel->forced_leader)
 808                 ret += scnprintf(buf + ret, size - ret, " }");
 809
 810         return ret;
 811 }
 812
 813 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 814                                       struct callchain_param *param)
 815 {
 816         bool function = evsel__is_function_event(evsel);
 817         struct perf_event_attr *attr = &evsel->core.attr;
 818
 819         evsel__set_sample_bit(evsel, CALLCHAIN);
 820
 821         attr->sample_max_stack = param->max_stack;
 822
 823         if (opts->kernel_callchains)
 824                 attr->exclude_callchain_user = 1;
 825         if (opts->user_callchains)
 826                 attr->exclude_callchain_kernel = 1;
 827         if (param->record_mode == CALLCHAIN_LBR) {
 828                 if (!opts->branch_stack) {
 829                         if (attr->exclude_user) {
 830                                 pr_warning("LBR callstack option is only available "
 831                                            "to get user callchain information. "
 832                                            "Falling back to framepointers.\n");
 833                         } else {
 834                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 835                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 836                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 837                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 838                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 839                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 840                         }
 841                 } else
 842                          pr_warning("Cannot use LBR callstack with branch stack. "
 843                                     "Falling back to framepointers.\n");
 844         }
 845
 846         if (param->record_mode == CALLCHAIN_DWARF) {
 847                 if (!function) {
 848                         evsel__set_sample_bit(evsel, REGS_USER);
 849                         evsel__set_sample_bit(evsel, STACK_USER);
 850                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 851                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 852                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 853                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 854                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 855                         } else {
 856                                 attr->sample_regs_user |= PERF_REGS_MASK;
 857                         }
 858                         attr->sample_stack_user = param->dump_size;
 859                         attr->exclude_callchain_user = 1;
 860                 } else {
 861                         pr_info("Cannot use DWARF unwind for function trace event,"
 862                                 " falling back to framepointers.\n");
 863                 }
 864         }
 865
 866         if (function) {
 867                 pr_info("Disabling user space callchains for function trace event.\n");
 868                 attr->exclude_callchain_user = 1;
 869         }
 870 }
 871
 872 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 873                              struct callchain_param *param)
 874 {
 875         if (param->enabled)
 876                 return __evsel__config_callchain(evsel, opts, param);
 877 }
 878
 879 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 880 {
 881         struct perf_event_attr *attr = &evsel->core.attr;
 882
 883         evsel__reset_sample_bit(evsel, CALLCHAIN);
 884         if (param->record_mode == CALLCHAIN_LBR) {
 885                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 886                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 887                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 888                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 889         }
 890         if (param->record_mode == CALLCHAIN_DWARF) {
 891                 evsel__reset_sample_bit(evsel, REGS_USER);
 892                 evsel__reset_sample_bit(evsel, STACK_USER);
 893         }
 894 }
 895
 896 static void evsel__apply_config_terms(struct evsel *evsel,
 897                                       struct record_opts *opts, bool track)
 898 {
 899         struct evsel_config_term *term;
 900         struct list_head *config_terms = &evsel->config_terms;
 901         struct perf_event_attr *attr = &evsel->core.attr;
 902         /* callgraph default */
 903         struct callchain_param param = {
 904                 .record_mode = callchain_param.record_mode,
 905         };
 906         u32 dump_size = 0;
 907         int max_stack = 0;
 908         const char *callgraph_buf = NULL;
 909
 910         list_for_each_entry(term, config_terms, list) {
 911                 switch (term->type) {
 912                 case EVSEL__CONFIG_TERM_PERIOD:
 913                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 914                                 attr->sample_period = term->val.period;
 915                                 attr->freq = 0;
 916                                 evsel__reset_sample_bit(evsel, PERIOD);
 917                         }
 918                         break;
 919                 case EVSEL__CONFIG_TERM_FREQ:
 920                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 921                                 attr->sample_freq = term->val.freq;
 922                                 attr->freq = 1;
 923                                 evsel__set_sample_bit(evsel, PERIOD);
 924                         }
 925                         break;
 926                 case EVSEL__CONFIG_TERM_TIME:
 927                         if (term->val.time)
 928                                 evsel__set_sample_bit(evsel, TIME);
 929                         else
 930                                 evsel__reset_sample_bit(evsel, TIME);
 931                         break;
 932                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 933                         callgraph_buf = term->val.str;
 934                         break;
 935                 case EVSEL__CONFIG_TERM_BRANCH:
 936                         if (term->val.str && strcmp(term->val.str, "no")) {
 937                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 938                                 parse_branch_str(term->val.str,
 939                                                  &attr->branch_sample_type);
 940                         } else
 941                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 942                         break;
 943                 case EVSEL__CONFIG_TERM_STACK_USER:
 944                         dump_size = term->val.stack_user;
 945                         break;
 946                 case EVSEL__CONFIG_TERM_MAX_STACK:
 947                         max_stack = term->val.max_stack;
 948                         break;
 949                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 950                         evsel->max_events = term->val.max_events;
 951                         break;
 952                 case EVSEL__CONFIG_TERM_INHERIT:
 953                         /*
 954                          * attr->inherit should has already been set by
 955                          * evsel__config. If user explicitly set
 956                          * inherit using config terms, override global
 957                          * opt->no_inherit setting.
 958                          */
 959                         attr->inherit = term->val.inherit ? 1 : 0;
 960                         break;
 961                 case EVSEL__CONFIG_TERM_OVERWRITE:
 962                         attr->write_backward = term->val.overwrite ? 1 : 0;
 963                         break;
 964                 case EVSEL__CONFIG_TERM_DRV_CFG:
 965                         break;
 966                 case EVSEL__CONFIG_TERM_PERCORE:
 967                         break;
 968                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
 969                         attr->aux_output = term->val.aux_output ? 1 : 0;
 970                         break;
 971                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 972                         /* Already applied by auxtrace */
 973                         break;
 974                 case EVSEL__CONFIG_TERM_CFG_CHG:
 975                         break;
 976                 default:
 977                         break;
 978                 }
 979         }
 980
 981         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 982         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 983                 bool sample_address = false;
 984
 985                 if (max_stack) {
 986                         param.max_stack = max_stack;
 987                         if (callgraph_buf == NULL)
 988                                 callgraph_buf = "fp";
 989                 }
 990
 991                 /* parse callgraph parameters */
 992                 if (callgraph_buf != NULL) {
 993                         if (!strcmp(callgraph_buf, "no")) {
 994                                 param.enabled = false;
 995                                 param.record_mode = CALLCHAIN_NONE;
 996                         } else {
 997                                 param.enabled = true;
 998                                 if (parse_callchain_record(callgraph_buf, &param)) {
 999                                         pr_err("per-event callgraph setting for %s failed. "
1000                                                "Apply callgraph global setting for it\n",
1001                                                evsel->name);
1002                                         return;
1003                                 }
1004                                 if (param.record_mode == CALLCHAIN_DWARF)
1005                                         sample_address = true;
1006                         }
1007                 }
1008                 if (dump_size > 0) {
1009                         dump_size = round_up(dump_size, sizeof(u64));
1010                         param.dump_size = dump_size;
1011                 }
1012
1013                 /* If global callgraph set, clear it */
1014                 if (callchain_param.enabled)
1015                         evsel__reset_callgraph(evsel, &callchain_param);
1016
1017                 /* set perf-event callgraph */
1018                 if (param.enabled) {
1019                         if (sample_address) {
1020                                 evsel__set_sample_bit(evsel, ADDR);
1021                                 evsel__set_sample_bit(evsel, DATA_SRC);
1022                                 evsel->core.attr.mmap_data = track;
1023                         }
1024                         evsel__config_callchain(evsel, opts, &param);
1025                 }
1026         }
1027 }
1028
1029 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1030 {
1031         struct evsel_config_term *term, *found_term = NULL;
1032
1033         list_for_each_entry(term, &evsel->config_terms, list) {
1034                 if (term->type == type)
1035                         found_term = term;
1036         }
1037
1038         return found_term;
1039 }
1040
1041 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1042 {
1043         evsel__set_sample_bit(evsel, WEIGHT);
1044 }
1045
1046 /*
1047  * The enable_on_exec/disabled value strategy:
1048  *
1049  *  1) For any type of traced program:
1050  *    - all independent events and group leaders are disabled
1051  *    - all group members are enabled
1052  *
1053  *     Group members are ruled by group leaders. They need to
1054  *     be enabled, because the group scheduling relies on that.
