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 "pmu-hybrid.h"
  50 #include "off_cpu.h"
  51 #include "../perf-sys.h"
  52 #include "util/parse-branch-options.h"
  53 #include "util/bpf-filter.h"
  54 #include <internal/xyarray.h>
  55 #include <internal/lib.h>
  56 #include <internal/threadmap.h>
  57
  58 #include <linux/ctype.h>
  59
  60 #ifdef HAVE_LIBTRACEEVENT
  61 #include <traceevent/event-parse.h>
  62 #endif
  63
  64 struct perf_missing_features perf_missing_features;
  65
  66 static clockid_t clockid;
  67
  68 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  69         NULL,
  70         "duration_time",
  71         "user_time",
  72         "system_time",
  73 };
  74
  75 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  76 {
  77         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  78                 return perf_tool_event__tool_names[ev];
  79
  80         return NULL;
  81 }
  82
  83 enum perf_tool_event perf_tool_event__from_str(const char *str)
  84 {
  85         int i;
  86
  87         perf_tool_event__for_each_event(i) {
  88                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  89                         return i;
  90         }
  91         return PERF_TOOL_NONE;
  92 }
  93
  94
  95 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  96 {
  97         return 0;
  98 }
  99
 100 void __weak test_attr__ready(void) { }
 101
 102 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 103 {
 104 }
 105
 106 static struct {
 107         size_t  size;
 108         int     (*init)(struct evsel *evsel);
 109         void    (*fini)(struct evsel *evsel);
 110 } perf_evsel__object = {
 111         .size = sizeof(struct evsel),
 112         .init = evsel__no_extra_init,
 113         .fini = evsel__no_extra_fini,
 114 };
 115
 116 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 117                          void (*fini)(struct evsel *evsel))
 118 {
 119
 120         if (object_size == 0)
 121                 goto set_methods;
 122
 123         if (perf_evsel__object.size > object_size)
 124                 return -EINVAL;
 125
 126         perf_evsel__object.size = object_size;
 127
 128 set_methods:
 129         if (init != NULL)
 130                 perf_evsel__object.init = init;
 131
 132         if (fini != NULL)
 133                 perf_evsel__object.fini = fini;
 134
 135         return 0;
 136 }
 137
 138 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 139
 140 int __evsel__sample_size(u64 sample_type)
 141 {
 142         u64 mask = sample_type & PERF_SAMPLE_MASK;
 143         int size = 0;
 144         int i;
 145
 146         for (i = 0; i < 64; i++) {
 147                 if (mask & (1ULL << i))
 148                         size++;
 149         }
 150
 151         size *= sizeof(u64);
 152
 153         return size;
 154 }
 155
 156 /**
 157  * __perf_evsel__calc_id_pos - calculate id_pos.
 158  * @sample_type: sample type
 159  *
 160  * This function returns the position of the event id (PERF_SAMPLE_ID or
 161  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 162  * perf_record_sample.
 163  */
 164 static int __perf_evsel__calc_id_pos(u64 sample_type)
 165 {
 166         int idx = 0;
 167
 168         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 169                 return 0;
 170
 171         if (!(sample_type & PERF_SAMPLE_ID))
 172                 return -1;
 173
 174         if (sample_type & PERF_SAMPLE_IP)
 175                 idx += 1;
 176
 177         if (sample_type & PERF_SAMPLE_TID)
 178                 idx += 1;
 179
 180         if (sample_type & PERF_SAMPLE_TIME)
 181                 idx += 1;
 182
 183         if (sample_type & PERF_SAMPLE_ADDR)
 184                 idx += 1;
 185
 186         return idx;
 187 }
 188
 189 /**
 190  * __perf_evsel__calc_is_pos - calculate is_pos.
 191  * @sample_type: sample type
 192  *
 193  * This function returns the position (counting backwards) of the event id
 194  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 195  * sample_id_all is used there is an id sample appended to non-sample events.
 196  */
 197 static int __perf_evsel__calc_is_pos(u64 sample_type)
 198 {
 199         int idx = 1;
 200
 201         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 202                 return 1;
 203
 204         if (!(sample_type & PERF_SAMPLE_ID))
 205                 return -1;
 206
 207         if (sample_type & PERF_SAMPLE_CPU)
 208                 idx += 1;
 209
 210         if (sample_type & PERF_SAMPLE_STREAM_ID)
 211                 idx += 1;
 212
 213         return idx;
 214 }
 215
 216 void evsel__calc_id_pos(struct evsel *evsel)
 217 {
 218         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 219         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 220 }
 221
 222 void __evsel__set_sample_bit(struct evsel *evsel,
 223                                   enum perf_event_sample_format bit)
 224 {
 225         if (!(evsel->core.attr.sample_type & bit)) {
 226                 evsel->core.attr.sample_type |= bit;
 227                 evsel->sample_size += sizeof(u64);
 228                 evsel__calc_id_pos(evsel);
 229         }
 230 }
 231
 232 void __evsel__reset_sample_bit(struct evsel *evsel,
 233                                     enum perf_event_sample_format bit)
 234 {
 235         if (evsel->core.attr.sample_type & bit) {
 236                 evsel->core.attr.sample_type &= ~bit;
 237                 evsel->sample_size -= sizeof(u64);
 238                 evsel__calc_id_pos(evsel);
 239         }
 240 }
 241
 242 void evsel__set_sample_id(struct evsel *evsel,
 243                                bool can_sample_identifier)
 244 {
 245         if (can_sample_identifier) {
 246                 evsel__reset_sample_bit(evsel, ID);
 247                 evsel__set_sample_bit(evsel, IDENTIFIER);
 248         } else {
 249                 evsel__set_sample_bit(evsel, ID);
 250         }
 251         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 252 }
 253
 254 /**
 255  * evsel__is_function_event - Return whether given evsel is a function
 256  * trace event
 257  *
 258  * @evsel - evsel selector to be tested
 259  *
 260  * Return %true if event is function trace event
 261  */
 262 bool evsel__is_function_event(struct evsel *evsel)
 263 {
 264 #define FUNCTION_EVENT "ftrace:function"
 265
 266         return evsel->name &&
 267                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 268
 269 #undef FUNCTION_EVENT
 270 }
 271
 272 void evsel__init(struct evsel *evsel,
 273                  struct perf_event_attr *attr, int idx)
 274 {
 275         perf_evsel__init(&evsel->core, attr, idx);
 276         evsel->tracking    = !idx;
 277         evsel->unit        = strdup("");
 278         evsel->scale       = 1.0;
 279         evsel->max_events  = ULONG_MAX;
 280         evsel->evlist      = NULL;
 281         evsel->bpf_obj     = NULL;
 282         evsel->bpf_fd      = -1;
 283         INIT_LIST_HEAD(&evsel->config_terms);
 284         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 285         INIT_LIST_HEAD(&evsel->bpf_filters);
 286         perf_evsel__object.init(evsel);
 287         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 288         evsel__calc_id_pos(evsel);
 289         evsel->cmdline_group_boundary = false;
 290         evsel->metric_events = NULL;
 291         evsel->per_pkg_mask  = NULL;
 292         evsel->collect_stat  = false;
 293         evsel->pmu_name      = NULL;
 294         evsel->skippable     = false;
 295 }
 296
 297 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 298 {
 299         struct evsel *evsel = zalloc(perf_evsel__object.size);
 300
 301         if (!evsel)
 302                 return NULL;
 303         evsel__init(evsel, attr, idx);
 304
 305         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 306                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 307                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 308                 evsel->core.attr.sample_period = 1;
 309         }
 310
 311         if (evsel__is_clock(evsel)) {
 312                 free((char *)evsel->unit);
 313                 evsel->unit = strdup("msec");
 314                 evsel->scale = 1e-6;
 315         }
 316
 317         return evsel;
 318 }
 319
 320 static bool perf_event_can_profile_kernel(void)
 321 {
 322         return perf_event_paranoid_check(1);
 323 }
 324
 325 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 326 {
 327         struct perf_event_attr attr = {
 328                 .type   = type,
 329                 .config = config,
 330                 .exclude_kernel = !perf_event_can_profile_kernel(),
 331         };
 332         struct evsel *evsel;
 333
 334         event_attr_init(&attr);
 335
 336         /*
 337          * Now let the usual logic to set up the perf_event_attr defaults
 338          * to kick in when we return and before perf_evsel__open() is called.
 339          */
 340         evsel = evsel__new(&attr);
 341         if (evsel == NULL)
 342                 goto out;
 343
 344         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 345
 346         evsel->precise_max = true;
 347
 348         /* use asprintf() because free(evsel) assumes name is allocated */
 349         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 350                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 351                      attr.exclude_kernel ? "u" : "",
 352                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 353                 goto error_free;
 354 out:
 355         return evsel;
 356 error_free:
 357         evsel__delete(evsel);
 358         evsel = NULL;
 359         goto out;
 360 }
 361
 362 int copy_config_terms(struct list_head *dst, struct list_head *src)
 363 {
 364         struct evsel_config_term *pos, *tmp;
 365
 366         list_for_each_entry(pos, src, list) {
 367                 tmp = malloc(sizeof(*tmp));
 368                 if (tmp == NULL)
 369                         return -ENOMEM;
 370
 371                 *tmp = *pos;
 372                 if (tmp->free_str) {
 373                         tmp->val.str = strdup(pos->val.str);
 374                         if (tmp->val.str == NULL) {
 375                                 free(tmp);
 376                                 return -ENOMEM;
 377                         }
 378                 }
 379                 list_add_tail(&tmp->list, dst);
 380         }
 381         return 0;
 382 }
 383
 384 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 385 {
 386         return copy_config_terms(&dst->config_terms, &src->config_terms);
 387 }
 388
 389 /**
 390  * evsel__clone - create a new evsel copied from @orig
 391  * @orig: original evsel
 392  *
 393  * The assumption is that @orig is not configured nor opened yet.
 394  * So we only care about the attributes that can be set while it's parsed.
 395  */
 396 struct evsel *evsel__clone(struct evsel *orig)
 397 {
 398         struct evsel *evsel;
 399
 400         BUG_ON(orig->core.fd);
 401         BUG_ON(orig->counts);
 402         BUG_ON(orig->priv);
 403         BUG_ON(orig->per_pkg_mask);
 404
 405         /* cannot handle BPF objects for now */
 406         if (orig->bpf_obj)
 407                 return NULL;
 408
 409         evsel = evsel__new(&orig->core.attr);
 410         if (evsel == NULL)
 411                 return NULL;
 412
 413         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 414         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 415         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 416         evsel->core.nr_members = orig->core.nr_members;
 417         evsel->core.system_wide = orig->core.system_wide;
 418         evsel->core.requires_cpu = orig->core.requires_cpu;
 419
 420         if (orig->name) {
 421                 evsel->name = strdup(orig->name);
 422                 if (evsel->name == NULL)
 423                         goto out_err;
 424         }
 425         if (orig->group_name) {
 426                 evsel->group_name = strdup(orig->group_name);
 427                 if (evsel->group_name == NULL)
 428                         goto out_err;
 429         }
 430         if (orig->pmu_name) {
 431                 evsel->pmu_name = strdup(orig->pmu_name);
 432                 if (evsel->pmu_name == NULL)
 433                         goto out_err;
 434         }
 435         if (orig->filter) {
 436                 evsel->filter = strdup(orig->filter);
 437                 if (evsel->filter == NULL)
 438                         goto out_err;
 439         }
 440         if (orig->metric_id) {
 441                 evsel->metric_id = strdup(orig->metric_id);
 442                 if (evsel->metric_id == NULL)
 443                         goto out_err;
 444         }
 445         evsel->cgrp = cgroup__get(orig->cgrp);
 446 #ifdef HAVE_LIBTRACEEVENT
 447         evsel->tp_format = orig->tp_format;
 448 #endif
 449         evsel->handler = orig->handler;
 450         evsel->core.leader = orig->core.leader;
 451
 452         evsel->max_events = orig->max_events;
 453         evsel->tool_event = orig->tool_event;
 454         free((char *)evsel->unit);
 455         evsel->unit = strdup(orig->unit);
 456         if (evsel->unit == NULL)
 457                 goto out_err;
 458
 459         evsel->scale = orig->scale;
 460         evsel->snapshot = orig->snapshot;
 461         evsel->per_pkg = orig->per_pkg;
 462         evsel->percore = orig->percore;
 463         evsel->precise_max = orig->precise_max;
 464         evsel->is_libpfm_event = orig->is_libpfm_event;
 465
 466         evsel->exclude_GH = orig->exclude_GH;
 467         evsel->sample_read = orig->sample_read;
 468         evsel->auto_merge_stats = orig->auto_merge_stats;
 469         evsel->collect_stat = orig->collect_stat;
 470         evsel->weak_group = orig->weak_group;
 471         evsel->use_config_name = orig->use_config_name;
 472         evsel->pmu = orig->pmu;
 473
 474         if (evsel__copy_config_terms(evsel, orig) < 0)
 475                 goto out_err;
 476
 477         return evsel;
 478
 479 out_err:
 480         evsel__delete(evsel);
 481         return NULL;
 482 }
 483
 484 /*
 485  * Returns pointer with encoded error via <linux/err.h> interface.
