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 <[email protected]>
   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 #ifdef HAVE_LIBBPF_SUPPORT
  50 #include <bpf/hashmap.h>
  51 #else
  52 #include "util/hashmap.h"
  53 #endif
  54 #include "pmu-hybrid.h"
  55 #include "off_cpu.h"
  56 #include "../perf-sys.h"
  57 #include "util/parse-branch-options.h"
  58 #include <internal/xyarray.h>
  59 #include <internal/lib.h>
  60
  61 #include <linux/ctype.h>
  62
  63 struct perf_missing_features perf_missing_features;
  64
  65 static clockid_t clockid;
  66
  67 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  68         NULL,
  69         "duration_time",
  70         "user_time",
  71         "system_time",
  72 };
  73
  74 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  75 {
  76         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  77                 return perf_tool_event__tool_names[ev];
  78
  79         return NULL;
  80 }
  81
  82 enum perf_tool_event perf_tool_event__from_str(const char *str)
  83 {
  84         int i;
  85
  86         perf_tool_event__for_each_event(i) {
  87                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  88                         return i;
  89         }
  90         return PERF_TOOL_NONE;
  91 }
  92
  93
  94 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  95 {
  96         return 0;
  97 }
  98
  99 void __weak test_attr__ready(void) { }
 100
 101 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 102 {
 103 }
 104
 105 static struct {
 106         size_t  size;
 107         int     (*init)(struct evsel *evsel);
 108         void    (*fini)(struct evsel *evsel);
 109 } perf_evsel__object = {
 110         .size = sizeof(struct evsel),
 111         .init = evsel__no_extra_init,
 112         .fini = evsel__no_extra_fini,
 113 };
 114
 115 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 116                          void (*fini)(struct evsel *evsel))
 117 {
 118
 119         if (object_size == 0)
 120                 goto set_methods;
 121
 122         if (perf_evsel__object.size > object_size)
 123                 return -EINVAL;
 124
 125         perf_evsel__object.size = object_size;
 126
 127 set_methods:
 128         if (init != NULL)
 129                 perf_evsel__object.init = init;
 130
 131         if (fini != NULL)
 132                 perf_evsel__object.fini = fini;
 133
 134         return 0;
 135 }
 136
 137 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 138
 139 int __evsel__sample_size(u64 sample_type)
 140 {
 141         u64 mask = sample_type & PERF_SAMPLE_MASK;
 142         int size = 0;
 143         int i;
 144
 145         for (i = 0; i < 64; i++) {
 146                 if (mask & (1ULL << i))
 147                         size++;
 148         }
 149
 150         size *= sizeof(u64);
 151
 152         return size;
 153 }
 154
 155 /**
 156  * __perf_evsel__calc_id_pos - calculate id_pos.
 157  * @sample_type: sample type
 158  *
 159  * This function returns the position of the event id (PERF_SAMPLE_ID or
 160  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 161  * perf_record_sample.
 162  */
 163 static int __perf_evsel__calc_id_pos(u64 sample_type)
 164 {
 165         int idx = 0;
 166
 167         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 168                 return 0;
 169
 170         if (!(sample_type & PERF_SAMPLE_ID))
 171                 return -1;
 172
 173         if (sample_type & PERF_SAMPLE_IP)
 174                 idx += 1;
 175
 176         if (sample_type & PERF_SAMPLE_TID)
 177                 idx += 1;
 178
 179         if (sample_type & PERF_SAMPLE_TIME)
 180                 idx += 1;
 181
 182         if (sample_type & PERF_SAMPLE_ADDR)
 183                 idx += 1;
 184
 185         return idx;
 186 }
 187
 188 /**
 189  * __perf_evsel__calc_is_pos - calculate is_pos.
 190  * @sample_type: sample type
 191  *
 192  * This function returns the position (counting backwards) of the event id
 193  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 194  * sample_id_all is used there is an id sample appended to non-sample events.
 195  */
 196 static int __perf_evsel__calc_is_pos(u64 sample_type)
 197 {
 198         int idx = 1;
 199
 200         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 201                 return 1;
 202
 203         if (!(sample_type & PERF_SAMPLE_ID))
 204                 return -1;
 205
 206         if (sample_type & PERF_SAMPLE_CPU)
 207                 idx += 1;
 208
 209         if (sample_type & PERF_SAMPLE_STREAM_ID)
 210                 idx += 1;
 211
 212         return idx;
 213 }
 214
 215 void evsel__calc_id_pos(struct evsel *evsel)
 216 {
 217         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 218         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 219 }
 220
 221 void __evsel__set_sample_bit(struct evsel *evsel,
 222                                   enum perf_event_sample_format bit)
 223 {
 224         if (!(evsel->core.attr.sample_type & bit)) {
 225                 evsel->core.attr.sample_type |= bit;
 226                 evsel->sample_size += sizeof(u64);
 227                 evsel__calc_id_pos(evsel);
 228         }
 229 }
 230
 231 void __evsel__reset_sample_bit(struct evsel *evsel,
 232                                     enum perf_event_sample_format bit)
 233 {
 234         if (evsel->core.attr.sample_type & bit) {
 235                 evsel->core.attr.sample_type &= ~bit;
 236                 evsel->sample_size -= sizeof(u64);
 237                 evsel__calc_id_pos(evsel);
 238         }
 239 }
 240
 241 void evsel__set_sample_id(struct evsel *evsel,
 242                                bool can_sample_identifier)
 243 {
 244         if (can_sample_identifier) {
 245                 evsel__reset_sample_bit(evsel, ID);
 246                 evsel__set_sample_bit(evsel, IDENTIFIER);
 247         } else {
 248                 evsel__set_sample_bit(evsel, ID);
 249         }
 250         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 251 }
 252
 253 /**
 254  * evsel__is_function_event - Return whether given evsel is a function
 255  * trace event
 256  *
 257  * @evsel - evsel selector to be tested
 258  *
 259  * Return %true if event is function trace event
 260  */
 261 bool evsel__is_function_event(struct evsel *evsel)
 262 {
 263 #define FUNCTION_EVENT "ftrace:function"
 264
 265         return evsel->name &&
 266                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 267
 268 #undef FUNCTION_EVENT
 269 }
 270
 271 void evsel__init(struct evsel *evsel,
 272                  struct perf_event_attr *attr, int idx)
 273 {
 274         perf_evsel__init(&evsel->core, attr, idx);
 275         evsel->tracking    = !idx;
 276         evsel->unit        = strdup("");
 277         evsel->scale       = 1.0;
 278         evsel->max_events  = ULONG_MAX;
 279         evsel->evlist      = NULL;
 280         evsel->bpf_obj     = NULL;
 281         evsel->bpf_fd      = -1;
 282         INIT_LIST_HEAD(&evsel->config_terms);
 283         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 284         perf_evsel__object.init(evsel);
 285         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 286         evsel__calc_id_pos(evsel);
 287         evsel->cmdline_group_boundary = false;
 288         evsel->metric_expr   = NULL;
 289         evsel->metric_name   = NULL;
 290         evsel->metric_events = NULL;
 291         evsel->per_pkg_mask  = NULL;
 292         evsel->collect_stat  = false;
 293         evsel->pmu_name      = NULL;
 294 }
 295
 296 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 297 {
 298         struct evsel *evsel = zalloc(perf_evsel__object.size);
 299
 300         if (!evsel)
 301                 return NULL;
 302         evsel__init(evsel, attr, idx);
 303
 304         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 305                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 306                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 307                 evsel->core.attr.sample_period = 1;
 308         }
 309
 310         if (evsel__is_clock(evsel)) {
 311                 free((char *)evsel->unit);
 312                 evsel->unit = strdup("msec");
 313                 evsel->scale = 1e-6;
 314         }
 315
 316         return evsel;
 317 }
 318
 319 static bool perf_event_can_profile_kernel(void)
 320 {
 321         return perf_event_paranoid_check(1);
 322 }
 323
 324 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 325 {
 326         struct perf_event_attr attr = {
 327                 .type   = type,
 328                 .config = config,
 329                 .exclude_kernel = !perf_event_can_profile_kernel(),
 330         };
 331         struct evsel *evsel;
 332
 333         event_attr_init(&attr);
 334
 335         /*
 336          * Now let the usual logic to set up the perf_event_attr defaults
 337          * to kick in when we return and before perf_evsel__open() is called.
 338          */
 339         evsel = evsel__new(&attr);
 340         if (evsel == NULL)
 341                 goto out;
 342
 343         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 344
 345         evsel->precise_max = true;
 346
 347         /* use asprintf() because free(evsel) assumes name is allocated */
 348         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 349                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 350                      attr.exclude_kernel ? "u" : "",
 351                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 352                 goto error_free;
 353 out:
 354         return evsel;
 355 error_free:
 356         evsel__delete(evsel);
 357         evsel = NULL;
 358         goto out;
 359 }
 360
 361 int copy_config_terms(struct list_head *dst, struct list_head *src)
 362 {
 363         struct evsel_config_term *pos, *tmp;
 364
 365         list_for_each_entry(pos, src, list) {
 366                 tmp = malloc(sizeof(*tmp));
 367                 if (tmp == NULL)
 368                         return -ENOMEM;
 369
 370                 *tmp = *pos;
 371                 if (tmp->free_str) {
 372                         tmp->val.str = strdup(pos->val.str);
 373                         if (tmp->val.str == NULL) {
 374                                 free(tmp);
 375                                 return -ENOMEM;
 376                         }
 377                 }
 378                 list_add_tail(&tmp->list, dst);
 379         }
 380         return 0;
 381 }
 382
 383 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 384 {
 385         return copy_config_terms(&dst->config_terms, &src->config_terms);
 386 }
 387
 388 /**
 389  * evsel__clone - create a new evsel copied from @orig
 390  * @orig: original evsel
 391  *
 392  * The assumption is that @orig is not configured nor opened yet.
 393  * So we only care about the attributes that can be set while it's parsed.
 394  */
 395 struct evsel *evsel__clone(struct evsel *orig)
 396 {
 397         struct evsel *evsel;
 398
 399         BUG_ON(orig->core.fd);
 400         BUG_ON(orig->counts);
 401         BUG_ON(orig->priv);
 402         BUG_ON(orig->per_pkg_mask);
 403
 404         /* cannot handle BPF objects for now */
 405         if (orig->bpf_obj)
 406                 return NULL;
 407
 408         evsel = evsel__new(&orig->core.attr);
 409         if (evsel == NULL)
 410                 return NULL;
 411
 412         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 413         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 414         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 415         evsel->core.nr_members = orig->core.nr_members;
 416         evsel->core.system_wide = orig->core.system_wide;
 417         evsel->core.requires_cpu = orig->core.requires_cpu;
 418
 419         if (orig->name) {
 420                 evsel->name = strdup(orig->name);
 421                 if (evsel->name == NULL)
 422                         goto out_err;
 423         }
 424         if (orig->group_name) {
 425                 evsel->group_name = strdup(orig->group_name);
 426                 if (evsel->group_name == NULL)
 427                         goto out_err;
 428         }
 429         if (orig->pmu_name) {
 430                 evsel->pmu_name = strdup(orig->pmu_name);
 431                 if (evsel->pmu_name == NULL)
 432                         goto out_err;
 433         }
 434         if (orig->filter) {
 435                 evsel->filter = strdup(orig->filter);
 436                 if (evsel->filter == NULL)
 437                         goto out_err;
 438         }
 439         if (orig->metric_id) {
 440                 evsel->metric_id = strdup(orig->metric_id);
 441                 if (evsel->metric_id == NULL)
 442                         goto out_err;
 443         }
 444         evsel->cgrp = cgroup__get(orig->cgrp);
 445         evsel->tp_format = orig->tp_format;
 446         evsel->handler = orig->handler;
 447         evsel->core.leader = orig->core.leader;
 448
 449         evsel->max_events = orig->max_events;
 450         evsel->tool_event = orig->tool_event;
 451         free((char *)evsel->unit);
 452         evsel->unit = strdup(orig->unit);
 453         if (evsel->unit == NULL)
 454                 goto out_err;
 455
 456         evsel->scale = orig->scale;
 457         evsel->snapshot = orig->snapshot;
 458         evsel->per_pkg = orig->per_pkg;
 459         evsel->percore = orig->percore;
 460         evsel->precise_max = orig->precise_max;
 461         evsel->use_uncore_alias = orig->use_uncore_alias;
 462         evsel->is_libpfm_event = orig->is_libpfm_event;
 463
 464         evsel->exclude_GH = orig->exclude_GH;
 465         evsel->sample_read = orig->sample_read;
 466         evsel->auto_merge_stats = orig->auto_merge_stats;
 467         evsel->collect_stat = orig->collect_stat;
 468         evsel->weak_group = orig->weak_group;
 469         evsel->use_config_name = orig->use_config_name;
 470
 471         if (evsel__copy_config_terms(evsel, orig) < 0)
 472                 goto out_err;
 473
 474         return evsel;
 475
 476 out_err:
 477         evsel__delete(evsel);
 478         return NULL;
 479 }
 480
 481 /*
 482  * Returns pointer with encoded error via <linux/err.h> interface.
