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->unit        = "";
 245         evsel->scale       = 1.0;
 246         evsel->max_events  = ULONG_MAX;
 247         evsel->evlist      = NULL;
 248         evsel->bpf_obj     = NULL;
 249         evsel->bpf_fd      = -1;
 250         INIT_LIST_HEAD(&evsel->config_terms);
 251         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 252         perf_evsel__object.init(evsel);
 253         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 254         evsel__calc_id_pos(evsel);
 255         evsel->cmdline_group_boundary = false;
 256         evsel->metric_expr   = NULL;
 257         evsel->metric_name   = NULL;
 258         evsel->metric_events = NULL;
 259         evsel->per_pkg_mask  = NULL;
 260         evsel->collect_stat  = false;
 261         evsel->pmu_name      = NULL;
 262 }
 263
 264 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 265 {
 266         struct evsel *evsel = zalloc(perf_evsel__object.size);
 267
 268         if (!evsel)
 269                 return NULL;
 270         evsel__init(evsel, attr, idx);
 271
 272         if (evsel__is_bpf_output(evsel)) {
 273                 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 274                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 275                 evsel->core.attr.sample_period = 1;
 276         }
 277
 278         if (evsel__is_clock(evsel)) {
 279                 /*
 280                  * The evsel->unit points to static alias->unit
 281                  * so it's ok to use static string in here.
 282                  */
 283                 static const char *unit = "msec";
 284
 285                 evsel->unit = unit;
 286                 evsel->scale = 1e-6;
 287         }
 288
 289         return evsel;
 290 }
 291
 292 static bool perf_event_can_profile_kernel(void)
 293 {
 294         return perf_event_paranoid_check(1);
 295 }
 296
 297 struct evsel *evsel__new_cycles(bool precise, __u32 type, __u64 config)
 298 {
 299         struct perf_event_attr attr = {
 300                 .type   = type,
 301                 .config = config,
 302                 .exclude_kernel = !perf_event_can_profile_kernel(),
 303         };
 304         struct evsel *evsel;
 305
 306         event_attr_init(&attr);
 307
 308         if (!precise)
 309                 goto new_event;
 310
 311         /*
 312          * Now let the usual logic to set up the perf_event_attr defaults
 313          * to kick in when we return and before perf_evsel__open() is called.
 314          */
 315 new_event:
 316         evsel = evsel__new(&attr);
 317         if (evsel == NULL)
 318                 goto out;
 319
 320         evsel->precise_max = true;
 321
 322         /* use asprintf() because free(evsel) assumes name is allocated */
 323         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 324                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 325                      attr.exclude_kernel ? "u" : "",
 326                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 327                 goto error_free;
 328 out:
 329         return evsel;
 330 error_free:
 331         evsel__delete(evsel);
 332         evsel = NULL;
 333         goto out;
 334 }
 335
 336 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 337 {
 338         struct evsel_config_term *pos, *tmp;
 339
 340         list_for_each_entry(pos, &src->config_terms, list) {
 341                 tmp = malloc(sizeof(*tmp));
 342                 if (tmp == NULL)
 343                         return -ENOMEM;
 344
 345                 *tmp = *pos;
 346                 if (tmp->free_str) {
 347                         tmp->val.str = strdup(pos->val.str);
 348                         if (tmp->val.str == NULL) {
 349                                 free(tmp);
 350                                 return -ENOMEM;
 351                         }
 352                 }
 353                 list_add_tail(&tmp->list, &dst->config_terms);
 354         }
 355         return 0;
 356 }
 357
 358 /**
 359  * evsel__clone - create a new evsel copied from @orig
 360  * @orig: original evsel
 361  *
 362  * The assumption is that @orig is not configured nor opened yet.
 363  * So we only care about the attributes that can be set while it's parsed.
 364  */
 365 struct evsel *evsel__clone(struct evsel *orig)
 366 {
 367         struct evsel *evsel;
 368
 369         BUG_ON(orig->core.fd);
 370         BUG_ON(orig->counts);
 371         BUG_ON(orig->priv);
 372         BUG_ON(orig->per_pkg_mask);
 373
 374         /* cannot handle BPF objects for now */
 375         if (orig->bpf_obj)
 376                 return NULL;
 377
 378         evsel = evsel__new(&orig->core.attr);
 379         if (evsel == NULL)
 380                 return NULL;
 381
 382         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 383         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 384         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 385         evsel->core.nr_members = orig->core.nr_members;
 386         evsel->core.system_wide = orig->core.system_wide;
 387
 388         if (orig->name) {
 389                 evsel->name = strdup(orig->name);
 390                 if (evsel->name == NULL)
 391                         goto out_err;
 392         }
 393         if (orig->group_name) {
 394                 evsel->group_name = strdup(orig->group_name);
 395                 if (evsel->group_name == NULL)
 396                         goto out_err;
 397         }
 398         if (orig->pmu_name) {
 399                 evsel->pmu_name = strdup(orig->pmu_name);
 400                 if (evsel->pmu_name == NULL)
 401                         goto out_err;
 402         }
 403         if (orig->filter) {
 404                 evsel->filter = strdup(orig->filter);
 405                 if (evsel->filter == NULL)
 406                         goto out_err;
 407         }
 408         evsel->cgrp = cgroup__get(orig->cgrp);
 409         evsel->tp_format = orig->tp_format;
 410         evsel->handler = orig->handler;
 411         evsel->core.leader = orig->core.leader;
 412
 413         evsel->max_events = orig->max_events;
 414         evsel->tool_event = orig->tool_event;
 415         evsel->unit = orig->unit;
 416         evsel->scale = orig->scale;
 417         evsel->snapshot = orig->snapshot;
 418         evsel->per_pkg = orig->per_pkg;
 419         evsel->percore = orig->percore;
 420         evsel->precise_max = orig->precise_max;
 421         evsel->use_uncore_alias = orig->use_uncore_alias;
 422         evsel->is_libpfm_event = orig->is_libpfm_event;
 423
 424         evsel->exclude_GH = orig->exclude_GH;
 425         evsel->sample_read = orig->sample_read;
 426         evsel->auto_merge_stats = orig->auto_merge_stats;
 427         evsel->collect_stat = orig->collect_stat;
 428         evsel->weak_group = orig->weak_group;
 429         evsel->use_config_name = orig->use_config_name;
 430
 431         if (evsel__copy_config_terms(evsel, orig) < 0)
 432                 goto out_err;
 433
 434         return evsel;
 435
 436 out_err:
 437         evsel__delete(evsel);
 438         return NULL;
 439 }
 440
 441 /*
 442  * Returns pointer with encoded error via <linux/err.h> interface.
 443  */
 444 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 445 {
 446         struct evsel *evsel = zalloc(perf_evsel__object.size);
 447         int err = -ENOMEM;
 448
 449         if (evsel == NULL) {
 450                 goto out_err;
 451         } else {
 452                 struct perf_event_attr attr = {
 453                         .type          = PERF_TYPE_TRACEPOINT,
 454                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 455                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 456                 };
 457
 458                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 459                         goto out_free;
 460
 461                 evsel->tp_format = trace_event__tp_format(sys, name);
 462                 if (IS_ERR(evsel->tp_format)) {
 463                         err = PTR_ERR(evsel->tp_format);
 464                         goto out_free;
 465                 }
 466
 467                 event_attr_init(&attr);
 468                 attr.config = evsel->tp_format->id;
 469                 attr.sample_period = 1;
 470                 evsel__init(evsel, &attr, idx);
 471         }
 472
 473         return evsel;
 474
 475 out_free:
 476         zfree(&evsel->name);
 477         free(evsel);
 478 out_err:
 479         return ERR_PTR(err);
 480 }
 481
 482 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
 483         "cycles",
 484         "instructions",
 485         "cache-references",
 486         "cache-misses",
 487         "branches",
 488         "branch-misses",
 489         "bus-cycles",
 490         "stalled-cycles-frontend",
 491         "stalled-cycles-backend",
 492         "ref-cycles",
 493 };
 494
 495 char *evsel__bpf_counter_events;
 496
 497 bool evsel__match_bpf_counter_events(const char *name)
 498 {
 499         int name_len;
 500         bool match;
 501         char *ptr;
 502
 503         if (!evsel__bpf_counter_events)
 504                 return false;
 505
 506         ptr = strstr(evsel__bpf_counter_events, name);
 507         name_len = strlen(name);
 508
 509         /* check name matches a full token in evsel__bpf_counter_events */
 510         match = (ptr != NULL) &&
 511                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 512                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 513
 514         return match;
 515 }
 516
 517 static const char *__evsel__hw_name(u64 config)
 518 {
 519         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 520                 return evsel__hw_names[config];
 521
 522         return "unknown-hardware";
 523 }
 524
 525 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 526 {
 527         int colon = 0, r = 0;
 528         struct perf_event_attr *attr = &evsel->core.attr;
 529         bool exclude_guest_default = false;
 530
 531 #define MOD_PRINT(context, mod) do {                                    \
 532                 if (!attr->exclude_##context) {                         \
 533                         if (!colon) colon = ++r;                        \
 534                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 535                 } } while(0)
 536
 537         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 538                 MOD_PRINT(kernel, 'k');
 539                 MOD_PRINT(user, 'u');
 540                 MOD_PRINT(hv, 'h');
 541                 exclude_guest_default = true;
 542         }
 543
 544         if (attr->precise_ip) {
 545                 if (!colon)
 546                         colon = ++r;
 547                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 548                 exclude_guest_default = true;
 549         }
 550
 551         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 552                 MOD_PRINT(host, 'H');
 553                 MOD_PRINT(guest, 'G');
 554         }
 555 #undef MOD_PRINT
 556         if (colon)
 557                 bf[colon - 1] = ':';
 558         return r;
 559 }
 560
 561 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 562 {
 563         int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 564         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 565 }
 566
 567 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
 568         "cpu-clock",
 569         "task-clock",
 570         "page-faults",
 571         "context-switches",
 572         "cpu-migrations",
 573         "minor-faults",
 574         "major-faults",
 575         "alignment-faults",
 576         "emulation-faults",
 577         "dummy",
 578 };
 579
 580 static const char *__evsel__sw_name(u64 config)
 581 {
 582         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 583                 return evsel__sw_names[config];
 584         return "unknown-software";
 585 }
 586
 587 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 588 {
 589         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 590         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 591 }
 592
 593 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 594 {
 595         int r;
 596
 597         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 598
 599         if (type & HW_BREAKPOINT_R)
 600                 r += scnprintf(bf + r, size - r, "r");
 601
 602         if (type & HW_BREAKPOINT_W)
 603                 r += scnprintf(bf + r, size - r, "w");
 604
 605         if (type & HW_BREAKPOINT_X)
 606                 r += scnprintf(bf + r, size - r, "x");
 607
 608         return r;
 609 }
 610
 611 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 612 {
 613         struct perf_event_attr *attr = &evsel->core.attr;
 614         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 615         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 616 }
 617
 618 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 619  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 620  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 621  { "LLC",       "L2",                                                   },
 622  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 623  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 624  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 625  { "node",                                                              },
 626 };
 627
