tools/perf/util/evsel.c

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