tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <linux/hw_breakpoint.h>
  16 #include <linux/perf_event.h>
  17 #include <linux/compiler.h>
  18 #include <linux/err.h>
  19 #include <linux/zalloc.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/resource.h>
  22 #include <sys/types.h>
  23 #include <dirent.h>
  24 #include <stdlib.h>
  25 #include <perf/evsel.h>
  26 #include "asm/bug.h"
  27 #include "bpf_counter.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "util/hashmap.h"
  49 #include "pmu-hybrid.h"
  50 #include "off_cpu.h"
  51 #include "../perf-sys.h"
  52 #include "util/parse-branch-options.h"
  53 #include <internal/xyarray.h>
  54 #include <internal/lib.h>
  55 #include <internal/threadmap.h>
  56
  57 #include <linux/ctype.h>
  58
  59 #ifdef HAVE_LIBTRACEEVENT
  60 #include <traceevent/event-parse.h>
  61 #endif
  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_events = NULL;
 289         evsel->per_pkg_mask  = NULL;
 290         evsel->collect_stat  = false;
 291         evsel->pmu_name      = NULL;
 292 }
 293
 294 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 295 {
 296         struct evsel *evsel = zalloc(perf_evsel__object.size);
 297
 298         if (!evsel)
 299                 return NULL;
 300         evsel__init(evsel, attr, idx);
 301
 302         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 303                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 304                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 305                 evsel->core.attr.sample_period = 1;
 306         }
 307
 308         if (evsel__is_clock(evsel)) {
 309                 free((char *)evsel->unit);
 310                 evsel->unit = strdup("msec");
 311                 evsel->scale = 1e-6;
 312         }
 313
 314         return evsel;
 315 }
 316
 317 static bool perf_event_can_profile_kernel(void)
 318 {
 319         return perf_event_paranoid_check(1);
 320 }
 321
 322 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 323 {
 324         struct perf_event_attr attr = {
 325                 .type   = type,
 326                 .config = config,
 327                 .exclude_kernel = !perf_event_can_profile_kernel(),
 328         };
 329         struct evsel *evsel;
 330
 331         event_attr_init(&attr);
 332
 333         /*
 334          * Now let the usual logic to set up the perf_event_attr defaults
 335          * to kick in when we return and before perf_evsel__open() is called.
 336          */
 337         evsel = evsel__new(&attr);
 338         if (evsel == NULL)
 339                 goto out;
 340
 341         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 342
 343         evsel->precise_max = true;
 344
 345         /* use asprintf() because free(evsel) assumes name is allocated */
 346         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 347                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 348                      attr.exclude_kernel ? "u" : "",
 349                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 350                 goto error_free;
 351 out:
 352         return evsel;
 353 error_free:
 354         evsel__delete(evsel);
 355         evsel = NULL;
 356         goto out;
 357 }
 358
 359 int copy_config_terms(struct list_head *dst, struct list_head *src)
 360 {
 361         struct evsel_config_term *pos, *tmp;
 362
 363         list_for_each_entry(pos, src, list) {
 364                 tmp = malloc(sizeof(*tmp));
 365                 if (tmp == NULL)
 366                         return -ENOMEM;
 367
 368                 *tmp = *pos;
 369                 if (tmp->free_str) {
 370                         tmp->val.str = strdup(pos->val.str);
 371                         if (tmp->val.str == NULL) {
 372                                 free(tmp);
 373                                 return -ENOMEM;
 374                         }
 375                 }
 376                 list_add_tail(&tmp->list, dst);
 377         }
 378         return 0;
 379 }
 380
 381 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 382 {
 383         return copy_config_terms(&dst->config_terms, &src->config_terms);
 384 }
 385
 386 /**
 387  * evsel__clone - create a new evsel copied from @orig
 388  * @orig: original evsel
 389  *
 390  * The assumption is that @orig is not configured nor opened yet.
 391  * So we only care about the attributes that can be set while it's parsed.
 392  */
 393 struct evsel *evsel__clone(struct evsel *orig)
 394 {
 395         struct evsel *evsel;
 396
 397         BUG_ON(orig->core.fd);
 398         BUG_ON(orig->counts);
 399         BUG_ON(orig->priv);
 400         BUG_ON(orig->per_pkg_mask);
 401
 402         /* cannot handle BPF objects for now */
 403         if (orig->bpf_obj)
 404                 return NULL;
 405
 406         evsel = evsel__new(&orig->core.attr);
 407         if (evsel == NULL)
 408                 return NULL;
 409
 410         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 411         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 412         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 413         evsel->core.nr_members = orig->core.nr_members;
 414         evsel->core.system_wide = orig->core.system_wide;
 415         evsel->core.requires_cpu = orig->core.requires_cpu;
 416
 417         if (orig->name) {
 418                 evsel->name = strdup(orig->name);
 419                 if (evsel->name == NULL)
 420                         goto out_err;
 421         }
 422         if (orig->group_name) {
 423                 evsel->group_name = strdup(orig->group_name);
 424                 if (evsel->group_name == NULL)
 425                         goto out_err;
 426         }
 427         if (orig->pmu_name) {
 428                 evsel->pmu_name = strdup(orig->pmu_name);
 429                 if (evsel->pmu_name == NULL)
 430                         goto out_err;
 431         }
 432         if (orig->filter) {
 433                 evsel->filter = strdup(orig->filter);
 434                 if (evsel->filter == NULL)
 435                         goto out_err;
 436         }
 437         if (orig->metric_id) {
 438                 evsel->metric_id = strdup(orig->metric_id);
 439                 if (evsel->metric_id == NULL)
 440                         goto out_err;
 441         }
 442         evsel->cgrp = cgroup__get(orig->cgrp);
 443 #ifdef HAVE_LIBTRACEEVENT
 444         evsel->tp_format = orig->tp_format;
 445 #endif
 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         evsel->pmu = orig->pmu;
 471
 472         if (evsel__copy_config_terms(evsel, orig) < 0)
 473                 goto out_err;
 474
 475         return evsel;
 476
 477 out_err:
 478         evsel__delete(evsel);
 479         return NULL;
 480 }
 481
 482 /*
 483  * Returns pointer with encoded error via <linux/err.h> interface.
 484  */
 485 #ifdef HAVE_LIBTRACEEVENT
 486 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 487 {
 488         struct evsel *evsel = zalloc(perf_evsel__object.size);
 489         int err = -ENOMEM;
 490
 491         if (evsel == NULL) {
 492                 goto out_err;
 493         } else {
 494                 struct perf_event_attr attr = {
 495                         .type          = PERF_TYPE_TRACEPOINT,
 496                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 497                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 498                 };
 499
 500                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 501                         goto out_free;
 502
 503                 evsel->tp_format = trace_event__tp_format(sys, name);
 504                 if (IS_ERR(evsel->tp_format)) {
 505                         err = PTR_ERR(evsel->tp_format);
 506                         goto out_free;
 507                 }
 508
 509                 event_attr_init(&attr);
 510                 attr.config = evsel->tp_format->id;
 511                 attr.sample_period = 1;
 512                 evsel__init(evsel, &attr, idx);
 513         }
 514
 515         return evsel;
 516
 517 out_free:
 518         zfree(&evsel->name);
 519         free(evsel);
 520 out_err:
 521         return ERR_PTR(err);
 522 }
 523 #endif
 524
 525 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 526         "cycles",
 527         "instructions",
 528         "cache-references",
 529         "cache-misses",
 530         "branches",
 531         "branch-misses",
 532         "bus-cycles",
 533         "stalled-cycles-frontend",
 534         "stalled-cycles-backend",
 535         "ref-cycles",
 536 };
 537
 538 char *evsel__bpf_counter_events;
 539
 540 bool evsel__match_bpf_counter_events(const char *name)
 541 {
 542         int name_len;
 543         bool match;
 544         char *ptr;
 545
 546         if (!evsel__bpf_counter_events)
 547                 return false;
 548
 549         ptr = strstr(evsel__bpf_counter_events, name);
 550         name_len = strlen(name);
 551
 552         /* check name matches a full token in evsel__bpf_counter_events */
 553         match = (ptr != NULL) &&
 554                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 555                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 556
 557         return match;
 558 }
 559
 560 static const char *__evsel__hw_name(u64 config)
 561 {
 562         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 563                 return evsel__hw_names[config];
 564
 565         return "unknown-hardware";
 566 }
 567
 568 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 569 {
 570         int colon = 0, r = 0;
 571         struct perf_event_attr *attr = &evsel->core.attr;
 572         bool exclude_guest_default = false;
 573
 574 #define MOD_PRINT(context, mod) do {                                    \
 575                 if (!attr->exclude_##context) {                         \
 576                         if (!colon) colon = ++r;                        \
 577                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 578                 } } while(0)
 579
 580         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 581                 MOD_PRINT(kernel, 'k');
 582                 MOD_PRINT(user, 'u');
 583                 MOD_PRINT(hv, 'h');
 584                 exclude_guest_default = true;
 585         }
 586
 587         if (attr->precise_ip) {
 588                 if (!colon)
 589                         colon = ++r;
 590                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 591                 exclude_guest_default = true;
 592         }
 593
 594         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 595                 MOD_PRINT(host, 'H');
 596                 MOD_PRINT(guest, 'G');
 597         }
 598 #undef MOD_PRINT
 599         if (colon)
 600                 bf[colon - 1] = ':';
 601         return r;
 602 }
 603
 604 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 605 {
 606         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 607 }
 608
 609 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 610 {
 611         int r = arch_evsel__hw_name(evsel, bf, size);
 612         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 613 }
 614
 615 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 616         "cpu-clock",
 617         "task-clock",
 618         "page-faults",
 619         "context-switches",
 620         "cpu-migrations",
 621         "minor-faults",
 622         "major-faults",
 623         "alignment-faults",
 624         "emulation-faults",
 625         "dummy",
 626 };
 627
 628 static const char *__evsel__sw_name(u64 config)
 629 {
 630         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 631                 return evsel__sw_names[config];
 632         return "unknown-software";
 633 }
 634
 635 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 636 {
 637         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 638         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 639 }
 640
 641 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 642 {
 643         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 644 }
 645
 646 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 647 {
 648         int r;
 649
 650         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 651
 652         if (type & HW_BREAKPOINT_R)
 653                 r += scnprintf(bf + r, size - r, "r");
 654
 655         if (type & HW_BREAKPOINT_W)
 656                 r += scnprintf(bf + r, size - r, "w");
 657
 658         if (type & HW_BREAKPOINT_X)
 659                 r += scnprintf(bf + r, size - r, "x");
 660
 661         return r;
 662 }
 663
 664 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 665 {
 666         struct perf_event_attr *attr = &evsel->core.attr;
 667         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 668         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 669 }
 670
 671 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 672  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 673  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 674  { "LLC",       "L2",                                                   },
 675  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 676  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 677  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 678  { "node",                                                              },
 679 };
 680
 681 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 682  { "load",      "loads",        "read",                                 },
