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 "thread_map.h"
  14 #include "target.h"
  15 #include "evlist.h"
  16 #include "evsel.h"
  17 #include "debug.h"
  18 #include "units.h"
  19 #include "asm/bug.h"
  20 #include "bpf-event.h"
  21 #include <signal.h>
  22 #include <unistd.h>
  23
  24 #include "parse-events.h"
  25 #include <subcmd/parse-options.h>
  26
  27 #include <fcntl.h>
  28 #include <sys/ioctl.h>
  29 #include <sys/mman.h>
  30
  31 #include <linux/bitops.h>
  32 #include <linux/hash.h>
  33 #include <linux/log2.h>
  34 #include <linux/err.h>
  35 #include <linux/zalloc.h>
  36
  37 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  38 int sigqueue(pid_t pid, int sig, const union sigval value);
  39 #endif
  40
  41 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  42 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  43
  44 void perf_evlist__init(struct perf_evlist *evlist, struct perf_cpu_map *cpus,
  45                        struct thread_map *threads)
  46 {
  47         int i;
  48
  49         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  50                 INIT_HLIST_HEAD(&evlist->heads[i]);
  51         INIT_LIST_HEAD(&evlist->entries);
  52         perf_evlist__set_maps(evlist, cpus, threads);
  53         fdarray__init(&evlist->pollfd, 64);
  54         evlist->workload.pid = -1;
  55         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  56 }
  57
  58 struct perf_evlist *perf_evlist__new(void)
  59 {
  60         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  61
  62         if (evlist != NULL)
  63                 perf_evlist__init(evlist, NULL, NULL);
  64
  65         return evlist;
  66 }
  67
  68 struct perf_evlist *perf_evlist__new_default(void)
  69 {
  70         struct perf_evlist *evlist = perf_evlist__new();
  71
  72         if (evlist && perf_evlist__add_default(evlist)) {
  73                 perf_evlist__delete(evlist);
  74                 evlist = NULL;
  75         }
  76
  77         return evlist;
  78 }
  79
  80 struct perf_evlist *perf_evlist__new_dummy(void)
  81 {
  82         struct perf_evlist *evlist = perf_evlist__new();
  83
  84         if (evlist && perf_evlist__add_dummy(evlist)) {
  85                 perf_evlist__delete(evlist);
  86                 evlist = NULL;
  87         }
  88
  89         return evlist;
  90 }
  91
  92 /**
  93  * perf_evlist__set_id_pos - set the positions of event ids.
  94  * @evlist: selected event list
  95  *
  96  * Events with compatible sample types all have the same id_pos
  97  * and is_pos.  For convenience, put a copy on evlist.
  98  */
  99 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
 100 {
 101         struct perf_evsel *first = perf_evlist__first(evlist);
 102
 103         evlist->id_pos = first->id_pos;
 104         evlist->is_pos = first->is_pos;
 105 }
 106
 107 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 108 {
 109         struct perf_evsel *evsel;
 110
 111         evlist__for_each_entry(evlist, evsel)
 112                 perf_evsel__calc_id_pos(evsel);
 113
 114         perf_evlist__set_id_pos(evlist);
 115 }
 116
 117 static void perf_evlist__purge(struct perf_evlist *evlist)
 118 {
 119         struct perf_evsel *pos, *n;
 120
 121         evlist__for_each_entry_safe(evlist, n, pos) {
 122                 list_del_init(&pos->node);
 123                 pos->evlist = NULL;
 124                 perf_evsel__delete(pos);
 125         }
 126
 127         evlist->nr_entries = 0;
 128 }
 129
 130 void perf_evlist__exit(struct perf_evlist *evlist)
 131 {
 132         zfree(&evlist->mmap);
 133         zfree(&evlist->overwrite_mmap);
 134         fdarray__exit(&evlist->pollfd);
 135 }
 136
 137 void perf_evlist__delete(struct perf_evlist *evlist)
 138 {
 139         if (evlist == NULL)
 140                 return;
 141
 142         perf_evlist__munmap(evlist);
 143         perf_evlist__close(evlist);
 144         cpu_map__put(evlist->cpus);
 145         thread_map__put(evlist->threads);
 146         evlist->cpus = NULL;
 147         evlist->threads = NULL;
 148         perf_evlist__purge(evlist);
 149         perf_evlist__exit(evlist);
 150         free(evlist);
 151 }
 152
 153 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 154                                           struct perf_evsel *evsel)
 155 {
 156         /*
 157          * We already have cpus for evsel (via PMU sysfs) so
 158          * keep it, if there's no target cpu list defined.
 159          */
 160         if (!evsel->own_cpus || evlist->has_user_cpus) {
 161                 cpu_map__put(evsel->cpus);
 162                 evsel->cpus = cpu_map__get(evlist->cpus);
 163         } else if (evsel->cpus != evsel->own_cpus) {
 164                 cpu_map__put(evsel->cpus);
 165                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 166         }
 167
 168         thread_map__put(evsel->threads);
 169         evsel->threads = thread_map__get(evlist->threads);
 170 }
 171
 172 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 173 {
 174         struct perf_evsel *evsel;
 175
 176         evlist__for_each_entry(evlist, evsel)
 177                 __perf_evlist__propagate_maps(evlist, evsel);
 178 }
 179
 180 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 181 {
 182         entry->evlist = evlist;
 183         list_add_tail(&entry->node, &evlist->entries);
 184         entry->idx = evlist->nr_entries;
 185         entry->tracking = !entry->idx;
 186
 187         if (!evlist->nr_entries++)
 188                 perf_evlist__set_id_pos(evlist);
 189
 190         __perf_evlist__propagate_maps(evlist, entry);
 191 }
 192
 193 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 194 {
 195         evsel->evlist = NULL;
 196         list_del_init(&evsel->node);
 197         evlist->nr_entries -= 1;
 198 }
 199
 200 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 201                                    struct list_head *list)
 202 {
 203         struct perf_evsel *evsel, *temp;
 204
 205         __evlist__for_each_entry_safe(list, temp, evsel) {
 206                 list_del_init(&evsel->node);
 207                 perf_evlist__add(evlist, evsel);
 208         }
 209 }
 210
 211 void __perf_evlist__set_leader(struct list_head *list)
 212 {
 213         struct perf_evsel *evsel, *leader;
 214
 215         leader = list_entry(list->next, struct perf_evsel, node);
 216         evsel = list_entry(list->prev, struct perf_evsel, node);
 217
 218         leader->nr_members = evsel->idx - leader->idx + 1;
 219
 220         __evlist__for_each_entry(list, evsel) {
 221                 evsel->leader = leader;
 222         }
 223 }
 224
 225 void perf_evlist__set_leader(struct perf_evlist *evlist)
 226 {
 227         if (evlist->nr_entries) {
 228                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 229                 __perf_evlist__set_leader(&evlist->entries);
 230         }
 231 }
 232
 233 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 234 {
 235         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 236
 237         if (evsel == NULL)
 238                 return -ENOMEM;
 239
 240         perf_evlist__add(evlist, evsel);
 241         return 0;
 242 }
 243
 244 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 245 {
 246         struct perf_event_attr attr = {
 247                 .type   = PERF_TYPE_SOFTWARE,
 248                 .config = PERF_COUNT_SW_DUMMY,
 249                 .size   = sizeof(attr), /* to capture ABI version */
 250         };
 251         struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
 252
 253         if (evsel == NULL)
 254                 return -ENOMEM;
 255
 256         perf_evlist__add(evlist, evsel);
 257         return 0;
 258 }
 259
 260 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 261                                   struct perf_event_attr *attrs, size_t nr_attrs)
 262 {
 263         struct perf_evsel *evsel, *n;
 264         LIST_HEAD(head);
 265         size_t i;
 266
 267         for (i = 0; i < nr_attrs; i++) {
 268                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 269                 if (evsel == NULL)
 270                         goto out_delete_partial_list;
