tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <traceevent/event-parse.h>
  16 #include <linux/hw_breakpoint.h>
  17 #include <linux/perf_event.h>
  18 #include <linux/compiler.h>
  19 #include <linux/err.h>
  20 #include <linux/zalloc.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include <stdlib.h>
  26 #include <perf/evsel.h>
  27 #include "asm/bug.h"
  28 #include "bpf_counter.h"
  29 #include "callchain.h"
  30 #include "cgroup.h"
  31 #include "counts.h"
  32 #include "event.h"
  33 #include "evsel.h"
  34 #include "util/env.h"
  35 #include "util/evsel_config.h"
  36 #include "util/evsel_fprintf.h"
  37 #include "evlist.h"
  38 #include <perf/cpumap.h>
  39 #include "thread_map.h"
  40 #include "target.h"
  41 #include "perf_regs.h"
  42 #include "record.h"
  43 #include "debug.h"
  44 #include "trace-event.h"
  45 #include "stat.h"
  46 #include "string2.h"
  47 #include "memswap.h"
  48 #include "util.h"
  49 #include "hashmap.h"
  50 #include "pmu-hybrid.h"
  51 #include "off_cpu.h"
  52 #include "../perf-sys.h"
  53 #include "util/parse-branch-options.h"
  54 #include <internal/xyarray.h>
  55 #include <internal/lib.h>
  56
  57 #include <linux/ctype.h>
  58
  59 struct perf_missing_features perf_missing_features;
  60
  61 static clockid_t clockid;
  62
  63 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  64         NULL,
  65         "duration_time",
  66         "user_time",
  67         "system_time",
  68 };
  69
  70 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  71 {
  72         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  73                 return perf_tool_event__tool_names[ev];
  74
  75         return NULL;
  76 }
  77
  78 enum perf_tool_event perf_tool_event__from_str(const char *str)
  79 {
  80         int i;
  81
  82         perf_tool_event__for_each_event(i) {
  83                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  84                         return i;
  85         }
  86         return PERF_TOOL_NONE;
  87 }
  88
  89
  90 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  91 {
  92         return 0;
  93 }
  94
  95 void __weak test_attr__ready(void) { }
  96
  97 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
  98 {
  99 }
 100
 101 static struct {
 102         size_t  size;
 103         int     (*init)(struct evsel *evsel);
 104         void    (*fini)(struct evsel *evsel);
 105 } perf_evsel__object = {
 106         .size = sizeof(struct evsel),
 107         .init = evsel__no_extra_init,
 108         .fini = evsel__no_extra_fini,
 109 };
 110
 111 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 112                          void (*fini)(struct evsel *evsel))
 113 {
 114
 115         if (object_size == 0)
 116                 goto set_methods;
 117
 118         if (perf_evsel__object.size > object_size)
 119                 return -EINVAL;
 120
 121         perf_evsel__object.size = object_size;
 122
 123 set_methods:
 124         if (init != NULL)
 125                 perf_evsel__object.init = init;
 126
 127         if (fini != NULL)
 128                 perf_evsel__object.fini = fini;
 129
 130         return 0;
 131 }
 132
 133 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 134
 135 int __evsel__sample_size(u64 sample_type)
 136 {
 137         u64 mask = sample_type & PERF_SAMPLE_MASK;
 138         int size = 0;
 139         int i;
 140
 141         for (i = 0; i < 64; i++) {
 142                 if (mask & (1ULL << i))
 143                         size++;
 144         }
 145
 146         size *= sizeof(u64);
 147
 148         return size;
 149 }
 150
 151 /**
 152  * __perf_evsel__calc_id_pos - calculate id_pos.
 153  * @sample_type: sample type
 154  *
 155  * This function returns the position of the event id (PERF_SAMPLE_ID or
 156  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 157  * perf_record_sample.
 158  */
 159 static int __perf_evsel__calc_id_pos(u64 sample_type)
 160 {
 161         int idx = 0;
 162
 163         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 164                 return 0;
 165
 166         if (!(sample_type & PERF_SAMPLE_ID))
 167                 return -1;
 168
 169         if (sample_type & PERF_SAMPLE_IP)
 170                 idx += 1;
 171
 172         if (sample_type & PERF_SAMPLE_TID)
 173                 idx += 1;
 174
 175         if (sample_type & PERF_SAMPLE_TIME)
 176                 idx += 1;
 177
 178         if (sample_type & PERF_SAMPLE_ADDR)
 179                 idx += 1;
 180
 181         return idx;
 182 }
 183
 184 /**
 185  * __perf_evsel__calc_is_pos - calculate is_pos.
 186  * @sample_type: sample type
 187  *
 188  * This function returns the position (counting backwards) of the event id
 189  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 190  * sample_id_all is used there is an id sample appended to non-sample events.
 191  */
 192 static int __perf_evsel__calc_is_pos(u64 sample_type)
 193 {
 194         int idx = 1;
 195
 196         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 197                 return 1;
 198
 199         if (!(sample_type & PERF_SAMPLE_ID))
 200                 return -1;
 201
 202         if (sample_type & PERF_SAMPLE_CPU)
 203                 idx += 1;
 204
 205         if (sample_type & PERF_SAMPLE_STREAM_ID)
 206                 idx += 1;
 207
 208         return idx;
 209 }
 210
 211 void evsel__calc_id_pos(struct evsel *evsel)
 212 {
 213         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 214         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 215 }
 216
 217 void __evsel__set_sample_bit(struct evsel *evsel,
 218                                   enum perf_event_sample_format bit)
 219 {
 220         if (!(evsel->core.attr.sample_type & bit)) {
 221                 evsel->core.attr.sample_type |= bit;
 222                 evsel->sample_size += sizeof(u64);
 223                 evsel__calc_id_pos(evsel);
 224         }
 225 }
 226
 227 void __evsel__reset_sample_bit(struct evsel *evsel,
 228                                     enum perf_event_sample_format bit)
 229 {
 230         if (evsel->core.attr.sample_type & bit) {
 231                 evsel->core.attr.sample_type &= ~bit;
 232                 evsel->sample_size -= sizeof(u64);
 233                 evsel__calc_id_pos(evsel);
 234         }
 235 }
 236
 237 void evsel__set_sample_id(struct evsel *evsel,
 238                                bool can_sample_identifier)
 239 {
 240         if (can_sample_identifier) {
 241                 evsel__reset_sample_bit(evsel, ID);
 242                 evsel__set_sample_bit(evsel, IDENTIFIER);
 243         } else {
 244                 evsel__set_sample_bit(evsel, ID);
 245         }
 246         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 247 }
 248
 249 /**
 250  * evsel__is_function_event - Return whether given evsel is a function
 251  * trace event
 252  *
 253  * @evsel - evsel selector to be tested
 254  *
 255  * Return %true if event is function trace event
 256  */
 257 bool evsel__is_function_event(struct evsel *evsel)
 258 {
 259 #define FUNCTION_EVENT "ftrace:function"
 260
 261         return evsel->name &&
 262                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 263
 264 #undef FUNCTION_EVENT
 265 }
 266
 267 void evsel__init(struct evsel *evsel,
 268                  struct perf_event_attr *attr, int idx)
 269 {
 270         perf_evsel__init(&evsel->core, attr, idx);
 271         evsel->tracking    = !idx;
 272         evsel->unit        = strdup("");
 273         evsel->scale       = 1.0;
 274         evsel->max_events  = ULONG_MAX;
 275         evsel->evlist      = NULL;
 276         evsel->bpf_obj     = NULL;
 277         evsel->bpf_fd      = -1;
 278         INIT_LIST_HEAD(&evsel->config_terms);
 279         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 280         perf_evsel__object.init(evsel);
 281         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 282         evsel__calc_id_pos(evsel);
 283         evsel->cmdline_group_boundary = false;
 284         evsel->metric_expr   = NULL;
 285         evsel->metric_name   = NULL;
 286         evsel->metric_events = NULL;
 287         evsel->per_pkg_mask  = NULL;
 288         evsel->collect_stat  = false;
 289         evsel->pmu_name      = NULL;
 290 }
 291
 292 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 293 {
 294         struct evsel *evsel = zalloc(perf_evsel__object.size);
 295
 296         if (!evsel)
 297                 return NULL;
 298         evsel__init(evsel, attr, idx);
 299
 300         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 301                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 302                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 303                 evsel->core.attr.sample_period = 1;
 304         }
 305
 306         if (evsel__is_clock(evsel)) {
 307                 free((char *)evsel->unit);
 308                 evsel->unit = strdup("msec");
 309                 evsel->scale = 1e-6;
 310         }
 311
 312         return evsel;
 313 }
 314
 315 static bool perf_event_can_profile_kernel(void)
 316 {
 317         return perf_event_paranoid_check(1);
 318 }
 319
 320 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 321 {
 322         struct perf_event_attr attr = {
 323                 .type   = type,
 324                 .config = config,
 325                 .exclude_kernel = !perf_event_can_profile_kernel(),
 326         };
 327         struct evsel *evsel;
 328
 329         event_attr_init(&attr);
 330
 331         /*
 332          * Now let the usual logic to set up the perf_event_attr defaults
 333          * to kick in when we return and before perf_evsel__open() is called.
 334          */
 335         evsel = evsel__new(&attr);
 336         if (evsel == NULL)
 337                 goto out;
 338
 339         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 340
 341         evsel->precise_max = true;
 342
 343         /* use asprintf() because free(evsel) assumes name is allocated */
 344         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 345                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 346                      attr.exclude_kernel ? "u" : "",
 347                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 348                 goto error_free;
 349 out:
 350         return evsel;
 351 error_free:
 352         evsel__delete(evsel);
 353         evsel = NULL;
 354         goto out;
 355 }
 356
 357 int copy_config_terms(struct list_head *dst, struct list_head *src)
 358 {
 359         struct evsel_config_term *pos, *tmp;
 360
 361         list_for_each_entry(pos, src, list) {
 362                 tmp = malloc(sizeof(*tmp));
 363                 if (tmp == NULL)
 364                         return -ENOMEM;
 365
 366                 *tmp = *pos;
 367                 if (tmp->free_str) {
 368                         tmp->val.str = strdup(pos->val.str);
 369                         if (tmp->val.str == NULL) {
 370                                 free(tmp);
 371                                 return -ENOMEM;
 372                         }
 373                 }
 374                 list_add_tail(&tmp->list, dst);
 375         }
 376         return 0;
 377 }
 378
 379 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 380 {
 381         return copy_config_terms(&dst->config_terms, &src->config_terms);
 382 }
 383
 384 /**
 385  * evsel__clone - create a new evsel copied from @orig
 386  * @orig: original evsel
 387  *
 388  * The assumption is that @orig is not configured nor opened yet.
 389  * So we only care about the attributes that can be set while it's parsed.