 628 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 629  { "load",      "loads",        "read",                                 },
 630  { "store",     "stores",       "write",                                },
 631  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 632 };
 633
 634 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 635  { "refs",      "Reference",    "ops",          "access",               },
 636  { "misses",    "miss",                                                 },
 637 };
 638
 639 #define C(x)            PERF_COUNT_HW_CACHE_##x
 640 #define CACHE_READ      (1 << C(OP_READ))
 641 #define CACHE_WRITE     (1 << C(OP_WRITE))
 642 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 643 #define COP(x)          (1 << x)
 644
 645 /*
 646  * cache operation stat
 647  * L1I : Read and prefetch only
 648  * ITLB and BPU : Read-only
 649  */
 650 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
 651  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 652  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 653  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 654  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 655  [C(ITLB)]      = (CACHE_READ),
 656  [C(BPU)]       = (CACHE_READ),
 657  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 658 };
 659
 660 bool evsel__is_cache_op_valid(u8 type, u8 op)
 661 {
 662         if (evsel__hw_cache_stat[type] & COP(op))
 663                 return true;    /* valid */
 664         else
 665                 return false;   /* invalid */
 666 }
 667
 668 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 669 {
 670         if (result) {
 671                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 672                                  evsel__hw_cache_op[op][0],
 673                                  evsel__hw_cache_result[result][0]);
 674         }
 675
 676         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 677                          evsel__hw_cache_op[op][1]);
 678 }
 679
 680 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 681 {
 682         u8 op, result, type = (config >>  0) & 0xff;
 683         const char *err = "unknown-ext-hardware-cache-type";
 684
 685         if (type >= PERF_COUNT_HW_CACHE_MAX)
 686                 goto out_err;
 687
 688         op = (config >>  8) & 0xff;
 689         err = "unknown-ext-hardware-cache-op";
 690         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 691                 goto out_err;
 692
 693         result = (config >> 16) & 0xff;
 694         err = "unknown-ext-hardware-cache-result";
 695         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 696                 goto out_err;
 697
 698         err = "invalid-cache";
 699         if (!evsel__is_cache_op_valid(type, op))
 700                 goto out_err;
 701
 702         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 703 out_err:
 704         return scnprintf(bf, size, "%s", err);
 705 }
 706
 707 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 708 {
 709         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 710         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 711 }
 712
 713 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 714 {
 715         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 716         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 717 }
 718
 719 static int evsel__tool_name(char *bf, size_t size)
 720 {
 721         int ret = scnprintf(bf, size, "duration_time");
 722         return ret;
 723 }
 724
 725 const char *evsel__name(struct evsel *evsel)
 726 {
 727         char bf[128];
 728
 729         if (!evsel)
 730                 goto out_unknown;
 731
 732         if (evsel->name)
 733                 return evsel->name;
 734
 735         switch (evsel->core.attr.type) {
 736         case PERF_TYPE_RAW:
 737                 evsel__raw_name(evsel, bf, sizeof(bf));
 738                 break;
 739
 740         case PERF_TYPE_HARDWARE:
 741                 evsel__hw_name(evsel, bf, sizeof(bf));
 742                 break;
 743
 744         case PERF_TYPE_HW_CACHE:
 745                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 746                 break;
 747
 748         case PERF_TYPE_SOFTWARE:
 749                 if (evsel->tool_event)
 750                         evsel__tool_name(bf, sizeof(bf));
 751                 else
 752                         evsel__sw_name(evsel, bf, sizeof(bf));
 753                 break;
 754
 755         case PERF_TYPE_TRACEPOINT:
 756                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 757                 break;
 758
 759         case PERF_TYPE_BREAKPOINT:
 760                 evsel__bp_name(evsel, bf, sizeof(bf));
 761                 break;
 762
 763         default:
 764                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 765                           evsel->core.attr.type);
 766                 break;
 767         }
 768
 769         evsel->name = strdup(bf);
 770
 771         if (evsel->name)
 772                 return evsel->name;
 773 out_unknown:
 774         return "unknown";
 775 }
 776
 777 const char *evsel__group_name(struct evsel *evsel)
 778 {
 779         return evsel->group_name ?: "anon group";
 780 }
 781
 782 /*
 783  * Returns the group details for the specified leader,
 784  * with following rules.
 785  *
 786  *  For record -e '{cycles,instructions}'
 787  *    'anon group { cycles:u, instructions:u }'
 788  *
 789  *  For record -e 'cycles,instructions' and report --group
 790  *    'cycles:u, instructions:u'
 791  */
 792 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 793 {
 794         int ret = 0;
 795         struct evsel *pos;
 796         const char *group_name = evsel__group_name(evsel);
 797
 798         if (!evsel->forced_leader)
 799                 ret = scnprintf(buf, size, "%s { ", group_name);
 800
 801         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 802
 803         for_each_group_member(pos, evsel)
 804                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 805
 806         if (!evsel->forced_leader)
 807                 ret += scnprintf(buf + ret, size - ret, " }");
 808
 809         return ret;
 810 }
 811
 812 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 813                                       struct callchain_param *param)
 814 {
 815         bool function = evsel__is_function_event(evsel);
 816         struct perf_event_attr *attr = &evsel->core.attr;
 817
 818         evsel__set_sample_bit(evsel, CALLCHAIN);
 819
 820         attr->sample_max_stack = param->max_stack;
 821
 822         if (opts->kernel_callchains)
 823                 attr->exclude_callchain_user = 1;
 824         if (opts->user_callchains)
 825                 attr->exclude_callchain_kernel = 1;
 826         if (param->record_mode == CALLCHAIN_LBR) {
 827                 if (!opts->branch_stack) {
 828                         if (attr->exclude_user) {
 829                                 pr_warning("LBR callstack option is only available "
 830                                            "to get user callchain information. "
 831                                            "Falling back to framepointers.\n");
 832                         } else {
 833                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 834                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 835                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 836                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 837                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 838                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 839                         }
 840                 } else
 841                          pr_warning("Cannot use LBR callstack with branch stack. "
 842                                     "Falling back to framepointers.\n");
 843         }
 844
 845         if (param->record_mode == CALLCHAIN_DWARF) {
 846                 if (!function) {
 847                         evsel__set_sample_bit(evsel, REGS_USER);
 848                         evsel__set_sample_bit(evsel, STACK_USER);
 849                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 850                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 851                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 852                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 853                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 854                         } else {
 855                                 attr->sample_regs_user |= PERF_REGS_MASK;
 856                         }
 857                         attr->sample_stack_user = param->dump_size;
 858                         attr->exclude_callchain_user = 1;
 859                 } else {
 860                         pr_info("Cannot use DWARF unwind for function trace event,"
 861                                 " falling back to framepointers.\n");
 862                 }
 863         }
 864
 865         if (function) {
 866                 pr_info("Disabling user space callchains for function trace event.\n");
 867                 attr->exclude_callchain_user = 1;
 868         }
 869 }
 870
 871 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 872                              struct callchain_param *param)
 873 {
 874         if (param->enabled)
 875                 return __evsel__config_callchain(evsel, opts, param);
 876 }
 877
 878 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 879 {
 880         struct perf_event_attr *attr = &evsel->core.attr;
 881
 882         evsel__reset_sample_bit(evsel, CALLCHAIN);
 883         if (param->record_mode == CALLCHAIN_LBR) {
 884                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 885                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 886                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 887                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 888         }
 889         if (param->record_mode == CALLCHAIN_DWARF) {
 890                 evsel__reset_sample_bit(evsel, REGS_USER);
 891                 evsel__reset_sample_bit(evsel, STACK_USER);
 892         }
 893 }
 894
 895 static void evsel__apply_config_terms(struct evsel *evsel,
 896                                       struct record_opts *opts, bool track)
 897 {
 898         struct evsel_config_term *term;
 899         struct list_head *config_terms = &evsel->config_terms;
 900         struct perf_event_attr *attr = &evsel->core.attr;
 901         /* callgraph default */
 902         struct callchain_param param = {
 903                 .record_mode = callchain_param.record_mode,
 904         };
 905         u32 dump_size = 0;
 906         int max_stack = 0;
 907         const char *callgraph_buf = NULL;
 908
 909         list_for_each_entry(term, config_terms, list) {
 910                 switch (term->type) {
 911                 case EVSEL__CONFIG_TERM_PERIOD:
 912                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 913                                 attr->sample_period = term->val.period;
 914                                 attr->freq = 0;
 915                                 evsel__reset_sample_bit(evsel, PERIOD);
 916                         }
 917                         break;
 918                 case EVSEL__CONFIG_TERM_FREQ:
 919                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 920                                 attr->sample_freq = term->val.freq;
 921                                 attr->freq = 1;
 922                                 evsel__set_sample_bit(evsel, PERIOD);
 923                         }
 924                         break;
 925                 case EVSEL__CONFIG_TERM_TIME:
 926                         if (term->val.time)
 927                                 evsel__set_sample_bit(evsel, TIME);
 928                         else
 929                                 evsel__reset_sample_bit(evsel, TIME);
 930                         break;
 931                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 932                         callgraph_buf = term->val.str;
 933                         break;
 934                 case EVSEL__CONFIG_TERM_BRANCH:
 935                         if (term->val.str && strcmp(term->val.str, "no")) {
 936                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 937                                 parse_branch_str(term->val.str,
 938                                                  &attr->branch_sample_type);
 939                         } else
 940                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 941                         break;
 942                 case EVSEL__CONFIG_TERM_STACK_USER:
 943                         dump_size = term->val.stack_user;
 944                         break;
 945                 case EVSEL__CONFIG_TERM_MAX_STACK:
 946                         max_stack = term->val.max_stack;
 947                         break;
 948                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 949                         evsel->max_events = term->val.max_events;
 950                         break;
 951                 case EVSEL__CONFIG_TERM_INHERIT:
 952                         /*
 953                          * attr->inherit should has already been set by
 954                          * evsel__config. If user explicitly set
 955                          * inherit using config terms, override global
 956                          * opt->no_inherit setting.
