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 <linux/hw_breakpoint.h>
  16 #include <linux/perf_event.h>
  17 #include <linux/compiler.h>
  18 #include <linux/err.h>
  19 #include <linux/zalloc.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/resource.h>
  22 #include <sys/types.h>
  23 #include <dirent.h>
  24 #include <stdlib.h>
  25 #include <perf/evsel.h>
  26 #include "asm/bug.h"
  27 #include "bpf_counter.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "util/hashmap.h"
  49 #include "off_cpu.h"
  50 #include "pmu.h"
  51 #include "pmus.h"
  52 #include "rlimit.h"
  53 #include "../perf-sys.h"
  54 #include "util/parse-branch-options.h"
  55 #include "util/bpf-filter.h"
  56 #include <internal/xyarray.h>
  57 #include <internal/lib.h>
  58 #include <internal/threadmap.h>
  59
  60 #include <linux/ctype.h>
  61
  62 #ifdef HAVE_LIBTRACEEVENT
  63 #include <traceevent/event-parse.h>
  64 #endif
  65
  66 struct perf_missing_features perf_missing_features;
  67
  68 static clockid_t clockid;
  69
  70 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  71         NULL,
  72         "duration_time",
  73         "user_time",
  74         "system_time",
  75 };
  76
  77 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  78 {
  79         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  80                 return perf_tool_event__tool_names[ev];
  81
  82         return NULL;
  83 }
  84
  85 enum perf_tool_event perf_tool_event__from_str(const char *str)
  86 {
  87         int i;
  88
  89         perf_tool_event__for_each_event(i) {
  90                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  91                         return i;
  92         }
  93         return PERF_TOOL_NONE;
  94 }
  95
  96
  97 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  98 {
  99         return 0;
 100 }
 101
 102 void __weak test_attr__ready(void) { }
 103
 104 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 105 {
 106 }
 107
 108 static struct {
 109         size_t  size;
 110         int     (*init)(struct evsel *evsel);
 111         void    (*fini)(struct evsel *evsel);
 112 } perf_evsel__object = {
 113         .size = sizeof(struct evsel),
 114         .init = evsel__no_extra_init,
 115         .fini = evsel__no_extra_fini,
 116 };
 117
 118 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 119                          void (*fini)(struct evsel *evsel))
 120 {
 121
 122         if (object_size == 0)
 123                 goto set_methods;
 124
 125         if (perf_evsel__object.size > object_size)
 126                 return -EINVAL;
 127
 128         perf_evsel__object.size = object_size;
 129
 130 set_methods:
 131         if (init != NULL)
 132                 perf_evsel__object.init = init;
 133
 134         if (fini != NULL)
 135                 perf_evsel__object.fini = fini;
 136
 137         return 0;
 138 }
 139
 140 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 141
 142 int __evsel__sample_size(u64 sample_type)
 143 {
 144         u64 mask = sample_type & PERF_SAMPLE_MASK;
 145         int size = 0;
 146         int i;
 147
 148         for (i = 0; i < 64; i++) {
 149                 if (mask & (1ULL << i))
 150                         size++;
 151         }
 152
 153         size *= sizeof(u64);
 154
 155         return size;
 156 }
 157
 158 /**
 159  * __perf_evsel__calc_id_pos - calculate id_pos.
 160  * @sample_type: sample type
 161  *
 162  * This function returns the position of the event id (PERF_SAMPLE_ID or
 163  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 164  * perf_record_sample.
 165  */
 166 static int __perf_evsel__calc_id_pos(u64 sample_type)
 167 {
 168         int idx = 0;
 169
 170         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 171                 return 0;
 172
 173         if (!(sample_type & PERF_SAMPLE_ID))
 174                 return -1;
 175
 176         if (sample_type & PERF_SAMPLE_IP)
 177                 idx += 1;
 178
 179         if (sample_type & PERF_SAMPLE_TID)
 180                 idx += 1;
 181
 182         if (sample_type & PERF_SAMPLE_TIME)
 183                 idx += 1;
 184
 185         if (sample_type & PERF_SAMPLE_ADDR)
 186                 idx += 1;
 187
 188         return idx;
 189 }
 190
 191 /**
 192  * __perf_evsel__calc_is_pos - calculate is_pos.
 193  * @sample_type: sample type
 194  *
 195  * This function returns the position (counting backwards) of the event id
 196  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 197  * sample_id_all is used there is an id sample appended to non-sample events.
 198  */
 199 static int __perf_evsel__calc_is_pos(u64 sample_type)
 200 {
 201         int idx = 1;
 202
 203         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 204                 return 1;
 205
 206         if (!(sample_type & PERF_SAMPLE_ID))
 207                 return -1;
 208
 209         if (sample_type & PERF_SAMPLE_CPU)
 210                 idx += 1;
 211
 212         if (sample_type & PERF_SAMPLE_STREAM_ID)
 213                 idx += 1;
 214
 215         return idx;
 216 }
 217
 218 void evsel__calc_id_pos(struct evsel *evsel)
 219 {
 220         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 221         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 222 }
 223
 224 void __evsel__set_sample_bit(struct evsel *evsel,
 225                                   enum perf_event_sample_format bit)
 226 {
 227         if (!(evsel->core.attr.sample_type & bit)) {
 228                 evsel->core.attr.sample_type |= bit;
 229                 evsel->sample_size += sizeof(u64);
 230                 evsel__calc_id_pos(evsel);
 231         }
 232 }
 233
 234 void __evsel__reset_sample_bit(struct evsel *evsel,
 235                                     enum perf_event_sample_format bit)
 236 {
 237         if (evsel->core.attr.sample_type & bit) {
 238                 evsel->core.attr.sample_type &= ~bit;
 239                 evsel->sample_size -= sizeof(u64);
 240                 evsel__calc_id_pos(evsel);
 241         }
 242 }
 243
 244 void evsel__set_sample_id(struct evsel *evsel,
 245                                bool can_sample_identifier)
 246 {
 247         if (can_sample_identifier) {
 248                 evsel__reset_sample_bit(evsel, ID);
 249                 evsel__set_sample_bit(evsel, IDENTIFIER);
 250         } else {
 251                 evsel__set_sample_bit(evsel, ID);
 252         }
 253         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 254 }
 255
 256 /**
 257  * evsel__is_function_event - Return whether given evsel is a function
 258  * trace event
 259  *
 260  * @evsel - evsel selector to be tested
 261  *
 262  * Return %true if event is function trace event
 263  */
 264 bool evsel__is_function_event(struct evsel *evsel)
 265 {
 266 #define FUNCTION_EVENT "ftrace:function"
 267
 268         return evsel->name &&
 269                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 270
 271 #undef FUNCTION_EVENT
 272 }
 273
 274 void evsel__init(struct evsel *evsel,
 275                  struct perf_event_attr *attr, int idx)
 276 {
 277         perf_evsel__init(&evsel->core, attr, idx);
 278         evsel->tracking    = !idx;
 279         evsel->unit        = strdup("");
 280         evsel->scale       = 1.0;
 281         evsel->max_events  = ULONG_MAX;
 282         evsel->evlist      = NULL;
 283         evsel->bpf_obj     = NULL;
 284         evsel->bpf_fd      = -1;
 285         INIT_LIST_HEAD(&evsel->config_terms);
 286         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 287         INIT_LIST_HEAD(&evsel->bpf_filters);
 288         perf_evsel__object.init(evsel);
 289         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 290         evsel__calc_id_pos(evsel);
 291         evsel->cmdline_group_boundary = false;
 292         evsel->metric_events = NULL;
 293         evsel->per_pkg_mask  = NULL;
 294         evsel->collect_stat  = false;
 295         evsel->pmu_name      = NULL;
 296         evsel->group_pmu_name = NULL;
 297         evsel->skippable     = false;
 298 }
 299
 300 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 301 {
 302         struct evsel *evsel = zalloc(perf_evsel__object.size);
 303
 304         if (!evsel)
 305                 return NULL;
 306         evsel__init(evsel, attr, idx);
 307
 308         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 309                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 310                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 311                 evsel->core.attr.sample_period = 1;
 312         }
 313
 314         if (evsel__is_clock(evsel)) {
 315                 free((char *)evsel->unit);
 316                 evsel->unit = strdup("msec");
 317                 evsel->scale = 1e-6;
 318         }
 319
 320         return evsel;
 321 }
 322
 323 int copy_config_terms(struct list_head *dst, struct list_head *src)
 324 {
 325         struct evsel_config_term *pos, *tmp;
 326
 327         list_for_each_entry(pos, src, list) {
 328                 tmp = malloc(sizeof(*tmp));
 329                 if (tmp == NULL)
 330                         return -ENOMEM;
 331
 332                 *tmp = *pos;
 333                 if (tmp->free_str) {
 334                         tmp->val.str = strdup(pos->val.str);
 335                         if (tmp->val.str == NULL) {
 336                                 free(tmp);
 337                                 return -ENOMEM;
 338                         }
 339                 }
 340                 list_add_tail(&tmp->list, dst);
 341         }
 342         return 0;
 343 }
 344
 345 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 346 {
 347         return copy_config_terms(&dst->config_terms, &src->config_terms);
 348 }
 349
 350 /**
 351  * evsel__clone - create a new evsel copied from @orig
 352  * @orig: original evsel
 353  *
 354  * The assumption is that @orig is not configured nor opened yet.
 355  * So we only care about the attributes that can be set while it's parsed.
 356  */
 357 struct evsel *evsel__clone(struct evsel *orig)
 358 {
 359         struct evsel *evsel;
 360
 361         BUG_ON(orig->core.fd);
 362         BUG_ON(orig->counts);
 363         BUG_ON(orig->priv);
 364         BUG_ON(orig->per_pkg_mask);
 365
 366         /* cannot handle BPF objects for now */
 367         if (orig->bpf_obj)
 368                 return NULL;
 369
 370         evsel = evsel__new(&orig->core.attr);
 371         if (evsel == NULL)
 372                 return NULL;
 373
 374         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 375         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 376         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 377         evsel->core.nr_members = orig->core.nr_members;
 378         evsel->core.system_wide = orig->core.system_wide;
 379         evsel->core.requires_cpu = orig->core.requires_cpu;
 380         evsel->core.is_pmu_core = orig->core.is_pmu_core;
 381
 382         if (orig->name) {
 383                 evsel->name = strdup(orig->name);
 384                 if (evsel->name == NULL)
 385                         goto out_err;
 386         }
 387         if (orig->group_name) {
 388                 evsel->group_name = strdup(orig->group_name);
 389                 if (evsel->group_name == NULL)
 390                         goto out_err;
 391         }
 392         if (orig->pmu_name) {
 393                 evsel->pmu_name = strdup(orig->pmu_name);
 394                 if (evsel->pmu_name == NULL)
 395                         goto out_err;
 396         }
 397         if (orig->group_pmu_name) {
 398                 evsel->group_pmu_name = strdup(orig->group_pmu_name);
 399                 if (evsel->group_pmu_name == NULL)
 400                         goto out_err;
 401         }
 402         if (orig->filter) {
 403                 evsel->filter = strdup(orig->filter);
 404                 if (evsel->filter == NULL)
 405                         goto out_err;
 406         }
 407         if (orig->metric_id) {
 408                 evsel->metric_id = strdup(orig->metric_id);
 409                 if (evsel->metric_id == NULL)
 410                         goto out_err;
 411         }
 412         evsel->cgrp = cgroup__get(orig->cgrp);
 413 #ifdef HAVE_LIBTRACEEVENT
 414         evsel->tp_format = orig->tp_format;
 415 #endif
 416         evsel->handler = orig->handler;
 417         evsel->core.leader = orig->core.leader;
 418
 419         evsel->max_events = orig->max_events;
 420         evsel->tool_event = orig->tool_event;
 421         free((char *)evsel->unit);
 422         evsel->unit = strdup(orig->unit);
 423         if (evsel->unit == NULL)
 424                 goto out_err;
 425
 426         evsel->scale = orig->scale;
 427         evsel->snapshot = orig->snapshot;
 428         evsel->per_pkg = orig->per_pkg;
 429         evsel->percore = orig->percore;
 430         evsel->precise_max = orig->precise_max;
 431         evsel->is_libpfm_event = orig->is_libpfm_event;
 432
 433         evsel->exclude_GH = orig->exclude_GH;
 434         evsel->sample_read = orig->sample_read;
 435         evsel->auto_merge_stats = orig->auto_merge_stats;
 436         evsel->collect_stat = orig->collect_stat;
 437         evsel->weak_group = orig->weak_group;
 438         evsel->use_config_name = orig->use_config_name;
 439         evsel->pmu = orig->pmu;
 440
 441         if (evsel__copy_config_terms(evsel, orig) < 0)
 442                 goto out_err;
 443
 444         return evsel;
 445
 446 out_err:
 447         evsel__delete(evsel);
 448         return NULL;
 449 }
 450
 451 /*
 452  * Returns pointer with encoded error via <linux/err.h> interface.
