tools/perf/util/evlist.c

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