tools/perf/util/evlist.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <[email protected]>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8 #include <api/fs/fs.h>
   9 #include <errno.h>
  10 #include <inttypes.h>
  11 #include <poll.h>
  12 #include "cpumap.h"
  13 #include "util/mmap.h"
  14 #include "thread_map.h"
  15 #include "target.h"
  16 #include "evlist.h"
  17 #include "evsel.h"
  18 #include "record.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "bpf_counter.h"
  22 #include <internal/lib.h> // page_size
  23 #include "affinity.h"
  24 #include "../perf.h"
  25 #include "asm/bug.h"
  26 #include "bpf-event.h"
  27 #include "util/string2.h"
  28 #include "util/perf_api_probe.h"
  29 #include "util/evsel_fprintf.h"
  30 #include "util/evlist-hybrid.h"
  31 #include "util/pmu.h"
  32 #include <signal.h>
  33 #include <unistd.h>
  34 #include <sched.h>
  35 #include <stdlib.h>
  36
  37 #include "parse-events.h"
  38 #include <subcmd/parse-options.h>
  39
  40 #include <fcntl.h>
  41 #include <sys/ioctl.h>
  42 #include <sys/mman.h>
  43 #include <sys/prctl.h>
  44 #include <sys/timerfd.h>
  45
  46 #include <linux/bitops.h>
  47 #include <linux/hash.h>
  48 #include <linux/log2.h>
  49 #include <linux/err.h>
  50 #include <linux/string.h>
  51 #include <linux/time64.h>
  52 #include <linux/zalloc.h>
  53 #include <perf/evlist.h>
  54 #include <perf/evsel.h>
  55 #include <perf/cpumap.h>
  56 #include <perf/mmap.h>
  57
  58 #include <internal/xyarray.h>
  59
  60 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  61 int sigqueue(pid_t pid, int sig, const union sigval value);
  62 #endif
  63
  64 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  65 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  66
  67 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  68                   struct perf_thread_map *threads)
  69 {
  70         perf_evlist__init(&evlist->core);
  71         perf_evlist__set_maps(&evlist->core, cpus, threads);
  72         evlist->workload.pid = -1;
  73         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  74         evlist->ctl_fd.fd = -1;
  75         evlist->ctl_fd.ack = -1;
  76         evlist->ctl_fd.pos = -1;
  77 }
  78
  79 struct evlist *evlist__new(void)
  80 {
  81         struct evlist *evlist = zalloc(sizeof(*evlist));
  82
  83         if (evlist != NULL)
  84                 evlist__init(evlist, NULL, NULL);
  85
  86         return evlist;
  87 }
  88
  89 struct evlist *evlist__new_default(void)
  90 {
  91         struct evlist *evlist = evlist__new();
  92
  93         if (evlist && evlist__add_default(evlist)) {
  94                 evlist__delete(evlist);
  95                 evlist = NULL;
  96         }
  97
  98         return evlist;
  99 }
 100
 101 struct evlist *evlist__new_dummy(void)
 102 {
 103         struct evlist *evlist = evlist__new();
 104
 105         if (evlist && evlist__add_dummy(evlist)) {
 106                 evlist__delete(evlist);
 107                 evlist = NULL;
 108         }
 109
 110         return evlist;
 111 }
 112
 113 /**
 114  * evlist__set_id_pos - set the positions of event ids.
 115  * @evlist: selected event list
 116  *
 117  * Events with compatible sample types all have the same id_pos
 118  * and is_pos.  For convenience, put a copy on evlist.
 119  */
 120 void evlist__set_id_pos(struct evlist *evlist)
 121 {
 122         struct evsel *first = evlist__first(evlist);
 123
 124         evlist->id_pos = first->id_pos;
 125         evlist->is_pos = first->is_pos;
 126 }
 127
 128 static void evlist__update_id_pos(struct evlist *evlist)
 129 {
 130         struct evsel *evsel;
 131
 132         evlist__for_each_entry(evlist, evsel)
 133                 evsel__calc_id_pos(evsel);
 134
 135         evlist__set_id_pos(evlist);
 136 }
 137
 138 static void evlist__purge(struct evlist *evlist)
 139 {
 140         struct evsel *pos, *n;
 141
 142         evlist__for_each_entry_safe(evlist, n, pos) {
 143                 list_del_init(&pos->core.node);
 144                 pos->evlist = NULL;
 145                 evsel__delete(pos);
 146         }
 147
 148         evlist->core.nr_entries = 0;
 149 }
 150
 151 void evlist__exit(struct evlist *evlist)
 152 {
 153         event_enable_timer__exit(&evlist->eet);
 154         zfree(&evlist->mmap);
 155         zfree(&evlist->overwrite_mmap);
 156         perf_evlist__exit(&evlist->core);
 157 }
 158
 159 void evlist__delete(struct evlist *evlist)
 160 {
 161         if (evlist == NULL)
 162                 return;
 163
 164         evlist__munmap(evlist);
 165         evlist__close(evlist);
 166         evlist__purge(evlist);
 167         evlist__exit(evlist);
 168         free(evlist);
 169 }
 170
 171 void evlist__add(struct evlist *evlist, struct evsel *entry)
 172 {
 173         perf_evlist__add(&evlist->core, &entry->core);
 174         entry->evlist = evlist;
 175         entry->tracking = !entry->core.idx;
 176
 177         if (evlist->core.nr_entries == 1)
 178                 evlist__set_id_pos(evlist);
 179 }
 180
 181 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 182 {
 183         evsel->evlist = NULL;
 184         perf_evlist__remove(&evlist->core, &evsel->core);
 185 }
 186
 187 void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
 188 {
 189         while (!list_empty(list)) {
 190                 struct evsel *evsel, *temp, *leader = NULL;
 191
 192                 __evlist__for_each_entry_safe(list, temp, evsel) {
 193                         list_del_init(&evsel->core.node);
 194                         evlist__add(evlist, evsel);
 195                         leader = evsel;
 196                         break;
 197                 }
 198
 199                 __evlist__for_each_entry_safe(list, temp, evsel) {
 200                         if (evsel__has_leader(evsel, leader)) {
 201                                 list_del_init(&evsel->core.node);
 202                                 evlist__add(evlist, evsel);
 203                         }
 204                 }
 205         }
 206 }
 207
 208 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 209                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 210 {
 211         size_t i;
 212         int err;
 213
 214         for (i = 0; i < nr_assocs; i++) {
 215                 // Adding a handler for an event not in this evlist, just ignore it.