 683  { "store",     "stores",       "write",                                },
 684  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 685 };
 686
 687 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 688  { "refs",      "Reference",    "ops",          "access",               },
 689  { "misses",    "miss",                                                 },
 690 };
 691
 692 #define C(x)            PERF_COUNT_HW_CACHE_##x
 693 #define CACHE_READ      (1 << C(OP_READ))
 694 #define CACHE_WRITE     (1 << C(OP_WRITE))
 695 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 696 #define COP(x)          (1 << x)
 697
 698 /*
 699  * cache operation stat
 700  * L1I : Read and prefetch only
 701  * ITLB and BPU : Read-only
 702  */
 703 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 704  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 705  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 706  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 707  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 708  [C(ITLB)]      = (CACHE_READ),
 709  [C(BPU)]       = (CACHE_READ),
 710  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 711 };
 712
 713 bool evsel__is_cache_op_valid(u8 type, u8 op)
 714 {
 715         if (evsel__hw_cache_stat[type] & COP(op))
 716                 return true;    /* valid */
 717         else
 718                 return false;   /* invalid */
 719 }
 720
 721 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 722 {
 723         if (result) {
 724                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 725                                  evsel__hw_cache_op[op][0],
 726                                  evsel__hw_cache_result[result][0]);
 727         }
 728
 729         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 730                          evsel__hw_cache_op[op][1]);
 731 }
 732
 733 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 734 {
 735         u8 op, result, type = (config >>  0) & 0xff;
 736         const char *err = "unknown-ext-hardware-cache-type";
 737
 738         if (type >= PERF_COUNT_HW_CACHE_MAX)
 739                 goto out_err;
 740
 741         op = (config >>  8) & 0xff;
 742         err = "unknown-ext-hardware-cache-op";
 743         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 744                 goto out_err;
 745
 746         result = (config >> 16) & 0xff;
 747         err = "unknown-ext-hardware-cache-result";
 748         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 749                 goto out_err;
 750
 751         err = "invalid-cache";
 752         if (!evsel__is_cache_op_valid(type, op))
 753                 goto out_err;
 754
 755         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 756 out_err:
 757         return scnprintf(bf, size, "%s", err);
 758 }
 759
 760 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 761 {
 762         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 763         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 764 }
 765
 766 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 767 {
 768         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 769         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 770 }
 771
 772 const char *evsel__name(struct evsel *evsel)
 773 {
 774         char bf[128];
 775
 776         if (!evsel)
 777                 goto out_unknown;
 778
 779         if (evsel->name)
 780                 return evsel->name;
 781
 782         switch (evsel->core.attr.type) {
 783         case PERF_TYPE_RAW:
 784                 evsel__raw_name(evsel, bf, sizeof(bf));
 785                 break;
 786
 787         case PERF_TYPE_HARDWARE:
 788                 evsel__hw_name(evsel, bf, sizeof(bf));
 789                 break;
 790
 791         case PERF_TYPE_HW_CACHE:
 792                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 793                 break;
 794
 795         case PERF_TYPE_SOFTWARE:
 796                 if (evsel__is_tool(evsel))
 797                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 798                 else
 799                         evsel__sw_name(evsel, bf, sizeof(bf));
 800                 break;
 801
 802         case PERF_TYPE_TRACEPOINT:
 803                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 804                 break;
 805
 806         case PERF_TYPE_BREAKPOINT:
 807                 evsel__bp_name(evsel, bf, sizeof(bf));
 808                 break;
 809
 810         default:
 811                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 812                           evsel->core.attr.type);
 813                 break;
 814         }
 815
 816         evsel->name = strdup(bf);
 817
 818         if (evsel->name)
 819                 return evsel->name;
 820 out_unknown:
 821         return "unknown";
 822 }
 823
 824 const char *evsel__metric_id(const struct evsel *evsel)
 825 {
 826         if (evsel->metric_id)
 827                 return evsel->metric_id;
 828
 829         if (evsel__is_tool(evsel))
 830                 return perf_tool_event__to_str(evsel->tool_event);
 831
 832         return "unknown";
 833 }
 834
 835 const char *evsel__group_name(struct evsel *evsel)
 836 {
 837         return evsel->group_name ?: "anon group";
 838 }
 839
 840 /*
 841  * Returns the group details for the specified leader,
 842  * with following rules.
 843  *
 844  *  For record -e '{cycles,instructions}'
 845  *    'anon group { cycles:u, instructions:u }'
 846  *
 847  *  For record -e 'cycles,instructions' and report --group
 848  *    'cycles:u, instructions:u'
 849  */
 850 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 851 {
 852         int ret = 0;
 853         struct evsel *pos;
 854         const char *group_name = evsel__group_name(evsel);
 855
 856         if (!evsel->forced_leader)
 857                 ret = scnprintf(buf, size, "%s { ", group_name);
 858
 859         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 860
 861         for_each_group_member(pos, evsel)
 862                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 863
 864         if (!evsel->forced_leader)
 865                 ret += scnprintf(buf + ret, size - ret, " }");
 866
 867         return ret;
 868 }
 869
 870 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 871                                       struct callchain_param *param)
 872 {
 873         bool function = evsel__is_function_event(evsel);
 874         struct perf_event_attr *attr = &evsel->core.attr;
 875
 876         evsel__set_sample_bit(evsel, CALLCHAIN);
 877
 878         attr->sample_max_stack = param->max_stack;
 879
 880         if (opts->kernel_callchains)
 881                 attr->exclude_callchain_user = 1;
 882         if (opts->user_callchains)
 883                 attr->exclude_callchain_kernel = 1;
 884         if (param->record_mode == CALLCHAIN_LBR) {
 885                 if (!opts->branch_stack) {
 886                         if (attr->exclude_user) {
 887                                 pr_warning("LBR callstack option is only available "
 888                                            "to get user callchain information. "
 889                                            "Falling back to framepointers.\n");
 890                         } else {
 891                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 892                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 893                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 894                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 895                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 896                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 897                         }
 898                 } else
 899                          pr_warning("Cannot use LBR callstack with branch stack. "
 900                                     "Falling back to framepointers.\n");
 901         }
 902
 903         if (param->record_mode == CALLCHAIN_DWARF) {
 904                 if (!function) {
 905                         evsel__set_sample_bit(evsel, REGS_USER);
 906                         evsel__set_sample_bit(evsel, STACK_USER);
 907                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 908                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 909                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 910                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 911                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 912                         } else {
 913                                 attr->sample_regs_user |= arch__user_reg_mask();
 914                         }
 915                         attr->sample_stack_user = param->dump_size;
 916                         attr->exclude_callchain_user = 1;
 917                 } else {
 918                         pr_info("Cannot use DWARF unwind for function trace event,"
 919                                 " falling back to framepointers.\n");
 920                 }
 921         }
 922
 923         if (function) {
 924                 pr_info("Disabling user space callchains for function trace event.\n");
 925                 attr->exclude_callchain_user = 1;
 926         }
 927 }
 928
 929 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 930                              struct callchain_param *param)
 931 {
 932         if (param->enabled)
 933                 return __evsel__config_callchain(evsel, opts, param);
 934 }
 935
 936 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 937 {
 938         struct perf_event_attr *attr = &evsel->core.attr;
 939
 940         evsel__reset_sample_bit(evsel, CALLCHAIN);
 941         if (param->record_mode == CALLCHAIN_LBR) {
 942                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 943                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 944                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 945                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 946         }
 947         if (param->record_mode == CALLCHAIN_DWARF) {
 948                 evsel__reset_sample_bit(evsel, REGS_USER);
 949                 evsel__reset_sample_bit(evsel, STACK_USER);
 950         }
 951 }
 952
 953 static void evsel__apply_config_terms(struct evsel *evsel,
 954                                       struct record_opts *opts, bool track)
 955 {
 956         struct evsel_config_term *term;
 957         struct list_head *config_terms = &evsel->config_terms;
 958         struct perf_event_attr *attr = &evsel->core.attr;
 959         /* callgraph default */
 960         struct callchain_param param = {
 961                 .record_mode = callchain_param.record_mode,
 962         };
 963         u32 dump_size = 0;
 964         int max_stack = 0;
 965         const char *callgraph_buf = NULL;
 966
 967         list_for_each_entry(term, config_terms, list) {
 968                 switch (term->type) {
 969                 case EVSEL__CONFIG_TERM_PERIOD:
 970                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 971                                 attr->sample_period = term->val.period;
 972                                 attr->freq = 0;
 973                                 evsel__reset_sample_bit(evsel, PERIOD);
 974                         }
 975                         break;
 976                 case EVSEL__CONFIG_TERM_FREQ:
 977                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 978                                 attr->sample_freq = term->val.freq;
 979                                 attr->freq = 1;
 980                                 evsel__set_sample_bit(evsel, PERIOD);
 981                         }
 982                         break;
 983                 case EVSEL__CONFIG_TERM_TIME:
 984                         if (term->val.time)
 985                                 evsel__set_sample_bit(evsel, TIME);
 986                         else
 987                                 evsel__reset_sample_bit(evsel, TIME);
 988                         break;
 989                 case EVSEL__CONFIG_TERM_CALLGRAPH:
 990                         callgraph_buf = term->val.str;
 991                         break;
 992                 case EVSEL__CONFIG_TERM_BRANCH:
 993                         if (term->val.str && strcmp(term->val.str, "no")) {
 994                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 995                                 parse_branch_str(term->val.str,
 996                                                  &attr->branch_sample_type);
 997                         } else
 998                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 999                         break;
1000                 case EVSEL__CONFIG_TERM_STACK_USER:
1001                         dump_size = term->val.stack_user;
1002                         break;
1003                 case EVSEL__CONFIG_TERM_MAX_STACK:
1004                         max_stack = term->val.max_stack;
1005                         break;
1006                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1007                         evsel->max_events = term->val.max_events;
1008                         break;
1009                 case EVSEL__CONFIG_TERM_INHERIT:
1010                         /*
1011                          * attr->inherit should has already been set by
1012                          * evsel__config. If user explicitly set
1013                          * inherit using config terms, override global
1014                          * opt->no_inherit setting.
