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        = "";
 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                 /*
 280                  * The evsel->unit points to static alias->unit
 281                  * so it's ok to use static string in here.
 282                  */
 283                 static const char *unit = "msec";
 284
 285                 evsel->unit = unit;
 286                 evsel->scale = 1e-6;
 287         }
 288
 289         return evsel;
 290 }
 291
 292 static bool perf_event_can_profile_kernel(void)
 293 {
 294         return perf_event_paranoid_check(1);
 295 }
 296
 297 struct evsel *evsel__new_cycles(bool precise, __u32 type, __u64 config)
 298 {
 299         struct perf_event_attr attr = {
 300                 .type   = type,
 301                 .config = config,
 302                 .exclude_kernel = !perf_event_can_profile_kernel(),
 303         };
 304         struct evsel *evsel;
 305
 306         event_attr_init(&attr);
 307
 308         if (!precise)
 309                 goto new_event;
 310
 311         /*
 312          * Now let the usual logic to set up the perf_event_attr defaults
 313          * to kick in when we return and before perf_evsel__open() is called.
 314          */
 315 new_event:
 316         evsel = evsel__new(&attr);
 317         if (evsel == NULL)
 318                 goto out;
 319
 320         evsel->precise_max = true;
 321
 322         /* use asprintf() because free(evsel) assumes name is allocated */
 323         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 324                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 325                      attr.exclude_kernel ? "u" : "",
 326                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 327                 goto error_free;
 328 out:
 329         return evsel;
 330 error_free:
 331         evsel__delete(evsel);
 332         evsel = NULL;
 333         goto out;
 334 }
 335
 336 int copy_config_terms(struct list_head *dst, struct list_head *src)
 337 {
 338         struct evsel_config_term *pos, *tmp;
 339
 340         list_for_each_entry(pos, src, list) {
 341                 tmp = malloc(sizeof(*tmp));
 342                 if (tmp == NULL)
 343                         return -ENOMEM;
 344
 345                 *tmp = *pos;
 346                 if (tmp->free_str) {
 347                         tmp->val.str = strdup(pos->val.str);
 348                         if (tmp->val.str == NULL) {
 349                                 free(tmp);
 350                                 return -ENOMEM;
 351                         }
 352                 }
 353                 list_add_tail(&tmp->list, dst);
 354         }
 355         return 0;
 356 }
 357
 358 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 359 {
 360         return copy_config_terms(&dst->config_terms, &src->config_terms);
 361 }
 362
 363 /**
 364  * evsel__clone - create a new evsel copied from @orig
 365  * @orig: original evsel
 366  *
 367  * The assumption is that @orig is not configured nor opened yet.
 368  * So we only care about the attributes that can be set while it's parsed.
 369  */
 370 struct evsel *evsel__clone(struct evsel *orig)
 371 {
 372         struct evsel *evsel;
 373
 374         BUG_ON(orig->core.fd);
 375         BUG_ON(orig->counts);
 376         BUG_ON(orig->priv);
 377         BUG_ON(orig->per_pkg_mask);
 378
 379         /* cannot handle BPF objects for now */
 380         if (orig->bpf_obj)
 381                 return NULL;
 382
 383         evsel = evsel__new(&orig->core.attr);
 384         if (evsel == NULL)
 385                 return NULL;
 386
 387         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 388         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 389         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 390         evsel->core.nr_members = orig->core.nr_members;
 391         evsel->core.system_wide = orig->core.system_wide;
 392
 393         if (orig->name) {
 394                 evsel->name = strdup(orig->name);
 395                 if (evsel->name == NULL)
 396                         goto out_err;
 397         }
 398         if (orig->group_name) {
 399                 evsel->group_name = strdup(orig->group_name);
 400                 if (evsel->group_name == NULL)
 401                         goto out_err;
 402         }
 403         if (orig->pmu_name) {
 404                 evsel->pmu_name = strdup(orig->pmu_name);
 405                 if (evsel->pmu_name == NULL)
 406                         goto out_err;
 407         }
 408         if (orig->filter) {
 409                 evsel->filter = strdup(orig->filter);
 410                 if (evsel->filter == NULL)
 411                         goto out_err;
 412         }
 413         evsel->cgrp = cgroup__get(orig->cgrp);
 414         evsel->tp_format = orig->tp_format;
 415         evsel->handler = orig->handler;
 416         evsel->core.leader = orig->core.leader;
 417
 418         evsel->max_events = orig->max_events;
 419         evsel->tool_event = orig->tool_event;
 420         evsel->unit = orig->unit;
 421         evsel->scale = orig->scale;
 422         evsel->snapshot = orig->snapshot;
 423         evsel->per_pkg = orig->per_pkg;
 424         evsel->percore = orig->percore;
 425         evsel->precise_max = orig->precise_max;
 426         evsel->use_uncore_alias = orig->use_uncore_alias;
 427         evsel->is_libpfm_event = orig->is_libpfm_event;
 428
 429         evsel->exclude_GH = orig->exclude_GH;
 430         evsel->sample_read = orig->sample_read;
 431         evsel->auto_merge_stats = orig->auto_merge_stats;
 432         evsel->collect_stat = orig->collect_stat;
 433         evsel->weak_group = orig->weak_group;
 434         evsel->use_config_name = orig->use_config_name;
 435
 436         if (evsel__copy_config_terms(evsel, orig) < 0)
 437                 goto out_err;
 438
 439         return evsel;
 440
 441 out_err:
 442         evsel__delete(evsel);
 443         return NULL;
 444 }
 445
 446 /*
 447  * Returns pointer with encoded error via <linux/err.h> interface.
 448  */
 449 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 450 {
 451         struct evsel *evsel = zalloc(perf_evsel__object.size);
 452         int err = -ENOMEM;
 453
 454         if (evsel == NULL) {
 455                 goto out_err;
 456         } else {
 457                 struct perf_event_attr attr = {
 458                         .type          = PERF_TYPE_TRACEPOINT,
 459                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 460                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 461                 };
 462
 463                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 464                         goto out_free;
 465
 466                 evsel->tp_format = trace_event__tp_format(sys, name);
 467                 if (IS_ERR(evsel->tp_format)) {
 468                         err = PTR_ERR(evsel->tp_format);
 469                         goto out_free;
 470                 }
 471
 472                 event_attr_init(&attr);
 473                 attr.config = evsel->tp_format->id;
 474                 attr.sample_period = 1;
 475                 evsel__init(evsel, &attr, idx);
 476         }
 477
 478         return evsel;
 479
 480 out_free:
 481         zfree(&evsel->name);
 482         free(evsel);
 483 out_err:
 484         return ERR_PTR(err);
 485 }
 486
 487 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = {
 488         "cycles",
 489         "instructions",
 490         "cache-references",
 491         "cache-misses",
 492         "branches",
 493         "branch-misses",
 494         "bus-cycles",
 495         "stalled-cycles-frontend",
 496         "stalled-cycles-backend",
 497         "ref-cycles",
 498 };
 499
 500 char *evsel__bpf_counter_events;
 501
 502 bool evsel__match_bpf_counter_events(const char *name)
 503 {
 504         int name_len;
 505         bool match;
 506         char *ptr;
 507
 508         if (!evsel__bpf_counter_events)
 509                 return false;
 510
 511         ptr = strstr(evsel__bpf_counter_events, name);
 512         name_len = strlen(name);
 513
 514         /* check name matches a full token in evsel__bpf_counter_events */
 515         match = (ptr != NULL) &&
 516                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 517                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 518
 519         return match;
 520 }
 521
 522 static const char *__evsel__hw_name(u64 config)
 523 {
 524         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 525                 return evsel__hw_names[config];
 526
 527         return "unknown-hardware";
 528 }
 529
 530 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 531 {
 532         int colon = 0, r = 0;
 533         struct perf_event_attr *attr = &evsel->core.attr;
 534         bool exclude_guest_default = false;
 535
 536 #define MOD_PRINT(context, mod) do {                                    \
 537                 if (!attr->exclude_##context) {                         \
 538                         if (!colon) colon = ++r;                        \
 539                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 540                 } } while(0)
 541
 542         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 543                 MOD_PRINT(kernel, 'k');
 544                 MOD_PRINT(user, 'u');
 545                 MOD_PRINT(hv, 'h');
 546                 exclude_guest_default = true;
 547         }
 548
 549         if (attr->precise_ip) {
 550                 if (!colon)
 551                         colon = ++r;
 552                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 553                 exclude_guest_default = true;
 554         }
 555
 556         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 557                 MOD_PRINT(host, 'H');
 558                 MOD_PRINT(guest, 'G');
 559         }
 560 #undef MOD_PRINT
 561         if (colon)
 562                 bf[colon - 1] = ':';
 563         return r;
 564 }
 565
 566 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 567 {
 568         int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 569         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 570 }
 571
 572 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = {
 573         "cpu-clock",
 574         "task-clock",
 575         "page-faults",
 576         "context-switches",
 577         "cpu-migrations",
 578         "minor-faults",
 579         "major-faults",
 580         "alignment-faults",
 581         "emulation-faults",
 582         "dummy",
 583 };
 584
 585 static const char *__evsel__sw_name(u64 config)
 586 {
 587         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 588                 return evsel__sw_names[config];
 589         return "unknown-software";
 590 }
 591
 592 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 593 {
 594         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 595         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 596 }
 597
 598 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 599 {
 600         int r;
 601
 602         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 603
 604         if (type & HW_BREAKPOINT_R)
 605                 r += scnprintf(bf + r, size - r, "r");
 606
 607         if (type & HW_BREAKPOINT_W)
 608                 r += scnprintf(bf + r, size - r, "w");
 609
 610         if (type & HW_BREAKPOINT_X)
 611                 r += scnprintf(bf + r, size - r, "x");
 612
 613         return r;
 614 }
 615
 616 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 617 {
 618         struct perf_event_attr *attr = &evsel->core.attr;
 619         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 620         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 621 }
 622
 623 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 624  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 625  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 626  { "LLC",       "L2",                                                   },
 627  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 628  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 629  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 630  { "node",                                                              },
 631 };
 632
 633 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 634  { "load",      "loads",        "read",                                 },
 635  { "store",     "stores",       "write",                                },
 636  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 637 };
 638
 639 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 640  { "refs",      "Reference",    "ops",          "access",               },
 641  { "misses",    "miss",                                                 },
 642 };
 643
 644 #define C(x)            PERF_COUNT_HW_CACHE_##x
 645 #define CACHE_READ      (1 << C(OP_READ))
