tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <poll.h>
  14 #include "cpumap.h"
  15 #include "thread_map.h"
  16 #include "target.h"
  17 #include "evlist.h"
  18 #include "evsel.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "asm/bug.h"
  22 #include "bpf-event.h"
  23 #include <signal.h>
  24 #include <unistd.h>
  25
  26 #include "parse-events.h"
  27 #include <subcmd/parse-options.h>
  28
  29 #include <fcntl.h>
  30 #include <sys/ioctl.h>
  31 #include <sys/mman.h>
  32
  33 #include <linux/bitops.h>
  34 #include <linux/hash.h>
  35 #include <linux/log2.h>
  36 #include <linux/err.h>
  37
  38 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  39 int sigqueue(pid_t pid, int sig, const union sigval value);
  40 #endif
  41
  42 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  43 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  44
  45 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  46                        struct thread_map *threads)
  47 {
  48         int i;
  49
  50         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  51                 INIT_HLIST_HEAD(&evlist->heads[i]);
  52         INIT_LIST_HEAD(&evlist->entries);
  53         perf_evlist__set_maps(evlist, cpus, threads);
  54         fdarray__init(&evlist->pollfd, 64);
  55         evlist->workload.pid = -1;
  56         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  57 }
  58
  59 struct perf_evlist *perf_evlist__new(void)
  60 {
  61         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  62
  63         if (evlist != NULL)
  64                 perf_evlist__init(evlist, NULL, NULL);
  65
  66         return evlist;
  67 }
  68
  69 struct perf_evlist *perf_evlist__new_default(void)
  70 {
  71         struct perf_evlist *evlist = perf_evlist__new();
  72
  73         if (evlist && perf_evlist__add_default(evlist)) {
  74                 perf_evlist__delete(evlist);
  75                 evlist = NULL;
  76         }
  77
  78         return evlist;
  79 }
  80
  81 struct perf_evlist *perf_evlist__new_dummy(void)
  82 {
  83         struct perf_evlist *evlist = perf_evlist__new();
  84
  85         if (evlist && perf_evlist__add_dummy(evlist)) {
  86                 perf_evlist__delete(evlist);
  87                 evlist = NULL;
  88         }
  89
  90         return evlist;
  91 }
  92
  93 /**
  94  * perf_evlist__set_id_pos - set the positions of event ids.
  95  * @evlist: selected event list
  96  *
  97  * Events with compatible sample types all have the same id_pos
  98  * and is_pos.  For convenience, put a copy on evlist.
  99  */
 100 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
 101 {
 102         struct perf_evsel *first = perf_evlist__first(evlist);
 103
 104         evlist->id_pos = first->id_pos;
 105         evlist->is_pos = first->is_pos;
 106 }
 107
 108 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 109 {
 110         struct perf_evsel *evsel;
 111
 112         evlist__for_each_entry(evlist, evsel)
 113                 perf_evsel__calc_id_pos(evsel);
 114
 115         perf_evlist__set_id_pos(evlist);
 116 }
 117
 118 static void perf_evlist__purge(struct perf_evlist *evlist)
 119 {
 120         struct perf_evsel *pos, *n;
 121
 122         evlist__for_each_entry_safe(evlist, n, pos) {
 123                 list_del_init(&pos->node);
 124                 pos->evlist = NULL;
 125                 perf_evsel__delete(pos);
 126         }
 127
 128         evlist->nr_entries = 0;
 129 }
 130
 131 void perf_evlist__exit(struct perf_evlist *evlist)
 132 {
 133         zfree(&evlist->mmap);
 134         zfree(&evlist->overwrite_mmap);
 135         fdarray__exit(&evlist->pollfd);
 136 }
 137
 138 void perf_evlist__delete(struct perf_evlist *evlist)
 139 {
 140         if (evlist == NULL)
 141                 return;
 142
 143         perf_evlist__munmap(evlist);
 144         perf_evlist__close(evlist);
 145         cpu_map__put(evlist->cpus);
 146         thread_map__put(evlist->threads);
 147         evlist->cpus = NULL;
 148         evlist->threads = NULL;
 149         perf_evlist__purge(evlist);
 150         perf_evlist__exit(evlist);
 151         free(evlist);
 152 }
 153
 154 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 155                                           struct perf_evsel *evsel)
 156 {
 157         /*
 158          * We already have cpus for evsel (via PMU sysfs) so
 159          * keep it, if there's no target cpu list defined.
 160          */
 161         if (!evsel->own_cpus || evlist->has_user_cpus) {
 162                 cpu_map__put(evsel->cpus);
 163                 evsel->cpus = cpu_map__get(evlist->cpus);
 164         } else if (evsel->cpus != evsel->own_cpus) {
 165                 cpu_map__put(evsel->cpus);
 166                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 167         }
 168
 169         thread_map__put(evsel->threads);
 170         evsel->threads = thread_map__get(evlist->threads);
 171 }
 172
 173 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 174 {
 175         struct perf_evsel *evsel;
 176
 177         evlist__for_each_entry(evlist, evsel)
 178                 __perf_evlist__propagate_maps(evlist, evsel);
 179 }
 180
 181 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 182 {
 183         entry->evlist = evlist;
 184         list_add_tail(&entry->node, &evlist->entries);
 185         entry->idx = evlist->nr_entries;
 186         entry->tracking = !entry->idx;
 187
 188         if (!evlist->nr_entries++)
 189                 perf_evlist__set_id_pos(evlist);
 190
 191         __perf_evlist__propagate_maps(evlist, entry);
 192 }
 193
 194 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 195 {
 196         evsel->evlist = NULL;
 197         list_del_init(&evsel->node);
 198         evlist->nr_entries -= 1;
 199 }
 200
 201 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 202                                    struct list_head *list)
 203 {
 204         struct perf_evsel *evsel, *temp;
 205
 206         __evlist__for_each_entry_safe(list, temp, evsel) {
 207                 list_del_init(&evsel->node);
 208                 perf_evlist__add(evlist, evsel);
 209         }
 210 }
 211
 212 void __perf_evlist__set_leader(struct list_head *list)
 213 {
 214         struct perf_evsel *evsel, *leader;
 215
 216         leader = list_entry(list->next, struct perf_evsel, node);
 217         evsel = list_entry(list->prev, struct perf_evsel, node);
 218
 219         leader->nr_members = evsel->idx - leader->idx + 1;
 220
 221         __evlist__for_each_entry(list, evsel) {
 222                 evsel->leader = leader;
 223         }
 224 }
 225
 226 void perf_evlist__set_leader(struct perf_evlist *evlist)
 227 {
 228         if (evlist->nr_entries) {
 229                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 230                 __perf_evlist__set_leader(&evlist->entries);
 231         }
 232 }
 233
 234 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 235 {
 236         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 237
 238         if (evsel == NULL)
 239                 return -ENOMEM;
 240
 241         perf_evlist__add(evlist, evsel);
 242         return 0;
 243 }
 244
 245 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 246 {
 247         struct perf_event_attr attr = {
 248                 .type   = PERF_TYPE_SOFTWARE,
 249                 .config = PERF_COUNT_SW_DUMMY,
 250                 .size   = sizeof(attr), /* to capture ABI version */
 251         };
 252         struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
 253
 254         if (evsel == NULL)
 255                 return -ENOMEM;
 256
 257         perf_evlist__add(evlist, evsel);
 258         return 0;
 259 }
 260
 261 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 262                                   struct perf_event_attr *attrs, size_t nr_attrs)
 263 {
 264         struct perf_evsel *evsel, *n;
 265         LIST_HEAD(head);
 266         size_t i;
 267
 268         for (i = 0; i < nr_attrs; i++) {
 269                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 270                 if (evsel == NULL)
 271                         goto out_delete_partial_list;
 272                 list_add_tail(&evsel->node, &head);