 216                 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 217                 if (evsel == NULL)
 218                         continue;
 219
 220                 err = -EEXIST;
 221                 if (evsel->handler != NULL)
 222                         goto out;
 223                 evsel->handler = assocs[i].handler;
 224         }
 225
 226         err = 0;
 227 out:
 228         return err;
 229 }
 230
 231 void evlist__set_leader(struct evlist *evlist)
 232 {
 233         perf_evlist__set_leader(&evlist->core);
 234 }
 235
 236 int __evlist__add_default(struct evlist *evlist, bool precise)
 237 {
 238         struct evsel *evsel;
 239
 240         evsel = evsel__new_cycles(precise, PERF_TYPE_HARDWARE,
 241                                   PERF_COUNT_HW_CPU_CYCLES);
 242         if (evsel == NULL)
 243                 return -ENOMEM;
 244
 245         evlist__add(evlist, evsel);
 246         return 0;
 247 }
 248
 249 static struct evsel *evlist__dummy_event(struct evlist *evlist)
 250 {
 251         struct perf_event_attr attr = {
 252                 .type   = PERF_TYPE_SOFTWARE,
 253                 .config = PERF_COUNT_SW_DUMMY,
 254                 .size   = sizeof(attr), /* to capture ABI version */
 255         };
 256
 257         return evsel__new_idx(&attr, evlist->core.nr_entries);
 258 }
 259
 260 int evlist__add_dummy(struct evlist *evlist)
 261 {
 262         struct evsel *evsel = evlist__dummy_event(evlist);
 263
 264         if (evsel == NULL)
 265                 return -ENOMEM;
 266
 267         evlist__add(evlist, evsel);
 268         return 0;
 269 }
 270
 271 struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
 272 {
 273         struct evsel *evsel = evlist__dummy_event(evlist);
 274
 275         if (!evsel)
 276                 return NULL;
 277
 278         evsel->core.attr.exclude_kernel = 1;
 279         evsel->core.attr.exclude_guest = 1;
 280         evsel->core.attr.exclude_hv = 1;
 281         evsel->core.attr.freq = 0;
 282         evsel->core.attr.sample_period = 1;
 283         evsel->core.system_wide = system_wide;
 284         evsel->no_aux_samples = true;
 285         evsel->name = strdup("dummy:u");
 286
 287         evlist__add(evlist, evsel);
 288         return evsel;
 289 }
 290
 291 struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
 292 {
 293         struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0);
 294
 295         if (IS_ERR(evsel))
 296                 return evsel;
 297
 298         evsel__set_sample_bit(evsel, CPU);
 299         evsel__set_sample_bit(evsel, TIME);
 300
 301         evsel->core.system_wide = system_wide;
 302         evsel->no_aux_samples = true;
 303
 304         evlist__add(evlist, evsel);
 305         return evsel;
 306 };
 307
 308 int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 309 {
 310         struct evsel *evsel, *n;
 311         LIST_HEAD(head);
 312         size_t i;
 313
 314         for (i = 0; i < nr_attrs; i++) {
 315                 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 316                 if (evsel == NULL)
 317                         goto out_delete_partial_list;
 318                 list_add_tail(&evsel->core.node, &head);
 319         }
 320
 321         evlist__splice_list_tail(evlist, &head);
 322
 323         return 0;
 324
 325 out_delete_partial_list:
 326         __evlist__for_each_entry_safe(&head, n, evsel)
 327                 evsel__delete(evsel);
 328         return -1;
 329 }
 330
 331 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 332 {
 333         size_t i;
 334
 335         for (i = 0; i < nr_attrs; i++)
 336                 event_attr_init(attrs + i);
 337
 338         return evlist__add_attrs(evlist, attrs, nr_attrs);
 339 }
 340
 341 __weak int arch_evlist__add_default_attrs(struct evlist *evlist,
 342                                           struct perf_event_attr *attrs,
 343                                           size_t nr_attrs)
 344 {
 345         if (!nr_attrs)
 346                 return 0;
 347
 348         return __evlist__add_default_attrs(evlist, attrs, nr_attrs);
 349 }
 350
 351 struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 352 {
 353         struct evsel *evsel;
 354
 355         evlist__for_each_entry(evlist, evsel) {
 356                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 357                     (int)evsel->core.attr.config == id)
 358                         return evsel;
 359         }
 360
 361         return NULL;
 362 }
 363
 364 struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
 365 {
 366         struct evsel *evsel;
 367
 368         evlist__for_each_entry(evlist, evsel) {
 369                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 370                     (strcmp(evsel->name, name) == 0))
 371                         return evsel;
 372         }
 373
 374         return NULL;
 375 }
 376
 377 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
 378 {
 379         struct evsel *evsel = evsel__newtp(sys, name);
 380
 381         if (IS_ERR(evsel))
 382                 return -1;
 383
 384         evsel->handler = handler;
 385         evlist__add(evlist, evsel);
 386         return 0;
 387 }
 388
 389 struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
 390 {
 391         struct evlist_cpu_iterator itr = {
 392                 .container = evlist,
 393                 .evsel = NULL,
 394                 .cpu_map_idx = 0,
 395                 .evlist_cpu_map_idx = 0,
 396                 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
 397                 .cpu = (struct perf_cpu){ .cpu = -1},
 398                 .affinity = affinity,
 399         };
 400
 401         if (evlist__empty(evlist)) {
 402                 /* Ensure the empty list doesn't iterate. */
 403                 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
 404         } else {
 405                 itr.evsel = evlist__first(evlist);
 406                 if (itr.affinity) {
 407                         itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
 408                         affinity__set(itr.affinity, itr.cpu.cpu);
 409                         itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
 410                         /*
 411                          * If this CPU isn't in the evsel's cpu map then advance
 412                          * through the list.
 413                          */
 414                         if (itr.cpu_map_idx == -1)
 415                                 evlist_cpu_iterator__next(&itr);
 416                 }
 417         }
 418         return itr;
 419 }
 420
 421 void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
 422 {
 423         while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
 424                 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
 425                 evlist_cpu_itr->cpu_map_idx =
 426                         perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 427                                           evlist_cpu_itr->cpu);
 428                 if (evlist_cpu_itr->cpu_map_idx != -1)
 429                         return;
 430         }
 431         evlist_cpu_itr->evlist_cpu_map_idx++;
 432         if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
 433                 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
 434                 evlist_cpu_itr->cpu =
 435                         perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
 436                                           evlist_cpu_itr->evlist_cpu_map_idx);
 437                 if (evlist_cpu_itr->affinity)
 438                         affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
 439                 evlist_cpu_itr->cpu_map_idx =
 440                         perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 441                                           evlist_cpu_itr->cpu);
 442                 /*
 443                  * If this CPU isn't in the evsel's cpu map then advance through
 444                  * the list.
 445                  */
 446                 if (evlist_cpu_itr->cpu_map_idx == -1)
 447                         evlist_cpu_iterator__next(evlist_cpu_itr);
 448         }
 449 }
 450
 451 bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
 452 {
 453         return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
 454 }
 455
 456 static int evsel__strcmp(struct evsel *pos, char *evsel_name)
 457 {
 458         if (!evsel_name)
 459                 return 0;
 460         if (evsel__is_dummy_event(pos))
 461                 return 1;
 462         return strcmp(pos->name, evsel_name);
 463 }
 464
 465 static int evlist__is_enabled(struct evlist *evlist)
 466 {
 467         struct evsel *pos;
 468
 469         evlist__for_each_entry(evlist, pos) {
 470                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 471                         continue;
 472                 /* If at least one event is enabled, evlist is enabled. */
 473                 if (!pos->disabled)
 474                         return true;
 475         }
 476         return false;
 477 }
 478
 479 static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 480 {
 481         struct evsel *pos;
 482         struct evlist_cpu_iterator evlist_cpu_itr;
 483         struct affinity saved_affinity, *affinity = NULL;
 484         bool has_imm = false;
 485
 486         // See explanation in evlist__close()
 487         if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 488                 if (affinity__setup(&saved_affinity) < 0)
 489                         return;
 490                 affinity = &saved_affinity;
 491         }
 492
 493         /* Disable 'immediate' events last */
 494         for (int imm = 0; imm <= 1; imm++) {
 495                 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 496                         pos = evlist_cpu_itr.evsel;
 497                         if (evsel__strcmp(pos, evsel_name))
 498                                 continue;
 499                         if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
 500                                 continue;
 501                         if (excl_dummy && evsel__is_dummy_event(pos))
 502                                 continue;
 503                         if (pos->immediate)
 504                                 has_imm = true;
 505                         if (pos->immediate != imm)
 506                                 continue;
 507                         evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 508                 }
 509                 if (!has_imm)
 510                         break;
 511         }
 512
 513         affinity__cleanup(affinity);
 514         evlist__for_each_entry(evlist, pos) {
 515                 if (evsel__strcmp(pos, evsel_name))
 516                         continue;
 517                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 518                         continue;
 519                 if (excl_dummy && evsel__is_dummy_event(pos))
 520                         continue;
 521                 pos->disabled = true;
 522         }
 523
 524         /*
 525          * If we disabled only single event, we need to check
 526          * the enabled state of the evlist manually.
 527          */
 528         if (evsel_name)
 529                 evlist->enabled = evlist__is_enabled(evlist);
 530         else
 531                 evlist->enabled = false;
 532 }
 533
 534 void evlist__disable(struct evlist *evlist)
 535 {
 536         __evlist__disable(evlist, NULL, false);
 537 }
 538
 539 void evlist__disable_non_dummy(struct evlist *evlist)
 540 {
 541         __evlist__disable(evlist, NULL, true);
 542 }
 543
 544 void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
 545 {
 546         __evlist__disable(evlist, evsel_name, false);
 547 }
 548
 549 static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 550 {
 551         struct evsel *pos;
 552         struct evlist_cpu_iterator evlist_cpu_itr;
 553         struct affinity saved_affinity, *affinity = NULL;
 554
 555         // See explanation in evlist__close()
 556         if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 557                 if (affinity__setup(&saved_affinity) < 0)
 558                         return;
 559                 affinity = &saved_affinity;
 560         }
 561
 562         evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 563                 pos = evlist_cpu_itr.evsel;
 564                 if (evsel__strcmp(pos, evsel_name))
 565                         continue;
 566                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 567                         continue;
 568                 if (excl_dummy && evsel__is_dummy_event(pos))
 569                         continue;
 570                 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 571         }
 572         affinity__cleanup(affinity);
 573         evlist__for_each_entry(evlist, pos) {
 574                 if (evsel__strcmp(pos, evsel_name))
 575                         continue;
 576                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 577                         continue;
 578                 if (excl_dummy && evsel__is_dummy_event(pos))
 579                         continue;
 580                 pos->disabled = false;
 581         }
 582
 583         /*
 584          * Even single event sets the 'enabled' for evlist,
 585          * so the toggle can work properly and toggle to
 586          * 'disabled' state.
