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 <signal.h>
  28 #include <unistd.h>
  29 #include <sched.h>
  30 #include <stdlib.h>
  31
  32 #include "parse-events.h"
  33 #include <subcmd/parse-options.h>
  34
  35 #include <fcntl.h>
  36 #include <sys/ioctl.h>
  37 #include <sys/mman.h>
  38
  39 #include <linux/bitops.h>
  40 #include <linux/hash.h>
  41 #include <linux/log2.h>
  42 #include <linux/err.h>
  43 #include <linux/string.h>
  44 #include <linux/zalloc.h>
  45 #include <perf/evlist.h>
  46 #include <perf/evsel.h>
  47 #include <perf/cpumap.h>
  48 #include <perf/mmap.h>
  49
  50 #include <internal/xyarray.h>
  51
  52 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  53 int sigqueue(pid_t pid, int sig, const union sigval value);
  54 #endif
  55
  56 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  57 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  58
  59 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  60                   struct perf_thread_map *threads)
  61 {
  62         perf_evlist__init(&evlist->core);
  63         perf_evlist__set_maps(&evlist->core, cpus, threads);
  64         evlist->workload.pid = -1;
  65         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  66 }
  67
  68 struct evlist *evlist__new(void)
  69 {
  70         struct evlist *evlist = zalloc(sizeof(*evlist));
  71
  72         if (evlist != NULL)
  73                 evlist__init(evlist, NULL, NULL);
  74
  75         return evlist;
  76 }
  77
  78 struct evlist *perf_evlist__new_default(void)
  79 {
  80         struct evlist *evlist = evlist__new();
  81
  82         if (evlist && perf_evlist__add_default(evlist)) {
  83                 evlist__delete(evlist);
  84                 evlist = NULL;
  85         }
  86
  87         return evlist;
  88 }
  89
  90 struct evlist *perf_evlist__new_dummy(void)
  91 {
  92         struct evlist *evlist = evlist__new();
  93
  94         if (evlist && perf_evlist__add_dummy(evlist)) {
  95                 evlist__delete(evlist);
  96                 evlist = NULL;
  97         }
  98
  99         return evlist;
 100 }
 101
 102 /**
 103  * perf_evlist__set_id_pos - set the positions of event ids.
 104  * @evlist: selected event list
 105  *
 106  * Events with compatible sample types all have the same id_pos
 107  * and is_pos.  For convenience, put a copy on evlist.
 108  */
 109 void perf_evlist__set_id_pos(struct evlist *evlist)
 110 {
 111         struct evsel *first = evlist__first(evlist);
 112
 113         evlist->id_pos = first->id_pos;
 114         evlist->is_pos = first->is_pos;
 115 }
 116
 117 static void perf_evlist__update_id_pos(struct evlist *evlist)
 118 {
 119         struct evsel *evsel;
 120
 121         evlist__for_each_entry(evlist, evsel)
 122                 perf_evsel__calc_id_pos(evsel);
 123
 124         perf_evlist__set_id_pos(evlist);
 125 }
 126
 127 static void evlist__purge(struct evlist *evlist)
 128 {
 129         struct evsel *pos, *n;
 130
 131         evlist__for_each_entry_safe(evlist, n, pos) {
 132                 list_del_init(&pos->core.node);
 133                 pos->evlist = NULL;
 134                 evsel__delete(pos);
 135         }
 136
 137         evlist->core.nr_entries = 0;
 138 }
 139
 140 void evlist__exit(struct evlist *evlist)
 141 {
 142         zfree(&evlist->mmap);
 143         zfree(&evlist->overwrite_mmap);
 144         perf_evlist__exit(&evlist->core);
 145 }
 146
 147 void evlist__delete(struct evlist *evlist)
 148 {
 149         if (evlist == NULL)
 150                 return;
 151
 152         evlist__munmap(evlist);
 153         evlist__close(evlist);
 154         evlist__purge(evlist);
 155         evlist__exit(evlist);
 156         free(evlist);
 157 }
 158
 159 void evlist__add(struct evlist *evlist, struct evsel *entry)
 160 {
 161         entry->evlist = evlist;
 162         entry->idx = evlist->core.nr_entries;
 163         entry->tracking = !entry->idx;
 164
 165         perf_evlist__add(&evlist->core, &entry->core);
 166
 167         if (evlist->core.nr_entries == 1)
 168                 perf_evlist__set_id_pos(evlist);
 169 }
 170
 171 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 172 {
 173         evsel->evlist = NULL;
 174         perf_evlist__remove(&evlist->core, &evsel->core);
 175 }
 176
 177 void perf_evlist__splice_list_tail(struct evlist *evlist,
 178                                    struct list_head *list)
 179 {
 180         struct evsel *evsel, *temp;
 181
 182         __evlist__for_each_entry_safe(list, temp, evsel) {
 183                 list_del_init(&evsel->core.node);
 184                 evlist__add(evlist, evsel);
 185         }
 186 }
 187
 188 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 189                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 190 {
 191         struct evsel *evsel;
 192         size_t i;
 193         int err;
 194
 195         for (i = 0; i < nr_assocs; i++) {
 196                 // Adding a handler for an event not in this evlist, just ignore it.
