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 "util/bpf-filter.h"
  54 #include <internal/xyarray.h>
  55 #include <internal/lib.h>
  56 #include <internal/threadmap.h>
  57
  58 #include <linux/ctype.h>
  59
  60 #ifdef HAVE_LIBTRACEEVENT
  61 #include <traceevent/event-parse.h>
  62 #endif
  63
  64 struct perf_missing_features perf_missing_features;
  65
  66 static clockid_t clockid;
  67
  68 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  69         NULL,
  70         "duration_time",
  71         "user_time",
  72         "system_time",
  73 };
  74
  75 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  76 {
  77         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  78                 return perf_tool_event__tool_names[ev];
  79
  80         return NULL;
  81 }
  82
  83 enum perf_tool_event perf_tool_event__from_str(const char *str)
  84 {
  85         int i;
  86
  87         perf_tool_event__for_each_event(i) {
  88                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  89                         return i;
  90         }
  91         return PERF_TOOL_NONE;
  92 }
  93
  94
  95 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  96 {
  97         return 0;
  98 }
  99
 100 void __weak test_attr__ready(void) { }
 101
 102 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 103 {
 104 }
 105
 106 static struct {
 107         size_t  size;
 108         int     (*init)(struct evsel *evsel);
 109         void    (*fini)(struct evsel *evsel);
 110 } perf_evsel__object = {
 111         .size = sizeof(struct evsel),
 112         .init = evsel__no_extra_init,
 113         .fini = evsel__no_extra_fini,
 114 };
 115
 116 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 117                          void (*fini)(struct evsel *evsel))
 118 {
 119
 120         if (object_size == 0)
 121                 goto set_methods;
 122
 123         if (perf_evsel__object.size > object_size)
 124                 return -EINVAL;
 125
 126         perf_evsel__object.size = object_size;
 127
 128 set_methods:
 129         if (init != NULL)
 130                 perf_evsel__object.init = init;
 131
 132         if (fini != NULL)
 133                 perf_evsel__object.fini = fini;
 134
 135         return 0;
 136 }
 137
 138 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 139
 140 int __evsel__sample_size(u64 sample_type)
 141 {
 142         u64 mask = sample_type & PERF_SAMPLE_MASK;
 143         int size = 0;
 144         int i;
 145
 146         for (i = 0; i < 64; i++) {
 147                 if (mask & (1ULL << i))
 148                         size++;
 149         }
 150
 151         size *= sizeof(u64);
 152
 153         return size;
 154 }
 155
 156 /**
 157  * __perf_evsel__calc_id_pos - calculate id_pos.
 158  * @sample_type: sample type
 159  *
 160  * This function returns the position of the event id (PERF_SAMPLE_ID or
 161  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 162  * perf_record_sample.
 163  */
 164 static int __perf_evsel__calc_id_pos(u64 sample_type)
 165 {
 166         int idx = 0;
 167
 168         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 169                 return 0;
 170
 171         if (!(sample_type & PERF_SAMPLE_ID))
 172                 return -1;
 173
 174         if (sample_type & PERF_SAMPLE_IP)
 175                 idx += 1;
 176
 177         if (sample_type & PERF_SAMPLE_TID)
 178                 idx += 1;
 179
 180         if (sample_type & PERF_SAMPLE_TIME)
 181                 idx += 1;
 182
 183         if (sample_type & PERF_SAMPLE_ADDR)
 184                 idx += 1;
 185
 186         return idx;
 187 }
 188
 189 /**
 190  * __perf_evsel__calc_is_pos - calculate is_pos.
 191  * @sample_type: sample type
 192  *
 193  * This function returns the position (counting backwards) of the event id
 194  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 195  * sample_id_all is used there is an id sample appended to non-sample events.
 196  */
 197 static int __perf_evsel__calc_is_pos(u64 sample_type)
 198 {
 199         int idx = 1;
 200
 201         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 202                 return 1;
 203
 204         if (!(sample_type & PERF_SAMPLE_ID))
 205                 return -1;
 206
 207         if (sample_type & PERF_SAMPLE_CPU)
 208                 idx += 1;
 209
 210         if (sample_type & PERF_SAMPLE_STREAM_ID)
 211                 idx += 1;
 212
 213         return idx;
 214 }
 215
 216 void evsel__calc_id_pos(struct evsel *evsel)
 217 {
 218         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 219         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 220 }
 221
 222 void __evsel__set_sample_bit(struct evsel *evsel,
 223                                   enum perf_event_sample_format bit)
 224 {
 225         if (!(evsel->core.attr.sample_type & bit)) {
 226                 evsel->core.attr.sample_type |= bit;
 227                 evsel->sample_size += sizeof(u64);
 228                 evsel__calc_id_pos(evsel);
 229         }
 230 }
 231
 232 void __evsel__reset_sample_bit(struct evsel *evsel,
 233                                     enum perf_event_sample_format bit)
 234 {
 235         if (evsel->core.attr.sample_type & bit) {
 236                 evsel->core.attr.sample_type &= ~bit;
 237                 evsel->sample_size -= sizeof(u64);
 238                 evsel__calc_id_pos(evsel);
 239         }
 240 }
 241
 242 void evsel__set_sample_id(struct evsel *evsel,
 243                                bool can_sample_identifier)
 244 {
 245         if (can_sample_identifier) {
 246                 evsel__reset_sample_bit(evsel, ID);
 247                 evsel__set_sample_bit(evsel, IDENTIFIER);
 248         } else {
 249                 evsel__set_sample_bit(evsel, ID);
 250         }
 251         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 252 }
 253
 254 /**
 255  * evsel__is_function_event - Return whether given evsel is a function
 256  * trace event
 257  *
 258  * @evsel - evsel selector to be tested
 259  *
 260  * Return %true if event is function trace event
 261  */
 262 bool evsel__is_function_event(struct evsel *evsel)
 263 {
 264 #define FUNCTION_EVENT "ftrace:function"
 265
 266         return evsel->name &&
 267                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 268
 269 #undef FUNCTION_EVENT
 270 }
 271
 272 void evsel__init(struct evsel *evsel,
 273                  struct perf_event_attr *attr, int idx)
 274 {
 275         perf_evsel__init(&evsel->core, attr, idx);
 276         evsel->tracking    = !idx;
 277         evsel->unit        = strdup("");
 278         evsel->scale       = 1.0;
 279         evsel->max_events  = ULONG_MAX;
 280         evsel->evlist      = NULL;
 281         evsel->bpf_obj     = NULL;
 282         evsel->bpf_fd      = -1;
 283         INIT_LIST_HEAD(&evsel->config_terms);
 284         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 285         perf_evsel__object.init(evsel);
 286         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 287         evsel__calc_id_pos(evsel);
 288         evsel->cmdline_group_boundary = false;
 289         evsel->metric_events = NULL;
 290         evsel->per_pkg_mask  = NULL;
 291         evsel->collect_stat  = false;
 292         evsel->pmu_name      = NULL;
 293 }
 294
 295 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 296 {
 297         struct evsel *evsel = zalloc(perf_evsel__object.size);
 298
 299         if (!evsel)
 300                 return NULL;
 301         evsel__init(evsel, attr, idx);
 302
 303         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 304                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 305                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 306                 evsel->core.attr.sample_period = 1;
 307         }
 308
 309         if (evsel__is_clock(evsel)) {
 310                 free((char *)evsel->unit);
 311                 evsel->unit = strdup("msec");
 312                 evsel->scale = 1e-6;
 313         }
 314
 315         return evsel;
 316 }
 317
 318 static bool perf_event_can_profile_kernel(void)
 319 {
 320         return perf_event_paranoid_check(1);
 321 }
 322
 323 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 324 {
 325         struct perf_event_attr attr = {
 326                 .type   = type,
 327                 .config = config,
 328                 .exclude_kernel = !perf_event_can_profile_kernel(),
 329         };
 330         struct evsel *evsel;
 331
 332         event_attr_init(&attr);
 333
 334         /*
 335          * Now let the usual logic to set up the perf_event_attr defaults
 336          * to kick in when we return and before perf_evsel__open() is called.
 337          */
 338         evsel = evsel__new(&attr);
 339         if (evsel == NULL)
 340                 goto out;
 341
 342         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 343
 344         evsel->precise_max = true;
 345
 346         /* use asprintf() because free(evsel) assumes name is allocated */
 347         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 348                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 349                      attr.exclude_kernel ? "u" : "",
 350                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 351                 goto error_free;
 352 out:
 353         return evsel;
 354 error_free:
 355         evsel__delete(evsel);
 356         evsel = NULL;
 357         goto out;
 358 }
 359
 360 int copy_config_terms(struct list_head *dst, struct list_head *src)
 361 {
 362         struct evsel_config_term *pos, *tmp;
 363
 364         list_for_each_entry(pos, src, list) {
 365                 tmp = malloc(sizeof(*tmp));
 366                 if (tmp == NULL)
 367                         return -ENOMEM;
 368
 369                 *tmp = *pos;
 370                 if (tmp->free_str) {
 371                         tmp->val.str = strdup(pos->val.str);
 372                         if (tmp->val.str == NULL) {
 373                                 free(tmp);
 374                                 return -ENOMEM;
 375                         }
 376                 }
 377                 list_add_tail(&tmp->list, dst);
 378         }
 379         return 0;
 380 }
 381
 382 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 383 {
 384         return copy_config_terms(&dst->config_terms, &src->config_terms);
 385 }
 386
 387 /**
 388  * evsel__clone - create a new evsel copied from @orig
 389  * @orig: original evsel
 390  *
 391  * The assumption is that @orig is not configured nor opened yet.
 392  * So we only care about the attributes that can be set while it's parsed.
 393  */
 394 struct evsel *evsel__clone(struct evsel *orig)
 395 {
 396         struct evsel *evsel;
 397
 398         BUG_ON(orig->core.fd);
 399         BUG_ON(orig->counts);
 400         BUG_ON(orig->priv);
 401         BUG_ON(orig->per_pkg_mask);
 402
 403         /* cannot handle BPF objects for now */
 404         if (orig->bpf_obj)
 405                 return NULL;
 406
 407         evsel = evsel__new(&orig->core.attr);
 408         if (evsel == NULL)
 409                 return NULL;
 410
 411         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 412         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 413         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 414         evsel->core.nr_members = orig->core.nr_members;
 415         evsel->core.system_wide = orig->core.system_wide;
 416         evsel->core.requires_cpu = orig->core.requires_cpu;
 417
 418         if (orig->name) {
 419                 evsel->name = strdup(orig->name);
 420                 if (evsel->name == NULL)
 421                         goto out_err;
 422         }
 423         if (orig->group_name) {
 424                 evsel->group_name = strdup(orig->group_name);
 425                 if (evsel->group_name == NULL)
 426                         goto out_err;
 427         }
 428         if (orig->pmu_name) {
 429                 evsel->pmu_name = strdup(orig->pmu_name);
 430                 if (evsel->pmu_name == NULL)
 431                         goto out_err;
 432         }
 433         if (orig->filter) {
 434                 evsel->filter = strdup(orig->filter);
 435                 if (evsel->filter == NULL)
 436                         goto out_err;
 437         }
 438         if (orig->metric_id) {
 439                 evsel->metric_id = strdup(orig->metric_id);
 440                 if (evsel->metric_id == NULL)
 441                         goto out_err;
 442         }
 443         evsel->cgrp = cgroup__get(orig->cgrp);
 444 #ifdef HAVE_LIBTRACEEVENT
 445         evsel->tp_format = orig->tp_format;
 446 #endif
 447         evsel->handler = orig->handler;
 448         evsel->core.leader = orig->core.leader;
 449
 450         evsel->max_events = orig->max_events;
 451         evsel->tool_event = orig->tool_event;
 452         free((char *)evsel->unit);
 453         evsel->unit = strdup(orig->unit);
 454         if (evsel->unit == NULL)
 455                 goto out_err;
 456
 457         evsel->scale = orig->scale;
 458         evsel->snapshot = orig->snapshot;
 459         evsel->per_pkg = orig->per_pkg;
 460         evsel->percore = orig->percore;
 461         evsel->precise_max = orig->precise_max;
 462         evsel->is_libpfm_event = orig->is_libpfm_event;
 463
 464         evsel->exclude_GH = orig->exclude_GH;
 465         evsel->sample_read = orig->sample_read;
 466         evsel->auto_merge_stats = orig->auto_merge_stats;
 467         evsel->collect_stat = orig->collect_stat;
 468         evsel->weak_group = orig->weak_group;
 469         evsel->use_config_name = orig->use_config_name;
 470         evsel->pmu = orig->pmu;
 471
 472         if (evsel__copy_config_terms(evsel, orig) < 0)
 473                 goto out_err;
 474
 475         return evsel;
 476
 477 out_err:
 478         evsel__delete(evsel);
 479         return NULL;
 480 }
 481
 482 /*
 483  * Returns pointer with encoded error via <linux/err.h> interface.
