tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <linux/hw_breakpoint.h>
  16 #include <linux/perf_event.h>
  17 #include <linux/compiler.h>
  18 #include <linux/err.h>
  19 #include <linux/zalloc.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/resource.h>
  22 #include <sys/types.h>
  23 #include <dirent.h>
  24 #include <stdlib.h>
  25 #include <perf/evsel.h>
  26 #include "asm/bug.h"
  27 #include "bpf_counter.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "util/hashmap.h"
  49 #include "off_cpu.h"
  50 #include "pmu.h"
  51 #include "pmus.h"
  52 #include "../perf-sys.h"
  53 #include "util/parse-branch-options.h"
  54 #include "util/bpf-filter.h"
  55 #include <internal/xyarray.h>
  56 #include <internal/lib.h>
  57 #include <internal/threadmap.h>
  58
  59 #include <linux/ctype.h>
  60
  61 #ifdef HAVE_LIBTRACEEVENT
  62 #include <traceevent/event-parse.h>
  63 #endif
  64
  65 struct perf_missing_features perf_missing_features;
  66
  67 static clockid_t clockid;
  68
  69 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  70         NULL,
  71         "duration_time",
  72         "user_time",
  73         "system_time",
  74 };
  75
  76 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  77 {
  78         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  79                 return perf_tool_event__tool_names[ev];
  80
  81         return NULL;
  82 }
  83
  84 enum perf_tool_event perf_tool_event__from_str(const char *str)
  85 {
  86         int i;
  87
  88         perf_tool_event__for_each_event(i) {
  89                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  90                         return i;
  91         }
  92         return PERF_TOOL_NONE;
  93 }
  94
  95
  96 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  97 {
  98         return 0;
  99 }
 100
 101 void __weak test_attr__ready(void) { }
 102
 103 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 104 {
 105 }
 106
 107 static struct {
 108         size_t  size;
 109         int     (*init)(struct evsel *evsel);
 110         void    (*fini)(struct evsel *evsel);
 111 } perf_evsel__object = {
 112         .size = sizeof(struct evsel),
 113         .init = evsel__no_extra_init,
 114         .fini = evsel__no_extra_fini,
 115 };
 116
 117 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 118                          void (*fini)(struct evsel *evsel))
 119 {
 120
 121         if (object_size == 0)
 122                 goto set_methods;
 123
 124         if (perf_evsel__object.size > object_size)
 125                 return -EINVAL;
 126
 127         perf_evsel__object.size = object_size;
 128
 129 set_methods:
 130         if (init != NULL)
 131                 perf_evsel__object.init = init;
 132
 133         if (fini != NULL)
 134                 perf_evsel__object.fini = fini;
 135
 136         return 0;
 137 }
 138
 139 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 140
 141 int __evsel__sample_size(u64 sample_type)
 142 {
 143         u64 mask = sample_type & PERF_SAMPLE_MASK;
 144         int size = 0;
 145         int i;
 146
 147         for (i = 0; i < 64; i++) {
 148                 if (mask & (1ULL << i))
 149                         size++;
 150         }
 151
 152         size *= sizeof(u64);
 153
 154         return size;
 155 }
 156
 157 /**
 158  * __perf_evsel__calc_id_pos - calculate id_pos.
 159  * @sample_type: sample type
 160  *
 161  * This function returns the position of the event id (PERF_SAMPLE_ID or
 162  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 163  * perf_record_sample.
 164  */
 165 static int __perf_evsel__calc_id_pos(u64 sample_type)
 166 {
 167         int idx = 0;
 168
 169         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 170                 return 0;
 171
 172         if (!(sample_type & PERF_SAMPLE_ID))
 173                 return -1;
 174
 175         if (sample_type & PERF_SAMPLE_IP)
 176                 idx += 1;
 177
 178         if (sample_type & PERF_SAMPLE_TID)
 179                 idx += 1;
 180
 181         if (sample_type & PERF_SAMPLE_TIME)
 182                 idx += 1;
 183
 184         if (sample_type & PERF_SAMPLE_ADDR)
 185                 idx += 1;
 186
 187         return idx;
 188 }
 189
 190 /**
 191  * __perf_evsel__calc_is_pos - calculate is_pos.
 192  * @sample_type: sample type
 193  *
 194  * This function returns the position (counting backwards) of the event id
 195  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 196  * sample_id_all is used there is an id sample appended to non-sample events.
 197  */
 198 static int __perf_evsel__calc_is_pos(u64 sample_type)
 199 {
 200         int idx = 1;
 201
 202         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 203                 return 1;
 204
 205         if (!(sample_type & PERF_SAMPLE_ID))
 206                 return -1;
 207
 208         if (sample_type & PERF_SAMPLE_CPU)
 209                 idx += 1;
 210
 211         if (sample_type & PERF_SAMPLE_STREAM_ID)
 212                 idx += 1;
 213
 214         return idx;
 215 }
 216
 217 void evsel__calc_id_pos(struct evsel *evsel)
 218 {
 219         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 220         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 221 }
 222
 223 void __evsel__set_sample_bit(struct evsel *evsel,
 224                                   enum perf_event_sample_format bit)
 225 {
 226         if (!(evsel->core.attr.sample_type & bit)) {
 227                 evsel->core.attr.sample_type |= bit;
 228                 evsel->sample_size += sizeof(u64);
 229                 evsel__calc_id_pos(evsel);
 230         }
 231 }
 232
 233 void __evsel__reset_sample_bit(struct evsel *evsel,
 234                                     enum perf_event_sample_format bit)
 235 {
 236         if (evsel->core.attr.sample_type & bit) {
 237                 evsel->core.attr.sample_type &= ~bit;
 238                 evsel->sample_size -= sizeof(u64);
 239                 evsel__calc_id_pos(evsel);
 240         }
 241 }
 242
 243 void evsel__set_sample_id(struct evsel *evsel,
 244                                bool can_sample_identifier)
 245 {
 246         if (can_sample_identifier) {
 247                 evsel__reset_sample_bit(evsel, ID);
 248                 evsel__set_sample_bit(evsel, IDENTIFIER);
 249         } else {
 250                 evsel__set_sample_bit(evsel, ID);
 251         }
 252         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 253 }
 254
 255 /**
 256  * evsel__is_function_event - Return whether given evsel is a function
 257  * trace event
 258  *
 259  * @evsel - evsel selector to be tested
 260  *
 261  * Return %true if event is function trace event
 262  */
 263 bool evsel__is_function_event(struct evsel *evsel)
 264 {
 265 #define FUNCTION_EVENT "ftrace:function"
 266
 267         return evsel->name &&
 268                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 269
 270 #undef FUNCTION_EVENT
 271 }
 272
 273 void evsel__init(struct evsel *evsel,
 274                  struct perf_event_attr *attr, int idx)
 275 {
 276         perf_evsel__init(&evsel->core, attr, idx);
 277         evsel->tracking    = !idx;
 278         evsel->unit        = strdup("");
 279         evsel->scale       = 1.0;
 280         evsel->max_events  = ULONG_MAX;
 281         evsel->evlist      = NULL;
 282         evsel->bpf_obj     = NULL;
 283         evsel->bpf_fd      = -1;
 284         INIT_LIST_HEAD(&evsel->config_terms);
 285         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 286         INIT_LIST_HEAD(&evsel->bpf_filters);
 287         perf_evsel__object.init(evsel);
 288         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 289         evsel__calc_id_pos(evsel);
 290         evsel->cmdline_group_boundary = false;
 291         evsel->metric_events = NULL;
 292         evsel->per_pkg_mask  = NULL;
 293         evsel->collect_stat  = false;
 294         evsel->pmu_name      = NULL;
 295         evsel->group_pmu_name = NULL;
 296         evsel->skippable     = false;
 297 }
 298
 299 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 300 {
 301         struct evsel *evsel = zalloc(perf_evsel__object.size);
 302
 303         if (!evsel)
 304                 return NULL;
 305         evsel__init(evsel, attr, idx);
 306
 307         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 308                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 309                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 310                 evsel->core.attr.sample_period = 1;
 311         }
 312
 313         if (evsel__is_clock(evsel)) {
 314                 free((char *)evsel->unit);
 315                 evsel->unit = strdup("msec");
 316                 evsel->scale = 1e-6;
 317         }
 318
 319         return evsel;
 320 }
 321
 322 int copy_config_terms(struct list_head *dst, struct list_head *src)
 323 {
 324         struct evsel_config_term *pos, *tmp;
 325
 326         list_for_each_entry(pos, src, list) {
 327                 tmp = malloc(sizeof(*tmp));
 328                 if (tmp == NULL)
 329                         return -ENOMEM;
 330
 331                 *tmp = *pos;
 332                 if (tmp->free_str) {
 333                         tmp->val.str = strdup(pos->val.str);
 334                         if (tmp->val.str == NULL) {
 335                                 free(tmp);
 336                                 return -ENOMEM;
 337                         }
 338                 }
 339                 list_add_tail(&tmp->list, dst);
 340         }
 341         return 0;
 342 }
 343
 344 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 345 {
 346         return copy_config_terms(&dst->config_terms, &src->config_terms);
 347 }
 348
 349 /**
 350  * evsel__clone - create a new evsel copied from @orig
 351  * @orig: original evsel
 352  *
 353  * The assumption is that @orig is not configured nor opened yet.
 354  * So we only care about the attributes that can be set while it's parsed.
 355  */
 356 struct evsel *evsel__clone(struct evsel *orig)
 357 {
 358         struct evsel *evsel;
 359
 360         BUG_ON(orig->core.fd);
 361         BUG_ON(orig->counts);
 362         BUG_ON(orig->priv);
 363         BUG_ON(orig->per_pkg_mask);
 364
 365         /* cannot handle BPF objects for now */
 366         if (orig->bpf_obj)
 367                 return NULL;
 368
 369         evsel = evsel__new(&orig->core.attr);
 370         if (evsel == NULL)
 371                 return NULL;
 372
 373         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 374         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 375         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 376         evsel->core.nr_members = orig->core.nr_members;
 377         evsel->core.system_wide = orig->core.system_wide;
 378         evsel->core.requires_cpu = orig->core.requires_cpu;
 379         evsel->core.is_pmu_core = orig->core.is_pmu_core;
 380
 381         if (orig->name) {
 382                 evsel->name = strdup(orig->name);
 383                 if (evsel->name == NULL)
 384                         goto out_err;
 385         }
 386         if (orig->group_name) {
 387                 evsel->group_name = strdup(orig->group_name);
 388                 if (evsel->group_name == NULL)
 389                         goto out_err;
 390         }
 391         if (orig->pmu_name) {
 392                 evsel->pmu_name = strdup(orig->pmu_name);
 393                 if (evsel->pmu_name == NULL)
 394                         goto out_err;
 395         }
 396         if (orig->group_pmu_name) {
 397                 evsel->group_pmu_name = strdup(orig->group_pmu_name);
 398                 if (evsel->group_pmu_name == NULL)
 399                         goto out_err;
 400         }
 401         if (orig->filter) {
 402                 evsel->filter = strdup(orig->filter);
 403                 if (evsel->filter == NULL)
 404                         goto out_err;
 405         }
 406         if (orig->metric_id) {
 407                 evsel->metric_id = strdup(orig->metric_id);
 408                 if (evsel->metric_id == NULL)
 409                         goto out_err;
 410         }
 411         evsel->cgrp = cgroup__get(orig->cgrp);
 412 #ifdef HAVE_LIBTRACEEVENT
 413         evsel->tp_format = orig->tp_format;
 414 #endif
 415         evsel->handler = orig->handler;
 416         evsel->core.leader = orig->core.leader;
 417
 418         evsel->max_events = orig->max_events;
 419         evsel->tool_event = orig->tool_event;
 420         free((char *)evsel->unit);
 421         evsel->unit = strdup(orig->unit);
 422         if (evsel->unit == NULL)
 423                 goto out_err;
 424
 425         evsel->scale = orig->scale;
 426         evsel->snapshot = orig->snapshot;
 427         evsel->per_pkg = orig->per_pkg;
 428         evsel->percore = orig->percore;
 429         evsel->precise_max = orig->precise_max;
 430         evsel->is_libpfm_event = orig->is_libpfm_event;
 431
 432         evsel->exclude_GH = orig->exclude_GH;
 433         evsel->sample_read = orig->sample_read;
 434         evsel->auto_merge_stats = orig->auto_merge_stats;
 435         evsel->collect_stat = orig->collect_stat;
 436         evsel->weak_group = orig->weak_group;
 437         evsel->use_config_name = orig->use_config_name;
 438         evsel->pmu = orig->pmu;
 439
 440         if (evsel__copy_config_terms(evsel, orig) < 0)
 441                 goto out_err;
 442
 443         return evsel;
 444
 445 out_err:
 446         evsel__delete(evsel);
 447         return NULL;
 448 }
 449
 450 /*
 451  * Returns pointer with encoded error via <linux/err.h> interface.
