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->is_libpfm_event = orig->is_libpfm_event;
 462
 463         evsel->exclude_GH = orig->exclude_GH;
 464         evsel->sample_read = orig->sample_read;
 465         evsel->auto_merge_stats = orig->auto_merge_stats;
 466         evsel->collect_stat = orig->collect_stat;
 467         evsel->weak_group = orig->weak_group;
 468         evsel->use_config_name = orig->use_config_name;
 469         evsel->pmu = orig->pmu;
 470
 471         if (evsel__copy_config_terms(evsel, orig) < 0)
 472                 goto out_err;
 473
 474         return evsel;
 475
 476 out_err:
 477         evsel__delete(evsel);
 478         return NULL;
 479 }
 480
 481 /*
 482  * Returns pointer with encoded error via <linux/err.h> interface.
 483  */
 484 #ifdef HAVE_LIBTRACEEVENT
 485 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 486 {
 487         struct evsel *evsel = zalloc(perf_evsel__object.size);
 488         int err = -ENOMEM;
 489
 490         if (evsel == NULL) {
 491                 goto out_err;
 492         } else {
 493                 struct perf_event_attr attr = {
 494                         .type          = PERF_TYPE_TRACEPOINT,
 495                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 496                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 497                 };
 498
 499                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 500                         goto out_free;
 501
 502                 evsel->tp_format = trace_event__tp_format(sys, name);
 503                 if (IS_ERR(evsel->tp_format)) {
 504                         err = PTR_ERR(evsel->tp_format);
 505                         goto out_free;
 506                 }
 507
 508                 event_attr_init(&attr);
 509                 attr.config = evsel->tp_format->id;
 510                 attr.sample_period = 1;
 511                 evsel__init(evsel, &attr, idx);
 512         }
 513
 514         return evsel;
 515
 516 out_free:
 517         zfree(&evsel->name);
 518         free(evsel);
 519 out_err:
 520         return ERR_PTR(err);
 521 }
 522 #endif
 523
 524 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 525         "cycles",
 526         "instructions",
 527         "cache-references",
 528         "cache-misses",
 529         "branches",
 530         "branch-misses",
 531         "bus-cycles",
 532         "stalled-cycles-frontend",
 533         "stalled-cycles-backend",
 534         "ref-cycles",
 535 };
 536
 537 char *evsel__bpf_counter_events;
 538
 539 bool evsel__match_bpf_counter_events(const char *name)
 540 {
 541         int name_len;
 542         bool match;
 543         char *ptr;
 544
 545         if (!evsel__bpf_counter_events)
 546                 return false;
 547
 548         ptr = strstr(evsel__bpf_counter_events, name);
 549         name_len = strlen(name);
 550
 551         /* check name matches a full token in evsel__bpf_counter_events */
 552         match = (ptr != NULL) &&
 553                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 554                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 555
 556         return match;
 557 }
 558
 559 static const char *__evsel__hw_name(u64 config)
 560 {
 561         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 562                 return evsel__hw_names[config];
 563
 564         return "unknown-hardware";
 565 }
 566
 567 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 568 {
 569         int colon = 0, r = 0;
 570         struct perf_event_attr *attr = &evsel->core.attr;
 571         bool exclude_guest_default = false;
 572
 573 #define MOD_PRINT(context, mod) do {                                    \
 574                 if (!attr->exclude_##context) {                         \
 575                         if (!colon) colon = ++r;                        \
 576                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 577                 } } while(0)
 578
 579         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 580                 MOD_PRINT(kernel, 'k');
 581                 MOD_PRINT(user, 'u');
 582                 MOD_PRINT(hv, 'h');
 583                 exclude_guest_default = true;
 584         }
 585
 586         if (attr->precise_ip) {
 587                 if (!colon)
 588                         colon = ++r;
 589                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 590                 exclude_guest_default = true;
 591         }
 592
 593         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 594                 MOD_PRINT(host, 'H');
 595                 MOD_PRINT(guest, 'G');
 596         }
 597 #undef MOD_PRINT
 598         if (colon)
 599                 bf[colon - 1] = ':';
 600         return r;
 601 }
 602
 603 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 604 {
 605         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 606 }
 607
 608 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 609 {
 610         int r = arch_evsel__hw_name(evsel, bf, size);
 611         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 612 }
 613
 614 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 615         "cpu-clock",
 616         "task-clock",
 617         "page-faults",
 618         "context-switches",
 619         "cpu-migrations",
 620         "minor-faults",
 621         "major-faults",
 622         "alignment-faults",
 623         "emulation-faults",
 624         "dummy",
 625 };
 626
 627 static const char *__evsel__sw_name(u64 config)
 628 {
 629         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 630                 return evsel__sw_names[config];
 631         return "unknown-software";
 632 }
 633
 634 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 635 {
 636         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 637         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 638 }
 639
 640 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 641 {
 642         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 643 }
 644
 645 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 646 {
 647         int r;
 648
 649         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 650
 651         if (type & HW_BREAKPOINT_R)
 652                 r += scnprintf(bf + r, size - r, "r");
 653
 654         if (type & HW_BREAKPOINT_W)
 655                 r += scnprintf(bf + r, size - r, "w");
 656
 657         if (type & HW_BREAKPOINT_X)
 658                 r += scnprintf(bf + r, size - r, "x");
 659
 660         return r;
 661 }
 662
 663 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 664 {
 665         struct perf_event_attr *attr = &evsel->core.attr;
 666         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 667         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 668 }
 669
 670 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 671  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 672  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 673  { "LLC",       "L2",                                                   },
 674  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 675  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 676  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 677  { "node",                                                              },
 678 };
 679
 680 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 681  { "load",      "loads",        "read",                                 },
 682  { "store",     "stores",       "write",                                },
 683  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 684 };
 685
 686 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 687  { "refs",      "Reference",    "ops",          "access",               },
 688  { "misses",    "miss",                                                 },
 689 };
 690
 691 #define C(x)            PERF_COUNT_HW_CACHE_##x
 692 #define CACHE_READ      (1 << C(OP_READ))
 693 #define CACHE_WRITE     (1 << C(OP_WRITE))
 694 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 695 #define COP(x)          (1 << x)
 696
 697 /*
 698  * cache operation stat
 699  * L1I : Read and prefetch only
 700  * ITLB and BPU : Read-only
 701  */
 702 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 703  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 704  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 705  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 706  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 707  [C(ITLB)]      = (CACHE_READ),
 708  [C(BPU)]       = (CACHE_READ),
 709  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 710 };
 711
 712 bool evsel__is_cache_op_valid(u8 type, u8 op)
 713 {
 714         if (evsel__hw_cache_stat[type] & COP(op))
 715                 return true;    /* valid */
 716         else
 717                 return false;   /* invalid */
 718 }
 719
 720 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 721 {
 722         if (result) {
 723                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 724                                  evsel__hw_cache_op[op][0],
 725                                  evsel__hw_cache_result[result][0]);
 726         }
 727
 728         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 729                          evsel__hw_cache_op[op][1]);
 730 }
 731
 732 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 733 {
 734         u8 op, result, type = (config >>  0) & 0xff;
 735         const char *err = "unknown-ext-hardware-cache-type";
 736
 737         if (type >= PERF_COUNT_HW_CACHE_MAX)
 738                 goto out_err;
 739
 740         op = (config >>  8) & 0xff;
 741         err = "unknown-ext-hardware-cache-op";
 742         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 743                 goto out_err;
 744
 745         result = (config >> 16) & 0xff;
 746         err = "unknown-ext-hardware-cache-result";
 747         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 748                 goto out_err;
 749
 750         err = "invalid-cache";
 751         if (!evsel__is_cache_op_valid(type, op))
 752                 goto out_err;
 753
 754         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 755 out_err:
 756         return scnprintf(bf, size, "%s", err);
 757 }
 758
 759 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 760 {
 761         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 762         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 763 }
 764
 765 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 766 {
 767         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 768         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 769 }
 770
 771 const char *evsel__name(struct evsel *evsel)
 772 {
 773         char bf[128];
 774
 775         if (!evsel)
 776                 goto out_unknown;
 777
 778         if (evsel->name)
 779                 return evsel->name;
 780
 781         switch (evsel->core.attr.type) {
 782         case PERF_TYPE_RAW:
 783                 evsel__raw_name(evsel, bf, sizeof(bf));
 784                 break;
 785
 786         case PERF_TYPE_HARDWARE:
 787                 evsel__hw_name(evsel, bf, sizeof(bf));
 788                 break;
 789
 790         case PERF_TYPE_HW_CACHE:
 791                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 792                 break;
 793
 794         case PERF_TYPE_SOFTWARE:
 795                 if (evsel__is_tool(evsel))
 796                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 797                 else
 798                         evsel__sw_name(evsel, bf, sizeof(bf));
 799                 break;
 800
 801         case PERF_TYPE_TRACEPOINT:
 802                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 803                 break;
 804
 805         case PERF_TYPE_BREAKPOINT:
 806                 evsel__bp_name(evsel, bf, sizeof(bf));
 807                 break;
 808
 809         default:
 810                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 811                           evsel->core.attr.type);
 812                 break;
 813         }
 814
 815         evsel->name = strdup(bf);
 816
 817         if (evsel->name)
 818                 return evsel->name;
 819 out_unknown:
 820         return "unknown";
 821 }
 822
 823 const char *evsel__group_pmu_name(const struct evsel *evsel)
 824 {
 825         const struct evsel *leader;
 826
 827         /* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
 828         if (evsel->pmu_name)
 829                 return evsel->pmu_name;
 830         /*
 831          * Software events may be in a group with other uncore PMU events. Use
 832          * the pmu_name of the group leader to avoid breaking the software event
 833          * out of the group.
 834          *
 835          * Aux event leaders, like intel_pt, expect a group with events from
 836          * other PMUs, so substitute the AUX event's PMU in this case.