 197                 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 198                 if (evsel == NULL)
 199                         continue;
 200
 201                 err = -EEXIST;
 202                 if (evsel->handler != NULL)
 203                         goto out;
 204                 evsel->handler = assocs[i].handler;
 205         }
 206
 207         err = 0;
 208 out:
 209         return err;
 210 }
 211
 212 void __perf_evlist__set_leader(struct list_head *list)
 213 {
 214         struct evsel *evsel, *leader;
 215
 216         leader = list_entry(list->next, struct evsel, core.node);
 217         evsel = list_entry(list->prev, struct evsel, core.node);
 218
 219         leader->core.nr_members = evsel->idx - leader->idx + 1;
 220
 221         __evlist__for_each_entry(list, evsel) {
 222                 evsel->leader = leader;
 223         }
 224 }
 225
 226 void perf_evlist__set_leader(struct evlist *evlist)
 227 {
 228         if (evlist->core.nr_entries) {
 229                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 230                 __perf_evlist__set_leader(&evlist->core.entries);
 231         }
 232 }
 233
 234 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
 235 {
 236         struct evsel *evsel = perf_evsel__new_cycles(precise);
 237
 238         if (evsel == NULL)
 239                 return -ENOMEM;
 240
 241         evlist__add(evlist, evsel);
 242         return 0;
 243 }
 244
 245 int perf_evlist__add_dummy(struct evlist *evlist)
 246 {
 247         struct perf_event_attr attr = {
 248                 .type   = PERF_TYPE_SOFTWARE,
 249                 .config = PERF_COUNT_SW_DUMMY,
 250                 .size   = sizeof(attr), /* to capture ABI version */
 251         };
 252         struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
 253
 254         if (evsel == NULL)
 255                 return -ENOMEM;
 256
 257         evlist__add(evlist, evsel);
 258         return 0;
 259 }
 260
 261 static int evlist__add_attrs(struct evlist *evlist,
 262                                   struct perf_event_attr *attrs, size_t nr_attrs)
 263 {
 264         struct evsel *evsel, *n;
 265         LIST_HEAD(head);
 266         size_t i;
 267
 268         for (i = 0; i < nr_attrs; i++) {
 269                 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 270                 if (evsel == NULL)
 271                         goto out_delete_partial_list;
 272                 list_add_tail(&evsel->core.node, &head);
 273         }
 274
 275         perf_evlist__splice_list_tail(evlist, &head);
 276
 277         return 0;
 278
 279 out_delete_partial_list:
 280         __evlist__for_each_entry_safe(&head, n, evsel)
 281                 evsel__delete(evsel);
 282         return -1;
 283 }
 284
 285 int __perf_evlist__add_default_attrs(struct evlist *evlist,
 286                                      struct perf_event_attr *attrs, size_t nr_attrs)
 287 {
 288         size_t i;
 289
 290         for (i = 0; i < nr_attrs; i++)
 291                 event_attr_init(attrs + i);
 292
 293         return evlist__add_attrs(evlist, attrs, nr_attrs);
 294 }
 295
 296 struct evsel *
 297 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 298 {
 299         struct evsel *evsel;
 300
 301         evlist__for_each_entry(evlist, evsel) {
 302                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 303                     (int)evsel->core.attr.config == id)
 304                         return evsel;
 305         }
 306
 307         return NULL;
 308 }
 309
 310 struct evsel *
 311 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
 312                                      const char *name)
 313 {
 314         struct evsel *evsel;
 315
 316         evlist__for_each_entry(evlist, evsel) {
 317                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 318                     (strcmp(evsel->name, name) == 0))
 319                         return evsel;
 320         }
 321
 322         return NULL;
 323 }
 324
 325 int perf_evlist__add_newtp(struct evlist *evlist,
 326                            const char *sys, const char *name, void *handler)
 327 {
 328         struct evsel *evsel = perf_evsel__newtp(sys, name);
 329
 330         if (IS_ERR(evsel))
 331                 return -1;
 332
 333         evsel->handler = handler;
 334         evlist__add(evlist, evsel);
 335         return 0;
 336 }
 337
 338 static int perf_evlist__nr_threads(struct evlist *evlist,
 339                                    struct evsel *evsel)
 340 {
 341         if (evsel->core.system_wide)
 342                 return 1;
 343         else
 344                 return perf_thread_map__nr(evlist->core.threads);
 345 }
 346
 347 void evlist__cpu_iter_start(struct evlist *evlist)
 348 {
 349         struct evsel *pos;
 350
 351         /*
 352          * Reset the per evsel cpu_iter. This is needed because
 353          * each evsel's cpumap may have a different index space,
 354          * and some operations need the index to modify
 355          * the FD xyarray (e.g. open, close)
 356          */
 357         evlist__for_each_entry(evlist, pos)
 358                 pos->cpu_iter = 0;
 359 }
 360
 361 bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
 362 {
 363         if (ev->cpu_iter >= ev->core.cpus->nr)
 364                 return true;
 365         if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
 366                 return true;
 367         return false;
 368 }
 369
 370 bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
 371 {
 372         if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
 373                 ev->cpu_iter++;
 374                 return false;
 375         }
 376         return true;
 377 }
 378
 379 void evlist__disable(struct evlist *evlist)
 380 {
 381         struct evsel *pos;
 382         struct affinity affinity;
 383         int cpu, i;
 384
 385         if (affinity__setup(&affinity) < 0)
 386                 return;
 387
 388         evlist__for_each_cpu(evlist, i, cpu) {
 389                 affinity__set(&affinity, cpu);
 390
 391                 evlist__for_each_entry(evlist, pos) {
 392                         if (evsel__cpu_iter_skip(pos, cpu))
 393                                 continue;
 394                         if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
 395                                 continue;
 396                         evsel__disable_cpu(pos, pos->cpu_iter - 1);
 397                 }
 398         }
 399         affinity__cleanup(&affinity);
 400         evlist__for_each_entry(evlist, pos) {
 401                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 402                         continue;
 403                 pos->disabled = true;
 404         }
 405
 406         evlist->enabled = false;
 407 }
 408
 409 void evlist__enable(struct evlist *evlist)
 410 {
 411         struct evsel *pos;
 412         struct affinity affinity;
 413         int cpu, i;
 414
 415         if (affinity__setup(&affinity) < 0)
 416                 return;
 417
 418         evlist__for_each_cpu(evlist, i, cpu) {
 419                 affinity__set(&affinity, cpu);
 420
 421                 evlist__for_each_entry(evlist, pos) {
 422                         if (evsel__cpu_iter_skip(pos, cpu))
 423                                 continue;
 424                         if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 425                                 continue;
 426                         evsel__enable_cpu(pos, pos->cpu_iter - 1);
 427                 }
 428         }
 429         affinity__cleanup(&affinity);
 430         evlist__for_each_entry(evlist, pos) {
 431                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 432                         continue;
 433                 pos->disabled = false;
 434         }
 435
 436         evlist->enabled = true;
 437 }
 438
 439 void perf_evlist__toggle_enable(struct evlist *evlist)
 440 {