1015                          */
1016                         attr->inherit = term->val.inherit ? 1 : 0;
1017                         break;
1018                 case EVSEL__CONFIG_TERM_OVERWRITE:
1019                         attr->write_backward = term->val.overwrite ? 1 : 0;
1020                         break;
1021                 case EVSEL__CONFIG_TERM_DRV_CFG:
1022                         break;
1023                 case EVSEL__CONFIG_TERM_PERCORE:
1024                         break;
1025                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1026                         attr->aux_output = term->val.aux_output ? 1 : 0;
1027                         break;
1028                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1029                         /* Already applied by auxtrace */
1030                         break;
1031                 case EVSEL__CONFIG_TERM_CFG_CHG:
1032                         break;
1033                 default:
1034                         break;
1035                 }
1036         }
1037
1038         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1039         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1040                 bool sample_address = false;
1041
1042                 if (max_stack) {
1043                         param.max_stack = max_stack;
1044                         if (callgraph_buf == NULL)
1045                                 callgraph_buf = "fp";
1046                 }
1047
1048                 /* parse callgraph parameters */
1049                 if (callgraph_buf != NULL) {
1050                         if (!strcmp(callgraph_buf, "no")) {
1051                                 param.enabled = false;
1052                                 param.record_mode = CALLCHAIN_NONE;
1053                         } else {
1054                                 param.enabled = true;
1055                                 if (parse_callchain_record(callgraph_buf, &param)) {
1056                                         pr_err("per-event callgraph setting for %s failed. "
1057                                                "Apply callgraph global setting for it\n",
1058                                                evsel->name);
1059                                         return;
1060                                 }
1061                                 if (param.record_mode == CALLCHAIN_DWARF)
1062                                         sample_address = true;
1063                         }
1064                 }
1065                 if (dump_size > 0) {
1066                         dump_size = round_up(dump_size, sizeof(u64));
1067                         param.dump_size = dump_size;
1068                 }
1069
1070                 /* If global callgraph set, clear it */
1071                 if (callchain_param.enabled)
1072                         evsel__reset_callgraph(evsel, &callchain_param);
1073
1074                 /* set perf-event callgraph */
1075                 if (param.enabled) {
1076                         if (sample_address) {
1077                                 evsel__set_sample_bit(evsel, ADDR);
1078                                 evsel__set_sample_bit(evsel, DATA_SRC);
1079                                 evsel->core.attr.mmap_data = track;
1080                         }
1081                         evsel__config_callchain(evsel, opts, &param);
1082                 }
1083         }
1084 }
1085
1086 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1087 {
1088         struct evsel_config_term *term, *found_term = NULL;
1089
1090         list_for_each_entry(term, &evsel->config_terms, list) {
1091                 if (term->type == type)
1092                         found_term = term;
1093         }
1094
1095         return found_term;
1096 }
1097
1098 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1099 {
1100         evsel__set_sample_bit(evsel, WEIGHT);
1101 }
1102
1103 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1104 {
1105 }
1106
1107 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1108                                     struct perf_event_attr *attr __maybe_unused)
1109 {
1110 }
1111
1112 static void evsel__set_default_freq_period(struct record_opts *opts,
1113                                            struct perf_event_attr *attr)
1114 {
1115         if (opts->freq) {
1116                 attr->freq = 1;
1117                 attr->sample_freq = opts->freq;
1118         } else {
1119                 attr->sample_period = opts->default_interval;
1120         }
1121 }
1122
1123 static bool evsel__is_offcpu_event(struct evsel *evsel)
1124 {
1125         return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1126 }
1127
1128 /*
1129  * The enable_on_exec/disabled value strategy:
1130  *
1131  *  1) For any type of traced program:
1132  *    - all independent events and group leaders are disabled
1133  *    - all group members are enabled
1134  *
1135  *     Group members are ruled by group leaders. They need to
1136  *     be enabled, because the group scheduling relies on that.
1137  *
1138  *  2) For traced programs executed by perf:
1139  *     - all independent events and group leaders have
1140  *       enable_on_exec set
1141  *     - we don't specifically enable or disable any event during
1142  *       the record command
1143  *
1144  *     Independent events and group leaders are initially disabled
1145  *     and get enabled by exec. Group members are ruled by group
1146  *     leaders as stated in 1).
1147  *
1148  *  3) For traced programs attached by perf (pid/tid):
1149  *     - we specifically enable or disable all events during
1150  *       the record command
1151  *
1152  *     When attaching events to already running traced we
1153  *     enable/disable events specifically, as there's no
1154  *     initial traced exec call.
1155  */
1156 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1157                    struct callchain_param *callchain)
1158 {
1159         struct evsel *leader = evsel__leader(evsel);
1160         struct perf_event_attr *attr = &evsel->core.attr;
1161         int track = evsel->tracking;
1162         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1163
1164         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1165         attr->inherit       = !opts->no_inherit;
1166         attr->write_backward = opts->overwrite ? 1 : 0;
1167         attr->read_format   = PERF_FORMAT_LOST;
1168
1169         evsel__set_sample_bit(evsel, IP);
1170         evsel__set_sample_bit(evsel, TID);
1171
1172         if (evsel->sample_read) {
1173                 evsel__set_sample_bit(evsel, READ);
1174
1175                 /*
1176                  * We need ID even in case of single event, because
1177                  * PERF_SAMPLE_READ process ID specific data.
1178                  */
1179                 evsel__set_sample_id(evsel, false);
1180
1181                 /*
1182                  * Apply group format only if we belong to group
1183                  * with more than one members.
1184                  */
1185                 if (leader->core.nr_members > 1) {
1186                         attr->read_format |= PERF_FORMAT_GROUP;
1187                         attr->inherit = 0;
1188                 }
1189         }
1190
1191         /*
1192          * We default some events to have a default interval. But keep
1193          * it a weak assumption overridable by the user.
1194          */
1195         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1196             (!evsel->is_libpfm_event && (!attr->sample_period ||
1197                                          opts->user_freq != UINT_MAX ||
1198                                          opts->user_interval != ULLONG_MAX)))
1199                 evsel__set_default_freq_period(opts, attr);
1200
1201         /*
1202          * If attr->freq was set (here or earlier), ask for period
1203          * to be sampled.
1204          */
1205         if (attr->freq)
1206                 evsel__set_sample_bit(evsel, PERIOD);
1207
1208         if (opts->no_samples)
1209                 attr->sample_freq = 0;
1210
1211         if (opts->inherit_stat) {
1212                 evsel->core.attr.read_format |=
1213                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1214                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1215                         PERF_FORMAT_ID;
1216                 attr->inherit_stat = 1;
1217         }
1218
1219         if (opts->sample_address) {
1220                 evsel__set_sample_bit(evsel, ADDR);
1221                 attr->mmap_data = track;
1222         }
1223
1224         /*
1225          * We don't allow user space callchains for  function trace
1226          * event, due to issues with page faults while tracing page
1227          * fault handler and its overall trickiness nature.
1228          */
1229         if (evsel__is_function_event(evsel))
1230                 evsel->core.attr.exclude_callchain_user = 1;
1231
1232         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1233                 evsel__config_callchain(evsel, opts, callchain);
1234
1235         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1236             !evsel__is_dummy_event(evsel)) {
1237                 attr->sample_regs_intr = opts->sample_intr_regs;
1238                 evsel__set_sample_bit(evsel, REGS_INTR);
1239         }
1240
1241         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1242             !evsel__is_dummy_event(evsel)) {
1243                 attr->sample_regs_user |= opts->sample_user_regs;
1244                 evsel__set_sample_bit(evsel, REGS_USER);
1245         }
1246
1247         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1248                 evsel__set_sample_bit(evsel, CPU);
1249
1250         /*
1251          * When the user explicitly disabled time don't force it here.
1252          */
1253         if (opts->sample_time &&
1254             (!perf_missing_features.sample_id_all &&
1255             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1256              opts->sample_time_set)))
1257                 evsel__set_sample_bit(evsel, TIME);
1258
1259         if (opts->raw_samples && !evsel->no_aux_samples) {
1260                 evsel__set_sample_bit(evsel, TIME);
1261                 evsel__set_sample_bit(evsel, RAW);
1262                 evsel__set_sample_bit(evsel, CPU);
1263         }
1264
1265         if (opts->sample_address)
1266                 evsel__set_sample_bit(evsel, DATA_SRC);
1267
1268         if (opts->sample_phys_addr)
1269                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1270
1271         if (opts->no_buffering) {
1272                 attr->watermark = 0;
1273                 attr->wakeup_events = 1;
1274         }
1275         if (opts->branch_stack && !evsel->no_aux_samples) {
1276                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1277                 attr->branch_sample_type = opts->branch_stack;
1278         }
1279
1280         if (opts->sample_weight)
1281                 arch_evsel__set_sample_weight(evsel);
1282
1283         attr->task     = track;
1284         attr->mmap     = track;
1285         attr->mmap2    = track && !perf_missing_features.mmap2;
1286         attr->comm     = track;
1287         attr->build_id = track && opts->build_id;
1288
1289         /*
1290          * ksymbol is tracked separately with text poke because it needs to be
1291          * system wide and enabled immediately.
1292          */
1293         if (!opts->text_poke)
1294                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1295         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1296
1297         if (opts->record_namespaces)
1298                 attr->namespaces  = track;
1299
1300         if (opts->record_cgroup) {
1301                 attr->cgroup = track && !perf_missing_features.cgroup;
1302                 evsel__set_sample_bit(evsel, CGROUP);
1303         }
1304
1305         if (opts->sample_data_page_size)
1306                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1307
1308         if (opts->sample_code_page_size)
1309                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1310
1311         if (opts->record_switch_events)
1312                 attr->context_switch = track;
1313
1314         if (opts->sample_transaction)
1315                 evsel__set_sample_bit(evsel, TRANSACTION);
1316
1317         if (opts->running_time) {
1318                 evsel->core.attr.read_format |=
1319                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1320                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1321         }
1322
1323         /*
1324          * XXX see the function comment above
1325          *
1326          * Disabling only independent events or group leaders,
1327          * keeping group members enabled.
1328          */
1329         if (evsel__is_group_leader(evsel))
1330                 attr->disabled = 1;
1331
1332         /*
1333          * Setting enable_on_exec for independent events and
1334          * group leaders for traced executed by perf.
1335          */
1336         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1337             !opts->target.initial_delay)
1338                 attr->enable_on_exec = 1;
1339
1340         if (evsel->immediate) {
1341                 attr->disabled = 0;
1342                 attr->enable_on_exec = 0;
1343         }
1344
1345         clockid = opts->clockid;
1346         if (opts->use_clockid) {
1347                 attr->use_clockid = 1;
1348                 attr->clockid = opts->clockid;
1349         }
1350
1351         if (evsel->precise_max)
1352                 attr->precise_ip = 3;
1353
1354         if (opts->all_user) {
1355                 attr->exclude_kernel = 1;
1356                 attr->exclude_user   = 0;
1357         }
1358
1359         if (opts->all_kernel) {
1360                 attr->exclude_kernel = 0;
1361                 attr->exclude_user   = 1;
1362         }
1363
1364         if (evsel->core.own_cpus || evsel->unit)
1365                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1366
1367         /*
1368          * Apply event specific term settings,
1369          * it overloads any global configuration.
