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 "../perf.h"
  22 #include "asm/bug.h"
  23 #include "bpf-event.h"
  24 #include "util/string2.h"
  25 #include <signal.h>
  26 #include <unistd.h>
  27 #include <sched.h>
  28 #include <stdlib.h>
  29
  30 #include "parse-events.h"
  31 #include <subcmd/parse-options.h>
  32
  33 #include <fcntl.h>
  34 #include <sys/ioctl.h>
  35 #include <sys/mman.h>
  36
  37 #include <linux/bitops.h>
  38 #include <linux/hash.h>
  39 #include <linux/log2.h>
  40 #include <linux/err.h>
  41 #include <linux/string.h>
  42 #include <linux/zalloc.h>
  43 #include <perf/evlist.h>
  44 #include <perf/evsel.h>
  45 #include <perf/cpumap.h>
  46 #include <perf/mmap.h>
  47
  48 #include <internal/xyarray.h>
  49
  50 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  51 int sigqueue(pid_t pid, int sig, const union sigval value);
  52 #endif
  53
  54 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  55 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  56
  57 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  58                   struct perf_thread_map *threads)
  59 {
  60         perf_evlist__init(&evlist->core);
  61         perf_evlist__set_maps(&evlist->core, cpus, threads);
  62         evlist->workload.pid = -1;
  63         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  64 }
  65
  66 struct evlist *evlist__new(void)
  67 {
  68         struct evlist *evlist = zalloc(sizeof(*evlist));
  69
  70         if (evlist != NULL)
  71                 evlist__init(evlist, NULL, NULL);
  72
  73         return evlist;
  74 }
  75
  76 struct evlist *perf_evlist__new_default(void)
  77 {
  78         struct evlist *evlist = evlist__new();
  79
  80         if (evlist && perf_evlist__add_default(evlist)) {
  81                 evlist__delete(evlist);
  82                 evlist = NULL;
  83         }
  84
  85         return evlist;
  86 }
  87
  88 struct evlist *perf_evlist__new_dummy(void)
  89 {
  90         struct evlist *evlist = evlist__new();
  91
  92         if (evlist && perf_evlist__add_dummy(evlist)) {
  93                 evlist__delete(evlist);
  94                 evlist = NULL;
  95         }
  96
  97         return evlist;
  98 }
  99
 100 /**
 101  * perf_evlist__set_id_pos - set the positions of event ids.
 102  * @evlist: selected event list
 103  *
 104  * Events with compatible sample types all have the same id_pos
 105  * and is_pos.  For convenience, put a copy on evlist.
 106  */
 107 void perf_evlist__set_id_pos(struct evlist *evlist)
 108 {
 109         struct evsel *first = evlist__first(evlist);
 110
 111         evlist->id_pos = first->id_pos;
 112         evlist->is_pos = first->is_pos;
 113 }
 114
 115 static void perf_evlist__update_id_pos(struct evlist *evlist)
 116 {
 117         struct evsel *evsel;
 118
 119         evlist__for_each_entry(evlist, evsel)
 120                 perf_evsel__calc_id_pos(evsel);
 121
 122         perf_evlist__set_id_pos(evlist);
 123 }
 124
 125 static void evlist__purge(struct evlist *evlist)
 126 {
 127         struct evsel *pos, *n;
 128
 129         evlist__for_each_entry_safe(evlist, n, pos) {
 130                 list_del_init(&pos->core.node);
 131                 pos->evlist = NULL;
 132                 evsel__delete(pos);
 133         }
 134
 135         evlist->core.nr_entries = 0;
 136 }
 137
 138 void evlist__exit(struct evlist *evlist)
 139 {
 140         zfree(&evlist->mmap);
 141         zfree(&evlist->overwrite_mmap);
 142         perf_evlist__exit(&evlist->core);
 143 }
 144
 145 void evlist__delete(struct evlist *evlist)
 146 {
 147         if (evlist == NULL)
 148                 return;
 149
 150         evlist__munmap(evlist);
 151         evlist__close(evlist);
 152         evlist__purge(evlist);
 153         evlist__exit(evlist);
 154         free(evlist);
 155 }
 156
 157 void evlist__add(struct evlist *evlist, struct evsel *entry)
 158 {
 159         entry->evlist = evlist;
 160         entry->idx = evlist->core.nr_entries;
 161         entry->tracking = !entry->idx;
 162
 163         perf_evlist__add(&evlist->core, &entry->core);
 164
 165         if (evlist->core.nr_entries == 1)
 166                 perf_evlist__set_id_pos(evlist);
 167 }
 168
 169 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 170 {
 171         evsel->evlist = NULL;
 172         perf_evlist__remove(&evlist->core, &evsel->core);
 173 }
 174
 175 void perf_evlist__splice_list_tail(struct evlist *evlist,
 176                                    struct list_head *list)
 177 {
 178         struct evsel *evsel, *temp;
 179
 180         __evlist__for_each_entry_safe(list, temp, evsel) {
 181                 list_del_init(&evsel->core.node);
 182                 evlist__add(evlist, evsel);
 183         }
 184 }
 185
 186 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 187                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 188 {
 189         struct evsel *evsel;
 190         size_t i;
 191         int err;
 192
 193         for (i = 0; i < nr_assocs; i++) {
 194                 // Adding a handler for an event not in this evlist, just ignore it.