 273         }
 274
 275         perf_evlist__splice_list_tail(evlist, &head);
 276
 277         return 0;
 278
 279 out_delete_partial_list:
 280         __evlist__for_each_entry_safe(&head, n, evsel)
 281                 perf_evsel__delete(evsel);
 282         return -1;
 283 }
 284
 285 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 286                                      struct perf_event_attr *attrs, size_t nr_attrs)
 287 {
 288         size_t i;
 289
 290         for (i = 0; i < nr_attrs; i++)
 291                 event_attr_init(attrs + i);
 292
 293         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 294 }
 295
 296 struct perf_evsel *
 297 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 298 {
 299         struct perf_evsel *evsel;
 300
 301         evlist__for_each_entry(evlist, evsel) {
 302                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 303                     (int)evsel->attr.config == id)
 304                         return evsel;
 305         }
 306
 307         return NULL;
 308 }
 309
 310 struct perf_evsel *
 311 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 312                                      const char *name)
 313 {
 314         struct perf_evsel *evsel;
 315
 316         evlist__for_each_entry(evlist, evsel) {
 317                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 318                     (strcmp(evsel->name, name) == 0))
 319                         return evsel;
 320         }
 321
 322         return NULL;
 323 }
 324
 325 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 326                            const char *sys, const char *name, void *handler)
 327 {
 328         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 329
 330         if (IS_ERR(evsel))
 331                 return -1;
 332
 333         evsel->handler = handler;
 334         perf_evlist__add(evlist, evsel);
 335         return 0;
 336 }
 337
 338 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 339                                    struct perf_evsel *evsel)
 340 {
 341         if (evsel->system_wide)
 342                 return 1;
 343         else
 344                 return thread_map__nr(evlist->threads);
 345 }
 346
 347 void perf_evlist__disable(struct perf_evlist *evlist)
 348 {
 349         struct perf_evsel *pos;
 350
 351         evlist__for_each_entry(evlist, pos) {
 352                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
 353                         continue;
 354                 perf_evsel__disable(pos);
 355         }
 356
 357         evlist->enabled = false;
 358 }
 359
 360 void perf_evlist__enable(struct perf_evlist *evlist)
 361 {
 362         struct perf_evsel *pos;
 363
 364         evlist__for_each_entry(evlist, pos) {
 365                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 366                         continue;
 367                 perf_evsel__enable(pos);
 368         }
 369
 370         evlist->enabled = true;
 371 }
 372
 373 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 374 {
 375         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 376 }
 377
 378 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 379                                          struct perf_evsel *evsel, int cpu)
 380 {
 381         int thread;
 382         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 383
 384         if (!evsel->fd)
 385                 return -EINVAL;
 386
 387         for (thread = 0; thread < nr_threads; thread++) {
 388                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 389                 if (err)
 390                         return err;
 391         }
 392         return 0;
 393 }
 394
 395 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 396                                             struct perf_evsel *evsel,
 397                                             int thread)
 398 {
 399         int cpu;
 400         int nr_cpus = cpu_map__nr(evlist->cpus);
 401
 402         if (!evsel->fd)
 403                 return -EINVAL;
 404
 405         for (cpu = 0; cpu < nr_cpus; cpu++) {
 406                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 407                 if (err)
 408                         return err;
 409         }
 410         return 0;
 411 }
 412
 413 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 414                                   struct perf_evsel *evsel, int idx)
 415 {
 416         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 417
 418         if (per_cpu_mmaps)
 419                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 420         else
 421                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 422 }
 423
 424 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 425 {
 426         int nr_cpus = cpu_map__nr(evlist->cpus);
 427         int nr_threads = thread_map__nr(evlist->threads);
 428         int nfds = 0;
 429         struct perf_evsel *evsel;
 430
 431         evlist__for_each_entry(evlist, evsel) {
 432                 if (evsel->system_wide)
 433                         nfds += nr_cpus;
 434                 else
 435                         nfds += nr_cpus * nr_threads;
 436         }
 437
 438         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 439             fdarray__grow(&evlist->pollfd, nfds) < 0)
 440                 return -ENOMEM;
 441
 442         return 0;
 443 }
 444
 445 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 446                                      struct perf_mmap *map, short revent)
 447 {
 448         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 449         /*
 450          * Save the idx so that when we filter out fds POLLHUP'ed we can
 451          * close the associated evlist->mmap[] entry.
 452          */
 453         if (pos >= 0) {
 454                 evlist->pollfd.priv[pos].ptr = map;
 455
 456                 fcntl(fd, F_SETFL, O_NONBLOCK);
 457         }
 458
 459         return pos;
 460 }
 461
 462 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 463 {
 464         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 465 }
 466
 467 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 468                                          void *arg __maybe_unused)
 469 {
 470         struct perf_mmap *map = fda->priv[fd].ptr;
 471
 472         if (map)
 473                 perf_mmap__put(map);
 474 }
 475
 476 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 477 {
 478         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 479                                perf_evlist__munmap_filtered, NULL);
 480 }
 481
 482 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 483 {
 484         return fdarray__poll(&evlist->pollfd, timeout);
 485 }
 486
 487 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 488                                  struct perf_evsel *evsel,
 489                                  int cpu, int thread, u64 id)
 490 {
 491         int hash;
 492         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 493
 494         sid->id = id;
 495         sid->evsel = evsel;
 496         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 497         hlist_add_head(&sid->node, &evlist->heads[hash]);
 498 }
 499
 500 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 501                          int cpu, int thread, u64 id)
 502 {
 503         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 504         evsel->id[evsel->ids++] = id;
 505 }
 506
 507 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 508                            struct perf_evsel *evsel,
 509                            int cpu, int thread, int fd)
 510 {
 511         u64 read_data[4] = { 0, };
 512         int id_idx = 1; /* The first entry is the counter value */
 513         u64 id;
 514         int ret;
 515
 516         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 517         if (!ret)
 518                 goto add;
 519
 520         if (errno != ENOTTY)
 521                 return -1;
 522
 523         /* Legacy way to get event id.. All hail to old kernels! */
 524
 525         /*
 526          * This way does not work with group format read, so bail
 527          * out in that case.