 486  */
 487 #ifdef HAVE_LIBTRACEEVENT
 488 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 489 {
 490         struct evsel *evsel = zalloc(perf_evsel__object.size);
 491         int err = -ENOMEM;
 492
 493         if (evsel == NULL) {
 494                 goto out_err;
 495         } else {
 496                 struct perf_event_attr attr = {
 497                         .type          = PERF_TYPE_TRACEPOINT,
 498                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 499                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 500                 };
 501
 502                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 503                         goto out_free;
 504
 505                 evsel->tp_format = trace_event__tp_format(sys, name);
 506                 if (IS_ERR(evsel->tp_format)) {
 507                         err = PTR_ERR(evsel->tp_format);
 508                         goto out_free;
 509                 }
 510
 511                 event_attr_init(&attr);
 512                 attr.config = evsel->tp_format->id;
 513                 attr.sample_period = 1;
 514                 evsel__init(evsel, &attr, idx);
 515         }
 516
 517         return evsel;
 518
 519 out_free:
 520         zfree(&evsel->name);
 521         free(evsel);
 522 out_err:
 523         return ERR_PTR(err);
 524 }
 525 #endif
 526
 527 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 528         "cycles",
 529         "instructions",
 530         "cache-references",
 531         "cache-misses",
 532         "branches",
 533         "branch-misses",
 534         "bus-cycles",
 535         "stalled-cycles-frontend",
 536         "stalled-cycles-backend",
 537         "ref-cycles",
 538 };
 539
 540 char *evsel__bpf_counter_events;
 541
 542 bool evsel__match_bpf_counter_events(const char *name)
 543 {
 544         int name_len;
 545         bool match;
 546         char *ptr;
 547
 548         if (!evsel__bpf_counter_events)
 549                 return false;
 550
 551         ptr = strstr(evsel__bpf_counter_events, name);
 552         name_len = strlen(name);
 553
 554         /* check name matches a full token in evsel__bpf_counter_events */
 555         match = (ptr != NULL) &&
 556                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 557                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 558
 559         return match;
 560 }
 561
 562 static const char *__evsel__hw_name(u64 config)
 563 {
 564         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 565                 return evsel__hw_names[config];
 566
 567         return "unknown-hardware";
 568 }
 569
 570 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 571 {
 572         int colon = 0, r = 0;
 573         struct perf_event_attr *attr = &evsel->core.attr;
 574         bool exclude_guest_default = false;
 575
 576 #define MOD_PRINT(context, mod) do {                                    \
 577                 if (!attr->exclude_##context) {                         \
 578                         if (!colon) colon = ++r;                        \
 579                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 580                 } } while(0)
 581
 582         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 583                 MOD_PRINT(kernel, 'k');
 584                 MOD_PRINT(user, 'u');
 585                 MOD_PRINT(hv, 'h');
 586                 exclude_guest_default = true;
 587         }
 588
 589         if (attr->precise_ip) {
 590                 if (!colon)
 591                         colon = ++r;
 592                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 593                 exclude_guest_default = true;
 594         }
 595
 596         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 597                 MOD_PRINT(host, 'H');
 598                 MOD_PRINT(guest, 'G');
 599         }
 600 #undef MOD_PRINT
 601         if (colon)
 602                 bf[colon - 1] = ':';
 603         return r;
 604 }
 605
 606 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 607 {
 608         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 609 }
 610
 611 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 612 {
 613         int r = arch_evsel__hw_name(evsel, bf, size);
 614         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 615 }
 616
 617 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 618         "cpu-clock",
 619         "task-clock",
 620         "page-faults",
 621         "context-switches",
 622         "cpu-migrations",
 623         "minor-faults",
 624         "major-faults",
 625         "alignment-faults",
 626         "emulation-faults",
 627         "dummy",
 628 };
 629
 630 static const char *__evsel__sw_name(u64 config)
 631 {
 632         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 633                 return evsel__sw_names[config];
 634         return "unknown-software";
 635 }
 636
 637 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 638 {
 639         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 640         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 641 }
 642
 643 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 644 {
 645         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 646 }
 647
 648 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 649 {
 650         int r;
 651
 652         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 653
 654         if (type & HW_BREAKPOINT_R)
 655                 r += scnprintf(bf + r, size - r, "r");
 656
 657         if (type & HW_BREAKPOINT_W)
 658                 r += scnprintf(bf + r, size - r, "w");
 659
 660         if (type & HW_BREAKPOINT_X)
 661                 r += scnprintf(bf + r, size - r, "x");
 662
 663         return r;
 664 }
 665
 666 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 667 {
 668         struct perf_event_attr *attr = &evsel->core.attr;
 669         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 670         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 671 }
 672
 673 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 674  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 675  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 676  { "LLC",       "L2",                                                   },
 677  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 678  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 679  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 680  { "node",                                                              },
 681 };
 682
 683 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 684  { "load",      "loads",        "read",                                 },
 685  { "store",     "stores",       "write",                                },
 686  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 687 };
 688
 689 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 690  { "refs",      "Reference",    "ops",          "access",               },
 691  { "misses",    "miss",                                                 },
 692 };
 693
 694 #define C(x)            PERF_COUNT_HW_CACHE_##x
 695 #define CACHE_READ      (1 << C(OP_READ))
 696 #define CACHE_WRITE     (1 << C(OP_WRITE))
 697 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 698 #define COP(x)          (1 << x)
 699
 700 /*
 701  * cache operation stat
 702  * L1I : Read and prefetch only
 703  * ITLB and BPU : Read-only
 704  */
 705 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 706  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 707  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 708  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 709  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 710  [C(ITLB)]      = (CACHE_READ),
 711  [C(BPU)]       = (CACHE_READ),
 712  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 713 };
 714
 715 bool evsel__is_cache_op_valid(u8 type, u8 op)
 716 {
 717         if (evsel__hw_cache_stat[type] & COP(op))
 718                 return true;    /* valid */
 719         else
 720                 return false;   /* invalid */
 721 }
 722
 723 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 724 {
 725         if (result) {
 726                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 727                                  evsel__hw_cache_op[op][0],
 728                                  evsel__hw_cache_result[result][0]);
 729         }
 730
 731         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 732                          evsel__hw_cache_op[op][1]);
 733 }
 734
 735 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 736 {
 737         u8 op, result, type = (config >>  0) & 0xff;
 738         const char *err = "unknown-ext-hardware-cache-type";
 739
 740         if (type >= PERF_COUNT_HW_CACHE_MAX)
 741                 goto out_err;
 742
 743         op = (config >>  8) & 0xff;
 744         err = "unknown-ext-hardware-cache-op";
 745         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 746                 goto out_err;
 747
 748         result = (config >> 16) & 0xff;
 749         err = "unknown-ext-hardware-cache-result";
 750         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 751                 goto out_err;
 752
 753         err = "invalid-cache";
 754         if (!evsel__is_cache_op_valid(type, op))
 755                 goto out_err;
 756
 757         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 758 out_err:
 759         return scnprintf(bf, size, "%s", err);
 760 }
 761
 762 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 763 {
 764         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 765         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 766 }
 767
 768 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 769 {
 770         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 771         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 772 }
 773
 774 const char *evsel__name(struct evsel *evsel)
 775 {
 776         char bf[128];
 777
 778         if (!evsel)
 779                 goto out_unknown;
 780
 781         if (evsel->name)
 782                 return evsel->name;
 783
 784         switch (evsel->core.attr.type) {
 785         case PERF_TYPE_RAW:
 786                 evsel__raw_name(evsel, bf, sizeof(bf));
 787                 break;
 788
 789         case PERF_TYPE_HARDWARE:
 790                 evsel__hw_name(evsel, bf, sizeof(bf));
 791                 break;
 792
 793         case PERF_TYPE_HW_CACHE:
 794                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 795                 break;
 796
 797         case PERF_TYPE_SOFTWARE:
 798                 if (evsel__is_tool(evsel))
 799                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 800                 else
 801                         evsel__sw_name(evsel, bf, sizeof(bf));
 802                 break;
 803
 804         case PERF_TYPE_TRACEPOINT:
 805                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 806                 break;
 807
 808         case PERF_TYPE_BREAKPOINT:
 809                 evsel__bp_name(evsel, bf, sizeof(bf));
 810                 break;
 811
 812         default:
 813                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 814                           evsel->core.attr.type);
 815                 break;
 816         }
 817
 818         evsel->name = strdup(bf);
 819
 820         if (evsel->name)
 821                 return evsel->name;
 822 out_unknown:
 823         return "unknown";
 824 }
 825
 826 bool evsel__name_is(struct evsel *evsel, const char *name)
 827 {
 828         return !strcmp(evsel__name(evsel), name);
 829 }
 830
 831 const char *evsel__group_pmu_name(const struct evsel *evsel)
 832 {
 833         struct evsel *leader = evsel__leader(evsel);
 834         struct evsel *pos;
 835
 836         /*
 837          * Software events may be in a group with other uncore PMU events. Use
 838          * the pmu_name of the first non-software event to avoid breaking the
 839          * software event out of the group.
 840          *
 841          * Aux event leaders, like intel_pt, expect a group with events from
 842          * other PMUs, so substitute the AUX event's PMU in this case.
 843          */
 844         if (evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) {
 845                 /* Starting with the leader, find the first event with a named PMU. */
 846                 for_each_group_evsel(pos, leader) {
 847                         if (pos->pmu_name)
 848                                 return pos->pmu_name;
 849                 }
 850         }
 851
 852         return evsel->pmu_name ?: "cpu";
 853 }
 854
 855 const char *evsel__metric_id(const struct evsel *evsel)
 856 {
 857         if (evsel->metric_id)
 858                 return evsel->metric_id;
 859
 860         if (evsel__is_tool(evsel))
 861                 return perf_tool_event__to_str(evsel->tool_event);
 862
 863         return "unknown";
 864 }
 865
 866 const char *evsel__group_name(struct evsel *evsel)
 867 {
 868         return evsel->group_name ?: "anon group";
 869 }
 870
 871 /*
 872  * Returns the group details for the specified leader,
 873  * with following rules.