 195                 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 196                 if (evsel == NULL)
 197                         continue;
 198
 199                 err = -EEXIST;
 200                 if (evsel->handler != NULL)
 201                         goto out;
 202                 evsel->handler = assocs[i].handler;
 203         }
 204
 205         err = 0;
 206 out:
 207         return err;
 208 }
 209
 210 void __perf_evlist__set_leader(struct list_head *list)
 211 {
 212         struct evsel *evsel, *leader;
 213
 214         leader = list_entry(list->next, struct evsel, core.node);
 215         evsel = list_entry(list->prev, struct evsel, core.node);
 216
 217         leader->core.nr_members = evsel->idx - leader->idx + 1;
 218
 219         __evlist__for_each_entry(list, evsel) {
 220                 evsel->leader = leader;
 221         }
 222 }
 223
 224 void perf_evlist__set_leader(struct evlist *evlist)
 225 {
 226         if (evlist->core.nr_entries) {
 227                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 228                 __perf_evlist__set_leader(&evlist->core.entries);
 229         }
 230 }
 231
 232 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
 233 {
 234         struct evsel *evsel = perf_evsel__new_cycles(precise);
 235
 236         if (evsel == NULL)
 237                 return -ENOMEM;
 238
 239         evlist__add(evlist, evsel);
 240         return 0;
 241 }
 242
 243 int perf_evlist__add_dummy(struct evlist *evlist)
 244 {
 245         struct perf_event_attr attr = {
 246                 .type   = PERF_TYPE_SOFTWARE,
 247                 .config = PERF_COUNT_SW_DUMMY,
 248                 .size   = sizeof(attr), /* to capture ABI version */
 249         };
 250         struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
 251
 252         if (evsel == NULL)
 253                 return -ENOMEM;
 254
 255         evlist__add(evlist, evsel);
 256         return 0;
 257 }
 258
 259 static int evlist__add_attrs(struct evlist *evlist,
 260                                   struct perf_event_attr *attrs, size_t nr_attrs)
 261 {
 262         struct evsel *evsel, *n;
 263         LIST_HEAD(head);
 264         size_t i;
 265
 266         for (i = 0; i < nr_attrs; i++) {
 267                 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 268                 if (evsel == NULL)
 269                         goto out_delete_partial_list;
 270                 list_add_tail(&evsel->core.node, &head);
 271         }
 272
 273         perf_evlist__splice_list_tail(evlist, &head);
 274
 275         return 0;
 276
 277 out_delete_partial_list:
 278         __evlist__for_each_entry_safe(&head, n, evsel)
 279                 evsel__delete(evsel);
 280         return -1;
 281 }
 282
 283 int __perf_evlist__add_default_attrs(struct evlist *evlist,
 284                                      struct perf_event_attr *attrs, size_t nr_attrs)
 285 {
 286         size_t i;
 287
 288         for (i = 0; i < nr_attrs; i++)
 289                 event_attr_init(attrs + i);
 290
 291         return evlist__add_attrs(evlist, attrs, nr_attrs);
 292 }
 293
 294 struct evsel *
 295 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 296 {
 297         struct evsel *evsel;
 298
 299         evlist__for_each_entry(evlist, evsel) {
 300                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 301                     (int)evsel->core.attr.config == id)
 302                         return evsel;
 303         }
 304
 305         return NULL;
 306 }
 307
 308 struct evsel *
 309 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
 310                                      const char *name)
 311 {
 312         struct evsel *evsel;
 313
 314         evlist__for_each_entry(evlist, evsel) {
 315                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 316                     (strcmp(evsel->name, name) == 0))
 317                         return evsel;
 318         }
 319
 320         return NULL;
 321 }
 322
 323 int perf_evlist__add_newtp(struct evlist *evlist,
 324                            const char *sys, const char *name, void *handler)
 325 {
 326         struct evsel *evsel = perf_evsel__newtp(sys, name);
 327
 328         if (IS_ERR(evsel))
 329                 return -1;
 330
 331         evsel->handler = handler;
 332         evlist__add(evlist, evsel);
 333         return 0;
 334 }
 335
 336 static int perf_evlist__nr_threads(struct evlist *evlist,
 337                                    struct evsel *evsel)
 338 {
 339         if (evsel->core.system_wide)
 340                 return 1;
 341         else
 342                 return perf_thread_map__nr(evlist->core.threads);
 343 }
 344
 345 void evlist__disable(struct evlist *evlist)
 346 {
 347         struct evsel *pos;
 348
 349         evlist__for_each_entry(evlist, pos) {
 350                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
 351                         continue;
 352                 evsel__disable(pos);
 353         }
 354
 355         evlist->enabled = false;
 356 }
 357
 358 void evlist__enable(struct evlist *evlist)
 359 {
 360         struct evsel *pos;
 361
 362         evlist__for_each_entry(evlist, pos) {
 363                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 364                         continue;
 365                 evsel__enable(pos);
 366         }
 367
 368         evlist->enabled = true;
 369 }
 370
 371 void perf_evlist__toggle_enable(struct evlist *evlist)
 372 {
 373         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 374 }
 375
 376 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
 377                                          struct evsel *evsel, int cpu)
 378 {
 379         int thread;
 380         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 381
 382         if (!evsel->core.fd)
 383                 return -EINVAL;
 384
 385         for (thread = 0; thread < nr_threads; thread++) {
 386                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 387                 if (err)
 388                         return err;
 389         }
 390         return 0;
 391 }
 392
 393 static int perf_evlist__enable_event_thread(struct evlist *evlist,
 394                                             struct evsel *evsel,
 395                                             int thread)
 396 {
 397         int cpu;
 398         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 399
 400         if (!evsel->core.fd)
 401                 return -EINVAL;
 402
 403         for (cpu = 0; cpu < nr_cpus; cpu++) {
 404                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 405                 if (err)
 406                         return err;
 407         }
 408         return 0;
 409 }
 410
 411 int perf_evlist__enable_event_idx(struct evlist *evlist,
 412                                   struct evsel *evsel, int idx)
 413 {
 414         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 415
 416         if (per_cpu_mmaps)
 417                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 418         else
 419                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 420 }
 421
 422 int evlist__add_pollfd(struct evlist *evlist, int fd)
 423 {
 424         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
 425 }
 426
 427 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 428 {
 429         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 430 }
 431
 432 int evlist__poll(struct evlist *evlist, int timeout)
 433 {
 434         return perf_evlist__poll(&evlist->core, timeout);
 435 }
 436
 437 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
 438 {
 439         struct hlist_head *head;
 440         struct perf_sample_id *sid;
 441         int hash;
 442