 484  */
 485 #ifdef HAVE_LIBTRACEEVENT
 486 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 487 {
 488         struct evsel *evsel = zalloc(perf_evsel__object.size);
 489         int err = -ENOMEM;
 490
 491         if (evsel == NULL) {
 492                 goto out_err;
 493         } else {
 494                 struct perf_event_attr attr = {
 495                         .type          = PERF_TYPE_TRACEPOINT,
 496                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 497                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 498                 };
 499
 500                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 501                         goto out_free;
 502
 503                 evsel->tp_format = trace_event__tp_format(sys, name);
 504                 if (IS_ERR(evsel->tp_format)) {
 505                         err = PTR_ERR(evsel->tp_format);
 506                         goto out_free;
 507                 }
 508
 509                 event_attr_init(&attr);
 510                 attr.config = evsel->tp_format->id;
 511                 attr.sample_period = 1;
 512                 evsel__init(evsel, &attr, idx);
 513         }
 514
 515         return evsel;
 516
 517 out_free:
 518         zfree(&evsel->name);
 519         free(evsel);
 520 out_err:
 521         return ERR_PTR(err);
 522 }
 523 #endif
 524
 525 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 526         "cycles",
 527         "instructions",
 528         "cache-references",
 529         "cache-misses",
 530         "branches",
 531         "branch-misses",
 532         "bus-cycles",
 533         "stalled-cycles-frontend",
 534         "stalled-cycles-backend",
 535         "ref-cycles",
 536 };
 537
 538 char *evsel__bpf_counter_events;
 539
 540 bool evsel__match_bpf_counter_events(const char *name)
 541 {
 542         int name_len;
 543         bool match;
 544         char *ptr;
 545
 546         if (!evsel__bpf_counter_events)
 547                 return false;
 548
 549         ptr = strstr(evsel__bpf_counter_events, name);
 550         name_len = strlen(name);
 551
 552         /* check name matches a full token in evsel__bpf_counter_events */
 553         match = (ptr != NULL) &&
 554                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 555                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 556
 557         return match;
 558 }
 559
 560 static const char *__evsel__hw_name(u64 config)
 561 {
 562         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 563                 return evsel__hw_names[config];
 564
 565         return "unknown-hardware";
 566 }
 567
 568 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 569 {
 570         int colon = 0, r = 0;
 571         struct perf_event_attr *attr = &evsel->core.attr;
 572         bool exclude_guest_default = false;
 573
 574 #define MOD_PRINT(context, mod) do {                                    \
 575                 if (!attr->exclude_##context) {                         \
 576                         if (!colon) colon = ++r;                        \
 577                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 578                 } } while(0)
 579
 580         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 581                 MOD_PRINT(kernel, 'k');
 582                 MOD_PRINT(user, 'u');
 583                 MOD_PRINT(hv, 'h');
 584                 exclude_guest_default = true;
 585         }
 586
 587         if (attr->precise_ip) {
 588                 if (!colon)
 589                         colon = ++r;
 590                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 591                 exclude_guest_default = true;
 592         }
 593
 594         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 595                 MOD_PRINT(host, 'H');
 596                 MOD_PRINT(guest, 'G');
 597         }
 598 #undef MOD_PRINT
 599         if (colon)
 600                 bf[colon - 1] = ':';
 601         return r;
 602 }
 603
 604 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 605 {
 606         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 607 }
 608
 609 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 610 {
 611         int r = arch_evsel__hw_name(evsel, bf, size);
 612         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 613 }
 614
 615 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 616         "cpu-clock",
 617         "task-clock",
 618         "page-faults",
 619         "context-switches",
 620         "cpu-migrations",
 621         "minor-faults",
 622         "major-faults",
 623         "alignment-faults",
 624         "emulation-faults",
 625         "dummy",
 626 };
 627
 628 static const char *__evsel__sw_name(u64 config)
 629 {
 630         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 631                 return evsel__sw_names[config];
 632         return "unknown-software";
 633 }
 634
 635 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 636 {
 637         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 638         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 639 }
 640
 641 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 642 {
 643         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 644 }
 645
 646 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 647 {
 648         int r;
 649
 650         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 651
 652         if (type & HW_BREAKPOINT_R)
 653                 r += scnprintf(bf + r, size - r, "r");
 654
 655         if (type & HW_BREAKPOINT_W)
 656                 r += scnprintf(bf + r, size - r, "w");
 657
 658         if (type & HW_BREAKPOINT_X)
 659                 r += scnprintf(bf + r, size - r, "x");
 660
 661         return r;
 662 }
 663
 664 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 665 {
 666         struct perf_event_attr *attr = &evsel->core.attr;
 667         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 668         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 669 }
 670
 671 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 672  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 673  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 674  { "LLC",       "L2",                                                   },
 675  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 676  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 677  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 678  { "node",                                                              },
 679 };
 680
 681 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 682  { "load",      "loads",        "read",                                 },
 683  { "store",     "stores",       "write",                                },
 684  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 685 };
 686
 687 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 688  { "refs",      "Reference",    "ops",          "access",               },
 689  { "misses",    "miss",                                                 },
 690 };
 691
 692 #define C(x)            PERF_COUNT_HW_CACHE_##x
 693 #define CACHE_READ      (1 << C(OP_READ))
 694 #define CACHE_WRITE     (1 << C(OP_WRITE))
 695 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 696 #define COP(x)          (1 << x)
 697
 698 /*
 699  * cache operation stat
 700  * L1I : Read and prefetch only
 701  * ITLB and BPU : Read-only
 702  */
 703 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 704  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 705  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 706  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 707  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 708  [C(ITLB)]      = (CACHE_READ),
 709  [C(BPU)]       = (CACHE_READ),
 710  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 711 };
 712
 713 bool evsel__is_cache_op_valid(u8 type, u8 op)
 714 {
 715         if (evsel__hw_cache_stat[type] & COP(op))
 716                 return true;    /* valid */
 717         else
 718                 return false;   /* invalid */
 719 }
 720
 721 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 722 {
 723         if (result) {
 724                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 725                                  evsel__hw_cache_op[op][0],
 726                                  evsel__hw_cache_result[result][0]);
 727         }
 728
 729         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 730                          evsel__hw_cache_op[op][1]);
 731 }
 732
 733 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 734 {
 735         u8 op, result, type = (config >>  0) & 0xff;
 736         const char *err = "unknown-ext-hardware-cache-type";
 737
 738         if (type >= PERF_COUNT_HW_CACHE_MAX)
 739                 goto out_err;
 740
 741         op = (config >>  8) & 0xff;
 742         err = "unknown-ext-hardware-cache-op";
 743         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 744                 goto out_err;
 745
 746         result = (config >> 16) & 0xff;
 747         err = "unknown-ext-hardware-cache-result";
 748         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 749                 goto out_err;
 750
 751         err = "invalid-cache";
 752         if (!evsel__is_cache_op_valid(type, op))
 753                 goto out_err;
 754
 755         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 756 out_err:
 757         return scnprintf(bf, size, "%s", err);
 758 }
 759
 760 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 761 {
 762         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 763         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 764 }
 765
 766 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 767 {
 768         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 769         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 770 }
 771
 772 const char *evsel__name(struct evsel *evsel)
 773 {
 774         char bf[128];
 775
 776         if (!evsel)
 777                 goto out_unknown;
 778
 779         if (evsel->name)
 780                 return evsel->name;
 781
 782         switch (evsel->core.attr.type) {
 783         case PERF_TYPE_RAW:
 784                 evsel__raw_name(evsel, bf, sizeof(bf));
 785                 break;
 786
 787         case PERF_TYPE_HARDWARE:
 788                 evsel__hw_name(evsel, bf, sizeof(bf));
 789                 break;
 790
 791         case PERF_TYPE_HW_CACHE:
 792                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 793                 break;
 794
 795         case PERF_TYPE_SOFTWARE:
 796                 if (evsel__is_tool(evsel))
 797                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 798                 else
 799                         evsel__sw_name(evsel, bf, sizeof(bf));
 800                 break;
 801
 802         case PERF_TYPE_TRACEPOINT:
 803                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 804                 break;
 805
 806         case PERF_TYPE_BREAKPOINT:
 807                 evsel__bp_name(evsel, bf, sizeof(bf));
 808                 break;
 809
 810         default:
 811                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 812                           evsel->core.attr.type);
 813                 break;
 814         }
 815
 816         evsel->name = strdup(bf);
 817
 818         if (evsel->name)
 819                 return evsel->name;
 820 out_unknown:
 821         return "unknown";
 822 }
 823
 824 bool evsel__name_is(struct evsel *evsel, const char *name)
 825 {
 826         return !strcmp(evsel__name(evsel), name);
 827 }
 828
 829 const char *evsel__group_pmu_name(const struct evsel *evsel)
 830 {
 831         const struct evsel *leader;
 832
 833         /* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
 834         if (evsel->pmu_name)
 835                 return evsel->pmu_name;
 836         /*
 837          * Software events may be in a group with other uncore PMU events. Use
 838          * the pmu_name of the group leader to avoid breaking the software event
 839          * out of the group.
 840          *
 841          * Aux event leaders, like intel_pt, expect a group with events from
 842          * other PMUs, so substitute the AUX event's PMU in this case.
 843          */
 844         leader  = evsel__leader(evsel);
 845         if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
 846             leader->pmu_name) {
 847                 return leader->pmu_name;
 848         }
 849
 850         return "cpu";
 851 }
 852
 853 const char *evsel__metric_id(const struct evsel *evsel)
 854 {
 855         if (evsel->metric_id)
 856                 return evsel->metric_id;
 857
 858         if (evsel__is_tool(evsel))
 859                 return perf_tool_event__to_str(evsel->tool_event);
 860
 861         return "unknown";
 862 }
 863
 864 const char *evsel__group_name(struct evsel *evsel)
 865 {
 866         return evsel->group_name ?: "anon group";
 867 }
 868
 869 /*
 870  * Returns the group details for the specified leader,
 871  * with following rules.