 483  */
 484 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 485 {
 486         struct evsel *evsel = zalloc(perf_evsel__object.size);
 487         int err = -ENOMEM;
 488
 489         if (evsel == NULL) {
 490                 goto out_err;
 491         } else {
 492                 struct perf_event_attr attr = {
 493                         .type          = PERF_TYPE_TRACEPOINT,
 494                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 495                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 496                 };
 497
 498                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 499                         goto out_free;
 500
 501                 evsel->tp_format = trace_event__tp_format(sys, name);
 502                 if (IS_ERR(evsel->tp_format)) {
 503                         err = PTR_ERR(evsel->tp_format);
 504                         goto out_free;
 505                 }
 506
 507                 event_attr_init(&attr);
 508                 attr.config = evsel->tp_format->id;
 509                 attr.sample_period = 1;
 510                 evsel__init(evsel, &attr, idx);
 511         }
 512
 513         return evsel;
 514
 515 out_free:
 516         zfree(&evsel->name);
 517         free(evsel);
 518 out_err:
 519         return ERR_PTR(err);
 520 }
 521
 522 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 523         "cycles",
 524         "instructions",
 525         "cache-references",
 526         "cache-misses",
 527         "branches",
 528         "branch-misses",
 529         "bus-cycles",
 530         "stalled-cycles-frontend",
 531         "stalled-cycles-backend",
 532         "ref-cycles",
 533 };
 534
 535 char *evsel__bpf_counter_events;
 536
 537 bool evsel__match_bpf_counter_events(const char *name)
 538 {
 539         int name_len;
 540         bool match;
 541         char *ptr;
 542
 543         if (!evsel__bpf_counter_events)
 544                 return false;
 545
 546         ptr = strstr(evsel__bpf_counter_events, name);
 547         name_len = strlen(name);
 548
 549         /* check name matches a full token in evsel__bpf_counter_events */
 550         match = (ptr != NULL) &&
 551                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 552                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 553
 554         return match;
 555 }
 556
 557 static const char *__evsel__hw_name(u64 config)
 558 {
 559         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 560                 return evsel__hw_names[config];
 561
 562         return "unknown-hardware";
 563 }
 564
 565 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 566 {
 567         int colon = 0, r = 0;
 568         struct perf_event_attr *attr = &evsel->core.attr;
 569         bool exclude_guest_default = false;
 570
 571 #define MOD_PRINT(context, mod) do {                                    \
 572                 if (!attr->exclude_##context) {                         \
 573                         if (!colon) colon = ++r;                        \
 574                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 575                 } } while(0)
 576
 577         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 578                 MOD_PRINT(kernel, 'k');
 579                 MOD_PRINT(user, 'u');
 580                 MOD_PRINT(hv, 'h');
 581                 exclude_guest_default = true;
 582         }
 583
 584         if (attr->precise_ip) {
 585                 if (!colon)
 586                         colon = ++r;
 587                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 588                 exclude_guest_default = true;
 589         }
 590
 591         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 592                 MOD_PRINT(host, 'H');
 593                 MOD_PRINT(guest, 'G');
 594         }
 595 #undef MOD_PRINT
 596         if (colon)
 597                 bf[colon - 1] = ':';
 598         return r;
 599 }
 600
 601 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 602 {
 603         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 604 }
 605
 606 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 607 {
 608         int r = arch_evsel__hw_name(evsel, bf, size);
 609         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 610 }
 611
 612 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 613         "cpu-clock",
 614         "task-clock",
 615         "page-faults",
 616         "context-switches",
 617         "cpu-migrations",
 618         "minor-faults",
 619         "major-faults",
 620         "alignment-faults",
 621         "emulation-faults",
 622         "dummy",
 623 };
 624
 625 static const char *__evsel__sw_name(u64 config)
 626 {
 627         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 628                 return evsel__sw_names[config];
 629         return "unknown-software";
 630 }
 631
 632 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 633 {
 634         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 635         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 636 }
 637
 638 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 639 {
 640         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 641 }
 642
 643 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 644 {
 645         int r;
 646
 647         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 648
 649         if (type & HW_BREAKPOINT_R)
 650                 r += scnprintf(bf + r, size - r, "r");
 651
 652         if (type & HW_BREAKPOINT_W)
 653                 r += scnprintf(bf + r, size - r, "w");
 654
 655         if (type & HW_BREAKPOINT_X)
 656                 r += scnprintf(bf + r, size - r, "x");
 657
 658         return r;
 659 }
 660
 661 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 662 {
 663         struct perf_event_attr *attr = &evsel->core.attr;
 664         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 665         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 666 }
 667
 668 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 669  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 670  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 671  { "LLC",       "L2",                                                   },
 672  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 673  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 674  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 675  { "node",                                                              },
 676 };
 677
 678 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 679  { "load",      "loads",        "read",                                 },
 680  { "store",     "stores",       "write",                                },
 681  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 682 };
 683
 684 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 685  { "refs",      "Reference",    "ops",          "access",               },
 686  { "misses",    "miss",                                                 },
 687 };
 688
 689 #define C(x)            PERF_COUNT_HW_CACHE_##x
 690 #define CACHE_READ      (1 << C(OP_READ))
 691 #define CACHE_WRITE     (1 << C(OP_WRITE))
 692 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 693 #define COP(x)          (1 << x)
 694
 695 /*
 696  * cache operation stat
 697  * L1I : Read and prefetch only
 698  * ITLB and BPU : Read-only
 699  */
 700 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 701  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 702  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 703  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 704  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 705  [C(ITLB)]      = (CACHE_READ),
 706  [C(BPU)]       = (CACHE_READ),
 707  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 708 };
 709
 710 bool evsel__is_cache_op_valid(u8 type, u8 op)
 711 {
 712         if (evsel__hw_cache_stat[type] & COP(op))
 713                 return true;    /* valid */
 714         else
 715                 return false;   /* invalid */
 716 }
 717
 718 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 719 {
 720         if (result) {
 721                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 722                                  evsel__hw_cache_op[op][0],
 723                                  evsel__hw_cache_result[result][0]);
 724         }
 725
 726         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 727                          evsel__hw_cache_op[op][1]);
 728 }
 729
 730 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 731 {
 732         u8 op, result, type = (config >>  0) & 0xff;
 733         const char *err = "unknown-ext-hardware-cache-type";
 734
 735         if (type >= PERF_COUNT_HW_CACHE_MAX)
 736                 goto out_err;
 737
 738         op = (config >>  8) & 0xff;
 739         err = "unknown-ext-hardware-cache-op";
 740         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 741                 goto out_err;
 742
 743         result = (config >> 16) & 0xff;
 744         err = "unknown-ext-hardware-cache-result";
 745         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 746                 goto out_err;
 747
 748         err = "invalid-cache";
 749         if (!evsel__is_cache_op_valid(type, op))
 750                 goto out_err;
 751
 752         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 753 out_err:
 754         return scnprintf(bf, size, "%s", err);
 755 }
 756
 757 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 758 {
 759         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 760         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 761 }
 762
 763 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 764 {
 765         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 766         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 767 }
 768
 769 const char *evsel__name(struct evsel *evsel)
 770 {
 771         char bf[128];
 772
 773         if (!evsel)
 774                 goto out_unknown;
 775
 776         if (evsel->name)
 777                 return evsel->name;
 778
 779         switch (evsel->core.attr.type) {
 780         case PERF_TYPE_RAW:
 781                 evsel__raw_name(evsel, bf, sizeof(bf));
 782                 break;
 783
 784         case PERF_TYPE_HARDWARE:
 785                 evsel__hw_name(evsel, bf, sizeof(bf));
 786                 break;
 787
 788         case PERF_TYPE_HW_CACHE:
 789                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 790                 break;
 791
 792         case PERF_TYPE_SOFTWARE:
 793                 if (evsel__is_tool(evsel))
 794                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 795                 else
 796                         evsel__sw_name(evsel, bf, sizeof(bf));
 797                 break;
 798
 799         case PERF_TYPE_TRACEPOINT:
 800                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 801                 break;
 802
 803         case PERF_TYPE_BREAKPOINT:
 804                 evsel__bp_name(evsel, bf, sizeof(bf));
 805                 break;
 806
 807         default:
 808                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 809                           evsel->core.attr.type);
 810                 break;
 811         }
 812
 813         evsel->name = strdup(bf);
 814
 815         if (evsel->name)
 816                 return evsel->name;
 817 out_unknown:
 818         return "unknown";
 819 }
 820
 821 const char *evsel__metric_id(const struct evsel *evsel)
 822 {
 823         if (evsel->metric_id)
 824                 return evsel->metric_id;
 825
 826         if (evsel__is_tool(evsel))
 827                 return perf_tool_event__to_str(evsel->tool_event);
 828
 829         return "unknown";
 830 }
 831
 832 const char *evsel__group_name(struct evsel *evsel)
 833 {
 834         return evsel->group_name ?: "anon group";
 835 }
 836
 837 /*
 838  * Returns the group details for the specified leader,
 839  * with following rules.