 957                          */
 958                         attr->inherit = term->val.inherit ? 1 : 0;
 959                         break;
 960                 case EVSEL__CONFIG_TERM_OVERWRITE:
 961                         attr->write_backward = term->val.overwrite ? 1 : 0;
 962                         break;
 963                 case EVSEL__CONFIG_TERM_DRV_CFG:
 964                         break;
 965                 case EVSEL__CONFIG_TERM_PERCORE:
 966                         break;
 967                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
 968                         attr->aux_output = term->val.aux_output ? 1 : 0;
 969                         break;
 970                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 971                         /* Already applied by auxtrace */
 972                         break;
 973                 case EVSEL__CONFIG_TERM_CFG_CHG:
 974                         break;
 975                 default:
 976                         break;
 977                 }
 978         }
 979
 980         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 981         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 982                 bool sample_address = false;
 983
 984                 if (max_stack) {
 985                         param.max_stack = max_stack;
 986                         if (callgraph_buf == NULL)
 987                                 callgraph_buf = "fp";
 988                 }
 989
 990                 /* parse callgraph parameters */
 991                 if (callgraph_buf != NULL) {
 992                         if (!strcmp(callgraph_buf, "no")) {
 993                                 param.enabled = false;
 994                                 param.record_mode = CALLCHAIN_NONE;
 995                         } else {
 996                                 param.enabled = true;
 997                                 if (parse_callchain_record(callgraph_buf, &param)) {
 998                                         pr_err("per-event callgraph setting for %s failed. "
 999                                                "Apply callgraph global setting for it\n",
1000                                                evsel->name);
1001                                         return;
1002                                 }
1003                                 if (param.record_mode == CALLCHAIN_DWARF)
1004                                         sample_address = true;
1005                         }
1006                 }
1007                 if (dump_size > 0) {
1008                         dump_size = round_up(dump_size, sizeof(u64));
1009                         param.dump_size = dump_size;
1010                 }
1011
1012                 /* If global callgraph set, clear it */
1013                 if (callchain_param.enabled)
1014                         evsel__reset_callgraph(evsel, &callchain_param);
1015
1016                 /* set perf-event callgraph */
1017                 if (param.enabled) {
1018                         if (sample_address) {
1019                                 evsel__set_sample_bit(evsel, ADDR);
1020                                 evsel__set_sample_bit(evsel, DATA_SRC);
1021                                 evsel->core.attr.mmap_data = track;
1022                         }
1023                         evsel__config_callchain(evsel, opts, &param);
1024                 }
1025         }
1026 }
1027
1028 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1029 {
1030         struct evsel_config_term *term, *found_term = NULL;
1031
1032         list_for_each_entry(term, &evsel->config_terms, list) {
1033                 if (term->type == type)
1034                         found_term = term;
1035         }
1036
1037         return found_term;
1038 }
1039
1040 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1041 {
1042         evsel__set_sample_bit(evsel, WEIGHT);
1043 }
1044
1045 /*
1046  * The enable_on_exec/disabled value strategy:
1047  *
1048  *  1) For any type of traced program:
1049  *    - all independent events and group leaders are disabled
1050  *    - all group members are enabled
1051  *
1052  *     Group members are ruled by group leaders. They need to
1053  *     be enabled, because the group scheduling relies on that.
1054  *
1055  *  2) For traced programs executed by perf:
1056  *     - all independent events and group leaders have
1057  *       enable_on_exec set
1058  *     - we don't specifically enable or disable any event during
1059  *       the record command
1060  *
1061  *     Independent events and group leaders are initially disabled
1062  *     and get enabled by exec. Group members are ruled by group
1063  *     leaders as stated in 1).
1064  *
1065  *  3) For traced programs attached by perf (pid/tid):
1066  *     - we specifically enable or disable all events during
1067  *       the record command
1068  *
1069  *     When attaching events to already running traced we
1070  *     enable/disable events specifically, as there's no
1071  *     initial traced exec call.
1072  */
1073 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1074                    struct callchain_param *callchain)
1075 {
1076         struct evsel *leader = evsel__leader(evsel);
1077         struct perf_event_attr *attr = &evsel->core.attr;
1078         int track = evsel->tracking;
1079         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1080
1081         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1082         attr->inherit       = !opts->no_inherit;
1083         attr->write_backward = opts->overwrite ? 1 : 0;
1084
1085         evsel__set_sample_bit(evsel, IP);
1086         evsel__set_sample_bit(evsel, TID);
1087
1088         if (evsel->sample_read) {
1089                 evsel__set_sample_bit(evsel, READ);
1090
1091                 /*
1092                  * We need ID even in case of single event, because
1093                  * PERF_SAMPLE_READ process ID specific data.
1094                  */
1095                 evsel__set_sample_id(evsel, false);
1096
1097                 /*
1098                  * Apply group format only if we belong to group
1099                  * with more than one members.
1100                  */
1101                 if (leader->core.nr_members > 1) {
1102                         attr->read_format |= PERF_FORMAT_GROUP;
1103                         attr->inherit = 0;
1104                 }
1105         }
1106
1107         /*
1108          * We default some events to have a default interval. But keep
1109          * it a weak assumption overridable by the user.
1110          */
1111         if (!attr->sample_period) {
1112                 if (opts->freq) {
1113                         attr->freq              = 1;
1114                         attr->sample_freq       = opts->freq;
1115                 } else {
1116                         attr->sample_period = opts->default_interval;
1117                 }
1118         }
1119         /*
1120          * If attr->freq was set (here or earlier), ask for period
1121          * to be sampled.
1122          */
1123         if (attr->freq)
1124                 evsel__set_sample_bit(evsel, PERIOD);
1125
1126         if (opts->no_samples)
1127                 attr->sample_freq = 0;
1128
1129         if (opts->inherit_stat) {
1130                 evsel->core.attr.read_format |=
1131                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1132                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1133                         PERF_FORMAT_ID;
1134                 attr->inherit_stat = 1;
1135         }
1136
1137         if (opts->sample_address) {
1138                 evsel__set_sample_bit(evsel, ADDR);
1139                 attr->mmap_data = track;
1140         }
1141
1142         /*
1143          * We don't allow user space callchains for  function trace
1144          * event, due to issues with page faults while tracing page
1145          * fault handler and its overall trickiness nature.
1146          */
1147         if (evsel__is_function_event(evsel))
1148                 evsel->core.attr.exclude_callchain_user = 1;
1149
1150         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1151                 evsel__config_callchain(evsel, opts, callchain);
1152
1153         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1154             !evsel__is_dummy_event(evsel)) {
1155                 attr->sample_regs_intr = opts->sample_intr_regs;
1156                 evsel__set_sample_bit(evsel, REGS_INTR);
1157         }
1158
1159         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1160             !evsel__is_dummy_event(evsel)) {
1161                 attr->sample_regs_user |= opts->sample_user_regs;
1162                 evsel__set_sample_bit(evsel, REGS_USER);
1163         }
1164
1165         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1166                 evsel__set_sample_bit(evsel, CPU);
1167
1168         /*
1169          * When the user explicitly disabled time don't force it here.
1170          */
1171         if (opts->sample_time &&
1172             (!perf_missing_features.sample_id_all &&
1173             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1174              opts->sample_time_set)))
1175                 evsel__set_sample_bit(evsel, TIME);
1176
1177         if (opts->raw_samples && !evsel->no_aux_samples) {
1178                 evsel__set_sample_bit(evsel, TIME);
1179                 evsel__set_sample_bit(evsel, RAW);
1180                 evsel__set_sample_bit(evsel, CPU);
1181         }
1182
1183         if (opts->sample_address)
1184                 evsel__set_sample_bit(evsel, DATA_SRC);
1185
1186         if (opts->sample_phys_addr)
1187                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1188
1189         if (opts->no_buffering) {
1190                 attr->watermark = 0;
1191                 attr->wakeup_events = 1;
1192         }
1193         if (opts->branch_stack && !evsel->no_aux_samples) {
1194                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1195                 attr->branch_sample_type = opts->branch_stack;
1196         }
1197
1198         if (opts->sample_weight)
1199                 arch_evsel__set_sample_weight(evsel);
1200
1201         attr->task     = track;
1202         attr->mmap     = track;
1203         attr->mmap2    = track && !perf_missing_features.mmap2;
1204         attr->comm     = track;
1205         attr->build_id = track && opts->build_id;
1206
1207         /*
1208          * ksymbol is tracked separately with text poke because it needs to be
1209          * system wide and enabled immediately.
