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_backward(struct perf_evlist *evlist, int idx)
 719 {
 720         struct perf_mmap *md = &evlist->mmap[idx];
 721
 722         /*
 723          * No need to check messup for backward ring buffer:
 724          * We can always read arbitrary long data from a backward
 725          * ring buffer unless we forget to pause it before reading.
 726          */
 727         return perf_mmap__read_backward(md);
 728 }
 729
 730 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 731 {
 732         return perf_evlist__mmap_read_forward(evlist, idx);
 733 }
 734
 735 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 736 {
 737         perf_mmap__read_catchup(&evlist->mmap[idx]);
 738 }
 739
 740 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 741 {
 742         perf_mmap__consume(&evlist->mmap[idx], false);
 743 }
 744
 745 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 746 {
 747         int i;
 748
 749         if (evlist->mmap)
 750                 for (i = 0; i < evlist->nr_mmaps; i++)
 751                         perf_mmap__munmap(&evlist->mmap[i]);
 752
 753         if (evlist->overwrite_mmap)
 754                 for (i = 0; i < evlist->nr_mmaps; i++)
 755                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 756 }
 757
 758 void perf_evlist__munmap(struct perf_evlist *evlist)
 759 {
 760         perf_evlist__munmap_nofree(evlist);
 761         zfree(&evlist->mmap);
 762         zfree(&evlist->overwrite_mmap);
 763 }
 764
 765 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 766 {
 767         int i;
 768         struct perf_mmap *map;
 769
 770         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 771         if (cpu_map__empty(evlist->cpus))
 772                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 773         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 774         if (!map)
 775                 return NULL;
 776
 777         for (i = 0; i < evlist->nr_mmaps; i++) {
 778                 map[i].fd = -1;
 779                 /*
 780                  * When the perf_mmap() call is made we grab one refcount, plus
 781                  * one extra to let perf_evlist__mmap_consume() get the last
 782                  * events after all real references (perf_mmap__get()) are
 783                  * dropped.
 784                  *
 785                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 786                  * thus does perf_mmap__get() on it.
 787                  */
 788                 refcount_set(&map[i].refcnt, 0);
 789         }
 790         return map;
 791 }
 792
 793 static bool
 794 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 795                          struct perf_evsel *evsel)
 796 {
 797         if (evsel->attr.write_backward)
 798                 return false;
 799         return true;
 800 }
 801
 802 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 803                                        struct mmap_params *mp, int cpu_idx,
 804                                        int thread, int *_output, int *_output_overwrite)
 805 {
 806         struct perf_evsel *evsel;
 807         int revent;
 808         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 809
 810         evlist__for_each_entry(evlist, evsel) {
 811                 struct perf_mmap *maps = evlist->mmap;
 812                 int *output = _output;
 813                 int fd;
 814                 int cpu;
 815
 816                 mp->prot = PROT_READ | PROT_WRITE;
 817                 if (evsel->attr.write_backward) {
 818                         output = _output_overwrite;
 819                         maps = evlist->overwrite_mmap;
 820
 821                         if (!maps) {
 822                                 maps = perf_evlist__alloc_mmap(evlist);
 823                                 if (!maps)
 824                                         return -1;
 825                                 evlist->overwrite_mmap = maps;
 826                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 827                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 828                         }
 829                         mp->prot &= ~PROT_WRITE;
 830                 }
 831
 832                 if (evsel->system_wide && thread)
 833                         continue;
 834
 835                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 836                 if (cpu == -1)
 837                         continue;
 838
 839                 fd = FD(evsel, cpu, thread);
 840
 841                 if (*output == -1) {
 842                         *output = fd;
 843
 844                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
 845                                 return -1;
 846                 } else {
 847                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 848                                 return -1;
 849
 850                         perf_mmap__get(&maps[idx]);
 851                 }
 852
 853                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 854
 855                 /*
 856                  * The system_wide flag causes a selected event to be opened
 857                  * always without a pid.  Consequently it will never get a
 858                  * POLLHUP, but it is used for tracking in combination with
 859                  * other events, so it should not need to be polled anyway.
 860                  * Therefore don't add it for polling.