 646 #define CACHE_WRITE     (1 << C(OP_WRITE))
 647 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 648 #define COP(x)          (1 << x)
 649
 650 /*
 651  * cache operation stat
 652  * L1I : Read and prefetch only
 653  * ITLB and BPU : Read-only
 654  */
 655 static unsigned long evsel__hw_cache_stat[C(MAX)] = {
 656  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 657  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 658  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 659  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 660  [C(ITLB)]      = (CACHE_READ),
 661  [C(BPU)]       = (CACHE_READ),
 662  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 663 };
 664
 665 bool evsel__is_cache_op_valid(u8 type, u8 op)
 666 {
 667         if (evsel__hw_cache_stat[type] & COP(op))
 668                 return true;    /* valid */
 669         else
 670                 return false;   /* invalid */
 671 }
 672
 673 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 674 {
 675         if (result) {
 676                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 677                                  evsel__hw_cache_op[op][0],
 678                                  evsel__hw_cache_result[result][0]);
 679         }
 680
 681         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 682                          evsel__hw_cache_op[op][1]);
 683 }
 684
 685 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 686 {
 687         u8 op, result, type = (config >>  0) & 0xff;
 688         const char *err = "unknown-ext-hardware-cache-type";
 689
 690         if (type >= PERF_COUNT_HW_CACHE_MAX)
 691                 goto out_err;
 692
 693         op = (config >>  8) & 0xff;
 694         err = "unknown-ext-hardware-cache-op";
 695         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 696                 goto out_err;
 697
 698         result = (config >> 16) & 0xff;
 699         err = "unknown-ext-hardware-cache-result";
 700         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 701                 goto out_err;
 702
 703         err = "invalid-cache";
 704         if (!evsel__is_cache_op_valid(type, op))
 705                 goto out_err;
 706
 707         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 708 out_err:
 709         return scnprintf(bf, size, "%s", err);
 710 }
 711
 712 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 713 {
 714         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 715         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 716 }
 717
 718 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 719 {
 720         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 721         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 722 }
 723
 724 static int evsel__tool_name(char *bf, size_t size)
 725 {
 726         int ret = scnprintf(bf, size, "duration_time");
 727         return ret;
 728 }
 729
 730 const char *evsel__name(struct evsel *evsel)
 731 {
 732         char bf[128];
 733
 734         if (!evsel)
 735                 goto out_unknown;
 736
 737         if (evsel->name)
 738                 return evsel->name;
 739
 740         switch (evsel->core.attr.type) {
 741         case PERF_TYPE_RAW:
 742                 evsel__raw_name(evsel, bf, sizeof(bf));
 743                 break;
 744
 745         case PERF_TYPE_HARDWARE:
 746                 evsel__hw_name(evsel, bf, sizeof(bf));
 747                 break;
 748
 749         case PERF_TYPE_HW_CACHE:
 750                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 751                 break;
 752
 753         case PERF_TYPE_SOFTWARE:
 754                 if (evsel->tool_event)
 755                         evsel__tool_name(bf, sizeof(bf));
 756                 else
 757                         evsel__sw_name(evsel, bf, sizeof(bf));
 758                 break;
 759
 760         case PERF_TYPE_TRACEPOINT:
 761                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 762                 break;
 763
 764         case PERF_TYPE_BREAKPOINT:
 765                 evsel__bp_name(evsel, bf, sizeof(bf));
 766                 break;
 767
 768         default:
 769                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 770                           evsel->core.attr.type);
 771                 break;
 772         }
 773
 774         evsel->name = strdup(bf);
 775
 776         if (evsel->name)
 777                 return evsel->name;
 778 out_unknown:
 779         return "unknown";
 780 }
 781
 782 const char *evsel__group_name(struct evsel *evsel)
 783 {
 784         return evsel->group_name ?: "anon group";
 785 }
 786
 787 /*
 788  * Returns the group details for the specified leader,
 789  * with following rules.
 790  *
 791  *  For record -e '{cycles,instructions}'
 792  *    'anon group { cycles:u, instructions:u }'
 793  *
 794  *  For record -e 'cycles,instructions' and report --group
 795  *    'cycles:u, instructions:u'
 796  */
 797 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 798 {
 799         int ret = 0;
 800         struct evsel *pos;
 801         const char *group_name = evsel__group_name(evsel);
 802
 803         if (!evsel->forced_leader)
 804                 ret = scnprintf(buf, size, "%s { ", group_name);
 805
 806         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 807
 808         for_each_group_member(pos, evsel)
 809                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 810
 811         if (!evsel->forced_leader)
 812                 ret += scnprintf(buf + ret, size - ret, " }");
 813
 814         return ret;
 815 }
 816
 817 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 818                                       struct callchain_param *param)
 819 {
 820         bool function = evsel__is_function_event(evsel);
 821         struct perf_event_attr *attr = &evsel->core.attr;
 822
 823         evsel__set_sample_bit(evsel, CALLCHAIN);
 824
 825         attr->sample_max_stack = param->max_stack;
 826
 827         if (opts->kernel_callchains)
 828                 attr->exclude_callchain_user = 1;
 829         if (opts->user_callchains)
 830                 attr->exclude_callchain_kernel = 1;
 831         if (param->record_mode == CALLCHAIN_LBR) {
 832                 if (!opts->branch_stack) {
 833                         if (attr->exclude_user) {
 834                                 pr_warning("LBR callstack option is only available "
 835                                            "to get user callchain information. "
 836                                            "Falling back to framepointers.\n");
 837                         } else {
 838                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 839                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 840                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 841                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 842                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 843                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 844                         }
 845                 } else
 846                          pr_warning("Cannot use LBR callstack with branch stack. "
 847                                     "Falling back to framepointers.\n");
 848         }
 849
 850         if (param->record_mode == CALLCHAIN_DWARF) {
 851                 if (!function) {
 852                         evsel__set_sample_bit(evsel, REGS_USER);
 853                         evsel__set_sample_bit(evsel, STACK_USER);
 854                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 855                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 856                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 857                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 858                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 859                         } else {
 860                                 attr->sample_regs_user |= PERF_REGS_MASK;
 861                         }
 862                         attr->sample_stack_user = param->dump_size;
 863                         attr->exclude_callchain_user = 1;
 864                 } else {
 865                         pr_info("Cannot use DWARF unwind for function trace event,"
 866                                 " falling back to framepointers.\n");
 867                 }
 868         }
 869
 870         if (function) {
 871                 pr_info("Disabling user space callchains for function trace event.\n");
 872                 attr->exclude_callchain_user = 1;
 873         }
 874 }
 875
 876 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 877                              struct callchain_param *param)
 878 {
 879         if (param->enabled)
 880                 return __evsel__config_callchain(evsel, opts, param);
 881 }
 882
 883 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 884 {
 885         struct perf_event_attr *attr = &evsel->core.attr;
 886
 887         evsel__reset_sample_bit(evsel, CALLCHAIN);
 888         if (param->record_mode == CALLCHAIN_LBR) {
 889                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 890                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 891                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 892                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 893         }
 894         if (param->record_mode == CALLCHAIN_DWARF) {
 895                 evsel__reset_sample_bit(evsel, REGS_USER);
 896                 evsel__reset_sample_bit(evsel, STACK_USER);
 897         }
 898 }
 899
 900 static void evsel__apply_config_terms(struct evsel *evsel,
 901                                       struct record_opts *opts, bool track)
 902 {
 903         struct evsel_config_term *term;
 904         struct list_head *config_terms = &evsel->config_terms;
 905         struct perf_event_attr *attr = &evsel->core.attr;
 906         /* callgraph default */
 907         struct callchain_param param = {
 908                 .record_mode = callchain_param.record_mode,
 909         };
 910         u32 dump_size = 0;
 911         int max_stack = 0;
 912         const char *callgraph_buf = NULL;
 913
 914         list_for_each_entry(term, config_terms, list) {
 915                 switch (term->type) {
 916                 case EVSEL__CONFIG_TERM_PERIOD:
 917                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 918                                 attr->sample_period = term->val.period;
 919                                 attr->freq = 0;
 920                                 evsel__reset_sample_bit(evsel, PERIOD);
 921                         }
 922                         break;
 923                 case EVSEL__CONFIG_TERM_FREQ:
 924                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 925                                 attr->sample_freq = term->val.freq;
 926                                 attr->freq = 1;
 927                                 evsel__set_sample_bit(evsel, PERIOD);
 928                         }
 929                         break;
 930                 case EVSEL__CONFIG_TERM_TIME:
 931                         if (term->val.time)
 932                                 evsel__set_sample_bit(evsel, TIME);
 933                         else
 934                                 evsel__reset_sample_bit(evsel, TIME);
 935                         break;
 936                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 937                         callgraph_buf = term->val.str;
 938                         break;
 939                 case EVSEL__CONFIG_TERM_BRANCH:
 940                         if (term->val.str && strcmp(term->val.str, "no")) {
 941                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 942                                 parse_branch_str(term->val.str,
 943                                                  &attr->branch_sample_type);
 944                         } else
 945                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 946                         break;
 947                 case EVSEL__CONFIG_TERM_STACK_USER:
 948                         dump_size = term->val.stack_user;
 949                         break;
 950                 case EVSEL__CONFIG_TERM_MAX_STACK:
 951                         max_stack = term->val.max_stack;
 952                         break;
 953                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
 954                         evsel->max_events = term->val.max_events;
 955                         break;
 956                 case EVSEL__CONFIG_TERM_INHERIT:
 957                         /*
 958                          * attr->inherit should has already been set by
 959                          * evsel__config. If user explicitly set
 960                          * inherit using config terms, override global
 961                          * opt->no_inherit setting.
