tools/perf/util/evsel.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8
   9 #include <byteswap.h>
  10 #include <errno.h>
  11 #include <inttypes.h>
  12 #include <linux/bitops.h>
  13 #include <api/fs/fs.h>
  14 #include <api/fs/tracing_path.h>
  15 #include <linux/hw_breakpoint.h>
  16 #include <linux/perf_event.h>
  17 #include <linux/compiler.h>
  18 #include <linux/err.h>
  19 #include <linux/zalloc.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/resource.h>
  22 #include <sys/types.h>
  23 #include <dirent.h>
  24 #include <stdlib.h>
  25 #include <perf/evsel.h>
  26 #include "asm/bug.h"
  27 #include "bpf_counter.h"
  28 #include "callchain.h"
  29 #include "cgroup.h"
  30 #include "counts.h"
  31 #include "event.h"
  32 #include "evsel.h"
  33 #include "util/env.h"
  34 #include "util/evsel_config.h"
  35 #include "util/evsel_fprintf.h"
  36 #include "evlist.h"
  37 #include <perf/cpumap.h>
  38 #include "thread_map.h"
  39 #include "target.h"
  40 #include "perf_regs.h"
  41 #include "record.h"
  42 #include "debug.h"
  43 #include "trace-event.h"
  44 #include "stat.h"
  45 #include "string2.h"
  46 #include "memswap.h"
  47 #include "util.h"
  48 #include "util/hashmap.h"
  49 #include "pmu-hybrid.h"
  50 #include "off_cpu.h"
  51 #include "../perf-sys.h"
  52 #include "util/parse-branch-options.h"
  53 #include <internal/xyarray.h>
  54 #include <internal/lib.h>
  55 #include <internal/threadmap.h>
  56
  57 #include <linux/ctype.h>
  58
  59 #ifdef HAVE_LIBTRACEEVENT
  60 #include <traceevent/event-parse.h>
  61 #endif
  62
  63 struct perf_missing_features perf_missing_features;
  64
  65 static clockid_t clockid;
  66
  67 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = {
  68         NULL,
  69         "duration_time",
  70         "user_time",
  71         "system_time",
  72 };
  73
  74 const char *perf_tool_event__to_str(enum perf_tool_event ev)
  75 {
  76         if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX)
  77                 return perf_tool_event__tool_names[ev];
  78
  79         return NULL;
  80 }
  81
  82 enum perf_tool_event perf_tool_event__from_str(const char *str)
  83 {
  84         int i;
  85
  86         perf_tool_event__for_each_event(i) {
  87                 if (!strcmp(str, perf_tool_event__tool_names[i]))
  88                         return i;
  89         }
  90         return PERF_TOOL_NONE;
  91 }
  92
  93
  94 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused)
  95 {
  96         return 0;
  97 }
  98
  99 void __weak test_attr__ready(void) { }
 100
 101 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused)
 102 {
 103 }
 104
 105 static struct {
 106         size_t  size;
 107         int     (*init)(struct evsel *evsel);
 108         void    (*fini)(struct evsel *evsel);
 109 } perf_evsel__object = {
 110         .size = sizeof(struct evsel),
 111         .init = evsel__no_extra_init,
 112         .fini = evsel__no_extra_fini,
 113 };
 114
 115 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel),
 116                          void (*fini)(struct evsel *evsel))
 117 {
 118
 119         if (object_size == 0)
 120                 goto set_methods;
 121
 122         if (perf_evsel__object.size > object_size)
 123                 return -EINVAL;
 124
 125         perf_evsel__object.size = object_size;
 126
 127 set_methods:
 128         if (init != NULL)
 129                 perf_evsel__object.init = init;
 130
 131         if (fini != NULL)
 132                 perf_evsel__object.fini = fini;
 133
 134         return 0;
 135 }
 136
 137 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
 138
 139 int __evsel__sample_size(u64 sample_type)
 140 {
 141         u64 mask = sample_type & PERF_SAMPLE_MASK;
 142         int size = 0;
 143         int i;
 144
 145         for (i = 0; i < 64; i++) {
 146                 if (mask & (1ULL << i))
 147                         size++;
 148         }
 149
 150         size *= sizeof(u64);
 151
 152         return size;
 153 }
 154
 155 /**
 156  * __perf_evsel__calc_id_pos - calculate id_pos.
 157  * @sample_type: sample type
 158  *
 159  * This function returns the position of the event id (PERF_SAMPLE_ID or
 160  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 161  * perf_record_sample.
 162  */
 163 static int __perf_evsel__calc_id_pos(u64 sample_type)
 164 {
 165         int idx = 0;
 166
 167         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 168                 return 0;
 169
 170         if (!(sample_type & PERF_SAMPLE_ID))
 171                 return -1;
 172
 173         if (sample_type & PERF_SAMPLE_IP)
 174                 idx += 1;
 175
 176         if (sample_type & PERF_SAMPLE_TID)
 177                 idx += 1;
 178
 179         if (sample_type & PERF_SAMPLE_TIME)
 180                 idx += 1;
 181
 182         if (sample_type & PERF_SAMPLE_ADDR)
 183                 idx += 1;
 184
 185         return idx;
 186 }
 187
 188 /**
 189  * __perf_evsel__calc_is_pos - calculate is_pos.
 190  * @sample_type: sample type
 191  *
 192  * This function returns the position (counting backwards) of the event id
 193  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 194  * sample_id_all is used there is an id sample appended to non-sample events.
 195  */
 196 static int __perf_evsel__calc_is_pos(u64 sample_type)
 197 {
 198         int idx = 1;
 199
 200         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 201                 return 1;
 202
 203         if (!(sample_type & PERF_SAMPLE_ID))
 204                 return -1;
 205
 206         if (sample_type & PERF_SAMPLE_CPU)
 207                 idx += 1;
 208
 209         if (sample_type & PERF_SAMPLE_STREAM_ID)
 210                 idx += 1;
 211
 212         return idx;
 213 }
 214
 215 void evsel__calc_id_pos(struct evsel *evsel)
 216 {
 217         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type);
 218         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type);
 219 }
 220
 221 void __evsel__set_sample_bit(struct evsel *evsel,
 222                                   enum perf_event_sample_format bit)
 223 {
 224         if (!(evsel->core.attr.sample_type & bit)) {
 225                 evsel->core.attr.sample_type |= bit;
 226                 evsel->sample_size += sizeof(u64);
 227                 evsel__calc_id_pos(evsel);
 228         }
 229 }
 230
 231 void __evsel__reset_sample_bit(struct evsel *evsel,
 232                                     enum perf_event_sample_format bit)
 233 {
 234         if (evsel->core.attr.sample_type & bit) {
 235                 evsel->core.attr.sample_type &= ~bit;
 236                 evsel->sample_size -= sizeof(u64);
 237                 evsel__calc_id_pos(evsel);
 238         }
 239 }
 240
 241 void evsel__set_sample_id(struct evsel *evsel,
 242                                bool can_sample_identifier)
 243 {
 244         if (can_sample_identifier) {
 245                 evsel__reset_sample_bit(evsel, ID);
 246                 evsel__set_sample_bit(evsel, IDENTIFIER);
 247         } else {
 248                 evsel__set_sample_bit(evsel, ID);
 249         }
 250         evsel->core.attr.read_format |= PERF_FORMAT_ID;
 251 }
 252
 253 /**
 254  * evsel__is_function_event - Return whether given evsel is a function
 255  * trace event
 256  *
 257  * @evsel - evsel selector to be tested
 258  *
 259  * Return %true if event is function trace event
 260  */
 261 bool evsel__is_function_event(struct evsel *evsel)
 262 {
 263 #define FUNCTION_EVENT "ftrace:function"
 264
 265         return evsel->name &&
 266                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 267
 268 #undef FUNCTION_EVENT
 269 }
 270
 271 void evsel__init(struct evsel *evsel,
 272                  struct perf_event_attr *attr, int idx)
 273 {
 274         perf_evsel__init(&evsel->core, attr, idx);
 275         evsel->tracking    = !idx;
 276         evsel->unit        = strdup("");
 277         evsel->scale       = 1.0;
 278         evsel->max_events  = ULONG_MAX;
 279         evsel->evlist      = NULL;
 280         evsel->bpf_obj     = NULL;
 281         evsel->bpf_fd      = -1;
 282         INIT_LIST_HEAD(&evsel->config_terms);
 283         INIT_LIST_HEAD(&evsel->bpf_counter_list);
 284         perf_evsel__object.init(evsel);
 285         evsel->sample_size = __evsel__sample_size(attr->sample_type);
 286         evsel__calc_id_pos(evsel);
 287         evsel->cmdline_group_boundary = false;
 288         evsel->metric_events = NULL;
 289         evsel->per_pkg_mask  = NULL;
 290         evsel->collect_stat  = false;
 291         evsel->pmu_name      = NULL;
 292 }
 293
 294 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx)
 295 {
 296         struct evsel *evsel = zalloc(perf_evsel__object.size);
 297
 298         if (!evsel)
 299                 return NULL;
 300         evsel__init(evsel, attr, idx);
 301
 302         if (evsel__is_bpf_output(evsel) && !attr->sample_type) {
 303                 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 304                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 305                 evsel->core.attr.sample_period = 1;
 306         }
 307
 308         if (evsel__is_clock(evsel)) {
 309                 free((char *)evsel->unit);
 310                 evsel->unit = strdup("msec");
 311                 evsel->scale = 1e-6;
 312         }
 313
 314         return evsel;
 315 }
 316
 317 static bool perf_event_can_profile_kernel(void)
 318 {
 319         return perf_event_paranoid_check(1);
 320 }
 321
 322 struct evsel *evsel__new_cycles(bool precise __maybe_unused, __u32 type, __u64 config)
 323 {
 324         struct perf_event_attr attr = {
 325                 .type   = type,
 326                 .config = config,
 327                 .exclude_kernel = !perf_event_can_profile_kernel(),
 328         };
 329         struct evsel *evsel;
 330
 331         event_attr_init(&attr);
 332
 333         /*
 334          * Now let the usual logic to set up the perf_event_attr defaults
 335          * to kick in when we return and before perf_evsel__open() is called.
 336          */
 337         evsel = evsel__new(&attr);
 338         if (evsel == NULL)
 339                 goto out;
 340
 341         arch_evsel__fixup_new_cycles(&evsel->core.attr);
 342
 343         evsel->precise_max = true;
 344
 345         /* use asprintf() because free(evsel) assumes name is allocated */
 346         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 347                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 348                      attr.exclude_kernel ? "u" : "",
 349                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 350                 goto error_free;
 351 out:
 352         return evsel;
 353 error_free:
 354         evsel__delete(evsel);
 355         evsel = NULL;
 356         goto out;
 357 }
 358
 359 int copy_config_terms(struct list_head *dst, struct list_head *src)
 360 {
 361         struct evsel_config_term *pos, *tmp;
 362
 363         list_for_each_entry(pos, src, list) {
 364                 tmp = malloc(sizeof(*tmp));
 365                 if (tmp == NULL)
 366                         return -ENOMEM;
 367
 368                 *tmp = *pos;
 369                 if (tmp->free_str) {
 370                         tmp->val.str = strdup(pos->val.str);
 371                         if (tmp->val.str == NULL) {
 372                                 free(tmp);
 373                                 return -ENOMEM;
 374                         }
 375                 }
 376                 list_add_tail(&tmp->list, dst);
 377         }
 378         return 0;
 379 }
 380
 381 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src)
 382 {
 383         return copy_config_terms(&dst->config_terms, &src->config_terms);
 384 }
 385
 386 /**
 387  * evsel__clone - create a new evsel copied from @orig
 388  * @orig: original evsel
 389  *
 390  * The assumption is that @orig is not configured nor opened yet.
 391  * So we only care about the attributes that can be set while it's parsed.