 837          */
 838         leader  = evsel__leader(evsel);
 839         if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
 840             leader->pmu_name) {
 841                 return leader->pmu_name;
 842         }
 843
 844         return "cpu";
 845 }
 846
 847 const char *evsel__metric_id(const struct evsel *evsel)
 848 {
 849         if (evsel->metric_id)
 850                 return evsel->metric_id;
 851
 852         if (evsel__is_tool(evsel))
 853                 return perf_tool_event__to_str(evsel->tool_event);
 854
 855         return "unknown";
 856 }
 857
 858 const char *evsel__group_name(struct evsel *evsel)
 859 {
 860         return evsel->group_name ?: "anon group";
 861 }
 862
 863 /*
 864  * Returns the group details for the specified leader,
 865  * with following rules.
 866  *
 867  *  For record -e '{cycles,instructions}'
 868  *    'anon group { cycles:u, instructions:u }'
 869  *
 870  *  For record -e 'cycles,instructions' and report --group
 871  *    'cycles:u, instructions:u'
 872  */
 873 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 874 {
 875         int ret = 0;
 876         struct evsel *pos;
 877         const char *group_name = evsel__group_name(evsel);
 878
 879         if (!evsel->forced_leader)
 880                 ret = scnprintf(buf, size, "%s { ", group_name);
 881
 882         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 883
 884         for_each_group_member(pos, evsel)
 885                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 886
 887         if (!evsel->forced_leader)
 888                 ret += scnprintf(buf + ret, size - ret, " }");
 889
 890         return ret;
 891 }
 892
 893 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 894                                       struct callchain_param *param)
 895 {
 896         bool function = evsel__is_function_event(evsel);
 897         struct perf_event_attr *attr = &evsel->core.attr;
 898
 899         evsel__set_sample_bit(evsel, CALLCHAIN);
 900
 901         attr->sample_max_stack = param->max_stack;
 902
 903         if (opts->kernel_callchains)
 904                 attr->exclude_callchain_user = 1;
 905         if (opts->user_callchains)
 906                 attr->exclude_callchain_kernel = 1;
 907         if (param->record_mode == CALLCHAIN_LBR) {
 908                 if (!opts->branch_stack) {
 909                         if (attr->exclude_user) {
 910                                 pr_warning("LBR callstack option is only available "
 911                                            "to get user callchain information. "
 912                                            "Falling back to framepointers.\n");
 913                         } else {
 914                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 915                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 916                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 917                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 918                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 919                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 920                         }
 921                 } else
 922                          pr_warning("Cannot use LBR callstack with branch stack. "
 923                                     "Falling back to framepointers.\n");
 924         }
 925
 926         if (param->record_mode == CALLCHAIN_DWARF) {
 927                 if (!function) {
 928                         evsel__set_sample_bit(evsel, REGS_USER);
 929                         evsel__set_sample_bit(evsel, STACK_USER);
 930                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 931                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 932                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 933                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 934                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 935                         } else {
 936                                 attr->sample_regs_user |= arch__user_reg_mask();
 937                         }
 938                         attr->sample_stack_user = param->dump_size;
 939                         attr->exclude_callchain_user = 1;
 940                 } else {
 941                         pr_info("Cannot use DWARF unwind for function trace event,"
 942                                 " falling back to framepointers.\n");
 943                 }
 944         }
 945
 946         if (function) {
 947                 pr_info("Disabling user space callchains for function trace event.\n");
 948                 attr->exclude_callchain_user = 1;
 949         }
 950 }
 951
 952 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 953                              struct callchain_param *param)
 954 {
 955         if (param->enabled)
 956                 return __evsel__config_callchain(evsel, opts, param);
 957 }
 958
 959 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 960 {
 961         struct perf_event_attr *attr = &evsel->core.attr;
 962
 963         evsel__reset_sample_bit(evsel, CALLCHAIN);
 964         if (param->record_mode == CALLCHAIN_LBR) {
 965                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 966                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 967                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 968                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 969         }
 970         if (param->record_mode == CALLCHAIN_DWARF) {
 971                 evsel__reset_sample_bit(evsel, REGS_USER);
 972                 evsel__reset_sample_bit(evsel, STACK_USER);
 973         }
 974 }
 975
 976 static void evsel__apply_config_terms(struct evsel *evsel,
 977                                       struct record_opts *opts, bool track)
 978 {
 979         struct evsel_config_term *term;
 980         struct list_head *config_terms = &evsel->config_terms;
 981         struct perf_event_attr *attr = &evsel->core.attr;
 982         /* callgraph default */
 983         struct callchain_param param = {
 984                 .record_mode = callchain_param.record_mode,
 985         };
 986         u32 dump_size = 0;
 987         int max_stack = 0;
 988         const char *callgraph_buf = NULL;
 989
 990         list_for_each_entry(term, config_terms, list) {
 991                 switch (term->type) {
 992                 case EVSEL__CONFIG_TERM_PERIOD:
 993                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 994                                 attr->sample_period = term->val.period;
 995                                 attr->freq = 0;
 996                                 evsel__reset_sample_bit(evsel, PERIOD);
 997                         }
 998                         break;
 999                 case EVSEL__CONFIG_TERM_FREQ:
1000                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
1001                                 attr->sample_freq = term->val.freq;
1002                                 attr->freq = 1;
1003                                 evsel__set_sample_bit(evsel, PERIOD);
1004                         }
1005                         break;
1006                 case EVSEL__CONFIG_TERM_TIME:
1007                         if (term->val.time)
1008                                 evsel__set_sample_bit(evsel, TIME);
1009                         else
1010                                 evsel__reset_sample_bit(evsel, TIME);
1011                         break;
1012                 case EVSEL__CONFIG_TERM_CALLGRAPH:
1013                         callgraph_buf = term->val.str;
1014                         break;
1015                 case EVSEL__CONFIG_TERM_BRANCH:
1016                         if (term->val.str && strcmp(term->val.str, "no")) {
1017                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1018                                 parse_branch_str(term->val.str,
1019                                                  &attr->branch_sample_type);
1020                         } else
1021                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1022                         break;
1023                 case EVSEL__CONFIG_TERM_STACK_USER:
1024                         dump_size = term->val.stack_user;
1025                         break;
1026                 case EVSEL__CONFIG_TERM_MAX_STACK:
1027                         max_stack = term->val.max_stack;
1028                         break;
1029                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1030                         evsel->max_events = term->val.max_events;
1031                         break;
1032                 case EVSEL__CONFIG_TERM_INHERIT:
1033                         /*
1034                          * attr->inherit should has already been set by
1035                          * evsel__config. If user explicitly set
1036                          * inherit using config terms, override global
1037                          * opt->no_inherit setting.
1038                          */
1039                         attr->inherit = term->val.inherit ? 1 : 0;
1040                         break;
1041                 case EVSEL__CONFIG_TERM_OVERWRITE:
1042                         attr->write_backward = term->val.overwrite ? 1 : 0;
1043                         break;
1044                 case EVSEL__CONFIG_TERM_DRV_CFG:
1045                         break;
1046                 case EVSEL__CONFIG_TERM_PERCORE:
1047                         break;
1048                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1049                         attr->aux_output = term->val.aux_output ? 1 : 0;
1050                         break;
1051                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1052                         /* Already applied by auxtrace */
1053                         break;
1054                 case EVSEL__CONFIG_TERM_CFG_CHG:
1055                         break;
1056                 default:
1057                         break;
1058                 }
1059         }
1060
1061         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1062         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1063                 bool sample_address = false;
1064
1065                 if (max_stack) {
1066                         param.max_stack = max_stack;
1067                         if (callgraph_buf == NULL)
1068                                 callgraph_buf = "fp";
1069                 }
1070
1071                 /* parse callgraph parameters */
1072                 if (callgraph_buf != NULL) {
1073                         if (!strcmp(callgraph_buf, "no")) {
1074                                 param.enabled = false;
1075                                 param.record_mode = CALLCHAIN_NONE;
1076                         } else {
1077                                 param.enabled = true;
1078                                 if (parse_callchain_record(callgraph_buf, &param)) {
1079                                         pr_err("per-event callgraph setting for %s failed. "
1080                                                "Apply callgraph global setting for it\n",
1081                                                evsel->name);
1082                                         return;
1083                                 }
1084                                 if (param.record_mode == CALLCHAIN_DWARF)
1085                                         sample_address = true;
1086                         }
1087                 }
1088                 if (dump_size > 0) {
1089                         dump_size = round_up(dump_size, sizeof(u64));
1090                         param.dump_size = dump_size;
1091                 }
1092
1093                 /* If global callgraph set, clear it */
1094                 if (callchain_param.enabled)
1095                         evsel__reset_callgraph(evsel, &callchain_param);
1096
1097                 /* set perf-event callgraph */
1098                 if (param.enabled) {
1099                         if (sample_address) {
1100                                 evsel__set_sample_bit(evsel, ADDR);
1101                                 evsel__set_sample_bit(evsel, DATA_SRC);
1102                                 evsel->core.attr.mmap_data = track;
1103                         }
1104                         evsel__config_callchain(evsel, opts, &param);
1105                 }
1106         }
1107 }
1108
1109 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1110 {
1111         struct evsel_config_term *term, *found_term = NULL;
1112
1113         list_for_each_entry(term, &evsel->config_terms, list) {
1114                 if (term->type == type)
1115                         found_term = term;
1116         }
1117
1118         return found_term;
1119 }
1120
1121 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1122 {
1123         evsel__set_sample_bit(evsel, WEIGHT);
1124 }
1125
1126 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1127 {
1128 }
1129
1130 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1131                                     struct perf_event_attr *attr __maybe_unused)
1132 {
1133 }
1134
1135 static void evsel__set_default_freq_period(struct record_opts *opts,
1136                                            struct perf_event_attr *attr)
1137 {
1138         if (opts->freq) {
1139                 attr->freq = 1;
1140                 attr->sample_freq = opts->freq;
1141         } else {
1142                 attr->sample_period = opts->default_interval;
1143         }
1144 }
1145
1146 static bool evsel__is_offcpu_event(struct evsel *evsel)
1147 {
1148         return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1149 }
1150
1151 /*
1152  * The enable_on_exec/disabled value strategy:
1153  *
1154  *  1) For any type of traced program:
1155  *    - all independent events and group leaders are disabled
1156  *    - all group members are enabled
1157  *
1158  *     Group members are ruled by group leaders. They need to
1159  *     be enabled, because the group scheduling relies on that.