 872  *
 873  *  For record -e '{cycles,instructions}'
 874  *    'anon group { cycles:u, instructions:u }'
 875  *
 876  *  For record -e 'cycles,instructions' and report --group
 877  *    'cycles:u, instructions:u'
 878  */
 879 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 880 {
 881         int ret = 0;
 882         struct evsel *pos;
 883         const char *group_name = evsel__group_name(evsel);
 884
 885         if (!evsel->forced_leader)
 886                 ret = scnprintf(buf, size, "%s { ", group_name);
 887
 888         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 889
 890         for_each_group_member(pos, evsel)
 891                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 892
 893         if (!evsel->forced_leader)
 894                 ret += scnprintf(buf + ret, size - ret, " }");
 895
 896         return ret;
 897 }
 898
 899 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 900                                       struct callchain_param *param)
 901 {
 902         bool function = evsel__is_function_event(evsel);
 903         struct perf_event_attr *attr = &evsel->core.attr;
 904
 905         evsel__set_sample_bit(evsel, CALLCHAIN);
 906
 907         attr->sample_max_stack = param->max_stack;
 908
 909         if (opts->kernel_callchains)
 910                 attr->exclude_callchain_user = 1;
 911         if (opts->user_callchains)
 912                 attr->exclude_callchain_kernel = 1;
 913         if (param->record_mode == CALLCHAIN_LBR) {
 914                 if (!opts->branch_stack) {
 915                         if (attr->exclude_user) {
 916                                 pr_warning("LBR callstack option is only available "
 917                                            "to get user callchain information. "
 918                                            "Falling back to framepointers.\n");
 919                         } else {
 920                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 921                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 922                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 923                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 924                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 925                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 926                         }
 927                 } else
 928                          pr_warning("Cannot use LBR callstack with branch stack. "
 929                                     "Falling back to framepointers.\n");
 930         }
 931
 932         if (param->record_mode == CALLCHAIN_DWARF) {
 933                 if (!function) {
 934                         evsel__set_sample_bit(evsel, REGS_USER);
 935                         evsel__set_sample_bit(evsel, STACK_USER);
 936                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 937                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 938                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 939                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 940                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 941                         } else {
 942                                 attr->sample_regs_user |= arch__user_reg_mask();
 943                         }
 944                         attr->sample_stack_user = param->dump_size;
 945                         attr->exclude_callchain_user = 1;
 946                 } else {
 947                         pr_info("Cannot use DWARF unwind for function trace event,"
 948                                 " falling back to framepointers.\n");
 949                 }
 950         }
 951
 952         if (function) {
 953                 pr_info("Disabling user space callchains for function trace event.\n");
 954                 attr->exclude_callchain_user = 1;
 955         }
 956 }
 957
 958 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 959                              struct callchain_param *param)
 960 {
 961         if (param->enabled)
 962                 return __evsel__config_callchain(evsel, opts, param);
 963 }
 964
 965 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 966 {
 967         struct perf_event_attr *attr = &evsel->core.attr;
 968
 969         evsel__reset_sample_bit(evsel, CALLCHAIN);
 970         if (param->record_mode == CALLCHAIN_LBR) {
 971                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 972                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 973                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 974                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 975         }
 976         if (param->record_mode == CALLCHAIN_DWARF) {
 977                 evsel__reset_sample_bit(evsel, REGS_USER);
 978                 evsel__reset_sample_bit(evsel, STACK_USER);
 979         }
 980 }
 981
 982 static void evsel__apply_config_terms(struct evsel *evsel,
 983                                       struct record_opts *opts, bool track)
 984 {
 985         struct evsel_config_term *term;
 986         struct list_head *config_terms = &evsel->config_terms;
 987         struct perf_event_attr *attr = &evsel->core.attr;
 988         /* callgraph default */
 989         struct callchain_param param = {
 990                 .record_mode = callchain_param.record_mode,
 991         };
 992         u32 dump_size = 0;
 993         int max_stack = 0;
 994         const char *callgraph_buf = NULL;
 995
 996         list_for_each_entry(term, config_terms, list) {
 997                 switch (term->type) {
 998                 case EVSEL__CONFIG_TERM_PERIOD:
 999                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
1000                                 attr->sample_period = term->val.period;
1001                                 attr->freq = 0;
1002                                 evsel__reset_sample_bit(evsel, PERIOD);
1003                         }
1004                         break;
1005                 case EVSEL__CONFIG_TERM_FREQ:
1006                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
1007                                 attr->sample_freq = term->val.freq;
1008                                 attr->freq = 1;
1009                                 evsel__set_sample_bit(evsel, PERIOD);
1010                         }
1011                         break;
1012                 case EVSEL__CONFIG_TERM_TIME:
1013                         if (term->val.time)
1014                                 evsel__set_sample_bit(evsel, TIME);
1015                         else
1016                                 evsel__reset_sample_bit(evsel, TIME);
1017                         break;
1018                 case EVSEL__CONFIG_TERM_CALLGRAPH:
1019                         callgraph_buf = term->val.str;
1020                         break;
1021                 case EVSEL__CONFIG_TERM_BRANCH:
1022                         if (term->val.str && strcmp(term->val.str, "no")) {
1023                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1024                                 parse_branch_str(term->val.str,
1025                                                  &attr->branch_sample_type);
1026                         } else
1027                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1028                         break;
1029                 case EVSEL__CONFIG_TERM_STACK_USER:
1030                         dump_size = term->val.stack_user;
1031                         break;
1032                 case EVSEL__CONFIG_TERM_MAX_STACK:
1033                         max_stack = term->val.max_stack;
1034                         break;
1035                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1036                         evsel->max_events = term->val.max_events;
1037                         break;
1038                 case EVSEL__CONFIG_TERM_INHERIT:
1039                         /*
1040                          * attr->inherit should has already been set by
1041                          * evsel__config. If user explicitly set
1042                          * inherit using config terms, override global
1043                          * opt->no_inherit setting.
1044                          */
1045                         attr->inherit = term->val.inherit ? 1 : 0;
1046                         break;
1047                 case EVSEL__CONFIG_TERM_OVERWRITE:
1048                         attr->write_backward = term->val.overwrite ? 1 : 0;
1049                         break;
1050                 case EVSEL__CONFIG_TERM_DRV_CFG:
1051                         break;
1052                 case EVSEL__CONFIG_TERM_PERCORE:
1053                         break;
1054                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1055                         attr->aux_output = term->val.aux_output ? 1 : 0;
1056                         break;
1057                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1058                         /* Already applied by auxtrace */
1059                         break;
1060                 case EVSEL__CONFIG_TERM_CFG_CHG:
1061                         break;
1062                 default:
1063                         break;
1064                 }
1065         }
1066
1067         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1068         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1069                 bool sample_address = false;
1070
1071                 if (max_stack) {
1072                         param.max_stack = max_stack;
1073                         if (callgraph_buf == NULL)
1074                                 callgraph_buf = "fp";
1075                 }
1076
1077                 /* parse callgraph parameters */
1078                 if (callgraph_buf != NULL) {
1079                         if (!strcmp(callgraph_buf, "no")) {
1080                                 param.enabled = false;
1081                                 param.record_mode = CALLCHAIN_NONE;
1082                         } else {
1083                                 param.enabled = true;
1084                                 if (parse_callchain_record(callgraph_buf, &param)) {
1085                                         pr_err("per-event callgraph setting for %s failed. "
1086                                                "Apply callgraph global setting for it\n",
1087                                                evsel->name);
1088                                         return;
1089                                 }
1090                                 if (param.record_mode == CALLCHAIN_DWARF)
1091                                         sample_address = true;
1092                         }
1093                 }
1094                 if (dump_size > 0) {
1095                         dump_size = round_up(dump_size, sizeof(u64));
1096                         param.dump_size = dump_size;
1097                 }
1098
1099                 /* If global callgraph set, clear it */
1100                 if (callchain_param.enabled)
1101                         evsel__reset_callgraph(evsel, &callchain_param);
1102
1103                 /* set perf-event callgraph */
1104                 if (param.enabled) {
1105                         if (sample_address) {
1106                                 evsel__set_sample_bit(evsel, ADDR);
1107                                 evsel__set_sample_bit(evsel, DATA_SRC);
1108                                 evsel->core.attr.mmap_data = track;
1109                         }
1110                         evsel__config_callchain(evsel, opts, &param);
1111                 }
1112         }
1113 }
1114
1115 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1116 {
1117         struct evsel_config_term *term, *found_term = NULL;
1118
1119         list_for_each_entry(term, &evsel->config_terms, list) {
1120                 if (term->type == type)
1121                         found_term = term;
1122         }
1123
1124         return found_term;
1125 }
1126
1127 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1128 {
1129         evsel__set_sample_bit(evsel, WEIGHT);
1130 }
1131
1132 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1133 {
1134 }
1135
1136 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1137                                     struct perf_event_attr *attr __maybe_unused)
1138 {
1139 }
1140
1141 static void evsel__set_default_freq_period(struct record_opts *opts,
1142                                            struct perf_event_attr *attr)
1143 {
1144         if (opts->freq) {
1145                 attr->freq = 1;
1146                 attr->sample_freq = opts->freq;
1147         } else {
1148                 attr->sample_period = opts->default_interval;
1149         }
1150 }
1151
1152 static bool evsel__is_offcpu_event(struct evsel *evsel)
1153 {
1154         return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT);
1155 }
1156
1157 /*
1158  * The enable_on_exec/disabled value strategy:
1159  *
1160  *  1) For any type of traced program:
1161  *    - all independent events and group leaders are disabled
1162  *    - all group members are enabled
1163  *
1164  *     Group members are ruled by group leaders. They need to
1165  *     be enabled, because the group scheduling relies on that.
1166  *
1167  *  2) For traced programs executed by perf:
1168  *     - all independent events and group leaders have
1169  *       enable_on_exec set
1170  *     - we don't specifically enable or disable any event during
1171  *       the record command
1172  *
1173  *     Independent events and group leaders are initially disabled
1174  *     and get enabled by exec. Group members are ruled by group
1175  *     leaders as stated in 1).
1176  *
1177  *  3) For traced programs attached by perf (pid/tid):
1178  *     - we specifically enable or disable all events during
1179  *       the record command
1180  *
1181  *     When attaching events to already running traced we
1182  *     enable/disable events specifically, as there's no
1183  *     initial traced exec call.
1184  */
1185 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1186                    struct callchain_param *callchain)
1187 {
1188         struct evsel *leader = evsel__leader(evsel);
1189         struct perf_event_attr *attr = &evsel->core.attr;
1190         int track = evsel->tracking;
1191         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1192
1193         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1194         attr->inherit       = !opts->no_inherit;
1195         attr->write_backward = opts->overwrite ? 1 : 0;
1196         attr->read_format   = PERF_FORMAT_LOST;
1197
1198         evsel__set_sample_bit(evsel, IP);
1199         evsel__set_sample_bit(evsel, TID);
1200
1201         if (evsel->sample_read) {
1202                 evsel__set_sample_bit(evsel, READ);
1203
1204                 /*
1205                  * We need ID even in case of single event, because
1206                  * PERF_SAMPLE_READ process ID specific data.
1207                  */
1208                 evsel__set_sample_id(evsel, false);
1209
1210                 /*
1211                  * Apply group format only if we belong to group
1212                  * with more than one members.
1213                  */
1214                 if (leader->core.nr_members > 1) {
1215                         attr->read_format |= PERF_FORMAT_GROUP;
1216                         attr->inherit = 0;
1217                 }
1218         }
1219
1220         /*
1221          * We default some events to have a default interval. But keep
1222          * it a weak assumption overridable by the user.
1223          */
1224         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1225             (!evsel->is_libpfm_event && (!attr->sample_period ||
1226                                          opts->user_freq != UINT_MAX ||
1227                                          opts->user_interval != ULLONG_MAX)))
1228                 evsel__set_default_freq_period(opts, attr);
1229
1230         /*
1231          * If attr->freq was set (here or earlier), ask for period
1232          * to be sampled.