 392  */
 393 struct evsel *evsel__clone(struct evsel *orig)
 394 {
 395         struct evsel *evsel;
 396
 397         BUG_ON(orig->core.fd);
 398         BUG_ON(orig->counts);
 399         BUG_ON(orig->priv);
 400         BUG_ON(orig->per_pkg_mask);
 401
 402         /* cannot handle BPF objects for now */
 403         if (orig->bpf_obj)
 404                 return NULL;
 405
 406         evsel = evsel__new(&orig->core.attr);
 407         if (evsel == NULL)
 408                 return NULL;
 409
 410         evsel->core.cpus = perf_cpu_map__get(orig->core.cpus);
 411         evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus);
 412         evsel->core.threads = perf_thread_map__get(orig->core.threads);
 413         evsel->core.nr_members = orig->core.nr_members;
 414         evsel->core.system_wide = orig->core.system_wide;
 415         evsel->core.requires_cpu = orig->core.requires_cpu;
 416
 417         if (orig->name) {
 418                 evsel->name = strdup(orig->name);
 419                 if (evsel->name == NULL)
 420                         goto out_err;
 421         }
 422         if (orig->group_name) {
 423                 evsel->group_name = strdup(orig->group_name);
 424                 if (evsel->group_name == NULL)
 425                         goto out_err;
 426         }
 427         if (orig->pmu_name) {
 428                 evsel->pmu_name = strdup(orig->pmu_name);
 429                 if (evsel->pmu_name == NULL)
 430                         goto out_err;
 431         }
 432         if (orig->filter) {
 433                 evsel->filter = strdup(orig->filter);
 434                 if (evsel->filter == NULL)
 435                         goto out_err;
 436         }
 437         if (orig->metric_id) {
 438                 evsel->metric_id = strdup(orig->metric_id);
 439                 if (evsel->metric_id == NULL)
 440                         goto out_err;
 441         }
 442         evsel->cgrp = cgroup__get(orig->cgrp);
 443 #ifdef HAVE_LIBTRACEEVENT
 444         evsel->tp_format = orig->tp_format;
 445 #endif
 446         evsel->handler = orig->handler;
 447         evsel->core.leader = orig->core.leader;
 448
 449         evsel->max_events = orig->max_events;
 450         evsel->tool_event = orig->tool_event;
 451         free((char *)evsel->unit);
 452         evsel->unit = strdup(orig->unit);
 453         if (evsel->unit == NULL)
 454                 goto out_err;
 455
 456         evsel->scale = orig->scale;
 457         evsel->snapshot = orig->snapshot;
 458         evsel->per_pkg = orig->per_pkg;
 459         evsel->percore = orig->percore;
 460         evsel->precise_max = orig->precise_max;
 461         evsel->use_uncore_alias = orig->use_uncore_alias;
 462         evsel->is_libpfm_event = orig->is_libpfm_event;
 463
 464         evsel->exclude_GH = orig->exclude_GH;
 465         evsel->sample_read = orig->sample_read;
 466         evsel->auto_merge_stats = orig->auto_merge_stats;
 467         evsel->collect_stat = orig->collect_stat;
 468         evsel->weak_group = orig->weak_group;
 469         evsel->use_config_name = orig->use_config_name;
 470         evsel->pmu = orig->pmu;
 471
 472         if (evsel__copy_config_terms(evsel, orig) < 0)
 473                 goto out_err;
 474
 475         return evsel;
 476
 477 out_err:
 478         evsel__delete(evsel);
 479         return NULL;
 480 }
 481
 482 /*
 483  * Returns pointer with encoded error via <linux/err.h> interface.
 484  */
 485 #ifdef HAVE_LIBTRACEEVENT
 486 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx)
 487 {
 488         struct evsel *evsel = zalloc(perf_evsel__object.size);
 489         int err = -ENOMEM;
 490
 491         if (evsel == NULL) {
 492                 goto out_err;
 493         } else {
 494                 struct perf_event_attr attr = {
 495                         .type          = PERF_TYPE_TRACEPOINT,
 496                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 497                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 498                 };
 499
 500                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 501                         goto out_free;
 502
 503                 evsel->tp_format = trace_event__tp_format(sys, name);
 504                 if (IS_ERR(evsel->tp_format)) {
 505                         err = PTR_ERR(evsel->tp_format);
 506                         goto out_free;
 507                 }
 508
 509                 event_attr_init(&attr);
 510                 attr.config = evsel->tp_format->id;
 511                 attr.sample_period = 1;
 512                 evsel__init(evsel, &attr, idx);
 513         }
 514
 515         return evsel;
 516
 517 out_free:
 518         zfree(&evsel->name);
 519         free(evsel);
 520 out_err:
 521         return ERR_PTR(err);
 522 }
 523 #endif
 524
 525 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = {
 526         "cycles",
 527         "instructions",
 528         "cache-references",
 529         "cache-misses",
 530         "branches",
 531         "branch-misses",
 532         "bus-cycles",
 533         "stalled-cycles-frontend",
 534         "stalled-cycles-backend",
 535         "ref-cycles",
 536 };
 537
 538 char *evsel__bpf_counter_events;
 539
 540 bool evsel__match_bpf_counter_events(const char *name)
 541 {
 542         int name_len;
 543         bool match;
 544         char *ptr;
 545
 546         if (!evsel__bpf_counter_events)
 547                 return false;
 548
 549         ptr = strstr(evsel__bpf_counter_events, name);
 550         name_len = strlen(name);
 551
 552         /* check name matches a full token in evsel__bpf_counter_events */
 553         match = (ptr != NULL) &&
 554                 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) &&
 555                 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0'));
 556
 557         return match;
 558 }
 559
 560 static const char *__evsel__hw_name(u64 config)
 561 {
 562         if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config])
 563                 return evsel__hw_names[config];
 564
 565         return "unknown-hardware";
 566 }
 567
 568 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size)
 569 {
 570         int colon = 0, r = 0;
 571         struct perf_event_attr *attr = &evsel->core.attr;
 572         bool exclude_guest_default = false;
 573
 574 #define MOD_PRINT(context, mod) do {                                    \
 575                 if (!attr->exclude_##context) {                         \
 576                         if (!colon) colon = ++r;                        \
 577                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 578                 } } while(0)
 579
 580         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 581                 MOD_PRINT(kernel, 'k');
 582                 MOD_PRINT(user, 'u');
 583                 MOD_PRINT(hv, 'h');
 584                 exclude_guest_default = true;
 585         }
 586
 587         if (attr->precise_ip) {
 588                 if (!colon)
 589                         colon = ++r;
 590                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 591                 exclude_guest_default = true;
 592         }
 593
 594         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 595                 MOD_PRINT(host, 'H');
 596                 MOD_PRINT(guest, 'G');
 597         }
 598 #undef MOD_PRINT
 599         if (colon)
 600                 bf[colon - 1] = ':';
 601         return r;
 602 }
 603
 604 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 605 {
 606         return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config));
 607 }
 608
 609 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size)
 610 {
 611         int r = arch_evsel__hw_name(evsel, bf, size);
 612         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 613 }
 614
 615 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = {
 616         "cpu-clock",
 617         "task-clock",
 618         "page-faults",
 619         "context-switches",
 620         "cpu-migrations",
 621         "minor-faults",
 622         "major-faults",
 623         "alignment-faults",
 624         "emulation-faults",
 625         "dummy",
 626 };
 627
 628 static const char *__evsel__sw_name(u64 config)
 629 {
 630         if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config])
 631                 return evsel__sw_names[config];
 632         return "unknown-software";
 633 }
 634
 635 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size)
 636 {
 637         int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config));
 638         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 639 }
 640
 641 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size)
 642 {
 643         return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev));
 644 }
 645
 646 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 647 {
 648         int r;
 649
 650         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 651
 652         if (type & HW_BREAKPOINT_R)
 653                 r += scnprintf(bf + r, size - r, "r");
 654
 655         if (type & HW_BREAKPOINT_W)
 656                 r += scnprintf(bf + r, size - r, "w");
 657
 658         if (type & HW_BREAKPOINT_X)
 659                 r += scnprintf(bf + r, size - r, "x");
 660
 661         return r;
 662 }
 663
 664 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size)
 665 {
 666         struct perf_event_attr *attr = &evsel->core.attr;
 667         int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 668         return r + evsel__add_modifiers(evsel, bf + r, size - r);
 669 }
 670
 671 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = {
 672  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 673  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 674  { "LLC",       "L2",                                                   },
 675  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 676  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 677  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 678  { "node",                                                              },
 679 };
 680
 681 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = {
 682  { "load",      "loads",        "read",                                 },
 683  { "store",     "stores",       "write",                                },
 684  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 685 };
 686
 687 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = {
 688  { "refs",      "Reference",    "ops",          "access",               },
 689  { "misses",    "miss",                                                 },
 690 };
 691
 692 #define C(x)            PERF_COUNT_HW_CACHE_##x
 693 #define CACHE_READ      (1 << C(OP_READ))
 694 #define CACHE_WRITE     (1 << C(OP_WRITE))
 695 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 696 #define COP(x)          (1 << x)
 697
 698 /*
 699  * cache operation stat
 700  * L1I : Read and prefetch only
 701  * ITLB and BPU : Read-only
 702  */
 703 static const unsigned long evsel__hw_cache_stat[C(MAX)] = {
 704  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 705  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 706  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 707  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 708  [C(ITLB)]      = (CACHE_READ),
 709  [C(BPU)]       = (CACHE_READ),
 710  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 711 };
 712
 713 bool evsel__is_cache_op_valid(u8 type, u8 op)
 714 {
 715         if (evsel__hw_cache_stat[type] & COP(op))
 716                 return true;    /* valid */
 717         else
 718                 return false;   /* invalid */
 719 }
 720
 721 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size)
 722 {
 723         if (result) {
 724                 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0],
 725                                  evsel__hw_cache_op[op][0],
 726                                  evsel__hw_cache_result[result][0]);
 727         }
 728
 729         return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0],
 730                          evsel__hw_cache_op[op][1]);
 731 }
 732
 733 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size)
 734 {
 735         u8 op, result, type = (config >>  0) & 0xff;
 736         const char *err = "unknown-ext-hardware-cache-type";
 737
 738         if (type >= PERF_COUNT_HW_CACHE_MAX)
 739                 goto out_err;
 740
 741         op = (config >>  8) & 0xff;
 742         err = "unknown-ext-hardware-cache-op";
 743         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 744                 goto out_err;
 745
 746         result = (config >> 16) & 0xff;
 747         err = "unknown-ext-hardware-cache-result";
 748         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 749                 goto out_err;
 750
 751         err = "invalid-cache";
 752         if (!evsel__is_cache_op_valid(type, op))
 753                 goto out_err;
 754
 755         return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 756 out_err:
 757         return scnprintf(bf, size, "%s", err);
 758 }
 759
 760 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size)
 761 {
 762         int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size);
 763         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 764 }
 765
 766 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size)
 767 {
 768         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config);
 769         return ret + evsel__add_modifiers(evsel, bf + ret, size - ret);
 770 }
 771
 772 const char *evsel__name(struct evsel *evsel)
 773 {
 774         char bf[128];
 775
 776         if (!evsel)
 777                 goto out_unknown;
 778
 779         if (evsel->name)
 780                 return evsel->name;
 781
 782         switch (evsel->core.attr.type) {
 783         case PERF_TYPE_RAW:
 784                 evsel__raw_name(evsel, bf, sizeof(bf));
 785                 break;
 786
 787         case PERF_TYPE_HARDWARE:
 788                 evsel__hw_name(evsel, bf, sizeof(bf));
 789                 break;
 790
 791         case PERF_TYPE_HW_CACHE:
 792                 evsel__hw_cache_name(evsel, bf, sizeof(bf));
 793                 break;
 794
 795         case PERF_TYPE_SOFTWARE:
 796                 if (evsel__is_tool(evsel))
 797                         evsel__tool_name(evsel->tool_event, bf, sizeof(bf));
 798                 else
 799                         evsel__sw_name(evsel, bf, sizeof(bf));
 800                 break;
 801
 802         case PERF_TYPE_TRACEPOINT:
 803                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 804                 break;
 805
 806         case PERF_TYPE_BREAKPOINT:
 807                 evsel__bp_name(evsel, bf, sizeof(bf));
 808                 break;
 809
 810         default:
 811                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 812                           evsel->core.attr.type);
 813                 break;
 814         }
 815
 816         evsel->name = strdup(bf);
 817
 818         if (evsel->name)
 819                 return evsel->name;
 820 out_unknown:
 821         return "unknown";
 822 }
 823
 824 const char *evsel__group_pmu_name(const struct evsel *evsel)
 825 {
 826         const struct evsel *leader;
 827
 828         /* If the pmu_name is set use it. pmu_name isn't set for CPU and software events. */
 829         if (evsel->pmu_name)
 830                 return evsel->pmu_name;
 831         /*
 832          * Software events may be in a group with other uncore PMU events. Use
 833          * the pmu_name of the group leader to avoid breaking the software event
 834          * out of the group.
 835          *
 836          * Aux event leaders, like intel_pt, expect a group with events from
 837          * other PMUs, so substitute the AUX event's PMU in this case.
 838          */
 839         leader  = evsel__leader(evsel);
 840         if ((evsel->core.attr.type == PERF_TYPE_SOFTWARE || evsel__is_aux_event(leader)) &&
 841             leader->pmu_name) {
 842                 return leader->pmu_name;
 843         }
 844
 845         return "cpu";
 846 }
 847
 848 const char *evsel__metric_id(const struct evsel *evsel)
 849 {
 850         if (evsel->metric_id)
 851                 return evsel->metric_id;
 852
 853         if (evsel__is_tool(evsel))
 854                 return perf_tool_event__to_str(evsel->tool_event);
 855
 856         return "unknown";
 857 }
 858
 859 const char *evsel__group_name(struct evsel *evsel)
 860 {
 861         return evsel->group_name ?: "anon group";
 862 }
 863
 864 /*
 865  * Returns the group details for the specified leader,
 866  * with following rules.