 271                 list_add_tail(&evsel->node, &head);
 272         }
 273
 274         perf_evlist__splice_list_tail(evlist, &head);
 275
 276         return 0;
 277
 278 out_delete_partial_list:
 279         __evlist__for_each_entry_safe(&head, n, evsel)
 280                 perf_evsel__delete(evsel);
 281         return -1;
 282 }
 283
 284 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 285                                      struct perf_event_attr *attrs, size_t nr_attrs)
 286 {
 287         size_t i;
 288
 289         for (i = 0; i < nr_attrs; i++)
 290                 event_attr_init(attrs + i);
 291
 292         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 293 }
 294
 295 struct perf_evsel *
 296 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 297 {
 298         struct perf_evsel *evsel;
 299
 300         evlist__for_each_entry(evlist, evsel) {
 301                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 302                     (int)evsel->attr.config == id)
 303                         return evsel;
 304         }
 305
 306         return NULL;
 307 }
 308
 309 struct perf_evsel *
 310 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 311                                      const char *name)
 312 {
 313         struct perf_evsel *evsel;
 314
 315         evlist__for_each_entry(evlist, evsel) {
 316                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 317                     (strcmp(evsel->name, name) == 0))
 318                         return evsel;
 319         }
 320
 321         return NULL;
 322 }
 323
 324 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 325                            const char *sys, const char *name, void *handler)
 326 {
 327         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 328
 329         if (IS_ERR(evsel))
 330                 return -1;
 331
 332         evsel->handler = handler;
 333         perf_evlist__add(evlist, evsel);
 334         return 0;
 335 }
 336
 337 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 338                                    struct perf_evsel *evsel)
 339 {
 340         if (evsel->system_wide)
 341                 return 1;
 342         else
 343                 return thread_map__nr(evlist->threads);
 344 }
 345
 346 void perf_evlist__disable(struct perf_evlist *evlist)
 347 {
 348         struct perf_evsel *pos;
 349
 350         evlist__for_each_entry(evlist, pos) {
 351                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
 352                         continue;
 353                 perf_evsel__disable(pos);
 354         }
 355
 356         evlist->enabled = false;
 357 }
 358
 359 void perf_evlist__enable(struct perf_evlist *evlist)
 360 {
 361         struct perf_evsel *pos;
 362
 363         evlist__for_each_entry(evlist, pos) {
 364                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 365                         continue;
 366                 perf_evsel__enable(pos);
 367         }
 368
 369         evlist->enabled = true;
 370 }
 371
 372 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 373 {
 374         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 375 }
 376
 377 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 378                                          struct perf_evsel *evsel, int cpu)
 379 {
 380         int thread;
 381         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 382
 383         if (!evsel->fd)
 384                 return -EINVAL;
 385
 386         for (thread = 0; thread < nr_threads; thread++) {
 387                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 388                 if (err)
 389                         return err;
 390         }
 391         return 0;
 392 }
 393
 394 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 395                                             struct perf_evsel *evsel,
 396                                             int thread)
 397 {
 398         int cpu;
 399         int nr_cpus = cpu_map__nr(evlist->cpus);
 400
 401         if (!evsel->fd)
 402                 return -EINVAL;
 403
 404         for (cpu = 0; cpu < nr_cpus; cpu++) {
 405                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 406                 if (err)
 407                         return err;
 408         }
 409         return 0;
 410 }
 411
 412 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 413                                   struct perf_evsel *evsel, int idx)
 414 {
 415         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 416
 417         if (per_cpu_mmaps)
 418                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 419         else
 420                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 421 }
 422
 423 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 424 {
 425         int nr_cpus = cpu_map__nr(evlist->cpus);
 426         int nr_threads = thread_map__nr(evlist->threads);
 427         int nfds = 0;
 428         struct perf_evsel *evsel;
 429
 430         evlist__for_each_entry(evlist, evsel) {
 431                 if (evsel->system_wide)
 432                         nfds += nr_cpus;
 433                 else
 434                         nfds += nr_cpus * nr_threads;
 435         }
 436
 437         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 438             fdarray__grow(&evlist->pollfd, nfds) < 0)
 439                 return -ENOMEM;
 440
 441         return 0;
 442 }
 443
 444 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 445                                      struct perf_mmap *map, short revent)
 446 {
 447         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 448         /*
 449          * Save the idx so that when we filter out fds POLLHUP'ed we can
 450          * close the associated evlist->mmap[] entry.
 451          */
 452         if (pos >= 0) {
 453                 evlist->pollfd.priv[pos].ptr = map;
 454
 455                 fcntl(fd, F_SETFL, O_NONBLOCK);
 456         }
 457
 458         return pos;
 459 }
 460
 461 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 462 {
 463         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 464 }
 465
 466 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 467                                          void *arg __maybe_unused)
 468 {
 469         struct perf_mmap *map = fda->priv[fd].ptr;
 470
 471         if (map)
 472                 perf_mmap__put(map);
 473 }
 474
 475 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 476 {
 477         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 478                                perf_evlist__munmap_filtered, NULL);
 479 }
 480
 481 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 482 {
 483         return fdarray__poll(&evlist->pollfd, timeout);
 484 }
 485
 486 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 487                                  struct perf_evsel *evsel,
 488                                  int cpu, int thread, u64 id)
 489 {
 490         int hash;
 491         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 492
 493         sid->id = id;
 494         sid->evsel = evsel;
 495         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 496         hlist_add_head(&sid->node, &evlist->heads[hash]);
 497 }
 498
 499 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 500                          int cpu, int thread, u64 id)
 501 {
 502         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 503         evsel->id[evsel->ids++] = id;
 504 }
 505
 506 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 507                            struct perf_evsel *evsel,
 508                            int cpu, int thread, int fd)
 509 {
 510         u64 read_data[4] = { 0, };
 511         int id_idx = 1; /* The first entry is the counter value */
 512         u64 id;
 513         int ret;
 514
 515         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 516         if (!ret)
 517                 goto add;
 518
 519         if (errno != ENOTTY)
 520                 return -1;
 521
 522         /* Legacy way to get event id.. All hail to old kernels! */
 523
 524         /*
 525          * This way does not work with group format read, so bail
 526          * out in that case.
