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