 867  *
 868  *  For record -e '{cycles,instructions}'
 869  *    'anon group { cycles:u, instructions:u }'
 870  *
 871  *  For record -e 'cycles,instructions' and report --group
 872  *    'cycles:u, instructions:u'
 873  */
 874 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size)
 875 {
 876         int ret = 0;
 877         struct evsel *pos;
 878         const char *group_name = evsel__group_name(evsel);
 879
 880         if (!evsel->forced_leader)
 881                 ret = scnprintf(buf, size, "%s { ", group_name);
 882
 883         ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel));
 884
 885         for_each_group_member(pos, evsel)
 886                 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos));
 887
 888         if (!evsel->forced_leader)
 889                 ret += scnprintf(buf + ret, size - ret, " }");
 890
 891         return ret;
 892 }
 893
 894 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 895                                       struct callchain_param *param)
 896 {
 897         bool function = evsel__is_function_event(evsel);
 898         struct perf_event_attr *attr = &evsel->core.attr;
 899
 900         evsel__set_sample_bit(evsel, CALLCHAIN);
 901
 902         attr->sample_max_stack = param->max_stack;
 903
 904         if (opts->kernel_callchains)
 905                 attr->exclude_callchain_user = 1;
 906         if (opts->user_callchains)
 907                 attr->exclude_callchain_kernel = 1;
 908         if (param->record_mode == CALLCHAIN_LBR) {
 909                 if (!opts->branch_stack) {
 910                         if (attr->exclude_user) {
 911                                 pr_warning("LBR callstack option is only available "
 912                                            "to get user callchain information. "
 913                                            "Falling back to framepointers.\n");
 914                         } else {
 915                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
 916                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 917                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 918                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 919                                                         PERF_SAMPLE_BRANCH_NO_FLAGS |
 920                                                         PERF_SAMPLE_BRANCH_HW_INDEX;
 921                         }
 922                 } else
 923                          pr_warning("Cannot use LBR callstack with branch stack. "
 924                                     "Falling back to framepointers.\n");
 925         }
 926
 927         if (param->record_mode == CALLCHAIN_DWARF) {
 928                 if (!function) {
 929                         evsel__set_sample_bit(evsel, REGS_USER);
 930                         evsel__set_sample_bit(evsel, STACK_USER);
 931                         if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) {
 932                                 attr->sample_regs_user |= DWARF_MINIMAL_REGS;
 933                                 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, "
 934                                            "specifying a subset with --user-regs may render DWARF unwinding unreliable, "
 935                                            "so the minimal registers set (IP, SP) is explicitly forced.\n");
 936                         } else {
 937                                 attr->sample_regs_user |= arch__user_reg_mask();
 938                         }
 939                         attr->sample_stack_user = param->dump_size;
 940                         attr->exclude_callchain_user = 1;
 941                 } else {
 942                         pr_info("Cannot use DWARF unwind for function trace event,"
 943                                 " falling back to framepointers.\n");
 944                 }
 945         }
 946
 947         if (function) {
 948                 pr_info("Disabling user space callchains for function trace event.\n");
 949                 attr->exclude_callchain_user = 1;
 950         }
 951 }
 952
 953 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts,
 954                              struct callchain_param *param)
 955 {
 956         if (param->enabled)
 957                 return __evsel__config_callchain(evsel, opts, param);
 958 }
 959
 960 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param)
 961 {
 962         struct perf_event_attr *attr = &evsel->core.attr;
 963
 964         evsel__reset_sample_bit(evsel, CALLCHAIN);
 965         if (param->record_mode == CALLCHAIN_LBR) {
 966                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
 967                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 968                                               PERF_SAMPLE_BRANCH_CALL_STACK |
 969                                               PERF_SAMPLE_BRANCH_HW_INDEX);
 970         }
 971         if (param->record_mode == CALLCHAIN_DWARF) {
 972                 evsel__reset_sample_bit(evsel, REGS_USER);
 973                 evsel__reset_sample_bit(evsel, STACK_USER);
 974         }
 975 }
 976
 977 static void evsel__apply_config_terms(struct evsel *evsel,
 978                                       struct record_opts *opts, bool track)
 979 {
 980         struct evsel_config_term *term;
 981         struct list_head *config_terms = &evsel->config_terms;
 982         struct perf_event_attr *attr = &evsel->core.attr;
 983         /* callgraph default */
 984         struct callchain_param param = {
 985                 .record_mode = callchain_param.record_mode,
 986         };
 987         u32 dump_size = 0;
 988         int max_stack = 0;
 989         const char *callgraph_buf = NULL;
 990
 991         list_for_each_entry(term, config_terms, list) {
 992                 switch (term->type) {
 993                 case EVSEL__CONFIG_TERM_PERIOD:
 994                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 995                                 attr->sample_period = term->val.period;
 996                                 attr->freq = 0;
 997                                 evsel__reset_sample_bit(evsel, PERIOD);
 998                         }
 999                         break;
1000                 case EVSEL__CONFIG_TERM_FREQ:
1001                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
1002                                 attr->sample_freq = term->val.freq;
1003                                 attr->freq = 1;
1004                                 evsel__set_sample_bit(evsel, PERIOD);
1005                         }
1006                         break;
1007                 case EVSEL__CONFIG_TERM_TIME:
1008                         if (term->val.time)
1009                                 evsel__set_sample_bit(evsel, TIME);
1010                         else
1011                                 evsel__reset_sample_bit(evsel, TIME);
1012                         break;
1013                 case EVSEL__CONFIG_TERM_CALLGRAPH:
1014                         callgraph_buf = term->val.str;
1015                         break;
1016                 case EVSEL__CONFIG_TERM_BRANCH:
1017                         if (term->val.str && strcmp(term->val.str, "no")) {
1018                                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1019                                 parse_branch_str(term->val.str,
1020                                                  &attr->branch_sample_type);
1021                         } else
1022                                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1023                         break;
1024                 case EVSEL__CONFIG_TERM_STACK_USER:
1025                         dump_size = term->val.stack_user;
1026                         break;
1027                 case EVSEL__CONFIG_TERM_MAX_STACK:
1028                         max_stack = term->val.max_stack;
1029                         break;
1030                 case EVSEL__CONFIG_TERM_MAX_EVENTS:
1031                         evsel->max_events = term->val.max_events;
1032                         break;
1033                 case EVSEL__CONFIG_TERM_INHERIT:
1034                         /*
1035                          * attr->inherit should has already been set by
1036                          * evsel__config. If user explicitly set
1037                          * inherit using config terms, override global
1038                          * opt->no_inherit setting.
1039                          */
1040                         attr->inherit = term->val.inherit ? 1 : 0;
1041                         break;
1042                 case EVSEL__CONFIG_TERM_OVERWRITE:
1043                         attr->write_backward = term->val.overwrite ? 1 : 0;
1044                         break;
1045                 case EVSEL__CONFIG_TERM_DRV_CFG:
1046                         break;
1047                 case EVSEL__CONFIG_TERM_PERCORE:
1048                         break;
1049                 case EVSEL__CONFIG_TERM_AUX_OUTPUT:
1050                         attr->aux_output = term->val.aux_output ? 1 : 0;
1051                         break;
1052                 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE:
1053                         /* Already applied by auxtrace */
1054                         break;
1055                 case EVSEL__CONFIG_TERM_CFG_CHG:
1056                         break;
1057                 default:
1058                         break;
1059                 }
1060         }
1061
1062         /* User explicitly set per-event callgraph, clear the old setting and reset. */
1063         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
1064                 bool sample_address = false;
1065
1066                 if (max_stack) {
1067                         param.max_stack = max_stack;
1068                         if (callgraph_buf == NULL)
1069                                 callgraph_buf = "fp";
1070                 }
1071
1072                 /* parse callgraph parameters */
1073                 if (callgraph_buf != NULL) {
1074                         if (!strcmp(callgraph_buf, "no")) {
1075                                 param.enabled = false;
1076                                 param.record_mode = CALLCHAIN_NONE;
1077                         } else {
1078                                 param.enabled = true;
1079                                 if (parse_callchain_record(callgraph_buf, &param)) {
1080                                         pr_err("per-event callgraph setting for %s failed. "
1081                                                "Apply callgraph global setting for it\n",
1082                                                evsel->name);
1083                                         return;
1084                                 }
1085                                 if (param.record_mode == CALLCHAIN_DWARF)
1086                                         sample_address = true;
1087                         }
1088                 }
1089                 if (dump_size > 0) {
1090                         dump_size = round_up(dump_size, sizeof(u64));
1091                         param.dump_size = dump_size;
1092                 }
1093
1094                 /* If global callgraph set, clear it */
1095                 if (callchain_param.enabled)
1096                         evsel__reset_callgraph(evsel, &callchain_param);
1097
1098                 /* set perf-event callgraph */
1099                 if (param.enabled) {
1100                         if (sample_address) {
1101                                 evsel__set_sample_bit(evsel, ADDR);
1102                                 evsel__set_sample_bit(evsel, DATA_SRC);
1103                                 evsel->core.attr.mmap_data = track;
1104                         }
1105                         evsel__config_callchain(evsel, opts, &param);
1106                 }
1107         }
1108 }
1109
1110 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type)
1111 {
1112         struct evsel_config_term *term, *found_term = NULL;
1113
1114         list_for_each_entry(term, &evsel->config_terms, list) {
1115                 if (term->type == type)
1116                         found_term = term;
1117         }
1118
1119         return found_term;
1120 }
1121
1122 void __weak arch_evsel__set_sample_weight(struct evsel *evsel)
1123 {
1124         evsel__set_sample_bit(evsel, WEIGHT);
1125 }
1126
1127 void __weak arch_evsel__fixup_new_cycles(struct perf_event_attr *attr __maybe_unused)
1128 {
1129 }
1130
1131 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused,
1132                                     struct perf_event_attr *attr __maybe_unused)
1133 {
1134 }
1135
1136 static void evsel__set_default_freq_period(struct record_opts *opts,
1137                                            struct perf_event_attr *attr)
1138 {
1139         if (opts->freq) {
1140                 attr->freq = 1;
1141                 attr->sample_freq = opts->freq;
1142         } else {
1143                 attr->sample_period = opts->default_interval;
1144         }
1145 }
1146
1147 static bool evsel__is_offcpu_event(struct evsel *evsel)
1148 {
1149         return evsel__is_bpf_output(evsel) && !strcmp(evsel->name, OFFCPU_EVENT);
1150 }
1151
1152 /*
1153  * The enable_on_exec/disabled value strategy:
1154  *
1155  *  1) For any type of traced program:
1156  *    - all independent events and group leaders are disabled
1157  *    - all group members are enabled
1158  *
1159  *     Group members are ruled by group leaders. They need to
1160  *     be enabled, because the group scheduling relies on that.
1161  *
1162  *  2) For traced programs executed by perf:
1163  *     - all independent events and group leaders have
1164  *       enable_on_exec set
1165  *     - we don't specifically enable or disable any event during
1166  *       the record command
1167  *
1168  *     Independent events and group leaders are initially disabled
1169  *     and get enabled by exec. Group members are ruled by group
1170  *     leaders as stated in 1).
1171  *
1172  *  3) For traced programs attached by perf (pid/tid):
1173  *     - we specifically enable or disable all events during
1174  *       the record command
1175  *
1176  *     When attaching events to already running traced we
1177  *     enable/disable events specifically, as there's no
1178  *     initial traced exec call.
1179  */
1180 void evsel__config(struct evsel *evsel, struct record_opts *opts,
1181                    struct callchain_param *callchain)
1182 {
1183         struct evsel *leader = evsel__leader(evsel);
1184         struct perf_event_attr *attr = &evsel->core.attr;
1185         int track = evsel->tracking;
1186         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
1187
1188         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
1189         attr->inherit       = !opts->no_inherit;
1190         attr->write_backward = opts->overwrite ? 1 : 0;
1191         attr->read_format   = PERF_FORMAT_LOST;
1192
1193         evsel__set_sample_bit(evsel, IP);
1194         evsel__set_sample_bit(evsel, TID);
1195
1196         if (evsel->sample_read) {
1197                 evsel__set_sample_bit(evsel, READ);
1198
1199                 /*
1200                  * We need ID even in case of single event, because
1201                  * PERF_SAMPLE_READ process ID specific data.
1202                  */
1203                 evsel__set_sample_id(evsel, false);
1204
1205                 /*
1206                  * Apply group format only if we belong to group
1207                  * with more than one members.
1208                  */
1209                 if (leader->core.nr_members > 1) {
1210                         attr->read_format |= PERF_FORMAT_GROUP;
1211                         attr->inherit = 0;
1212                 }
1213         }
1214
1215         /*
1216          * We default some events to have a default interval. But keep
1217          * it a weak assumption overridable by the user.
1218          */
1219         if ((evsel->is_libpfm_event && !attr->sample_period) ||
1220             (!evsel->is_libpfm_event && (!attr->sample_period ||
1221                                          opts->user_freq != UINT_MAX ||
1222                                          opts->user_interval != ULLONG_MAX)))
1223                 evsel__set_default_freq_period(opts, attr);
1224
1225         /*
1226          * If attr->freq was set (here or earlier), ask for period
1227          * to be sampled.