 441         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 442 }
 443
 444 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
 445                                          struct evsel *evsel, int cpu)
 446 {
 447         int thread;
 448         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 449
 450         if (!evsel->core.fd)
 451                 return -EINVAL;
 452
 453         for (thread = 0; thread < nr_threads; thread++) {
 454                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 455                 if (err)
 456                         return err;
 457         }
 458         return 0;
 459 }
 460
 461 static int perf_evlist__enable_event_thread(struct evlist *evlist,
 462                                             struct evsel *evsel,
 463                                             int thread)
 464 {
 465         int cpu;
 466         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 467
 468         if (!evsel->core.fd)
 469                 return -EINVAL;
 470
 471         for (cpu = 0; cpu < nr_cpus; cpu++) {
 472                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 473                 if (err)
 474                         return err;
 475         }
 476         return 0;
 477 }
 478
 479 int perf_evlist__enable_event_idx(struct evlist *evlist,
 480                                   struct evsel *evsel, int idx)
 481 {
 482         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 483
 484         if (per_cpu_mmaps)
 485                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 486         else
 487                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 488 }
 489
 490 int evlist__add_pollfd(struct evlist *evlist, int fd)
 491 {
 492         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
 493 }
 494
 495 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 496 {
 497         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 498 }
 499
 500 int evlist__poll(struct evlist *evlist, int timeout)
 501 {
 502         return perf_evlist__poll(&evlist->core, timeout);
 503 }
 504
 505 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
 506 {
 507         struct hlist_head *head;
 508         struct perf_sample_id *sid;
 509         int hash;
 510
 511         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 512         head = &evlist->core.heads[hash];
 513
 514         hlist_for_each_entry(sid, head, node)
 515                 if (sid->id == id)
 516                         return sid;
 517
 518         return NULL;
 519 }
 520
 521 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
 522 {
 523         struct perf_sample_id *sid;
 524
 525         if (evlist->core.nr_entries == 1 || !id)
 526                 return evlist__first(evlist);
 527
 528         sid = perf_evlist__id2sid(evlist, id);
 529         if (sid)
 530                 return container_of(sid->evsel, struct evsel, core);
 531
 532         if (!perf_evlist__sample_id_all(evlist))
 533                 return evlist__first(evlist);
 534
 535         return NULL;
 536 }
 537
 538 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
 539                                                 u64 id)
 540 {
 541         struct perf_sample_id *sid;
 542
 543         if (!id)
 544                 return NULL;
 545
 546         sid = perf_evlist__id2sid(evlist, id);
 547         if (sid)
 548                 return container_of(sid->evsel, struct evsel, core);
 549
 550         return NULL;
 551 }
 552
 553 static int perf_evlist__event2id(struct evlist *evlist,
 554                                  union perf_event *event, u64 *id)
 555 {
 556         const __u64 *array = event->sample.array;
 557         ssize_t n;
 558
 559         n = (event->header.size - sizeof(event->header)) >> 3;
 560
 561         if (event->header.type == PERF_RECORD_SAMPLE) {
 562                 if (evlist->id_pos >= n)
 563                         return -1;
 564                 *id = array[evlist->id_pos];
 565         } else {
 566                 if (evlist->is_pos > n)
 567                         return -1;
 568                 n -= evlist->is_pos;
 569                 *id = array[n];
 570         }
 571         return 0;
 572 }
 573
 574 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
 575                                             union perf_event *event)
 576 {
 577         struct evsel *first = evlist__first(evlist);
 578         struct hlist_head *head;
 579         struct perf_sample_id *sid;
 580         int hash;
 581         u64 id;
 582
 583         if (evlist->core.nr_entries == 1)
 584                 return first;
 585
 586         if (!first->core.attr.sample_id_all &&
 587             event->header.type != PERF_RECORD_SAMPLE)
 588                 return first;
 589
 590         if (perf_evlist__event2id(evlist, event, &id))
 591                 return NULL;
 592
 593         /* Synthesized events have an id of zero */
 594         if (!id)
 595                 return first;
 596
 597         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 598         head = &evlist->core.heads[hash];
 599
 600         hlist_for_each_entry(sid, head, node) {
 601                 if (sid->id == id)
 602                         return container_of(sid->evsel, struct evsel, core);
 603         }
 604         return NULL;
 605 }
 606
 607 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
 608 {
 609         int i;
 610
 611         if (!evlist->overwrite_mmap)
 612                 return 0;
 613
 614         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 615                 int fd = evlist->overwrite_mmap[i].core.fd;
 616                 int err;
 617
 618                 if (fd < 0)
 619                         continue;
 620                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 621                 if (err)
 622                         return err;
 623         }
 624         return 0;
 625 }
 626
 627 static int perf_evlist__pause(struct evlist *evlist)
 628 {
 629         return perf_evlist__set_paused(evlist, true);
 630 }
 631
 632 static int perf_evlist__resume(struct evlist *evlist)
 633 {
 634         return perf_evlist__set_paused(evlist, false);
 635 }
 636
 637 static void evlist__munmap_nofree(struct evlist *evlist)
 638 {
 639         int i;
 640
 641         if (evlist->mmap)
 642                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 643                         perf_mmap__munmap(&evlist->mmap[i].core);
 644
 645         if (evlist->overwrite_mmap)
 646                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 647                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 648 }
 649
 650 void evlist__munmap(struct evlist *evlist)
 651 {
 652         evlist__munmap_nofree(evlist);
 653         zfree(&evlist->mmap);
 654         zfree(&evlist->overwrite_mmap);
 655 }
 656
 657 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 658 {
 659         struct mmap *m = container_of(map, struct mmap, core);
 660
 661         mmap__munmap(m);
 662 }
 663
 664 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 665                                        bool overwrite)
 666 {
 667         int i;
 668         struct mmap *map;
 669
 670         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 671         if (!map)
 672                 return NULL;
 673
 674         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 675                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 676
 677                 /*
 678                  * When the perf_mmap() call is made we grab one refcount, plus
 679                  * one extra to let perf_mmap__consume() get the last
 680                  * events after all real references (perf_mmap__get()) are
 681                  * dropped.
 682                  *
 683                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 684                  * thus does perf_mmap__get() on it.