 452  */
 453 #ifdef HAVE_LIBTRACEEVENT
 454 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 455 {
 456         struct evsel *evsel = zalloc(perf_evsel__object.size);
 457         int err = -ENOMEM;
 458
 459         if (evsel == NULL) {
 460                 goto out_err;
 461         } else {
 462                 struct perf_event_attr attr = {
 463                         .type          = PERF_TYPE_TRACEPOINT,
 464                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 465                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 466                 };
 467
 468                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 469                         goto out_free;
 470
 471                 evsel->tp_format = trace_event__tp_format(sys, name);
 472                 if (IS_ERR(evsel->tp_format)) {
 473                         err = PTR_ERR(evsel->tp_format);
 474                         goto out_free;
 475                 }
 476
 477                 event_attr_init(&attr);
 478                 attr.config = evsel->tp_format->id;
 479                 attr.sample_period = 1;
 480                 evsel__init(evsel, &attr, idx);
 481         }
 482
 483         return evsel;
 484
 485 out_free:
 486         zfree(&evsel->name);
 487         free(evsel);
 488 out_err:
 489         return ERR_PTR(err);
 490 }
 491 #endif
 492
 493 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 494         "cycles",
 495         "instructions",
 496         "cache-references",
 497         "cache-misses",
 498         "branches",
 499         "branch-misses",
 500         "bus-cycles",
 501         "stalled-cycles-frontend",
 502         "stalled-cycles-backend",
 503         "ref-cycles",
 504 };
 505
 506 char *evsel__bpf_counter_events;
 507
 508 bool evsel__match_bpf_counter_events(const char *name)
 509 {
 510         int name_len;
 511         bool match;
 512         char *ptr;
 513
 514         if (!evsel__bpf_counter_events)
 515                 return false;
 516
 517         ptr = strstr(evsel__bpf_counter_events, name);
 518         name_len = strlen(name);
 519
 520         /* check name matches a full token in evsel__bpf_counter_events */
 521         match = (ptr != NULL) &&
 522                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 523                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 524
 525         return match;
 526 }
 527
 528 static const char *__evsel__hw_name(u64 config)
 529 {
 530         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 531                 return evsel__hw_names[config];
 532
 533         return "unknown-hardware";
 534 }
 535
 536 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 537 {
 538         int colon = 0, r = 0;
 539         struct perf_event_attr *attr = &evsel->core.attr;
 540         bool exclude_guest_default = false;
 541
 542 #define MOD_PRINT(context, mod) do {                                    \
 543                 if (!attr->exclude_##context) {                         \
 544                         if (!colon) colon = ++r;                        \
 545                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 546                 } } while(0)
 547
 548         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 549                 MOD_PRINT(kernel, 'k');
 550                 MOD_PRINT(user, 'u');
 551                 MOD_PRINT(hv, 'h');
 552                 exclude_guest_default = true;
 553         }
 554
 555         if (attr->precise_ip) {
 556                 if (!colon)
 557                         colon = ++r;
 558                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 559                 exclude_guest_default = true;
 560         }
 561
 562         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 563                 MOD_PRINT(host, 'H');
 564                 MOD_PRINT(guest, 'G');
 565         }
 566 #undef MOD_PRINT
 567         if (colon)
 568                 bf[colon - 1] = ':';
 569         return r;
 570 }
 571
 572 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 573 {
 574         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 575 }
 576
 577 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 578 {
 579         int r = arch_evsel__hw_name(evsel, bf, size);
 580         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 581 }
 582
 583 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 584         "cpu-clock",
 585         "task-clock",
 586         "page-faults",
 587         "context-switches",
 588         "cpu-migrations",
 589         "minor-faults",
 590         "major-faults",
 591         "alignment-faults",
 592         "emulation-faults",
 593         "dummy",
 594 };
 595
 596 static const char *__evsel__sw_name(u64 config)
 597 {
 598         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 599                 return evsel__sw_names[config];
 600         return "unknown-software";
 601 }
 602
 603 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 604 {
 605         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 606         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 607 }
 608
 609 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 610 {
 611         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 612 }
 613
 614 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 615 {
 616         int r;
 617
 618         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 619
 620         if (type & HW_BREAKPOINT_R)
 621                 r += scnprintf(bf + r, size - r, "r");
 622
 623         if (type & HW_BREAKPOINT_W)
 624                 r += scnprintf(bf + r, size - r, "w");
 625
 626         if (type & HW_BREAKPOINT_X)
 627                 r += scnprintf(bf + r, size - r, "x");
 628
 629         return r;
 630 }
 631
 632 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 633 {
 634         struct perf_event_attr *attr = &evsel->core.attr;
 635         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 636         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 637 }
 638
 639 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 640  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 641  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 642  { "LLC",       "L2",                                                   },
 643  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 644  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 645  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 646  { "node",                                                              },
 647 };
 648
 649 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 650  { "load",      "loads",        "read",                                 },
 651  { "store",     "stores",       "write",                                },
 652  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 653 };
 654
 655 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 656  { "refs",      "Reference",    "ops",          "access",               },
 657  { "misses",    "miss",                                                 },
 658 };
 659
 660 #define C(x)            PERF_COUNT_HW_CACHE_##x
 661 #define CACHE_READ      (1 << C(OP_READ))
 662 #define CACHE_WRITE     (1 << C(OP_WRITE))
 663 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 664 #define COP(x)          (1 << x)
 665
 666 /*
 667  * cache operation stat
 668  * L1I : Read and prefetch only
 669  * ITLB and BPU : Read-only
 670  */
 671 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 672  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 673  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 674  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 675  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 676  [C(ITLB)]      = (CACHE_READ),
 677  [C(BPU)]       = (CACHE_READ),
 678  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 679 };
 680
 681 bool evsel__is_cache_op_valid(u8 type, u8 op)
 682 {
 683         if (evsel__hw_cache_stat[type] & COP(op))
 684                 return true;    /* valid */
 685         else
 686                 return false;   /* invalid */
 687 }
 688
 689 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 690 {
 691         if (result) {
 692                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 693                                  evsel__hw_cache_op[op][0],
 694                                  evsel__hw_cache_result[result][0]);
 695         }
 696
 697         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 698                          evsel__hw_cache_op[op][1]);
 699 }
 700
 701 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 702 {
 703         u8 op, result, type = (config >>  0) & 0xff;
 704         const char *err = "unknown-ext-hardware-cache-type";
 705
 706         if (type >= PERF_COUNT_HW_CACHE_MAX)
 707                 goto out_err;
 708
 709         op = (config >>  8) & 0xff;
 710         err = "unknown-ext-hardware-cache-op";
 711         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 712                 goto out_err;
 713
 714         result = (config >> 16) & 0xff;
 715         err = "unknown-ext-hardware-cache-result";
 716         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 717                 goto out_err;
 718
 719         err = "invalid-cache";
 720         if (!evsel__is_cache_op_valid(type, op))
 721                 goto out_err;
 722
 723         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 724 out_err:
 725         return scnprintf(bf, size, "%s", err);
 726 }
 727
 728 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 729 {
 730         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 731         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 732 }
 733
 734 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 735 {
 736         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 737         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 738 }
 739
 740 const char *evsel__name(struct evsel *evsel)
 741 {
 742         char bf[128];
 743
 744         if (!evsel)
 745                 goto out_unknown;
 746
 747         if (evsel->name)
 748                 return evsel->name;
 749
 750         switch (evsel->core.attr.type) {
 751         case PERF_TYPE_RAW:
 752                 evsel__raw_name(evsel, bf, sizeof(bf));
 753                 break;
 754
 755         case PERF_TYPE_HARDWARE:
 756                 evsel__hw_name(evsel, bf, sizeof(bf));
 757                 break;
 758
 759         case PERF_TYPE_HW_CACHE:
 760                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 761                 break;
 762
 763         case PERF_TYPE_SOFTWARE:
 764                 if (evsel__is_tool(evsel))
 765                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 766                 else
 767                         evsel__sw_name(evsel, bf, sizeof(bf));
 768                 break;
 769
 770         case PERF_TYPE_TRACEPOINT:
 771                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 772                 break;
 773
 774         case PERF_TYPE_BREAKPOINT:
 775                 evsel__bp_name(evsel, bf, sizeof(bf));
 776                 break;
 777
 778         default:
 779                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 780                           evsel->core.attr.type);
 781                 break;
 782         }
 783
 784         evsel->name = strdup(bf);
 785
 786         if (evsel->name)
 787                 return evsel->name;
 788 out_unknown:
 789         return "unknown";
 790 }
 791
 792 bool evsel__name_is(struct evsel *evsel, const char *name)
 793 {
 794         return !strcmp(evsel__name(evsel), name);
 795 }
 796
 797 const char *evsel__metric_id(const struct evsel *evsel)
 798 {
 799         if (evsel->metric_id)
 800                 return evsel->metric_id;
 801
 802         if (evsel__is_tool(evsel))
 803                 return perf_tool_event__to_str(evsel->tool_event);
 804
 805         return "unknown";
 806 }
 807
 808 const char *evsel__group_name(struct evsel *evsel)
 809 {
 810         return evsel->group_name ?: "anon group";
 811 }
 812
 813 /*
 814  * Returns the group details for the specified leader,
 815  * with following rules.