 390  */
 391 struct evsel *evsel__clone(struct evsel *orig)
 392 {
 393         struct evsel *evsel;
 394
 395         BUG_ON(orig->core.fd);
 396         BUG_ON(orig->counts);
 397         BUG_ON(orig->priv);
 398         BUG_ON(orig->per_pkg_mask);
 399
 400         /* cannot handle BPF objects for now */
 401         if (orig->bpf_obj)
 402                 return NULL;
 403
 404         evsel = evsel__new(&orig->core.attr);
 405         if (evsel == NULL)
 406                 return NULL;
 407
 408         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 409         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 410         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 411         evsel->core.nr_members = orig->core.nr_members;
 412         evsel->core.system_wide = orig->core.system_wide;
 413         evsel->core.requires_cpu = orig->core.requires_cpu;
 414
 415         if (orig->name) {
 416                 evsel->name = strdup(orig->name);
 417                 if (evsel->name == NULL)
 418                         goto out_err;
 419         }
 420         if (orig->group_name) {
 421                 evsel->group_name = strdup(orig->group_name);
 422                 if (evsel->group_name == NULL)
 423                         goto out_err;
 424         }
 425         if (orig->pmu_name) {
 426                 evsel->pmu_name = strdup(orig->pmu_name);
 427                 if (evsel->pmu_name == NULL)
 428                         goto out_err;
 429         }
 430         if (orig->filter) {
 431                 evsel->filter = strdup(orig->filter);
 432                 if (evsel->filter == NULL)
 433                         goto out_err;
 434         }
 435         if (orig->metric_id) {
 436                 evsel->metric_id = strdup(orig->metric_id);
 437                 if (evsel->metric_id == NULL)
 438                         goto out_err;
 439         }
 440         evsel->cgrp = cgroup__get(orig->cgrp);
 441         evsel->tp_format = orig->tp_format;
 442         evsel->handler = orig->handler;
 443         evsel->core.leader = orig->core.leader;
 444
 445         evsel->max_events = orig->max_events;
 446         evsel->tool_event = orig->tool_event;
 447         free((char *)evsel->unit);
 448         evsel->unit = strdup(orig->unit);
 449         if (evsel->unit == NULL)
 450                 goto out_err;
 451
 452         evsel->scale = orig->scale;
 453         evsel->snapshot = orig->snapshot;
 454         evsel->per_pkg = orig->per_pkg;
 455         evsel->percore = orig->percore;
 456         evsel->precise_max = orig->precise_max;
 457         evsel->use_uncore_alias = orig->use_uncore_alias;
 458         evsel->is_libpfm_event = orig->is_libpfm_event;
 459
 460         evsel->exclude_GH = orig->exclude_GH;
 461         evsel->sample_read = orig->sample_read;
 462         evsel->auto_merge_stats = orig->auto_merge_stats;
 463         evsel->collect_stat = orig->collect_stat;
 464         evsel->weak_group = orig->weak_group;
 465         evsel->use_config_name = orig->use_config_name;
 466
 467         if (evsel__copy_config_terms(evsel, orig) < 0)
 468                 goto out_err;
 469
 470         return evsel;
 471
 472 out_err:
 473         evsel__delete(evsel);
 474         return NULL;
 475 }
 476
 477 /*
 478  * Returns pointer with encoded error via <linux/err.h> interface.
 479  */
 480 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 481 {
 482         struct evsel *evsel = zalloc(perf_evsel__object.size);
 483         int err = -ENOMEM;
 484
 485         if (evsel == NULL) {
 486                 goto out_err;
 487         } else {
 488                 struct perf_event_attr attr = {
 489                         .type          = PERF_TYPE_TRACEPOINT,
 490                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 491                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 492                 };
 493
 494                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 495                         goto out_free;
 496
 497                 evsel->tp_format = trace_event__tp_format(sys, name);
 498                 if (IS_ERR(evsel->tp_format)) {
 499                         err = PTR_ERR(evsel->tp_format);
 500                         goto out_free;
 501                 }
 502
 503                 event_attr_init(&attr);
 504                 attr.config = evsel->tp_format->id;
 505                 attr.sample_period = 1;
 506                 evsel__init(evsel, &attr, idx);
 507         }
 508
 509         return evsel;
 510
 511 out_free:
 512         zfree(&evsel->name);
 513         free(evsel);
 514 out_err:
 515         return ERR_PTR(err);
 516 }
 517
 518 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 519         "cycles",
 520         "instructions",
 521         "cache-references",
 522         "cache-misses",
 523         "branches",
 524         "branch-misses",
 525         "bus-cycles",
 526         "stalled-cycles-frontend",
 527         "stalled-cycles-backend",
 528         "ref-cycles",
 529 };
 530
 531 char *evsel__bpf_counter_events;
 532
 533 bool evsel__match_bpf_counter_events(const char *name)
 534 {
 535         int name_len;
 536         bool match;
 537         char *ptr;
 538
 539         if (!evsel__bpf_counter_events)
 540                 return false;
 541
 542         ptr = strstr(evsel__bpf_counter_events, name);
 543         name_len = strlen(name);
 544
 545         /* check name matches a full token in evsel__bpf_counter_events */
 546         match = (ptr != NULL) &&
 547                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 548                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 549
 550         return match;
 551 }
 552
 553 static const char *__evsel__hw_name(u64 config)
 554 {
 555         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 556                 return evsel__hw_names[config];
 557
 558         return "unknown-hardware";
 559 }
 560
 561 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 562 {
 563         int colon = 0, r = 0;
 564         struct perf_event_attr *attr = &evsel->core.attr;
 565         bool exclude_guest_default = false;
 566
 567 #define MOD_PRINT(context, mod) do {                                    \
 568                 if (!attr->exclude_##context) {                         \
 569                         if (!colon) colon = ++r;                        \
 570                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 571                 } } while(0)
 572
 573         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 574                 MOD_PRINT(kernel, 'k');
 575                 MOD_PRINT(user, 'u');
 576                 MOD_PRINT(hv, 'h');
 577                 exclude_guest_default = true;
 578         }
 579
 580         if (attr->precise_ip) {
 581                 if (!colon)
 582                         colon = ++r;
 583                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 584                 exclude_guest_default = true;
 585         }
 586
 587         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 588                 MOD_PRINT(host, 'H');
 589                 MOD_PRINT(guest, 'G');
 590         }
 591 #undef MOD_PRINT
 592         if (colon)
 593                 bf[colon - 1] = ':';
 594         return r;
 595 }
 596
 597 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 598 {
 599         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 600 }
 601
 602 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 603 {
 604         int r = arch_evsel__hw_name(evsel, bf, size);
 605         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 606 }
 607
 608 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 609         "cpu-clock",
 610         "task-clock",
 611         "page-faults",
 612         "context-switches",
 613         "cpu-migrations",
 614         "minor-faults",
 615         "major-faults",
 616         "alignment-faults",
 617         "emulation-faults",
 618         "dummy",
 619 };
 620
 621 static const char *__evsel__sw_name(u64 config)
 622 {
 623         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 624                 return evsel__sw_names[config];
 625         return "unknown-software";
 626 }
 627
 628 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 629 {
 630         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 631         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 632 }
 633
 634 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 635 {
 636         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 637 }
 638
 639 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 640 {
 641         int r;
 642
 643         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 644
 645         if (type & HW_BREAKPOINT_R)
 646                 r += scnprintf(bf + r, size - r, "r");
 647
 648         if (type & HW_BREAKPOINT_W)
 649                 r += scnprintf(bf + r, size - r, "w");
 650
 651         if (type & HW_BREAKPOINT_X)
 652                 r += scnprintf(bf + r, size - r, "x");
 653
 654         return r;
 655 }
 656
 657 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 658 {
 659         struct perf_event_attr *attr = &evsel->core.attr;
 660         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 661         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 662 }
 663
 664 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 665  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 666  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 667  { "LLC",       "L2",                                                   },
 668  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 669  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 670  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 671  { "node",                                                              },
 672 };
 673
 674 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 675  { "load",      "loads",        "read",                                 },
 676  { "store",     "stores",       "write",                                },
 677  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 678 };
 679
 680 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 681  { "refs",      "Reference",    "ops",          "access",               },
 682  { "misses",    "miss",                                                 },
 683 };
 684
 685 #define C(x)            PERF_COUNT_HW_CACHE_##x
 686 #define CACHE_READ      (1 << C(OP_READ))
 687 #define CACHE_WRITE     (1 << C(OP_WRITE))
 688 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 689 #define COP(x)          (1 << x)
 690
 691 /*
 692  * cache operation stat
 693  * L1I : Read and prefetch only
 694  * ITLB and BPU : Read-only
 695  */
 696 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 697  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 698  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 699  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 700  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 701  [C(ITLB)]      = (CACHE_READ),
 702  [C(BPU)]       = (CACHE_READ),
 703  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 704 };
 705
 706 bool evsel__is_cache_op_valid(u8 type, u8 op)
 707 {
 708         if (evsel__hw_cache_stat[type] & COP(op))
 709                 return true;    /* valid */
 710         else
 711                 return false;   /* invalid */
 712 }
 713
 714 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 715 {
 716         if (result) {
 717                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 718                                  evsel__hw_cache_op[op][0],
 719                                  evsel__hw_cache_result[result][0]);
 720         }
 721
 722         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 723                          evsel__hw_cache_op[op][1]);
 724 }
 725
 726 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 727 {
 728         u8 op, result, type = (config >>  0) & 0xff;
 729         const char *err = "unknown-ext-hardware-cache-type";
 730
 731         if (type >= PERF_COUNT_HW_CACHE_MAX)
 732                 goto out_err;
 733
 734         op = (config >>  8) & 0xff;
 735         err = "unknown-ext-hardware-cache-op";
 736         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 737                 goto out_err;
 738
 739         result = (config >> 16) & 0xff;
 740         err = "unknown-ext-hardware-cache-result";
 741         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 742                 goto out_err;
 743
 744         err = "invalid-cache";
 745         if (!evsel__is_cache_op_valid(type, op))
 746                 goto out_err;
 747
 748         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 749 out_err:
 750         return scnprintf(bf, size, "%s", err);
 751 }
 752
 753 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 754 {
 755         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 756         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 757 }
 758
 759 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 760 {
 761         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 762         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 763 }
 764
 765 const char *evsel__name(struct evsel *evsel)
 766 {
 767         char bf[128];
 768
 769         if (!evsel)
 770                 goto out_unknown;
 771
 772         if (evsel->name)
 773                 return evsel->name;
 774
 775         switch (evsel->core.attr.type) {
 776         case PERF_TYPE_RAW:
 777                 evsel__raw_name(evsel, bf, sizeof(bf));
 778                 break;
 779
 780         case PERF_TYPE_HARDWARE:
 781                 evsel__hw_name(evsel, bf, sizeof(bf));
 782                 break;
 783
 784         case PERF_TYPE_HW_CACHE:
 785                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 786                 break;
 787
 788         case PERF_TYPE_SOFTWARE:
 789                 if (evsel__is_tool(evsel))
 790                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 791                 else
 792                         evsel__sw_name(evsel, bf, sizeof(bf));
 793                 break;
 794
 795         case PERF_TYPE_TRACEPOINT:
 796                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 797                 break;
 798
 799         case PERF_TYPE_BREAKPOINT:
 800                 evsel__bp_name(evsel, bf, sizeof(bf));
 801                 break;
 802
 803         default:
 804                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 805                           evsel->core.attr.type);
 806                 break;
 807         }
 808
 809         evsel->name = strdup(bf);
 810
 811         if (evsel->name)
 812                 return evsel->name;
 813 out_unknown:
 814         return "unknown";
 815 }
 816
 817 const char *evsel__metric_id(const struct evsel *evsel)
 818 {
 819         if (evsel->metric_id)
 820                 return evsel->metric_id;
 821
 822         if (evsel__is_tool(evsel))
 823                 return perf_tool_event__to_str(evsel->tool_event);
 824
 825         return "unknown";
 826 }
 827
 828 const char *evsel__group_name(struct evsel *evsel)
 829 {
 830         return evsel->group_name ?: "anon group";
 831 }
 832
 833 /*
 834  * Returns the group details for the specified leader,
 835  * with following rules.
