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