 816  *
 817  *  For record -e '{cycles,instructions}'
 818  *    'anon group { cycles:u, instructions:u }'
 819  *
 820  *  For record -e 'cycles,instructions' and report --group
 821  *    'cycles:u, instructions:u'
 822  */
 823 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 824 {
 825         int ret = 0;
 826         struct evsel *pos;
 827         const char *group_name = evsel__group_name(evsel);
 828
 829         if (!evsel->forced_leader)
 830                 ret = scnprintf(buf, size, "%s { ", group_name);
 831
 832         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 833
 834         for_each_group_member(pos, evsel)
 835                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 836
 837         if (!evsel->forced_leader)
 838                 ret += scnprintf(buf + ret, size - ret, " }");
 839
 840         return ret;
 841 }
 842
 843 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 844                                       struct callchain_param *param)
 845 {
 846         bool function = evsel__is_function_event(evsel);
 847         struct perf_event_attr *attr = &evsel->core.attr;
 848         const char *arch = perf_env__arch(evsel__env(evsel));
 849
 850         evsel__set_sample_bit(evsel, CALLCHAIN);
 851
 852         attr->sample_max_stack = param->max_stack;
 853
 854         if (opts->kernel_callchains)
 855                 attr->exclude_callchain_user = 1;
 856         if (opts->user_callchains)
 857                 attr->exclude_callchain_kernel = 1;
 858         if (param->record_mode == CALLCHAIN_LBR) {
 859                 if (!opts->branch_stack) {
 860                         if (attr->exclude_user) {
 861                                 pr_warning("LBR callstack option is only available "
 862                                            "to get user callchain information. "
 863                                            "Falling back to framepointers.\n");
 864                         } else {
 865                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 866                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 867                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 868                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 869                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 870                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 871                         }
 872                 } else
 873                          pr_warning("Cannot use LBR callstack with branch stack. "
 874                                     "Falling back to framepointers.\n");
 875         }
 876
 877         if (param->record_mode == CALLCHAIN_DWARF) {
 878                 if (!function) {
 879                         evsel__set_sample_bit(evsel, REGS_USER);
 880                         evsel__set_sample_bit(evsel, STACK_USER);
 881                         if (opts->sample_user_regs &&
 882                             DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) {
 883                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch);
 884                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 885                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 886                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 887                         } else {
 888                                 attr->sample_regs_user |= arch__user_reg_mask();
 889                         }
 890                         attr->sample_stack_user = param->dump_size;
 891                         attr->exclude_callchain_user = 1;
 892                 } else {
 893                         pr_info("Cannot use DWARF unwind for function trace event,"
 894                                 " falling back to framepointers.\n");
 895                 }
 896         }
 897
 898         if (function) {
 899                 pr_info("Disabling user space callchains for function trace event.\n");
 900                 attr->exclude_callchain_user = 1;
 901         }
 902 }
 903
 904 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 905                              struct callchain_param *param)
 906 {
 907         if (param->enabled)
 908                 return __evsel__config_callchain(evsel, opts, param);
 909 }
 910
 911 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 912 {
 913         struct perf_event_attr *attr = &evsel->core.attr;
 914
 915         evsel__reset_sample_bit(evsel, CALLCHAIN);
 916         if (param->record_mode == CALLCHAIN_LBR) {
 917                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 918                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 919                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 920                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 921         }
 922         if (param->record_mode == CALLCHAIN_DWARF) {
 923                 evsel__reset_sample_bit(evsel, REGS_USER);
 924                 evsel__reset_sample_bit(evsel, STACK_USER);
 925         }
 926 }
 927
 928 static void evsel__apply_config_terms(struct evsel *evsel,
 929                                       struct record_opts *opts, bool track)
 930 {
 931         struct evsel_config_term *term;
 932         struct list_head *config_terms = &evsel->config_terms;
 933         struct perf_event_attr *attr = &evsel->core.attr;
 934         /* callgraph default */
 935         struct callchain_param param = {
 936                 .record_mode = callchain_param.record_mode,
 937         };
 938         u32 dump_size = 0;
 939         int max_stack = 0;
 940         const char *callgraph_buf = NULL;
 941
 942         list_for_each_entry(term, config_terms, list) {
 943                 switch (term->type) {
 944                 case EVSEL__CONFIG_TERM_PERIOD:
 945                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 946                                 attr->sample_period = term->val.period;
 947                                 attr->freq = 0;
 948                                 evsel__reset_sample_bit(evsel, PERIOD);
 949                         }
 950                         break;
 951                 case EVSEL__CONFIG_TERM_FREQ:
 952                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 953                                 attr->sample_freq = term->val.freq;
 954                                 attr->freq = 1;
 955                                 evsel__set_sample_bit(evsel, PERIOD);
 956                         }
 957                         break;
 958                 case EVSEL__CONFIG_TERM_TIME:
 959                         if (term->val.time)
 960                                 evsel__set_sample_bit(evsel, TIME);
 961                         else
 962                                 evsel__reset_sample_bit(evsel, TIME);
 963                         break;
 964                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 965                         callgraph_buf = term->val.str;
 966                         break;
 967                 case EVSEL__CONFIG_TERM_BRANCH:
 968                         if (term->val.str && strcmp(term->val.str, "no")) {
 969                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 970                                 parse_branch_str(term->val.str,
 971                                                  &attr->branch_sample_type);
 972                         } else
 973                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 974                         break;
 975                 case EVSEL__CONFIG_TERM_STACK_USER:
 976                         dump_size = term->val.stack_user;
 977                         break;
 978                 case EVSEL__CONFIG_TERM_MAX_STACK:
 979                         max_stack = term->val.max_stack;
 980                         break;
 981                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 982                         evsel->max_events = term->val.max_events;
 983                         break;
 984                 case EVSEL__CONFIG_TERM_INHERIT:
 985                         /*
 986                          * attr->inherit should has already been set by
 987                          * evsel__config. If user explicitly set
 988                          * inherit using config terms, override global
 989                          * opt->no_inherit setting.
 990                          */
 991                         attr->inherit = term->val.inherit ? 1 : 0;
 992                         break;
 993                 case EVSEL__CONFIG_TERM_OVERWRITE:
 994                         attr->write_backward = term->val.overwrite ? 1 : 0;
 995                         break;
 996                 case EVSEL__CONFIG_TERM_DRV_CFG:
 997                         break;
 998                 case EVSEL__CONFIG_TERM_PERCORE:
 999                         break;
1000                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1001                         attr->aux_output = term->val.aux_output ? 1 : 0;
1002                         break;
1003                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1004                         /* Already applied by auxtrace */
1005                         break;
1006                 case EVSEL__CONFIG_TERM_CFG_CHG:
1007                         break;
1008                 default:
1009                         break;
1010                 }
1011         }
1012
1013         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1014         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1015                 bool sample_address = false;
1016
1017                 if (max_stack) {
1018                         param.max_stack = max_stack;
1019                         if (callgraph_buf == NULL)
1020                                 callgraph_buf = "fp";
1021                 }
1022
1023                 /* parse callgraph parameters */
1024                 if (callgraph_buf != NULL) {
1025                         if (!strcmp(callgraph_buf, "no")) {
1026                                 param.enabled = false;
1027                                 param.record_mode = CALLCHAIN_NONE;
1028                         } else {
1029                                 param.enabled = true;
1030                                 if (parse_callchain_record(callgraph_buf, &param)) {
1031                                         pr_err("per-event callgraph setting for %s failed. "
1032                                                "Apply callgraph global setting for it\n",
1033                                                evsel->name);
1034                                         return;
1035                                 }
1036                                 if (param.record_mode == CALLCHAIN_DWARF)
1037                                         sample_address = true;
1038                         }
1039                 }
1040                 if (dump_size > 0) {
1041                         dump_size = round_up(dump_size, sizeof(u64));
1042                         param.dump_size = dump_size;
1043                 }
1044
1045                 /* If global callgraph set, clear it */
1046                 if (callchain_param.enabled)
1047                         evsel__reset_callgraph(evsel, &callchain_param);
1048
1049                 /* set perf-event callgraph */
1050                 if (param.enabled) {
1051                         if (sample_address) {
1052                                 evsel__set_sample_bit(evsel, ADDR);
1053                                 evsel__set_sample_bit(evsel, DATA_SRC);
1054                                 evsel->core.attr.mmap_data = track;
1055                         }
1056                         evsel__config_callchain(evsel, opts, &param);
1057                 }
1058         }
1059 }
1060
1061 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1062 {
1063         struct evsel_config_term *term, *found_term = NULL;
1064
1065         list_for_each_entry(term, &evsel->config_terms, list) {
1066                 if (term->type == type)
1067                         found_term = term;
1068         }
1069
1070         return found_term;
1071 }
1072
1073 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1074 {
1075         evsel__set_sample_bit(evsel, WEIGHT);
1076 }
1077
1078 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1079                                     struct perf_event_attr *attr __maybe_unused)
1080 {
1081 }
1082
1083 static void evsel__set_default_freq_period(struct record_opts *opts,
1084                                            struct perf_event_attr *attr)
1085 {
1086         if (opts->freq) {
1087                 attr->freq = 1;
1088                 attr->sample_freq = opts->freq;
1089         } else {
1090                 attr->sample_period = opts->default_interval;
1091         }
1092 }
1093
1094 static bool evsel__is_offcpu_event(struct evsel *evsel)
1095 {
1096         return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1097 }
1098
1099 /*
1100  * The enable_on_exec/disabled value strategy:
1101  *
1102  *  1) For any type of traced program:
1103  *    - all independent events and group leaders are disabled
1104  *    - all group members are enabled
1105  *
1106  *     Group members are ruled by group leaders. They need to
1107  *     be enabled, because the group scheduling relies on that.
1108  *
1109  *  2) For traced programs executed by perf:
1110  *     - all independent events and group leaders have
1111  *       enable_on_exec set
1112  *     - we don't specifically enable or disable any event during
1113  *       the record command
1114  *
1115  *     Independent events and group leaders are initially disabled
1116  *     and get enabled by exec. Group members are ruled by group
1117  *     leaders as stated in 1).
1118  *
1119  *  3) For traced programs attached by perf (pid/tid):
1120  *     - we specifically enable or disable all events during
1121  *       the record command
1122  *
1123  *     When attaching events to already running traced we
1124  *     enable/disable events specifically, as there's no
1125  *     initial traced exec call.
1126  */
1127 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1128                    struct callchain_param *callchain)
1129 {
1130         struct evsel *leader = evsel__leader(evsel);
1131         struct perf_event_attr *attr = &evsel->core.attr;
1132         int track = evsel->tracking;
1133         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1134
1135         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1136         attr->inherit       = !opts->no_inherit;
1137         attr->write_backward = opts->overwrite ? 1 : 0;
1138         attr->read_format   = PERF_FORMAT_LOST;
1139
1140         evsel__set_sample_bit(evsel, IP);
1141         evsel__set_sample_bit(evsel, TID);
1142
1143         if (evsel->sample_read) {
1144                 evsel__set_sample_bit(evsel, READ);
1145
1146                 /*
1147                  * We need ID even in case of single event, because
1148                  * PERF_SAMPLE_READ process ID specific data.
1149                  */
1150                 evsel__set_sample_id(evsel, false);
1151
1152                 /*
1153                  * Apply group format only if we belong to group
1154                  * with more than one members.
1155                  */
1156                 if (leader->core.nr_members > 1) {
1157                         attr->read_format |= PERF_FORMAT_GROUP;
1158                         attr->inherit = 0;
1159                 }
1160         }
1161
1162         /*
1163          * We default some events to have a default interval. But keep
1164          * it a weak assumption overridable by the user.
1165          */
1166         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1167             (!evsel->is_libpfm_event && (!attr->sample_period ||
1168                                          opts->user_freq != UINT_MAX ||
1169                                          opts->user_interval != ULLONG_MAX)))
1170                 evsel__set_default_freq_period(opts, attr);
1171
1172         /*
1173          * If attr->freq was set (here or earlier), ask for period
1174          * to be sampled.
1175          */
1176         if (attr->freq)
1177                 evsel__set_sample_bit(evsel, PERIOD);
1178
1179         if (opts->no_samples)
1180                 attr->sample_freq = 0;
1181
1182         if (opts->inherit_stat) {
1183                 evsel->core.attr.read_format |=
1184                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1185                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1186                         PERF_FORMAT_ID;
1187                 attr->inherit_stat = 1;
1188         }
1189
1190         if (opts->sample_address) {
1191                 evsel__set_sample_bit(evsel, ADDR);
1192                 attr->mmap_data = track;
1193         }
1194
1195         /*
1196          * We don't allow user space callchains for  function trace
1197          * event, due to issues with page faults while tracing page
1198          * fault handler and its overall trickiness nature.