 861                  */
 862                 if (!evsel->system_wide &&
 863                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 864                         perf_mmap__put(&maps[idx]);
 865                         return -1;
 866                 }
 867
 868                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 869                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 870                                                    fd) < 0)
 871                                 return -1;
 872                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 873                                                  thread);
 874                 }
 875         }
 876
 877         return 0;
 878 }
 879
 880 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 881                                      struct mmap_params *mp)
 882 {
 883         int cpu, thread;
 884         int nr_cpus = cpu_map__nr(evlist->cpus);
 885         int nr_threads = thread_map__nr(evlist->threads);
 886
 887         pr_debug2("perf event ring buffer mmapped per cpu\n");
 888         for (cpu = 0; cpu < nr_cpus; cpu++) {
 889                 int output = -1;
 890                 int output_overwrite = -1;
 891
 892                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 893                                               true);
 894
 895                 for (thread = 0; thread < nr_threads; thread++) {
 896                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 897                                                         thread, &output, &output_overwrite))
 898                                 goto out_unmap;
 899                 }
 900         }
 901
 902         return 0;
 903
 904 out_unmap:
 905         perf_evlist__munmap_nofree(evlist);
 906         return -1;
 907 }
 908
 909 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 910                                         struct mmap_params *mp)
 911 {
 912         int thread;
 913         int nr_threads = thread_map__nr(evlist->threads);
 914
 915         pr_debug2("perf event ring buffer mmapped per thread\n");
 916         for (thread = 0; thread < nr_threads; thread++) {
 917                 int output = -1;
 918                 int output_overwrite = -1;
 919
 920                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 921                                               false);
 922
 923                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 924                                                 &output, &output_overwrite))
 925                         goto out_unmap;
 926         }
 927
 928         return 0;
 929
 930 out_unmap:
 931         perf_evlist__munmap_nofree(evlist);
 932         return -1;
 933 }
 934
 935 unsigned long perf_event_mlock_kb_in_pages(void)
 936 {
 937         unsigned long pages;
 938         int max;
 939
 940         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 941                 /*
 942                  * Pick a once upon a time good value, i.e. things look
 943                  * strange since we can't read a sysctl value, but lets not
 944                  * die yet...
 945                  */
 946                 max = 512;
 947         } else {
 948                 max -= (page_size / 1024);
 949         }
 950
 951         pages = (max * 1024) / page_size;
 952         if (!is_power_of_2(pages))
 953                 pages = rounddown_pow_of_two(pages);
 954
 955         return pages;
 956 }
 957
 958 size_t perf_evlist__mmap_size(unsigned long pages)
 959 {
 960         if (pages == UINT_MAX)
 961                 pages = perf_event_mlock_kb_in_pages();
 962         else if (!is_power_of_2(pages))
 963                 return 0;
 964
 965         return (pages + 1) * page_size;
 966 }
 967
 968 static long parse_pages_arg(const char *str, unsigned long min,
 969                             unsigned long max)
 970 {
 971         unsigned long pages, val;
 972         static struct parse_tag tags[] = {
 973                 { .tag  = 'B', .mult = 1       },
 974                 { .tag  = 'K', .mult = 1 << 10 },
 975                 { .tag  = 'M', .mult = 1 << 20 },
 976                 { .tag  = 'G', .mult = 1 << 30 },
 977                 { .tag  = 0 },
 978         };
 979
 980         if (str == NULL)
 981                 return -EINVAL;
 982
 983         val = parse_tag_value(str, tags);
 984         if (val != (unsigned long) -1) {
 985                 /* we got file size value */
 986                 pages = PERF_ALIGN(val, page_size) / page_size;
 987         } else {
 988                 /* we got pages count value */
 989                 char *eptr;
 990                 pages = strtoul(str, &eptr, 10);
 991                 if (*eptr != '\0')
 992                         return -EINVAL;
 993         }
 994
 995         if (pages == 0 && min == 0) {
 996                 /* leave number of pages at 0 */
 997         } else if (!is_power_of_2(pages)) {
 998                 char buf[100];
 999
1000                 /* round pages up to next power of 2 */
1001                 pages = roundup_pow_of_two(pages);
1002                 if (!pages)
1003                         return -EINVAL;
1004
1005                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1006                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1007                         buf, pages);
1008         }
1009
1010         if (pages > max)
1011                 return -EINVAL;
1012
1013         return pages;
1014 }
1015
1016 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1017 {
1018         unsigned long max = UINT_MAX;
1019         long pages;
1020
1021         if (max > SIZE_MAX / page_size)
1022                 max = SIZE_MAX / page_size;
1023
1024         pages = parse_pages_arg(str, 1, max);
1025         if (pages < 0) {
1026                 pr_err("Invalid argument for --mmap_pages/-m\n");
1027                 return -1;
1028         }
1029
1030         *mmap_pages = pages;
1031         return 0;
1032 }
1033
1034 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1035                                   int unset __maybe_unused)
1036 {
1037         return __perf_evlist__parse_mmap_pages(opt->value, str);
1038 }
1039
1040 /**
1041  * perf_evlist__mmap_ex - Create mmaps to receive events.