1160  *
1161  *  2) For traced programs executed by perf:
1162  *     - all independent events and group leaders have
1163  *       enable_on_exec set
1164  *     - we don't specifically enable or disable any event during
1165  *       the record command
1166  *
1167  *     Independent events and group leaders are initially disabled
1168  *     and get enabled by exec. Group members are ruled by group
1169  *     leaders as stated in 1).
1170  *
1171  *  3) For traced programs attached by perf (pid/tid):
1172  *     - we specifically enable or disable all events during
1173  *       the record command
1174  *
1175  *     When attaching events to already running traced we
1176  *     enable/disable events specifically, as there's no
1177  *     initial traced exec call.
1178  */
1179 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1180                    struct callchain_param *callchain)
1181 {
1182         struct evsel *leader = evsel__leader(evsel);
1183         struct perf_event_attr *attr = &evsel->core.attr;
1184         int track = evsel->tracking;
1185         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1186
1187         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1188         attr->inherit       = !opts->no_inherit;
1189         attr->write_backward = opts->overwrite ? 1 : 0;
1190         attr->read_format   = PERF_FORMAT_LOST;
1191
1192         evsel__set_sample_bit(evsel, IP);
1193         evsel__set_sample_bit(evsel, TID);
1194
1195         if (evsel->sample_read) {
1196                 evsel__set_sample_bit(evsel, READ);
1197
1198                 /*
1199                  * We need ID even in case of single event, because
1200                  * PERF_SAMPLE_READ process ID specific data.
1201                  */
1202                 evsel__set_sample_id(evsel, false);
1203
1204                 /*
1205                  * Apply group format only if we belong to group
1206                  * with more than one members.
1207                  */
1208                 if (leader->core.nr_members > 1) {
1209                         attr->read_format |= PERF_FORMAT_GROUP;
1210                         attr->inherit = 0;
1211                 }
1212         }
1213
1214         /*
1215          * We default some events to have a default interval. But keep
1216          * it a weak assumption overridable by the user.
1217          */
1218         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1219             (!evsel->is_libpfm_event && (!attr->sample_period ||
1220                                          opts->user_freq != UINT_MAX ||
1221                                          opts->user_interval != ULLONG_MAX)))
1222                 evsel__set_default_freq_period(opts, attr);
1223
1224         /*
1225          * If attr->freq was set (here or earlier), ask for period
1226          * to be sampled.
1227          */
1228         if (attr->freq)
1229                 evsel__set_sample_bit(evsel, PERIOD);
1230
1231         if (opts->no_samples)
1232                 attr->sample_freq = 0;
1233
1234         if (opts->inherit_stat) {
1235                 evsel->core.attr.read_format |=
1236                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1237                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1238                         PERF_FORMAT_ID;
1239                 attr->inherit_stat = 1;
1240         }
1241
1242         if (opts->sample_address) {
1243                 evsel__set_sample_bit(evsel, ADDR);
1244                 attr->mmap_data = track;
1245         }
1246
1247         /*
1248          * We don't allow user space callchains for  function trace
1249          * event, due to issues with page faults while tracing page
1250          * fault handler and its overall trickiness nature.
1251          */
1252         if (evsel__is_function_event(evsel))
1253                 evsel->core.attr.exclude_callchain_user = 1;
1254
1255         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1256                 evsel__config_callchain(evsel, opts, callchain);
1257
1258         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1259             !evsel__is_dummy_event(evsel)) {
1260                 attr->sample_regs_intr = opts->sample_intr_regs;
1261                 evsel__set_sample_bit(evsel, REGS_INTR);
1262         }
1263
1264         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1265             !evsel__is_dummy_event(evsel)) {
1266                 attr->sample_regs_user |= opts->sample_user_regs;
1267                 evsel__set_sample_bit(evsel, REGS_USER);
1268         }
1269
1270         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1271                 evsel__set_sample_bit(evsel, CPU);
1272
1273         /*
1274          * When the user explicitly disabled time don't force it here.
1275          */
1276         if (opts->sample_time &&
1277             (!perf_missing_features.sample_id_all &&
1278             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1279              opts->sample_time_set)))
1280                 evsel__set_sample_bit(evsel, TIME);
1281
1282         if (opts->raw_samples && !evsel->no_aux_samples) {
1283                 evsel__set_sample_bit(evsel, TIME);
1284                 evsel__set_sample_bit(evsel, RAW);
1285                 evsel__set_sample_bit(evsel, CPU);
1286         }
1287
1288         if (opts->sample_address)
1289                 evsel__set_sample_bit(evsel, DATA_SRC);
1290
1291         if (opts->sample_phys_addr)
1292                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1293
1294         if (opts->no_buffering) {
1295                 attr->watermark = 0;
1296                 attr->wakeup_events = 1;
1297         }
1298         if (opts->branch_stack && !evsel->no_aux_samples) {
1299                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1300                 attr->branch_sample_type = opts->branch_stack;
1301         }
1302
1303         if (opts->sample_weight)
1304                 arch_evsel__set_sample_weight(evsel);
1305
1306         attr->task     = track;
1307         attr->mmap     = track;
1308         attr->mmap2    = track && !perf_missing_features.mmap2;
1309         attr->comm     = track;
1310         attr->build_id = track && opts->build_id;
1311
1312         /*
1313          * ksymbol is tracked separately with text poke because it needs to be
1314          * system wide and enabled immediately.
1315          */
1316         if (!opts->text_poke)
1317                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1318         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1319
1320         if (opts->record_namespaces)
1321                 attr->namespaces  = track;
1322
1323         if (opts->record_cgroup) {
1324                 attr->cgroup = track && !perf_missing_features.cgroup;
1325                 evsel__set_sample_bit(evsel, CGROUP);
1326         }
1327
1328         if (opts->sample_data_page_size)
1329                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1330
1331         if (opts->sample_code_page_size)
1332                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1333
1334         if (opts->record_switch_events)
1335                 attr->context_switch = track;
1336
1337         if (opts->sample_transaction)
1338                 evsel__set_sample_bit(evsel, TRANSACTION);
1339
1340         if (opts->running_time) {
1341                 evsel->core.attr.read_format |=
1342                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1343                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1344         }
1345
1346         /*
1347          * XXX see the function comment above
1348          *
1349          * Disabling only independent events or group leaders,
1350          * keeping group members enabled.
1351          */
1352         if (evsel__is_group_leader(evsel))
1353                 attr->disabled = 1;
1354
1355         /*
1356          * Setting enable_on_exec for independent events and
1357          * group leaders for traced executed by perf.
1358          */
1359         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1360             !opts->target.initial_delay)
1361                 attr->enable_on_exec = 1;
1362
1363         if (evsel->immediate) {
1364                 attr->disabled = 0;
1365                 attr->enable_on_exec = 0;
1366         }
1367
1368         clockid = opts->clockid;
1369         if (opts->use_clockid) {
1370                 attr->use_clockid = 1;
1371                 attr->clockid = opts->clockid;
1372         }
1373
1374         if (evsel->precise_max)
1375                 attr->precise_ip = 3;
1376
1377         if (opts->all_user) {
1378                 attr->exclude_kernel = 1;
1379                 attr->exclude_user   = 0;
1380         }
1381
1382         if (opts->all_kernel) {
1383                 attr->exclude_kernel = 0;
1384                 attr->exclude_user   = 1;
1385         }
1386
1387         if (evsel->core.own_cpus || evsel->unit)
1388                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1389
1390         /*
1391          * Apply event specific term settings,
1392          * it overloads any global configuration.
1393          */
1394         evsel__apply_config_terms(evsel, opts, track);
1395
1396         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1397
1398         /* The --period option takes the precedence. */
1399         if (opts->period_set) {
1400                 if (opts->period)
1401                         evsel__set_sample_bit(evsel, PERIOD);
1402                 else
1403                         evsel__reset_sample_bit(evsel, PERIOD);
1404         }
1405
1406         /*
1407          * A dummy event never triggers any actual counter and therefore
1408          * cannot be used with branch_stack.
1409          *
1410          * For initial_delay, a dummy event is added implicitly.
1411          * The software event will trigger -EOPNOTSUPP error out,
1412          * if BRANCH_STACK bit is set.
1413          */
1414         if (evsel__is_dummy_event(evsel))
1415                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1416
1417         if (evsel__is_offcpu_event(evsel))
1418                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1419
1420         arch__post_evsel_config(evsel, attr);
1421 }
1422
1423 int evsel__set_filter(struct evsel *evsel, const char *filter)
1424 {
1425         char *new_filter = strdup(filter);
1426
1427         if (new_filter != NULL) {
1428                 free(evsel->filter);
1429                 evsel->filter = new_filter;
1430                 return 0;
1431         }
1432
1433         return -1;
1434 }
1435
1436 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1437 {
1438         char *new_filter;
1439
1440         if (evsel->filter == NULL)
1441                 return evsel__set_filter(evsel, filter);
1442
1443         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1444                 free(evsel->filter);
1445                 evsel->filter = new_filter;
1446                 return 0;
1447         }
1448
1449         return -1;
1450 }
1451
1452 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1453 {
1454         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1455 }
1456
1457 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1458 {
1459         return evsel__append_filter(evsel, "%s,%s", filter);
1460 }
1461
1462 /* Caller has to clear disabled after going through all CPUs. */
1463 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1464 {
1465         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1466 }
1467
1468 int evsel__enable(struct evsel *evsel)
1469 {
1470         int err = perf_evsel__enable(&evsel->core);
1471
1472         if (!err)
1473                 evsel->disabled = false;
1474         return err;
1475 }
1476
1477 /* Caller has to set disabled after going through all CPUs. */
1478 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1479 {
1480         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1481 }
1482
1483 int evsel__disable(struct evsel *evsel)
1484 {
1485         int err = perf_evsel__disable(&evsel->core);
1486         /*
1487          * We mark it disabled here so that tools that disable a event can
1488          * ignore events after they disable it. I.e. the ring buffer may have
1489          * already a few more events queued up before the kernel got the stop
1490          * request.
