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