 527          */
 528         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 529                 return -1;
 530
 531         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 532             read(fd, &read_data, sizeof(read_data)) == -1)
 533                 return -1;
 534
 535         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 536                 ++id_idx;
 537         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 538                 ++id_idx;
 539
 540         id = read_data[id_idx];
 541
 542  add:
 543         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 544         return 0;
 545 }
 546
 547 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 548                                      struct perf_evsel *evsel, int idx, int cpu,
 549                                      int thread)
 550 {
 551         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 552         sid->idx = idx;
 553         if (evlist->cpus && cpu >= 0)
 554                 sid->cpu = evlist->cpus->map[cpu];
 555         else
 556                 sid->cpu = -1;
 557         if (!evsel->system_wide && evlist->threads && thread >= 0)
 558                 sid->tid = thread_map__pid(evlist->threads, thread);
 559         else
 560                 sid->tid = -1;
 561 }
 562
 563 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 564 {
 565         struct hlist_head *head;
 566         struct perf_sample_id *sid;
 567         int hash;
 568
 569         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 570         head = &evlist->heads[hash];
 571
 572         hlist_for_each_entry(sid, head, node)
 573                 if (sid->id == id)
 574                         return sid;
 575
 576         return NULL;
 577 }
 578
 579 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 580 {
 581         struct perf_sample_id *sid;
 582
 583         if (evlist->nr_entries == 1 || !id)
 584                 return perf_evlist__first(evlist);
 585
 586         sid = perf_evlist__id2sid(evlist, id);
 587         if (sid)
 588                 return sid->evsel;
 589
 590         if (!perf_evlist__sample_id_all(evlist))
 591                 return perf_evlist__first(evlist);
 592
 593         return NULL;
 594 }
 595
 596 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 597                                                 u64 id)
 598 {
 599         struct perf_sample_id *sid;
 600
 601         if (!id)
 602                 return NULL;
 603
 604         sid = perf_evlist__id2sid(evlist, id);
 605         if (sid)
 606                 return sid->evsel;
 607
 608         return NULL;
 609 }
 610
 611 static int perf_evlist__event2id(struct perf_evlist *evlist,
 612                                  union perf_event *event, u64 *id)
 613 {
 614         const u64 *array = event->sample.array;
 615         ssize_t n;
 616
 617         n = (event->header.size - sizeof(event->header)) >> 3;
 618
 619         if (event->header.type == PERF_RECORD_SAMPLE) {
 620                 if (evlist->id_pos >= n)
 621                         return -1;
 622                 *id = array[evlist->id_pos];
 623         } else {
 624                 if (evlist->is_pos > n)
 625                         return -1;
 626                 n -= evlist->is_pos;
 627                 *id = array[n];
 628         }
 629         return 0;
 630 }
 631
 632 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 633                                             union perf_event *event)
 634 {
 635         struct perf_evsel *first = perf_evlist__first(evlist);
 636         struct hlist_head *head;
 637         struct perf_sample_id *sid;
 638         int hash;
 639         u64 id;
 640
 641         if (evlist->nr_entries == 1)
 642                 return first;
 643
 644         if (!first->attr.sample_id_all &&
 645             event->header.type != PERF_RECORD_SAMPLE)
 646                 return first;
 647
 648         if (perf_evlist__event2id(evlist, event, &id))
 649                 return NULL;
 650
 651         /* Synthesized events have an id of zero */
 652         if (!id)
 653                 return first;
 654
 655         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 656         head = &evlist->heads[hash];
 657
 658         hlist_for_each_entry(sid, head, node) {
 659                 if (sid->id == id)
 660                         return sid->evsel;
 661         }
 662         return NULL;
 663 }
 664
 665 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 666 {
 667         int i;
 668
 669         if (!evlist->overwrite_mmap)
 670                 return 0;
 671
 672         for (i = 0; i < evlist->nr_mmaps; i++) {
 673                 int fd = evlist->overwrite_mmap[i].fd;
 674                 int err;
 675
 676                 if (fd < 0)
 677                         continue;
 678                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 679                 if (err)
 680                         return err;
 681         }
 682         return 0;
 683 }
 684
 685 static int perf_evlist__pause(struct perf_evlist *evlist)
 686 {
 687         return perf_evlist__set_paused(evlist, true);
 688 }
 689
 690 static int perf_evlist__resume(struct perf_evlist *evlist)
 691 {
 692         return perf_evlist__set_paused(evlist, false);
 693 }
 694
 695 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 696 {
 697         int i;
 698
 699         if (evlist->mmap)
 700                 for (i = 0; i < evlist->nr_mmaps; i++)
 701                         perf_mmap__munmap(&evlist->mmap[i]);
 702
 703         if (evlist->overwrite_mmap)
 704                 for (i = 0; i < evlist->nr_mmaps; i++)
 705                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 706 }
 707
 708 void perf_evlist__munmap(struct perf_evlist *evlist)
 709 {
 710         perf_evlist__munmap_nofree(evlist);
 711         zfree(&evlist->mmap);
 712         zfree(&evlist->overwrite_mmap);
 713 }
 714
 715 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
 716                                                  bool overwrite)
 717 {
 718         int i;
 719         struct perf_mmap *map;
 720
 721         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 722         if (cpu_map__empty(evlist->cpus))
 723                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 724         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 725         if (!map)
 726                 return NULL;
 727
 728         for (i = 0; i < evlist->nr_mmaps; i++) {
 729                 map[i].fd = -1;
 730                 map[i].overwrite = overwrite;
 731                 /*
 732                  * When the perf_mmap() call is made we grab one refcount, plus
 733                  * one extra to let perf_mmap__consume() get the last
 734                  * events after all real references (perf_mmap__get()) are
 735                  * dropped.
 736                  *
 737                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 738                  * thus does perf_mmap__get() on it.