 453  */
 454 #ifdef HAVE_LIBTRACEEVENT
 455 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 456 {
 457         struct evsel *evsel = zalloc(perf_evsel__object.size);
 458         int err = -ENOMEM;
 459
 460         if (evsel == NULL) {
 461                 goto out_err;
 462         } else {
 463                 struct perf_event_attr attr = {
 464                         .type          = PERF_TYPE_TRACEPOINT,
 465                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 466                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 467                 };
 468
 469                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 470                         goto out_free;
 471
 472                 evsel->tp_format = trace_event__tp_format(sys, name);
 473                 if (IS_ERR(evsel->tp_format)) {
 474                         err = PTR_ERR(evsel->tp_format);
 475                         goto out_free;
 476                 }
 477
 478                 event_attr_init(&attr);
 479                 attr.config = evsel->tp_format->id;
 480                 attr.sample_period = 1;
 481                 evsel__init(evsel, &attr, idx);
 482         }
 483
 484         return evsel;
 485
 486 out_free:
 487         zfree(&evsel->name);
 488         free(evsel);
 489 out_err:
 490         return ERR_PTR(err);
 491 }
 492 #endif
 493
 494 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 495         "cycles",
 496         "instructions",
 497         "cache-references",
 498         "cache-misses",
 499         "branches",
 500         "branch-misses",
 501         "bus-cycles",
 502         "stalled-cycles-frontend",
 503         "stalled-cycles-backend",
 504         "ref-cycles",
 505 };
 506
 507 char *evsel__bpf_counter_events;
 508
 509 bool evsel__match_bpf_counter_events(const char *name)
 510 {
 511         int name_len;
 512         bool match;
 513         char *ptr;
 514
 515         if (!evsel__bpf_counter_events)
 516                 return false;
 517
 518         ptr = strstr(evsel__bpf_counter_events, name);
 519         name_len = strlen(name);
 520
 521         /* check name matches a full token in evsel__bpf_counter_events */
 522         match = (ptr != NULL) &&
 523                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 524                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 525
 526         return match;
 527 }
 528
 529 static const char *__evsel__hw_name(u64 config)
 530 {
 531         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 532                 return evsel__hw_names[config];
 533
 534         return "unknown-hardware";
 535 }
 536
 537 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 538 {
 539         int colon = 0, r = 0;
 540         struct perf_event_attr *attr = &evsel->core.attr;
 541         bool exclude_guest_default = false;
 542
 543 #define MOD_PRINT(context, mod) do {                                    \
 544                 if (!attr->exclude_##context) {                         \
 545                         if (!colon) colon = ++r;                        \
 546                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 547                 } } while(0)
 548
 549         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 550                 MOD_PRINT(kernel, 'k');
 551                 MOD_PRINT(user, 'u');
 552                 MOD_PRINT(hv, 'h');
 553                 exclude_guest_default = true;
 554         }
 555
 556         if (attr->precise_ip) {
 557                 if (!colon)
 558                         colon = ++r;
 559                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 560                 exclude_guest_default = true;
 561         }
 562
 563         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 564                 MOD_PRINT(host, 'H');
 565                 MOD_PRINT(guest, 'G');
 566         }
 567 #undef MOD_PRINT
 568         if (colon)
 569                 bf[colon - 1] = ':';
 570         return r;
 571 }
 572
 573 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 574 {
 575         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 576 }
 577
 578 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 579 {
 580         int r = arch_evsel__hw_name(evsel, bf, size);
 581         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 582 }
 583
 584 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 585         "cpu-clock",
 586         "task-clock",
 587         "page-faults",
 588         "context-switches",
 589         "cpu-migrations",
 590         "minor-faults",
 591         "major-faults",
 592         "alignment-faults",
 593         "emulation-faults",
 594         "dummy",
 595 };
 596
 597 static const char *__evsel__sw_name(u64 config)
 598 {
 599         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 600                 return evsel__sw_names[config];
 601         return "unknown-software";
 602 }
 603
 604 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 605 {
 606         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 607         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 608 }
 609
 610 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 611 {
 612         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 613 }
 614
 615 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 616 {
 617         int r;
 618
 619         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 620
 621         if (type & HW_BREAKPOINT_R)
 622                 r += scnprintf(bf + r, size - r, "r");
 623
 624         if (type & HW_BREAKPOINT_W)
 625                 r += scnprintf(bf + r, size - r, "w");
 626
 627         if (type & HW_BREAKPOINT_X)
 628                 r += scnprintf(bf + r, size - r, "x");
 629
 630         return r;
 631 }
 632
 633 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 634 {
 635         struct perf_event_attr *attr = &evsel->core.attr;
 636         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 637         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 638 }
 639
 640 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 641  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 642  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 643  { "LLC",       "L2",                                                   },
 644  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 645  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 646  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 647  { "node",                                                              },
 648 };
 649
 650 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 651  { "load",      "loads",        "read",                                 },
 652  { "store",     "stores",       "write",                                },
 653  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 654 };
 655
 656 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 657  { "refs",      "Reference",    "ops",          "access",               },
 658  { "misses",    "miss",                                                 },
 659 };
 660
 661 #define C(x)            PERF_COUNT_HW_CACHE_##x
 662 #define CACHE_READ      (1 << C(OP_READ))
 663 #define CACHE_WRITE     (1 << C(OP_WRITE))
 664 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 665 #define COP(x)          (1 << x)
 666
 667 /*
 668  * cache operation stat
 669  * L1I : Read and prefetch only
 670  * ITLB and BPU : Read-only
 671  */
 672 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 673  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 674  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 675  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 676  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 677  [C(ITLB)]      = (CACHE_READ),
 678  [C(BPU)]       = (CACHE_READ),
 679  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 680 };
 681
 682 bool evsel__is_cache_op_valid(u8 type, u8 op)
 683 {
 684         if (evsel__hw_cache_stat[type] & COP(op))
 685                 return true;    /* valid */
 686         else
 687                 return false;   /* invalid */
 688 }
 689
 690 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 691 {
 692         if (result) {
 693                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 694                                  evsel__hw_cache_op[op][0],
 695                                  evsel__hw_cache_result[result][0]);
 696         }
 697
 698         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 699                          evsel__hw_cache_op[op][1]);
 700 }
 701
 702 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 703 {
 704         u8 op, result, type = (config >>  0) & 0xff;
 705         const char *err = "unknown-ext-hardware-cache-type";
 706
 707         if (type >= PERF_COUNT_HW_CACHE_MAX)
 708                 goto out_err;
 709
 710         op = (config >>  8) & 0xff;
 711         err = "unknown-ext-hardware-cache-op";
 712         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 713                 goto out_err;
 714
 715         result = (config >> 16) & 0xff;
 716         err = "unknown-ext-hardware-cache-result";
 717         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 718                 goto out_err;
 719
 720         err = "invalid-cache";
 721         if (!evsel__is_cache_op_valid(type, op))
 722                 goto out_err;
 723
 724         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 725 out_err:
 726         return scnprintf(bf, size, "%s", err);
 727 }
 728
 729 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 730 {
 731         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 732         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 733 }
 734
 735 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 736 {
 737         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 738         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 739 }
 740
 741 const char *evsel__name(struct evsel *evsel)
 742 {
 743         char bf[128];
 744
 745         if (!evsel)
 746                 goto out_unknown;
 747
 748         if (evsel->name)
 749                 return evsel->name;
 750
 751         switch (evsel->core.attr.type) {
 752         case PERF_TYPE_RAW:
 753                 evsel__raw_name(evsel, bf, sizeof(bf));
 754                 break;
 755
 756         case PERF_TYPE_HARDWARE:
 757                 evsel__hw_name(evsel, bf, sizeof(bf));
 758                 break;
 759
 760         case PERF_TYPE_HW_CACHE:
 761                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 762                 break;
 763
 764         case PERF_TYPE_SOFTWARE:
 765                 if (evsel__is_tool(evsel))
 766                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 767                 else
 768                         evsel__sw_name(evsel, bf, sizeof(bf));
 769                 break;
 770
 771         case PERF_TYPE_TRACEPOINT:
 772                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 773                 break;
 774
 775         case PERF_TYPE_BREAKPOINT:
 776                 evsel__bp_name(evsel, bf, sizeof(bf));
 777                 break;
 778
 779         default:
 780                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 781                           evsel->core.attr.type);
 782                 break;
 783         }
 784
 785         evsel->name = strdup(bf);
 786
 787         if (evsel->name)
 788                 return evsel->name;
 789 out_unknown:
 790         return "unknown";
 791 }
 792
 793 bool evsel__name_is(struct evsel *evsel, const char *name)
 794 {
 795         return !strcmp(evsel__name(evsel), name);
 796 }
 797
 798 const char *evsel__metric_id(const struct evsel *evsel)
 799 {
 800         if (evsel->metric_id)
 801                 return evsel->metric_id;
 802
 803         if (evsel__is_tool(evsel))
 804                 return perf_tool_event__to_str(evsel->tool_event);
 805
 806         return "unknown";
 807 }
 808
 809 const char *evsel__group_name(struct evsel *evsel)
 810 {
 811         return evsel->group_name ?: "anon group";
 812 }
 813
 814 /*
 815  * Returns the group details for the specified leader,
 816  * with following rules.