 685                  */
 686                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 687         }
 688
 689         return map;
 690 }
 691
 692 static void
 693 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 694                          struct perf_mmap_param *_mp,
 695                          int idx, bool per_cpu)
 696 {
 697         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 698         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 699
 700         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
 701 }
 702
 703 static struct perf_mmap*
 704 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 705 {
 706         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 707         struct mmap *maps;
 708
 709         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 710
 711         if (!maps) {
 712                 maps = evlist__alloc_mmap(evlist, overwrite);
 713                 if (!maps)
 714                         return NULL;
 715
 716                 if (overwrite) {
 717                         evlist->overwrite_mmap = maps;
 718                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 719                                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 720                 } else {
 721                         evlist->mmap = maps;
 722                 }
 723         }
 724
 725         return &maps[idx].core;
 726 }
 727
 728 static int
 729 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 730                           int output, int cpu)
 731 {
 732         struct mmap *map = container_of(_map, struct mmap, core);
 733         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 734
 735         return mmap__mmap(map, mp, output, cpu);
 736 }
 737
 738 unsigned long perf_event_mlock_kb_in_pages(void)
 739 {
 740         unsigned long pages;
 741         int max;
 742
 743         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 744                 /*
 745                  * Pick a once upon a time good value, i.e. things look
 746                  * strange since we can't read a sysctl value, but lets not
 747                  * die yet...
 748                  */
 749                 max = 512;
 750         } else {
 751                 max -= (page_size / 1024);
 752         }
 753
 754         pages = (max * 1024) / page_size;
 755         if (!is_power_of_2(pages))
 756                 pages = rounddown_pow_of_two(pages);
 757
 758         return pages;
 759 }
 760
 761 size_t evlist__mmap_size(unsigned long pages)
 762 {
 763         if (pages == UINT_MAX)
 764                 pages = perf_event_mlock_kb_in_pages();
 765         else if (!is_power_of_2(pages))
 766                 return 0;
 767
 768         return (pages + 1) * page_size;
 769 }
 770
 771 static long parse_pages_arg(const char *str, unsigned long min,
 772                             unsigned long max)
 773 {
 774         unsigned long pages, val;
 775         static struct parse_tag tags[] = {
 776                 { .tag  = 'B', .mult = 1       },
 777                 { .tag  = 'K', .mult = 1 << 10 },
 778                 { .tag  = 'M', .mult = 1 << 20 },
 779                 { .tag  = 'G', .mult = 1 << 30 },
 780                 { .tag  = 0 },
 781         };
 782
 783         if (str == NULL)
 784                 return -EINVAL;
 785
 786         val = parse_tag_value(str, tags);
 787         if (val != (unsigned long) -1) {
 788                 /* we got file size value */
 789                 pages = PERF_ALIGN(val, page_size) / page_size;
 790         } else {
 791                 /* we got pages count value */
 792                 char *eptr;
 793                 pages = strtoul(str, &eptr, 10);
 794                 if (*eptr != '\0')
 795                         return -EINVAL;
 796         }
 797
 798         if (pages == 0 && min == 0) {
 799                 /* leave number of pages at 0 */
 800         } else if (!is_power_of_2(pages)) {
 801                 char buf[100];
 802
 803                 /* round pages up to next power of 2 */
 804                 pages = roundup_pow_of_two(pages);
 805                 if (!pages)
 806                         return -EINVAL;
 807
 808                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 809                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 810                         buf, pages);
 811         }
 812
 813         if (pages > max)
 814                 return -EINVAL;
 815
 816         return pages;
 817 }
 818
 819 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 820 {
 821         unsigned long max = UINT_MAX;
 822         long pages;
 823
 824         if (max > SIZE_MAX / page_size)
 825                 max = SIZE_MAX / page_size;
 826
 827         pages = parse_pages_arg(str, 1, max);
 828         if (pages < 0) {
 829                 pr_err("Invalid argument for --mmap_pages/-m\n");
 830                 return -1;
 831         }
 832
 833         *mmap_pages = pages;
 834         return 0;
 835 }
 836
 837 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 838                                   int unset __maybe_unused)
 839 {
 840         return __perf_evlist__parse_mmap_pages(opt->value, str);
 841 }
 842
 843 /**
 844  * evlist__mmap_ex - Create mmaps to receive events.
 845  * @evlist: list of events
 846  * @pages: map length in pages
 847  * @overwrite: overwrite older events?
 848  * @auxtrace_pages - auxtrace map length in pages
 849  * @auxtrace_overwrite - overwrite older auxtrace data?
 850  *
 851  * If @overwrite is %false the user needs to signal event consumption using
 852  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 853  * automatically.
 854  *
 855  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 856  * consumption using auxtrace_mmap__write_tail().
 857  *
 858  * Return: %0 on success, negative error code otherwise.
 859  */
 860 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 861                          unsigned int auxtrace_pages,
 862                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 863                          int comp_level)
 864 {
 865         /*
 866          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 867          * Its value is decided by evsel's write_backward.
 868          * So &mp should not be passed through const pointer.
 869          */
 870         struct mmap_params mp = {
 871                 .nr_cblocks     = nr_cblocks,
 872                 .affinity       = affinity,
 873                 .flush          = flush,
 874                 .comp_level     = comp_level
 875         };
 876         struct perf_evlist_mmap_ops ops = {
 877                 .idx  = perf_evlist__mmap_cb_idx,
 878                 .get  = perf_evlist__mmap_cb_get,
 879                 .mmap = perf_evlist__mmap_cb_mmap,
 880         };
 881
 882         evlist->core.mmap_len = evlist__mmap_size(pages);
 883         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 884
 885         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 886                                    auxtrace_pages, auxtrace_overwrite);
 887
 888         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 889 }
 890
 891 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 892 {
 893         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 894 }
 895
 896 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
 897 {
 898         bool all_threads = (target->per_thread && target->system_wide);
 899         struct perf_cpu_map *cpus;
 900         struct perf_thread_map *threads;
 901
 902         /*
 903          * If specify '-a' and '--per-thread' to perf record, perf record
 904          * will override '--per-thread'. target->per_thread = false and
 905          * target->system_wide = true.
 906          *
 907          * If specify '--per-thread' only to perf record,
 908          * target->per_thread = true and target->system_wide = false.
 909          *
 910          * So target->per_thread && target->system_wide is false.
 911          * For perf record, thread_map__new_str doesn't call
 912          * thread_map__new_all_cpus. That will keep perf record's
 913          * current behavior.
 914          *
 915          * For perf stat, it allows the case that target->per_thread and
 916          * target->system_wide are all true. It means to collect system-wide
 917          * per-thread data. thread_map__new_str will call
 918          * thread_map__new_all_cpus to enumerate all threads.
