tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <poll.h>
  14 #include "cpumap.h"
  15 #include "thread_map.h"
  16 #include "target.h"
  17 #include "evlist.h"
  18 #include "evsel.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "asm/bug.h"
  22 #include "bpf-event.h"
  23 #include <signal.h>
  24 #include <unistd.h>
  25
  26 #include "parse-events.h"
  27 #include <subcmd/parse-options.h>
  28
  29 #include <fcntl.h>
  30 #include <sys/ioctl.h>
  31 #include <sys/mman.h>
  32
  33 #include <linux/bitops.h>
  34 #include <linux/hash.h>
  35 #include <linux/log2.h>
  36 #include <linux/err.h>
  37
  38 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  39 int sigqueue(pid_t pid, int sig, const union sigval value);
  40 #endif
  41
  42 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  43 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  44
  45 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  46                        struct thread_map *threads)
  47 {
  48         int i;
  49
  50         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  51                 INIT_HLIST_HEAD(&evlist->heads[i]);
  52         INIT_LIST_HEAD(&evlist->entries);
  53         perf_evlist__set_maps(evlist, cpus, threads);
  54         fdarray__init(&evlist->pollfd, 64);
  55         evlist->workload.pid = -1;
  56         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  57 }
  58
  59 struct perf_evlist *perf_evlist__new(void)
  60 {
  61         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  62
  63         if (evlist != NULL)
  64                 perf_evlist__init(evlist, NULL, NULL);
  65
  66         return evlist;
  67 }
  68
  69 struct perf_evlist *perf_evlist__new_default(void)
  70 {
  71         struct perf_evlist *evlist = perf_evlist__new();
  72
  73         if (evlist && perf_evlist__add_default(evlist)) {
  74                 perf_evlist__delete(evlist);
  75                 evlist = NULL;
  76         }
  77
  78         return evlist;
  79 }
  80
  81 struct perf_evlist *perf_evlist__new_dummy(void)
  82 {
  83         struct perf_evlist *evlist = perf_evlist__new();
  84
  85         if (evlist && perf_evlist__add_dummy(evlist)) {
  86                 perf_evlist__delete(evlist);
  87                 evlist = NULL;
  88         }
  89
  90         return evlist;
  91 }
  92
  93 /**
  94  * perf_evlist__set_id_pos - set the positions of event ids.
  95  * @evlist: selected event list
  96  *
  97  * Events with compatible sample types all have the same id_pos
  98  * and is_pos.  For convenience, put a copy on evlist.
  99  */
 100 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
 101 {
 102         struct perf_evsel *first = perf_evlist__first(evlist);
 103
 104         evlist->id_pos = first->id_pos;
 105         evlist->is_pos = first->is_pos;
 106 }
 107
 108 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 109 {
 110         struct perf_evsel *evsel;
 111
 112         evlist__for_each_entry(evlist, evsel)
 113                 perf_evsel__calc_id_pos(evsel);
 114
 115         perf_evlist__set_id_pos(evlist);
 116 }
 117
 118 static void perf_evlist__purge(struct perf_evlist *evlist)
 119 {
 120         struct perf_evsel *pos, *n;
 121
 122         evlist__for_each_entry_safe(evlist, n, pos) {
 123                 list_del_init(&pos->node);
 124                 pos->evlist = NULL;
 125                 perf_evsel__delete(pos);
 126         }
 127
 128         evlist->nr_entries = 0;
 129 }
 130
 131 void perf_evlist__exit(struct perf_evlist *evlist)
 132 {
 133         zfree(&evlist->mmap);
 134         zfree(&evlist->overwrite_mmap);
 135         fdarray__exit(&evlist->pollfd);
 136 }
 137
 138 void perf_evlist__delete(struct perf_evlist *evlist)
 139 {
 140         if (evlist == NULL)
 141                 return;
 142
 143         perf_evlist__munmap(evlist);
 144         perf_evlist__close(evlist);
 145         cpu_map__put(evlist->cpus);
 146         thread_map__put(evlist->threads);
 147         evlist->cpus = NULL;
 148         evlist->threads = NULL;
 149         perf_evlist__purge(evlist);
 150         perf_evlist__exit(evlist);
 151         free(evlist);
 152 }
 153
 154 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 155                                           struct perf_evsel *evsel)
 156 {
 157         /*
 158          * We already have cpus for evsel (via PMU sysfs) so
 159          * keep it, if there's no target cpu list defined.
 160          */
 161         if (!evsel->own_cpus || evlist->has_user_cpus) {
 162                 cpu_map__put(evsel->cpus);
 163                 evsel->cpus = cpu_map__get(evlist->cpus);
 164         } else if (evsel->cpus != evsel->own_cpus) {
 165                 cpu_map__put(evsel->cpus);
 166                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 167         }
 168
 169         thread_map__put(evsel->threads);
 170         evsel->threads = thread_map__get(evlist->threads);
 171 }
 172
 173 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 174 {
 175         struct perf_evsel *evsel;
 176
 177         evlist__for_each_entry(evlist, evsel)
 178                 __perf_evlist__propagate_maps(evlist, evsel);
 179 }
 180
 181 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 182 {
 183         entry->evlist = evlist;
 184         list_add_tail(&entry->node, &evlist->entries);
 185         entry->idx = evlist->nr_entries;
 186         entry->tracking = !entry->idx;
 187
 188         if (!evlist->nr_entries++)
 189                 perf_evlist__set_id_pos(evlist);
 190
 191         __perf_evlist__propagate_maps(evlist, entry);
 192 }
 193
 194 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 195 {
 196         evsel->evlist = NULL;
 197         list_del_init(&evsel->node);
 198         evlist->nr_entries -= 1;
 199 }
 200
 201 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 202                                    struct list_head *list)
 203 {
 204         struct perf_evsel *evsel, *temp;
 205
 206         __evlist__for_each_entry_safe(list, temp, evsel) {
 207                 list_del_init(&evsel->node);
 208                 perf_evlist__add(evlist, evsel);
 209         }
 210 }
 211
 212 void __perf_evlist__set_leader(struct list_head *list)
 213 {
 214         struct perf_evsel *evsel, *leader;
 215
 216         leader = list_entry(list->next, struct perf_evsel, node);
 217         evsel = list_entry(list->prev, struct perf_evsel, node);
 218
 219         leader->nr_members = evsel->idx - leader->idx + 1;
 220
 221         __evlist__for_each_entry(list, evsel) {
 222                 evsel->leader = leader;
 223         }
 224 }
 225
 226 void perf_evlist__set_leader(struct perf_evlist *evlist)
 227 {
 228         if (evlist->nr_entries) {
 229                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 230                 __perf_evlist__set_leader(&evlist->entries);
 231         }
 232 }
 233
 234 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *pattr)
 235 {
 236         struct perf_event_attr attr = {
 237                 .type           = PERF_TYPE_HARDWARE,
 238                 .config         = PERF_COUNT_HW_CPU_CYCLES,
 239                 .exclude_kernel = 1,
 240                 .precise_ip     = 3,
 241         };
 242
 243         event_attr_init(&attr);
 244
 245         /*
 246          * Unnamed union member, not supported as struct member named
 247          * initializer in older compilers such as gcc 4.4.7
 248          */
 249         attr.sample_period = 1;
