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 <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 <unistd.h>
  19
  20 #include "parse-events.h"
  21 #include "parse-options.h"
  22
  23 #include <sys/mman.h>
  24
  25 #include <linux/bitops.h>
  26 #include <linux/hash.h>
  27 #include <linux/log2.h>
  28
  29 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
  30 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
  31
  32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  33 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  34
  35 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  36                        struct thread_map *threads)
  37 {
  38         int i;
  39
  40         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  41                 INIT_HLIST_HEAD(&evlist->heads[i]);
  42         INIT_LIST_HEAD(&evlist->entries);
  43         perf_evlist__set_maps(evlist, cpus, threads);
  44         fdarray__init(&evlist->pollfd, 64);
  45         evlist->workload.pid = -1;
  46 }
  47
  48 struct perf_evlist *perf_evlist__new(void)
  49 {
  50         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  51
  52         if (evlist != NULL)
  53                 perf_evlist__init(evlist, NULL, NULL);
  54
  55         return evlist;
  56 }
  57
  58 struct perf_evlist *perf_evlist__new_default(void)
  59 {
  60         struct perf_evlist *evlist = perf_evlist__new();
  61
  62         if (evlist && perf_evlist__add_default(evlist)) {
  63                 perf_evlist__delete(evlist);
  64                 evlist = NULL;
  65         }
  66
  67         return evlist;
  68 }
  69
  70 /**
  71  * perf_evlist__set_id_pos - set the positions of event ids.
  72  * @evlist: selected event list
  73  *
  74  * Events with compatible sample types all have the same id_pos
  75  * and is_pos.  For convenience, put a copy on evlist.
  76  */
  77 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  78 {
  79         struct perf_evsel *first = perf_evlist__first(evlist);
  80
  81         evlist->id_pos = first->id_pos;
  82         evlist->is_pos = first->is_pos;
  83 }
  84
  85 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  86 {
  87         struct perf_evsel *evsel;
  88
  89         evlist__for_each(evlist, evsel)
  90                 perf_evsel__calc_id_pos(evsel);
  91
  92         perf_evlist__set_id_pos(evlist);
  93 }
  94
  95 static void perf_evlist__purge(struct perf_evlist *evlist)
  96 {
  97         struct perf_evsel *pos, *n;
  98
  99         evlist__for_each_safe(evlist, n, pos) {
 100                 list_del_init(&pos->node);
 101                 perf_evsel__delete(pos);
 102         }
 103
 104         evlist->nr_entries = 0;
 105 }
 106
 107 void perf_evlist__exit(struct perf_evlist *evlist)
 108 {
 109         zfree(&evlist->mmap);
 110         fdarray__exit(&evlist->pollfd);
 111 }
 112
 113 void perf_evlist__delete(struct perf_evlist *evlist)
 114 {
 115         perf_evlist__munmap(evlist);
 116         perf_evlist__close(evlist);
 117         cpu_map__put(evlist->cpus);
 118         thread_map__put(evlist->threads);
 119         evlist->cpus = NULL;
 120         evlist->threads = NULL;
 121         perf_evlist__purge(evlist);
 122         perf_evlist__exit(evlist);
 123         free(evlist);
 124 }
 125
 126 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 127 {
 128         list_add_tail(&entry->node, &evlist->entries);
 129         entry->idx = evlist->nr_entries;
 130         entry->tracking = !entry->idx;
 131
 132         if (!evlist->nr_entries++)
 133                 perf_evlist__set_id_pos(evlist);
 134 }
 135
 136 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 137                                    struct list_head *list,
 138                                    int nr_entries)
 139 {
 140         bool set_id_pos = !evlist->nr_entries;
 141
 142         list_splice_tail(list, &evlist->entries);
 143         evlist->nr_entries += nr_entries;
 144         if (set_id_pos)
 145                 perf_evlist__set_id_pos(evlist);
 146 }
 147
 148 void __perf_evlist__set_leader(struct list_head *list)
 149 {
 150         struct perf_evsel *evsel, *leader;
 151
 152         leader = list_entry(list->next, struct perf_evsel, node);
 153         evsel = list_entry(list->prev, struct perf_evsel, node);
 154
 155         leader->nr_members = evsel->idx - leader->idx + 1;
 156
 157         __evlist__for_each(list, evsel) {
 158                 evsel->leader = leader;
 159         }
 160 }
 161
 162 void perf_evlist__set_leader(struct perf_evlist *evlist)
 163 {
 164         if (evlist->nr_entries) {
 165                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 166                 __perf_evlist__set_leader(&evlist->entries);
 167         }
 168 }
 169
 170 int perf_evlist__add_default(struct perf_evlist *evlist)
 171 {
 172         struct perf_event_attr attr = {
 173                 .type = PERF_TYPE_HARDWARE,
 174                 .config = PERF_COUNT_HW_CPU_CYCLES,
 175         };
 176         struct perf_evsel *evsel;
 177
 178         event_attr_init(&attr);
 179
 180         evsel = perf_evsel__new(&attr);
 181         if (evsel == NULL)
 182                 goto error;
 183
 184         /* use strdup() because free(evsel) assumes name is allocated */
 185         evsel->name = strdup("cycles");
 186         if (!evsel->name)
 187                 goto error_free;
 188
 189         perf_evlist__add(evlist, evsel);
 190         return 0;
 191 error_free:
 192         perf_evsel__delete(evsel);
 193 error:
 194         return -ENOMEM;
 195 }
 196
 197 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 198                                   struct perf_event_attr *attrs, size_t nr_attrs)
 199 {
 200         struct perf_evsel *evsel, *n;
 201         LIST_HEAD(head);
 202         size_t i;
 203
 204         for (i = 0; i < nr_attrs; i++) {
 205                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 206                 if (evsel == NULL)
 207                         goto out_delete_partial_list;
 208                 list_add_tail(&evsel->node, &head);
 209         }
 210
 211         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 212
 213         return 0;
 214
 215 out_delete_partial_list:
 216         __evlist__for_each_safe(&head, n, evsel)
 217                 perf_evsel__delete(evsel);
 218         return -1;
 219 }
 220
 221 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 222                                      struct perf_event_attr *attrs, size_t nr_attrs)
 223 {
 224         size_t i;
 225
 226         for (i = 0; i < nr_attrs; i++)
 227                 event_attr_init(attrs + i);
 228
 229         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 230 }
 231
 232 struct perf_evsel *
 233 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 234 {
 235         struct perf_evsel *evsel;
 236
 237         evlist__for_each(evlist, evsel) {
 238                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 239                     (int)evsel->attr.config == id)
 240                         return evsel;
 241         }
 242