 739                  */
 740                 refcount_set(&map[i].refcnt, 0);
 741         }
 742         return map;
 743 }
 744
 745 static bool
 746 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 747                          struct perf_evsel *evsel)
 748 {
 749         if (evsel->attr.write_backward)
 750                 return false;
 751         return true;
 752 }
 753
 754 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 755                                        struct mmap_params *mp, int cpu_idx,
 756                                        int thread, int *_output, int *_output_overwrite)
 757 {
 758         struct perf_evsel *evsel;
 759         int revent;
 760         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 761
 762         evlist__for_each_entry(evlist, evsel) {
 763                 struct perf_mmap *maps = evlist->mmap;
 764                 int *output = _output;
 765                 int fd;
 766                 int cpu;
 767
 768                 mp->prot = PROT_READ | PROT_WRITE;
 769                 if (evsel->attr.write_backward) {
 770                         output = _output_overwrite;
 771                         maps = evlist->overwrite_mmap;
 772
 773                         if (!maps) {
 774                                 maps = perf_evlist__alloc_mmap(evlist, true);
 775                                 if (!maps)
 776                                         return -1;
 777                                 evlist->overwrite_mmap = maps;
 778                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 779                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 780                         }
 781                         mp->prot &= ~PROT_WRITE;
 782                 }
 783
 784                 if (evsel->system_wide && thread)
 785                         continue;
 786
 787                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 788                 if (cpu == -1)
 789                         continue;
 790
 791                 fd = FD(evsel, cpu, thread);
 792
 793                 if (*output == -1) {
 794                         *output = fd;
 795
 796                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 797                                 return -1;
 798                 } else {
 799                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 800                                 return -1;
 801
 802                         perf_mmap__get(&maps[idx]);
 803                 }
 804
 805                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 806
 807                 /*
 808                  * The system_wide flag causes a selected event to be opened
 809                  * always without a pid.  Consequently it will never get a
 810                  * POLLHUP, but it is used for tracking in combination with
 811                  * other events, so it should not need to be polled anyway.
 812                  * Therefore don't add it for polling.
 813                  */
 814                 if (!evsel->system_wide &&
 815                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 816                         perf_mmap__put(&maps[idx]);
 817                         return -1;
 818                 }
 819
 820                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 821                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 822                                                    fd) < 0)
 823                                 return -1;
 824                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 825                                                  thread);
 826                 }
 827         }
 828
 829         return 0;
 830 }
 831
 832 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 833                                      struct mmap_params *mp)
 834 {
 835         int cpu, thread;
 836         int nr_cpus = cpu_map__nr(evlist->cpus);
 837         int nr_threads = thread_map__nr(evlist->threads);
 838
 839         pr_debug2("perf event ring buffer mmapped per cpu\n");
 840         for (cpu = 0; cpu < nr_cpus; cpu++) {
 841                 int output = -1;
 842                 int output_overwrite = -1;
 843
 844                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 845                                               true);
 846
 847                 for (thread = 0; thread < nr_threads; thread++) {
 848                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 849                                                         thread, &output, &output_overwrite))
 850                                 goto out_unmap;
 851                 }
 852         }
 853
 854         return 0;
 855
 856 out_unmap:
 857         perf_evlist__munmap_nofree(evlist);
 858         return -1;
 859 }
 860
 861 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 862                                         struct mmap_params *mp)
 863 {
 864         int thread;
 865         int nr_threads = thread_map__nr(evlist->threads);
 866
 867         pr_debug2("perf event ring buffer mmapped per thread\n");
 868         for (thread = 0; thread < nr_threads; thread++) {
 869                 int output = -1;
 870                 int output_overwrite = -1;
 871
 872                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 873                                               false);
 874
 875                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 876                                                 &output, &output_overwrite))
 877                         goto out_unmap;
 878         }
 879
 880         return 0;
 881
 882 out_unmap:
 883         perf_evlist__munmap_nofree(evlist);
 884         return -1;
 885 }
 886
 887 unsigned long perf_event_mlock_kb_in_pages(void)
 888 {
 889         unsigned long pages;
 890         int max;
 891
 892         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 893                 /*
 894                  * Pick a once upon a time good value, i.e. things look
 895                  * strange since we can't read a sysctl value, but lets not
 896                  * die yet...
 897                  */
 898                 max = 512;
 899         } else {
 900                 max -= (page_size / 1024);
 901         }
 902
 903         pages = (max * 1024) / page_size;
 904         if (!is_power_of_2(pages))
 905                 pages = rounddown_pow_of_two(pages);
 906
 907         return pages;
 908 }
 909
 910 size_t perf_evlist__mmap_size(unsigned long pages)
 911 {
 912         if (pages == UINT_MAX)
 913                 pages = perf_event_mlock_kb_in_pages();
 914         else if (!is_power_of_2(pages))
 915                 return 0;
 916
 917         return (pages + 1) * page_size;
 918 }
 919
 920 static long parse_pages_arg(const char *str, unsigned long min,
 921                             unsigned long max)
 922 {
 923         unsigned long pages, val;
 924         static struct parse_tag tags[] = {
 925                 { .tag  = 'B', .mult = 1       },
 926                 { .tag  = 'K', .mult = 1 << 10 },
 927                 { .tag  = 'M', .mult = 1 << 20 },
 928                 { .tag  = 'G', .mult = 1 << 30 },
 929                 { .tag  = 0 },
 930         };
 931
 932         if (str == NULL)
 933                 return -EINVAL;
 934
 935         val = parse_tag_value(str, tags);
 936         if (val != (unsigned long) -1) {
 937                 /* we got file size value */
 938                 pages = PERF_ALIGN(val, page_size) / page_size;
 939         } else {
 940                 /* we got pages count value */
 941                 char *eptr;
 942                 pages = strtoul(str, &eptr, 10);
 943                 if (*eptr != '\0')
 944                         return -EINVAL;
 945         }
 946
 947         if (pages == 0 && min == 0) {
 948                 /* leave number of pages at 0 */
 949         } else if (!is_power_of_2(pages)) {
 950                 char buf[100];
 951
 952                 /* round pages up to next power of 2 */
 953                 pages = roundup_pow_of_two(pages);
 954                 if (!pages)
 955                         return -EINVAL;
 956
 957                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 958                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 959                         buf, pages);
 960         }
 961
 962         if (pages > max)
 963                 return -EINVAL;
 964
 965         return pages;
 966 }
 967
 968 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 969 {
 970         unsigned long max = UINT_MAX;
 971         long pages;
 972
 973         if (max > SIZE_MAX / page_size)
 974                 max = SIZE_MAX / page_size;
 975
 976         pages = parse_pages_arg(str, 1, max);
 977         if (pages < 0) {
 978                 pr_err("Invalid argument for --mmap_pages/-m\n");
 979                 return -1;
 980         }
 981
 982         *mmap_pages = pages;
 983         return 0;
 984 }
 985
 986 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 987                                   int unset __maybe_unused)
 988 {
 989         return __perf_evlist__parse_mmap_pages(opt->value, str);
 990 }
 991
 992 /**
 993  * perf_evlist__mmap_ex - Create mmaps to receive events.
 994  * @evlist: list of events
 995  * @pages: map length in pages
 996  * @overwrite: overwrite older events?
 997  * @auxtrace_pages - auxtrace map length in pages
 998  * @auxtrace_overwrite - overwrite older auxtrace data?