1370          */
1371         evsel__apply_config_terms(evsel, opts, track);
1372
1373         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1374
1375         /* The --period option takes the precedence. */
1376         if (opts->period_set) {
1377                 if (opts->period)
1378                         evsel__set_sample_bit(evsel, PERIOD);
1379                 else
1380                         evsel__reset_sample_bit(evsel, PERIOD);
1381         }
1382
1383         /*
1384          * A dummy event never triggers any actual counter and therefore
1385          * cannot be used with branch_stack.
1386          *
1387          * For initial_delay, a dummy event is added implicitly.
1388          * The software event will trigger -EOPNOTSUPP error out,
1389          * if BRANCH_STACK bit is set.
1390          */
1391         if (evsel__is_dummy_event(evsel))
1392                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1393
1394         if (evsel__is_offcpu_event(evsel))
1395                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1396
1397         arch__post_evsel_config(evsel, attr);
1398 }
1399
1400 int evsel__set_filter(struct evsel *evsel, const char *filter)
1401 {
1402         char *new_filter = strdup(filter);
1403
1404         if (new_filter != NULL) {
1405                 free(evsel->filter);
1406                 evsel->filter = new_filter;
1407                 return 0;
1408         }
1409
1410         return -1;
1411 }
1412
1413 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1414 {
1415         char *new_filter;
1416
1417         if (evsel->filter == NULL)
1418                 return evsel__set_filter(evsel, filter);
1419
1420         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1421                 free(evsel->filter);
1422                 evsel->filter = new_filter;
1423                 return 0;
1424         }
1425
1426         return -1;
1427 }
1428
1429 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1430 {
1431         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1432 }
1433
1434 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1435 {
1436         return evsel__append_filter(evsel, "%s,%s", filter);
1437 }
1438
1439 /* Caller has to clear disabled after going through all CPUs. */
1440 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1441 {
1442         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1443 }
1444
1445 int evsel__enable(struct evsel *evsel)
1446 {
1447         int err = perf_evsel__enable(&evsel->core);
1448
1449         if (!err)
1450                 evsel->disabled = false;
1451         return err;
1452 }
1453
1454 /* Caller has to set disabled after going through all CPUs. */
1455 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1456 {
1457         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1458 }
1459
1460 int evsel__disable(struct evsel *evsel)
1461 {
1462         int err = perf_evsel__disable(&evsel->core);
1463         /*
1464          * We mark it disabled here so that tools that disable a event can
1465          * ignore events after they disable it. I.e. the ring buffer may have
1466          * already a few more events queued up before the kernel got the stop
1467          * request.
1468          */
1469         if (!err)
1470                 evsel->disabled = true;
1471
1472         return err;
1473 }
1474
1475 void free_config_terms(struct list_head *config_terms)
1476 {
1477         struct evsel_config_term *term, *h;
1478
1479         list_for_each_entry_safe(term, h, config_terms, list) {
1480                 list_del_init(&term->list);
1481                 if (term->free_str)
1482                         zfree(&term->val.str);
1483                 free(term);
1484         }
1485 }
1486
1487 static void evsel__free_config_terms(struct evsel *evsel)
1488 {
1489         free_config_terms(&evsel->config_terms);
1490 }
1491
1492 void evsel__exit(struct evsel *evsel)
1493 {
1494         assert(list_empty(&evsel->core.node));
1495         assert(evsel->evlist == NULL);
1496         bpf_counter__destroy(evsel);
1497         evsel__free_counts(evsel);
1498         perf_evsel__free_fd(&evsel->core);
1499         perf_evsel__free_id(&evsel->core);
1500         evsel__free_config_terms(evsel);
1501         cgroup__put(evsel->cgrp);
1502         perf_cpu_map__put(evsel->core.cpus);
1503         perf_cpu_map__put(evsel->core.own_cpus);
1504         perf_thread_map__put(evsel->core.threads);
1505         zfree(&evsel->group_name);
1506         zfree(&evsel->name);
1507         zfree(&evsel->pmu_name);
1508         zfree(&evsel->unit);
1509         zfree(&evsel->metric_id);
1510         evsel__zero_per_pkg(evsel);
1511         hashmap__free(evsel->per_pkg_mask);
1512         evsel->per_pkg_mask = NULL;
1513         zfree(&evsel->metric_events);
1514         perf_evsel__object.fini(evsel);
1515 }
1516
1517 void evsel__delete(struct evsel *evsel)
1518 {
1519         evsel__exit(evsel);
1520         free(evsel);
1521 }
1522
1523 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1524                            struct perf_counts_values *count)
1525 {
1526         struct perf_counts_values tmp;
1527
1528         if (!evsel->prev_raw_counts)
1529                 return;
1530
1531         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1532         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1533
1534         count->val = count->val - tmp.val;
1535         count->ena = count->ena - tmp.ena;
1536         count->run = count->run - tmp.run;
1537 }
1538
1539 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1540 {
1541         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1542
1543         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1544 }
1545
1546 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1547                              u64 val, u64 ena, u64 run, u64 lost)
1548 {
1549         struct perf_counts_values *count;
1550
1551         count = perf_counts(counter->counts, cpu_map_idx, thread);
1552
1553         count->val    = val;
1554         count->ena    = ena;
1555         count->run    = run;
1556         count->lost   = lost;
1557
1558         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1559 }
1560
1561 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1562 {
1563         u64 read_format = leader->core.attr.read_format;
1564         struct sample_read_value *v;
1565         u64 nr, ena = 0, run = 0, lost = 0;
1566
1567         nr = *data++;
1568
1569         if (nr != (u64) leader->core.nr_members)
1570                 return -EINVAL;
1571
1572         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1573                 ena = *data++;
1574
1575         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1576                 run = *data++;
1577
1578         v = (void *)data;
1579         sample_read_group__for_each(v, nr, read_format) {
1580                 struct evsel *counter;
1581
1582                 counter = evlist__id2evsel(leader->evlist, v->id);
1583                 if (!counter)
1584                         return -EINVAL;
1585
1586                 if (read_format & PERF_FORMAT_LOST)
1587                         lost = v->lost;
1588
1589                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1590         }
1591
1592         return 0;
1593 }
1594
1595 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1596 {
1597         struct perf_stat_evsel *ps = leader->stats;
1598         u64 read_format = leader->core.attr.read_format;
1599         int size = perf_evsel__read_size(&leader->core);
1600         u64 *data = ps->group_data;
1601
1602         if (!(read_format & PERF_FORMAT_ID))
1603                 return -EINVAL;
1604
1605         if (!evsel__is_group_leader(leader))
1606                 return -EINVAL;
1607
1608         if (!data) {
1609                 data = zalloc(size);
1610                 if (!data)
1611                         return -ENOMEM;
1612
1613                 ps->group_data = data;
1614         }
1615
1616         if (FD(leader, cpu_map_idx, thread) < 0)
1617                 return -EINVAL;
1618
1619         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1620                 return -errno;
1621
1622         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1623 }
1624
1625 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1626 {
1627         u64 read_format = evsel->core.attr.read_format;
1628
1629         if (read_format & PERF_FORMAT_GROUP)
1630                 return evsel__read_group(evsel, cpu_map_idx, thread);
1631
1632         return evsel__read_one(evsel, cpu_map_idx, thread);
1633 }
1634
1635 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1636 {
1637         struct perf_counts_values count;
1638         size_t nv = scale ? 3 : 1;
1639
1640         if (FD(evsel, cpu_map_idx, thread) < 0)
1641                 return -EINVAL;
1642
1643         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1644                 return -ENOMEM;
1645
1646         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1647                 return -errno;
1648
1649         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1650         perf_counts_values__scale(&count, scale, NULL);
1651         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1652         return 0;
1653 }
1654
1655 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1656                                   int cpu_map_idx)
1657 {
1658         struct perf_cpu cpu;
1659
1660         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1661         return perf_cpu_map__idx(other->core.cpus, cpu);
1662 }
1663
1664 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1665 {
1666         struct evsel *leader = evsel__leader(evsel);
1667
1668         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1669             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1670                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1671         }
1672
1673         return cpu_map_idx;
1674 }
1675
1676 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1677 {
1678         struct evsel *leader = evsel__leader(evsel);
1679         int fd;
1680
1681         if (evsel__is_group_leader(evsel))
1682                 return -1;
1683
1684         /*
1685          * Leader must be already processed/open,
1686          * if not it's a bug.
1687          */
1688         BUG_ON(!leader->core.fd);
1689
1690         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1691         if (cpu_map_idx == -1)
1692                 return -1;
1693
1694         fd = FD(leader, cpu_map_idx, thread);
1695         BUG_ON(fd == -1);
1696
1697         return fd;
1698 }
1699
1700 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1701 {
1702         for (int cpu = 0; cpu < nr_cpus; cpu++)
1703                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1704                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1705 }
1706
1707 static int update_fds(struct evsel *evsel,
1708                       int nr_cpus, int cpu_map_idx,
1709                       int nr_threads, int thread_idx)
1710 {
1711         struct evsel *pos;
1712
1713         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1714                 return -EINVAL;
1715
1716         evlist__for_each_entry(evsel->evlist, pos) {
1717                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1718
1719                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1720
1721                 /*
1722                  * Since fds for next evsel has not been created,
1723                  * there is no need to iterate whole event list.
1724                  */
1725                 if (pos == evsel)
1726                         break;
1727         }
1728         return 0;
1729 }
1730
1731 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1732                                          int nr_cpus, int cpu_map_idx,
1733                                          struct perf_thread_map *threads,
1734                                          int thread, int err)
1735 {
1736         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1737
1738         if (!evsel->ignore_missing_thread)
1739                 return false;
1740
1741         /* The system wide setup does not work with threads. */
1742         if (evsel->core.system_wide)
1743                 return false;
1744
1745         /* The -ESRCH is perf event syscall errno for pid's not found. */
1746         if (err != -ESRCH)
1747                 return false;
1748
1749         /* If there's only one thread, let it fail. */
1750         if (threads->nr == 1)
1751                 return false;
1752
1753         /*
1754          * We should remove fd for missing_thread first
1755          * because thread_map__remove() will decrease threads->nr.