1210          */
1211         if (!opts->text_poke)
1212                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1213         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1214
1215         if (opts->record_namespaces)
1216                 attr->namespaces  = track;
1217
1218         if (opts->record_cgroup) {
1219                 attr->cgroup = track && !perf_missing_features.cgroup;
1220                 evsel__set_sample_bit(evsel, CGROUP);
1221         }
1222
1223         if (opts->sample_data_page_size)
1224                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1225
1226         if (opts->sample_code_page_size)
1227                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1228
1229         if (opts->record_switch_events)
1230                 attr->context_switch = track;
1231
1232         if (opts->sample_transaction)
1233                 evsel__set_sample_bit(evsel, TRANSACTION);
1234
1235         if (opts->running_time) {
1236                 evsel->core.attr.read_format |=
1237                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1238                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1239         }
1240
1241         /*
1242          * XXX see the function comment above
1243          *
1244          * Disabling only independent events or group leaders,
1245          * keeping group members enabled.
1246          */
1247         if (evsel__is_group_leader(evsel))
1248                 attr->disabled = 1;
1249
1250         /*
1251          * Setting enable_on_exec for independent events and
1252          * group leaders for traced executed by perf.
1253          */
1254         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1255             !opts->initial_delay)
1256                 attr->enable_on_exec = 1;
1257
1258         if (evsel->immediate) {
1259                 attr->disabled = 0;
1260                 attr->enable_on_exec = 0;
1261         }
1262
1263         clockid = opts->clockid;
1264         if (opts->use_clockid) {
1265                 attr->use_clockid = 1;
1266                 attr->clockid = opts->clockid;
1267         }
1268
1269         if (evsel->precise_max)
1270                 attr->precise_ip = 3;
1271
1272         if (opts->all_user) {
1273                 attr->exclude_kernel = 1;
1274                 attr->exclude_user   = 0;
1275         }
1276
1277         if (opts->all_kernel) {
1278                 attr->exclude_kernel = 0;
1279                 attr->exclude_user   = 1;
1280         }
1281
1282         if (evsel->core.own_cpus || evsel->unit)
1283                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1284
1285         /*
1286          * Apply event specific term settings,
1287          * it overloads any global configuration.
1288          */
1289         evsel__apply_config_terms(evsel, opts, track);
1290
1291         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1292
1293         /* The --period option takes the precedence. */
1294         if (opts->period_set) {
1295                 if (opts->period)
1296                         evsel__set_sample_bit(evsel, PERIOD);
1297                 else
1298                         evsel__reset_sample_bit(evsel, PERIOD);
1299         }
1300
1301         /*
1302          * A dummy event never triggers any actual counter and therefore
1303          * cannot be used with branch_stack.
1304          *
1305          * For initial_delay, a dummy event is added implicitly.
1306          * The software event will trigger -EOPNOTSUPP error out,
1307          * if BRANCH_STACK bit is set.
1308          */
1309         if (evsel__is_dummy_event(evsel))
1310                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1311 }
1312
1313 int evsel__set_filter(struct evsel *evsel, const char *filter)
1314 {
1315         char *new_filter = strdup(filter);
1316
1317         if (new_filter != NULL) {
1318                 free(evsel->filter);
1319                 evsel->filter = new_filter;
1320                 return 0;
1321         }
1322
1323         return -1;
1324 }
1325
1326 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1327 {
1328         char *new_filter;
1329
1330         if (evsel->filter == NULL)
1331                 return evsel__set_filter(evsel, filter);
1332
1333         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1334                 free(evsel->filter);
1335                 evsel->filter = new_filter;
1336                 return 0;
1337         }
1338
1339         return -1;
1340 }
1341
1342 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1343 {
1344         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1345 }
1346
1347 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1348 {
1349         return evsel__append_filter(evsel, "%s,%s", filter);
1350 }
1351
1352 /* Caller has to clear disabled after going through all CPUs. */
1353 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1354 {
1355         return perf_evsel__enable_cpu(&evsel->core, cpu);
1356 }
1357
1358 int evsel__enable(struct evsel *evsel)
1359 {
1360         int err = perf_evsel__enable(&evsel->core);
1361
1362         if (!err)
1363                 evsel->disabled = false;
1364         return err;
1365 }
1366
1367 /* Caller has to set disabled after going through all CPUs. */
1368 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1369 {
1370         return perf_evsel__disable_cpu(&evsel->core, cpu);
1371 }
1372
1373 int evsel__disable(struct evsel *evsel)
1374 {
1375         int err = perf_evsel__disable(&evsel->core);
1376         /*
1377          * We mark it disabled here so that tools that disable a event can
1378          * ignore events after they disable it. I.e. the ring buffer may have
1379          * already a few more events queued up before the kernel got the stop
1380          * request.
1381          */
1382         if (!err)
1383                 evsel->disabled = true;
1384
1385         return err;
1386 }
1387
1388 static void evsel__free_config_terms(struct evsel *evsel)
1389 {
1390         struct evsel_config_term *term, *h;
1391
1392         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1393                 list_del_init(&term->list);
1394                 if (term->free_str)
1395                         zfree(&term->val.str);
1396                 free(term);
1397         }
1398 }
1399
1400 void evsel__exit(struct evsel *evsel)
1401 {
1402         assert(list_empty(&evsel->core.node));
1403         assert(evsel->evlist == NULL);
1404         bpf_counter__destroy(evsel);
1405         evsel__free_counts(evsel);
1406         perf_evsel__free_fd(&evsel->core);
1407         perf_evsel__free_id(&evsel->core);
1408         evsel__free_config_terms(evsel);
1409         cgroup__put(evsel->cgrp);
1410         perf_cpu_map__put(evsel->core.cpus);
1411         perf_cpu_map__put(evsel->core.own_cpus);
1412         perf_thread_map__put(evsel->core.threads);
1413         zfree(&evsel->group_name);
1414         zfree(&evsel->name);
1415         zfree(&evsel->pmu_name);
1416         evsel__zero_per_pkg(evsel);
1417         hashmap__free(evsel->per_pkg_mask);
1418         evsel->per_pkg_mask = NULL;
1419         zfree(&evsel->metric_events);
1420         perf_evsel__object.fini(evsel);
1421 }
1422
1423 void evsel__delete(struct evsel *evsel)
1424 {
1425         evsel__exit(evsel);
1426         free(evsel);
1427 }
1428
1429 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1430                            struct perf_counts_values *count)
1431 {
1432         struct perf_counts_values tmp;
1433
1434         if (!evsel->prev_raw_counts)
1435                 return;
1436
1437         if (cpu == -1) {
1438                 tmp = evsel->prev_raw_counts->aggr;
1439                 evsel->prev_raw_counts->aggr = *count;
1440         } else {
1441                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1442                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1443         }
1444
1445         count->val = count->val - tmp.val;
1446         count->ena = count->ena - tmp.ena;
1447         count->run = count->run - tmp.run;
1448 }
1449
1450 void perf_counts_values__scale(struct perf_counts_values *count,
1451                                bool scale, s8 *pscaled)
1452 {
1453         s8 scaled = 0;
1454
1455         if (scale) {
1456                 if (count->run == 0) {
1457                         scaled = -1;
1458                         count->val = 0;
1459                 } else if (count->run < count->ena) {
1460                         scaled = 1;
1461                         count->val = (u64)((double) count->val * count->ena / count->run);
1462                 }
1463         }
1464
1465         if (pscaled)
1466                 *pscaled = scaled;
1467 }
1468
1469 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1470 {
1471         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1472
1473         return perf_evsel__read(&evsel->core, cpu, thread, count);
1474 }
1475
1476 static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val, u64 ena, u64 run)
1477 {
1478         struct perf_counts_values *count;
1479
1480         count = perf_counts(counter->counts, cpu, thread);
1481
1482         count->val    = val;
1483         count->ena    = ena;
1484         count->run    = run;
1485
1486         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1487 }
1488
1489 static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data)
1490 {
1491         u64 read_format = leader->core.attr.read_format;
1492         struct sample_read_value *v;
1493         u64 nr, ena = 0, run = 0, i;
1494
1495         nr = *data++;
1496
1497         if (nr != (u64) leader->core.nr_members)
1498                 return -EINVAL;
1499
1500         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1501                 ena = *data++;
1502
1503         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1504                 run = *data++;
1505
1506         v = (struct sample_read_value *) data;
1507
1508         evsel__set_count(leader, cpu, thread, v[0].value, ena, run);
1509
1510         for (i = 1; i < nr; i++) {
1511                 struct evsel *counter;
1512
1513                 counter = evlist__id2evsel(leader->evlist, v[i].id);
1514                 if (!counter)
1515                         return -EINVAL;
1516
1517                 evsel__set_count(counter, cpu, thread, v[i].value, ena, run);
1518         }
1519
1520         return 0;
1521 }
1522
1523 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1524 {
1525         struct perf_stat_evsel *ps = leader->stats;
1526         u64 read_format = leader->core.attr.read_format;
1527         int size = perf_evsel__read_size(&leader->core);
1528         u64 *data = ps->group_data;
1529
1530         if (!(read_format & PERF_FORMAT_ID))
1531                 return -EINVAL;
1532
1533         if (!evsel__is_group_leader(leader))
1534                 return -EINVAL;
1535
1536         if (!data) {
1537                 data = zalloc(size);
1538                 if (!data)
1539                         return -ENOMEM;
1540
1541                 ps->group_data = data;
1542         }
1543
1544         if (FD(leader, cpu, thread) < 0)
1545                 return -EINVAL;
1546
1547         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1548                 return -errno;
1549
1550         return evsel__process_group_data(leader, cpu, thread, data);
1551 }
1552
1553 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1554 {
1555         u64 read_format = evsel->core.attr.read_format;
1556
1557         if (read_format & PERF_FORMAT_GROUP)
1558                 return evsel__read_group(evsel, cpu, thread);
1559
1560         return evsel__read_one(evsel, cpu, thread);
1561 }
1562
1563 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1564 {
1565         struct perf_counts_values count;
1566         size_t nv = scale ? 3 : 1;
1567
1568         if (FD(evsel, cpu, thread) < 0)
1569                 return -EINVAL;
1570
1571         if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1572                 return -ENOMEM;
1573
1574         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1575                 return -errno;
1576
1577         evsel__compute_deltas(evsel, cpu, thread, &count);
1578         perf_counts_values__scale(&count, scale, NULL);
1579         *perf_counts(evsel->counts, cpu, thread) = count;
1580         return 0;
1581 }
1582
1583 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1584                                   int cpu)
1585 {
1586         int cpuid;
1587
1588         cpuid = perf_cpu_map__cpu(evsel->core.cpus, cpu);
1589         return perf_cpu_map__idx(other->core.cpus, cpuid);
1590 }
1591
1592 static int evsel__hybrid_group_cpu(struct evsel *evsel, int cpu)
1593 {
1594         struct evsel *leader = evsel__leader(evsel);
1595
1596         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1597             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1598                 return evsel__match_other_cpu(evsel, leader, cpu);
1599         }
1600
1601         return cpu;
1602 }
1603
1604 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1605 {
1606         struct evsel *leader = evsel__leader(evsel);
1607         int fd;
1608
1609         if (evsel__is_group_leader(evsel))
1610                 return -1;
1611
1612         /*
1613          * Leader must be already processed/open,
1614          * if not it's a bug.