1491          */
1492         if (!err)
1493                 evsel->disabled = true;
1494
1495         return err;
1496 }
1497
1498 void free_config_terms(struct list_head *config_terms)
1499 {
1500         struct evsel_config_term *term, *h;
1501
1502         list_for_each_entry_safe(term, h, config_terms, list) {
1503                 list_del_init(&term->list);
1504                 if (term->free_str)
1505                         zfree(&term->val.str);
1506                 free(term);
1507         }
1508 }
1509
1510 static void evsel__free_config_terms(struct evsel *evsel)
1511 {
1512         free_config_terms(&evsel->config_terms);
1513 }
1514
1515 void evsel__exit(struct evsel *evsel)
1516 {
1517         assert(list_empty(&evsel->core.node));
1518         assert(evsel->evlist == NULL);
1519         bpf_counter__destroy(evsel);
1520         evsel__free_counts(evsel);
1521         perf_evsel__free_fd(&evsel->core);
1522         perf_evsel__free_id(&evsel->core);
1523         evsel__free_config_terms(evsel);
1524         cgroup__put(evsel->cgrp);
1525         perf_cpu_map__put(evsel->core.cpus);
1526         perf_cpu_map__put(evsel->core.own_cpus);
1527         perf_thread_map__put(evsel->core.threads);
1528         zfree(&evsel->group_name);
1529         zfree(&evsel->name);
1530         zfree(&evsel->pmu_name);
1531         zfree(&evsel->unit);
1532         zfree(&evsel->metric_id);
1533         evsel__zero_per_pkg(evsel);
1534         hashmap__free(evsel->per_pkg_mask);
1535         evsel->per_pkg_mask = NULL;
1536         zfree(&evsel->metric_events);
1537         perf_evsel__object.fini(evsel);
1538 }
1539
1540 void evsel__delete(struct evsel *evsel)
1541 {
1542         evsel__exit(evsel);
1543         free(evsel);
1544 }
1545
1546 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1547                            struct perf_counts_values *count)
1548 {
1549         struct perf_counts_values tmp;
1550
1551         if (!evsel->prev_raw_counts)
1552                 return;
1553
1554         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1555         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1556
1557         count->val = count->val - tmp.val;
1558         count->ena = count->ena - tmp.ena;
1559         count->run = count->run - tmp.run;
1560 }
1561
1562 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1563 {
1564         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1565
1566         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1567 }
1568
1569 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1570                              u64 val, u64 ena, u64 run, u64 lost)
1571 {
1572         struct perf_counts_values *count;
1573
1574         count = perf_counts(counter->counts, cpu_map_idx, thread);
1575
1576         count->val    = val;
1577         count->ena    = ena;
1578         count->run    = run;
1579         count->lost   = lost;
1580
1581         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1582 }
1583
1584 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1585 {
1586         u64 read_format = leader->core.attr.read_format;
1587         struct sample_read_value *v;
1588         u64 nr, ena = 0, run = 0, lost = 0;
1589
1590         nr = *data++;
1591
1592         if (nr != (u64) leader->core.nr_members)
1593                 return -EINVAL;
1594
1595         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1596                 ena = *data++;
1597
1598         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1599                 run = *data++;
1600
1601         v = (void *)data;
1602         sample_read_group__for_each(v, nr, read_format) {
1603                 struct evsel *counter;
1604
1605                 counter = evlist__id2evsel(leader->evlist, v->id);
1606                 if (!counter)
1607                         return -EINVAL;
1608
1609                 if (read_format & PERF_FORMAT_LOST)
1610                         lost = v->lost;
1611
1612                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1613         }
1614
1615         return 0;
1616 }
1617
1618 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1619 {
1620         struct perf_stat_evsel *ps = leader->stats;
1621         u64 read_format = leader->core.attr.read_format;
1622         int size = perf_evsel__read_size(&leader->core);
1623         u64 *data = ps->group_data;
1624
1625         if (!(read_format & PERF_FORMAT_ID))
1626                 return -EINVAL;
1627
1628         if (!evsel__is_group_leader(leader))
1629                 return -EINVAL;
1630
1631         if (!data) {
1632                 data = zalloc(size);
1633                 if (!data)
1634                         return -ENOMEM;
1635
1636                 ps->group_data = data;
1637         }
1638
1639         if (FD(leader, cpu_map_idx, thread) < 0)
1640                 return -EINVAL;
1641
1642         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1643                 return -errno;
1644
1645         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1646 }
1647
1648 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1649 {
1650         u64 read_format = evsel->core.attr.read_format;
1651
1652         if (read_format & PERF_FORMAT_GROUP)
1653                 return evsel__read_group(evsel, cpu_map_idx, thread);
1654
1655         return evsel__read_one(evsel, cpu_map_idx, thread);
1656 }
1657
1658 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1659 {
1660         struct perf_counts_values count;
1661         size_t nv = scale ? 3 : 1;
1662
1663         if (FD(evsel, cpu_map_idx, thread) < 0)
1664                 return -EINVAL;
1665
1666         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1667                 return -ENOMEM;
1668
1669         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1670                 return -errno;
1671
1672         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1673         perf_counts_values__scale(&count, scale, NULL);
1674         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1675         return 0;
1676 }
1677
1678 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1679                                   int cpu_map_idx)
1680 {
1681         struct perf_cpu cpu;
1682
1683         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1684         return perf_cpu_map__idx(other->core.cpus, cpu);
1685 }
1686
1687 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1688 {
1689         struct evsel *leader = evsel__leader(evsel);
1690
1691         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1692             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1693                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1694         }
1695
1696         return cpu_map_idx;
1697 }
1698
1699 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1700 {
1701         struct evsel *leader = evsel__leader(evsel);
1702         int fd;
1703
1704         if (evsel__is_group_leader(evsel))
1705                 return -1;
1706
1707         /*
1708          * Leader must be already processed/open,
1709          * if not it's a bug.
1710          */
1711         BUG_ON(!leader->core.fd);
1712
1713         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1714         if (cpu_map_idx == -1)
1715                 return -1;
1716
1717         fd = FD(leader, cpu_map_idx, thread);
1718         BUG_ON(fd == -1);
1719
1720         return fd;
1721 }
1722
1723 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1724 {
1725         for (int cpu = 0; cpu < nr_cpus; cpu++)
1726                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1727                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1728 }
1729
1730 static int update_fds(struct evsel *evsel,
1731                       int nr_cpus, int cpu_map_idx,
1732                       int nr_threads, int thread_idx)
1733 {
1734         struct evsel *pos;
1735
1736         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1737                 return -EINVAL;
1738
1739         evlist__for_each_entry(evsel->evlist, pos) {
1740                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1741
1742                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1743
1744                 /*
1745                  * Since fds for next evsel has not been created,
1746                  * there is no need to iterate whole event list.
1747                  */
1748                 if (pos == evsel)
1749                         break;
1750         }
1751         return 0;
1752 }
1753
1754 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1755                                          int nr_cpus, int cpu_map_idx,
1756                                          struct perf_thread_map *threads,
1757                                          int thread, int err)
1758 {
1759         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1760
1761         if (!evsel->ignore_missing_thread)
1762                 return false;
1763
1764         /* The system wide setup does not work with threads. */
1765         if (evsel->core.system_wide)
1766                 return false;
1767
1768         /* The -ESRCH is perf event syscall errno for pid's not found. */
1769         if (err != -ESRCH)
1770                 return false;
1771
1772         /* If there's only one thread, let it fail. */
1773         if (threads->nr == 1)
1774                 return false;
1775
1776         /*
1777          * We should remove fd for missing_thread first
1778          * because thread_map__remove() will decrease threads->nr.
1779          */
1780         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1781                 return false;
1782
1783         if (thread_map__remove(threads, thread))
1784                 return false;
1785
1786         pr_warning("WARNING: Ignored open failure for pid %d\n",
1787                    ignore_pid);
1788         return true;
1789 }
1790
1791 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1792                                 void *priv __maybe_unused)
1793 {
1794         return fprintf(fp, "  %-32s %s\n", name, val);
1795 }
1796
1797 static void display_attr(struct perf_event_attr *attr)
1798 {
1799         if (verbose >= 2 || debug_peo_args) {
1800                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1801                 fprintf(stderr, "perf_event_attr:\n");
1802                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1803                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1804         }
1805 }
1806
1807 bool evsel__precise_ip_fallback(struct evsel *evsel)
1808 {
1809         /* Do not try less precise if not requested. */
1810         if (!evsel->precise_max)
1811                 return false;
1812
1813         /*
1814          * We tried all the precise_ip values, and it's
1815          * still failing, so leave it to standard fallback.
1816          */
1817         if (!evsel->core.attr.precise_ip) {
1818                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1819                 return false;
1820         }
1821
1822         if (!evsel->precise_ip_original)
1823                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1824
1825         evsel->core.attr.precise_ip--;
1826         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1827         display_attr(&evsel->core.attr);
1828         return true;
1829 }
1830
1831 static struct perf_cpu_map *empty_cpu_map;
1832 static struct perf_thread_map *empty_thread_map;
1833
1834 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1835                 struct perf_thread_map *threads)
1836 {
1837         int nthreads = perf_thread_map__nr(threads);
1838
1839         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1840             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1841                 return -EINVAL;
1842
1843         if (cpus == NULL) {
1844                 if (empty_cpu_map == NULL) {
1845                         empty_cpu_map = perf_cpu_map__dummy_new();
1846                         if (empty_cpu_map == NULL)
1847                                 return -ENOMEM;
1848                 }
1849
1850                 cpus = empty_cpu_map;
1851         }
1852
1853         if (threads == NULL) {
1854                 if (empty_thread_map == NULL) {
1855                         empty_thread_map = thread_map__new_by_tid(-1);
1856                         if (empty_thread_map == NULL)
1857                                 return -ENOMEM;
1858                 }
1859
1860                 threads = empty_thread_map;
1861         }
1862
1863         if (evsel->core.fd == NULL &&
1864             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1865                 return -ENOMEM;
1866
1867         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1868         if (evsel->cgrp)
1869                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1870
1871         return 0;
1872 }
1873
1874 static void evsel__disable_missing_features(struct evsel *evsel)
1875 {
1876         if (perf_missing_features.read_lost)
1877                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1878         if (perf_missing_features.weight_struct) {
1879                 evsel__set_sample_bit(evsel, WEIGHT);
1880                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1881         }
1882         if (perf_missing_features.clockid_wrong)
1883                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1884         if (perf_missing_features.clockid) {
1885                 evsel->core.attr.use_clockid = 0;
1886                 evsel->core.attr.clockid = 0;
1887         }
1888         if (perf_missing_features.cloexec)
1889                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1890         if (perf_missing_features.mmap2)
1891                 evsel->core.attr.mmap2 = 0;
1892         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1893                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1894         if (perf_missing_features.lbr_flags)
1895                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1896                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1897         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1898                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1899         if (perf_missing_features.ksymbol)
1900                 evsel->core.attr.ksymbol = 0;
1901         if (perf_missing_features.bpf)
1902                 evsel->core.attr.bpf_event = 0;
1903         if (perf_missing_features.branch_hw_idx)
1904                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1905         if (perf_missing_features.sample_id_all)
1906                 evsel->core.attr.sample_id_all = 0;
1907 }
1908
1909 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1910                         struct perf_thread_map *threads)
1911 {
1912         int err;
1913
1914         err = __evsel__prepare_open(evsel, cpus, threads);
1915         if (err)
1916                 return err;
1917
1918         evsel__disable_missing_features(evsel);
1919
1920         return err;
1921 }
1922
1923 bool evsel__detect_missing_features(struct evsel *evsel)
1924 {
1925         /*
1926          * Must probe features in the order they were added to the
1927          * perf_event_attr interface.