1228          */
1229         if (attr->freq)
1230                 evsel__set_sample_bit(evsel, PERIOD);
1231
1232         if (opts->no_samples)
1233                 attr->sample_freq = 0;
1234
1235         if (opts->inherit_stat) {
1236                 evsel->core.attr.read_format |=
1237                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1238                         PERF_FORMAT_TOTAL_TIME_RUNNING |
1239                         PERF_FORMAT_ID;
1240                 attr->inherit_stat = 1;
1241         }
1242
1243         if (opts->sample_address) {
1244                 evsel__set_sample_bit(evsel, ADDR);
1245                 attr->mmap_data = track;
1246         }
1247
1248         /*
1249          * We don't allow user space callchains for  function trace
1250          * event, due to issues with page faults while tracing page
1251          * fault handler and its overall trickiness nature.
1252          */
1253         if (evsel__is_function_event(evsel))
1254                 evsel->core.attr.exclude_callchain_user = 1;
1255
1256         if (callchain && callchain->enabled && !evsel->no_aux_samples)
1257                 evsel__config_callchain(evsel, opts, callchain);
1258
1259         if (opts->sample_intr_regs && !evsel->no_aux_samples &&
1260             !evsel__is_dummy_event(evsel)) {
1261                 attr->sample_regs_intr = opts->sample_intr_regs;
1262                 evsel__set_sample_bit(evsel, REGS_INTR);
1263         }
1264
1265         if (opts->sample_user_regs && !evsel->no_aux_samples &&
1266             !evsel__is_dummy_event(evsel)) {
1267                 attr->sample_regs_user |= opts->sample_user_regs;
1268                 evsel__set_sample_bit(evsel, REGS_USER);
1269         }
1270
1271         if (target__has_cpu(&opts->target) || opts->sample_cpu)
1272                 evsel__set_sample_bit(evsel, CPU);
1273
1274         /*
1275          * When the user explicitly disabled time don't force it here.
1276          */
1277         if (opts->sample_time &&
1278             (!perf_missing_features.sample_id_all &&
1279             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
1280              opts->sample_time_set)))
1281                 evsel__set_sample_bit(evsel, TIME);
1282
1283         if (opts->raw_samples && !evsel->no_aux_samples) {
1284                 evsel__set_sample_bit(evsel, TIME);
1285                 evsel__set_sample_bit(evsel, RAW);
1286                 evsel__set_sample_bit(evsel, CPU);
1287         }
1288
1289         if (opts->sample_address)
1290                 evsel__set_sample_bit(evsel, DATA_SRC);
1291
1292         if (opts->sample_phys_addr)
1293                 evsel__set_sample_bit(evsel, PHYS_ADDR);
1294
1295         if (opts->no_buffering) {
1296                 attr->watermark = 0;
1297                 attr->wakeup_events = 1;
1298         }
1299         if (opts->branch_stack && !evsel->no_aux_samples) {
1300                 evsel__set_sample_bit(evsel, BRANCH_STACK);
1301                 attr->branch_sample_type = opts->branch_stack;
1302         }
1303
1304         if (opts->sample_weight)
1305                 arch_evsel__set_sample_weight(evsel);
1306
1307         attr->task     = track;
1308         attr->mmap     = track;
1309         attr->mmap2    = track && !perf_missing_features.mmap2;
1310         attr->comm     = track;
1311         attr->build_id = track && opts->build_id;
1312
1313         /*
1314          * ksymbol is tracked separately with text poke because it needs to be
1315          * system wide and enabled immediately.
1316          */
1317         if (!opts->text_poke)
1318                 attr->ksymbol = track && !perf_missing_features.ksymbol;
1319         attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf;
1320
1321         if (opts->record_namespaces)
1322                 attr->namespaces  = track;
1323
1324         if (opts->record_cgroup) {
1325                 attr->cgroup = track && !perf_missing_features.cgroup;
1326                 evsel__set_sample_bit(evsel, CGROUP);
1327         }
1328
1329         if (opts->sample_data_page_size)
1330                 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE);
1331
1332         if (opts->sample_code_page_size)
1333                 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE);
1334
1335         if (opts->record_switch_events)
1336                 attr->context_switch = track;
1337
1338         if (opts->sample_transaction)
1339                 evsel__set_sample_bit(evsel, TRANSACTION);
1340
1341         if (opts->running_time) {
1342                 evsel->core.attr.read_format |=
1343                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1344                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1345         }
1346
1347         /*
1348          * XXX see the function comment above
1349          *
1350          * Disabling only independent events or group leaders,
1351          * keeping group members enabled.
1352          */
1353         if (evsel__is_group_leader(evsel))
1354                 attr->disabled = 1;
1355
1356         /*
1357          * Setting enable_on_exec for independent events and
1358          * group leaders for traced executed by perf.
1359          */
1360         if (target__none(&opts->target) && evsel__is_group_leader(evsel) &&
1361             !opts->target.initial_delay)
1362                 attr->enable_on_exec = 1;
1363
1364         if (evsel->immediate) {
1365                 attr->disabled = 0;
1366                 attr->enable_on_exec = 0;
1367         }
1368
1369         clockid = opts->clockid;
1370         if (opts->use_clockid) {
1371                 attr->use_clockid = 1;
1372                 attr->clockid = opts->clockid;
1373         }
1374
1375         if (evsel->precise_max)
1376                 attr->precise_ip = 3;
1377
1378         if (opts->all_user) {
1379                 attr->exclude_kernel = 1;
1380                 attr->exclude_user   = 0;
1381         }
1382
1383         if (opts->all_kernel) {
1384                 attr->exclude_kernel = 0;
1385                 attr->exclude_user   = 1;
1386         }
1387
1388         if (evsel->core.own_cpus || evsel->unit)
1389                 evsel->core.attr.read_format |= PERF_FORMAT_ID;
1390
1391         /*
1392          * Apply event specific term settings,
1393          * it overloads any global configuration.
1394          */
1395         evsel__apply_config_terms(evsel, opts, track);
1396
1397         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1398
1399         /* The --period option takes the precedence. */
1400         if (opts->period_set) {
1401                 if (opts->period)
1402                         evsel__set_sample_bit(evsel, PERIOD);
1403                 else
1404                         evsel__reset_sample_bit(evsel, PERIOD);
1405         }
1406
1407         /*
1408          * A dummy event never triggers any actual counter and therefore
1409          * cannot be used with branch_stack.
1410          *
1411          * For initial_delay, a dummy event is added implicitly.
1412          * The software event will trigger -EOPNOTSUPP error out,
1413          * if BRANCH_STACK bit is set.
1414          */
1415         if (evsel__is_dummy_event(evsel))
1416                 evsel__reset_sample_bit(evsel, BRANCH_STACK);
1417
1418         if (evsel__is_offcpu_event(evsel))
1419                 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES;
1420
1421         arch__post_evsel_config(evsel, attr);
1422 }
1423
1424 int evsel__set_filter(struct evsel *evsel, const char *filter)
1425 {
1426         char *new_filter = strdup(filter);
1427
1428         if (new_filter != NULL) {
1429                 free(evsel->filter);
1430                 evsel->filter = new_filter;
1431                 return 0;
1432         }
1433
1434         return -1;
1435 }
1436
1437 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter)
1438 {
1439         char *new_filter;
1440
1441         if (evsel->filter == NULL)
1442                 return evsel__set_filter(evsel, filter);
1443
1444         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1445                 free(evsel->filter);
1446                 evsel->filter = new_filter;
1447                 return 0;
1448         }
1449
1450         return -1;
1451 }
1452
1453 int evsel__append_tp_filter(struct evsel *evsel, const char *filter)
1454 {
1455         return evsel__append_filter(evsel, "(%s) && (%s)", filter);
1456 }
1457
1458 int evsel__append_addr_filter(struct evsel *evsel, const char *filter)
1459 {
1460         return evsel__append_filter(evsel, "%s,%s", filter);
1461 }
1462
1463 /* Caller has to clear disabled after going through all CPUs. */
1464 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx)
1465 {
1466         return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx);
1467 }
1468
1469 int evsel__enable(struct evsel *evsel)
1470 {
1471         int err = perf_evsel__enable(&evsel->core);
1472
1473         if (!err)
1474                 evsel->disabled = false;
1475         return err;
1476 }
1477
1478 /* Caller has to set disabled after going through all CPUs. */
1479 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx)
1480 {
1481         return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx);
1482 }
1483
1484 int evsel__disable(struct evsel *evsel)
1485 {
1486         int err = perf_evsel__disable(&evsel->core);
1487         /*
1488          * We mark it disabled here so that tools that disable a event can
1489          * ignore events after they disable it. I.e. the ring buffer may have
1490          * already a few more events queued up before the kernel got the stop
1491          * request.
1492          */
1493         if (!err)
1494                 evsel->disabled = true;
1495
1496         return err;
1497 }
1498
1499 void free_config_terms(struct list_head *config_terms)
1500 {
1501         struct evsel_config_term *term, *h;
1502
1503         list_for_each_entry_safe(term, h, config_terms, list) {
1504                 list_del_init(&term->list);
1505                 if (term->free_str)
1506                         zfree(&term->val.str);
1507                 free(term);
1508         }
1509 }
1510
1511 static void evsel__free_config_terms(struct evsel *evsel)
1512 {
1513         free_config_terms(&evsel->config_terms);
1514 }
1515
1516 void evsel__exit(struct evsel *evsel)
1517 {
1518         assert(list_empty(&evsel->core.node));
1519         assert(evsel->evlist == NULL);
1520         bpf_counter__destroy(evsel);
1521         evsel__free_counts(evsel);
1522         perf_evsel__free_fd(&evsel->core);
1523         perf_evsel__free_id(&evsel->core);
1524         evsel__free_config_terms(evsel);
1525         cgroup__put(evsel->cgrp);
1526         perf_cpu_map__put(evsel->core.cpus);
1527         perf_cpu_map__put(evsel->core.own_cpus);
1528         perf_thread_map__put(evsel->core.threads);
1529         zfree(&evsel->group_name);
1530         zfree(&evsel->name);
1531         zfree(&evsel->pmu_name);
1532         zfree(&evsel->unit);
1533         zfree(&evsel->metric_id);
1534         evsel__zero_per_pkg(evsel);
1535         hashmap__free(evsel->per_pkg_mask);
1536         evsel->per_pkg_mask = NULL;
1537         zfree(&evsel->metric_events);
1538         perf_evsel__object.fini(evsel);
1539 }
1540
1541 void evsel__delete(struct evsel *evsel)
1542 {
1543         evsel__exit(evsel);
1544         free(evsel);
1545 }
1546
1547 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread,
1548                            struct perf_counts_values *count)
1549 {
1550         struct perf_counts_values tmp;
1551
1552         if (!evsel->prev_raw_counts)
1553                 return;
1554
1555         tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread);
1556         *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count;
1557
1558         count->val = count->val - tmp.val;
1559         count->ena = count->ena - tmp.ena;
1560         count->run = count->run - tmp.run;
1561 }
1562
1563 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
1564 {
1565         struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread);
1566
1567         return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
1568 }
1569
1570 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
1571                              u64 val, u64 ena, u64 run, u64 lost)
1572 {
1573         struct perf_counts_values *count;
1574
1575         count = perf_counts(counter->counts, cpu_map_idx, thread);
1576
1577         count->val    = val;
1578         count->ena    = ena;
1579         count->run    = run;
1580         count->lost   = lost;
1581
1582         perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
1583 }
1584
1585 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data)
1586 {
1587         u64 read_format = leader->core.attr.read_format;
1588         struct sample_read_value *v;
1589         u64 nr, ena = 0, run = 0, lost = 0;
1590
1591         nr = *data++;
1592
1593         if (nr != (u64) leader->core.nr_members)
1594                 return -EINVAL;
1595
1596         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1597                 ena = *data++;
1598
1599         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1600                 run = *data++;
1601
1602         v = (void *)data;
1603         sample_read_group__for_each(v, nr, read_format) {
1604                 struct evsel *counter;
1605
1606                 counter = evlist__id2evsel(leader->evlist, v->id);
1607                 if (!counter)
1608                         return -EINVAL;
1609
1610                 if (read_format & PERF_FORMAT_LOST)
1611                         lost = v->lost;
1612
1613                 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost);
1614         }
1615
1616         return 0;
1617 }
1618
1619 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread)
1620 {
1621         struct perf_stat_evsel *ps = leader->stats;
1622         u64 read_format = leader->core.attr.read_format;
1623         int size = perf_evsel__read_size(&leader->core);
1624         u64 *data = ps->group_data;
1625
1626         if (!(read_format & PERF_FORMAT_ID))
1627                 return -EINVAL;
1628
1629         if (!evsel__is_group_leader(leader))
1630                 return -EINVAL;
1631
1632         if (!data) {
1633                 data = zalloc(size);
1634                 if (!data)
1635                         return -ENOMEM;
1636
1637                 ps->group_data = data;
1638         }
1639
1640         if (FD(leader, cpu_map_idx, thread) < 0)
1641                 return -EINVAL;
1642
1643         if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0)
1644                 return -errno;
1645
1646         return evsel__process_group_data(leader, cpu_map_idx, thread, data);
1647 }
1648
1649 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
1650 {
1651         u64 read_format = evsel->core.attr.read_format;
1652
1653         if (read_format & PERF_FORMAT_GROUP)
1654                 return evsel__read_group(evsel, cpu_map_idx, thread);
1655
1656         return evsel__read_one(evsel, cpu_map_idx, thread);
1657 }
1658
1659 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale)
1660 {
1661         struct perf_counts_values count;
1662         size_t nv = scale ? 3 : 1;
1663
1664         if (FD(evsel, cpu_map_idx, thread) < 0)
1665                 return -EINVAL;
1666
1667         if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0)
1668                 return -ENOMEM;
1669
1670         if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0)
1671                 return -errno;
1672
1673         evsel__compute_deltas(evsel, cpu_map_idx, thread, &count);
1674         perf_counts_values__scale(&count, scale, NULL);
1675         *perf_counts(evsel->counts, cpu_map_idx, thread) = count;
1676         return 0;
1677 }
1678
1679 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other,
1680                                   int cpu_map_idx)
1681 {
1682         struct perf_cpu cpu;
1683
1684         cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx);
1685         return perf_cpu_map__idx(other->core.cpus, cpu);
1686 }
1687
1688 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx)
1689 {
1690         struct evsel *leader = evsel__leader(evsel);
1691
1692         if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) ||
1693             (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) {
1694                 return evsel__match_other_cpu(evsel, leader, cpu_map_idx);
1695         }
1696
1697         return cpu_map_idx;
1698 }
1699
1700 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread)
1701 {
1702         struct evsel *leader = evsel__leader(evsel);
1703         int fd;
1704
1705         if (evsel__is_group_leader(evsel))
1706                 return -1;
1707
1708         /*
1709          * Leader must be already processed/open,
1710          * if not it's a bug.