 528          */
 529         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 530                 return -1;
 531
 532         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 533             read(fd, &read_data, sizeof(read_data)) == -1)
 534                 return -1;
 535
 536         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 537                 ++id_idx;
 538         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 539                 ++id_idx;
 540
 541         id = read_data[id_idx];
 542
 543  add:
 544         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 545         return 0;
 546 }
 547
 548 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 549                                      struct perf_evsel *evsel, int idx, int cpu,
 550                                      int thread)
 551 {
 552         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 553         sid->idx = idx;
 554         if (evlist->cpus && cpu >= 0)
 555                 sid->cpu = evlist->cpus->map[cpu];
 556         else
 557                 sid->cpu = -1;
 558         if (!evsel->system_wide && evlist->threads && thread >= 0)
 559                 sid->tid = thread_map__pid(evlist->threads, thread);
 560         else
 561                 sid->tid = -1;
 562 }
 563
 564 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 565 {
 566         struct hlist_head *head;
 567         struct perf_sample_id *sid;
 568         int hash;
 569
 570         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 571         head = &evlist->heads[hash];
 572
 573         hlist_for_each_entry(sid, head, node)
 574                 if (sid->id == id)
 575                         return sid;
 576
 577         return NULL;
 578 }
 579
 580 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 581 {
 582         struct perf_sample_id *sid;
 583
 584         if (evlist->nr_entries == 1 || !id)
 585                 return perf_evlist__first(evlist);
 586
 587         sid = perf_evlist__id2sid(evlist, id);
 588         if (sid)
 589                 return sid->evsel;
 590
 591         if (!perf_evlist__sample_id_all(evlist))
 592                 return perf_evlist__first(evlist);
 593
 594         return NULL;
 595 }
 596
 597 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 598                                                 u64 id)
 599 {
 600         struct perf_sample_id *sid;
 601
 602         if (!id)
 603                 return NULL;
 604
 605         sid = perf_evlist__id2sid(evlist, id);
 606         if (sid)
 607                 return sid->evsel;
 608
 609         return NULL;
 610 }
 611
 612 static int perf_evlist__event2id(struct perf_evlist *evlist,
 613                                  union perf_event *event, u64 *id)
 614 {
 615         const u64 *array = event->sample.array;
 616         ssize_t n;
 617
 618         n = (event->header.size - sizeof(event->header)) >> 3;
 619
 620         if (event->header.type == PERF_RECORD_SAMPLE) {
 621                 if (evlist->id_pos >= n)
 622                         return -1;
 623                 *id = array[evlist->id_pos];
 624         } else {
 625                 if (evlist->is_pos > n)
 626                         return -1;
 627                 n -= evlist->is_pos;
 628                 *id = array[n];
 629         }
 630         return 0;
 631 }
 632
 633 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 634                                             union perf_event *event)
 635 {
 636         struct perf_evsel *first = perf_evlist__first(evlist);
 637         struct hlist_head *head;
 638         struct perf_sample_id *sid;
 639         int hash;
 640         u64 id;
 641
 642         if (evlist->nr_entries == 1)
 643                 return first;
 644
 645         if (!first->attr.sample_id_all &&
 646             event->header.type != PERF_RECORD_SAMPLE)
 647                 return first;
 648
 649         if (perf_evlist__event2id(evlist, event, &id))
 650                 return NULL;
 651
 652         /* Synthesized events have an id of zero */
 653         if (!id)
 654                 return first;
 655
 656         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 657         head = &evlist->heads[hash];
 658
 659         hlist_for_each_entry(sid, head, node) {
 660                 if (sid->id == id)
 661                         return sid->evsel;
 662         }
 663         return NULL;
 664 }
 665
 666 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 667 {
 668         int i;
 669
 670         if (!evlist->overwrite_mmap)
 671                 return 0;
 672
 673         for (i = 0; i < evlist->nr_mmaps; i++) {
 674                 int fd = evlist->overwrite_mmap[i].fd;
 675                 int err;
 676
 677                 if (fd < 0)
 678                         continue;
 679                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 680                 if (err)
 681                         return err;
 682         }
 683         return 0;
 684 }
 685
 686 static int perf_evlist__pause(struct perf_evlist *evlist)
 687 {
 688         return perf_evlist__set_paused(evlist, true);
 689 }
 690
 691 static int perf_evlist__resume(struct perf_evlist *evlist)
 692 {
 693         return perf_evlist__set_paused(evlist, false);
 694 }
 695
 696 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 697 {
 698         int i;
 699
 700         if (evlist->mmap)
 701                 for (i = 0; i < evlist->nr_mmaps; i++)
 702                         perf_mmap__munmap(&evlist->mmap[i]);
 703
 704         if (evlist->overwrite_mmap)
 705                 for (i = 0; i < evlist->nr_mmaps; i++)
 706                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 707 }
 708
 709 void perf_evlist__munmap(struct perf_evlist *evlist)
 710 {
 711         perf_evlist__munmap_nofree(evlist);
 712         zfree(&evlist->mmap);
 713         zfree(&evlist->overwrite_mmap);
 714 }
 715
 716 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
 717                                                  bool overwrite)
 718 {
 719         int i;
 720         struct perf_mmap *map;
 721
 722         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 723         if (cpu_map__empty(evlist->cpus))
 724                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 725         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 726         if (!map)
 727                 return NULL;
 728
 729         for (i = 0; i < evlist->nr_mmaps; i++) {
 730                 map[i].fd = -1;
 731                 map[i].overwrite = overwrite;
 732                 /*
 733                  * When the perf_mmap() call is made we grab one refcount, plus
 734                  * one extra to let perf_mmap__consume() get the last
 735                  * events after all real references (perf_mmap__get()) are
 736                  * dropped.
 737                  *
 738                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 739                  * thus does perf_mmap__get() on it.