 836  *
 837  *  For record -e '{cycles,instructions}'
 838  *    'anon group { cycles:u, instructions:u }'
 839  *
 840  *  For record -e 'cycles,instructions' and report --group
 841  *    'cycles:u, instructions:u'
 842  */
 843 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 844 {
 845         int ret = 0;
 846         struct evsel *pos;
 847         const char *group_name = evsel__group_name(evsel);
 848
 849         if (!evsel->forced_leader)
 850                 ret = scnprintf(buf, size, "%s { ", group_name);
 851
 852         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 853
 854         for_each_group_member(pos, evsel)
 855                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 856
 857         if (!evsel->forced_leader)
 858                 ret += scnprintf(buf + ret, size - ret, " }");
 859
 860         return ret;
 861 }
 862
 863 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 864                                       struct callchain_param *param)
 865 {
 866         bool function = evsel__is_function_event(evsel);
 867         struct perf_event_attr *attr = &evsel->core.attr;
 868
 869         evsel__set_sample_bit(evsel, CALLCHAIN);
 870
 871         attr->sample_max_stack = param->max_stack;
 872
 873         if (opts->kernel_callchains)
 874                 attr->exclude_callchain_user = 1;
 875         if (opts->user_callchains)
 876                 attr->exclude_callchain_kernel = 1;
 877         if (param->record_mode == CALLCHAIN_LBR) {
 878                 if (!opts->branch_stack) {
 879                         if (attr->exclude_user) {
 880                                 pr_warning("LBR callstack option is only available "
 881                                            "to get user callchain information. "
 882                                            "Falling back to framepointers.\n");
 883                         } else {
 884                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 885                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 886                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 887                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 888                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 889                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 890                         }
 891                 } else
 892                          pr_warning("Cannot use LBR callstack with branch stack. "
 893                                     "Falling back to framepointers.\n");
 894         }
 895
 896         if (param->record_mode == CALLCHAIN_DWARF) {
 897                 if (!function) {
 898                         evsel__set_sample_bit(evsel, REGS_USER);
 899                         evsel__set_sample_bit(evsel, STACK_USER);
 900                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 901                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 902                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 903                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 904                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 905                         } else {
 906                                 attr->sample_regs_user |= arch__user_reg_mask();
 907                         }
 908                         attr->sample_stack_user = param->dump_size;
 909                         attr->exclude_callchain_user = 1;
 910                 } else {
 911                         pr_info("Cannot use DWARF unwind for function trace event,"
 912                                 " falling back to framepointers.\n");
 913                 }
 914         }
 915
 916         if (function) {
 917                 pr_info("Disabling user space callchains for function trace event.\n");
 918                 attr->exclude_callchain_user = 1;
 919         }
 920 }
 921
 922 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 923                              struct callchain_param *param)
 924 {
 925         if (param->enabled)
 926                 return __evsel__config_callchain(evsel, opts, param);
 927 }
 928
 929 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 930 {
 931         struct perf_event_attr *attr = &evsel->core.attr;
 932
 933         evsel__reset_sample_bit(evsel, CALLCHAIN);
 934         if (param->record_mode == CALLCHAIN_LBR) {
 935                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 936                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 937                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 938                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 939         }
 940         if (param->record_mode == CALLCHAIN_DWARF) {
 941                 evsel__reset_sample_bit(evsel, REGS_USER);
 942                 evsel__reset_sample_bit(evsel, STACK_USER);
 943         }
 944 }
 945
 946 static void evsel__apply_config_terms(struct evsel *evsel,
 947                                       struct record_opts *opts, bool track)
 948 {
 949         struct evsel_config_term *term;
 950         struct list_head *config_terms = &evsel->config_terms;
 951         struct perf_event_attr *attr = &evsel->core.attr;
 952         /* callgraph default */
 953         struct callchain_param param = {
 954                 .record_mode = callchain_param.record_mode,
 955         };
 956         u32 dump_size = 0;
 957         int max_stack = 0;
 958         const char *callgraph_buf = NULL;
 959
 960         list_for_each_entry(term, config_terms, list) {
 961                 switch (term->type) {
 962                 case EVSEL__CONFIG_TERM_PERIOD:
 963                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 964                                 attr->sample_period = term->val.period;
 965                                 attr->freq = 0;
 966                                 evsel__reset_sample_bit(evsel, PERIOD);
 967                         }
 968                         break;
 969                 case EVSEL__CONFIG_TERM_FREQ:
 970                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 971                                 attr->sample_freq = term->val.freq;
 972                                 attr->freq = 1;
 973                                 evsel__set_sample_bit(evsel, PERIOD);
 974                         }
 975                         break;
 976                 case EVSEL__CONFIG_TERM_TIME:
 977                         if (term->val.time)
 978                                 evsel__set_sample_bit(evsel, TIME);
 979                         else
 980                                 evsel__reset_sample_bit(evsel, TIME);
 981                         break;
 982                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 983                         callgraph_buf = term->val.str;
 984                         break;
 985                 case EVSEL__CONFIG_TERM_BRANCH:
 986                         if (term->val.str && strcmp(term->val.str, "no")) {
 987                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 988                                 parse_branch_str(term->val.str,
 989                                                  &attr->branch_sample_type);
 990                         } else
 991                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 992                         break;
 993                 case EVSEL__CONFIG_TERM_STACK_USER:
 994                         dump_size = term->val.stack_user;
 995                         break;
 996                 case EVSEL__CONFIG_TERM_MAX_STACK:
 997                         max_stack = term->val.max_stack;
 998                         break;
 999                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1000                         evsel->max_events = term->val.max_events;
1001                         break;
1002                 case EVSEL__CONFIG_TERM_INHERIT:
1003                         /*
1004                          * attr->inherit should has already been set by
1005                          * evsel__config. If user explicitly set
1006                          * inherit using config terms, override global
1007                          * opt->no_inherit setting.
1008                          */
1009                         attr->inherit = term->val.inherit ? 1 : 0;
1010                         break;
1011                 case EVSEL__CONFIG_TERM_OVERWRITE:
1012                         attr->write_backward = term->val.overwrite ? 1 : 0;
1013                         break;
1014                 case EVSEL__CONFIG_TERM_DRV_CFG:
1015                         break;
1016                 case EVSEL__CONFIG_TERM_PERCORE:
1017                         break;
1018                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1019                         attr->aux_output = term->val.aux_output ? 1 : 0;
1020                         break;
1021                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1022                         /* Already applied by auxtrace */
1023                         break;
1024                 case EVSEL__CONFIG_TERM_CFG_CHG:
1025                         break;
1026                 default:
1027                         break;
1028                 }
1029         }
1030
1031         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1032         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1033                 bool sample_address = false;
1034
1035                 if (max_stack) {
1036                         param.max_stack = max_stack;
1037                         if (callgraph_buf == NULL)
1038                                 callgraph_buf = "fp";
1039                 }
1040
1041                 /* parse callgraph parameters */
1042                 if (callgraph_buf != NULL) {
1043                         if (!strcmp(callgraph_buf, "no")) {
1044                                 param.enabled = false;
1045                                 param.record_mode = CALLCHAIN_NONE;
1046                         } else {
1047                                 param.enabled = true;
1048                                 if (parse_callchain_record(callgraph_buf, &param)) {
1049                                         pr_err("per-event callgraph setting for %s failed. "
1050                                                "Apply callgraph global setting for it\n",
1051                                                evsel->name);
1052                                         return;
1053                                 }
1054                                 if (param.record_mode == CALLCHAIN_DWARF)
1055                                         sample_address = true;
1056                         }
1057                 }
1058                 if (dump_size > 0) {
1059                         dump_size = round_up(dump_size, sizeof(u64));
1060                         param.dump_size = dump_size;
1061                 }
1062
1063                 /* If global callgraph set, clear it */
1064                 if (callchain_param.enabled)
1065                         evsel__reset_callgraph(evsel, &callchain_param);
1066
1067                 /* set perf-event callgraph */
1068                 if (param.enabled) {
1069                         if (sample_address) {
1070                                 evsel__set_sample_bit(evsel, ADDR);
1071                                 evsel__set_sample_bit(evsel, DATA_SRC);
1072                                 evsel->core.attr.mmap_data = track;
1073                         }
1074                         evsel__config_callchain(evsel, opts, &param);
1075                 }
1076         }
1077 }
1078
1079 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1080 {
1081         struct evsel_config_term *term, *found_term = NULL;
1082
1083         list_for_each_entry(term, &evsel->config_terms, list) {
1084                 if (term->type == type)
1085                         found_term = term;
1086         }
1087
1088         return found_term;
1089 }
1090
1091 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1092 {
1093         evsel__set_sample_bit(evsel, WEIGHT);
1094 }
1095
1096 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1097 {
1098 }
1099
1100 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1101                                     struct perf_event_attr *attr __maybe_unused)
1102 {
1103 }
1104
1105 static void evsel__set_default_freq_period(struct record_opts *opts,
1106                                            struct perf_event_attr *attr)
1107 {
1108         if (opts->freq) {
1109                 attr->freq = 1;
1110                 attr->sample_freq = opts->freq;
1111         } else {
1112                 attr->sample_period = opts->default_interval;
1113         }
1114 }
1115
1116 static bool evsel__is_offcpu_event(struct evsel *evsel)
1117 {
1118         return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1119 }
1120
1121 /*
1122  * The enable_on_exec/disabled value strategy:
1123  *
1124  *  1) For any type of traced program:
1125  *    - all independent events and group leaders are disabled
1126  *    - all group members are enabled
1127  *
1128  *     Group members are ruled by group leaders. They need to
1129  *     be enabled, because the group scheduling relies on that.
1130  *
1131  *  2) For traced programs executed by perf:
1132  *     - all independent events and group leaders have
1133  *       enable_on_exec set
1134  *     - we don't specifically enable or disable any event during
1135  *       the record command
1136  *
1137  *     Independent events and group leaders are initially disabled
1138  *     and get enabled by exec. Group members are ruled by group
1139  *     leaders as stated in 1).
1140  *
1141  *  3) For traced programs attached by perf (pid/tid):
1142  *     - we specifically enable or disable all events during
1143  *       the record command
1144  *
1145  *     When attaching events to already running traced we
1146  *     enable/disable events specifically, as there's no
1147  *     initial traced exec call.
1148  */
1149 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1150                    struct callchain_param *callchain)
1151 {
1152         struct evsel *leader = evsel__leader(evsel);
1153         struct perf_event_attr *attr = &evsel->core.attr;
1154         int track = evsel->tracking;
1155         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1156
1157         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1158         attr->inherit       = !opts->no_inherit;
1159         attr->write_backward = opts->overwrite ? 1 : 0;
1160
1161         evsel__set_sample_bit(evsel, IP);
1162         evsel__set_sample_bit(evsel, TID);
1163
1164         if (evsel->sample_read) {
1165                 evsel__set_sample_bit(evsel, READ);
1166
1167                 /*
1168                  * We need ID even in case of single event, because
1169                  * PERF_SAMPLE_READ process ID specific data.
1170                  */
1171                 evsel__set_sample_id(evsel, false);
1172
1173                 /*
1174                  * Apply group format only if we belong to group
1175                  * with more than one members.
1176                  */
1177                 if (leader->core.nr_members > 1) {
1178                         attr->read_format |= PERF_FORMAT_GROUP;
1179                         attr->inherit = 0;
1180                 }
1181         }
1182
1183         /*
1184          * We default some events to have a default interval. But keep
1185          * it a weak assumption overridable by the user.
1186          */
1187         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1188             (!evsel->is_libpfm_event && (!attr->sample_period ||
1189                                          opts->user_freq != UINT_MAX ||
1190                                          opts->user_interval != ULLONG_MAX)))
1191                 evsel__set_default_freq_period(opts, attr);
1192
1193         /*
1194          * If attr->freq was set (here or earlier), ask for period
1195          * to be sampled.