1928          */
1929         if (!perf_missing_features.read_lost &&
1930             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1931                 perf_missing_features.read_lost = true;
1932                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1933                 return true;
1934         } else if (!perf_missing_features.weight_struct &&
1935             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1936                 perf_missing_features.weight_struct = true;
1937                 pr_debug2("switching off weight struct support\n");
1938                 return true;
1939         } else if (!perf_missing_features.code_page_size &&
1940             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1941                 perf_missing_features.code_page_size = true;
1942                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1943                 return false;
1944         } else if (!perf_missing_features.data_page_size &&
1945             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1946                 perf_missing_features.data_page_size = true;
1947                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1948                 return false;
1949         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1950                 perf_missing_features.cgroup = true;
1951                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1952                 return false;
1953         } else if (!perf_missing_features.branch_hw_idx &&
1954             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1955                 perf_missing_features.branch_hw_idx = true;
1956                 pr_debug2("switching off branch HW index support\n");
1957                 return true;
1958         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1959                 perf_missing_features.aux_output = true;
1960                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1961                 return false;
1962         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1963                 perf_missing_features.bpf = true;
1964                 pr_debug2_peo("switching off bpf_event\n");
1965                 return true;
1966         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1967                 perf_missing_features.ksymbol = true;
1968                 pr_debug2_peo("switching off ksymbol\n");
1969                 return true;
1970         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1971                 perf_missing_features.write_backward = true;
1972                 pr_debug2_peo("switching off write_backward\n");
1973                 return false;
1974         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1975                 perf_missing_features.clockid_wrong = true;
1976                 pr_debug2_peo("switching off clockid\n");
1977                 return true;
1978         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1979                 perf_missing_features.clockid = true;
1980                 pr_debug2_peo("switching off use_clockid\n");
1981                 return true;
1982         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1983                 perf_missing_features.cloexec = true;
1984                 pr_debug2_peo("switching off cloexec flag\n");
1985                 return true;
1986         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1987                 perf_missing_features.mmap2 = true;
1988                 pr_debug2_peo("switching off mmap2\n");
1989                 return true;
1990         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1991                 if (evsel->pmu == NULL)
1992                         evsel->pmu = evsel__find_pmu(evsel);
1993
1994                 if (evsel->pmu)
1995                         evsel->pmu->missing_features.exclude_guest = true;
1996                 else {
1997                         /* we cannot find PMU, disable attrs now */
1998                         evsel->core.attr.exclude_host = false;
1999                         evsel->core.attr.exclude_guest = false;
2000                 }
2001
2002                 if (evsel->exclude_GH) {
2003                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2004                         return false;
2005                 }
2006                 if (!perf_missing_features.exclude_guest) {
2007                         perf_missing_features.exclude_guest = true;
2008                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2009                 }
2010                 return true;
2011         } else if (!perf_missing_features.sample_id_all) {
2012                 perf_missing_features.sample_id_all = true;
2013                 pr_debug2_peo("switching off sample_id_all\n");
2014                 return true;
2015         } else if (!perf_missing_features.lbr_flags &&
2016                         (evsel->core.attr.branch_sample_type &
2017                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2018                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2019                 perf_missing_features.lbr_flags = true;
2020                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2021                 return true;
2022         } else if (!perf_missing_features.group_read &&
2023                     evsel->core.attr.inherit &&
2024                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2025                    evsel__is_group_leader(evsel)) {
2026                 perf_missing_features.group_read = true;
2027                 pr_debug2_peo("switching off group read\n");
2028                 return true;
2029         } else {
2030                 return false;
2031         }
2032 }
2033
2034 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2035 {
2036         int old_errno;
2037         struct rlimit l;
2038
2039         if (*set_rlimit < INCREASED_MAX) {
2040                 old_errno = errno;
2041
2042                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2043                         if (*set_rlimit == NO_CHANGE) {
2044                                 l.rlim_cur = l.rlim_max;
2045                         } else {
2046                                 l.rlim_cur = l.rlim_max + 1000;
2047                                 l.rlim_max = l.rlim_cur;
2048                         }
2049                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2050                                 (*set_rlimit) += 1;
2051                                 errno = old_errno;
2052                                 return true;
2053                         }
2054                 }
2055                 errno = old_errno;
2056         }
2057
2058         return false;
2059 }
2060
2061 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2062                 struct perf_thread_map *threads,
2063                 int start_cpu_map_idx, int end_cpu_map_idx)
2064 {
2065         int idx, thread, nthreads;
2066         int pid = -1, err, old_errno;
2067         enum rlimit_action set_rlimit = NO_CHANGE;
2068
2069         err = __evsel__prepare_open(evsel, cpus, threads);
2070         if (err)
2071                 return err;
2072
2073         if (cpus == NULL)
2074                 cpus = empty_cpu_map;
2075
2076         if (threads == NULL)
2077                 threads = empty_thread_map;
2078
2079         nthreads = perf_thread_map__nr(threads);
2080
2081         if (evsel->cgrp)
2082                 pid = evsel->cgrp->fd;
2083
2084 fallback_missing_features:
2085         evsel__disable_missing_features(evsel);
2086
2087         display_attr(&evsel->core.attr);
2088
2089         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2090
2091                 for (thread = 0; thread < nthreads; thread++) {
2092                         int fd, group_fd;
2093 retry_open:
2094                         if (thread >= nthreads)
2095                                 break;
2096
2097                         if (!evsel->cgrp && !evsel->core.system_wide)
2098                                 pid = perf_thread_map__pid(threads, thread);
2099
2100                         group_fd = get_group_fd(evsel, idx, thread);
2101
2102                         test_attr__ready();
2103
2104                         /* Debug message used by test scripts */
2105                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2106                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2107
2108                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2109                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2110                                                 group_fd, evsel->open_flags);
2111
2112                         FD(evsel, idx, thread) = fd;
2113
2114                         if (fd < 0) {
2115                                 err = -errno;
2116
2117                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2118                                           err);
2119                                 goto try_fallback;
2120                         }
2121
2122                         bpf_counter__install_pe(evsel, idx, fd);
2123
2124                         if (unlikely(test_attr__enabled)) {
2125                                 test_attr__open(&evsel->core.attr, pid,
2126                                                 perf_cpu_map__cpu(cpus, idx),
2127                                                 fd, group_fd, evsel->open_flags);
2128                         }
2129
2130                         /* Debug message used by test scripts */
2131                         pr_debug2_peo(" = %d\n", fd);
2132
2133                         if (evsel->bpf_fd >= 0) {
2134                                 int evt_fd = fd;
2135                                 int bpf_fd = evsel->bpf_fd;
2136
2137                                 err = ioctl(evt_fd,
2138                                             PERF_EVENT_IOC_SET_BPF,
2139                                             bpf_fd);
2140                                 if (err && errno != EEXIST) {
2141                                         pr_err("failed to attach bpf fd %d: %s\n",
2142                                                bpf_fd, strerror(errno));
2143                                         err = -EINVAL;
2144                                         goto out_close;
2145                                 }
2146                         }
2147
2148                         set_rlimit = NO_CHANGE;
2149
2150                         /*
2151                          * If we succeeded but had to kill clockid, fail and
2152                          * have evsel__open_strerror() print us a nice error.
2153                          */
2154                         if (perf_missing_features.clockid ||
2155                             perf_missing_features.clockid_wrong) {
2156                                 err = -EINVAL;
2157                                 goto out_close;
2158                         }
2159                 }
2160         }
2161
2162         return 0;
2163
2164 try_fallback:
2165         if (evsel__precise_ip_fallback(evsel))
2166                 goto retry_open;
2167
2168         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2169                                          idx, threads, thread, err)) {
2170                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2171                 nthreads--;
2172
2173                 /* ... and pretend like nothing have happened. */
2174                 err = 0;
2175                 goto retry_open;
2176         }
2177         /*
2178          * perf stat needs between 5 and 22 fds per CPU. When we run out
2179          * of them try to increase the limits.