 587          */
 588         evlist->enabled = true;
 589 }
 590
 591 void evlist__enable(struct evlist *evlist)
 592 {
 593         __evlist__enable(evlist, NULL, false);
 594 }
 595
 596 void evlist__enable_non_dummy(struct evlist *evlist)
 597 {
 598         __evlist__enable(evlist, NULL, true);
 599 }
 600
 601 void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
 602 {
 603         __evlist__enable(evlist, evsel_name, false);
 604 }
 605
 606 void evlist__toggle_enable(struct evlist *evlist)
 607 {
 608         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 609 }
 610
 611 int evlist__add_pollfd(struct evlist *evlist, int fd)
 612 {
 613         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
 614 }
 615
 616 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 617 {
 618         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 619 }
 620
 621 #ifdef HAVE_EVENTFD_SUPPORT
 622 int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
 623 {
 624         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
 625                                        fdarray_flag__nonfilterable |
 626                                        fdarray_flag__non_perf_event);
 627 }
 628 #endif
 629
 630 int evlist__poll(struct evlist *evlist, int timeout)
 631 {
 632         return perf_evlist__poll(&evlist->core, timeout);
 633 }
 634
 635 struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
 636 {
 637         struct hlist_head *head;
 638         struct perf_sample_id *sid;
 639         int hash;
 640
 641         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 642         head = &evlist->core.heads[hash];
 643
 644         hlist_for_each_entry(sid, head, node)
 645                 if (sid->id == id)
 646                         return sid;
 647
 648         return NULL;
 649 }
 650
 651 struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
 652 {
 653         struct perf_sample_id *sid;
 654
 655         if (evlist->core.nr_entries == 1 || !id)
 656                 return evlist__first(evlist);
 657
 658         sid = evlist__id2sid(evlist, id);
 659         if (sid)
 660                 return container_of(sid->evsel, struct evsel, core);
 661
 662         if (!evlist__sample_id_all(evlist))
 663                 return evlist__first(evlist);
 664
 665         return NULL;
 666 }
 667
 668 struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
 669 {
 670         struct perf_sample_id *sid;
 671
 672         if (!id)
 673                 return NULL;
 674
 675         sid = evlist__id2sid(evlist, id);
 676         if (sid)
 677                 return container_of(sid->evsel, struct evsel, core);
 678
 679         return NULL;
 680 }
 681
 682 static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
 683 {
 684         const __u64 *array = event->sample.array;
 685         ssize_t n;
 686
 687         n = (event->header.size - sizeof(event->header)) >> 3;
 688
 689         if (event->header.type == PERF_RECORD_SAMPLE) {
 690                 if (evlist->id_pos >= n)
 691                         return -1;
 692                 *id = array[evlist->id_pos];
 693         } else {
 694                 if (evlist->is_pos > n)
 695                         return -1;
 696                 n -= evlist->is_pos;
 697                 *id = array[n];
 698         }
 699         return 0;
 700 }
 701
 702 struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
 703 {
 704         struct evsel *first = evlist__first(evlist);
 705         struct hlist_head *head;
 706         struct perf_sample_id *sid;
 707         int hash;
 708         u64 id;
 709
 710         if (evlist->core.nr_entries == 1)
 711                 return first;
 712
 713         if (!first->core.attr.sample_id_all &&
 714             event->header.type != PERF_RECORD_SAMPLE)
 715                 return first;
 716
 717         if (evlist__event2id(evlist, event, &id))
 718                 return NULL;
 719
 720         /* Synthesized events have an id of zero */
 721         if (!id)
 722                 return first;
 723
 724         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 725         head = &evlist->core.heads[hash];
 726
 727         hlist_for_each_entry(sid, head, node) {
 728                 if (sid->id == id)
 729                         return container_of(sid->evsel, struct evsel, core);
 730         }
 731         return NULL;
 732 }
 733
 734 static int evlist__set_paused(struct evlist *evlist, bool value)
 735 {
 736         int i;
 737
 738         if (!evlist->overwrite_mmap)
 739                 return 0;
 740
 741         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 742                 int fd = evlist->overwrite_mmap[i].core.fd;
 743                 int err;
 744
 745                 if (fd < 0)
 746                         continue;
 747                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 748                 if (err)
 749                         return err;
 750         }
 751         return 0;
 752 }
 753
 754 static int evlist__pause(struct evlist *evlist)
 755 {
 756         return evlist__set_paused(evlist, true);
 757 }
 758
 759 static int evlist__resume(struct evlist *evlist)
 760 {
 761         return evlist__set_paused(evlist, false);
 762 }
 763
 764 static void evlist__munmap_nofree(struct evlist *evlist)
 765 {
 766         int i;
 767
 768         if (evlist->mmap)
 769                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 770                         perf_mmap__munmap(&evlist->mmap[i].core);
 771
 772         if (evlist->overwrite_mmap)
 773                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 774                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 775 }
 776
 777 void evlist__munmap(struct evlist *evlist)
 778 {
 779         evlist__munmap_nofree(evlist);
 780         zfree(&evlist->mmap);
 781         zfree(&evlist->overwrite_mmap);
 782 }
 783
 784 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 785 {
 786         struct mmap *m = container_of(map, struct mmap, core);
 787
 788         mmap__munmap(m);
 789 }
 790
 791 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 792                                        bool overwrite)
 793 {
 794         int i;
 795         struct mmap *map;
 796
 797         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 798         if (!map)
 799                 return NULL;
 800
 801         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 802                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 803
 804                 /*
 805                  * When the perf_mmap() call is made we grab one refcount, plus
 806                  * one extra to let perf_mmap__consume() get the last
 807                  * events after all real references (perf_mmap__get()) are
 808                  * dropped.
 809                  *
 810                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 811                  * thus does perf_mmap__get() on it.
 812                  */
 813                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 814         }
 815
 816         return map;
 817 }
 818
 819 static void
 820 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 821                          struct perf_evsel *_evsel,
 822                          struct perf_mmap_param *_mp,
 823                          int idx)
 824 {
 825         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 826         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 827         struct evsel *evsel = container_of(_evsel, struct evsel, core);
 828
 829         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
 830 }
 831
 832 static struct perf_mmap*
 833 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 834 {
 835         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 836         struct mmap *maps;
 837
 838         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 839
 840         if (!maps) {
 841                 maps = evlist__alloc_mmap(evlist, overwrite);
 842                 if (!maps)
 843                         return NULL;
 844
 845                 if (overwrite) {
 846                         evlist->overwrite_mmap = maps;
 847                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 848                                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 849                 } else {
 850                         evlist->mmap = maps;
 851                 }
 852         }
 853
 854         return &maps[idx].core;
 855 }
 856
 857 static int
 858 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 859                           int output, struct perf_cpu cpu)
 860 {
 861         struct mmap *map = container_of(_map, struct mmap, core);
 862         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 863
 864         return mmap__mmap(map, mp, output, cpu);
 865 }
 866
 867 unsigned long perf_event_mlock_kb_in_pages(void)
 868 {
 869         unsigned long pages;
 870         int max;
 871
 872         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 873                 /*
 874                  * Pick a once upon a time good value, i.e. things look
 875                  * strange since we can't read a sysctl value, but lets not
 876                  * die yet...