 817  *
 818  *  For record -e '{cycles,instructions}'
 819  *    'anon group { cycles:u, instructions:u }'
 820  *
 821  *  For record -e 'cycles,instructions' and report --group
 822  *    'cycles:u, instructions:u'
 823  */
 824 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 825 {
 826         int ret = 0;
 827         struct evsel *pos;
 828         const char *group_name = evsel__group_name(evsel);
 829
 830         if (!evsel->forced_leader)
 831                 ret = scnprintf(buf, size, "%s { ", group_name);
 832
 833         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 834
 835         for_each_group_member(pos, evsel)
 836                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 837
 838         if (!evsel->forced_leader)
 839                 ret += scnprintf(buf + ret, size - ret, " }");
 840
 841         return ret;
 842 }
 843
 844 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 845                                       struct callchain_param *param)
 846 {
 847         bool function = evsel__is_function_event(evsel);
 848         struct perf_event_attr *attr = &evsel->core.attr;
 849         const char *arch = perf_env__arch(evsel__env(evsel));
 850
 851         evsel__set_sample_bit(evsel, CALLCHAIN);
 852
 853         attr->sample_max_stack = param->max_stack;
 854
 855         if (opts->kernel_callchains)
 856                 attr->exclude_callchain_user = 1;
 857         if (opts->user_callchains)
 858                 attr->exclude_callchain_kernel = 1;
 859         if (param->record_mode == CALLCHAIN_LBR) {
 860                 if (!opts->branch_stack) {
 861                         if (attr->exclude_user) {
 862                                 pr_warning("LBR callstack option is only available "
 863                                            "to get user callchain information. "
 864                                            "Falling back to framepointers.\n");
 865                         } else {
 866                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 867                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 868                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 869                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 870                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 871                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 872                         }
 873                 } else
 874                          pr_warning("Cannot use LBR callstack with branch stack. "
 875                                     "Falling back to framepointers.\n");
 876         }
 877
 878         if (param->record_mode == CALLCHAIN_DWARF) {
 879                 if (!function) {
 880                         evsel__set_sample_bit(evsel, REGS_USER);
 881                         evsel__set_sample_bit(evsel, STACK_USER);
 882                         if (opts->sample_user_regs &&
 883                             DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) {
 884                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch);
 885                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 886                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 887                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 888                         } else {
 889                                 attr->sample_regs_user |= arch__user_reg_mask();
 890                         }
 891                         attr->sample_stack_user = param->dump_size;
 892                         attr->exclude_callchain_user = 1;
 893                 } else {
 894                         pr_info("Cannot use DWARF unwind for function trace event,"
 895                                 " falling back to framepointers.\n");
 896                 }
 897         }
 898
 899         if (function) {
 900                 pr_info("Disabling user space callchains for function trace event.\n");
 901                 attr->exclude_callchain_user = 1;
 902         }
 903 }
 904
 905 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 906                              struct callchain_param *param)
 907 {
 908         if (param->enabled)
 909                 return __evsel__config_callchain(evsel, opts, param);
 910 }
 911
 912 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 913 {
 914         struct perf_event_attr *attr = &evsel->core.attr;
 915
 916         evsel__reset_sample_bit(evsel, CALLCHAIN);
 917         if (param->record_mode == CALLCHAIN_LBR) {
 918                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 919                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 920                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 921                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 922         }
 923         if (param->record_mode == CALLCHAIN_DWARF) {
 924                 evsel__reset_sample_bit(evsel, REGS_USER);
 925                 evsel__reset_sample_bit(evsel, STACK_USER);
 926         }
 927 }
 928
 929 static void evsel__apply_config_terms(struct evsel *evsel,
 930                                       struct record_opts *opts, bool track)
 931 {
 932         struct evsel_config_term *term;
 933         struct list_head *config_terms = &evsel->config_terms;
 934         struct perf_event_attr *attr = &evsel->core.attr;
 935         /* callgraph default */
 936         struct callchain_param param = {
 937                 .record_mode = callchain_param.record_mode,
 938         };
 939         u32 dump_size = 0;
 940         int max_stack = 0;
 941         const char *callgraph_buf = NULL;
 942
 943         list_for_each_entry(term, config_terms, list) {
 944                 switch (term->type) {
 945                 case EVSEL__CONFIG_TERM_PERIOD:
 946                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 947                                 attr->sample_period = term->val.period;
 948                                 attr->freq = 0;
 949                                 evsel__reset_sample_bit(evsel, PERIOD);
 950                         }
 951                         break;
 952                 case EVSEL__CONFIG_TERM_FREQ:
 953                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 954                                 attr->sample_freq = term->val.freq;
 955                                 attr->freq = 1;
 956                                 evsel__set_sample_bit(evsel, PERIOD);
 957                         }
 958                         break;
 959                 case EVSEL__CONFIG_TERM_TIME:
 960                         if (term->val.time)
 961                                 evsel__set_sample_bit(evsel, TIME);
 962                         else
 963                                 evsel__reset_sample_bit(evsel, TIME);
 964                         break;
 965                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 966                         callgraph_buf = term->val.str;
 967                         break;
 968                 case EVSEL__CONFIG_TERM_BRANCH:
 969                         if (term->val.str && strcmp(term->val.str, "no")) {
 970                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 971                                 parse_branch_str(term->val.str,
 972                                                  &attr->branch_sample_type);
 973                         } else
 974                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 975                         break;
 976                 case EVSEL__CONFIG_TERM_STACK_USER:
 977                         dump_size = term->val.stack_user;
 978                         break;
 979                 case EVSEL__CONFIG_TERM_MAX_STACK:
 980                         max_stack = term->val.max_stack;
 981                         break;
 982                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 983                         evsel->max_events = term->val.max_events;
 984                         break;
 985                 case EVSEL__CONFIG_TERM_INHERIT:
 986                         /*
 987                          * attr->inherit should has already been set by
 988                          * evsel__config. If user explicitly set
 989                          * inherit using config terms, override global
 990                          * opt->no_inherit setting.
 991                          */
 992                         attr->inherit = term->val.inherit ? 1 : 0;
 993                         break;
 994                 case EVSEL__CONFIG_TERM_OVERWRITE:
 995                         attr->write_backward = term->val.overwrite ? 1 : 0;
 996                         break;
 997                 case EVSEL__CONFIG_TERM_DRV_CFG:
 998                         break;
 999                 case EVSEL__CONFIG_TERM_PERCORE:
1000                         break;
1001                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1002                         attr->aux_output = term->val.aux_output ? 1 : 0;
1003                         break;
1004                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1005                         /* Already applied by auxtrace */
1006                         break;
1007                 case EVSEL__CONFIG_TERM_CFG_CHG:
1008                         break;
1009                 default:
1010                         break;
1011                 }
1012         }
1013
1014         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1015         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1016                 bool sample_address = false;
1017
1018                 if (max_stack) {
1019                         param.max_stack = max_stack;
1020                         if (callgraph_buf == NULL)
1021                                 callgraph_buf = "fp";
1022                 }
1023
1024                 /* parse callgraph parameters */
1025                 if (callgraph_buf != NULL) {
1026                         if (!strcmp(callgraph_buf, "no")) {
1027                                 param.enabled = false;
1028                                 param.record_mode = CALLCHAIN_NONE;
1029                         } else {
1030                                 param.enabled = true;
1031                                 if (parse_callchain_record(callgraph_buf, &param)) {
1032                                         pr_err("per-event callgraph setting for %s failed. "
1033                                                "Apply callgraph global setting for it\n",
1034                                                evsel->name);
1035                                         return;
1036                                 }
1037                                 if (param.record_mode == CALLCHAIN_DWARF)
1038                                         sample_address = true;
1039                         }
1040                 }
1041                 if (dump_size > 0) {
1042                         dump_size = round_up(dump_size, sizeof(u64));
1043                         param.dump_size = dump_size;
1044                 }
1045
1046                 /* If global callgraph set, clear it */
1047                 if (callchain_param.enabled)
1048                         evsel__reset_callgraph(evsel, &callchain_param);
1049
1050                 /* set perf-event callgraph */
1051                 if (param.enabled) {
1052                         if (sample_address) {
1053                                 evsel__set_sample_bit(evsel, ADDR);
1054                                 evsel__set_sample_bit(evsel, DATA_SRC);
1055                                 evsel->core.attr.mmap_data = track;
1056                         }
1057                         evsel__config_callchain(evsel, opts, &param);
1058                 }
1059         }
1060 }
1061
1062 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1063 {
1064         struct evsel_config_term *term, *found_term = NULL;
1065
1066         list_for_each_entry(term, &evsel->config_terms, list) {
1067                 if (term->type == type)
1068                         found_term = term;
1069         }
1070
1071         return found_term;
1072 }
1073
1074 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1075 {
1076         evsel__set_sample_bit(evsel, WEIGHT);
1077 }
1078
1079 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1080                                     struct perf_event_attr *attr __maybe_unused)
1081 {
1082 }
1083
1084 static void evsel__set_default_freq_period(struct record_opts *opts,
1085                                            struct perf_event_attr *attr)
1086 {
1087         if (opts->freq) {
1088                 attr->freq = 1;
1089                 attr->sample_freq = opts->freq;
1090         } else {
1091                 attr->sample_period = opts->default_interval;
1092         }
1093 }
1094
1095 static bool evsel__is_offcpu_event(struct evsel *evsel)
1096 {
1097         return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1098 }
1099
1100 /*
1101  * The enable_on_exec/disabled value strategy:
1102  *
1103  *  1) For any type of traced program:
1104  *    - all independent events and group leaders are disabled
1105  *    - all group members are enabled
1106  *
1107  *     Group members are ruled by group leaders. They need to
1108  *     be enabled, because the group scheduling relies on that.
1109  *
1110  *  2) For traced programs executed by perf:
1111  *     - all independent events and group leaders have
1112  *       enable_on_exec set
1113  *     - we don't specifically enable or disable any event during
1114  *       the record command
1115  *
1116  *     Independent events and group leaders are initially disabled
1117  *     and get enabled by exec. Group members are ruled by group
1118  *     leaders as stated in 1).
1119  *
1120  *  3) For traced programs attached by perf (pid/tid):
1121  *     - we specifically enable or disable all events during
1122  *       the record command
1123  *
1124  *     When attaching events to already running traced we
1125  *     enable/disable events specifically, as there's no
1126  *     initial traced exec call.
1127  */
1128 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1129                    struct callchain_param *callchain)
1130 {
1131         struct evsel *leader = evsel__leader(evsel);
1132         struct perf_event_attr *attr = &evsel->core.attr;
1133         int track = evsel->tracking;
1134         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1135
1136         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1137         attr->inherit       = !opts->no_inherit;
1138         attr->write_backward = opts->overwrite ? 1 : 0;
1139         attr->read_format   = PERF_FORMAT_LOST;
1140
1141         evsel__set_sample_bit(evsel, IP);
1142         evsel__set_sample_bit(evsel, TID);
1143
1144         if (evsel->sample_read) {
1145                 evsel__set_sample_bit(evsel, READ);
1146
1147                 /*
1148                  * We need ID even in case of single event, because
1149                  * PERF_SAMPLE_READ process ID specific data.
1150                  */
1151                 evsel__set_sample_id(evsel, false);
1152
1153                 /*
1154                  * Apply group format only if we belong to group
1155                  * with more than one members.
1156                  */
1157                 if (leader->core.nr_members > 1) {
1158                         attr->read_format |= PERF_FORMAT_GROUP;
1159                         attr->inherit = 0;
1160                 }
1161         }
1162
1163         /*
1164          * We default some events to have a default interval. But keep
1165          * it a weak assumption overridable by the user.
1166          */
1167         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1168             (!evsel->is_libpfm_event && (!attr->sample_period ||
1169                                          opts->user_freq != UINT_MAX ||
1170                                          opts->user_interval != ULLONG_MAX)))
1171                 evsel__set_default_freq_period(opts, attr);
1172
1173         /*
1174          * If attr->freq was set (here or earlier), ask for period
1175          * to be sampled.
1176          */
1177         if (attr->freq)
1178                 evsel__set_sample_bit(evsel, PERIOD);
1179
1180         if (opts->no_samples)
1181                 attr->sample_freq = 0;
1182
1183         if (opts->inherit_stat) {
1184                 evsel->core.attr.read_format |=
1185                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1186                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1187                         PERF_FORMAT_ID;
1188                 attr->inherit_stat = 1;
1189         }
1190
1191         if (opts->sample_address) {
1192                 evsel__set_sample_bit(evsel, ADDR);
1193                 attr->mmap_data = track;
1194         }
1195
1196         /*
1197          * We don't allow user space callchains for  function trace
1198          * event, due to issues with page faults while tracing page
1199          * fault handler and its overall trickiness nature.