 250
 251         while (attr.precise_ip != 0) {
 252                 int fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 253                 if (fd != -1) {
 254                         close(fd);
 255                         break;
 256                 }
 257                 --attr.precise_ip;
 258         }
 259
 260         pattr->precise_ip = attr.precise_ip;
 261 }
 262
 263 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 264 {
 265         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 266
 267         if (evsel == NULL)
 268                 return -ENOMEM;
 269
 270         perf_evlist__add(evlist, evsel);
 271         return 0;
 272 }
 273
 274 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 275 {
 276         struct perf_event_attr attr = {
 277                 .type   = PERF_TYPE_SOFTWARE,
 278                 .config = PERF_COUNT_SW_DUMMY,
 279                 .size   = sizeof(attr), /* to capture ABI version */
 280         };
 281         struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
 282
 283         if (evsel == NULL)
 284                 return -ENOMEM;
 285
 286         perf_evlist__add(evlist, evsel);
 287         return 0;
 288 }
 289
 290 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 291                                   struct perf_event_attr *attrs, size_t nr_attrs)
 292 {
 293         struct perf_evsel *evsel, *n;
 294         LIST_HEAD(head);
 295         size_t i;
 296
 297         for (i = 0; i < nr_attrs; i++) {
 298                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 299                 if (evsel == NULL)
 300                         goto out_delete_partial_list;
 301                 list_add_tail(&evsel->node, &head);
 302         }
 303
 304         perf_evlist__splice_list_tail(evlist, &head);
 305
 306         return 0;
 307
 308 out_delete_partial_list:
 309         __evlist__for_each_entry_safe(&head, n, evsel)
 310                 perf_evsel__delete(evsel);
 311         return -1;
 312 }
 313
 314 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 315                                      struct perf_event_attr *attrs, size_t nr_attrs)
 316 {
 317         size_t i;
 318
 319         for (i = 0; i < nr_attrs; i++)
 320                 event_attr_init(attrs + i);
 321
 322         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 323 }
 324
 325 struct perf_evsel *
 326 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 327 {
 328         struct perf_evsel *evsel;
 329
 330         evlist__for_each_entry(evlist, evsel) {
 331                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 332                     (int)evsel->attr.config == id)
 333                         return evsel;
 334         }
 335
 336         return NULL;
 337 }
 338
 339 struct perf_evsel *
 340 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 341                                      const char *name)
 342 {
 343         struct perf_evsel *evsel;
 344
 345         evlist__for_each_entry(evlist, evsel) {
 346                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 347                     (strcmp(evsel->name, name) == 0))
 348                         return evsel;
 349         }
 350
 351         return NULL;
 352 }
 353
 354 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 355                            const char *sys, const char *name, void *handler)
 356 {
 357         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 358
 359         if (IS_ERR(evsel))
 360                 return -1;
 361
 362         evsel->handler = handler;
 363         perf_evlist__add(evlist, evsel);
 364         return 0;
 365 }
 366
 367 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 368                                    struct perf_evsel *evsel)
 369 {
 370         if (evsel->system_wide)
 371                 return 1;
 372         else
 373                 return thread_map__nr(evlist->threads);
 374 }
 375
 376 void perf_evlist__disable(struct perf_evlist *evlist)
 377 {
 378         struct perf_evsel *pos;
 379
 380         evlist__for_each_entry(evlist, pos) {
 381                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
 382                         continue;
 383                 perf_evsel__disable(pos);
 384         }
 385
 386         evlist->enabled = false;
 387 }
 388
 389 void perf_evlist__enable(struct perf_evlist *evlist)
 390 {
 391         struct perf_evsel *pos;
 392
 393         evlist__for_each_entry(evlist, pos) {
 394                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 395                         continue;
 396                 perf_evsel__enable(pos);
 397         }
 398
 399         evlist->enabled = true;
 400 }
 401
 402 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 403 {
 404         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 405 }
 406
 407 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 408                                          struct perf_evsel *evsel, int cpu)
 409 {
 410         int thread;
 411         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 412
 413         if (!evsel->fd)
 414                 return -EINVAL;
 415
 416         for (thread = 0; thread < nr_threads; thread++) {
 417                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 418                 if (err)
 419                         return err;
 420         }
 421         return 0;
 422 }
 423
 424 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 425                                             struct perf_evsel *evsel,
 426                                             int thread)
 427 {
 428         int cpu;
 429         int nr_cpus = cpu_map__nr(evlist->cpus);
 430
 431         if (!evsel->fd)
 432                 return -EINVAL;
 433
 434         for (cpu = 0; cpu < nr_cpus; cpu++) {
 435                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 436                 if (err)
 437                         return err;
 438         }
 439         return 0;
 440 }
 441
 442 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 443                                   struct perf_evsel *evsel, int idx)
 444 {
 445         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 446
 447         if (per_cpu_mmaps)
 448                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 449         else
 450                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 451 }
 452
 453 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 454 {
 455         int nr_cpus = cpu_map__nr(evlist->cpus);
 456         int nr_threads = thread_map__nr(evlist->threads);
 457         int nfds = 0;
 458         struct perf_evsel *evsel;
 459
 460         evlist__for_each_entry(evlist, evsel) {
 461                 if (evsel->system_wide)
 462                         nfds += nr_cpus;
 463                 else
 464                         nfds += nr_cpus * nr_threads;
 465         }
 466
 467         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 468             fdarray__grow(&evlist->pollfd, nfds) < 0)
 469                 return -ENOMEM;
 470
 471         return 0;
 472 }
 473
 474 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 475                                      struct perf_mmap *map, short revent)
 476 {
 477         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 478         /*
 479          * Save the idx so that when we filter out fds POLLHUP'ed we can
 480          * close the associated evlist->mmap[] entry.