 874  *
 875  *  For record -e '{cycles,instructions}'
 876  *    'anon group { cycles:u, instructions:u }'
 877  *
 878  *  For record -e 'cycles,instructions' and report --group
 879  *    'cycles:u, instructions:u'
 880  */
 881 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 882 {
 883         int ret = 0;
 884         struct evsel *pos;
 885         const char *group_name = evsel__group_name(evsel);
 886
 887         if (!evsel->forced_leader)
 888                 ret = scnprintf(buf, size, "%s { ", group_name);
 889
 890         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 891
 892         for_each_group_member(pos, evsel)
 893                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 894
 895         if (!evsel->forced_leader)
 896                 ret += scnprintf(buf + ret, size - ret, " }");
 897
 898         return ret;
 899 }
 900
 901 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 902                                       struct callchain_param *param)
 903 {
 904         bool function = evsel__is_function_event(evsel);
 905         struct perf_event_attr *attr = &evsel->core.attr;
 906
 907         evsel__set_sample_bit(evsel, CALLCHAIN);
 908
 909         attr->sample_max_stack = param->max_stack;
 910
 911         if (opts->kernel_callchains)
 912                 attr->exclude_callchain_user = 1;
 913         if (opts->user_callchains)
 914                 attr->exclude_callchain_kernel = 1;
 915         if (param->record_mode == CALLCHAIN_LBR) {
 916                 if (!opts->branch_stack) {
 917                         if (attr->exclude_user) {
 918                                 pr_warning("LBR callstack option is only available "
 919                                            "to get user callchain information. "
 920                                            "Falling back to framepointers.\n");
 921                         } else {
 922                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 923                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 924                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 925                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 926                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 927                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 928                         }
 929                 } else
 930                          pr_warning("Cannot use LBR callstack with branch stack. "
 931                                     "Falling back to framepointers.\n");
 932         }
 933
 934         if (param->record_mode == CALLCHAIN_DWARF) {
 935                 if (!function) {
 936                         evsel__set_sample_bit(evsel, REGS_USER);
 937                         evsel__set_sample_bit(evsel, STACK_USER);
 938                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 939                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 940                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 941                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 942                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 943                         } else {
 944                                 attr->sample_regs_user |= arch__user_reg_mask();
 945                         }
 946                         attr->sample_stack_user = param->dump_size;
 947                         attr->exclude_callchain_user = 1;
 948                 } else {
 949                         pr_info("Cannot use DWARF unwind for function trace event,"
 950                                 " falling back to framepointers.\n");
 951                 }
 952         }
 953
 954         if (function) {
 955                 pr_info("Disabling user space callchains for function trace event.\n");
 956                 attr->exclude_callchain_user = 1;
 957         }
 958 }
 959
 960 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 961                              struct callchain_param *param)
 962 {
 963         if (param->enabled)
 964                 return __evsel__config_callchain(evsel, opts, param);
 965 }
 966
 967 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 968 {
 969         struct perf_event_attr *attr = &evsel->core.attr;
 970
 971         evsel__reset_sample_bit(evsel, CALLCHAIN);
 972         if (param->record_mode == CALLCHAIN_LBR) {
 973                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 974                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 975                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 976                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 977         }
 978         if (param->record_mode == CALLCHAIN_DWARF) {
 979                 evsel__reset_sample_bit(evsel, REGS_USER);
 980                 evsel__reset_sample_bit(evsel, STACK_USER);
 981         }
 982 }
 983
 984 static void evsel__apply_config_terms(struct evsel *evsel,
 985                                       struct record_opts *opts, bool track)
 986 {
 987         struct evsel_config_term *term;
 988         struct list_head *config_terms = &evsel->config_terms;
 989         struct perf_event_attr *attr = &evsel->core.attr;
 990         /* callgraph default */
 991         struct callchain_param param = {
 992                 .record_mode = callchain_param.record_mode,
 993         };
 994         u32 dump_size = 0;
 995         int max_stack = 0;
 996         const char *callgraph_buf = NULL;
 997
 998         list_for_each_entry(term, config_terms, list) {
 999                 switch (term->type) {
1000                 case EVSEL__CONFIG_TERM_PERIOD:
1001                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
1002                                 attr->sample_period = term->val.period;
1003                                 attr->freq = 0;
1004                                 evsel__reset_sample_bit(evsel, PERIOD);
1005                         }
1006                         break;
1007                 case EVSEL__CONFIG_TERM_FREQ:
1008                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
1009                                 attr->sample_freq = term->val.freq;
1010                                 attr->freq = 1;
1011                                 evsel__set_sample_bit(evsel, PERIOD);
1012                         }
1013                         break;
1014                 case EVSEL__CONFIG_TERM_TIME:
1015                         if (term->val.time)
1016                                 evsel__set_sample_bit(evsel, TIME);
1017                         else
1018                                 evsel__reset_sample_bit(evsel, TIME);
1019                         break;
1020                 case EVSEL__CONFIG_TERM_CALLGRAPH:
1021                         callgraph_buf = term->val.str;
1022                         break;
1023                 case EVSEL__CONFIG_TERM_BRANCH:
1024                         if (term->val.str && strcmp(term->val.str, "no")) {
1025                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1026                                 parse_branch_str(term->val.str,
1027                                                  &attr->branch_sample_type);
1028                         } else
1029                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1030                         break;
1031                 case EVSEL__CONFIG_TERM_STACK_USER:
1032                         dump_size = term->val.stack_user;
1033                         break;
1034                 case EVSEL__CONFIG_TERM_MAX_STACK:
1035                         max_stack = term->val.max_stack;
1036                         break;
1037                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1038                         evsel->max_events = term->val.max_events;
1039                         break;
1040                 case EVSEL__CONFIG_TERM_INHERIT:
1041                         /*
1042                          * attr->inherit should has already been set by
1043                          * evsel__config. If user explicitly set
1044                          * inherit using config terms, override global
1045                          * opt->no_inherit setting.
1046                          */
1047                         attr->inherit = term->val.inherit ? 1 : 0;
1048                         break;
1049                 case EVSEL__CONFIG_TERM_OVERWRITE:
1050                         attr->write_backward = term->val.overwrite ? 1 : 0;
1051                         break;
1052                 case EVSEL__CONFIG_TERM_DRV_CFG:
1053                         break;
1054                 case EVSEL__CONFIG_TERM_PERCORE:
1055                         break;
1056                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1057                         attr->aux_output = term->val.aux_output ? 1 : 0;
1058                         break;
1059                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1060                         /* Already applied by auxtrace */
1061                         break;
1062                 case EVSEL__CONFIG_TERM_CFG_CHG:
1063                         break;
1064                 default:
1065                         break;
1066                 }
1067         }
1068
1069         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1070         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1071                 bool sample_address = false;
1072
1073                 if (max_stack) {
1074                         param.max_stack = max_stack;
1075                         if (callgraph_buf == NULL)
1076                                 callgraph_buf = "fp";
1077                 }
1078
1079                 /* parse callgraph parameters */
1080                 if (callgraph_buf != NULL) {
1081                         if (!strcmp(callgraph_buf, "no")) {
1082                                 param.enabled = false;
1083                                 param.record_mode = CALLCHAIN_NONE;
1084                         } else {
1085                                 param.enabled = true;
1086                                 if (parse_callchain_record(callgraph_buf, &param)) {
1087                                         pr_err("per-event callgraph setting for %s failed. "
1088                                                "Apply callgraph global setting for it\n",
1089                                                evsel->name);
1090                                         return;
1091                                 }
1092                                 if (param.record_mode == CALLCHAIN_DWARF)
1093                                         sample_address = true;
1094                         }
1095                 }
1096                 if (dump_size > 0) {
1097                         dump_size = round_up(dump_size, sizeof(u64));
1098                         param.dump_size = dump_size;
1099                 }
1100
1101                 /* If global callgraph set, clear it */
1102                 if (callchain_param.enabled)
1103                         evsel__reset_callgraph(evsel, &callchain_param);
1104
1105                 /* set perf-event callgraph */
1106                 if (param.enabled) {
1107                         if (sample_address) {
1108                                 evsel__set_sample_bit(evsel, ADDR);
1109                                 evsel__set_sample_bit(evsel, DATA_SRC);
1110                                 evsel->core.attr.mmap_data = track;
1111                         }
1112                         evsel__config_callchain(evsel, opts, &param);
1113                 }
1114         }
1115 }
1116
1117 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1118 {
1119         struct evsel_config_term *term, *found_term = NULL;
1120
1121         list_for_each_entry(term, &evsel->config_terms, list) {
1122                 if (term->type == type)
1123                         found_term = term;
1124         }
1125
1126         return found_term;
1127 }
1128
1129 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1130 {
1131         evsel__set_sample_bit(evsel, WEIGHT);
1132 }
1133
1134 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1135 {
1136 }
1137
1138 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1139                                     struct perf_event_attr *attr __maybe_unused)
1140 {
1141 }
1142
1143 static void evsel__set_default_freq_period(struct record_opts *opts,
1144                                            struct perf_event_attr *attr)
1145 {
1146         if (opts->freq) {
1147                 attr->freq = 1;
1148                 attr->sample_freq = opts->freq;
1149         } else {
1150                 attr->sample_period = opts->default_interval;
1151         }
1152 }
1153
1154 static bool evsel__is_offcpu_event(struct evsel *evsel)
1155 {
1156         return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1157 }
1158
1159 /*
1160  * The enable_on_exec/disabled value strategy:
1161  *
1162  *  1) For any type of traced program:
1163  *    - all independent events and group leaders are disabled
1164  *    - all group members are enabled
1165  *
1166  *     Group members are ruled by group leaders. They need to
1167  *     be enabled, because the group scheduling relies on that.
1168  *
1169  *  2) For traced programs executed by perf:
1170  *     - all independent events and group leaders have
1171  *       enable_on_exec set
1172  *     - we don't specifically enable or disable any event during
1173  *       the record command
1174  *
1175  *     Independent events and group leaders are initially disabled
1176  *     and get enabled by exec. Group members are ruled by group
1177  *     leaders as stated in 1).
1178  *
1179  *  3) For traced programs attached by perf (pid/tid):
1180  *     - we specifically enable or disable all events during
1181  *       the record command
1182  *
1183  *     When attaching events to already running traced we
1184  *     enable/disable events specifically, as there's no
1185  *     initial traced exec call.
1186  */
1187 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1188                    struct callchain_param *callchain)
1189 {
1190         struct evsel *leader = evsel__leader(evsel);
1191         struct perf_event_attr *attr = &evsel->core.attr;
1192         int track = evsel->tracking;
1193         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1194
1195         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1196         attr->inherit       = !opts->no_inherit;
1197         attr->write_backward = opts->overwrite ? 1 : 0;
1198         attr->read_format   = PERF_FORMAT_LOST;
1199
1200         evsel__set_sample_bit(evsel, IP);
1201         evsel__set_sample_bit(evsel, TID);
1202
1203         if (evsel->sample_read) {
1204                 evsel__set_sample_bit(evsel, READ);
1205
1206                 /*
1207                  * We need ID even in case of single event, because
1208                  * PERF_SAMPLE_READ process ID specific data.
1209                  */
1210                 evsel__set_sample_id(evsel, false);
1211
1212                 /*
1213                  * Apply group format only if we belong to group
1214                  * with more than one members.
1215                  */
1216                 if (leader->core.nr_members > 1) {
1217                         attr->read_format |= PERF_FORMAT_GROUP;
1218                         attr->inherit = 0;
1219                 }
1220         }
1221
1222         /*
1223          * We default some events to have a default interval. But keep
1224          * it a weak assumption overridable by the user.
1225          */
1226         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1227             (!evsel->is_libpfm_event && (!attr->sample_period ||
1228                                          opts->user_freq != UINT_MAX ||
1229                                          opts->user_interval != ULLONG_MAX)))
1230                 evsel__set_default_freq_period(opts, attr);
1231
1232         /*
1233          * If attr->freq was set (here or earlier), ask for period
1234          * to be sampled.