 962                          */
 963                         attr->inherit = term->val.inherit ? 1 : 0;
 964                         break;
 965                 case EVSEL__CONFIG_TERM_OVERWRITE:
 966                         attr->write_backward = term->val.overwrite ? 1 : 0;
 967                         break;
 968                 case EVSEL__CONFIG_TERM_DRV_CFG:
 969                         break;
 970                 case EVSEL__CONFIG_TERM_PERCORE:
 971                         break;
 972                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
 973                         attr->aux_output = term->val.aux_output ? 1 : 0;
 974                         break;
 975                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
 976                         /* Already applied by auxtrace */
 977                         break;
 978                 case EVSEL__CONFIG_TERM_CFG_CHG:
 979                         break;
 980                 default:
 981                         break;
 982                 }
 983         }
 984
 985         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 986         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 987                 bool sample_address = false;
 988
 989                 if (max_stack) {
 990                         param.max_stack = max_stack;
 991                         if (callgraph_buf == NULL)
 992                                 callgraph_buf = "fp";
 993                 }
 994
 995                 /* parse callgraph parameters */
 996                 if (callgraph_buf != NULL) {
 997                         if (!strcmp(callgraph_buf, "no")) {
 998                                 param.enabled = false;
 999                                 param.record_mode = CALLCHAIN_NONE;
1000                         } else {
1001                                 param.enabled = true;
1002                                 if (parse_callchain_record(callgraph_buf, &param)) {
1003                                         pr_err("per-event callgraph setting for %s failed. "
1004                                                "Apply callgraph global setting for it\n",
1005                                                evsel->name);
1006                                         return;
1007                                 }
1008                                 if (param.record_mode == CALLCHAIN_DWARF)
1009                                         sample_address = true;
1010                         }
1011                 }
1012                 if (dump_size > 0) {
1013                         dump_size = round_up(dump_size, sizeof(u64));
1014                         param.dump_size = dump_size;
1015                 }
1016
1017                 /* If global callgraph set, clear it */
1018                 if (callchain_param.enabled)
1019                         evsel__reset_callgraph(evsel, &callchain_param);
1020
1021                 /* set perf-event callgraph */
1022                 if (param.enabled) {
1023                         if (sample_address) {
1024                                 evsel__set_sample_bit(evsel, ADDR);
1025                                 evsel__set_sample_bit(evsel, DATA_SRC);
1026                                 evsel->core.attr.mmap_data = track;
1027                         }
1028                         evsel__config_callchain(evsel, opts, &param);
1029                 }
1030         }
1031 }
1032
1033 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1034 {
1035         struct evsel_config_term *term, *found_term = NULL;
1036
1037         list_for_each_entry(term, &evsel->config_terms, list) {
1038                 if (term->type == type)
1039                         found_term = term;
1040         }
1041
1042         return found_term;
1043 }
1044
1045 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1046 {
1047         evsel__set_sample_bit(evsel, WEIGHT);
1048 }
1049
1050 /*
1051  * The enable_on_exec/disabled value strategy:
1052  *
1053  *  1) For any type of traced program:
1054  *    - all independent events and group leaders are disabled
1055  *    - all group members are enabled
1056  *
1057  *     Group members are ruled by group leaders. They need to
1058  *     be enabled, because the group scheduling relies on that.
1059  *
1060  *  2) For traced programs executed by perf:
1061  *     - all independent events and group leaders have
1062  *       enable_on_exec set
1063  *     - we don't specifically enable or disable any event during
1064  *       the record command
1065  *
1066  *     Independent events and group leaders are initially disabled
1067  *     and get enabled by exec. Group members are ruled by group
1068  *     leaders as stated in 1).
1069  *
1070  *  3) For traced programs attached by perf (pid/tid):
1071  *     - we specifically enable or disable all events during
1072  *       the record command
1073  *
1074  *     When attaching events to already running traced we
1075  *     enable/disable events specifically, as there's no
1076  *     initial traced exec call.
1077  */
1078 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1079                    struct callchain_param *callchain)
1080 {
1081         struct evsel *leader = evsel__leader(evsel);
1082         struct perf_event_attr *attr = &evsel->core.attr;
1083         int track = evsel->tracking;
1084         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1085
1086         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1087         attr->inherit       = !opts->no_inherit;
1088         attr->write_backward = opts->overwrite ? 1 : 0;
1089
1090         evsel__set_sample_bit(evsel, IP);
1091         evsel__set_sample_bit(evsel, TID);
1092
1093         if (evsel->sample_read) {
1094                 evsel__set_sample_bit(evsel, READ);
1095
1096                 /*
1097                  * We need ID even in case of single event, because
1098                  * PERF_SAMPLE_READ process ID specific data.
1099                  */
1100                 evsel__set_sample_id(evsel, false);
1101
1102                 /*
1103                  * Apply group format only if we belong to group
1104                  * with more than one members.
1105                  */
1106                 if (leader->core.nr_members > 1) {
1107                         attr->read_format |= PERF_FORMAT_GROUP;
1108                         attr->inherit = 0;
1109                 }
1110         }
1111
1112         /*
1113          * We default some events to have a default interval. But keep
1114          * it a weak assumption overridable by the user.
1115          */
1116         if (!attr->sample_period) {
1117                 if (opts->freq) {
1118                         attr->freq              = 1;
1119                         attr->sample_freq       = opts->freq;
1120                 } else {
1121                         attr->sample_period = opts->default_interval;
1122                 }
1123         }
1124         /*
1125          * If attr->freq was set (here or earlier), ask for period
1126          * to be sampled.
1127          */
1128         if (attr->freq)
1129                 evsel__set_sample_bit(evsel, PERIOD);
1130
1131         if (opts->no_samples)
1132                 attr->sample_freq = 0;
1133
1134         if (opts->inherit_stat) {
1135                 evsel->core.attr.read_format |=
1136                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1137                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1138                         PERF_FORMAT_ID;
1139                 attr->inherit_stat = 1;
1140         }
1141
1142         if (opts->sample_address) {
1143                 evsel__set_sample_bit(evsel, ADDR);
1144                 attr->mmap_data = track;
1145         }
1146
1147         /*
1148          * We don't allow user space callchains for  function trace
1149          * event, due to issues with page faults while tracing page
1150          * fault handler and its overall trickiness nature.
1151          */
1152         if (evsel__is_function_event(evsel))
1153                 evsel->core.attr.exclude_callchain_user = 1;
1154
1155         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1156                 evsel__config_callchain(evsel, opts, callchain);
1157
1158         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1159             !evsel__is_dummy_event(evsel)) {
1160                 attr->sample_regs_intr = opts->sample_intr_regs;
1161                 evsel__set_sample_bit(evsel, REGS_INTR);
1162         }
1163
1164         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1165             !evsel__is_dummy_event(evsel)) {
1166                 attr->sample_regs_user |= opts->sample_user_regs;
1167                 evsel__set_sample_bit(evsel, REGS_USER);
1168         }
1169
1170         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1171                 evsel__set_sample_bit(evsel, CPU);
1172
1173         /*
1174          * When the user explicitly disabled time don't force it here.
1175          */
1176         if (opts->sample_time &&
1177             (!perf_missing_features.sample_id_all &&
1178             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1179              opts->sample_time_set)))
1180                 evsel__set_sample_bit(evsel, TIME);
1181
1182         if (opts->raw_samples && !evsel->no_aux_samples) {
1183                 evsel__set_sample_bit(evsel, TIME);
1184                 evsel__set_sample_bit(evsel, RAW);
1185                 evsel__set_sample_bit(evsel, CPU);
1186         }
1187
1188         if (opts->sample_address)
1189                 evsel__set_sample_bit(evsel, DATA_SRC);
1190
1191         if (opts->sample_phys_addr)
1192                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1193
1194         if (opts->no_buffering) {
1195                 attr->watermark = 0;
1196                 attr->wakeup_events = 1;
1197         }
1198         if (opts->branch_stack && !evsel->no_aux_samples) {
1199                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1200                 attr->branch_sample_type = opts->branch_stack;
1201         }
1202
1203         if (opts->sample_weight)
1204                 arch_evsel__set_sample_weight(evsel);
1205
1206         attr->task     = track;
1207         attr->mmap     = track;
1208         attr->mmap2    = track && !perf_missing_features.mmap2;
1209         attr->comm     = track;
1210         attr->build_id = track && opts->build_id;
1211
1212         /*
1213          * ksymbol is tracked separately with text poke because it needs to be
1214          * system wide and enabled immediately.
1215          */
1216         if (!opts->text_poke)
1217                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1218         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1219
1220         if (opts->record_namespaces)
1221                 attr->namespaces  = track;
1222
1223         if (opts->record_cgroup) {
1224                 attr->cgroup = track && !perf_missing_features.cgroup;
1225                 evsel__set_sample_bit(evsel, CGROUP);
1226         }
1227
1228         if (opts->sample_data_page_size)
1229                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1230
1231         if (opts->sample_code_page_size)
1232                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1233
1234         if (opts->record_switch_events)
1235                 attr->context_switch = track;
1236
1237         if (opts->sample_transaction)
1238                 evsel__set_sample_bit(evsel, TRANSACTION);
1239
1240         if (opts->running_time) {
1241                 evsel->core.attr.read_format |=
1242                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1243                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1244         }
1245
1246         /*
1247          * XXX see the function comment above
1248          *
1249          * Disabling only independent events or group leaders,
1250          * keeping group members enabled.
1251          */
1252         if (evsel__is_group_leader(evsel))
1253                 attr->disabled = 1;
1254
1255         /*
1256          * Setting enable_on_exec for independent events and
1257          * group leaders for traced executed by perf.