1615          */
1616         BUG_ON(!leader->core.fd);
1617
1618         cpu = evsel__hybrid_group_cpu(evsel, cpu);
1619         if (cpu == -1)
1620                 return -1;
1621
1622         fd = FD(leader, cpu, thread);
1623         BUG_ON(fd == -1);
1624
1625         return fd;
1626 }
1627
1628 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1629 {
1630         for (int cpu = 0; cpu < nr_cpus; cpu++)
1631                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1632                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1633 }
1634
1635 static int update_fds(struct evsel *evsel,
1636                       int nr_cpus, int cpu_idx,
1637                       int nr_threads, int thread_idx)
1638 {
1639         struct evsel *pos;
1640
1641         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1642                 return -EINVAL;
1643
1644         evlist__for_each_entry(evsel->evlist, pos) {
1645                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1646
1647                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1648
1649                 /*
1650                  * Since fds for next evsel has not been created,
1651                  * there is no need to iterate whole event list.
1652                  */
1653                 if (pos == evsel)
1654                         break;
1655         }
1656         return 0;
1657 }
1658
1659 static bool ignore_missing_thread(struct evsel *evsel,
1660                                   int nr_cpus, int cpu,
1661                                   struct perf_thread_map *threads,
1662                                   int thread, int err)
1663 {
1664         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1665
1666         if (!evsel->ignore_missing_thread)
1667                 return false;
1668
1669         /* The system wide setup does not work with threads. */
1670         if (evsel->core.system_wide)
1671                 return false;
1672
1673         /* The -ESRCH is perf event syscall errno for pid's not found. */
1674         if (err != -ESRCH)
1675                 return false;
1676
1677         /* If there's only one thread, let it fail. */
1678         if (threads->nr == 1)
1679                 return false;
1680
1681         /*
1682          * We should remove fd for missing_thread first
1683          * because thread_map__remove() will decrease threads->nr.
1684          */
1685         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1686                 return false;
1687
1688         if (thread_map__remove(threads, thread))
1689                 return false;
1690
1691         pr_warning("WARNING: Ignored open failure for pid %d\n",
1692                    ignore_pid);
1693         return true;
1694 }
1695
1696 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1697                                 void *priv __maybe_unused)
1698 {
1699         return fprintf(fp, "  %-32s %s\n", name, val);
1700 }
1701
1702 static void display_attr(struct perf_event_attr *attr)
1703 {
1704         if (verbose >= 2 || debug_peo_args) {
1705                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1706                 fprintf(stderr, "perf_event_attr:\n");
1707                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1708                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1709         }
1710 }
1711
1712 static int perf_event_open(struct evsel *evsel,
1713                            pid_t pid, int cpu, int group_fd,
1714                            unsigned long flags)
1715 {
1716         int precise_ip = evsel->core.attr.precise_ip;
1717         int fd;
1718
1719         while (1) {
1720                 pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1721                           pid, cpu, group_fd, flags);
1722
1723                 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags);
1724                 if (fd >= 0)
1725                         break;
1726
1727                 /* Do not try less precise if not requested. */
1728                 if (!evsel->precise_max)
1729                         break;
1730
1731                 /*
1732                  * We tried all the precise_ip values, and it's
1733                  * still failing, so leave it to standard fallback.
1734                  */
1735                 if (!evsel->core.attr.precise_ip) {
1736                         evsel->core.attr.precise_ip = precise_ip;
1737                         break;
1738                 }
1739
1740                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP);
1741                 evsel->core.attr.precise_ip--;
1742                 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1743                 display_attr(&evsel->core.attr);
1744         }
1745
1746         return fd;
1747 }
1748
1749 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1750                 struct perf_thread_map *threads,
1751                 int start_cpu, int end_cpu)
1752 {
1753         int cpu, thread, nthreads;
1754         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1755         int pid = -1, err, old_errno;
1756         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1757
1758         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1759             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1760                 return -EINVAL;
1761
1762         if (cpus == NULL) {
1763                 static struct perf_cpu_map *empty_cpu_map;
1764
1765                 if (empty_cpu_map == NULL) {
1766                         empty_cpu_map = perf_cpu_map__dummy_new();
1767                         if (empty_cpu_map == NULL)
1768                                 return -ENOMEM;
1769                 }
1770
1771                 cpus = empty_cpu_map;
1772         }
1773
1774         if (threads == NULL) {
1775                 static struct perf_thread_map *empty_thread_map;
1776
1777                 if (empty_thread_map == NULL) {
1778                         empty_thread_map = thread_map__new_by_tid(-1);
1779                         if (empty_thread_map == NULL)
1780                                 return -ENOMEM;
1781                 }
1782
1783                 threads = empty_thread_map;
1784         }
1785
1786         if (evsel->core.system_wide)
1787                 nthreads = 1;
1788         else
1789                 nthreads = threads->nr;
1790
1791         if (evsel->core.fd == NULL &&
1792             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1793                 return -ENOMEM;
1794
1795         if (evsel->cgrp) {
1796                 flags |= PERF_FLAG_PID_CGROUP;
1797                 pid = evsel->cgrp->fd;
1798         }
1799
1800 fallback_missing_features:
1801         if (perf_missing_features.weight_struct) {
1802                 evsel__set_sample_bit(evsel, WEIGHT);
1803                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1804         }
1805         if (perf_missing_features.clockid_wrong)
1806                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1807         if (perf_missing_features.clockid) {
1808                 evsel->core.attr.use_clockid = 0;
1809                 evsel->core.attr.clockid = 0;
1810         }
1811         if (perf_missing_features.cloexec)
1812                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1813         if (perf_missing_features.mmap2)
1814                 evsel->core.attr.mmap2 = 0;
1815         if (perf_missing_features.exclude_guest)
1816                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1817         if (perf_missing_features.lbr_flags)
1818                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1819                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1820         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1821                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1822         if (perf_missing_features.ksymbol)
1823                 evsel->core.attr.ksymbol = 0;
1824         if (perf_missing_features.bpf)
1825                 evsel->core.attr.bpf_event = 0;
1826         if (perf_missing_features.branch_hw_idx)
1827                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1828 retry_sample_id:
1829         if (perf_missing_features.sample_id_all)
1830                 evsel->core.attr.sample_id_all = 0;
1831
1832         display_attr(&evsel->core.attr);
1833
1834         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1835
1836                 for (thread = 0; thread < nthreads; thread++) {
1837                         int fd, group_fd;
1838
1839                         if (!evsel->cgrp && !evsel->core.system_wide)
1840                                 pid = perf_thread_map__pid(threads, thread);
1841
1842                         group_fd = get_group_fd(evsel, cpu, thread);
1843 retry_open:
1844                         test_attr__ready();
1845
1846                         fd = perf_event_open(evsel, pid, cpus->map[cpu],
1847                                              group_fd, flags);
1848
1849                         FD(evsel, cpu, thread) = fd;
1850
1851                         bpf_counter__install_pe(evsel, cpu, fd);
1852
1853                         if (unlikely(test_attr__enabled)) {
1854                                 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
1855                                                 fd, group_fd, flags);
1856                         }
1857
1858                         if (fd < 0) {
1859                                 err = -errno;
1860
1861                                 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
1862                                         /*
1863                                          * We just removed 1 thread, so take a step
1864                                          * back on thread index and lower the upper
1865                                          * nthreads limit.
1866                                          */
1867                                         nthreads--;
1868                                         thread--;
1869
1870                                         /* ... and pretend like nothing have happened. */
1871                                         err = 0;
1872                                         continue;
1873                                 }
1874
1875                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
1876                                           err);
1877                                 goto try_fallback;
1878                         }
1879
1880                         pr_debug2_peo(" = %d\n", fd);
1881
1882                         if (evsel->bpf_fd >= 0) {
1883                                 int evt_fd = fd;
1884                                 int bpf_fd = evsel->bpf_fd;
1885
1886                                 err = ioctl(evt_fd,
1887                                             PERF_EVENT_IOC_SET_BPF,
1888                                             bpf_fd);
1889                                 if (err && errno != EEXIST) {
1890                                         pr_err("failed to attach bpf fd %d: %s\n",
1891                                                bpf_fd, strerror(errno));
1892                                         err = -EINVAL;
1893                                         goto out_close;
1894                                 }
1895                         }
1896
1897                         set_rlimit = NO_CHANGE;
1898
1899                         /*
1900                          * If we succeeded but had to kill clockid, fail and
1901                          * have evsel__open_strerror() print us a nice error.