1235          */
1236         if (attr->freq)
1237                 evsel__set_sample_bit(evsel, PERIOD);
1238
1239         if (opts->no_samples)
1240                 attr->sample_freq = 0;
1241
1242         if (opts->inherit_stat) {
1243                 evsel->core.attr.read_format |=
1244                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1245                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1246                         PERF_FORMAT_ID;
1247                 attr->inherit_stat = 1;
1248         }
1249
1250         if (opts->sample_address) {
1251                 evsel__set_sample_bit(evsel, ADDR);
1252                 attr->mmap_data = track;
1253         }
1254
1255         /*
1256          * We don't allow user space callchains for  function trace
1257          * event, due to issues with page faults while tracing page
1258          * fault handler and its overall trickiness nature.
1259          */
1260         if (evsel__is_function_event(evsel))
1261                 evsel->core.attr.exclude_callchain_user = 1;
1262
1263         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1264                 evsel__config_callchain(evsel, opts, callchain);
1265
1266         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1267             !evsel__is_dummy_event(evsel)) {
1268                 attr->sample_regs_intr = opts->sample_intr_regs;
1269                 evsel__set_sample_bit(evsel, REGS_INTR);
1270         }
1271
1272         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1273             !evsel__is_dummy_event(evsel)) {
1274                 attr->sample_regs_user |= opts->sample_user_regs;
1275                 evsel__set_sample_bit(evsel, REGS_USER);
1276         }
1277
1278         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1279                 evsel__set_sample_bit(evsel, CPU);
1280
1281         /*
1282          * When the user explicitly disabled time don't force it here.
1283          */
1284         if (opts->sample_time &&
1285             (!perf_missing_features.sample_id_all &&
1286             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1287              opts->sample_time_set)))
1288                 evsel__set_sample_bit(evsel, TIME);
1289
1290         if (opts->raw_samples && !evsel->no_aux_samples) {
1291                 evsel__set_sample_bit(evsel, TIME);
1292                 evsel__set_sample_bit(evsel, RAW);
1293                 evsel__set_sample_bit(evsel, CPU);
1294         }
1295
1296         if (opts->sample_address)
1297                 evsel__set_sample_bit(evsel, DATA_SRC);
1298
1299         if (opts->sample_phys_addr)
1300                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1301
1302         if (opts->no_buffering) {
1303                 attr->watermark = 0;
1304                 attr->wakeup_events = 1;
1305         }
1306         if (opts->branch_stack && !evsel->no_aux_samples) {
1307                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1308                 attr->branch_sample_type = opts->branch_stack;
1309         }
1310
1311         if (opts->sample_weight)
1312                 arch_evsel__set_sample_weight(evsel);
1313
1314         attr->task     = track;
1315         attr->mmap     = track;
1316         attr->mmap2    = track && !perf_missing_features.mmap2;
1317         attr->comm     = track;
1318         attr->build_id = track && opts->build_id;
1319
1320         /*
1321          * ksymbol is tracked separately with text poke because it needs to be
1322          * system wide and enabled immediately.
1323          */
1324         if (!opts->text_poke)
1325                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1326         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1327
1328         if (opts->record_namespaces)
1329                 attr->namespaces  = track;
1330
1331         if (opts->record_cgroup) {
1332                 attr->cgroup = track && !perf_missing_features.cgroup;
1333                 evsel__set_sample_bit(evsel, CGROUP);
1334         }
1335
1336         if (opts->sample_data_page_size)
1337                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1338
1339         if (opts->sample_code_page_size)
1340                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1341
1342         if (opts->record_switch_events)
1343                 attr->context_switch = track;
1344
1345         if (opts->sample_transaction)
1346                 evsel__set_sample_bit(evsel, TRANSACTION);
1347
1348         if (opts->running_time) {
1349                 evsel->core.attr.read_format |=
1350                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1351                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1352         }
1353
1354         /*
1355          * XXX see the function comment above
1356          *
1357          * Disabling only independent events or group leaders,
1358          * keeping group members enabled.
1359          */
1360         if (evsel__is_group_leader(evsel))
1361                 attr->disabled = 1;
1362
1363         /*
1364          * Setting enable_on_exec for independent events and
1365          * group leaders for traced executed by perf.
1366          */
1367         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1368             !opts->target.initial_delay)
1369                 attr->enable_on_exec = 1;
1370
1371         if (evsel->immediate) {
1372                 attr->disabled = 0;
1373                 attr->enable_on_exec = 0;
1374         }
1375
1376         clockid = opts->clockid;
1377         if (opts->use_clockid) {
1378                 attr->use_clockid = 1;
1379                 attr->clockid = opts->clockid;
1380         }
1381
1382         if (evsel->precise_max)
1383                 attr->precise_ip = 3;
1384
1385         if (opts->all_user) {
1386                 attr->exclude_kernel = 1;
1387                 attr->exclude_user   = 0;
1388         }
1389
1390         if (opts->all_kernel) {
1391                 attr->exclude_kernel = 0;
1392                 attr->exclude_user   = 1;
1393         }
1394
1395         if (evsel->core.own_cpus || evsel->unit)
1396                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1397
1398         /*
1399          * Apply event specific term settings,
1400          * it overloads any global configuration.
1401          */
1402         evsel__apply_config_terms(evsel, opts, track);
1403
1404         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1405
1406         /* The --period option takes the precedence. */
1407         if (opts->period_set) {
1408                 if (opts->period)
1409                         evsel__set_sample_bit(evsel, PERIOD);
1410                 else
1411                         evsel__reset_sample_bit(evsel, PERIOD);
1412         }
1413
1414         /*
1415          * A dummy event never triggers any actual counter and therefore
1416          * cannot be used with branch_stack.
1417          *
1418          * For initial_delay, a dummy event is added implicitly.
1419          * The software event will trigger -EOPNOTSUPP error out,
1420          * if BRANCH_STACK bit is set.
1421          */
1422         if (evsel__is_dummy_event(evsel))
1423                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1424
1425         if (evsel__is_offcpu_event(evsel))
1426                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1427
1428         arch__post_evsel_config(evsel, attr);
1429 }
1430
1431 int evsel__set_filter(struct evsel *evsel, const char *filter)
1432 {
1433         char *new_filter = strdup(filter);
1434
1435         if (new_filter != NULL) {
1436                 free(evsel->filter);
1437                 evsel->filter = new_filter;
1438                 return 0;
1439         }
1440
1441         return -1;
1442 }
1443
1444 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1445 {
1446         char *new_filter;
1447
1448         if (evsel->filter == NULL)
1449                 return evsel__set_filter(evsel, filter);
1450
1451         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1452                 free(evsel->filter);
1453                 evsel->filter = new_filter;
1454                 return 0;
1455         }
1456
1457         return -1;
1458 }
1459
1460 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1461 {
1462         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1463 }
1464
1465 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1466 {
1467         return evsel__append_filter(evsel, "%s,%s", filter);
1468 }
1469
1470 /* Caller has to clear disabled after going through all CPUs. */
1471 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1472 {
1473         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1474 }
1475
1476 int evsel__enable(struct evsel *evsel)
1477 {
1478         int err = perf_evsel__enable(&evsel->core);
1479
1480         if (!err)
1481                 evsel->disabled = false;
1482         return err;
1483 }
1484
1485 /* Caller has to set disabled after going through all CPUs. */
1486 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1487 {
1488         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1489 }
1490
1491 int evsel__disable(struct evsel *evsel)
1492 {
1493         int err = perf_evsel__disable(&evsel->core);
1494         /*
1495          * We mark it disabled here so that tools that disable a event can
1496          * ignore events after they disable it. I.e. the ring buffer may have
1497          * already a few more events queued up before the kernel got the stop
1498          * request.
1499          */
1500         if (!err)
1501                 evsel->disabled = true;
1502
1503         return err;
1504 }
1505
1506 void free_config_terms(struct list_head *config_terms)
1507 {
1508         struct evsel_config_term *term, *h;
1509
1510         list_for_each_entry_safe(term, h, config_terms, list) {
1511                 list_del_init(&term->list);
1512                 if (term->free_str)
1513                         zfree(&term->val.str);
1514                 free(term);
1515         }
1516 }
1517
1518 static void evsel__free_config_terms(struct evsel *evsel)
1519 {
1520         free_config_terms(&evsel->config_terms);
1521 }
1522
1523 void evsel__exit(struct evsel *evsel)
1524 {
1525         assert(list_empty(&evsel->core.node));
1526         assert(evsel->evlist == NULL);
1527         bpf_counter__destroy(evsel);
1528         perf_bpf_filter__destroy(evsel);
1529         evsel__free_counts(evsel);
1530         perf_evsel__free_fd(&evsel->core);
1531         perf_evsel__free_id(&evsel->core);
1532         evsel__free_config_terms(evsel);
1533         cgroup__put(evsel->cgrp);
1534         perf_cpu_map__put(evsel->core.cpus);
1535         perf_cpu_map__put(evsel->core.own_cpus);
1536         perf_thread_map__put(evsel->core.threads);
1537         zfree(&evsel->group_name);
1538         zfree(&evsel->name);
1539         zfree(&evsel->pmu_name);
1540         zfree(&evsel->unit);
1541         zfree(&evsel->metric_id);
1542         evsel__zero_per_pkg(evsel);
1543         hashmap__free(evsel->per_pkg_mask);
1544         evsel->per_pkg_mask = NULL;
1545         zfree(&evsel->metric_events);
1546         perf_evsel__object.fini(evsel);
1547 }
1548
1549 void evsel__delete(struct evsel *evsel)
1550 {
1551         if (!evsel)
1552                 return;
1553
1554         evsel__exit(evsel);
1555         free(evsel);
1556 }
1557
1558 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1559                            struct perf_counts_values *count)
1560 {
1561         struct perf_counts_values tmp;
1562
1563         if (!evsel->prev_raw_counts)
1564                 return;
1565
1566         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1567         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1568
1569         count->val = count->val - tmp.val;
1570         count->ena = count->ena - tmp.ena;
1571         count->run = count->run - tmp.run;
1572 }
1573
1574 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1575 {
1576         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1577
1578         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1579 }
1580
1581 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1582                              u64 val, u64 ena, u64 run, u64 lost)
1583 {
1584         struct perf_counts_values *count;
1585
1586         count = perf_counts(counter->counts, cpu_map_idx, thread);
1587
1588         count->val    = val;
1589         count->ena    = ena;
1590         count->run    = run;
1591         count->lost   = lost;
1592
1593         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1594 }
1595
1596 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1597 {
1598         u64 read_format = leader->core.attr.read_format;
1599         struct sample_read_value *v;
1600         u64 nr, ena = 0, run = 0, lost = 0;
1601
1602         nr = *data++;
1603
1604         if (nr != (u64) leader->core.nr_members)
1605                 return -EINVAL;
1606
1607         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1608                 ena = *data++;
1609
1610         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1611                 run = *data++;
1612
1613         v = (void *)data;
1614         sample_read_group__for_each(v, nr, read_format) {
1615                 struct evsel *counter;
1616
1617                 counter = evlist__id2evsel(leader->evlist, v->id);
1618                 if (!counter)
1619                         return -EINVAL;
1620
1621                 if (read_format & PERF_FORMAT_LOST)
1622                         lost = v->lost;
1623
1624                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1625         }
1626
1627         return 0;
1628 }
1629
1630 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1631 {
1632         struct perf_stat_evsel *ps = leader->stats;
1633         u64 read_format = leader->core.attr.read_format;
1634         int size = perf_evsel__read_size(&leader->core);
1635         u64 *data = ps->group_data;
1636
1637         if (!(read_format & PERF_FORMAT_ID))
1638                 return -EINVAL;
1639
1640         if (!evsel__is_group_leader(leader))
1641                 return -EINVAL;
1642
1643         if (!data) {
1644                 data = zalloc(size);
1645                 if (!data)
1646                         return -ENOMEM;
1647
1648                 ps->group_data = data;
1649         }
1650
1651         if (FD(leader, cpu_map_idx, thread) < 0)
1652                 return -EINVAL;
1653
1654         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1655                 return -errno;
1656
1657         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1658 }
1659
1660 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1661 {
1662         u64 read_format = evsel->core.attr.read_format;
1663
1664         if (read_format & PERF_FORMAT_GROUP)
1665                 return evsel__read_group(evsel, cpu_map_idx, thread);
1666
1667         return evsel__read_one(evsel, cpu_map_idx, thread);
1668 }
1669
1670 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1671 {
1672         struct perf_counts_values count;
1673         size_t nv = scale ? 3 : 1;
1674
1675         if (FD(evsel, cpu_map_idx, thread) < 0)
1676                 return -EINVAL;
1677
1678         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1679                 return -ENOMEM;
1680
1681         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1682                 return -errno;
1683
1684         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1685         perf_counts_values__scale(&count, scale, NULL);
1686         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1687         return 0;
1688 }
1689
1690 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1691                                   int cpu_map_idx)
1692 {
1693         struct perf_cpu cpu;
1694
1695         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1696         return perf_cpu_map__idx(other->core.cpus, cpu);
1697 }
1698
1699 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1700 {
1701         struct evsel *leader = evsel__leader(evsel);
1702
1703         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1704             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1705                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1706         }
1707
1708         return cpu_map_idx;
1709 }
1710
1711 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1712 {
1713         struct evsel *leader = evsel__leader(evsel);
1714         int fd;
1715
1716         if (evsel__is_group_leader(evsel))
1717                 return -1;
1718
1719         /*
1720          * Leader must be already processed/open,
1721          * if not it's a bug.