 443         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 444         head = &evlist->core.heads[hash];
 445
 446         hlist_for_each_entry(sid, head, node)
 447                 if (sid->id == id)
 448                         return sid;
 449
 450         return NULL;
 451 }
 452
 453 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
 454 {
 455         struct perf_sample_id *sid;
 456
 457         if (evlist->core.nr_entries == 1 || !id)
 458                 return evlist__first(evlist);
 459
 460         sid = perf_evlist__id2sid(evlist, id);
 461         if (sid)
 462                 return container_of(sid->evsel, struct evsel, core);
 463
 464         if (!perf_evlist__sample_id_all(evlist))
 465                 return evlist__first(evlist);
 466
 467         return NULL;
 468 }
 469
 470 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
 471                                                 u64 id)
 472 {
 473         struct perf_sample_id *sid;
 474
 475         if (!id)
 476                 return NULL;
 477
 478         sid = perf_evlist__id2sid(evlist, id);
 479         if (sid)
 480                 return container_of(sid->evsel, struct evsel, core);
 481
 482         return NULL;
 483 }
 484
 485 static int perf_evlist__event2id(struct evlist *evlist,
 486                                  union perf_event *event, u64 *id)
 487 {
 488         const __u64 *array = event->sample.array;
 489         ssize_t n;
 490
 491         n = (event->header.size - sizeof(event->header)) >> 3;
 492
 493         if (event->header.type == PERF_RECORD_SAMPLE) {
 494                 if (evlist->id_pos >= n)
 495                         return -1;
 496                 *id = array[evlist->id_pos];
 497         } else {
 498                 if (evlist->is_pos > n)
 499                         return -1;
 500                 n -= evlist->is_pos;
 501                 *id = array[n];
 502         }
 503         return 0;
 504 }
 505
 506 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
 507                                             union perf_event *event)
 508 {
 509         struct evsel *first = evlist__first(evlist);
 510         struct hlist_head *head;
 511         struct perf_sample_id *sid;
 512         int hash;
 513         u64 id;
 514
 515         if (evlist->core.nr_entries == 1)
 516                 return first;
 517
 518         if (!first->core.attr.sample_id_all &&
 519             event->header.type != PERF_RECORD_SAMPLE)
 520                 return first;
 521
 522         if (perf_evlist__event2id(evlist, event, &id))
 523                 return NULL;
 524
 525         /* Synthesized events have an id of zero */
 526         if (!id)
 527                 return first;
 528
 529         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 530         head = &evlist->core.heads[hash];
 531
 532         hlist_for_each_entry(sid, head, node) {
 533                 if (sid->id == id)
 534                         return container_of(sid->evsel, struct evsel, core);
 535         }
 536         return NULL;
 537 }
 538
 539 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
 540 {
 541         int i;
 542
 543         if (!evlist->overwrite_mmap)
 544                 return 0;
 545
 546         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 547                 int fd = evlist->overwrite_mmap[i].core.fd;
 548                 int err;
 549
 550                 if (fd < 0)
 551                         continue;
 552                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 553                 if (err)
 554                         return err;
 555         }
 556         return 0;
 557 }
 558
 559 static int perf_evlist__pause(struct evlist *evlist)
 560 {
 561         return perf_evlist__set_paused(evlist, true);
 562 }
 563
 564 static int perf_evlist__resume(struct evlist *evlist)
 565 {
 566         return perf_evlist__set_paused(evlist, false);
 567 }
 568
 569 static void evlist__munmap_nofree(struct evlist *evlist)
 570 {
 571         int i;
 572
 573         if (evlist->mmap)
 574                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 575                         perf_mmap__munmap(&evlist->mmap[i].core);
 576
 577         if (evlist->overwrite_mmap)
 578                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 579                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 580 }
 581
 582 void evlist__munmap(struct evlist *evlist)
 583 {
 584         evlist__munmap_nofree(evlist);
 585         zfree(&evlist->mmap);
 586         zfree(&evlist->overwrite_mmap);
 587 }
 588
 589 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 590 {
 591         struct mmap *m = container_of(map, struct mmap, core);
 592
 593         mmap__munmap(m);
 594 }
 595
 596 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 597                                        bool overwrite)
 598 {
 599         int i;
 600         struct mmap *map;
 601
 602         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 603         if (!map)
 604                 return NULL;
 605
 606         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 607                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 608
 609                 /*
 610                  * When the perf_mmap() call is made we grab one refcount, plus
 611                  * one extra to let perf_mmap__consume() get the last
 612                  * events after all real references (perf_mmap__get()) are
 613                  * dropped.
 614                  *
 615                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 616                  * thus does perf_mmap__get() on it.
 617                  */
 618                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 619         }
 620
 621         return map;
 622 }
 623
 624 static void
 625 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 626                          struct perf_mmap_param *_mp,
 627                          int idx, bool per_cpu)
 628 {
 629         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 630         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 631
 632         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
 633 }
 634
 635 static struct perf_mmap*
 636 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 637 {
 638         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 639         struct mmap *maps;
 640
 641         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 642
 643         if (!maps) {
 644                 maps = evlist__alloc_mmap(evlist, overwrite);
 645                 if (!maps)
 646                         return NULL;
 647
 648                 if (overwrite) {
 649                         evlist->overwrite_mmap = maps;
 650                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 651                                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 652                 } else {
 653                         evlist->mmap = maps;
 654                 }
 655         }
 656
 657         return &maps[idx].core;
 658 }
 659
 660 static int
 661 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 662                           int output, int cpu)
 663 {
 664         struct mmap *map = container_of(_map, struct mmap, core);
 665         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 666
 667         return mmap__mmap(map, mp, output, cpu);
 668 }
 669
 670 unsigned long perf_event_mlock_kb_in_pages(void)
 671 {
 672         unsigned long pages;
 673         int max;
 674
 675         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 676                 /*
 677                  * Pick a once upon a time good value, i.e. things look
 678                  * strange since we can't read a sysctl value, but lets not
 679                  * die yet...