1200          */
1201         if (evsel__is_function_event(evsel))
1202                 evsel->core.attr.exclude_callchain_user = 1;
1203
1204         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1205                 evsel__config_callchain(evsel, opts, callchain);
1206
1207         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1208             !evsel__is_dummy_event(evsel)) {
1209                 attr->sample_regs_intr = opts->sample_intr_regs;
1210                 evsel__set_sample_bit(evsel, REGS_INTR);
1211         }
1212
1213         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1214             !evsel__is_dummy_event(evsel)) {
1215                 attr->sample_regs_user |= opts->sample_user_regs;
1216                 evsel__set_sample_bit(evsel, REGS_USER);
1217         }
1218
1219         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1220                 evsel__set_sample_bit(evsel, CPU);
1221
1222         /*
1223          * When the user explicitly disabled time don't force it here.
1224          */
1225         if (opts->sample_time &&
1226             (!perf_missing_features.sample_id_all &&
1227             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1228              opts->sample_time_set)))
1229                 evsel__set_sample_bit(evsel, TIME);
1230
1231         if (opts->raw_samples && !evsel->no_aux_samples) {
1232                 evsel__set_sample_bit(evsel, TIME);
1233                 evsel__set_sample_bit(evsel, RAW);
1234                 evsel__set_sample_bit(evsel, CPU);
1235         }
1236
1237         if (opts->sample_address)
1238                 evsel__set_sample_bit(evsel, DATA_SRC);
1239
1240         if (opts->sample_phys_addr)
1241                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1242
1243         if (opts->no_buffering) {
1244                 attr->watermark = 0;
1245                 attr->wakeup_events = 1;
1246         }
1247         if (opts->branch_stack && !evsel->no_aux_samples) {
1248                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1249                 attr->branch_sample_type = opts->branch_stack;
1250         }
1251
1252         if (opts->sample_weight)
1253                 arch_evsel__set_sample_weight(evsel);
1254
1255         attr->task     = track;
1256         attr->mmap     = track;
1257         attr->mmap2    = track && !perf_missing_features.mmap2;
1258         attr->comm     = track;
1259         attr->build_id = track && opts->build_id;
1260
1261         /*
1262          * ksymbol is tracked separately with text poke because it needs to be
1263          * system wide and enabled immediately.
1264          */
1265         if (!opts->text_poke)
1266                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1267         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1268
1269         if (opts->record_namespaces)
1270                 attr->namespaces  = track;
1271
1272         if (opts->record_cgroup) {
1273                 attr->cgroup = track && !perf_missing_features.cgroup;
1274                 evsel__set_sample_bit(evsel, CGROUP);
1275         }
1276
1277         if (opts->sample_data_page_size)
1278                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1279
1280         if (opts->sample_code_page_size)
1281                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1282
1283         if (opts->record_switch_events)
1284                 attr->context_switch = track;
1285
1286         if (opts->sample_transaction)
1287                 evsel__set_sample_bit(evsel, TRANSACTION);
1288
1289         if (opts->running_time) {
1290                 evsel->core.attr.read_format |=
1291                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1292                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1293         }
1294
1295         /*
1296          * XXX see the function comment above
1297          *
1298          * Disabling only independent events or group leaders,
1299          * keeping group members enabled.
1300          */
1301         if (evsel__is_group_leader(evsel))
1302                 attr->disabled = 1;
1303
1304         /*
1305          * Setting enable_on_exec for independent events and
1306          * group leaders for traced executed by perf.
1307          */
1308         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1309             !opts->target.initial_delay)
1310                 attr->enable_on_exec = 1;
1311
1312         if (evsel->immediate) {
1313                 attr->disabled = 0;
1314                 attr->enable_on_exec = 0;
1315         }
1316
1317         clockid = opts->clockid;
1318         if (opts->use_clockid) {
1319                 attr->use_clockid = 1;
1320                 attr->clockid = opts->clockid;
1321         }
1322
1323         if (evsel->precise_max)
1324                 attr->precise_ip = 3;
1325
1326         if (opts->all_user) {
1327                 attr->exclude_kernel = 1;
1328                 attr->exclude_user   = 0;
1329         }
1330
1331         if (opts->all_kernel) {
1332                 attr->exclude_kernel = 0;
1333                 attr->exclude_user   = 1;
1334         }
1335
1336         if (evsel->core.own_cpus || evsel->unit)
1337                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1338
1339         /*
1340          * Apply event specific term settings,
1341          * it overloads any global configuration.
1342          */
1343         evsel__apply_config_terms(evsel, opts, track);
1344
1345         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1346
1347         /* The --period option takes the precedence. */
1348         if (opts->period_set) {
1349                 if (opts->period)
1350                         evsel__set_sample_bit(evsel, PERIOD);
1351                 else
1352                         evsel__reset_sample_bit(evsel, PERIOD);
1353         }
1354
1355         /*
1356          * A dummy event never triggers any actual counter and therefore
1357          * cannot be used with branch_stack.
1358          *
1359          * For initial_delay, a dummy event is added implicitly.
1360          * The software event will trigger -EOPNOTSUPP error out,
1361          * if BRANCH_STACK bit is set.
1362          */
1363         if (evsel__is_dummy_event(evsel))
1364                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1365
1366         if (evsel__is_offcpu_event(evsel))
1367                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1368
1369         arch__post_evsel_config(evsel, attr);
1370 }
1371
1372 int evsel__set_filter(struct evsel *evsel, const char *filter)
1373 {
1374         char *new_filter = strdup(filter);
1375
1376         if (new_filter != NULL) {
1377                 free(evsel->filter);
1378                 evsel->filter = new_filter;
1379                 return 0;
1380         }
1381
1382         return -1;
1383 }
1384
1385 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1386 {
1387         char *new_filter;
1388
1389         if (evsel->filter == NULL)
1390                 return evsel__set_filter(evsel, filter);
1391
1392         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1393                 free(evsel->filter);
1394                 evsel->filter = new_filter;
1395                 return 0;
1396         }
1397
1398         return -1;
1399 }
1400
1401 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1402 {
1403         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1404 }
1405
1406 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1407 {
1408         return evsel__append_filter(evsel, "%s,%s", filter);
1409 }
1410
1411 /* Caller has to clear disabled after going through all CPUs. */
1412 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1413 {
1414         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1415 }
1416
1417 int evsel__enable(struct evsel *evsel)
1418 {
1419         int err = perf_evsel__enable(&evsel->core);
1420
1421         if (!err)
1422                 evsel->disabled = false;
1423         return err;
1424 }
1425
1426 /* Caller has to set disabled after going through all CPUs. */
1427 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1428 {
1429         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1430 }
1431
1432 int evsel__disable(struct evsel *evsel)
1433 {
1434         int err = perf_evsel__disable(&evsel->core);
1435         /*
1436          * We mark it disabled here so that tools that disable a event can
1437          * ignore events after they disable it. I.e. the ring buffer may have
1438          * already a few more events queued up before the kernel got the stop
1439          * request.
1440          */
1441         if (!err)
1442                 evsel->disabled = true;
1443
1444         return err;
1445 }
1446
1447 void free_config_terms(struct list_head *config_terms)
1448 {
1449         struct evsel_config_term *term, *h;
1450
1451         list_for_each_entry_safe(term, h, config_terms, list) {
1452                 list_del_init(&term->list);
1453                 if (term->free_str)
1454                         zfree(&term->val.str);
1455                 free(term);
1456         }
1457 }
1458
1459 static void evsel__free_config_terms(struct evsel *evsel)
1460 {
1461         free_config_terms(&evsel->config_terms);
1462 }
1463
1464 void evsel__exit(struct evsel *evsel)
1465 {
1466         assert(list_empty(&evsel->core.node));
1467         assert(evsel->evlist == NULL);
1468         bpf_counter__destroy(evsel);
1469         perf_bpf_filter__destroy(evsel);
1470         evsel__free_counts(evsel);
1471         perf_evsel__free_fd(&evsel->core);
1472         perf_evsel__free_id(&evsel->core);
1473         evsel__free_config_terms(evsel);
1474         cgroup__put(evsel->cgrp);
1475         perf_cpu_map__put(evsel->core.cpus);
1476         perf_cpu_map__put(evsel->core.own_cpus);
1477         perf_thread_map__put(evsel->core.threads);
1478         zfree(&evsel->group_name);
1479         zfree(&evsel->name);
1480         zfree(&evsel->filter);
1481         zfree(&evsel->pmu_name);
1482         zfree(&evsel->group_pmu_name);
1483         zfree(&evsel->unit);
1484         zfree(&evsel->metric_id);
1485         evsel__zero_per_pkg(evsel);
1486         hashmap__free(evsel->per_pkg_mask);
1487         evsel->per_pkg_mask = NULL;
1488         zfree(&evsel->metric_events);
1489         perf_evsel__object.fini(evsel);
1490 }
1491
1492 void evsel__delete(struct evsel *evsel)
1493 {
1494         if (!evsel)
1495                 return;
1496
1497         evsel__exit(evsel);
1498         free(evsel);
1499 }
1500
1501 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1502                            struct perf_counts_values *count)
1503 {
1504         struct perf_counts_values tmp;
1505
1506         if (!evsel->prev_raw_counts)
1507                 return;
1508
1509         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1510         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1511
1512         count->val = count->val - tmp.val;
1513         count->ena = count->ena - tmp.ena;
1514         count->run = count->run - tmp.run;
1515 }
1516
1517 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1518 {
1519         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1520
1521         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1522 }
1523
1524 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1525                              u64 val, u64 ena, u64 run, u64 lost)
1526 {
1527         struct perf_counts_values *count;
1528
1529         count = perf_counts(counter->counts, cpu_map_idx, thread);
1530
1531         count->val    = val;
1532         count->ena    = ena;
1533         count->run    = run;
1534         count->lost   = lost;
1535
1536         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1537 }
1538
1539 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1540 {
1541         u64 read_format = leader->core.attr.read_format;
1542         struct sample_read_value *v;
1543         u64 nr, ena = 0, run = 0, lost = 0;
1544
1545         nr = *data++;
1546
1547         if (nr != (u64) leader->core.nr_members)
1548                 return -EINVAL;
1549
1550         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1551                 ena = *data++;
1552
1553         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1554                 run = *data++;
1555
1556         v = (void *)data;
1557         sample_read_group__for_each(v, nr, read_format) {
1558                 struct evsel *counter;
1559
1560                 counter = evlist__id2evsel(leader->evlist, v->id);
1561                 if (!counter)
1562                         return -EINVAL;
1563
1564                 if (read_format & PERF_FORMAT_LOST)
1565                         lost = v->lost;
1566
1567                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1568         }
1569
1570         return 0;
1571 }
1572
1573 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1574 {
1575         struct perf_stat_evsel *ps = leader->stats;
1576         u64 read_format = leader->core.attr.read_format;
1577         int size = perf_evsel__read_size(&leader->core);
1578         u64 *data = ps->group_data;
1579
1580         if (!(read_format & PERF_FORMAT_ID))
1581                 return -EINVAL;
1582
1583         if (!evsel__is_group_leader(leader))
1584                 return -EINVAL;
1585
1586         if (!data) {
1587                 data = zalloc(size);
1588                 if (!data)
1589                         return -ENOMEM;
1590
1591                 ps->group_data = data;
1592         }
1593
1594         if (FD(leader, cpu_map_idx, thread) < 0)
1595                 return -EINVAL;
1596
1597         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1598                 return -errno;
1599
1600         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1601 }
1602
1603 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1604 {
1605         u64 read_format = evsel->core.attr.read_format;
1606
1607         if (read_format & PERF_FORMAT_GROUP)
1608                 return evsel__read_group(evsel, cpu_map_idx, thread);
1609
1610         return evsel__read_one(evsel, cpu_map_idx, thread);
1611 }
1612
1613 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1614 {
1615         struct perf_counts_values count;
1616         size_t nv = scale ? 3 : 1;
1617
1618         if (FD(evsel, cpu_map_idx, thread) < 0)
1619                 return -EINVAL;
1620
1621         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1622                 return -ENOMEM;
1623
1624         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1625                 return -errno;
1626
1627         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1628         perf_counts_values__scale(&count, scale, NULL);
1629         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1630         return 0;
1631 }
1632
1633 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1634                                   int cpu_map_idx)
1635 {
1636         struct perf_cpu cpu;
1637
1638         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1639         return perf_cpu_map__idx(other->core.cpus, cpu);
1640 }
1641
1642 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1643 {
1644         struct evsel *leader = evsel__leader(evsel);
1645
1646         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1647             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1648                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1649         }
1650
1651         return cpu_map_idx;
1652 }
1653
1654 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1655 {
1656         struct evsel *leader = evsel__leader(evsel);
1657         int fd;
1658
1659         if (evsel__is_group_leader(evsel))
1660                 return -1;
1661
1662         /*
1663          * Leader must be already processed/open,
1664          * if not it's a bug.
