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