1055  *
1056  *  2) For traced programs executed by perf:
1057  *     - all independent events and group leaders have
1058  *       enable_on_exec set
1059  *     - we don't specifically enable or disable any event during
1060  *       the record command
1061  *
1062  *     Independent events and group leaders are initially disabled
1063  *     and get enabled by exec. Group members are ruled by group
1064  *     leaders as stated in 1).
1065  *
1066  *  3) For traced programs attached by perf (pid/tid):
1067  *     - we specifically enable or disable all events during
1068  *       the record command
1069  *
1070  *     When attaching events to already running traced we
1071  *     enable/disable events specifically, as there's no
1072  *     initial traced exec call.
1073  */
1074 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1075                    struct callchain_param *callchain)
1076 {
1077         struct evsel *leader = evsel->leader;
1078         struct perf_event_attr *attr = &evsel->core.attr;
1079         int track = evsel->tracking;
1080         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1081
1082         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1083         attr->inherit       = !opts->no_inherit;
1084         attr->write_backward = opts->overwrite ? 1 : 0;
1085
1086         evsel__set_sample_bit(evsel, IP);
1087         evsel__set_sample_bit(evsel, TID);
1088
1089         if (evsel->sample_read) {
1090                 evsel__set_sample_bit(evsel, READ);
1091
1092                 /*
1093                  * We need ID even in case of single event, because
1094                  * PERF_SAMPLE_READ process ID specific data.
1095                  */
1096                 evsel__set_sample_id(evsel, false);
1097
1098                 /*
1099                  * Apply group format only if we belong to group
1100                  * with more than one members.
1101                  */
1102                 if (leader->core.nr_members > 1) {
1103                         attr->read_format |= PERF_FORMAT_GROUP;
1104                         attr->inherit = 0;
1105                 }
1106         }
1107
1108         /*
1109          * We default some events to have a default interval. But keep
1110          * it a weak assumption overridable by the user.
1111          */
1112         if (!attr->sample_period) {
1113                 if (opts->freq) {
1114                         attr->freq              = 1;
1115                         attr->sample_freq       = opts->freq;
1116                 } else {
1117                         attr->sample_period = opts->default_interval;
1118                 }
1119         }
1120         /*
1121          * If attr->freq was set (here or earlier), ask for period
1122          * to be sampled.
1123          */
1124         if (attr->freq)
1125                 evsel__set_sample_bit(evsel, PERIOD);
1126
1127         if (opts->no_samples)
1128                 attr->sample_freq = 0;
1129
1130         if (opts->inherit_stat) {
1131                 evsel->core.attr.read_format |=
1132                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1133                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1134                         PERF_FORMAT_ID;
1135                 attr->inherit_stat = 1;
1136         }
1137
1138         if (opts->sample_address) {
1139                 evsel__set_sample_bit(evsel, ADDR);
1140                 attr->mmap_data = track;
1141         }
1142
1143         /*
1144          * We don't allow user space callchains for  function trace
1145          * event, due to issues with page faults while tracing page
1146          * fault handler and its overall trickiness nature.
1147          */
1148         if (evsel__is_function_event(evsel))
1149                 evsel->core.attr.exclude_callchain_user = 1;
1150
1151         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1152                 evsel__config_callchain(evsel, opts, callchain);
1153
1154         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1155             !evsel__is_dummy_event(evsel)) {
1156                 attr->sample_regs_intr = opts->sample_intr_regs;
1157                 evsel__set_sample_bit(evsel, REGS_INTR);
1158         }
1159
1160         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1161             !evsel__is_dummy_event(evsel)) {
1162                 attr->sample_regs_user |= opts->sample_user_regs;
1163                 evsel__set_sample_bit(evsel, REGS_USER);
1164         }
1165
1166         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1167                 evsel__set_sample_bit(evsel, CPU);
1168
1169         /*
1170          * When the user explicitly disabled time don't force it here.
1171          */
1172         if (opts->sample_time &&
1173             (!perf_missing_features.sample_id_all &&
1174             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1175              opts->sample_time_set)))
1176                 evsel__set_sample_bit(evsel, TIME);
1177
1178         if (opts->raw_samples && !evsel->no_aux_samples) {
1179                 evsel__set_sample_bit(evsel, TIME);
1180                 evsel__set_sample_bit(evsel, RAW);
1181                 evsel__set_sample_bit(evsel, CPU);
1182         }
1183
1184         if (opts->sample_address)
1185                 evsel__set_sample_bit(evsel, DATA_SRC);
1186
1187         if (opts->sample_phys_addr)
1188                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1189
1190         if (opts->no_buffering) {
1191                 attr->watermark = 0;
1192                 attr->wakeup_events = 1;
1193         }
1194         if (opts->branch_stack && !evsel->no_aux_samples) {
1195                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1196                 attr->branch_sample_type = opts->branch_stack;
1197         }
1198
1199         if (opts->sample_weight)
1200                 arch_evsel__set_sample_weight(evsel);
1201
1202         attr->task     = track;
1203         attr->mmap     = track;
1204         attr->mmap2    = track && !perf_missing_features.mmap2;
1205         attr->comm     = track;
1206         attr->build_id = track && opts->build_id;
1207
1208         /*
1209          * ksymbol is tracked separately with text poke because it needs to be
1210          * system wide and enabled immediately.
1211          */
1212         if (!opts->text_poke)
1213                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1214         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1215
1216         if (opts->record_namespaces)
1217                 attr->namespaces  = track;
1218
1219         if (opts->record_cgroup) {
1220                 attr->cgroup = track && !perf_missing_features.cgroup;
1221                 evsel__set_sample_bit(evsel, CGROUP);
1222         }
1223
1224         if (opts->sample_data_page_size)
1225                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1226
1227         if (opts->sample_code_page_size)
1228                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1229
1230         if (opts->record_switch_events)
1231                 attr->context_switch = track;
1232
1233         if (opts->sample_transaction)
1234                 evsel__set_sample_bit(evsel, TRANSACTION);
1235
1236         if (opts->running_time) {
1237                 evsel->core.attr.read_format |=
1238                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1239                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1240         }
1241
1242         /*
1243          * XXX see the function comment above
1244          *
1245          * Disabling only independent events or group leaders,
1246          * keeping group members enabled.
1247          */
1248         if (evsel__is_group_leader(evsel))
1249                 attr->disabled = 1;
1250
1251         /*
1252          * Setting enable_on_exec for independent events and
1253          * group leaders for traced executed by perf.
1254          */
1255         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1256             !opts->initial_delay)
1257                 attr->enable_on_exec = 1;
1258
1259         if (evsel->immediate) {
1260                 attr->disabled = 0;
1261                 attr->enable_on_exec = 0;
1262         }
1263
1264         clockid = opts->clockid;
1265         if (opts->use_clockid) {
1266                 attr->use_clockid = 1;
1267                 attr->clockid = opts->clockid;
1268         }
1269
1270         if (evsel->precise_max)
1271                 attr->precise_ip = 3;
1272
1273         if (opts->all_user) {
1274                 attr->exclude_kernel = 1;
1275                 attr->exclude_user   = 0;
1276         }
1277
1278         if (opts->all_kernel) {
1279                 attr->exclude_kernel = 0;
1280                 attr->exclude_user   = 1;
1281         }
1282
1283         if (evsel->core.own_cpus || evsel->unit)
1284                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1285
1286         /*
1287          * Apply event specific term settings,
1288          * it overloads any global configuration.