1196          */
1197         if (attr->freq)
1198                 evsel__set_sample_bit(evsel, PERIOD);
1199
1200         if (opts->no_samples)
1201                 attr->sample_freq = 0;
1202
1203         if (opts->inherit_stat) {
1204                 evsel->core.attr.read_format |=
1205                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1206                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1207                         PERF_FORMAT_ID;
1208                 attr->inherit_stat = 1;
1209         }
1210
1211         if (opts->sample_address) {
1212                 evsel__set_sample_bit(evsel, ADDR);
1213                 attr->mmap_data = track;
1214         }
1215
1216         /*
1217          * We don't allow user space callchains for  function trace
1218          * event, due to issues with page faults while tracing page
1219          * fault handler and its overall trickiness nature.
1220          */
1221         if (evsel__is_function_event(evsel))
1222                 evsel->core.attr.exclude_callchain_user = 1;
1223
1224         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1225                 evsel__config_callchain(evsel, opts, callchain);
1226
1227         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1228             !evsel__is_dummy_event(evsel)) {
1229                 attr->sample_regs_intr = opts->sample_intr_regs;
1230                 evsel__set_sample_bit(evsel, REGS_INTR);
1231         }
1232
1233         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1234             !evsel__is_dummy_event(evsel)) {
1235                 attr->sample_regs_user |= opts->sample_user_regs;
1236                 evsel__set_sample_bit(evsel, REGS_USER);
1237         }
1238
1239         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1240                 evsel__set_sample_bit(evsel, CPU);
1241
1242         /*
1243          * When the user explicitly disabled time don't force it here.
1244          */
1245         if (opts->sample_time &&
1246             (!perf_missing_features.sample_id_all &&
1247             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1248              opts->sample_time_set)))
1249                 evsel__set_sample_bit(evsel, TIME);
1250
1251         if (opts->raw_samples && !evsel->no_aux_samples) {
1252                 evsel__set_sample_bit(evsel, TIME);
1253                 evsel__set_sample_bit(evsel, RAW);
1254                 evsel__set_sample_bit(evsel, CPU);
1255         }
1256
1257         if (opts->sample_address)
1258                 evsel__set_sample_bit(evsel, DATA_SRC);
1259
1260         if (opts->sample_phys_addr)
1261                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1262
1263         if (opts->no_buffering) {
1264                 attr->watermark = 0;
1265                 attr->wakeup_events = 1;
1266         }
1267         if (opts->branch_stack && !evsel->no_aux_samples) {
1268                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1269                 attr->branch_sample_type = opts->branch_stack;
1270         }
1271
1272         if (opts->sample_weight)
1273                 arch_evsel__set_sample_weight(evsel);
1274
1275         attr->task     = track;
1276         attr->mmap     = track;
1277         attr->mmap2    = track && !perf_missing_features.mmap2;
1278         attr->comm     = track;
1279         attr->build_id = track && opts->build_id;
1280
1281         /*
1282          * ksymbol is tracked separately with text poke because it needs to be
1283          * system wide and enabled immediately.
1284          */
1285         if (!opts->text_poke)
1286                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1287         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1288
1289         if (opts->record_namespaces)
1290                 attr->namespaces  = track;
1291
1292         if (opts->record_cgroup) {
1293                 attr->cgroup = track && !perf_missing_features.cgroup;
1294                 evsel__set_sample_bit(evsel, CGROUP);
1295         }
1296
1297         if (opts->sample_data_page_size)
1298                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1299
1300         if (opts->sample_code_page_size)
1301                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1302
1303         if (opts->record_switch_events)
1304                 attr->context_switch = track;
1305
1306         if (opts->sample_transaction)
1307                 evsel__set_sample_bit(evsel, TRANSACTION);
1308
1309         if (opts->running_time) {
1310                 evsel->core.attr.read_format |=
1311                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1312                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1313         }
1314
1315         /*
1316          * XXX see the function comment above
1317          *
1318          * Disabling only independent events or group leaders,
1319          * keeping group members enabled.
1320          */
1321         if (evsel__is_group_leader(evsel))
1322                 attr->disabled = 1;
1323
1324         /*
1325          * Setting enable_on_exec for independent events and
1326          * group leaders for traced executed by perf.
1327          */
1328         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1329             !opts->initial_delay)
1330                 attr->enable_on_exec = 1;
1331
1332         if (evsel->immediate) {
1333                 attr->disabled = 0;
1334                 attr->enable_on_exec = 0;
1335         }
1336
1337         clockid = opts->clockid;
1338         if (opts->use_clockid) {
1339                 attr->use_clockid = 1;
1340                 attr->clockid = opts->clockid;
1341         }
1342
1343         if (evsel->precise_max)
1344                 attr->precise_ip = 3;
1345
1346         if (opts->all_user) {
1347                 attr->exclude_kernel = 1;
1348                 attr->exclude_user   = 0;
1349         }
1350
1351         if (opts->all_kernel) {
1352                 attr->exclude_kernel = 0;
1353                 attr->exclude_user   = 1;
1354         }
1355
1356         if (evsel->core.own_cpus || evsel->unit)
1357                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1358
1359         /*
1360          * Apply event specific term settings,
1361          * it overloads any global configuration.
1362          */
1363         evsel__apply_config_terms(evsel, opts, track);
1364
1365         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1366
1367         /* The --period option takes the precedence. */
1368         if (opts->period_set) {
1369                 if (opts->period)
1370                         evsel__set_sample_bit(evsel, PERIOD);
1371                 else
1372                         evsel__reset_sample_bit(evsel, PERIOD);
1373         }
1374
1375         /*
1376          * A dummy event never triggers any actual counter and therefore
1377          * cannot be used with branch_stack.
1378          *
1379          * For initial_delay, a dummy event is added implicitly.
1380          * The software event will trigger -EOPNOTSUPP error out,
1381          * if BRANCH_STACK bit is set.
1382          */
1383         if (evsel__is_dummy_event(evsel))
1384                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1385
1386         if (evsel__is_offcpu_event(evsel))
1387                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1388
1389         arch__post_evsel_config(evsel, attr);
1390 }
1391
1392 int evsel__set_filter(struct evsel *evsel, const char *filter)
1393 {
1394         char *new_filter = strdup(filter);
1395
1396         if (new_filter != NULL) {
1397                 free(evsel->filter);
1398                 evsel->filter = new_filter;
1399                 return 0;
1400         }
1401
1402         return -1;
1403 }
1404
1405 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1406 {
1407         char *new_filter;
1408
1409         if (evsel->filter == NULL)
1410                 return evsel__set_filter(evsel, filter);
1411
1412         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1413                 free(evsel->filter);
1414                 evsel->filter = new_filter;
1415                 return 0;
1416         }
1417
1418         return -1;
1419 }
1420
1421 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1422 {
1423         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1424 }
1425
1426 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1427 {
1428         return evsel__append_filter(evsel, "%s,%s", filter);
1429 }
1430
1431 /* Caller has to clear disabled after going through all CPUs. */
1432 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1433 {
1434         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1435 }
1436
1437 int evsel__enable(struct evsel *evsel)
1438 {
1439         int err = perf_evsel__enable(&evsel->core);
1440
1441         if (!err)
1442                 evsel->disabled = false;
1443         return err;
1444 }
1445
1446 /* Caller has to set disabled after going through all CPUs. */
1447 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1448 {
1449         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1450 }
1451
1452 int evsel__disable(struct evsel *evsel)
1453 {
1454         int err = perf_evsel__disable(&evsel->core);
1455         /*
1456          * We mark it disabled here so that tools that disable a event can
1457          * ignore events after they disable it. I.e. the ring buffer may have
1458          * already a few more events queued up before the kernel got the stop
1459          * request.
1460          */
1461         if (!err)
1462                 evsel->disabled = true;
1463
1464         return err;
1465 }
1466
1467 void free_config_terms(struct list_head *config_terms)
1468 {
1469         struct evsel_config_term *term, *h;
1470
1471         list_for_each_entry_safe(term, h, config_terms, list) {
1472                 list_del_init(&term->list);
1473                 if (term->free_str)
1474                         zfree(&term->val.str);
1475                 free(term);
1476         }
1477 }
1478
1479 static void evsel__free_config_terms(struct evsel *evsel)
1480 {
1481         free_config_terms(&evsel->config_terms);
1482 }
1483
1484 void evsel__exit(struct evsel *evsel)
1485 {
1486         assert(list_empty(&evsel->core.node));
1487         assert(evsel->evlist == NULL);
1488         bpf_counter__destroy(evsel);
1489         evsel__free_counts(evsel);
1490         perf_evsel__free_fd(&evsel->core);
1491         perf_evsel__free_id(&evsel->core);
1492         evsel__free_config_terms(evsel);
1493         cgroup__put(evsel->cgrp);
1494         perf_cpu_map__put(evsel->core.cpus);
1495         perf_cpu_map__put(evsel->core.own_cpus);
1496         perf_thread_map__put(evsel->core.threads);
1497         zfree(&evsel->group_name);
1498         zfree(&evsel->name);
1499         zfree(&evsel->pmu_name);
1500         zfree(&evsel->unit);
1501         zfree(&evsel->metric_id);
1502         evsel__zero_per_pkg(evsel);
1503         hashmap__free(evsel->per_pkg_mask);
1504         evsel->per_pkg_mask = NULL;
1505         zfree(&evsel->metric_events);
1506         perf_evsel__object.fini(evsel);
1507 }
1508
1509 void evsel__delete(struct evsel *evsel)
1510 {
1511         evsel__exit(evsel);
1512         free(evsel);
1513 }
1514
1515 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1516                            struct perf_counts_values *count)
1517 {
1518         struct perf_counts_values tmp;
1519
1520         if (!evsel->prev_raw_counts)
1521                 return;
1522
1523         if (cpu_map_idx == -1) {
1524                 tmp = evsel->prev_raw_counts->aggr;
1525                 evsel->prev_raw_counts->aggr = *count;
1526         } else {
1527                 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1528                 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1529         }
1530
1531         count->val = count->val - tmp.val;
1532         count->ena = count->ena - tmp.ena;
1533         count->run = count->run - tmp.run;
1534 }
1535
1536 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1537 {
1538         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1539
1540         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1541 }
1542
1543 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1544                              u64 val, u64 ena, u64 run)
1545 {
1546         struct perf_counts_values *count;
1547
1548         count = perf_counts(counter->counts, cpu_map_idx, thread);
1549
1550         count->val    = val;
1551         count->ena    = ena;
1552         count->run    = run;
1553
1554         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1555 }
1556
1557 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1558 {
1559         u64 read_format = leader->core.attr.read_format;
1560         struct sample_read_value *v;
1561         u64 nr, ena = 0, run = 0, i;
1562
1563         nr = *data++;
1564
1565         if (nr != (u64) leader->core.nr_members)
1566                 return -EINVAL;
1567
1568         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1569                 ena = *data++;
1570
1571         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1572                 run = *data++;
1573
1574         v = (struct sample_read_value *) data;
1575
1576         evsel__set_count(leader, cpu_map_idx, thread, v[0].value, ena, run);
1577
1578         for (i = 1; i < nr; i++) {
1579                 struct evsel *counter;
1580
1581                 counter = evlist__id2evsel(leader->evlist, v[i].id);
1582                 if (!counter)
1583                         return -EINVAL;
1584
1585                 evsel__set_count(counter, cpu_map_idx, thread, v[i].value, ena, run);
1586         }
1587
1588         return 0;
1589 }
1590
1591 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1592 {
1593         struct perf_stat_evsel *ps = leader->stats;
1594         u64 read_format = leader->core.attr.read_format;
1595         int size = perf_evsel__read_size(&leader->core);
1596         u64 *data = ps->group_data;
1597
1598         if (!(read_format & PERF_FORMAT_ID))
1599                 return -EINVAL;
1600
1601         if (!evsel__is_group_leader(leader))
1602                 return -EINVAL;
1603
1604         if (!data) {
1605                 data = zalloc(size);
1606                 if (!data)
1607                         return -ENOMEM;
1608
1609                 ps->group_data = data;
1610         }
1611
1612         if (FD(leader, cpu_map_idx, thread) < 0)
1613                 return -EINVAL;
1614
1615         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1616                 return -errno;
1617
1618         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1619 }
1620
1621 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1622 {
1623         u64 read_format = evsel->core.attr.read_format;
1624
1625         if (read_format & PERF_FORMAT_GROUP)
1626                 return evsel__read_group(evsel, cpu_map_idx, thread);
1627
1628         return evsel__read_one(evsel, cpu_map_idx, thread);
1629 }
1630
1631 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1632 {
1633         struct perf_counts_values count;
1634         size_t nv = scale ? 3 : 1;
1635
1636         if (FD(evsel, cpu_map_idx, thread) < 0)
1637                 return -EINVAL;
1638
1639         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1640                 return -ENOMEM;
1641
1642         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1643                 return -errno;
1644
1645         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1646         perf_counts_values__scale(&count, scale, NULL);
1647         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1648         return 0;
1649 }
1650
1651 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1652                                   int cpu_map_idx)
1653 {
1654         struct perf_cpu cpu;
1655
1656         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1657         return perf_cpu_map__idx(other->core.cpus, cpu);
1658 }
1659
1660 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1661 {
1662         struct evsel *leader = evsel__leader(evsel);
1663
1664         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1665             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1666                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1667         }
1668
1669         return cpu_map_idx;
1670 }
1671
1672 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1673 {
1674         struct evsel *leader = evsel__leader(evsel);
1675         int fd;
1676
1677         if (evsel__is_group_leader(evsel))
1678                 return -1;
1679
1680         /*
1681          * Leader must be already processed/open,
1682          * if not it's a bug.