 877                  */
 878                 max = 512;
 879         } else {
 880                 max -= (page_size / 1024);
 881         }
 882
 883         pages = (max * 1024) / page_size;
 884         if (!is_power_of_2(pages))
 885                 pages = rounddown_pow_of_two(pages);
 886
 887         return pages;
 888 }
 889
 890 size_t evlist__mmap_size(unsigned long pages)
 891 {
 892         if (pages == UINT_MAX)
 893                 pages = perf_event_mlock_kb_in_pages();
 894         else if (!is_power_of_2(pages))
 895                 return 0;
 896
 897         return (pages + 1) * page_size;
 898 }
 899
 900 static long parse_pages_arg(const char *str, unsigned long min,
 901                             unsigned long max)
 902 {
 903         unsigned long pages, val;
 904         static struct parse_tag tags[] = {
 905                 { .tag  = 'B', .mult = 1       },
 906                 { .tag  = 'K', .mult = 1 << 10 },
 907                 { .tag  = 'M', .mult = 1 << 20 },
 908                 { .tag  = 'G', .mult = 1 << 30 },
 909                 { .tag  = 0 },
 910         };
 911
 912         if (str == NULL)
 913                 return -EINVAL;
 914
 915         val = parse_tag_value(str, tags);
 916         if (val != (unsigned long) -1) {
 917                 /* we got file size value */
 918                 pages = PERF_ALIGN(val, page_size) / page_size;
 919         } else {
 920                 /* we got pages count value */
 921                 char *eptr;
 922                 pages = strtoul(str, &eptr, 10);
 923                 if (*eptr != '\0')
 924                         return -EINVAL;
 925         }
 926
 927         if (pages == 0 && min == 0) {
 928                 /* leave number of pages at 0 */
 929         } else if (!is_power_of_2(pages)) {
 930                 char buf[100];
 931
 932                 /* round pages up to next power of 2 */
 933                 pages = roundup_pow_of_two(pages);
 934                 if (!pages)
 935                         return -EINVAL;
 936
 937                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 938                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 939                         buf, pages);
 940         }
 941
 942         if (pages > max)
 943                 return -EINVAL;
 944
 945         return pages;
 946 }
 947
 948 int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 949 {
 950         unsigned long max = UINT_MAX;
 951         long pages;
 952
 953         if (max > SIZE_MAX / page_size)
 954                 max = SIZE_MAX / page_size;
 955
 956         pages = parse_pages_arg(str, 1, max);
 957         if (pages < 0) {
 958                 pr_err("Invalid argument for --mmap_pages/-m\n");
 959                 return -1;
 960         }
 961
 962         *mmap_pages = pages;
 963         return 0;
 964 }
 965
 966 int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
 967 {
 968         return __evlist__parse_mmap_pages(opt->value, str);
 969 }
 970
 971 /**
 972  * evlist__mmap_ex - Create mmaps to receive events.
 973  * @evlist: list of events
 974  * @pages: map length in pages
 975  * @overwrite: overwrite older events?
 976  * @auxtrace_pages - auxtrace map length in pages
 977  * @auxtrace_overwrite - overwrite older auxtrace data?
 978  *
 979  * If @overwrite is %false the user needs to signal event consumption using
 980  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 981  * automatically.
 982  *
 983  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 984  * consumption using auxtrace_mmap__write_tail().
 985  *
 986  * Return: %0 on success, negative error code otherwise.
 987  */
 988 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 989                          unsigned int auxtrace_pages,
 990                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 991                          int comp_level)
 992 {
 993         /*
 994          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 995          * Its value is decided by evsel's write_backward.
 996          * So &mp should not be passed through const pointer.
 997          */
 998         struct mmap_params mp = {
 999                 .nr_cblocks     = nr_cblocks,
1000                 .affinity       = affinity,
1001                 .flush          = flush,
1002                 .comp_level     = comp_level
1003         };
1004         struct perf_evlist_mmap_ops ops = {
1005                 .idx  = perf_evlist__mmap_cb_idx,
1006                 .get  = perf_evlist__mmap_cb_get,
1007                 .mmap = perf_evlist__mmap_cb_mmap,
1008         };
1009
1010         evlist->core.mmap_len = evlist__mmap_size(pages);
1011         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
1012
1013         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
1014                                    auxtrace_pages, auxtrace_overwrite);
1015
1016         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
1017 }
1018
1019 int evlist__mmap(struct evlist *evlist, unsigned int pages)
1020 {
1021         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1022 }
1023
1024 int evlist__create_maps(struct evlist *evlist, struct target *target)
1025 {
1026         bool all_threads = (target->per_thread && target->system_wide);
1027         struct perf_cpu_map *cpus;
1028         struct perf_thread_map *threads;
1029
1030         /*
1031          * If specify '-a' and '--per-thread' to perf record, perf record
1032          * will override '--per-thread'. target->per_thread = false and
1033          * target->system_wide = true.
1034          *
1035          * If specify '--per-thread' only to perf record,
1036          * target->per_thread = true and target->system_wide = false.
1037          *
1038          * So target->per_thread && target->system_wide is false.
1039          * For perf record, thread_map__new_str doesn't call
1040          * thread_map__new_all_cpus. That will keep perf record's
1041          * current behavior.
1042          *
1043          * For perf stat, it allows the case that target->per_thread and
1044          * target->system_wide are all true. It means to collect system-wide
1045          * per-thread data. thread_map__new_str will call
1046          * thread_map__new_all_cpus to enumerate all threads.
1047          */
1048         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1049                                       all_threads);
1050
1051         if (!threads)
1052                 return -1;
1053
1054         if (target__uses_dummy_map(target))
1055                 cpus = perf_cpu_map__dummy_new();
1056         else
1057                 cpus = perf_cpu_map__new(target->cpu_list);
1058
1059         if (!cpus)
1060                 goto out_delete_threads;
1061
1062         evlist->core.has_user_cpus = !!target->cpu_list && !target->hybrid;
1063
1064         perf_evlist__set_maps(&evlist->core, cpus, threads);
1065
1066         /* as evlist now has references, put count here */
1067         perf_cpu_map__put(cpus);
1068         perf_thread_map__put(threads);
1069
1070         return 0;
1071
1072 out_delete_threads:
1073         perf_thread_map__put(threads);
1074         return -1;
1075 }
1076
1077 int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1078 {
1079         struct evsel *evsel;
1080         int err = 0;
1081
1082         evlist__for_each_entry(evlist, evsel) {
1083                 if (evsel->filter == NULL)
1084                         continue;
1085
1086                 /*
1087                  * filters only work for tracepoint event, which doesn't have cpu limit.
1088                  * So evlist and evsel should always be same.
1089                  */
1090                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1091                 if (err) {
1092                         *err_evsel = evsel;
1093                         break;
1094                 }
1095         }
1096
1097         return err;
1098 }
1099
1100 int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1101 {
1102         struct evsel *evsel;
1103         int err = 0;
1104
1105         if (filter == NULL)
1106                 return -1;
1107
1108         evlist__for_each_entry(evlist, evsel) {
1109                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1110                         continue;
1111
1112                 err = evsel__set_filter(evsel, filter);
1113                 if (err)
1114                         break;
1115         }
1116
1117         return err;
1118 }
1119
1120 int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1121 {
1122         struct evsel *evsel;
1123         int err = 0;
1124
1125         if (filter == NULL)
1126                 return -1;
1127
1128         evlist__for_each_entry(evlist, evsel) {
1129                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1130                         continue;
1131
1132                 err = evsel__append_tp_filter(evsel, filter);
1133                 if (err)
1134                         break;
1135         }
1136
1137         return err;
1138 }
1139
1140 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1141 {
1142         char *filter;
1143         size_t i;
1144
1145         for (i = 0; i < npids; ++i) {
1146                 if (i == 0) {
1147                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1148                                 return NULL;
1149                 } else {
1150                         char *tmp;
1151
1152                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1153                                 goto out_free;
1154
1155                         free(filter);
1156                         filter = tmp;
1157                 }
1158         }
1159
1160         return filter;
1161 out_free:
1162         free(filter);
1163         return NULL;
1164 }
1165
1166 int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1167 {
1168         char *filter = asprintf__tp_filter_pids(npids, pids);
1169         int ret = evlist__set_tp_filter(evlist, filter);
1170
1171         free(filter);
1172         return ret;
1173 }
1174
1175 int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1176 {
1177         return evlist__set_tp_filter_pids(evlist, 1, &pid);
1178 }
1179
1180 int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1181 {
1182         char *filter = asprintf__tp_filter_pids(npids, pids);
1183         int ret = evlist__append_tp_filter(evlist, filter);
1184
1185         free(filter);
1186         return ret;
1187 }
1188
1189 int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1190 {
1191         return evlist__append_tp_filter_pids(evlist, 1, &pid);
1192 }
1193
1194 bool evlist__valid_sample_type(struct evlist *evlist)
1195 {
1196         struct evsel *pos;
1197
1198         if (evlist->core.nr_entries == 1)
1199                 return true;
1200
1201         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1202                 return false;
1203
1204         evlist__for_each_entry(evlist, pos) {
1205                 if (pos->id_pos != evlist->id_pos ||
1206                     pos->is_pos != evlist->is_pos)
1207                         return false;
1208         }
1209
1210         return true;
1211 }
1212
1213 u64 __evlist__combined_sample_type(struct evlist *evlist)
1214 {
1215         struct evsel *evsel;
1216
1217         if (evlist->combined_sample_type)
1218                 return evlist->combined_sample_type;
1219
1220         evlist__for_each_entry(evlist, evsel)
1221                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1222
1223         return evlist->combined_sample_type;
1224 }
1225
1226 u64 evlist__combined_sample_type(struct evlist *evlist)
1227 {
1228         evlist->combined_sample_type = 0;
1229         return __evlist__combined_sample_type(evlist);
1230 }
1231
1232 u64 evlist__combined_branch_type(struct evlist *evlist)
1233 {
1234         struct evsel *evsel;
1235         u64 branch_type = 0;
1236
1237         evlist__for_each_entry(evlist, evsel)
1238                 branch_type |= evsel->core.attr.branch_sample_type;
1239         return branch_type;
1240 }
1241
1242 bool evlist__valid_read_format(struct evlist *evlist)
1243 {
1244         struct evsel *first = evlist__first(evlist), *pos = first;
1245         u64 read_format = first->core.attr.read_format;
1246         u64 sample_type = first->core.attr.sample_type;
1247
1248         evlist__for_each_entry(evlist, pos) {
1249                 if (read_format != pos->core.attr.read_format) {
1250                         pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1251                                  read_format, (u64)pos->core.attr.read_format);
1252                 }
1253         }
1254
1255         /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1256         if ((sample_type & PERF_SAMPLE_READ) &&
1257             !(read_format & PERF_FORMAT_ID)) {
1258                 return false;
1259         }
1260
1261         return true;
1262 }
1263
1264 u16 evlist__id_hdr_size(struct evlist *evlist)
1265 {
1266         struct evsel *first = evlist__first(evlist);
1267
1268         return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1269 }
1270
1271 bool evlist__valid_sample_id_all(struct evlist *evlist)
1272 {
1273         struct evsel *first = evlist__first(evlist), *pos = first;
1274
1275         evlist__for_each_entry_continue(evlist, pos) {
1276                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1277                         return false;
1278         }
1279
1280         return true;
1281 }
1282
1283 bool evlist__sample_id_all(struct evlist *evlist)
1284 {
1285         struct evsel *first = evlist__first(evlist);
1286         return first->core.attr.sample_id_all;
1287 }
1288
1289 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1290 {
1291         evlist->selected = evsel;
1292 }
1293
1294 void evlist__close(struct evlist *evlist)
1295 {
1296         struct evsel *evsel;
1297         struct evlist_cpu_iterator evlist_cpu_itr;
1298         struct affinity affinity;
1299
1300         /*
1301          * With perf record core.user_requested_cpus is usually NULL.
