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