1289          */
1290         evsel__apply_config_terms(evsel, opts, track);
1291
1292         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1293
1294         /* The --period option takes the precedence. */
1295         if (opts->period_set) {
1296                 if (opts->period)
1297                         evsel__set_sample_bit(evsel, PERIOD);
1298                 else
1299                         evsel__reset_sample_bit(evsel, PERIOD);
1300         }
1301
1302         /*
1303          * A dummy event never triggers any actual counter and therefore
1304          * cannot be used with branch_stack.
1305          *
1306          * For initial_delay, a dummy event is added implicitly.
1307          * The software event will trigger -EOPNOTSUPP error out,
1308          * if BRANCH_STACK bit is set.
1309          */
1310         if (evsel__is_dummy_event(evsel))
1311                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1312 }
1313
1314 int evsel__set_filter(struct evsel *evsel, const char *filter)
1315 {
1316         char *new_filter = strdup(filter);
1317
1318         if (new_filter != NULL) {
1319                 free(evsel->filter);
1320                 evsel->filter = new_filter;
1321                 return 0;
1322         }
1323
1324         return -1;
1325 }
1326
1327 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1328 {
1329         char *new_filter;
1330
1331         if (evsel->filter == NULL)
1332                 return evsel__set_filter(evsel, filter);
1333
1334         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1335                 free(evsel->filter);
1336                 evsel->filter = new_filter;
1337                 return 0;
1338         }
1339
1340         return -1;
1341 }
1342
1343 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1344 {
1345         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1346 }
1347
1348 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1349 {
1350         return evsel__append_filter(evsel, "%s,%s", filter);
1351 }
1352
1353 /* Caller has to clear disabled after going through all CPUs. */
1354 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1355 {
1356         return perf_evsel__enable_cpu(&evsel->core, cpu);
1357 }
1358
1359 int evsel__enable(struct evsel *evsel)
1360 {
1361         int err = perf_evsel__enable(&evsel->core);
1362
1363         if (!err)
1364                 evsel->disabled = false;
1365         return err;
1366 }
1367
1368 /* Caller has to set disabled after going through all CPUs. */
1369 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1370 {
1371         return perf_evsel__disable_cpu(&evsel->core, cpu);
1372 }
1373
1374 int evsel__disable(struct evsel *evsel)
1375 {
1376         int err = perf_evsel__disable(&evsel->core);
1377         /*
1378          * We mark it disabled here so that tools that disable a event can
1379          * ignore events after they disable it. I.e. the ring buffer may have
1380          * already a few more events queued up before the kernel got the stop
1381          * request.
1382          */
1383         if (!err)
1384                 evsel->disabled = true;
1385
1386         return err;
1387 }
1388
1389 static void evsel__free_config_terms(struct evsel *evsel)
1390 {
1391         struct evsel_config_term *term, *h;
1392
1393         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1394                 list_del_init(&term->list);
1395                 if (term->free_str)
1396                         zfree(&term->val.str);
1397                 free(term);
1398         }
1399 }
1400
1401 void evsel__exit(struct evsel *evsel)
1402 {
1403         assert(list_empty(&evsel->core.node));
1404         assert(evsel->evlist == NULL);
1405         bpf_counter__destroy(evsel);
1406         evsel__free_counts(evsel);
1407         perf_evsel__free_fd(&evsel->core);
1408         perf_evsel__free_id(&evsel->core);
1409         evsel__free_config_terms(evsel);
1410         cgroup__put(evsel->cgrp);
1411         perf_cpu_map__put(evsel->core.cpus);
1412         perf_cpu_map__put(evsel->core.own_cpus);
1413         perf_thread_map__put(evsel->core.threads);
1414         zfree(&evsel->group_name);
1415         zfree(&evsel->name);
1416         zfree(&evsel->pmu_name);
1417         evsel__zero_per_pkg(evsel);
1418         hashmap__free(evsel->per_pkg_mask);
1419         evsel->per_pkg_mask = NULL;
1420         zfree(&evsel->metric_events);
1421         perf_evsel__object.fini(evsel);
1422 }
1423
1424 void evsel__delete(struct evsel *evsel)
1425 {
1426         evsel__exit(evsel);
1427         free(evsel);
1428 }
1429
1430 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1431                            struct perf_counts_values *count)
1432 {
1433         struct perf_counts_values tmp;
1434
1435         if (!evsel->prev_raw_counts)
1436                 return;
1437
1438         if (cpu == -1) {
1439                 tmp = evsel->prev_raw_counts->aggr;
1440                 evsel->prev_raw_counts->aggr = *count;
1441         } else {
1442                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1443                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1444         }
1445
1446         count->val = count->val - tmp.val;
1447         count->ena = count->ena - tmp.ena;
1448         count->run = count->run - tmp.run;
1449 }
1450
1451 void perf_counts_values__scale(struct perf_counts_values *count,
1452                                bool scale, s8 *pscaled)
1453 {
1454         s8 scaled = 0;
1455
1456         if (scale) {
1457                 if (count->run == 0) {
1458                         scaled = -1;
1459                         count->val = 0;
1460                 } else if (count->run < count->ena) {
1461                         scaled = 1;
1462                         count->val = (u64)((double) count->val * count->ena / count->run);
1463                 }
1464         }
1465
1466         if (pscaled)
1467                 *pscaled = scaled;
1468 }
1469
1470 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1471 {
1472         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1473
1474         return perf_evsel__read(&evsel->core, cpu, thread, count);
1475 }
1476
1477 static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val, u64 ena, u64 run)
1478 {
1479         struct perf_counts_values *count;
1480
1481         count = perf_counts(counter->counts, cpu, thread);
1482
1483         count->val    = val;
1484         count->ena    = ena;
1485         count->run    = run;
1486
1487         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1488 }
1489
1490 static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data)
1491 {
1492         u64 read_format = leader->core.attr.read_format;
1493         struct sample_read_value *v;
1494         u64 nr, ena = 0, run = 0, i;
1495
1496         nr = *data++;
1497
1498         if (nr != (u64) leader->core.nr_members)
1499                 return -EINVAL;
1500
1501         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1502                 ena = *data++;
1503
1504         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1505                 run = *data++;
1506
1507         v = (struct sample_read_value *) data;
1508
1509         evsel__set_count(leader, cpu, thread, v[0].value, ena, run);
1510
1511         for (i = 1; i < nr; i++) {
1512                 struct evsel *counter;
1513
1514                 counter = evlist__id2evsel(leader->evlist, v[i].id);
1515                 if (!counter)
1516                         return -EINVAL;
1517
1518                 evsel__set_count(counter, cpu, thread, v[i].value, ena, run);
1519         }
1520
1521         return 0;
1522 }
1523
1524 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1525 {
1526         struct perf_stat_evsel *ps = leader->stats;
1527         u64 read_format = leader->core.attr.read_format;
1528         int size = perf_evsel__read_size(&leader->core);
1529         u64 *data = ps->group_data;
1530
1531         if (!(read_format & PERF_FORMAT_ID))
1532                 return -EINVAL;
1533
1534         if (!evsel__is_group_leader(leader))
1535                 return -EINVAL;
1536
1537         if (!data) {
1538                 data = zalloc(size);
1539                 if (!data)
1540                         return -ENOMEM;
1541
1542                 ps->group_data = data;
1543         }
1544
1545         if (FD(leader, cpu, thread) < 0)
1546                 return -EINVAL;
1547
1548         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1549                 return -errno;
1550
1551         return evsel__process_group_data(leader, cpu, thread, data);
1552 }
1553
1554 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1555 {
1556         u64 read_format = evsel->core.attr.read_format;
1557
1558         if (read_format & PERF_FORMAT_GROUP)
1559                 return evsel__read_group(evsel, cpu, thread);
1560
1561         return evsel__read_one(evsel, cpu, thread);
1562 }
1563
1564 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1565 {
1566         struct perf_counts_values count;
1567         size_t nv = scale ? 3 : 1;
1568
1569         if (FD(evsel, cpu, thread) < 0)
1570                 return -EINVAL;
1571
1572         if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1573                 return -ENOMEM;
1574
1575         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1576                 return -errno;
1577
1578         evsel__compute_deltas(evsel, cpu, thread, &count);
1579         perf_counts_values__scale(&count, scale, NULL);
1580         *perf_counts(evsel->counts, cpu, thread) = count;
1581         return 0;
1582 }
1583
1584 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1585                                   int cpu)
1586 {
1587         int cpuid;
1588
1589         cpuid = perf_cpu_map__cpu(evsel->core.cpus, cpu);
1590         return perf_cpu_map__idx(other->core.cpus, cpuid);
1591 }
1592
1593 static int evsel__hybrid_group_cpu(struct evsel *evsel, int cpu)
1594 {
1595         struct evsel *leader = evsel->leader;
1596
1597         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1598             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1599                 return evsel__match_other_cpu(evsel, leader, cpu);
1600         }
1601
1602         return cpu;
1603 }
1604
1605 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1606 {
1607         struct evsel *leader = evsel->leader;
1608         int fd;
1609
1610         if (evsel__is_group_leader(evsel))
1611                 return -1;
1612
1613         /*
1614          * Leader must be already processed/open,
1615          * if not it's a bug.