1756          */
1757         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1758                 return false;
1759
1760         if (thread_map__remove(threads, thread))
1761                 return false;
1762
1763         pr_warning("WARNING: Ignored open failure for pid %d\n",
1764                    ignore_pid);
1765         return true;
1766 }
1767
1768 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1769                                 void *priv __maybe_unused)
1770 {
1771         return fprintf(fp, "  %-32s %s\n", name, val);
1772 }
1773
1774 static void display_attr(struct perf_event_attr *attr)
1775 {
1776         if (verbose >= 2 || debug_peo_args) {
1777                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1778                 fprintf(stderr, "perf_event_attr:\n");
1779                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1780                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1781         }
1782 }
1783
1784 bool evsel__precise_ip_fallback(struct evsel *evsel)
1785 {
1786         /* Do not try less precise if not requested. */
1787         if (!evsel->precise_max)
1788                 return false;
1789
1790         /*
1791          * We tried all the precise_ip values, and it's
1792          * still failing, so leave it to standard fallback.
1793          */
1794         if (!evsel->core.attr.precise_ip) {
1795                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1796                 return false;
1797         }
1798
1799         if (!evsel->precise_ip_original)
1800                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1801
1802         evsel->core.attr.precise_ip--;
1803         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1804         display_attr(&evsel->core.attr);
1805         return true;
1806 }
1807
1808 static struct perf_cpu_map *empty_cpu_map;
1809 static struct perf_thread_map *empty_thread_map;
1810
1811 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1812                 struct perf_thread_map *threads)
1813 {
1814         int nthreads = perf_thread_map__nr(threads);
1815
1816         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1817             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1818                 return -EINVAL;
1819
1820         if (cpus == NULL) {
1821                 if (empty_cpu_map == NULL) {
1822                         empty_cpu_map = perf_cpu_map__dummy_new();
1823                         if (empty_cpu_map == NULL)
1824                                 return -ENOMEM;
1825                 }
1826
1827                 cpus = empty_cpu_map;
1828         }
1829
1830         if (threads == NULL) {
1831                 if (empty_thread_map == NULL) {
1832                         empty_thread_map = thread_map__new_by_tid(-1);
1833                         if (empty_thread_map == NULL)
1834                                 return -ENOMEM;
1835                 }
1836
1837                 threads = empty_thread_map;
1838         }
1839
1840         if (evsel->core.fd == NULL &&
1841             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1842                 return -ENOMEM;
1843
1844         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1845         if (evsel->cgrp)
1846                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1847
1848         return 0;
1849 }
1850
1851 static void evsel__disable_missing_features(struct evsel *evsel)
1852 {
1853         if (perf_missing_features.read_lost)
1854                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1855         if (perf_missing_features.weight_struct) {
1856                 evsel__set_sample_bit(evsel, WEIGHT);
1857                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1858         }
1859         if (perf_missing_features.clockid_wrong)
1860                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1861         if (perf_missing_features.clockid) {
1862                 evsel->core.attr.use_clockid = 0;
1863                 evsel->core.attr.clockid = 0;
1864         }
1865         if (perf_missing_features.cloexec)
1866                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1867         if (perf_missing_features.mmap2)
1868                 evsel->core.attr.mmap2 = 0;
1869         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1870                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1871         if (perf_missing_features.lbr_flags)
1872                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1873                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1874         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1875                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1876         if (perf_missing_features.ksymbol)
1877                 evsel->core.attr.ksymbol = 0;
1878         if (perf_missing_features.bpf)
1879                 evsel->core.attr.bpf_event = 0;
1880         if (perf_missing_features.branch_hw_idx)
1881                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1882         if (perf_missing_features.sample_id_all)
1883                 evsel->core.attr.sample_id_all = 0;
1884 }
1885
1886 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1887                         struct perf_thread_map *threads)
1888 {
1889         int err;
1890
1891         err = __evsel__prepare_open(evsel, cpus, threads);
1892         if (err)
1893                 return err;
1894
1895         evsel__disable_missing_features(evsel);
1896
1897         return err;
1898 }
1899
1900 bool evsel__detect_missing_features(struct evsel *evsel)
1901 {
1902         /*
1903          * Must probe features in the order they were added to the
1904          * perf_event_attr interface.
1905          */
1906         if (!perf_missing_features.read_lost &&
1907             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1908                 perf_missing_features.read_lost = true;
1909                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1910                 return true;
1911         } else if (!perf_missing_features.weight_struct &&
1912             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1913                 perf_missing_features.weight_struct = true;
1914                 pr_debug2("switching off weight struct support\n");
1915                 return true;
1916         } else if (!perf_missing_features.code_page_size &&
1917             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1918                 perf_missing_features.code_page_size = true;
1919                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1920                 return false;
1921         } else if (!perf_missing_features.data_page_size &&
1922             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1923                 perf_missing_features.data_page_size = true;
1924                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1925                 return false;
1926         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1927                 perf_missing_features.cgroup = true;
1928                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1929                 return false;
1930         } else if (!perf_missing_features.branch_hw_idx &&
1931             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1932                 perf_missing_features.branch_hw_idx = true;
1933                 pr_debug2("switching off branch HW index support\n");
1934                 return true;
1935         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1936                 perf_missing_features.aux_output = true;
1937                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1938                 return false;
1939         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1940                 perf_missing_features.bpf = true;
1941                 pr_debug2_peo("switching off bpf_event\n");
1942                 return true;
1943         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1944                 perf_missing_features.ksymbol = true;
1945                 pr_debug2_peo("switching off ksymbol\n");
1946                 return true;
1947         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1948                 perf_missing_features.write_backward = true;
1949                 pr_debug2_peo("switching off write_backward\n");
1950                 return false;
1951         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1952                 perf_missing_features.clockid_wrong = true;
1953                 pr_debug2_peo("switching off clockid\n");
1954                 return true;
1955         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1956                 perf_missing_features.clockid = true;
1957                 pr_debug2_peo("switching off use_clockid\n");
1958                 return true;
1959         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1960                 perf_missing_features.cloexec = true;
1961                 pr_debug2_peo("switching off cloexec flag\n");
1962                 return true;
1963         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1964                 perf_missing_features.mmap2 = true;
1965                 pr_debug2_peo("switching off mmap2\n");
1966                 return true;
1967         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1968                 if (evsel->pmu == NULL)
1969                         evsel->pmu = evsel__find_pmu(evsel);
1970
1971                 if (evsel->pmu)
1972                         evsel->pmu->missing_features.exclude_guest = true;
1973                 else {
1974                         /* we cannot find PMU, disable attrs now */
1975                         evsel->core.attr.exclude_host = false;
1976                         evsel->core.attr.exclude_guest = false;
1977                 }
1978
1979                 if (evsel->exclude_GH) {
1980                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
1981                         return false;
1982                 }
1983                 if (!perf_missing_features.exclude_guest) {
1984                         perf_missing_features.exclude_guest = true;
1985                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
1986                 }
1987                 return true;
1988         } else if (!perf_missing_features.sample_id_all) {
1989                 perf_missing_features.sample_id_all = true;
1990                 pr_debug2_peo("switching off sample_id_all\n");
1991                 return true;
1992         } else if (!perf_missing_features.lbr_flags &&
1993                         (evsel->core.attr.branch_sample_type &
1994                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1995                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1996                 perf_missing_features.lbr_flags = true;
1997                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
1998                 return true;
1999         } else if (!perf_missing_features.group_read &&
2000                     evsel->core.attr.inherit &&
2001                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2002                    evsel__is_group_leader(evsel)) {
2003                 perf_missing_features.group_read = true;
2004                 pr_debug2_peo("switching off group read\n");
2005                 return true;
2006         } else {
2007                 return false;
2008         }
2009 }
2010
2011 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2012 {
2013         int old_errno;
2014         struct rlimit l;
2015
2016         if (*set_rlimit < INCREASED_MAX) {
2017                 old_errno = errno;
2018
2019                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2020                         if (*set_rlimit == NO_CHANGE) {
2021                                 l.rlim_cur = l.rlim_max;
2022                         } else {
2023                                 l.rlim_cur = l.rlim_max + 1000;
2024                                 l.rlim_max = l.rlim_cur;
2025                         }
2026                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2027                                 (*set_rlimit) += 1;
2028                                 errno = old_errno;
2029                                 return true;
2030                         }
2031                 }
2032                 errno = old_errno;
2033         }
2034
2035         return false;
2036 }
2037
2038 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2039                 struct perf_thread_map *threads,
2040                 int start_cpu_map_idx, int end_cpu_map_idx)
2041 {
2042         int idx, thread, nthreads;
2043         int pid = -1, err, old_errno;
2044         enum rlimit_action set_rlimit = NO_CHANGE;
2045
2046         err = __evsel__prepare_open(evsel, cpus, threads);
2047         if (err)
2048                 return err;
2049
2050         if (cpus == NULL)
2051                 cpus = empty_cpu_map;
2052
2053         if (threads == NULL)
2054                 threads = empty_thread_map;
2055
2056         nthreads = perf_thread_map__nr(threads);
2057
2058         if (evsel->cgrp)
2059                 pid = evsel->cgrp->fd;
2060
2061 fallback_missing_features:
2062         evsel__disable_missing_features(evsel);
2063
2064         display_attr(&evsel->core.attr);
2065
2066         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2067
2068                 for (thread = 0; thread < nthreads; thread++) {
2069                         int fd, group_fd;
2070 retry_open:
2071                         if (thread >= nthreads)
2072                                 break;
2073
2074                         if (!evsel->cgrp && !evsel->core.system_wide)
2075                                 pid = perf_thread_map__pid(threads, thread);
2076
2077                         group_fd = get_group_fd(evsel, idx, thread);
2078
2079                         test_attr__ready();
2080
2081                         /* Debug message used by test scripts */
2082                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2083                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2084
2085                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2086                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2087                                                 group_fd, evsel->open_flags);
2088
2089                         FD(evsel, idx, thread) = fd;
2090
2091                         if (fd < 0) {
2092                                 err = -errno;
2093
2094                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2095                                           err);
2096                                 goto try_fallback;
2097                         }
2098
2099                         bpf_counter__install_pe(evsel, idx, fd);
2100
2101                         if (unlikely(test_attr__enabled)) {
2102                                 test_attr__open(&evsel->core.attr, pid,
2103                                                 perf_cpu_map__cpu(cpus, idx),
2104                                                 fd, group_fd, evsel->open_flags);
2105                         }
2106
2107                         /* Debug message used by test scripts */
2108                         pr_debug2_peo(" = %d\n", fd);
2109
2110                         if (evsel->bpf_fd >= 0) {
2111                                 int evt_fd = fd;
2112                                 int bpf_fd = evsel->bpf_fd;
2113
2114                                 err = ioctl(evt_fd,
2115                                             PERF_EVENT_IOC_SET_BPF,
2116                                             bpf_fd);
2117                                 if (err && errno != EEXIST) {
2118                                         pr_err("failed to attach bpf fd %d: %s\n",
2119                                                bpf_fd, strerror(errno));
2120                                         err = -EINVAL;
2121                                         goto out_close;
2122                                 }
2123                         }
2124
2125                         set_rlimit = NO_CHANGE;
2126
2127                         /*
2128                          * If we succeeded but had to kill clockid, fail and
2129                          * have evsel__open_strerror() print us a nice error.