1683          */
1684         BUG_ON(!leader->core.fd);
1685
1686         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1687         if (cpu_map_idx == -1)
1688                 return -1;
1689
1690         fd = FD(leader, cpu_map_idx, thread);
1691         BUG_ON(fd == -1);
1692
1693         return fd;
1694 }
1695
1696 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1697 {
1698         for (int cpu = 0; cpu < nr_cpus; cpu++)
1699                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1700                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1701 }
1702
1703 static int update_fds(struct evsel *evsel,
1704                       int nr_cpus, int cpu_map_idx,
1705                       int nr_threads, int thread_idx)
1706 {
1707         struct evsel *pos;
1708
1709         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1710                 return -EINVAL;
1711
1712         evlist__for_each_entry(evsel->evlist, pos) {
1713                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1714
1715                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1716
1717                 /*
1718                  * Since fds for next evsel has not been created,
1719                  * there is no need to iterate whole event list.
1720                  */
1721                 if (pos == evsel)
1722                         break;
1723         }
1724         return 0;
1725 }
1726
1727 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1728                                          int nr_cpus, int cpu_map_idx,
1729                                          struct perf_thread_map *threads,
1730                                          int thread, int err)
1731 {
1732         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1733
1734         if (!evsel->ignore_missing_thread)
1735                 return false;
1736
1737         /* The system wide setup does not work with threads. */
1738         if (evsel->core.system_wide)
1739                 return false;
1740
1741         /* The -ESRCH is perf event syscall errno for pid's not found. */
1742         if (err != -ESRCH)
1743                 return false;
1744
1745         /* If there's only one thread, let it fail. */
1746         if (threads->nr == 1)
1747                 return false;
1748
1749         /*
1750          * We should remove fd for missing_thread first
1751          * because thread_map__remove() will decrease threads->nr.
1752          */
1753         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1754                 return false;
1755
1756         if (thread_map__remove(threads, thread))
1757                 return false;
1758
1759         pr_warning("WARNING: Ignored open failure for pid %d\n",
1760                    ignore_pid);
1761         return true;
1762 }
1763
1764 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1765                                 void *priv __maybe_unused)
1766 {
1767         return fprintf(fp, "  %-32s %s\n", name, val);
1768 }
1769
1770 static void display_attr(struct perf_event_attr *attr)
1771 {
1772         if (verbose >= 2 || debug_peo_args) {
1773                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1774                 fprintf(stderr, "perf_event_attr:\n");
1775                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1776                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1777         }
1778 }
1779
1780 bool evsel__precise_ip_fallback(struct evsel *evsel)
1781 {
1782         /* Do not try less precise if not requested. */
1783         if (!evsel->precise_max)
1784                 return false;
1785
1786         /*
1787          * We tried all the precise_ip values, and it's
1788          * still failing, so leave it to standard fallback.
1789          */
1790         if (!evsel->core.attr.precise_ip) {
1791                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1792                 return false;
1793         }
1794
1795         if (!evsel->precise_ip_original)
1796                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1797
1798         evsel->core.attr.precise_ip--;
1799         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1800         display_attr(&evsel->core.attr);
1801         return true;
1802 }
1803
1804 static struct perf_cpu_map *empty_cpu_map;
1805 static struct perf_thread_map *empty_thread_map;
1806
1807 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1808                 struct perf_thread_map *threads)
1809 {
1810         int nthreads;
1811
1812         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1813             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1814                 return -EINVAL;
1815
1816         if (cpus == NULL) {
1817                 if (empty_cpu_map == NULL) {
1818                         empty_cpu_map = perf_cpu_map__dummy_new();
1819                         if (empty_cpu_map == NULL)
1820                                 return -ENOMEM;
1821                 }
1822
1823                 cpus = empty_cpu_map;
1824         }
1825
1826         if (threads == NULL) {
1827                 if (empty_thread_map == NULL) {
1828                         empty_thread_map = thread_map__new_by_tid(-1);
1829                         if (empty_thread_map == NULL)
1830                                 return -ENOMEM;
1831                 }
1832
1833                 threads = empty_thread_map;
1834         }
1835
1836         if (evsel->core.system_wide)
1837                 nthreads = 1;
1838         else
1839                 nthreads = threads->nr;
1840
1841         if (evsel->core.fd == NULL &&
1842             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1843                 return -ENOMEM;
1844
1845         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1846         if (evsel->cgrp)
1847                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1848
1849         return 0;
1850 }
1851
1852 static void evsel__disable_missing_features(struct evsel *evsel)
1853 {
1854         if (perf_missing_features.weight_struct) {
1855                 evsel__set_sample_bit(evsel, WEIGHT);
1856                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1857         }
1858         if (perf_missing_features.clockid_wrong)
1859                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1860         if (perf_missing_features.clockid) {
1861                 evsel->core.attr.use_clockid = 0;
1862                 evsel->core.attr.clockid = 0;
1863         }
1864         if (perf_missing_features.cloexec)
1865                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1866         if (perf_missing_features.mmap2)
1867                 evsel->core.attr.mmap2 = 0;
1868         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1869                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1870         if (perf_missing_features.lbr_flags)
1871                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1872                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1873         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1874                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1875         if (perf_missing_features.ksymbol)
1876                 evsel->core.attr.ksymbol = 0;
1877         if (perf_missing_features.bpf)
1878                 evsel->core.attr.bpf_event = 0;
1879         if (perf_missing_features.branch_hw_idx)
1880                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1881         if (perf_missing_features.sample_id_all)
1882                 evsel->core.attr.sample_id_all = 0;
1883 }
1884
1885 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1886                         struct perf_thread_map *threads)
1887 {
1888         int err;
1889
1890         err = __evsel__prepare_open(evsel, cpus, threads);
1891         if (err)
1892                 return err;
1893
1894         evsel__disable_missing_features(evsel);
1895
1896         return err;
1897 }
1898
1899 bool evsel__detect_missing_features(struct evsel *evsel)
1900 {
1901         /*
1902          * Must probe features in the order they were added to the
1903          * perf_event_attr interface.
1904          */
1905         if (!perf_missing_features.weight_struct &&
1906             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1907                 perf_missing_features.weight_struct = true;
1908                 pr_debug2("switching off weight struct support\n");
1909                 return true;
1910         } else if (!perf_missing_features.code_page_size &&
1911             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1912                 perf_missing_features.code_page_size = true;
1913                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1914                 return false;
1915         } else if (!perf_missing_features.data_page_size &&
1916             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1917                 perf_missing_features.data_page_size = true;
1918                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1919                 return false;
1920         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1921                 perf_missing_features.cgroup = true;
1922                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1923                 return false;
1924         } else if (!perf_missing_features.branch_hw_idx &&
1925             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1926                 perf_missing_features.branch_hw_idx = true;
1927                 pr_debug2("switching off branch HW index support\n");
1928                 return true;
1929         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1930                 perf_missing_features.aux_output = true;
1931                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1932                 return false;
1933         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1934                 perf_missing_features.bpf = true;
1935                 pr_debug2_peo("switching off bpf_event\n");
1936                 return true;
1937         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1938                 perf_missing_features.ksymbol = true;
1939                 pr_debug2_peo("switching off ksymbol\n");
1940                 return true;
1941         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1942                 perf_missing_features.write_backward = true;
1943                 pr_debug2_peo("switching off write_backward\n");
1944                 return false;
1945         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1946                 perf_missing_features.clockid_wrong = true;
1947                 pr_debug2_peo("switching off clockid\n");
1948                 return true;
1949         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1950                 perf_missing_features.clockid = true;
1951                 pr_debug2_peo("switching off use_clockid\n");
1952                 return true;
1953         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1954                 perf_missing_features.cloexec = true;
1955                 pr_debug2_peo("switching off cloexec flag\n");
1956                 return true;
1957         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1958                 perf_missing_features.mmap2 = true;
1959                 pr_debug2_peo("switching off mmap2\n");
1960                 return true;
1961         } else if ((evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) &&
1962                    (evsel->pmu == NULL || evsel->pmu->missing_features.exclude_guest)) {
1963                 if (evsel->pmu == NULL) {
1964                         evsel->pmu = evsel__find_pmu(evsel);
1965                         if (evsel->pmu)
1966                                 evsel->pmu->missing_features.exclude_guest = true;
1967                         else {
1968                                 /* we cannot find PMU, disable attrs now */
1969                                 evsel->core.attr.exclude_host = false;
1970                                 evsel->core.attr.exclude_guest = false;
1971                         }
1972                 }
1973
1974                 if (evsel->exclude_GH) {
1975                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1976                         return false;
1977                 }
1978                 if (!perf_missing_features.exclude_guest) {
1979                         perf_missing_features.exclude_guest = true;
1980                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1981                 }
1982                 return true;
1983         } else if (!perf_missing_features.sample_id_all) {
1984                 perf_missing_features.sample_id_all = true;
1985                 pr_debug2_peo("switching off sample_id_all\n");
1986                 return true;
1987         } else if (!perf_missing_features.lbr_flags &&
1988                         (evsel->core.attr.branch_sample_type &
1989                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1990                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1991                 perf_missing_features.lbr_flags = true;
1992                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1993                 return true;
1994         } else if (!perf_missing_features.group_read &&
1995                     evsel->core.attr.inherit &&
1996                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1997                    evsel__is_group_leader(evsel)) {
1998                 perf_missing_features.group_read = true;