1616          */
1617         BUG_ON(!leader->core.fd);
1618
1619         cpu = evsel__hybrid_group_cpu(evsel, cpu);
1620         if (cpu == -1)
1621                 return -1;
1622
1623         fd = FD(leader, cpu, thread);
1624         BUG_ON(fd == -1);
1625
1626         return fd;
1627 }
1628
1629 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1630 {
1631         for (int cpu = 0; cpu < nr_cpus; cpu++)
1632                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1633                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1634 }
1635
1636 static int update_fds(struct evsel *evsel,
1637                       int nr_cpus, int cpu_idx,
1638                       int nr_threads, int thread_idx)
1639 {
1640         struct evsel *pos;
1641
1642         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1643                 return -EINVAL;
1644
1645         evlist__for_each_entry(evsel->evlist, pos) {
1646                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1647
1648                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1649
1650                 /*
1651                  * Since fds for next evsel has not been created,
1652                  * there is no need to iterate whole event list.
1653                  */
1654                 if (pos == evsel)
1655                         break;
1656         }
1657         return 0;
1658 }
1659
1660 static bool ignore_missing_thread(struct evsel *evsel,
1661                                   int nr_cpus, int cpu,
1662                                   struct perf_thread_map *threads,
1663                                   int thread, int err)
1664 {
1665         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1666
1667         if (!evsel->ignore_missing_thread)
1668                 return false;
1669
1670         /* The system wide setup does not work with threads. */
1671         if (evsel->core.system_wide)
1672                 return false;
1673
1674         /* The -ESRCH is perf event syscall errno for pid's not found. */
1675         if (err != -ESRCH)
1676                 return false;
1677
1678         /* If there's only one thread, let it fail. */
1679         if (threads->nr == 1)
1680                 return false;
1681
1682         /*
1683          * We should remove fd for missing_thread first
1684          * because thread_map__remove() will decrease threads->nr.
1685          */
1686         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1687                 return false;
1688
1689         if (thread_map__remove(threads, thread))
1690                 return false;
1691
1692         pr_warning("WARNING: Ignored open failure for pid %d\n",
1693                    ignore_pid);
1694         return true;
1695 }
1696
1697 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1698                                 void *priv __maybe_unused)
1699 {
1700         return fprintf(fp, "  %-32s %s\n", name, val);
1701 }
1702
1703 static void display_attr(struct perf_event_attr *attr)
1704 {
1705         if (verbose >= 2 || debug_peo_args) {
1706                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1707                 fprintf(stderr, "perf_event_attr:\n");
1708                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1709                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1710         }
1711 }
1712
1713 static int perf_event_open(struct evsel *evsel,
1714                            pid_t pid, int cpu, int group_fd,
1715                            unsigned long flags)
1716 {
1717         int precise_ip = evsel->core.attr.precise_ip;
1718         int fd;
1719
1720         while (1) {
1721                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1722                           pid, cpu, group_fd, flags);
1723
1724                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1725                 if (fd >= 0)
1726                         break;
1727
1728                 /* Do not try less precise if not requested. */
1729                 if (!evsel->precise_max)
1730                         break;
1731
1732                 /*
1733                  * We tried all the precise_ip values, and it's
1734                  * still failing, so leave it to standard fallback.
1735                  */
1736                 if (!evsel->core.attr.precise_ip) {
1737                         evsel->core.attr.precise_ip = precise_ip;
1738                         break;
1739                 }
1740
1741                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1742                 evsel->core.attr.precise_ip--;
1743                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1744                 display_attr(&evsel->core.attr);
1745         }
1746
1747         return fd;
1748 }
1749
1750 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1751                 struct perf_thread_map *threads,
1752                 int start_cpu, int end_cpu)
1753 {
1754         int cpu, thread, nthreads;
1755         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1756         int pid = -1, err, old_errno;
1757         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1758
1759         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1760             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1761                 return -EINVAL;
1762
1763         if (cpus == NULL) {
1764                 static struct perf_cpu_map *empty_cpu_map;
1765
1766                 if (empty_cpu_map == NULL) {
1767                         empty_cpu_map = perf_cpu_map__dummy_new();
1768                         if (empty_cpu_map == NULL)
1769                                 return -ENOMEM;
1770                 }
1771
1772                 cpus = empty_cpu_map;
1773         }
1774
1775         if (threads == NULL) {
1776                 static struct perf_thread_map *empty_thread_map;
1777
1778                 if (empty_thread_map == NULL) {
1779                         empty_thread_map = thread_map__new_by_tid(-1);
1780                         if (empty_thread_map == NULL)
1781                                 return -ENOMEM;
1782                 }
1783
1784                 threads = empty_thread_map;
1785         }
1786
1787         if (evsel->core.system_wide)
1788                 nthreads = 1;
1789         else
1790                 nthreads = threads->nr;
1791
1792         if (evsel->core.fd == NULL &&
1793             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1794                 return -ENOMEM;
1795
1796         if (evsel->cgrp) {
1797                 flags |= PERF_FLAG_PID_CGROUP;
1798                 pid = evsel->cgrp->fd;
1799         }
1800
1801 fallback_missing_features:
1802         if (perf_missing_features.weight_struct) {
1803                 evsel__set_sample_bit(evsel, WEIGHT);
1804                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1805         }
1806         if (perf_missing_features.clockid_wrong)
1807                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1808         if (perf_missing_features.clockid) {
1809                 evsel->core.attr.use_clockid = 0;
1810                 evsel->core.attr.clockid = 0;
1811         }
1812         if (perf_missing_features.cloexec)
1813                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1814         if (perf_missing_features.mmap2)
1815                 evsel->core.attr.mmap2 = 0;
1816         if (perf_missing_features.exclude_guest)
1817                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1818         if (perf_missing_features.lbr_flags)
1819                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1820                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1821         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1822                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1823         if (perf_missing_features.ksymbol)
1824                 evsel->core.attr.ksymbol = 0;
1825         if (perf_missing_features.bpf)
1826                 evsel->core.attr.bpf_event = 0;
1827         if (perf_missing_features.branch_hw_idx)
1828                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1829 retry_sample_id:
1830         if (perf_missing_features.sample_id_all)
1831                 evsel->core.attr.sample_id_all = 0;
1832
1833         display_attr(&evsel->core.attr);
1834
1835         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1836
1837                 for (thread = 0; thread < nthreads; thread++) {
1838                         int fd, group_fd;
1839
1840                         if (!evsel->cgrp && !evsel->core.system_wide)
1841                                 pid = perf_thread_map__pid(threads, thread);
1842
1843                         group_fd = get_group_fd(evsel, cpu, thread);
1844 retry_open:
1845                         test_attr__ready();
1846
1847                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1848                                              group_fd, flags);
1849
1850                         FD(evsel, cpu, thread) = fd;
1851
1852                         bpf_counter__install_pe(evsel, cpu, fd);
1853
1854                         if (unlikely(test_attr__enabled)) {
1855                                 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1856                                                 fd, group_fd, flags);
1857                         }
1858
1859                         if (fd < 0) {
1860                                 err = -errno;
1861
1862                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1863                                         /*
1864                                          * We just removed 1 thread, so take a step
1865                                          * back on thread index and lower the upper
1866                                          * nthreads limit.