1258          */
1259         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1260             !opts->initial_delay)
1261                 attr->enable_on_exec = 1;
1262
1263         if (evsel->immediate) {
1264                 attr->disabled = 0;
1265                 attr->enable_on_exec = 0;
1266         }
1267
1268         clockid = opts->clockid;
1269         if (opts->use_clockid) {
1270                 attr->use_clockid = 1;
1271                 attr->clockid = opts->clockid;
1272         }
1273
1274         if (evsel->precise_max)
1275                 attr->precise_ip = 3;
1276
1277         if (opts->all_user) {
1278                 attr->exclude_kernel = 1;
1279                 attr->exclude_user   = 0;
1280         }
1281
1282         if (opts->all_kernel) {
1283                 attr->exclude_kernel = 0;
1284                 attr->exclude_user   = 1;
1285         }
1286
1287         if (evsel->core.own_cpus || evsel->unit)
1288                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1289
1290         /*
1291          * Apply event specific term settings,
1292          * it overloads any global configuration.
1293          */
1294         evsel__apply_config_terms(evsel, opts, track);
1295
1296         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1297
1298         /* The --period option takes the precedence. */
1299         if (opts->period_set) {
1300                 if (opts->period)
1301                         evsel__set_sample_bit(evsel, PERIOD);
1302                 else
1303                         evsel__reset_sample_bit(evsel, PERIOD);
1304         }
1305
1306         /*
1307          * A dummy event never triggers any actual counter and therefore
1308          * cannot be used with branch_stack.
1309          *
1310          * For initial_delay, a dummy event is added implicitly.
1311          * The software event will trigger -EOPNOTSUPP error out,
1312          * if BRANCH_STACK bit is set.
1313          */
1314         if (evsel__is_dummy_event(evsel))
1315                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1316 }
1317
1318 int evsel__set_filter(struct evsel *evsel, const char *filter)
1319 {
1320         char *new_filter = strdup(filter);
1321
1322         if (new_filter != NULL) {
1323                 free(evsel->filter);
1324                 evsel->filter = new_filter;
1325                 return 0;
1326         }
1327
1328         return -1;
1329 }
1330
1331 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1332 {
1333         char *new_filter;
1334
1335         if (evsel->filter == NULL)
1336                 return evsel__set_filter(evsel, filter);
1337
1338         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1339                 free(evsel->filter);
1340                 evsel->filter = new_filter;
1341                 return 0;
1342         }
1343
1344         return -1;
1345 }
1346
1347 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1348 {
1349         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1350 }
1351
1352 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1353 {
1354         return evsel__append_filter(evsel, "%s,%s", filter);
1355 }
1356
1357 /* Caller has to clear disabled after going through all CPUs. */
1358 int evsel__enable_cpu(struct evsel *evsel, int cpu)
1359 {
1360         return perf_evsel__enable_cpu(&evsel->core, cpu);
1361 }
1362
1363 int evsel__enable(struct evsel *evsel)
1364 {
1365         int err = perf_evsel__enable(&evsel->core);
1366
1367         if (!err)
1368                 evsel->disabled = false;
1369         return err;
1370 }
1371
1372 /* Caller has to set disabled after going through all CPUs. */
1373 int evsel__disable_cpu(struct evsel *evsel, int cpu)
1374 {
1375         return perf_evsel__disable_cpu(&evsel->core, cpu);
1376 }
1377
1378 int evsel__disable(struct evsel *evsel)
1379 {
1380         int err = perf_evsel__disable(&evsel->core);
1381         /*
1382          * We mark it disabled here so that tools that disable a event can
1383          * ignore events after they disable it. I.e. the ring buffer may have
1384          * already a few more events queued up before the kernel got the stop
1385          * request.
1386          */
1387         if (!err)
1388                 evsel->disabled = true;
1389
1390         return err;
1391 }
1392
1393 void free_config_terms(struct list_head *config_terms)
1394 {
1395         struct evsel_config_term *term, *h;
1396
1397         list_for_each_entry_safe(term, h, config_terms, list) {
1398                 list_del_init(&term->list);
1399                 if (term->free_str)
1400                         zfree(&term->val.str);
1401                 free(term);
1402         }
1403 }
1404
1405 static void evsel__free_config_terms(struct evsel *evsel)
1406 {
1407         free_config_terms(&evsel->config_terms);
1408 }
1409
1410 void evsel__exit(struct evsel *evsel)
1411 {
1412         assert(list_empty(&evsel->core.node));
1413         assert(evsel->evlist == NULL);
1414         bpf_counter__destroy(evsel);
1415         evsel__free_counts(evsel);
1416         perf_evsel__free_fd(&evsel->core);
1417         perf_evsel__free_id(&evsel->core);
1418         evsel__free_config_terms(evsel);
1419         cgroup__put(evsel->cgrp);
1420         perf_cpu_map__put(evsel->core.cpus);
1421         perf_cpu_map__put(evsel->core.own_cpus);
1422         perf_thread_map__put(evsel->core.threads);
1423         zfree(&evsel->group_name);
1424         zfree(&evsel->name);
1425         zfree(&evsel->pmu_name);
1426         evsel__zero_per_pkg(evsel);
1427         hashmap__free(evsel->per_pkg_mask);
1428         evsel->per_pkg_mask = NULL;
1429         zfree(&evsel->metric_events);
1430         perf_evsel__object.fini(evsel);
1431 }
1432
1433 void evsel__delete(struct evsel *evsel)
1434 {
1435         evsel__exit(evsel);
1436         free(evsel);
1437 }
1438
1439 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread,
1440                            struct perf_counts_values *count)
1441 {
1442         struct perf_counts_values tmp;
1443
1444         if (!evsel->prev_raw_counts)
1445                 return;
1446
1447         if (cpu == -1) {
1448                 tmp = evsel->prev_raw_counts->aggr;
1449                 evsel->prev_raw_counts->aggr = *count;
1450         } else {
1451                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1452                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1453         }
1454
1455         count->val = count->val - tmp.val;
1456         count->ena = count->ena - tmp.ena;
1457         count->run = count->run - tmp.run;
1458 }
1459
1460 void perf_counts_values__scale(struct perf_counts_values *count,
1461                                bool scale, s8 *pscaled)
1462 {
1463         s8 scaled = 0;
1464
1465         if (scale) {
1466                 if (count->run == 0) {
1467                         scaled = -1;
1468                         count->val = 0;
1469                 } else if (count->run < count->ena) {
1470                         scaled = 1;
1471                         count->val = (u64)((double) count->val * count->ena / count->run);
1472                 }
1473         }
1474
1475         if (pscaled)
1476                 *pscaled = scaled;
1477 }
1478
1479 static int evsel__read_one(struct evsel *evsel, int cpu, int thread)
1480 {
1481         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1482
1483         return perf_evsel__read(&evsel->core, cpu, thread, count);
1484 }
1485
1486 static void evsel__set_count(struct evsel *counter, int cpu, int thread, u64 val, u64 ena, u64 run)
1487 {
1488         struct perf_counts_values *count;
1489
1490         count = perf_counts(counter->counts, cpu, thread);
1491
1492         count->val    = val;
1493         count->ena    = ena;
1494         count->run    = run;
1495
1496         perf_counts__set_loaded(counter->counts, cpu, thread, true);
1497 }
1498
1499 static int evsel__process_group_data(struct evsel *leader, int cpu, int thread, u64 *data)
1500 {
1501         u64 read_format = leader->core.attr.read_format;
1502         struct sample_read_value *v;
1503         u64 nr, ena = 0, run = 0, i;
1504
1505         nr = *data++;
1506
1507         if (nr != (u64) leader->core.nr_members)
1508                 return -EINVAL;
1509
1510         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1511                 ena = *data++;
1512
1513         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1514                 run = *data++;
1515
1516         v = (struct sample_read_value *) data;
1517
1518         evsel__set_count(leader, cpu, thread, v[0].value, ena, run);
1519
1520         for (i = 1; i < nr; i++) {
1521                 struct evsel *counter;
1522
1523                 counter = evlist__id2evsel(leader->evlist, v[i].id);
1524                 if (!counter)
1525                         return -EINVAL;
1526
1527                 evsel__set_count(counter, cpu, thread, v[i].value, ena, run);
1528         }
1529
1530         return 0;
1531 }
1532
1533 static int evsel__read_group(struct evsel *leader, int cpu, int thread)
1534 {
1535         struct perf_stat_evsel *ps = leader->stats;
1536         u64 read_format = leader->core.attr.read_format;
1537         int size = perf_evsel__read_size(&leader->core);
1538         u64 *data = ps->group_data;
1539
1540         if (!(read_format & PERF_FORMAT_ID))
1541                 return -EINVAL;
1542
1543         if (!evsel__is_group_leader(leader))
1544                 return -EINVAL;
1545
1546         if (!data) {
1547                 data = zalloc(size);
1548                 if (!data)
1549                         return -ENOMEM;
1550
1551                 ps->group_data = data;
1552         }
1553
1554         if (FD(leader, cpu, thread) < 0)
1555                 return -EINVAL;
1556
1557         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1558                 return -errno;
1559
1560         return evsel__process_group_data(leader, cpu, thread, data);
1561 }
1562
1563 int evsel__read_counter(struct evsel *evsel, int cpu, int thread)
1564 {
1565         u64 read_format = evsel->core.attr.read_format;
1566
1567         if (read_format & PERF_FORMAT_GROUP)
1568                 return evsel__read_group(evsel, cpu, thread);
1569
1570         return evsel__read_one(evsel, cpu, thread);
1571 }
1572
1573 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale)
1574 {
1575         struct perf_counts_values count;
1576         size_t nv = scale ? 3 : 1;
1577
1578         if (FD(evsel, cpu, thread) < 0)
1579                 return -EINVAL;
1580
1581         if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1582                 return -ENOMEM;
1583
1584         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1585                 return -errno;
1586
1587         evsel__compute_deltas(evsel, cpu, thread, &count);
1588         perf_counts_values__scale(&count, scale, NULL);
1589         *perf_counts(evsel->counts, cpu, thread) = count;
1590         return 0;
1591 }
1592
1593 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1594                                   int cpu)
1595 {
1596         int cpuid;
1597
1598         cpuid = perf_cpu_map__cpu(evsel->core.cpus, cpu);
1599         return perf_cpu_map__idx(other->core.cpus, cpuid);
1600 }
1601
1602 static int evsel__hybrid_group_cpu(struct evsel *evsel, int cpu)
1603 {
1604         struct evsel *leader = evsel__leader(evsel);
1605
1606         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1607             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1608                 return evsel__match_other_cpu(evsel, leader, cpu);
1609         }
1610
1611         return cpu;
1612 }
1613
1614 static int get_group_fd(struct evsel *evsel, int cpu, int thread)
1615 {
1616         struct evsel *leader = evsel__leader(evsel);
1617         int fd;
1618
1619         if (evsel__is_group_leader(evsel))
1620                 return -1;
1621
1622         /*
1623          * Leader must be already processed/open,
1624          * if not it's a bug.