 919          */
 920         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 921                                       all_threads);
 922
 923         if (!threads)
 924                 return -1;
 925
 926         if (target__uses_dummy_map(target))
 927                 cpus = perf_cpu_map__dummy_new();
 928         else
 929                 cpus = perf_cpu_map__new(target->cpu_list);
 930
 931         if (!cpus)
 932                 goto out_delete_threads;
 933
 934         evlist->core.has_user_cpus = !!target->cpu_list;
 935
 936         perf_evlist__set_maps(&evlist->core, cpus, threads);
 937
 938         return 0;
 939
 940 out_delete_threads:
 941         perf_thread_map__put(threads);
 942         return -1;
 943 }
 944
 945 void __perf_evlist__set_sample_bit(struct evlist *evlist,
 946                                    enum perf_event_sample_format bit)
 947 {
 948         struct evsel *evsel;
 949
 950         evlist__for_each_entry(evlist, evsel)
 951                 __perf_evsel__set_sample_bit(evsel, bit);
 952 }
 953
 954 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
 955                                      enum perf_event_sample_format bit)
 956 {
 957         struct evsel *evsel;
 958
 959         evlist__for_each_entry(evlist, evsel)
 960                 __perf_evsel__reset_sample_bit(evsel, bit);
 961 }
 962
 963 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 964 {
 965         struct evsel *evsel;
 966         int err = 0;
 967
 968         evlist__for_each_entry(evlist, evsel) {
 969                 if (evsel->filter == NULL)
 970                         continue;
 971
 972                 /*
 973                  * filters only work for tracepoint event, which doesn't have cpu limit.
 974                  * So evlist and evsel should always be same.
 975                  */
 976                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
 977                 if (err) {
 978                         *err_evsel = evsel;
 979                         break;
 980                 }
 981         }
 982
 983         return err;
 984 }
 985
 986 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
 987 {
 988         struct evsel *evsel;
 989         int err = 0;
 990
 991         if (filter == NULL)
 992                 return -1;
 993
 994         evlist__for_each_entry(evlist, evsel) {
 995                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
 996                         continue;
 997
 998                 err = perf_evsel__set_filter(evsel, filter);
 999                 if (err)
1000                         break;
1001         }
1002
1003         return err;
1004 }
1005
1006 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1007 {
1008         struct evsel *evsel;
1009         int err = 0;
1010
1011         if (filter == NULL)
1012                 return -1;
1013
1014         evlist__for_each_entry(evlist, evsel) {
1015                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1016                         continue;
1017
1018                 err = perf_evsel__append_tp_filter(evsel, filter);
1019                 if (err)
1020                         break;
1021         }
1022
1023         return err;
1024 }
1025
1026 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1027 {
1028         char *filter;
1029         size_t i;
1030
1031         for (i = 0; i < npids; ++i) {
1032                 if (i == 0) {
1033                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1034                                 return NULL;
1035                 } else {
1036                         char *tmp;
1037
1038                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1039                                 goto out_free;
1040
1041                         free(filter);
1042                         filter = tmp;
1043                 }
1044         }
1045
1046         return filter;
1047 out_free:
1048         free(filter);
1049         return NULL;
1050 }
1051
1052 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1053 {
1054         char *filter = asprintf__tp_filter_pids(npids, pids);
1055         int ret = perf_evlist__set_tp_filter(evlist, filter);
1056
1057         free(filter);
1058         return ret;
1059 }
1060
1061 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1062 {
1063         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1064 }
1065
1066 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1067 {
1068         char *filter = asprintf__tp_filter_pids(npids, pids);
1069         int ret = perf_evlist__append_tp_filter(evlist, filter);
1070
1071         free(filter);
1072         return ret;
1073 }
1074
1075 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1076 {
1077         return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1078 }
1079
1080 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1081 {
1082         struct evsel *pos;
1083
1084         if (evlist->core.nr_entries == 1)
1085                 return true;
1086
1087         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1088                 return false;
1089
1090         evlist__for_each_entry(evlist, pos) {
1091                 if (pos->id_pos != evlist->id_pos ||
1092                     pos->is_pos != evlist->is_pos)
1093                         return false;
1094         }
1095
1096         return true;
1097 }
1098
1099 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1100 {
1101         struct evsel *evsel;
1102
1103         if (evlist->combined_sample_type)
1104                 return evlist->combined_sample_type;
1105
1106         evlist__for_each_entry(evlist, evsel)
1107                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1108
1109         return evlist->combined_sample_type;
1110 }
1111
1112 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1113 {
1114         evlist->combined_sample_type = 0;
1115         return __perf_evlist__combined_sample_type(evlist);
1116 }
1117
1118 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1119 {
1120         struct evsel *evsel;
1121         u64 branch_type = 0;
1122
1123         evlist__for_each_entry(evlist, evsel)
1124                 branch_type |= evsel->core.attr.branch_sample_type;
1125         return branch_type;
1126 }
1127
1128 bool perf_evlist__valid_read_format(struct evlist *evlist)
1129 {
1130         struct evsel *first = evlist__first(evlist), *pos = first;
1131         u64 read_format = first->core.attr.read_format;
1132         u64 sample_type = first->core.attr.sample_type;
1133
1134         evlist__for_each_entry(evlist, pos) {
1135                 if (read_format != pos->core.attr.read_format) {
1136                         pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1137                                  read_format, (u64)pos->core.attr.read_format);
1138                 }
1139         }
1140
1141         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1142         if ((sample_type & PERF_SAMPLE_READ) &&
1143             !(read_format & PERF_FORMAT_ID)) {
1144                 return false;
1145         }
1146
1147         return true;
1148 }
1149
1150 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1151 {
1152         struct evsel *first = evlist__first(evlist);
1153         struct perf_sample *data;
1154         u64 sample_type;
1155         u16 size = 0;
1156
1157         if (!first->core.attr.sample_id_all)
1158                 goto out;
1159
1160         sample_type = first->core.attr.sample_type;
1161
1162         if (sample_type & PERF_SAMPLE_TID)
1163                 size += sizeof(data->tid) * 2;
1164
1165        if (sample_type & PERF_SAMPLE_TIME)
1166                 size += sizeof(data->time);
1167
1168         if (sample_type & PERF_SAMPLE_ID)
1169                 size += sizeof(data->id);
1170
1171         if (sample_type & PERF_SAMPLE_STREAM_ID)
1172                 size += sizeof(data->stream_id);
1173
1174         if (sample_type & PERF_SAMPLE_CPU)
1175                 size += sizeof(data->cpu) * 2;
1176
1177         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1178                 size += sizeof(data->id);
1179 out:
1180         return size;
1181 }
1182
1183 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1184 {
1185         struct evsel *first = evlist__first(evlist), *pos = first;
1186
1187         evlist__for_each_entry_continue(evlist, pos) {
1188                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1189                         return false;
1190         }
1191
1192         return true;
1193 }
1194
1195 bool perf_evlist__sample_id_all(struct evlist *evlist)
1196 {
1197         struct evsel *first = evlist__first(evlist);
1198         return first->core.attr.sample_id_all;
1199 }
1200
1201 void perf_evlist__set_selected(struct evlist *evlist,
1202                                struct evsel *evsel)
1203 {
1204         evlist->selected = evsel;
1205 }
1206
1207 void evlist__close(struct evlist *evlist)
1208 {
1209         struct evsel *evsel;
1210         struct affinity affinity;
1211         int cpu, i;
1212
1213         /*
1214          * With perf record core.cpus is usually NULL.