 243         return NULL;
 244 }
 245
 246 struct perf_evsel *
 247 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 248                                      const char *name)
 249 {
 250         struct perf_evsel *evsel;
 251
 252         evlist__for_each(evlist, evsel) {
 253                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 254                     (strcmp(evsel->name, name) == 0))
 255                         return evsel;
 256         }
 257
 258         return NULL;
 259 }
 260
 261 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 262                            const char *sys, const char *name, void *handler)
 263 {
 264         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 265
 266         if (evsel == NULL)
 267                 return -1;
 268
 269         evsel->handler = handler;
 270         perf_evlist__add(evlist, evsel);
 271         return 0;
 272 }
 273
 274 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 275                                    struct perf_evsel *evsel)
 276 {
 277         if (evsel->system_wide)
 278                 return 1;
 279         else
 280                 return thread_map__nr(evlist->threads);
 281 }
 282
 283 void perf_evlist__disable(struct perf_evlist *evlist)
 284 {
 285         int cpu, thread;
 286         struct perf_evsel *pos;
 287         int nr_cpus = cpu_map__nr(evlist->cpus);
 288         int nr_threads;
 289
 290         for (cpu = 0; cpu < nr_cpus; cpu++) {
 291                 evlist__for_each(evlist, pos) {
 292                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 293                                 continue;
 294                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 295                         for (thread = 0; thread < nr_threads; thread++)
 296                                 ioctl(FD(pos, cpu, thread),
 297                                       PERF_EVENT_IOC_DISABLE, 0);
 298                 }
 299         }
 300
 301         evlist->enabled = false;
 302 }
 303
 304 void perf_evlist__enable(struct perf_evlist *evlist)
 305 {
 306         int cpu, thread;
 307         struct perf_evsel *pos;
 308         int nr_cpus = cpu_map__nr(evlist->cpus);
 309         int nr_threads;
 310
 311         for (cpu = 0; cpu < nr_cpus; cpu++) {
 312                 evlist__for_each(evlist, pos) {
 313                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 314                                 continue;
 315                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 316                         for (thread = 0; thread < nr_threads; thread++)
 317                                 ioctl(FD(pos, cpu, thread),
 318                                       PERF_EVENT_IOC_ENABLE, 0);
 319                 }
 320         }
 321
 322         evlist->enabled = true;
 323 }
 324
 325 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 326 {
 327         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 328 }
 329
 330 int perf_evlist__disable_event(struct perf_evlist *evlist,
 331                                struct perf_evsel *evsel)
 332 {
 333         int cpu, thread, err;
 334         int nr_cpus = cpu_map__nr(evlist->cpus);
 335         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 336
 337         if (!evsel->fd)
 338                 return 0;
 339
 340         for (cpu = 0; cpu < nr_cpus; cpu++) {
 341                 for (thread = 0; thread < nr_threads; thread++) {
 342                         err = ioctl(FD(evsel, cpu, thread),
 343                                     PERF_EVENT_IOC_DISABLE, 0);
 344                         if (err)
 345                                 return err;
 346                 }
 347         }
 348         return 0;
 349 }
 350
 351 int perf_evlist__enable_event(struct perf_evlist *evlist,
 352                               struct perf_evsel *evsel)
 353 {
 354         int cpu, thread, err;
 355         int nr_cpus = cpu_map__nr(evlist->cpus);
 356         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 357
 358         if (!evsel->fd)
 359                 return -EINVAL;
 360
 361         for (cpu = 0; cpu < nr_cpus; cpu++) {
 362                 for (thread = 0; thread < nr_threads; thread++) {
 363                         err = ioctl(FD(evsel, cpu, thread),
 364                                     PERF_EVENT_IOC_ENABLE, 0);
 365                         if (err)
 366                                 return err;
 367                 }
 368         }
 369         return 0;
 370 }
 371
 372 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 373                                          struct perf_evsel *evsel, int cpu)
 374 {
 375         int thread, err;
 376         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 377
 378         if (!evsel->fd)
 379                 return -EINVAL;
 380
 381         for (thread = 0; thread < nr_threads; thread++) {
 382                 err = ioctl(FD(evsel, cpu, thread),
 383                             PERF_EVENT_IOC_ENABLE, 0);
 384                 if (err)
 385                         return err;
 386         }
 387         return 0;
 388 }
 389
 390 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 391                                             struct perf_evsel *evsel,
 392                                             int thread)
 393 {
 394         int cpu, err;
 395         int nr_cpus = cpu_map__nr(evlist->cpus);
 396
 397         if (!evsel->fd)
 398                 return -EINVAL;
 399
 400         for (cpu = 0; cpu < nr_cpus; cpu++) {
 401                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 402                 if (err)
 403                         return err;
 404         }
 405         return 0;
 406 }
 407
 408 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 409                                   struct perf_evsel *evsel, int idx)
 410 {
 411         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 412
 413         if (per_cpu_mmaps)
 414                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 415         else
 416                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 417 }
 418
 419 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 420 {
 421         int nr_cpus = cpu_map__nr(evlist->cpus);
 422         int nr_threads = thread_map__nr(evlist->threads);
 423         int nfds = 0;
 424         struct perf_evsel *evsel;
 425
 426         evlist__for_each(evlist, evsel) {
 427                 if (evsel->system_wide)
 428                         nfds += nr_cpus;
 429                 else
 430                         nfds += nr_cpus * nr_threads;
 431         }
 432
 433         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 434             fdarray__grow(&evlist->pollfd, nfds) < 0)
 435                 return -ENOMEM;
 436
 437         return 0;
 438 }
 439
 440 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
 441 {
 442         int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
 443         /*
 444          * Save the idx so that when we filter out fds POLLHUP'ed we can
 445          * close the associated evlist->mmap[] entry.