1999                 pr_debug2_peo("switching off group read\n");
2000                 return true;
2001         } else {
2002                 return false;
2003         }
2004 }
2005
2006 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2007 {
2008         int old_errno;
2009         struct rlimit l;
2010
2011         if (*set_rlimit < INCREASED_MAX) {
2012                 old_errno = errno;
2013
2014                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2015                         if (*set_rlimit == NO_CHANGE) {
2016                                 l.rlim_cur = l.rlim_max;
2017                         } else {
2018                                 l.rlim_cur = l.rlim_max + 1000;
2019                                 l.rlim_max = l.rlim_cur;
2020                         }
2021                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2022                                 (*set_rlimit) += 1;
2023                                 errno = old_errno;
2024                                 return true;
2025                         }
2026                 }
2027                 errno = old_errno;
2028         }
2029
2030         return false;
2031 }
2032
2033 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2034                 struct perf_thread_map *threads,
2035                 int start_cpu_map_idx, int end_cpu_map_idx)
2036 {
2037         int idx, thread, nthreads;
2038         int pid = -1, err, old_errno;
2039         enum rlimit_action set_rlimit = NO_CHANGE;
2040
2041         err = __evsel__prepare_open(evsel, cpus, threads);
2042         if (err)
2043                 return err;
2044
2045         if (cpus == NULL)
2046                 cpus = empty_cpu_map;
2047
2048         if (threads == NULL)
2049                 threads = empty_thread_map;
2050
2051         if (evsel->core.system_wide)
2052                 nthreads = 1;
2053         else
2054                 nthreads = threads->nr;
2055
2056         if (evsel->cgrp)
2057                 pid = evsel->cgrp->fd;
2058
2059 fallback_missing_features:
2060         evsel__disable_missing_features(evsel);
2061
2062         display_attr(&evsel->core.attr);
2063
2064         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2065
2066                 for (thread = 0; thread < nthreads; thread++) {
2067                         int fd, group_fd;
2068 retry_open:
2069                         if (thread >= nthreads)
2070                                 break;
2071
2072                         if (!evsel->cgrp && !evsel->core.system_wide)
2073                                 pid = perf_thread_map__pid(threads, thread);
2074
2075                         group_fd = get_group_fd(evsel, idx, thread);
2076
2077                         test_attr__ready();
2078
2079                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2080                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2081
2082                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2083                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2084                                                 group_fd, evsel->open_flags);
2085
2086                         FD(evsel, idx, thread) = fd;
2087
2088                         if (fd < 0) {
2089                                 err = -errno;
2090
2091                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2092                                           err);
2093                                 goto try_fallback;
2094                         }
2095
2096                         bpf_counter__install_pe(evsel, idx, fd);
2097
2098                         if (unlikely(test_attr__enabled)) {
2099                                 test_attr__open(&evsel->core.attr, pid,
2100                                                 perf_cpu_map__cpu(cpus, idx),
2101                                                 fd, group_fd, evsel->open_flags);
2102                         }
2103
2104                         pr_debug2_peo(" = %d\n", fd);
2105
2106                         if (evsel->bpf_fd >= 0) {
2107                                 int evt_fd = fd;
2108                                 int bpf_fd = evsel->bpf_fd;
2109
2110                                 err = ioctl(evt_fd,
2111                                             PERF_EVENT_IOC_SET_BPF,
2112                                             bpf_fd);
2113                                 if (err && errno != EEXIST) {
2114                                         pr_err("failed to attach bpf fd %d: %s\n",
2115                                                bpf_fd, strerror(errno));
2116                                         err = -EINVAL;
2117                                         goto out_close;
2118                                 }
2119                         }
2120
2121                         set_rlimit = NO_CHANGE;
2122
2123                         /*
2124                          * If we succeeded but had to kill clockid, fail and
2125                          * have evsel__open_strerror() print us a nice error.
2126                          */
2127                         if (perf_missing_features.clockid ||
2128                             perf_missing_features.clockid_wrong) {
2129                                 err = -EINVAL;
2130                                 goto out_close;
2131                         }
2132                 }
2133         }
2134
2135         return 0;
2136
2137 try_fallback:
2138         if (evsel__precise_ip_fallback(evsel))
2139                 goto retry_open;
2140
2141         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2142                                          idx, threads, thread, err)) {
2143                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2144                 nthreads--;
2145
2146                 /* ... and pretend like nothing have happened. */
2147                 err = 0;
2148                 goto retry_open;
2149         }
2150         /*
2151          * perf stat needs between 5 and 22 fds per CPU. When we run out
2152          * of them try to increase the limits.
2153          */
2154         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2155                 goto retry_open;
2156
2157         if (err != -EINVAL || idx > 0 || thread > 0)
2158                 goto out_close;
2159
2160         if (evsel__detect_missing_features(evsel))
2161                 goto fallback_missing_features;
2162 out_close:
2163         if (err)
2164                 threads->err_thread = thread;
2165
2166         old_errno = errno;
2167         do {
2168                 while (--thread >= 0) {
2169                         if (FD(evsel, idx, thread) >= 0)
2170                                 close(FD(evsel, idx, thread));
2171                         FD(evsel, idx, thread) = -1;
2172                 }
2173                 thread = nthreads;
2174         } while (--idx >= 0);
2175         errno = old_errno;
2176         return err;
2177 }
2178
2179 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2180                 struct perf_thread_map *threads)
2181 {
2182         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2183 }
2184
2185 void evsel__close(struct evsel *evsel)
2186 {
2187         perf_evsel__close(&evsel->core);
2188         perf_evsel__free_id(&evsel->core);
2189 }
2190
2191 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2192 {
2193         if (cpu_map_idx == -1)
2194                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2195
2196         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2197 }
2198
2199 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2200 {
2201         return evsel__open(evsel, NULL, threads);
2202 }
2203
2204 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2205                                        const union perf_event *event,
2206                                        struct perf_sample *sample)
2207 {
2208         u64 type = evsel->core.attr.sample_type;
2209         const __u64 *array = event->sample.array;
2210         bool swapped = evsel->needs_swap;
2211         union u64_swap u;
2212
2213         array += ((event->header.size -
2214                    sizeof(event->header)) / sizeof(u64)) - 1;
2215
2216         if (type & PERF_SAMPLE_IDENTIFIER) {
2217                 sample->id = *array;
2218                 array--;
2219         }
2220
2221         if (type & PERF_SAMPLE_CPU) {
2222                 u.val64 = *array;
2223                 if (swapped) {
2224                         /* undo swap of u64, then swap on individual u32s */
2225                         u.val64 = bswap_64(u.val64);
2226                         u.val32[0] = bswap_32(u.val32[0]);
2227                 }
2228
2229                 sample->cpu = u.val32[0];
2230                 array--;
2231         }
2232
2233         if (type & PERF_SAMPLE_STREAM_ID) {
2234                 sample->stream_id = *array;
2235                 array--;
2236         }
2237
2238         if (type & PERF_SAMPLE_ID) {
2239                 sample->id = *array;
2240                 array--;
2241         }
2242
2243         if (type & PERF_SAMPLE_TIME) {
2244                 sample->time = *array;
2245                 array--;
2246         }
2247
2248         if (type & PERF_SAMPLE_TID) {
2249                 u.val64 = *array;
2250                 if (swapped) {
2251                         /* undo swap of u64, then swap on individual u32s */
2252                         u.val64 = bswap_64(u.val64);
2253                         u.val32[0] = bswap_32(u.val32[0]);
2254                         u.val32[1] = bswap_32(u.val32[1]);
2255                 }
2256
2257                 sample->pid = u.val32[0];
2258                 sample->tid = u.val32[1];
2259                 array--;
2260         }
2261
2262         return 0;
2263 }
2264
2265 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2266                             u64 size)
2267 {
2268         return size > max_size || offset + size > endp;
2269 }
2270
2271 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2272         do {                                                            \
2273                 if (overflow(endp, (max_size), (offset), (size)))       \
2274                         return -EFAULT;                                 \
2275         } while (0)
2276
2277 #define OVERFLOW_CHECK_u64(offset) \
2278         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2279
2280 static int
2281 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2282 {
2283         /*
2284          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2285          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2286          * check the format does not go past the end of the event.
2287          */
2288         if (sample_size + sizeof(event->header) > event->header.size)
2289                 return -EFAULT;
2290
2291         return 0;
2292 }
2293
2294 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2295                                           const __u64 *array,
2296                                           u64 type __maybe_unused)
2297 {
2298         data->weight = *array;
2299 }
2300
2301 u64 evsel__bitfield_swap_branch_flags(u64 value)
2302 {
2303         u64 new_val = 0;
2304
2305         /*
2306          * branch_flags
2307          * union {
2308          *      u64 values;
2309          *      struct {
2310          *              mispred:1       //target mispredicted
2311          *              predicted:1     //target predicted
2312          *              in_tx:1         //in transaction
2313          *              abort:1         //transaction abort
2314          *              cycles:16       //cycle count to last branch
2315          *              type:4          //branch type
2316          *              reserved:40
2317          *      }
2318          * }
2319          *
2320          * Avoid bswap64() the entire branch_flag.value,
2321          * as it has variable bit-field sizes. Instead the
2322          * macro takes the bit-field position/size,
2323          * swaps it based on the host endianness.
2324          *
2325          * tep_is_bigendian() is used here instead of
2326          * bigendian() to avoid python test fails.
2327          */
2328         if (tep_is_bigendian()) {
2329                 new_val = bitfield_swap(value, 0, 1);
2330                 new_val |= bitfield_swap(value, 1, 1);
2331                 new_val |= bitfield_swap(value, 2, 1);
2332                 new_val |= bitfield_swap(value, 3, 1);
2333                 new_val |= bitfield_swap(value, 4, 16);
2334                 new_val |= bitfield_swap(value, 20, 4);
2335                 new_val |= bitfield_swap(value, 24, 40);
2336         } else {
2337                 new_val = bitfield_swap(value, 63, 1);
2338                 new_val |= bitfield_swap(value, 62, 1);
2339                 new_val |= bitfield_swap(value, 61, 1);
2340                 new_val |= bitfield_swap(value, 60, 1);
2341                 new_val |= bitfield_swap(value, 44, 16);
2342                 new_val |= bitfield_swap(value, 40, 4);
2343                 new_val |= bitfield_swap(value, 0, 40);
2344         }
2345
2346         return new_val;
2347 }
2348
2349 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2350                         struct perf_sample *data)
2351 {
2352         u64 type = evsel->core.attr.sample_type;
2353         bool swapped = evsel->needs_swap;
2354         const __u64 *array;
2355         u16 max_size = event->header.size;
2356         const void *endp = (void *)event + max_size;
2357         u64 sz;
2358
2359         /*
2360          * used for cross-endian analysis. See git commit 65014ab3
2361          * for why this goofiness is needed.