1711          */
1712         BUG_ON(!leader->core.fd);
1713
1714         cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx);
1715         if (cpu_map_idx == -1)
1716                 return -1;
1717
1718         fd = FD(leader, cpu_map_idx, thread);
1719         BUG_ON(fd == -1);
1720
1721         return fd;
1722 }
1723
1724 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx)
1725 {
1726         for (int cpu = 0; cpu < nr_cpus; cpu++)
1727                 for (int thread = thread_idx; thread < nr_threads - 1; thread++)
1728                         FD(pos, cpu, thread) = FD(pos, cpu, thread + 1);
1729 }
1730
1731 static int update_fds(struct evsel *evsel,
1732                       int nr_cpus, int cpu_map_idx,
1733                       int nr_threads, int thread_idx)
1734 {
1735         struct evsel *pos;
1736
1737         if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads)
1738                 return -EINVAL;
1739
1740         evlist__for_each_entry(evsel->evlist, pos) {
1741                 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx;
1742
1743                 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx);
1744
1745                 /*
1746                  * Since fds for next evsel has not been created,
1747                  * there is no need to iterate whole event list.
1748                  */
1749                 if (pos == evsel)
1750                         break;
1751         }
1752         return 0;
1753 }
1754
1755 static bool evsel__ignore_missing_thread(struct evsel *evsel,
1756                                          int nr_cpus, int cpu_map_idx,
1757                                          struct perf_thread_map *threads,
1758                                          int thread, int err)
1759 {
1760         pid_t ignore_pid = perf_thread_map__pid(threads, thread);
1761
1762         if (!evsel->ignore_missing_thread)
1763                 return false;
1764
1765         /* The system wide setup does not work with threads. */
1766         if (evsel->core.system_wide)
1767                 return false;
1768
1769         /* The -ESRCH is perf event syscall errno for pid's not found. */
1770         if (err != -ESRCH)
1771                 return false;
1772
1773         /* If there's only one thread, let it fail. */
1774         if (threads->nr == 1)
1775                 return false;
1776
1777         /*
1778          * We should remove fd for missing_thread first
1779          * because thread_map__remove() will decrease threads->nr.
1780          */
1781         if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread))
1782                 return false;
1783
1784         if (thread_map__remove(threads, thread))
1785                 return false;
1786
1787         pr_warning("WARNING: Ignored open failure for pid %d\n",
1788                    ignore_pid);
1789         return true;
1790 }
1791
1792 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1793                                 void *priv __maybe_unused)
1794 {
1795         return fprintf(fp, "  %-32s %s\n", name, val);
1796 }
1797
1798 static void display_attr(struct perf_event_attr *attr)
1799 {
1800         if (verbose >= 2 || debug_peo_args) {
1801                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1802                 fprintf(stderr, "perf_event_attr:\n");
1803                 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL);
1804                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1805         }
1806 }
1807
1808 bool evsel__precise_ip_fallback(struct evsel *evsel)
1809 {
1810         /* Do not try less precise if not requested. */
1811         if (!evsel->precise_max)
1812                 return false;
1813
1814         /*
1815          * We tried all the precise_ip values, and it's
1816          * still failing, so leave it to standard fallback.
1817          */
1818         if (!evsel->core.attr.precise_ip) {
1819                 evsel->core.attr.precise_ip = evsel->precise_ip_original;
1820                 return false;
1821         }
1822
1823         if (!evsel->precise_ip_original)
1824                 evsel->precise_ip_original = evsel->core.attr.precise_ip;
1825
1826         evsel->core.attr.precise_ip--;
1827         pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip);
1828         display_attr(&evsel->core.attr);
1829         return true;
1830 }
1831
1832 static struct perf_cpu_map *empty_cpu_map;
1833 static struct perf_thread_map *empty_thread_map;
1834
1835 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1836                 struct perf_thread_map *threads)
1837 {
1838         int nthreads = perf_thread_map__nr(threads);
1839
1840         if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) ||
1841             (perf_missing_features.aux_output     && evsel->core.attr.aux_output))
1842                 return -EINVAL;
1843
1844         if (cpus == NULL) {
1845                 if (empty_cpu_map == NULL) {
1846                         empty_cpu_map = perf_cpu_map__dummy_new();
1847                         if (empty_cpu_map == NULL)
1848                                 return -ENOMEM;
1849                 }
1850
1851                 cpus = empty_cpu_map;
1852         }
1853
1854         if (threads == NULL) {
1855                 if (empty_thread_map == NULL) {
1856                         empty_thread_map = thread_map__new_by_tid(-1);
1857                         if (empty_thread_map == NULL)
1858                                 return -ENOMEM;
1859                 }
1860
1861                 threads = empty_thread_map;
1862         }
1863
1864         if (evsel->core.fd == NULL &&
1865             perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0)
1866                 return -ENOMEM;
1867
1868         evsel->open_flags = PERF_FLAG_FD_CLOEXEC;
1869         if (evsel->cgrp)
1870                 evsel->open_flags |= PERF_FLAG_PID_CGROUP;
1871
1872         return 0;
1873 }
1874
1875 static void evsel__disable_missing_features(struct evsel *evsel)
1876 {
1877         if (perf_missing_features.read_lost)
1878                 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST;
1879         if (perf_missing_features.weight_struct) {
1880                 evsel__set_sample_bit(evsel, WEIGHT);
1881                 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT);
1882         }
1883         if (perf_missing_features.clockid_wrong)
1884                 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */
1885         if (perf_missing_features.clockid) {
1886                 evsel->core.attr.use_clockid = 0;
1887                 evsel->core.attr.clockid = 0;
1888         }
1889         if (perf_missing_features.cloexec)
1890                 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1891         if (perf_missing_features.mmap2)
1892                 evsel->core.attr.mmap2 = 0;
1893         if (evsel->pmu && evsel->pmu->missing_features.exclude_guest)
1894                 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0;
1895         if (perf_missing_features.lbr_flags)
1896                 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1897                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1898         if (perf_missing_features.group_read && evsel->core.attr.inherit)
1899                 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1900         if (perf_missing_features.ksymbol)
1901                 evsel->core.attr.ksymbol = 0;
1902         if (perf_missing_features.bpf)
1903                 evsel->core.attr.bpf_event = 0;
1904         if (perf_missing_features.branch_hw_idx)
1905                 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX;
1906         if (perf_missing_features.sample_id_all)
1907                 evsel->core.attr.sample_id_all = 0;
1908 }
1909
1910 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus,
1911                         struct perf_thread_map *threads)
1912 {
1913         int err;
1914
1915         err = __evsel__prepare_open(evsel, cpus, threads);
1916         if (err)
1917                 return err;
1918
1919         evsel__disable_missing_features(evsel);
1920
1921         return err;
1922 }
1923
1924 bool evsel__detect_missing_features(struct evsel *evsel)
1925 {
1926         /*
1927          * Must probe features in the order they were added to the
1928          * perf_event_attr interface.
1929          */
1930         if (!perf_missing_features.read_lost &&
1931             (evsel->core.attr.read_format & PERF_FORMAT_LOST)) {
1932                 perf_missing_features.read_lost = true;
1933                 pr_debug2("switching off PERF_FORMAT_LOST support\n");
1934                 return true;
1935         } else if (!perf_missing_features.weight_struct &&
1936             (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) {
1937                 perf_missing_features.weight_struct = true;
1938                 pr_debug2("switching off weight struct support\n");
1939                 return true;
1940         } else if (!perf_missing_features.code_page_size &&
1941             (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) {
1942                 perf_missing_features.code_page_size = true;
1943                 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n");
1944                 return false;
1945         } else if (!perf_missing_features.data_page_size &&
1946             (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) {
1947                 perf_missing_features.data_page_size = true;
1948                 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n");
1949                 return false;
1950         } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) {
1951                 perf_missing_features.cgroup = true;
1952                 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n");
1953                 return false;
1954         } else if (!perf_missing_features.branch_hw_idx &&
1955             (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) {
1956                 perf_missing_features.branch_hw_idx = true;
1957                 pr_debug2("switching off branch HW index support\n");
1958                 return true;
1959         } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) {
1960                 perf_missing_features.aux_output = true;
1961                 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n");
1962                 return false;
1963         } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) {
1964                 perf_missing_features.bpf = true;
1965                 pr_debug2_peo("switching off bpf_event\n");
1966                 return true;
1967         } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) {
1968                 perf_missing_features.ksymbol = true;
1969                 pr_debug2_peo("switching off ksymbol\n");
1970                 return true;
1971         } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) {
1972                 perf_missing_features.write_backward = true;
1973                 pr_debug2_peo("switching off write_backward\n");
1974                 return false;
1975         } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) {
1976                 perf_missing_features.clockid_wrong = true;
1977                 pr_debug2_peo("switching off clockid\n");
1978                 return true;
1979         } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) {
1980                 perf_missing_features.clockid = true;
1981                 pr_debug2_peo("switching off use_clockid\n");
1982                 return true;
1983         } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) {
1984                 perf_missing_features.cloexec = true;
1985                 pr_debug2_peo("switching off cloexec flag\n");
1986                 return true;
1987         } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) {
1988                 perf_missing_features.mmap2 = true;
1989                 pr_debug2_peo("switching off mmap2\n");
1990                 return true;
1991         } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) {
1992                 if (evsel->pmu == NULL)
1993                         evsel->pmu = evsel__find_pmu(evsel);
1994
1995                 if (evsel->pmu)
1996                         evsel->pmu->missing_features.exclude_guest = true;
1997                 else {
1998                         /* we cannot find PMU, disable attrs now */
1999                         evsel->core.attr.exclude_host = false;
2000                         evsel->core.attr.exclude_guest = false;
2001                 }
2002
2003                 if (evsel->exclude_GH) {
2004                         pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n");
2005                         return false;
2006                 }
2007                 if (!perf_missing_features.exclude_guest) {
2008                         perf_missing_features.exclude_guest = true;
2009                         pr_debug2_peo("switching off exclude_guest, exclude_host\n");
2010                 }
2011                 return true;
2012         } else if (!perf_missing_features.sample_id_all) {
2013                 perf_missing_features.sample_id_all = true;
2014                 pr_debug2_peo("switching off sample_id_all\n");
2015                 return true;
2016         } else if (!perf_missing_features.lbr_flags &&
2017                         (evsel->core.attr.branch_sample_type &
2018                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
2019                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
2020                 perf_missing_features.lbr_flags = true;
2021                 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n");
2022                 return true;
2023         } else if (!perf_missing_features.group_read &&
2024                     evsel->core.attr.inherit &&
2025                    (evsel->core.attr.read_format & PERF_FORMAT_GROUP) &&
2026                    evsel__is_group_leader(evsel)) {
2027                 perf_missing_features.group_read = true;
2028                 pr_debug2_peo("switching off group read\n");
2029                 return true;
2030         } else {
2031                 return false;
2032         }
2033 }
2034