1867                                          */
1868                                         nthreads--;
1869                                         thread--;
1870
1871                                         /* ... and pretend like nothing have happened. */
1872                                         err = 0;
1873                                         continue;
1874                                 }
1875
1876                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1877                                           err);
1878                                 goto try_fallback;
1879                         }
1880
1881                         pr_debug2_peo(" = %d\n", fd);
1882
1883                         if (evsel->bpf_fd >= 0) {
1884                                 int evt_fd = fd;
1885                                 int bpf_fd = evsel->bpf_fd;
1886
1887                                 err = ioctl(evt_fd,
1888                                             PERF_EVENT_IOC_SET_BPF,
1889                                             bpf_fd);
1890                                 if (err && errno != EEXIST) {
1891                                         pr_err("failed to attach bpf fd %d: %s\n",
1892                                                bpf_fd, strerror(errno));
1893                                         err = -EINVAL;
1894                                         goto out_close;
1895                                 }
1896                         }
1897
1898                         set_rlimit = NO_CHANGE;
1899
1900                         /*
1901                          * If we succeeded but had to kill clockid, fail and
1902                          * have evsel__open_strerror() print us a nice error.
1903                          */
1904                         if (perf_missing_features.clockid ||
1905                             perf_missing_features.clockid_wrong) {
1906                                 err = -EINVAL;
1907                                 goto out_close;
1908                         }
1909                 }
1910         }
1911
1912         return 0;
1913
1914 try_fallback:
1915         /*
1916          * perf stat needs between 5 and 22 fds per CPU. When we run out
1917          * of them try to increase the limits.
1918          */
1919         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1920                 struct rlimit l;
1921
1922                 old_errno = errno;
1923                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1924                         if (set_rlimit == NO_CHANGE)
1925                                 l.rlim_cur = l.rlim_max;
1926                         else {
1927                                 l.rlim_cur = l.rlim_max + 1000;
1928                                 l.rlim_max = l.rlim_cur;
1929                         }
1930                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1931                                 set_rlimit++;
1932                                 errno = old_errno;
1933                                 goto retry_open;
1934                         }
1935                 }
1936                 errno = old_errno;
1937         }
1938
1939         if (err != -EINVAL || cpu > 0 || thread > 0)
1940                 goto out_close;
1941
1942         /*
1943          * Must probe features in the order they were added to the
1944          * perf_event_attr interface.
1945          */
1946         if (!perf_missing_features.weight_struct &&
1947             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1948                 perf_missing_features.weight_struct = true;
1949                 pr_debug2("switching off weight struct support\n");
1950                 goto fallback_missing_features;
1951         } else if (!perf_missing_features.code_page_size &&
1952             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1953                 perf_missing_features.code_page_size = true;
1954                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1955                 goto out_close;
1956         } else if (!perf_missing_features.data_page_size &&
1957             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1958                 perf_missing_features.data_page_size = true;
1959                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1960                 goto out_close;
1961         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1962                 perf_missing_features.cgroup = true;
1963                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1964                 goto out_close;
1965         } else if (!perf_missing_features.branch_hw_idx &&
1966             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1967                 perf_missing_features.branch_hw_idx = true;
1968                 pr_debug2("switching off branch HW index support\n");
1969                 goto fallback_missing_features;
1970         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1971                 perf_missing_features.aux_output = true;
1972                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1973                 goto out_close;
1974         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1975                 perf_missing_features.bpf = true;
1976                 pr_debug2_peo("switching off bpf_event\n");
1977                 goto fallback_missing_features;
1978         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1979                 perf_missing_features.ksymbol = true;
1980                 pr_debug2_peo("switching off ksymbol\n");
1981                 goto fallback_missing_features;
1982         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1983                 perf_missing_features.write_backward = true;
1984                 pr_debug2_peo("switching off write_backward\n");
1985                 goto out_close;
1986         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1987                 perf_missing_features.clockid_wrong = true;
1988                 pr_debug2_peo("switching off clockid\n");
1989                 goto fallback_missing_features;
1990         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1991                 perf_missing_features.clockid = true;
1992                 pr_debug2_peo("switching off use_clockid\n");
1993                 goto fallback_missing_features;
1994         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1995                 perf_missing_features.cloexec = true;
1996                 pr_debug2_peo("switching off cloexec flag\n");
1997                 goto fallback_missing_features;
1998         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1999                 perf_missing_features.mmap2 = true;
2000                 pr_debug2_peo("switching off mmap2\n");
2001                 goto fallback_missing_features;
2002         } else if (!perf_missing_features.exclude_guest &&
2003                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
2004                 perf_missing_features.exclude_guest = true;
2005                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2006                 goto fallback_missing_features;
2007         } else if (!perf_missing_features.sample_id_all) {
2008                 perf_missing_features.sample_id_all = true;
2009                 pr_debug2_peo("switching off sample_id_all\n");
2010                 goto retry_sample_id;
2011         } else if (!perf_missing_features.lbr_flags &&
2012                         (evsel->core.attr.branch_sample_type &
2013                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2014                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2015                 perf_missing_features.lbr_flags = true;
2016                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2017                 goto fallback_missing_features;
2018         } else if (!perf_missing_features.group_read &&
2019                     evsel->core.attr.inherit &&
2020                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2021                    evsel__is_group_leader(evsel)) {
2022                 perf_missing_features.group_read = true;
2023                 pr_debug2_peo("switching off group read\n");
2024                 goto fallback_missing_features;
2025         }
2026 out_close:
2027         if (err)
2028                 threads->err_thread = thread;
2029
2030         old_errno = errno;
2031         do {
2032                 while (--thread >= 0) {
2033                         if (FD(evsel, cpu, thread) >= 0)
2034                                 close(FD(evsel, cpu, thread));
2035                         FD(evsel, cpu, thread) = -1;
2036                 }
2037                 thread = nthreads;
2038         } while (--cpu >= 0);
2039         errno = old_errno;
2040         return err;
2041 }
2042
2043 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2044                 struct perf_thread_map *threads)
2045 {
2046         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
2047 }
2048
2049 void evsel__close(struct evsel *evsel)
2050 {
2051         perf_evsel__close(&evsel->core);
2052         perf_evsel__free_id(&evsel->core);
2053 }
2054
2055 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
2056 {
2057         if (cpu == -1)
2058                 return evsel__open_cpu(evsel, cpus, NULL, 0,
2059                                         cpus ? cpus->nr : 1);
2060
2061         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
2062 }
2063
2064 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2065 {
2066         return evsel__open(evsel, NULL, threads);
2067 }
2068
2069 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2070                                        const union perf_event *event,
2071                                        struct perf_sample *sample)
2072 {
2073         u64 type = evsel->core.attr.sample_type;
2074         const __u64 *array = event->sample.array;
2075         bool swapped = evsel->needs_swap;
2076         union u64_swap u;
2077
2078         array += ((event->header.size -
2079                    sizeof(event->header)) / sizeof(u64)) - 1;
2080
2081         if (type & PERF_SAMPLE_IDENTIFIER) {
2082                 sample->id = *array;
2083                 array--;
2084         }
2085
2086         if (type & PERF_SAMPLE_CPU) {
2087                 u.val64 = *array;
2088                 if (swapped) {
2089                         /* undo swap of u64, then swap on individual u32s */
2090                         u.val64 = bswap_64(u.val64);
2091                         u.val32[0] = bswap_32(u.val32[0]);
2092                 }
2093
2094                 sample->cpu = u.val32[0];
2095                 array--;
2096         }
2097
2098         if (type & PERF_SAMPLE_STREAM_ID) {
2099                 sample->stream_id = *array;
2100                 array--;
2101         }
2102
2103         if (type & PERF_SAMPLE_ID) {
2104                 sample->id = *array;
2105                 array--;
2106         }
2107
2108         if (type & PERF_SAMPLE_TIME) {
2109                 sample->time = *array;
2110                 array--;
2111         }
2112
2113         if (type & PERF_SAMPLE_TID) {
2114                 u.val64 = *array;
2115                 if (swapped) {
2116                         /* undo swap of u64, then swap on individual u32s */
2117                         u.val64 = bswap_64(u.val64);
2118                         u.val32[0] = bswap_32(u.val32[0]);
2119                         u.val32[1] = bswap_32(u.val32[1]);
2120                 }
2121
2122                 sample->pid = u.val32[0];
2123                 sample->tid = u.val32[1];
2124                 array--;
2125         }
2126
2127         return 0;
2128 }
2129
2130 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2131                             u64 size)
2132 {
2133         return size > max_size || offset + size > endp;
2134 }
2135
2136 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2137         do {                                                            \
2138                 if (overflow(endp, (max_size), (offset), (size)))       \
2139                         return -EFAULT;                                 \
2140         } while (0)
2141
2142 #define OVERFLOW_CHECK_u64(offset) \
2143         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2144
2145 static int
2146 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2147 {
2148         /*
2149          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2150          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2151          * check the format does not go past the end of the event.