2180          */
2181         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2182                 goto retry_open;
2183
2184         if (err != -EINVAL || idx > 0 || thread > 0)
2185                 goto out_close;
2186
2187         if (evsel__detect_missing_features(evsel))
2188                 goto fallback_missing_features;
2189 out_close:
2190         if (err)
2191                 threads->err_thread = thread;
2192
2193         old_errno = errno;
2194         do {
2195                 while (--thread >= 0) {
2196                         if (FD(evsel, idx, thread) >= 0)
2197                                 close(FD(evsel, idx, thread));
2198                         FD(evsel, idx, thread) = -1;
2199                 }
2200                 thread = nthreads;
2201         } while (--idx >= 0);
2202         errno = old_errno;
2203         return err;
2204 }
2205
2206 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2207                 struct perf_thread_map *threads)
2208 {
2209         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2210 }
2211
2212 void evsel__close(struct evsel *evsel)
2213 {
2214         perf_evsel__close(&evsel->core);
2215         perf_evsel__free_id(&evsel->core);
2216 }
2217
2218 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2219 {
2220         if (cpu_map_idx == -1)
2221                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2222
2223         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2224 }
2225
2226 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2227 {
2228         return evsel__open(evsel, NULL, threads);
2229 }
2230
2231 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2232                                        const union perf_event *event,
2233                                        struct perf_sample *sample)
2234 {
2235         u64 type = evsel->core.attr.sample_type;
2236         const __u64 *array = event->sample.array;
2237         bool swapped = evsel->needs_swap;
2238         union u64_swap u;
2239
2240         array += ((event->header.size -
2241                    sizeof(event->header)) / sizeof(u64)) - 1;
2242
2243         if (type & PERF_SAMPLE_IDENTIFIER) {
2244                 sample->id = *array;
2245                 array--;
2246         }
2247
2248         if (type & PERF_SAMPLE_CPU) {
2249                 u.val64 = *array;
2250                 if (swapped) {
2251                         /* undo swap of u64, then swap on individual u32s */
2252                         u.val64 = bswap_64(u.val64);
2253                         u.val32[0] = bswap_32(u.val32[0]);
2254                 }
2255
2256                 sample->cpu = u.val32[0];
2257                 array--;
2258         }
2259
2260         if (type & PERF_SAMPLE_STREAM_ID) {
2261                 sample->stream_id = *array;
2262                 array--;
2263         }
2264
2265         if (type & PERF_SAMPLE_ID) {
2266                 sample->id = *array;
2267                 array--;
2268         }
2269
2270         if (type & PERF_SAMPLE_TIME) {
2271                 sample->time = *array;
2272                 array--;
2273         }
2274
2275         if (type & PERF_SAMPLE_TID) {
2276                 u.val64 = *array;
2277                 if (swapped) {
2278                         /* undo swap of u64, then swap on individual u32s */
2279                         u.val64 = bswap_64(u.val64);
2280                         u.val32[0] = bswap_32(u.val32[0]);
2281                         u.val32[1] = bswap_32(u.val32[1]);
2282                 }
2283
2284                 sample->pid = u.val32[0];
2285                 sample->tid = u.val32[1];
2286                 array--;
2287         }
2288
2289         return 0;
2290 }
2291
2292 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2293                             u64 size)
2294 {
2295         return size > max_size || offset + size > endp;
2296 }
2297
2298 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2299         do {                                                            \
2300                 if (overflow(endp, (max_size), (offset), (size)))       \
2301                         return -EFAULT;                                 \
2302         } while (0)
2303
2304 #define OVERFLOW_CHECK_u64(offset) \
2305         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2306
2307 static int
2308 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2309 {
2310         /*
2311          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2312          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2313          * check the format does not go past the end of the event.
2314          */
2315         if (sample_size + sizeof(event->header) > event->header.size)
2316                 return -EFAULT;
2317
2318         return 0;
2319 }
2320
2321 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2322                                           const __u64 *array,
2323                                           u64 type __maybe_unused)
2324 {
2325         data->weight = *array;
2326 }
2327
2328 u64 evsel__bitfield_swap_branch_flags(u64 value)
2329 {
2330         u64 new_val = 0;
2331
2332         /*
2333          * branch_flags
2334          * union {
2335          *      u64 values;
2336          *      struct {
2337          *              mispred:1       //target mispredicted
2338          *              predicted:1     //target predicted
2339          *              in_tx:1         //in transaction
2340          *              abort:1         //transaction abort
2341          *              cycles:16       //cycle count to last branch
2342          *              type:4          //branch type
2343          *              spec:2          //branch speculation info
2344          *              new_type:4      //additional branch type
2345          *              priv:3          //privilege level
2346          *              reserved:31
2347          *      }
2348          * }
2349          *
2350          * Avoid bswap64() the entire branch_flag.value,
2351          * as it has variable bit-field sizes. Instead the
2352          * macro takes the bit-field position/size,
2353          * swaps it based on the host endianness.
2354          */
2355         if (host_is_bigendian()) {
2356                 new_val = bitfield_swap(value, 0, 1);
2357                 new_val |= bitfield_swap(value, 1, 1);
2358                 new_val |= bitfield_swap(value, 2, 1);
2359                 new_val |= bitfield_swap(value, 3, 1);
2360                 new_val |= bitfield_swap(value, 4, 16);
2361                 new_val |= bitfield_swap(value, 20, 4);
2362                 new_val |= bitfield_swap(value, 24, 2);
2363                 new_val |= bitfield_swap(value, 26, 4);
2364                 new_val |= bitfield_swap(value, 30, 3);
2365                 new_val |= bitfield_swap(value, 33, 31);
2366         } else {
2367                 new_val = bitfield_swap(value, 63, 1);
2368                 new_val |= bitfield_swap(value, 62, 1);
2369                 new_val |= bitfield_swap(value, 61, 1);
2370                 new_val |= bitfield_swap(value, 60, 1);
2371                 new_val |= bitfield_swap(value, 44, 16);
2372                 new_val |= bitfield_swap(value, 40, 4);
2373                 new_val |= bitfield_swap(value, 38, 2);
2374                 new_val |= bitfield_swap(value, 34, 4);
2375                 new_val |= bitfield_swap(value, 31, 3);
2376                 new_val |= bitfield_swap(value, 0, 31);
2377         }
2378
2379         return new_val;
2380 }
2381
2382 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2383                         struct perf_sample *data)
2384 {
2385         u64 type = evsel->core.attr.sample_type;
2386         bool swapped = evsel->needs_swap;
2387         const __u64 *array;
2388         u16 max_size = event->header.size;
2389         const void *endp = (void *)event + max_size;
2390         u64 sz;
2391
2392         /*
2393          * used for cross-endian analysis. See git commit 65014ab3
2394          * for why this goofiness is needed.
2395          */
2396         union u64_swap u;
2397
2398         memset(data, 0, sizeof(*data));
2399         data->cpu = data->pid = data->tid = -1;
2400         data->stream_id = data->id = data->time = -1ULL;
2401         data->period = evsel->core.attr.sample_period;
2402         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2403         data->misc    = event->header.misc;
2404         data->id = -1ULL;
2405         data->data_src = PERF_MEM_DATA_SRC_NONE;
2406         data->vcpu = -1;
2407
2408         if (event->header.type != PERF_RECORD_SAMPLE) {
2409                 if (!evsel->core.attr.sample_id_all)
2410                         return 0;
2411                 return perf_evsel__parse_id_sample(evsel, event, data);
2412         }
2413
2414         array = event->sample.array;
2415
2416         if (perf_event__check_size(event, evsel->sample_size))
2417                 return -EFAULT;
2418
2419         if (type & PERF_SAMPLE_IDENTIFIER) {
2420                 data->id = *array;
2421                 array++;
2422         }
2423
2424         if (type & PERF_SAMPLE_IP) {
2425                 data->ip = *array;
2426                 array++;
2427         }
2428
2429         if (type & PERF_SAMPLE_TID) {
2430                 u.val64 = *array;
2431                 if (swapped) {
2432                         /* undo swap of u64, then swap on individual u32s */
2433                         u.val64 = bswap_64(u.val64);
2434                         u.val32[0] = bswap_32(u.val32[0]);
2435                         u.val32[1] = bswap_32(u.val32[1]);
2436                 }
2437
2438                 data->pid = u.val32[0];
2439                 data->tid = u.val32[1];
2440                 array++;
2441         }
2442
2443         if (type & PERF_SAMPLE_TIME) {
2444                 data->time = *array;
2445                 array++;
2446         }
2447
2448         if (type & PERF_SAMPLE_ADDR) {
2449                 data->addr = *array;
2450                 array++;
2451         }
2452
2453         if (type & PERF_SAMPLE_ID) {
2454                 data->id = *array;
2455                 array++;
2456         }
2457
2458         if (type & PERF_SAMPLE_STREAM_ID) {
2459                 data->stream_id = *array;
2460                 array++;
2461         }
2462
2463         if (type & PERF_SAMPLE_CPU) {
2464
2465                 u.val64 = *array;
2466                 if (swapped) {
2467                         /* undo swap of u64, then swap on individual u32s */
2468                         u.val64 = bswap_64(u.val64);
2469                         u.val32[0] = bswap_32(u.val32[0]);
2470                 }
2471
2472                 data->cpu = u.val32[0];
2473                 array++;
2474         }
2475
2476         if (type & PERF_SAMPLE_PERIOD) {
2477                 data->period = *array;
2478                 array++;
2479         }
2480
2481         if (type & PERF_SAMPLE_READ) {
2482                 u64 read_format = evsel->core.attr.read_format;
2483
2484                 OVERFLOW_CHECK_u64(array);
2485                 if (read_format & PERF_FORMAT_GROUP)
2486                         data->read.group.nr = *array;
2487                 else
2488                         data->read.one.value = *array;
2489
2490                 array++;
2491
2492                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2493                         OVERFLOW_CHECK_u64(array);
2494                         data->read.time_enabled = *array;
2495                         array++;
2496                 }
2497
2498                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2499                         OVERFLOW_CHECK_u64(array);
2500                         data->read.time_running = *array;
2501                         array++;
2502                 }
2503
2504                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2505                 if (read_format & PERF_FORMAT_GROUP) {
2506                         const u64 max_group_nr = UINT64_MAX /
2507                                         sizeof(struct sample_read_value);
2508
2509                         if (data->read.group.nr > max_group_nr)
2510                                 return -EFAULT;
2511
2512                         sz = data->read.group.nr * sample_read_value_size(read_format);
2513                         OVERFLOW_CHECK(array, sz, max_size);
2514                         data->read.group.values =
2515                                         (struct sample_read_value *)array;
2516                         array = (void *)array + sz;
2517                 } else {
2518                         OVERFLOW_CHECK_u64(array);
2519                         data->read.one.id = *array;
2520                         array++;
2521
2522                         if (read_format & PERF_FORMAT_LOST) {
2523                                 OVERFLOW_CHECK_u64(array);
2524                                 data->read.one.lost = *array;
2525                                 array++;
2526                         }
2527                 }
2528         }
2529
2530         if (type & PERF_SAMPLE_CALLCHAIN) {
2531                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2532
2533                 OVERFLOW_CHECK_u64(array);
2534                 data->callchain = (struct ip_callchain *)array++;
2535                 if (data->callchain->nr > max_callchain_nr)
2536                         return -EFAULT;
2537                 sz = data->callchain->nr * sizeof(u64);
2538                 OVERFLOW_CHECK(array, sz, max_size);
2539                 array = (void *)array + sz;
2540         }
2541
2542         if (type & PERF_SAMPLE_RAW) {
2543                 OVERFLOW_CHECK_u64(array);
2544                 u.val64 = *array;
2545
2546                 /*
2547                  * Undo swap of u64, then swap on individual u32s,
2548                  * get the size of the raw area and undo all of the
2549                  * swap. The pevent interface handles endianness by
2550                  * itself.