 840  *
 841  *  For record -e '{cycles,instructions}'
 842  *    'anon group { cycles:u, instructions:u }'
 843  *
 844  *  For record -e 'cycles,instructions' and report --group
 845  *    'cycles:u, instructions:u'
 846  */
 847 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 848 {
 849         int ret = 0;
 850         struct evsel *pos;
 851         const char *group_name = evsel__group_name(evsel);
 852
 853         if (!evsel->forced_leader)
 854                 ret = scnprintf(buf, size, "%s { ", group_name);
 855
 856         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 857
 858         for_each_group_member(pos, evsel)
 859                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 860
 861         if (!evsel->forced_leader)
 862                 ret += scnprintf(buf + ret, size - ret, " }");
 863
 864         return ret;
 865 }
 866
 867 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 868                                       struct callchain_param *param)
 869 {
 870         bool function = evsel__is_function_event(evsel);
 871         struct perf_event_attr *attr = &evsel->core.attr;
 872
 873         evsel__set_sample_bit(evsel, CALLCHAIN);
 874
 875         attr->sample_max_stack = param->max_stack;
 876
 877         if (opts->kernel_callchains)
 878                 attr->exclude_callchain_user = 1;
 879         if (opts->user_callchains)
 880                 attr->exclude_callchain_kernel = 1;
 881         if (param->record_mode == CALLCHAIN_LBR) {
 882                 if (!opts->branch_stack) {
 883                         if (attr->exclude_user) {
 884                                 pr_warning("LBR callstack option is only available "
 885                                            "to get user callchain information. "
 886                                            "Falling back to framepointers.\n");
 887                         } else {
 888                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 889                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 890                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 891                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 892                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 893                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 894                         }
 895                 } else
 896                          pr_warning("Cannot use LBR callstack with branch stack. "
 897                                     "Falling back to framepointers.\n");
 898         }
 899
 900         if (param->record_mode == CALLCHAIN_DWARF) {
 901                 if (!function) {
 902                         evsel__set_sample_bit(evsel, REGS_USER);
 903                         evsel__set_sample_bit(evsel, STACK_USER);
 904                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 905                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 906                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 907                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 908                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 909                         } else {
 910                                 attr->sample_regs_user |= arch__user_reg_mask();
 911                         }
 912                         attr->sample_stack_user = param->dump_size;
 913                         attr->exclude_callchain_user = 1;
 914                 } else {
 915                         pr_info("Cannot use DWARF unwind for function trace event,"
 916                                 " falling back to framepointers.\n");
 917                 }
 918         }
 919
 920         if (function) {
 921                 pr_info("Disabling user space callchains for function trace event.\n");
 922                 attr->exclude_callchain_user = 1;
 923         }
 924 }
 925
 926 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 927                              struct callchain_param *param)
 928 {
 929         if (param->enabled)
 930                 return __evsel__config_callchain(evsel, opts, param);
 931 }
 932
 933 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 934 {
 935         struct perf_event_attr *attr = &evsel->core.attr;
 936
 937         evsel__reset_sample_bit(evsel, CALLCHAIN);
 938         if (param->record_mode == CALLCHAIN_LBR) {
 939                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 940                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 941                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 942                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 943         }
 944         if (param->record_mode == CALLCHAIN_DWARF) {
 945                 evsel__reset_sample_bit(evsel, REGS_USER);
 946                 evsel__reset_sample_bit(evsel, STACK_USER);
 947         }
 948 }
 949
 950 static void evsel__apply_config_terms(struct evsel *evsel,
 951                                       struct record_opts *opts, bool track)
 952 {
 953         struct evsel_config_term *term;
 954         struct list_head *config_terms = &evsel->config_terms;
 955         struct perf_event_attr *attr = &evsel->core.attr;
 956         /* callgraph default */
 957         struct callchain_param param = {
 958                 .record_mode = callchain_param.record_mode,
 959         };
 960         u32 dump_size = 0;
 961         int max_stack = 0;
 962         const char *callgraph_buf = NULL;
 963
 964         list_for_each_entry(term, config_terms, list) {
 965                 switch (term->type) {
 966                 case EVSEL__CONFIG_TERM_PERIOD:
 967                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 968                                 attr->sample_period = term->val.period;
 969                                 attr->freq = 0;
 970                                 evsel__reset_sample_bit(evsel, PERIOD);
 971                         }
 972                         break;
 973                 case EVSEL__CONFIG_TERM_FREQ:
 974                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 975                                 attr->sample_freq = term->val.freq;
 976                                 attr->freq = 1;
 977                                 evsel__set_sample_bit(evsel, PERIOD);
 978                         }
 979                         break;
 980                 case EVSEL__CONFIG_TERM_TIME:
 981                         if (term->val.time)
 982                                 evsel__set_sample_bit(evsel, TIME);
 983                         else
 984                                 evsel__reset_sample_bit(evsel, TIME);
 985                         break;
 986                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 987                         callgraph_buf = term->val.str;
 988                         break;
 989                 case EVSEL__CONFIG_TERM_BRANCH:
 990                         if (term->val.str && strcmp(term->val.str, "no")) {
 991                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 992                                 parse_branch_str(term->val.str,
 993                                                  &attr->branch_sample_type);
 994                         } else
 995                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 996                         break;
 997                 case EVSEL__CONFIG_TERM_STACK_USER:
 998                         dump_size = term->val.stack_user;
 999                         break;
1000                 case EVSEL__CONFIG_TERM_MAX_STACK:
1001                         max_stack = term->val.max_stack;
1002                         break;
1003                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1004                         evsel->max_events = term->val.max_events;
1005                         break;
1006                 case EVSEL__CONFIG_TERM_INHERIT:
1007                         /*
1008                          * attr->inherit should has already been set by
1009                          * evsel__config. If user explicitly set
1010                          * inherit using config terms, override global
1011                          * opt->no_inherit setting.
1012                          */
1013                         attr->inherit = term->val.inherit ? 1 : 0;
1014                         break;
1015                 case EVSEL__CONFIG_TERM_OVERWRITE:
1016                         attr->write_backward = term->val.overwrite ? 1 : 0;
1017                         break;
1018                 case EVSEL__CONFIG_TERM_DRV_CFG:
1019                         break;
1020                 case EVSEL__CONFIG_TERM_PERCORE:
1021                         break;
1022                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1023                         attr->aux_output = term->val.aux_output ? 1 : 0;
1024                         break;
1025                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1026                         /* Already applied by auxtrace */
1027                         break;
1028                 case EVSEL__CONFIG_TERM_CFG_CHG:
1029                         break;
1030                 default:
1031                         break;
1032                 }
1033         }
1034
1035         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1036         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1037                 bool sample_address = false;
1038
1039                 if (max_stack) {
1040                         param.max_stack = max_stack;
1041                         if (callgraph_buf == NULL)
1042                                 callgraph_buf = "fp";
1043                 }
1044
1045                 /* parse callgraph parameters */
1046                 if (callgraph_buf != NULL) {
1047                         if (!strcmp(callgraph_buf, "no")) {
1048                                 param.enabled = false;
1049                                 param.record_mode = CALLCHAIN_NONE;
1050                         } else {
1051                                 param.enabled = true;
1052                                 if (parse_callchain_record(callgraph_buf, &param)) {
1053                                         pr_err("per-event callgraph setting for %s failed. "
1054                                                "Apply callgraph global setting for it\n",
1055                                                evsel->name);
1056                                         return;
1057                                 }
1058                                 if (param.record_mode == CALLCHAIN_DWARF)
1059                                         sample_address = true;
1060                         }
1061                 }
1062                 if (dump_size > 0) {
1063                         dump_size = round_up(dump_size, sizeof(u64));
1064                         param.dump_size = dump_size;
1065                 }
1066
1067                 /* If global callgraph set, clear it */
1068                 if (callchain_param.enabled)
1069                         evsel__reset_callgraph(evsel, &callchain_param);
1070
1071                 /* set perf-event callgraph */
1072                 if (param.enabled) {
1073                         if (sample_address) {
1074                                 evsel__set_sample_bit(evsel, ADDR);
1075                                 evsel__set_sample_bit(evsel, DATA_SRC);
1076                                 evsel->core.attr.mmap_data = track;
1077                         }
1078                         evsel__config_callchain(evsel, opts, &param);
1079                 }
1080         }
1081 }
1082
1083 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1084 {
1085         struct evsel_config_term *term, *found_term = NULL;
1086
1087         list_for_each_entry(term, &evsel->config_terms, list) {
1088                 if (term->type == type)
1089                         found_term = term;
1090         }
1091
1092         return found_term;
1093 }
1094
1095 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1096 {
1097         evsel__set_sample_bit(evsel, WEIGHT);
1098 }
1099
1100 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1101 {
1102 }
1103
1104 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1105                                     struct perf_event_attr *attr __maybe_unused)
1106 {
1107 }
1108
1109 static void evsel__set_default_freq_period(struct record_opts *opts,
1110                                            struct perf_event_attr *attr)
1111 {
1112         if (opts->freq) {
1113                 attr->freq = 1;
1114                 attr->sample_freq = opts->freq;
1115         } else {
1116                 attr->sample_period = opts->default_interval;
1117         }
1118 }
1119
1120 static bool evsel__is_offcpu_event(struct evsel *evsel)
1121 {
1122         return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1123 }
1124
1125 /*
1126  * The enable_on_exec/disabled value strategy:
1127  *
1128  *  1) For any type of traced program:
1129  *    - all independent events and group leaders are disabled
1130  *    - all group members are enabled
1131  *
1132  *     Group members are ruled by group leaders. They need to
1133  *     be enabled, because the group scheduling relies on that.
1134  *
1135  *  2) For traced programs executed by perf:
1136  *     - all independent events and group leaders have
1137  *       enable_on_exec set
1138  *     - we don't specifically enable or disable any event during
1139  *       the record command
1140  *
1141  *     Independent events and group leaders are initially disabled
1142  *     and get enabled by exec. Group members are ruled by group
1143  *     leaders as stated in 1).
1144  *
1145  *  3) For traced programs attached by perf (pid/tid):
1146  *     - we specifically enable or disable all events during
1147  *       the record command
1148  *
1149  *     When attaching events to already running traced we
1150  *     enable/disable events specifically, as there's no
1151  *     initial traced exec call.
1152  */
1153 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1154                    struct callchain_param *callchain)
1155 {
1156         struct evsel *leader = evsel__leader(evsel);
1157         struct perf_event_attr *attr = &evsel->core.attr;
1158         int track = evsel->tracking;
1159         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1160
1161         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1162         attr->inherit       = !opts->no_inherit;
1163         attr->write_backward = opts->overwrite ? 1 : 0;
1164         attr->read_format   = PERF_FORMAT_LOST;
1165
1166         evsel__set_sample_bit(evsel, IP);
1167         evsel__set_sample_bit(evsel, TID);
1168
1169         if (evsel->sample_read) {
1170                 evsel__set_sample_bit(evsel, READ);
1171
1172                 /*
1173                  * We need ID even in case of single event, because
1174                  * PERF_SAMPLE_READ process ID specific data.
1175                  */
1176                 evsel__set_sample_id(evsel, false);
1177
1178                 /*
1179                  * Apply group format only if we belong to group
1180                  * with more than one members.
1181                  */
1182                 if (leader->core.nr_members > 1) {
1183                         attr->read_format |= PERF_FORMAT_GROUP;
1184                         attr->inherit = 0;
1185                 }
1186         }
1187
1188         /*
1189          * We default some events to have a default interval. But keep
1190          * it a weak assumption overridable by the user.
1191          */
1192         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1193             (!evsel->is_libpfm_event && (!attr->sample_period ||
1194                                          opts->user_freq != UINT_MAX ||
1195                                          opts->user_interval != ULLONG_MAX)))
1196                 evsel__set_default_freq_period(opts, attr);
1197
1198         /*
1199          * If attr->freq was set (here or earlier), ask for period
1200          * to be sampled.
1201          */
1202         if (attr->freq)
1203                 evsel__set_sample_bit(evsel, PERIOD);
1204
1205         if (opts->no_samples)
1206                 attr->sample_freq = 0;
1207
1208         if (opts->inherit_stat) {
1209                 evsel->core.attr.read_format |=
1210                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1211                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1212                         PERF_FORMAT_ID;
1213                 attr->inherit_stat = 1;
1214         }
1215
1216         if (opts->sample_address) {
1217                 evsel__set_sample_bit(evsel, ADDR);
1218                 attr->mmap_data = track;
1219         }
1220
1221         /*
1222          * We don't allow user space callchains for  function trace
1223          * event, due to issues with page faults while tracing page
1224          * fault handler and its overall trickiness nature.