 680                  */
 681                 max = 512;
 682         } else {
 683                 max -= (page_size / 1024);
 684         }
 685
 686         pages = (max * 1024) / page_size;
 687         if (!is_power_of_2(pages))
 688                 pages = rounddown_pow_of_two(pages);
 689
 690         return pages;
 691 }
 692
 693 size_t evlist__mmap_size(unsigned long pages)
 694 {
 695         if (pages == UINT_MAX)
 696                 pages = perf_event_mlock_kb_in_pages();
 697         else if (!is_power_of_2(pages))
 698                 return 0;
 699
 700         return (pages + 1) * page_size;
 701 }
 702
 703 static long parse_pages_arg(const char *str, unsigned long min,
 704                             unsigned long max)
 705 {
 706         unsigned long pages, val;
 707         static struct parse_tag tags[] = {
 708                 { .tag  = 'B', .mult = 1       },
 709                 { .tag  = 'K', .mult = 1 << 10 },
 710                 { .tag  = 'M', .mult = 1 << 20 },
 711                 { .tag  = 'G', .mult = 1 << 30 },
 712                 { .tag  = 0 },
 713         };
 714
 715         if (str == NULL)
 716                 return -EINVAL;
 717
 718         val = parse_tag_value(str, tags);
 719         if (val != (unsigned long) -1) {
 720                 /* we got file size value */
 721                 pages = PERF_ALIGN(val, page_size) / page_size;
 722         } else {
 723                 /* we got pages count value */
 724                 char *eptr;
 725                 pages = strtoul(str, &eptr, 10);
 726                 if (*eptr != '\0')
 727                         return -EINVAL;
 728         }
 729
 730         if (pages == 0 && min == 0) {
 731                 /* leave number of pages at 0 */
 732         } else if (!is_power_of_2(pages)) {
 733                 char buf[100];
 734
 735                 /* round pages up to next power of 2 */
 736                 pages = roundup_pow_of_two(pages);
 737                 if (!pages)
 738                         return -EINVAL;
 739
 740                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 741                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 742                         buf, pages);
 743         }
 744
 745         if (pages > max)
 746                 return -EINVAL;
 747
 748         return pages;
 749 }
 750
 751 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 752 {
 753         unsigned long max = UINT_MAX;
 754         long pages;
 755
 756         if (max > SIZE_MAX / page_size)
 757                 max = SIZE_MAX / page_size;
 758
 759         pages = parse_pages_arg(str, 1, max);
 760         if (pages < 0) {
 761                 pr_err("Invalid argument for --mmap_pages/-m\n");
 762                 return -1;
 763         }
 764
 765         *mmap_pages = pages;
 766         return 0;
 767 }
 768
 769 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 770                                   int unset __maybe_unused)
 771 {
 772         return __perf_evlist__parse_mmap_pages(opt->value, str);
 773 }
 774
 775 /**
 776  * evlist__mmap_ex - Create mmaps to receive events.
 777  * @evlist: list of events
 778  * @pages: map length in pages
 779  * @overwrite: overwrite older events?
 780  * @auxtrace_pages - auxtrace map length in pages
 781  * @auxtrace_overwrite - overwrite older auxtrace data?
 782  *
 783  * If @overwrite is %false the user needs to signal event consumption using
 784  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 785  * automatically.
 786  *
 787  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 788  * consumption using auxtrace_mmap__write_tail().
 789  *
 790  * Return: %0 on success, negative error code otherwise.
 791  */
 792 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 793                          unsigned int auxtrace_pages,
 794                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 795                          int comp_level)
 796 {
 797         /*
 798          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 799          * Its value is decided by evsel's write_backward.
 800          * So &mp should not be passed through const pointer.
 801          */
 802         struct mmap_params mp = {
 803                 .nr_cblocks     = nr_cblocks,
 804                 .affinity       = affinity,
 805                 .flush          = flush,
 806                 .comp_level     = comp_level
 807         };
 808         struct perf_evlist_mmap_ops ops = {
 809                 .idx  = perf_evlist__mmap_cb_idx,
 810                 .get  = perf_evlist__mmap_cb_get,
 811                 .mmap = perf_evlist__mmap_cb_mmap,
 812         };
 813
 814         evlist->core.mmap_len = evlist__mmap_size(pages);
 815         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 816
 817         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 818                                    auxtrace_pages, auxtrace_overwrite);
 819
 820         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 821 }
 822
 823 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 824 {
 825         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 826 }
 827
 828 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
 829 {
 830         bool all_threads = (target->per_thread && target->system_wide);
 831         struct perf_cpu_map *cpus;
 832         struct perf_thread_map *threads;
 833
 834         /*
 835          * If specify '-a' and '--per-thread' to perf record, perf record
 836          * will override '--per-thread'. target->per_thread = false and
 837          * target->system_wide = true.
 838          *
 839          * If specify '--per-thread' only to perf record,
 840          * target->per_thread = true and target->system_wide = false.
 841          *
 842          * So target->per_thread && target->system_wide is false.
 843          * For perf record, thread_map__new_str doesn't call
 844          * thread_map__new_all_cpus. That will keep perf record's
 845          * current behavior.
 846          *
 847          * For perf stat, it allows the case that target->per_thread and
 848          * target->system_wide are all true. It means to collect system-wide
 849          * per-thread data. thread_map__new_str will call
 850          * thread_map__new_all_cpus to enumerate all threads.
