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