1199          */
1200         if (evsel__is_function_event(evsel))
1201                 evsel->core.attr.exclude_callchain_user = 1;
1202
1203         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1204                 evsel__config_callchain(evsel, opts, callchain);
1205
1206         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1207             !evsel__is_dummy_event(evsel)) {
1208                 attr->sample_regs_intr = opts->sample_intr_regs;
1209                 evsel__set_sample_bit(evsel, REGS_INTR);
1210         }
1211
1212         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1213             !evsel__is_dummy_event(evsel)) {
1214                 attr->sample_regs_user |= opts->sample_user_regs;
1215                 evsel__set_sample_bit(evsel, REGS_USER);
1216         }
1217
1218         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1219                 evsel__set_sample_bit(evsel, CPU);
1220
1221         /*
1222          * When the user explicitly disabled time don't force it here.
1223          */
1224         if (opts->sample_time &&
1225             (!perf_missing_features.sample_id_all &&
1226             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1227              opts->sample_time_set)))
1228                 evsel__set_sample_bit(evsel, TIME);
1229
1230         if (opts->raw_samples && !evsel->no_aux_samples) {
1231                 evsel__set_sample_bit(evsel, TIME);
1232                 evsel__set_sample_bit(evsel, RAW);
1233                 evsel__set_sample_bit(evsel, CPU);
1234         }
1235
1236         if (opts->sample_address)
1237                 evsel__set_sample_bit(evsel, DATA_SRC);
1238
1239         if (opts->sample_phys_addr)
1240                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1241
1242         if (opts->no_buffering) {
1243                 attr->watermark = 0;
1244                 attr->wakeup_events = 1;
1245         }
1246         if (opts->branch_stack && !evsel->no_aux_samples) {
1247                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1248                 attr->branch_sample_type = opts->branch_stack;
1249         }
1250
1251         if (opts->sample_weight)
1252                 arch_evsel__set_sample_weight(evsel);
1253
1254         attr->task     = track;
1255         attr->mmap     = track;
1256         attr->mmap2    = track && !perf_missing_features.mmap2;
1257         attr->comm     = track;
1258         attr->build_id = track && opts->build_id;
1259
1260         /*
1261          * ksymbol is tracked separately with text poke because it needs to be
1262          * system wide and enabled immediately.
1263          */
1264         if (!opts->text_poke)
1265                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1266         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1267
1268         if (opts->record_namespaces)
1269                 attr->namespaces  = track;
1270
1271         if (opts->record_cgroup) {
1272                 attr->cgroup = track && !perf_missing_features.cgroup;
1273                 evsel__set_sample_bit(evsel, CGROUP);
1274         }
1275
1276         if (opts->sample_data_page_size)
1277                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1278
1279         if (opts->sample_code_page_size)
1280                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1281
1282         if (opts->record_switch_events)
1283                 attr->context_switch = track;
1284
1285         if (opts->sample_transaction)
1286                 evsel__set_sample_bit(evsel, TRANSACTION);
1287
1288         if (opts->running_time) {
1289                 evsel->core.attr.read_format |=
1290                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1291                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1292         }
1293
1294         /*
1295          * XXX see the function comment above
1296          *
1297          * Disabling only independent events or group leaders,
1298          * keeping group members enabled.
1299          */
1300         if (evsel__is_group_leader(evsel))
1301                 attr->disabled = 1;
1302
1303         /*
1304          * Setting enable_on_exec for independent events and
1305          * group leaders for traced executed by perf.
1306          */
1307         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1308             !opts->target.initial_delay)
1309                 attr->enable_on_exec = 1;
1310
1311         if (evsel->immediate) {
1312                 attr->disabled = 0;
1313                 attr->enable_on_exec = 0;
1314         }
1315
1316         clockid = opts->clockid;
1317         if (opts->use_clockid) {
1318                 attr->use_clockid = 1;
1319                 attr->clockid = opts->clockid;
1320         }
1321
1322         if (evsel->precise_max)
1323                 attr->precise_ip = 3;
1324
1325         if (opts->all_user) {
1326                 attr->exclude_kernel = 1;
1327                 attr->exclude_user   = 0;
1328         }
1329
1330         if (opts->all_kernel) {
1331                 attr->exclude_kernel = 0;
1332                 attr->exclude_user   = 1;
1333         }
1334
1335         if (evsel->core.own_cpus || evsel->unit)
1336                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1337
1338         /*
1339          * Apply event specific term settings,
1340          * it overloads any global configuration.
1341          */
1342         evsel__apply_config_terms(evsel, opts, track);
1343
1344         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1345
1346         /* The --period option takes the precedence. */
1347         if (opts->period_set) {
1348                 if (opts->period)
1349                         evsel__set_sample_bit(evsel, PERIOD);
1350                 else
1351                         evsel__reset_sample_bit(evsel, PERIOD);
1352         }
1353
1354         /*
1355          * A dummy event never triggers any actual counter and therefore
1356          * cannot be used with branch_stack.
1357          *
1358          * For initial_delay, a dummy event is added implicitly.
1359          * The software event will trigger -EOPNOTSUPP error out,
1360          * if BRANCH_STACK bit is set.
1361          */
1362         if (evsel__is_dummy_event(evsel))
1363                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1364
1365         if (evsel__is_offcpu_event(evsel))
1366                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1367
1368         arch__post_evsel_config(evsel, attr);
1369 }
1370
1371 int evsel__set_filter(struct evsel *evsel, const char *filter)
1372 {
1373         char *new_filter = strdup(filter);
1374
1375         if (new_filter != NULL) {
1376                 free(evsel->filter);
1377                 evsel->filter = new_filter;
1378                 return 0;
1379         }
1380
1381         return -1;
1382 }
1383
1384 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1385 {
1386         char *new_filter;
1387
1388         if (evsel->filter == NULL)
1389                 return evsel__set_filter(evsel, filter);
1390
1391         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1392                 free(evsel->filter);
1393                 evsel->filter = new_filter;
1394                 return 0;
1395         }
1396
1397         return -1;
1398 }
1399
1400 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1401 {
1402         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1403 }
1404
1405 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1406 {
1407         return evsel__append_filter(evsel, "%s,%s", filter);
1408 }
1409
1410 /* Caller has to clear disabled after going through all CPUs. */
1411 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1412 {
1413         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1414 }
1415
1416 int evsel__enable(struct evsel *evsel)
1417 {
1418         int err = perf_evsel__enable(&evsel->core);
1419
1420         if (!err)
1421                 evsel->disabled = false;
1422         return err;
1423 }
1424
1425 /* Caller has to set disabled after going through all CPUs. */
1426 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1427 {
1428         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1429 }
1430
1431 int evsel__disable(struct evsel *evsel)
1432 {
1433         int err = perf_evsel__disable(&evsel->core);
1434         /*
1435          * We mark it disabled here so that tools that disable a event can
1436          * ignore events after they disable it. I.e. the ring buffer may have
1437          * already a few more events queued up before the kernel got the stop
1438          * request.
1439          */
1440         if (!err)
1441                 evsel->disabled = true;
1442
1443         return err;
1444 }
1445
1446 void free_config_terms(struct list_head *config_terms)
1447 {
1448         struct evsel_config_term *term, *h;
1449
1450         list_for_each_entry_safe(term, h, config_terms, list) {
1451                 list_del_init(&term->list);
1452                 if (term->free_str)
1453                         zfree(&term->val.str);
1454                 free(term);
1455         }
1456 }
1457
1458 static void evsel__free_config_terms(struct evsel *evsel)
1459 {
1460         free_config_terms(&evsel->config_terms);
1461 }
1462
1463 void evsel__exit(struct evsel *evsel)
1464 {
1465         assert(list_empty(&evsel->core.node));
1466         assert(evsel->evlist == NULL);
1467         bpf_counter__destroy(evsel);
1468         perf_bpf_filter__destroy(evsel);
1469         evsel__free_counts(evsel);
1470         perf_evsel__free_fd(&evsel->core);
1471         perf_evsel__free_id(&evsel->core);
1472         evsel__free_config_terms(evsel);
1473         cgroup__put(evsel->cgrp);
1474         perf_cpu_map__put(evsel->core.cpus);
1475         perf_cpu_map__put(evsel->core.own_cpus);
1476         perf_thread_map__put(evsel->core.threads);
1477         zfree(&evsel->group_name);
1478         zfree(&evsel->name);
1479         zfree(&evsel->filter);
1480         zfree(&evsel->pmu_name);
1481         zfree(&evsel->group_pmu_name);
1482         zfree(&evsel->unit);
1483         zfree(&evsel->metric_id);
1484         evsel__zero_per_pkg(evsel);
1485         hashmap__free(evsel->per_pkg_mask);
1486         evsel->per_pkg_mask = NULL;
1487         zfree(&evsel->metric_events);
1488         perf_evsel__object.fini(evsel);
1489 }
1490
1491 void evsel__delete(struct evsel *evsel)
1492 {
1493         if (!evsel)
1494                 return;
1495
1496         evsel__exit(evsel);
1497         free(evsel);
1498 }
1499
1500 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1501                            struct perf_counts_values *count)
1502 {
1503         struct perf_counts_values tmp;
1504
1505         if (!evsel->prev_raw_counts)
1506                 return;
1507
1508         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1509         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1510
1511         count->val = count->val - tmp.val;
1512         count->ena = count->ena - tmp.ena;
1513         count->run = count->run - tmp.run;
1514 }
1515
1516 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1517 {
1518         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1519
1520         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1521 }
1522
1523 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1524                              u64 val, u64 ena, u64 run, u64 lost)
1525 {
1526         struct perf_counts_values *count;
1527
1528         count = perf_counts(counter->counts, cpu_map_idx, thread);
1529
1530         count->val    = val;
1531         count->ena    = ena;
1532         count->run    = run;
1533         count->lost   = lost;
1534
1535         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1536 }
1537
1538 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1539 {
1540         u64 read_format = leader->core.attr.read_format;
1541         struct sample_read_value *v;
1542         u64 nr, ena = 0, run = 0, lost = 0;
1543
1544         nr = *data++;
1545
1546         if (nr != (u64) leader->core.nr_members)
1547                 return -EINVAL;
1548
1549         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1550                 ena = *data++;
1551
1552         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1553                 run = *data++;
1554
1555         v = (void *)data;
1556         sample_read_group__for_each(v, nr, read_format) {
1557                 struct evsel *counter;
1558
1559                 counter = evlist__id2evsel(leader->evlist, v->id);
1560                 if (!counter)
1561                         return -EINVAL;
1562
1563                 if (read_format & PERF_FORMAT_LOST)
1564                         lost = v->lost;
1565
1566                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1567         }
1568
1569         return 0;
1570 }
1571
1572 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1573 {
1574         struct perf_stat_evsel *ps = leader->stats;
1575         u64 read_format = leader->core.attr.read_format;
1576         int size = perf_evsel__read_size(&leader->core);
1577         u64 *data = ps->group_data;
1578
1579         if (!(read_format & PERF_FORMAT_ID))
1580                 return -EINVAL;
1581
1582         if (!evsel__is_group_leader(leader))
1583                 return -EINVAL;
1584
1585         if (!data) {
1586                 data = zalloc(size);
1587                 if (!data)
1588                         return -ENOMEM;
1589
1590                 ps->group_data = data;
1591         }
1592
1593         if (FD(leader, cpu_map_idx, thread) < 0)
1594                 return -EINVAL;
1595
1596         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1597                 return -errno;
1598
1599         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1600 }
1601
1602 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1603 {
1604         u64 read_format = evsel->core.attr.read_format;
1605
1606         if (read_format & PERF_FORMAT_GROUP)
1607                 return evsel__read_group(evsel, cpu_map_idx, thread);
1608
1609         return evsel__read_one(evsel, cpu_map_idx, thread);
1610 }
1611
1612 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1613 {
1614         struct perf_counts_values count;
1615         size_t nv = scale ? 3 : 1;
1616
1617         if (FD(evsel, cpu_map_idx, thread) < 0)
1618                 return -EINVAL;
1619
1620         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1621                 return -ENOMEM;
1622
1623         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1624                 return -errno;
1625
1626         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1627         perf_counts_values__scale(&count, scale, NULL);
1628         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1629         return 0;
1630 }
1631
1632 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1633                                   int cpu_map_idx)
1634 {
1635         struct perf_cpu cpu;
1636
1637         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1638         return perf_cpu_map__idx(other->core.cpus, cpu);
1639 }
1640
1641 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1642 {
1643         struct evsel *leader = evsel__leader(evsel);
1644
1645         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1646             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1647                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1648         }
1649
1650         return cpu_map_idx;
1651 }
1652
1653 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1654 {
1655         struct evsel *leader = evsel__leader(evsel);
1656         int fd;
1657
1658         if (evsel__is_group_leader(evsel))
1659                 return -1;
1660
1661         /*
1662          * Leader must be already processed/open,
1663          * if not it's a bug.