2551                  */
2552                 if (swapped) {
2553                         u.val64 = bswap_64(u.val64);
2554                         u.val32[0] = bswap_32(u.val32[0]);
2555                         u.val32[1] = bswap_32(u.val32[1]);
2556                 }
2557                 data->raw_size = u.val32[0];
2558
2559                 /*
2560                  * The raw data is aligned on 64bits including the
2561                  * u32 size, so it's safe to use mem_bswap_64.
2562                  */
2563                 if (swapped)
2564                         mem_bswap_64((void *) array, data->raw_size);
2565
2566                 array = (void *)array + sizeof(u32);
2567
2568                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2569                 data->raw_data = (void *)array;
2570                 array = (void *)array + data->raw_size;
2571         }
2572
2573         if (type & PERF_SAMPLE_BRANCH_STACK) {
2574                 const u64 max_branch_nr = UINT64_MAX /
2575                                           sizeof(struct branch_entry);
2576                 struct branch_entry *e;
2577                 unsigned int i;
2578
2579                 OVERFLOW_CHECK_u64(array);
2580                 data->branch_stack = (struct branch_stack *)array++;
2581
2582                 if (data->branch_stack->nr > max_branch_nr)
2583                         return -EFAULT;
2584
2585                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2586                 if (evsel__has_branch_hw_idx(evsel)) {
2587                         sz += sizeof(u64);
2588                         e = &data->branch_stack->entries[0];
2589                 } else {
2590                         data->no_hw_idx = true;
2591                         /*
2592                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2593                          * only nr and entries[] will be output by kernel.
2594                          */
2595                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2596                 }
2597
2598                 if (swapped) {
2599                         /*
2600                          * struct branch_flag does not have endian
2601                          * specific bit field definition. And bswap
2602                          * will not resolve the issue, since these
2603                          * are bit fields.
2604                          *
2605                          * evsel__bitfield_swap_branch_flags() uses a
2606                          * bitfield_swap macro to swap the bit position
2607                          * based on the host endians.
2608                          */
2609                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2610                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2611                 }
2612
2613                 OVERFLOW_CHECK(array, sz, max_size);
2614                 array = (void *)array + sz;
2615         }
2616
2617         if (type & PERF_SAMPLE_REGS_USER) {
2618                 OVERFLOW_CHECK_u64(array);
2619                 data->user_regs.abi = *array;
2620                 array++;
2621
2622                 if (data->user_regs.abi) {
2623                         u64 mask = evsel->core.attr.sample_regs_user;
2624
2625                         sz = hweight64(mask) * sizeof(u64);
2626                         OVERFLOW_CHECK(array, sz, max_size);
2627                         data->user_regs.mask = mask;
2628                         data->user_regs.regs = (u64 *)array;
2629                         array = (void *)array + sz;
2630                 }
2631         }
2632
2633         if (type & PERF_SAMPLE_STACK_USER) {
2634                 OVERFLOW_CHECK_u64(array);
2635                 sz = *array++;
2636
2637                 data->user_stack.offset = ((char *)(array - 1)
2638                                           - (char *) event);
2639
2640                 if (!sz) {
2641                         data->user_stack.size = 0;
2642                 } else {
2643                         OVERFLOW_CHECK(array, sz, max_size);
2644                         data->user_stack.data = (char *)array;
2645                         array = (void *)array + sz;
2646                         OVERFLOW_CHECK_u64(array);
2647                         data->user_stack.size = *array++;
2648                         if (WARN_ONCE(data->user_stack.size > sz,
2649                                       "user stack dump failure\n"))
2650                                 return -EFAULT;
2651                 }
2652         }
2653
2654         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2655                 OVERFLOW_CHECK_u64(array);
2656                 arch_perf_parse_sample_weight(data, array, type);
2657                 array++;
2658         }
2659
2660         if (type & PERF_SAMPLE_DATA_SRC) {
2661                 OVERFLOW_CHECK_u64(array);
2662                 data->data_src = *array;
2663                 array++;
2664         }
2665
2666         if (type & PERF_SAMPLE_TRANSACTION) {
2667                 OVERFLOW_CHECK_u64(array);
2668                 data->transaction = *array;
2669                 array++;
2670         }
2671
2672         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2673         if (type & PERF_SAMPLE_REGS_INTR) {
2674                 OVERFLOW_CHECK_u64(array);
2675                 data->intr_regs.abi = *array;
2676                 array++;
2677
2678                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2679                         u64 mask = evsel->core.attr.sample_regs_intr;
2680
2681                         sz = hweight64(mask) * sizeof(u64);
2682                         OVERFLOW_CHECK(array, sz, max_size);
2683                         data->intr_regs.mask = mask;
2684                         data->intr_regs.regs = (u64 *)array;
2685                         array = (void *)array + sz;
2686                 }
2687         }
2688
2689         data->phys_addr = 0;
2690         if (type & PERF_SAMPLE_PHYS_ADDR) {
2691                 data->phys_addr = *array;
2692                 array++;
2693         }
2694
2695         data->cgroup = 0;
2696         if (type & PERF_SAMPLE_CGROUP) {
2697                 data->cgroup = *array;
2698                 array++;
2699         }
2700
2701         data->data_page_size = 0;
2702         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2703                 data->data_page_size = *array;
2704                 array++;
2705         }
2706
2707         data->code_page_size = 0;
2708         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2709                 data->code_page_size = *array;
2710                 array++;
2711         }
2712
2713         if (type & PERF_SAMPLE_AUX) {
2714                 OVERFLOW_CHECK_u64(array);
2715                 sz = *array++;
2716
2717                 OVERFLOW_CHECK(array, sz, max_size);
2718                 /* Undo swap of data */
2719                 if (swapped)
2720                         mem_bswap_64((char *)array, sz);
2721                 data->aux_sample.size = sz;
2722                 data->aux_sample.data = (char *)array;
2723                 array = (void *)array + sz;
2724         }
2725
2726         return 0;
2727 }
2728
2729 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2730                                   u64 *timestamp)
2731 {
2732         u64 type = evsel->core.attr.sample_type;
2733         const __u64 *array;
2734
2735         if (!(type & PERF_SAMPLE_TIME))
2736                 return -1;
2737
2738         if (event->header.type != PERF_RECORD_SAMPLE) {
2739                 struct perf_sample data = {
2740                         .time = -1ULL,
2741                 };
2742
2743                 if (!evsel->core.attr.sample_id_all)
2744                         return -1;
2745                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2746                         return -1;
2747
2748                 *timestamp = data.time;
2749                 return 0;
2750         }
2751
2752         array = event->sample.array;
2753
2754         if (perf_event__check_size(event, evsel->sample_size))
2755                 return -EFAULT;
2756
2757         if (type & PERF_SAMPLE_IDENTIFIER)
2758                 array++;
2759
2760         if (type & PERF_SAMPLE_IP)
2761                 array++;
2762
2763         if (type & PERF_SAMPLE_TID)
2764                 array++;
2765
2766         if (type & PERF_SAMPLE_TIME)
2767                 *timestamp = *array;
2768
2769         return 0;
2770 }
2771
2772 u16 evsel__id_hdr_size(struct evsel *evsel)
2773 {
2774         u64 sample_type = evsel->core.attr.sample_type;
2775         u16 size = 0;
2776
2777         if (sample_type & PERF_SAMPLE_TID)
2778                 size += sizeof(u64);
2779
2780         if (sample_type & PERF_SAMPLE_TIME)
2781                 size += sizeof(u64);
2782
2783         if (sample_type & PERF_SAMPLE_ID)
2784                 size += sizeof(u64);
2785
2786         if (sample_type & PERF_SAMPLE_STREAM_ID)
2787                 size += sizeof(u64);
2788
2789         if (sample_type & PERF_SAMPLE_CPU)
2790                 size += sizeof(u64);
2791
2792         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2793                 size += sizeof(u64);
2794
2795         return size;
2796 }
2797
2798 #ifdef HAVE_LIBTRACEEVENT
2799 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2800 {
2801         return tep_find_field(evsel->tp_format, name);
2802 }
2803
2804 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2805 {
2806         struct tep_format_field *field = evsel__field(evsel, name);
2807         int offset;
2808
2809         if (!field)
2810                 return NULL;
2811
2812         offset = field->offset;
2813
2814         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2815                 offset = *(int *)(sample->raw_data + field->offset);
2816                 offset &= 0xffff;
2817                 if (tep_field_is_relative(field->flags))
2818                         offset += field->offset + field->size;
2819         }
2820
2821         return sample->raw_data + offset;
2822 }
2823
2824 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2825                          bool needs_swap)
2826 {
2827         u64 value;
2828         void *ptr = sample->raw_data + field->offset;
2829
2830         switch (field->size) {
2831         case 1:
2832                 return *(u8 *)ptr;
2833         case 2:
2834                 value = *(u16 *)ptr;
2835                 break;
2836         case 4:
2837                 value = *(u32 *)ptr;
2838                 break;
2839         case 8:
2840                 memcpy(&value, ptr, sizeof(u64));
2841                 break;
2842         default:
2843                 return 0;
2844         }
2845
2846         if (!needs_swap)
2847                 return value;
2848
2849         switch (field->size) {
2850         case 2:
2851                 return bswap_16(value);
2852         case 4:
2853                 return bswap_32(value);
2854         case 8:
2855                 return bswap_64(value);
2856         default:
2857                 return 0;
2858         }
2859
2860         return 0;
2861 }
2862
2863 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2864 {
2865         struct tep_format_field *field = evsel__field(evsel, name);
2866
2867         if (!field)
2868                 return 0;
2869
2870         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2871 }
2872 #endif
2873
2874 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2875 {
2876         int paranoid;
2877
2878         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2879             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2880             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2881                 /*
2882                  * If it's cycles then fall back to hrtimer based
2883                  * cpu-clock-tick sw counter, which is always available even if
2884                  * no PMU support.