 446          */
 447         if (pos >= 0) {
 448                 evlist->pollfd.priv[pos].idx = idx;
 449
 450                 fcntl(fd, F_SETFL, O_NONBLOCK);
 451         }
 452
 453         return pos;
 454 }
 455
 456 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 457 {
 458         return __perf_evlist__add_pollfd(evlist, fd, -1);
 459 }
 460
 461 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
 462 {
 463         struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
 464
 465         perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
 466 }
 467
 468 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 469 {
 470         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 471                                perf_evlist__munmap_filtered);
 472 }
 473
 474 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 475 {
 476         return fdarray__poll(&evlist->pollfd, timeout);
 477 }
 478
 479 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 480                                  struct perf_evsel *evsel,
 481                                  int cpu, int thread, u64 id)
 482 {
 483         int hash;
 484         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 485
 486         sid->id = id;
 487         sid->evsel = evsel;
 488         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 489         hlist_add_head(&sid->node, &evlist->heads[hash]);
 490 }
 491
 492 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 493                          int cpu, int thread, u64 id)
 494 {
 495         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 496         evsel->id[evsel->ids++] = id;
 497 }
 498
 499 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 500                                   struct perf_evsel *evsel,
 501                                   int cpu, int thread, int fd)
 502 {
 503         u64 read_data[4] = { 0, };
 504         int id_idx = 1; /* The first entry is the counter value */
 505         u64 id;
 506         int ret;
 507
 508         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 509         if (!ret)
 510                 goto add;
 511
 512         if (errno != ENOTTY)
 513                 return -1;
 514
 515         /* Legacy way to get event id.. All hail to old kernels! */
 516
 517         /*
 518          * This way does not work with group format read, so bail
 519          * out in that case.
 520          */
 521         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 522                 return -1;
 523
 524         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 525             read(fd, &read_data, sizeof(read_data)) == -1)
 526                 return -1;
 527
 528         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 529                 ++id_idx;
 530         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 531                 ++id_idx;
 532
 533         id = read_data[id_idx];
 534
 535  add:
 536         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 537         return 0;
 538 }
 539
 540 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 541                                      struct perf_evsel *evsel, int idx, int cpu,
 542                                      int thread)
 543 {
 544         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 545         sid->idx = idx;
 546         if (evlist->cpus && cpu >= 0)
 547                 sid->cpu = evlist->cpus->map[cpu];
 548         else
 549                 sid->cpu = -1;
 550         if (!evsel->system_wide && evlist->threads && thread >= 0)
 551                 sid->tid = thread_map__pid(evlist->threads, thread);
 552         else
 553                 sid->tid = -1;
 554 }
 555
 556 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 557 {
 558         struct hlist_head *head;
 559         struct perf_sample_id *sid;
 560         int hash;
 561
 562         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 563         head = &evlist->heads[hash];
 564
 565         hlist_for_each_entry(sid, head, node)
 566                 if (sid->id == id)
 567                         return sid;
 568
 569         return NULL;
 570 }
 571
 572 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 573 {
 574         struct perf_sample_id *sid;
 575
 576         if (evlist->nr_entries == 1)
 577                 return perf_evlist__first(evlist);
 578
 579         sid = perf_evlist__id2sid(evlist, id);
 580         if (sid)
 581                 return sid->evsel;
 582
 583         if (!perf_evlist__sample_id_all(evlist))
 584                 return perf_evlist__first(evlist);
 585
 586         return NULL;
 587 }
 588
 589 static int perf_evlist__event2id(struct perf_evlist *evlist,
 590                                  union perf_event *event, u64 *id)
 591 {
 592         const u64 *array = event->sample.array;
 593         ssize_t n;
 594
 595         n = (event->header.size - sizeof(event->header)) >> 3;
 596
 597         if (event->header.type == PERF_RECORD_SAMPLE) {
 598                 if (evlist->id_pos >= n)
 599                         return -1;
 600                 *id = array[evlist->id_pos];
 601         } else {
 602                 if (evlist->is_pos > n)
 603                         return -1;
 604                 n -= evlist->is_pos;
 605                 *id = array[n];
 606         }
 607         return 0;
 608 }
 609
 610 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 611                                                    union perf_event *event)
 612 {
 613         struct perf_evsel *first = perf_evlist__first(evlist);
 614         struct hlist_head *head;
 615         struct perf_sample_id *sid;
 616         int hash;
 617         u64 id;
 618
 619         if (evlist->nr_entries == 1)
 620                 return first;
 621
 622         if (!first->attr.sample_id_all &&
 623             event->header.type != PERF_RECORD_SAMPLE)
 624                 return first;
 625
 626         if (perf_evlist__event2id(evlist, event, &id))
 627                 return NULL;
 628
 629         /* Synthesized events have an id of zero */
 630         if (!id)
 631                 return first;
 632
 633         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 634         head = &evlist->heads[hash];
 635
 636         hlist_for_each_entry(sid, head, node) {
 637                 if (sid->id == id)
 638                         return sid->evsel;
 639         }
 640         return NULL;
 641 }
 642
 643 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 644 {
 645         struct perf_mmap *md = &evlist->mmap[idx];
 646         u64 head;
 647         u64 old = md->prev;
 648         unsigned char *data = md->base + page_size;
 649         union perf_event *event = NULL;
 650
 651         /*
 652          * Check if event was unmapped due to a POLLHUP/POLLERR.