1215          * Use the old method to handle this for now.
1216          */
1217         if (!evlist->core.cpus) {
1218                 evlist__for_each_entry_reverse(evlist, evsel)
1219                         evsel__close(evsel);
1220                 return;
1221         }
1222
1223         if (affinity__setup(&affinity) < 0)
1224                 return;
1225         evlist__for_each_cpu(evlist, i, cpu) {
1226                 affinity__set(&affinity, cpu);
1227
1228                 evlist__for_each_entry_reverse(evlist, evsel) {
1229                         if (evsel__cpu_iter_skip(evsel, cpu))
1230                             continue;
1231                         perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1232                 }
1233         }
1234         affinity__cleanup(&affinity);
1235         evlist__for_each_entry_reverse(evlist, evsel) {
1236                 perf_evsel__free_fd(&evsel->core);
1237                 perf_evsel__free_id(&evsel->core);
1238         }
1239 }
1240
1241 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1242 {
1243         struct perf_cpu_map *cpus;
1244         struct perf_thread_map *threads;
1245         int err = -ENOMEM;
1246
1247         /*
1248          * Try reading /sys/devices/system/cpu/online to get
1249          * an all cpus map.
1250          *
1251          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1252          * code needs an overhaul to properly forward the
1253          * error, and we may not want to do that fallback to a
1254          * default cpu identity map :-\
1255          */
1256         cpus = perf_cpu_map__new(NULL);
1257         if (!cpus)
1258                 goto out;
1259
1260         threads = perf_thread_map__new_dummy();
1261         if (!threads)
1262                 goto out_put;
1263
1264         perf_evlist__set_maps(&evlist->core, cpus, threads);
1265 out:
1266         return err;
1267 out_put:
1268         perf_cpu_map__put(cpus);
1269         goto out;
1270 }
1271
1272 int evlist__open(struct evlist *evlist)
1273 {
1274         struct evsel *evsel;
1275         int err;
1276
1277         /*
1278          * Default: one fd per CPU, all threads, aka systemwide
1279          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1280          */
1281         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1282                 err = perf_evlist__create_syswide_maps(evlist);
1283                 if (err < 0)
1284                         goto out_err;
1285         }
1286
1287         perf_evlist__update_id_pos(evlist);
1288
1289         evlist__for_each_entry(evlist, evsel) {
1290                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1291                 if (err < 0)
1292                         goto out_err;
1293         }
1294
1295         return 0;
1296 out_err:
1297         evlist__close(evlist);
1298         errno = -err;
1299         return err;
1300 }
1301
1302 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1303                                   const char *argv[], bool pipe_output,
1304                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1305 {
1306         int child_ready_pipe[2], go_pipe[2];
1307         char bf;
1308
1309         if (pipe(child_ready_pipe) < 0) {
1310                 perror("failed to create 'ready' pipe");
1311                 return -1;
1312         }
1313
1314         if (pipe(go_pipe) < 0) {
1315                 perror("failed to create 'go' pipe");
1316                 goto out_close_ready_pipe;
1317         }
1318
1319         evlist->workload.pid = fork();
1320         if (evlist->workload.pid < 0) {
1321                 perror("failed to fork");
1322                 goto out_close_pipes;
1323         }
1324
1325         if (!evlist->workload.pid) {
1326                 int ret;
1327
1328                 if (pipe_output)
1329                         dup2(2, 1);
1330
1331                 signal(SIGTERM, SIG_DFL);
1332
1333                 close(child_ready_pipe[0]);
1334                 close(go_pipe[1]);
1335                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1336
1337                 /*
1338                  * Tell the parent we're ready to go
1339                  */
1340                 close(child_ready_pipe[1]);
1341
1342                 /*
1343                  * Wait until the parent tells us to go.
1344                  */
1345                 ret = read(go_pipe[0], &bf, 1);
1346                 /*
1347                  * The parent will ask for the execvp() to be performed by
1348                  * writing exactly one byte, in workload.cork_fd, usually via
1349                  * perf_evlist__start_workload().
1350                  *
1351                  * For cancelling the workload without actually running it,
1352                  * the parent will just close workload.cork_fd, without writing
1353                  * anything, i.e. read will return zero and we just exit()
1354                  * here.