2885                  *
2886                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2887                  * b0a873e).
2888                  */
2889                 scnprintf(msg, msgsize, "%s",
2890 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2891
2892                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2893                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2894
2895                 zfree(&evsel->name);
2896                 return true;
2897         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2898                    (paranoid = perf_event_paranoid()) > 1) {
2899                 const char *name = evsel__name(evsel);
2900                 char *new_name;
2901                 const char *sep = ":";
2902
2903                 /* If event has exclude user then don't exclude kernel. */
2904                 if (evsel->core.attr.exclude_user)
2905                         return false;
2906
2907                 /* Is there already the separator in the name. */
2908                 if (strchr(name, '/') ||
2909                     (strchr(name, ':') && !evsel->is_libpfm_event))
2910                         sep = "";
2911
2912                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2913                         return false;
2914
2915                 if (evsel->name)
2916                         free(evsel->name);
2917                 evsel->name = new_name;
2918                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2919                           "to fall back to excluding kernel and hypervisor "
2920                           " samples", paranoid);
2921                 evsel->core.attr.exclude_kernel = 1;
2922                 evsel->core.attr.exclude_hv     = 1;
2923
2924                 return true;
2925         }
2926
2927         return false;
2928 }
2929
2930 static bool find_process(const char *name)
2931 {
2932         size_t len = strlen(name);
2933         DIR *dir;
2934         struct dirent *d;
2935         int ret = -1;
2936
2937         dir = opendir(procfs__mountpoint());
2938         if (!dir)
2939                 return false;
2940
2941         /* Walk through the directory. */
2942         while (ret && (d = readdir(dir)) != NULL) {
2943                 char path[PATH_MAX];
2944                 char *data;
2945                 size_t size;
2946
2947                 if ((d->d_type != DT_DIR) ||
2948                      !strcmp(".", d->d_name) ||
2949                      !strcmp("..", d->d_name))
2950                         continue;
2951
2952                 scnprintf(path, sizeof(path), "%s/%s/comm",
2953                           procfs__mountpoint(), d->d_name);
2954
2955                 if (filename__read_str(path, &data, &size))
2956                         continue;
2957
2958                 ret = strncmp(name, data, len);
2959                 free(data);
2960         }
2961
2962         closedir(dir);
2963         return ret ? false : true;
2964 }
2965
2966 static bool is_amd(const char *arch, const char *cpuid)
2967 {
2968         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2969 }
2970
2971 static bool is_amd_ibs(struct evsel *evsel)
2972 {
2973         return evsel->core.attr.precise_ip
2974             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2975 }
2976
2977 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2978                          int err, char *msg, size_t size)
2979 {
2980         struct perf_env *env = evsel__env(evsel);
2981         const char *arch = perf_env__arch(env);
2982         const char *cpuid = perf_env__cpuid(env);
2983         char sbuf[STRERR_BUFSIZE];
2984         int printed = 0, enforced = 0;
2985
2986         switch (err) {
2987         case EPERM:
2988         case EACCES:
2989                 printed += scnprintf(msg + printed, size - printed,
2990                         "Access to performance monitoring and observability operations is limited.\n");
2991
2992                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2993                         if (enforced) {
2994                                 printed += scnprintf(msg + printed, size - printed,
2995                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2996                                         "monitoring and observability operations. Inspect system audit records for\n"
2997                                         "more perf_event access control information and adjusting the policy.\n");
2998                         }
2999                 }
3000
3001                 if (err == EPERM)
3002                         printed += scnprintf(msg, size,
3003                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
3004
3005                 return scnprintf(msg + printed, size - printed,
3006                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3007                  "access to performance monitoring and observability operations for processes\n"
3008                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3009                  "More information can be found at 'Perf events and tool security' document:\n"
3010                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3011                  "perf_event_paranoid setting is %d:\n"
3012                  "  -1: Allow use of (almost) all events by all users\n"
3013                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3014                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
3015                  ">= 1: Disallow CPU event access\n"
3016                  ">= 2: Disallow kernel profiling\n"
3017                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3018                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3019                  perf_event_paranoid());
3020         case ENOENT:
3021                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3022         case EMFILE:
3023                 return scnprintf(msg, size, "%s",
3024                          "Too many events are opened.\n"
3025                          "Probably the maximum number of open file descriptors has been reached.\n"
3026                          "Hint: Try again after reducing the number of events.\n"
3027                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3028         case ENOMEM:
3029                 if (evsel__has_callchain(evsel) &&
3030                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3031                         return scnprintf(msg, size,
3032                                          "Not enough memory to setup event with callchain.\n"
3033                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3034                                          "Hint: Current value: %d", sysctl__max_stack());
3035                 break;
3036         case ENODEV:
3037                 if (target->cpu_list)
3038                         return scnprintf(msg, size, "%s",
3039          "No such device - did you specify an out-of-range profile CPU?");
3040                 break;
3041         case EOPNOTSUPP:
3042                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3043                         return scnprintf(msg, size,
3044         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3045                                          evsel__name(evsel));
3046                 if (evsel->core.attr.aux_output)
3047                         return scnprintf(msg, size,
3048         "%s: PMU Hardware doesn't support 'aux_output' feature",
3049                                          evsel__name(evsel));
3050                 if (evsel->core.attr.sample_period != 0)
3051                         return scnprintf(msg, size,
3052         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3053                                          evsel__name(evsel));
3054                 if (evsel->core.attr.precise_ip)
3055                         return scnprintf(msg, size, "%s",
3056         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3057 #if defined(__i386__) || defined(__x86_64__)
3058                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3059                         return scnprintf(msg, size, "%s",
3060         "No hardware sampling interrupt available.\n");
3061 #endif
3062                 break;
3063         case EBUSY:
3064                 if (find_process("oprofiled"))
3065                         return scnprintf(msg, size,
3066         "The PMU counters are busy/taken by another profiler.\n"
3067         "We found oprofile daemon running, please stop it and try again.");
3068                 break;
3069         case EINVAL:
3070                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3071                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3072                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3073                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3074                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3075                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3076                 if (perf_missing_features.clockid)
3077                         return scnprintf(msg, size, "clockid feature not supported.");
3078                 if (perf_missing_features.clockid_wrong)
3079                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3080                 if (perf_missing_features.aux_output)
3081                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3082                 if (!target__has_cpu(target))
3083                         return scnprintf(msg, size,
3084         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3085                                         evsel__name(evsel));
3086                 if (is_amd(arch, cpuid)) {
3087                         if (is_amd_ibs(evsel)) {
3088                                 if (evsel->core.attr.exclude_kernel)
3089                                         return scnprintf(msg, size,
3090         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3091                                 if (!evsel->core.system_wide)
3092                                         return scnprintf(msg, size,
3093         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3094                         }
3095                 }
3096
3097                 break;
3098         case ENODATA:
3099                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3100                                  "Please add an auxiliary event in front of the load latency event.");
3101         default:
3102                 break;
3103         }
3104
3105         return scnprintf(msg, size,
3106         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3107         "/bin/dmesg | grep -i perf may provide additional information.\n",
3108                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3109 }
3110
3111 struct perf_env *evsel__env(struct evsel *evsel)
3112 {
3113         if (evsel && evsel->evlist && evsel->evlist->env)
3114                 return evsel->evlist->env;
3115         return &perf_env;
3116 }
3117
3118 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3119 {
3120         int cpu_map_idx, thread;
3121
3122         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3123                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3124                      thread++) {
3125                         int fd = FD(evsel, cpu_map_idx, thread);
3126
3127                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3128                                                    cpu_map_idx, thread, fd) < 0)
3129                                 return -1;
3130                 }
3131         }
3132
3133         return 0;
3134 }
3135
3136 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3137 {
3138         struct perf_cpu_map *cpus = evsel->core.cpus;
3139         struct perf_thread_map *threads = evsel->core.threads;
3140
3141         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3142                 return -ENOMEM;
3143
3144         return store_evsel_ids(evsel, evlist);
3145 }
3146
3147 void evsel__zero_per_pkg(struct evsel *evsel)
3148 {
3149         struct hashmap_entry *cur;
3150         size_t bkt;
3151
3152         if (evsel->per_pkg_mask) {
3153                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3154                         free((void *)cur->pkey);
3155
3156                 hashmap__clear(evsel->per_pkg_mask);
3157         }
3158 }
3159
3160 bool evsel__is_hybrid(const struct evsel *evsel)
3161 {
3162         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3163 }
3164
3165 struct evsel *evsel__leader(const struct evsel *evsel)
3166 {
3167         return container_of(evsel->core.leader, struct evsel, core);
3168 }
3169
3170 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3171 {
3172         return evsel->core.leader == &leader->core;
3173 }
3174
3175 bool evsel__is_leader(struct evsel *evsel)
3176 {
3177         return evsel__has_leader(evsel, evsel);
3178 }
3179
3180 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3181 {
3182         evsel->core.leader = &leader->core;
3183 }
3184
3185 int evsel__source_count(const struct evsel *evsel)
3186 {
3187         struct evsel *pos;
3188         int count = 0;
3189
3190         evlist__for_each_entry(evsel->evlist, pos) {
3191                 if (pos->metric_leader == evsel)
3192                         count++;
3193         }
3194         return count;
3195 }
3196
3197 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3198 {
3199         return false;
3200 }
3201
3202 /*
3203  * Remove an event from a given group (leader).
3204  * Some events, e.g., perf metrics Topdown events,
3205  * must always be grouped. Ignore the events.
3206  */
3207 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3208 {
3209         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3210                 evsel__set_leader(evsel, evsel);
3211                 evsel->core.nr_members = 0;
3212                 leader->core.nr_members--;
3213         }
3214 }