1625          */
1626         BUG_ON(!leader->core.fd);
1627
1628         cpu = evsel__hybrid_group_cpu(evsel, cpu);
1629         if (cpu == -1)
1630                 return -1;
1631
1632         fd = FD(leader, cpu, thread);
1633         BUG_ON(fd == -1);
1634
1635         return fd;
1636 }
1637
1638 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1639 {
1640         for (int cpu = 0; cpu < nr_cpus; cpu++)
1641                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1642                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1643 }
1644
1645 static int update_fds(struct evsel *evsel,
1646                       int nr_cpus, int cpu_idx,
1647                       int nr_threads, int thread_idx)
1648 {
1649         struct evsel *pos;
1650
1651         if (cpu_idx >= nr_cpus || thread_idx >= nr_threads)
1652                 return -EINVAL;
1653
1654         evlist__for_each_entry(evsel->evlist, pos) {
1655                 nr_cpus = pos != evsel ? nr_cpus : cpu_idx;
1656
1657                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1658
1659                 /*
1660                  * Since fds for next evsel has not been created,
1661                  * there is no need to iterate whole event list.
1662                  */
1663                 if (pos == evsel)
1664                         break;
1665         }
1666         return 0;
1667 }
1668
1669 bool evsel__ignore_missing_thread(struct evsel *evsel,
1670                                   int nr_cpus, int cpu,
1671                                   struct perf_thread_map *threads,
1672                                   int thread, int err)
1673 {
1674         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1675
1676         if (!evsel->ignore_missing_thread)
1677                 return false;
1678
1679         /* The system wide setup does not work with threads. */
1680         if (evsel->core.system_wide)
1681                 return false;
1682
1683         /* The -ESRCH is perf event syscall errno for pid's not found. */
1684         if (err != -ESRCH)
1685                 return false;
1686
1687         /* If there's only one thread, let it fail. */
1688         if (threads->nr == 1)
1689                 return false;
1690
1691         /*
1692          * We should remove fd for missing_thread first
1693          * because thread_map__remove() will decrease threads->nr.
1694          */
1695         if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread))
1696                 return false;
1697
1698         if (thread_map__remove(threads, thread))
1699                 return false;
1700
1701         pr_warning("WARNING: Ignored open failure for pid %d\n",
1702                    ignore_pid);
1703         return true;
1704 }
1705
1706 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1707                                 void *priv __maybe_unused)
1708 {
1709         return fprintf(fp, "  %-32s %s\n", name, val);
1710 }
1711
1712 static void display_attr(struct perf_event_attr *attr)
1713 {
1714         if (verbose >= 2 || debug_peo_args) {
1715                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1716                 fprintf(stderr, "perf_event_attr:\n");
1717                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1718                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1719         }
1720 }
1721
1722 bool evsel__precise_ip_fallback(struct evsel *evsel)
1723 {
1724         /* Do not try less precise if not requested. */
1725         if (!evsel->precise_max)
1726                 return false;
1727
1728         /*
1729          * We tried all the precise_ip values, and it's
1730          * still failing, so leave it to standard fallback.
1731          */
1732         if (!evsel->core.attr.precise_ip) {
1733                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1734                 return false;
1735         }
1736
1737         if (!evsel->precise_ip_original)
1738                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1739
1740         evsel->core.attr.precise_ip--;
1741         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1742         display_attr(&evsel->core.attr);
1743         return true;
1744 }
1745
1746 static struct perf_cpu_map *empty_cpu_map;
1747 static struct perf_thread_map *empty_thread_map;
1748
1749 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1750                 struct perf_thread_map *threads)
1751 {
1752         int nthreads;
1753
1754         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1755             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1756                 return -EINVAL;
1757
1758         if (cpus == NULL) {
1759                 if (empty_cpu_map == NULL) {
1760                         empty_cpu_map = perf_cpu_map__dummy_new();
1761                         if (empty_cpu_map == NULL)
1762                                 return -ENOMEM;
1763                 }
1764
1765                 cpus = empty_cpu_map;
1766         }
1767
1768         if (threads == NULL) {
1769                 if (empty_thread_map == NULL) {
1770                         empty_thread_map = thread_map__new_by_tid(-1);
1771                         if (empty_thread_map == NULL)
1772                                 return -ENOMEM;
1773                 }
1774
1775                 threads = empty_thread_map;
1776         }
1777
1778         if (evsel->core.system_wide)
1779                 nthreads = 1;
1780         else
1781                 nthreads = threads->nr;
1782
1783         if (evsel->core.fd == NULL &&
1784             perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0)
1785                 return -ENOMEM;
1786
1787         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1788         if (evsel->cgrp)
1789                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1790
1791         return 0;
1792 }
1793
1794 static void evsel__disable_missing_features(struct evsel *evsel)
1795 {
1796         if (perf_missing_features.weight_struct) {
1797                 evsel__set_sample_bit(evsel, WEIGHT);
1798                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1799         }
1800         if (perf_missing_features.clockid_wrong)
1801                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1802         if (perf_missing_features.clockid) {
1803                 evsel->core.attr.use_clockid = 0;
1804                 evsel->core.attr.clockid = 0;
1805         }
1806         if (perf_missing_features.cloexec)
1807                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1808         if (perf_missing_features.mmap2)
1809                 evsel->core.attr.mmap2 = 0;
1810         if (perf_missing_features.exclude_guest)
1811                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1812         if (perf_missing_features.lbr_flags)
1813                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1814                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1815         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1816                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1817         if (perf_missing_features.ksymbol)
1818                 evsel->core.attr.ksymbol = 0;
1819         if (perf_missing_features.bpf)
1820                 evsel->core.attr.bpf_event = 0;
1821         if (perf_missing_features.branch_hw_idx)
1822                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1823         if (perf_missing_features.sample_id_all)
1824                 evsel->core.attr.sample_id_all = 0;
1825 }
1826
1827 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1828                         struct perf_thread_map *threads)
1829 {
1830         int err;
1831
1832         err = __evsel__prepare_open(evsel, cpus, threads);
1833         if (err)
1834                 return err;
1835
1836         evsel__disable_missing_features(evsel);
1837
1838         return err;
1839 }
1840
1841 bool evsel__detect_missing_features(struct evsel *evsel)
1842 {
1843         /*
1844          * Must probe features in the order they were added to the
1845          * perf_event_attr interface.