 653          */
 654         if (!atomic_read(&md->refcnt))
 655                 return NULL;
 656
 657         head = perf_mmap__read_head(md);
 658         if (evlist->overwrite) {
 659                 /*
 660                  * If we're further behind than half the buffer, there's a chance
 661                  * the writer will bite our tail and mess up the samples under us.
 662                  *
 663                  * If we somehow ended up ahead of the head, we got messed up.
 664                  *
 665                  * In either case, truncate and restart at head.
 666                  */
 667                 int diff = head - old;
 668                 if (diff > md->mask / 2 || diff < 0) {
 669                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 670
 671                         /*
 672                          * head points to a known good entry, start there.
 673                          */
 674                         old = head;
 675                 }
 676         }
 677
 678         if (old != head) {
 679                 size_t size;
 680
 681                 event = (union perf_event *)&data[old & md->mask];
 682                 size = event->header.size;
 683
 684                 /*
 685                  * Event straddles the mmap boundary -- header should always
 686                  * be inside due to u64 alignment of output.
 687                  */
 688                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 689                         unsigned int offset = old;
 690                         unsigned int len = min(sizeof(*event), size), cpy;
 691                         void *dst = md->event_copy;
 692
 693                         do {
 694                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 695                                 memcpy(dst, &data[offset & md->mask], cpy);
 696                                 offset += cpy;
 697                                 dst += cpy;
 698                                 len -= cpy;
 699                         } while (len);
 700
 701                         event = (union perf_event *) md->event_copy;
 702                 }
 703
 704                 old += size;
 705         }
 706
 707         md->prev = old;
 708
 709         return event;
 710 }
 711
 712 static bool perf_mmap__empty(struct perf_mmap *md)
 713 {
 714         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 715 }
 716
 717 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
 718 {
 719         atomic_inc(&evlist->mmap[idx].refcnt);
 720 }
 721
 722 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
 723 {
 724         BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
 725
 726         if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
 727                 __perf_evlist__munmap(evlist, idx);
 728 }
 729
 730 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 731 {
 732         struct perf_mmap *md = &evlist->mmap[idx];
 733
 734         if (!evlist->overwrite) {
 735                 u64 old = md->prev;
 736
 737                 perf_mmap__write_tail(md, old);
 738         }
 739
 740         if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 741                 perf_evlist__mmap_put(evlist, idx);
 742 }
 743
 744 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 745                                struct auxtrace_mmap_params *mp __maybe_unused,
 746                                void *userpg __maybe_unused,
 747                                int fd __maybe_unused)
 748 {
 749         return 0;
 750 }
 751
 752 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 753 {
 754 }
 755
 756 void __weak auxtrace_mmap_params__init(
 757                         struct auxtrace_mmap_params *mp __maybe_unused,
 758                         off_t auxtrace_offset __maybe_unused,
 759                         unsigned int auxtrace_pages __maybe_unused,
 760                         bool auxtrace_overwrite __maybe_unused)
 761 {
 762 }
 763
 764 void __weak auxtrace_mmap_params__set_idx(
 765                         struct auxtrace_mmap_params *mp __maybe_unused,
 766                         struct perf_evlist *evlist __maybe_unused,
 767                         int idx __maybe_unused,
 768                         bool per_cpu __maybe_unused)
 769 {
 770 }
 771
 772 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 773 {
 774         if (evlist->mmap[idx].base != NULL) {
 775                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 776                 evlist->mmap[idx].base = NULL;
 777                 atomic_set(&evlist->mmap[idx].refcnt, 0);
 778         }
 779         auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
 780 }
 781
 782 void perf_evlist__munmap(struct perf_evlist *evlist)
 783 {
 784         int i;
 785
 786         if (evlist->mmap == NULL)
 787                 return;
 788
 789         for (i = 0; i < evlist->nr_mmaps; i++)
 790                 __perf_evlist__munmap(evlist, i);
 791
 792         zfree(&evlist->mmap);
 793 }
 794
 795 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 796 {
 797         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 798         if (cpu_map__empty(evlist->cpus))
 799                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 800         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 801         return evlist->mmap != NULL ? 0 : -ENOMEM;
 802 }
 803
 804 struct mmap_params {
 805         int prot;
 806         int mask;
 807         struct auxtrace_mmap_params auxtrace_mp;
 808 };
 809
 810 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 811                                struct mmap_params *mp, int fd)
 812 {
 813         /*
 814          * The last one will be done at perf_evlist__mmap_consume(), so that we
 815          * make sure we don't prevent tools from consuming every last event in
 816          * the ring buffer.
 817          *
 818          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 819          * anymore, but the last events for it are still in the ring buffer,
 820          * waiting to be consumed.
 821          *
 822          * Tools can chose to ignore this at their own discretion, but the
 823          * evlist layer can't just drop it when filtering events in
 824          * perf_evlist__filter_pollfd().