1722          */
1723         BUG_ON(!leader->core.fd);
1724
1725         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1726         if (cpu_map_idx == -1)
1727                 return -1;
1728
1729         fd = FD(leader, cpu_map_idx, thread);
1730         BUG_ON(fd == -1 && !leader->skippable);
1731
1732         /*
1733          * When the leader has been skipped, return -2 to distinguish from no
1734          * group leader case.
1735          */
1736         return fd == -1 ? -2 : fd;
1737 }
1738
1739 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1740 {
1741         for (int cpu = 0; cpu < nr_cpus; cpu++)
1742                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1743                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1744 }
1745
1746 static int update_fds(struct evsel *evsel,
1747                       int nr_cpus, int cpu_map_idx,
1748                       int nr_threads, int thread_idx)
1749 {
1750         struct evsel *pos;
1751
1752         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1753                 return -EINVAL;
1754
1755         evlist__for_each_entry(evsel->evlist, pos) {
1756                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1757
1758                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1759
1760                 /*
1761                  * Since fds for next evsel has not been created,
1762                  * there is no need to iterate whole event list.
1763                  */
1764                 if (pos == evsel)
1765                         break;
1766         }
1767         return 0;
1768 }
1769
1770 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1771                                          int nr_cpus, int cpu_map_idx,
1772                                          struct perf_thread_map *threads,
1773                                          int thread, int err)
1774 {
1775         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1776
1777         if (!evsel->ignore_missing_thread)
1778                 return false;
1779
1780         /* The system wide setup does not work with threads. */
1781         if (evsel->core.system_wide)
1782                 return false;
1783
1784         /* The -ESRCH is perf event syscall errno for pid's not found. */
1785         if (err != -ESRCH)
1786                 return false;
1787
1788         /* If there's only one thread, let it fail. */
1789         if (threads->nr == 1)
1790                 return false;
1791
1792         /*
1793          * We should remove fd for missing_thread first
1794          * because thread_map__remove() will decrease threads->nr.
1795          */
1796         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1797                 return false;
1798
1799         if (thread_map__remove(threads, thread))
1800                 return false;
1801
1802         pr_warning("WARNING: Ignored open failure for pid %d\n",
1803                    ignore_pid);
1804         return true;
1805 }
1806
1807 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1808                                 void *priv __maybe_unused)
1809 {
1810         return fprintf(fp, "  %-32s %s\n", name, val);
1811 }
1812
1813 static void display_attr(struct perf_event_attr *attr)
1814 {
1815         if (verbose >= 2 || debug_peo_args) {
1816                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1817                 fprintf(stderr, "perf_event_attr:\n");
1818                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1819                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1820         }
1821 }
1822
1823 bool evsel__precise_ip_fallback(struct evsel *evsel)
1824 {
1825         /* Do not try less precise if not requested. */
1826         if (!evsel->precise_max)
1827                 return false;
1828
1829         /*
1830          * We tried all the precise_ip values, and it's
1831          * still failing, so leave it to standard fallback.
1832          */
1833         if (!evsel->core.attr.precise_ip) {
1834                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1835                 return false;
1836         }
1837
1838         if (!evsel->precise_ip_original)
1839                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1840
1841         evsel->core.attr.precise_ip--;
1842         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1843         display_attr(&evsel->core.attr);
1844         return true;
1845 }
1846
1847 static struct perf_cpu_map *empty_cpu_map;
1848 static struct perf_thread_map *empty_thread_map;
1849
1850 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1851                 struct perf_thread_map *threads)
1852 {
1853         int nthreads = perf_thread_map__nr(threads);
1854
1855         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1856             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1857                 return -EINVAL;
1858
1859         if (cpus == NULL) {
1860                 if (empty_cpu_map == NULL) {
1861                         empty_cpu_map = perf_cpu_map__dummy_new();
1862                         if (empty_cpu_map == NULL)
1863                                 return -ENOMEM;
1864                 }
1865
1866                 cpus = empty_cpu_map;
1867         }
1868
1869         if (threads == NULL) {
1870                 if (empty_thread_map == NULL) {
1871                         empty_thread_map = thread_map__new_by_tid(-1);
1872                         if (empty_thread_map == NULL)
1873                                 return -ENOMEM;
1874                 }
1875
1876                 threads = empty_thread_map;
1877         }
1878
1879         if (evsel->core.fd == NULL &&
1880             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1881                 return -ENOMEM;
1882
1883         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1884         if (evsel->cgrp)
1885                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1886
1887         return 0;
1888 }
1889
1890 static void evsel__disable_missing_features(struct evsel *evsel)
1891 {
1892         if (perf_missing_features.read_lost)
1893                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1894         if (perf_missing_features.weight_struct) {
1895                 evsel__set_sample_bit(evsel, WEIGHT);
1896                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1897         }
1898         if (perf_missing_features.clockid_wrong)
1899                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1900         if (perf_missing_features.clockid) {
1901                 evsel->core.attr.use_clockid = 0;
1902                 evsel->core.attr.clockid = 0;
1903         }
1904         if (perf_missing_features.cloexec)
1905                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1906         if (perf_missing_features.mmap2)
1907                 evsel->core.attr.mmap2 = 0;
1908         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1909                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1910         if (perf_missing_features.lbr_flags)
1911                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1912                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1913         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1914                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1915         if (perf_missing_features.ksymbol)
1916                 evsel->core.attr.ksymbol = 0;
1917         if (perf_missing_features.bpf)
1918                 evsel->core.attr.bpf_event = 0;
1919         if (perf_missing_features.branch_hw_idx)
1920                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1921         if (perf_missing_features.sample_id_all)
1922                 evsel->core.attr.sample_id_all = 0;
1923 }
1924
1925 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1926                         struct perf_thread_map *threads)
1927 {
1928         int err;
1929
1930         err = __evsel__prepare_open(evsel, cpus, threads);
1931         if (err)
1932                 return err;
1933
1934         evsel__disable_missing_features(evsel);
1935
1936         return err;
1937 }
1938
1939 bool evsel__detect_missing_features(struct evsel *evsel)
1940 {
1941         /*
1942          * Must probe features in the order they were added to the
1943          * perf_event_attr interface.
1944          */
1945         if (!perf_missing_features.read_lost &&
1946             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1947                 perf_missing_features.read_lost = true;
1948                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1949                 return true;
1950         } else if (!perf_missing_features.weight_struct &&
1951             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1952                 perf_missing_features.weight_struct = true;
1953                 pr_debug2("switching off weight struct support\n");
1954                 return true;
1955         } else if (!perf_missing_features.code_page_size &&
1956             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1957                 perf_missing_features.code_page_size = true;
1958                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1959                 return false;
1960         } else if (!perf_missing_features.data_page_size &&
1961             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1962                 perf_missing_features.data_page_size = true;
1963                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1964                 return false;
1965         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1966                 perf_missing_features.cgroup = true;
1967                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1968                 return false;
1969         } else if (!perf_missing_features.branch_hw_idx &&
1970             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1971                 perf_missing_features.branch_hw_idx = true;
1972                 pr_debug2("switching off branch HW index support\n");
1973                 return true;
1974         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1975                 perf_missing_features.aux_output = true;
1976                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1977                 return false;
1978         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1979                 perf_missing_features.bpf = true;
1980                 pr_debug2_peo("switching off bpf_event\n");
1981                 return true;
1982         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1983                 perf_missing_features.ksymbol = true;
1984                 pr_debug2_peo("switching off ksymbol\n");
1985                 return true;
1986         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1987                 perf_missing_features.write_backward = true;
1988                 pr_debug2_peo("switching off write_backward\n");
1989                 return false;
1990         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1991                 perf_missing_features.clockid_wrong = true;
1992                 pr_debug2_peo("switching off clockid\n");
1993                 return true;
1994         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1995                 perf_missing_features.clockid = true;
1996                 pr_debug2_peo("switching off use_clockid\n");
1997                 return true;
1998         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1999                 perf_missing_features.cloexec = true;
2000                 pr_debug2_peo("switching off cloexec flag\n");
2001                 return true;
2002         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
2003                 perf_missing_features.mmap2 = true;
2004                 pr_debug2_peo("switching off mmap2\n");
2005                 return true;
2006         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
2007                 if (evsel->pmu == NULL)
2008                         evsel->pmu = evsel__find_pmu(evsel);
2009
2010                 if (evsel->pmu)
2011                         evsel->pmu->missing_features.exclude_guest = true;
2012                 else {
2013                         /* we cannot find PMU, disable attrs now */
2014                         evsel->core.attr.exclude_host = false;
2015                         evsel->core.attr.exclude_guest = false;
2016                 }
2017
2018                 if (evsel->exclude_GH) {
2019                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2020                         return false;
2021                 }
2022                 if (!perf_missing_features.exclude_guest) {
2023                         perf_missing_features.exclude_guest = true;
2024                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2025                 }
2026                 return true;
2027         } else if (!perf_missing_features.sample_id_all) {
2028                 perf_missing_features.sample_id_all = true;
2029                 pr_debug2_peo("switching off sample_id_all\n");
2030                 return true;
2031         } else if (!perf_missing_features.lbr_flags &&
2032                         (evsel->core.attr.branch_sample_type &
2033                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2034                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2035                 perf_missing_features.lbr_flags = true;
2036                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2037                 return true;
2038         } else if (!perf_missing_features.group_read &&
2039                     evsel->core.attr.inherit &&
2040                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2041                    evsel__is_group_leader(evsel)) {
2042                 perf_missing_features.group_read = true;
2043                 pr_debug2_peo("switching off group read\n");
2044                 return true;
2045         } else {
2046                 return false;
2047         }
2048 }
2049
2050 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2051 {
2052         int old_errno;
2053         struct rlimit l;
2054
2055         if (*set_rlimit < INCREASED_MAX) {
2056                 old_errno = errno;
2057
2058                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2059                         if (*set_rlimit == NO_CHANGE) {
2060                                 l.rlim_cur = l.rlim_max;
2061                         } else {
2062                                 l.rlim_cur = l.rlim_max + 1000;
2063                                 l.rlim_max = l.rlim_cur;
2064                         }
2065                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2066                                 (*set_rlimit) += 1;
2067                                 errno = old_errno;
2068                                 return true;
2069                         }
2070                 }
2071                 errno = old_errno;
2072         }
2073
2074         return false;
2075 }
2076
2077 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2078                 struct perf_thread_map *threads,
2079                 int start_cpu_map_idx, int end_cpu_map_idx)
2080 {
2081         int idx, thread, nthreads;
2082         int pid = -1, err, old_errno;
2083         enum rlimit_action set_rlimit = NO_CHANGE;
2084
2085         err = __evsel__prepare_open(evsel, cpus, threads);
2086         if (err)
2087                 return err;
2088
2089         if (cpus == NULL)
2090                 cpus = empty_cpu_map;
2091
2092         if (threads == NULL)
2093                 threads = empty_thread_map;
2094
2095         nthreads = perf_thread_map__nr(threads);
2096
2097         if (evsel->cgrp)
2098                 pid = evsel->cgrp->fd;
2099
2100 fallback_missing_features:
2101         evsel__disable_missing_features(evsel);
2102
2103         display_attr(&evsel->core.attr);
2104
2105         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2106
2107                 for (thread = 0; thread < nthreads; thread++) {
2108                         int fd, group_fd;
2109 retry_open:
2110                         if (thread >= nthreads)
2111                                 break;
2112
2113                         if (!evsel->cgrp && !evsel->core.system_wide)
2114                                 pid = perf_thread_map__pid(threads, thread);
2115
2116                         group_fd = get_group_fd(evsel, idx, thread);
2117
2118                         if (group_fd == -2) {
2119                                 pr_debug("broken group leader for %s\n", evsel->name);
2120                                 err = -EINVAL;
2121                                 goto out_close;
2122                         }
2123
2124                         test_attr__ready();
2125
2126                         /* Debug message used by test scripts */
2127                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2128                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2129
2130                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2131                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2132                                                 group_fd, evsel->open_flags);
2133
2134                         FD(evsel, idx, thread) = fd;
2135
2136                         if (fd < 0) {
2137                                 err = -errno;
2138
2139                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2140                                           err);
2141                                 goto try_fallback;
2142                         }
2143
2144                         bpf_counter__install_pe(evsel, idx, fd);
2145
2146                         if (unlikely(test_attr__enabled)) {
2147                                 test_attr__open(&evsel->core.attr, pid,
2148                                                 perf_cpu_map__cpu(cpus, idx),
2149                                                 fd, group_fd, evsel->open_flags);
2150                         }
2151
2152                         /* Debug message used by test scripts */
2153                         pr_debug2_peo(" = %d\n", fd);
2154
2155                         if (evsel->bpf_fd >= 0) {
2156                                 int evt_fd = fd;
2157                                 int bpf_fd = evsel->bpf_fd;
2158
2159                                 err = ioctl(evt_fd,
2160                                             PERF_EVENT_IOC_SET_BPF,
2161                                             bpf_fd);
2162                                 if (err && errno != EEXIST) {
2163                                         pr_err("failed to attach bpf fd %d: %s\n",
2164                                                bpf_fd, strerror(errno));
2165                                         err = -EINVAL;
2166                                         goto out_close;
2167                                 }
2168                         }
2169
2170                         set_rlimit = NO_CHANGE;
2171
2172                         /*
2173                          * If we succeeded but had to kill clockid, fail and
2174                          * have evsel__open_strerror() print us a nice error.
