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