1233          */
1234         if (attr->freq)
1235                 evsel__set_sample_bit(evsel, PERIOD);
1236
1237         if (opts->no_samples)
1238                 attr->sample_freq = 0;
1239
1240         if (opts->inherit_stat) {
1241                 evsel->core.attr.read_format |=
1242                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1243                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1244                         PERF_FORMAT_ID;
1245                 attr->inherit_stat = 1;
1246         }
1247
1248         if (opts->sample_address) {
1249                 evsel__set_sample_bit(evsel, ADDR);
1250                 attr->mmap_data = track;
1251         }
1252
1253         /*
1254          * We don't allow user space callchains for  function trace
1255          * event, due to issues with page faults while tracing page
1256          * fault handler and its overall trickiness nature.
1257          */
1258         if (evsel__is_function_event(evsel))
1259                 evsel->core.attr.exclude_callchain_user = 1;
1260
1261         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1262                 evsel__config_callchain(evsel, opts, callchain);
1263
1264         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1265             !evsel__is_dummy_event(evsel)) {
1266                 attr->sample_regs_intr = opts->sample_intr_regs;
1267                 evsel__set_sample_bit(evsel, REGS_INTR);
1268         }
1269
1270         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1271             !evsel__is_dummy_event(evsel)) {
1272                 attr->sample_regs_user |= opts->sample_user_regs;
1273                 evsel__set_sample_bit(evsel, REGS_USER);
1274         }
1275
1276         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1277                 evsel__set_sample_bit(evsel, CPU);
1278
1279         /*
1280          * When the user explicitly disabled time don't force it here.
1281          */
1282         if (opts->sample_time &&
1283             (!perf_missing_features.sample_id_all &&
1284             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1285              opts->sample_time_set)))
1286                 evsel__set_sample_bit(evsel, TIME);
1287
1288         if (opts->raw_samples && !evsel->no_aux_samples) {
1289                 evsel__set_sample_bit(evsel, TIME);
1290                 evsel__set_sample_bit(evsel, RAW);
1291                 evsel__set_sample_bit(evsel, CPU);
1292         }
1293
1294         if (opts->sample_address)
1295                 evsel__set_sample_bit(evsel, DATA_SRC);
1296
1297         if (opts->sample_phys_addr)
1298                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1299
1300         if (opts->no_buffering) {
1301                 attr->watermark = 0;
1302                 attr->wakeup_events = 1;
1303         }
1304         if (opts->branch_stack && !evsel->no_aux_samples) {
1305                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1306                 attr->branch_sample_type = opts->branch_stack;
1307         }
1308
1309         if (opts->sample_weight)
1310                 arch_evsel__set_sample_weight(evsel);
1311
1312         attr->task     = track;
1313         attr->mmap     = track;
1314         attr->mmap2    = track && !perf_missing_features.mmap2;
1315         attr->comm     = track;
1316         attr->build_id = track && opts->build_id;
1317
1318         /*
1319          * ksymbol is tracked separately with text poke because it needs to be
1320          * system wide and enabled immediately.
1321          */
1322         if (!opts->text_poke)
1323                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1324         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1325
1326         if (opts->record_namespaces)
1327                 attr->namespaces  = track;
1328
1329         if (opts->record_cgroup) {
1330                 attr->cgroup = track && !perf_missing_features.cgroup;
1331                 evsel__set_sample_bit(evsel, CGROUP);
1332         }
1333
1334         if (opts->sample_data_page_size)
1335                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1336
1337         if (opts->sample_code_page_size)
1338                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1339
1340         if (opts->record_switch_events)
1341                 attr->context_switch = track;
1342
1343         if (opts->sample_transaction)
1344                 evsel__set_sample_bit(evsel, TRANSACTION);
1345
1346         if (opts->running_time) {
1347                 evsel->core.attr.read_format |=
1348                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1349                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1350         }
1351
1352         /*
1353          * XXX see the function comment above
1354          *
1355          * Disabling only independent events or group leaders,
1356          * keeping group members enabled.
1357          */
1358         if (evsel__is_group_leader(evsel))
1359                 attr->disabled = 1;
1360
1361         /*
1362          * Setting enable_on_exec for independent events and
1363          * group leaders for traced executed by perf.
1364          */
1365         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1366             !opts->target.initial_delay)
1367                 attr->enable_on_exec = 1;
1368
1369         if (evsel->immediate) {
1370                 attr->disabled = 0;
1371                 attr->enable_on_exec = 0;
1372         }
1373
1374         clockid = opts->clockid;
1375         if (opts->use_clockid) {
1376                 attr->use_clockid = 1;
1377                 attr->clockid = opts->clockid;
1378         }
1379
1380         if (evsel->precise_max)
1381                 attr->precise_ip = 3;
1382
1383         if (opts->all_user) {
1384                 attr->exclude_kernel = 1;
1385                 attr->exclude_user   = 0;
1386         }
1387
1388         if (opts->all_kernel) {
1389                 attr->exclude_kernel = 0;
1390                 attr->exclude_user   = 1;
1391         }
1392
1393         if (evsel->core.own_cpus || evsel->unit)
1394                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1395
1396         /*
1397          * Apply event specific term settings,
1398          * it overloads any global configuration.
1399          */
1400         evsel__apply_config_terms(evsel, opts, track);
1401
1402         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1403
1404         /* The --period option takes the precedence. */
1405         if (opts->period_set) {
1406                 if (opts->period)
1407                         evsel__set_sample_bit(evsel, PERIOD);
1408                 else
1409                         evsel__reset_sample_bit(evsel, PERIOD);
1410         }
1411
1412         /*
1413          * A dummy event never triggers any actual counter and therefore
1414          * cannot be used with branch_stack.
1415          *
1416          * For initial_delay, a dummy event is added implicitly.
1417          * The software event will trigger -EOPNOTSUPP error out,
1418          * if BRANCH_STACK bit is set.
1419          */
1420         if (evsel__is_dummy_event(evsel))
1421                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1422
1423         if (evsel__is_offcpu_event(evsel))
1424                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1425
1426         arch__post_evsel_config(evsel, attr);
1427 }
1428
1429 int evsel__set_filter(struct evsel *evsel, const char *filter)
1430 {
1431         char *new_filter = strdup(filter);
1432
1433         if (new_filter != NULL) {
1434                 free(evsel->filter);
1435                 evsel->filter = new_filter;
1436                 return 0;
1437         }
1438
1439         return -1;
1440 }
1441
1442 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1443 {
1444         char *new_filter;
1445
1446         if (evsel->filter == NULL)
1447                 return evsel__set_filter(evsel, filter);
1448
1449         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1450                 free(evsel->filter);
1451                 evsel->filter = new_filter;
1452                 return 0;
1453         }
1454
1455         return -1;
1456 }
1457
1458 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1459 {
1460         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1461 }
1462
1463 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1464 {
1465         return evsel__append_filter(evsel, "%s,%s", filter);
1466 }
1467
1468 /* Caller has to clear disabled after going through all CPUs. */
1469 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1470 {
1471         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1472 }
1473
1474 int evsel__enable(struct evsel *evsel)
1475 {
1476         int err = perf_evsel__enable(&evsel->core);
1477
1478         if (!err)
1479                 evsel->disabled = false;
1480         return err;
1481 }
1482
1483 /* Caller has to set disabled after going through all CPUs. */
1484 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1485 {
1486         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1487 }
1488
1489 int evsel__disable(struct evsel *evsel)
1490 {
1491         int err = perf_evsel__disable(&evsel->core);
1492         /*
1493          * We mark it disabled here so that tools that disable a event can
1494          * ignore events after they disable it. I.e. the ring buffer may have
1495          * already a few more events queued up before the kernel got the stop
1496          * request.
1497          */
1498         if (!err)
1499                 evsel->disabled = true;
1500
1501         return err;
1502 }
1503
1504 void free_config_terms(struct list_head *config_terms)
1505 {
1506         struct evsel_config_term *term, *h;
1507
1508         list_for_each_entry_safe(term, h, config_terms, list) {
1509                 list_del_init(&term->list);
1510                 if (term->free_str)
1511                         zfree(&term->val.str);
1512                 free(term);
1513         }
1514 }
1515
1516 static void evsel__free_config_terms(struct evsel *evsel)
1517 {
1518         free_config_terms(&evsel->config_terms);
1519 }
1520
1521 void evsel__exit(struct evsel *evsel)
1522 {
1523         assert(list_empty(&evsel->core.node));
1524         assert(evsel->evlist == NULL);
1525         bpf_counter__destroy(evsel);
1526         perf_bpf_filter__destroy(evsel);
1527         evsel__free_counts(evsel);
1528         perf_evsel__free_fd(&evsel->core);
1529         perf_evsel__free_id(&evsel->core);
1530         evsel__free_config_terms(evsel);
1531         cgroup__put(evsel->cgrp);
1532         perf_cpu_map__put(evsel->core.cpus);
1533         perf_cpu_map__put(evsel->core.own_cpus);
1534         perf_thread_map__put(evsel->core.threads);
1535         zfree(&evsel->group_name);
1536         zfree(&evsel->name);
1537         zfree(&evsel->pmu_name);
1538         zfree(&evsel->unit);
1539         zfree(&evsel->metric_id);
1540         evsel__zero_per_pkg(evsel);
1541         hashmap__free(evsel->per_pkg_mask);
1542         evsel->per_pkg_mask = NULL;
1543         zfree(&evsel->metric_events);
1544         perf_evsel__object.fini(evsel);
1545 }
1546
1547 void evsel__delete(struct evsel *evsel)
1548 {
1549         if (!evsel)
1550                 return;
1551
1552         evsel__exit(evsel);
1553         free(evsel);
1554 }
1555
1556 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1557                            struct perf_counts_values *count)
1558 {
1559         struct perf_counts_values tmp;
1560
1561         if (!evsel->prev_raw_counts)
1562                 return;
1563
1564         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1565         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1566
1567         count->val = count->val - tmp.val;
1568         count->ena = count->ena - tmp.ena;
1569         count->run = count->run - tmp.run;
1570 }
1571
1572 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1573 {
1574         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1575
1576         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1577 }
1578
1579 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1580                              u64 val, u64 ena, u64 run, u64 lost)
1581 {
1582         struct perf_counts_values *count;
1583
1584         count = perf_counts(counter->counts, cpu_map_idx, thread);
1585
1586         count->val    = val;
1587         count->ena    = ena;
1588         count->run    = run;
1589         count->lost   = lost;
1590
1591         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1592 }
1593
1594 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1595 {
1596         u64 read_format = leader->core.attr.read_format;
1597         struct sample_read_value *v;
1598         u64 nr, ena = 0, run = 0, lost = 0;
1599
1600         nr = *data++;
1601
1602         if (nr != (u64) leader->core.nr_members)
1603                 return -EINVAL;
1604
1605         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1606                 ena = *data++;
1607
1608         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1609                 run = *data++;
1610
1611         v = (void *)data;
1612         sample_read_group__for_each(v, nr, read_format) {
1613                 struct evsel *counter;
1614
1615                 counter = evlist__id2evsel(leader->evlist, v->id);
1616                 if (!counter)
1617                         return -EINVAL;
1618
1619                 if (read_format & PERF_FORMAT_LOST)
1620                         lost = v->lost;
1621
1622                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1623         }
1624
1625         return 0;
1626 }
1627
1628 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1629 {
1630         struct perf_stat_evsel *ps = leader->stats;
1631         u64 read_format = leader->core.attr.read_format;
1632         int size = perf_evsel__read_size(&leader->core);
1633         u64 *data = ps->group_data;
1634
1635         if (!(read_format & PERF_FORMAT_ID))
1636                 return -EINVAL;
1637
1638         if (!evsel__is_group_leader(leader))
1639                 return -EINVAL;
1640
1641         if (!data) {
1642                 data = zalloc(size);
1643                 if (!data)
1644                         return -ENOMEM;
1645
1646                 ps->group_data = data;
1647         }
1648
1649         if (FD(leader, cpu_map_idx, thread) < 0)
1650                 return -EINVAL;
1651
1652         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1653                 return -errno;
1654
1655         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1656 }
1657
1658 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1659 {
1660         u64 read_format = evsel->core.attr.read_format;
1661
1662         if (read_format & PERF_FORMAT_GROUP)
1663                 return evsel__read_group(evsel, cpu_map_idx, thread);
1664
1665         return evsel__read_one(evsel, cpu_map_idx, thread);
1666 }
1667
1668 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1669 {
1670         struct perf_counts_values count;
1671         size_t nv = scale ? 3 : 1;
1672
1673         if (FD(evsel, cpu_map_idx, thread) < 0)
1674                 return -EINVAL;
1675
1676         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1677                 return -ENOMEM;
1678
1679         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1680                 return -errno;
1681
1682         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1683         perf_counts_values__scale(&count, scale, NULL);
1684         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1685         return 0;
1686 }
1687
1688 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1689                                   int cpu_map_idx)
1690 {
1691         struct perf_cpu cpu;
1692
1693         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1694         return perf_cpu_map__idx(other->core.cpus, cpu);
1695 }
1696
1697 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1698 {
1699         struct evsel *leader = evsel__leader(evsel);
1700
1701         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1702             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1703                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1704         }
1705
1706         return cpu_map_idx;
1707 }
1708
1709 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1710 {
1711         struct evsel *leader = evsel__leader(evsel);
1712         int fd;
1713
1714         if (evsel__is_group_leader(evsel))
1715                 return -1;
1716
1717         /*
1718          * Leader must be already processed/open,
1719          * if not it's a bug.