2152          */
2153         if (sample_size + sizeof(event->header) > event->header.size)
2154                 return -EFAULT;
2155
2156         return 0;
2157 }
2158
2159 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2160                                           const __u64 *array,
2161                                           u64 type __maybe_unused)
2162 {
2163         data->weight = *array;
2164 }
2165
2166 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2167                         struct perf_sample *data)
2168 {
2169         u64 type = evsel->core.attr.sample_type;
2170         bool swapped = evsel->needs_swap;
2171         const __u64 *array;
2172         u16 max_size = event->header.size;
2173         const void *endp = (void *)event + max_size;
2174         u64 sz;
2175
2176         /*
2177          * used for cross-endian analysis. See git commit 65014ab3
2178          * for why this goofiness is needed.
2179          */
2180         union u64_swap u;
2181
2182         memset(data, 0, sizeof(*data));
2183         data->cpu = data->pid = data->tid = -1;
2184         data->stream_id = data->id = data->time = -1ULL;
2185         data->period = evsel->core.attr.sample_period;
2186         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2187         data->misc    = event->header.misc;
2188         data->id = -1ULL;
2189         data->data_src = PERF_MEM_DATA_SRC_NONE;
2190
2191         if (event->header.type != PERF_RECORD_SAMPLE) {
2192                 if (!evsel->core.attr.sample_id_all)
2193                         return 0;
2194                 return perf_evsel__parse_id_sample(evsel, event, data);
2195         }
2196
2197         array = event->sample.array;
2198
2199         if (perf_event__check_size(event, evsel->sample_size))
2200                 return -EFAULT;
2201
2202         if (type & PERF_SAMPLE_IDENTIFIER) {
2203                 data->id = *array;
2204                 array++;
2205         }
2206
2207         if (type & PERF_SAMPLE_IP) {
2208                 data->ip = *array;
2209                 array++;
2210         }
2211
2212         if (type & PERF_SAMPLE_TID) {
2213                 u.val64 = *array;
2214                 if (swapped) {
2215                         /* undo swap of u64, then swap on individual u32s */
2216                         u.val64 = bswap_64(u.val64);
2217                         u.val32[0] = bswap_32(u.val32[0]);
2218                         u.val32[1] = bswap_32(u.val32[1]);
2219                 }
2220
2221                 data->pid = u.val32[0];
2222                 data->tid = u.val32[1];
2223                 array++;
2224         }
2225
2226         if (type & PERF_SAMPLE_TIME) {
2227                 data->time = *array;
2228                 array++;
2229         }
2230
2231         if (type & PERF_SAMPLE_ADDR) {
2232                 data->addr = *array;
2233                 array++;
2234         }
2235
2236         if (type & PERF_SAMPLE_ID) {
2237                 data->id = *array;
2238                 array++;
2239         }
2240
2241         if (type & PERF_SAMPLE_STREAM_ID) {
2242                 data->stream_id = *array;
2243                 array++;
2244         }
2245
2246         if (type & PERF_SAMPLE_CPU) {
2247
2248                 u.val64 = *array;
2249                 if (swapped) {
2250                         /* undo swap of u64, then swap on individual u32s */
2251                         u.val64 = bswap_64(u.val64);
2252                         u.val32[0] = bswap_32(u.val32[0]);
2253                 }
2254
2255                 data->cpu = u.val32[0];
2256                 array++;
2257         }
2258
2259         if (type & PERF_SAMPLE_PERIOD) {
2260                 data->period = *array;
2261                 array++;
2262         }
2263
2264         if (type & PERF_SAMPLE_READ) {
2265                 u64 read_format = evsel->core.attr.read_format;
2266
2267                 OVERFLOW_CHECK_u64(array);
2268                 if (read_format & PERF_FORMAT_GROUP)
2269                         data->read.group.nr = *array;
2270                 else
2271                         data->read.one.value = *array;
2272
2273                 array++;
2274
2275                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2276                         OVERFLOW_CHECK_u64(array);
2277                         data->read.time_enabled = *array;
2278                         array++;
2279                 }
2280
2281                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2282                         OVERFLOW_CHECK_u64(array);
2283                         data->read.time_running = *array;
2284                         array++;
2285                 }
2286
2287                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2288                 if (read_format & PERF_FORMAT_GROUP) {
2289                         const u64 max_group_nr = UINT64_MAX /
2290                                         sizeof(struct sample_read_value);
2291
2292                         if (data->read.group.nr > max_group_nr)
2293                                 return -EFAULT;
2294                         sz = data->read.group.nr *
2295                              sizeof(struct sample_read_value);
2296                         OVERFLOW_CHECK(array, sz, max_size);
2297                         data->read.group.values =
2298                                         (struct sample_read_value *)array;
2299                         array = (void *)array + sz;
2300                 } else {
2301                         OVERFLOW_CHECK_u64(array);
2302                         data->read.one.id = *array;
2303                         array++;
2304                 }
2305         }
2306
2307         if (type & PERF_SAMPLE_CALLCHAIN) {
2308                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2309
2310                 OVERFLOW_CHECK_u64(array);
2311                 data->callchain = (struct ip_callchain *)array++;
2312                 if (data->callchain->nr > max_callchain_nr)
2313                         return -EFAULT;
2314                 sz = data->callchain->nr * sizeof(u64);
2315                 OVERFLOW_CHECK(array, sz, max_size);
2316                 array = (void *)array + sz;
2317         }
2318
2319         if (type & PERF_SAMPLE_RAW) {
2320                 OVERFLOW_CHECK_u64(array);
2321                 u.val64 = *array;
2322
2323                 /*
2324                  * Undo swap of u64, then swap on individual u32s,
2325                  * get the size of the raw area and undo all of the
2326                  * swap. The pevent interface handles endianness by
2327                  * itself.