1302          * Use the old method to handle this for now.
1303          */
1304         if (!evlist->core.user_requested_cpus ||
1305             cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1306                 evlist__for_each_entry_reverse(evlist, evsel)
1307                         evsel__close(evsel);
1308                 return;
1309         }
1310
1311         if (affinity__setup(&affinity) < 0)
1312                 return;
1313
1314         evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1315                 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1316                                       evlist_cpu_itr.cpu_map_idx);
1317         }
1318
1319         affinity__cleanup(&affinity);
1320         evlist__for_each_entry_reverse(evlist, evsel) {
1321                 perf_evsel__free_fd(&evsel->core);
1322                 perf_evsel__free_id(&evsel->core);
1323         }
1324         perf_evlist__reset_id_hash(&evlist->core);
1325 }
1326
1327 static int evlist__create_syswide_maps(struct evlist *evlist)
1328 {
1329         struct perf_cpu_map *cpus;
1330         struct perf_thread_map *threads;
1331
1332         /*
1333          * Try reading /sys/devices/system/cpu/online to get
1334          * an all cpus map.
1335          *
1336          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1337          * code needs an overhaul to properly forward the
1338          * error, and we may not want to do that fallback to a
1339          * default cpu identity map :-\
1340          */
1341         cpus = perf_cpu_map__new(NULL);
1342         if (!cpus)
1343                 goto out;
1344
1345         threads = perf_thread_map__new_dummy();
1346         if (!threads)
1347                 goto out_put;
1348
1349         perf_evlist__set_maps(&evlist->core, cpus, threads);
1350
1351         perf_thread_map__put(threads);
1352 out_put:
1353         perf_cpu_map__put(cpus);
1354 out:
1355         return -ENOMEM;
1356 }
1357
1358 int evlist__open(struct evlist *evlist)
1359 {
1360         struct evsel *evsel;
1361         int err;
1362
1363         /*
1364          * Default: one fd per CPU, all threads, aka systemwide
1365          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1366          */
1367         if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1368                 err = evlist__create_syswide_maps(evlist);
1369                 if (err < 0)
1370                         goto out_err;
1371         }
1372
1373         evlist__update_id_pos(evlist);
1374
1375         evlist__for_each_entry(evlist, evsel) {
1376                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1377                 if (err < 0)
1378                         goto out_err;
1379         }
1380
1381         return 0;
1382 out_err:
1383         evlist__close(evlist);
1384         errno = -err;
1385         return err;
1386 }
1387
1388 int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1389                              bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1390 {
1391         int child_ready_pipe[2], go_pipe[2];
1392         char bf;
1393
1394         if (pipe(child_ready_pipe) < 0) {
1395                 perror("failed to create 'ready' pipe");
1396                 return -1;
1397         }
1398
1399         if (pipe(go_pipe) < 0) {
1400                 perror("failed to create 'go' pipe");
1401                 goto out_close_ready_pipe;
1402         }
1403
1404         evlist->workload.pid = fork();
1405         if (evlist->workload.pid < 0) {
1406                 perror("failed to fork");
1407                 goto out_close_pipes;
1408         }
1409
1410         if (!evlist->workload.pid) {
1411                 int ret;
1412
1413                 if (pipe_output)
1414                         dup2(2, 1);
1415
1416                 signal(SIGTERM, SIG_DFL);
1417
1418                 close(child_ready_pipe[0]);
1419                 close(go_pipe[1]);
1420                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1421
1422                 /*
1423                  * Change the name of this process not to confuse --exclude-perf users
1424                  * that sees 'perf' in the window up to the execvp() and thinks that
1425                  * perf samples are not being excluded.
1426                  */
1427                 prctl(PR_SET_NAME, "perf-exec");
1428
1429                 /*
1430                  * Tell the parent we're ready to go
1431                  */
1432                 close(child_ready_pipe[1]);
1433
1434                 /*
1435                  * Wait until the parent tells us to go.
1436                  */
1437                 ret = read(go_pipe[0], &bf, 1);
1438                 /*
1439                  * The parent will ask for the execvp() to be performed by
1440                  * writing exactly one byte, in workload.cork_fd, usually via
1441                  * evlist__start_workload().
1442                  *
1443                  * For cancelling the workload without actually running it,
1444                  * the parent will just close workload.cork_fd, without writing
1445                  * anything, i.e. read will return zero and we just exit()
1446                  * here.
1447                  */
1448                 if (ret != 1) {
1449                         if (ret == -1)
1450                                 perror("unable to read pipe");
1451                         exit(ret);
1452                 }
1453
1454                 execvp(argv[0], (char **)argv);
1455
1456                 if (exec_error) {
1457                         union sigval val;
1458
1459                         val.sival_int = errno;
1460                         if (sigqueue(getppid(), SIGUSR1, val))
1461                                 perror(argv[0]);
1462                 } else
1463                         perror(argv[0]);
1464                 exit(-1);
1465         }
1466
1467         if (exec_error) {
1468                 struct sigaction act = {
1469                         .sa_flags     = SA_SIGINFO,
1470                         .sa_sigaction = exec_error,
1471                 };
1472                 sigaction(SIGUSR1, &act, NULL);
1473         }
1474
1475         if (target__none(target)) {
1476                 if (evlist->core.threads == NULL) {
1477                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1478                                 __func__, __LINE__);
1479                         goto out_close_pipes;
1480                 }
1481                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1482         }
1483
1484         close(child_ready_pipe[1]);
1485         close(go_pipe[0]);
1486         /*
1487          * wait for child to settle
1488          */
1489         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1490                 perror("unable to read pipe");
1491                 goto out_close_pipes;
1492         }
1493
1494         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1495         evlist->workload.cork_fd = go_pipe[1];
1496         close(child_ready_pipe[0]);
1497         return 0;
1498
1499 out_close_pipes:
1500         close(go_pipe[0]);
1501         close(go_pipe[1]);
1502 out_close_ready_pipe:
1503         close(child_ready_pipe[0]);
1504         close(child_ready_pipe[1]);
1505         return -1;
1506 }
1507
1508 int evlist__start_workload(struct evlist *evlist)
1509 {
1510         if (evlist->workload.cork_fd > 0) {
1511                 char bf = 0;
1512                 int ret;
1513                 /*
1514                  * Remove the cork, let it rip!