 851          */
 852         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 853                                       all_threads);
 854
 855         if (!threads)
 856                 return -1;
 857
 858         if (target__uses_dummy_map(target))
 859                 cpus = perf_cpu_map__dummy_new();
 860         else
 861                 cpus = perf_cpu_map__new(target->cpu_list);
 862
 863         if (!cpus)
 864                 goto out_delete_threads;
 865
 866         evlist->core.has_user_cpus = !!target->cpu_list;
 867
 868         perf_evlist__set_maps(&evlist->core, cpus, threads);
 869
 870         return 0;
 871
 872 out_delete_threads:
 873         perf_thread_map__put(threads);
 874         return -1;
 875 }
 876
 877 void __perf_evlist__set_sample_bit(struct evlist *evlist,
 878                                    enum perf_event_sample_format bit)
 879 {
 880         struct evsel *evsel;
 881
 882         evlist__for_each_entry(evlist, evsel)
 883                 __perf_evsel__set_sample_bit(evsel, bit);
 884 }
 885
 886 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
 887                                      enum perf_event_sample_format bit)
 888 {
 889         struct evsel *evsel;
 890
 891         evlist__for_each_entry(evlist, evsel)
 892                 __perf_evsel__reset_sample_bit(evsel, bit);
 893 }
 894
 895 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 896 {
 897         struct evsel *evsel;
 898         int err = 0;
 899
 900         evlist__for_each_entry(evlist, evsel) {
 901                 if (evsel->filter == NULL)
 902                         continue;
 903
 904                 /*
 905                  * filters only work for tracepoint event, which doesn't have cpu limit.
 906                  * So evlist and evsel should always be same.
 907                  */
 908                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
 909                 if (err) {
 910                         *err_evsel = evsel;
 911                         break;
 912                 }
 913         }
 914
 915         return err;
 916 }
 917
 918 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
 919 {
 920         struct evsel *evsel;
 921         int err = 0;
 922
 923         if (filter == NULL)
 924                 return -1;
 925
 926         evlist__for_each_entry(evlist, evsel) {
 927                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
 928                         continue;
 929
 930                 err = perf_evsel__set_filter(evsel, filter);
 931                 if (err)
 932                         break;
 933         }
 934
 935         return err;
 936 }
 937
 938 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
 939 {
 940         struct evsel *evsel;
 941         int err = 0;
 942
 943         if (filter == NULL)
 944                 return -1;
 945
 946         evlist__for_each_entry(evlist, evsel) {
 947                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
 948                         continue;
 949
 950                 err = perf_evsel__append_tp_filter(evsel, filter);
 951                 if (err)
 952                         break;
 953         }
 954
 955         return err;
 956 }
 957
 958 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
 959 {
 960         char *filter;
 961         size_t i;
 962
 963         for (i = 0; i < npids; ++i) {
 964                 if (i == 0) {
 965                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
 966                                 return NULL;
 967                 } else {
 968                         char *tmp;
 969
 970                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
 971                                 goto out_free;
 972
 973                         free(filter);
 974                         filter = tmp;
 975                 }
 976         }
 977
 978         return filter;
 979 out_free:
 980         free(filter);
 981         return NULL;
 982 }
 983
 984 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
 985 {
 986         char *filter = asprintf__tp_filter_pids(npids, pids);
 987         int ret = perf_evlist__set_tp_filter(evlist, filter);
 988
 989         free(filter);
 990         return ret;
 991 }
 992
 993 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
 994 {
 995         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
 996 }
 997
 998 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
 999 {
1000         char *filter = asprintf__tp_filter_pids(npids, pids);
1001         int ret = perf_evlist__append_tp_filter(evlist, filter);
1002
1003         free(filter);
1004         return ret;
1005 }
1006
1007 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1008 {
1009         return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1010 }
1011
1012 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1013 {
1014         struct evsel *pos;
1015
1016         if (evlist->core.nr_entries == 1)
1017                 return true;
1018
1019         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1020                 return false;
1021
1022         evlist__for_each_entry(evlist, pos) {
1023                 if (pos->id_pos != evlist->id_pos ||
1024                     pos->is_pos != evlist->is_pos)
1025                         return false;
1026         }
1027
1028         return true;
1029 }
1030
1031 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1032 {
1033         struct evsel *evsel;
1034
1035         if (evlist->combined_sample_type)
1036                 return evlist->combined_sample_type;
1037
1038         evlist__for_each_entry(evlist, evsel)
1039                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1040
1041         return evlist->combined_sample_type;
1042 }
1043
1044 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1045 {
1046         evlist->combined_sample_type = 0;
1047         return __perf_evlist__combined_sample_type(evlist);
1048 }
1049
1050 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1051 {
1052         struct evsel *evsel;
1053         u64 branch_type = 0;
1054
1055         evlist__for_each_entry(evlist, evsel)
1056                 branch_type |= evsel->core.attr.branch_sample_type;
1057         return branch_type;
1058 }
1059
1060 bool perf_evlist__valid_read_format(struct evlist *evlist)
1061 {
1062         struct evsel *first = evlist__first(evlist), *pos = first;
1063         u64 read_format = first->core.attr.read_format;
1064         u64 sample_type = first->core.attr.sample_type;
1065
1066         evlist__for_each_entry(evlist, pos) {
1067                 if (read_format != pos->core.attr.read_format)
1068                         return false;
1069         }
1070
1071         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1072         if ((sample_type & PERF_SAMPLE_READ) &&
1073             !(read_format & PERF_FORMAT_ID)) {
1074                 return false;
1075         }
1076
1077         return true;
1078 }
1079
1080 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1081 {
1082         struct evsel *first = evlist__first(evlist);
1083         struct perf_sample *data;
1084         u64 sample_type;
1085         u16 size = 0;
1086
1087         if (!first->core.attr.sample_id_all)
1088                 goto out;
1089
1090         sample_type = first->core.attr.sample_type;
1091
1092         if (sample_type & PERF_SAMPLE_TID)
1093                 size += sizeof(data->tid) * 2;
1094
1095        if (sample_type & PERF_SAMPLE_TIME)
1096                 size += sizeof(data->time);
1097
1098         if (sample_type & PERF_SAMPLE_ID)
1099                 size += sizeof(data->id);
1100
1101         if (sample_type & PERF_SAMPLE_STREAM_ID)
1102                 size += sizeof(data->stream_id);
1103
1104         if (sample_type & PERF_SAMPLE_CPU)
1105                 size += sizeof(data->cpu) * 2;
1106
1107         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1108                 size += sizeof(data->id);
1109 out:
1110         return size;
1111 }
1112
1113 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1114 {
1115         struct evsel *first = evlist__first(evlist), *pos = first;
1116
1117         evlist__for_each_entry_continue(evlist, pos) {
1118                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1119                         return false;
1120         }
1121
1122         return true;
1123 }
1124
1125 bool perf_evlist__sample_id_all(struct evlist *evlist)
1126 {
1127         struct evsel *first = evlist__first(evlist);
1128         return first->core.attr.sample_id_all;
1129 }
1130
1131 void perf_evlist__set_selected(struct evlist *evlist,
1132                                struct evsel *evsel)
1133 {
1134         evlist->selected = evsel;
1135 }
1136
1137 void evlist__close(struct evlist *evlist)
1138 {
1139         struct evsel *evsel;
1140
1141         evlist__for_each_entry_reverse(evlist, evsel)
1142                 evsel__close(evsel);
1143 }
1144
1145 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1146 {
1147         struct perf_cpu_map *cpus;
1148         struct perf_thread_map *threads;
1149         int err = -ENOMEM;
1150
1151         /*
1152          * Try reading /sys/devices/system/cpu/online to get
1153          * an all cpus map.