2328                  */
2329                 if (swapped) {
2330                         u.val64 = bswap_64(u.val64);
2331                         u.val32[0] = bswap_32(u.val32[0]);
2332                         u.val32[1] = bswap_32(u.val32[1]);
2333                 }
2334                 data->raw_size = u.val32[0];
2335
2336                 /*
2337                  * The raw data is aligned on 64bits including the
2338                  * u32 size, so it's safe to use mem_bswap_64.
2339                  */
2340                 if (swapped)
2341                         mem_bswap_64((void *) array, data->raw_size);
2342
2343                 array = (void *)array + sizeof(u32);
2344
2345                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2346                 data->raw_data = (void *)array;
2347                 array = (void *)array + data->raw_size;
2348         }
2349
2350         if (type & PERF_SAMPLE_BRANCH_STACK) {
2351                 const u64 max_branch_nr = UINT64_MAX /
2352                                           sizeof(struct branch_entry);
2353
2354                 OVERFLOW_CHECK_u64(array);
2355                 data->branch_stack = (struct branch_stack *)array++;
2356
2357                 if (data->branch_stack->nr > max_branch_nr)
2358                         return -EFAULT;
2359
2360                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2361                 if (evsel__has_branch_hw_idx(evsel))
2362                         sz += sizeof(u64);
2363                 else
2364                         data->no_hw_idx = true;
2365                 OVERFLOW_CHECK(array, sz, max_size);
2366                 array = (void *)array + sz;
2367         }
2368
2369         if (type & PERF_SAMPLE_REGS_USER) {
2370                 OVERFLOW_CHECK_u64(array);
2371                 data->user_regs.abi = *array;
2372                 array++;
2373
2374                 if (data->user_regs.abi) {
2375                         u64 mask = evsel->core.attr.sample_regs_user;
2376
2377                         sz = hweight64(mask) * sizeof(u64);
2378                         OVERFLOW_CHECK(array, sz, max_size);
2379                         data->user_regs.mask = mask;
2380                         data->user_regs.regs = (u64 *)array;
2381                         array = (void *)array + sz;
2382                 }
2383         }
2384
2385         if (type & PERF_SAMPLE_STACK_USER) {
2386                 OVERFLOW_CHECK_u64(array);
2387                 sz = *array++;
2388
2389                 data->user_stack.offset = ((char *)(array - 1)
2390                                           - (char *) event);
2391
2392                 if (!sz) {
2393                         data->user_stack.size = 0;
2394                 } else {
2395                         OVERFLOW_CHECK(array, sz, max_size);
2396                         data->user_stack.data = (char *)array;
2397                         array = (void *)array + sz;
2398                         OVERFLOW_CHECK_u64(array);
2399                         data->user_stack.size = *array++;
2400                         if (WARN_ONCE(data->user_stack.size > sz,
2401                                       "user stack dump failure\n"))
2402                                 return -EFAULT;
2403                 }
2404         }
2405
2406         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2407                 OVERFLOW_CHECK_u64(array);
2408                 arch_perf_parse_sample_weight(data, array, type);
2409                 array++;
2410         }
2411
2412         if (type & PERF_SAMPLE_DATA_SRC) {
2413                 OVERFLOW_CHECK_u64(array);
2414                 data->data_src = *array;
2415                 array++;
2416         }
2417
2418         if (type & PERF_SAMPLE_TRANSACTION) {
2419                 OVERFLOW_CHECK_u64(array);
2420                 data->transaction = *array;
2421                 array++;
2422         }
2423
2424         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2425         if (type & PERF_SAMPLE_REGS_INTR) {
2426                 OVERFLOW_CHECK_u64(array);
2427                 data->intr_regs.abi = *array;
2428                 array++;
2429
2430                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2431                         u64 mask = evsel->core.attr.sample_regs_intr;
2432
2433                         sz = hweight64(mask) * sizeof(u64);
2434                         OVERFLOW_CHECK(array, sz, max_size);
2435                         data->intr_regs.mask = mask;
2436                         data->intr_regs.regs = (u64 *)array;
2437                         array = (void *)array + sz;
2438                 }
2439         }
2440
2441         data->phys_addr = 0;
2442         if (type & PERF_SAMPLE_PHYS_ADDR) {
2443                 data->phys_addr = *array;
2444                 array++;
2445         }
2446
2447         data->cgroup = 0;
2448         if (type & PERF_SAMPLE_CGROUP) {
2449                 data->cgroup = *array;
2450                 array++;
2451         }
2452
2453         data->data_page_size = 0;
2454         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2455                 data->data_page_size = *array;
2456                 array++;
2457         }
2458
2459         data->code_page_size = 0;
2460         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2461                 data->code_page_size = *array;
2462                 array++;
2463         }
2464
2465         if (type & PERF_SAMPLE_AUX) {
2466                 OVERFLOW_CHECK_u64(array);
2467                 sz = *array++;
2468
2469                 OVERFLOW_CHECK(array, sz, max_size);
2470                 /* Undo swap of data */
2471                 if (swapped)
2472                         mem_bswap_64((char *)array, sz);
2473                 data->aux_sample.size = sz;
2474                 data->aux_sample.data = (char *)array;
2475                 array = (void *)array + sz;
2476         }
2477
2478         return 0;
2479 }
2480
2481 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2482                                   u64 *timestamp)
2483 {
2484         u64 type = evsel->core.attr.sample_type;
2485         const __u64 *array;
2486
2487         if (!(type & PERF_SAMPLE_TIME))
2488                 return -1;
2489
2490         if (event->header.type != PERF_RECORD_SAMPLE) {
2491                 struct perf_sample data = {
2492                         .time = -1ULL,
2493                 };
2494
2495                 if (!evsel->core.attr.sample_id_all)
2496                         return -1;
2497                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2498                         return -1;
2499
2500                 *timestamp = data.time;
2501                 return 0;
2502         }
2503
2504         array = event->sample.array;
2505
2506         if (perf_event__check_size(event, evsel->sample_size))
2507                 return -EFAULT;
2508
2509         if (type & PERF_SAMPLE_IDENTIFIER)
2510                 array++;
2511
2512         if (type & PERF_SAMPLE_IP)
2513                 array++;
2514
2515         if (type & PERF_SAMPLE_TID)
2516                 array++;
2517
2518         if (type & PERF_SAMPLE_TIME)
2519                 *timestamp = *array;
2520
2521         return 0;
2522 }
2523
2524 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2525 {
2526         return tep_find_field(evsel->tp_format, name);
2527 }
2528
2529 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2530 {
2531         struct tep_format_field *field = evsel__field(evsel, name);
2532         int offset;
2533
2534         if (!field)
2535                 return NULL;
2536
2537         offset = field->offset;
2538
2539         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2540                 offset = *(int *)(sample->raw_data + field->offset);
2541                 offset &= 0xffff;
2542         }
2543
2544         return sample->raw_data + offset;
2545 }
2546
2547 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2548                          bool needs_swap)
2549 {
2550         u64 value;
2551         void *ptr = sample->raw_data + field->offset;
2552
2553         switch (field->size) {
2554         case 1:
2555                 return *(u8 *)ptr;
2556         case 2:
2557                 value = *(u16 *)ptr;
2558                 break;
2559         case 4:
2560                 value = *(u32 *)ptr;
2561                 break;
2562         case 8:
2563                 memcpy(&value, ptr, sizeof(u64));
2564                 break;
2565         default:
2566                 return 0;
2567         }
2568
2569         if (!needs_swap)
2570                 return value;
2571
2572         switch (field->size) {
2573         case 2:
2574                 return bswap_16(value);
2575         case 4:
2576                 return bswap_32(value);
2577         case 8:
2578                 return bswap_64(value);
2579         default:
2580                 return 0;
2581         }
2582
2583         return 0;
2584 }
2585
2586 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2587 {
2588         struct tep_format_field *field = evsel__field(evsel, name);
2589
2590         if (!field)
2591                 return 0;
2592
2593         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2594 }
2595
2596 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2597 {
2598         int paranoid;
2599
2600         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2601             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2602             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2603                 /*
2604                  * If it's cycles then fall back to hrtimer based
2605                  * cpu-clock-tick sw counter, which is always available even if
2606                  * no PMU support.