2362          */
2363         union u64_swap u;
2364
2365         memset(data, 0, sizeof(*data));
2366         data->cpu = data->pid = data->tid = -1;
2367         data->stream_id = data->id = data->time = -1ULL;
2368         data->period = evsel->core.attr.sample_period;
2369         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2370         data->misc    = event->header.misc;
2371         data->id = -1ULL;
2372         data->data_src = PERF_MEM_DATA_SRC_NONE;
2373         data->vcpu = -1;
2374
2375         if (event->header.type != PERF_RECORD_SAMPLE) {
2376                 if (!evsel->core.attr.sample_id_all)
2377                         return 0;
2378                 return perf_evsel__parse_id_sample(evsel, event, data);
2379         }
2380
2381         array = event->sample.array;
2382
2383         if (perf_event__check_size(event, evsel->sample_size))
2384                 return -EFAULT;
2385
2386         if (type & PERF_SAMPLE_IDENTIFIER) {
2387                 data->id = *array;
2388                 array++;
2389         }
2390
2391         if (type & PERF_SAMPLE_IP) {
2392                 data->ip = *array;
2393                 array++;
2394         }
2395
2396         if (type & PERF_SAMPLE_TID) {
2397                 u.val64 = *array;
2398                 if (swapped) {
2399                         /* undo swap of u64, then swap on individual u32s */
2400                         u.val64 = bswap_64(u.val64);
2401                         u.val32[0] = bswap_32(u.val32[0]);
2402                         u.val32[1] = bswap_32(u.val32[1]);
2403                 }
2404
2405                 data->pid = u.val32[0];
2406                 data->tid = u.val32[1];
2407                 array++;
2408         }
2409
2410         if (type & PERF_SAMPLE_TIME) {
2411                 data->time = *array;
2412                 array++;
2413         }
2414
2415         if (type & PERF_SAMPLE_ADDR) {
2416                 data->addr = *array;
2417                 array++;
2418         }
2419
2420         if (type & PERF_SAMPLE_ID) {
2421                 data->id = *array;
2422                 array++;
2423         }
2424
2425         if (type & PERF_SAMPLE_STREAM_ID) {
2426                 data->stream_id = *array;
2427                 array++;
2428         }
2429
2430         if (type & PERF_SAMPLE_CPU) {
2431
2432                 u.val64 = *array;
2433                 if (swapped) {
2434                         /* undo swap of u64, then swap on individual u32s */
2435                         u.val64 = bswap_64(u.val64);
2436                         u.val32[0] = bswap_32(u.val32[0]);
2437                 }
2438
2439                 data->cpu = u.val32[0];
2440                 array++;
2441         }
2442
2443         if (type & PERF_SAMPLE_PERIOD) {
2444                 data->period = *array;
2445                 array++;
2446         }
2447
2448         if (type & PERF_SAMPLE_READ) {
2449                 u64 read_format = evsel->core.attr.read_format;
2450
2451                 OVERFLOW_CHECK_u64(array);
2452                 if (read_format & PERF_FORMAT_GROUP)
2453                         data->read.group.nr = *array;
2454                 else
2455                         data->read.one.value = *array;
2456
2457                 array++;
2458
2459                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2460                         OVERFLOW_CHECK_u64(array);
2461                         data->read.time_enabled = *array;
2462                         array++;
2463                 }
2464
2465                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2466                         OVERFLOW_CHECK_u64(array);
2467                         data->read.time_running = *array;
2468                         array++;
2469                 }
2470
2471                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2472                 if (read_format & PERF_FORMAT_GROUP) {
2473                         const u64 max_group_nr = UINT64_MAX /
2474                                         sizeof(struct sample_read_value);
2475
2476                         if (data->read.group.nr > max_group_nr)
2477                                 return -EFAULT;
2478                         sz = data->read.group.nr *
2479                              sizeof(struct sample_read_value);
2480                         OVERFLOW_CHECK(array, sz, max_size);
2481                         data->read.group.values =
2482                                         (struct sample_read_value *)array;
2483                         array = (void *)array + sz;
2484                 } else {
2485                         OVERFLOW_CHECK_u64(array);
2486                         data->read.one.id = *array;
2487                         array++;
2488                 }
2489         }
2490
2491         if (type & PERF_SAMPLE_CALLCHAIN) {
2492                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2493
2494                 OVERFLOW_CHECK_u64(array);
2495                 data->callchain = (struct ip_callchain *)array++;
2496                 if (data->callchain->nr > max_callchain_nr)
2497                         return -EFAULT;
2498                 sz = data->callchain->nr * sizeof(u64);
2499                 OVERFLOW_CHECK(array, sz, max_size);
2500                 array = (void *)array + sz;
2501         }
2502
2503         if (type & PERF_SAMPLE_RAW) {
2504                 OVERFLOW_CHECK_u64(array);
2505                 u.val64 = *array;
2506
2507                 /*
2508                  * Undo swap of u64, then swap on individual u32s,
2509                  * get the size of the raw area and undo all of the
2510                  * swap. The pevent interface handles endianness by
2511                  * itself.
2512                  */
2513                 if (swapped) {
2514                         u.val64 = bswap_64(u.val64);
2515                         u.val32[0] = bswap_32(u.val32[0]);
2516                         u.val32[1] = bswap_32(u.val32[1]);
2517                 }
2518                 data->raw_size = u.val32[0];
2519
2520                 /*
2521                  * The raw data is aligned on 64bits including the
2522                  * u32 size, so it's safe to use mem_bswap_64.
2523                  */
2524                 if (swapped)
2525                         mem_bswap_64((void *) array, data->raw_size);
2526
2527                 array = (void *)array + sizeof(u32);
2528
2529                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2530                 data->raw_data = (void *)array;
2531                 array = (void *)array + data->raw_size;
2532         }
2533
2534         if (type & PERF_SAMPLE_BRANCH_STACK) {
2535                 const u64 max_branch_nr = UINT64_MAX /
2536                                           sizeof(struct branch_entry);
2537                 struct branch_entry *e;
2538                 unsigned int i;
2539
2540                 OVERFLOW_CHECK_u64(array);
2541                 data->branch_stack = (struct branch_stack *)array++;
2542
2543                 if (data->branch_stack->nr > max_branch_nr)
2544                         return -EFAULT;
2545
2546                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2547                 if (evsel__has_branch_hw_idx(evsel)) {
2548                         sz += sizeof(u64);
2549                         e = &data->branch_stack->entries[0];
2550                 } else {
2551                         data->no_hw_idx = true;
2552                         /*
2553                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2554                          * only nr and entries[] will be output by kernel.
2555                          */
2556                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2557                 }
2558
2559                 if (swapped) {
2560                         /*
2561                          * struct branch_flag does not have endian
2562                          * specific bit field definition. And bswap
2563                          * will not resolve the issue, since these
2564                          * are bit fields.
2565                          *
2566                          * evsel__bitfield_swap_branch_flags() uses a
2567                          * bitfield_swap macro to swap the bit position
2568                          * based on the host endians.
2569                          */
2570                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2571                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2572                 }
2573
2574                 OVERFLOW_CHECK(array, sz, max_size);
2575                 array = (void *)array + sz;
2576         }
2577
2578         if (type & PERF_SAMPLE_REGS_USER) {
2579                 OVERFLOW_CHECK_u64(array);
2580                 data->user_regs.abi = *array;
2581                 array++;
2582
2583                 if (data->user_regs.abi) {
2584                         u64 mask = evsel->core.attr.sample_regs_user;
2585
2586                         sz = hweight64(mask) * sizeof(u64);
2587                         OVERFLOW_CHECK(array, sz, max_size);
2588                         data->user_regs.mask = mask;
2589                         data->user_regs.regs = (u64 *)array;
2590                         array = (void *)array + sz;
2591                 }
2592         }
2593
2594         if (type & PERF_SAMPLE_STACK_USER) {
2595                 OVERFLOW_CHECK_u64(array);
2596                 sz = *array++;
2597
2598                 data->user_stack.offset = ((char *)(array - 1)
2599                                           - (char *) event);
2600
2601                 if (!sz) {
2602                         data->user_stack.size = 0;
2603                 } else {
2604                         OVERFLOW_CHECK(array, sz, max_size);
2605                         data->user_stack.data = (char *)array;
2606                         array = (void *)array + sz;
2607                         OVERFLOW_CHECK_u64(array);
2608                         data->user_stack.size = *array++;
2609                         if (WARN_ONCE(data->user_stack.size > sz,
2610                                       "user stack dump failure\n"))
2611                                 return -EFAULT;
2612                 }
2613         }
2614
2615         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2616                 OVERFLOW_CHECK_u64(array);
2617                 arch_perf_parse_sample_weight(data, array, type);
2618                 array++;
2619         }
2620
2621         if (type & PERF_SAMPLE_DATA_SRC) {
2622                 OVERFLOW_CHECK_u64(array);
2623                 data->data_src = *array;
2624                 array++;
2625         }
2626
2627         if (type & PERF_SAMPLE_TRANSACTION) {
2628                 OVERFLOW_CHECK_u64(array);
2629                 data->transaction = *array;
2630                 array++;
2631         }
2632
2633         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2634         if (type & PERF_SAMPLE_REGS_INTR) {
2635                 OVERFLOW_CHECK_u64(array);
2636                 data->intr_regs.abi = *array;
2637                 array++;
2638
2639                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2640                         u64 mask = evsel->core.attr.sample_regs_intr;
2641
2642                         sz = hweight64(mask) * sizeof(u64);
2643                         OVERFLOW_CHECK(array, sz, max_size);
2644                         data->intr_regs.mask = mask;
2645                         data->intr_regs.regs = (u64 *)array;
2646                         array = (void *)array + sz;
2647                 }
2648         }
2649
2650         data->phys_addr = 0;
2651         if (type & PERF_SAMPLE_PHYS_ADDR) {
2652                 data->phys_addr = *array;
2653                 array++;
2654         }
2655
2656         data->cgroup = 0;
2657         if (type & PERF_SAMPLE_CGROUP) {
2658                 data->cgroup = *array;
2659                 array++;
2660         }
2661
2662         data->data_page_size = 0;
2663         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2664                 data->data_page_size = *array;
2665                 array++;
2666         }
2667
2668         data->code_page_size = 0;
2669         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2670                 data->code_page_size = *array;
2671                 array++;
2672         }
2673
2674         if (type & PERF_SAMPLE_AUX) {
2675                 OVERFLOW_CHECK_u64(array);
2676                 sz = *array++;
2677
2678                 OVERFLOW_CHECK(array, sz, max_size);
2679                 /* Undo swap of data */
2680                 if (swapped)
2681                         mem_bswap_64((char *)array, sz);
2682                 data->aux_sample.size = sz;
2683                 data->aux_sample.data = (char *)array;
2684                 array = (void *)array + sz;
2685         }
2686
2687         return 0;
2688 }
2689
2690 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2691                                   u64 *timestamp)
2692 {
2693         u64 type = evsel->core.attr.sample_type;
2694         const __u64 *array;
2695
2696         if (!(type & PERF_SAMPLE_TIME))
2697                 return -1;
2698
2699         if (event->header.type != PERF_RECORD_SAMPLE) {
2700                 struct perf_sample data = {
2701                         .time = -1ULL,
2702                 };
2703
2704                 if (!evsel->core.attr.sample_id_all)
2705                         return -1;
2706                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2707                         return -1;
2708
2709                 *timestamp = data.time;
2710                 return 0;
2711         }
2712
2713         array = event->sample.array;
2714
2715         if (perf_event__check_size(event, evsel->sample_size))
2716                 return -EFAULT;
2717
2718         if (type & PERF_SAMPLE_IDENTIFIER)
2719                 array++;
2720
2721         if (type & PERF_SAMPLE_IP)
2722                 array++;
2723
2724         if (type & PERF_SAMPLE_TID)
2725                 array++;
2726
2727         if (type & PERF_SAMPLE_TIME)
2728                 *timestamp = *array;
2729
2730         return 0;
2731 }
2732
2733 u16 evsel__id_hdr_size(struct evsel *evsel)
2734 {
2735         u64 sample_type = evsel->core.attr.sample_type;
2736         u16 size = 0;
2737
2738         if (sample_type & PERF_SAMPLE_TID)
2739                 size += sizeof(u64);
2740
2741         if (sample_type & PERF_SAMPLE_TIME)
2742                 size += sizeof(u64);
2743
2744         if (sample_type & PERF_SAMPLE_ID)
2745                 size += sizeof(u64);
2746
2747         if (sample_type & PERF_SAMPLE_STREAM_ID)
2748                 size += sizeof(u64);
2749
2750         if (sample_type & PERF_SAMPLE_CPU)
2751                 size += sizeof(u64);
2752
2753         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2754                 size += sizeof(u64);
2755
2756         return size;
2757 }
2758
2759 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2760 {
2761         return tep_find_field(evsel->tp_format, name);
2762 }
2763
2764 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2765 {
2766         struct tep_format_field *field = evsel__field(evsel, name);
2767         int offset;
2768
2769         if (!field)
2770                 return NULL;
2771
2772         offset = field->offset;
2773
2774         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2775                 offset = *(int *)(sample->raw_data + field->offset);
2776                 offset &= 0xffff;
2777                 if (field->flags & TEP_FIELD_IS_RELATIVE)
2778                         offset += field->offset + field->size;
2779         }
2780
2781         return sample->raw_data + offset;
2782 }
2783
2784 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2785                          bool needs_swap)
2786 {
2787         u64 value;
2788         void *ptr = sample->raw_data + field->offset;
2789
2790         switch (field->size) {
2791         case 1:
2792                 return *(u8 *)ptr;
2793         case 2:
2794                 value = *(u16 *)ptr;
2795                 break;
2796         case 4:
2797                 value = *(u32 *)ptr;
2798                 break;
2799         case 8:
2800                 memcpy(&value, ptr, sizeof(u64));
2801                 break;
2802         default:
2803                 return 0;
2804         }
2805
2806         if (!needs_swap)
2807                 return value;
2808
2809         switch (field->size) {
2810         case 2:
2811                 return bswap_16(value);
2812         case 4:
2813                 return bswap_32(value);
2814         case 8:
2815                 return bswap_64(value);
2816         default:
2817                 return 0;
2818         }
2819
2820         return 0;
2821 }
2822
2823 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2824 {
2825         struct tep_format_field *field = evsel__field(evsel, name);
2826
2827         if (!field)
2828                 return 0;
2829
2830         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2831 }
2832
2833 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2834 {
2835         int paranoid;
2836
2837         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2838             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2839             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2840                 /*
2841                  * If it's cycles then fall back to hrtimer based
2842                  * cpu-clock-tick sw counter, which is always available even if
2843                  * no PMU support.