 740                  */
 741                 refcount_set(&map[i].refcnt, 0);
 742         }
 743         return map;
 744 }
 745
 746 static bool
 747 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 748                          struct perf_evsel *evsel)
 749 {
 750         if (evsel->attr.write_backward)
 751                 return false;
 752         return true;
 753 }
 754
 755 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 756                                        struct mmap_params *mp, int cpu_idx,
 757                                        int thread, int *_output, int *_output_overwrite)
 758 {
 759         struct perf_evsel *evsel;
 760         int revent;
 761         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 762
 763         evlist__for_each_entry(evlist, evsel) {
 764                 struct perf_mmap *maps = evlist->mmap;
 765                 int *output = _output;
 766                 int fd;
 767                 int cpu;
 768
 769                 mp->prot = PROT_READ | PROT_WRITE;
 770                 if (evsel->attr.write_backward) {
 771                         output = _output_overwrite;
 772                         maps = evlist->overwrite_mmap;
 773
 774                         if (!maps) {
 775                                 maps = perf_evlist__alloc_mmap(evlist, true);
 776                                 if (!maps)
 777                                         return -1;
 778                                 evlist->overwrite_mmap = maps;
 779                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 780                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 781                         }
 782                         mp->prot &= ~PROT_WRITE;
 783                 }
 784
 785                 if (evsel->system_wide && thread)
 786                         continue;
 787
 788                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 789                 if (cpu == -1)
 790                         continue;
 791
 792                 fd = FD(evsel, cpu, thread);
 793
 794                 if (*output == -1) {
 795                         *output = fd;
 796
 797                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 798                                 return -1;
 799                 } else {
 800                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 801                                 return -1;
 802
 803                         perf_mmap__get(&maps[idx]);
 804                 }
 805
 806                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 807
 808                 /*
 809                  * The system_wide flag causes a selected event to be opened
 810                  * always without a pid.  Consequently it will never get a
 811                  * POLLHUP, but it is used for tracking in combination with
 812                  * other events, so it should not need to be polled anyway.
 813                  * Therefore don't add it for polling.
 814                  */
 815                 if (!evsel->system_wide &&
 816                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 817                         perf_mmap__put(&maps[idx]);
 818                         return -1;
 819                 }
 820
 821                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 822                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 823                                                    fd) < 0)
 824                                 return -1;
 825                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 826                                                  thread);
 827                 }
 828         }
 829
 830         return 0;
 831 }
 832
 833 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 834                                      struct mmap_params *mp)
 835 {
 836         int cpu, thread;
 837         int nr_cpus = cpu_map__nr(evlist->cpus);
 838         int nr_threads = thread_map__nr(evlist->threads);
 839
 840         pr_debug2("perf event ring buffer mmapped per cpu\n");
 841         for (cpu = 0; cpu < nr_cpus; cpu++) {
 842                 int output = -1;
 843                 int output_overwrite = -1;
 844
 845                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 846                                               true);
 847
 848                 for (thread = 0; thread < nr_threads; thread++) {
 849                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 850                                                         thread, &output, &output_overwrite))
 851                                 goto out_unmap;
 852                 }
 853         }
 854
 855         return 0;
 856
 857 out_unmap:
 858         perf_evlist__munmap_nofree(evlist);
 859         return -1;
 860 }
 861
 862 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 863                                         struct mmap_params *mp)
 864 {
 865         int thread;
 866         int nr_threads = thread_map__nr(evlist->threads);
 867
 868         pr_debug2("perf event ring buffer mmapped per thread\n");
 869         for (thread = 0; thread < nr_threads; thread++) {
 870                 int output = -1;
 871                 int output_overwrite = -1;
 872
 873                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 874                                               false);
 875
 876                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 877                                                 &output, &output_overwrite))
 878                         goto out_unmap;
 879         }
 880
 881         return 0;
 882
 883 out_unmap:
 884         perf_evlist__munmap_nofree(evlist);
 885         return -1;
 886 }
 887
 888 unsigned long perf_event_mlock_kb_in_pages(void)
 889 {
 890         unsigned long pages;
 891         int max;
 892
 893         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 894                 /*
 895                  * Pick a once upon a time good value, i.e. things look
 896                  * strange since we can't read a sysctl value, but lets not
 897                  * die yet...
 898                  */
 899                 max = 512;
 900         } else {
 901                 max -= (page_size / 1024);
 902         }
 903
 904         pages = (max * 1024) / page_size;
 905         if (!is_power_of_2(pages))
 906                 pages = rounddown_pow_of_two(pages);
 907
 908         return pages;
 909 }
 910
 911 size_t perf_evlist__mmap_size(unsigned long pages)
 912 {
 913         if (pages == UINT_MAX)
 914                 pages = perf_event_mlock_kb_in_pages();
 915         else if (!is_power_of_2(pages))
 916                 return 0;
 917
 918         return (pages + 1) * page_size;
 919 }
 920
 921 static long parse_pages_arg(const char *str, unsigned long min,
 922                             unsigned long max)
 923 {
 924         unsigned long pages, val;
 925         static struct parse_tag tags[] = {
 926                 { .tag  = 'B', .mult = 1       },
 927                 { .tag  = 'K', .mult = 1 << 10 },
 928                 { .tag  = 'M', .mult = 1 << 20 },
 929                 { .tag  = 'G', .mult = 1 << 30 },
 930                 { .tag  = 0 },
 931         };
 932
 933         if (str == NULL)
 934                 return -EINVAL;
 935
 936         val = parse_tag_value(str, tags);
 937         if (val != (unsigned long) -1) {
 938                 /* we got file size value */
 939                 pages = PERF_ALIGN(val, page_size) / page_size;
 940         } else {
 941                 /* we got pages count value */
 942                 char *eptr;
 943                 pages = strtoul(str, &eptr, 10);
 944                 if (*eptr != '\0')
 945                         return -EINVAL;
 946         }
 947
 948         if (pages == 0 && min == 0) {
 949                 /* leave number of pages at 0 */
 950         } else if (!is_power_of_2(pages)) {
 951                 char buf[100];
 952
 953                 /* round pages up to next power of 2 */
 954                 pages = roundup_pow_of_two(pages);
 955                 if (!pages)
 956                         return -EINVAL;
 957
 958                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 959                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 960                         buf, pages);
 961         }
 962
 963         if (pages > max)
 964                 return -EINVAL;
 965
 966         return pages;
 967 }
 968
 969 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 970 {
 971         unsigned long max = UINT_MAX;
 972         long pages;
 973
 974         if (max > SIZE_MAX / page_size)
 975                 max = SIZE_MAX / page_size;
 976
 977         pages = parse_pages_arg(str, 1, max);
 978         if (pages < 0) {
 979                 pr_err("Invalid argument for --mmap_pages/-m\n");
 980                 return -1;
 981         }
 982
 983         *mmap_pages = pages;
 984         return 0;
 985 }
 986
 987 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 988                                   int unset __maybe_unused)
 989 {
 990         return __perf_evlist__parse_mmap_pages(opt->value, str);
 991 }
 992
 993 /**
 994  * perf_evlist__mmap_ex - Create mmaps to receive events.
 995  * @evlist: list of events
 996  * @pages: map length in pages
 997  * @overwrite: overwrite older events?
 998  * @auxtrace_pages - auxtrace map length in pages
 999  * @auxtrace_overwrite - overwrite older auxtrace data?