 481          */
 482         if (pos >= 0) {
 483                 evlist->pollfd.priv[pos].ptr = map;
 484
 485                 fcntl(fd, F_SETFL, O_NONBLOCK);
 486         }
 487
 488         return pos;
 489 }
 490
 491 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 492 {
 493         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 494 }
 495
 496 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 497                                          void *arg __maybe_unused)
 498 {
 499         struct perf_mmap *map = fda->priv[fd].ptr;
 500
 501         if (map)
 502                 perf_mmap__put(map);
 503 }
 504
 505 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 506 {
 507         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 508                                perf_evlist__munmap_filtered, NULL);
 509 }
 510
 511 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 512 {
 513         return fdarray__poll(&evlist->pollfd, timeout);
 514 }
 515
 516 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 517                                  struct perf_evsel *evsel,
 518                                  int cpu, int thread, u64 id)
 519 {
 520         int hash;
 521         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 522
 523         sid->id = id;
 524         sid->evsel = evsel;
 525         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 526         hlist_add_head(&sid->node, &evlist->heads[hash]);
 527 }
 528
 529 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 530                          int cpu, int thread, u64 id)
 531 {
 532         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 533         evsel->id[evsel->ids++] = id;
 534 }
 535
 536 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 537                            struct perf_evsel *evsel,
 538                            int cpu, int thread, int fd)
 539 {
 540         u64 read_data[4] = { 0, };
 541         int id_idx = 1; /* The first entry is the counter value */
 542         u64 id;
 543         int ret;
 544
 545         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 546         if (!ret)
 547                 goto add;
 548
 549         if (errno != ENOTTY)
 550                 return -1;
 551
 552         /* Legacy way to get event id.. All hail to old kernels! */
 553
 554         /*
 555          * This way does not work with group format read, so bail
 556          * out in that case.
 557          */
 558         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 559                 return -1;
 560
 561         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 562             read(fd, &read_data, sizeof(read_data)) == -1)
 563                 return -1;
 564
 565         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 566                 ++id_idx;
 567         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 568                 ++id_idx;
 569
 570         id = read_data[id_idx];
 571
 572  add:
 573         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 574         return 0;
 575 }
 576
 577 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 578                                      struct perf_evsel *evsel, int idx, int cpu,
 579                                      int thread)
 580 {
 581         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 582         sid->idx = idx;
 583         if (evlist->cpus && cpu >= 0)
 584                 sid->cpu = evlist->cpus->map[cpu];
 585         else
 586                 sid->cpu = -1;
 587         if (!evsel->system_wide && evlist->threads && thread >= 0)
 588                 sid->tid = thread_map__pid(evlist->threads, thread);
 589         else
 590                 sid->tid = -1;
 591 }
 592
 593 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 594 {
 595         struct hlist_head *head;
 596         struct perf_sample_id *sid;
 597         int hash;
 598
 599         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 600         head = &evlist->heads[hash];
 601
 602         hlist_for_each_entry(sid, head, node)
 603                 if (sid->id == id)
 604                         return sid;
 605
 606         return NULL;
 607 }
 608
 609 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 610 {
 611         struct perf_sample_id *sid;
 612
 613         if (evlist->nr_entries == 1 || !id)
 614                 return perf_evlist__first(evlist);
 615
 616         sid = perf_evlist__id2sid(evlist, id);
 617         if (sid)
 618                 return sid->evsel;
 619
 620         if (!perf_evlist__sample_id_all(evlist))
 621                 return perf_evlist__first(evlist);
 622
 623         return NULL;
 624 }
 625
 626 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 627                                                 u64 id)
 628 {
 629         struct perf_sample_id *sid;
 630
 631         if (!id)
 632                 return NULL;
 633
 634         sid = perf_evlist__id2sid(evlist, id);
 635         if (sid)
 636                 return sid->evsel;
 637
 638         return NULL;
 639 }
 640
 641 static int perf_evlist__event2id(struct perf_evlist *evlist,
 642                                  union perf_event *event, u64 *id)
 643 {
 644         const u64 *array = event->sample.array;
 645         ssize_t n;
 646
 647         n = (event->header.size - sizeof(event->header)) >> 3;
 648
 649         if (event->header.type == PERF_RECORD_SAMPLE) {
 650                 if (evlist->id_pos >= n)
 651                         return -1;
 652                 *id = array[evlist->id_pos];
 653         } else {
 654                 if (evlist->is_pos > n)
 655                         return -1;
 656                 n -= evlist->is_pos;
 657                 *id = array[n];
 658         }
 659         return 0;
 660 }
 661
 662 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 663                                             union perf_event *event)
 664 {
 665         struct perf_evsel *first = perf_evlist__first(evlist);
 666         struct hlist_head *head;
 667         struct perf_sample_id *sid;
 668         int hash;
 669         u64 id;
 670
 671         if (evlist->nr_entries == 1)
 672                 return first;
 673
 674         if (!first->attr.sample_id_all &&
 675             event->header.type != PERF_RECORD_SAMPLE)
 676                 return first;
 677
 678         if (perf_evlist__event2id(evlist, event, &id))
 679                 return NULL;
 680
 681         /* Synthesized events have an id of zero */
 682         if (!id)
 683                 return first;
 684
 685         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 686         head = &evlist->heads[hash];
 687
 688         hlist_for_each_entry(sid, head, node) {
 689                 if (sid->id == id)
 690                         return sid->evsel;
 691         }
 692         return NULL;
 693 }
 694
 695 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 696 {
 697         int i;
 698
 699         if (!evlist->overwrite_mmap)
 700                 return 0;
 701
 702         for (i = 0; i < evlist->nr_mmaps; i++) {
 703                 int fd = evlist->overwrite_mmap[i].fd;
 704                 int err;
 705
 706                 if (fd < 0)
 707                         continue;
 708                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 709                 if (err)
 710                         return err;
 711         }
 712         return 0;
 713 }
 714
 715 static int perf_evlist__pause(struct perf_evlist *evlist)
 716 {
 717         return perf_evlist__set_paused(evlist, true);
 718 }
 719
 720 static int perf_evlist__resume(struct perf_evlist *evlist)
 721 {
 722         return perf_evlist__set_paused(evlist, false);
 723 }
 724
 725 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 726 {
 727         int i;
 728
 729         if (evlist->mmap)
 730                 for (i = 0; i < evlist->nr_mmaps; i++)
 731                         perf_mmap__munmap(&evlist->mmap[i]);
 732
 733         if (evlist->overwrite_mmap)
 734                 for (i = 0; i < evlist->nr_mmaps; i++)
 735                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 736 }
 737
 738 void perf_evlist__munmap(struct perf_evlist *evlist)
 739 {
 740         perf_evlist__munmap_nofree(evlist);
 741         zfree(&evlist->mmap);
 742         zfree(&evlist->overwrite_mmap);
 743 }
 744
 745 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
 746                                                  bool overwrite)
 747 {
 748         int i;
 749         struct perf_mmap *map;
 750
 751         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 752         if (cpu_map__empty(evlist->cpus))
 753                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 754         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 755         if (!map)
 756                 return NULL;
 757
 758         for (i = 0; i < evlist->nr_mmaps; i++) {
 759                 map[i].fd = -1;
 760                 map[i].overwrite = overwrite;
 761                 /*
 762                  * When the perf_mmap() call is made we grab one refcount, plus
 763                  * one extra to let perf_mmap__consume() get the last
 764                  * events after all real references (perf_mmap__get()) are
 765                  * dropped.
 766                  *
 767                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 768                  * thus does perf_mmap__get() on it.
