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