2844                  *
2845                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2846                  * b0a873e).
2847                  */
2848                 scnprintf(msg, msgsize, "%s",
2849 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2850
2851                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2852                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2853
2854                 zfree(&evsel->name);
2855                 return true;
2856         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2857                    (paranoid = perf_event_paranoid()) > 1) {
2858                 const char *name = evsel__name(evsel);
2859                 char *new_name;
2860                 const char *sep = ":";
2861
2862                 /* If event has exclude user then don't exclude kernel. */
2863                 if (evsel->core.attr.exclude_user)
2864                         return false;
2865
2866                 /* Is there already the separator in the name. */
2867                 if (strchr(name, '/') ||
2868                     (strchr(name, ':') && !evsel->is_libpfm_event))
2869                         sep = "";
2870
2871                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2872                         return false;
2873
2874                 if (evsel->name)
2875                         free(evsel->name);
2876                 evsel->name = new_name;
2877                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2878                           "to fall back to excluding kernel and hypervisor "
2879                           " samples", paranoid);
2880                 evsel->core.attr.exclude_kernel = 1;
2881                 evsel->core.attr.exclude_hv     = 1;
2882
2883                 return true;
2884         }
2885
2886         return false;
2887 }
2888
2889 static bool find_process(const char *name)
2890 {
2891         size_t len = strlen(name);
2892         DIR *dir;
2893         struct dirent *d;
2894         int ret = -1;
2895
2896         dir = opendir(procfs__mountpoint());
2897         if (!dir)
2898                 return false;
2899
2900         /* Walk through the directory. */
2901         while (ret && (d = readdir(dir)) != NULL) {
2902                 char path[PATH_MAX];
2903                 char *data;
2904                 size_t size;
2905
2906                 if ((d->d_type != DT_DIR) ||
2907                      !strcmp(".", d->d_name) ||
2908                      !strcmp("..", d->d_name))
2909                         continue;
2910
2911                 scnprintf(path, sizeof(path), "%s/%s/comm",
2912                           procfs__mountpoint(), d->d_name);
2913
2914                 if (filename__read_str(path, &data, &size))
2915                         continue;
2916
2917                 ret = strncmp(name, data, len);
2918                 free(data);
2919         }
2920
2921         closedir(dir);
2922         return ret ? false : true;
2923 }
2924
2925 static bool is_amd(const char *arch, const char *cpuid)
2926 {
2927         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2928 }
2929
2930 static bool is_amd_ibs(struct evsel *evsel)
2931 {
2932         return evsel->core.attr.precise_ip
2933             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2934 }
2935
2936 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2937                          int err, char *msg, size_t size)
2938 {
2939         struct perf_env *env = evsel__env(evsel);
2940         const char *arch = perf_env__arch(env);
2941         const char *cpuid = perf_env__cpuid(env);
2942         char sbuf[STRERR_BUFSIZE];
2943         int printed = 0, enforced = 0;
2944
2945         switch (err) {
2946         case EPERM:
2947         case EACCES:
2948                 printed += scnprintf(msg + printed, size - printed,
2949                         "Access to performance monitoring and observability operations is limited.\n");
2950
2951                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2952                         if (enforced) {
2953                                 printed += scnprintf(msg + printed, size - printed,
2954                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2955                                         "monitoring and observability operations. Inspect system audit records for\n"
2956                                         "more perf_event access control information and adjusting the policy.\n");
2957                         }
2958                 }
2959
2960                 if (err == EPERM)
2961                         printed += scnprintf(msg, size,
2962                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2963
2964                 return scnprintf(msg + printed, size - printed,
2965                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2966                  "access to performance monitoring and observability operations for processes\n"
2967                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2968                  "More information can be found at 'Perf events and tool security' document:\n"
2969                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2970                  "perf_event_paranoid setting is %d:\n"
2971                  "  -1: Allow use of (almost) all events by all users\n"
2972                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2973                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2974                  ">= 1: Disallow CPU event access\n"
2975                  ">= 2: Disallow kernel profiling\n"
2976                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2977                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2978                  perf_event_paranoid());
2979         case ENOENT:
2980                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2981         case EMFILE:
2982                 return scnprintf(msg, size, "%s",
2983                          "Too many events are opened.\n"
2984                          "Probably the maximum number of open file descriptors has been reached.\n"
2985                          "Hint: Try again after reducing the number of events.\n"
2986                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2987         case ENOMEM:
2988                 if (evsel__has_callchain(evsel) &&
2989                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2990                         return scnprintf(msg, size,
2991                                          "Not enough memory to setup event with callchain.\n"
2992                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2993                                          "Hint: Current value: %d", sysctl__max_stack());
2994                 break;
2995         case ENODEV:
2996                 if (target->cpu_list)
2997                         return scnprintf(msg, size, "%s",
2998          "No such device - did you specify an out-of-range profile CPU?");
2999                 break;
3000         case EOPNOTSUPP:
3001                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3002                         return scnprintf(msg, size,
3003         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3004                                          evsel__name(evsel));
3005                 if (evsel->core.attr.aux_output)
3006                         return scnprintf(msg, size,
3007         "%s: PMU Hardware doesn't support 'aux_output' feature",
3008                                          evsel__name(evsel));
3009                 if (evsel->core.attr.sample_period != 0)
3010                         return scnprintf(msg, size,
3011         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3012                                          evsel__name(evsel));
3013                 if (evsel->core.attr.precise_ip)
3014                         return scnprintf(msg, size, "%s",
3015         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3016 #if defined(__i386__) || defined(__x86_64__)
3017                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3018                         return scnprintf(msg, size, "%s",
3019         "No hardware sampling interrupt available.\n");
3020 #endif
3021                 break;
3022         case EBUSY:
3023                 if (find_process("oprofiled"))
3024                         return scnprintf(msg, size,
3025         "The PMU counters are busy/taken by another profiler.\n"
3026         "We found oprofile daemon running, please stop it and try again.");
3027                 break;
3028         case EINVAL:
3029                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3030                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3031                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3032                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3033                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3034                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3035                 if (perf_missing_features.clockid)
3036                         return scnprintf(msg, size, "clockid feature not supported.");
3037                 if (perf_missing_features.clockid_wrong)
3038                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3039                 if (perf_missing_features.aux_output)
3040                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3041                 if (!target__has_cpu(target))
3042                         return scnprintf(msg, size,
3043         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3044                                         evsel__name(evsel));
3045                 if (is_amd(arch, cpuid)) {
3046                         if (is_amd_ibs(evsel)) {
3047                                 if (evsel->core.attr.exclude_kernel)
3048                                         return scnprintf(msg, size,
3049         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3050                                 if (!evsel->core.system_wide)
3051                                         return scnprintf(msg, size,
3052         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3053                         }
3054                 }
3055
3056                 break;
3057         case ENODATA:
3058                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3059                                  "Please add an auxiliary event in front of the load latency event.");
3060         default:
3061                 break;
3062         }
3063
3064         return scnprintf(msg, size,
3065         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3066         "/bin/dmesg | grep -i perf may provide additional information.\n",
3067                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3068 }
3069
3070 struct perf_env *evsel__env(struct evsel *evsel)
3071 {
3072         if (evsel && evsel->evlist && evsel->evlist->env)
3073                 return evsel->evlist->env;
3074         return &perf_env;
3075 }
3076
3077 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3078 {
3079         int cpu_map_idx, thread;
3080
3081         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3082                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3083                      thread++) {
3084                         int fd = FD(evsel, cpu_map_idx, thread);
3085
3086                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3087                                                    cpu_map_idx, thread, fd) < 0)
3088                                 return -1;
3089                 }
3090         }
3091
3092         return 0;
3093 }
3094
3095 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3096 {
3097         struct perf_cpu_map *cpus = evsel->core.cpus;
3098         struct perf_thread_map *threads = evsel->core.threads;
3099
3100         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3101                 return -ENOMEM;
3102
3103         return store_evsel_ids(evsel, evlist);
3104 }
3105
3106 void evsel__zero_per_pkg(struct evsel *evsel)
3107 {
3108         struct hashmap_entry *cur;
3109         size_t bkt;
3110
3111         if (evsel->per_pkg_mask) {
3112                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3113                         free((char *)cur->key);
3114
3115                 hashmap__clear(evsel->per_pkg_mask);
3116         }
3117 }
3118
3119 bool evsel__is_hybrid(struct evsel *evsel)
3120 {
3121         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3122 }
3123
3124 struct evsel *evsel__leader(struct evsel *evsel)
3125 {
3126         return container_of(evsel->core.leader, struct evsel, core);
3127 }
3128
3129 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3130 {
3131         return evsel->core.leader == &leader->core;
3132 }
3133
3134 bool evsel__is_leader(struct evsel *evsel)
3135 {
3136         return evsel__has_leader(evsel, evsel);
3137 }
3138
3139 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3140 {
3141         evsel->core.leader = &leader->core;
3142 }
3143
3144 int evsel__source_count(const struct evsel *evsel)
3145 {
3146         struct evsel *pos;
3147         int count = 0;
3148
3149         evlist__for_each_entry(evsel->evlist, pos) {
3150                 if (pos->metric_leader == evsel)
3151                         count++;
3152         }
3153         return count;
3154 }
3155
3156 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3157 {
3158         return false;
3159 }
3160
3161 /*
3162  * Remove an event from a given group (leader).
3163  * Some events, e.g., perf metrics Topdown events,
3164  * must always be grouped. Ignore the events.
3165  */
3166 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3167 {
3168         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3169                 evsel__set_leader(evsel, evsel);
3170                 evsel->core.nr_members = 0;
3171                 leader->core.nr_members--;
3172         }
3173 }