1225          */
1226         if (evsel__is_function_event(evsel))
1227                 evsel->core.attr.exclude_callchain_user = 1;
1228
1229         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1230                 evsel__config_callchain(evsel, opts, callchain);
1231
1232         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1233             !evsel__is_dummy_event(evsel)) {
1234                 attr->sample_regs_intr = opts->sample_intr_regs;
1235                 evsel__set_sample_bit(evsel, REGS_INTR);
1236         }
1237
1238         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1239             !evsel__is_dummy_event(evsel)) {
1240                 attr->sample_regs_user |= opts->sample_user_regs;
1241                 evsel__set_sample_bit(evsel, REGS_USER);
1242         }
1243
1244         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1245                 evsel__set_sample_bit(evsel, CPU);
1246
1247         /*
1248          * When the user explicitly disabled time don't force it here.
1249          */
1250         if (opts->sample_time &&
1251             (!perf_missing_features.sample_id_all &&
1252             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1253              opts->sample_time_set)))
1254                 evsel__set_sample_bit(evsel, TIME);
1255
1256         if (opts->raw_samples && !evsel->no_aux_samples) {
1257                 evsel__set_sample_bit(evsel, TIME);
1258                 evsel__set_sample_bit(evsel, RAW);
1259                 evsel__set_sample_bit(evsel, CPU);
1260         }
1261
1262         if (opts->sample_address)
1263                 evsel__set_sample_bit(evsel, DATA_SRC);
1264
1265         if (opts->sample_phys_addr)
1266                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1267
1268         if (opts->no_buffering) {
1269                 attr->watermark = 0;
1270                 attr->wakeup_events = 1;
1271         }
1272         if (opts->branch_stack && !evsel->no_aux_samples) {
1273                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1274                 attr->branch_sample_type = opts->branch_stack;
1275         }
1276
1277         if (opts->sample_weight)
1278                 arch_evsel__set_sample_weight(evsel);
1279
1280         attr->task     = track;
1281         attr->mmap     = track;
1282         attr->mmap2    = track && !perf_missing_features.mmap2;
1283         attr->comm     = track;
1284         attr->build_id = track && opts->build_id;
1285
1286         /*
1287          * ksymbol is tracked separately with text poke because it needs to be
1288          * system wide and enabled immediately.
1289          */
1290         if (!opts->text_poke)
1291                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1292         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1293
1294         if (opts->record_namespaces)
1295                 attr->namespaces  = track;
1296
1297         if (opts->record_cgroup) {
1298                 attr->cgroup = track && !perf_missing_features.cgroup;
1299                 evsel__set_sample_bit(evsel, CGROUP);
1300         }
1301
1302         if (opts->sample_data_page_size)
1303                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1304
1305         if (opts->sample_code_page_size)
1306                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1307
1308         if (opts->record_switch_events)
1309                 attr->context_switch = track;
1310
1311         if (opts->sample_transaction)
1312                 evsel__set_sample_bit(evsel, TRANSACTION);
1313
1314         if (opts->running_time) {
1315                 evsel->core.attr.read_format |=
1316                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1317                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1318         }
1319
1320         /*
1321          * XXX see the function comment above
1322          *
1323          * Disabling only independent events or group leaders,
1324          * keeping group members enabled.
1325          */
1326         if (evsel__is_group_leader(evsel))
1327                 attr->disabled = 1;
1328
1329         /*
1330          * Setting enable_on_exec for independent events and
1331          * group leaders for traced executed by perf.
1332          */
1333         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1334             !opts->initial_delay)
1335                 attr->enable_on_exec = 1;
1336
1337         if (evsel->immediate) {
1338                 attr->disabled = 0;
1339                 attr->enable_on_exec = 0;
1340         }
1341
1342         clockid = opts->clockid;
1343         if (opts->use_clockid) {
1344                 attr->use_clockid = 1;
1345                 attr->clockid = opts->clockid;
1346         }
1347
1348         if (evsel->precise_max)
1349                 attr->precise_ip = 3;
1350
1351         if (opts->all_user) {
1352                 attr->exclude_kernel = 1;
1353                 attr->exclude_user   = 0;
1354         }
1355
1356         if (opts->all_kernel) {
1357                 attr->exclude_kernel = 0;
1358                 attr->exclude_user   = 1;
1359         }
1360
1361         if (evsel->core.own_cpus || evsel->unit)
1362                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1363
1364         /*
1365          * Apply event specific term settings,
1366          * it overloads any global configuration.
1367          */
1368         evsel__apply_config_terms(evsel, opts, track);
1369
1370         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1371
1372         /* The --period option takes the precedence. */
1373         if (opts->period_set) {
1374                 if (opts->period)
1375                         evsel__set_sample_bit(evsel, PERIOD);
1376                 else
1377                         evsel__reset_sample_bit(evsel, PERIOD);
1378         }
1379
1380         /*
1381          * A dummy event never triggers any actual counter and therefore
1382          * cannot be used with branch_stack.
1383          *
1384          * For initial_delay, a dummy event is added implicitly.
1385          * The software event will trigger -EOPNOTSUPP error out,
1386          * if BRANCH_STACK bit is set.
1387          */
1388         if (evsel__is_dummy_event(evsel))
1389                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1390
1391         if (evsel__is_offcpu_event(evsel))
1392                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1393
1394         arch__post_evsel_config(evsel, attr);
1395 }
1396
1397 int evsel__set_filter(struct evsel *evsel, const char *filter)
1398 {
1399         char *new_filter = strdup(filter);
1400
1401         if (new_filter != NULL) {
1402                 free(evsel->filter);
1403                 evsel->filter = new_filter;
1404                 return 0;
1405         }
1406
1407         return -1;
1408 }
1409
1410 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1411 {
1412         char *new_filter;
1413
1414         if (evsel->filter == NULL)
1415                 return evsel__set_filter(evsel, filter);
1416
1417         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1418                 free(evsel->filter);
1419                 evsel->filter = new_filter;
1420                 return 0;
1421         }
1422
1423         return -1;
1424 }
1425
1426 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1427 {
1428         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1429 }
1430
1431 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1432 {
1433         return evsel__append_filter(evsel, "%s,%s", filter);
1434 }
1435
1436 /* Caller has to clear disabled after going through all CPUs. */
1437 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1438 {
1439         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1440 }
1441
1442 int evsel__enable(struct evsel *evsel)
1443 {
1444         int err = perf_evsel__enable(&evsel->core);
1445
1446         if (!err)
1447                 evsel->disabled = false;
1448         return err;
1449 }
1450
1451 /* Caller has to set disabled after going through all CPUs. */
1452 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1453 {
1454         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1455 }
1456
1457 int evsel__disable(struct evsel *evsel)
1458 {
1459         int err = perf_evsel__disable(&evsel->core);
1460         /*
1461          * We mark it disabled here so that tools that disable a event can
1462          * ignore events after they disable it. I.e. the ring buffer may have
1463          * already a few more events queued up before the kernel got the stop
1464          * request.
1465          */
1466         if (!err)
1467                 evsel->disabled = true;
1468
1469         return err;
1470 }
1471
1472 void free_config_terms(struct list_head *config_terms)
1473 {
1474         struct evsel_config_term *term, *h;
1475
1476         list_for_each_entry_safe(term, h, config_terms, list) {
1477                 list_del_init(&term->list);
1478                 if (term->free_str)
1479                         zfree(&term->val.str);
1480                 free(term);
1481         }
1482 }
1483
1484 static void evsel__free_config_terms(struct evsel *evsel)
1485 {
1486         free_config_terms(&evsel->config_terms);
1487 }
1488
1489 void evsel__exit(struct evsel *evsel)
1490 {
1491         assert(list_empty(&evsel->core.node));
1492         assert(evsel->evlist == NULL);
1493         bpf_counter__destroy(evsel);
1494         evsel__free_counts(evsel);
1495         perf_evsel__free_fd(&evsel->core);
1496         perf_evsel__free_id(&evsel->core);
1497         evsel__free_config_terms(evsel);
1498         cgroup__put(evsel->cgrp);
1499         perf_cpu_map__put(evsel->core.cpus);
1500         perf_cpu_map__put(evsel->core.own_cpus);
1501         perf_thread_map__put(evsel->core.threads);
1502         zfree(&evsel->group_name);
1503         zfree(&evsel->name);
1504         zfree(&evsel->pmu_name);
1505         zfree(&evsel->unit);
1506         zfree(&evsel->metric_id);
1507         evsel__zero_per_pkg(evsel);
1508         hashmap__free(evsel->per_pkg_mask);
1509         evsel->per_pkg_mask = NULL;
1510         zfree(&evsel->metric_events);
1511         perf_evsel__object.fini(evsel);
1512 }
1513
1514 void evsel__delete(struct evsel *evsel)
1515 {
1516         evsel__exit(evsel);
1517         free(evsel);
1518 }
1519
1520 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1521                            struct perf_counts_values *count)
1522 {
1523         struct perf_counts_values tmp;
1524
1525         if (!evsel->prev_raw_counts)
1526                 return;
1527
1528         if (cpu_map_idx == -1) {
1529                 tmp = evsel->prev_raw_counts->aggr;
1530                 evsel->prev_raw_counts->aggr = *count;
1531         } else {
1532                 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1533                 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1534         }
1535
1536         count->val = count->val - tmp.val;
1537         count->ena = count->ena - tmp.ena;
1538         count->run = count->run - tmp.run;
1539 }
1540
1541 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1542 {
1543         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1544
1545         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1546 }
1547
1548 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1549                              u64 val, u64 ena, u64 run, u64 lost)
1550 {
1551         struct perf_counts_values *count;
1552
1553         count = perf_counts(counter->counts, cpu_map_idx, thread);
1554
1555         count->val    = val;
1556         count->ena    = ena;
1557         count->run    = run;
1558         count->lost   = lost;
1559
1560         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1561 }
1562
1563 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1564 {
1565         u64 read_format = leader->core.attr.read_format;
1566         struct sample_read_value *v;
1567         u64 nr, ena = 0, run = 0, lost = 0;
1568
1569         nr = *data++;
1570
1571         if (nr != (u64) leader->core.nr_members)
1572                 return -EINVAL;
1573
1574         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1575                 ena = *data++;
1576
1577         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1578                 run = *data++;
1579
1580         v = (void *)data;
1581         sample_read_group__for_each(v, nr, read_format) {
1582                 struct evsel *counter;
1583
1584                 counter = evlist__id2evsel(leader->evlist, v->id);
1585                 if (!counter)
1586                         return -EINVAL;
1587
1588                 if (read_format & PERF_FORMAT_LOST)
1589                         lost = v->lost;
1590
1591                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1592         }
1593
1594         return 0;
1595 }
1596
1597 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1598 {
1599         struct perf_stat_evsel *ps = leader->stats;
1600         u64 read_format = leader->core.attr.read_format;
1601         int size = perf_evsel__read_size(&leader->core);
1602         u64 *data = ps->group_data;
1603
1604         if (!(read_format & PERF_FORMAT_ID))
1605                 return -EINVAL;
1606
1607         if (!evsel__is_group_leader(leader))
1608                 return -EINVAL;
1609
1610         if (!data) {
1611                 data = zalloc(size);
1612                 if (!data)
1613                         return -ENOMEM;
1614
1615                 ps->group_data = data;
1616         }
1617
1618         if (FD(leader, cpu_map_idx, thread) < 0)
1619                 return -EINVAL;
1620
1621         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1622                 return -errno;
1623
1624         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1625 }
1626
1627 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1628 {
1629         u64 read_format = evsel->core.attr.read_format;
1630
1631         if (read_format & PERF_FORMAT_GROUP)
1632                 return evsel__read_group(evsel, cpu_map_idx, thread);
1633
1634         return evsel__read_one(evsel, cpu_map_idx, thread);
1635 }
1636
1637 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1638 {
1639         struct perf_counts_values count;
1640         size_t nv = scale ? 3 : 1;
1641
1642         if (FD(evsel, cpu_map_idx, thread) < 0)
1643                 return -EINVAL;
1644
1645         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1646                 return -ENOMEM;
1647
1648         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1649                 return -errno;
1650
1651         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1652         perf_counts_values__scale(&count, scale, NULL);
1653         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1654         return 0;
1655 }
1656
1657 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1658                                   int cpu_map_idx)
1659 {
1660         struct perf_cpu cpu;
1661
1662         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1663         return perf_cpu_map__idx(other->core.cpus, cpu);
1664 }
1665
1666 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1667 {
1668         struct evsel *leader = evsel__leader(evsel);
1669
1670         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1671             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1672                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1673         }
1674
1675         return cpu_map_idx;
1676 }
1677
1678 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1679 {
1680         struct evsel *leader = evsel__leader(evsel);
1681         int fd;
1682
1683         if (evsel__is_group_leader(evsel))
1684                 return -1;
1685
1686         /*
1687          * Leader must be already processed/open,
1688          * if not it's a bug.