1664          */
1665         BUG_ON(!leader->core.fd);
1666
1667         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1668         if (cpu_map_idx == -1)
1669                 return -1;
1670
1671         fd = FD(leader, cpu_map_idx, thread);
1672         BUG_ON(fd == -1 && !leader->skippable);
1673
1674         /*
1675          * When the leader has been skipped, return -2 to distinguish from no
1676          * group leader case.
1677          */
1678         return fd == -1 ? -2 : fd;
1679 }
1680
1681 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1682 {
1683         for (int cpu = 0; cpu < nr_cpus; cpu++)
1684                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1685                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1686 }
1687
1688 static int update_fds(struct evsel *evsel,
1689                       int nr_cpus, int cpu_map_idx,
1690                       int nr_threads, int thread_idx)
1691 {
1692         struct evsel *pos;
1693
1694         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1695                 return -EINVAL;
1696
1697         evlist__for_each_entry(evsel->evlist, pos) {
1698                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1699
1700                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1701
1702                 /*
1703                  * Since fds for next evsel has not been created,
1704                  * there is no need to iterate whole event list.
1705                  */
1706                 if (pos == evsel)
1707                         break;
1708         }
1709         return 0;
1710 }
1711
1712 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1713                                          int nr_cpus, int cpu_map_idx,
1714                                          struct perf_thread_map *threads,
1715                                          int thread, int err)
1716 {
1717         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1718
1719         if (!evsel->ignore_missing_thread)
1720                 return false;
1721
1722         /* The system wide setup does not work with threads. */
1723         if (evsel->core.system_wide)
1724                 return false;
1725
1726         /* The -ESRCH is perf event syscall errno for pid's not found. */
1727         if (err != -ESRCH)
1728                 return false;
1729
1730         /* If there's only one thread, let it fail. */
1731         if (threads->nr == 1)
1732                 return false;
1733
1734         /*
1735          * We should remove fd for missing_thread first
1736          * because thread_map__remove() will decrease threads->nr.
1737          */
1738         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1739                 return false;
1740
1741         if (thread_map__remove(threads, thread))
1742                 return false;
1743
1744         pr_warning("WARNING: Ignored open failure for pid %d\n",
1745                    ignore_pid);
1746         return true;
1747 }
1748
1749 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1750                                 void *priv __maybe_unused)
1751 {
1752         return fprintf(fp, "  %-32s %s\n", name, val);
1753 }
1754
1755 static void display_attr(struct perf_event_attr *attr)
1756 {
1757         if (verbose >= 2 || debug_peo_args) {
1758                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1759                 fprintf(stderr, "perf_event_attr:\n");
1760                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1761                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1762         }
1763 }
1764
1765 bool evsel__precise_ip_fallback(struct evsel *evsel)
1766 {
1767         /* Do not try less precise if not requested. */
1768         if (!evsel->precise_max)
1769                 return false;
1770
1771         /*
1772          * We tried all the precise_ip values, and it's
1773          * still failing, so leave it to standard fallback.
1774          */
1775         if (!evsel->core.attr.precise_ip) {
1776                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1777                 return false;
1778         }
1779
1780         if (!evsel->precise_ip_original)
1781                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1782
1783         evsel->core.attr.precise_ip--;
1784         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1785         display_attr(&evsel->core.attr);
1786         return true;
1787 }
1788
1789 static struct perf_cpu_map *empty_cpu_map;
1790 static struct perf_thread_map *empty_thread_map;
1791
1792 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1793                 struct perf_thread_map *threads)
1794 {
1795         int nthreads = perf_thread_map__nr(threads);
1796
1797         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1798             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1799                 return -EINVAL;
1800
1801         if (cpus == NULL) {
1802                 if (empty_cpu_map == NULL) {
1803                         empty_cpu_map = perf_cpu_map__dummy_new();
1804                         if (empty_cpu_map == NULL)
1805                                 return -ENOMEM;
1806                 }
1807
1808                 cpus = empty_cpu_map;
1809         }
1810
1811         if (threads == NULL) {
1812                 if (empty_thread_map == NULL) {
1813                         empty_thread_map = thread_map__new_by_tid(-1);
1814                         if (empty_thread_map == NULL)
1815                                 return -ENOMEM;
1816                 }
1817
1818                 threads = empty_thread_map;
1819         }
1820
1821         if (evsel->core.fd == NULL &&
1822             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1823                 return -ENOMEM;
1824
1825         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1826         if (evsel->cgrp)
1827                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1828
1829         return 0;
1830 }
1831
1832 static void evsel__disable_missing_features(struct evsel *evsel)
1833 {
1834         if (perf_missing_features.read_lost)
1835                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1836         if (perf_missing_features.weight_struct) {
1837                 evsel__set_sample_bit(evsel, WEIGHT);
1838                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1839         }
1840         if (perf_missing_features.clockid_wrong)
1841                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1842         if (perf_missing_features.clockid) {
1843                 evsel->core.attr.use_clockid = 0;
1844                 evsel->core.attr.clockid = 0;
1845         }
1846         if (perf_missing_features.cloexec)
1847                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1848         if (perf_missing_features.mmap2)
1849                 evsel->core.attr.mmap2 = 0;
1850         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1851                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1852         if (perf_missing_features.lbr_flags)
1853                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1854                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1855         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1856                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1857         if (perf_missing_features.ksymbol)
1858                 evsel->core.attr.ksymbol = 0;
1859         if (perf_missing_features.bpf)
1860                 evsel->core.attr.bpf_event = 0;
1861         if (perf_missing_features.branch_hw_idx)
1862                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1863         if (perf_missing_features.sample_id_all)
1864                 evsel->core.attr.sample_id_all = 0;
1865 }
1866
1867 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1868                         struct perf_thread_map *threads)
1869 {
1870         int err;
1871
1872         err = __evsel__prepare_open(evsel, cpus, threads);
1873         if (err)
1874                 return err;
1875
1876         evsel__disable_missing_features(evsel);
1877
1878         return err;
1879 }
1880
1881 bool evsel__detect_missing_features(struct evsel *evsel)
1882 {
1883         /*
1884          * Must probe features in the order they were added to the
1885          * perf_event_attr interface.
1886          */
1887         if (!perf_missing_features.read_lost &&
1888             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1889                 perf_missing_features.read_lost = true;
1890                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1891                 return true;
1892         } else if (!perf_missing_features.weight_struct &&
1893             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1894                 perf_missing_features.weight_struct = true;
1895                 pr_debug2("switching off weight struct support\n");
1896                 return true;
1897         } else if (!perf_missing_features.code_page_size &&
1898             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1899                 perf_missing_features.code_page_size = true;
1900                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1901                 return false;
1902         } else if (!perf_missing_features.data_page_size &&
1903             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1904                 perf_missing_features.data_page_size = true;
1905                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1906                 return false;
1907         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1908                 perf_missing_features.cgroup = true;
1909                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1910                 return false;
1911         } else if (!perf_missing_features.branch_hw_idx &&
1912             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1913                 perf_missing_features.branch_hw_idx = true;
1914                 pr_debug2("switching off branch HW index support\n");
1915                 return true;
1916         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1917                 perf_missing_features.aux_output = true;
1918                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1919                 return false;
1920         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1921                 perf_missing_features.bpf = true;
1922                 pr_debug2_peo("switching off bpf_event\n");
1923                 return true;
1924         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1925                 perf_missing_features.ksymbol = true;
1926                 pr_debug2_peo("switching off ksymbol\n");
1927                 return true;
1928         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1929                 perf_missing_features.write_backward = true;
1930                 pr_debug2_peo("switching off write_backward\n");
1931                 return false;
1932         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1933                 perf_missing_features.clockid_wrong = true;
1934                 pr_debug2_peo("switching off clockid\n");
1935                 return true;
1936         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1937                 perf_missing_features.clockid = true;
1938                 pr_debug2_peo("switching off use_clockid\n");
1939                 return true;
1940         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1941                 perf_missing_features.cloexec = true;
1942                 pr_debug2_peo("switching off cloexec flag\n");
1943                 return true;
1944         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1945                 perf_missing_features.mmap2 = true;
1946                 pr_debug2_peo("switching off mmap2\n");
1947                 return true;
1948         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1949                 if (evsel->pmu == NULL)
1950                         evsel->pmu = evsel__find_pmu(evsel);
1951
1952                 if (evsel->pmu)
1953                         evsel->pmu->missing_features.exclude_guest = true;
1954                 else {
1955                         /* we cannot find PMU, disable attrs now */
1956                         evsel->core.attr.exclude_host = false;
1957                         evsel->core.attr.exclude_guest = false;
1958                 }
1959
1960                 if (evsel->exclude_GH) {
1961                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1962                         return false;
1963                 }
1964                 if (!perf_missing_features.exclude_guest) {
1965                         perf_missing_features.exclude_guest = true;
1966                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1967                 }
1968                 return true;
1969         } else if (!perf_missing_features.sample_id_all) {
1970                 perf_missing_features.sample_id_all = true;
1971                 pr_debug2_peo("switching off sample_id_all\n");
1972                 return true;
1973         } else if (!perf_missing_features.lbr_flags &&
1974                         (evsel->core.attr.branch_sample_type &
1975                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1976                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1977                 perf_missing_features.lbr_flags = true;
1978                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1979                 return true;
1980         } else if (!perf_missing_features.group_read &&
1981                     evsel->core.attr.inherit &&
1982                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1983                    evsel__is_group_leader(evsel)) {
1984                 perf_missing_features.group_read = true;
1985                 pr_debug2_peo("switching off group read\n");
1986                 return true;
1987         } else {
1988                 return false;
1989         }
1990 }
1991
1992 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
1993 {
1994         int old_errno;
1995         struct rlimit l;
1996
1997         if (*set_rlimit < INCREASED_MAX) {
1998                 old_errno = errno;
1999
2000                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2001                         if (*set_rlimit == NO_CHANGE) {
2002                                 l.rlim_cur = l.rlim_max;
2003                         } else {
2004                                 l.rlim_cur = l.rlim_max + 1000;
2005                                 l.rlim_max = l.rlim_cur;
2006                         }
2007                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2008                                 (*set_rlimit) += 1;
2009                                 errno = old_errno;
2010                                 return true;
2011                         }
2012                 }
2013                 errno = old_errno;
2014         }
2015
2016         return false;
2017 }
2018
2019 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2020                 struct perf_thread_map *threads,
2021                 int start_cpu_map_idx, int end_cpu_map_idx)
2022 {
2023         int idx, thread, nthreads;
2024         int pid = -1, err, old_errno;
2025         enum rlimit_action set_rlimit = NO_CHANGE;
2026
2027         err = __evsel__prepare_open(evsel, cpus, threads);
2028         if (err)
2029                 return err;
2030
2031         if (cpus == NULL)
2032                 cpus = empty_cpu_map;
2033
2034         if (threads == NULL)
2035                 threads = empty_thread_map;
2036
2037         nthreads = perf_thread_map__nr(threads);
2038
2039         if (evsel->cgrp)
2040                 pid = evsel->cgrp->fd;
2041
2042 fallback_missing_features:
2043         evsel__disable_missing_features(evsel);
2044
2045         pr_debug3("Opening: %s\n", evsel__name(evsel));
2046         display_attr(&evsel->core.attr);
2047
2048         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2049
2050                 for (thread = 0; thread < nthreads; thread++) {
2051                         int fd, group_fd;
2052 retry_open:
2053                         if (thread >= nthreads)
2054                                 break;
2055
2056                         if (!evsel->cgrp && !evsel->core.system_wide)
2057                                 pid = perf_thread_map__pid(threads, thread);
2058
2059                         group_fd = get_group_fd(evsel, idx, thread);
2060
2061                         if (group_fd == -2) {
2062                                 pr_debug("broken group leader for %s\n", evsel->name);
2063                                 err = -EINVAL;
2064                                 goto out_close;
2065                         }
2066
2067                         test_attr__ready();
2068
2069                         /* Debug message used by test scripts */
2070                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2071                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2072
2073                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2074                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2075                                                 group_fd, evsel->open_flags);
2076
2077                         FD(evsel, idx, thread) = fd;
2078
2079                         if (fd < 0) {
2080                                 err = -errno;
2081
2082                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2083                                           err);
2084                                 goto try_fallback;
2085                         }
2086
2087                         bpf_counter__install_pe(evsel, idx, fd);
2088
2089                         if (unlikely(test_attr__enabled)) {
2090                                 test_attr__open(&evsel->core.attr, pid,
2091                                                 perf_cpu_map__cpu(cpus, idx),
2092                                                 fd, group_fd, evsel->open_flags);
2093                         }
2094
2095                         /* Debug message used by test scripts */
2096                         pr_debug2_peo(" = %d\n", fd);
2097
2098                         if (evsel->bpf_fd >= 0) {
2099                                 int evt_fd = fd;
2100                                 int bpf_fd = evsel->bpf_fd;
2101
2102                                 err = ioctl(evt_fd,
2103                                             PERF_EVENT_IOC_SET_BPF,
2104                                             bpf_fd);
2105                                 if (err && errno != EEXIST) {
2106                                         pr_err("failed to attach bpf fd %d: %s\n",
2107                                                bpf_fd, strerror(errno));
2108                                         err = -EINVAL;
2109                                         goto out_close;
2110                                 }
2111                         }
2112
2113                         set_rlimit = NO_CHANGE;
2114
2115                         /*
2116                          * If we succeeded but had to kill clockid, fail and
2117                          * have evsel__open_strerror() print us a nice error.