 769                  */
 770                 refcount_set(&map[i].refcnt, 0);
 771         }
 772         return map;
 773 }
 774
 775 static bool
 776 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 777                          struct perf_evsel *evsel)
 778 {
 779         if (evsel->attr.write_backward)
 780                 return false;
 781         return true;
 782 }
 783
 784 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 785                                        struct mmap_params *mp, int cpu_idx,
 786                                        int thread, int *_output, int *_output_overwrite)
 787 {
 788         struct perf_evsel *evsel;
 789         int revent;
 790         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 791
 792         evlist__for_each_entry(evlist, evsel) {
 793                 struct perf_mmap *maps = evlist->mmap;
 794                 int *output = _output;
 795                 int fd;
 796                 int cpu;
 797
 798                 mp->prot = PROT_READ | PROT_WRITE;
 799                 if (evsel->attr.write_backward) {
 800                         output = _output_overwrite;
 801                         maps = evlist->overwrite_mmap;
 802
 803                         if (!maps) {
 804                                 maps = perf_evlist__alloc_mmap(evlist, true);
 805                                 if (!maps)
 806                                         return -1;
 807                                 evlist->overwrite_mmap = maps;
 808                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 809                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 810                         }
 811                         mp->prot &= ~PROT_WRITE;
 812                 }
 813
 814                 if (evsel->system_wide && thread)
 815                         continue;
 816
 817                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 818                 if (cpu == -1)
 819                         continue;
 820
 821                 fd = FD(evsel, cpu, thread);
 822
 823                 if (*output == -1) {
 824                         *output = fd;
 825
 826                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 827                                 return -1;
 828                 } else {
 829                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 830                                 return -1;
 831
 832                         perf_mmap__get(&maps[idx]);
 833                 }
 834
 835                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 836
 837                 /*
 838                  * The system_wide flag causes a selected event to be opened
 839                  * always without a pid.  Consequently it will never get a
 840                  * POLLHUP, but it is used for tracking in combination with
 841                  * other events, so it should not need to be polled anyway.
 842                  * Therefore don't add it for polling.
 843                  */
 844                 if (!evsel->system_wide &&
 845                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 846                         perf_mmap__put(&maps[idx]);
 847                         return -1;
 848                 }
 849
 850                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 851                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 852                                                    fd) < 0)
 853                                 return -1;
 854                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 855                                                  thread);
 856                 }
 857         }
 858
 859         return 0;
 860 }
 861
 862 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 863                                      struct mmap_params *mp)
 864 {
 865         int cpu, thread;
 866         int nr_cpus = cpu_map__nr(evlist->cpus);
 867         int nr_threads = thread_map__nr(evlist->threads);
 868
 869         pr_debug2("perf event ring buffer mmapped per cpu\n");
 870         for (cpu = 0; cpu < nr_cpus; cpu++) {
 871                 int output = -1;
 872                 int output_overwrite = -1;
 873
 874                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 875                                               true);
 876
 877                 for (thread = 0; thread < nr_threads; thread++) {
 878                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 879                                                         thread, &output, &output_overwrite))
 880                                 goto out_unmap;
 881                 }
 882         }
 883
 884         return 0;
 885
 886 out_unmap:
 887         perf_evlist__munmap_nofree(evlist);
 888         return -1;
 889 }
 890
 891 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 892                                         struct mmap_params *mp)
 893 {
 894         int thread;
 895         int nr_threads = thread_map__nr(evlist->threads);
 896
 897         pr_debug2("perf event ring buffer mmapped per thread\n");
 898         for (thread = 0; thread < nr_threads; thread++) {
 899                 int output = -1;
 900                 int output_overwrite = -1;
 901
 902                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 903                                               false);
 904
 905                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 906                                                 &output, &output_overwrite))
 907                         goto out_unmap;
 908         }
 909
 910         return 0;
 911
 912 out_unmap:
 913         perf_evlist__munmap_nofree(evlist);
 914         return -1;
 915 }
 916
 917 unsigned long perf_event_mlock_kb_in_pages(void)
 918 {
 919         unsigned long pages;
 920         int max;
 921
 922         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 923                 /*
 924                  * Pick a once upon a time good value, i.e. things look
 925                  * strange since we can't read a sysctl value, but lets not
 926                  * die yet...
 927                  */
 928                 max = 512;
 929         } else {
 930                 max -= (page_size / 1024);
 931         }
 932
 933         pages = (max * 1024) / page_size;
 934         if (!is_power_of_2(pages))
 935                 pages = rounddown_pow_of_two(pages);
 936
 937         return pages;
 938 }
 939
 940 size_t perf_evlist__mmap_size(unsigned long pages)
 941 {
 942         if (pages == UINT_MAX)
 943                 pages = perf_event_mlock_kb_in_pages();
 944         else if (!is_power_of_2(pages))
 945                 return 0;
 946
 947         return (pages + 1) * page_size;
 948 }
 949
 950 static long parse_pages_arg(const char *str, unsigned long min,
 951                             unsigned long max)
 952 {
 953         unsigned long pages, val;
 954         static struct parse_tag tags[] = {
 955                 { .tag  = 'B', .mult = 1       },
 956                 { .tag  = 'K', .mult = 1 << 10 },
 957                 { .tag  = 'M', .mult = 1 << 20 },
 958                 { .tag  = 'G', .mult = 1 << 30 },
 959                 { .tag  = 0 },
 960         };
 961
 962         if (str == NULL)
 963                 return -EINVAL;
 964
 965         val = parse_tag_value(str, tags);
 966         if (val != (unsigned long) -1) {
 967                 /* we got file size value */
 968                 pages = PERF_ALIGN(val, page_size) / page_size;
 969         } else {
 970                 /* we got pages count value */
 971                 char *eptr;
 972                 pages = strtoul(str, &eptr, 10);
 973                 if (*eptr != '\0')
 974                         return -EINVAL;
 975         }
 976
 977         if (pages == 0 && min == 0) {
 978                 /* leave number of pages at 0 */
 979         } else if (!is_power_of_2(pages)) {
 980                 char buf[100];
 981
 982                 /* round pages up to next power of 2 */
 983                 pages = roundup_pow_of_two(pages);
 984                 if (!pages)
 985                         return -EINVAL;
 986
 987                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 988                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 989                         buf, pages);
 990         }
 991
 992         if (pages > max)
 993                 return -EINVAL;
 994
 995         return pages;
 996 }
 997
 998 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 999 {
1000         unsigned long max = UINT_MAX;
1001         long pages;
1002
1003         if (max > SIZE_MAX / page_size)
1004                 max = SIZE_MAX / page_size;
1005
1006         pages = parse_pages_arg(str, 1, max);
1007         if (pages < 0) {
1008                 pr_err("Invalid argument for --mmap_pages/-m\n");
1009                 return -1;
1010         }
1011
1012         *mmap_pages = pages;
1013         return 0;
1014 }
1015
1016 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1017                                   int unset __maybe_unused)
1018 {
1019         return __perf_evlist__parse_mmap_pages(opt->value, str);
1020 }
1021
1022 /**
1023  * perf_evlist__mmap_ex - Create mmaps to receive events.