1000  *
1001  * If @overwrite is %false the user needs to signal event consumption using
1002  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1003  * automatically.
1004  *
1005  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1006  * consumption using auxtrace_mmap__write_tail().
1007  *
1008  * Return: %0 on success, negative error code otherwise.
1009  */
1010 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1011                          unsigned int auxtrace_pages,
1012                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
1013                          int comp_level)
1014 {
1015         struct perf_evsel *evsel;
1016         const struct cpu_map *cpus = evlist->cpus;
1017         const struct thread_map *threads = evlist->threads;
1018         /*
1019          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1020          * Its value is decided by evsel's write_backward.
1021          * So &mp should not be passed through const pointer.
1022          */
1023         struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
1024                                   .comp_level = comp_level };
1025
1026         if (!evlist->mmap)
1027                 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1028         if (!evlist->mmap)
1029                 return -ENOMEM;
1030
1031         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1032                 return -ENOMEM;
1033
1034         evlist->mmap_len = perf_evlist__mmap_size(pages);
1035         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1036         mp.mask = evlist->mmap_len - page_size - 1;
1037
1038         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1039                                    auxtrace_pages, auxtrace_overwrite);
1040
1041         evlist__for_each_entry(evlist, evsel) {
1042                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1043                     evsel->sample_id == NULL &&
1044                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1045                         return -ENOMEM;
1046         }
1047
1048         if (cpu_map__empty(cpus))
1049                 return perf_evlist__mmap_per_thread(evlist, &mp);
1050
1051         return perf_evlist__mmap_per_cpu(evlist, &mp);
1052 }
1053
1054 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1055 {
1056         return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1057 }
1058
1059 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1060 {
1061         bool all_threads = (target->per_thread && target->system_wide);
1062         struct cpu_map *cpus;
1063         struct thread_map *threads;
1064
1065         /*
1066          * If specify '-a' and '--per-thread' to perf record, perf record
1067          * will override '--per-thread'. target->per_thread = false and
1068          * target->system_wide = true.
1069          *
1070          * If specify '--per-thread' only to perf record,
1071          * target->per_thread = true and target->system_wide = false.
1072          *
1073          * So target->per_thread && target->system_wide is false.
1074          * For perf record, thread_map__new_str doesn't call
1075          * thread_map__new_all_cpus. That will keep perf record's
1076          * current behavior.
1077          *
1078          * For perf stat, it allows the case that target->per_thread and
1079          * target->system_wide are all true. It means to collect system-wide
1080          * per-thread data. thread_map__new_str will call
1081          * thread_map__new_all_cpus to enumerate all threads.
1082          */
1083         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1084                                       all_threads);
1085
1086         if (!threads)
1087                 return -1;
1088
1089         if (target__uses_dummy_map(target))
1090                 cpus = cpu_map__dummy_new();
1091         else
1092                 cpus = cpu_map__new(target->cpu_list);
1093
1094         if (!cpus)
1095                 goto out_delete_threads;
1096
1097         evlist->has_user_cpus = !!target->cpu_list;
1098
1099         perf_evlist__set_maps(evlist, cpus, threads);
1100
1101         return 0;
1102
1103 out_delete_threads:
1104         thread_map__put(threads);
1105         return -1;
1106 }
1107
1108 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1109                            struct thread_map *threads)
1110 {
1111         /*
1112          * Allow for the possibility that one or another of the maps isn't being
1113          * changed i.e. don't put it.  Note we are assuming the maps that are
1114          * being applied are brand new and evlist is taking ownership of the
1115          * original reference count of 1.  If that is not the case it is up to
1116          * the caller to increase the reference count.
1117          */
1118         if (cpus != evlist->cpus) {
1119                 cpu_map__put(evlist->cpus);
1120                 evlist->cpus = cpu_map__get(cpus);
1121         }
1122
1123         if (threads != evlist->threads) {
1124                 thread_map__put(evlist->threads);
1125                 evlist->threads = thread_map__get(threads);
1126         }
1127
1128         perf_evlist__propagate_maps(evlist);
1129 }
1130
1131 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1132                                    enum perf_event_sample_format bit)
1133 {
1134         struct perf_evsel *evsel;
1135
1136         evlist__for_each_entry(evlist, evsel)
1137                 __perf_evsel__set_sample_bit(evsel, bit);
1138 }
1139
1140 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1141                                      enum perf_event_sample_format bit)
1142 {
1143         struct perf_evsel *evsel;
1144
1145         evlist__for_each_entry(evlist, evsel)
1146                 __perf_evsel__reset_sample_bit(evsel, bit);
1147 }
1148
1149 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1150 {
1151         struct perf_evsel *evsel;
1152         int err = 0;
1153
1154         evlist__for_each_entry(evlist, evsel) {
1155                 if (evsel->filter == NULL)
1156                         continue;
1157
1158                 /*
1159                  * filters only work for tracepoint event, which doesn't have cpu limit.
1160                  * So evlist and evsel should always be same.