2118                          */
2119                         if (perf_missing_features.clockid ||
2120                             perf_missing_features.clockid_wrong) {
2121                                 err = -EINVAL;
2122                                 goto out_close;
2123                         }
2124                 }
2125         }
2126
2127         return 0;
2128
2129 try_fallback:
2130         if (evsel__precise_ip_fallback(evsel))
2131                 goto retry_open;
2132
2133         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2134                                          idx, threads, thread, err)) {
2135                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2136                 nthreads--;
2137
2138                 /* ... and pretend like nothing have happened. */
2139                 err = 0;
2140                 goto retry_open;
2141         }
2142         /*
2143          * perf stat needs between 5 and 22 fds per CPU. When we run out
2144          * of them try to increase the limits.
2145          */
2146         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2147                 goto retry_open;
2148
2149         if (err != -EINVAL || idx > 0 || thread > 0)
2150                 goto out_close;
2151
2152         if (evsel__detect_missing_features(evsel))
2153                 goto fallback_missing_features;
2154 out_close:
2155         if (err)
2156                 threads->err_thread = thread;
2157
2158         old_errno = errno;
2159         do {
2160                 while (--thread >= 0) {
2161                         if (FD(evsel, idx, thread) >= 0)
2162                                 close(FD(evsel, idx, thread));
2163                         FD(evsel, idx, thread) = -1;
2164                 }
2165                 thread = nthreads;
2166         } while (--idx >= 0);
2167         errno = old_errno;
2168         return err;
2169 }
2170
2171 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2172                 struct perf_thread_map *threads)
2173 {
2174         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2175 }
2176
2177 void evsel__close(struct evsel *evsel)
2178 {
2179         perf_evsel__close(&evsel->core);
2180         perf_evsel__free_id(&evsel->core);
2181 }
2182
2183 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2184 {
2185         if (cpu_map_idx == -1)
2186                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2187
2188         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2189 }
2190
2191 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2192 {
2193         return evsel__open(evsel, NULL, threads);
2194 }
2195
2196 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2197                                        const union perf_event *event,
2198                                        struct perf_sample *sample)
2199 {
2200         u64 type = evsel->core.attr.sample_type;
2201         const __u64 *array = event->sample.array;
2202         bool swapped = evsel->needs_swap;
2203         union u64_swap u;
2204
2205         array += ((event->header.size -
2206                    sizeof(event->header)) / sizeof(u64)) - 1;
2207
2208         if (type & PERF_SAMPLE_IDENTIFIER) {
2209                 sample->id = *array;
2210                 array--;
2211         }
2212
2213         if (type & PERF_SAMPLE_CPU) {
2214                 u.val64 = *array;
2215                 if (swapped) {
2216                         /* undo swap of u64, then swap on individual u32s */
2217                         u.val64 = bswap_64(u.val64);
2218                         u.val32[0] = bswap_32(u.val32[0]);
2219                 }
2220
2221                 sample->cpu = u.val32[0];
2222                 array--;
2223         }
2224
2225         if (type & PERF_SAMPLE_STREAM_ID) {
2226                 sample->stream_id = *array;
2227                 array--;
2228         }
2229
2230         if (type & PERF_SAMPLE_ID) {
2231                 sample->id = *array;
2232                 array--;
2233         }
2234
2235         if (type & PERF_SAMPLE_TIME) {
2236                 sample->time = *array;
2237                 array--;
2238         }
2239
2240         if (type & PERF_SAMPLE_TID) {
2241                 u.val64 = *array;
2242                 if (swapped) {
2243                         /* undo swap of u64, then swap on individual u32s */
2244                         u.val64 = bswap_64(u.val64);
2245                         u.val32[0] = bswap_32(u.val32[0]);
2246                         u.val32[1] = bswap_32(u.val32[1]);
2247                 }
2248
2249                 sample->pid = u.val32[0];
2250                 sample->tid = u.val32[1];
2251                 array--;
2252         }
2253
2254         return 0;
2255 }
2256
2257 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2258                             u64 size)
2259 {
2260         return size > max_size || offset + size > endp;
2261 }
2262
2263 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2264         do {                                                            \
2265                 if (overflow(endp, (max_size), (offset), (size)))       \
2266                         return -EFAULT;                                 \
2267         } while (0)
2268
2269 #define OVERFLOW_CHECK_u64(offset) \
2270         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2271
2272 static int
2273 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2274 {
2275         /*
2276          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2277          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2278          * check the format does not go past the end of the event.
2279          */
2280         if (sample_size + sizeof(event->header) > event->header.size)
2281                 return -EFAULT;
2282
2283         return 0;
2284 }
2285
2286 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2287                                           const __u64 *array,
2288                                           u64 type __maybe_unused)
2289 {
2290         data->weight = *array;
2291 }
2292
2293 u64 evsel__bitfield_swap_branch_flags(u64 value)
2294 {
2295         u64 new_val = 0;
2296
2297         /*
2298          * branch_flags
2299          * union {
2300          *      u64 values;
2301          *      struct {
2302          *              mispred:1       //target mispredicted
2303          *              predicted:1     //target predicted
2304          *              in_tx:1         //in transaction
2305          *              abort:1         //transaction abort
2306          *              cycles:16       //cycle count to last branch
2307          *              type:4          //branch type
2308          *              spec:2          //branch speculation info
2309          *              new_type:4      //additional branch type
2310          *              priv:3          //privilege level
2311          *              reserved:31
2312          *      }
2313          * }
2314          *
2315          * Avoid bswap64() the entire branch_flag.value,
2316          * as it has variable bit-field sizes. Instead the
2317          * macro takes the bit-field position/size,
2318          * swaps it based on the host endianness.
2319          */
2320         if (host_is_bigendian()) {
2321                 new_val = bitfield_swap(value, 0, 1);
2322                 new_val |= bitfield_swap(value, 1, 1);
2323                 new_val |= bitfield_swap(value, 2, 1);
2324                 new_val |= bitfield_swap(value, 3, 1);
2325                 new_val |= bitfield_swap(value, 4, 16);
2326                 new_val |= bitfield_swap(value, 20, 4);
2327                 new_val |= bitfield_swap(value, 24, 2);
2328                 new_val |= bitfield_swap(value, 26, 4);
2329                 new_val |= bitfield_swap(value, 30, 3);
2330                 new_val |= bitfield_swap(value, 33, 31);
2331         } else {
2332                 new_val = bitfield_swap(value, 63, 1);
2333                 new_val |= bitfield_swap(value, 62, 1);
2334                 new_val |= bitfield_swap(value, 61, 1);
2335                 new_val |= bitfield_swap(value, 60, 1);
2336                 new_val |= bitfield_swap(value, 44, 16);
2337                 new_val |= bitfield_swap(value, 40, 4);
2338                 new_val |= bitfield_swap(value, 38, 2);
2339                 new_val |= bitfield_swap(value, 34, 4);
2340                 new_val |= bitfield_swap(value, 31, 3);
2341                 new_val |= bitfield_swap(value, 0, 31);
2342         }
2343
2344         return new_val;
2345 }
2346
2347 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2348                         struct perf_sample *data)
2349 {
2350         u64 type = evsel->core.attr.sample_type;
2351         bool swapped = evsel->needs_swap;
2352         const __u64 *array;
2353         u16 max_size = event->header.size;
2354         const void *endp = (void *)event + max_size;
2355         u64 sz;
2356
2357         /*
2358          * used for cross-endian analysis. See git commit 65014ab3
2359          * for why this goofiness is needed.