1689          */
1690         BUG_ON(!leader->core.fd);
1691
1692         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1693         if (cpu_map_idx == -1)
1694                 return -1;
1695
1696         fd = FD(leader, cpu_map_idx, thread);
1697         BUG_ON(fd == -1);
1698
1699         return fd;
1700 }
1701
1702 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1703 {
1704         for (int cpu = 0; cpu < nr_cpus; cpu++)
1705                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1706                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1707 }
1708
1709 static int update_fds(struct evsel *evsel,
1710                       int nr_cpus, int cpu_map_idx,
1711                       int nr_threads, int thread_idx)
1712 {
1713         struct evsel *pos;
1714
1715         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1716                 return -EINVAL;
1717
1718         evlist__for_each_entry(evsel->evlist, pos) {
1719                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1720
1721                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1722
1723                 /*
1724                  * Since fds for next evsel has not been created,
1725                  * there is no need to iterate whole event list.
1726                  */
1727                 if (pos == evsel)
1728                         break;
1729         }
1730         return 0;
1731 }
1732
1733 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1734                                          int nr_cpus, int cpu_map_idx,
1735                                          struct perf_thread_map *threads,
1736                                          int thread, int err)
1737 {
1738         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1739
1740         if (!evsel->ignore_missing_thread)
1741                 return false;
1742
1743         /* The system wide setup does not work with threads. */
1744         if (evsel->core.system_wide)
1745                 return false;
1746
1747         /* The -ESRCH is perf event syscall errno for pid's not found. */
1748         if (err != -ESRCH)
1749                 return false;
1750
1751         /* If there's only one thread, let it fail. */
1752         if (threads->nr == 1)
1753                 return false;
1754
1755         /*
1756          * We should remove fd for missing_thread first
1757          * because thread_map__remove() will decrease threads->nr.
1758          */
1759         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1760                 return false;
1761
1762         if (thread_map__remove(threads, thread))
1763                 return false;
1764
1765         pr_warning("WARNING: Ignored open failure for pid %d\n",
1766                    ignore_pid);
1767         return true;
1768 }
1769
1770 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1771                                 void *priv __maybe_unused)
1772 {
1773         return fprintf(fp, "  %-32s %s\n", name, val);
1774 }
1775
1776 static void display_attr(struct perf_event_attr *attr)
1777 {
1778         if (verbose >= 2 || debug_peo_args) {
1779                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1780                 fprintf(stderr, "perf_event_attr:\n");
1781                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1782                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1783         }
1784 }
1785
1786 bool evsel__precise_ip_fallback(struct evsel *evsel)
1787 {
1788         /* Do not try less precise if not requested. */
1789         if (!evsel->precise_max)
1790                 return false;
1791
1792         /*
1793          * We tried all the precise_ip values, and it's
1794          * still failing, so leave it to standard fallback.
1795          */
1796         if (!evsel->core.attr.precise_ip) {
1797                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1798                 return false;
1799         }
1800
1801         if (!evsel->precise_ip_original)
1802                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1803
1804         evsel->core.attr.precise_ip--;
1805         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1806         display_attr(&evsel->core.attr);
1807         return true;
1808 }
1809
1810 static struct perf_cpu_map *empty_cpu_map;
1811 static struct perf_thread_map *empty_thread_map;
1812
1813 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1814                 struct perf_thread_map *threads)
1815 {
1816         int nthreads = perf_thread_map__nr(threads);
1817
1818         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1819             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1820                 return -EINVAL;
1821
1822         if (cpus == NULL) {
1823                 if (empty_cpu_map == NULL) {
1824                         empty_cpu_map = perf_cpu_map__dummy_new();
1825                         if (empty_cpu_map == NULL)
1826                                 return -ENOMEM;
1827                 }
1828
1829                 cpus = empty_cpu_map;
1830         }
1831
1832         if (threads == NULL) {
1833                 if (empty_thread_map == NULL) {
1834                         empty_thread_map = thread_map__new_by_tid(-1);
1835                         if (empty_thread_map == NULL)
1836                                 return -ENOMEM;
1837                 }
1838
1839                 threads = empty_thread_map;
1840         }
1841
1842         if (evsel->core.fd == NULL &&
1843             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1844                 return -ENOMEM;
1845
1846         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1847         if (evsel->cgrp)
1848                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1849
1850         return 0;
1851 }
1852
1853 static void evsel__disable_missing_features(struct evsel *evsel)
1854 {
1855         if (perf_missing_features.read_lost)
1856                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1857         if (perf_missing_features.weight_struct) {
1858                 evsel__set_sample_bit(evsel, WEIGHT);
1859                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1860         }
1861         if (perf_missing_features.clockid_wrong)
1862                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1863         if (perf_missing_features.clockid) {
1864                 evsel->core.attr.use_clockid = 0;
1865                 evsel->core.attr.clockid = 0;
1866         }
1867         if (perf_missing_features.cloexec)
1868                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1869         if (perf_missing_features.mmap2)
1870                 evsel->core.attr.mmap2 = 0;
1871         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1872                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1873         if (perf_missing_features.lbr_flags)
1874                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1875                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1876         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1877                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1878         if (perf_missing_features.ksymbol)
1879                 evsel->core.attr.ksymbol = 0;
1880         if (perf_missing_features.bpf)
1881                 evsel->core.attr.bpf_event = 0;
1882         if (perf_missing_features.branch_hw_idx)
1883                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1884         if (perf_missing_features.sample_id_all)
1885                 evsel->core.attr.sample_id_all = 0;
1886 }
1887
1888 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1889                         struct perf_thread_map *threads)
1890 {
1891         int err;
1892
1893         err = __evsel__prepare_open(evsel, cpus, threads);
1894         if (err)
1895                 return err;
1896
1897         evsel__disable_missing_features(evsel);
1898
1899         return err;
1900 }
1901
1902 bool evsel__detect_missing_features(struct evsel *evsel)
1903 {
1904         /*
1905          * Must probe features in the order they were added to the
1906          * perf_event_attr interface.
1907          */
1908         if (!perf_missing_features.read_lost &&
1909             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1910                 perf_missing_features.read_lost = true;
1911                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1912                 return true;
1913         } else if (!perf_missing_features.weight_struct &&
1914             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1915                 perf_missing_features.weight_struct = true;
1916                 pr_debug2("switching off weight struct support\n");
1917                 return true;
1918         } else if (!perf_missing_features.code_page_size &&
1919             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1920                 perf_missing_features.code_page_size = true;
1921                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1922                 return false;
1923         } else if (!perf_missing_features.data_page_size &&
1924             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1925                 perf_missing_features.data_page_size = true;
1926                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1927                 return false;
1928         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1929                 perf_missing_features.cgroup = true;
1930                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1931                 return false;
1932         } else if (!perf_missing_features.branch_hw_idx &&
1933             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1934                 perf_missing_features.branch_hw_idx = true;
1935                 pr_debug2("switching off branch HW index support\n");
1936                 return true;
1937         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1938                 perf_missing_features.aux_output = true;
1939                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1940                 return false;
1941         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1942                 perf_missing_features.bpf = true;
1943                 pr_debug2_peo("switching off bpf_event\n");
1944                 return true;
1945         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1946                 perf_missing_features.ksymbol = true;
1947                 pr_debug2_peo("switching off ksymbol\n");
1948                 return true;
1949         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1950                 perf_missing_features.write_backward = true;
1951                 pr_debug2_peo("switching off write_backward\n");
1952                 return false;
1953         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1954                 perf_missing_features.clockid_wrong = true;
1955                 pr_debug2_peo("switching off clockid\n");
1956                 return true;
1957         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1958                 perf_missing_features.clockid = true;
1959                 pr_debug2_peo("switching off use_clockid\n");
1960                 return true;
1961         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1962                 perf_missing_features.cloexec = true;
1963                 pr_debug2_peo("switching off cloexec flag\n");
1964                 return true;
1965         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1966                 perf_missing_features.mmap2 = true;
1967                 pr_debug2_peo("switching off mmap2\n");
1968                 return true;
1969         } else if ((evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) &&
1970                    (evsel->pmu == NULL || evsel->pmu->missing_features.exclude_guest)) {
1971                 if (evsel->pmu == NULL) {
1972                         evsel->pmu = evsel__find_pmu(evsel);
1973                         if (evsel->pmu)
1974                                 evsel->pmu->missing_features.exclude_guest = true;
1975                         else {
1976                                 /* we cannot find PMU, disable attrs now */
1977                                 evsel->core.attr.exclude_host = false;
1978                                 evsel->core.attr.exclude_guest = false;
1979                         }
1980                 }
1981
1982                 if (evsel->exclude_GH) {
1983                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1984                         return false;
1985                 }
1986                 if (!perf_missing_features.exclude_guest) {
1987                         perf_missing_features.exclude_guest = true;
1988                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1989                 }
1990                 return true;
1991         } else if (!perf_missing_features.sample_id_all) {
1992                 perf_missing_features.sample_id_all = true;
1993                 pr_debug2_peo("switching off sample_id_all\n");
1994                 return true;
1995         } else if (!perf_missing_features.lbr_flags &&
1996                         (evsel->core.attr.branch_sample_type &
1997                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1998                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1999                 perf_missing_features.lbr_flags = true;
2000                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2001                 return true;
2002         } else if (!perf_missing_features.group_read &&
2003                     evsel->core.attr.inherit &&
2004                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2005                    evsel__is_group_leader(evsel)) {
2006                 perf_missing_features.group_read = true;
2007                 pr_debug2_peo("switching off group read\n");
2008                 return true;
2009         } else {
2010                 return false;
2011         }
2012 }
2013
2014 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2015 {
2016         int old_errno;
2017         struct rlimit l;
2018
2019         if (*set_rlimit < INCREASED_MAX) {
2020                 old_errno = errno;
2021
2022                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2023                         if (*set_rlimit == NO_CHANGE) {
2024                                 l.rlim_cur = l.rlim_max;
2025                         } else {
2026                                 l.rlim_cur = l.rlim_max + 1000;
2027                                 l.rlim_max = l.rlim_cur;
2028                         }
2029                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2030                                 (*set_rlimit) += 1;
2031                                 errno = old_errno;
2032                                 return true;
2033                         }
2034                 }
2035                 errno = old_errno;
2036         }
2037
2038         return false;
2039 }
2040
2041 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2042                 struct perf_thread_map *threads,
2043                 int start_cpu_map_idx, int end_cpu_map_idx)
2044 {
2045         int idx, thread, nthreads;
2046         int pid = -1, err, old_errno;
2047         enum rlimit_action set_rlimit = NO_CHANGE;
2048
2049         err = __evsel__prepare_open(evsel, cpus, threads);
2050         if (err)
2051                 return err;
2052
2053         if (cpus == NULL)
2054                 cpus = empty_cpu_map;
2055
2056         if (threads == NULL)
2057                 threads = empty_thread_map;
2058
2059         nthreads = perf_thread_map__nr(threads);
2060
2061         if (evsel->cgrp)
2062                 pid = evsel->cgrp->fd;
2063
2064 fallback_missing_features:
2065         evsel__disable_missing_features(evsel);
2066
2067         display_attr(&evsel->core.attr);
2068
2069         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2070
2071                 for (thread = 0; thread < nthreads; thread++) {
2072                         int fd, group_fd;
2073 retry_open:
2074                         if (thread >= nthreads)
2075                                 break;
2076
2077                         if (!evsel->cgrp && !evsel->core.system_wide)
2078                                 pid = perf_thread_map__pid(threads, thread);
2079
2080                         group_fd = get_group_fd(evsel, idx, thread);
2081
2082                         test_attr__ready();
2083
2084                         /* Debug message used by test scripts */
2085                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2086                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2087
2088                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2089                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2090                                                 group_fd, evsel->open_flags);
2091
2092                         FD(evsel, idx, thread) = fd;
2093
2094                         if (fd < 0) {
2095                                 err = -errno;
2096
2097                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2098                                           err);
2099                                 goto try_fallback;
2100                         }
2101
2102                         bpf_counter__install_pe(evsel, idx, fd);
2103
2104                         if (unlikely(test_attr__enabled)) {
2105                                 test_attr__open(&evsel->core.attr, pid,
2106                                                 perf_cpu_map__cpu(cpus, idx),
2107                                                 fd, group_fd, evsel->open_flags);
2108                         }
2109
2110                         /* Debug message used by test scripts */
2111                         pr_debug2_peo(" = %d\n", fd);
2112
2113                         if (evsel->bpf_fd >= 0) {
2114                                 int evt_fd = fd;
2115                                 int bpf_fd = evsel->bpf_fd;
2116
2117                                 err = ioctl(evt_fd,
2118                                             PERF_EVENT_IOC_SET_BPF,
2119                                             bpf_fd);
2120                                 if (err && errno != EEXIST) {
2121                                         pr_err("failed to attach bpf fd %d: %s\n",
2122                                                bpf_fd, strerror(errno));
2123                                         err = -EINVAL;
2124                                         goto out_close;
2125                                 }
2126                         }
2127
2128                         set_rlimit = NO_CHANGE;
2129
2130                         /*
2131                          * If we succeeded but had to kill clockid, fail and
2132                          * have evsel__open_strerror() print us a nice error.