1846          */
1847         if (!perf_missing_features.weight_struct &&
1848             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1849                 perf_missing_features.weight_struct = true;
1850                 pr_debug2("switching off weight struct support\n");
1851                 return true;
1852         } else if (!perf_missing_features.code_page_size &&
1853             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1854                 perf_missing_features.code_page_size = true;
1855                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1856                 return false;
1857         } else if (!perf_missing_features.data_page_size &&
1858             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1859                 perf_missing_features.data_page_size = true;
1860                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1861                 return false;
1862         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1863                 perf_missing_features.cgroup = true;
1864                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1865                 return false;
1866         } else if (!perf_missing_features.branch_hw_idx &&
1867             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1868                 perf_missing_features.branch_hw_idx = true;
1869                 pr_debug2("switching off branch HW index support\n");
1870                 return true;
1871         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1872                 perf_missing_features.aux_output = true;
1873                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1874                 return false;
1875         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1876                 perf_missing_features.bpf = true;
1877                 pr_debug2_peo("switching off bpf_event\n");
1878                 return true;
1879         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1880                 perf_missing_features.ksymbol = true;
1881                 pr_debug2_peo("switching off ksymbol\n");
1882                 return true;
1883         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1884                 perf_missing_features.write_backward = true;
1885                 pr_debug2_peo("switching off write_backward\n");
1886                 return false;
1887         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1888                 perf_missing_features.clockid_wrong = true;
1889                 pr_debug2_peo("switching off clockid\n");
1890                 return true;
1891         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1892                 perf_missing_features.clockid = true;
1893                 pr_debug2_peo("switching off use_clockid\n");
1894                 return true;
1895         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1896                 perf_missing_features.cloexec = true;
1897                 pr_debug2_peo("switching off cloexec flag\n");
1898                 return true;
1899         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1900                 perf_missing_features.mmap2 = true;
1901                 pr_debug2_peo("switching off mmap2\n");
1902                 return true;
1903         } else if (!perf_missing_features.exclude_guest &&
1904                    (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) {
1905                 perf_missing_features.exclude_guest = true;
1906                 pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1907                 return true;
1908         } else if (!perf_missing_features.sample_id_all) {
1909                 perf_missing_features.sample_id_all = true;
1910                 pr_debug2_peo("switching off sample_id_all\n");
1911                 return true;
1912         } else if (!perf_missing_features.lbr_flags &&
1913                         (evsel->core.attr.branch_sample_type &
1914                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1915                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1916                 perf_missing_features.lbr_flags = true;
1917                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1918                 return true;
1919         } else if (!perf_missing_features.group_read &&
1920                     evsel->core.attr.inherit &&
1921                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
1922                    evsel__is_group_leader(evsel)) {
1923                 perf_missing_features.group_read = true;
1924                 pr_debug2_peo("switching off group read\n");
1925                 return true;
1926         } else {
1927                 return false;
1928         }
1929 }
1930
1931 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
1932 {
1933         int old_errno;
1934         struct rlimit l;
1935
1936         if (*set_rlimit < INCREASED_MAX) {
1937                 old_errno = errno;
1938
1939                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1940                         if (*set_rlimit == NO_CHANGE) {
1941                                 l.rlim_cur = l.rlim_max;
1942                         } else {
1943                                 l.rlim_cur = l.rlim_max + 1000;
1944                                 l.rlim_max = l.rlim_cur;
1945                         }
1946                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1947                                 (*set_rlimit) += 1;
1948                                 errno = old_errno;
1949                                 return true;
1950                         }
1951                 }
1952                 errno = old_errno;
1953         }
1954
1955         return false;
1956 }
1957
1958 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
1959                 struct perf_thread_map *threads,
1960                 int start_cpu, int end_cpu)
1961 {
1962         int cpu, thread, nthreads;
1963         int pid = -1, err, old_errno;
1964         enum rlimit_action set_rlimit = NO_CHANGE;
1965
1966         err = __evsel__prepare_open(evsel, cpus, threads);
1967         if (err)
1968                 return err;
1969
1970         if (cpus == NULL)
1971                 cpus = empty_cpu_map;
1972
1973         if (threads == NULL)
1974                 threads = empty_thread_map;
1975
1976         if (evsel->core.system_wide)
1977                 nthreads = 1;
1978         else
1979                 nthreads = threads->nr;
1980
1981         if (evsel->cgrp)
1982                 pid = evsel->cgrp->fd;
1983
1984 fallback_missing_features:
1985         evsel__disable_missing_features(evsel);
1986
1987         display_attr(&evsel->core.attr);
1988
1989         for (cpu = start_cpu; cpu < end_cpu; cpu++) {
1990
1991                 for (thread = 0; thread < nthreads; thread++) {
1992                         int fd, group_fd;
1993 retry_open:
1994                         if (thread >= nthreads)
1995                                 break;
1996
1997                         if (!evsel->cgrp && !evsel->core.system_wide)
1998                                 pid = perf_thread_map__pid(threads, thread);
1999
2000                         group_fd = get_group_fd(evsel, cpu, thread);
2001
2002                         test_attr__ready();
2003
2004                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2005                                 pid, cpus->map[cpu], group_fd, evsel->open_flags);
2006
2007                         fd = sys_perf_event_open(&evsel->core.attr, pid, cpus->map[cpu],
2008                                                 group_fd, evsel->open_flags);
2009
2010                         FD(evsel, cpu, thread) = fd;
2011
2012                         if (fd < 0) {
2013                                 err = -errno;
2014
2015                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2016                                           err);
2017                                 goto try_fallback;
2018                         }
2019
2020                         bpf_counter__install_pe(evsel, cpu, fd);
2021
2022                         if (unlikely(test_attr__enabled)) {
2023                                 test_attr__open(&evsel->core.attr, pid, cpus->map[cpu],
2024                                                 fd, group_fd, evsel->open_flags);
2025                         }
2026
2027                         pr_debug2_peo(" = %d\n", fd);
2028
2029                         if (evsel->bpf_fd >= 0) {
2030                                 int evt_fd = fd;
2031                                 int bpf_fd = evsel->bpf_fd;
2032
2033                                 err = ioctl(evt_fd,
2034                                             PERF_EVENT_IOC_SET_BPF,
2035                                             bpf_fd);
2036                                 if (err && errno != EEXIST) {
2037                                         pr_err("failed to attach bpf fd %d: %s\n",
2038                                                bpf_fd, strerror(errno));
2039                                         err = -EINVAL;
2040                                         goto out_close;
2041                                 }
2042                         }
2043
2044                         set_rlimit = NO_CHANGE;
2045
2046                         /*
2047                          * If we succeeded but had to kill clockid, fail and
2048                          * have evsel__open_strerror() print us a nice error.
2049                          */
2050                         if (perf_missing_features.clockid ||
2051                             perf_missing_features.clockid_wrong) {
2052                                 err = -EINVAL;
2053                                 goto out_close;
2054                         }
2055                 }
2056         }
2057
2058         return 0;
2059
2060 try_fallback:
2061         if (evsel__precise_ip_fallback(evsel))
2062                 goto retry_open;
2063
2064         if (evsel__ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) {
2065                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2066                 nthreads--;
2067
2068                 /* ... and pretend like nothing have happened. */
2069                 err = 0;
2070                 goto retry_open;
2071         }
2072         /*
2073          * perf stat needs between 5 and 22 fds per CPU. When we run out
2074          * of them try to increase the limits.
2075          */
2076         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2077                 goto retry_open;
2078
2079         if (err != -EINVAL || cpu > 0 || thread > 0)
2080                 goto out_close;
2081
2082         if (evsel__detect_missing_features(evsel))
2083                 goto fallback_missing_features;
2084 out_close:
2085         if (err)
2086                 threads->err_thread = thread;
2087
2088         old_errno = errno;
2089         do {
2090                 while (--thread >= 0) {
2091                         if (FD(evsel, cpu, thread) >= 0)
2092                                 close(FD(evsel, cpu, thread));
2093                         FD(evsel, cpu, thread) = -1;
2094                 }
2095                 thread = nthreads;
2096         } while (--cpu >= 0);
2097         errno = old_errno;
2098         return err;
2099 }
2100
2101 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2102                 struct perf_thread_map *threads)
2103 {
2104         return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1);
2105 }
2106
2107 void evsel__close(struct evsel *evsel)
2108 {
2109         perf_evsel__close(&evsel->core);
2110         perf_evsel__free_id(&evsel->core);
2111 }
2112
2113 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu)
2114 {
2115         if (cpu == -1)
2116                 return evsel__open_cpu(evsel, cpus, NULL, 0,
2117                                         cpus ? cpus->nr : 1);
2118
2119         return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1);
2120 }
2121
2122 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2123 {
2124         return evsel__open(evsel, NULL, threads);
2125 }
2126
2127 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2128                                        const union perf_event *event,
2129                                        struct perf_sample *sample)
2130 {
2131         u64 type = evsel->core.attr.sample_type;
2132         const __u64 *array = event->sample.array;
2133         bool swapped = evsel->needs_swap;
2134         union u64_swap u;
2135
2136         array += ((event->header.size -
2137                    sizeof(event->header)) / sizeof(u64)) - 1;
2138
2139         if (type & PERF_SAMPLE_IDENTIFIER) {
2140                 sample->id = *array;
2141                 array--;
2142         }
2143
2144         if (type & PERF_SAMPLE_CPU) {
2145                 u.val64 = *array;
2146                 if (swapped) {
2147                         /* undo swap of u64, then swap on individual u32s */
2148                         u.val64 = bswap_64(u.val64);
2149                         u.val32[0] = bswap_32(u.val32[0]);
2150                 }
2151
2152                 sample->cpu = u.val32[0];
2153                 array--;
2154         }
2155
2156         if (type & PERF_SAMPLE_STREAM_ID) {
2157                 sample->stream_id = *array;
2158                 array--;
2159         }
2160
2161         if (type & PERF_SAMPLE_ID) {
2162                 sample->id = *array;
2163                 array--;
2164         }
2165
2166         if (type & PERF_SAMPLE_TIME) {
2167                 sample->time = *array;
2168                 array--;
2169         }
2170
2171         if (type & PERF_SAMPLE_TID) {
2172                 u.val64 = *array;
2173                 if (swapped) {
2174                         /* undo swap of u64, then swap on individual u32s */
2175                         u.val64 = bswap_64(u.val64);
2176                         u.val32[0] = bswap_32(u.val32[0]);
2177                         u.val32[1] = bswap_32(u.val32[1]);
2178                 }
2179
2180                 sample->pid = u.val32[0];
2181                 sample->tid = u.val32[1];
2182                 array--;
2183         }
2184
2185         return 0;
2186 }
2187
2188 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2189                             u64 size)
2190 {
2191         return size > max_size || offset + size > endp;
2192 }
2193
2194 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2195         do {                                                            \
2196                 if (overflow(endp, (max_size), (offset), (size)))       \
2197                         return -EFAULT;                                 \
2198         } while (0)
2199
2200 #define OVERFLOW_CHECK_u64(offset) \
2201         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2202
2203 static int
2204 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2205 {
2206         /*
2207          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2208          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2209          * check the format does not go past the end of the event.
2210          */
2211         if (sample_size + sizeof(event->header) > event->header.size)
2212                 return -EFAULT;
2213
2214         return 0;
2215 }
2216
2217 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2218                                           const __u64 *array,
2219                                           u64 type __maybe_unused)
2220 {
2221         data->weight = *array;
2222 }
2223
2224 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2225                         struct perf_sample *data)
2226 {
2227         u64 type = evsel->core.attr.sample_type;
2228         bool swapped = evsel->needs_swap;
2229         const __u64 *array;
2230         u16 max_size = event->header.size;
2231         const void *endp = (void *)event + max_size;
2232         u64 sz;
2233
2234         /*
2235          * used for cross-endian analysis. See git commit 65014ab3
2236          * for why this goofiness is needed.