1154          *
1155          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1156          * code needs an overhaul to properly forward the
1157          * error, and we may not want to do that fallback to a
1158          * default cpu identity map :-\
1159          */
1160         cpus = perf_cpu_map__new(NULL);
1161         if (!cpus)
1162                 goto out;
1163
1164         threads = perf_thread_map__new_dummy();
1165         if (!threads)
1166                 goto out_put;
1167
1168         perf_evlist__set_maps(&evlist->core, cpus, threads);
1169 out:
1170         return err;
1171 out_put:
1172         perf_cpu_map__put(cpus);
1173         goto out;
1174 }
1175
1176 int evlist__open(struct evlist *evlist)
1177 {
1178         struct evsel *evsel;
1179         int err;
1180
1181         /*
1182          * Default: one fd per CPU, all threads, aka systemwide
1183          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1184          */
1185         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1186                 err = perf_evlist__create_syswide_maps(evlist);
1187                 if (err < 0)
1188                         goto out_err;
1189         }
1190
1191         perf_evlist__update_id_pos(evlist);
1192
1193         evlist__for_each_entry(evlist, evsel) {
1194                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1195                 if (err < 0)
1196                         goto out_err;
1197         }
1198
1199         return 0;
1200 out_err:
1201         evlist__close(evlist);
1202         errno = -err;
1203         return err;
1204 }
1205
1206 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1207                                   const char *argv[], bool pipe_output,
1208                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1209 {
1210         int child_ready_pipe[2], go_pipe[2];
1211         char bf;
1212
1213         if (pipe(child_ready_pipe) < 0) {
1214                 perror("failed to create 'ready' pipe");
1215                 return -1;
1216         }
1217
1218         if (pipe(go_pipe) < 0) {
1219                 perror("failed to create 'go' pipe");
1220                 goto out_close_ready_pipe;
1221         }
1222
1223         evlist->workload.pid = fork();
1224         if (evlist->workload.pid < 0) {
1225                 perror("failed to fork");
1226                 goto out_close_pipes;
1227         }
1228
1229         if (!evlist->workload.pid) {
1230                 int ret;
1231
1232                 if (pipe_output)
1233                         dup2(2, 1);
1234
1235                 signal(SIGTERM, SIG_DFL);
1236
1237                 close(child_ready_pipe[0]);
1238                 close(go_pipe[1]);
1239                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1240
1241                 /*
1242                  * Tell the parent we're ready to go
1243                  */
1244                 close(child_ready_pipe[1]);
1245
1246                 /*
1247                  * Wait until the parent tells us to go.
1248                  */
1249                 ret = read(go_pipe[0], &bf, 1);
1250                 /*
1251                  * The parent will ask for the execvp() to be performed by
1252                  * writing exactly one byte, in workload.cork_fd, usually via
1253                  * perf_evlist__start_workload().
1254                  *
1255                  * For cancelling the workload without actually running it,
1256                  * the parent will just close workload.cork_fd, without writing
1257                  * anything, i.e. read will return zero and we just exit()
1258                  * here.
1259                  */
1260                 if (ret != 1) {
1261                         if (ret == -1)
1262                                 perror("unable to read pipe");
1263                         exit(ret);
1264                 }
1265
1266                 execvp(argv[0], (char **)argv);
1267
1268                 if (exec_error) {
1269                         union sigval val;
1270
1271                         val.sival_int = errno;
1272                         if (sigqueue(getppid(), SIGUSR1, val))
1273                                 perror(argv[0]);
1274                 } else
1275                         perror(argv[0]);
1276                 exit(-1);
1277         }
1278
1279         if (exec_error) {
1280                 struct sigaction act = {
1281                         .sa_flags     = SA_SIGINFO,
1282                         .sa_sigaction = exec_error,
1283                 };
1284                 sigaction(SIGUSR1, &act, NULL);
1285         }
1286
1287         if (target__none(target)) {
1288                 if (evlist->core.threads == NULL) {
1289                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1290                                 __func__, __LINE__);
1291                         goto out_close_pipes;
1292                 }
1293                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1294         }
1295
1296         close(child_ready_pipe[1]);
1297         close(go_pipe[0]);
1298         /*
1299          * wait for child to settle
1300          */
1301         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1302                 perror("unable to read pipe");
1303                 goto out_close_pipes;
1304         }
1305
1306         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1307         evlist->workload.cork_fd = go_pipe[1];
1308         close(child_ready_pipe[0]);
1309         return 0;
1310
1311 out_close_pipes:
1312         close(go_pipe[0]);
1313         close(go_pipe[1]);
1314 out_close_ready_pipe:
1315         close(child_ready_pipe[0]);
1316         close(child_ready_pipe[1]);
1317         return -1;
1318 }
1319
1320 int perf_evlist__start_workload(struct evlist *evlist)
1321 {
1322         if (evlist->workload.cork_fd > 0) {
1323                 char bf = 0;
1324                 int ret;
1325                 /*
1326                  * Remove the cork, let it rip!