1515                  */
1516                 ret = write(evlist->workload.cork_fd, &bf, 1);
1517                 if (ret < 0)
1518                         perror("unable to write to pipe");
1519
1520                 close(evlist->workload.cork_fd);
1521                 return ret;
1522         }
1523
1524         return 0;
1525 }
1526
1527 int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1528 {
1529         struct evsel *evsel = evlist__event2evsel(evlist, event);
1530         int ret;
1531
1532         if (!evsel)
1533                 return -EFAULT;
1534         ret = evsel__parse_sample(evsel, event, sample);
1535         if (ret)
1536                 return ret;
1537         if (perf_guest && sample->id) {
1538                 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1539
1540                 if (sid) {
1541                         sample->machine_pid = sid->machine_pid;
1542                         sample->vcpu = sid->vcpu.cpu;
1543                 }
1544         }
1545         return 0;
1546 }
1547
1548 int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1549 {
1550         struct evsel *evsel = evlist__event2evsel(evlist, event);
1551
1552         if (!evsel)
1553                 return -EFAULT;
1554         return evsel__parse_sample_timestamp(evsel, event, timestamp);
1555 }
1556
1557 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1558 {
1559         int printed, value;
1560         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1561
1562         switch (err) {
1563         case EACCES:
1564         case EPERM:
1565                 printed = scnprintf(buf, size,
1566                                     "Error:\t%s.\n"
1567                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1568
1569                 value = perf_event_paranoid();
1570
1571                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1572
1573                 if (value >= 2) {
1574                         printed += scnprintf(buf + printed, size - printed,
1575                                              "For your workloads it needs to be <= 1\nHint:\t");
1576                 }
1577                 printed += scnprintf(buf + printed, size - printed,
1578                                      "For system wide tracing it needs to be set to -1.\n");
1579
1580                 printed += scnprintf(buf + printed, size - printed,
1581                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1582                                     "Hint:\tThe current value is %d.", value);
1583                 break;
1584         case EINVAL: {
1585                 struct evsel *first = evlist__first(evlist);
1586                 int max_freq;
1587
1588                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1589                         goto out_default;
1590
1591                 if (first->core.attr.sample_freq < (u64)max_freq)
1592                         goto out_default;
1593
1594                 printed = scnprintf(buf, size,
1595                                     "Error:\t%s.\n"
1596                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1597                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1598                                     emsg, max_freq, first->core.attr.sample_freq);
1599                 break;
1600         }
1601         default:
1602 out_default:
1603                 scnprintf(buf, size, "%s", emsg);
1604                 break;
1605         }
1606
1607         return 0;
1608 }
1609
1610 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1611 {
1612         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1613         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1614
1615         switch (err) {
1616         case EPERM:
1617                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1618                 printed += scnprintf(buf + printed, size - printed,
1619                                      "Error:\t%s.\n"
1620                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1621                                      "Hint:\tTried using %zd kB.\n",
1622                                      emsg, pages_max_per_user, pages_attempted);
1623
1624                 if (pages_attempted >= pages_max_per_user) {
1625                         printed += scnprintf(buf + printed, size - printed,
1626                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1627                                              pages_max_per_user + pages_attempted);
1628                 }
1629
1630                 printed += scnprintf(buf + printed, size - printed,
1631                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1632                 break;
1633         default:
1634                 scnprintf(buf, size, "%s", emsg);
1635                 break;
1636         }
1637
1638         return 0;
1639 }
1640
1641 void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1642 {
1643         struct evsel *evsel, *n;
1644         LIST_HEAD(move);
1645
1646         if (move_evsel == evlist__first(evlist))
1647                 return;
1648
1649         evlist__for_each_entry_safe(evlist, n, evsel) {
1650                 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1651                         list_move_tail(&evsel->core.node, &move);
1652         }
1653
1654         list_splice(&move, &evlist->core.entries);
1655 }
1656
1657 struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1658 {
1659         struct evsel *evsel;
1660
1661         evlist__for_each_entry(evlist, evsel) {
1662                 if (evsel->tracking)
1663                         return evsel;
1664         }
1665
1666         return evlist__first(evlist);
1667 }
1668
1669 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1670 {
1671         struct evsel *evsel;
1672
1673         if (tracking_evsel->tracking)
1674                 return;
1675
1676         evlist__for_each_entry(evlist, evsel) {
1677                 if (evsel != tracking_evsel)
1678                         evsel->tracking = false;
1679         }
1680
1681         tracking_evsel->tracking = true;
1682 }
1683
1684 struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1685 {
1686         struct evsel *evsel;
1687
1688         evlist__for_each_entry(evlist, evsel) {
1689                 if (!evsel->name)
1690                         continue;
1691                 if (strcmp(str, evsel->name) == 0)
1692                         return evsel;
1693         }
1694
1695         return NULL;
1696 }
1697
1698 void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1699 {
1700         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1701         enum action {
1702                 NONE,
1703                 PAUSE,
1704                 RESUME,
1705         } action = NONE;
1706
1707         if (!evlist->overwrite_mmap)
1708                 return;
1709
1710         switch (old_state) {
1711         case BKW_MMAP_NOTREADY: {
1712                 if (state != BKW_MMAP_RUNNING)
1713                         goto state_err;
1714                 break;
1715         }
1716         case BKW_MMAP_RUNNING: {
1717                 if (state != BKW_MMAP_DATA_PENDING)
1718                         goto state_err;
1719                 action = PAUSE;
1720                 break;
1721         }
1722         case BKW_MMAP_DATA_PENDING: {
1723                 if (state != BKW_MMAP_EMPTY)
1724                         goto state_err;
1725                 break;
1726         }
1727         case BKW_MMAP_EMPTY: {
1728                 if (state != BKW_MMAP_RUNNING)
1729                         goto state_err;
1730                 action = RESUME;
1731                 break;
1732         }
1733         default:
1734                 WARN_ONCE(1, "Shouldn't get there\n");
1735         }
1736
1737         evlist->bkw_mmap_state = state;
1738
1739         switch (action) {
1740         case PAUSE:
1741                 evlist__pause(evlist);
1742                 break;
1743         case RESUME:
1744                 evlist__resume(evlist);
1745                 break;
1746         case NONE:
1747         default:
1748                 break;
1749         }
1750
1751 state_err:
1752         return;
1753 }
1754
1755 bool evlist__exclude_kernel(struct evlist *evlist)
1756 {
1757         struct evsel *evsel;
1758
1759         evlist__for_each_entry(evlist, evsel) {
1760                 if (!evsel->core.attr.exclude_kernel)
1761                         return false;
1762         }
1763
1764         return true;
1765 }
1766
1767 /*
1768  * Events in data file are not collect in groups, but we still want
1769  * the group display. Set the artificial group and set the leader's
1770  * forced_leader flag to notify the display code.
1771  */
1772 void evlist__force_leader(struct evlist *evlist)
1773 {
1774         if (!evlist->core.nr_groups) {
1775                 struct evsel *leader = evlist__first(evlist);
1776
1777                 evlist__set_leader(evlist);
1778                 leader->forced_leader = true;
1779         }
1780 }
1781
1782 struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1783 {
1784         struct evsel *c2, *leader;
1785         bool is_open = true;
1786
1787         leader = evsel__leader(evsel);
1788
1789         pr_debug("Weak group for %s/%d failed\n",
1790                         leader->name, leader->core.nr_members);
1791
1792         /*
1793          * for_each_group_member doesn't work here because it doesn't
1794          * include the first entry.
1795          */
1796         evlist__for_each_entry(evsel_list, c2) {
1797                 if (c2 == evsel)
1798                         is_open = false;
1799                 if (evsel__has_leader(c2, leader)) {
1800                         if (is_open && close)
1801                                 perf_evsel__close(&c2->core);
1802                         /*
1803                          * We want to close all members of the group and reopen
1804                          * them. Some events, like Intel topdown, require being
1805                          * in a group and so keep these in the group.
1806                          */
1807                         evsel__remove_from_group(c2, leader);
1808
1809                         /*
1810                          * Set this for all former members of the group
1811                          * to indicate they get reopened.
1812                          */
1813                         c2->reset_group = true;
1814                 }
1815         }
1816         /* Reset the leader count if all entries were removed. */
1817         if (leader->core.nr_members == 1)
1818                 leader->core.nr_members = 0;
1819         return leader;
1820 }
1821
1822 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1823 {
1824         char *s, *p;
1825         int ret = 0, fd;
1826
1827         if (strncmp(str, "fifo:", 5))
1828                 return -EINVAL;
1829
1830         str += 5;
1831         if (!*str || *str == ',')
1832                 return -EINVAL;
1833
1834         s = strdup(str);
1835         if (!s)
1836                 return -ENOMEM;
1837
1838         p = strchr(s, ',');
1839         if (p)
1840                 *p = '\0';
1841
1842         /*
1843          * O_RDWR avoids POLLHUPs which is necessary to allow the other
1844          * end of a FIFO to be repeatedly opened and closed.