1665          */
1666         BUG_ON(!leader->core.fd);
1667
1668         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1669         if (cpu_map_idx == -1)
1670                 return -1;
1671
1672         fd = FD(leader, cpu_map_idx, thread);
1673         BUG_ON(fd == -1 && !leader->skippable);
1674
1675         /*
1676          * When the leader has been skipped, return -2 to distinguish from no
1677          * group leader case.
1678          */
1679         return fd == -1 ? -2 : fd;
1680 }
1681
1682 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1683 {
1684         for (int cpu = 0; cpu < nr_cpus; cpu++)
1685                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1686                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1687 }
1688
1689 static int update_fds(struct evsel *evsel,
1690                       int nr_cpus, int cpu_map_idx,
1691                       int nr_threads, int thread_idx)
1692 {
1693         struct evsel *pos;
1694
1695         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1696                 return -EINVAL;
1697
1698         evlist__for_each_entry(evsel->evlist, pos) {
1699                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1700
1701                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1702
1703                 /*
1704                  * Since fds for next evsel has not been created,
1705                  * there is no need to iterate whole event list.
1706                  */
1707                 if (pos == evsel)
1708                         break;
1709         }
1710         return 0;
1711 }
1712
1713 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1714                                          int nr_cpus, int cpu_map_idx,
1715                                          struct perf_thread_map *threads,
1716                                          int thread, int err)
1717 {
1718         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1719
1720         if (!evsel->ignore_missing_thread)
1721                 return false;
1722
1723         /* The system wide setup does not work with threads. */
1724         if (evsel->core.system_wide)
1725                 return false;
1726
1727         /* The -ESRCH is perf event syscall errno for pid's not found. */
1728         if (err != -ESRCH)
1729                 return false;
1730
1731         /* If there's only one thread, let it fail. */
1732         if (threads->nr == 1)
1733                 return false;
1734
1735         /*
1736          * We should remove fd for missing_thread first
1737          * because thread_map__remove() will decrease threads->nr.
1738          */
1739         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1740                 return false;
1741
1742         if (thread_map__remove(threads, thread))
1743                 return false;
1744
1745         pr_warning("WARNING: Ignored open failure for pid %d\n",
1746                    ignore_pid);
1747         return true;
1748 }
1749
1750 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1751                                 void *priv __maybe_unused)
1752 {
1753         return fprintf(fp, "  %-32s %s\n", name, val);
1754 }
1755
1756 static void display_attr(struct perf_event_attr *attr)
1757 {
1758         if (verbose >= 2 || debug_peo_args) {
1759                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1760                 fprintf(stderr, "perf_event_attr:\n");
1761                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1762                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1763         }
1764 }
1765
1766 bool evsel__precise_ip_fallback(struct evsel *evsel)
1767 {
1768         /* Do not try less precise if not requested. */
1769         if (!evsel->precise_max)
1770                 return false;
1771
1772         /*
1773          * We tried all the precise_ip values, and it's
1774          * still failing, so leave it to standard fallback.
1775          */
1776         if (!evsel->core.attr.precise_ip) {
1777                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1778                 return false;
1779         }
1780
1781         if (!evsel->precise_ip_original)
1782                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1783
1784         evsel->core.attr.precise_ip--;
1785         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1786         display_attr(&evsel->core.attr);
1787         return true;
1788 }
1789
1790 static struct perf_cpu_map *empty_cpu_map;
1791 static struct perf_thread_map *empty_thread_map;
1792
1793 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1794                 struct perf_thread_map *threads)
1795 {
1796         int nthreads = perf_thread_map__nr(threads);
1797
1798         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1799             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1800                 return -EINVAL;
1801
1802         if (cpus == NULL) {
1803                 if (empty_cpu_map == NULL) {
1804                         empty_cpu_map = perf_cpu_map__dummy_new();
1805                         if (empty_cpu_map == NULL)
1806                                 return -ENOMEM;
1807                 }
1808
1809                 cpus = empty_cpu_map;
1810         }
1811
1812         if (threads == NULL) {
1813                 if (empty_thread_map == NULL) {
1814                         empty_thread_map = thread_map__new_by_tid(-1);
1815                         if (empty_thread_map == NULL)
1816                                 return -ENOMEM;
1817                 }
1818
1819                 threads = empty_thread_map;
1820         }
1821
1822         if (evsel->core.fd == NULL &&
1823             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1824                 return -ENOMEM;
1825
1826         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1827         if (evsel->cgrp)
1828                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1829
1830         return 0;
1831 }
1832
1833 static void evsel__disable_missing_features(struct evsel *evsel)
1834 {
1835         if (perf_missing_features.read_lost)
1836                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1837         if (perf_missing_features.weight_struct) {
1838                 evsel__set_sample_bit(evsel, WEIGHT);
1839                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1840         }
1841         if (perf_missing_features.clockid_wrong)
1842                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1843         if (perf_missing_features.clockid) {
1844                 evsel->core.attr.use_clockid = 0;
1845                 evsel->core.attr.clockid = 0;
1846         }
1847         if (perf_missing_features.cloexec)
1848                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1849         if (perf_missing_features.mmap2)
1850                 evsel->core.attr.mmap2 = 0;
1851         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1852                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1853         if (perf_missing_features.lbr_flags)
1854                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1855                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1856         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1857                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1858         if (perf_missing_features.ksymbol)
1859                 evsel->core.attr.ksymbol = 0;
1860         if (perf_missing_features.bpf)
1861                 evsel->core.attr.bpf_event = 0;
1862         if (perf_missing_features.branch_hw_idx)
1863                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1864         if (perf_missing_features.sample_id_all)
1865                 evsel->core.attr.sample_id_all = 0;
1866 }
1867
1868 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1869                         struct perf_thread_map *threads)
1870 {
1871         int err;
1872
1873         err = __evsel__prepare_open(evsel, cpus, threads);
1874         if (err)
1875                 return err;
1876
1877         evsel__disable_missing_features(evsel);
1878
1879         return err;
1880 }
1881
1882 bool evsel__detect_missing_features(struct evsel *evsel)
1883 {
1884         /*
1885          * Must probe features in the order they were added to the
1886          * perf_event_attr interface.
1887          */
1888         if (!perf_missing_features.read_lost &&
1889             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1890                 perf_missing_features.read_lost = true;
1891                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1892                 return true;
1893         } else if (!perf_missing_features.weight_struct &&
1894             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1895                 perf_missing_features.weight_struct = true;
1896                 pr_debug2("switching off weight struct support\n");
1897                 return true;
1898         } else if (!perf_missing_features.code_page_size &&
1899             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1900                 perf_missing_features.code_page_size = true;
1901                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1902                 return false;
1903         } else if (!perf_missing_features.data_page_size &&
1904             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1905                 perf_missing_features.data_page_size = true;
1906                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1907                 return false;
1908         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1909                 perf_missing_features.cgroup = true;
1910                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1911                 return false;
1912         } else if (!perf_missing_features.branch_hw_idx &&
1913             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1914                 perf_missing_features.branch_hw_idx = true;
1915                 pr_debug2("switching off branch HW index support\n");
1916                 return true;
1917         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1918                 perf_missing_features.aux_output = true;
1919                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1920                 return false;
1921         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1922                 perf_missing_features.bpf = true;
1923                 pr_debug2_peo("switching off bpf_event\n");
1924                 return true;
1925         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1926                 perf_missing_features.ksymbol = true;
1927                 pr_debug2_peo("switching off ksymbol\n");
1928                 return true;
1929         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1930                 perf_missing_features.write_backward = true;
1931                 pr_debug2_peo("switching off write_backward\n");
1932                 return false;
1933         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1934                 perf_missing_features.clockid_wrong = true;
1935                 pr_debug2_peo("switching off clockid\n");
1936                 return true;
1937         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1938                 perf_missing_features.clockid = true;
1939                 pr_debug2_peo("switching off use_clockid\n");
1940                 return true;
1941         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1942                 perf_missing_features.cloexec = true;
1943                 pr_debug2_peo("switching off cloexec flag\n");
1944                 return true;
1945         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1946                 perf_missing_features.mmap2 = true;
1947                 pr_debug2_peo("switching off mmap2\n");
1948                 return true;
1949         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1950                 if (evsel->pmu == NULL)
1951                         evsel->pmu = evsel__find_pmu(evsel);
1952
1953                 if (evsel->pmu)
1954                         evsel->pmu->missing_features.exclude_guest = true;
1955                 else {
1956                         /* we cannot find PMU, disable attrs now */
1957                         evsel->core.attr.exclude_host = false;
1958                         evsel->core.attr.exclude_guest = false;
1959                 }
1960
1961                 if (evsel->exclude_GH) {
1962                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1963                         return false;
1964                 }
1965                 if (!perf_missing_features.exclude_guest) {
1966                         perf_missing_features.exclude_guest = true;
1967                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1968                 }
1969                 return true;
1970         } else if (!perf_missing_features.sample_id_all) {
1971                 perf_missing_features.sample_id_all = true;
1972                 pr_debug2_peo("switching off sample_id_all\n");
1973                 return true;
1974         } else if (!perf_missing_features.lbr_flags &&
1975                         (evsel->core.attr.branch_sample_type &
1976                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1977                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1978                 perf_missing_features.lbr_flags = true;
1979                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1980                 return true;
1981         } else if (!perf_missing_features.group_read &&
1982                     evsel->core.attr.inherit &&
1983                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1984                    evsel__is_group_leader(evsel)) {
1985                 perf_missing_features.group_read = true;
1986                 pr_debug2_peo("switching off group read\n");
1987                 return true;
1988         } else {
1989                 return false;
1990         }
1991 }
1992
1993 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1994                 struct perf_thread_map *threads,
1995                 int start_cpu_map_idx, int end_cpu_map_idx)
1996 {
1997         int idx, thread, nthreads;
1998         int pid = -1, err, old_errno;
1999         enum rlimit_action set_rlimit = NO_CHANGE;
2000
2001         err = __evsel__prepare_open(evsel, cpus, threads);
2002         if (err)
2003                 return err;
2004
2005         if (cpus == NULL)
2006                 cpus = empty_cpu_map;
2007
2008         if (threads == NULL)
2009                 threads = empty_thread_map;
2010
2011         nthreads = perf_thread_map__nr(threads);
2012
2013         if (evsel->cgrp)
2014                 pid = evsel->cgrp->fd;
2015
2016 fallback_missing_features:
2017         evsel__disable_missing_features(evsel);
2018
2019         pr_debug3("Opening: %s\n", evsel__name(evsel));
2020         display_attr(&evsel->core.attr);
2021
2022         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2023
2024                 for (thread = 0; thread < nthreads; thread++) {
2025                         int fd, group_fd;
2026 retry_open:
2027                         if (thread >= nthreads)
2028                                 break;
2029
2030                         if (!evsel->cgrp && !evsel->core.system_wide)
2031                                 pid = perf_thread_map__pid(threads, thread);
2032
2033                         group_fd = get_group_fd(evsel, idx, thread);
2034
2035                         if (group_fd == -2) {
2036                                 pr_debug("broken group leader for %s\n", evsel->name);
2037                                 err = -EINVAL;
2038                                 goto out_close;
2039                         }
2040
2041                         test_attr__ready();
2042
2043                         /* Debug message used by test scripts */
2044                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2045                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2046
2047                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2048                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2049                                                 group_fd, evsel->open_flags);
2050
2051                         FD(evsel, idx, thread) = fd;
2052
2053                         if (fd < 0) {
2054                                 err = -errno;
2055
2056                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2057                                           err);
2058                                 goto try_fallback;
2059                         }
2060
2061                         bpf_counter__install_pe(evsel, idx, fd);
2062
2063                         if (unlikely(test_attr__enabled)) {
2064                                 test_attr__open(&evsel->core.attr, pid,
2065                                                 perf_cpu_map__cpu(cpus, idx),
2066                                                 fd, group_fd, evsel->open_flags);
2067                         }
2068
2069                         /* Debug message used by test scripts */
2070                         pr_debug2_peo(" = %d\n", fd);
2071
2072                         if (evsel->bpf_fd >= 0) {
2073                                 int evt_fd = fd;
2074                                 int bpf_fd = evsel->bpf_fd;
2075
2076                                 err = ioctl(evt_fd,
2077                                             PERF_EVENT_IOC_SET_BPF,
2078                                             bpf_fd);
2079                                 if (err && errno != EEXIST) {
2080                                         pr_err("failed to attach bpf fd %d: %s\n",
2081                                                bpf_fd, strerror(errno));
2082                                         err = -EINVAL;
2083                                         goto out_close;
2084                                 }
2085                         }
2086
2087                         set_rlimit = NO_CHANGE;
2088
2089                         /*
2090                          * If we succeeded but had to kill clockid, fail and
2091                          * have evsel__open_strerror() print us a nice error.