 999  *
1000  * If @overwrite is %false the user needs to signal event consumption using
1001  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1002  * automatically.
1003  *
1004  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1005  * consumption using auxtrace_mmap__write_tail().
1006  *
1007  * Return: %0 on success, negative error code otherwise.
1008  */
1009 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1010                          unsigned int auxtrace_pages,
1011                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1012                          int comp_level)
1013 {
1014         struct perf_evsel *evsel;
1015         const struct perf_cpu_map *cpus = evlist->cpus;
1016         const struct thread_map *threads = evlist->threads;
1017         /*
1018          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1019          * Its value is decided by evsel's write_backward.
1020          * So &mp should not be passed through const pointer.
1021          */
1022         struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
1023                                   .comp_level = comp_level };
1024
1025         if (!evlist->mmap)
1026                 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1027         if (!evlist->mmap)
1028                 return -ENOMEM;
1029
1030         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1031                 return -ENOMEM;
1032
1033         evlist->mmap_len = perf_evlist__mmap_size(pages);
1034         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1035         mp.mask = evlist->mmap_len - page_size - 1;
1036
1037         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1038                                    auxtrace_pages, auxtrace_overwrite);
1039
1040         evlist__for_each_entry(evlist, evsel) {
1041                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1042                     evsel->sample_id == NULL &&
1043                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1044                         return -ENOMEM;
1045         }
1046
1047         if (cpu_map__empty(cpus))
1048                 return perf_evlist__mmap_per_thread(evlist, &mp);
1049
1050         return perf_evlist__mmap_per_cpu(evlist, &mp);
1051 }
1052
1053 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1054 {
1055         return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1056 }
1057
1058 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1059 {
1060         bool all_threads = (target->per_thread && target->system_wide);
1061         struct perf_cpu_map *cpus;
1062         struct thread_map *threads;
1063
1064         /*
1065          * If specify '-a' and '--per-thread' to perf record, perf record
1066          * will override '--per-thread'. target->per_thread = false and
1067          * target->system_wide = true.
1068          *
1069          * If specify '--per-thread' only to perf record,
1070          * target->per_thread = true and target->system_wide = false.
1071          *
1072          * So target->per_thread && target->system_wide is false.
1073          * For perf record, thread_map__new_str doesn't call
1074          * thread_map__new_all_cpus. That will keep perf record's
1075          * current behavior.
1076          *
1077          * For perf stat, it allows the case that target->per_thread and
1078          * target->system_wide are all true. It means to collect system-wide
1079          * per-thread data. thread_map__new_str will call
1080          * thread_map__new_all_cpus to enumerate all threads.
1081          */
1082         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1083                                       all_threads);
1084
1085         if (!threads)
1086                 return -1;
1087
1088         if (target__uses_dummy_map(target))
1089                 cpus = cpu_map__dummy_new();
1090         else
1091                 cpus = cpu_map__new(target->cpu_list);
1092
1093         if (!cpus)
1094                 goto out_delete_threads;
1095
1096         evlist->has_user_cpus = !!target->cpu_list;
1097
1098         perf_evlist__set_maps(evlist, cpus, threads);
1099
1100         return 0;
1101
1102 out_delete_threads:
1103         thread_map__put(threads);
1104         return -1;
1105 }
1106
1107 void perf_evlist__set_maps(struct perf_evlist *evlist, struct perf_cpu_map *cpus,
1108                            struct thread_map *threads)
1109 {
1110         /*
1111          * Allow for the possibility that one or another of the maps isn't being
1112          * changed i.e. don't put it.  Note we are assuming the maps that are
1113          * being applied are brand new and evlist is taking ownership of the
1114          * original reference count of 1.  If that is not the case it is up to
1115          * the caller to increase the reference count.
1116          */
1117         if (cpus != evlist->cpus) {
1118                 cpu_map__put(evlist->cpus);
1119                 evlist->cpus = cpu_map__get(cpus);
1120         }
1121
1122         if (threads != evlist->threads) {
1123                 thread_map__put(evlist->threads);
1124                 evlist->threads = thread_map__get(threads);
1125         }
1126
1127         perf_evlist__propagate_maps(evlist);
1128 }
1129
1130 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1131                                    enum perf_event_sample_format bit)
1132 {
1133         struct perf_evsel *evsel;
1134
1135         evlist__for_each_entry(evlist, evsel)
1136                 __perf_evsel__set_sample_bit(evsel, bit);
1137 }
1138
1139 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1140                                      enum perf_event_sample_format bit)
1141 {
1142         struct perf_evsel *evsel;
1143
1144         evlist__for_each_entry(evlist, evsel)
1145                 __perf_evsel__reset_sample_bit(evsel, bit);
1146 }
1147
1148 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1149 {
1150         struct perf_evsel *evsel;
1151         int err = 0;
1152
1153         evlist__for_each_entry(evlist, evsel) {
1154                 if (evsel->filter == NULL)
1155                         continue;
1156
1157                 /*
1158                  * filters only work for tracepoint event, which doesn't have cpu limit.
1159                  * So evlist and evsel should always be same.