1720          */
1721         BUG_ON(!leader->core.fd);
1722
1723         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1724         if (cpu_map_idx == -1)
1725                 return -1;
1726
1727         fd = FD(leader, cpu_map_idx, thread);
1728         BUG_ON(fd == -1);
1729
1730         return fd;
1731 }
1732
1733 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1734 {
1735         for (int cpu = 0; cpu < nr_cpus; cpu++)
1736                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1737                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1738 }
1739
1740 static int update_fds(struct evsel *evsel,
1741                       int nr_cpus, int cpu_map_idx,
1742                       int nr_threads, int thread_idx)
1743 {
1744         struct evsel *pos;
1745
1746         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1747                 return -EINVAL;
1748
1749         evlist__for_each_entry(evsel->evlist, pos) {
1750                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1751
1752                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1753
1754                 /*
1755                  * Since fds for next evsel has not been created,
1756                  * there is no need to iterate whole event list.
1757                  */
1758                 if (pos == evsel)
1759                         break;
1760         }
1761         return 0;
1762 }
1763
1764 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1765                                          int nr_cpus, int cpu_map_idx,
1766                                          struct perf_thread_map *threads,
1767                                          int thread, int err)
1768 {
1769         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1770
1771         if (!evsel->ignore_missing_thread)
1772                 return false;
1773
1774         /* The system wide setup does not work with threads. */
1775         if (evsel->core.system_wide)
1776                 return false;
1777
1778         /* The -ESRCH is perf event syscall errno for pid's not found. */
1779         if (err != -ESRCH)
1780                 return false;
1781
1782         /* If there's only one thread, let it fail. */
1783         if (threads->nr == 1)
1784                 return false;
1785
1786         /*
1787          * We should remove fd for missing_thread first
1788          * because thread_map__remove() will decrease threads->nr.
1789          */
1790         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1791                 return false;
1792
1793         if (thread_map__remove(threads, thread))
1794                 return false;
1795
1796         pr_warning("WARNING: Ignored open failure for pid %d\n",
1797                    ignore_pid);
1798         return true;
1799 }
1800
1801 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1802                                 void *priv __maybe_unused)
1803 {
1804         return fprintf(fp, "  %-32s %s\n", name, val);
1805 }
1806
1807 static void display_attr(struct perf_event_attr *attr)
1808 {
1809         if (verbose >= 2 || debug_peo_args) {
1810                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1811                 fprintf(stderr, "perf_event_attr:\n");
1812                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1813                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1814         }
1815 }
1816
1817 bool evsel__precise_ip_fallback(struct evsel *evsel)
1818 {
1819         /* Do not try less precise if not requested. */
1820         if (!evsel->precise_max)
1821                 return false;
1822
1823         /*
1824          * We tried all the precise_ip values, and it's
1825          * still failing, so leave it to standard fallback.
1826          */
1827         if (!evsel->core.attr.precise_ip) {
1828                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1829                 return false;
1830         }
1831
1832         if (!evsel->precise_ip_original)
1833                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1834
1835         evsel->core.attr.precise_ip--;
1836         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1837         display_attr(&evsel->core.attr);
1838         return true;
1839 }
1840
1841 static struct perf_cpu_map *empty_cpu_map;
1842 static struct perf_thread_map *empty_thread_map;
1843
1844 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1845                 struct perf_thread_map *threads)
1846 {
1847         int nthreads = perf_thread_map__nr(threads);
1848
1849         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1850             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1851                 return -EINVAL;
1852
1853         if (cpus == NULL) {
1854                 if (empty_cpu_map == NULL) {
1855                         empty_cpu_map = perf_cpu_map__dummy_new();
1856                         if (empty_cpu_map == NULL)
1857                                 return -ENOMEM;
1858                 }
1859
1860                 cpus = empty_cpu_map;
1861         }
1862
1863         if (threads == NULL) {
1864                 if (empty_thread_map == NULL) {
1865                         empty_thread_map = thread_map__new_by_tid(-1);
1866                         if (empty_thread_map == NULL)
1867                                 return -ENOMEM;
1868                 }
1869
1870                 threads = empty_thread_map;
1871         }
1872
1873         if (evsel->core.fd == NULL &&
1874             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1875                 return -ENOMEM;
1876
1877         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1878         if (evsel->cgrp)
1879                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1880
1881         return 0;
1882 }
1883
1884 static void evsel__disable_missing_features(struct evsel *evsel)
1885 {
1886         if (perf_missing_features.read_lost)
1887                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1888         if (perf_missing_features.weight_struct) {
1889                 evsel__set_sample_bit(evsel, WEIGHT);
1890                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1891         }
1892         if (perf_missing_features.clockid_wrong)
1893                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1894         if (perf_missing_features.clockid) {
1895                 evsel->core.attr.use_clockid = 0;
1896                 evsel->core.attr.clockid = 0;
1897         }
1898         if (perf_missing_features.cloexec)
1899                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1900         if (perf_missing_features.mmap2)
1901                 evsel->core.attr.mmap2 = 0;
1902         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1903                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1904         if (perf_missing_features.lbr_flags)
1905                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1906                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1907         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1908                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1909         if (perf_missing_features.ksymbol)
1910                 evsel->core.attr.ksymbol = 0;
1911         if (perf_missing_features.bpf)
1912                 evsel->core.attr.bpf_event = 0;
1913         if (perf_missing_features.branch_hw_idx)
1914                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1915         if (perf_missing_features.sample_id_all)
1916                 evsel->core.attr.sample_id_all = 0;
1917 }
1918
1919 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1920                         struct perf_thread_map *threads)
1921 {
1922         int err;
1923
1924         err = __evsel__prepare_open(evsel, cpus, threads);
1925         if (err)
1926                 return err;
1927
1928         evsel__disable_missing_features(evsel);
1929
1930         return err;
1931 }
1932
1933 bool evsel__detect_missing_features(struct evsel *evsel)
1934 {
1935         /*
1936          * Must probe features in the order they were added to the
1937          * perf_event_attr interface.
1938          */
1939         if (!perf_missing_features.read_lost &&
1940             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1941                 perf_missing_features.read_lost = true;
1942                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1943                 return true;
1944         } else if (!perf_missing_features.weight_struct &&
1945             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1946                 perf_missing_features.weight_struct = true;
1947                 pr_debug2("switching off weight struct support\n");
1948                 return true;
1949         } else if (!perf_missing_features.code_page_size &&
1950             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1951                 perf_missing_features.code_page_size = true;
1952                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1953                 return false;
1954         } else if (!perf_missing_features.data_page_size &&
1955             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1956                 perf_missing_features.data_page_size = true;
1957                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1958                 return false;
1959         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1960                 perf_missing_features.cgroup = true;
1961                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1962                 return false;
1963         } else if (!perf_missing_features.branch_hw_idx &&
1964             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1965                 perf_missing_features.branch_hw_idx = true;
1966                 pr_debug2("switching off branch HW index support\n");
1967                 return true;
1968         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1969                 perf_missing_features.aux_output = true;
1970                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1971                 return false;
1972         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1973                 perf_missing_features.bpf = true;
1974                 pr_debug2_peo("switching off bpf_event\n");
1975                 return true;
1976         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1977                 perf_missing_features.ksymbol = true;
1978                 pr_debug2_peo("switching off ksymbol\n");
1979                 return true;
1980         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1981                 perf_missing_features.write_backward = true;
1982                 pr_debug2_peo("switching off write_backward\n");
1983                 return false;
1984         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1985                 perf_missing_features.clockid_wrong = true;
1986                 pr_debug2_peo("switching off clockid\n");
1987                 return true;
1988         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1989                 perf_missing_features.clockid = true;
1990                 pr_debug2_peo("switching off use_clockid\n");
1991                 return true;
1992         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1993                 perf_missing_features.cloexec = true;
1994                 pr_debug2_peo("switching off cloexec flag\n");
1995                 return true;
1996         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1997                 perf_missing_features.mmap2 = true;
1998                 pr_debug2_peo("switching off mmap2\n");
1999                 return true;
2000         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
2001                 if (evsel->pmu == NULL)
2002                         evsel->pmu = evsel__find_pmu(evsel);
2003
2004                 if (evsel->pmu)
2005                         evsel->pmu->missing_features.exclude_guest = true;
2006                 else {
2007                         /* we cannot find PMU, disable attrs now */
2008                         evsel->core.attr.exclude_host = false;
2009                         evsel->core.attr.exclude_guest = false;
2010                 }
2011
2012                 if (evsel->exclude_GH) {
2013                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2014                         return false;
2015                 }
2016                 if (!perf_missing_features.exclude_guest) {
2017                         perf_missing_features.exclude_guest = true;
2018                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2019                 }
2020                 return true;
2021         } else if (!perf_missing_features.sample_id_all) {
2022                 perf_missing_features.sample_id_all = true;
2023                 pr_debug2_peo("switching off sample_id_all\n");
2024                 return true;
2025         } else if (!perf_missing_features.lbr_flags &&
2026                         (evsel->core.attr.branch_sample_type &
2027                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2028                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2029                 perf_missing_features.lbr_flags = true;
2030                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2031                 return true;
2032         } else if (!perf_missing_features.group_read &&
2033                     evsel->core.attr.inherit &&
2034                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2035                    evsel__is_group_leader(evsel)) {
2036                 perf_missing_features.group_read = true;
2037                 pr_debug2_peo("switching off group read\n");
2038                 return true;
2039         } else {
2040                 return false;
2041         }
2042 }
2043
2044 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2045 {
2046         int old_errno;
2047         struct rlimit l;
2048
2049         if (*set_rlimit < INCREASED_MAX) {
2050                 old_errno = errno;
2051
2052                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2053                         if (*set_rlimit == NO_CHANGE) {
2054                                 l.rlim_cur = l.rlim_max;
2055                         } else {
2056                                 l.rlim_cur = l.rlim_max + 1000;
2057                                 l.rlim_max = l.rlim_cur;
2058                         }
2059                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2060                                 (*set_rlimit) += 1;
2061                                 errno = old_errno;
2062                                 return true;
2063                         }
2064                 }
2065                 errno = old_errno;
2066         }
2067
2068         return false;
2069 }
2070
2071 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2072                 struct perf_thread_map *threads,
2073                 int start_cpu_map_idx, int end_cpu_map_idx)
2074 {
2075         int idx, thread, nthreads;
2076         int pid = -1, err, old_errno;
2077         enum rlimit_action set_rlimit = NO_CHANGE;
2078
2079         err = __evsel__prepare_open(evsel, cpus, threads);
2080         if (err)
2081                 return err;
2082
2083         if (cpus == NULL)
2084                 cpus = empty_cpu_map;
2085
2086         if (threads == NULL)
2087                 threads = empty_thread_map;
2088
2089         nthreads = perf_thread_map__nr(threads);
2090
2091         if (evsel->cgrp)
2092                 pid = evsel->cgrp->fd;
2093
2094 fallback_missing_features:
2095         evsel__disable_missing_features(evsel);
2096
2097         display_attr(&evsel->core.attr);
2098
2099         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2100
2101                 for (thread = 0; thread < nthreads; thread++) {
2102                         int fd, group_fd;
2103 retry_open:
2104                         if (thread >= nthreads)
2105                                 break;
2106
2107                         if (!evsel->cgrp && !evsel->core.system_wide)
2108                                 pid = perf_thread_map__pid(threads, thread);
2109
2110                         group_fd = get_group_fd(evsel, idx, thread);
2111
2112                         test_attr__ready();
2113
2114                         /* Debug message used by test scripts */
2115                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2116                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2117
2118                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2119                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2120                                                 group_fd, evsel->open_flags);
2121
2122                         FD(evsel, idx, thread) = fd;
2123
2124                         if (fd < 0) {
2125                                 err = -errno;
2126
2127                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2128                                           err);
2129                                 goto try_fallback;
2130                         }
2131
2132                         bpf_counter__install_pe(evsel, idx, fd);
2133
2134                         if (unlikely(test_attr__enabled)) {
2135                                 test_attr__open(&evsel->core.attr, pid,
2136                                                 perf_cpu_map__cpu(cpus, idx),
2137                                                 fd, group_fd, evsel->open_flags);
2138                         }
2139
2140                         /* Debug message used by test scripts */
2141                         pr_debug2_peo(" = %d\n", fd);
2142
2143                         if (evsel->bpf_fd >= 0) {
2144                                 int evt_fd = fd;
2145                                 int bpf_fd = evsel->bpf_fd;
2146
2147                                 err = ioctl(evt_fd,
2148                                             PERF_EVENT_IOC_SET_BPF,
2149                                             bpf_fd);
2150                                 if (err && errno != EEXIST) {
2151                                         pr_err("failed to attach bpf fd %d: %s\n",
2152                                                bpf_fd, strerror(errno));
2153                                         err = -EINVAL;
2154                                         goto out_close;
2155                                 }
2156                         }
2157
2158                         set_rlimit = NO_CHANGE;
2159
2160                         /*
2161                          * If we succeeded but had to kill clockid, fail and
2162                          * have evsel__open_strerror() print us a nice error.