1355                  */
1356                 if (ret != 1) {
1357                         if (ret == -1)
1358                                 perror("unable to read pipe");
1359                         exit(ret);
1360                 }
1361
1362                 execvp(argv[0], (char **)argv);
1363
1364                 if (exec_error) {
1365                         union sigval val;
1366
1367                         val.sival_int = errno;
1368                         if (sigqueue(getppid(), SIGUSR1, val))
1369                                 perror(argv[0]);
1370                 } else
1371                         perror(argv[0]);
1372                 exit(-1);
1373         }
1374
1375         if (exec_error) {
1376                 struct sigaction act = {
1377                         .sa_flags     = SA_SIGINFO,
1378                         .sa_sigaction = exec_error,
1379                 };
1380                 sigaction(SIGUSR1, &act, NULL);
1381         }
1382
1383         if (target__none(target)) {
1384                 if (evlist->core.threads == NULL) {
1385                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1386                                 __func__, __LINE__);
1387                         goto out_close_pipes;
1388                 }
1389                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1390         }
1391
1392         close(child_ready_pipe[1]);
1393         close(go_pipe[0]);
1394         /*
1395          * wait for child to settle
1396          */
1397         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1398                 perror("unable to read pipe");
1399                 goto out_close_pipes;
1400         }
1401
1402         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1403         evlist->workload.cork_fd = go_pipe[1];
1404         close(child_ready_pipe[0]);
1405         return 0;
1406
1407 out_close_pipes:
1408         close(go_pipe[0]);
1409         close(go_pipe[1]);
1410 out_close_ready_pipe:
1411         close(child_ready_pipe[0]);
1412         close(child_ready_pipe[1]);
1413         return -1;
1414 }
1415
1416 int perf_evlist__start_workload(struct evlist *evlist)
1417 {
1418         if (evlist->workload.cork_fd > 0) {
1419                 char bf = 0;
1420                 int ret;
1421                 /*
1422                  * Remove the cork, let it rip!
1423                  */
1424                 ret = write(evlist->workload.cork_fd, &bf, 1);
1425                 if (ret < 0)
1426                         perror("unable to write to pipe");
1427
1428                 close(evlist->workload.cork_fd);
1429                 return ret;
1430         }
1431
1432         return 0;
1433 }
1434
1435 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1436                               struct perf_sample *sample)
1437 {
1438         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1439
1440         if (!evsel)
1441                 return -EFAULT;
1442         return perf_evsel__parse_sample(evsel, event, sample);
1443 }
1444
1445 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1446                                         union perf_event *event,
1447                                         u64 *timestamp)
1448 {
1449         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1450
1451         if (!evsel)
1452                 return -EFAULT;
1453         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1454 }
1455
1456 int perf_evlist__strerror_open(struct evlist *evlist,
1457                                int err, char *buf, size_t size)
1458 {
1459         int printed, value;
1460         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1461
1462         switch (err) {
1463         case EACCES:
1464         case EPERM:
1465                 printed = scnprintf(buf, size,
1466                                     "Error:\t%s.\n"
1467                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1468
1469                 value = perf_event_paranoid();
1470
1471                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1472
1473                 if (value >= 2) {
1474                         printed += scnprintf(buf + printed, size - printed,
1475                                              "For your workloads it needs to be <= 1\nHint:\t");
1476                 }
1477                 printed += scnprintf(buf + printed, size - printed,
1478                                      "For system wide tracing it needs to be set to -1.\n");
1479
1480                 printed += scnprintf(buf + printed, size - printed,
1481                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1482                                     "Hint:\tThe current value is %d.", value);
1483                 break;
1484         case EINVAL: {
1485                 struct evsel *first = evlist__first(evlist);
1486                 int max_freq;
1487
1488                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1489                         goto out_default;
1490
1491                 if (first->core.attr.sample_freq < (u64)max_freq)
1492                         goto out_default;
1493
1494                 printed = scnprintf(buf, size,
1495                                     "Error:\t%s.\n"
1496                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1497                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1498                                     emsg, max_freq, first->core.attr.sample_freq);
1499                 break;
1500         }
1501         default:
1502 out_default:
1503                 scnprintf(buf, size, "%s", emsg);
1504                 break;
1505         }
1506
1507         return 0;
1508 }
1509
1510 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1511 {
1512         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1513         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1514
1515         switch (err) {
1516         case EPERM:
1517                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1518                 printed += scnprintf(buf + printed, size - printed,
1519                                      "Error:\t%s.\n"
1520                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1521                                      "Hint:\tTried using %zd kB.\n",
1522                                      emsg, pages_max_per_user, pages_attempted);
1523
1524                 if (pages_attempted >= pages_max_per_user) {
1525                         printed += scnprintf(buf + printed, size - printed,
1526                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1527                                              pages_max_per_user + pages_attempted);
1528                 }
1529
1530                 printed += scnprintf(buf + printed, size - printed,
1531                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1532                 break;
1533         default:
1534                 scnprintf(buf, size, "%s", emsg);
1535                 break;
1536         }
1537
1538         return 0;
1539 }
1540
1541 void perf_evlist__to_front(struct evlist *evlist,
1542                            struct evsel *move_evsel)
1543 {
1544         struct evsel *evsel, *n;
1545         LIST_HEAD(move);
1546
1547         if (move_evsel == evlist__first(evlist))
1548                 return;
1549
1550         evlist__for_each_entry_safe(evlist, n, evsel) {
1551                 if (evsel->leader == move_evsel->leader)
1552                         list_move_tail(&evsel->core.node, &move);
1553         }
1554
1555         list_splice(&move, &evlist->core.entries);
1556 }
1557
1558 void perf_evlist__set_tracking_event(struct evlist *evlist,
1559                                      struct evsel *tracking_evsel)
1560 {
1561         struct evsel *evsel;
1562
1563         if (tracking_evsel->tracking)
1564                 return;
1565
1566         evlist__for_each_entry(evlist, evsel) {
1567                 if (evsel != tracking_evsel)
1568                         evsel->tracking = false;
1569         }
1570
1571         tracking_evsel->tracking = true;
1572 }
1573
1574 struct evsel *
1575 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1576                                const char *str)
1577 {
1578         struct evsel *evsel;
1579
1580         evlist__for_each_entry(evlist, evsel) {
1581                 if (!evsel->name)
1582                         continue;
1583                 if (strcmp(str, evsel->name) == 0)
1584                         return evsel;
1585         }
1586
1587         return NULL;
1588 }
1589
1590 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1591                                   enum bkw_mmap_state state)
1592 {
1593         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1594         enum action {
1595                 NONE,
1596                 PAUSE,
1597                 RESUME,
1598         } action = NONE;
1599
1600         if (!evlist->overwrite_mmap)
1601                 return;
1602
1603         switch (old_state) {
1604         case BKW_MMAP_NOTREADY: {
1605                 if (state != BKW_MMAP_RUNNING)
1606                         goto state_err;
1607                 break;
1608         }
1609         case BKW_MMAP_RUNNING: {
1610                 if (state != BKW_MMAP_DATA_PENDING)
1611                         goto state_err;
1612                 action = PAUSE;
1613                 break;
1614         }
1615         case BKW_MMAP_DATA_PENDING: {
1616                 if (state != BKW_MMAP_EMPTY)
1617                         goto state_err;
1618                 break;
1619         }
1620         case BKW_MMAP_EMPTY: {
1621                 if (state != BKW_MMAP_RUNNING)
1622                         goto state_err;
1623                 action = RESUME;
1624                 break;
1625         }
1626         default:
1627                 WARN_ONCE(1, "Shouldn't get there\n");
1628         }
1629
1630         evlist->bkw_mmap_state = state;
1631
1632         switch (action) {
1633         case PAUSE:
1634                 perf_evlist__pause(evlist);
1635                 break;
1636         case RESUME:
1637                 perf_evlist__resume(evlist);
1638                 break;
1639         case NONE:
1640         default:
1641                 break;
1642         }
1643
1644 state_err:
1645         return;
1646 }
1647
1648 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1649 {
1650         struct evsel *evsel;
1651
1652         evlist__for_each_entry(evlist, evsel) {
1653                 if (!evsel->core.attr.exclude_kernel)
1654                         return false;
1655         }
1656
1657         return true;
1658 }
1659
1660 /*
1661  * Events in data file are not collect in groups, but we still want
1662  * the group display. Set the artificial group and set the leader's
1663  * forced_leader flag to notify the display code.