1042  * @evlist: list of events
1043  * @pages: map length in pages
1044  * @overwrite: overwrite older events?
1045  * @auxtrace_pages - auxtrace map length in pages
1046  * @auxtrace_overwrite - overwrite older auxtrace data?
1047  *
1048  * If @overwrite is %false the user needs to signal event consumption using
1049  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1050  * automatically.
1051  *
1052  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1053  * consumption using auxtrace_mmap__write_tail().
1054  *
1055  * Return: %0 on success, negative error code otherwise.
1056  */
1057 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1058                          unsigned int auxtrace_pages,
1059                          bool auxtrace_overwrite)
1060 {
1061         struct perf_evsel *evsel;
1062         const struct cpu_map *cpus = evlist->cpus;
1063         const struct thread_map *threads = evlist->threads;
1064         /*
1065          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1066          * Its value is decided by evsel's write_backward.
1067          * So &mp should not be passed through const pointer.
1068          */
1069         struct mmap_params mp;
1070
1071         if (!evlist->mmap)
1072                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1073         if (!evlist->mmap)
1074                 return -ENOMEM;
1075
1076         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1077                 return -ENOMEM;
1078
1079         evlist->mmap_len = perf_evlist__mmap_size(pages);
1080         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1081         mp.mask = evlist->mmap_len - page_size - 1;
1082
1083         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1084                                    auxtrace_pages, auxtrace_overwrite);
1085
1086         evlist__for_each_entry(evlist, evsel) {
1087                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1088                     evsel->sample_id == NULL &&
1089                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1090                         return -ENOMEM;
1091         }
1092
1093         if (cpu_map__empty(cpus))
1094                 return perf_evlist__mmap_per_thread(evlist, &mp);
1095
1096         return perf_evlist__mmap_per_cpu(evlist, &mp);
1097 }
1098
1099 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1100 {
1101         return perf_evlist__mmap_ex(evlist, pages, 0, false);
1102 }
1103
1104 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1105 {
1106         struct cpu_map *cpus;
1107         struct thread_map *threads;
1108
1109         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1110                                       target->per_thread);
1111
1112         if (!threads)
1113                 return -1;
1114
1115         if (target__uses_dummy_map(target))
1116                 cpus = cpu_map__dummy_new();
1117         else
1118                 cpus = cpu_map__new(target->cpu_list);
1119
1120         if (!cpus)
1121                 goto out_delete_threads;
1122
1123         evlist->has_user_cpus = !!target->cpu_list;
1124
1125         perf_evlist__set_maps(evlist, cpus, threads);
1126
1127         return 0;
1128
1129 out_delete_threads:
1130         thread_map__put(threads);
1131         return -1;
1132 }
1133
1134 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1135                            struct thread_map *threads)
1136 {
1137         /*
1138          * Allow for the possibility that one or another of the maps isn't being
1139          * changed i.e. don't put it.  Note we are assuming the maps that are
1140          * being applied are brand new and evlist is taking ownership of the
1141          * original reference count of 1.  If that is not the case it is up to
1142          * the caller to increase the reference count.
1143          */
1144         if (cpus != evlist->cpus) {
1145                 cpu_map__put(evlist->cpus);
1146                 evlist->cpus = cpu_map__get(cpus);
1147         }
1148
1149         if (threads != evlist->threads) {
1150                 thread_map__put(evlist->threads);
1151                 evlist->threads = thread_map__get(threads);
1152         }
1153
1154         perf_evlist__propagate_maps(evlist);
1155 }
1156
1157 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1158                                    enum perf_event_sample_format bit)
1159 {
1160         struct perf_evsel *evsel;
1161
1162         evlist__for_each_entry(evlist, evsel)
1163                 __perf_evsel__set_sample_bit(evsel, bit);
1164 }
1165
1166 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1167                                      enum perf_event_sample_format bit)
1168 {
1169         struct perf_evsel *evsel;
1170
1171         evlist__for_each_entry(evlist, evsel)
1172                 __perf_evsel__reset_sample_bit(evsel, bit);
1173 }
1174
1175 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1176 {
1177         struct perf_evsel *evsel;
1178         int err = 0;
1179
1180         evlist__for_each_entry(evlist, evsel) {
1181                 if (evsel->filter == NULL)
1182                         continue;
1183
1184                 /*
1185                  * filters only work for tracepoint event, which doesn't have cpu limit.