1161                  */
1162                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1163                 if (err) {
1164                         *err_evsel = evsel;
1165                         break;
1166                 }
1167         }
1168
1169         return err;
1170 }
1171
1172 int perf_evlist__set_tp_filter(struct perf_evlist *evlist, const char *filter)
1173 {
1174         struct perf_evsel *evsel;
1175         int err = 0;
1176
1177         evlist__for_each_entry(evlist, evsel) {
1178                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1179                         continue;
1180
1181                 err = perf_evsel__set_filter(evsel, filter);
1182                 if (err)
1183                         break;
1184         }
1185
1186         return err;
1187 }
1188
1189 int perf_evlist__set_tp_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1190 {
1191         char *filter;
1192         int ret = -1;
1193         size_t i;
1194
1195         for (i = 0; i < npids; ++i) {
1196                 if (i == 0) {
1197                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1198                                 return -1;
1199                 } else {
1200                         char *tmp;
1201
1202                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1203                                 goto out_free;
1204
1205                         free(filter);
1206                         filter = tmp;
1207                 }
1208         }
1209
1210         ret = perf_evlist__set_tp_filter(evlist, filter);
1211 out_free:
1212         free(filter);
1213         return ret;
1214 }
1215
1216 int perf_evlist__set_tp_filter_pid(struct perf_evlist *evlist, pid_t pid)
1217 {
1218         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1219 }
1220
1221 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1222 {
1223         struct perf_evsel *pos;
1224
1225         if (evlist->nr_entries == 1)
1226                 return true;
1227
1228         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1229                 return false;
1230
1231         evlist__for_each_entry(evlist, pos) {
1232                 if (pos->id_pos != evlist->id_pos ||
1233                     pos->is_pos != evlist->is_pos)
1234                         return false;
1235         }
1236
1237         return true;
1238 }
1239
1240 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1241 {
1242         struct perf_evsel *evsel;
1243
1244         if (evlist->combined_sample_type)
1245                 return evlist->combined_sample_type;
1246
1247         evlist__for_each_entry(evlist, evsel)
1248                 evlist->combined_sample_type |= evsel->attr.sample_type;
1249
1250         return evlist->combined_sample_type;
1251 }
1252
1253 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1254 {
1255         evlist->combined_sample_type = 0;
1256         return __perf_evlist__combined_sample_type(evlist);
1257 }
1258
1259 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1260 {
1261         struct perf_evsel *evsel;
1262         u64 branch_type = 0;
1263
1264         evlist__for_each_entry(evlist, evsel)
1265                 branch_type |= evsel->attr.branch_sample_type;
1266         return branch_type;
1267 }
1268
1269 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1270 {
1271         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1272         u64 read_format = first->attr.read_format;
1273         u64 sample_type = first->attr.sample_type;
1274
1275         evlist__for_each_entry(evlist, pos) {
1276                 if (read_format != pos->attr.read_format)
1277                         return false;
1278         }
1279
1280         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1281         if ((sample_type & PERF_SAMPLE_READ) &&
1282             !(read_format & PERF_FORMAT_ID)) {
1283                 return false;
1284         }
1285
1286         return true;
1287 }
1288
1289 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1290 {
1291         struct perf_evsel *first = perf_evlist__first(evlist);
1292         return first->attr.read_format;
1293 }
1294
1295 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1296 {
1297         struct perf_evsel *first = perf_evlist__first(evlist);
1298         struct perf_sample *data;
1299         u64 sample_type;
1300         u16 size = 0;
1301
1302         if (!first->attr.sample_id_all)
1303                 goto out;
1304
1305         sample_type = first->attr.sample_type;
1306
1307         if (sample_type & PERF_SAMPLE_TID)
1308                 size += sizeof(data->tid) * 2;
1309
1310        if (sample_type & PERF_SAMPLE_TIME)
1311                 size += sizeof(data->time);
1312
1313         if (sample_type & PERF_SAMPLE_ID)
1314                 size += sizeof(data->id);
1315
1316         if (sample_type & PERF_SAMPLE_STREAM_ID)
1317                 size += sizeof(data->stream_id);
1318
1319         if (sample_type & PERF_SAMPLE_CPU)
1320                 size += sizeof(data->cpu) * 2;
1321
1322         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1323                 size += sizeof(data->id);
1324 out:
1325         return size;
1326 }
1327
1328 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1329 {
1330         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1331
1332         evlist__for_each_entry_continue(evlist, pos) {
1333                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1334                         return false;
1335         }
1336
1337         return true;
1338 }
1339
1340 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1341 {
1342         struct perf_evsel *first = perf_evlist__first(evlist);
1343         return first->attr.sample_id_all;
1344 }
1345
1346 void perf_evlist__set_selected(struct perf_evlist *evlist,
1347                                struct perf_evsel *evsel)
1348 {
1349         evlist->selected = evsel;
1350 }
1351
1352 void perf_evlist__close(struct perf_evlist *evlist)
1353 {
1354         struct perf_evsel *evsel;
1355
1356         evlist__for_each_entry_reverse(evlist, evsel)
1357                 perf_evsel__close(evsel);
1358 }
1359
1360 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1361 {
1362         struct cpu_map    *cpus;
1363         struct thread_map *threads;
1364         int err = -ENOMEM;
1365
1366         /*
1367          * Try reading /sys/devices/system/cpu/online to get
1368          * an all cpus map.
1369          *
1370          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1371          * code needs an overhaul to properly forward the
1372          * error, and we may not want to do that fallback to a
1373          * default cpu identity map :-\
1374          */
1375         cpus = cpu_map__new(NULL);
1376         if (!cpus)
1377                 goto out;
1378
1379         threads = thread_map__new_dummy();
1380         if (!threads)
1381                 goto out_put;
1382
1383         perf_evlist__set_maps(evlist, cpus, threads);
1384 out:
1385         return err;
1386 out_put:
1387         cpu_map__put(cpus);
1388         goto out;
1389 }
1390
1391 int perf_evlist__open(struct perf_evlist *evlist)
1392 {
1393         struct perf_evsel *evsel;
1394         int err;
1395
1396         /*
1397          * Default: one fd per CPU, all threads, aka systemwide
1398          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1399          */
1400         if (evlist->threads == NULL && evlist->cpus == NULL) {
1401                 err = perf_evlist__create_syswide_maps(evlist);
1402                 if (err < 0)
1403                         goto out_err;
1404         }
1405
1406         perf_evlist__update_id_pos(evlist);
1407
1408         evlist__for_each_entry(evlist, evsel) {
1409                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1410                 if (err < 0)
1411                         goto out_err;
1412         }
1413
1414         return 0;
1415 out_err:
1416         perf_evlist__close(evlist);
1417         errno = -err;
1418         return err;
1419 }
1420
1421 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1422                                   const char *argv[], bool pipe_output,
1423                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1424 {
1425         int child_ready_pipe[2], go_pipe[2];
1426         char bf;
1427
1428         if (pipe(child_ready_pipe) < 0) {
1429                 perror("failed to create 'ready' pipe");
1430                 return -1;
1431         }
1432
1433         if (pipe(go_pipe) < 0) {
1434                 perror("failed to create 'go' pipe");
1435                 goto out_close_ready_pipe;
1436         }
1437
1438         evlist->workload.pid = fork();
1439         if (evlist->workload.pid < 0) {
1440                 perror("failed to fork");
1441                 goto out_close_pipes;
1442         }
1443
1444         if (!evlist->workload.pid) {
1445                 int ret;
1446
1447                 if (pipe_output)
1448                         dup2(2, 1);
1449
1450                 signal(SIGTERM, SIG_DFL);
1451
1452                 close(child_ready_pipe[0]);
1453                 close(go_pipe[1]);
1454                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1455
1456                 /*
1457                  * Tell the parent we're ready to go
1458                  */
1459                 close(child_ready_pipe[1]);
1460
1461                 /*
1462                  * Wait until the parent tells us to go.
1463                  */
1464                 ret = read(go_pipe[0], &bf, 1);
1465                 /*
1466                  * The parent will ask for the execvp() to be performed by
1467                  * writing exactly one byte, in workload.cork_fd, usually via
1468                  * perf_evlist__start_workload().
1469                  *
1470                  * For cancelling the workload without actually running it,
1471                  * the parent will just close workload.cork_fd, without writing
1472                  * anything, i.e. read will return zero and we just exit()
1473                  * here.