2607                  *
2608                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2609                  * b0a873e).
2610                  */
2611                 scnprintf(msg, msgsize, "%s",
2612 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2613
2614                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2615                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2616
2617                 zfree(&evsel->name);
2618                 return true;
2619         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2620                    (paranoid = perf_event_paranoid()) > 1) {
2621                 const char *name = evsel__name(evsel);
2622                 char *new_name;
2623                 const char *sep = ":";
2624
2625                 /* If event has exclude user then don't exclude kernel. */
2626                 if (evsel->core.attr.exclude_user)
2627                         return false;
2628
2629                 /* Is there already the separator in the name. */
2630                 if (strchr(name, '/') ||
2631                     (strchr(name, ':') && !evsel->is_libpfm_event))
2632                         sep = "";
2633
2634                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2635                         return false;
2636
2637                 if (evsel->name)
2638                         free(evsel->name);
2639                 evsel->name = new_name;
2640                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2641                           "to fall back to excluding kernel and hypervisor "
2642                           " samples", paranoid);
2643                 evsel->core.attr.exclude_kernel = 1;
2644                 evsel->core.attr.exclude_hv     = 1;
2645
2646                 return true;
2647         }
2648
2649         return false;
2650 }
2651
2652 static bool find_process(const char *name)
2653 {
2654         size_t len = strlen(name);
2655         DIR *dir;
2656         struct dirent *d;
2657         int ret = -1;
2658
2659         dir = opendir(procfs__mountpoint());
2660         if (!dir)
2661                 return false;
2662
2663         /* Walk through the directory. */
2664         while (ret && (d = readdir(dir)) != NULL) {
2665                 char path[PATH_MAX];
2666                 char *data;
2667                 size_t size;
2668
2669                 if ((d->d_type != DT_DIR) ||
2670                      !strcmp(".", d->d_name) ||
2671                      !strcmp("..", d->d_name))
2672                         continue;
2673
2674                 scnprintf(path, sizeof(path), "%s/%s/comm",
2675                           procfs__mountpoint(), d->d_name);
2676
2677                 if (filename__read_str(path, &data, &size))
2678                         continue;
2679
2680                 ret = strncmp(name, data, len);
2681                 free(data);
2682         }
2683
2684         closedir(dir);
2685         return ret ? false : true;
2686 }
2687
2688 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2689                          int err, char *msg, size_t size)
2690 {
2691         char sbuf[STRERR_BUFSIZE];
2692         int printed = 0, enforced = 0;
2693
2694         switch (err) {
2695         case EPERM:
2696         case EACCES:
2697                 printed += scnprintf(msg + printed, size - printed,
2698                         "Access to performance monitoring and observability operations is limited.\n");
2699
2700                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2701                         if (enforced) {
2702                                 printed += scnprintf(msg + printed, size - printed,
2703                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2704                                         "monitoring and observability operations. Inspect system audit records for\n"
2705                                         "more perf_event access control information and adjusting the policy.\n");
2706                         }
2707                 }
2708
2709                 if (err == EPERM)
2710                         printed += scnprintf(msg, size,
2711                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2712
2713                 return scnprintf(msg + printed, size - printed,
2714                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2715                  "access to performance monitoring and observability operations for processes\n"
2716                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2717                  "More information can be found at 'Perf events and tool security' document:\n"
2718                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2719                  "perf_event_paranoid setting is %d:\n"
2720                  "  -1: Allow use of (almost) all events by all users\n"
2721                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2722                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2723                  ">= 1: Disallow CPU event access\n"
2724                  ">= 2: Disallow kernel profiling\n"
2725                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2726                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2727                  perf_event_paranoid());
2728         case ENOENT:
2729                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2730         case EMFILE:
2731                 return scnprintf(msg, size, "%s",
2732                          "Too many events are opened.\n"
2733                          "Probably the maximum number of open file descriptors has been reached.\n"
2734                          "Hint: Try again after reducing the number of events.\n"
2735                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2736         case ENOMEM:
2737                 if (evsel__has_callchain(evsel) &&
2738                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2739                         return scnprintf(msg, size,
2740                                          "Not enough memory to setup event with callchain.\n"
2741                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2742                                          "Hint: Current value: %d", sysctl__max_stack());
2743                 break;
2744         case ENODEV:
2745                 if (target->cpu_list)
2746                         return scnprintf(msg, size, "%s",
2747          "No such device - did you specify an out-of-range profile CPU?");
2748                 break;
2749         case EOPNOTSUPP:
2750                 if (evsel->core.attr.aux_output)
2751                         return scnprintf(msg, size,
2752         "%s: PMU Hardware doesn't support 'aux_output' feature",
2753                                          evsel__name(evsel));
2754                 if (evsel->core.attr.sample_period != 0)
2755                         return scnprintf(msg, size,
2756         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2757                                          evsel__name(evsel));
2758                 if (evsel->core.attr.precise_ip)
2759                         return scnprintf(msg, size, "%s",
2760         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2761 #if defined(__i386__) || defined(__x86_64__)
2762                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2763                         return scnprintf(msg, size, "%s",
2764         "No hardware sampling interrupt available.\n");
2765 #endif
2766                 break;
2767         case EBUSY:
2768                 if (find_process("oprofiled"))
2769                         return scnprintf(msg, size,
2770         "The PMU counters are busy/taken by another profiler.\n"
2771         "We found oprofile daemon running, please stop it and try again.");
2772                 break;
2773         case EINVAL:
2774                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
2775                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
2776                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2777                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2778                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2779                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2780                 if (perf_missing_features.clockid)
2781                         return scnprintf(msg, size, "clockid feature not supported.");
2782                 if (perf_missing_features.clockid_wrong)
2783                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2784                 if (perf_missing_features.aux_output)
2785                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2786                 break;
2787         case ENODATA:
2788                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
2789                                  "Please add an auxiliary event in front of the load latency event.");
2790         default:
2791                 break;
2792         }
2793
2794         return scnprintf(msg, size,
2795         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2796         "/bin/dmesg | grep -i perf may provide additional information.\n",
2797                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2798 }
2799
2800 struct perf_env *evsel__env(struct evsel *evsel)
2801 {
2802         if (evsel && evsel->evlist)
2803                 return evsel->evlist->env;
2804         return &perf_env;
2805 }
2806
2807 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2808 {
2809         int cpu, thread;
2810
2811         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2812                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2813                      thread++) {
2814                         int fd = FD(evsel, cpu, thread);
2815
2816                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2817                                                    cpu, thread, fd) < 0)
2818                                 return -1;
2819                 }
2820         }
2821
2822         return 0;
2823 }
2824
2825 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2826 {
2827         struct perf_cpu_map *cpus = evsel->core.cpus;
2828         struct perf_thread_map *threads = evsel->core.threads;
2829
2830         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2831                 return -ENOMEM;
2832
2833         return store_evsel_ids(evsel, evlist);
2834 }
2835
2836 void evsel__zero_per_pkg(struct evsel *evsel)
2837 {
2838         struct hashmap_entry *cur;
2839         size_t bkt;
2840
2841         if (evsel->per_pkg_mask) {
2842                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
2843                         free((char *)cur->key);
2844
2845                 hashmap__clear(evsel->per_pkg_mask);
2846         }
2847 }
2848
2849 bool evsel__is_hybrid(struct evsel *evsel)
2850 {
2851         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
2852 }