2133                          */
2134                         if (perf_missing_features.clockid ||
2135                             perf_missing_features.clockid_wrong) {
2136                                 err = -EINVAL;
2137                                 goto out_close;
2138                         }
2139                 }
2140         }
2141
2142         return 0;
2143
2144 try_fallback:
2145         if (evsel__precise_ip_fallback(evsel))
2146                 goto retry_open;
2147
2148         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2149                                          idx, threads, thread, err)) {
2150                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2151                 nthreads--;
2152
2153                 /* ... and pretend like nothing have happened. */
2154                 err = 0;
2155                 goto retry_open;
2156         }
2157         /*
2158          * perf stat needs between 5 and 22 fds per CPU. When we run out
2159          * of them try to increase the limits.
2160          */
2161         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2162                 goto retry_open;
2163
2164         if (err != -EINVAL || idx > 0 || thread > 0)
2165                 goto out_close;
2166
2167         if (evsel__detect_missing_features(evsel))
2168                 goto fallback_missing_features;
2169 out_close:
2170         if (err)
2171                 threads->err_thread = thread;
2172
2173         old_errno = errno;
2174         do {
2175                 while (--thread >= 0) {
2176                         if (FD(evsel, idx, thread) >= 0)
2177                                 close(FD(evsel, idx, thread));
2178                         FD(evsel, idx, thread) = -1;
2179                 }
2180                 thread = nthreads;
2181         } while (--idx >= 0);
2182         errno = old_errno;
2183         return err;
2184 }
2185
2186 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2187                 struct perf_thread_map *threads)
2188 {
2189         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2190 }
2191
2192 void evsel__close(struct evsel *evsel)
2193 {
2194         perf_evsel__close(&evsel->core);
2195         perf_evsel__free_id(&evsel->core);
2196 }
2197
2198 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2199 {
2200         if (cpu_map_idx == -1)
2201                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2202
2203         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2204 }
2205
2206 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2207 {
2208         return evsel__open(evsel, NULL, threads);
2209 }
2210
2211 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2212                                        const union perf_event *event,
2213                                        struct perf_sample *sample)
2214 {
2215         u64 type = evsel->core.attr.sample_type;
2216         const __u64 *array = event->sample.array;
2217         bool swapped = evsel->needs_swap;
2218         union u64_swap u;
2219
2220         array += ((event->header.size -
2221                    sizeof(event->header)) / sizeof(u64)) - 1;
2222
2223         if (type & PERF_SAMPLE_IDENTIFIER) {
2224                 sample->id = *array;
2225                 array--;
2226         }
2227
2228         if (type & PERF_SAMPLE_CPU) {
2229                 u.val64 = *array;
2230                 if (swapped) {
2231                         /* undo swap of u64, then swap on individual u32s */
2232                         u.val64 = bswap_64(u.val64);
2233                         u.val32[0] = bswap_32(u.val32[0]);
2234                 }
2235
2236                 sample->cpu = u.val32[0];
2237                 array--;
2238         }
2239
2240         if (type & PERF_SAMPLE_STREAM_ID) {
2241                 sample->stream_id = *array;
2242                 array--;
2243         }
2244
2245         if (type & PERF_SAMPLE_ID) {
2246                 sample->id = *array;
2247                 array--;
2248         }
2249
2250         if (type & PERF_SAMPLE_TIME) {
2251                 sample->time = *array;
2252                 array--;
2253         }
2254
2255         if (type & PERF_SAMPLE_TID) {
2256                 u.val64 = *array;
2257                 if (swapped) {
2258                         /* undo swap of u64, then swap on individual u32s */
2259                         u.val64 = bswap_64(u.val64);
2260                         u.val32[0] = bswap_32(u.val32[0]);
2261                         u.val32[1] = bswap_32(u.val32[1]);
2262                 }
2263
2264                 sample->pid = u.val32[0];
2265                 sample->tid = u.val32[1];
2266                 array--;
2267         }
2268
2269         return 0;
2270 }
2271
2272 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2273                             u64 size)
2274 {
2275         return size > max_size || offset + size > endp;
2276 }
2277
2278 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2279         do {                                                            \
2280                 if (overflow(endp, (max_size), (offset), (size)))       \
2281                         return -EFAULT;                                 \
2282         } while (0)
2283
2284 #define OVERFLOW_CHECK_u64(offset) \
2285         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2286
2287 static int
2288 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2289 {
2290         /*
2291          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2292          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2293          * check the format does not go past the end of the event.
2294          */
2295         if (sample_size + sizeof(event->header) > event->header.size)
2296                 return -EFAULT;
2297
2298         return 0;
2299 }
2300
2301 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2302                                           const __u64 *array,
2303                                           u64 type __maybe_unused)
2304 {
2305         data->weight = *array;
2306 }
2307
2308 u64 evsel__bitfield_swap_branch_flags(u64 value)
2309 {
2310         u64 new_val = 0;
2311
2312         /*
2313          * branch_flags
2314          * union {
2315          *      u64 values;
2316          *      struct {
2317          *              mispred:1       //target mispredicted
2318          *              predicted:1     //target predicted
2319          *              in_tx:1         //in transaction
2320          *              abort:1         //transaction abort
2321          *              cycles:16       //cycle count to last branch
2322          *              type:4          //branch type
2323          *              reserved:40
2324          *      }
2325          * }
2326          *
2327          * Avoid bswap64() the entire branch_flag.value,
2328          * as it has variable bit-field sizes. Instead the
2329          * macro takes the bit-field position/size,
2330          * swaps it based on the host endianness.
2331          *
2332          * tep_is_bigendian() is used here instead of
2333          * bigendian() to avoid python test fails.
2334          */
2335         if (tep_is_bigendian()) {
2336                 new_val = bitfield_swap(value, 0, 1);
2337                 new_val |= bitfield_swap(value, 1, 1);
2338                 new_val |= bitfield_swap(value, 2, 1);
2339                 new_val |= bitfield_swap(value, 3, 1);
2340                 new_val |= bitfield_swap(value, 4, 16);
2341                 new_val |= bitfield_swap(value, 20, 4);
2342                 new_val |= bitfield_swap(value, 24, 40);
2343         } else {
2344                 new_val = bitfield_swap(value, 63, 1);
2345                 new_val |= bitfield_swap(value, 62, 1);
2346                 new_val |= bitfield_swap(value, 61, 1);
2347                 new_val |= bitfield_swap(value, 60, 1);
2348                 new_val |= bitfield_swap(value, 44, 16);
2349                 new_val |= bitfield_swap(value, 40, 4);
2350                 new_val |= bitfield_swap(value, 0, 40);
2351         }
2352
2353         return new_val;
2354 }
2355
2356 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2357                         struct perf_sample *data)
2358 {
2359         u64 type = evsel->core.attr.sample_type;
2360         bool swapped = evsel->needs_swap;
2361         const __u64 *array;
2362         u16 max_size = event->header.size;
2363         const void *endp = (void *)event + max_size;
2364         u64 sz;
2365
2366         /*
2367          * used for cross-endian analysis. See git commit 65014ab3
2368          * for why this goofiness is needed.