1024  * @evlist: list of events
1025  * @pages: map length in pages
1026  * @overwrite: overwrite older events?
1027  * @auxtrace_pages - auxtrace map length in pages
1028  * @auxtrace_overwrite - overwrite older auxtrace data?
1029  *
1030  * If @overwrite is %false the user needs to signal event consumption using
1031  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1032  * automatically.
1033  *
1034  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1035  * consumption using auxtrace_mmap__write_tail().
1036  *
1037  * Return: %0 on success, negative error code otherwise.
1038  */
1039 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1040                          unsigned int auxtrace_pages,
1041                          bool auxtrace_overwrite, int nr_cblocks, int affinity)
1042 {
1043         struct perf_evsel *evsel;
1044         const struct cpu_map *cpus = evlist->cpus;
1045         const struct thread_map *threads = evlist->threads;
1046         /*
1047          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1048          * Its value is decided by evsel's write_backward.
1049          * So &mp should not be passed through const pointer.
1050          */
1051         struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity };
1052
1053         if (!evlist->mmap)
1054                 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1055         if (!evlist->mmap)
1056                 return -ENOMEM;
1057
1058         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1059                 return -ENOMEM;
1060
1061         evlist->mmap_len = perf_evlist__mmap_size(pages);
1062         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1063         mp.mask = evlist->mmap_len - page_size - 1;
1064
1065         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1066                                    auxtrace_pages, auxtrace_overwrite);
1067
1068         evlist__for_each_entry(evlist, evsel) {
1069                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1070                     evsel->sample_id == NULL &&
1071                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1072                         return -ENOMEM;
1073         }
1074
1075         if (cpu_map__empty(cpus))
1076                 return perf_evlist__mmap_per_thread(evlist, &mp);
1077
1078         return perf_evlist__mmap_per_cpu(evlist, &mp);
1079 }
1080
1081 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1082 {
1083         return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS);
1084 }
1085
1086 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1087 {
1088         bool all_threads = (target->per_thread && target->system_wide);
1089         struct cpu_map *cpus;
1090         struct thread_map *threads;
1091
1092         /*
1093          * If specify '-a' and '--per-thread' to perf record, perf record
1094          * will override '--per-thread'. target->per_thread = false and
1095          * target->system_wide = true.
1096          *
1097          * If specify '--per-thread' only to perf record,
1098          * target->per_thread = true and target->system_wide = false.
1099          *
1100          * So target->per_thread && target->system_wide is false.
1101          * For perf record, thread_map__new_str doesn't call
1102          * thread_map__new_all_cpus. That will keep perf record's
1103          * current behavior.
1104          *
1105          * For perf stat, it allows the case that target->per_thread and
1106          * target->system_wide are all true. It means to collect system-wide
1107          * per-thread data. thread_map__new_str will call
1108          * thread_map__new_all_cpus to enumerate all threads.
1109          */
1110         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1111                                       all_threads);
1112
1113         if (!threads)
1114                 return -1;
1115
1116         if (target__uses_dummy_map(target))
1117                 cpus = cpu_map__dummy_new();
1118         else
1119                 cpus = cpu_map__new(target->cpu_list);
1120
1121         if (!cpus)
1122                 goto out_delete_threads;
1123
1124         evlist->has_user_cpus = !!target->cpu_list;
1125
1126         perf_evlist__set_maps(evlist, cpus, threads);
1127
1128         return 0;
1129
1130 out_delete_threads:
1131         thread_map__put(threads);
1132         return -1;
1133 }
1134
1135 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1136                            struct thread_map *threads)
1137 {
1138         /*
1139          * Allow for the possibility that one or another of the maps isn't being
1140          * changed i.e. don't put it.  Note we are assuming the maps that are
1141          * being applied are brand new and evlist is taking ownership of the
1142          * original reference count of 1.  If that is not the case it is up to
1143          * the caller to increase the reference count.
1144          */
1145         if (cpus != evlist->cpus) {
1146                 cpu_map__put(evlist->cpus);
1147                 evlist->cpus = cpu_map__get(cpus);
1148         }
1149
1150         if (threads != evlist->threads) {
1151                 thread_map__put(evlist->threads);
1152                 evlist->threads = thread_map__get(threads);
1153         }
1154
1155         perf_evlist__propagate_maps(evlist);
1156 }
1157
1158 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1159                                    enum perf_event_sample_format bit)
1160 {
1161         struct perf_evsel *evsel;
1162
1163         evlist__for_each_entry(evlist, evsel)
1164                 __perf_evsel__set_sample_bit(evsel, bit);
1165 }
1166
1167 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1168                                      enum perf_event_sample_format bit)
1169 {
1170         struct perf_evsel *evsel;
1171
1172         evlist__for_each_entry(evlist, evsel)
1173                 __perf_evsel__reset_sample_bit(evsel, bit);
1174 }
1175
1176 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1177 {
1178         struct perf_evsel *evsel;
1179         int err = 0;
1180
1181         evlist__for_each_entry(evlist, evsel) {
1182                 if (evsel->filter == NULL)
1183                         continue;
1184
1185                 /*
1186                  * filters only work for tracepoint event, which doesn't have cpu limit.
1187                  * So evlist and evsel should always be same.