1664  */
1665 void perf_evlist__force_leader(struct evlist *evlist)
1666 {
1667         if (!evlist->nr_groups) {
1668                 struct evsel *leader = evlist__first(evlist);
1669
1670                 perf_evlist__set_leader(evlist);
1671                 leader->forced_leader = true;
1672         }
1673 }
1674
1675 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1676                                                  struct evsel *evsel,
1677                                                 bool close)
1678 {
1679         struct evsel *c2, *leader;
1680         bool is_open = true;
1681
1682         leader = evsel->leader;
1683         pr_debug("Weak group for %s/%d failed\n",
1684                         leader->name, leader->core.nr_members);
1685
1686         /*
1687          * for_each_group_member doesn't work here because it doesn't
1688          * include the first entry.
1689          */
1690         evlist__for_each_entry(evsel_list, c2) {
1691                 if (c2 == evsel)
1692                         is_open = false;
1693                 if (c2->leader == leader) {
1694                         if (is_open && close)
1695                                 perf_evsel__close(&c2->core);
1696                         c2->leader = c2;
1697                         c2->core.nr_members = 0;
1698                         /*
1699                          * Set this for all former members of the group
1700                          * to indicate they get reopened.
1701                          */
1702                         c2->reset_group = true;
1703                 }
1704         }
1705         return leader;
1706 }
1707
1708 int perf_evlist__add_sb_event(struct evlist *evlist,
1709                               struct perf_event_attr *attr,
1710                               perf_evsel__sb_cb_t cb,
1711                               void *data)
1712 {
1713         struct evsel *evsel;
1714
1715         if (!attr->sample_id_all) {
1716                 pr_warning("enabling sample_id_all for all side band events\n");
1717                 attr->sample_id_all = 1;
1718         }
1719
1720         evsel = perf_evsel__new_idx(attr, evlist->core.nr_entries);
1721         if (!evsel)
1722                 return -1;
1723
1724         evsel->side_band.cb = cb;
1725         evsel->side_band.data = data;
1726         evlist__add(evlist, evsel);
1727         return 0;
1728 }
1729
1730 static void *perf_evlist__poll_thread(void *arg)
1731 {
1732         struct evlist *evlist = arg;
1733         bool draining = false;
1734         int i, done = 0;
1735         /*
1736          * In order to read symbols from other namespaces perf to needs to call
1737          * setns(2).  This isn't permitted if the struct_fs has multiple users.
1738          * unshare(2) the fs so that we may continue to setns into namespaces
1739          * that we're observing when, for instance, reading the build-ids at
1740          * the end of a 'perf record' session.
1741          */
1742         unshare(CLONE_FS);
1743
1744         while (!done) {
1745                 bool got_data = false;
1746
1747                 if (evlist->thread.done)
1748                         draining = true;
1749
1750                 if (!draining)
1751                         evlist__poll(evlist, 1000);
1752
1753                 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1754                         struct mmap *map = &evlist->mmap[i];
1755                         union perf_event *event;
1756
1757                         if (perf_mmap__read_init(&map->core))
1758                                 continue;
1759                         while ((event = perf_mmap__read_event(&map->core)) != NULL) {
1760                                 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1761
1762                                 if (evsel && evsel->side_band.cb)
1763                                         evsel->side_band.cb(event, evsel->side_band.data);
1764                                 else
1765                                         pr_warning("cannot locate proper evsel for the side band event\n");
1766
1767                                 perf_mmap__consume(&map->core);
1768                                 got_data = true;
1769                         }
1770                         perf_mmap__read_done(&map->core);
1771                 }
1772
1773                 if (draining && !got_data)
1774                         break;
1775         }
1776         return NULL;
1777 }
1778
1779 int perf_evlist__start_sb_thread(struct evlist *evlist,
1780                                  struct target *target)
1781 {
1782         struct evsel *counter;
1783
1784         if (!evlist)
1785                 return 0;
1786
1787         if (perf_evlist__create_maps(evlist, target))
1788                 goto out_delete_evlist;
1789
1790         evlist__for_each_entry(evlist, counter) {
1791                 if (evsel__open(counter, evlist->core.cpus,
1792                                      evlist->core.threads) < 0)
1793                         goto out_delete_evlist;
1794         }
1795
1796         if (evlist__mmap(evlist, UINT_MAX))
1797                 goto out_delete_evlist;
1798
1799         evlist__for_each_entry(evlist, counter) {
1800                 if (evsel__enable(counter))
1801                         goto out_delete_evlist;
1802         }
1803
1804         evlist->thread.done = 0;
1805         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1806                 goto out_delete_evlist;
1807
1808         return 0;
1809
1810 out_delete_evlist:
1811         evlist__delete(evlist);
1812         evlist = NULL;
1813         return -1;
1814 }
1815
1816 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1817 {
1818         if (!evlist)
1819                 return;
1820         evlist->thread.done = 1;
1821         pthread_join(evlist->thread.th, NULL);
1822         evlist__delete(evlist);
1823 }