2163                          */
2164                         if (perf_missing_features.clockid ||
2165                             perf_missing_features.clockid_wrong) {
2166                                 err = -EINVAL;
2167                                 goto out_close;
2168                         }
2169                 }
2170         }
2171
2172         return 0;
2173
2174 try_fallback:
2175         if (evsel__precise_ip_fallback(evsel))
2176                 goto retry_open;
2177
2178         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2179                                          idx, threads, thread, err)) {
2180                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2181                 nthreads--;
2182
2183                 /* ... and pretend like nothing have happened. */
2184                 err = 0;
2185                 goto retry_open;
2186         }
2187         /*
2188          * perf stat needs between 5 and 22 fds per CPU. When we run out
2189          * of them try to increase the limits.
2190          */
2191         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2192                 goto retry_open;
2193
2194         if (err != -EINVAL || idx > 0 || thread > 0)
2195                 goto out_close;
2196
2197         if (evsel__detect_missing_features(evsel))
2198                 goto fallback_missing_features;
2199 out_close:
2200         if (err)
2201                 threads->err_thread = thread;
2202
2203         old_errno = errno;
2204         do {
2205                 while (--thread >= 0) {
2206                         if (FD(evsel, idx, thread) >= 0)
2207                                 close(FD(evsel, idx, thread));
2208                         FD(evsel, idx, thread) = -1;
2209                 }
2210                 thread = nthreads;
2211         } while (--idx >= 0);
2212         errno = old_errno;
2213         return err;
2214 }
2215
2216 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2217                 struct perf_thread_map *threads)
2218 {
2219         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2220 }
2221
2222 void evsel__close(struct evsel *evsel)
2223 {
2224         perf_evsel__close(&evsel->core);
2225         perf_evsel__free_id(&evsel->core);
2226 }
2227
2228 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2229 {
2230         if (cpu_map_idx == -1)
2231                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2232
2233         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2234 }
2235
2236 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2237 {
2238         return evsel__open(evsel, NULL, threads);
2239 }
2240
2241 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2242                                        const union perf_event *event,
2243                                        struct perf_sample *sample)
2244 {
2245         u64 type = evsel->core.attr.sample_type;
2246         const __u64 *array = event->sample.array;
2247         bool swapped = evsel->needs_swap;
2248         union u64_swap u;
2249
2250         array += ((event->header.size -
2251                    sizeof(event->header)) / sizeof(u64)) - 1;
2252
2253         if (type & PERF_SAMPLE_IDENTIFIER) {
2254                 sample->id = *array;
2255                 array--;
2256         }
2257
2258         if (type & PERF_SAMPLE_CPU) {
2259                 u.val64 = *array;
2260                 if (swapped) {
2261                         /* undo swap of u64, then swap on individual u32s */
2262                         u.val64 = bswap_64(u.val64);
2263                         u.val32[0] = bswap_32(u.val32[0]);
2264                 }
2265
2266                 sample->cpu = u.val32[0];
2267                 array--;
2268         }
2269
2270         if (type & PERF_SAMPLE_STREAM_ID) {
2271                 sample->stream_id = *array;
2272                 array--;
2273         }
2274
2275         if (type & PERF_SAMPLE_ID) {
2276                 sample->id = *array;
2277                 array--;
2278         }
2279
2280         if (type & PERF_SAMPLE_TIME) {
2281                 sample->time = *array;
2282                 array--;
2283         }
2284
2285         if (type & PERF_SAMPLE_TID) {
2286                 u.val64 = *array;
2287                 if (swapped) {
2288                         /* undo swap of u64, then swap on individual u32s */
2289                         u.val64 = bswap_64(u.val64);
2290                         u.val32[0] = bswap_32(u.val32[0]);
2291                         u.val32[1] = bswap_32(u.val32[1]);
2292                 }
2293
2294                 sample->pid = u.val32[0];
2295                 sample->tid = u.val32[1];
2296                 array--;
2297         }
2298
2299         return 0;
2300 }
2301
2302 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2303                             u64 size)
2304 {
2305         return size > max_size || offset + size > endp;
2306 }
2307
2308 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2309         do {                                                            \
2310                 if (overflow(endp, (max_size), (offset), (size)))       \
2311                         return -EFAULT;                                 \
2312         } while (0)
2313
2314 #define OVERFLOW_CHECK_u64(offset) \
2315         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2316
2317 static int
2318 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2319 {
2320         /*
2321          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2322          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2323          * check the format does not go past the end of the event.
2324          */
2325         if (sample_size + sizeof(event->header) > event->header.size)
2326                 return -EFAULT;
2327
2328         return 0;
2329 }
2330
2331 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2332                                           const __u64 *array,
2333                                           u64 type __maybe_unused)
2334 {
2335         data->weight = *array;
2336 }
2337
2338 u64 evsel__bitfield_swap_branch_flags(u64 value)
2339 {
2340         u64 new_val = 0;
2341
2342         /*
2343          * branch_flags
2344          * union {
2345          *      u64 values;
2346          *      struct {
2347          *              mispred:1       //target mispredicted
2348          *              predicted:1     //target predicted
2349          *              in_tx:1         //in transaction
2350          *              abort:1         //transaction abort
2351          *              cycles:16       //cycle count to last branch
2352          *              type:4          //branch type
2353          *              spec:2          //branch speculation info
2354          *              new_type:4      //additional branch type
2355          *              priv:3          //privilege level
2356          *              reserved:31
2357          *      }
2358          * }
2359          *
2360          * Avoid bswap64() the entire branch_flag.value,
2361          * as it has variable bit-field sizes. Instead the
2362          * macro takes the bit-field position/size,
2363          * swaps it based on the host endianness.
2364          */
2365         if (host_is_bigendian()) {
2366                 new_val = bitfield_swap(value, 0, 1);
2367                 new_val |= bitfield_swap(value, 1, 1);
2368                 new_val |= bitfield_swap(value, 2, 1);
2369                 new_val |= bitfield_swap(value, 3, 1);
2370                 new_val |= bitfield_swap(value, 4, 16);
2371                 new_val |= bitfield_swap(value, 20, 4);
2372                 new_val |= bitfield_swap(value, 24, 2);
2373                 new_val |= bitfield_swap(value, 26, 4);
2374                 new_val |= bitfield_swap(value, 30, 3);
2375                 new_val |= bitfield_swap(value, 33, 31);
2376         } else {
2377                 new_val = bitfield_swap(value, 63, 1);
2378                 new_val |= bitfield_swap(value, 62, 1);
2379                 new_val |= bitfield_swap(value, 61, 1);
2380                 new_val |= bitfield_swap(value, 60, 1);
2381                 new_val |= bitfield_swap(value, 44, 16);
2382                 new_val |= bitfield_swap(value, 40, 4);
2383                 new_val |= bitfield_swap(value, 38, 2);
2384                 new_val |= bitfield_swap(value, 34, 4);
2385                 new_val |= bitfield_swap(value, 31, 3);
2386                 new_val |= bitfield_swap(value, 0, 31);
2387         }
2388
2389         return new_val;
2390 }
2391
2392 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2393                         struct perf_sample *data)
2394 {
2395         u64 type = evsel->core.attr.sample_type;
2396         bool swapped = evsel->needs_swap;
2397         const __u64 *array;
2398         u16 max_size = event->header.size;
2399         const void *endp = (void *)event + max_size;
2400         u64 sz;
2401
2402         /*
2403          * used for cross-endian analysis. See git commit 65014ab3
2404          * for why this goofiness is needed.