2175                          */
2176                         if (perf_missing_features.clockid ||
2177                             perf_missing_features.clockid_wrong) {
2178                                 err = -EINVAL;
2179                                 goto out_close;
2180                         }
2181                 }
2182         }
2183
2184         return 0;
2185
2186 try_fallback:
2187         if (evsel__precise_ip_fallback(evsel))
2188                 goto retry_open;
2189
2190         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2191                                          idx, threads, thread, err)) {
2192                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2193                 nthreads--;
2194
2195                 /* ... and pretend like nothing have happened. */
2196                 err = 0;
2197                 goto retry_open;
2198         }
2199         /*
2200          * perf stat needs between 5 and 22 fds per CPU. When we run out
2201          * of them try to increase the limits.
2202          */
2203         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2204                 goto retry_open;
2205
2206         if (err != -EINVAL || idx > 0 || thread > 0)
2207                 goto out_close;
2208
2209         if (evsel__detect_missing_features(evsel))
2210                 goto fallback_missing_features;
2211 out_close:
2212         if (err)
2213                 threads->err_thread = thread;
2214
2215         old_errno = errno;
2216         do {
2217                 while (--thread >= 0) {
2218                         if (FD(evsel, idx, thread) >= 0)
2219                                 close(FD(evsel, idx, thread));
2220                         FD(evsel, idx, thread) = -1;
2221                 }
2222                 thread = nthreads;
2223         } while (--idx >= 0);
2224         errno = old_errno;
2225         return err;
2226 }
2227
2228 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2229                 struct perf_thread_map *threads)
2230 {
2231         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2232 }
2233
2234 void evsel__close(struct evsel *evsel)
2235 {
2236         perf_evsel__close(&evsel->core);
2237         perf_evsel__free_id(&evsel->core);
2238 }
2239
2240 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2241 {
2242         if (cpu_map_idx == -1)
2243                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2244
2245         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2246 }
2247
2248 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2249 {
2250         return evsel__open(evsel, NULL, threads);
2251 }
2252
2253 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2254                                        const union perf_event *event,
2255                                        struct perf_sample *sample)
2256 {
2257         u64 type = evsel->core.attr.sample_type;
2258         const __u64 *array = event->sample.array;
2259         bool swapped = evsel->needs_swap;
2260         union u64_swap u;
2261
2262         array += ((event->header.size -
2263                    sizeof(event->header)) / sizeof(u64)) - 1;
2264
2265         if (type & PERF_SAMPLE_IDENTIFIER) {
2266                 sample->id = *array;
2267                 array--;
2268         }
2269
2270         if (type & PERF_SAMPLE_CPU) {
2271                 u.val64 = *array;
2272                 if (swapped) {
2273                         /* undo swap of u64, then swap on individual u32s */
2274                         u.val64 = bswap_64(u.val64);
2275                         u.val32[0] = bswap_32(u.val32[0]);
2276                 }
2277
2278                 sample->cpu = u.val32[0];
2279                 array--;
2280         }
2281
2282         if (type & PERF_SAMPLE_STREAM_ID) {
2283                 sample->stream_id = *array;
2284                 array--;
2285         }
2286
2287         if (type & PERF_SAMPLE_ID) {
2288                 sample->id = *array;
2289                 array--;
2290         }
2291
2292         if (type & PERF_SAMPLE_TIME) {
2293                 sample->time = *array;
2294                 array--;
2295         }
2296
2297         if (type & PERF_SAMPLE_TID) {
2298                 u.val64 = *array;
2299                 if (swapped) {
2300                         /* undo swap of u64, then swap on individual u32s */
2301                         u.val64 = bswap_64(u.val64);
2302                         u.val32[0] = bswap_32(u.val32[0]);
2303                         u.val32[1] = bswap_32(u.val32[1]);
2304                 }
2305
2306                 sample->pid = u.val32[0];
2307                 sample->tid = u.val32[1];
2308                 array--;
2309         }
2310
2311         return 0;
2312 }
2313
2314 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2315                             u64 size)
2316 {
2317         return size > max_size || offset + size > endp;
2318 }
2319
2320 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2321         do {                                                            \
2322                 if (overflow(endp, (max_size), (offset), (size)))       \
2323                         return -EFAULT;                                 \
2324         } while (0)
2325
2326 #define OVERFLOW_CHECK_u64(offset) \
2327         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2328
2329 static int
2330 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2331 {
2332         /*
2333          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2334          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2335          * check the format does not go past the end of the event.
2336          */
2337         if (sample_size + sizeof(event->header) > event->header.size)
2338                 return -EFAULT;
2339
2340         return 0;
2341 }
2342
2343 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2344                                           const __u64 *array,
2345                                           u64 type __maybe_unused)
2346 {
2347         data->weight = *array;
2348 }
2349
2350 u64 evsel__bitfield_swap_branch_flags(u64 value)
2351 {
2352         u64 new_val = 0;
2353
2354         /*
2355          * branch_flags
2356          * union {
2357          *      u64 values;
2358          *      struct {
2359          *              mispred:1       //target mispredicted
2360          *              predicted:1     //target predicted
2361          *              in_tx:1         //in transaction
2362          *              abort:1         //transaction abort
2363          *              cycles:16       //cycle count to last branch
2364          *              type:4          //branch type
2365          *              spec:2          //branch speculation info
2366          *              new_type:4      //additional branch type
2367          *              priv:3          //privilege level
2368          *              reserved:31
2369          *      }
2370          * }
2371          *
2372          * Avoid bswap64() the entire branch_flag.value,
2373          * as it has variable bit-field sizes. Instead the
2374          * macro takes the bit-field position/size,
2375          * swaps it based on the host endianness.
2376          */
2377         if (host_is_bigendian()) {
2378                 new_val = bitfield_swap(value, 0, 1);
2379                 new_val |= bitfield_swap(value, 1, 1);
2380                 new_val |= bitfield_swap(value, 2, 1);
2381                 new_val |= bitfield_swap(value, 3, 1);
2382                 new_val |= bitfield_swap(value, 4, 16);
2383                 new_val |= bitfield_swap(value, 20, 4);
2384                 new_val |= bitfield_swap(value, 24, 2);
2385                 new_val |= bitfield_swap(value, 26, 4);
2386                 new_val |= bitfield_swap(value, 30, 3);
2387                 new_val |= bitfield_swap(value, 33, 31);
2388         } else {
2389                 new_val = bitfield_swap(value, 63, 1);
2390                 new_val |= bitfield_swap(value, 62, 1);
2391                 new_val |= bitfield_swap(value, 61, 1);
2392                 new_val |= bitfield_swap(value, 60, 1);
2393                 new_val |= bitfield_swap(value, 44, 16);
2394                 new_val |= bitfield_swap(value, 40, 4);
2395                 new_val |= bitfield_swap(value, 38, 2);
2396                 new_val |= bitfield_swap(value, 34, 4);
2397                 new_val |= bitfield_swap(value, 31, 3);
2398                 new_val |= bitfield_swap(value, 0, 31);
2399         }
2400
2401         return new_val;
2402 }
2403
2404 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2405                         struct perf_sample *data)
2406 {
2407         u64 type = evsel->core.attr.sample_type;
2408         bool swapped = evsel->needs_swap;
2409         const __u64 *array;
2410         u16 max_size = event->header.size;
2411         const void *endp = (void *)event + max_size;
2412         u64 sz;
2413
2414         /*
2415          * used for cross-endian analysis. See git commit 65014ab3
2416          * for why this goofiness is needed.