 825          */
 826         atomic_set(&evlist->mmap[idx].refcnt, 2);
 827         evlist->mmap[idx].prev = 0;
 828         evlist->mmap[idx].mask = mp->mask;
 829         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 830                                       MAP_SHARED, fd, 0);
 831         if (evlist->mmap[idx].base == MAP_FAILED) {
 832                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 833                           errno);
 834                 evlist->mmap[idx].base = NULL;
 835                 return -1;
 836         }
 837
 838         if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
 839                                 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
 840                 return -1;
 841
 842         return 0;
 843 }
 844
 845 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 846                                        struct mmap_params *mp, int cpu,
 847                                        int thread, int *output)
 848 {
 849         struct perf_evsel *evsel;
 850
 851         evlist__for_each(evlist, evsel) {
 852                 int fd;
 853
 854                 if (evsel->system_wide && thread)
 855                         continue;
 856
 857                 fd = FD(evsel, cpu, thread);
 858
 859                 if (*output == -1) {
 860                         *output = fd;
 861                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 862                                 return -1;
 863                 } else {
 864                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 865                                 return -1;
 866
 867                         perf_evlist__mmap_get(evlist, idx);
 868                 }
 869
 870                 /*
 871                  * The system_wide flag causes a selected event to be opened
 872                  * always without a pid.  Consequently it will never get a
 873                  * POLLHUP, but it is used for tracking in combination with
 874                  * other events, so it should not need to be polled anyway.
 875                  * Therefore don't add it for polling.
 876                  */
 877                 if (!evsel->system_wide &&
 878                     __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
 879                         perf_evlist__mmap_put(evlist, idx);
 880                         return -1;
 881                 }
 882
 883                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 884                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 885                                                    fd) < 0)
 886                                 return -1;
 887                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 888                                                  thread);
 889                 }
 890         }
 891
 892         return 0;
 893 }
 894
 895 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 896                                      struct mmap_params *mp)
 897 {
 898         int cpu, thread;
 899         int nr_cpus = cpu_map__nr(evlist->cpus);
 900         int nr_threads = thread_map__nr(evlist->threads);
 901
 902         pr_debug2("perf event ring buffer mmapped per cpu\n");
 903         for (cpu = 0; cpu < nr_cpus; cpu++) {
 904                 int output = -1;
 905
 906                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 907                                               true);
 908
 909                 for (thread = 0; thread < nr_threads; thread++) {
 910                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 911                                                         thread, &output))
 912                                 goto out_unmap;
 913                 }
 914         }
 915
 916         return 0;
 917
 918 out_unmap:
 919         for (cpu = 0; cpu < nr_cpus; cpu++)
 920                 __perf_evlist__munmap(evlist, cpu);
 921         return -1;
 922 }
 923
 924 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 925                                         struct mmap_params *mp)
 926 {
 927         int thread;
 928         int nr_threads = thread_map__nr(evlist->threads);
 929
 930         pr_debug2("perf event ring buffer mmapped per thread\n");
 931         for (thread = 0; thread < nr_threads; thread++) {
 932                 int output = -1;
 933
 934                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 935                                               false);
 936
 937                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 938                                                 &output))
 939                         goto out_unmap;
 940         }
 941
 942         return 0;
 943
 944 out_unmap:
 945         for (thread = 0; thread < nr_threads; thread++)
 946                 __perf_evlist__munmap(evlist, thread);
 947         return -1;
 948 }
 949
 950 static size_t perf_evlist__mmap_size(unsigned long pages)
 951 {
 952         if (pages == UINT_MAX) {
 953                 int max;
 954
 955                 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 956                         /*
 957                          * Pick a once upon a time good value, i.e. things look
 958                          * strange since we can't read a sysctl value, but lets not
 959                          * die yet...
 960                          */
 961                         max = 512;
 962                 } else {
 963                         max -= (page_size / 1024);
 964                 }
 965
 966                 pages = (max * 1024) / page_size;
 967                 if (!is_power_of_2(pages))
 968                         pages = rounddown_pow_of_two(pages);
 969         } else if (!is_power_of_2(pages))
 970                 return 0;
 971
 972         return (pages + 1) * page_size;
 973 }
 974
 975 static long parse_pages_arg(const char *str, unsigned long min,
 976                             unsigned long max)
 977 {
 978         unsigned long pages, val;
 979         static struct parse_tag tags[] = {
 980                 { .tag  = 'B', .mult = 1       },
 981                 { .tag  = 'K', .mult = 1 << 10 },
 982                 { .tag  = 'M', .mult = 1 << 20 },
 983                 { .tag  = 'G', .mult = 1 << 30 },
 984                 { .tag  = 0 },
 985         };
 986
 987         if (str == NULL)
 988                 return -EINVAL;
 989
 990         val = parse_tag_value(str, tags);
 991         if (val != (unsigned long) -1) {
 992                 /* we got file size value */
 993                 pages = PERF_ALIGN(val, page_size) / page_size;
 994         } else {
 995                 /* we got pages count value */
 996                 char *eptr;
 997                 pages = strtoul(str, &eptr, 10);
 998                 if (*eptr != '\0')
 999                         return -EINVAL;
1000         }
1001
1002         if (pages == 0 && min == 0) {
1003                 /* leave number of pages at 0 */
1004         } else if (!is_power_of_2(pages)) {
1005                 /* round pages up to next power of 2 */
1006                 pages = roundup_pow_of_two(pages);
1007                 if (!pages)
1008                         return -EINVAL;
1009                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1010                         pages * page_size, pages);
1011         }
1012
1013         if (pages > max)
1014                 return -EINVAL;
1015
1016         return pages;
1017 }
1018
1019 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1020 {
1021         unsigned long max = UINT_MAX;
1022         long pages;
1023
1024         if (max > SIZE_MAX / page_size)
1025                 max = SIZE_MAX / page_size;
1026
1027         pages = parse_pages_arg(str, 1, max);
1028         if (pages < 0) {
1029                 pr_err("Invalid argument for --mmap_pages/-m\n");
1030                 return -1;
1031         }
1032
1033         *mmap_pages = pages;
1034         return 0;
1035 }
1036
1037 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1038                                   int unset __maybe_unused)
1039 {
1040         return __perf_evlist__parse_mmap_pages(opt->value, str);
1041 }
1042
1043 /**
1044  * perf_evlist__mmap_ex - Create mmaps to receive events.