2035 bool evsel__increase_rlimit(enum rlimit_action *set_rlimit)
2036 {
2037         int old_errno;
2038         struct rlimit l;
2039
2040         if (*set_rlimit < INCREASED_MAX) {
2041                 old_errno = errno;
2042
2043                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
2044                         if (*set_rlimit == NO_CHANGE) {
2045                                 l.rlim_cur = l.rlim_max;
2046                         } else {
2047                                 l.rlim_cur = l.rlim_max + 1000;
2048                                 l.rlim_max = l.rlim_cur;
2049                         }
2050                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
2051                                 (*set_rlimit) += 1;
2052                                 errno = old_errno;
2053                                 return true;
2054                         }
2055                 }
2056                 errno = old_errno;
2057         }
2058
2059         return false;
2060 }
2061
2062 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
2063                 struct perf_thread_map *threads,
2064                 int start_cpu_map_idx, int end_cpu_map_idx)
2065 {
2066         int idx, thread, nthreads;
2067         int pid = -1, err, old_errno;
2068         enum rlimit_action set_rlimit = NO_CHANGE;
2069
2070         err = __evsel__prepare_open(evsel, cpus, threads);
2071         if (err)
2072                 return err;
2073
2074         if (cpus == NULL)
2075                 cpus = empty_cpu_map;
2076
2077         if (threads == NULL)
2078                 threads = empty_thread_map;
2079
2080         nthreads = perf_thread_map__nr(threads);
2081
2082         if (evsel->cgrp)
2083                 pid = evsel->cgrp->fd;
2084
2085 fallback_missing_features:
2086         evsel__disable_missing_features(evsel);
2087
2088         display_attr(&evsel->core.attr);
2089
2090         for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) {
2091
2092                 for (thread = 0; thread < nthreads; thread++) {
2093                         int fd, group_fd;
2094 retry_open:
2095                         if (thread >= nthreads)
2096                                 break;
2097
2098                         if (!evsel->cgrp && !evsel->core.system_wide)
2099                                 pid = perf_thread_map__pid(threads, thread);
2100
2101                         group_fd = get_group_fd(evsel, idx, thread);
2102
2103                         test_attr__ready();
2104
2105                         /* Debug message used by test scripts */
2106                         pr_debug2_peo("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
2107                                 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags);
2108
2109                         fd = sys_perf_event_open(&evsel->core.attr, pid,
2110                                                 perf_cpu_map__cpu(cpus, idx).cpu,
2111                                                 group_fd, evsel->open_flags);
2112
2113                         FD(evsel, idx, thread) = fd;
2114
2115                         if (fd < 0) {
2116                                 err = -errno;
2117
2118                                 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n",
2119                                           err);
2120                                 goto try_fallback;
2121                         }
2122
2123                         bpf_counter__install_pe(evsel, idx, fd);
2124
2125                         if (unlikely(test_attr__enabled)) {
2126                                 test_attr__open(&evsel->core.attr, pid,
2127                                                 perf_cpu_map__cpu(cpus, idx),
2128                                                 fd, group_fd, evsel->open_flags);
2129                         }
2130
2131                         /* Debug message used by test scripts */
2132                         pr_debug2_peo(" = %d\n", fd);
2133
2134                         if (evsel->bpf_fd >= 0) {
2135                                 int evt_fd = fd;
2136                                 int bpf_fd = evsel->bpf_fd;
2137
2138                                 err = ioctl(evt_fd,
2139                                             PERF_EVENT_IOC_SET_BPF,
2140                                             bpf_fd);
2141                                 if (err && errno != EEXIST) {
2142                                         pr_err("failed to attach bpf fd %d: %s\n",
2143                                                bpf_fd, strerror(errno));
2144                                         err = -EINVAL;
2145                                         goto out_close;
2146                                 }
2147                         }
2148
2149                         set_rlimit = NO_CHANGE;
2150
2151                         /*
2152                          * If we succeeded but had to kill clockid, fail and
2153                          * have evsel__open_strerror() print us a nice error.
2154                          */
2155                         if (perf_missing_features.clockid ||
2156                             perf_missing_features.clockid_wrong) {
2157                                 err = -EINVAL;
2158                                 goto out_close;
2159                         }
2160                 }
2161         }
2162
2163         return 0;
2164
2165 try_fallback:
2166         if (evsel__precise_ip_fallback(evsel))
2167                 goto retry_open;
2168
2169         if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus),
2170                                          idx, threads, thread, err)) {
2171                 /* We just removed 1 thread, so lower the upper nthreads limit. */
2172                 nthreads--;
2173
2174                 /* ... and pretend like nothing have happened. */
2175                 err = 0;
2176                 goto retry_open;
2177         }
2178         /*
2179          * perf stat needs between 5 and 22 fds per CPU. When we run out
2180          * of them try to increase the limits.
2181          */
2182         if (err == -EMFILE && evsel__increase_rlimit(&set_rlimit))
2183                 goto retry_open;
2184
2185         if (err != -EINVAL || idx > 0 || thread > 0)
2186                 goto out_close;
2187
2188         if (evsel__detect_missing_features(evsel))
2189                 goto fallback_missing_features;
2190 out_close:
2191         if (err)
2192                 threads->err_thread = thread;
2193
2194         old_errno = errno;
2195         do {
2196                 while (--thread >= 0) {
2197                         if (FD(evsel, idx, thread) >= 0)
2198                                 close(FD(evsel, idx, thread));
2199                         FD(evsel, idx, thread) = -1;
2200                 }
2201                 thread = nthreads;
2202         } while (--idx >= 0);
2203         errno = old_errno;
2204         return err;
2205 }
2206
2207 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
2208                 struct perf_thread_map *threads)
2209 {
2210         return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus));
2211 }
2212
2213 void evsel__close(struct evsel *evsel)
2214 {
2215         perf_evsel__close(&evsel->core);
2216         perf_evsel__free_id(&evsel->core);
2217 }
2218
2219 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx)
2220 {
2221         if (cpu_map_idx == -1)
2222                 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus));
2223
2224         return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1);
2225 }
2226
2227 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads)
2228 {
2229         return evsel__open(evsel, NULL, threads);
2230 }
2231
2232 static int perf_evsel__parse_id_sample(const struct evsel *evsel,
2233                                        const union perf_event *event,
2234                                        struct perf_sample *sample)
2235 {
2236         u64 type = evsel->core.attr.sample_type;
2237         const __u64 *array = event->sample.array;
2238         bool swapped = evsel->needs_swap;
2239         union u64_swap u;
2240
2241         array += ((event->header.size -
2242                    sizeof(event->header)) / sizeof(u64)) - 1;
2243
2244         if (type & PERF_SAMPLE_IDENTIFIER) {
2245                 sample->id = *array;
2246                 array--;
2247         }
2248
2249         if (type & PERF_SAMPLE_CPU) {
2250                 u.val64 = *array;
2251                 if (swapped) {
2252                         /* undo swap of u64, then swap on individual u32s */
2253                         u.val64 = bswap_64(u.val64);
2254                         u.val32[0] = bswap_32(u.val32[0]);
2255                 }
2256
2257                 sample->cpu = u.val32[0];
2258                 array--;
2259         }
2260
2261         if (type & PERF_SAMPLE_STREAM_ID) {
2262                 sample->stream_id = *array;
2263                 array--;
2264         }
2265
2266         if (type & PERF_SAMPLE_ID) {
2267                 sample->id = *array;
2268                 array--;
2269         }
2270
2271         if (type & PERF_SAMPLE_TIME) {
2272                 sample->time = *array;
2273                 array--;
2274         }
2275
2276         if (type & PERF_SAMPLE_TID) {
2277                 u.val64 = *array;
2278                 if (swapped) {
2279                         /* undo swap of u64, then swap on individual u32s */
2280                         u.val64 = bswap_64(u.val64);
2281                         u.val32[0] = bswap_32(u.val32[0]);
2282                         u.val32[1] = bswap_32(u.val32[1]);
2283                 }
2284
2285                 sample->pid = u.val32[0];
2286                 sample->tid = u.val32[1];
2287                 array--;
2288         }
2289
2290         return 0;
2291 }
2292
2293 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
2294                             u64 size)
2295 {
2296         return size > max_size || offset + size > endp;
2297 }
2298
2299 #define OVERFLOW_CHECK(offset, size, max_size)                          \
2300         do {                                                            \
2301                 if (overflow(endp, (max_size), (offset), (size)))       \
2302                         return -EFAULT;                                 \
2303         } while (0)
2304
2305 #define OVERFLOW_CHECK_u64(offset) \
2306         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
2307
2308 static int
2309 perf_event__check_size(union perf_event *event, unsigned int sample_size)
2310 {
2311         /*
2312          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
2313          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
2314          * check the format does not go past the end of the event.
2315          */
2316         if (sample_size + sizeof(event->header) > event->header.size)
2317                 return -EFAULT;
2318
2319         return 0;
2320 }
2321
2322 void __weak arch_perf_parse_sample_weight(struct perf_sample *data,
2323                                           const __u64 *array,
2324                                           u64 type __maybe_unused)
2325 {
2326         data->weight = *array;
2327 }
2328
2329 u64 evsel__bitfield_swap_branch_flags(u64 value)
2330 {
2331         u64 new_val = 0;
2332
2333         /*
2334          * branch_flags
2335          * union {
2336          *      u64 values;
2337          *      struct {
2338          *              mispred:1       //target mispredicted
2339          *              predicted:1     //target predicted
2340          *              in_tx:1         //in transaction
2341          *              abort:1         //transaction abort
2342          *              cycles:16       //cycle count to last branch
2343          *              type:4          //branch type
2344          *              spec:2          //branch speculation info
2345          *              new_type:4      //additional branch type
2346          *              priv:3          //privilege level
2347          *              reserved:31
2348          *      }
2349          * }
2350          *
2351          * Avoid bswap64() the entire branch_flag.value,
2352          * as it has variable bit-field sizes. Instead the
2353          * macro takes the bit-field position/size,
2354          * swaps it based on the host endianness.
2355          */
2356         if (host_is_bigendian()) {
2357                 new_val = bitfield_swap(value, 0, 1);
2358                 new_val |= bitfield_swap(value, 1, 1);
2359                 new_val |= bitfield_swap(value, 2, 1);
2360                 new_val |= bitfield_swap(value, 3, 1);
2361                 new_val |= bitfield_swap(value, 4, 16);
2362                 new_val |= bitfield_swap(value, 20, 4);
2363                 new_val |= bitfield_swap(value, 24, 2);
2364                 new_val |= bitfield_swap(value, 26, 4);
2365                 new_val |= bitfield_swap(value, 30, 3);
2366                 new_val |= bitfield_swap(value, 33, 31);
2367         } else {
2368                 new_val = bitfield_swap(value, 63, 1);
2369                 new_val |= bitfield_swap(value, 62, 1);
2370                 new_val |= bitfield_swap(value, 61, 1);
2371                 new_val |= bitfield_swap(value, 60, 1);
2372                 new_val |= bitfield_swap(value, 44, 16);
2373                 new_val |= bitfield_swap(value, 40, 4);
2374                 new_val |= bitfield_swap(value, 38, 2);
2375                 new_val |= bitfield_swap(value, 34, 4);
2376                 new_val |= bitfield_swap(value, 31, 3);
2377                 new_val |= bitfield_swap(value, 0, 31);
2378         }
2379
2380         return new_val;
2381 }
2382
2383 int evsel__parse_sample(struct evsel *evsel, union perf_event *event,
2384                         struct perf_sample *data)
2385 {
2386         u64 type = evsel->core.attr.sample_type;
2387         bool swapped = evsel->needs_swap;
2388         const __u64 *array;
2389         u16 max_size = event->header.size;
2390         const void *endp = (void *)event + max_size;
2391         u64 sz;
2392
2393         /*
2394          * used for cross-endian analysis. See git commit 65014ab3
2395          * for why this goofiness is needed.