1474                  */
1475                 if (ret != 1) {
1476                         if (ret == -1)
1477                                 perror("unable to read pipe");
1478                         exit(ret);
1479                 }
1480
1481                 execvp(argv[0], (char **)argv);
1482
1483                 if (exec_error) {
1484                         union sigval val;
1485
1486                         val.sival_int = errno;
1487                         if (sigqueue(getppid(), SIGUSR1, val))
1488                                 perror(argv[0]);
1489                 } else
1490                         perror(argv[0]);
1491                 exit(-1);
1492         }
1493
1494         if (exec_error) {
1495                 struct sigaction act = {
1496                         .sa_flags     = SA_SIGINFO,
1497                         .sa_sigaction = exec_error,
1498                 };
1499                 sigaction(SIGUSR1, &act, NULL);
1500         }
1501
1502         if (target__none(target)) {
1503                 if (evlist->threads == NULL) {
1504                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1505                                 __func__, __LINE__);
1506                         goto out_close_pipes;
1507                 }
1508                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1509         }
1510
1511         close(child_ready_pipe[1]);
1512         close(go_pipe[0]);
1513         /*
1514          * wait for child to settle
1515          */
1516         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1517                 perror("unable to read pipe");
1518                 goto out_close_pipes;
1519         }
1520
1521         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1522         evlist->workload.cork_fd = go_pipe[1];
1523         close(child_ready_pipe[0]);
1524         return 0;
1525
1526 out_close_pipes:
1527         close(go_pipe[0]);
1528         close(go_pipe[1]);
1529 out_close_ready_pipe:
1530         close(child_ready_pipe[0]);
1531         close(child_ready_pipe[1]);
1532         return -1;
1533 }
1534
1535 int perf_evlist__start_workload(struct perf_evlist *evlist)
1536 {
1537         if (evlist->workload.cork_fd > 0) {
1538                 char bf = 0;
1539                 int ret;
1540                 /*
1541                  * Remove the cork, let it rip!
1542                  */
1543                 ret = write(evlist->workload.cork_fd, &bf, 1);
1544                 if (ret < 0)
1545                         perror("unable to write to pipe");
1546
1547                 close(evlist->workload.cork_fd);
1548                 return ret;
1549         }
1550
1551         return 0;
1552 }
1553
1554 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1555                               struct perf_sample *sample)
1556 {
1557         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1558
1559         if (!evsel)
1560                 return -EFAULT;
1561         return perf_evsel__parse_sample(evsel, event, sample);
1562 }
1563
1564 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1565                                         union perf_event *event,
1566                                         u64 *timestamp)
1567 {
1568         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1569
1570         if (!evsel)
1571                 return -EFAULT;
1572         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1573 }
1574
1575 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1576 {
1577         struct perf_evsel *evsel;
1578         size_t printed = 0;
1579
1580         evlist__for_each_entry(evlist, evsel) {
1581                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1582                                    perf_evsel__name(evsel));
1583         }
1584
1585         return printed + fprintf(fp, "\n");
1586 }
1587
1588 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1589                                int err, char *buf, size_t size)
1590 {
1591         int printed, value;
1592         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1593
1594         switch (err) {
1595         case EACCES:
1596         case EPERM:
1597                 printed = scnprintf(buf, size,
1598                                     "Error:\t%s.\n"
1599                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1600
1601                 value = perf_event_paranoid();
1602
1603                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1604
1605                 if (value >= 2) {
1606                         printed += scnprintf(buf + printed, size - printed,
1607                                              "For your workloads it needs to be <= 1\nHint:\t");
1608                 }
1609                 printed += scnprintf(buf + printed, size - printed,
1610                                      "For system wide tracing it needs to be set to -1.\n");
1611
1612                 printed += scnprintf(buf + printed, size - printed,
1613                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1614                                     "Hint:\tThe current value is %d.", value);
1615                 break;
1616         case EINVAL: {
1617                 struct perf_evsel *first = perf_evlist__first(evlist);
1618                 int max_freq;
1619
1620                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1621                         goto out_default;
1622
1623                 if (first->attr.sample_freq < (u64)max_freq)
1624                         goto out_default;
1625
1626                 printed = scnprintf(buf, size,
1627                                     "Error:\t%s.\n"
1628                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1629                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1630                                     emsg, max_freq, first->attr.sample_freq);
1631                 break;
1632         }
1633         default:
1634 out_default:
1635                 scnprintf(buf, size, "%s", emsg);
1636                 break;
1637         }
1638
1639         return 0;
1640 }
1641
1642 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1643 {
1644         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1645         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1646
1647         switch (err) {
1648         case EPERM:
1649                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1650                 printed += scnprintf(buf + printed, size - printed,
1651                                      "Error:\t%s.\n"
1652                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1653                                      "Hint:\tTried using %zd kB.\n",
1654                                      emsg, pages_max_per_user, pages_attempted);
1655
1656                 if (pages_attempted >= pages_max_per_user) {
1657                         printed += scnprintf(buf + printed, size - printed,
1658                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1659                                              pages_max_per_user + pages_attempted);
1660                 }
1661
1662                 printed += scnprintf(buf + printed, size - printed,
1663                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1664                 break;
1665         default:
1666                 scnprintf(buf, size, "%s", emsg);
1667                 break;
1668         }
1669
1670         return 0;
1671 }
1672
1673 void perf_evlist__to_front(struct perf_evlist *evlist,
1674                            struct perf_evsel *move_evsel)
1675 {
1676         struct perf_evsel *evsel, *n;
1677         LIST_HEAD(move);
1678
1679         if (move_evsel == perf_evlist__first(evlist))
1680                 return;
1681
1682         evlist__for_each_entry_safe(evlist, n, evsel) {
1683                 if (evsel->leader == move_evsel->leader)
1684                         list_move_tail(&evsel->node, &move);
1685         }
1686
1687         list_splice(&move, &evlist->entries);
1688 }
1689
1690 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1691                                      struct perf_evsel *tracking_evsel)
1692 {
1693         struct perf_evsel *evsel;
1694
1695         if (tracking_evsel->tracking)
1696                 return;
1697
1698         evlist__for_each_entry(evlist, evsel) {
1699                 if (evsel != tracking_evsel)
1700                         evsel->tracking = false;
1701         }
1702
1703         tracking_evsel->tracking = true;
1704 }
1705
1706 struct perf_evsel *
1707 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1708                                const char *str)
1709 {
1710         struct perf_evsel *evsel;
1711
1712         evlist__for_each_entry(evlist, evsel) {
1713                 if (!evsel->name)
1714                         continue;
1715                 if (strcmp(str, evsel->name) == 0)
1716                         return evsel;
1717         }
1718
1719         return NULL;
1720 }
1721
1722 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1723                                   enum bkw_mmap_state state)
1724 {
1725         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1726         enum action {
1727                 NONE,
1728                 PAUSE,
1729                 RESUME,
1730         } action = NONE;
1731
1732         if (!evlist->overwrite_mmap)
1733                 return;
1734
1735         switch (old_state) {
1736         case BKW_MMAP_NOTREADY: {
1737                 if (state != BKW_MMAP_RUNNING)
1738                         goto state_err;
1739                 break;
1740         }
1741         case BKW_MMAP_RUNNING: {
1742                 if (state != BKW_MMAP_DATA_PENDING)
1743                         goto state_err;
1744                 action = PAUSE;
1745                 break;
1746         }
1747         case BKW_MMAP_DATA_PENDING: {
1748                 if (state != BKW_MMAP_EMPTY)
1749                         goto state_err;
1750                 break;
1751         }
1752         case BKW_MMAP_EMPTY: {
1753                 if (state != BKW_MMAP_RUNNING)
1754                         goto state_err;
1755                 action = RESUME;
1756                 break;
1757         }
1758         default:
1759                 WARN_ONCE(1, "Shouldn't get there\n");
1760         }
1761
1762         evlist->bkw_mmap_state = state;
1763
1764         switch (action) {
1765         case PAUSE:
1766                 perf_evlist__pause(evlist);
1767                 break;
1768         case RESUME:
1769                 perf_evlist__resume(evlist);
1770                 break;
1771         case NONE:
1772         default:
1773                 break;
1774         }
1775
1776 state_err:
1777         return;
1778 }
1779
1780 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1781 {
1782         struct perf_evsel *evsel;
1783
1784         evlist__for_each_entry(evlist, evsel) {
1785                 if (!evsel->attr.exclude_kernel)
1786                         return false;
1787         }
1788
1789         return true;
1790 }
1791
1792 /*
1793  * Events in data file are not collect in groups, but we still want
1794  * the group display. Set the artificial group and set the leader's
1795  * forced_leader flag to notify the display code.