1045  * @evlist: list of events
1046  * @pages: map length in pages
1047  * @overwrite: overwrite older events?
1048  * @auxtrace_pages - auxtrace map length in pages
1049  * @auxtrace_overwrite - overwrite older auxtrace data?
1050  *
1051  * If @overwrite is %false the user needs to signal event consumption using
1052  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1053  * automatically.
1054  *
1055  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1056  * consumption using auxtrace_mmap__write_tail().
1057  *
1058  * Return: %0 on success, negative error code otherwise.
1059  */
1060 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1061                          bool overwrite, unsigned int auxtrace_pages,
1062                          bool auxtrace_overwrite)
1063 {
1064         struct perf_evsel *evsel;
1065         const struct cpu_map *cpus = evlist->cpus;
1066         const struct thread_map *threads = evlist->threads;
1067         struct mmap_params mp = {
1068                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1069         };
1070
1071         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1072                 return -ENOMEM;
1073
1074         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1075                 return -ENOMEM;
1076
1077         evlist->overwrite = overwrite;
1078         evlist->mmap_len = perf_evlist__mmap_size(pages);
1079         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1080         mp.mask = evlist->mmap_len - page_size - 1;
1081
1082         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1083                                    auxtrace_pages, auxtrace_overwrite);
1084
1085         evlist__for_each(evlist, evsel) {
1086                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1087                     evsel->sample_id == NULL &&
1088                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1089                         return -ENOMEM;
1090         }
1091
1092         if (cpu_map__empty(cpus))
1093                 return perf_evlist__mmap_per_thread(evlist, &mp);
1094
1095         return perf_evlist__mmap_per_cpu(evlist, &mp);
1096 }
1097
1098 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1099                       bool overwrite)
1100 {
1101         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1102 }
1103
1104 static int perf_evlist__propagate_maps(struct perf_evlist *evlist,
1105                                        struct target *target)
1106 {
1107         struct perf_evsel *evsel;
1108
1109         evlist__for_each(evlist, evsel) {
1110                 /*
1111                  * We already have cpus for evsel (via PMU sysfs) so
1112                  * keep it, if there's no target cpu list defined.
1113                  */
1114                 if (evsel->cpus && target->cpu_list)
1115                         cpu_map__put(evsel->cpus);
1116
1117                 if (!evsel->cpus || target->cpu_list)
1118                         evsel->cpus = cpu_map__get(evlist->cpus);
1119
1120                 evsel->threads = thread_map__get(evlist->threads);
1121
1122                 if (!evsel->cpus || !evsel->threads)
1123                         return -ENOMEM;
1124         }
1125
1126         return 0;
1127 }
1128
1129 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1130 {
1131         evlist->threads = thread_map__new_str(target->pid, target->tid,
1132                                               target->uid);
1133
1134         if (evlist->threads == NULL)
1135                 return -1;
1136
1137         if (target__uses_dummy_map(target))
1138                 evlist->cpus = cpu_map__dummy_new();
1139         else
1140                 evlist->cpus = cpu_map__new(target->cpu_list);
1141
1142         if (evlist->cpus == NULL)
1143                 goto out_delete_threads;
1144
1145         return perf_evlist__propagate_maps(evlist, target);
1146
1147 out_delete_threads:
1148         thread_map__put(evlist->threads);
1149         evlist->threads = NULL;
1150         return -1;
1151 }
1152
1153 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1154 {
1155         struct perf_evsel *evsel;
1156         int err = 0;
1157         const int ncpus = cpu_map__nr(evlist->cpus),
1158                   nthreads = thread_map__nr(evlist->threads);
1159
1160         evlist__for_each(evlist, evsel) {
1161                 if (evsel->filter == NULL)
1162                         continue;
1163
1164                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1165                 if (err) {
1166                         *err_evsel = evsel;
1167                         break;
1168                 }
1169         }
1170
1171         return err;
1172 }
1173
1174 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1175 {
1176         struct perf_evsel *evsel;
1177         int err = 0;
1178         const int ncpus = cpu_map__nr(evlist->cpus),
1179                   nthreads = thread_map__nr(evlist->threads);
1180
1181         evlist__for_each(evlist, evsel) {
1182                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1183                 if (err)
1184                         break;
1185         }
1186
1187         return err;
1188 }
1189
1190 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1191 {
1192         char *filter;
1193         int ret = -1;
1194         size_t i;
1195
1196         for (i = 0; i < npids; ++i) {
1197                 if (i == 0) {
1198                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1199                                 return -1;
1200                 } else {
1201                         char *tmp;
1202
1203                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1204                                 goto out_free;
1205
1206                         free(filter);
1207                         filter = tmp;
1208                 }
1209         }
1210
1211         ret = perf_evlist__set_filter(evlist, filter);
1212 out_free:
1213         free(filter);
1214         return ret;
1215 }
1216
1217 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1218 {
1219         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1220 }
1221
1222 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1223 {
1224         struct perf_evsel *pos;
1225
1226         if (evlist->nr_entries == 1)
1227                 return true;
1228
1229         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1230                 return false;
1231
1232         evlist__for_each(evlist, pos) {
1233                 if (pos->id_pos != evlist->id_pos ||
1234                     pos->is_pos != evlist->is_pos)
1235                         return false;
1236         }
1237
1238         return true;
1239 }
1240
1241 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1242 {
1243         struct perf_evsel *evsel;
1244
1245         if (evlist->combined_sample_type)
1246                 return evlist->combined_sample_type;
1247
1248         evlist__for_each(evlist, evsel)