1902                          */
1903                         if (perf_missing_features.clockid ||
1904                             perf_missing_features.clockid_wrong) {
1905                                 err = -EINVAL;
1906                                 goto out_close;
1907                         }
1908                 }
1909         }
1910
1911         return 0;
1912
1913 try_fallback:
1914         /*
1915          * perf stat needs between 5 and 22 fds per CPU. When we run out
1916          * of them try to increase the limits.
1917          */
1918         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1919                 struct rlimit l;
1920
1921                 old_errno = errno;
1922                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1923                         if (set_rlimit == NO_CHANGE)
1924                                 l.rlim_cur = l.rlim_max;
1925                         else {
1926                                 l.rlim_cur = l.rlim_max + 1000;
1927                                 l.rlim_max = l.rlim_cur;
1928                         }
1929                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1930                                 set_rlimit++;
1931                                 errno = old_errno;
1932                                 goto retry_open;
1933                         }
1934                 }
1935                 errno = old_errno;
1936         }
1937
1938         if (err != -EINVAL || cpu > 0 || thread > 0)
1939                 goto out_close;
1940
1941         /*
1942          * Must probe features in the order they were added to the
1943          * perf_event_attr interface.
1944          */
1945         if (!perf_missing_features.weight_struct &&
1946             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1947                 perf_missing_features.weight_struct = true;
1948                 pr_debug2("switching off weight struct support\n");
1949                 goto fallback_missing_features;
1950         } else if (!perf_missing_features.code_page_size &&
1951             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1952                 perf_missing_features.code_page_size = true;
1953                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1954                 goto out_close;
1955         } else if (!perf_missing_features.data_page_size &&
1956             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1957                 perf_missing_features.data_page_size = true;
1958                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1959                 goto out_close;
1960         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1961                 perf_missing_features.cgroup = true;
1962                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1963                 goto out_close;
1964         } else if (!perf_missing_features.branch_hw_idx &&
1965             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1966                 perf_missing_features.branch_hw_idx = true;
1967                 pr_debug2("switching off branch HW index support\n");
1968                 goto fallback_missing_features;
1969         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1970                 perf_missing_features.aux_output = true;
1971                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1972                 goto out_close;
1973         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1974                 perf_missing_features.bpf = true;
1975                 pr_debug2_peo("switching off bpf_event\n");
1976                 goto fallback_missing_features;
1977         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1978                 perf_missing_features.ksymbol = true;
1979                 pr_debug2_peo("switching off ksymbol\n");
1980                 goto fallback_missing_features;
1981         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1982                 perf_missing_features.write_backward = true;
1983                 pr_debug2_peo("switching off write_backward\n");
1984                 goto out_close;
1985         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1986                 perf_missing_features.clockid_wrong = true;
1987                 pr_debug2_peo("switching off clockid\n");
1988                 goto fallback_missing_features;
1989         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1990                 perf_missing_features.clockid = true;
1991                 pr_debug2_peo("switching off use_clockid\n");
1992                 goto fallback_missing_features;
1993         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1994                 perf_missing_features.cloexec = true;
1995                 pr_debug2_peo("switching off cloexec flag\n");
1996                 goto fallback_missing_features;
1997         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1998                 perf_missing_features.mmap2 = true;
1999                 pr_debug2_peo("switching off mmap2\n");
2000                 goto fallback_missing_features;
2001         } else if (!perf_missing_features.exclude_guest &&
2002                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
2003                 perf_missing_features.exclude_guest = true;
2004                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2005                 goto fallback_missing_features;
2006         } else if (!perf_missing_features.sample_id_all) {
2007                 perf_missing_features.sample_id_all = true;
2008                 pr_debug2_peo("switching off sample_id_all\n");
2009                 goto retry_sample_id;
2010         } else if (!perf_missing_features.lbr_flags &&
2011                         (evsel->core.attr.branch_sample_type &
2012                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2013                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2014                 perf_missing_features.lbr_flags = true;
2015                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2016                 goto fallback_missing_features;
2017         } else if (!perf_missing_features.group_read &&
2018                     evsel->core.attr.inherit &&
2019                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2020                    evsel__is_group_leader(evsel)) {
2021                 perf_missing_features.group_read = true;
2022                 pr_debug2_peo("switching off group read\n");
2023                 goto fallback_missing_features;
2024         }
2025 out_close:
2026         if (err)
2027                 threads->err_thread = thread;
2028
2029         old_errno = errno;
2030         do {
2031                 while (--thread >= 0) {
2032                         if (FD(evsel, cpu, thread) >= 0)
2033                                 close(FD(evsel, cpu, thread));
2034                         FD(evsel, cpu, thread) = -1;
2035                 }
2036                 thread = nthreads;
2037         } while (--cpu >= 0);
2038         errno = old_errno;
2039         return err;
2040 }
2041
2042 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2043                 struct perf_thread_map *threads)
2044 {
2045         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
2046 }
2047
2048 void evsel__close(struct evsel *evsel)
2049 {
2050         perf_evsel__close(&evsel->core);
2051         perf_evsel__free_id(&evsel->core);
2052 }
2053
2054 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
2055 {
2056         if (cpu == -1)
2057                 return evsel__open_cpu(evsel, cpus, NULL, 0,
2058                                         cpus ? cpus->nr : 1);
2059
2060         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
2061 }
2062
2063 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2064 {
2065         return evsel__open(evsel, NULL, threads);
2066 }
2067
2068 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2069                                        const union perf_event *event,
2070                                        struct perf_sample *sample)
2071 {
2072         u64 type = evsel->core.attr.sample_type;
2073         const __u64 *array = event->sample.array;
2074         bool swapped = evsel->needs_swap;
2075         union u64_swap u;
2076
2077         array += ((event->header.size -
2078                    sizeof(event->header)) / sizeof(u64)) - 1;
2079
2080         if (type & PERF_SAMPLE_IDENTIFIER) {
2081                 sample->id = *array;
2082                 array--;
2083         }
2084
2085         if (type & PERF_SAMPLE_CPU) {
2086                 u.val64 = *array;
2087                 if (swapped) {
2088                         /* undo swap of u64, then swap on individual u32s */
2089                         u.val64 = bswap_64(u.val64);
2090                         u.val32[0] = bswap_32(u.val32[0]);
2091                 }
2092
2093                 sample->cpu = u.val32[0];
2094                 array--;
2095         }
2096
2097         if (type & PERF_SAMPLE_STREAM_ID) {
2098                 sample->stream_id = *array;
2099                 array--;
2100         }
2101
2102         if (type & PERF_SAMPLE_ID) {
2103                 sample->id = *array;
2104                 array--;
2105         }
2106
2107         if (type & PERF_SAMPLE_TIME) {
2108                 sample->time = *array;
2109                 array--;
2110         }
2111
2112         if (type & PERF_SAMPLE_TID) {
2113                 u.val64 = *array;
2114                 if (swapped) {
2115                         /* undo swap of u64, then swap on individual u32s */
2116                         u.val64 = bswap_64(u.val64);
2117                         u.val32[0] = bswap_32(u.val32[0]);
2118                         u.val32[1] = bswap_32(u.val32[1]);
2119                 }
2120
2121                 sample->pid = u.val32[0];
2122                 sample->tid = u.val32[1];
2123                 array--;
2124         }
2125
2126         return 0;
2127 }
2128
2129 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2130                             u64 size)
2131 {
2132         return size > max_size || offset + size > endp;
2133 }
2134
2135 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2136         do {                                                            \
2137                 if (overflow(endp, (max_size), (offset), (size)))       \
2138                         return -EFAULT;                                 \
2139         } while (0)
2140
2141 #define OVERFLOW_CHECK_u64(offset) \
2142         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2143
2144 static int
2145 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2146 {
2147         /*
2148          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2149          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2150          * check the format does not go past the end of the event.
2151          */
2152         if (sample_size + sizeof(event->header) > event->header.size)
2153                 return -EFAULT;
2154
2155         return 0;
2156 }
2157
2158 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2159                                           const __u64 *array,
2160                                           u64 type __maybe_unused)
2161 {
2162         data->weight = *array;
2163 }
2164
2165 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2166                         struct perf_sample *data)
2167 {
2168         u64 type = evsel->core.attr.sample_type;
2169         bool swapped = evsel->needs_swap;
2170         const __u64 *array;
2171         u16 max_size = event->header.size;
2172         const void *endp = (void *)event + max_size;
2173         u64 sz;
2174
2175         /*
2176          * used for cross-endian analysis. See git commit 65014ab3
2177          * for why this goofiness is needed.