1327                  */
1328                 ret = write(evlist->workload.cork_fd, &bf, 1);
1329                 if (ret < 0)
1330                         perror("unable to write to pipe");
1331
1332                 close(evlist->workload.cork_fd);
1333                 return ret;
1334         }
1335
1336         return 0;
1337 }
1338
1339 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1340                               struct perf_sample *sample)
1341 {
1342         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1343
1344         if (!evsel)
1345                 return -EFAULT;
1346         return perf_evsel__parse_sample(evsel, event, sample);
1347 }
1348
1349 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1350                                         union perf_event *event,
1351                                         u64 *timestamp)
1352 {
1353         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1354
1355         if (!evsel)
1356                 return -EFAULT;
1357         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1358 }
1359
1360 int perf_evlist__strerror_open(struct evlist *evlist,
1361                                int err, char *buf, size_t size)
1362 {
1363         int printed, value;
1364         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1365
1366         switch (err) {
1367         case EACCES:
1368         case EPERM:
1369                 printed = scnprintf(buf, size,
1370                                     "Error:\t%s.\n"
1371                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1372
1373                 value = perf_event_paranoid();
1374
1375                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1376
1377                 if (value >= 2) {
1378                         printed += scnprintf(buf + printed, size - printed,
1379                                              "For your workloads it needs to be <= 1\nHint:\t");
1380                 }
1381                 printed += scnprintf(buf + printed, size - printed,
1382                                      "For system wide tracing it needs to be set to -1.\n");
1383
1384                 printed += scnprintf(buf + printed, size - printed,
1385                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1386                                     "Hint:\tThe current value is %d.", value);
1387                 break;
1388         case EINVAL: {
1389                 struct evsel *first = evlist__first(evlist);
1390                 int max_freq;
1391
1392                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1393                         goto out_default;
1394
1395                 if (first->core.attr.sample_freq < (u64)max_freq)
1396                         goto out_default;
1397
1398                 printed = scnprintf(buf, size,
1399                                     "Error:\t%s.\n"
1400                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1401                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1402                                     emsg, max_freq, first->core.attr.sample_freq);
1403                 break;
1404         }
1405         default:
1406 out_default:
1407                 scnprintf(buf, size, "%s", emsg);
1408                 break;
1409         }
1410
1411         return 0;
1412 }
1413
1414 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1415 {
1416         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1417         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1418
1419         switch (err) {
1420         case EPERM:
1421                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1422                 printed += scnprintf(buf + printed, size - printed,
1423                                      "Error:\t%s.\n"
1424                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1425                                      "Hint:\tTried using %zd kB.\n",
1426                                      emsg, pages_max_per_user, pages_attempted);
1427
1428                 if (pages_attempted >= pages_max_per_user) {
1429                         printed += scnprintf(buf + printed, size - printed,
1430                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1431                                              pages_max_per_user + pages_attempted);
1432                 }
1433
1434                 printed += scnprintf(buf + printed, size - printed,
1435                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1436                 break;
1437         default:
1438                 scnprintf(buf, size, "%s", emsg);
1439                 break;
1440         }
1441
1442         return 0;
1443 }
1444
1445 void perf_evlist__to_front(struct evlist *evlist,
1446                            struct evsel *move_evsel)
1447 {
1448         struct evsel *evsel, *n;
1449         LIST_HEAD(move);
1450
1451         if (move_evsel == evlist__first(evlist))
1452                 return;
1453
1454         evlist__for_each_entry_safe(evlist, n, evsel) {
1455                 if (evsel->leader == move_evsel->leader)
1456                         list_move_tail(&evsel->core.node, &move);
1457         }
1458
1459         list_splice(&move, &evlist->core.entries);
1460 }
1461
1462 void perf_evlist__set_tracking_event(struct evlist *evlist,
1463                                      struct evsel *tracking_evsel)
1464 {
1465         struct evsel *evsel;
1466
1467         if (tracking_evsel->tracking)
1468                 return;
1469
1470         evlist__for_each_entry(evlist, evsel) {
1471                 if (evsel != tracking_evsel)
1472                         evsel->tracking = false;
1473         }
1474
1475         tracking_evsel->tracking = true;
1476 }
1477
1478 struct evsel *
1479 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1480                                const char *str)
1481 {
1482         struct evsel *evsel;
1483
1484         evlist__for_each_entry(evlist, evsel) {
1485                 if (!evsel->name)
1486                         continue;
1487                 if (strcmp(str, evsel->name) == 0)
1488                         return evsel;
1489         }
1490
1491         return NULL;
1492 }
1493
1494 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1495                                   enum bkw_mmap_state state)
1496 {
1497         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1498         enum action {
1499                 NONE,
1500                 PAUSE,
1501                 RESUME,
1502         } action = NONE;
1503
1504         if (!evlist->overwrite_mmap)
1505                 return;
1506
1507         switch (old_state) {
1508         case BKW_MMAP_NOTREADY: {
1509                 if (state != BKW_MMAP_RUNNING)
1510                         goto state_err;
1511                 break;
1512         }
1513         case BKW_MMAP_RUNNING: {
1514                 if (state != BKW_MMAP_DATA_PENDING)
1515                         goto state_err;
1516                 action = PAUSE;
1517                 break;
1518         }
1519         case BKW_MMAP_DATA_PENDING: {
1520                 if (state != BKW_MMAP_EMPTY)
1521                         goto state_err;
1522                 break;
1523         }
1524         case BKW_MMAP_EMPTY: {
1525                 if (state != BKW_MMAP_RUNNING)
1526                         goto state_err;
1527                 action = RESUME;
1528                 break;
1529         }
1530         default:
1531                 WARN_ONCE(1, "Shouldn't get there\n");
1532         }
1533
1534         evlist->bkw_mmap_state = state;
1535
1536         switch (action) {
1537         case PAUSE:
1538                 perf_evlist__pause(evlist);
1539                 break;
1540         case RESUME:
1541                 perf_evlist__resume(evlist);
1542                 break;
1543         case NONE:
1544         default:
1545                 break;
1546         }
1547
1548 state_err:
1549         return;
1550 }
1551
1552 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1553 {
1554         struct evsel *evsel;
1555
1556         evlist__for_each_entry(evlist, evsel) {
1557                 if (!evsel->core.attr.exclude_kernel)
1558                         return false;
1559         }
1560
1561         return true;
1562 }
1563
1564 /*
1565  * Events in data file are not collect in groups, but we still want
1566  * the group display. Set the artificial group and set the leader's
1567  * forced_leader flag to notify the display code.