1186                  * So evlist and evsel should always be same.
1187                  */
1188                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1189                 if (err) {
1190                         *err_evsel = evsel;
1191                         break;
1192                 }
1193         }
1194
1195         return err;
1196 }
1197
1198 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1199 {
1200         struct perf_evsel *evsel;
1201         int err = 0;
1202
1203         evlist__for_each_entry(evlist, evsel) {
1204                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1205                         continue;
1206
1207                 err = perf_evsel__set_filter(evsel, filter);
1208                 if (err)
1209                         break;
1210         }
1211
1212         return err;
1213 }
1214
1215 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1216 {
1217         char *filter;
1218         int ret = -1;
1219         size_t i;
1220
1221         for (i = 0; i < npids; ++i) {
1222                 if (i == 0) {
1223                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1224                                 return -1;
1225                 } else {
1226                         char *tmp;
1227
1228                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1229                                 goto out_free;
1230
1231                         free(filter);
1232                         filter = tmp;
1233                 }
1234         }
1235
1236         ret = perf_evlist__set_filter(evlist, filter);
1237 out_free:
1238         free(filter);
1239         return ret;
1240 }
1241
1242 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1243 {
1244         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1245 }
1246
1247 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1248 {
1249         struct perf_evsel *pos;
1250
1251         if (evlist->nr_entries == 1)
1252                 return true;
1253
1254         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1255                 return false;
1256
1257         evlist__for_each_entry(evlist, pos) {
1258                 if (pos->id_pos != evlist->id_pos ||
1259                     pos->is_pos != evlist->is_pos)
1260                         return false;
1261         }
1262
1263         return true;
1264 }
1265
1266 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1267 {
1268         struct perf_evsel *evsel;
1269
1270         if (evlist->combined_sample_type)
1271                 return evlist->combined_sample_type;
1272
1273         evlist__for_each_entry(evlist, evsel)
1274                 evlist->combined_sample_type |= evsel->attr.sample_type;
1275
1276         return evlist->combined_sample_type;
1277 }
1278
1279 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1280 {
1281         evlist->combined_sample_type = 0;
1282         return __perf_evlist__combined_sample_type(evlist);
1283 }
1284
1285 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1286 {
1287         struct perf_evsel *evsel;
1288         u64 branch_type = 0;
1289
1290         evlist__for_each_entry(evlist, evsel)
1291                 branch_type |= evsel->attr.branch_sample_type;
1292         return branch_type;
1293 }
1294
1295 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1296 {
1297         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1298         u64 read_format = first->attr.read_format;
1299         u64 sample_type = first->attr.sample_type;
1300
1301         evlist__for_each_entry(evlist, pos) {
1302                 if (read_format != pos->attr.read_format)
1303                         return false;
1304         }
1305
1306         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1307         if ((sample_type & PERF_SAMPLE_READ) &&
1308             !(read_format & PERF_FORMAT_ID)) {
1309                 return false;
1310         }
1311
1312         return true;
1313 }
1314
1315 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1316 {
1317         struct perf_evsel *first = perf_evlist__first(evlist);
1318         return first->attr.read_format;
1319 }
1320
1321 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1322 {
1323         struct perf_evsel *first = perf_evlist__first(evlist);
1324         struct perf_sample *data;
1325         u64 sample_type;
1326         u16 size = 0;
1327
1328         if (!first->attr.sample_id_all)
1329                 goto out;
1330
1331         sample_type = first->attr.sample_type;
1332
1333         if (sample_type & PERF_SAMPLE_TID)
1334                 size += sizeof(data->tid) * 2;
1335
1336        if (sample_type & PERF_SAMPLE_TIME)
1337                 size += sizeof(data->time);
1338
1339         if (sample_type & PERF_SAMPLE_ID)
1340                 size += sizeof(data->id);
1341
1342         if (sample_type & PERF_SAMPLE_STREAM_ID)
1343                 size += sizeof(data->stream_id);
