tools/perf/util/evlist.c

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