2092                          */
2093                         if (perf_missing_features.clockid ||
2094                             perf_missing_features.clockid_wrong) {
2095                                 err = -EINVAL;
2096                                 goto out_close;
2097                         }
2098                 }
2099         }
2100
2101         return 0;
2102
2103 try_fallback:
2104         if (evsel__precise_ip_fallback(evsel))
2105                 goto retry_open;
2106
2107         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2108                                          idx, threads, thread, err)) {
2109                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2110                 nthreads--;
2111
2112                 /* ... and pretend like nothing have happened. */
2113                 err = 0;
2114                 goto retry_open;
2115         }
2116         /*
2117          * perf stat needs between 5 and 22 fds per CPU. When we run out
2118          * of them try to increase the limits.
2119          */
2120         if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit))
2121                 goto retry_open;
2122
2123         if (err != -EINVAL || idx > 0 || thread > 0)
2124                 goto out_close;
2125
2126         if (evsel__detect_missing_features(evsel))
2127                 goto fallback_missing_features;
2128 out_close:
2129         if (err)
2130                 threads->err_thread = thread;
2131
2132         old_errno = errno;
2133         do {
2134                 while (--thread >= 0) {
2135                         if (FD(evsel, idx, thread) >= 0)
2136                                 close(FD(evsel, idx, thread));
2137                         FD(evsel, idx, thread) = -1;
2138                 }
2139                 thread = nthreads;
2140         } while (--idx >= 0);
2141         errno = old_errno;
2142         return err;
2143 }
2144
2145 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2146                 struct perf_thread_map *threads)
2147 {
2148         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2149 }
2150
2151 void evsel__close(struct evsel *evsel)
2152 {
2153         perf_evsel__close(&evsel->core);
2154         perf_evsel__free_id(&evsel->core);
2155 }
2156
2157 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2158 {
2159         if (cpu_map_idx == -1)
2160                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2161
2162         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2163 }
2164
2165 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2166 {
2167         return evsel__open(evsel, NULL, threads);
2168 }
2169
2170 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2171                                        const union perf_event *event,
2172                                        struct perf_sample *sample)
2173 {
2174         u64 type = evsel->core.attr.sample_type;
2175         const __u64 *array = event->sample.array;
2176         bool swapped = evsel->needs_swap;
2177         union u64_swap u;
2178
2179         array += ((event->header.size -
2180                    sizeof(event->header)) / sizeof(u64)) - 1;
2181
2182         if (type & PERF_SAMPLE_IDENTIFIER) {
2183                 sample->id = *array;
2184                 array--;
2185         }
2186
2187         if (type & PERF_SAMPLE_CPU) {
2188                 u.val64 = *array;
2189                 if (swapped) {
2190                         /* undo swap of u64, then swap on individual u32s */
2191                         u.val64 = bswap_64(u.val64);
2192                         u.val32[0] = bswap_32(u.val32[0]);
2193                 }
2194
2195                 sample->cpu = u.val32[0];
2196                 array--;
2197         }
2198
2199         if (type & PERF_SAMPLE_STREAM_ID) {
2200                 sample->stream_id = *array;
2201                 array--;
2202         }
2203
2204         if (type & PERF_SAMPLE_ID) {
2205                 sample->id = *array;
2206                 array--;
2207         }
2208
2209         if (type & PERF_SAMPLE_TIME) {
2210                 sample->time = *array;
2211                 array--;
2212         }
2213
2214         if (type & PERF_SAMPLE_TID) {
2215                 u.val64 = *array;
2216                 if (swapped) {
2217                         /* undo swap of u64, then swap on individual u32s */
2218                         u.val64 = bswap_64(u.val64);
2219                         u.val32[0] = bswap_32(u.val32[0]);
2220                         u.val32[1] = bswap_32(u.val32[1]);
2221                 }
2222
2223                 sample->pid = u.val32[0];
2224                 sample->tid = u.val32[1];
2225                 array--;
2226         }
2227
2228         return 0;
2229 }
2230
2231 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2232                             u64 size)
2233 {
2234         return size > max_size || offset + size > endp;
2235 }
2236
2237 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2238         do {                                                            \
2239                 if (overflow(endp, (max_size), (offset), (size)))       \
2240                         return -EFAULT;                                 \
2241         } while (0)
2242
2243 #define OVERFLOW_CHECK_u64(offset) \
2244         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2245
2246 static int
2247 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2248 {
2249         /*
2250          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2251          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2252          * check the format does not go past the end of the event.
2253          */
2254         if (sample_size + sizeof(event->header) > event->header.size)
2255                 return -EFAULT;
2256
2257         return 0;
2258 }
2259
2260 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2261                                           const __u64 *array,
2262                                           u64 type __maybe_unused)
2263 {
2264         data->weight = *array;
2265 }
2266
2267 u64 evsel__bitfield_swap_branch_flags(u64 value)
2268 {
2269         u64 new_val = 0;
2270
2271         /*
2272          * branch_flags
2273          * union {
2274          *      u64 values;
2275          *      struct {
2276          *              mispred:1       //target mispredicted
2277          *              predicted:1     //target predicted
2278          *              in_tx:1         //in transaction
2279          *              abort:1         //transaction abort
2280          *              cycles:16       //cycle count to last branch
2281          *              type:4          //branch type
2282          *              spec:2          //branch speculation info
2283          *              new_type:4      //additional branch type
2284          *              priv:3          //privilege level
2285          *              reserved:31
2286          *      }
2287          * }
2288          *
2289          * Avoid bswap64() the entire branch_flag.value,
2290          * as it has variable bit-field sizes. Instead the
2291          * macro takes the bit-field position/size,
2292          * swaps it based on the host endianness.
2293          */
2294         if (host_is_bigendian()) {
2295                 new_val = bitfield_swap(value, 0, 1);
2296                 new_val |= bitfield_swap(value, 1, 1);
2297                 new_val |= bitfield_swap(value, 2, 1);
2298                 new_val |= bitfield_swap(value, 3, 1);
2299                 new_val |= bitfield_swap(value, 4, 16);
2300                 new_val |= bitfield_swap(value, 20, 4);
2301                 new_val |= bitfield_swap(value, 24, 2);
2302                 new_val |= bitfield_swap(value, 26, 4);
2303                 new_val |= bitfield_swap(value, 30, 3);
2304                 new_val |= bitfield_swap(value, 33, 31);
2305         } else {
2306                 new_val = bitfield_swap(value, 63, 1);
2307                 new_val |= bitfield_swap(value, 62, 1);
2308                 new_val |= bitfield_swap(value, 61, 1);
2309                 new_val |= bitfield_swap(value, 60, 1);
2310                 new_val |= bitfield_swap(value, 44, 16);
2311                 new_val |= bitfield_swap(value, 40, 4);
2312                 new_val |= bitfield_swap(value, 38, 2);
2313                 new_val |= bitfield_swap(value, 34, 4);
2314                 new_val |= bitfield_swap(value, 31, 3);
2315                 new_val |= bitfield_swap(value, 0, 31);
2316         }
2317
2318         return new_val;
2319 }
2320
2321 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2322                         struct perf_sample *data)
2323 {
2324         u64 type = evsel->core.attr.sample_type;
2325         bool swapped = evsel->needs_swap;
2326         const __u64 *array;
2327         u16 max_size = event->header.size;
2328         const void *endp = (void *)event + max_size;
2329         u64 sz;
2330
2331         /*
2332          * used for cross-endian analysis. See git commit 65014ab3
2333          * for why this goofiness is needed.