2237          */
2238         union u64_swap u;
2239
2240         memset(data, 0, sizeof(*data));
2241         data->cpu = data->pid = data->tid = -1;
2242         data->stream_id = data->id = data->time = -1ULL;
2243         data->period = evsel->core.attr.sample_period;
2244         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2245         data->misc    = event->header.misc;
2246         data->id = -1ULL;
2247         data->data_src = PERF_MEM_DATA_SRC_NONE;
2248
2249         if (event->header.type != PERF_RECORD_SAMPLE) {
2250                 if (!evsel->core.attr.sample_id_all)
2251                         return 0;
2252                 return perf_evsel__parse_id_sample(evsel, event, data);
2253         }
2254
2255         array = event->sample.array;
2256
2257         if (perf_event__check_size(event, evsel->sample_size))
2258                 return -EFAULT;
2259
2260         if (type & PERF_SAMPLE_IDENTIFIER) {
2261                 data->id = *array;
2262                 array++;
2263         }
2264
2265         if (type & PERF_SAMPLE_IP) {
2266                 data->ip = *array;
2267                 array++;
2268         }
2269
2270         if (type & PERF_SAMPLE_TID) {
2271                 u.val64 = *array;
2272                 if (swapped) {
2273                         /* undo swap of u64, then swap on individual u32s */
2274                         u.val64 = bswap_64(u.val64);
2275                         u.val32[0] = bswap_32(u.val32[0]);
2276                         u.val32[1] = bswap_32(u.val32[1]);
2277                 }
2278
2279                 data->pid = u.val32[0];
2280                 data->tid = u.val32[1];
2281                 array++;
2282         }
2283
2284         if (type & PERF_SAMPLE_TIME) {
2285                 data->time = *array;
2286                 array++;
2287         }
2288
2289         if (type & PERF_SAMPLE_ADDR) {
2290                 data->addr = *array;
2291                 array++;
2292         }
2293
2294         if (type & PERF_SAMPLE_ID) {
2295                 data->id = *array;
2296                 array++;
2297         }
2298
2299         if (type & PERF_SAMPLE_STREAM_ID) {
2300                 data->stream_id = *array;
2301                 array++;
2302         }
2303
2304         if (type & PERF_SAMPLE_CPU) {
2305
2306                 u.val64 = *array;
2307                 if (swapped) {
2308                         /* undo swap of u64, then swap on individual u32s */
2309                         u.val64 = bswap_64(u.val64);
2310                         u.val32[0] = bswap_32(u.val32[0]);
2311                 }
2312
2313                 data->cpu = u.val32[0];
2314                 array++;
2315         }
2316
2317         if (type & PERF_SAMPLE_PERIOD) {
2318                 data->period = *array;
2319                 array++;
2320         }
2321
2322         if (type & PERF_SAMPLE_READ) {
2323                 u64 read_format = evsel->core.attr.read_format;
2324
2325                 OVERFLOW_CHECK_u64(array);
2326                 if (read_format & PERF_FORMAT_GROUP)
2327                         data->read.group.nr = *array;
2328                 else
2329                         data->read.one.value = *array;
2330
2331                 array++;
2332
2333                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2334                         OVERFLOW_CHECK_u64(array);
2335                         data->read.time_enabled = *array;
2336                         array++;
2337                 }
2338
2339                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2340                         OVERFLOW_CHECK_u64(array);
2341                         data->read.time_running = *array;
2342                         array++;
2343                 }
2344
2345                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2346                 if (read_format & PERF_FORMAT_GROUP) {
2347                         const u64 max_group_nr = UINT64_MAX /
2348                                         sizeof(struct sample_read_value);
2349
2350                         if (data->read.group.nr > max_group_nr)
2351                                 return -EFAULT;
2352                         sz = data->read.group.nr *
2353                              sizeof(struct sample_read_value);
2354                         OVERFLOW_CHECK(array, sz, max_size);
2355                         data->read.group.values =
2356                                         (struct sample_read_value *)array;
2357                         array = (void *)array + sz;
2358                 } else {
2359                         OVERFLOW_CHECK_u64(array);
2360                         data->read.one.id = *array;
2361                         array++;
2362                 }
2363         }
2364
2365         if (type & PERF_SAMPLE_CALLCHAIN) {
2366                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2367
2368                 OVERFLOW_CHECK_u64(array);
2369                 data->callchain = (struct ip_callchain *)array++;
2370                 if (data->callchain->nr > max_callchain_nr)
2371                         return -EFAULT;
2372                 sz = data->callchain->nr * sizeof(u64);
2373                 OVERFLOW_CHECK(array, sz, max_size);
2374                 array = (void *)array + sz;
2375         }
2376
2377         if (type & PERF_SAMPLE_RAW) {
2378                 OVERFLOW_CHECK_u64(array);
2379                 u.val64 = *array;
2380
2381                 /*
2382                  * Undo swap of u64, then swap on individual u32s,
2383                  * get the size of the raw area and undo all of the
2384                  * swap. The pevent interface handles endianness by
2385                  * itself.
2386                  */
2387                 if (swapped) {
2388                         u.val64 = bswap_64(u.val64);
2389                         u.val32[0] = bswap_32(u.val32[0]);
2390                         u.val32[1] = bswap_32(u.val32[1]);
2391                 }
2392                 data->raw_size = u.val32[0];
2393
2394                 /*
2395                  * The raw data is aligned on 64bits including the
2396                  * u32 size, so it's safe to use mem_bswap_64.
2397                  */
2398                 if (swapped)
2399                         mem_bswap_64((void *) array, data->raw_size);
2400
2401                 array = (void *)array + sizeof(u32);
2402
2403                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2404                 data->raw_data = (void *)array;
2405                 array = (void *)array + data->raw_size;
2406         }
2407
2408         if (type & PERF_SAMPLE_BRANCH_STACK) {
2409                 const u64 max_branch_nr = UINT64_MAX /
2410                                           sizeof(struct branch_entry);
2411
2412                 OVERFLOW_CHECK_u64(array);
2413                 data->branch_stack = (struct branch_stack *)array++;
2414
2415                 if (data->branch_stack->nr > max_branch_nr)
2416                         return -EFAULT;
2417
2418                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2419                 if (evsel__has_branch_hw_idx(evsel))
2420                         sz += sizeof(u64);
2421                 else
2422                         data->no_hw_idx = true;
2423                 OVERFLOW_CHECK(array, sz, max_size);
2424                 array = (void *)array + sz;
2425         }
2426
2427         if (type & PERF_SAMPLE_REGS_USER) {
2428                 OVERFLOW_CHECK_u64(array);
2429                 data->user_regs.abi = *array;
2430                 array++;
2431
2432                 if (data->user_regs.abi) {
2433                         u64 mask = evsel->core.attr.sample_regs_user;
2434
2435                         sz = hweight64(mask) * sizeof(u64);
2436                         OVERFLOW_CHECK(array, sz, max_size);
2437                         data->user_regs.mask = mask;
2438                         data->user_regs.regs = (u64 *)array;
2439                         array = (void *)array + sz;
2440                 }
2441         }
2442
2443         if (type & PERF_SAMPLE_STACK_USER) {
2444                 OVERFLOW_CHECK_u64(array);
2445                 sz = *array++;
2446
2447                 data->user_stack.offset = ((char *)(array - 1)
2448                                           - (char *) event);
2449
2450                 if (!sz) {
2451                         data->user_stack.size = 0;
2452                 } else {
2453                         OVERFLOW_CHECK(array, sz, max_size);
2454                         data->user_stack.data = (char *)array;
2455                         array = (void *)array + sz;
2456                         OVERFLOW_CHECK_u64(array);
2457                         data->user_stack.size = *array++;
2458                         if (WARN_ONCE(data->user_stack.size > sz,
2459                                       "user stack dump failure\n"))
2460                                 return -EFAULT;
2461                 }
2462         }
2463
2464         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2465                 OVERFLOW_CHECK_u64(array);
2466                 arch_perf_parse_sample_weight(data, array, type);
2467                 array++;
2468         }
2469
2470         if (type & PERF_SAMPLE_DATA_SRC) {
2471                 OVERFLOW_CHECK_u64(array);
2472                 data->data_src = *array;
2473                 array++;
2474         }
2475
2476         if (type & PERF_SAMPLE_TRANSACTION) {
2477                 OVERFLOW_CHECK_u64(array);
2478                 data->transaction = *array;
2479                 array++;
2480         }
2481
2482         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2483         if (type & PERF_SAMPLE_REGS_INTR) {
2484                 OVERFLOW_CHECK_u64(array);
2485                 data->intr_regs.abi = *array;
2486                 array++;
2487
2488                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2489                         u64 mask = evsel->core.attr.sample_regs_intr;
2490
2491                         sz = hweight64(mask) * sizeof(u64);
2492                         OVERFLOW_CHECK(array, sz, max_size);
2493                         data->intr_regs.mask = mask;
2494                         data->intr_regs.regs = (u64 *)array;
2495                         array = (void *)array + sz;
2496                 }
2497         }
2498
2499         data->phys_addr = 0;
2500         if (type & PERF_SAMPLE_PHYS_ADDR) {
2501                 data->phys_addr = *array;
2502                 array++;
2503         }
2504
2505         data->cgroup = 0;
2506         if (type & PERF_SAMPLE_CGROUP) {
2507                 data->cgroup = *array;
2508                 array++;
2509         }
2510
2511         data->data_page_size = 0;
2512         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2513                 data->data_page_size = *array;
2514                 array++;
2515         }
2516
2517         data->code_page_size = 0;
2518         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2519                 data->code_page_size = *array;
2520                 array++;
2521         }
2522
2523         if (type & PERF_SAMPLE_AUX) {
2524                 OVERFLOW_CHECK_u64(array);
2525                 sz = *array++;
2526
2527                 OVERFLOW_CHECK(array, sz, max_size);
2528                 /* Undo swap of data */
2529                 if (swapped)
2530                         mem_bswap_64((char *)array, sz);
2531                 data->aux_sample.size = sz;
2532                 data->aux_sample.data = (char *)array;
2533                 array = (void *)array + sz;
2534         }
2535
2536         return 0;
2537 }
2538
2539 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2540                                   u64 *timestamp)
2541 {
2542         u64 type = evsel->core.attr.sample_type;
2543         const __u64 *array;
2544
2545         if (!(type & PERF_SAMPLE_TIME))
2546                 return -1;
2547
2548         if (event->header.type != PERF_RECORD_SAMPLE) {
2549                 struct perf_sample data = {
2550                         .time = -1ULL,
2551                 };
2552
2553                 if (!evsel->core.attr.sample_id_all)
2554                         return -1;
2555                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2556                         return -1;
2557
2558                 *timestamp = data.time;
2559                 return 0;
2560         }
2561
2562         array = event->sample.array;
2563
2564         if (perf_event__check_size(event, evsel->sample_size))
2565                 return -EFAULT;
2566
2567         if (type & PERF_SAMPLE_IDENTIFIER)
2568                 array++;
2569
2570         if (type & PERF_SAMPLE_IP)
2571                 array++;
2572
2573         if (type & PERF_SAMPLE_TID)
2574                 array++;
2575
2576         if (type & PERF_SAMPLE_TIME)
2577                 *timestamp = *array;
2578
2579         return 0;
2580 }
2581
2582 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2583 {
2584         return tep_find_field(evsel->tp_format, name);
2585 }
2586
2587 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2588 {
2589         struct tep_format_field *field = evsel__field(evsel, name);
2590         int offset;
2591
2592         if (!field)
2593                 return NULL;
2594
2595         offset = field->offset;
2596
2597         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2598                 offset = *(int *)(sample->raw_data + field->offset);
2599                 offset &= 0xffff;
2600         }
2601
2602         return sample->raw_data + offset;
2603 }
2604
2605 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2606                          bool needs_swap)
2607 {
2608         u64 value;
2609         void *ptr = sample->raw_data + field->offset;
2610
2611         switch (field->size) {
2612         case 1:
2613                 return *(u8 *)ptr;
2614         case 2:
2615                 value = *(u16 *)ptr;
2616                 break;
2617         case 4:
2618                 value = *(u32 *)ptr;
2619                 break;
2620         case 8:
2621                 memcpy(&value, ptr, sizeof(u64));
2622                 break;
2623         default:
2624                 return 0;
2625         }
2626
2627         if (!needs_swap)
2628                 return value;
2629
2630         switch (field->size) {
2631         case 2:
2632                 return bswap_16(value);
2633         case 4:
2634                 return bswap_32(value);
2635         case 8:
2636                 return bswap_64(value);
2637         default:
2638                 return 0;
2639         }
2640
2641         return 0;
2642 }
2643
2644 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2645 {
2646         struct tep_format_field *field = evsel__field(evsel, name);
2647
2648         if (!field)
2649                 return 0;
2650
2651         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2652 }
2653
2654 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2655 {
2656         int paranoid;
2657
2658         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2659             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2660             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2661                 /*
2662                  * If it's cycles then fall back to hrtimer based
2663                  * cpu-clock-tick sw counter, which is always available even if
2664                  * no PMU support.