2360          */
2361         union u64_swap u;
2362
2363         memset(data, 0, sizeof(*data));
2364         data->cpu = data->pid = data->tid = -1;
2365         data->stream_id = data->id = data->time = -1ULL;
2366         data->period = evsel->core.attr.sample_period;
2367         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2368         data->misc    = event->header.misc;
2369         data->id = -1ULL;
2370         data->data_src = PERF_MEM_DATA_SRC_NONE;
2371         data->vcpu = -1;
2372
2373         if (event->header.type != PERF_RECORD_SAMPLE) {
2374                 if (!evsel->core.attr.sample_id_all)
2375                         return 0;
2376                 return perf_evsel__parse_id_sample(evsel, event, data);
2377         }
2378
2379         array = event->sample.array;
2380
2381         if (perf_event__check_size(event, evsel->sample_size))
2382                 return -EFAULT;
2383
2384         if (type & PERF_SAMPLE_IDENTIFIER) {
2385                 data->id = *array;
2386                 array++;
2387         }
2388
2389         if (type & PERF_SAMPLE_IP) {
2390                 data->ip = *array;
2391                 array++;
2392         }
2393
2394         if (type & PERF_SAMPLE_TID) {
2395                 u.val64 = *array;
2396                 if (swapped) {
2397                         /* undo swap of u64, then swap on individual u32s */
2398                         u.val64 = bswap_64(u.val64);
2399                         u.val32[0] = bswap_32(u.val32[0]);
2400                         u.val32[1] = bswap_32(u.val32[1]);
2401                 }
2402
2403                 data->pid = u.val32[0];
2404                 data->tid = u.val32[1];
2405                 array++;
2406         }
2407
2408         if (type & PERF_SAMPLE_TIME) {
2409                 data->time = *array;
2410                 array++;
2411         }
2412
2413         if (type & PERF_SAMPLE_ADDR) {
2414                 data->addr = *array;
2415                 array++;
2416         }
2417
2418         if (type & PERF_SAMPLE_ID) {
2419                 data->id = *array;
2420                 array++;
2421         }
2422
2423         if (type & PERF_SAMPLE_STREAM_ID) {
2424                 data->stream_id = *array;
2425                 array++;
2426         }
2427
2428         if (type & PERF_SAMPLE_CPU) {
2429
2430                 u.val64 = *array;
2431                 if (swapped) {
2432                         /* undo swap of u64, then swap on individual u32s */
2433                         u.val64 = bswap_64(u.val64);
2434                         u.val32[0] = bswap_32(u.val32[0]);
2435                 }
2436
2437                 data->cpu = u.val32[0];
2438                 array++;
2439         }
2440
2441         if (type & PERF_SAMPLE_PERIOD) {
2442                 data->period = *array;
2443                 array++;
2444         }
2445
2446         if (type & PERF_SAMPLE_READ) {
2447                 u64 read_format = evsel->core.attr.read_format;
2448
2449                 OVERFLOW_CHECK_u64(array);
2450                 if (read_format & PERF_FORMAT_GROUP)
2451                         data->read.group.nr = *array;
2452                 else
2453                         data->read.one.value = *array;
2454
2455                 array++;
2456
2457                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2458                         OVERFLOW_CHECK_u64(array);
2459                         data->read.time_enabled = *array;
2460                         array++;
2461                 }
2462
2463                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2464                         OVERFLOW_CHECK_u64(array);
2465                         data->read.time_running = *array;
2466                         array++;
2467                 }
2468
2469                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2470                 if (read_format & PERF_FORMAT_GROUP) {
2471                         const u64 max_group_nr = UINT64_MAX /
2472                                         sizeof(struct sample_read_value);
2473
2474                         if (data->read.group.nr > max_group_nr)
2475                                 return -EFAULT;
2476
2477                         sz = data->read.group.nr * sample_read_value_size(read_format);
2478                         OVERFLOW_CHECK(array, sz, max_size);
2479                         data->read.group.values =
2480                                         (struct sample_read_value *)array;
2481                         array = (void *)array + sz;
2482                 } else {
2483                         OVERFLOW_CHECK_u64(array);
2484                         data->read.one.id = *array;
2485                         array++;
2486
2487                         if (read_format & PERF_FORMAT_LOST) {
2488                                 OVERFLOW_CHECK_u64(array);
2489                                 data->read.one.lost = *array;
2490                                 array++;
2491                         }
2492                 }
2493         }
2494
2495         if (type & PERF_SAMPLE_CALLCHAIN) {
2496                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2497
2498                 OVERFLOW_CHECK_u64(array);
2499                 data->callchain = (struct ip_callchain *)array++;
2500                 if (data->callchain->nr > max_callchain_nr)
2501                         return -EFAULT;
2502                 sz = data->callchain->nr * sizeof(u64);
2503                 OVERFLOW_CHECK(array, sz, max_size);
2504                 array = (void *)array + sz;
2505         }
2506
2507         if (type & PERF_SAMPLE_RAW) {
2508                 OVERFLOW_CHECK_u64(array);
2509                 u.val64 = *array;
2510
2511                 /*
2512                  * Undo swap of u64, then swap on individual u32s,
2513                  * get the size of the raw area and undo all of the
2514                  * swap. The pevent interface handles endianness by
2515                  * itself.
2516                  */
2517                 if (swapped) {
2518                         u.val64 = bswap_64(u.val64);
2519                         u.val32[0] = bswap_32(u.val32[0]);
2520                         u.val32[1] = bswap_32(u.val32[1]);
2521                 }
2522                 data->raw_size = u.val32[0];
2523
2524                 /*
2525                  * The raw data is aligned on 64bits including the
2526                  * u32 size, so it's safe to use mem_bswap_64.
2527                  */
2528                 if (swapped)
2529                         mem_bswap_64((void *) array, data->raw_size);
2530
2531                 array = (void *)array + sizeof(u32);
2532
2533                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2534                 data->raw_data = (void *)array;
2535                 array = (void *)array + data->raw_size;
2536         }
2537
2538         if (type & PERF_SAMPLE_BRANCH_STACK) {
2539                 const u64 max_branch_nr = UINT64_MAX /
2540                                           sizeof(struct branch_entry);
2541                 struct branch_entry *e;
2542                 unsigned int i;
2543
2544                 OVERFLOW_CHECK_u64(array);
2545                 data->branch_stack = (struct branch_stack *)array++;
2546
2547                 if (data->branch_stack->nr > max_branch_nr)
2548                         return -EFAULT;
2549
2550                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2551                 if (evsel__has_branch_hw_idx(evsel)) {
2552                         sz += sizeof(u64);
2553                         e = &data->branch_stack->entries[0];
2554                 } else {
2555                         data->no_hw_idx = true;
2556                         /*
2557                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2558                          * only nr and entries[] will be output by kernel.
2559                          */
2560                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2561                 }
2562
2563                 if (swapped) {
2564                         /*
2565                          * struct branch_flag does not have endian
2566                          * specific bit field definition. And bswap
2567                          * will not resolve the issue, since these
2568                          * are bit fields.
2569                          *
2570                          * evsel__bitfield_swap_branch_flags() uses a
2571                          * bitfield_swap macro to swap the bit position
2572                          * based on the host endians.
2573                          */
2574                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2575                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2576                 }
2577
2578                 OVERFLOW_CHECK(array, sz, max_size);
2579                 array = (void *)array + sz;
2580         }
2581
2582         if (type & PERF_SAMPLE_REGS_USER) {
2583                 OVERFLOW_CHECK_u64(array);
2584                 data->user_regs.abi = *array;
2585                 array++;
2586
2587                 if (data->user_regs.abi) {
2588                         u64 mask = evsel->core.attr.sample_regs_user;
2589
2590                         sz = hweight64(mask) * sizeof(u64);
2591                         OVERFLOW_CHECK(array, sz, max_size);
2592                         data->user_regs.mask = mask;
2593                         data->user_regs.regs = (u64 *)array;
2594                         array = (void *)array + sz;
2595                 }
2596         }
2597
2598         if (type & PERF_SAMPLE_STACK_USER) {
2599                 OVERFLOW_CHECK_u64(array);
2600                 sz = *array++;
2601
2602                 data->user_stack.offset = ((char *)(array - 1)
2603                                           - (char *) event);
2604
2605                 if (!sz) {
2606                         data->user_stack.size = 0;
2607                 } else {
2608                         OVERFLOW_CHECK(array, sz, max_size);
2609                         data->user_stack.data = (char *)array;
2610                         array = (void *)array + sz;
2611                         OVERFLOW_CHECK_u64(array);
2612                         data->user_stack.size = *array++;
2613                         if (WARN_ONCE(data->user_stack.size > sz,
2614                                       "user stack dump failure\n"))
2615                                 return -EFAULT;
2616                 }
2617         }
2618
2619         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2620                 OVERFLOW_CHECK_u64(array);
2621                 arch_perf_parse_sample_weight(data, array, type);
2622                 array++;
2623         }
2624
2625         if (type & PERF_SAMPLE_DATA_SRC) {
2626                 OVERFLOW_CHECK_u64(array);
2627                 data->data_src = *array;
2628                 array++;
2629         }
2630
2631         if (type & PERF_SAMPLE_TRANSACTION) {
2632                 OVERFLOW_CHECK_u64(array);
2633                 data->transaction = *array;
2634                 array++;
2635         }
2636
2637         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2638         if (type & PERF_SAMPLE_REGS_INTR) {
2639                 OVERFLOW_CHECK_u64(array);
2640                 data->intr_regs.abi = *array;
2641                 array++;
2642
2643                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2644                         u64 mask = evsel->core.attr.sample_regs_intr;
2645
2646                         sz = hweight64(mask) * sizeof(u64);
2647                         OVERFLOW_CHECK(array, sz, max_size);
2648                         data->intr_regs.mask = mask;
2649                         data->intr_regs.regs = (u64 *)array;
2650                         array = (void *)array + sz;
2651                 }
2652         }
2653
2654         data->phys_addr = 0;
2655         if (type & PERF_SAMPLE_PHYS_ADDR) {
2656                 data->phys_addr = *array;
2657                 array++;
2658         }
2659
2660         data->cgroup = 0;
2661         if (type & PERF_SAMPLE_CGROUP) {
2662                 data->cgroup = *array;
2663                 array++;
2664         }
2665
2666         data->data_page_size = 0;
2667         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2668                 data->data_page_size = *array;
2669                 array++;
2670         }
2671
2672         data->code_page_size = 0;
2673         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2674                 data->code_page_size = *array;
2675                 array++;
2676         }
2677
2678         if (type & PERF_SAMPLE_AUX) {
2679                 OVERFLOW_CHECK_u64(array);
2680                 sz = *array++;
2681
2682                 OVERFLOW_CHECK(array, sz, max_size);
2683                 /* Undo swap of data */
2684                 if (swapped)
2685                         mem_bswap_64((char *)array, sz);
2686                 data->aux_sample.size = sz;
2687                 data->aux_sample.data = (char *)array;
2688                 array = (void *)array + sz;
2689         }
2690
2691         return 0;
2692 }
2693
2694 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2695                                   u64 *timestamp)
2696 {
2697         u64 type = evsel->core.attr.sample_type;
2698         const __u64 *array;
2699
2700         if (!(type & PERF_SAMPLE_TIME))
2701                 return -1;
2702
2703         if (event->header.type != PERF_RECORD_SAMPLE) {
2704                 struct perf_sample data = {
2705                         .time = -1ULL,
2706                 };
2707
2708                 if (!evsel->core.attr.sample_id_all)
2709                         return -1;
2710                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2711                         return -1;
2712
2713                 *timestamp = data.time;
2714                 return 0;
2715         }
2716
2717         array = event->sample.array;
2718
2719         if (perf_event__check_size(event, evsel->sample_size))
2720                 return -EFAULT;
2721
2722         if (type & PERF_SAMPLE_IDENTIFIER)
2723                 array++;
2724
2725         if (type & PERF_SAMPLE_IP)
2726                 array++;
2727
2728         if (type & PERF_SAMPLE_TID)
2729                 array++;
2730
2731         if (type & PERF_SAMPLE_TIME)
2732                 *timestamp = *array;
2733
2734         return 0;
2735 }
2736
2737 u16 evsel__id_hdr_size(struct evsel *evsel)
2738 {
2739         u64 sample_type = evsel->core.attr.sample_type;
2740         u16 size = 0;
2741
2742         if (sample_type & PERF_SAMPLE_TID)
2743                 size += sizeof(u64);
2744
2745         if (sample_type & PERF_SAMPLE_TIME)
2746                 size += sizeof(u64);
2747
2748         if (sample_type & PERF_SAMPLE_ID)
2749                 size += sizeof(u64);
2750
2751         if (sample_type & PERF_SAMPLE_STREAM_ID)
2752                 size += sizeof(u64);
2753
2754         if (sample_type & PERF_SAMPLE_CPU)
2755                 size += sizeof(u64);
2756
2757         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2758                 size += sizeof(u64);
2759
2760         return size;
2761 }
2762
2763 #ifdef HAVE_LIBTRACEEVENT
2764 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2765 {
2766         return tep_find_field(evsel->tp_format, name);
2767 }
2768
2769 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2770 {
2771         struct tep_format_field *field = evsel__field(evsel, name);
2772         int offset;
2773
2774         if (!field)
2775                 return NULL;
2776
2777         offset = field->offset;
2778
2779         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2780                 offset = *(int *)(sample->raw_data + field->offset);
2781                 offset &= 0xffff;
2782                 if (tep_field_is_relative(field->flags))
2783                         offset += field->offset + field->size;
2784         }
2785
2786         return sample->raw_data + offset;
2787 }
2788
2789 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2790                          bool needs_swap)
2791 {
2792         u64 value;
2793         void *ptr = sample->raw_data + field->offset;
2794
2795         switch (field->size) {
2796         case 1:
2797                 return *(u8 *)ptr;
2798         case 2:
2799                 value = *(u16 *)ptr;
2800                 break;
2801         case 4:
2802                 value = *(u32 *)ptr;
2803                 break;
2804         case 8:
2805                 memcpy(&value, ptr, sizeof(u64));
2806                 break;
2807         default:
2808                 return 0;
2809         }
2810
2811         if (!needs_swap)
2812                 return value;
2813
2814         switch (field->size) {
2815         case 2:
2816                 return bswap_16(value);
2817         case 4:
2818                 return bswap_32(value);
2819         case 8:
2820                 return bswap_64(value);
2821         default:
2822                 return 0;
2823         }
2824
2825         return 0;
2826 }
2827
2828 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2829 {
2830         struct tep_format_field *field = evsel__field(evsel, name);
2831
2832         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2833 }
2834 #endif
2835
2836 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2837 {
2838         int paranoid;
2839
2840         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2841             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2842             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2843                 /*
2844                  * If it's cycles then fall back to hrtimer based
2845                  * cpu-clock-tick sw counter, which is always available even if
2846                  * no PMU support.