2369          */
2370         union u64_swap u;
2371
2372         memset(data, 0, sizeof(*data));
2373         data->cpu = data->pid = data->tid = -1;
2374         data->stream_id = data->id = data->time = -1ULL;
2375         data->period = evsel->core.attr.sample_period;
2376         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2377         data->misc    = event->header.misc;
2378         data->id = -1ULL;
2379         data->data_src = PERF_MEM_DATA_SRC_NONE;
2380         data->vcpu = -1;
2381
2382         if (event->header.type != PERF_RECORD_SAMPLE) {
2383                 if (!evsel->core.attr.sample_id_all)
2384                         return 0;
2385                 return perf_evsel__parse_id_sample(evsel, event, data);
2386         }
2387
2388         array = event->sample.array;
2389
2390         if (perf_event__check_size(event, evsel->sample_size))
2391                 return -EFAULT;
2392
2393         if (type & PERF_SAMPLE_IDENTIFIER) {
2394                 data->id = *array;
2395                 array++;
2396         }
2397
2398         if (type & PERF_SAMPLE_IP) {
2399                 data->ip = *array;
2400                 array++;
2401         }
2402
2403         if (type & PERF_SAMPLE_TID) {
2404                 u.val64 = *array;
2405                 if (swapped) {
2406                         /* undo swap of u64, then swap on individual u32s */
2407                         u.val64 = bswap_64(u.val64);
2408                         u.val32[0] = bswap_32(u.val32[0]);
2409                         u.val32[1] = bswap_32(u.val32[1]);
2410                 }
2411
2412                 data->pid = u.val32[0];
2413                 data->tid = u.val32[1];
2414                 array++;
2415         }
2416
2417         if (type & PERF_SAMPLE_TIME) {
2418                 data->time = *array;
2419                 array++;
2420         }
2421
2422         if (type & PERF_SAMPLE_ADDR) {
2423                 data->addr = *array;
2424                 array++;
2425         }
2426
2427         if (type & PERF_SAMPLE_ID) {
2428                 data->id = *array;
2429                 array++;
2430         }
2431
2432         if (type & PERF_SAMPLE_STREAM_ID) {
2433                 data->stream_id = *array;
2434                 array++;
2435         }
2436
2437         if (type & PERF_SAMPLE_CPU) {
2438
2439                 u.val64 = *array;
2440                 if (swapped) {
2441                         /* undo swap of u64, then swap on individual u32s */
2442                         u.val64 = bswap_64(u.val64);
2443                         u.val32[0] = bswap_32(u.val32[0]);
2444                 }
2445
2446                 data->cpu = u.val32[0];
2447                 array++;
2448         }
2449
2450         if (type & PERF_SAMPLE_PERIOD) {
2451                 data->period = *array;
2452                 array++;
2453         }
2454
2455         if (type & PERF_SAMPLE_READ) {
2456                 u64 read_format = evsel->core.attr.read_format;
2457
2458                 OVERFLOW_CHECK_u64(array);
2459                 if (read_format & PERF_FORMAT_GROUP)
2460                         data->read.group.nr = *array;
2461                 else
2462                         data->read.one.value = *array;
2463
2464                 array++;
2465
2466                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2467                         OVERFLOW_CHECK_u64(array);
2468                         data->read.time_enabled = *array;
2469                         array++;
2470                 }
2471
2472                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2473                         OVERFLOW_CHECK_u64(array);
2474                         data->read.time_running = *array;
2475                         array++;
2476                 }
2477
2478                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2479                 if (read_format & PERF_FORMAT_GROUP) {
2480                         const u64 max_group_nr = UINT64_MAX /
2481                                         sizeof(struct sample_read_value);
2482
2483                         if (data->read.group.nr > max_group_nr)
2484                                 return -EFAULT;
2485
2486                         sz = data->read.group.nr * sample_read_value_size(read_format);
2487                         OVERFLOW_CHECK(array, sz, max_size);
2488                         data->read.group.values =
2489                                         (struct sample_read_value *)array;
2490                         array = (void *)array + sz;
2491                 } else {
2492                         OVERFLOW_CHECK_u64(array);
2493                         data->read.one.id = *array;
2494                         array++;
2495
2496                         if (read_format & PERF_FORMAT_LOST) {
2497                                 OVERFLOW_CHECK_u64(array);
2498                                 data->read.one.lost = *array;
2499                                 array++;
2500                         }
2501                 }
2502         }
2503
2504         if (type & PERF_SAMPLE_CALLCHAIN) {
2505                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2506
2507                 OVERFLOW_CHECK_u64(array);
2508                 data->callchain = (struct ip_callchain *)array++;
2509                 if (data->callchain->nr > max_callchain_nr)
2510                         return -EFAULT;
2511                 sz = data->callchain->nr * sizeof(u64);
2512                 OVERFLOW_CHECK(array, sz, max_size);
2513                 array = (void *)array + sz;
2514         }
2515
2516         if (type & PERF_SAMPLE_RAW) {
2517                 OVERFLOW_CHECK_u64(array);
2518                 u.val64 = *array;
2519
2520                 /*
2521                  * Undo swap of u64, then swap on individual u32s,
2522                  * get the size of the raw area and undo all of the
2523                  * swap. The pevent interface handles endianness by
2524                  * itself.
2525                  */
2526                 if (swapped) {
2527                         u.val64 = bswap_64(u.val64);
2528                         u.val32[0] = bswap_32(u.val32[0]);
2529                         u.val32[1] = bswap_32(u.val32[1]);
2530                 }
2531                 data->raw_size = u.val32[0];
2532
2533                 /*
2534                  * The raw data is aligned on 64bits including the
2535                  * u32 size, so it's safe to use mem_bswap_64.
2536                  */
2537                 if (swapped)
2538                         mem_bswap_64((void *) array, data->raw_size);
2539
2540                 array = (void *)array + sizeof(u32);
2541
2542                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2543                 data->raw_data = (void *)array;
2544                 array = (void *)array + data->raw_size;
2545         }
2546
2547         if (type & PERF_SAMPLE_BRANCH_STACK) {
2548                 const u64 max_branch_nr = UINT64_MAX /
2549                                           sizeof(struct branch_entry);
2550                 struct branch_entry *e;
2551                 unsigned int i;
2552
2553                 OVERFLOW_CHECK_u64(array);
2554                 data->branch_stack = (struct branch_stack *)array++;
2555
2556                 if (data->branch_stack->nr > max_branch_nr)
2557                         return -EFAULT;
2558
2559                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2560                 if (evsel__has_branch_hw_idx(evsel)) {
2561                         sz += sizeof(u64);
2562                         e = &data->branch_stack->entries[0];
2563                 } else {
2564                         data->no_hw_idx = true;
2565                         /*
2566                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2567                          * only nr and entries[] will be output by kernel.
2568                          */
2569                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2570                 }
2571
2572                 if (swapped) {
2573                         /*
2574                          * struct branch_flag does not have endian
2575                          * specific bit field definition. And bswap
2576                          * will not resolve the issue, since these
2577                          * are bit fields.
2578                          *
2579                          * evsel__bitfield_swap_branch_flags() uses a
2580                          * bitfield_swap macro to swap the bit position
2581                          * based on the host endians.
2582                          */
2583                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2584                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2585                 }
2586
2587                 OVERFLOW_CHECK(array, sz, max_size);
2588                 array = (void *)array + sz;
2589         }
2590
2591         if (type & PERF_SAMPLE_REGS_USER) {
2592                 OVERFLOW_CHECK_u64(array);
2593                 data->user_regs.abi = *array;
2594                 array++;
2595
2596                 if (data->user_regs.abi) {
2597                         u64 mask = evsel->core.attr.sample_regs_user;
2598
2599                         sz = hweight64(mask) * sizeof(u64);
2600                         OVERFLOW_CHECK(array, sz, max_size);
2601                         data->user_regs.mask = mask;
2602                         data->user_regs.regs = (u64 *)array;
2603                         array = (void *)array + sz;
2604                 }
2605         }
2606
2607         if (type & PERF_SAMPLE_STACK_USER) {
2608                 OVERFLOW_CHECK_u64(array);
2609                 sz = *array++;
2610
2611                 data->user_stack.offset = ((char *)(array - 1)
2612                                           - (char *) event);
2613
2614                 if (!sz) {
2615                         data->user_stack.size = 0;
2616                 } else {
2617                         OVERFLOW_CHECK(array, sz, max_size);
2618                         data->user_stack.data = (char *)array;
2619                         array = (void *)array + sz;
2620                         OVERFLOW_CHECK_u64(array);
2621                         data->user_stack.size = *array++;
2622                         if (WARN_ONCE(data->user_stack.size > sz,
2623                                       "user stack dump failure\n"))
2624                                 return -EFAULT;
2625                 }
2626         }
2627
2628         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2629                 OVERFLOW_CHECK_u64(array);
2630                 arch_perf_parse_sample_weight(data, array, type);
2631                 array++;
2632         }
2633
2634         if (type & PERF_SAMPLE_DATA_SRC) {
2635                 OVERFLOW_CHECK_u64(array);
2636                 data->data_src = *array;
2637                 array++;
2638         }
2639
2640         if (type & PERF_SAMPLE_TRANSACTION) {
2641                 OVERFLOW_CHECK_u64(array);
2642                 data->transaction = *array;
2643                 array++;
2644         }
2645
2646         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2647         if (type & PERF_SAMPLE_REGS_INTR) {
2648                 OVERFLOW_CHECK_u64(array);
2649                 data->intr_regs.abi = *array;
2650                 array++;
2651
2652                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2653                         u64 mask = evsel->core.attr.sample_regs_intr;
2654
2655                         sz = hweight64(mask) * sizeof(u64);
2656                         OVERFLOW_CHECK(array, sz, max_size);
2657                         data->intr_regs.mask = mask;
2658                         data->intr_regs.regs = (u64 *)array;
2659                         array = (void *)array + sz;
2660                 }
2661         }
2662
2663         data->phys_addr = 0;
2664         if (type & PERF_SAMPLE_PHYS_ADDR) {
2665                 data->phys_addr = *array;
2666                 array++;
2667         }
2668
2669         data->cgroup = 0;
2670         if (type & PERF_SAMPLE_CGROUP) {
2671                 data->cgroup = *array;
2672                 array++;
2673         }
2674
2675         data->data_page_size = 0;
2676         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2677                 data->data_page_size = *array;
2678                 array++;
2679         }
2680
2681         data->code_page_size = 0;
2682         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2683                 data->code_page_size = *array;
2684                 array++;
2685         }
2686
2687         if (type & PERF_SAMPLE_AUX) {
2688                 OVERFLOW_CHECK_u64(array);
2689                 sz = *array++;
2690
2691                 OVERFLOW_CHECK(array, sz, max_size);
2692                 /* Undo swap of data */
2693                 if (swapped)
2694                         mem_bswap_64((char *)array, sz);
2695                 data->aux_sample.size = sz;
2696                 data->aux_sample.data = (char *)array;
2697                 array = (void *)array + sz;
2698         }
2699
2700         return 0;
2701 }
2702
2703 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2704                                   u64 *timestamp)
2705 {
2706         u64 type = evsel->core.attr.sample_type;
2707         const __u64 *array;
2708
2709         if (!(type & PERF_SAMPLE_TIME))
2710                 return -1;
2711
2712         if (event->header.type != PERF_RECORD_SAMPLE) {
2713                 struct perf_sample data = {
2714                         .time = -1ULL,
2715                 };
2716
2717                 if (!evsel->core.attr.sample_id_all)
2718                         return -1;
2719                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2720                         return -1;
2721
2722                 *timestamp = data.time;
2723                 return 0;
2724         }
2725
2726         array = event->sample.array;
2727
2728         if (perf_event__check_size(event, evsel->sample_size))
2729                 return -EFAULT;
2730
2731         if (type & PERF_SAMPLE_IDENTIFIER)
2732                 array++;
2733
2734         if (type & PERF_SAMPLE_IP)
2735                 array++;
2736
2737         if (type & PERF_SAMPLE_TID)
2738                 array++;
2739
2740         if (type & PERF_SAMPLE_TIME)
2741                 *timestamp = *array;
2742
2743         return 0;
2744 }
2745
2746 u16 evsel__id_hdr_size(struct evsel *evsel)
2747 {
2748         u64 sample_type = evsel->core.attr.sample_type;
2749         u16 size = 0;
2750
2751         if (sample_type & PERF_SAMPLE_TID)
2752                 size += sizeof(u64);
2753
2754         if (sample_type & PERF_SAMPLE_TIME)
2755                 size += sizeof(u64);
2756
2757         if (sample_type & PERF_SAMPLE_ID)
2758                 size += sizeof(u64);
2759
2760         if (sample_type & PERF_SAMPLE_STREAM_ID)
2761                 size += sizeof(u64);
2762
2763         if (sample_type & PERF_SAMPLE_CPU)
2764                 size += sizeof(u64);
2765
2766         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2767                 size += sizeof(u64);
2768
2769         return size;
2770 }
2771
2772 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2773 {
2774         return tep_find_field(evsel->tp_format, name);
2775 }
2776
2777 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2778 {
2779         struct tep_format_field *field = evsel__field(evsel, name);
2780         int offset;
2781
2782         if (!field)
2783                 return NULL;
2784
2785         offset = field->offset;
2786
2787         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2788                 offset = *(int *)(sample->raw_data + field->offset);
2789                 offset &= 0xffff;
2790                 if (field->flags & TEP_FIELD_IS_RELATIVE)
2791                         offset += field->offset + field->size;
2792         }
2793
2794         return sample->raw_data + offset;
2795 }
2796
2797 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2798                          bool needs_swap)
2799 {
2800         u64 value;
2801         void *ptr = sample->raw_data + field->offset;
2802
2803         switch (field->size) {
2804         case 1:
2805                 return *(u8 *)ptr;
2806         case 2:
2807                 value = *(u16 *)ptr;
2808                 break;
2809         case 4:
2810                 value = *(u32 *)ptr;
2811                 break;
2812         case 8:
2813                 memcpy(&value, ptr, sizeof(u64));
2814                 break;
2815         default:
2816                 return 0;
2817         }
2818
2819         if (!needs_swap)
2820                 return value;
2821
2822         switch (field->size) {
2823         case 2:
2824                 return bswap_16(value);
2825         case 4:
2826                 return bswap_32(value);
2827         case 8:
2828                 return bswap_64(value);
2829         default:
2830                 return 0;
2831         }
2832
2833         return 0;
2834 }
2835
2836 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2837 {
2838         struct tep_format_field *field = evsel__field(evsel, name);
2839
2840         if (!field)
2841                 return 0;
2842
2843         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2844 }
2845
2846 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2847 {
2848         int paranoid;
2849
2850         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2851             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2852             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2853                 /*
2854                  * If it's cycles then fall back to hrtimer based
2855                  * cpu-clock-tick sw counter, which is always available even if
2856                  * no PMU support.