1160                  */
1161                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1162                 if (err) {
1163                         *err_evsel = evsel;
1164                         break;
1165                 }
1166         }
1167
1168         return err;
1169 }
1170
1171 int perf_evlist__set_tp_filter(struct perf_evlist *evlist, const char *filter)
1172 {
1173         struct perf_evsel *evsel;
1174         int err = 0;
1175
1176         evlist__for_each_entry(evlist, evsel) {
1177                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1178                         continue;
1179
1180                 err = perf_evsel__set_filter(evsel, filter);
1181                 if (err)
1182                         break;
1183         }
1184
1185         return err;
1186 }
1187
1188 int perf_evlist__set_tp_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1189 {
1190         char *filter;
1191         int ret = -1;
1192         size_t i;
1193
1194         for (i = 0; i < npids; ++i) {
1195                 if (i == 0) {
1196                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1197                                 return -1;
1198                 } else {
1199                         char *tmp;
1200
1201                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1202                                 goto out_free;
1203
1204                         free(filter);
1205                         filter = tmp;
1206                 }
1207         }
1208
1209         ret = perf_evlist__set_tp_filter(evlist, filter);
1210 out_free:
1211         free(filter);
1212         return ret;
1213 }
1214
1215 int perf_evlist__set_tp_filter_pid(struct perf_evlist *evlist, pid_t pid)
1216 {
1217         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1218 }
1219
1220 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1221 {
1222         struct perf_evsel *pos;
1223
1224         if (evlist->nr_entries == 1)
1225                 return true;
1226
1227         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1228                 return false;
1229
1230         evlist__for_each_entry(evlist, pos) {
1231                 if (pos->id_pos != evlist->id_pos ||
1232                     pos->is_pos != evlist->is_pos)
1233                         return false;
1234         }
1235
1236         return true;
1237 }
1238
1239 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1240 {
1241         struct perf_evsel *evsel;
1242
1243         if (evlist->combined_sample_type)
1244                 return evlist->combined_sample_type;
1245
1246         evlist__for_each_entry(evlist, evsel)
1247                 evlist->combined_sample_type |= evsel->attr.sample_type;
1248
1249         return evlist->combined_sample_type;
1250 }
1251
1252 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1253 {
1254         evlist->combined_sample_type = 0;
1255         return __perf_evlist__combined_sample_type(evlist);
1256 }
1257
1258 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1259 {
1260         struct perf_evsel *evsel;
1261         u64 branch_type = 0;
1262
1263         evlist__for_each_entry(evlist, evsel)
1264                 branch_type |= evsel->attr.branch_sample_type;
1265         return branch_type;
1266 }
1267
1268 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1269 {
1270         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1271         u64 read_format = first->attr.read_format;
1272         u64 sample_type = first->attr.sample_type;
1273
1274         evlist__for_each_entry(evlist, pos) {
1275                 if (read_format != pos->attr.read_format)
1276                         return false;
1277         }
1278
1279         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1280         if ((sample_type & PERF_SAMPLE_READ) &&
1281             !(read_format & PERF_FORMAT_ID)) {
1282                 return false;
1283         }
1284
1285         return true;
1286 }
1287
1288 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1289 {
1290         struct perf_evsel *first = perf_evlist__first(evlist);
1291         return first->attr.read_format;
1292 }
1293
1294 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1295 {
1296         struct perf_evsel *first = perf_evlist__first(evlist);
1297         struct perf_sample *data;
1298         u64 sample_type;
1299         u16 size = 0;
1300
1301         if (!first->attr.sample_id_all)
1302                 goto out;
1303
1304         sample_type = first->attr.sample_type;
1305
1306         if (sample_type & PERF_SAMPLE_TID)
1307                 size += sizeof(data->tid) * 2;
1308
1309        if (sample_type & PERF_SAMPLE_TIME)
1310                 size += sizeof(data->time);
1311
1312         if (sample_type & PERF_SAMPLE_ID)
1313                 size += sizeof(data->id);
1314
1315         if (sample_type & PERF_SAMPLE_STREAM_ID)
1316                 size += sizeof(data->stream_id);
1317
1318         if (sample_type & PERF_SAMPLE_CPU)
1319                 size += sizeof(data->cpu) * 2;
1320
1321         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1322                 size += sizeof(data->id);
1323 out:
1324         return size;
1325 }
1326
1327 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1328 {
1329         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1330
1331         evlist__for_each_entry_continue(evlist, pos) {
1332                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1333                         return false;
1334         }
1335
1336         return true;
1337 }
1338
1339 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1340 {
1341         struct perf_evsel *first = perf_evlist__first(evlist);
1342         return first->attr.sample_id_all;
1343 }
1344
1345 void perf_evlist__set_selected(struct perf_evlist *evlist,
1346                                struct perf_evsel *evsel)
1347 {
1348         evlist->selected = evsel;
1349 }
1350
1351 void perf_evlist__close(struct perf_evlist *evlist)
1352 {
1353         struct perf_evsel *evsel;
1354
1355         evlist__for_each_entry_reverse(evlist, evsel)
1356                 perf_evsel__close(evsel);
1357 }
1358
1359 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1360 {
1361         struct perf_cpu_map *cpus;
1362         struct thread_map *threads;
1363         int err = -ENOMEM;
1364
1365         /*
1366          * Try reading /sys/devices/system/cpu/online to get
1367          * an all cpus map.
1368          *
1369          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1370          * code needs an overhaul to properly forward the
1371          * error, and we may not want to do that fallback to a
1372          * default cpu identity map :-\
1373          */
1374         cpus = cpu_map__new(NULL);
1375         if (!cpus)
1376                 goto out;
1377
1378         threads = thread_map__new_dummy();
1379         if (!threads)
1380                 goto out_put;
1381
1382         perf_evlist__set_maps(evlist, cpus, threads);
1383 out:
1384         return err;
1385 out_put:
1386         cpu_map__put(cpus);
1387         goto out;
1388 }
1389
1390 int perf_evlist__open(struct perf_evlist *evlist)
1391 {
1392         struct perf_evsel *evsel;
1393         int err;
1394
1395         /*
1396          * Default: one fd per CPU, all threads, aka systemwide
1397          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1398          */
1399         if (evlist->threads == NULL && evlist->cpus == NULL) {
1400                 err = perf_evlist__create_syswide_maps(evlist);
1401                 if (err < 0)
1402                         goto out_err;
1403         }
1404
1405         perf_evlist__update_id_pos(evlist);
1406
1407         evlist__for_each_entry(evlist, evsel) {
1408                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1409                 if (err < 0)
1410                         goto out_err;
1411         }
1412
1413         return 0;
1414 out_err:
1415         perf_evlist__close(evlist);
1416         errno = -err;
1417         return err;
1418 }
1419
1420 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1421                                   const char *argv[], bool pipe_output,
1422                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1423 {
1424         int child_ready_pipe[2], go_pipe[2];
1425         char bf;
1426
1427         if (pipe(child_ready_pipe) < 0) {
1428                 perror("failed to create 'ready' pipe");
1429                 return -1;
1430         }
1431
1432         if (pipe(go_pipe) < 0) {
1433                 perror("failed to create 'go' pipe");
1434                 goto out_close_ready_pipe;
1435         }
1436
1437         evlist->workload.pid = fork();
1438         if (evlist->workload.pid < 0) {
1439                 perror("failed to fork");
1440                 goto out_close_pipes;
1441         }
1442
1443         if (!evlist->workload.pid) {
1444                 int ret;
1445
1446                 if (pipe_output)
1447                         dup2(2, 1);
1448
1449                 signal(SIGTERM, SIG_DFL);
1450
1451                 close(child_ready_pipe[0]);
1452                 close(go_pipe[1]);
1453                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1454
1455                 /*
1456                  * Tell the parent we're ready to go
1457                  */
1458                 close(child_ready_pipe[1]);
1459
1460                 /*
1461                  * Wait until the parent tells us to go.
1462                  */
1463                 ret = read(go_pipe[0], &bf, 1);
1464                 /*
1465                  * The parent will ask for the execvp() to be performed by
1466                  * writing exactly one byte, in workload.cork_fd, usually via
1467                  * perf_evlist__start_workload().
1468                  *
1469                  * For cancelling the workload without actually running it,
1470                  * the parent will just close workload.cork_fd, without writing
1471                  * anything, i.e. read will return zero and we just exit()
1472                  * here.