2396          */
2397         union u64_swap u;
2398
2399         memset(data, 0, sizeof(*data));
2400         data->cpu = data->pid = data->tid = -1;
2401         data->stream_id = data->id = data->time = -1ULL;
2402         data->period = evsel->core.attr.sample_period;
2403         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
2404         data->misc    = event->header.misc;
2405         data->id = -1ULL;
2406         data->data_src = PERF_MEM_DATA_SRC_NONE;
2407         data->vcpu = -1;
2408
2409         if (event->header.type != PERF_RECORD_SAMPLE) {
2410                 if (!evsel->core.attr.sample_id_all)
2411                         return 0;
2412                 return perf_evsel__parse_id_sample(evsel, event, data);
2413         }
2414
2415         array = event->sample.array;
2416
2417         if (perf_event__check_size(event, evsel->sample_size))
2418                 return -EFAULT;
2419
2420         if (type & PERF_SAMPLE_IDENTIFIER) {
2421                 data->id = *array;
2422                 array++;
2423         }
2424
2425         if (type & PERF_SAMPLE_IP) {
2426                 data->ip = *array;
2427                 array++;
2428         }
2429
2430         if (type & PERF_SAMPLE_TID) {
2431                 u.val64 = *array;
2432                 if (swapped) {
2433                         /* undo swap of u64, then swap on individual u32s */
2434                         u.val64 = bswap_64(u.val64);
2435                         u.val32[0] = bswap_32(u.val32[0]);
2436                         u.val32[1] = bswap_32(u.val32[1]);
2437                 }
2438
2439                 data->pid = u.val32[0];
2440                 data->tid = u.val32[1];
2441                 array++;
2442         }
2443
2444         if (type & PERF_SAMPLE_TIME) {
2445                 data->time = *array;
2446                 array++;
2447         }
2448
2449         if (type & PERF_SAMPLE_ADDR) {
2450                 data->addr = *array;
2451                 array++;
2452         }
2453
2454         if (type & PERF_SAMPLE_ID) {
2455                 data->id = *array;
2456                 array++;
2457         }
2458
2459         if (type & PERF_SAMPLE_STREAM_ID) {
2460                 data->stream_id = *array;
2461                 array++;
2462         }
2463
2464         if (type & PERF_SAMPLE_CPU) {
2465
2466                 u.val64 = *array;
2467                 if (swapped) {
2468                         /* undo swap of u64, then swap on individual u32s */
2469                         u.val64 = bswap_64(u.val64);
2470                         u.val32[0] = bswap_32(u.val32[0]);
2471                 }
2472
2473                 data->cpu = u.val32[0];
2474                 array++;
2475         }
2476
2477         if (type & PERF_SAMPLE_PERIOD) {
2478                 data->period = *array;
2479                 array++;
2480         }
2481
2482         if (type & PERF_SAMPLE_READ) {
2483                 u64 read_format = evsel->core.attr.read_format;
2484
2485                 OVERFLOW_CHECK_u64(array);
2486                 if (read_format & PERF_FORMAT_GROUP)
2487                         data->read.group.nr = *array;
2488                 else
2489                         data->read.one.value = *array;
2490
2491                 array++;
2492
2493                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2494                         OVERFLOW_CHECK_u64(array);
2495                         data->read.time_enabled = *array;
2496                         array++;
2497                 }
2498
2499                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2500                         OVERFLOW_CHECK_u64(array);
2501                         data->read.time_running = *array;
2502                         array++;
2503                 }
2504
2505                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2506                 if (read_format & PERF_FORMAT_GROUP) {
2507                         const u64 max_group_nr = UINT64_MAX /
2508                                         sizeof(struct sample_read_value);
2509
2510                         if (data->read.group.nr > max_group_nr)
2511                                 return -EFAULT;
2512
2513                         sz = data->read.group.nr * sample_read_value_size(read_format);
2514                         OVERFLOW_CHECK(array, sz, max_size);
2515                         data->read.group.values =
2516                                         (struct sample_read_value *)array;
2517                         array = (void *)array + sz;
2518                 } else {
2519                         OVERFLOW_CHECK_u64(array);
2520                         data->read.one.id = *array;
2521                         array++;
2522
2523                         if (read_format & PERF_FORMAT_LOST) {
2524                                 OVERFLOW_CHECK_u64(array);
2525                                 data->read.one.lost = *array;
2526                                 array++;
2527                         }
2528                 }
2529         }
2530
2531         if (type & PERF_SAMPLE_CALLCHAIN) {
2532                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2533
2534                 OVERFLOW_CHECK_u64(array);
2535                 data->callchain = (struct ip_callchain *)array++;
2536                 if (data->callchain->nr > max_callchain_nr)
2537                         return -EFAULT;
2538                 sz = data->callchain->nr * sizeof(u64);
2539                 OVERFLOW_CHECK(array, sz, max_size);
2540                 array = (void *)array + sz;
2541         }
2542
2543         if (type & PERF_SAMPLE_RAW) {
2544                 OVERFLOW_CHECK_u64(array);
2545                 u.val64 = *array;
2546
2547                 /*
2548                  * Undo swap of u64, then swap on individual u32s,
2549                  * get the size of the raw area and undo all of the
2550                  * swap. The pevent interface handles endianness by
2551                  * itself.
2552                  */
2553                 if (swapped) {
2554                         u.val64 = bswap_64(u.val64);
2555                         u.val32[0] = bswap_32(u.val32[0]);
2556                         u.val32[1] = bswap_32(u.val32[1]);
2557                 }
2558                 data->raw_size = u.val32[0];
2559
2560                 /*
2561                  * The raw data is aligned on 64bits including the
2562                  * u32 size, so it's safe to use mem_bswap_64.
2563                  */
2564                 if (swapped)
2565                         mem_bswap_64((void *) array, data->raw_size);
2566
2567                 array = (void *)array + sizeof(u32);
2568
2569                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2570                 data->raw_data = (void *)array;
2571                 array = (void *)array + data->raw_size;
2572         }
2573
2574         if (type & PERF_SAMPLE_BRANCH_STACK) {
2575                 const u64 max_branch_nr = UINT64_MAX /
2576                                           sizeof(struct branch_entry);
2577                 struct branch_entry *e;
2578                 unsigned int i;
2579
2580                 OVERFLOW_CHECK_u64(array);
2581                 data->branch_stack = (struct branch_stack *)array++;
2582
2583                 if (data->branch_stack->nr > max_branch_nr)
2584                         return -EFAULT;
2585
2586                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2587                 if (evsel__has_branch_hw_idx(evsel)) {
2588                         sz += sizeof(u64);
2589                         e = &data->branch_stack->entries[0];
2590                 } else {
2591                         data->no_hw_idx = true;
2592                         /*
2593                          * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied,
2594                          * only nr and entries[] will be output by kernel.
2595                          */
2596                         e = (struct branch_entry *)&data->branch_stack->hw_idx;
2597                 }
2598
2599                 if (swapped) {
2600                         /*
2601                          * struct branch_flag does not have endian
2602                          * specific bit field definition. And bswap
2603                          * will not resolve the issue, since these
2604                          * are bit fields.
2605                          *
2606                          * evsel__bitfield_swap_branch_flags() uses a
2607                          * bitfield_swap macro to swap the bit position
2608                          * based on the host endians.
2609                          */
2610                         for (i = 0; i < data->branch_stack->nr; i++, e++)
2611                                 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value);
2612                 }
2613
2614                 OVERFLOW_CHECK(array, sz, max_size);
2615                 array = (void *)array + sz;
2616         }
2617
2618         if (type & PERF_SAMPLE_REGS_USER) {
2619                 OVERFLOW_CHECK_u64(array);
2620                 data->user_regs.abi = *array;
2621                 array++;
2622
2623                 if (data->user_regs.abi) {
2624                         u64 mask = evsel->core.attr.sample_regs_user;
2625
2626                         sz = hweight64(mask) * sizeof(u64);
2627                         OVERFLOW_CHECK(array, sz, max_size);
2628                         data->user_regs.mask = mask;
2629                         data->user_regs.regs = (u64 *)array;
2630                         array = (void *)array + sz;
2631                 }
2632         }
2633
2634         if (type & PERF_SAMPLE_STACK_USER) {
2635                 OVERFLOW_CHECK_u64(array);
2636                 sz = *array++;
2637
2638                 data->user_stack.offset = ((char *)(array - 1)
2639                                           - (char *) event);
2640
2641                 if (!sz) {
2642                         data->user_stack.size = 0;
2643                 } else {
2644                         OVERFLOW_CHECK(array, sz, max_size);
2645                         data->user_stack.data = (char *)array;
2646                         array = (void *)array + sz;
2647                         OVERFLOW_CHECK_u64(array);
2648                         data->user_stack.size = *array++;
2649                         if (WARN_ONCE(data->user_stack.size > sz,
2650                                       "user stack dump failure\n"))
2651                                 return -EFAULT;
2652                 }
2653         }
2654
2655         if (type & PERF_SAMPLE_WEIGHT_TYPE) {
2656                 OVERFLOW_CHECK_u64(array);
2657                 arch_perf_parse_sample_weight(data, array, type);
2658                 array++;
2659         }
2660
2661         if (type & PERF_SAMPLE_DATA_SRC) {
2662                 OVERFLOW_CHECK_u64(array);
2663                 data->data_src = *array;
2664                 array++;
2665         }
2666
2667         if (type & PERF_SAMPLE_TRANSACTION) {
2668                 OVERFLOW_CHECK_u64(array);
2669                 data->transaction = *array;
2670                 array++;
2671         }
2672
2673         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2674         if (type & PERF_SAMPLE_REGS_INTR) {
2675                 OVERFLOW_CHECK_u64(array);
2676                 data->intr_regs.abi = *array;
2677                 array++;
2678
2679                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2680                         u64 mask = evsel->core.attr.sample_regs_intr;
2681
2682                         sz = hweight64(mask) * sizeof(u64);
2683                         OVERFLOW_CHECK(array, sz, max_size);
2684                         data->intr_regs.mask = mask;
2685                         data->intr_regs.regs = (u64 *)array;
2686                         array = (void *)array + sz;
2687                 }
2688         }
2689
2690         data->phys_addr = 0;
2691         if (type & PERF_SAMPLE_PHYS_ADDR) {
2692                 data->phys_addr = *array;
2693                 array++;
2694         }
2695
2696         data->cgroup = 0;
2697         if (type & PERF_SAMPLE_CGROUP) {
2698                 data->cgroup = *array;
2699                 array++;
2700         }
2701
2702         data->data_page_size = 0;
2703         if (type & PERF_SAMPLE_DATA_PAGE_SIZE) {
2704                 data->data_page_size = *array;
2705                 array++;
2706         }
2707
2708         data->code_page_size = 0;
2709         if (type & PERF_SAMPLE_CODE_PAGE_SIZE) {
2710                 data->code_page_size = *array;
2711                 array++;
2712         }
2713
2714         if (type & PERF_SAMPLE_AUX) {
2715                 OVERFLOW_CHECK_u64(array);
2716                 sz = *array++;
2717
2718                 OVERFLOW_CHECK(array, sz, max_size);
2719                 /* Undo swap of data */
2720                 if (swapped)
2721                         mem_bswap_64((char *)array, sz);
2722                 data->aux_sample.size = sz;
2723                 data->aux_sample.data = (char *)array;
2724                 array = (void *)array + sz;
2725         }
2726
2727         return 0;
2728 }
2729
2730 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event,
2731                                   u64 *timestamp)
2732 {
2733         u64 type = evsel->core.attr.sample_type;
2734         const __u64 *array;
2735
2736         if (!(type & PERF_SAMPLE_TIME))
2737                 return -1;
2738
2739         if (event->header.type != PERF_RECORD_SAMPLE) {
2740                 struct perf_sample data = {
2741                         .time = -1ULL,
2742                 };
2743
2744                 if (!evsel->core.attr.sample_id_all)
2745                         return -1;
2746                 if (perf_evsel__parse_id_sample(evsel, event, &data))
2747                         return -1;
2748
2749                 *timestamp = data.time;
2750                 return 0;
2751         }
2752
2753         array = event->sample.array;
2754
2755         if (perf_event__check_size(event, evsel->sample_size))
2756                 return -EFAULT;
2757
2758         if (type & PERF_SAMPLE_IDENTIFIER)
2759                 array++;
2760
2761         if (type & PERF_SAMPLE_IP)
2762                 array++;
2763
2764         if (type & PERF_SAMPLE_TID)
2765                 array++;
2766
2767         if (type & PERF_SAMPLE_TIME)
2768                 *timestamp = *array;
2769
2770         return 0;
2771 }
2772
2773 u16 evsel__id_hdr_size(struct evsel *evsel)
2774 {
2775         u64 sample_type = evsel->core.attr.sample_type;
2776         u16 size = 0;
2777
2778         if (sample_type & PERF_SAMPLE_TID)
2779                 size += sizeof(u64);
2780
2781         if (sample_type & PERF_SAMPLE_TIME)
2782                 size += sizeof(u64);
2783
2784         if (sample_type & PERF_SAMPLE_ID)
2785                 size += sizeof(u64);
2786
2787         if (sample_type & PERF_SAMPLE_STREAM_ID)
2788                 size += sizeof(u64);
2789
2790         if (sample_type & PERF_SAMPLE_CPU)
2791                 size += sizeof(u64);
2792
2793         if (sample_type & PERF_SAMPLE_IDENTIFIER)
2794                 size += sizeof(u64);
2795
2796         return size;
2797 }
2798
2799 #ifdef HAVE_LIBTRACEEVENT
2800 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name)
2801 {
2802         return tep_find_field(evsel->tp_format, name);
2803 }
2804
2805 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name)
2806 {
2807         struct tep_format_field *field = evsel__field(evsel, name);
2808         int offset;
2809
2810         if (!field)
2811                 return NULL;
2812
2813         offset = field->offset;
2814
2815         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2816                 offset = *(int *)(sample->raw_data + field->offset);
2817                 offset &= 0xffff;
2818                 if (tep_field_is_relative(field->flags))
2819                         offset += field->offset + field->size;
2820         }
2821
2822         return sample->raw_data + offset;
2823 }
2824
2825 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample,
2826                          bool needs_swap)
2827 {
2828         u64 value;
2829         void *ptr = sample->raw_data + field->offset;
2830
2831         switch (field->size) {
2832         case 1:
2833                 return *(u8 *)ptr;
2834         case 2:
2835                 value = *(u16 *)ptr;
2836                 break;
2837         case 4:
2838                 value = *(u32 *)ptr;
2839                 break;
2840         case 8:
2841                 memcpy(&value, ptr, sizeof(u64));
2842                 break;
2843         default:
2844                 return 0;
2845         }
2846
2847         if (!needs_swap)
2848                 return value;
2849
2850         switch (field->size) {
2851         case 2:
2852                 return bswap_16(value);
2853         case 4:
2854                 return bswap_32(value);
2855         case 8:
2856                 return bswap_64(value);
2857         default:
2858                 return 0;
2859         }
2860
2861         return 0;
2862 }
2863
2864 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name)
2865 {
2866         struct tep_format_field *field = evsel__field(evsel, name);
2867
2868         if (!field)
2869                 return 0;
2870
2871         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2872 }
2873 #endif
2874
2875 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
2876 {
2877         int paranoid;
2878
2879         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2880             evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
2881             evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2882                 /*
2883                  * If it's cycles then fall back to hrtimer based
2884                  * cpu-clock-tick sw counter, which is always available even if
2885                  * no PMU support.