2665                  *
2666                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2667                  * b0a873e).
2668                  */
2669                 scnprintf(msg, msgsize, "%s",
2670 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2671
2672                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2673                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2674
2675                 zfree(&evsel->name);
2676                 return true;
2677         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2678                    (paranoid = perf_event_paranoid()) > 1) {
2679                 const char *name = evsel__name(evsel);
2680                 char *new_name;
2681                 const char *sep = ":";
2682
2683                 /* If event has exclude user then don't exclude kernel. */
2684                 if (evsel->core.attr.exclude_user)
2685                         return false;
2686
2687                 /* Is there already the separator in the name. */
2688                 if (strchr(name, '/') ||
2689                     (strchr(name, ':') && !evsel->is_libpfm_event))
2690                         sep = "";
2691
2692                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2693                         return false;
2694
2695                 if (evsel->name)
2696                         free(evsel->name);
2697                 evsel->name = new_name;
2698                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2699                           "to fall back to excluding kernel and hypervisor "
2700                           " samples", paranoid);
2701                 evsel->core.attr.exclude_kernel = 1;
2702                 evsel->core.attr.exclude_hv     = 1;
2703
2704                 return true;
2705         }
2706
2707         return false;
2708 }
2709
2710 static bool find_process(const char *name)
2711 {
2712         size_t len = strlen(name);
2713         DIR *dir;
2714         struct dirent *d;
2715         int ret = -1;
2716
2717         dir = opendir(procfs__mountpoint());
2718         if (!dir)
2719                 return false;
2720
2721         /* Walk through the directory. */
2722         while (ret && (d = readdir(dir)) != NULL) {
2723                 char path[PATH_MAX];
2724                 char *data;
2725                 size_t size;
2726
2727                 if ((d->d_type != DT_DIR) ||
2728                      !strcmp(".", d->d_name) ||
2729                      !strcmp("..", d->d_name))
2730                         continue;
2731
2732                 scnprintf(path, sizeof(path), "%s/%s/comm",
2733                           procfs__mountpoint(), d->d_name);
2734
2735                 if (filename__read_str(path, &data, &size))
2736                         continue;
2737
2738                 ret = strncmp(name, data, len);
2739                 free(data);
2740         }
2741
2742         closedir(dir);
2743         return ret ? false : true;
2744 }
2745
2746 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2747                          int err, char *msg, size_t size)
2748 {
2749         char sbuf[STRERR_BUFSIZE];
2750         int printed = 0, enforced = 0;
2751
2752         switch (err) {
2753         case EPERM:
2754         case EACCES:
2755                 printed += scnprintf(msg + printed, size - printed,
2756                         "Access to performance monitoring and observability operations is limited.\n");
2757
2758                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2759                         if (enforced) {
2760                                 printed += scnprintf(msg + printed, size - printed,
2761                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2762                                         "monitoring and observability operations. Inspect system audit records for\n"
2763                                         "more perf_event access control information and adjusting the policy.\n");
2764                         }
2765                 }
2766
2767                 if (err == EPERM)
2768                         printed += scnprintf(msg, size,
2769                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2770
2771                 return scnprintf(msg + printed, size - printed,
2772                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2773                  "access to performance monitoring and observability operations for processes\n"
2774                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2775                  "More information can be found at 'Perf events and tool security' document:\n"
2776                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2777                  "perf_event_paranoid setting is %d:\n"
2778                  "  -1: Allow use of (almost) all events by all users\n"
2779                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2780                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2781                  ">= 1: Disallow CPU event access\n"
2782                  ">= 2: Disallow kernel profiling\n"
2783                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2784                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2785                  perf_event_paranoid());
2786         case ENOENT:
2787                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2788         case EMFILE:
2789                 return scnprintf(msg, size, "%s",
2790                          "Too many events are opened.\n"
2791                          "Probably the maximum number of open file descriptors has been reached.\n"
2792                          "Hint: Try again after reducing the number of events.\n"
2793                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2794         case ENOMEM:
2795                 if (evsel__has_callchain(evsel) &&
2796                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2797                         return scnprintf(msg, size,
2798                                          "Not enough memory to setup event with callchain.\n"
2799                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2800                                          "Hint: Current value: %d", sysctl__max_stack());
2801                 break;
2802         case ENODEV:
2803                 if (target->cpu_list)
2804                         return scnprintf(msg, size, "%s",
2805          "No such device - did you specify an out-of-range profile CPU?");
2806                 break;
2807         case EOPNOTSUPP:
2808                 if (evsel->core.attr.aux_output)
2809                         return scnprintf(msg, size,
2810         "%s: PMU Hardware doesn't support 'aux_output' feature",
2811                                          evsel__name(evsel));
2812                 if (evsel->core.attr.sample_period != 0)
2813                         return scnprintf(msg, size,
2814         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
2815                                          evsel__name(evsel));
2816                 if (evsel->core.attr.precise_ip)
2817                         return scnprintf(msg, size, "%s",
2818         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2819 #if defined(__i386__) || defined(__x86_64__)
2820                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
2821                         return scnprintf(msg, size, "%s",
2822         "No hardware sampling interrupt available.\n");
2823 #endif
2824                 break;
2825         case EBUSY:
2826                 if (find_process("oprofiled"))
2827                         return scnprintf(msg, size,
2828         "The PMU counters are busy/taken by another profiler.\n"
2829         "We found oprofile daemon running, please stop it and try again.");
2830                 break;
2831         case EINVAL:
2832                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
2833                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
2834                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
2835                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
2836                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
2837                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2838                 if (perf_missing_features.clockid)
2839                         return scnprintf(msg, size, "clockid feature not supported.");
2840                 if (perf_missing_features.clockid_wrong)
2841                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2842                 if (perf_missing_features.aux_output)
2843                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
2844                 break;
2845         case ENODATA:
2846                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
2847                                  "Please add an auxiliary event in front of the load latency event.");
2848         default:
2849                 break;
2850         }
2851
2852         return scnprintf(msg, size,
2853         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2854         "/bin/dmesg | grep -i perf may provide additional information.\n",
2855                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
2856 }
2857
2858 struct perf_env *evsel__env(struct evsel *evsel)
2859 {
2860         if (evsel && evsel->evlist)
2861                 return evsel->evlist->env;
2862         return &perf_env;
2863 }
2864
2865 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
2866 {
2867         int cpu, thread;
2868
2869         for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) {
2870                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
2871                      thread++) {
2872                         int fd = FD(evsel, cpu, thread);
2873
2874                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
2875                                                    cpu, thread, fd) < 0)
2876                                 return -1;
2877                 }
2878         }
2879
2880         return 0;
2881 }
2882
2883 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
2884 {
2885         struct perf_cpu_map *cpus = evsel->core.cpus;
2886         struct perf_thread_map *threads = evsel->core.threads;
2887
2888         if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr))
2889                 return -ENOMEM;
2890
2891         return store_evsel_ids(evsel, evlist);
2892 }
2893
2894 void evsel__zero_per_pkg(struct evsel *evsel)
2895 {
2896         struct hashmap_entry *cur;
2897         size_t bkt;
2898
2899         if (evsel->per_pkg_mask) {
2900                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
2901                         free((char *)cur->key);
2902
2903                 hashmap__clear(evsel->per_pkg_mask);
2904         }
2905 }
2906
2907 bool evsel__is_hybrid(struct evsel *evsel)
2908 {
2909         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
2910 }
2911
2912 struct evsel *evsel__leader(struct evsel *evsel)
2913 {
2914         return container_of(evsel->core.leader, struct evsel, core);
2915 }
2916
2917 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
2918 {
2919         return evsel->core.leader == &leader->core;
2920 }
2921
2922 bool evsel__is_leader(struct evsel *evsel)
2923 {
2924         return evsel__has_leader(evsel, evsel);
2925 }
2926
2927 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
2928 {
2929         evsel->core.leader = &leader->core;
2930 }