1473                  */
1474                 if (ret != 1) {
1475                         if (ret == -1)
1476                                 perror("unable to read pipe");
1477                         exit(ret);
1478                 }
1479
1480                 execvp(argv[0], (char **)argv);
1481
1482                 if (exec_error) {
1483                         union sigval val;
1484
1485                         val.sival_int = errno;
1486                         if (sigqueue(getppid(), SIGUSR1, val))
1487                                 perror(argv[0]);
1488                 } else
1489                         perror(argv[0]);
1490                 exit(-1);
1491         }
1492
1493         if (exec_error) {
1494                 struct sigaction act = {
1495                         .sa_flags     = SA_SIGINFO,
1496                         .sa_sigaction = exec_error,
1497                 };
1498                 sigaction(SIGUSR1, &act, NULL);
1499         }
1500
1501         if (target__none(target)) {
1502                 if (evlist->threads == NULL) {
1503                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1504                                 __func__, __LINE__);
1505                         goto out_close_pipes;
1506                 }
1507                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1508         }
1509
1510         close(child_ready_pipe[1]);
1511         close(go_pipe[0]);
1512         /*
1513          * wait for child to settle
1514          */
1515         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1516                 perror("unable to read pipe");
1517                 goto out_close_pipes;
1518         }
1519
1520         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1521         evlist->workload.cork_fd = go_pipe[1];
1522         close(child_ready_pipe[0]);
1523         return 0;
1524
1525 out_close_pipes:
1526         close(go_pipe[0]);
1527         close(go_pipe[1]);
1528 out_close_ready_pipe:
1529         close(child_ready_pipe[0]);
1530         close(child_ready_pipe[1]);
1531         return -1;
1532 }
1533
1534 int perf_evlist__start_workload(struct perf_evlist *evlist)
1535 {
1536         if (evlist->workload.cork_fd > 0) {
1537                 char bf = 0;
1538                 int ret;
1539                 /*
1540                  * Remove the cork, let it rip!
1541                  */
1542                 ret = write(evlist->workload.cork_fd, &bf, 1);
1543                 if (ret < 0)
1544                         perror("unable to write to pipe");
1545
1546                 close(evlist->workload.cork_fd);
1547                 return ret;
1548         }
1549
1550         return 0;
1551 }
1552
1553 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1554                               struct perf_sample *sample)
1555 {
1556         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1557
1558         if (!evsel)
1559                 return -EFAULT;
1560         return perf_evsel__parse_sample(evsel, event, sample);
1561 }
1562
1563 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1564                                         union perf_event *event,
1565                                         u64 *timestamp)
1566 {
1567         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1568
1569         if (!evsel)
1570                 return -EFAULT;
1571         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1572 }
1573
1574 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1575 {
1576         struct perf_evsel *evsel;
1577         size_t printed = 0;
1578
1579         evlist__for_each_entry(evlist, evsel) {
1580                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1581                                    perf_evsel__name(evsel));
1582         }
1583
1584         return printed + fprintf(fp, "\n");
1585 }
1586
1587 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1588                                int err, char *buf, size_t size)
1589 {
1590         int printed, value;
1591         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1592
1593         switch (err) {
1594         case EACCES:
1595         case EPERM:
1596                 printed = scnprintf(buf, size,
1597                                     "Error:\t%s.\n"
1598                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1599
1600                 value = perf_event_paranoid();
1601
1602                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1603
1604                 if (value >= 2) {
1605                         printed += scnprintf(buf + printed, size - printed,
1606                                              "For your workloads it needs to be <= 1\nHint:\t");
1607                 }
1608                 printed += scnprintf(buf + printed, size - printed,
1609                                      "For system wide tracing it needs to be set to -1.\n");
1610
1611                 printed += scnprintf(buf + printed, size - printed,
1612                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1613                                     "Hint:\tThe current value is %d.", value);
1614                 break;
1615         case EINVAL: {
1616                 struct perf_evsel *first = perf_evlist__first(evlist);
1617                 int max_freq;
1618
1619                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1620                         goto out_default;
1621
1622                 if (first->attr.sample_freq < (u64)max_freq)
1623                         goto out_default;
1624
1625                 printed = scnprintf(buf, size,
1626                                     "Error:\t%s.\n"
1627                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1628                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1629                                     emsg, max_freq, first->attr.sample_freq);
1630                 break;
1631         }
1632         default:
1633 out_default:
1634                 scnprintf(buf, size, "%s", emsg);
1635                 break;
1636         }
1637
1638         return 0;
1639 }
1640
1641 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1642 {
1643         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1644         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1645
1646         switch (err) {
1647         case EPERM:
1648                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1649                 printed += scnprintf(buf + printed, size - printed,
1650                                      "Error:\t%s.\n"
1651                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1652                                      "Hint:\tTried using %zd kB.\n",
1653                                      emsg, pages_max_per_user, pages_attempted);
1654
1655                 if (pages_attempted >= pages_max_per_user) {
1656                         printed += scnprintf(buf + printed, size - printed,
1657                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1658                                              pages_max_per_user + pages_attempted);
1659                 }
1660
1661                 printed += scnprintf(buf + printed, size - printed,
1662                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1663                 break;
1664         default:
1665                 scnprintf(buf, size, "%s", emsg);
1666                 break;
1667         }
1668
1669         return 0;
1670 }
1671
1672 void perf_evlist__to_front(struct perf_evlist *evlist,
1673                            struct perf_evsel *move_evsel)
1674 {
1675         struct perf_evsel *evsel, *n;
1676         LIST_HEAD(move);
1677
1678         if (move_evsel == perf_evlist__first(evlist))
1679                 return;
1680
1681         evlist__for_each_entry_safe(evlist, n, evsel) {
1682                 if (evsel->leader == move_evsel->leader)
1683                         list_move_tail(&evsel->node, &move);
1684         }
1685
1686         list_splice(&move, &evlist->entries);
1687 }
1688
1689 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1690                                      struct perf_evsel *tracking_evsel)
1691 {
1692         struct perf_evsel *evsel;
1693
1694         if (tracking_evsel->tracking)
1695                 return;
1696
1697         evlist__for_each_entry(evlist, evsel) {
1698                 if (evsel != tracking_evsel)
1699                         evsel->tracking = false;
1700         }
1701
1702         tracking_evsel->tracking = true;
1703 }
1704
1705 struct perf_evsel *
1706 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1707                                const char *str)
1708 {
1709         struct perf_evsel *evsel;
1710
1711         evlist__for_each_entry(evlist, evsel) {
1712                 if (!evsel->name)
1713                         continue;
1714                 if (strcmp(str, evsel->name) == 0)
1715                         return evsel;
1716         }
1717
1718         return NULL;
1719 }
1720
1721 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1722                                   enum bkw_mmap_state state)
1723 {
1724         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1725         enum action {
1726                 NONE,
1727                 PAUSE,
1728                 RESUME,
1729         } action = NONE;
1730
1731         if (!evlist->overwrite_mmap)
1732                 return;
1733
1734         switch (old_state) {
1735         case BKW_MMAP_NOTREADY: {
1736                 if (state != BKW_MMAP_RUNNING)
1737                         goto state_err;
1738                 break;
1739         }
1740         case BKW_MMAP_RUNNING: {
1741                 if (state != BKW_MMAP_DATA_PENDING)
1742                         goto state_err;
1743                 action = PAUSE;
1744                 break;
1745         }
1746         case BKW_MMAP_DATA_PENDING: {
1747                 if (state != BKW_MMAP_EMPTY)
1748                         goto state_err;
1749                 break;
1750         }
1751         case BKW_MMAP_EMPTY: {
1752                 if (state != BKW_MMAP_RUNNING)
1753                         goto state_err;
1754                 action = RESUME;
1755                 break;
1756         }
1757         default:
1758                 WARN_ONCE(1, "Shouldn't get there\n");
1759         }
1760
1761         evlist->bkw_mmap_state = state;
1762
1763         switch (action) {
1764         case PAUSE:
1765                 perf_evlist__pause(evlist);
1766                 break;
1767         case RESUME:
1768                 perf_evlist__resume(evlist);
1769                 break;
1770         case NONE:
1771         default:
1772                 break;
1773         }
1774
1775 state_err:
1776         return;
1777 }
1778
1779 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1780 {
1781         struct perf_evsel *evsel;
1782
1783         evlist__for_each_entry(evlist, evsel) {
1784                 if (!evsel->attr.exclude_kernel)
1785                         return false;
1786         }
1787
1788         return true;
1789 }
1790
1791 /*
1792  * Events in data file are not collect in groups, but we still want
1793  * the group display. Set the artificial group and set the leader's
1794  * forced_leader flag to notify the display code.