2178          */
2179         union u64_swap u;
2180
2181         memset(data, 0, sizeof(*data));
2182         data->cpu = data->pid = data->tid = -1;
2183         data->stream_id = data->id = data->time = -1ULL;
2184         data->period = evsel->core.attr.sample_period;
2185         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2186         data->misc    = event->header.misc;
2187         data->id = -1ULL;
2188         data->data_src = PERF_MEM_DATA_SRC_NONE;
2189
2190         if (event->header.type != PERF_RECORD_SAMPLE) {
2191                 if (!evsel->core.attr.sample_id_all)
2192                         return 0;
2193                 return perf_evsel__parse_id_sample(evsel, event, data);
2194         }
2195
2196         array = event->sample.array;
2197
2198         if (perf_event__check_size(event, evsel->sample_size))
2199                 return -EFAULT;
2200
2201         if (type & PERF_SAMPLE_IDENTIFIER) {
2202                 data->id = *array;
2203                 array++;
2204         }
2205
2206         if (type & PERF_SAMPLE_IP) {
2207                 data->ip = *array;
2208                 array++;
2209         }
2210
2211         if (type & PERF_SAMPLE_TID) {
2212                 u.val64 = *array;
2213                 if (swapped) {
2214                         /* undo swap of u64, then swap on individual u32s */
2215                         u.val64 = bswap_64(u.val64);
2216                         u.val32[0] = bswap_32(u.val32[0]);
2217                         u.val32[1] = bswap_32(u.val32[1]);
2218                 }
2219
2220                 data->pid = u.val32[0];
2221                 data->tid = u.val32[1];
2222                 array++;
2223         }
2224
2225         if (type & PERF_SAMPLE_TIME) {
2226                 data->time = *array;
2227                 array++;
2228         }
2229
2230         if (type & PERF_SAMPLE_ADDR) {
2231                 data->addr = *array;
2232                 array++;
2233         }
2234
2235         if (type & PERF_SAMPLE_ID) {
2236                 data->id = *array;
2237                 array++;
2238         }
2239
2240         if (type & PERF_SAMPLE_STREAM_ID) {
2241                 data->stream_id = *array;
2242                 array++;
2243         }
2244
2245         if (type & PERF_SAMPLE_CPU) {
2246
2247                 u.val64 = *array;
2248                 if (swapped) {
2249                         /* undo swap of u64, then swap on individual u32s */
2250                         u.val64 = bswap_64(u.val64);
2251                         u.val32[0] = bswap_32(u.val32[0]);
2252                 }
2253
2254                 data->cpu = u.val32[0];
2255                 array++;
2256         }
2257
2258         if (type & PERF_SAMPLE_PERIOD) {
2259                 data->period = *array;
2260                 array++;
2261         }
2262
2263         if (type & PERF_SAMPLE_READ) {
2264                 u64 read_format = evsel->core.attr.read_format;
2265
2266                 OVERFLOW_CHECK_u64(array);
2267                 if (read_format & PERF_FORMAT_GROUP)
2268                         data->read.group.nr = *array;
2269                 else
2270                         data->read.one.value = *array;
2271
2272                 array++;
2273
2274                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2275                         OVERFLOW_CHECK_u64(array);
2276                         data->read.time_enabled = *array;
2277                         array++;
2278                 }
2279
2280                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2281                         OVERFLOW_CHECK_u64(array);
2282                         data->read.time_running = *array;
2283                         array++;
2284                 }
2285
2286                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2287                 if (read_format & PERF_FORMAT_GROUP) {
2288                         const u64 max_group_nr = UINT64_MAX /
2289                                         sizeof(struct sample_read_value);
2290
2291                         if (data->read.group.nr > max_group_nr)
2292                                 return -EFAULT;
2293                         sz = data->read.group.nr *
2294                              sizeof(struct sample_read_value);
2295                         OVERFLOW_CHECK(array, sz, max_size);
2296                         data->read.group.values =
2297                                         (struct sample_read_value *)array;
2298                         array = (void *)array + sz;
2299                 } else {
2300                         OVERFLOW_CHECK_u64(array);
2301                         data->read.one.id = *array;
2302                         array++;
2303                 }
2304         }
2305
2306         if (type & PERF_SAMPLE_CALLCHAIN) {
2307                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2308
2309                 OVERFLOW_CHECK_u64(array);
2310                 data->callchain = (struct ip_callchain *)array++;
2311                 if (data->callchain->nr > max_callchain_nr)
2312                         return -EFAULT;
2313                 sz = data->callchain->nr * sizeof(u64);
2314                 OVERFLOW_CHECK(array, sz, max_size);
2315                 array = (void *)array + sz;
2316         }
2317
2318         if (type & PERF_SAMPLE_RAW) {
2319                 OVERFLOW_CHECK_u64(array);
2320                 u.val64 = *array;
2321
2322                 /*
2323                  * Undo swap of u64, then swap on individual u32s,
2324                  * get the size of the raw area and undo all of the
2325                  * swap. The pevent interface handles endianness by
2326                  * itself.
2327                  */
2328                 if (swapped) {
2329                         u.val64 = bswap_64(u.val64);
2330                         u.val32[0] = bswap_32(u.val32[0]);
2331                         u.val32[1] = bswap_32(u.val32[1]);
2332                 }
2333                 data->raw_size = u.val32[0];
2334
2335                 /*
2336                  * The raw data is aligned on 64bits including the
2337                  * u32 size, so it's safe to use mem_bswap_64.
2338                  */
2339                 if (swapped)
2340                         mem_bswap_64((void *) array, data->raw_size);
2341
2342                 array = (void *)array + sizeof(u32);
2343
2344                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2345                 data->raw_data = (void *)array;
2346                 array = (void *)array + data->raw_size;
2347         }
2348
2349         if (type & PERF_SAMPLE_BRANCH_STACK) {
2350                 const u64 max_branch_nr = UINT64_MAX /
2351                                           sizeof(struct branch_entry);
2352
2353                 OVERFLOW_CHECK_u64(array);
2354                 data->branch_stack = (struct branch_stack *)array++;
2355
2356                 if (data->branch_stack->nr > max_branch_nr)
2357                         return -EFAULT;
2358
2359                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2360                 if (evsel__has_branch_hw_idx(evsel))
2361                         sz += sizeof(u64);
2362                 else
2363                         data->no_hw_idx = true;
2364                 OVERFLOW_CHECK(array, sz, max_size);
2365                 array = (void *)array + sz;
2366         }
2367
2368         if (type & PERF_SAMPLE_REGS_USER) {
2369                 OVERFLOW_CHECK_u64(array);
2370                 data->user_regs.abi = *array;
2371                 array++;
2372
2373                 if (data->user_regs.abi) {
2374                         u64 mask = evsel->core.attr.sample_regs_user;
2375
2376                         sz = hweight64(mask) * sizeof(u64);
2377                         OVERFLOW_CHECK(array, sz, max_size);
2378                         data->user_regs.mask = mask;
2379                         data->user_regs.regs = (u64 *)array;
2380                         array = (void *)array + sz;
2381                 }
2382         }
2383
2384         if (type & PERF_SAMPLE_STACK_USER) {
2385                 OVERFLOW_CHECK_u64(array);
2386                 sz = *array++;
2387
2388                 data->user_stack.offset = ((char *)(array - 1)
2389                                           - (char *) event);
2390
2391                 if (!sz) {
2392                         data->user_stack.size = 0;
2393                 } else {
2394                         OVERFLOW_CHECK(array, sz, max_size);
2395                         data->user_stack.data = (char *)array;
2396                         array = (void *)array + sz;
2397                         OVERFLOW_CHECK_u64(array);
2398                         data->user_stack.size = *array++;
2399                         if (WARN_ONCE(data->user_stack.size > sz,
2400                                       "user stack dump failure\n"))
2401                                 return -EFAULT;
2402                 }
2403         }
2404
2405         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2406                 OVERFLOW_CHECK_u64(array);
2407                 arch_perf_parse_sample_weight(data, array, type);
2408                 array++;
2409         }
2410
2411         if (type & PERF_SAMPLE_DATA_SRC) {
2412                 OVERFLOW_CHECK_u64(array);
2413                 data->data_src = *array;
2414                 array++;
2415         }
2416
2417         if (type & PERF_SAMPLE_TRANSACTION) {
2418                 OVERFLOW_CHECK_u64(array);
2419                 data->transaction = *array;
2420                 array++;
2421         }
2422
2423         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2424         if (type & PERF_SAMPLE_REGS_INTR) {
2425                 OVERFLOW_CHECK_u64(array);
2426                 data->intr_regs.abi = *array;
2427                 array++;
2428
2429                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2430                         u64 mask = evsel->core.attr.sample_regs_intr;
2431
2432                         sz = hweight64(mask) * sizeof(u64);
2433                         OVERFLOW_CHECK(array, sz, max_size);
2434                         data->intr_regs.mask = mask;
2435                         data->intr_regs.regs = (u64 *)array;
2436                         array = (void *)array + sz;
2437                 }
2438         }
2439
2440         data->phys_addr = 0;
2441         if (type & PERF_SAMPLE_PHYS_ADDR) {
2442                 data->phys_addr = *array;
2443                 array++;
2444         }
2445
2446         data->cgroup = 0;
2447         if (type & PERF_SAMPLE_CGROUP) {
2448                 data->cgroup = *array;
2449                 array++;
2450         }
2451
2452         data->data_page_size = 0;
2453         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2454                 data->data_page_size = *array;
2455                 array++;
2456         }
2457
2458         data->code_page_size = 0;
2459         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2460                 data->code_page_size = *array;
2461                 array++;
2462         }
2463
2464         if (type & PERF_SAMPLE_AUX) {
2465                 OVERFLOW_CHECK_u64(array);
2466                 sz = *array++;
2467
2468                 OVERFLOW_CHECK(array, sz, max_size);
2469                 /* Undo swap of data */
2470                 if (swapped)
2471                         mem_bswap_64((char *)array, sz);
2472                 data->aux_sample.size = sz;
2473                 data->aux_sample.data = (char *)array;
2474                 array = (void *)array + sz;
2475         }
2476
2477         return 0;
2478 }
2479
2480 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2481                                   u64 *timestamp)
2482 {
2483         u64 type = evsel->core.attr.sample_type;
2484         const __u64 *array;
2485
2486         if (!(type & PERF_SAMPLE_TIME))
2487                 return -1;
2488
2489         if (event->header.type != PERF_RECORD_SAMPLE) {
2490                 struct perf_sample data = {
2491                         .time = -1ULL,
2492                 };
2493
2494                 if (!evsel->core.attr.sample_id_all)
2495                         return -1;
2496                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2497                         return -1;
2498
2499                 *timestamp = data.time;
2500                 return 0;
2501         }
2502
2503         array = event->sample.array;
2504
2505         if (perf_event__check_size(event, evsel->sample_size))
2506                 return -EFAULT;
2507
2508         if (type & PERF_SAMPLE_IDENTIFIER)
2509                 array++;
2510
2511         if (type & PERF_SAMPLE_IP)
2512                 array++;
2513
2514         if (type & PERF_SAMPLE_TID)
2515                 array++;
2516
2517         if (type & PERF_SAMPLE_TIME)
2518                 *timestamp = *array;
2519
2520         return 0;
2521 }
2522
2523 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2524 {
2525         return tep_find_field(evsel->tp_format, name);
2526 }
2527
2528 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2529 {
2530         struct tep_format_field *field = evsel__field(evsel, name);
2531         int offset;
2532
2533         if (!field)
2534                 return NULL;
2535
2536         offset = field->offset;
2537
2538         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2539                 offset = *(int *)(sample->raw_data + field->offset);
2540                 offset &= 0xffff;
2541         }
2542
2543         return sample->raw_data + offset;
2544 }
2545
2546 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2547                          bool needs_swap)
2548 {
2549         u64 value;
2550         void *ptr = sample->raw_data + field->offset;
2551
2552         switch (field->size) {
2553         case 1:
2554                 return *(u8 *)ptr;
2555         case 2:
2556                 value = *(u16 *)ptr;
2557                 break;
2558         case 4:
2559                 value = *(u32 *)ptr;
2560                 break;
2561         case 8:
2562                 memcpy(&value, ptr, sizeof(u64));
2563                 break;
2564         default:
2565                 return 0;
2566         }
2567
2568         if (!needs_swap)
2569                 return value;
2570
2571         switch (field->size) {
2572         case 2:
2573                 return bswap_16(value);
2574         case 4:
2575                 return bswap_32(value);
2576         case 8:
2577                 return bswap_64(value);
2578         default:
2579                 return 0;
2580         }
2581
2582         return 0;
2583 }
2584
2585 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2586 {
2587         struct tep_format_field *field = evsel__field(evsel, name);
2588
2589         if (!field)
2590                 return 0;
2591
2592         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2593 }
2594
2595 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2596 {
2597         int paranoid;
2598
2599         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2600             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2601             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2602                 /*
2603                  * If it's cycles then fall back to hrtimer based
2604                  * cpu-clock-tick sw counter, which is always available even if
2605                  * no PMU support.