1188                  */
1189                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1190                 if (err) {
1191                         *err_evsel = evsel;
1192                         break;
1193                 }
1194         }
1195
1196         return err;
1197 }
1198
1199 int perf_evlist__set_tp_filter(struct perf_evlist *evlist, const char *filter)
1200 {
1201         struct perf_evsel *evsel;
1202         int err = 0;
1203
1204         evlist__for_each_entry(evlist, evsel) {
1205                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1206                         continue;
1207
1208                 err = perf_evsel__set_filter(evsel, filter);
1209                 if (err)
1210                         break;
1211         }
1212
1213         return err;
1214 }
1215
1216 int perf_evlist__set_tp_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1217 {
1218         char *filter;
1219         int ret = -1;
1220         size_t i;
1221
1222         for (i = 0; i < npids; ++i) {
1223                 if (i == 0) {
1224                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1225                                 return -1;
1226                 } else {
1227                         char *tmp;
1228
1229                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1230                                 goto out_free;
1231
1232                         free(filter);
1233                         filter = tmp;
1234                 }
1235         }
1236
1237         ret = perf_evlist__set_tp_filter(evlist, filter);
1238 out_free:
1239         free(filter);
1240         return ret;
1241 }
1242
1243 int perf_evlist__set_tp_filter_pid(struct perf_evlist *evlist, pid_t pid)
1244 {
1245         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1246 }
1247
1248 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1249 {
1250         struct perf_evsel *pos;
1251
1252         if (evlist->nr_entries == 1)
1253                 return true;
1254
1255         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1256                 return false;
1257
1258         evlist__for_each_entry(evlist, pos) {
1259                 if (pos->id_pos != evlist->id_pos ||
1260                     pos->is_pos != evlist->is_pos)
1261                         return false;
1262         }
1263
1264         return true;
1265 }
1266
1267 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1268 {
1269         struct perf_evsel *evsel;
1270
1271         if (evlist->combined_sample_type)
1272                 return evlist->combined_sample_type;
1273
1274         evlist__for_each_entry(evlist, evsel)
1275                 evlist->combined_sample_type |= evsel->attr.sample_type;
1276
1277         return evlist->combined_sample_type;
1278 }
1279
1280 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1281 {
1282         evlist->combined_sample_type = 0;
1283         return __perf_evlist__combined_sample_type(evlist);
1284 }
1285
1286 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1287 {
1288         struct perf_evsel *evsel;
1289         u64 branch_type = 0;
1290
1291         evlist__for_each_entry(evlist, evsel)
1292                 branch_type |= evsel->attr.branch_sample_type;
1293         return branch_type;
1294 }
1295
1296 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1297 {
1298         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1299         u64 read_format = first->attr.read_format;
1300         u64 sample_type = first->attr.sample_type;
1301
1302         evlist__for_each_entry(evlist, pos) {
1303                 if (read_format != pos->attr.read_format)
1304                         return false;
1305         }
1306
1307         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1308         if ((sample_type & PERF_SAMPLE_READ) &&
1309             !(read_format & PERF_FORMAT_ID)) {
1310                 return false;
1311         }
1312
1313         return true;
1314 }
1315
1316 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1317 {
1318         struct perf_evsel *first = perf_evlist__first(evlist);
1319         return first->attr.read_format;
1320 }
1321
1322 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1323 {
1324         struct perf_evsel *first = perf_evlist__first(evlist);
1325         struct perf_sample *data;
1326         u64 sample_type;
1327         u16 size = 0;
1328
1329         if (!first->attr.sample_id_all)
1330                 goto out;
1331
1332         sample_type = first->attr.sample_type;
1333
1334         if (sample_type & PERF_SAMPLE_TID)
1335                 size += sizeof(data->tid) * 2;
1336
1337        if (sample_type & PERF_SAMPLE_TIME)
1338                 size += sizeof(data->time);
1339
1340         if (sample_type & PERF_SAMPLE_ID)
1341                 size += sizeof(data->id);
1342
1343         if (sample_type & PERF_SAMPLE_STREAM_ID)
1344                 size += sizeof(data->stream_id);
1345
1346         if (sample_type & PERF_SAMPLE_CPU)
1347                 size += sizeof(data->cpu) * 2;
1348
1349         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1350                 size += sizeof(data->id);
1351 out:
1352         return size;
1353 }
1354
1355 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1356 {
1357         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1358
1359         evlist__for_each_entry_continue(evlist, pos) {
1360                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1361                         return false;
1362         }
1363
1364         return true;
1365 }
1366
1367 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1368 {
1369         struct perf_evsel *first = perf_evlist__first(evlist);
1370         return first->attr.sample_id_all;
1371 }
1372
1373 void perf_evlist__set_selected(struct perf_evlist *evlist,
1374                                struct perf_evsel *evsel)
1375 {
1376         evlist->selected = evsel;
1377 }
1378
1379 void perf_evlist__close(struct perf_evlist *evlist)
1380 {
1381         struct perf_evsel *evsel;
1382
1383         evlist__for_each_entry_reverse(evlist, evsel)
1384                 perf_evsel__close(evsel);
1385 }
1386
1387 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1388 {
1389         struct cpu_map    *cpus;
1390         struct thread_map *threads;
1391         int err = -ENOMEM;
1392
1393         /*
1394          * Try reading /sys/devices/system/cpu/online to get
1395          * an all cpus map.
1396          *
1397          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1398          * code needs an overhaul to properly forward the
1399          * error, and we may not want to do that fallback to a
1400          * default cpu identity map :-\
1401          */
1402         cpus = cpu_map__new(NULL);
1403         if (!cpus)
1404                 goto out;
1405
1406         threads = thread_map__new_dummy();
1407         if (!threads)
1408                 goto out_put;
1409
1410         perf_evlist__set_maps(evlist, cpus, threads);
1411 out:
1412         return err;
1413 out_put:
1414         cpu_map__put(cpus);
1415         goto out;
1416 }
1417
1418 int perf_evlist__open(struct perf_evlist *evlist)
1419 {
1420         struct perf_evsel *evsel;
1421         int err;
1422
1423         /*
1424          * Default: one fd per CPU, all threads, aka systemwide
1425          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1426          */
1427         if (evlist->threads == NULL && evlist->cpus == NULL) {
1428                 err = perf_evlist__create_syswide_maps(evlist);
1429                 if (err < 0)
1430                         goto out_err;
1431         }
1432
1433         perf_evlist__update_id_pos(evlist);
1434
1435         evlist__for_each_entry(evlist, evsel) {
1436                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1437                 if (err < 0)
1438                         goto out_err;
1439         }
1440
1441         return 0;
1442 out_err:
1443         perf_evlist__close(evlist);
1444         errno = -err;
1445         return err;
1446 }
1447
1448 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1449                                   const char *argv[], bool pipe_output,
1450                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1451 {
1452         int child_ready_pipe[2], go_pipe[2];
1453         char bf;
1454
1455         if (pipe(child_ready_pipe) < 0) {
1456                 perror("failed to create 'ready' pipe");
1457                 return -1;
1458         }
1459
1460         if (pipe(go_pipe) < 0) {
1461                 perror("failed to create 'go' pipe");
1462                 goto out_close_ready_pipe;
1463         }
1464
1465         evlist->workload.pid = fork();
1466         if (evlist->workload.pid < 0) {
1467                 perror("failed to fork");
1468                 goto out_close_pipes;
1469         }
1470
1471         if (!evlist->workload.pid) {
1472                 int ret;
1473
1474                 if (pipe_output)
1475                         dup2(2, 1);
1476
1477                 signal(SIGTERM, SIG_DFL);
1478
1479                 close(child_ready_pipe[0]);
1480                 close(go_pipe[1]);
1481                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1482
1483                 /*
1484                  * Tell the parent we're ready to go
1485                  */
1486                 close(child_ready_pipe[1]);
1487
1488                 /*
1489                  * Wait until the parent tells us to go.
1490                  */
1491                 ret = read(go_pipe[0], &bf, 1);
1492                 /*
1493                  * The parent will ask for the execvp() to be performed by
1494                  * writing exactly one byte, in workload.cork_fd, usually via
1495                  * perf_evlist__start_workload().
1496                  *
1497                  * For cancelling the workload without actually running it,
1498                  * the parent will just close workload.cork_fd, without writing
1499                  * anything, i.e. read will return zero and we just exit()
1500                  * here.