2405          */
2406         union u64_swap u;
2407
2408         memset(data, 0, sizeof(*data));
2409         data->cpu = data->pid = data->tid = -1;
2410         data->stream_id = data->id = data->time = -1ULL;
2411         data->period = evsel->core.attr.sample_period;
2412         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2413         data->misc    = event->header.misc;
2414         data->id = -1ULL;
2415         data->data_src = PERF_MEM_DATA_SRC_NONE;
2416         data->vcpu = -1;
2417
2418         if (event->header.type != PERF_RECORD_SAMPLE) {
2419                 if (!evsel->core.attr.sample_id_all)
2420                         return 0;
2421                 return perf_evsel__parse_id_sample(evsel, event, data);
2422         }
2423
2424         array = event->sample.array;
2425
2426         if (perf_event__check_size(event, evsel->sample_size))
2427                 return -EFAULT;
2428
2429         if (type & PERF_SAMPLE_IDENTIFIER) {
2430                 data->id = *array;
2431                 array++;
2432         }
2433
2434         if (type & PERF_SAMPLE_IP) {
2435                 data->ip = *array;
2436                 array++;
2437         }
2438
2439         if (type & PERF_SAMPLE_TID) {
2440                 u.val64 = *array;
2441                 if (swapped) {
2442                         /* undo swap of u64, then swap on individual u32s */
2443                         u.val64 = bswap_64(u.val64);
2444                         u.val32[0] = bswap_32(u.val32[0]);
2445                         u.val32[1] = bswap_32(u.val32[1]);
2446                 }
2447
2448                 data->pid = u.val32[0];
2449                 data->tid = u.val32[1];
2450                 array++;
2451         }
2452
2453         if (type & PERF_SAMPLE_TIME) {
2454                 data->time = *array;
2455                 array++;
2456         }
2457
2458         if (type & PERF_SAMPLE_ADDR) {
2459                 data->addr = *array;
2460                 array++;
2461         }
2462
2463         if (type & PERF_SAMPLE_ID) {
2464                 data->id = *array;
2465                 array++;
2466         }
2467
2468         if (type & PERF_SAMPLE_STREAM_ID) {
2469                 data->stream_id = *array;
2470                 array++;
2471         }
2472
2473         if (type & PERF_SAMPLE_CPU) {
2474
2475                 u.val64 = *array;
2476                 if (swapped) {
2477                         /* undo swap of u64, then swap on individual u32s */
2478                         u.val64 = bswap_64(u.val64);
2479                         u.val32[0] = bswap_32(u.val32[0]);
2480                 }
2481
2482                 data->cpu = u.val32[0];
2483                 array++;
2484         }
2485
2486         if (type & PERF_SAMPLE_PERIOD) {
2487                 data->period = *array;
2488                 array++;
2489         }
2490
2491         if (type & PERF_SAMPLE_READ) {
2492                 u64 read_format = evsel->core.attr.read_format;
2493
2494                 OVERFLOW_CHECK_u64(array);
2495                 if (read_format & PERF_FORMAT_GROUP)
2496                         data->read.group.nr = *array;
2497                 else
2498                         data->read.one.value = *array;
2499
2500                 array++;
2501
2502                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2503                         OVERFLOW_CHECK_u64(array);
2504                         data->read.time_enabled = *array;
2505                         array++;
2506                 }
2507
2508                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2509                         OVERFLOW_CHECK_u64(array);
2510                         data->read.time_running = *array;
2511                         array++;
2512                 }
2513
2514                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2515                 if (read_format & PERF_FORMAT_GROUP) {
2516                         const u64 max_group_nr = UINT64_MAX /
2517                                         sizeof(struct sample_read_value);
2518
2519                         if (data->read.group.nr > max_group_nr)
2520                                 return -EFAULT;
2521
2522                         sz = data->read.group.nr * sample_read_value_size(read_format);
2523                         OVERFLOW_CHECK(array, sz, max_size);
2524                         data->read.group.values =
2525                                         (struct sample_read_value *)array;
2526                         array = (void *)array + sz;
2527                 } else {
2528                         OVERFLOW_CHECK_u64(array);
2529                         data->read.one.id = *array;
2530                         array++;
2531
2532                         if (read_format & PERF_FORMAT_LOST) {
2533                                 OVERFLOW_CHECK_u64(array);
2534                                 data->read.one.lost = *array;
2535                                 array++;
2536                         }
2537                 }
2538         }
2539
2540         if (type & PERF_SAMPLE_CALLCHAIN) {
2541                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2542
2543                 OVERFLOW_CHECK_u64(array);
2544                 data->callchain = (struct ip_callchain *)array++;
2545                 if (data->callchain->nr > max_callchain_nr)
2546                         return -EFAULT;
2547                 sz = data->callchain->nr * sizeof(u64);
2548                 OVERFLOW_CHECK(array, sz, max_size);
2549                 array = (void *)array + sz;
2550         }
2551
2552         if (type & PERF_SAMPLE_RAW) {
2553                 OVERFLOW_CHECK_u64(array);
2554                 u.val64 = *array;
2555
2556                 /*
2557                  * Undo swap of u64, then swap on individual u32s,
2558                  * get the size of the raw area and undo all of the
2559                  * swap. The pevent interface handles endianness by
2560                  * itself.
2561                  */
2562                 if (swapped) {
2563                         u.val64 = bswap_64(u.val64);
2564                         u.val32[0] = bswap_32(u.val32[0]);
2565                         u.val32[1] = bswap_32(u.val32[1]);
2566                 }
2567                 data->raw_size = u.val32[0];
2568
2569                 /*
2570                  * The raw data is aligned on 64bits including the
2571                  * u32 size, so it's safe to use mem_bswap_64.
2572                  */
2573                 if (swapped)
2574                         mem_bswap_64((void *) array, data->raw_size);
2575
2576                 array = (void *)array + sizeof(u32);
2577
2578                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2579                 data->raw_data = (void *)array;
2580                 array = (void *)array + data->raw_size;
2581         }
2582
2583         if (type & PERF_SAMPLE_BRANCH_STACK) {
2584                 const u64 max_branch_nr = UINT64_MAX /
2585                                           sizeof(struct branch_entry);
2586                 struct branch_entry *e;
2587                 unsigned int i;
2588
2589                 OVERFLOW_CHECK_u64(array);
2590                 data->branch_stack = (struct branch_stack *)array++;
2591
2592                 if (data->branch_stack->nr > max_branch_nr)
2593                         return -EFAULT;
2594
2595                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2596                 if (evsel__has_branch_hw_idx(evsel)) {
2597                         sz += sizeof(u64);
2598                         e = &data->branch_stack->entries[0];
2599                 } else {
2600                         data->no_hw_idx = true;
2601                         /*
2602                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2603                          * only nr and entries[] will be output by kernel.
2604                          */
2605                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2606                 }
2607
2608                 if (swapped) {
2609                         /*
2610                          * struct branch_flag does not have endian
2611                          * specific bit field definition. And bswap
2612                          * will not resolve the issue, since these
2613                          * are bit fields.
2614                          *
2615                          * evsel__bitfield_swap_branch_flags() uses a
2616                          * bitfield_swap macro to swap the bit position
2617                          * based on the host endians.
2618                          */
2619                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2620                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2621                 }
2622
2623                 OVERFLOW_CHECK(array, sz, max_size);
2624                 array = (void *)array + sz;
2625         }
2626
2627         if (type & PERF_SAMPLE_REGS_USER) {
2628                 OVERFLOW_CHECK_u64(array);
2629                 data->user_regs.abi = *array;
2630                 array++;
2631
2632                 if (data->user_regs.abi) {
2633                         u64 mask = evsel->core.attr.sample_regs_user;
2634
2635                         sz = hweight64(mask) * sizeof(u64);
2636                         OVERFLOW_CHECK(array, sz, max_size);
2637                         data->user_regs.mask = mask;
2638                         data->user_regs.regs = (u64 *)array;
2639                         array = (void *)array + sz;
2640                 }
2641         }
2642
2643         if (type & PERF_SAMPLE_STACK_USER) {
2644                 OVERFLOW_CHECK_u64(array);
2645                 sz = *array++;
2646
2647                 data->user_stack.offset = ((char *)(array - 1)
2648                                           - (char *) event);
2649
2650                 if (!sz) {
2651                         data->user_stack.size = 0;
2652                 } else {
2653                         OVERFLOW_CHECK(array, sz, max_size);
2654                         data->user_stack.data = (char *)array;
2655                         array = (void *)array + sz;
2656                         OVERFLOW_CHECK_u64(array);
2657                         data->user_stack.size = *array++;
2658                         if (WARN_ONCE(data->user_stack.size > sz,
2659                                       "user stack dump failure\n"))
2660                                 return -EFAULT;
2661                 }
2662         }
2663
2664         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2665                 OVERFLOW_CHECK_u64(array);
2666                 arch_perf_parse_sample_weight(data, array, type);
2667                 array++;
2668         }
2669
2670         if (type & PERF_SAMPLE_DATA_SRC) {
2671                 OVERFLOW_CHECK_u64(array);
2672                 data->data_src = *array;
2673                 array++;
2674         }
2675
2676         if (type & PERF_SAMPLE_TRANSACTION) {
2677                 OVERFLOW_CHECK_u64(array);
2678                 data->transaction = *array;
2679                 array++;
2680         }
2681
2682         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2683         if (type & PERF_SAMPLE_REGS_INTR) {
2684                 OVERFLOW_CHECK_u64(array);
2685                 data->intr_regs.abi = *array;
2686                 array++;
2687
2688                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2689                         u64 mask = evsel->core.attr.sample_regs_intr;
2690
2691                         sz = hweight64(mask) * sizeof(u64);
2692                         OVERFLOW_CHECK(array, sz, max_size);
2693                         data->intr_regs.mask = mask;
2694                         data->intr_regs.regs = (u64 *)array;
2695                         array = (void *)array + sz;
2696                 }
2697         }
2698
2699         data->phys_addr = 0;
2700         if (type & PERF_SAMPLE_PHYS_ADDR) {
2701                 data->phys_addr = *array;
2702                 array++;
2703         }
2704
2705         data->cgroup = 0;
2706         if (type & PERF_SAMPLE_CGROUP) {
2707                 data->cgroup = *array;
2708                 array++;
2709         }
2710
2711         data->data_page_size = 0;
2712         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2713                 data->data_page_size = *array;
2714                 array++;
2715         }
2716
2717         data->code_page_size = 0;
2718         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2719                 data->code_page_size = *array;
2720                 array++;
2721         }
2722
2723         if (type & PERF_SAMPLE_AUX) {
2724                 OVERFLOW_CHECK_u64(array);
2725                 sz = *array++;
2726
2727                 OVERFLOW_CHECK(array, sz, max_size);
2728                 /* Undo swap of data */
2729                 if (swapped)
2730                         mem_bswap_64((char *)array, sz);
2731                 data->aux_sample.size = sz;
2732                 data->aux_sample.data = (char *)array;
2733                 array = (void *)array + sz;
2734         }
2735
2736         return 0;
2737 }
2738
2739 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2740                                   u64 *timestamp)
2741 {
2742         u64 type = evsel->core.attr.sample_type;
2743         const __u64 *array;
2744
2745         if (!(type & PERF_SAMPLE_TIME))
2746                 return -1;
2747
2748         if (event->header.type != PERF_RECORD_SAMPLE) {
2749                 struct perf_sample data = {
2750                         .time = -1ULL,
2751                 };
2752
2753                 if (!evsel->core.attr.sample_id_all)
2754                         return -1;
2755                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2756                         return -1;
2757
2758                 *timestamp = data.time;
2759                 return 0;
2760         }
2761
2762         array = event->sample.array;
2763
2764         if (perf_event__check_size(event, evsel->sample_size))
2765                 return -EFAULT;
2766
2767         if (type & PERF_SAMPLE_IDENTIFIER)
2768                 array++;
2769
2770         if (type & PERF_SAMPLE_IP)
2771                 array++;
2772
2773         if (type & PERF_SAMPLE_TID)
2774                 array++;
2775
2776         if (type & PERF_SAMPLE_TIME)
2777                 *timestamp = *array;
2778
2779         return 0;
2780 }
2781
2782 u16 evsel__id_hdr_size(struct evsel *evsel)
2783 {
2784         u64 sample_type = evsel->core.attr.sample_type;
2785         u16 size = 0;
2786
2787         if (sample_type & PERF_SAMPLE_TID)
2788                 size += sizeof(u64);
2789
2790         if (sample_type & PERF_SAMPLE_TIME)
2791                 size += sizeof(u64);
2792
2793         if (sample_type & PERF_SAMPLE_ID)
2794                 size += sizeof(u64);
2795
2796         if (sample_type & PERF_SAMPLE_STREAM_ID)
2797                 size += sizeof(u64);
2798
2799         if (sample_type & PERF_SAMPLE_CPU)
2800                 size += sizeof(u64);
2801
2802         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2803                 size += sizeof(u64);
2804
2805         return size;
2806 }
2807
2808 #ifdef HAVE_LIBTRACEEVENT
2809 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2810 {
2811         return tep_find_field(evsel->tp_format, name);
2812 }
2813
2814 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2815 {
2816         struct tep_format_field *field = evsel__field(evsel, name);
2817         int offset;
2818
2819         if (!field)
2820                 return NULL;
2821
2822         offset = field->offset;
2823
2824         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2825                 offset = *(int *)(sample->raw_data + field->offset);
2826                 offset &= 0xffff;
2827                 if (tep_field_is_relative(field->flags))
2828                         offset += field->offset + field->size;
2829         }
2830
2831         return sample->raw_data + offset;
2832 }
2833
2834 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2835                          bool needs_swap)
2836 {
2837         u64 value;
2838         void *ptr = sample->raw_data + field->offset;
2839
2840         switch (field->size) {
2841         case 1:
2842                 return *(u8 *)ptr;
2843         case 2:
2844                 value = *(u16 *)ptr;
2845                 break;
2846         case 4:
2847                 value = *(u32 *)ptr;
2848                 break;
2849         case 8:
2850                 memcpy(&value, ptr, sizeof(u64));
2851                 break;
2852         default:
2853                 return 0;
2854         }
2855
2856         if (!needs_swap)
2857                 return value;
2858
2859         switch (field->size) {
2860         case 2:
2861                 return bswap_16(value);
2862         case 4:
2863                 return bswap_32(value);
2864         case 8:
2865                 return bswap_64(value);
2866         default:
2867                 return 0;
2868         }
2869
2870         return 0;
2871 }
2872
2873 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2874 {
2875         struct tep_format_field *field = evsel__field(evsel, name);
2876
2877         if (!field)
2878                 return 0;
2879
2880         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2881 }
2882 #endif
2883
2884 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2885 {
2886         int paranoid;
2887
2888         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2889             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2890             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2891                 /*
2892                  * If it's cycles then fall back to hrtimer based
2893                  * cpu-clock-tick sw counter, which is always available even if
2894                  * no PMU support.
