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 "asm/bug.h"
  19 #include <unistd.h>
  20
  21 #include "parse-events.h"
  22 #include <subcmd/parse-options.h>
  23
  24 #include <sys/mman.h>
  25
  26 #include <linux/bitops.h>
  27 #include <linux/hash.h>
  28 #include <linux/log2.h>
  29 #include <linux/err.h>
  30
  31 static void perf_mmap__munmap(struct perf_mmap *map);
  32 static void perf_mmap__put(struct perf_mmap *map);
  33
  34 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  35 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  36
  37 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  38                        struct thread_map *threads)
  39 {
  40         int i;
  41
  42         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  43                 INIT_HLIST_HEAD(&evlist->heads[i]);
  44         INIT_LIST_HEAD(&evlist->entries);
  45         perf_evlist__set_maps(evlist, cpus, threads);
  46         fdarray__init(&evlist->pollfd, 64);
  47         evlist->workload.pid = -1;
  48         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  49 }
  50
  51 struct perf_evlist *perf_evlist__new(void)
  52 {
  53         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  54
  55         if (evlist != NULL)
  56                 perf_evlist__init(evlist, NULL, NULL);
  57
  58         return evlist;
  59 }
  60
  61 struct perf_evlist *perf_evlist__new_default(void)
  62 {
  63         struct perf_evlist *evlist = perf_evlist__new();
  64
  65         if (evlist && perf_evlist__add_default(evlist)) {
  66                 perf_evlist__delete(evlist);
  67                 evlist = NULL;
  68         }
  69
  70         return evlist;
  71 }
  72
  73 struct perf_evlist *perf_evlist__new_dummy(void)
  74 {
  75         struct perf_evlist *evlist = perf_evlist__new();
  76
  77         if (evlist && perf_evlist__add_dummy(evlist)) {
  78                 perf_evlist__delete(evlist);
  79                 evlist = NULL;
  80         }
  81
  82         return evlist;
  83 }
  84
  85 /**
  86  * perf_evlist__set_id_pos - set the positions of event ids.
  87  * @evlist: selected event list
  88  *
  89  * Events with compatible sample types all have the same id_pos
  90  * and is_pos.  For convenience, put a copy on evlist.
  91  */
  92 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  93 {
  94         struct perf_evsel *first = perf_evlist__first(evlist);
  95
  96         evlist->id_pos = first->id_pos;
  97         evlist->is_pos = first->is_pos;
  98 }
  99
 100 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 101 {
 102         struct perf_evsel *evsel;
 103
 104         evlist__for_each_entry(evlist, evsel)
 105                 perf_evsel__calc_id_pos(evsel);
 106
 107         perf_evlist__set_id_pos(evlist);
 108 }
 109
 110 static void perf_evlist__purge(struct perf_evlist *evlist)
 111 {
 112         struct perf_evsel *pos, *n;
 113
 114         evlist__for_each_entry_safe(evlist, n, pos) {
 115                 list_del_init(&pos->node);
 116                 pos->evlist = NULL;
 117                 perf_evsel__delete(pos);
 118         }
 119
 120         evlist->nr_entries = 0;
 121 }
 122
 123 void perf_evlist__exit(struct perf_evlist *evlist)
 124 {
 125         zfree(&evlist->mmap);
 126         zfree(&evlist->backward_mmap);
 127         fdarray__exit(&evlist->pollfd);
 128 }
 129
 130 void perf_evlist__delete(struct perf_evlist *evlist)
 131 {
 132         if (evlist == NULL)
 133                 return;
 134
 135         perf_evlist__munmap(evlist);
 136         perf_evlist__close(evlist);
 137         cpu_map__put(evlist->cpus);
 138         thread_map__put(evlist->threads);
 139         evlist->cpus = NULL;
 140         evlist->threads = NULL;
 141         perf_evlist__purge(evlist);
 142         perf_evlist__exit(evlist);
 143         free(evlist);
 144 }
 145
 146 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 147                                           struct perf_evsel *evsel)
 148 {
 149         /*
 150          * We already have cpus for evsel (via PMU sysfs) so
 151          * keep it, if there's no target cpu list defined.