1501                  */
1502                 if (ret != 1) {
1503                         if (ret == -1)
1504                                 perror("unable to read pipe");
1505                         exit(ret);
1506                 }
1507
1508                 execvp(argv[0], (char **)argv);
1509
1510                 if (exec_error) {
1511                         union sigval val;
1512
1513                         val.sival_int = errno;
1514                         if (sigqueue(getppid(), SIGUSR1, val))
1515                                 perror(argv[0]);
1516                 } else
1517                         perror(argv[0]);
1518                 exit(-1);
1519         }
1520
1521         if (exec_error) {
1522                 struct sigaction act = {
1523                         .sa_flags     = SA_SIGINFO,
1524                         .sa_sigaction = exec_error,
1525                 };
1526                 sigaction(SIGUSR1, &act, NULL);
1527         }
1528
1529         if (target__none(target)) {
1530                 if (evlist->threads == NULL) {
1531                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1532                                 __func__, __LINE__);
1533                         goto out_close_pipes;
1534                 }
1535                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1536         }
1537
1538         close(child_ready_pipe[1]);
1539         close(go_pipe[0]);
1540         /*
1541          * wait for child to settle
1542          */
1543         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1544                 perror("unable to read pipe");
1545                 goto out_close_pipes;
1546         }
1547
1548         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1549         evlist->workload.cork_fd = go_pipe[1];
1550         close(child_ready_pipe[0]);
1551         return 0;
1552
1553 out_close_pipes:
1554         close(go_pipe[0]);
1555         close(go_pipe[1]);
1556 out_close_ready_pipe:
1557         close(child_ready_pipe[0]);
1558         close(child_ready_pipe[1]);
1559         return -1;
1560 }
1561
1562 int perf_evlist__start_workload(struct perf_evlist *evlist)
1563 {
1564         if (evlist->workload.cork_fd > 0) {
1565                 char bf = 0;
1566                 int ret;
1567                 /*
1568                  * Remove the cork, let it rip!
1569                  */
1570                 ret = write(evlist->workload.cork_fd, &bf, 1);
1571                 if (ret < 0)
1572                         perror("unable to write to pipe");
1573
1574                 close(evlist->workload.cork_fd);
1575                 return ret;
1576         }
1577
1578         return 0;
1579 }
1580
1581 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1582                               struct perf_sample *sample)
1583 {
1584         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1585
1586         if (!evsel)
1587                 return -EFAULT;
1588         return perf_evsel__parse_sample(evsel, event, sample);
1589 }
1590
1591 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1592                                         union perf_event *event,
1593                                         u64 *timestamp)
1594 {
1595         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1596
1597         if (!evsel)
1598                 return -EFAULT;
1599         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1600 }
1601
1602 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1603 {
1604         struct perf_evsel *evsel;
1605         size_t printed = 0;
1606
1607         evlist__for_each_entry(evlist, evsel) {
1608                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1609                                    perf_evsel__name(evsel));
1610         }
1611
1612         return printed + fprintf(fp, "\n");
1613 }
1614
1615 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1616                                int err, char *buf, size_t size)
1617 {
1618         int printed, value;
1619         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1620
1621         switch (err) {
1622         case EACCES:
1623         case EPERM:
1624                 printed = scnprintf(buf, size,
1625                                     "Error:\t%s.\n"
1626                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1627
1628                 value = perf_event_paranoid();
1629
1630                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1631
1632                 if (value >= 2) {
1633                         printed += scnprintf(buf + printed, size - printed,
1634                                              "For your workloads it needs to be <= 1\nHint:\t");
1635                 }
1636                 printed += scnprintf(buf + printed, size - printed,
1637                                      "For system wide tracing it needs to be set to -1.\n");
1638
1639                 printed += scnprintf(buf + printed, size - printed,
1640                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1641                                     "Hint:\tThe current value is %d.", value);
1642                 break;
1643         case EINVAL: {
1644                 struct perf_evsel *first = perf_evlist__first(evlist);
1645                 int max_freq;
1646
1647                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1648                         goto out_default;
1649
1650                 if (first->attr.sample_freq < (u64)max_freq)
1651                         goto out_default;
1652
1653                 printed = scnprintf(buf, size,
1654                                     "Error:\t%s.\n"
1655                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1656                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1657                                     emsg, max_freq, first->attr.sample_freq);
1658                 break;
1659         }
1660         default:
1661 out_default:
1662                 scnprintf(buf, size, "%s", emsg);
1663                 break;
1664         }
1665
1666         return 0;
1667 }
1668
1669 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1670 {
1671         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1672         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1673
1674         switch (err) {
1675         case EPERM:
1676                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1677                 printed += scnprintf(buf + printed, size - printed,
1678                                      "Error:\t%s.\n"
1679                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1680                                      "Hint:\tTried using %zd kB.\n",
1681                                      emsg, pages_max_per_user, pages_attempted);
1682
1683                 if (pages_attempted >= pages_max_per_user) {
1684                         printed += scnprintf(buf + printed, size - printed,
1685                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1686                                              pages_max_per_user + pages_attempted);
1687                 }
1688
1689                 printed += scnprintf(buf + printed, size - printed,
1690                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1691                 break;
1692         default:
1693                 scnprintf(buf, size, "%s", emsg);
1694                 break;
1695         }
1696
1697         return 0;
1698 }
1699
1700 void perf_evlist__to_front(struct perf_evlist *evlist,
1701                            struct perf_evsel *move_evsel)
1702 {
1703         struct perf_evsel *evsel, *n;
1704         LIST_HEAD(move);
1705
1706         if (move_evsel == perf_evlist__first(evlist))
1707                 return;
1708
1709         evlist__for_each_entry_safe(evlist, n, evsel) {
1710                 if (evsel->leader == move_evsel->leader)
1711                         list_move_tail(&evsel->node, &move);
1712         }
1713
1714         list_splice(&move, &evlist->entries);
1715 }
1716
1717 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1718                                      struct perf_evsel *tracking_evsel)
1719 {
1720         struct perf_evsel *evsel;
1721
1722         if (tracking_evsel->tracking)
1723                 return;
1724
1725         evlist__for_each_entry(evlist, evsel) {
1726                 if (evsel != tracking_evsel)
1727                         evsel->tracking = false;
1728         }
1729
1730         tracking_evsel->tracking = true;
1731 }
1732
1733 struct perf_evsel *
1734 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1735                                const char *str)
1736 {
1737         struct perf_evsel *evsel;
1738
1739         evlist__for_each_entry(evlist, evsel) {
1740                 if (!evsel->name)
1741                         continue;
1742                 if (strcmp(str, evsel->name) == 0)
1743                         return evsel;
1744         }
1745
1746         return NULL;
1747 }
1748
1749 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1750                                   enum bkw_mmap_state state)
1751 {
1752         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1753         enum action {
1754                 NONE,
1755                 PAUSE,
1756                 RESUME,
1757         } action = NONE;
1758
1759         if (!evlist->overwrite_mmap)
1760                 return;
1761
1762         switch (old_state) {
1763         case BKW_MMAP_NOTREADY: {
1764                 if (state != BKW_MMAP_RUNNING)
1765                         goto state_err;
1766                 break;
1767         }
1768         case BKW_MMAP_RUNNING: {
1769                 if (state != BKW_MMAP_DATA_PENDING)
1770                         goto state_err;
1771                 action = PAUSE;
1772                 break;
1773         }
1774         case BKW_MMAP_DATA_PENDING: {
1775                 if (state != BKW_MMAP_EMPTY)
1776                         goto state_err;
1777                 break;
1778         }
1779         case BKW_MMAP_EMPTY: {
1780                 if (state != BKW_MMAP_RUNNING)
1781                         goto state_err;
1782                 action = RESUME;
1783                 break;
1784         }
1785         default:
1786                 WARN_ONCE(1, "Shouldn't get there\n");
1787         }
1788
1789         evlist->bkw_mmap_state = state;
1790
1791         switch (action) {
1792         case PAUSE:
1793                 perf_evlist__pause(evlist);
1794                 break;
1795         case RESUME:
1796                 perf_evlist__resume(evlist);
1797                 break;
1798         case NONE:
1799         default:
1800                 break;
1801         }
1802
1803 state_err:
1804         return;
1805 }
1806
1807 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1808 {
1809         struct perf_evsel *evsel;
1810
1811         evlist__for_each_entry(evlist, evsel) {
1812                 if (!evsel->attr.exclude_kernel)
1813                         return false;
1814         }
1815
1816         return true;
1817 }
1818
1819 /*
1820  * Events in data file are not collect in groups, but we still want
1821  * the group display. Set the artificial group and set the leader's
1822  * forced_leader flag to notify the display code.