1795  */
1796 void perf_evlist__force_leader(struct perf_evlist *evlist)
1797 {
1798         if (!evlist->nr_groups) {
1799                 struct perf_evsel *leader = perf_evlist__first(evlist);
1800
1801                 perf_evlist__set_leader(evlist);
1802                 leader->forced_leader = true;
1803         }
1804 }
1805
1806 struct perf_evsel *perf_evlist__reset_weak_group(struct perf_evlist *evsel_list,
1807                                                  struct perf_evsel *evsel)
1808 {
1809         struct perf_evsel *c2, *leader;
1810         bool is_open = true;
1811
1812         leader = evsel->leader;
1813         pr_debug("Weak group for %s/%d failed\n",
1814                         leader->name, leader->nr_members);
1815
1816         /*
1817          * for_each_group_member doesn't work here because it doesn't
1818          * include the first entry.
1819          */
1820         evlist__for_each_entry(evsel_list, c2) {
1821                 if (c2 == evsel)
1822                         is_open = false;
1823                 if (c2->leader == leader) {
1824                         if (is_open)
1825                                 perf_evsel__close(c2);
1826                         c2->leader = c2;
1827                         c2->nr_members = 0;
1828                 }
1829         }
1830         return leader;
1831 }
1832
1833 int perf_evlist__add_sb_event(struct perf_evlist **evlist,
1834                               struct perf_event_attr *attr,
1835                               perf_evsel__sb_cb_t cb,
1836                               void *data)
1837 {
1838         struct perf_evsel *evsel;
1839         bool new_evlist = (*evlist) == NULL;
1840
1841         if (*evlist == NULL)
1842                 *evlist = perf_evlist__new();
1843         if (*evlist == NULL)
1844                 return -1;
1845
1846         if (!attr->sample_id_all) {
1847                 pr_warning("enabling sample_id_all for all side band events\n");
1848                 attr->sample_id_all = 1;
1849         }
1850
1851         evsel = perf_evsel__new_idx(attr, (*evlist)->nr_entries);
1852         if (!evsel)
1853                 goto out_err;
1854
1855         evsel->side_band.cb = cb;
1856         evsel->side_band.data = data;
1857         perf_evlist__add(*evlist, evsel);
1858         return 0;
1859
1860 out_err:
1861         if (new_evlist) {
1862                 perf_evlist__delete(*evlist);
1863                 *evlist = NULL;
1864         }
1865         return -1;
1866 }
1867
1868 static void *perf_evlist__poll_thread(void *arg)
1869 {
1870         struct perf_evlist *evlist = arg;
1871         bool draining = false;
1872         int i, done = 0;
1873
1874         while (!done) {
1875                 bool got_data = false;
1876
1877                 if (evlist->thread.done)
1878                         draining = true;
1879
1880                 if (!draining)
1881                         perf_evlist__poll(evlist, 1000);
1882
1883                 for (i = 0; i < evlist->nr_mmaps; i++) {
1884                         struct perf_mmap *map = &evlist->mmap[i];
1885                         union perf_event *event;
1886
1887                         if (perf_mmap__read_init(map))
1888                                 continue;
1889                         while ((event = perf_mmap__read_event(map)) != NULL) {
1890                                 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1891
1892                                 if (evsel && evsel->side_band.cb)
1893                                         evsel->side_band.cb(event, evsel->side_band.data);
1894                                 else
1895                                         pr_warning("cannot locate proper evsel for the side band event\n");
1896
1897                                 perf_mmap__consume(map);
1898                                 got_data = true;
1899                         }
1900                         perf_mmap__read_done(map);
1901                 }
1902
1903                 if (draining && !got_data)
1904                         break;
1905         }
1906         return NULL;
1907 }
1908
1909 int perf_evlist__start_sb_thread(struct perf_evlist *evlist,
1910                                  struct target *target)
1911 {
1912         struct perf_evsel *counter;
1913
1914         if (!evlist)
1915                 return 0;
1916
1917         if (perf_evlist__create_maps(evlist, target))
1918                 goto out_delete_evlist;
1919
1920         evlist__for_each_entry(evlist, counter) {
1921                 if (perf_evsel__open(counter, evlist->cpus,
1922                                      evlist->threads) < 0)
1923                         goto out_delete_evlist;
1924         }
1925
1926         if (perf_evlist__mmap(evlist, UINT_MAX))
1927                 goto out_delete_evlist;
1928
1929         evlist__for_each_entry(evlist, counter) {
1930                 if (perf_evsel__enable(counter))
1931                         goto out_delete_evlist;
1932         }
1933
1934         evlist->thread.done = 0;
1935         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1936                 goto out_delete_evlist;
1937
1938         return 0;
1939
1940 out_delete_evlist:
1941         perf_evlist__delete(evlist);
1942         evlist = NULL;
1943         return -1;
1944 }
1945
1946 void perf_evlist__stop_sb_thread(struct perf_evlist *evlist)
1947 {
1948         if (!evlist)
1949                 return;
1950         evlist->thread.done = 1;
1951         pthread_join(evlist->thread.th, NULL);
1952         perf_evlist__delete(evlist);
1953 }