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