2417          */
2418         union u64_swap u;
2419
2420         memset(data, 0, sizeof(*data));
2421         data->cpu = data->pid = data->tid = -1;
2422         data->stream_id = data->id = data->time = -1ULL;
2423         data->period = evsel->core.attr.sample_period;
2424         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2425         data->misc    = event->header.misc;
2426         data->id = -1ULL;
2427         data->data_src = PERF_MEM_DATA_SRC_NONE;
2428         data->vcpu = -1;
2429
2430         if (event->header.type != PERF_RECORD_SAMPLE) {
2431                 if (!evsel->core.attr.sample_id_all)
2432                         return 0;
2433                 return perf_evsel__parse_id_sample(evsel, event, data);
2434         }
2435
2436         array = event->sample.array;
2437
2438         if (perf_event__check_size(event, evsel->sample_size))
2439                 return -EFAULT;
2440
2441         if (type & PERF_SAMPLE_IDENTIFIER) {
2442                 data->id = *array;
2443                 array++;
2444         }
2445
2446         if (type & PERF_SAMPLE_IP) {
2447                 data->ip = *array;
2448                 array++;
2449         }
2450
2451         if (type & PERF_SAMPLE_TID) {
2452                 u.val64 = *array;
2453                 if (swapped) {
2454                         /* undo swap of u64, then swap on individual u32s */
2455                         u.val64 = bswap_64(u.val64);
2456                         u.val32[0] = bswap_32(u.val32[0]);
2457                         u.val32[1] = bswap_32(u.val32[1]);
2458                 }
2459
2460                 data->pid = u.val32[0];
2461                 data->tid = u.val32[1];
2462                 array++;
2463         }
2464
2465         if (type & PERF_SAMPLE_TIME) {
2466                 data->time = *array;
2467                 array++;
2468         }
2469
2470         if (type & PERF_SAMPLE_ADDR) {
2471                 data->addr = *array;
2472                 array++;
2473         }
2474
2475         if (type & PERF_SAMPLE_ID) {
2476                 data->id = *array;
2477                 array++;
2478         }
2479
2480         if (type & PERF_SAMPLE_STREAM_ID) {
2481                 data->stream_id = *array;
2482                 array++;
2483         }
2484
2485         if (type & PERF_SAMPLE_CPU) {
2486
2487                 u.val64 = *array;
2488                 if (swapped) {
2489                         /* undo swap of u64, then swap on individual u32s */
2490                         u.val64 = bswap_64(u.val64);
2491                         u.val32[0] = bswap_32(u.val32[0]);
2492                 }
2493
2494                 data->cpu = u.val32[0];
2495                 array++;
2496         }
2497
2498         if (type & PERF_SAMPLE_PERIOD) {
2499                 data->period = *array;
2500                 array++;
2501         }
2502
2503         if (type & PERF_SAMPLE_READ) {
2504                 u64 read_format = evsel->core.attr.read_format;
2505
2506                 OVERFLOW_CHECK_u64(array);
2507                 if (read_format & PERF_FORMAT_GROUP)
2508                         data->read.group.nr = *array;
2509                 else
2510                         data->read.one.value = *array;
2511
2512                 array++;
2513
2514                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2515                         OVERFLOW_CHECK_u64(array);
2516                         data->read.time_enabled = *array;
2517                         array++;
2518                 }
2519
2520                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2521                         OVERFLOW_CHECK_u64(array);
2522                         data->read.time_running = *array;
2523                         array++;
2524                 }
2525
2526                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2527                 if (read_format & PERF_FORMAT_GROUP) {
2528                         const u64 max_group_nr = UINT64_MAX /
2529                                         sizeof(struct sample_read_value);
2530
2531                         if (data->read.group.nr > max_group_nr)
2532                                 return -EFAULT;
2533
2534                         sz = data->read.group.nr * sample_read_value_size(read_format);
2535                         OVERFLOW_CHECK(array, sz, max_size);
2536                         data->read.group.values =
2537                                         (struct sample_read_value *)array;
2538                         array = (void *)array + sz;
2539                 } else {
2540                         OVERFLOW_CHECK_u64(array);
2541                         data->read.one.id = *array;
2542                         array++;
2543
2544                         if (read_format & PERF_FORMAT_LOST) {
2545                                 OVERFLOW_CHECK_u64(array);
2546                                 data->read.one.lost = *array;
2547                                 array++;
2548                         }
2549                 }
2550         }
2551
2552         if (type & PERF_SAMPLE_CALLCHAIN) {
2553                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2554
2555                 OVERFLOW_CHECK_u64(array);
2556                 data->callchain = (struct ip_callchain *)array++;
2557                 if (data->callchain->nr > max_callchain_nr)
2558                         return -EFAULT;
2559                 sz = data->callchain->nr * sizeof(u64);
2560                 OVERFLOW_CHECK(array, sz, max_size);
2561                 array = (void *)array + sz;
2562         }
2563
2564         if (type & PERF_SAMPLE_RAW) {
2565                 OVERFLOW_CHECK_u64(array);
2566                 u.val64 = *array;
2567
2568                 /*
2569                  * Undo swap of u64, then swap on individual u32s,
2570                  * get the size of the raw area and undo all of the
2571                  * swap. The pevent interface handles endianness by
2572                  * itself.
2573                  */
2574                 if (swapped) {
2575                         u.val64 = bswap_64(u.val64);
2576                         u.val32[0] = bswap_32(u.val32[0]);
2577                         u.val32[1] = bswap_32(u.val32[1]);
2578                 }
2579                 data->raw_size = u.val32[0];
2580
2581                 /*
2582                  * The raw data is aligned on 64bits including the
2583                  * u32 size, so it's safe to use mem_bswap_64.
2584                  */
2585                 if (swapped)
2586                         mem_bswap_64((void *) array, data->raw_size);
2587
2588                 array = (void *)array + sizeof(u32);
2589
2590                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2591                 data->raw_data = (void *)array;
2592                 array = (void *)array + data->raw_size;
2593         }
2594
2595         if (type & PERF_SAMPLE_BRANCH_STACK) {
2596                 const u64 max_branch_nr = UINT64_MAX /
2597                                           sizeof(struct branch_entry);
2598                 struct branch_entry *e;
2599                 unsigned int i;
2600
2601                 OVERFLOW_CHECK_u64(array);
2602                 data->branch_stack = (struct branch_stack *)array++;
2603
2604                 if (data->branch_stack->nr > max_branch_nr)
2605                         return -EFAULT;
2606
2607                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2608                 if (evsel__has_branch_hw_idx(evsel)) {
2609                         sz += sizeof(u64);
2610                         e = &data->branch_stack->entries[0];
2611                 } else {
2612                         data->no_hw_idx = true;
2613                         /*
2614                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2615                          * only nr and entries[] will be output by kernel.
2616                          */
2617                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2618                 }
2619
2620                 if (swapped) {
2621                         /*
2622                          * struct branch_flag does not have endian
2623                          * specific bit field definition. And bswap
2624                          * will not resolve the issue, since these
2625                          * are bit fields.
2626                          *
2627                          * evsel__bitfield_swap_branch_flags() uses a
2628                          * bitfield_swap macro to swap the bit position
2629                          * based on the host endians.
2630                          */
2631                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2632                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2633                 }
2634
2635                 OVERFLOW_CHECK(array, sz, max_size);
2636                 array = (void *)array + sz;
2637         }
2638
2639         if (type & PERF_SAMPLE_REGS_USER) {
2640                 OVERFLOW_CHECK_u64(array);
2641                 data->user_regs.abi = *array;
2642                 array++;
2643
2644                 if (data->user_regs.abi) {
2645                         u64 mask = evsel->core.attr.sample_regs_user;
2646
2647                         sz = hweight64(mask) * sizeof(u64);
2648                         OVERFLOW_CHECK(array, sz, max_size);
2649                         data->user_regs.mask = mask;
2650                         data->user_regs.regs = (u64 *)array;
2651                         array = (void *)array + sz;
2652                 }
2653         }
2654
2655         if (type & PERF_SAMPLE_STACK_USER) {
2656                 OVERFLOW_CHECK_u64(array);
2657                 sz = *array++;
2658
2659                 data->user_stack.offset = ((char *)(array - 1)
2660                                           - (char *) event);
2661
2662                 if (!sz) {
2663                         data->user_stack.size = 0;
2664                 } else {
2665                         OVERFLOW_CHECK(array, sz, max_size);
2666                         data->user_stack.data = (char *)array;
2667                         array = (void *)array + sz;
2668                         OVERFLOW_CHECK_u64(array);
2669                         data->user_stack.size = *array++;
2670                         if (WARN_ONCE(data->user_stack.size > sz,
2671                                       "user stack dump failure\n"))
2672                                 return -EFAULT;
2673                 }
2674         }
2675
2676         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2677                 OVERFLOW_CHECK_u64(array);
2678                 arch_perf_parse_sample_weight(data, array, type);
2679                 array++;
2680         }
2681
2682         if (type & PERF_SAMPLE_DATA_SRC) {
2683                 OVERFLOW_CHECK_u64(array);
2684                 data->data_src = *array;
2685                 array++;
2686         }
2687
2688         if (type & PERF_SAMPLE_TRANSACTION) {
2689                 OVERFLOW_CHECK_u64(array);
2690                 data->transaction = *array;
2691                 array++;
2692         }
2693
2694         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2695         if (type & PERF_SAMPLE_REGS_INTR) {
2696                 OVERFLOW_CHECK_u64(array);
2697                 data->intr_regs.abi = *array;
2698                 array++;
2699
2700                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2701                         u64 mask = evsel->core.attr.sample_regs_intr;
2702
2703                         sz = hweight64(mask) * sizeof(u64);
2704                         OVERFLOW_CHECK(array, sz, max_size);
2705                         data->intr_regs.mask = mask;
2706                         data->intr_regs.regs = (u64 *)array;
2707                         array = (void *)array + sz;
2708                 }
2709         }
2710
2711         data->phys_addr = 0;
2712         if (type & PERF_SAMPLE_PHYS_ADDR) {
2713                 data->phys_addr = *array;
2714                 array++;
2715         }
2716
2717         data->cgroup = 0;
2718         if (type & PERF_SAMPLE_CGROUP) {
2719                 data->cgroup = *array;
2720                 array++;
2721         }
2722
2723         data->data_page_size = 0;
2724         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2725                 data->data_page_size = *array;
2726                 array++;
2727         }
2728
2729         data->code_page_size = 0;
2730         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2731                 data->code_page_size = *array;
2732                 array++;
2733         }
2734
2735         if (type & PERF_SAMPLE_AUX) {
2736                 OVERFLOW_CHECK_u64(array);
2737                 sz = *array++;
2738
2739                 OVERFLOW_CHECK(array, sz, max_size);
2740                 /* Undo swap of data */
2741                 if (swapped)
2742                         mem_bswap_64((char *)array, sz);
2743                 data->aux_sample.size = sz;
2744                 data->aux_sample.data = (char *)array;
2745                 array = (void *)array + sz;
2746         }
2747
2748         return 0;
2749 }
2750
2751 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2752                                   u64 *timestamp)
2753 {
2754         u64 type = evsel->core.attr.sample_type;
2755         const __u64 *array;
2756
2757         if (!(type & PERF_SAMPLE_TIME))
2758                 return -1;
2759
2760         if (event->header.type != PERF_RECORD_SAMPLE) {
2761                 struct perf_sample data = {
2762                         .time = -1ULL,
2763                 };
2764
2765                 if (!evsel->core.attr.sample_id_all)
2766                         return -1;
2767                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2768                         return -1;
2769
2770                 *timestamp = data.time;
2771                 return 0;
2772         }
2773
2774         array = event->sample.array;
2775
2776         if (perf_event__check_size(event, evsel->sample_size))
2777                 return -EFAULT;
2778
2779         if (type & PERF_SAMPLE_IDENTIFIER)
2780                 array++;
2781
2782         if (type & PERF_SAMPLE_IP)
2783                 array++;
2784
2785         if (type & PERF_SAMPLE_TID)
2786                 array++;
2787
2788         if (type & PERF_SAMPLE_TIME)
2789                 *timestamp = *array;
2790
2791         return 0;
2792 }
2793
2794 u16 evsel__id_hdr_size(struct evsel *evsel)
2795 {
2796         u64 sample_type = evsel->core.attr.sample_type;
2797         u16 size = 0;
2798
2799         if (sample_type & PERF_SAMPLE_TID)
2800                 size += sizeof(u64);
2801
2802         if (sample_type & PERF_SAMPLE_TIME)
2803                 size += sizeof(u64);
2804
2805         if (sample_type & PERF_SAMPLE_ID)
2806                 size += sizeof(u64);
2807
2808         if (sample_type & PERF_SAMPLE_STREAM_ID)
2809                 size += sizeof(u64);
2810
2811         if (sample_type & PERF_SAMPLE_CPU)
2812                 size += sizeof(u64);
2813
2814         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2815                 size += sizeof(u64);
2816
2817         return size;
2818 }
2819
2820 #ifdef HAVE_LIBTRACEEVENT
2821 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2822 {
2823         return tep_find_field(evsel->tp_format, name);
2824 }
2825
2826 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2827 {
2828         struct tep_format_field *field = evsel__field(evsel, name);
2829         int offset;
2830
2831         if (!field)
2832                 return NULL;
2833
2834         offset = field->offset;
2835
2836         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2837                 offset = *(int *)(sample->raw_data + field->offset);
2838                 offset &= 0xffff;
2839                 if (tep_field_is_relative(field->flags))
2840                         offset += field->offset + field->size;
2841         }
2842
2843         return sample->raw_data + offset;
2844 }
2845
2846 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2847                          bool needs_swap)
2848 {
2849         u64 value;
2850         void *ptr = sample->raw_data + field->offset;
2851
2852         switch (field->size) {
2853         case 1:
2854                 return *(u8 *)ptr;
2855         case 2:
2856                 value = *(u16 *)ptr;
2857                 break;
2858         case 4:
2859                 value = *(u32 *)ptr;
2860                 break;
2861         case 8:
2862                 memcpy(&value, ptr, sizeof(u64));
2863                 break;
2864         default:
2865                 return 0;
2866         }
2867
2868         if (!needs_swap)
2869                 return value;
2870
2871         switch (field->size) {
2872         case 2:
2873                 return bswap_16(value);
2874         case 4:
2875                 return bswap_32(value);
2876         case 8:
2877                 return bswap_64(value);
2878         default:
2879                 return 0;
2880         }
2881
2882         return 0;
2883 }
2884
2885 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2886 {
2887         struct tep_format_field *field = evsel__field(evsel, name);
2888
2889         if (!field)
2890                 return 0;
2891
2892         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2893 }
2894 #endif
2895
2896 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2897 {
2898         int paranoid;
2899
2900         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2901             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2902             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2903                 /*
2904                  * If it's cycles then fall back to hrtimer based
2905                  * cpu-clock-tick sw counter, which is always available even if
2906                  * no PMU support.