1344
1345         if (sample_type & PERF_SAMPLE_CPU)
1346                 size += sizeof(data->cpu) * 2;
1347
1348         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1349                 size += sizeof(data->id);
1350 out:
1351         return size;
1352 }
1353
1354 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1355 {
1356         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1357
1358         evlist__for_each_entry_continue(evlist, pos) {
1359                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1360                         return false;
1361         }
1362
1363         return true;
1364 }
1365
1366 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1367 {
1368         struct perf_evsel *first = perf_evlist__first(evlist);
1369         return first->attr.sample_id_all;
1370 }
1371
1372 void perf_evlist__set_selected(struct perf_evlist *evlist,
1373                                struct perf_evsel *evsel)
1374 {
1375         evlist->selected = evsel;
1376 }
1377
1378 void perf_evlist__close(struct perf_evlist *evlist)
1379 {
1380         struct perf_evsel *evsel;
1381
1382         evlist__for_each_entry_reverse(evlist, evsel)
1383                 perf_evsel__close(evsel);
1384 }
1385
1386 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1387 {
1388         struct cpu_map    *cpus;
1389         struct thread_map *threads;
1390         int err = -ENOMEM;
1391
1392         /*
1393          * Try reading /sys/devices/system/cpu/online to get
1394          * an all cpus map.
1395          *
1396          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1397          * code needs an overhaul to properly forward the
1398          * error, and we may not want to do that fallback to a
1399          * default cpu identity map :-\
1400          */
1401         cpus = cpu_map__new(NULL);
1402         if (!cpus)
1403                 goto out;
1404
1405         threads = thread_map__new_dummy();
1406         if (!threads)
1407                 goto out_put;
1408
1409         perf_evlist__set_maps(evlist, cpus, threads);
1410 out:
1411         return err;
1412 out_put:
1413         cpu_map__put(cpus);
1414         goto out;
1415 }
1416
1417 int perf_evlist__open(struct perf_evlist *evlist)
1418 {
1419         struct perf_evsel *evsel;
1420         int err;
1421
1422         /*
1423          * Default: one fd per CPU, all threads, aka systemwide
1424          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1425          */
1426         if (evlist->threads == NULL && evlist->cpus == NULL) {
1427                 err = perf_evlist__create_syswide_maps(evlist);
1428                 if (err < 0)
1429                         goto out_err;
1430         }
1431
1432         perf_evlist__update_id_pos(evlist);
1433
1434         evlist__for_each_entry(evlist, evsel) {
1435                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1436                 if (err < 0)
1437                         goto out_err;
1438         }
1439
1440         return 0;
1441 out_err:
1442         perf_evlist__close(evlist);
1443         errno = -err;
1444         return err;
1445 }
1446
1447 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1448                                   const char *argv[], bool pipe_output,
1449                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1450 {
1451         int child_ready_pipe[2], go_pipe[2];
1452         char bf;
1453
1454         if (pipe(child_ready_pipe) < 0) {
1455                 perror("failed to create 'ready' pipe");
1456                 return -1;
1457         }
1458
1459         if (pipe(go_pipe) < 0) {
1460                 perror("failed to create 'go' pipe");
1461                 goto out_close_ready_pipe;
1462         }
1463
1464         evlist->workload.pid = fork();
1465         if (evlist->workload.pid < 0) {
1466                 perror("failed to fork");
1467                 goto out_close_pipes;
1468         }
1469
1470         if (!evlist->workload.pid) {
1471                 int ret;
1472
1473                 if (pipe_output)
1474                         dup2(2, 1);
1475
1476                 signal(SIGTERM, SIG_DFL);
1477
1478                 close(child_ready_pipe[0]);
1479                 close(go_pipe[1]);
1480                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1481
1482                 /*
1483                  * Tell the parent we're ready to go
1484                  */
1485                 close(child_ready_pipe[1]);
1486
1487                 /*
1488                  * Wait until the parent tells us to go.
1489                  */
1490                 ret = read(go_pipe[0], &bf, 1);
1491                 /*
1492                  * The parent will ask for the execvp() to be performed by
1493                  * writing exactly one byte, in workload.cork_fd, usually via
1494                  * perf_evlist__start_workload().