1568  */
1569 void perf_evlist__force_leader(struct evlist *evlist)
1570 {
1571         if (!evlist->nr_groups) {
1572                 struct evsel *leader = evlist__first(evlist);
1573
1574                 perf_evlist__set_leader(evlist);
1575                 leader->forced_leader = true;
1576         }
1577 }
1578
1579 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1580                                                  struct evsel *evsel)
1581 {
1582         struct evsel *c2, *leader;
1583         bool is_open = true;
1584
1585         leader = evsel->leader;
1586         pr_debug("Weak group for %s/%d failed\n",
1587                         leader->name, leader->core.nr_members);
1588
1589         /*
1590          * for_each_group_member doesn't work here because it doesn't
1591          * include the first entry.
1592          */
1593         evlist__for_each_entry(evsel_list, c2) {
1594                 if (c2 == evsel)
1595                         is_open = false;
1596                 if (c2->leader == leader) {
1597                         if (is_open)
1598                                 perf_evsel__close(&c2->core);
1599                         c2->leader = c2;
1600                         c2->core.nr_members = 0;
1601                 }
1602         }
1603         return leader;
1604 }
1605
1606 int perf_evlist__add_sb_event(struct evlist **evlist,
1607                               struct perf_event_attr *attr,
1608                               perf_evsel__sb_cb_t cb,
1609                               void *data)
1610 {
1611         struct evsel *evsel;
1612         bool new_evlist = (*evlist) == NULL;
1613
1614         if (*evlist == NULL)
1615                 *evlist = evlist__new();
1616         if (*evlist == NULL)
1617                 return -1;
1618
1619         if (!attr->sample_id_all) {
1620                 pr_warning("enabling sample_id_all for all side band events\n");
1621                 attr->sample_id_all = 1;
1622         }
1623
1624         evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1625         if (!evsel)
1626                 goto out_err;
1627
1628         evsel->side_band.cb = cb;
1629         evsel->side_band.data = data;
1630         evlist__add(*evlist, evsel);
1631         return 0;
1632
1633 out_err:
1634         if (new_evlist) {
1635                 evlist__delete(*evlist);
1636                 *evlist = NULL;
1637         }
1638         return -1;
1639 }
1640
1641 static void *perf_evlist__poll_thread(void *arg)
1642 {
1643         struct evlist *evlist = arg;
1644         bool draining = false;
1645         int i, done = 0;
1646         /*
1647          * In order to read symbols from other namespaces perf to needs to call
1648          * setns(2).  This isn't permitted if the struct_fs has multiple users.
1649          * unshare(2) the fs so that we may continue to setns into namespaces
1650          * that we're observing when, for instance, reading the build-ids at
1651          * the end of a 'perf record' session.
1652          */
1653         unshare(CLONE_FS);
1654
1655         while (!done) {
1656                 bool got_data = false;
1657
1658                 if (evlist->thread.done)
1659                         draining = true;
1660
1661                 if (!draining)
1662                         evlist__poll(evlist, 1000);
1663
1664                 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1665                         struct mmap *map = &evlist->mmap[i];
1666                         union perf_event *event;
1667
1668                         if (perf_mmap__read_init(&map->core))
1669                                 continue;
1670                         while ((event = perf_mmap__read_event(&map->core)) != NULL) {
1671                                 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1672
1673                                 if (evsel && evsel->side_band.cb)
1674                                         evsel->side_band.cb(event, evsel->side_band.data);
1675                                 else
1676                                         pr_warning("cannot locate proper evsel for the side band event\n");
1677
1678                                 perf_mmap__consume(&map->core);
1679                                 got_data = true;
1680                         }
1681                         perf_mmap__read_done(&map->core);
1682                 }
1683
1684                 if (draining && !got_data)
1685                         break;
1686         }
1687         return NULL;
1688 }
1689
1690 int perf_evlist__start_sb_thread(struct evlist *evlist,
1691                                  struct target *target)
1692 {
1693         struct evsel *counter;
1694
1695         if (!evlist)
1696                 return 0;
1697
1698         if (perf_evlist__create_maps(evlist, target))
1699                 goto out_delete_evlist;
1700
1701         evlist__for_each_entry(evlist, counter) {
1702                 if (evsel__open(counter, evlist->core.cpus,
1703                                      evlist->core.threads) < 0)
1704                         goto out_delete_evlist;
1705         }
1706
1707         if (evlist__mmap(evlist, UINT_MAX))
1708                 goto out_delete_evlist;
1709
1710         evlist__for_each_entry(evlist, counter) {
1711                 if (evsel__enable(counter))
1712                         goto out_delete_evlist;
1713         }
1714
1715         evlist->thread.done = 0;
1716         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1717                 goto out_delete_evlist;
1718
1719         return 0;
1720
1721 out_delete_evlist:
1722         evlist__delete(evlist);
1723         evlist = NULL;
1724         return -1;
1725 }
1726
1727 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1728 {
1729         if (!evlist)
1730                 return;
1731         evlist->thread.done = 1;
1732         pthread_join(evlist->thread.th, NULL);
1733         evlist__delete(evlist);
1734 }