2130                          */
2131                         if (perf_missing_features.clockid ||
2132                             perf_missing_features.clockid_wrong) {
2133                                 err = -EINVAL;
2134                                 goto out_close;
2135                         }
2136                 }
2137         }
2138
2139         return 0;
2140
2141 try_fallback:
2142         if (evsel__precise_ip_fallback(evsel))
2143                 goto retry_open;
2144
2145         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2146                                          idx, threads, thread, err)) {
2147                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2148                 nthreads--;
2149
2150                 /* ... and pretend like nothing have happened. */
2151                 err = 0;
2152                 goto retry_open;
2153         }
2154         /*
2155          * perf stat needs between 5 and 22 fds per CPU. When we run out
2156          * of them try to increase the limits.
2157          */
2158         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2159                 goto retry_open;
2160
2161         if (err != -EINVAL || idx > 0 || thread > 0)
2162                 goto out_close;
2163
2164         if (evsel__detect_missing_features(evsel))
2165                 goto fallback_missing_features;
2166 out_close:
2167         if (err)
2168                 threads->err_thread = thread;
2169
2170         old_errno = errno;
2171         do {
2172                 while (--thread >= 0) {
2173                         if (FD(evsel, idx, thread) >= 0)
2174                                 close(FD(evsel, idx, thread));
2175                         FD(evsel, idx, thread) = -1;
2176                 }
2177                 thread = nthreads;
2178         } while (--idx >= 0);
2179         errno = old_errno;
2180         return err;
2181 }
2182
2183 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2184                 struct perf_thread_map *threads)
2185 {
2186         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2187 }
2188
2189 void evsel__close(struct evsel *evsel)
2190 {
2191         perf_evsel__close(&evsel->core);
2192         perf_evsel__free_id(&evsel->core);
2193 }
2194
2195 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2196 {
2197         if (cpu_map_idx == -1)
2198                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2199
2200         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2201 }
2202
2203 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2204 {
2205         return evsel__open(evsel, NULL, threads);
2206 }
2207
2208 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2209                                        const union perf_event *event,
2210                                        struct perf_sample *sample)
2211 {
2212         u64 type = evsel->core.attr.sample_type;
2213         const __u64 *array = event->sample.array;
2214         bool swapped = evsel->needs_swap;
2215         union u64_swap u;
2216
2217         array += ((event->header.size -
2218                    sizeof(event->header)) / sizeof(u64)) - 1;
2219
2220         if (type & PERF_SAMPLE_IDENTIFIER) {
2221                 sample->id = *array;
2222                 array--;
2223         }
2224
2225         if (type & PERF_SAMPLE_CPU) {
2226                 u.val64 = *array;
2227                 if (swapped) {
2228                         /* undo swap of u64, then swap on individual u32s */
2229                         u.val64 = bswap_64(u.val64);
2230                         u.val32[0] = bswap_32(u.val32[0]);
2231                 }
2232
2233                 sample->cpu = u.val32[0];
2234                 array--;
2235         }
2236
2237         if (type & PERF_SAMPLE_STREAM_ID) {
2238                 sample->stream_id = *array;
2239                 array--;
2240         }
2241
2242         if (type & PERF_SAMPLE_ID) {
2243                 sample->id = *array;
2244                 array--;
2245         }
2246
2247         if (type & PERF_SAMPLE_TIME) {
2248                 sample->time = *array;
2249                 array--;
2250         }
2251
2252         if (type & PERF_SAMPLE_TID) {
2253                 u.val64 = *array;
2254                 if (swapped) {
2255                         /* undo swap of u64, then swap on individual u32s */
2256                         u.val64 = bswap_64(u.val64);
2257                         u.val32[0] = bswap_32(u.val32[0]);
2258                         u.val32[1] = bswap_32(u.val32[1]);
2259                 }
2260
2261                 sample->pid = u.val32[0];
2262                 sample->tid = u.val32[1];
2263                 array--;
2264         }
2265
2266         return 0;
2267 }
2268
2269 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2270                             u64 size)
2271 {
2272         return size > max_size || offset + size > endp;
2273 }
2274
2275 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2276         do {                                                            \
2277                 if (overflow(endp, (max_size), (offset), (size)))       \
2278                         return -EFAULT;                                 \
2279         } while (0)
2280
2281 #define OVERFLOW_CHECK_u64(offset) \
2282         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2283
2284 static int
2285 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2286 {
2287         /*
2288          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2289          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2290          * check the format does not go past the end of the event.
2291          */
2292         if (sample_size + sizeof(event->header) > event->header.size)
2293                 return -EFAULT;
2294
2295         return 0;
2296 }
2297
2298 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2299                                           const __u64 *array,
2300                                           u64 type __maybe_unused)
2301 {
2302         data->weight = *array;
2303 }
2304
2305 u64 evsel__bitfield_swap_branch_flags(u64 value)
2306 {
2307         u64 new_val = 0;
2308
2309         /*
2310          * branch_flags
2311          * union {
2312          *      u64 values;
2313          *      struct {
2314          *              mispred:1       //target mispredicted
2315          *              predicted:1     //target predicted
2316          *              in_tx:1         //in transaction
2317          *              abort:1         //transaction abort
2318          *              cycles:16       //cycle count to last branch
2319          *              type:4          //branch type
2320          *              spec:2          //branch speculation info
2321          *              new_type:4      //additional branch type
2322          *              priv:3          //privilege level
2323          *              reserved:31
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         if (host_is_bigendian()) {
2333                 new_val = bitfield_swap(value, 0, 1);
2334                 new_val |= bitfield_swap(value, 1, 1);
2335                 new_val |= bitfield_swap(value, 2, 1);
2336                 new_val |= bitfield_swap(value, 3, 1);
2337                 new_val |= bitfield_swap(value, 4, 16);
2338                 new_val |= bitfield_swap(value, 20, 4);
2339                 new_val |= bitfield_swap(value, 24, 2);
2340                 new_val |= bitfield_swap(value, 26, 4);
2341                 new_val |= bitfield_swap(value, 30, 3);
2342                 new_val |= bitfield_swap(value, 33, 31);
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, 38, 2);
2351                 new_val |= bitfield_swap(value, 34, 4);
2352                 new_val |= bitfield_swap(value, 31, 3);
2353                 new_val |= bitfield_swap(value, 0, 31);
2354         }
2355
2356         return new_val;
2357 }
2358
2359 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2360                         struct perf_sample *data)
2361 {
2362         u64 type = evsel->core.attr.sample_type;
2363         bool swapped = evsel->needs_swap;
2364         const __u64 *array;
2365         u16 max_size = event->header.size;
2366         const void *endp = (void *)event + max_size;
2367         u64 sz;
2368
2369         /*
2370          * used for cross-endian analysis. See git commit 65014ab3
2371          * for why this goofiness is needed.
2372          */
2373         union u64_swap u;
2374
2375         memset(data, 0, sizeof(*data));
2376         data->cpu = data->pid = data->tid = -1;
2377         data->stream_id = data->id = data->time = -1ULL;
2378         data->period = evsel->core.attr.sample_period;
2379         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2380         data->misc    = event->header.misc;
2381         data->id = -1ULL;
2382         data->data_src = PERF_MEM_DATA_SRC_NONE;
2383         data->vcpu = -1;
2384
2385         if (event->header.type != PERF_RECORD_SAMPLE) {
2386                 if (!evsel->core.attr.sample_id_all)
2387                         return 0;
2388                 return perf_evsel__parse_id_sample(evsel, event, data);
2389         }
2390
2391         array = event->sample.array;
2392
2393         if (perf_event__check_size(event, evsel->sample_size))
2394                 return -EFAULT;
2395
2396         if (type & PERF_SAMPLE_IDENTIFIER) {
2397                 data->id = *array;
2398                 array++;
2399         }
2400
2401         if (type & PERF_SAMPLE_IP) {
2402                 data->ip = *array;
2403                 array++;
2404         }
2405
2406         if (type & PERF_SAMPLE_TID) {
2407                 u.val64 = *array;
2408                 if (swapped) {
2409                         /* undo swap of u64, then swap on individual u32s */
2410                         u.val64 = bswap_64(u.val64);
2411                         u.val32[0] = bswap_32(u.val32[0]);
2412                         u.val32[1] = bswap_32(u.val32[1]);
2413                 }
2414
2415                 data->pid = u.val32[0];
2416                 data->tid = u.val32[1];
2417                 array++;
2418         }
2419
2420         if (type & PERF_SAMPLE_TIME) {
2421                 data->time = *array;
2422                 array++;
2423         }
2424
2425         if (type & PERF_SAMPLE_ADDR) {
2426                 data->addr = *array;
2427                 array++;
2428         }
2429
2430         if (type & PERF_SAMPLE_ID) {
2431                 data->id = *array;
2432                 array++;
2433         }
2434
2435         if (type & PERF_SAMPLE_STREAM_ID) {
2436                 data->stream_id = *array;
2437                 array++;
2438         }
2439
2440         if (type & PERF_SAMPLE_CPU) {
2441
2442                 u.val64 = *array;
2443                 if (swapped) {
2444                         /* undo swap of u64, then swap on individual u32s */
2445                         u.val64 = bswap_64(u.val64);
2446                         u.val32[0] = bswap_32(u.val32[0]);
2447                 }
2448
2449                 data->cpu = u.val32[0];
2450                 array++;
2451         }
2452
2453         if (type & PERF_SAMPLE_PERIOD) {
2454                 data->period = *array;
2455                 array++;
2456         }
2457
2458         if (type & PERF_SAMPLE_READ) {
2459                 u64 read_format = evsel->core.attr.read_format;
2460
2461                 OVERFLOW_CHECK_u64(array);
2462                 if (read_format & PERF_FORMAT_GROUP)
2463                         data->read.group.nr = *array;
2464                 else
2465                         data->read.one.value = *array;
2466
2467                 array++;
2468
2469                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2470                         OVERFLOW_CHECK_u64(array);
2471                         data->read.time_enabled = *array;
2472                         array++;
2473                 }
2474
2475                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2476                         OVERFLOW_CHECK_u64(array);
2477                         data->read.time_running = *array;
2478                         array++;
2479                 }
2480
2481                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2482                 if (read_format & PERF_FORMAT_GROUP) {
2483                         const u64 max_group_nr = UINT64_MAX /
2484                                         sizeof(struct sample_read_value);
2485
2486                         if (data->read.group.nr > max_group_nr)
2487                                 return -EFAULT;
2488
2489                         sz = data->read.group.nr * sample_read_value_size(read_format);
2490                         OVERFLOW_CHECK(array, sz, max_size);
2491                         data->read.group.values =
2492                                         (struct sample_read_value *)array;
2493                         array = (void *)array + sz;
2494                 } else {
2495                         OVERFLOW_CHECK_u64(array);
2496                         data->read.one.id = *array;
2497                         array++;
2498
2499                         if (read_format & PERF_FORMAT_LOST) {
2500                                 OVERFLOW_CHECK_u64(array);
2501                                 data->read.one.lost = *array;
2502                                 array++;
2503                         }
2504                 }
2505         }
2506
2507         if (type & PERF_SAMPLE_CALLCHAIN) {
2508                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2509
2510                 OVERFLOW_CHECK_u64(array);
2511                 data->callchain = (struct ip_callchain *)array++;
2512                 if (data->callchain->nr > max_callchain_nr)
2513                         return -EFAULT;
2514                 sz = data->callchain->nr * sizeof(u64);
2515                 OVERFLOW_CHECK(array, sz, max_size);
2516                 array = (void *)array + sz;
2517         }
2518
2519         if (type & PERF_SAMPLE_RAW) {
2520                 OVERFLOW_CHECK_u64(array);
2521                 u.val64 = *array;
2522
2523                 /*
2524                  * Undo swap of u64, then swap on individual u32s,
2525                  * get the size of the raw area and undo all of the
2526                  * swap. The pevent interface handles endianness by
2527                  * itself.