1495                  *
1496                  * For cancelling the workload without actually running it,
1497                  * the parent will just close workload.cork_fd, without writing
1498                  * anything, i.e. read will return zero and we just exit()
1499                  * here.
1500                  */
1501                 if (ret != 1) {
1502                         if (ret == -1)
1503                                 perror("unable to read pipe");
1504                         exit(ret);
1505                 }
1506
1507                 execvp(argv[0], (char **)argv);
1508
1509                 if (exec_error) {
1510                         union sigval val;
1511
1512                         val.sival_int = errno;
1513                         if (sigqueue(getppid(), SIGUSR1, val))
1514                                 perror(argv[0]);
1515                 } else
1516                         perror(argv[0]);
1517                 exit(-1);
1518         }
1519
1520         if (exec_error) {
1521                 struct sigaction act = {
1522                         .sa_flags     = SA_SIGINFO,
1523                         .sa_sigaction = exec_error,
1524                 };
1525                 sigaction(SIGUSR1, &act, NULL);
1526         }
1527
1528         if (target__none(target)) {
1529                 if (evlist->threads == NULL) {
1530                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1531                                 __func__, __LINE__);
1532                         goto out_close_pipes;
1533                 }
1534                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1535         }
1536
1537         close(child_ready_pipe[1]);
1538         close(go_pipe[0]);
1539         /*
1540          * wait for child to settle
1541          */
1542         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1543                 perror("unable to read pipe");
1544                 goto out_close_pipes;
1545         }
1546
1547         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1548         evlist->workload.cork_fd = go_pipe[1];
1549         close(child_ready_pipe[0]);
1550         return 0;
1551
1552 out_close_pipes:
1553         close(go_pipe[0]);
1554         close(go_pipe[1]);
1555 out_close_ready_pipe:
1556         close(child_ready_pipe[0]);
1557         close(child_ready_pipe[1]);
1558         return -1;
1559 }
1560
1561 int perf_evlist__start_workload(struct perf_evlist *evlist)
1562 {
1563         if (evlist->workload.cork_fd > 0) {
1564                 char bf = 0;
1565                 int ret;
1566                 /*
1567                  * Remove the cork, let it rip!
1568                  */
1569                 ret = write(evlist->workload.cork_fd, &bf, 1);
1570                 if (ret < 0)
1571                         perror("unable to write to pipe");
1572
1573                 close(evlist->workload.cork_fd);
1574                 return ret;
1575         }
1576
1577         return 0;
1578 }
1579
1580 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1581                               struct perf_sample *sample)
1582 {
1583         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1584
1585         if (!evsel)
1586                 return -EFAULT;
1587         return perf_evsel__parse_sample(evsel, event, sample);
1588 }
1589
1590 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1591                                         union perf_event *event,
1592                                         u64 *timestamp)
1593 {
1594         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1595
1596         if (!evsel)
1597                 return -EFAULT;
1598         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1599 }
1600
1601 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1602 {
1603         struct perf_evsel *evsel;
1604         size_t printed = 0;
1605
1606         evlist__for_each_entry(evlist, evsel) {
1607                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1608                                    perf_evsel__name(evsel));
1609         }
1610
1611         return printed + fprintf(fp, "\n");
1612 }
1613
1614 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1615                                int err, char *buf, size_t size)
1616 {
1617         int printed, value;
1618         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1619
1620         switch (err) {
1621         case EACCES:
1622         case EPERM:
1623                 printed = scnprintf(buf, size,
1624                                     "Error:\t%s.\n"
1625                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1626
1627                 value = perf_event_paranoid();
1628
1629                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1630
1631                 if (value >= 2) {
1632                         printed += scnprintf(buf + printed, size - printed,
1633                                              "For your workloads it needs to be <= 1\nHint:\t");
1634                 }
1635                 printed += scnprintf(buf + printed, size - printed,
1636                                      "For system wide tracing it needs to be set to -1.\n");
1637
1638                 printed += scnprintf(buf + printed, size - printed,
1639                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1640                                     "Hint:\tThe current value is %d.", value);
1641                 break;
1642         case EINVAL: {
1643                 struct perf_evsel *first = perf_evlist__first(evlist);
1644                 int max_freq;
1645
1646                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1647                         goto out_default;