1845          */
1846         fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1847         if (fd < 0) {
1848                 pr_err("Failed to open '%s'\n", s);
1849                 ret = -errno;
1850                 goto out_free;
1851         }
1852         *ctl_fd = fd;
1853         *ctl_fd_close = true;
1854
1855         if (p && *++p) {
1856                 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1857                 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1858                 if (fd < 0) {
1859                         pr_err("Failed to open '%s'\n", p);
1860                         ret = -errno;
1861                         goto out_free;
1862                 }
1863                 *ctl_fd_ack = fd;
1864         }
1865
1866 out_free:
1867         free(s);
1868         return ret;
1869 }
1870
1871 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1872 {
1873         char *comma = NULL, *endptr = NULL;
1874
1875         *ctl_fd_close = false;
1876
1877         if (strncmp(str, "fd:", 3))
1878                 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1879
1880         *ctl_fd = strtoul(&str[3], &endptr, 0);
1881         if (endptr == &str[3])
1882                 return -EINVAL;
1883
1884         comma = strchr(str, ',');
1885         if (comma) {
1886                 if (endptr != comma)
1887                         return -EINVAL;
1888
1889                 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1890                 if (endptr == comma + 1 || *endptr != '\0')
1891                         return -EINVAL;
1892         }
1893
1894         return 0;
1895 }
1896
1897 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1898 {
1899         if (*ctl_fd_close) {
1900                 *ctl_fd_close = false;
1901                 close(ctl_fd);
1902                 if (ctl_fd_ack >= 0)
1903                         close(ctl_fd_ack);
1904         }
1905 }
1906
1907 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1908 {
1909         if (fd == -1) {
1910                 pr_debug("Control descriptor is not initialized\n");
1911                 return 0;
1912         }
1913
1914         evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1915                                                      fdarray_flag__nonfilterable |
1916                                                      fdarray_flag__non_perf_event);
1917         if (evlist->ctl_fd.pos < 0) {
1918                 evlist->ctl_fd.pos = -1;
1919                 pr_err("Failed to add ctl fd entry: %m\n");
1920                 return -1;
1921         }
1922
1923         evlist->ctl_fd.fd = fd;
1924         evlist->ctl_fd.ack = ack;
1925
1926         return 0;
1927 }
1928
1929 bool evlist__ctlfd_initialized(struct evlist *evlist)
1930 {
1931         return evlist->ctl_fd.pos >= 0;
1932 }
1933
1934 int evlist__finalize_ctlfd(struct evlist *evlist)
1935 {
1936         struct pollfd *entries = evlist->core.pollfd.entries;
1937
1938         if (!evlist__ctlfd_initialized(evlist))
1939                 return 0;
1940
1941         entries[evlist->ctl_fd.pos].fd = -1;
1942         entries[evlist->ctl_fd.pos].events = 0;
1943         entries[evlist->ctl_fd.pos].revents = 0;
1944
1945         evlist->ctl_fd.pos = -1;
1946         evlist->ctl_fd.ack = -1;
1947         evlist->ctl_fd.fd = -1;
1948
1949         return 0;
1950 }
1951
1952 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1953                               char *cmd_data, size_t data_size)
1954 {
1955         int err;
1956         char c;
1957         size_t bytes_read = 0;
1958
1959         *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1960         memset(cmd_data, 0, data_size);
1961         data_size--;
1962
1963         do {
1964                 err = read(evlist->ctl_fd.fd, &c, 1);
1965                 if (err > 0) {
1966                         if (c == '\n' || c == '\0')
1967                                 break;
1968                         cmd_data[bytes_read++] = c;
1969                         if (bytes_read == data_size)
1970                                 break;
1971                         continue;
1972                 } else if (err == -1) {
1973                         if (errno == EINTR)
1974                                 continue;
1975                         if (errno == EAGAIN || errno == EWOULDBLOCK)
1976                                 err = 0;
1977                         else
1978                                 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1979                 }
1980                 break;
1981         } while (1);
1982
1983         pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1984                  bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1985
1986         if (bytes_read > 0) {
1987                 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1988                              (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1989                         *cmd = EVLIST_CTL_CMD_ENABLE;
1990                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1991                                     (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1992                         *cmd = EVLIST_CTL_CMD_DISABLE;
1993                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
1994                                     (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
1995                         *cmd = EVLIST_CTL_CMD_SNAPSHOT;
1996                         pr_debug("is snapshot\n");
1997                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
1998                                     (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
1999                         *cmd = EVLIST_CTL_CMD_EVLIST;
2000                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2001                                     (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2002                         *cmd = EVLIST_CTL_CMD_STOP;
2003                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2004                                     (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2005                         *cmd = EVLIST_CTL_CMD_PING;
2006                 }
2007         }
2008
2009         return bytes_read ? (int)bytes_read : err;
2010 }
2011
2012 int evlist__ctlfd_ack(struct evlist *evlist)
2013 {
2014         int err;
2015
2016         if (evlist->ctl_fd.ack == -1)
2017                 return 0;
2018
2019         err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2020                     sizeof(EVLIST_CTL_CMD_ACK_TAG));
2021         if (err == -1)
2022                 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2023
2024         return err;
2025 }
2026
2027 static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2028 {
2029         char *data = cmd_data + cmd_size;
2030
2031         /* no argument */
2032         if (!*data)
2033                 return 0;
2034
2035         /* there's argument */
2036         if (*data == ' ') {
2037                 *arg = data + 1;
2038                 return 1;
2039         }
2040
2041         /* malformed */
2042         return -1;
2043 }
2044
2045 static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2046 {
2047         struct evsel *evsel;
2048         char *name;
2049         int err;
2050
2051         err = get_cmd_arg(cmd_data,
2052                           enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2053                                    sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2054                           &name);
2055         if (err < 0) {
2056                 pr_info("failed: wrong command\n");
2057                 return -1;
2058         }
2059
2060         if (err) {
2061                 evsel = evlist__find_evsel_by_str(evlist, name);
2062                 if (evsel) {
2063                         if (enable)
2064                                 evlist__enable_evsel(evlist, name);
2065                         else
2066                                 evlist__disable_evsel(evlist, name);
2067                         pr_info("Event %s %s\n", evsel->name,
2068                                 enable ? "enabled" : "disabled");
2069                 } else {
2070                         pr_info("failed: can't find '%s' event\n", name);
2071                 }
2072         } else {
2073                 if (enable) {
2074                         evlist__enable(evlist);
2075                         pr_info(EVLIST_ENABLED_MSG);
2076                 } else {
2077                         evlist__disable(evlist);
2078                         pr_info(EVLIST_DISABLED_MSG);
2079                 }
2080         }
2081
2082         return 0;
2083 }
2084
2085 static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2086 {
2087         struct perf_attr_details details = { .verbose = false, };
2088         struct evsel *evsel;
2089         char *arg;
2090         int err;
2091
2092         err = get_cmd_arg(cmd_data,
2093                           sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2094                           &arg);
2095         if (err < 0) {
2096                 pr_info("failed: wrong command\n");
2097                 return -1;
2098         }
2099
2100         if (err) {
2101                 if (!strcmp(arg, "-v")) {
2102                         details.verbose = true;
2103                 } else if (!strcmp(arg, "-g")) {
2104                         details.event_group = true;
2105                 } else if (!strcmp(arg, "-F")) {
2106                         details.freq = true;
2107                 } else {
2108                         pr_info("failed: wrong command\n");
2109                         return -1;
2110                 }
2111         }
2112
2113         evlist__for_each_entry(evlist, evsel)
2114                 evsel__fprintf(evsel, &details, stderr);
2115
2116         return 0;
2117 }
2118
2119 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2120 {
2121         int err = 0;
2122         char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2123         int ctlfd_pos = evlist->ctl_fd.pos;
2124         struct pollfd *entries = evlist->core.pollfd.entries;
2125
2126         if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2127                 return 0;
2128
2129         if (entries[ctlfd_pos].revents & POLLIN) {
2130                 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2131                                          EVLIST_CTL_CMD_MAX_LEN);
2132                 if (err > 0) {
2133                         switch (*cmd) {
2134                         case EVLIST_CTL_CMD_ENABLE:
2135                         case EVLIST_CTL_CMD_DISABLE:
2136                                 err = evlist__ctlfd_enable(evlist, cmd_data,
2137                                                            *cmd == EVLIST_CTL_CMD_ENABLE);
2138                                 break;
2139                         case EVLIST_CTL_CMD_EVLIST:
2140                                 err = evlist__ctlfd_list(evlist, cmd_data);
2141                                 break;
2142                         case EVLIST_CTL_CMD_SNAPSHOT:
2143                         case EVLIST_CTL_CMD_STOP:
2144                         case EVLIST_CTL_CMD_PING:
2145                                 break;
2146                         case EVLIST_CTL_CMD_ACK:
2147                         case EVLIST_CTL_CMD_UNSUPPORTED:
2148                         default:
2149                                 pr_debug("ctlfd: unsupported %d\n", *cmd);
2150                                 break;
2151                         }
2152                         if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2153                               *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2154                                 evlist__ctlfd_ack(evlist);
2155                 }
2156         }
2157
2158         if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2159                 evlist__finalize_ctlfd(evlist);
2160         else
2161                 entries[ctlfd_pos].revents = 0;
2162
2163         return err;
2164 }
2165
2166 /**
2167  * struct event_enable_time - perf record -D/--delay single time range.