1249                 evlist->combined_sample_type |= evsel->attr.sample_type;
1250
1251         return evlist->combined_sample_type;
1252 }
1253
1254 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1255 {
1256         evlist->combined_sample_type = 0;
1257         return __perf_evlist__combined_sample_type(evlist);
1258 }
1259
1260 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1261 {
1262         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1263         u64 read_format = first->attr.read_format;
1264         u64 sample_type = first->attr.sample_type;
1265
1266         evlist__for_each(evlist, pos) {
1267                 if (read_format != pos->attr.read_format)
1268                         return false;
1269         }
1270
1271         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1272         if ((sample_type & PERF_SAMPLE_READ) &&
1273             !(read_format & PERF_FORMAT_ID)) {
1274                 return false;
1275         }
1276
1277         return true;
1278 }
1279
1280 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1281 {
1282         struct perf_evsel *first = perf_evlist__first(evlist);
1283         return first->attr.read_format;
1284 }
1285
1286 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1287 {
1288         struct perf_evsel *first = perf_evlist__first(evlist);
1289         struct perf_sample *data;
1290         u64 sample_type;
1291         u16 size = 0;
1292
1293         if (!first->attr.sample_id_all)
1294                 goto out;
1295
1296         sample_type = first->attr.sample_type;
1297
1298         if (sample_type & PERF_SAMPLE_TID)
1299                 size += sizeof(data->tid) * 2;
1300
1301        if (sample_type & PERF_SAMPLE_TIME)
1302                 size += sizeof(data->time);
1303
1304         if (sample_type & PERF_SAMPLE_ID)
1305                 size += sizeof(data->id);
1306
1307         if (sample_type & PERF_SAMPLE_STREAM_ID)
1308                 size += sizeof(data->stream_id);
1309
1310         if (sample_type & PERF_SAMPLE_CPU)
1311                 size += sizeof(data->cpu) * 2;
1312
1313         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1314                 size += sizeof(data->id);
1315 out:
1316         return size;
1317 }
1318
1319 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1320 {
1321         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1322
1323         evlist__for_each_continue(evlist, pos) {
1324                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1325                         return false;
1326         }
1327
1328         return true;
1329 }
1330
1331 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1332 {
1333         struct perf_evsel *first = perf_evlist__first(evlist);
1334         return first->attr.sample_id_all;
1335 }
1336
1337 void perf_evlist__set_selected(struct perf_evlist *evlist,
1338                                struct perf_evsel *evsel)
1339 {
1340         evlist->selected = evsel;
1341 }
1342
1343 void perf_evlist__close(struct perf_evlist *evlist)
1344 {
1345         struct perf_evsel *evsel;
1346         int ncpus = cpu_map__nr(evlist->cpus);
1347         int nthreads = thread_map__nr(evlist->threads);
1348         int n;
1349
1350         evlist__for_each_reverse(evlist, evsel) {
1351                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1352                 perf_evsel__close(evsel, n, nthreads);
1353         }
1354 }
1355
1356 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1357 {
1358         int err = -ENOMEM;
1359
1360         /*
1361          * Try reading /sys/devices/system/cpu/online to get
1362          * an all cpus map.
1363          *
1364          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1365          * code needs an overhaul to properly forward the
1366          * error, and we may not want to do that fallback to a
1367          * default cpu identity map :-\
1368          */
1369         evlist->cpus = cpu_map__new(NULL);
1370         if (evlist->cpus == NULL)
1371                 goto out;
1372
1373         evlist->threads = thread_map__new_dummy();
1374         if (evlist->threads == NULL)
1375                 goto out_free_cpus;
1376
1377         err = 0;
1378 out:
1379         return err;
1380 out_free_cpus:
1381         cpu_map__put(evlist->cpus);
1382         evlist->cpus = NULL;
1383         goto out;
1384 }
1385
1386 int perf_evlist__open(struct perf_evlist *evlist)
1387 {
1388         struct perf_evsel *evsel;
1389         int err;
1390
1391         /*
1392          * Default: one fd per CPU, all threads, aka systemwide
1393          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1394          */
1395         if (evlist->threads == NULL && evlist->cpus == NULL) {
1396                 err = perf_evlist__create_syswide_maps(evlist);
1397                 if (err < 0)
1398                         goto out_err;
1399         }
1400
1401         perf_evlist__update_id_pos(evlist);
1402
1403         evlist__for_each(evlist, evsel) {
1404                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1405                 if (err < 0)
1406                         goto out_err;
1407         }
1408
1409         return 0;
1410 out_err:
1411         perf_evlist__close(evlist);
1412         errno = -err;
1413         return err;
1414 }
1415
1416 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1417                                   const char *argv[], bool pipe_output,
1418                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1419 {
1420         int child_ready_pipe[2], go_pipe[2];
1421         char bf;
1422
1423         if (pipe(child_ready_pipe) < 0) {
1424                 perror("failed to create 'ready' pipe");
1425                 return -1;
1426         }
1427
1428         if (pipe(go_pipe) < 0) {
1429                 perror("failed to create 'go' pipe");
1430                 goto out_close_ready_pipe;
1431         }
1432
1433         evlist->workload.pid = fork();
1434         if (evlist->workload.pid < 0) {
1435                 perror("failed to fork");
1436                 goto out_close_pipes;
1437         }
1438
1439         if (!evlist->workload.pid) {
1440                 int ret;
1441
1442                 if (pipe_output)
1443                         dup2(2, 1);
1444
1445                 signal(SIGTERM, SIG_DFL);
1446
1447                 close(child_ready_pipe[0]);
1448                 close(go_pipe[1]);
1449                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1450
1451                 /*
1452                  * Tell the parent we're ready to go
1453                  */
1454                 close(child_ready_pipe[1]);
1455
1456                 /*
1457                  * Wait until the parent tells us to go.
1458                  */
1459                 ret = read(go_pipe[0], &bf, 1);
1460                 /*
1461                  * The parent will ask for the execvp() to be performed by
1462                  * writing exactly one byte, in workload.cork_fd, usually via
1463                  * perf_evlist__start_workload().