1648
1649                 if (first->attr.sample_freq < (u64)max_freq)
1650                         goto out_default;
1651
1652                 printed = scnprintf(buf, size,
1653                                     "Error:\t%s.\n"
1654                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1655                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1656                                     emsg, max_freq, first->attr.sample_freq);
1657                 break;
1658         }
1659         default:
1660 out_default:
1661                 scnprintf(buf, size, "%s", emsg);
1662                 break;
1663         }
1664
1665         return 0;
1666 }
1667
1668 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1669 {
1670         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1671         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1672
1673         switch (err) {
1674         case EPERM:
1675                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1676                 printed += scnprintf(buf + printed, size - printed,
1677                                      "Error:\t%s.\n"
1678                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1679                                      "Hint:\tTried using %zd kB.\n",
1680                                      emsg, pages_max_per_user, pages_attempted);
1681
1682                 if (pages_attempted >= pages_max_per_user) {
1683                         printed += scnprintf(buf + printed, size - printed,
1684                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1685                                              pages_max_per_user + pages_attempted);
1686                 }
1687
1688                 printed += scnprintf(buf + printed, size - printed,
1689                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1690                 break;
1691         default:
1692                 scnprintf(buf, size, "%s", emsg);
1693                 break;
1694         }
1695
1696         return 0;
1697 }
1698
1699 void perf_evlist__to_front(struct perf_evlist *evlist,
1700                            struct perf_evsel *move_evsel)
1701 {
1702         struct perf_evsel *evsel, *n;
1703         LIST_HEAD(move);
1704
1705         if (move_evsel == perf_evlist__first(evlist))
1706                 return;
1707
1708         evlist__for_each_entry_safe(evlist, n, evsel) {
1709                 if (evsel->leader == move_evsel->leader)
1710                         list_move_tail(&evsel->node, &move);
1711         }
1712
1713         list_splice(&move, &evlist->entries);
1714 }
1715
1716 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1717                                      struct perf_evsel *tracking_evsel)
1718 {
1719         struct perf_evsel *evsel;
1720
1721         if (tracking_evsel->tracking)
1722                 return;
1723
1724         evlist__for_each_entry(evlist, evsel) {
1725                 if (evsel != tracking_evsel)
1726                         evsel->tracking = false;
1727         }
1728
1729         tracking_evsel->tracking = true;
1730 }
1731
1732 struct perf_evsel *
1733 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1734                                const char *str)
1735 {
1736         struct perf_evsel *evsel;
1737
1738         evlist__for_each_entry(evlist, evsel) {
1739                 if (!evsel->name)
1740                         continue;
1741                 if (strcmp(str, evsel->name) == 0)
1742                         return evsel;
1743         }
1744
1745         return NULL;
1746 }
1747
1748 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1749                                   enum bkw_mmap_state state)
1750 {
1751         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1752         enum action {
1753                 NONE,
1754                 PAUSE,
1755                 RESUME,
1756         } action = NONE;
1757
1758         if (!evlist->overwrite_mmap)
1759                 return;
1760
1761         switch (old_state) {
1762         case BKW_MMAP_NOTREADY: {
1763                 if (state != BKW_MMAP_RUNNING)
1764                         goto state_err;
1765                 break;
1766         }
1767         case BKW_MMAP_RUNNING: {
1768                 if (state != BKW_MMAP_DATA_PENDING)
1769                         goto state_err;
1770                 action = PAUSE;
1771                 break;
1772         }
1773         case BKW_MMAP_DATA_PENDING: {
1774                 if (state != BKW_MMAP_EMPTY)
1775                         goto state_err;
1776                 break;
1777         }
1778         case BKW_MMAP_EMPTY: {
1779                 if (state != BKW_MMAP_RUNNING)
1780                         goto state_err;
1781                 action = RESUME;
1782                 break;
1783         }
1784         default:
1785                 WARN_ONCE(1, "Shouldn't get there\n");
1786         }
1787
1788         evlist->bkw_mmap_state = state;
1789
1790         switch (action) {
1791         case PAUSE:
1792                 perf_evlist__pause(evlist);
1793                 break;
1794         case RESUME:
1795                 perf_evlist__resume(evlist);
1796                 break;
1797         case NONE:
1798         default:
1799                 break;
1800         }
1801
1802 state_err:
1803         return;
1804 }
1805
1806 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1807 {
1808         struct perf_evsel *evsel;
1809
1810         evlist__for_each_entry(evlist, evsel) {
1811                 if (!evsel->attr.exclude_kernel)
1812                         return false;
1813         }
1814
1815         return true;
1816 }