2528                  */
2529                 if (swapped) {
2530                         u.val64 = bswap_64(u.val64);
2531                         u.val32[0] = bswap_32(u.val32[0]);
2532                         u.val32[1] = bswap_32(u.val32[1]);
2533                 }
2534                 data->raw_size = u.val32[0];
2535
2536                 /*
2537                  * The raw data is aligned on 64bits including the
2538                  * u32 size, so it's safe to use mem_bswap_64.
2539                  */
2540                 if (swapped)
2541                         mem_bswap_64((void *) array, data->raw_size);
2542
2543                 array = (void *)array + sizeof(u32);
2544
2545                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2546                 data->raw_data = (void *)array;
2547                 array = (void *)array + data->raw_size;
2548         }
2549
2550         if (type & PERF_SAMPLE_BRANCH_STACK) {
2551                 const u64 max_branch_nr = UINT64_MAX /
2552                                           sizeof(struct branch_entry);
2553                 struct branch_entry *e;
2554                 unsigned int i;
2555
2556                 OVERFLOW_CHECK_u64(array);
2557                 data->branch_stack = (struct branch_stack *)array++;
2558
2559                 if (data->branch_stack->nr > max_branch_nr)
2560                         return -EFAULT;
2561
2562                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2563                 if (evsel__has_branch_hw_idx(evsel)) {
2564                         sz += sizeof(u64);
2565                         e = &data->branch_stack->entries[0];
2566                 } else {
2567                         data->no_hw_idx = true;
2568                         /*
2569                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2570                          * only nr and entries[] will be output by kernel.
2571                          */
2572                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2573                 }
2574
2575                 if (swapped) {
2576                         /*
2577                          * struct branch_flag does not have endian
2578                          * specific bit field definition. And bswap
2579                          * will not resolve the issue, since these
2580                          * are bit fields.
2581                          *
2582                          * evsel__bitfield_swap_branch_flags() uses a
2583                          * bitfield_swap macro to swap the bit position
2584                          * based on the host endians.
2585                          */
2586                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2587                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2588                 }
2589
2590                 OVERFLOW_CHECK(array, sz, max_size);
2591                 array = (void *)array + sz;
2592         }
2593
2594         if (type & PERF_SAMPLE_REGS_USER) {
2595                 OVERFLOW_CHECK_u64(array);
2596                 data->user_regs.abi = *array;
2597                 array++;
2598
2599                 if (data->user_regs.abi) {
2600                         u64 mask = evsel->core.attr.sample_regs_user;
2601
2602                         sz = hweight64(mask) * sizeof(u64);
2603                         OVERFLOW_CHECK(array, sz, max_size);
2604                         data->user_regs.mask = mask;
2605                         data->user_regs.regs = (u64 *)array;
2606                         array = (void *)array + sz;
2607                 }
2608         }
2609
2610         if (type & PERF_SAMPLE_STACK_USER) {
2611                 OVERFLOW_CHECK_u64(array);
2612                 sz = *array++;
2613
2614                 data->user_stack.offset = ((char *)(array - 1)
2615                                           - (char *) event);
2616
2617                 if (!sz) {
2618                         data->user_stack.size = 0;
2619                 } else {
2620                         OVERFLOW_CHECK(array, sz, max_size);
2621                         data->user_stack.data = (char *)array;
2622                         array = (void *)array + sz;
2623                         OVERFLOW_CHECK_u64(array);
2624                         data->user_stack.size = *array++;
2625                         if (WARN_ONCE(data->user_stack.size > sz,
2626                                       "user stack dump failure\n"))
2627                                 return -EFAULT;
2628                 }
2629         }
2630
2631         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2632                 OVERFLOW_CHECK_u64(array);
2633                 arch_perf_parse_sample_weight(data, array, type);
2634                 array++;
2635         }
2636
2637         if (type & PERF_SAMPLE_DATA_SRC) {
2638                 OVERFLOW_CHECK_u64(array);
2639                 data->data_src = *array;
2640                 array++;
2641         }
2642
2643         if (type & PERF_SAMPLE_TRANSACTION) {
2644                 OVERFLOW_CHECK_u64(array);
2645                 data->transaction = *array;
2646                 array++;
2647         }
2648
2649         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2650         if (type & PERF_SAMPLE_REGS_INTR) {
2651                 OVERFLOW_CHECK_u64(array);
2652                 data->intr_regs.abi = *array;
2653                 array++;
2654
2655                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2656                         u64 mask = evsel->core.attr.sample_regs_intr;
2657
2658                         sz = hweight64(mask) * sizeof(u64);
2659                         OVERFLOW_CHECK(array, sz, max_size);
2660                         data->intr_regs.mask = mask;
2661                         data->intr_regs.regs = (u64 *)array;
2662                         array = (void *)array + sz;
2663                 }
2664         }
2665
2666         data->phys_addr = 0;
2667         if (type & PERF_SAMPLE_PHYS_ADDR) {
2668                 data->phys_addr = *array;
2669                 array++;
2670         }
2671
2672         data->cgroup = 0;
2673         if (type & PERF_SAMPLE_CGROUP) {
2674                 data->cgroup = *array;
2675                 array++;
2676         }
2677
2678         data->data_page_size = 0;
2679         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2680                 data->data_page_size = *array;
2681                 array++;
2682         }
2683
2684         data->code_page_size = 0;
2685         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2686                 data->code_page_size = *array;
2687                 array++;
2688         }
2689
2690         if (type & PERF_SAMPLE_AUX) {
2691                 OVERFLOW_CHECK_u64(array);
2692                 sz = *array++;
2693
2694                 OVERFLOW_CHECK(array, sz, max_size);
2695                 /* Undo swap of data */
2696                 if (swapped)
2697                         mem_bswap_64((char *)array, sz);
2698                 data->aux_sample.size = sz;
2699                 data->aux_sample.data = (char *)array;
2700                 array = (void *)array + sz;
2701         }
2702
2703         return 0;
2704 }
2705
2706 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2707                                   u64 *timestamp)
2708 {
2709         u64 type = evsel->core.attr.sample_type;
2710         const __u64 *array;
2711
2712         if (!(type & PERF_SAMPLE_TIME))
2713                 return -1;
2714
2715         if (event->header.type != PERF_RECORD_SAMPLE) {
2716                 struct perf_sample data = {
2717                         .time = -1ULL,
2718                 };
2719
2720                 if (!evsel->core.attr.sample_id_all)
2721                         return -1;
2722                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2723                         return -1;
2724
2725                 *timestamp = data.time;
2726                 return 0;
2727         }
2728
2729         array = event->sample.array;
2730
2731         if (perf_event__check_size(event, evsel->sample_size))
2732                 return -EFAULT;
2733
2734         if (type & PERF_SAMPLE_IDENTIFIER)
2735                 array++;
2736
2737         if (type & PERF_SAMPLE_IP)
2738                 array++;
2739
2740         if (type & PERF_SAMPLE_TID)
2741                 array++;
2742
2743         if (type & PERF_SAMPLE_TIME)
2744                 *timestamp = *array;
2745
2746         return 0;
2747 }
2748
2749 u16 evsel__id_hdr_size(struct evsel *evsel)
2750 {
2751         u64 sample_type = evsel->core.attr.sample_type;
2752         u16 size = 0;
2753
2754         if (sample_type & PERF_SAMPLE_TID)
2755                 size += sizeof(u64);
2756
2757         if (sample_type & PERF_SAMPLE_TIME)
2758                 size += sizeof(u64);
2759
2760         if (sample_type & PERF_SAMPLE_ID)
2761                 size += sizeof(u64);
2762
2763         if (sample_type & PERF_SAMPLE_STREAM_ID)
2764                 size += sizeof(u64);
2765
2766         if (sample_type & PERF_SAMPLE_CPU)
2767                 size += sizeof(u64);
2768
2769         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2770                 size += sizeof(u64);
2771
2772         return size;
2773 }
2774
2775 #ifdef HAVE_LIBTRACEEVENT
2776 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2777 {
2778         return tep_find_field(evsel->tp_format, name);
2779 }
2780
2781 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2782 {
2783         struct tep_format_field *field = evsel__field(evsel, name);
2784         int offset;
2785
2786         if (!field)
2787                 return NULL;
2788
2789         offset = field->offset;
2790
2791         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2792                 offset = *(int *)(sample->raw_data + field->offset);
2793                 offset &= 0xffff;
2794                 if (tep_field_is_relative(field->flags))
2795                         offset += field->offset + field->size;
2796         }
2797
2798         return sample->raw_data + offset;
2799 }
2800
2801 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2802                          bool needs_swap)
2803 {
2804         u64 value;
2805         void *ptr = sample->raw_data + field->offset;
2806
2807         switch (field->size) {
2808         case 1:
2809                 return *(u8 *)ptr;
2810         case 2:
2811                 value = *(u16 *)ptr;
2812                 break;
2813         case 4:
2814                 value = *(u32 *)ptr;
2815                 break;
2816         case 8:
2817                 memcpy(&value, ptr, sizeof(u64));
2818                 break;
2819         default:
2820                 return 0;
2821         }
2822
2823         if (!needs_swap)
2824                 return value;
2825
2826         switch (field->size) {
2827         case 2:
2828                 return bswap_16(value);
2829         case 4:
2830                 return bswap_32(value);
2831         case 8:
2832                 return bswap_64(value);
2833         default:
2834                 return 0;
2835         }
2836
2837         return 0;
2838 }
2839
2840 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2841 {
2842         struct tep_format_field *field = evsel__field(evsel, name);
2843
2844         if (!field)
2845                 return 0;
2846
2847         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2848 }
2849 #endif
2850
2851 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2852 {
2853         int paranoid;
2854
2855         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2856             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2857             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2858                 /*
2859                  * If it's cycles then fall back to hrtimer based
2860                  * cpu-clock-tick sw counter, which is always available even if
2861                  * no PMU support.