2847                  *
2848                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2849                  * b0a873e).
2850                  */
2851                 scnprintf(msg, msgsize, "%s",
2852 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2853
2854                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2855                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2856
2857                 zfree(&evsel->name);
2858                 return true;
2859         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2860                    (paranoid = perf_event_paranoid()) > 1) {
2861                 const char *name = evsel__name(evsel);
2862                 char *new_name;
2863                 const char *sep = ":";
2864
2865                 /* If event has exclude user then don't exclude kernel. */
2866                 if (evsel->core.attr.exclude_user)
2867                         return false;
2868
2869                 /* Is there already the separator in the name. */
2870                 if (strchr(name, '/') ||
2871                     (strchr(name, ':') && !evsel->is_libpfm_event))
2872                         sep = "";
2873
2874                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2875                         return false;
2876
2877                 free(evsel->name);
2878                 evsel->name = new_name;
2879                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2880                           "to fall back to excluding kernel and hypervisor "
2881                           " samples", paranoid);
2882                 evsel->core.attr.exclude_kernel = 1;
2883                 evsel->core.attr.exclude_hv     = 1;
2884
2885                 return true;
2886         }
2887
2888         return false;
2889 }
2890
2891 static bool find_process(const char *name)
2892 {
2893         size_t len = strlen(name);
2894         DIR *dir;
2895         struct dirent *d;
2896         int ret = -1;
2897
2898         dir = opendir(procfs__mountpoint());
2899         if (!dir)
2900                 return false;
2901
2902         /* Walk through the directory. */
2903         while (ret && (d = readdir(dir)) != NULL) {
2904                 char path[PATH_MAX];
2905                 char *data;
2906                 size_t size;
2907
2908                 if ((d->d_type != DT_DIR) ||
2909                      !strcmp(".", d->d_name) ||
2910                      !strcmp("..", d->d_name))
2911                         continue;
2912
2913                 scnprintf(path, sizeof(path), "%s/%s/comm",
2914                           procfs__mountpoint(), d->d_name);
2915
2916                 if (filename__read_str(path, &data, &size))
2917                         continue;
2918
2919                 ret = strncmp(name, data, len);
2920                 free(data);
2921         }
2922
2923         closedir(dir);
2924         return ret ? false : true;
2925 }
2926
2927 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused,
2928                                      char *msg __maybe_unused,
2929                                      size_t size __maybe_unused)
2930 {
2931         return 0;
2932 }
2933
2934 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2935                          int err, char *msg, size_t size)
2936 {
2937         char sbuf[STRERR_BUFSIZE];
2938         int printed = 0, enforced = 0;
2939         int ret;
2940
2941         switch (err) {
2942         case EPERM:
2943         case EACCES:
2944                 printed += scnprintf(msg + printed, size - printed,
2945                         "Access to performance monitoring and observability operations is limited.\n");
2946
2947                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2948                         if (enforced) {
2949                                 printed += scnprintf(msg + printed, size - printed,
2950                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2951                                         "monitoring and observability operations. Inspect system audit records for\n"
2952                                         "more perf_event access control information and adjusting the policy.\n");
2953                         }
2954                 }
2955
2956                 if (err == EPERM)
2957                         printed += scnprintf(msg, size,
2958                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2959
2960                 return scnprintf(msg + printed, size - printed,
2961                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2962                  "access to performance monitoring and observability operations for processes\n"
2963                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2964                  "More information can be found at 'Perf events and tool security' document:\n"
2965                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2966                  "perf_event_paranoid setting is %d:\n"
2967                  "  -1: Allow use of (almost) all events by all users\n"
2968                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2969                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2970                  ">= 1: Disallow CPU event access\n"
2971                  ">= 2: Disallow kernel profiling\n"
2972                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2973                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2974                  perf_event_paranoid());
2975         case ENOENT:
2976                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2977         case EMFILE:
2978                 return scnprintf(msg, size, "%s",
2979                          "Too many events are opened.\n"
2980                          "Probably the maximum number of open file descriptors has been reached.\n"
2981                          "Hint: Try again after reducing the number of events.\n"
2982                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2983         case ENOMEM:
2984                 if (evsel__has_callchain(evsel) &&
2985                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2986                         return scnprintf(msg, size,
2987                                          "Not enough memory to setup event with callchain.\n"
2988                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2989                                          "Hint: Current value: %d", sysctl__max_stack());
2990                 break;
2991         case ENODEV:
2992                 if (target->cpu_list)
2993                         return scnprintf(msg, size, "%s",
2994          "No such device - did you specify an out-of-range profile CPU?");
2995                 break;
2996         case EOPNOTSUPP:
2997                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
2998                         return scnprintf(msg, size,
2999         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3000                                          evsel__name(evsel));
3001                 if (evsel->core.attr.aux_output)
3002                         return scnprintf(msg, size,
3003         "%s: PMU Hardware doesn't support 'aux_output' feature",
3004                                          evsel__name(evsel));
3005                 if (evsel->core.attr.sample_period != 0)
3006                         return scnprintf(msg, size,
3007         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3008                                          evsel__name(evsel));
3009                 if (evsel->core.attr.precise_ip)
3010                         return scnprintf(msg, size, "%s",
3011         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3012 #if defined(__i386__) || defined(__x86_64__)
3013                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3014                         return scnprintf(msg, size, "%s",
3015         "No hardware sampling interrupt available.\n");
3016 #endif
3017                 break;
3018         case EBUSY:
3019                 if (find_process("oprofiled"))
3020                         return scnprintf(msg, size,
3021         "The PMU counters are busy/taken by another profiler.\n"
3022         "We found oprofile daemon running, please stop it and try again.");
3023                 break;
3024         case EINVAL:
3025                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3026                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3027                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3028                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3029                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3030                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3031                 if (perf_missing_features.clockid)
3032                         return scnprintf(msg, size, "clockid feature not supported.");
3033                 if (perf_missing_features.clockid_wrong)
3034                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3035                 if (perf_missing_features.aux_output)
3036                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3037                 if (!target__has_cpu(target))
3038                         return scnprintf(msg, size,
3039         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3040                                         evsel__name(evsel));
3041
3042                 break;
3043         case ENODATA:
3044                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3045                                  "Please add an auxiliary event in front of the load latency event.");
3046         default:
3047                 break;
3048         }
3049
3050         ret = arch_evsel__open_strerror(evsel, msg, size);
3051         if (ret)
3052                 return ret;
3053
3054         return scnprintf(msg, size,
3055         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3056         "/bin/dmesg | grep -i perf may provide additional information.\n",
3057                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3058 }
3059
3060 struct perf_env *evsel__env(struct evsel *evsel)
3061 {
3062         if (evsel && evsel->evlist && evsel->evlist->env)
3063                 return evsel->evlist->env;
3064         return &perf_env;
3065 }
3066
3067 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3068 {
3069         int cpu_map_idx, thread;
3070
3071         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3072                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3073                      thread++) {
3074                         int fd = FD(evsel, cpu_map_idx, thread);
3075
3076                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3077                                                    cpu_map_idx, thread, fd) < 0)
3078                                 return -1;
3079                 }
3080         }
3081
3082         return 0;
3083 }
3084
3085 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3086 {
3087         struct perf_cpu_map *cpus = evsel->core.cpus;
3088         struct perf_thread_map *threads = evsel->core.threads;
3089
3090         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3091                 return -ENOMEM;
3092
3093         return store_evsel_ids(evsel, evlist);
3094 }
3095
3096 void evsel__zero_per_pkg(struct evsel *evsel)
3097 {
3098         struct hashmap_entry *cur;
3099         size_t bkt;
3100
3101         if (evsel->per_pkg_mask) {
3102                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3103                         zfree(&cur->pkey);
3104
3105                 hashmap__clear(evsel->per_pkg_mask);
3106         }
3107 }
3108
3109 /**
3110  * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this
3111  *                    will be false on hybrid systems for hardware and legacy
3112  *                    cache events.
3113  */
3114 bool evsel__is_hybrid(const struct evsel *evsel)
3115 {
3116         if (perf_pmus__num_core_pmus() == 1)
3117                 return false;
3118
3119         return evsel->core.is_pmu_core;
3120 }
3121
3122 struct evsel *evsel__leader(const struct evsel *evsel)
3123 {
3124         return container_of(evsel->core.leader, struct evsel, core);
3125 }
3126
3127 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3128 {
3129         return evsel->core.leader == &leader->core;
3130 }
3131
3132 bool evsel__is_leader(struct evsel *evsel)
3133 {
3134         return evsel__has_leader(evsel, evsel);
3135 }
3136
3137 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3138 {
3139         evsel->core.leader = &leader->core;
3140 }
3141
3142 int evsel__source_count(const struct evsel *evsel)
3143 {
3144         struct evsel *pos;
3145         int count = 0;
3146
3147         evlist__for_each_entry(evsel->evlist, pos) {
3148                 if (pos->metric_leader == evsel)
3149                         count++;
3150         }
3151         return count;
3152 }
3153
3154 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3155 {
3156         return false;
3157 }
3158
3159 /*
3160  * Remove an event from a given group (leader).
3161  * Some events, e.g., perf metrics Topdown events,
3162  * must always be grouped. Ignore the events.
3163  */
3164 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3165 {
3166         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3167                 evsel__set_leader(evsel, evsel);
3168                 evsel->core.nr_members = 0;
3169                 leader->core.nr_members--;
3170         }
3171 }