2886                  *
2887                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2888                  * b0a873e).
2889                  */
2890                 scnprintf(msg, msgsize, "%s",
2891 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2892
2893                 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
2894                 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
2895
2896                 zfree(&evsel->name);
2897                 return true;
2898         } else if (err == EACCES && !evsel->core.attr.exclude_kernel &&
2899                    (paranoid = perf_event_paranoid()) > 1) {
2900                 const char *name = evsel__name(evsel);
2901                 char *new_name;
2902                 const char *sep = ":";
2903
2904                 /* If event has exclude user then don't exclude kernel. */
2905                 if (evsel->core.attr.exclude_user)
2906                         return false;
2907
2908                 /* Is there already the separator in the name. */
2909                 if (strchr(name, '/') ||
2910                     (strchr(name, ':') && !evsel->is_libpfm_event))
2911                         sep = "";
2912
2913                 if (asprintf(&new_name, "%s%su", name, sep) < 0)
2914                         return false;
2915
2916                 if (evsel->name)
2917                         free(evsel->name);
2918                 evsel->name = new_name;
2919                 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying "
2920                           "to fall back to excluding kernel and hypervisor "
2921                           " samples", paranoid);
2922                 evsel->core.attr.exclude_kernel = 1;
2923                 evsel->core.attr.exclude_hv     = 1;
2924
2925                 return true;
2926         }
2927
2928         return false;
2929 }
2930
2931 static bool find_process(const char *name)
2932 {
2933         size_t len = strlen(name);
2934         DIR *dir;
2935         struct dirent *d;
2936         int ret = -1;
2937
2938         dir = opendir(procfs__mountpoint());
2939         if (!dir)
2940                 return false;
2941
2942         /* Walk through the directory. */
2943         while (ret && (d = readdir(dir)) != NULL) {
2944                 char path[PATH_MAX];
2945                 char *data;
2946                 size_t size;
2947
2948                 if ((d->d_type != DT_DIR) ||
2949                      !strcmp(".", d->d_name) ||
2950                      !strcmp("..", d->d_name))
2951                         continue;
2952
2953                 scnprintf(path, sizeof(path), "%s/%s/comm",
2954                           procfs__mountpoint(), d->d_name);
2955
2956                 if (filename__read_str(path, &data, &size))
2957                         continue;
2958
2959                 ret = strncmp(name, data, len);
2960                 free(data);
2961         }
2962
2963         closedir(dir);
2964         return ret ? false : true;
2965 }
2966
2967 static bool is_amd(const char *arch, const char *cpuid)
2968 {
2969         return arch && !strcmp("x86", arch) && cpuid && strstarts(cpuid, "AuthenticAMD");
2970 }
2971
2972 static bool is_amd_ibs(struct evsel *evsel)
2973 {
2974         return evsel->core.attr.precise_ip
2975             || (evsel->pmu_name && !strncmp(evsel->pmu_name, "ibs", 3));
2976 }
2977
2978 int evsel__open_strerror(struct evsel *evsel, struct target *target,
2979                          int err, char *msg, size_t size)
2980 {
2981         struct perf_env *env = evsel__env(evsel);
2982         const char *arch = perf_env__arch(env);
2983         const char *cpuid = perf_env__cpuid(env);
2984         char sbuf[STRERR_BUFSIZE];
2985         int printed = 0, enforced = 0;
2986
2987         switch (err) {
2988         case EPERM:
2989         case EACCES:
2990                 printed += scnprintf(msg + printed, size - printed,
2991                         "Access to performance monitoring and observability operations is limited.\n");
2992
2993                 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) {
2994                         if (enforced) {
2995                                 printed += scnprintf(msg + printed, size - printed,
2996                                         "Enforced MAC policy settings (SELinux) can limit access to performance\n"
2997                                         "monitoring and observability operations. Inspect system audit records for\n"
2998                                         "more perf_event access control information and adjusting the policy.\n");
2999                         }
3000                 }
3001
3002                 if (err == EPERM)
3003                         printed += scnprintf(msg, size,
3004                                 "No permission to enable %s event.\n\n", evsel__name(evsel));
3005
3006                 return scnprintf(msg + printed, size - printed,
3007                  "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n"
3008                  "access to performance monitoring and observability operations for processes\n"
3009                  "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n"
3010                  "More information can be found at 'Perf events and tool security' document:\n"
3011                  "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n"
3012                  "perf_event_paranoid setting is %d:\n"
3013                  "  -1: Allow use of (almost) all events by all users\n"
3014                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
3015                  ">= 0: Disallow raw and ftrace function tracepoint access\n"
3016                  ">= 1: Disallow CPU event access\n"
3017                  ">= 2: Disallow kernel profiling\n"
3018                  "To make the adjusted perf_event_paranoid setting permanent preserve it\n"
3019                  "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)",
3020                  perf_event_paranoid());
3021         case ENOENT:
3022                 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel));
3023         case EMFILE:
3024                 return scnprintf(msg, size, "%s",
3025                          "Too many events are opened.\n"
3026                          "Probably the maximum number of open file descriptors has been reached.\n"
3027                          "Hint: Try again after reducing the number of events.\n"
3028                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
3029         case ENOMEM:
3030                 if (evsel__has_callchain(evsel) &&
3031                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
3032                         return scnprintf(msg, size,
3033                                          "Not enough memory to setup event with callchain.\n"
3034                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
3035                                          "Hint: Current value: %d", sysctl__max_stack());
3036                 break;
3037         case ENODEV:
3038                 if (target->cpu_list)
3039                         return scnprintf(msg, size, "%s",
3040          "No such device - did you specify an out-of-range profile CPU?");
3041                 break;
3042         case EOPNOTSUPP:
3043                 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK)
3044                         return scnprintf(msg, size,
3045         "%s: PMU Hardware or event type doesn't support branch stack sampling.",
3046                                          evsel__name(evsel));
3047                 if (evsel->core.attr.aux_output)
3048                         return scnprintf(msg, size,
3049         "%s: PMU Hardware doesn't support 'aux_output' feature",
3050                                          evsel__name(evsel));
3051                 if (evsel->core.attr.sample_period != 0)
3052                         return scnprintf(msg, size,
3053         "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'",
3054                                          evsel__name(evsel));
3055                 if (evsel->core.attr.precise_ip)
3056                         return scnprintf(msg, size, "%s",
3057         "\'precise\' request may not be supported. Try removing 'p' modifier.");
3058 #if defined(__i386__) || defined(__x86_64__)
3059                 if (evsel->core.attr.type == PERF_TYPE_HARDWARE)
3060                         return scnprintf(msg, size, "%s",
3061         "No hardware sampling interrupt available.\n");
3062 #endif
3063                 break;
3064         case EBUSY:
3065                 if (find_process("oprofiled"))
3066                         return scnprintf(msg, size,
3067         "The PMU counters are busy/taken by another profiler.\n"
3068         "We found oprofile daemon running, please stop it and try again.");
3069                 break;
3070         case EINVAL:
3071                 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size)
3072                         return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel.");
3073                 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size)
3074                         return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel.");
3075                 if (evsel->core.attr.write_backward && perf_missing_features.write_backward)
3076                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
3077                 if (perf_missing_features.clockid)
3078                         return scnprintf(msg, size, "clockid feature not supported.");
3079                 if (perf_missing_features.clockid_wrong)
3080                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
3081                 if (perf_missing_features.aux_output)
3082                         return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel.");
3083                 if (!target__has_cpu(target))
3084                         return scnprintf(msg, size,
3085         "Invalid event (%s) in per-thread mode, enable system wide with '-a'.",
3086                                         evsel__name(evsel));
3087                 if (is_amd(arch, cpuid)) {
3088                         if (is_amd_ibs(evsel)) {
3089                                 if (evsel->core.attr.exclude_kernel)
3090                                         return scnprintf(msg, size,
3091         "AMD IBS can't exclude kernel events.  Try running at a higher privilege level.");
3092                                 if (!evsel->core.system_wide)
3093                                         return scnprintf(msg, size,
3094         "AMD IBS may only be available in system-wide/per-cpu mode.  Try using -a, or -C and workload affinity");
3095                         }
3096                 }
3097
3098                 break;
3099         case ENODATA:
3100                 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. "
3101                                  "Please add an auxiliary event in front of the load latency event.");
3102         default:
3103                 break;
3104         }
3105
3106         return scnprintf(msg, size,
3107         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
3108         "/bin/dmesg | grep -i perf may provide additional information.\n",
3109                          err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel));
3110 }
3111
3112 struct perf_env *evsel__env(struct evsel *evsel)
3113 {
3114         if (evsel && evsel->evlist && evsel->evlist->env)
3115                 return evsel->evlist->env;
3116         return &perf_env;
3117 }
3118
3119 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist)
3120 {
3121         int cpu_map_idx, thread;
3122
3123         for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) {
3124                 for (thread = 0; thread < xyarray__max_y(evsel->core.fd);
3125                      thread++) {
3126                         int fd = FD(evsel, cpu_map_idx, thread);
3127
3128                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core,
3129                                                    cpu_map_idx, thread, fd) < 0)
3130                                 return -1;
3131                 }
3132         }
3133
3134         return 0;
3135 }
3136
3137 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist)
3138 {
3139         struct perf_cpu_map *cpus = evsel->core.cpus;
3140         struct perf_thread_map *threads = evsel->core.threads;
3141
3142         if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr))
3143                 return -ENOMEM;
3144
3145         return store_evsel_ids(evsel, evlist);
3146 }
3147
3148 void evsel__zero_per_pkg(struct evsel *evsel)
3149 {
3150         struct hashmap_entry *cur;
3151         size_t bkt;
3152
3153         if (evsel->per_pkg_mask) {
3154                 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt)
3155                         free((void *)cur->pkey);
3156
3157                 hashmap__clear(evsel->per_pkg_mask);
3158         }
3159 }
3160
3161 bool evsel__is_hybrid(const struct evsel *evsel)
3162 {
3163         return evsel->pmu_name && perf_pmu__is_hybrid(evsel->pmu_name);
3164 }
3165
3166 struct evsel *evsel__leader(const struct evsel *evsel)
3167 {
3168         return container_of(evsel->core.leader, struct evsel, core);
3169 }
3170
3171 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader)
3172 {
3173         return evsel->core.leader == &leader->core;
3174 }
3175
3176 bool evsel__is_leader(struct evsel *evsel)
3177 {
3178         return evsel__has_leader(evsel, evsel);
3179 }
3180
3181 void evsel__set_leader(struct evsel *evsel, struct evsel *leader)
3182 {
3183         evsel->core.leader = &leader->core;
3184 }
3185
3186 int evsel__source_count(const struct evsel *evsel)
3187 {
3188         struct evsel *pos;
3189         int count = 0;
3190
3191         evlist__for_each_entry(evsel->evlist, pos) {
3192                 if (pos->metric_leader == evsel)
3193                         count++;
3194         }
3195         return count;
3196 }
3197
3198 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused)
3199 {
3200         return false;
3201 }
3202
3203 /*
3204  * Remove an event from a given group (leader).
3205  * Some events, e.g., perf metrics Topdown events,
3206  * must always be grouped. Ignore the events.
3207  */
3208 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader)
3209 {
3210         if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) {
3211                 evsel__set_leader(evsel, evsel);
3212                 evsel->core.nr_members = 0;
3213                 leader->core.nr_members--;
3214         }
3215 }