2857                  *
2858                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2859                  * b0a873e).
2860                  */
2861                 scnprintf(msg, msgsize, "%s",
2862 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2863
2864                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2865                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2866
2867                 zfree(&evsel->name);
2868                 return true;
2869         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2870                    (paranoid = perf_event_paranoid()) > 1) {
2871                 const char *name = evsel__name(evsel);
2872                 char *new_name;
2873                 const char *sep = ":";
2874
2875                 /* If event has exclude user then don't exclude kernel. */
2876                 if (evsel->core.attr.exclude_user)
2877                         return false;
2878
2879                 /* Is there already the separator in the name. */
2880                 if (strchr(name, '/') ||
2881                     (strchr(name, ':') && !evsel->is_libpfm_event))
2882                         sep = "";
2883
2884                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2885                         return false;
2886
2887                 if (evsel->name)
2888                         free(evsel->name);
2889                 evsel->name = new_name;
2890                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2891                           "to fall back to excluding kernel and hypervisor "
2892                           " samples", paranoid);
2893                 evsel->core.attr.exclude_kernel = 1;
2894                 evsel->core.attr.exclude_hv     = 1;
2895
2896                 return true;
2897         }
2898
2899         return false;
2900 }
2901
2902 static bool find_process(const char *name)
2903 {
2904         size_t len = strlen(name);
2905         DIR *dir;
2906         struct dirent *d;
2907         int ret = -1;
2908
2909         dir = opendir(procfs__mountpoint());
2910         if (!dir)
2911                 return false;
2912
2913         /* Walk through the directory. */
2914         while (ret && (d = readdir(dir)) != NULL) {
2915                 char path[PATH_MAX];
2916                 char *data;
2917                 size_t size;
2918
2919                 if ((d->d_type != DT_DIR) ||
2920                      !strcmp(".", d->d_name) ||
2921                      !strcmp("..", d->d_name))
2922                         continue;
2923
2924                 scnprintf(path, sizeof(path), "%s/%s/comm",
2925                           procfs__mountpoint(), d->d_name);
2926
2927                 if (filename__read_str(path, &data, &size))
2928                         continue;
2929
2930                 ret = strncmp(name, data, len);
2931                 free(data);
2932         }
2933
2934         closedir(dir);
2935         return ret ? false : true;
2936 }
2937
2938 static bool is_amd(const char *arch, const char *cpuid)
2939 {
2940         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2941 }
2942
2943 static bool is_amd_ibs(struct evsel *evsel)
2944 {
2945         return evsel->core.attr.precise_ip
2946             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2947 }
2948
2949 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2950                          int err, char *msg, size_t size)
2951 {
2952         struct perf_env *env = evsel__env(evsel);
2953         const char *arch = perf_env__arch(env);
2954         const char *cpuid = perf_env__cpuid(env);
2955         char sbuf[STRERR_BUFSIZE];
2956         int printed = 0, enforced = 0;
2957
2958         switch (err) {
2959         case EPERM:
2960         case EACCES:
2961                 printed += scnprintf(msg + printed, size - printed,
2962                         "Access to performance monitoring and observability operations is limited.\n");
2963
2964                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2965                         if (enforced) {
2966                                 printed += scnprintf(msg + printed, size - printed,
2967                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2968                                         "monitoring and observability operations. Inspect system audit records for\n"
2969                                         "more perf_event access control information and adjusting the policy.\n");
2970                         }
2971                 }
2972
2973                 if (err == EPERM)
2974                         printed += scnprintf(msg, size,
2975                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2976
2977                 return scnprintf(msg + printed, size - printed,
2978                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2979                  "access to performance monitoring and observability operations for processes\n"
2980                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2981                  "More information can be found at 'Perf events and tool security' document:\n"
2982                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2983                  "perf_event_paranoid setting is %d:\n"
2984                  "  -1: Allow use of (almost) all events by all users\n"
2985                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2986                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2987                  ">= 1: Disallow CPU event access\n"
2988                  ">= 2: Disallow kernel profiling\n"
2989                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2990                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2991                  perf_event_paranoid());
2992         case ENOENT:
2993                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2994         case EMFILE:
2995                 return scnprintf(msg, size, "%s",
2996                          "Too many events are opened.\n"
2997                          "Probably the maximum number of open file descriptors has been reached.\n"
2998                          "Hint: Try again after reducing the number of events.\n"
2999                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3000         case ENOMEM:
3001                 if (evsel__has_callchain(evsel) &&
3002                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3003                         return scnprintf(msg, size,
3004                                          "Not enough memory to setup event with callchain.\n"
3005                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3006                                          "Hint: Current value: %d", sysctl__max_stack());
3007                 break;
3008         case ENODEV:
3009                 if (target->cpu_list)
3010                         return scnprintf(msg, size, "%s",
3011          "No such device - did you specify an out-of-range profile CPU?");
3012                 break;
3013         case EOPNOTSUPP:
3014                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3015                         return scnprintf(msg, size,
3016         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3017                                          evsel__name(evsel));
3018                 if (evsel->core.attr.aux_output)
3019                         return scnprintf(msg, size,
3020         "%s: PMU Hardware doesn't support 'aux_output' feature",
3021                                          evsel__name(evsel));
3022                 if (evsel->core.attr.sample_period != 0)
3023                         return scnprintf(msg, size,
3024         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3025                                          evsel__name(evsel));
3026                 if (evsel->core.attr.precise_ip)
3027                         return scnprintf(msg, size, "%s",
3028         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3029 #if defined(__i386__) || defined(__x86_64__)
3030                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3031                         return scnprintf(msg, size, "%s",
3032         "No hardware sampling interrupt available.\n");
3033 #endif
3034                 break;
3035         case EBUSY:
3036                 if (find_process("oprofiled"))
3037                         return scnprintf(msg, size,
3038         "The PMU counters are busy/taken by another profiler.\n"
3039         "We found oprofile daemon running, please stop it and try again.");
3040                 break;
3041         case EINVAL:
3042                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3043                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3044                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3045                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3046                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3047                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3048                 if (perf_missing_features.clockid)
3049                         return scnprintf(msg, size, "clockid feature not supported.");
3050                 if (perf_missing_features.clockid_wrong)
3051                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3052                 if (perf_missing_features.aux_output)
3053                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3054                 if (!target__has_cpu(target))
3055                         return scnprintf(msg, size,
3056         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3057                                         evsel__name(evsel));
3058                 if (is_amd(arch, cpuid)) {
3059                         if (is_amd_ibs(evsel)) {
3060                                 if (evsel->core.attr.exclude_kernel)
3061                                         return scnprintf(msg, size,
3062         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3063                                 if (!evsel->core.system_wide)
3064                                         return scnprintf(msg, size,
3065         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3066                         }
3067                 }
3068
3069                 break;
3070         case ENODATA:
3071                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3072                                  "Please add an auxiliary event in front of the load latency event.");
3073         default:
3074                 break;
3075         }
3076
3077         return scnprintf(msg, size,
3078         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3079         "/bin/dmesg | grep -i perf may provide additional information.\n",
3080                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3081 }
3082
3083 struct perf_env *evsel__env(struct evsel *evsel)
3084 {
3085         if (evsel && evsel->evlist && evsel->evlist->env)
3086                 return evsel->evlist->env;
3087         return &perf_env;
3088 }
3089
3090 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3091 {
3092         int cpu_map_idx, thread;
3093
3094         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3095                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3096                      thread++) {
3097                         int fd = FD(evsel, cpu_map_idx, thread);
3098
3099                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3100                                                    cpu_map_idx, thread, fd) < 0)
3101                                 return -1;
3102                 }
3103         }
3104
3105         return 0;
3106 }
3107
3108 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3109 {
3110         struct perf_cpu_map *cpus = evsel->core.cpus;
3111         struct perf_thread_map *threads = evsel->core.threads;
3112
3113         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3114                 return -ENOMEM;
3115
3116         return store_evsel_ids(evsel, evlist);
3117 }
3118
3119 void evsel__zero_per_pkg(struct evsel *evsel)
3120 {
3121         struct hashmap_entry *cur;
3122         size_t bkt;
3123
3124         if (evsel->per_pkg_mask) {
3125                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3126                         free((char *)cur->key);
3127
3128                 hashmap__clear(evsel->per_pkg_mask);
3129         }
3130 }
3131
3132 bool evsel__is_hybrid(struct evsel *evsel)
3133 {
3134         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3135 }
3136
3137 struct evsel *evsel__leader(struct evsel *evsel)
3138 {
3139         return container_of(evsel->core.leader, struct evsel, core);
3140 }
3141
3142 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3143 {
3144         return evsel->core.leader == &leader->core;
3145 }
3146
3147 bool evsel__is_leader(struct evsel *evsel)
3148 {
3149         return evsel__has_leader(evsel, evsel);
3150 }
3151
3152 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3153 {
3154         evsel->core.leader = &leader->core;
3155 }
3156
3157 int evsel__source_count(const struct evsel *evsel)
3158 {
3159         struct evsel *pos;
3160         int count = 0;
3161
3162         evlist__for_each_entry(evsel->evlist, pos) {
3163                 if (pos->metric_leader == evsel)
3164                         count++;
3165         }
3166         return count;
3167 }
3168
3169 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3170 {
3171         return false;
3172 }
3173
3174 /*
3175  * Remove an event from a given group (leader).
3176  * Some events, e.g., perf metrics Topdown events,
3177  * must always be grouped. Ignore the events.
3178  */
3179 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3180 {
3181         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3182                 evsel__set_leader(evsel, evsel);
3183                 evsel->core.nr_members = 0;
3184                 leader->core.nr_members--;
3185         }
3186 }