1823  */
1824 void perf_evlist__force_leader(struct perf_evlist *evlist)
1825 {
1826         if (!evlist->nr_groups) {
1827                 struct perf_evsel *leader = perf_evlist__first(evlist);
1828
1829                 perf_evlist__set_leader(evlist);
1830                 leader->forced_leader = true;
1831         }
1832 }
1833
1834 struct perf_evsel *perf_evlist__reset_weak_group(struct perf_evlist *evsel_list,
1835                                                  struct perf_evsel *evsel)
1836 {
1837         struct perf_evsel *c2, *leader;
1838         bool is_open = true;
1839
1840         leader = evsel->leader;
1841         pr_debug("Weak group for %s/%d failed\n",
1842                         leader->name, leader->nr_members);
1843
1844         /*
1845          * for_each_group_member doesn't work here because it doesn't
1846          * include the first entry.
1847          */
1848         evlist__for_each_entry(evsel_list, c2) {
1849                 if (c2 == evsel)
1850                         is_open = false;
1851                 if (c2->leader == leader) {
1852                         if (is_open)
1853                                 perf_evsel__close(c2);
1854                         c2->leader = c2;
1855                         c2->nr_members = 0;
1856                 }
1857         }
1858         return leader;
1859 }
1860
1861 int perf_evlist__add_sb_event(struct perf_evlist **evlist,
1862                               struct perf_event_attr *attr,
1863                               perf_evsel__sb_cb_t cb,
1864                               void *data)
1865 {
1866         struct perf_evsel *evsel;
1867         bool new_evlist = (*evlist) == NULL;
1868
1869         if (*evlist == NULL)
1870                 *evlist = perf_evlist__new();
1871         if (*evlist == NULL)
1872                 return -1;
1873
1874         if (!attr->sample_id_all) {
1875                 pr_warning("enabling sample_id_all for all side band events\n");
1876                 attr->sample_id_all = 1;
1877         }
1878
1879         evsel = perf_evsel__new_idx(attr, (*evlist)->nr_entries);
1880         if (!evsel)
1881                 goto out_err;
1882
1883         evsel->side_band.cb = cb;
1884         evsel->side_band.data = data;
1885         perf_evlist__add(*evlist, evsel);
1886         return 0;
1887
1888 out_err:
1889         if (new_evlist) {
1890                 perf_evlist__delete(*evlist);
1891                 *evlist = NULL;
1892         }
1893         return -1;
1894 }
1895
1896 static void *perf_evlist__poll_thread(void *arg)
1897 {
1898         struct perf_evlist *evlist = arg;
1899         bool draining = false;
1900         int i;
1901
1902         while (draining || !(evlist->thread.done)) {
1903                 if (draining)
1904                         draining = false;
1905                 else if (evlist->thread.done)
1906                         draining = true;
1907
1908                 if (!draining)
1909                         perf_evlist__poll(evlist, 1000);
1910
1911                 for (i = 0; i < evlist->nr_mmaps; i++) {
1912                         struct perf_mmap *map = &evlist->mmap[i];
1913                         union perf_event *event;
1914
1915                         if (perf_mmap__read_init(map))
1916                                 continue;
1917                         while ((event = perf_mmap__read_event(map)) != NULL) {
1918                                 struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1919
1920                                 if (evsel && evsel->side_band.cb)
1921                                         evsel->side_band.cb(event, evsel->side_band.data);
1922                                 else
1923                                         pr_warning("cannot locate proper evsel for the side band event\n");
1924
1925                                 perf_mmap__consume(map);
1926                         }
1927                         perf_mmap__read_done(map);
1928                 }
1929         }
1930         return NULL;
1931 }
1932
1933 int perf_evlist__start_sb_thread(struct perf_evlist *evlist,
1934                                  struct target *target)
1935 {
1936         struct perf_evsel *counter;
1937
1938         if (!evlist)
1939                 return 0;
1940
1941         if (perf_evlist__create_maps(evlist, target))
1942                 goto out_delete_evlist;
1943
1944         evlist__for_each_entry(evlist, counter) {
1945                 if (perf_evsel__open(counter, evlist->cpus,
1946                                      evlist->threads) < 0)
1947                         goto out_delete_evlist;
1948         }
1949
1950         if (perf_evlist__mmap(evlist, UINT_MAX))
1951                 goto out_delete_evlist;
1952
1953         evlist__for_each_entry(evlist, counter) {
1954                 if (perf_evsel__enable(counter))
1955                         goto out_delete_evlist;
1956         }
1957
1958         evlist->thread.done = 0;
1959         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1960                 goto out_delete_evlist;
1961
1962         return 0;
1963
1964 out_delete_evlist:
1965         perf_evlist__delete(evlist);
1966         evlist = NULL;
1967         return -1;
1968 }
1969
1970 void perf_evlist__stop_sb_thread(struct perf_evlist *evlist)
1971 {
1972         if (!evlist)
1973                 return;
1974         evlist->thread.done = 1;
1975         pthread_join(evlist->thread.th, NULL);
1976         perf_evlist__delete(evlist);
1977 }