1464                  *
1465                  * For cancelling the workload without actually running it,
1466                  * the parent will just close workload.cork_fd, without writing
1467                  * anything, i.e. read will return zero and we just exit()
1468                  * here.
1469                  */
1470                 if (ret != 1) {
1471                         if (ret == -1)
1472                                 perror("unable to read pipe");
1473                         exit(ret);
1474                 }
1475
1476                 execvp(argv[0], (char **)argv);
1477
1478                 if (exec_error) {
1479                         union sigval val;
1480
1481                         val.sival_int = errno;
1482                         if (sigqueue(getppid(), SIGUSR1, val))
1483                                 perror(argv[0]);
1484                 } else
1485                         perror(argv[0]);
1486                 exit(-1);
1487         }
1488
1489         if (exec_error) {
1490                 struct sigaction act = {
1491                         .sa_flags     = SA_SIGINFO,
1492                         .sa_sigaction = exec_error,
1493                 };
1494                 sigaction(SIGUSR1, &act, NULL);
1495         }
1496
1497         if (target__none(target)) {
1498                 if (evlist->threads == NULL) {
1499                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1500                                 __func__, __LINE__);
1501                         goto out_close_pipes;
1502                 }
1503                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1504         }
1505
1506         close(child_ready_pipe[1]);
1507         close(go_pipe[0]);
1508         /*
1509          * wait for child to settle
1510          */
1511         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1512                 perror("unable to read pipe");
1513                 goto out_close_pipes;
1514         }
1515
1516         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1517         evlist->workload.cork_fd = go_pipe[1];
1518         close(child_ready_pipe[0]);
1519         return 0;
1520
1521 out_close_pipes:
1522         close(go_pipe[0]);
1523         close(go_pipe[1]);
1524 out_close_ready_pipe:
1525         close(child_ready_pipe[0]);
1526         close(child_ready_pipe[1]);
1527         return -1;
1528 }
1529
1530 int perf_evlist__start_workload(struct perf_evlist *evlist)
1531 {
1532         if (evlist->workload.cork_fd > 0) {
1533                 char bf = 0;
1534                 int ret;
1535                 /*
1536                  * Remove the cork, let it rip!
1537                  */
1538                 ret = write(evlist->workload.cork_fd, &bf, 1);
1539                 if (ret < 0)
1540                         perror("enable to write to pipe");
1541
1542                 close(evlist->workload.cork_fd);
1543                 return ret;
1544         }
1545
1546         return 0;
1547 }
1548
1549 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1550                               struct perf_sample *sample)
1551 {
1552         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1553
1554         if (!evsel)
1555                 return -EFAULT;
1556         return perf_evsel__parse_sample(evsel, event, sample);
1557 }
1558
1559 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1560 {
1561         struct perf_evsel *evsel;
1562         size_t printed = 0;
1563
1564         evlist__for_each(evlist, evsel) {
1565                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1566                                    perf_evsel__name(evsel));
1567         }
1568
1569         return printed + fprintf(fp, "\n");
1570 }
1571
1572 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1573                                int err, char *buf, size_t size)
1574 {
1575         int printed, value;
1576         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1577
1578         switch (err) {
1579         case EACCES:
1580         case EPERM:
1581                 printed = scnprintf(buf, size,
1582                                     "Error:\t%s.\n"
1583                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1584
1585                 value = perf_event_paranoid();
1586
1587                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1588
1589                 if (value >= 2) {
1590                         printed += scnprintf(buf + printed, size - printed,
1591                                              "For your workloads it needs to be <= 1\nHint:\t");
1592                 }
1593                 printed += scnprintf(buf + printed, size - printed,
1594                                      "For system wide tracing it needs to be set to -1.\n");
1595
1596                 printed += scnprintf(buf + printed, size - printed,
1597                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1598                                     "Hint:\tThe current value is %d.", value);
1599                 break;
1600         default:
1601                 scnprintf(buf, size, "%s", emsg);
1602                 break;
1603         }
1604
1605         return 0;
1606 }
1607
1608 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1609 {
1610         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1611         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1612
1613         switch (err) {
1614         case EPERM:
1615                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1616                 printed += scnprintf(buf + printed, size - printed,
1617                                      "Error:\t%s.\n"
1618                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1619                                      "Hint:\tTried using %zd kB.\n",
1620                                      emsg, pages_max_per_user, pages_attempted);
1621
1622                 if (pages_attempted >= pages_max_per_user) {
1623                         printed += scnprintf(buf + printed, size - printed,
1624                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1625                                              pages_max_per_user + pages_attempted);
1626                 }
1627
1628                 printed += scnprintf(buf + printed, size - printed,
1629                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1630                 break;
1631         default:
1632                 scnprintf(buf, size, "%s", emsg);
1633                 break;
1634         }
1635
1636         return 0;
1637 }
1638
1639 void perf_evlist__to_front(struct perf_evlist *evlist,
1640                            struct perf_evsel *move_evsel)
1641 {
1642         struct perf_evsel *evsel, *n;
1643         LIST_HEAD(move);
1644
1645         if (move_evsel == perf_evlist__first(evlist))
1646                 return;
1647
1648         evlist__for_each_safe(evlist, n, evsel) {
1649                 if (evsel->leader == move_evsel->leader)
1650                         list_move_tail(&evsel->node, &move);
1651         }
1652
1653         list_splice(&move, &evlist->entries);
1654 }
1655
1656 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1657                                      struct perf_evsel *tracking_evsel)
1658 {
1659         struct perf_evsel *evsel;
1660
1661         if (tracking_evsel->tracking)
1662                 return;
1663
1664         evlist__for_each(evlist, evsel) {
1665                 if (evsel != tracking_evsel)
1666                         evsel->tracking = false;
1667         }
1668
1669         tracking_evsel->tracking = true;
1670 }