2907                  *
2908                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2909                  * b0a873e).
2910                  */
2911                 scnprintf(msg, msgsize, "%s",
2912 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2913
2914                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2915                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2916
2917                 zfree(&evsel->name);
2918                 return true;
2919         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2920                    (paranoid = perf_event_paranoid()) > 1) {
2921                 const char *name = evsel__name(evsel);
2922                 char *new_name;
2923                 const char *sep = ":";
2924
2925                 /* If event has exclude user then don't exclude kernel. */
2926                 if (evsel->core.attr.exclude_user)
2927                         return false;
2928
2929                 /* Is there already the separator in the name. */
2930                 if (strchr(name, '/') ||
2931                     (strchr(name, ':') && !evsel->is_libpfm_event))
2932                         sep = "";
2933
2934                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2935                         return false;
2936
2937                 free(evsel->name);
2938                 evsel->name = new_name;
2939                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2940                           "to fall back to excluding kernel and hypervisor "
2941                           " samples", paranoid);
2942                 evsel->core.attr.exclude_kernel = 1;
2943                 evsel->core.attr.exclude_hv     = 1;
2944
2945                 return true;
2946         }
2947
2948         return false;
2949 }
2950
2951 static bool find_process(const char *name)
2952 {
2953         size_t len = strlen(name);
2954         DIR *dir;
2955         struct dirent *d;
2956         int ret = -1;
2957
2958         dir = opendir(procfs__mountpoint());
2959         if (!dir)
2960                 return false;
2961
2962         /* Walk through the directory. */
2963         while (ret && (d = readdir(dir)) != NULL) {
2964                 char path[PATH_MAX];
2965                 char *data;
2966                 size_t size;
2967
2968                 if ((d->d_type != DT_DIR) ||
2969                      !strcmp(".", d->d_name) ||
2970                      !strcmp("..", d->d_name))
2971                         continue;
2972
2973                 scnprintf(path, sizeof(path), "%s/%s/comm",
2974                           procfs__mountpoint(), d->d_name);
2975
2976                 if (filename__read_str(path, &data, &size))
2977                         continue;
2978
2979                 ret = strncmp(name, data, len);
2980                 free(data);
2981         }
2982
2983         closedir(dir);
2984         return ret ? false : true;
2985 }
2986
2987 static bool is_amd(const char *arch, const char *cpuid)
2988 {
2989         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2990 }
2991
2992 static bool is_amd_ibs(struct evsel *evsel)
2993 {
2994         return evsel->core.attr.precise_ip
2995             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2996 }
2997
2998 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2999                          int err, char *msg, size_t size)
3000 {
3001         struct perf_env *env = evsel__env(evsel);
3002         const char *arch = perf_env__arch(env);
3003         const char *cpuid = perf_env__cpuid(env);
3004         char sbuf[STRERR_BUFSIZE];
3005         int printed = 0, enforced = 0;
3006
3007         switch (err) {
3008         case EPERM:
3009         case EACCES:
3010                 printed += scnprintf(msg + printed, size - printed,
3011                         "Access to performance monitoring and observability operations is limited.\n");
3012
3013                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
3014                         if (enforced) {
3015                                 printed += scnprintf(msg + printed, size - printed,
3016                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
3017                                         "monitoring and observability operations. Inspect system audit records for\n"
3018                                         "more perf_event access control information and adjusting the policy.\n");
3019                         }
3020                 }
3021
3022                 if (err == EPERM)
3023                         printed += scnprintf(msg, size,
3024                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
3025
3026                 return scnprintf(msg + printed, size - printed,
3027                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3028                  "access to performance monitoring and observability operations for processes\n"
3029                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3030                  "More information can be found at 'Perf events and tool security' document:\n"
3031                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3032                  "perf_event_paranoid setting is %d:\n"
3033                  "  -1: Allow use of (almost) all events by all users\n"
3034                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3035                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
3036                  ">= 1: Disallow CPU event access\n"
3037                  ">= 2: Disallow kernel profiling\n"
3038                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3039                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3040                  perf_event_paranoid());
3041         case ENOENT:
3042                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3043         case EMFILE:
3044                 return scnprintf(msg, size, "%s",
3045                          "Too many events are opened.\n"
3046                          "Probably the maximum number of open file descriptors has been reached.\n"
3047                          "Hint: Try again after reducing the number of events.\n"
3048                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3049         case ENOMEM:
3050                 if (evsel__has_callchain(evsel) &&
3051                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3052                         return scnprintf(msg, size,
3053                                          "Not enough memory to setup event with callchain.\n"
3054                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3055                                          "Hint: Current value: %d", sysctl__max_stack());
3056                 break;
3057         case ENODEV:
3058                 if (target->cpu_list)
3059                         return scnprintf(msg, size, "%s",
3060          "No such device - did you specify an out-of-range profile CPU?");
3061                 break;
3062         case EOPNOTSUPP:
3063                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3064                         return scnprintf(msg, size,
3065         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3066                                          evsel__name(evsel));
3067                 if (evsel->core.attr.aux_output)
3068                         return scnprintf(msg, size,
3069         "%s: PMU Hardware doesn't support 'aux_output' feature",
3070                                          evsel__name(evsel));
3071                 if (evsel->core.attr.sample_period != 0)
3072                         return scnprintf(msg, size,
3073         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3074                                          evsel__name(evsel));
3075                 if (evsel->core.attr.precise_ip)
3076                         return scnprintf(msg, size, "%s",
3077         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3078 #if defined(__i386__) || defined(__x86_64__)
3079                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3080                         return scnprintf(msg, size, "%s",
3081         "No hardware sampling interrupt available.\n");
3082 #endif
3083                 break;
3084         case EBUSY:
3085                 if (find_process("oprofiled"))
3086                         return scnprintf(msg, size,
3087         "The PMU counters are busy/taken by another profiler.\n"
3088         "We found oprofile daemon running, please stop it and try again.");
3089                 break;
3090         case EINVAL:
3091                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3092                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3093                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3094                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3095                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3096                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3097                 if (perf_missing_features.clockid)
3098                         return scnprintf(msg, size, "clockid feature not supported.");
3099                 if (perf_missing_features.clockid_wrong)
3100                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3101                 if (perf_missing_features.aux_output)
3102                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3103                 if (!target__has_cpu(target))
3104                         return scnprintf(msg, size,
3105         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3106                                         evsel__name(evsel));
3107                 if (is_amd(arch, cpuid)) {
3108                         if (is_amd_ibs(evsel)) {
3109                                 if (evsel->core.attr.exclude_kernel)
3110                                         return scnprintf(msg, size,
3111         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3112                                 if (!evsel->core.system_wide)
3113                                         return scnprintf(msg, size,
3114         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3115                         }
3116                 }
3117
3118                 break;
3119         case ENODATA:
3120                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3121                                  "Please add an auxiliary event in front of the load latency event.");
3122         default:
3123                 break;
3124         }
3125
3126         return scnprintf(msg, size,
3127         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3128         "/bin/dmesg | grep -i perf may provide additional information.\n",
3129                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3130 }
3131
3132 struct perf_env *evsel__env(struct evsel *evsel)
3133 {
3134         if (evsel && evsel->evlist && evsel->evlist->env)
3135                 return evsel->evlist->env;
3136         return &perf_env;
3137 }
3138
3139 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3140 {
3141         int cpu_map_idx, thread;
3142
3143         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3144                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3145                      thread++) {
3146                         int fd = FD(evsel, cpu_map_idx, thread);
3147
3148                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3149                                                    cpu_map_idx, thread, fd) < 0)
3150                                 return -1;
3151                 }
3152         }
3153
3154         return 0;
3155 }
3156
3157 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3158 {
3159         struct perf_cpu_map *cpus = evsel->core.cpus;
3160         struct perf_thread_map *threads = evsel->core.threads;
3161
3162         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3163                 return -ENOMEM;
3164
3165         return store_evsel_ids(evsel, evlist);
3166 }
3167
3168 void evsel__zero_per_pkg(struct evsel *evsel)
3169 {
3170         struct hashmap_entry *cur;
3171         size_t bkt;
3172
3173         if (evsel->per_pkg_mask) {
3174                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3175                         zfree(&cur->pkey);
3176
3177                 hashmap__clear(evsel->per_pkg_mask);
3178         }
3179 }
3180
3181 bool evsel__is_hybrid(const struct evsel *evsel)
3182 {
3183         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3184 }
3185
3186 struct evsel *evsel__leader(const struct evsel *evsel)
3187 {
3188         return container_of(evsel->core.leader, struct evsel, core);
3189 }
3190
3191 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3192 {
3193         return evsel->core.leader == &leader->core;
3194 }
3195
3196 bool evsel__is_leader(struct evsel *evsel)
3197 {
3198         return evsel__has_leader(evsel, evsel);
3199 }
3200
3201 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3202 {
3203         evsel->core.leader = &leader->core;
3204 }
3205
3206 int evsel__source_count(const struct evsel *evsel)
3207 {
3208         struct evsel *pos;
3209         int count = 0;
3210
3211         evlist__for_each_entry(evsel->evlist, pos) {
3212                 if (pos->metric_leader == evsel)
3213                         count++;
3214         }
3215         return count;
3216 }
3217
3218 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3219 {
3220         return false;
3221 }
3222
3223 /*
3224  * Remove an event from a given group (leader).
3225  * Some events, e.g., perf metrics Topdown events,
3226  * must always be grouped. Ignore the events.
3227  */
3228 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3229 {
3230         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3231                 evsel__set_leader(evsel, evsel);
3232                 evsel->core.nr_members = 0;
3233                 leader->core.nr_members--;
3234         }
3235 }