2168  * @start: start of time range to enable events in milliseconds
2169  * @end: end of time range to enable events in milliseconds
2170  *
2171  * N.B. this structure is also accessed as an array of int.
2172  */
2173 struct event_enable_time {
2174         int     start;
2175         int     end;
2176 };
2177
2178 static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2179 {
2180         const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2181         int ret, start, end, n;
2182
2183         ret = sscanf(str, fmt, &start, &end, &n);
2184         if (ret != 2 || end <= start)
2185                 return -EINVAL;
2186         if (range) {
2187                 range->start = start;
2188                 range->end = end;
2189         }
2190         return n;
2191 }
2192
2193 static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2194 {
2195         int incr = !!range;
2196         bool first = true;
2197         ssize_t ret, cnt;
2198
2199         for (cnt = 0; *str; cnt++) {
2200                 ret = parse_event_enable_time(str, range, first);
2201                 if (ret < 0)
2202                         return ret;
2203                 /* Check no overlap */
2204                 if (!first && range && range->start <= range[-1].end)
2205                         return -EINVAL;
2206                 str += ret;
2207                 range += incr;
2208                 first = false;
2209         }
2210         return cnt;
2211 }
2212
2213 /**
2214  * struct event_enable_timer - control structure for perf record -D/--delay.
2215  * @evlist: event list
2216  * @times: time ranges that events are enabled (N.B. this is also accessed as an
2217  *         array of int)
2218  * @times_cnt: number of time ranges
2219  * @timerfd: timer file descriptor
2220  * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2221  * @times_step: current position in (int *)@times)[],
2222  *              refer event_enable_timer__process()
2223  *
2224  * Note, this structure is only used when there are time ranges, not when there
2225  * is only an initial delay.
2226  */
2227 struct event_enable_timer {
2228         struct evlist *evlist;
2229         struct event_enable_time *times;
2230         size_t  times_cnt;
2231         int     timerfd;
2232         int     pollfd_pos;
2233         size_t  times_step;
2234 };
2235
2236 static int str_to_delay(const char *str)
2237 {
2238         char *endptr;
2239         long d;
2240
2241         d = strtol(str, &endptr, 10);
2242         if (*endptr || d > INT_MAX || d < -1)
2243                 return 0;
2244         return d;
2245 }
2246
2247 int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2248                                     const char *str, int unset)
2249 {
2250         enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2251         struct event_enable_timer *eet;
2252         ssize_t times_cnt;
2253         ssize_t ret;
2254         int err;
2255
2256         if (unset)
2257                 return 0;
2258
2259         opts->initial_delay = str_to_delay(str);
2260         if (opts->initial_delay)
2261                 return 0;
2262
2263         ret = parse_event_enable_times(str, NULL);
2264         if (ret < 0)
2265                 return ret;
2266
2267         times_cnt = ret;
2268         if (times_cnt == 0)
2269                 return -EINVAL;
2270
2271         eet = zalloc(sizeof(*eet));
2272         if (!eet)
2273                 return -ENOMEM;
2274
2275         eet->times = calloc(times_cnt, sizeof(*eet->times));
2276         if (!eet->times) {
2277                 err = -ENOMEM;
2278                 goto free_eet;
2279         }
2280
2281         if (parse_event_enable_times(str, eet->times) != times_cnt) {
2282                 err = -EINVAL;
2283                 goto free_eet_times;
2284         }
2285
2286         eet->times_cnt = times_cnt;
2287
2288         eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2289         if (eet->timerfd == -1) {
2290                 err = -errno;
2291                 pr_err("timerfd_create failed: %s\n", strerror(errno));
2292                 goto free_eet_times;
2293         }
2294
2295         eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2296         if (eet->pollfd_pos < 0) {
2297                 err = eet->pollfd_pos;
2298                 goto close_timerfd;
2299         }
2300
2301         eet->evlist = evlist;
2302         evlist->eet = eet;
2303         opts->initial_delay = eet->times[0].start;
2304
2305         return 0;
2306
2307 close_timerfd:
2308         close(eet->timerfd);
2309 free_eet_times:
2310         free(eet->times);
2311 free_eet:
2312         free(eet);
2313         return err;
2314 }
2315
2316 static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2317 {
2318         struct itimerspec its = {
2319                 .it_value.tv_sec = ms / MSEC_PER_SEC,
2320                 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2321         };
2322         int err = 0;
2323
2324         if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2325                 err = -errno;
2326                 pr_err("timerfd_settime failed: %s\n", strerror(errno));
2327         }
2328         return err;
2329 }
2330
2331 int event_enable_timer__start(struct event_enable_timer *eet)
2332 {
2333         int ms;
2334
2335         if (!eet)
2336                 return 0;
2337
2338         ms = eet->times[0].end - eet->times[0].start;
2339         eet->times_step = 1;
2340
2341         return event_enable_timer__set_timer(eet, ms);
2342 }
2343
2344 int event_enable_timer__process(struct event_enable_timer *eet)
2345 {
2346         struct pollfd *entries;
2347         short revents;
2348
2349         if (!eet)
2350                 return 0;
2351
2352         entries = eet->evlist->core.pollfd.entries;
2353         revents = entries[eet->pollfd_pos].revents;
2354         entries[eet->pollfd_pos].revents = 0;
2355
2356         if (revents & POLLIN) {
2357                 size_t step = eet->times_step;
2358                 size_t pos = step / 2;
2359
2360                 if (step & 1) {
2361                         evlist__disable_non_dummy(eet->evlist);
2362                         pr_info(EVLIST_DISABLED_MSG);
2363                         if (pos >= eet->times_cnt - 1) {
2364                                 /* Disarm timer */
2365                                 event_enable_timer__set_timer(eet, 0);
2366                                 return 1; /* Stop */
2367                         }
2368                 } else {
2369                         evlist__enable_non_dummy(eet->evlist);
2370                         pr_info(EVLIST_ENABLED_MSG);
2371                 }
2372
2373                 step += 1;
2374                 pos = step / 2;
2375
2376                 if (pos < eet->times_cnt) {
2377                         int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2378                         int ms = times[step] - times[step - 1];
2379
2380                         eet->times_step = step;
2381                         return event_enable_timer__set_timer(eet, ms);
2382                 }
2383         }
2384
2385         return 0;
2386 }
2387
2388 void event_enable_timer__exit(struct event_enable_timer **ep)
2389 {
2390         if (!ep || !*ep)
2391                 return;
2392         free((*ep)->times);
2393         zfree(ep);
2394 }
2395
2396 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2397 {
2398         struct evsel *evsel;
2399
2400         evlist__for_each_entry(evlist, evsel) {
2401                 if (evsel->core.idx == idx)
2402                         return evsel;
2403         }
2404         return NULL;
2405 }
2406
2407 int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
2408 {
2409         struct evsel *evsel;
2410         int printed = 0;
2411
2412         evlist__for_each_entry(evlist, evsel) {
2413                 if (evsel__is_dummy_event(evsel))
2414                         continue;
2415                 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) {
2416                         printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel));
2417                 } else {
2418                         printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : "");
2419                         break;
2420                 }
2421         }
2422
2423         return printed;
2424 }
2425
2426 void evlist__check_mem_load_aux(struct evlist *evlist)
2427 {
2428         struct evsel *leader, *evsel, *pos;
2429
2430         /*
2431          * For some platforms, the 'mem-loads' event is required to use
2432          * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2433          * must be the group leader. Now we disable this group before reporting
2434          * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2435          * any valid memory load information.
2436          */
2437         evlist__for_each_entry(evlist, evsel) {
2438                 leader = evsel__leader(evsel);
2439                 if (leader == evsel)
2440                         continue;
2441
2442                 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2443                         for_each_group_evsel(pos, leader) {
2444                                 evsel__set_leader(pos, pos);
2445                                 pos->core.nr_members = 0;
2446                         }
2447                 }
2448         }
2449 }