2606                  *
2607                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2608                  * b0a873e).
2609                  */
2610                 scnprintf(msg, msgsize, "%s",
2611 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2612
2613                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2614                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2615
2616                 zfree(&evsel->name);
2617                 return true;
2618         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2619                    (paranoid = perf_event_paranoid()) > 1) {
2620                 const char *name = evsel__name(evsel);
2621                 char *new_name;
2622                 const char *sep = ":";
2623
2624                 /* If event has exclude user then don't exclude kernel. */
2625                 if (evsel->core.attr.exclude_user)
2626                         return false;
2627
2628                 /* Is there already the separator in the name. */
2629                 if (strchr(name, '/') ||
2630                     (strchr(name, ':') && !evsel->is_libpfm_event))
2631                         sep = "";
2632
2633                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2634                         return false;
2635
2636                 if (evsel->name)
2637                         free(evsel->name);
2638                 evsel->name = new_name;
2639                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2640                           "to fall back to excluding kernel and hypervisor "
2641                           " samples", paranoid);
2642                 evsel->core.attr.exclude_kernel = 1;
2643                 evsel->core.attr.exclude_hv     = 1;
2644
2645                 return true;
2646         }
2647
2648         return false;
2649 }
2650
2651 static bool find_process(const char *name)
2652 {
2653         size_t len = strlen(name);
2654         DIR *dir;
2655         struct dirent *d;
2656         int ret = -1;
2657
2658         dir = opendir(procfs__mountpoint());
2659         if (!dir)
2660                 return false;
2661
2662         /* Walk through the directory. */
2663         while (ret && (d = readdir(dir)) != NULL) {
2664                 char path[PATH_MAX];
2665                 char *data;
2666                 size_t size;
2667
2668                 if ((d->d_type != DT_DIR) ||
2669                      !strcmp(".", d->d_name) ||
2670                      !strcmp("..", d->d_name))
2671                         continue;
2672
2673                 scnprintf(path, sizeof(path), "%s/%s/comm",
2674                           procfs__mountpoint(), d->d_name);
2675
2676                 if (filename__read_str(path, &data, &size))
2677                         continue;
2678
2679                 ret = strncmp(name, data, len);
2680                 free(data);
2681         }
2682
2683         closedir(dir);
2684         return ret ? false : true;
2685 }
2686
2687 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2688                          int err, char *msg, size_t size)
2689 {
2690         char sbuf[STRERR_BUFSIZE];
2691         int printed = 0, enforced = 0;
2692
2693         switch (err) {
2694         case EPERM:
2695         case EACCES:
2696                 printed += scnprintf(msg + printed, size - printed,
2697                         "Access to performance monitoring and observability operations is limited.\n");
2698
2699                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2700                         if (enforced) {
2701                                 printed += scnprintf(msg + printed, size - printed,
2702                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2703                                         "monitoring and observability operations. Inspect system audit records for\n"
2704                                         "more perf_event access control information and adjusting the policy.\n");
2705                         }
2706                 }
2707
2708                 if (err == EPERM)
2709                         printed += scnprintf(msg, size,
2710                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2711
2712                 return scnprintf(msg + printed, size - printed,
2713                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2714                  "access to performance monitoring and observability operations for processes\n"
2715                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2716                  "More information can be found at 'Perf events and tool security' document:\n"
2717                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2718                  "perf_event_paranoid setting is %d:\n"
2719                  "  -1: Allow use of (almost) all events by all users\n"
2720                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2721                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2722                  ">= 1: Disallow CPU event access\n"
2723                  ">= 2: Disallow kernel profiling\n"
2724                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2725                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2726                  perf_event_paranoid());
2727         case ENOENT:
2728                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2729         case EMFILE:
2730                 return scnprintf(msg, size, "%s",
2731                          "Too many events are opened.\n"
2732                          "Probably the maximum number of open file descriptors has been reached.\n"
2733                          "Hint: Try again after reducing the number of events.\n"
2734                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2735         case ENOMEM:
2736                 if (evsel__has_callchain(evsel) &&
2737                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2738                         return scnprintf(msg, size,
2739                                          "Not enough memory to setup event with callchain.\n"
2740                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2741                                          "Hint: Current value: %d", sysctl__max_stack());
2742                 break;
2743         case ENODEV:
2744                 if (target->cpu_list)
2745                         return scnprintf(msg, size, "%s",
2746          "No such device - did you specify an out-of-range profile CPU?");
2747                 break;
2748         case EOPNOTSUPP:
2749                 if (evsel->core.attr.aux_output)
2750                         return scnprintf(msg, size,
2751         "%s: PMU Hardware doesn't support 'aux_output' feature",
2752                                          evsel__name(evsel));
2753                 if (evsel->core.attr.sample_period != 0)
2754                         return scnprintf(msg, size,
2755         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2756                                          evsel__name(evsel));
2757                 if (evsel->core.attr.precise_ip)
2758                         return scnprintf(msg, size, "%s",
2759         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2760 #if defined(__i386__) || defined(__x86_64__)
2761                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2762                         return scnprintf(msg, size, "%s",
2763         "No hardware sampling interrupt available.\n");
2764 #endif
2765                 break;
2766         case EBUSY:
2767                 if (find_process("oprofiled"))
2768                         return scnprintf(msg, size,
2769         "The PMU counters are busy/taken by another profiler.\n"
2770         "We found oprofile daemon running, please stop it and try again.");
2771                 break;
2772         case EINVAL:
2773                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
2774                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
2775                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2776                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2777                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2778                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2779                 if (perf_missing_features.clockid)
2780                         return scnprintf(msg, size, "clockid feature not supported.");
2781                 if (perf_missing_features.clockid_wrong)
2782                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2783                 if (perf_missing_features.aux_output)
2784                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2785                 break;
2786         case ENODATA:
2787                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
2788                                  "Please add an auxiliary event in front of the load latency event.");
2789         default:
2790                 break;
2791         }
2792
2793         return scnprintf(msg, size,
2794         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2795         "/bin/dmesg | grep -i perf may provide additional information.\n",
2796                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2797 }
2798
2799 struct perf_env *evsel__env(struct evsel *evsel)
2800 {
2801         if (evsel && evsel->evlist)
2802                 return evsel->evlist->env;
2803         return &perf_env;
2804 }
2805
2806 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2807 {
2808         int cpu, thread;
2809
2810         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2811                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2812                      thread++) {
2813                         int fd = FD(evsel, cpu, thread);
2814
2815                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2816                                                    cpu, thread, fd) < 0)
2817                                 return -1;
2818                 }
2819         }
2820
2821         return 0;
2822 }
2823
2824 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2825 {
2826         struct perf_cpu_map *cpus = evsel->core.cpus;
2827         struct perf_thread_map *threads = evsel->core.threads;
2828
2829         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2830                 return -ENOMEM;
2831
2832         return store_evsel_ids(evsel, evlist);
2833 }
2834
2835 void evsel__zero_per_pkg(struct evsel *evsel)
2836 {
2837         struct hashmap_entry *cur;
2838         size_t bkt;
2839
2840         if (evsel->per_pkg_mask) {
2841                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
2842                         free((char *)cur->key);
2843
2844                 hashmap__clear(evsel->per_pkg_mask);
2845         }
2846 }
2847
2848 bool evsel__is_hybrid(struct evsel *evsel)
2849 {
2850         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
2851 }
2852
2853 struct evsel *evsel__leader(struct evsel *evsel)
2854 {
2855         return container_of(evsel->core.leader, struct evsel, core);
2856 }
2857
2858 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
2859 {
2860         return evsel->core.leader == &leader->core;
2861 }
2862
2863 bool evsel__is_leader(struct evsel *evsel)
2864 {
2865         return evsel__has_leader(evsel, evsel);
2866 }
2867
2868 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
2869 {
2870         evsel->core.leader = &leader->core;
2871 }