2334          */
2335         union u64_swap u;
2336
2337         memset(data, 0, sizeof(*data));
2338         data->cpu = data->pid = data->tid = -1;
2339         data->stream_id = data->id = data->time = -1ULL;
2340         data->period = evsel->core.attr.sample_period;
2341         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2342         data->misc    = event->header.misc;
2343         data->id = -1ULL;
2344         data->data_src = PERF_MEM_DATA_SRC_NONE;
2345         data->vcpu = -1;
2346
2347         if (event->header.type != PERF_RECORD_SAMPLE) {
2348                 if (!evsel->core.attr.sample_id_all)
2349                         return 0;
2350                 return perf_evsel__parse_id_sample(evsel, event, data);
2351         }
2352
2353         array = event->sample.array;
2354
2355         if (perf_event__check_size(event, evsel->sample_size))
2356                 return -EFAULT;
2357
2358         if (type & PERF_SAMPLE_IDENTIFIER) {
2359                 data->id = *array;
2360                 array++;
2361         }
2362
2363         if (type & PERF_SAMPLE_IP) {
2364                 data->ip = *array;
2365                 array++;
2366         }
2367
2368         if (type & PERF_SAMPLE_TID) {
2369                 u.val64 = *array;
2370                 if (swapped) {
2371                         /* undo swap of u64, then swap on individual u32s */
2372                         u.val64 = bswap_64(u.val64);
2373                         u.val32[0] = bswap_32(u.val32[0]);
2374                         u.val32[1] = bswap_32(u.val32[1]);
2375                 }
2376
2377                 data->pid = u.val32[0];
2378                 data->tid = u.val32[1];
2379                 array++;
2380         }
2381
2382         if (type & PERF_SAMPLE_TIME) {
2383                 data->time = *array;
2384                 array++;
2385         }
2386
2387         if (type & PERF_SAMPLE_ADDR) {
2388                 data->addr = *array;
2389                 array++;
2390         }
2391
2392         if (type & PERF_SAMPLE_ID) {
2393                 data->id = *array;
2394                 array++;
2395         }
2396
2397         if (type & PERF_SAMPLE_STREAM_ID) {
2398                 data->stream_id = *array;
2399                 array++;
2400         }
2401
2402         if (type & PERF_SAMPLE_CPU) {
2403
2404                 u.val64 = *array;
2405                 if (swapped) {
2406                         /* undo swap of u64, then swap on individual u32s */
2407                         u.val64 = bswap_64(u.val64);
2408                         u.val32[0] = bswap_32(u.val32[0]);
2409                 }
2410
2411                 data->cpu = u.val32[0];
2412                 array++;
2413         }
2414
2415         if (type & PERF_SAMPLE_PERIOD) {
2416                 data->period = *array;
2417                 array++;
2418         }
2419
2420         if (type & PERF_SAMPLE_READ) {
2421                 u64 read_format = evsel->core.attr.read_format;
2422
2423                 OVERFLOW_CHECK_u64(array);
2424                 if (read_format & PERF_FORMAT_GROUP)
2425                         data->read.group.nr = *array;
2426                 else
2427                         data->read.one.value = *array;
2428
2429                 array++;
2430
2431                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2432                         OVERFLOW_CHECK_u64(array);
2433                         data->read.time_enabled = *array;
2434                         array++;
2435                 }
2436
2437                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2438                         OVERFLOW_CHECK_u64(array);
2439                         data->read.time_running = *array;
2440                         array++;
2441                 }
2442
2443                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2444                 if (read_format & PERF_FORMAT_GROUP) {
2445                         const u64 max_group_nr = UINT64_MAX /
2446                                         sizeof(struct sample_read_value);
2447
2448                         if (data->read.group.nr > max_group_nr)
2449                                 return -EFAULT;
2450
2451                         sz = data->read.group.nr * sample_read_value_size(read_format);
2452                         OVERFLOW_CHECK(array, sz, max_size);
2453                         data->read.group.values =
2454                                         (struct sample_read_value *)array;
2455                         array = (void *)array + sz;
2456                 } else {
2457                         OVERFLOW_CHECK_u64(array);
2458                         data->read.one.id = *array;
2459                         array++;
2460
2461                         if (read_format & PERF_FORMAT_LOST) {
2462                                 OVERFLOW_CHECK_u64(array);
2463                                 data->read.one.lost = *array;
2464                                 array++;
2465                         }
2466                 }
2467         }
2468
2469         if (type & PERF_SAMPLE_CALLCHAIN) {
2470                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2471
2472                 OVERFLOW_CHECK_u64(array);
2473                 data->callchain = (struct ip_callchain *)array++;
2474                 if (data->callchain->nr > max_callchain_nr)
2475                         return -EFAULT;
2476                 sz = data->callchain->nr * sizeof(u64);
2477                 OVERFLOW_CHECK(array, sz, max_size);
2478                 array = (void *)array + sz;
2479         }
2480
2481         if (type & PERF_SAMPLE_RAW) {
2482                 OVERFLOW_CHECK_u64(array);
2483                 u.val64 = *array;
2484
2485                 /*
2486                  * Undo swap of u64, then swap on individual u32s,
2487                  * get the size of the raw area and undo all of the
2488                  * swap. The pevent interface handles endianness by
2489                  * itself.
2490                  */
2491                 if (swapped) {
2492                         u.val64 = bswap_64(u.val64);
2493                         u.val32[0] = bswap_32(u.val32[0]);
2494                         u.val32[1] = bswap_32(u.val32[1]);
2495                 }
2496                 data->raw_size = u.val32[0];
2497
2498                 /*
2499                  * The raw data is aligned on 64bits including the
2500                  * u32 size, so it's safe to use mem_bswap_64.
2501                  */
2502                 if (swapped)
2503                         mem_bswap_64((void *) array, data->raw_size);
2504
2505                 array = (void *)array + sizeof(u32);
2506
2507                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2508                 data->raw_data = (void *)array;
2509                 array = (void *)array + data->raw_size;
2510         }
2511
2512         if (type & PERF_SAMPLE_BRANCH_STACK) {
2513                 const u64 max_branch_nr = UINT64_MAX /
2514                                           sizeof(struct branch_entry);
2515                 struct branch_entry *e;
2516                 unsigned int i;
2517
2518                 OVERFLOW_CHECK_u64(array);
2519                 data->branch_stack = (struct branch_stack *)array++;
2520
2521                 if (data->branch_stack->nr > max_branch_nr)
2522                         return -EFAULT;
2523
2524                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2525                 if (evsel__has_branch_hw_idx(evsel)) {
2526                         sz += sizeof(u64);
2527                         e = &data->branch_stack->entries[0];
2528                 } else {
2529                         data->no_hw_idx = true;
2530                         /*
2531                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2532                          * only nr and entries[] will be output by kernel.
2533                          */
2534                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2535                 }
2536
2537                 if (swapped) {
2538                         /*
2539                          * struct branch_flag does not have endian
2540                          * specific bit field definition. And bswap
2541                          * will not resolve the issue, since these
2542                          * are bit fields.
2543                          *
2544                          * evsel__bitfield_swap_branch_flags() uses a
2545                          * bitfield_swap macro to swap the bit position
2546                          * based on the host endians.
2547                          */
2548                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2549                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2550                 }
2551
2552                 OVERFLOW_CHECK(array, sz, max_size);
2553                 array = (void *)array + sz;
2554         }
2555
2556         if (type & PERF_SAMPLE_REGS_USER) {
2557                 OVERFLOW_CHECK_u64(array);
2558                 data->user_regs.abi = *array;
2559                 array++;
2560
2561                 if (data->user_regs.abi) {
2562                         u64 mask = evsel->core.attr.sample_regs_user;
2563
2564                         sz = hweight64(mask) * sizeof(u64);
2565                         OVERFLOW_CHECK(array, sz, max_size);
2566                         data->user_regs.mask = mask;
2567                         data->user_regs.regs = (u64 *)array;
2568                         array = (void *)array + sz;
2569                 }
2570         }
2571
2572         if (type & PERF_SAMPLE_STACK_USER) {
2573                 OVERFLOW_CHECK_u64(array);
2574                 sz = *array++;
2575
2576                 data->user_stack.offset = ((char *)(array - 1)
2577                                           - (char *) event);
2578
2579                 if (!sz) {
2580                         data->user_stack.size = 0;
2581                 } else {
2582                         OVERFLOW_CHECK(array, sz, max_size);
2583                         data->user_stack.data = (char *)array;
2584                         array = (void *)array + sz;
2585                         OVERFLOW_CHECK_u64(array);
2586                         data->user_stack.size = *array++;
2587                         if (WARN_ONCE(data->user_stack.size > sz,
2588                                       "user stack dump failure\n"))
2589                                 return -EFAULT;
2590                 }
2591         }
2592
2593         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2594                 OVERFLOW_CHECK_u64(array);
2595                 arch_perf_parse_sample_weight(data, array, type);
2596                 array++;
2597         }
2598
2599         if (type & PERF_SAMPLE_DATA_SRC) {
2600                 OVERFLOW_CHECK_u64(array);
2601                 data->data_src = *array;
2602                 array++;
2603         }
2604
2605         if (type & PERF_SAMPLE_TRANSACTION) {
2606                 OVERFLOW_CHECK_u64(array);
2607                 data->transaction = *array;
2608                 array++;
2609         }
2610
2611         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2612         if (type & PERF_SAMPLE_REGS_INTR) {
2613                 OVERFLOW_CHECK_u64(array);
2614                 data->intr_regs.abi = *array;
2615                 array++;
2616
2617                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2618                         u64 mask = evsel->core.attr.sample_regs_intr;
2619
2620                         sz = hweight64(mask) * sizeof(u64);
2621                         OVERFLOW_CHECK(array, sz, max_size);
2622                         data->intr_regs.mask = mask;
2623                         data->intr_regs.regs = (u64 *)array;
2624                         array = (void *)array + sz;
2625                 }
2626         }
2627
2628         data->phys_addr = 0;
2629         if (type & PERF_SAMPLE_PHYS_ADDR) {
2630                 data->phys_addr = *array;
2631                 array++;
2632         }
2633
2634         data->cgroup = 0;
2635         if (type & PERF_SAMPLE_CGROUP) {
2636                 data->cgroup = *array;
2637                 array++;
2638         }
2639
2640         data->data_page_size = 0;
2641         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2642                 data->data_page_size = *array;
2643                 array++;
2644         }
2645
2646         data->code_page_size = 0;
2647         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2648                 data->code_page_size = *array;
2649                 array++;
2650         }
2651
2652         if (type & PERF_SAMPLE_AUX) {
2653                 OVERFLOW_CHECK_u64(array);
2654                 sz = *array++;
2655
2656                 OVERFLOW_CHECK(array, sz, max_size);
2657                 /* Undo swap of data */
2658                 if (swapped)
2659                         mem_bswap_64((char *)array, sz);
2660                 data->aux_sample.size = sz;
2661                 data->aux_sample.data = (char *)array;
2662                 array = (void *)array + sz;
2663         }
2664
2665         return 0;
2666 }
2667
2668 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2669                                   u64 *timestamp)
2670 {
2671         u64 type = evsel->core.attr.sample_type;
2672         const __u64 *array;
2673
2674         if (!(type & PERF_SAMPLE_TIME))
2675                 return -1;
2676
2677         if (event->header.type != PERF_RECORD_SAMPLE) {
2678                 struct perf_sample data = {
2679                         .time = -1ULL,
2680                 };
2681
2682                 if (!evsel->core.attr.sample_id_all)
2683                         return -1;
2684                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2685                         return -1;
2686
2687                 *timestamp = data.time;
2688                 return 0;
2689         }
2690
2691         array = event->sample.array;
2692
2693         if (perf_event__check_size(event, evsel->sample_size))
2694                 return -EFAULT;
2695
2696         if (type & PERF_SAMPLE_IDENTIFIER)
2697                 array++;
2698
2699         if (type & PERF_SAMPLE_IP)
2700                 array++;
2701
2702         if (type & PERF_SAMPLE_TID)
2703                 array++;
2704
2705         if (type & PERF_SAMPLE_TIME)
2706                 *timestamp = *array;
2707
2708         return 0;
2709 }
2710
2711 u16 evsel__id_hdr_size(struct evsel *evsel)
2712 {
2713         u64 sample_type = evsel->core.attr.sample_type;
2714         u16 size = 0;
2715
2716         if (sample_type & PERF_SAMPLE_TID)
2717                 size += sizeof(u64);
2718
2719         if (sample_type & PERF_SAMPLE_TIME)
2720                 size += sizeof(u64);
2721
2722         if (sample_type & PERF_SAMPLE_ID)
2723                 size += sizeof(u64);
2724
2725         if (sample_type & PERF_SAMPLE_STREAM_ID)
2726                 size += sizeof(u64);
2727
2728         if (sample_type & PERF_SAMPLE_CPU)
2729                 size += sizeof(u64);
2730
2731         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2732                 size += sizeof(u64);
2733
2734         return size;
2735 }
2736
2737 #ifdef HAVE_LIBTRACEEVENT
2738 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2739 {
2740         return tep_find_field(evsel->tp_format, name);
2741 }
2742
2743 struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name)
2744 {
2745         return tep_find_common_field(evsel->tp_format, name);
2746 }
2747
2748 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2749 {
2750         struct tep_format_field *field = evsel__field(evsel, name);
2751         int offset;
2752
2753         if (!field)
2754                 return NULL;
2755
2756         offset = field->offset;
2757
2758         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2759                 offset = *(int *)(sample->raw_data + field->offset);
2760                 offset &= 0xffff;
2761                 if (tep_field_is_relative(field->flags))
2762                         offset += field->offset + field->size;
2763         }
2764
2765         return sample->raw_data + offset;
2766 }
2767
2768 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2769                          bool needs_swap)
2770 {
2771         u64 value;
2772         void *ptr = sample->raw_data + field->offset;
2773
2774         switch (field->size) {
2775         case 1:
2776                 return *(u8 *)ptr;
2777         case 2:
2778                 value = *(u16 *)ptr;
2779                 break;
2780         case 4:
2781                 value = *(u32 *)ptr;
2782                 break;
2783         case 8:
2784                 memcpy(&value, ptr, sizeof(u64));
2785                 break;
2786         default:
2787                 return 0;
2788         }
2789
2790         if (!needs_swap)
2791                 return value;
2792
2793         switch (field->size) {
2794         case 2:
2795                 return bswap_16(value);
2796         case 4:
2797                 return bswap_32(value);
2798         case 8:
2799                 return bswap_64(value);
2800         default:
2801                 return 0;
2802         }
2803
2804         return 0;
2805 }
2806
2807 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2808 {
2809         struct tep_format_field *field = evsel__field(evsel, name);
2810
2811         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2812 }
2813
2814 u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name)
2815 {
2816         struct tep_format_field *field = evsel__common_field(evsel, name);
2817
2818         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2819 }
2820
2821 #endif
2822
2823 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2824 {
2825         int paranoid;
2826
2827         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2828             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2829             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2830                 /*
2831                  * If it's cycles then fall back to hrtimer based
2832                  * cpu-clock-tick sw counter, which is always available even if
2833                  * no PMU support.