2895                  *
2896                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2897                  * b0a873e).
2898                  */
2899                 scnprintf(msg, msgsize, "%s",
2900 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2901
2902                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2903                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2904
2905                 zfree(&evsel->name);
2906                 return true;
2907         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2908                    (paranoid = perf_event_paranoid()) > 1) {
2909                 const char *name = evsel__name(evsel);
2910                 char *new_name;
2911                 const char *sep = ":";
2912
2913                 /* If event has exclude user then don't exclude kernel. */
2914                 if (evsel->core.attr.exclude_user)
2915                         return false;
2916
2917                 /* Is there already the separator in the name. */
2918                 if (strchr(name, '/') ||
2919                     (strchr(name, ':') && !evsel->is_libpfm_event))
2920                         sep = "";
2921
2922                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2923                         return false;
2924
2925                 free(evsel->name);
2926                 evsel->name = new_name;
2927                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2928                           "to fall back to excluding kernel and hypervisor "
2929                           " samples", paranoid);
2930                 evsel->core.attr.exclude_kernel = 1;
2931                 evsel->core.attr.exclude_hv     = 1;
2932
2933                 return true;
2934         }
2935
2936         return false;
2937 }
2938
2939 static bool find_process(const char *name)
2940 {
2941         size_t len = strlen(name);
2942         DIR *dir;
2943         struct dirent *d;
2944         int ret = -1;
2945
2946         dir = opendir(procfs__mountpoint());
2947         if (!dir)
2948                 return false;
2949
2950         /* Walk through the directory. */
2951         while (ret && (d = readdir(dir)) != NULL) {
2952                 char path[PATH_MAX];
2953                 char *data;
2954                 size_t size;
2955
2956                 if ((d->d_type != DT_DIR) ||
2957                      !strcmp(".", d->d_name) ||
2958                      !strcmp("..", d->d_name))
2959                         continue;
2960
2961                 scnprintf(path, sizeof(path), "%s/%s/comm",
2962                           procfs__mountpoint(), d->d_name);
2963
2964                 if (filename__read_str(path, &data, &size))
2965                         continue;
2966
2967                 ret = strncmp(name, data, len);
2968                 free(data);
2969         }
2970
2971         closedir(dir);
2972         return ret ? false : true;
2973 }
2974
2975 static bool is_amd(const char *arch, const char *cpuid)
2976 {
2977         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2978 }
2979
2980 static bool is_amd_ibs(struct evsel *evsel)
2981 {
2982         return evsel->core.attr.precise_ip
2983             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2984 }
2985
2986 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2987                          int err, char *msg, size_t size)
2988 {
2989         struct perf_env *env = evsel__env(evsel);
2990         const char *arch = perf_env__arch(env);
2991         const char *cpuid = perf_env__cpuid(env);
2992         char sbuf[STRERR_BUFSIZE];
2993         int printed = 0, enforced = 0;
2994
2995         switch (err) {
2996         case EPERM:
2997         case EACCES:
2998                 printed += scnprintf(msg + printed, size - printed,
2999                         "Access to performance monitoring and observability operations is limited.\n");
3000
3001                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
3002                         if (enforced) {
3003                                 printed += scnprintf(msg + printed, size - printed,
3004                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
3005                                         "monitoring and observability operations. Inspect system audit records for\n"
3006                                         "more perf_event access control information and adjusting the policy.\n");
3007                         }
3008                 }
3009
3010                 if (err == EPERM)
3011                         printed += scnprintf(msg, size,
3012                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
3013
3014                 return scnprintf(msg + printed, size - printed,
3015                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3016                  "access to performance monitoring and observability operations for processes\n"
3017                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3018                  "More information can be found at 'Perf events and tool security' document:\n"
3019                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3020                  "perf_event_paranoid setting is %d:\n"
3021                  "  -1: Allow use of (almost) all events by all users\n"
3022                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3023                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
3024                  ">= 1: Disallow CPU event access\n"
3025                  ">= 2: Disallow kernel profiling\n"
3026                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3027                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3028                  perf_event_paranoid());
3029         case ENOENT:
3030                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3031         case EMFILE:
3032                 return scnprintf(msg, size, "%s",
3033                          "Too many events are opened.\n"
3034                          "Probably the maximum number of open file descriptors has been reached.\n"
3035                          "Hint: Try again after reducing the number of events.\n"
3036                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3037         case ENOMEM:
3038                 if (evsel__has_callchain(evsel) &&
3039                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3040                         return scnprintf(msg, size,
3041                                          "Not enough memory to setup event with callchain.\n"
3042                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3043                                          "Hint: Current value: %d", sysctl__max_stack());
3044                 break;
3045         case ENODEV:
3046                 if (target->cpu_list)
3047                         return scnprintf(msg, size, "%s",
3048          "No such device - did you specify an out-of-range profile CPU?");
3049                 break;
3050         case EOPNOTSUPP:
3051                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3052                         return scnprintf(msg, size,
3053         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3054                                          evsel__name(evsel));
3055                 if (evsel->core.attr.aux_output)
3056                         return scnprintf(msg, size,
3057         "%s: PMU Hardware doesn't support 'aux_output' feature",
3058                                          evsel__name(evsel));
3059                 if (evsel->core.attr.sample_period != 0)
3060                         return scnprintf(msg, size,
3061         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3062                                          evsel__name(evsel));
3063                 if (evsel->core.attr.precise_ip)
3064                         return scnprintf(msg, size, "%s",
3065         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3066 #if defined(__i386__) || defined(__x86_64__)
3067                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3068                         return scnprintf(msg, size, "%s",
3069         "No hardware sampling interrupt available.\n");
3070 #endif
3071                 break;
3072         case EBUSY:
3073                 if (find_process("oprofiled"))
3074                         return scnprintf(msg, size,
3075         "The PMU counters are busy/taken by another profiler.\n"
3076         "We found oprofile daemon running, please stop it and try again.");
3077                 break;
3078         case EINVAL:
3079                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3080                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3081                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3082                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3083                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3084                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3085                 if (perf_missing_features.clockid)
3086                         return scnprintf(msg, size, "clockid feature not supported.");
3087                 if (perf_missing_features.clockid_wrong)
3088                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3089                 if (perf_missing_features.aux_output)
3090                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3091                 if (!target__has_cpu(target))
3092                         return scnprintf(msg, size,
3093         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3094                                         evsel__name(evsel));
3095                 if (is_amd(arch, cpuid)) {
3096                         if (is_amd_ibs(evsel)) {
3097                                 if (evsel->core.attr.exclude_kernel)
3098                                         return scnprintf(msg, size,
3099         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3100                                 if (!evsel->core.system_wide)
3101                                         return scnprintf(msg, size,
3102         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3103                         }
3104                 }
3105
3106                 break;
3107         case ENODATA:
3108                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3109                                  "Please add an auxiliary event in front of the load latency event.");
3110         default:
3111                 break;
3112         }
3113
3114         return scnprintf(msg, size,
3115         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3116         "/bin/dmesg | grep -i perf may provide additional information.\n",
3117                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3118 }
3119
3120 struct perf_env *evsel__env(struct evsel *evsel)
3121 {
3122         if (evsel && evsel->evlist && evsel->evlist->env)
3123                 return evsel->evlist->env;
3124         return &perf_env;
3125 }
3126
3127 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3128 {
3129         int cpu_map_idx, thread;
3130
3131         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3132                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3133                      thread++) {
3134                         int fd = FD(evsel, cpu_map_idx, thread);
3135
3136                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3137                                                    cpu_map_idx, thread, fd) < 0)
3138                                 return -1;
3139                 }
3140         }
3141
3142         return 0;
3143 }
3144
3145 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3146 {
3147         struct perf_cpu_map *cpus = evsel->core.cpus;
3148         struct perf_thread_map *threads = evsel->core.threads;
3149
3150         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3151                 return -ENOMEM;
3152
3153         return store_evsel_ids(evsel, evlist);
3154 }
3155
3156 void evsel__zero_per_pkg(struct evsel *evsel)
3157 {
3158         struct hashmap_entry *cur;
3159         size_t bkt;
3160
3161         if (evsel->per_pkg_mask) {
3162                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3163                         zfree(&cur->pkey);
3164
3165                 hashmap__clear(evsel->per_pkg_mask);
3166         }
3167 }
3168
3169 bool evsel__is_hybrid(const struct evsel *evsel)
3170 {
3171         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3172 }
3173
3174 struct evsel *evsel__leader(const struct evsel *evsel)
3175 {
3176         return container_of(evsel->core.leader, struct evsel, core);
3177 }
3178
3179 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3180 {
3181         return evsel->core.leader == &leader->core;
3182 }
3183
3184 bool evsel__is_leader(struct evsel *evsel)
3185 {
3186         return evsel__has_leader(evsel, evsel);
3187 }
3188
3189 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3190 {
3191         evsel->core.leader = &leader->core;
3192 }
3193
3194 int evsel__source_count(const struct evsel *evsel)
3195 {
3196         struct evsel *pos;
3197         int count = 0;
3198
3199         evlist__for_each_entry(evsel->evlist, pos) {
3200                 if (pos->metric_leader == evsel)
3201                         count++;
3202         }
3203         return count;
3204 }
3205
3206 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3207 {
3208         return false;
3209 }
3210
3211 /*
3212  * Remove an event from a given group (leader).
3213  * Some events, e.g., perf metrics Topdown events,
3214  * must always be grouped. Ignore the events.
3215  */
3216 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3217 {
3218         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3219                 evsel__set_leader(evsel, evsel);
3220                 evsel->core.nr_members = 0;
3221                 leader->core.nr_members--;
3222         }
3223 }