1796  */
1797 void perf_evlist__force_leader(struct perf_evlist *evlist)
1798 {
1799         if (!evlist->nr_groups) {
1800                 struct perf_evsel *leader = perf_evlist__first(evlist);
1801
1802                 perf_evlist__set_leader(evlist);
1803                 leader->forced_leader = true;
1804         }
1805 }
1806
1807 struct perf_evsel *perf_evlist__reset_weak_group(struct perf_evlist *evsel_list,
1808                                                  struct perf_evsel *evsel)
1809 {
1810         struct perf_evsel *c2, *leader;
1811         bool is_open = true;
1812
1813         leader = evsel->leader;
1814         pr_debug("Weak group for %s/%d failed\n",
1815                         leader->name, leader->nr_members);
1816
1817         /*
1818          * for_each_group_member doesn't work here because it doesn't
1819          * include the first entry.
1820          */
1821         evlist__for_each_entry(evsel_list, c2) {
1822                 if (c2 == evsel)
1823                         is_open = false;
1824                 if (c2->leader == leader) {
1825                         if (is_open)
1826                                 perf_evsel__close(c2);
1827                         c2->leader = c2;
1828                         c2->nr_members = 0;
1829                 }
1830         }
1831         return leader;
1832 }
1833
1834 int perf_evlist__add_sb_event(struct perf_evlist **evlist,
1835                               struct perf_event_attr *attr,
1836                               perf_evsel__sb_cb_t cb,
1837                               void *data)
1838 {
1839         struct perf_evsel *evsel;
1840         bool new_evlist = (*evlist) == NULL;
1841
1842         if (*evlist == NULL)
1843                 *evlist = perf_evlist__new();
1844         if (*evlist == NULL)
1845                 return -1;
1846
1847         if (!attr->sample_id_all) {
1848                 pr_warning("enabling sample_id_all for all side band events\n");
1849                 attr->sample_id_all = 1;
1850         }
1851
1852         evsel = perf_evsel__new_idx(attr, (*evlist)->nr_entries);
1853         if (!evsel)
1854                 goto out_err;
1855
1856         evsel->side_band.cb = cb;
1857         evsel->side_band.data = data;
1858         perf_evlist__add(*evlist, evsel);
1859         return 0;
1860
1861 out_err:
1862         if (new_evlist) {
1863                 perf_evlist__delete(*evlist);
1864                 *evlist = NULL;
1865         }
1866         return -1;
1867 }
1868
1869 static void *perf_evlist__poll_thread(void *arg)
1870 {
1871         struct perf_evlist *evlist = arg;
1872         bool draining = false;
1873         int i, done = 0;
1874
1875         while (!done) {
1876                 bool got_data = false;
1877
1878                 if (evlist->thread.done)
1879                         draining = true;
1880
1881                 if (!draining)
1882                         perf_evlist__poll(evlist, 1000);
1883
1884                 for (i = 0; i < evlist->nr_mmaps; i++) {
1885                         struct perf_mmap *map = &evlist->mmap[i];
1886                         union perf_event *event;
1887
1888                         if (perf_mmap__read_init(map))
1889                                 continue;
1890                         while ((event = perf_mmap__read_event(map)) != NULL) {
1891                                 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1892
1893                                 if (evsel && evsel->side_band.cb)
1894                                         evsel->side_band.cb(event, evsel->side_band.data);
1895                                 else
1896                                         pr_warning("cannot locate proper evsel for the side band event\n");
1897
1898                                 perf_mmap__consume(map);
1899                                 got_data = true;
1900                         }
1901                         perf_mmap__read_done(map);
1902                 }
1903
1904                 if (draining && !got_data)
1905                         break;
1906         }
1907         return NULL;
1908 }
1909
1910 int perf_evlist__start_sb_thread(struct perf_evlist *evlist,
1911                                  struct target *target)
1912 {
1913         struct perf_evsel *counter;
1914
1915         if (!evlist)
1916                 return 0;
1917
1918         if (perf_evlist__create_maps(evlist, target))
1919                 goto out_delete_evlist;
1920
1921         evlist__for_each_entry(evlist, counter) {
1922                 if (perf_evsel__open(counter, evlist->cpus,
1923                                      evlist->threads) < 0)
1924                         goto out_delete_evlist;
1925         }
1926
1927         if (perf_evlist__mmap(evlist, UINT_MAX))
1928                 goto out_delete_evlist;
1929
1930         evlist__for_each_entry(evlist, counter) {
1931                 if (perf_evsel__enable(counter))
1932                         goto out_delete_evlist;
1933         }
1934
1935         evlist->thread.done = 0;
1936         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1937                 goto out_delete_evlist;
1938
1939         return 0;
1940
1941 out_delete_evlist:
1942         perf_evlist__delete(evlist);
1943         evlist = NULL;
1944         return -1;
1945 }
1946
1947 void perf_evlist__stop_sb_thread(struct perf_evlist *evlist)
1948 {
1949         if (!evlist)
1950                 return;
1951         evlist->thread.done = 1;
1952         pthread_join(evlist->thread.th, NULL);
1953         perf_evlist__delete(evlist);
1954 }