2862                  *
2863                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2864                  * b0a873e).
2865                  */
2866                 scnprintf(msg, msgsize, "%s",
2867 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2868
2869                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2870                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2871
2872                 zfree(&evsel->name);
2873                 return true;
2874         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2875                    (paranoid = perf_event_paranoid()) > 1) {
2876                 const char *name = evsel__name(evsel);
2877                 char *new_name;
2878                 const char *sep = ":";
2879
2880                 /* If event has exclude user then don't exclude kernel. */
2881                 if (evsel->core.attr.exclude_user)
2882                         return false;
2883
2884                 /* Is there already the separator in the name. */
2885                 if (strchr(name, '/') ||
2886                     (strchr(name, ':') && !evsel->is_libpfm_event))
2887                         sep = "";
2888
2889                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2890                         return false;
2891
2892                 if (evsel->name)
2893                         free(evsel->name);
2894                 evsel->name = new_name;
2895                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2896                           "to fall back to excluding kernel and hypervisor "
2897                           " samples", paranoid);
2898                 evsel->core.attr.exclude_kernel = 1;
2899                 evsel->core.attr.exclude_hv     = 1;
2900
2901                 return true;
2902         }
2903
2904         return false;
2905 }
2906
2907 static bool find_process(const char *name)
2908 {
2909         size_t len = strlen(name);
2910         DIR *dir;
2911         struct dirent *d;
2912         int ret = -1;
2913
2914         dir = opendir(procfs__mountpoint());
2915         if (!dir)
2916                 return false;
2917
2918         /* Walk through the directory. */
2919         while (ret && (d = readdir(dir)) != NULL) {
2920                 char path[PATH_MAX];
2921                 char *data;
2922                 size_t size;
2923
2924                 if ((d->d_type != DT_DIR) ||
2925                      !strcmp(".", d->d_name) ||
2926                      !strcmp("..", d->d_name))
2927                         continue;
2928
2929                 scnprintf(path, sizeof(path), "%s/%s/comm",
2930                           procfs__mountpoint(), d->d_name);
2931
2932                 if (filename__read_str(path, &data, &size))
2933                         continue;
2934
2935                 ret = strncmp(name, data, len);
2936                 free(data);
2937         }
2938
2939         closedir(dir);
2940         return ret ? false : true;
2941 }
2942
2943 static bool is_amd(const char *arch, const char *cpuid)
2944 {
2945         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2946 }
2947
2948 static bool is_amd_ibs(struct evsel *evsel)
2949 {
2950         return evsel->core.attr.precise_ip
2951             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2952 }
2953
2954 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2955                          int err, char *msg, size_t size)
2956 {
2957         struct perf_env *env = evsel__env(evsel);
2958         const char *arch = perf_env__arch(env);
2959         const char *cpuid = perf_env__cpuid(env);
2960         char sbuf[STRERR_BUFSIZE];
2961         int printed = 0, enforced = 0;
2962
2963         switch (err) {
2964         case EPERM:
2965         case EACCES:
2966                 printed += scnprintf(msg + printed, size - printed,
2967                         "Access to performance monitoring and observability operations is limited.\n");
2968
2969                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2970                         if (enforced) {
2971                                 printed += scnprintf(msg + printed, size - printed,
2972                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2973                                         "monitoring and observability operations. Inspect system audit records for\n"
2974                                         "more perf_event access control information and adjusting the policy.\n");
2975                         }
2976                 }
2977
2978                 if (err == EPERM)
2979                         printed += scnprintf(msg, size,
2980                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
2981
2982                 return scnprintf(msg + printed, size - printed,
2983                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
2984                  "access to performance monitoring and observability operations for processes\n"
2985                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
2986                  "More information can be found at 'Perf events and tool security' document:\n"
2987                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
2988                  "perf_event_paranoid setting is %d:\n"
2989                  "  -1: Allow use of (almost) all events by all users\n"
2990                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2991                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
2992                  ">= 1: Disallow CPU event access\n"
2993                  ">= 2: Disallow kernel profiling\n"
2994                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
2995                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
2996                  perf_event_paranoid());
2997         case ENOENT:
2998                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
2999         case EMFILE:
3000                 return scnprintf(msg, size, "%s",
3001                          "Too many events are opened.\n"
3002                          "Probably the maximum number of open file descriptors has been reached.\n"
3003                          "Hint: Try again after reducing the number of events.\n"
3004                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3005         case ENOMEM:
3006                 if (evsel__has_callchain(evsel) &&
3007                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3008                         return scnprintf(msg, size,
3009                                          "Not enough memory to setup event with callchain.\n"
3010                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3011                                          "Hint: Current value: %d", sysctl__max_stack());
3012                 break;
3013         case ENODEV:
3014                 if (target->cpu_list)
3015                         return scnprintf(msg, size, "%s",
3016          "No such device - did you specify an out-of-range profile CPU?");
3017                 break;
3018         case EOPNOTSUPP:
3019                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3020                         return scnprintf(msg, size,
3021         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3022                                          evsel__name(evsel));
3023                 if (evsel->core.attr.aux_output)
3024                         return scnprintf(msg, size,
3025         "%s: PMU Hardware doesn't support 'aux_output' feature",
3026                                          evsel__name(evsel));
3027                 if (evsel->core.attr.sample_period != 0)
3028                         return scnprintf(msg, size,
3029         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3030                                          evsel__name(evsel));
3031                 if (evsel->core.attr.precise_ip)
3032                         return scnprintf(msg, size, "%s",
3033         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3034 #if defined(__i386__) || defined(__x86_64__)
3035                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3036                         return scnprintf(msg, size, "%s",
3037         "No hardware sampling interrupt available.\n");
3038 #endif
3039                 break;
3040         case EBUSY:
3041                 if (find_process("oprofiled"))
3042                         return scnprintf(msg, size,
3043         "The PMU counters are busy/taken by another profiler.\n"
3044         "We found oprofile daemon running, please stop it and try again.");
3045                 break;
3046         case EINVAL:
3047                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3048                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3049                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3050                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3051                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3052                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3053                 if (perf_missing_features.clockid)
3054                         return scnprintf(msg, size, "clockid feature not supported.");
3055                 if (perf_missing_features.clockid_wrong)
3056                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3057                 if (perf_missing_features.aux_output)
3058                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3059                 if (!target__has_cpu(target))
3060                         return scnprintf(msg, size,
3061         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3062                                         evsel__name(evsel));
3063                 if (is_amd(arch, cpuid)) {
3064                         if (is_amd_ibs(evsel)) {
3065                                 if (evsel->core.attr.exclude_kernel)
3066                                         return scnprintf(msg, size,
3067         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3068                                 if (!evsel->core.system_wide)
3069                                         return scnprintf(msg, size,
3070         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3071                         }
3072                 }
3073
3074                 break;
3075         case ENODATA:
3076                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3077                                  "Please add an auxiliary event in front of the load latency event.");
3078         default:
3079                 break;
3080         }
3081
3082         return scnprintf(msg, size,
3083         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3084         "/bin/dmesg | grep -i perf may provide additional information.\n",
3085                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3086 }
3087
3088 struct perf_env *evsel__env(struct evsel *evsel)
3089 {
3090         if (evsel && evsel->evlist && evsel->evlist->env)
3091                 return evsel->evlist->env;
3092         return &perf_env;
3093 }
3094
3095 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3096 {
3097         int cpu_map_idx, thread;
3098
3099         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3100                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3101                      thread++) {
3102                         int fd = FD(evsel, cpu_map_idx, thread);
3103
3104                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3105                                                    cpu_map_idx, thread, fd) < 0)
3106                                 return -1;
3107                 }
3108         }
3109
3110         return 0;
3111 }
3112
3113 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3114 {
3115         struct perf_cpu_map *cpus = evsel->core.cpus;
3116         struct perf_thread_map *threads = evsel->core.threads;
3117
3118         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3119                 return -ENOMEM;
3120
3121         return store_evsel_ids(evsel, evlist);
3122 }
3123
3124 void evsel__zero_per_pkg(struct evsel *evsel)
3125 {
3126         struct hashmap_entry *cur;
3127         size_t bkt;
3128
3129         if (evsel->per_pkg_mask) {
3130                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3131                         free((void *)cur->pkey);
3132
3133                 hashmap__clear(evsel->per_pkg_mask);
3134         }
3135 }
3136
3137 bool evsel__is_hybrid(const struct evsel *evsel)
3138 {
3139         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3140 }
3141
3142 struct evsel *evsel__leader(const struct evsel *evsel)
3143 {
3144         return container_of(evsel->core.leader, struct evsel, core);
3145 }
3146
3147 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3148 {
3149         return evsel->core.leader == &leader->core;
3150 }
3151
3152 bool evsel__is_leader(struct evsel *evsel)
3153 {
3154         return evsel__has_leader(evsel, evsel);
3155 }
3156
3157 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3158 {
3159         evsel->core.leader = &leader->core;
3160 }
3161
3162 int evsel__source_count(const struct evsel *evsel)
3163 {
3164         struct evsel *pos;
3165         int count = 0;
3166
3167         evlist__for_each_entry(evsel->evlist, pos) {
3168                 if (pos->metric_leader == evsel)
3169                         count++;
3170         }
3171         return count;
3172 }
3173
3174 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3175 {
3176         return false;
3177 }
3178
3179 /*
3180  * Remove an event from a given group (leader).
3181  * Some events, e.g., perf metrics Topdown events,
3182  * must always be grouped. Ignore the events.
3183  */
3184 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3185 {
3186         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3187                 evsel__set_leader(evsel, evsel);
3188                 evsel->core.nr_members = 0;
3189                 leader->core.nr_members--;
3190         }
3191 }