2834                  *
2835                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2836                  * b0a873e).
2837                  */
2838                 scnprintf(msg, msgsize, "%s",
2839 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2840
2841                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2842                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2843
2844                 zfree(&evsel->name);
2845                 return true;
2846         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2847                    (paranoid = perf_event_paranoid()) > 1) {
2848                 const char *name = evsel__name(evsel);
2849                 char *new_name;
2850                 const char *sep = ":";
2851
2852                 /* If event has exclude user then don't exclude kernel. */
2853                 if (evsel->core.attr.exclude_user)
2854                         return false;
2855
2856                 /* Is there already the separator in the name. */
2857                 if (strchr(name, '/') ||
2858                     (strchr(name, ':') && !evsel->is_libpfm_event))
2859                         sep = "";
2860
2861                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2862                         return false;
2863
2864                 free(evsel->name);
2865                 evsel->name = new_name;
2866                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2867                           "to fall back to excluding kernel and hypervisor "
2868                           " samples", paranoid);
2869                 evsel->core.attr.exclude_kernel = 1;
2870                 evsel->core.attr.exclude_hv     = 1;
2871
2872                 return true;
2873         }
2874
2875         return false;
2876 }
2877
2878 static bool find_process(const char *name)
2879 {
2880         size_t len = strlen(name);
2881         DIR *dir;
2882         struct dirent *d;
2883         int ret = -1;
2884
2885         dir = opendir(procfs__mountpoint());
2886         if (!dir)
2887                 return false;
2888
2889         /* Walk through the directory. */
2890         while (ret && (d = readdir(dir)) != NULL) {
2891                 char path[PATH_MAX];
2892                 char *data;
2893                 size_t size;
2894
2895                 if ((d->d_type != DT_DIR) ||
2896                      !strcmp(".", d->d_name) ||
2897                      !strcmp("..", d->d_name))
2898                         continue;
2899
2900                 scnprintf(path, sizeof(path), "%s/%s/comm",
2901                           procfs__mountpoint(), d->d_name);
2902
2903                 if (filename__read_str(path, &data, &size))
2904                         continue;
2905
2906                 ret = strncmp(name, data, len);
2907                 free(data);
2908         }
2909
2910         closedir(dir);
2911         return ret ? false : true;
2912 }
2913
2914 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
2915                                      char *msg __maybe_unused,
2916                                      size_t size __maybe_unused)
2917 {
2918         return 0;
2919 }
2920
2921 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2922                          int err, char *msg, size_t size)
2923 {
2924         char sbuf[STRERR_BUFSIZE];
2925         int printed = 0, enforced = 0;
2926         int ret;
2927
2928         switch (err) {
2929         case EPERM:
2930         case EACCES:
2931                 printed += scnprintf(msg + printed, size - printed,
2932                         "Access to performance monitoring and observability operations is limited.\n");
2933
2934                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2935                         if (enforced) {
2936                                 printed += scnprintf(msg + printed, size - printed,
2937                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2938                                         "monitoring and observability operations. Inspect system audit records for\n"
2939                                         "more perf_event access control information and adjusting the policy.\n");
2940                         }
2941                 }
2942
2943                 if (err == EPERM)
2944                         printed += scnprintf(msg, size,
2945                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2946
2947                 return scnprintf(msg + printed, size - printed,
2948                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2949                  "access to performance monitoring and observability operations for processes\n"
2950                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2951                  "More information can be found at 'Perf events and tool security' document:\n"
2952                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2953                  "perf_event_paranoid setting is %d:\n"
2954                  "  -1: Allow use of (almost) all events by all users\n"
2955                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2956                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2957                  ">= 1: Disallow CPU event access\n"
2958                  ">= 2: Disallow kernel profiling\n"
2959                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2960                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2961                  perf_event_paranoid());
2962         case ENOENT:
2963                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2964         case EMFILE:
2965                 return scnprintf(msg, size, "%s",
2966                          "Too many events are opened.\n"
2967                          "Probably the maximum number of open file descriptors has been reached.\n"
2968                          "Hint: Try again after reducing the number of events.\n"
2969                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2970         case ENOMEM:
2971                 if (evsel__has_callchain(evsel) &&
2972                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2973                         return scnprintf(msg, size,
2974                                          "Not enough memory to setup event with callchain.\n"
2975                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2976                                          "Hint: Current value: %d", sysctl__max_stack());
2977                 break;
2978         case ENODEV:
2979                 if (target->cpu_list)
2980                         return scnprintf(msg, size, "%s",
2981          "No such device - did you specify an out-of-range profile CPU?");
2982                 break;
2983         case EOPNOTSUPP:
2984                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
2985                         return scnprintf(msg, size,
2986         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
2987                                          evsel__name(evsel));
2988                 if (evsel->core.attr.aux_output)
2989                         return scnprintf(msg, size,
2990         "%s: PMU Hardware doesn't support 'aux_output' feature",
2991                                          evsel__name(evsel));
2992                 if (evsel->core.attr.sample_period != 0)
2993                         return scnprintf(msg, size,
2994         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2995                                          evsel__name(evsel));
2996                 if (evsel->core.attr.precise_ip)
2997                         return scnprintf(msg, size, "%s",
2998         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2999 #if defined(__i386__) || defined(__x86_64__)
3000                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3001                         return scnprintf(msg, size, "%s",
3002         "No hardware sampling interrupt available.\n");
3003 #endif
3004                 break;
3005         case EBUSY:
3006                 if (find_process("oprofiled"))
3007                         return scnprintf(msg, size,
3008         "The PMU counters are busy/taken by another profiler.\n"
3009         "We found oprofile daemon running, please stop it and try again.");
3010                 break;
3011         case EINVAL:
3012                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3013                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3014                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3015                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3016                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3017                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3018                 if (perf_missing_features.clockid)
3019                         return scnprintf(msg, size, "clockid feature not supported.");
3020                 if (perf_missing_features.clockid_wrong)
3021                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3022                 if (perf_missing_features.aux_output)
3023                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3024                 if (!target__has_cpu(target))
3025                         return scnprintf(msg, size,
3026         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3027                                         evsel__name(evsel));
3028
3029                 break;
3030         case ENODATA:
3031                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3032                                  "Please add an auxiliary event in front of the load latency event.");
3033         default:
3034                 break;
3035         }
3036
3037         ret = arch_evsel__open_strerror(evsel, msg, size);
3038         if (ret)
3039                 return ret;
3040
3041         return scnprintf(msg, size,
3042         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3043         "/bin/dmesg | grep -i perf may provide additional information.\n",
3044                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3045 }
3046
3047 struct perf_env *evsel__env(struct evsel *evsel)
3048 {
3049         if (evsel && evsel->evlist && evsel->evlist->env)
3050                 return evsel->evlist->env;
3051         return &perf_env;
3052 }
3053
3054 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3055 {
3056         int cpu_map_idx, thread;
3057
3058         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3059                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3060                      thread++) {
3061                         int fd = FD(evsel, cpu_map_idx, thread);
3062
3063                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3064                                                    cpu_map_idx, thread, fd) < 0)
3065                                 return -1;
3066                 }
3067         }
3068
3069         return 0;
3070 }
3071
3072 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3073 {
3074         struct perf_cpu_map *cpus = evsel->core.cpus;
3075         struct perf_thread_map *threads = evsel->core.threads;
3076
3077         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3078                 return -ENOMEM;
3079
3080         return store_evsel_ids(evsel, evlist);
3081 }
3082
3083 void evsel__zero_per_pkg(struct evsel *evsel)
3084 {
3085         struct hashmap_entry *cur;
3086         size_t bkt;
3087
3088         if (evsel->per_pkg_mask) {
3089                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3090                         zfree(&cur->pkey);
3091
3092                 hashmap__clear(evsel->per_pkg_mask);
3093         }
3094 }
3095
3096 /**
3097  * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3098  *                    will be false on hybrid systems for hardware and legacy
3099  *                    cache events.
3100  */
3101 bool evsel__is_hybrid(const struct evsel *evsel)
3102 {
3103         if (perf_pmus__num_core_pmus() == 1)
3104                 return false;
3105
3106         return evsel->core.is_pmu_core;
3107 }
3108
3109 struct evsel *evsel__leader(const struct evsel *evsel)
3110 {
3111         return container_of(evsel->core.leader, struct evsel, core);
3112 }
3113
3114 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3115 {
3116         return evsel->core.leader == &leader->core;
3117 }
3118
3119 bool evsel__is_leader(struct evsel *evsel)
3120 {
3121         return evsel__has_leader(evsel, evsel);
3122 }
3123
3124 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3125 {
3126         evsel->core.leader = &leader->core;
3127 }
3128
3129 int evsel__source_count(const struct evsel *evsel)
3130 {
3131         struct evsel *pos;
3132         int count = 0;
3133
3134         evlist__for_each_entry(evsel->evlist, pos) {
3135                 if (pos->metric_leader == evsel)
3136                         count++;
3137         }
3138         return count;
3139 }
3140
3141 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3142 {
3143         return false;
3144 }
3145
3146 /*
3147  * Remove an event from a given group (leader).
3148  * Some events, e.g., perf metrics Topdown events,
3149  * must always be grouped. Ignore the events.
3150  */
3151 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3152 {
3153         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3154                 evsel__set_leader(evsel, evsel);
3155                 evsel->core.nr_members = 0;
3156                 leader->core.nr_members--;
3157         }
3158 }