 152          */
 153         if (!evsel->own_cpus || evlist->has_user_cpus) {
 154                 cpu_map__put(evsel->cpus);
 155                 evsel->cpus = cpu_map__get(evlist->cpus);
 156         } else if (evsel->cpus != evsel->own_cpus) {
 157                 cpu_map__put(evsel->cpus);
 158                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 159         }
 160
 161         thread_map__put(evsel->threads);
 162         evsel->threads = thread_map__get(evlist->threads);
 163 }
 164
 165 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 166 {
 167         struct perf_evsel *evsel;
 168
 169         evlist__for_each_entry(evlist, evsel)
 170                 __perf_evlist__propagate_maps(evlist, evsel);
 171 }
 172
 173 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 174 {
 175         entry->evlist = evlist;
 176         list_add_tail(&entry->node, &evlist->entries);
 177         entry->idx = evlist->nr_entries;
 178         entry->tracking = !entry->idx;
 179
 180         if (!evlist->nr_entries++)
 181                 perf_evlist__set_id_pos(evlist);
 182
 183         __perf_evlist__propagate_maps(evlist, entry);
 184 }
 185
 186 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 187 {
 188         evsel->evlist = NULL;
 189         list_del_init(&evsel->node);
 190         evlist->nr_entries -= 1;
 191 }
 192
 193 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 194                                    struct list_head *list)
 195 {
 196         struct perf_evsel *evsel, *temp;
 197
 198         __evlist__for_each_entry_safe(list, temp, evsel) {
 199                 list_del_init(&evsel->node);
 200                 perf_evlist__add(evlist, evsel);
 201         }
 202 }
 203
 204 void __perf_evlist__set_leader(struct list_head *list)
 205 {
 206         struct perf_evsel *evsel, *leader;
 207
 208         leader = list_entry(list->next, struct perf_evsel, node);
 209         evsel = list_entry(list->prev, struct perf_evsel, node);
 210
 211         leader->nr_members = evsel->idx - leader->idx + 1;
 212
 213         __evlist__for_each_entry(list, evsel) {
 214                 evsel->leader = leader;
 215         }
 216 }
 217
 218 void perf_evlist__set_leader(struct perf_evlist *evlist)
 219 {
 220         if (evlist->nr_entries) {
 221                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 222                 __perf_evlist__set_leader(&evlist->entries);
 223         }
 224 }
 225
 226 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 227 {
 228         attr->precise_ip = 3;
 229
 230         while (attr->precise_ip != 0) {
 231                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 232                 if (fd != -1) {
 233                         close(fd);
 234                         break;
 235                 }
 236                 --attr->precise_ip;
 237         }
 238 }
 239
 240 int perf_evlist__add_default(struct perf_evlist *evlist)
 241 {
 242         struct perf_event_attr attr = {
 243                 .type = PERF_TYPE_HARDWARE,
 244                 .config = PERF_COUNT_HW_CPU_CYCLES,
 245         };
 246         struct perf_evsel *evsel;
 247
 248         event_attr_init(&attr);
 249
 250         perf_event_attr__set_max_precise_ip(&attr);
 251
 252         evsel = perf_evsel__new(&attr);
 253         if (evsel == NULL)
 254                 goto error;
 255
 256         /* use asprintf() because free(evsel) assumes name is allocated */
 257         if (asprintf(&evsel->name, "cycles%.*s",
 258                      attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
 259                 goto error_free;
 260
 261         perf_evlist__add(evlist, evsel);
 262         return 0;
 263 error_free:
 264         perf_evsel__delete(evsel);
 265 error:
 266         return -ENOMEM;
 267 }
 268
 269 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 270 {
 271         struct perf_event_attr attr = {
 272                 .type   = PERF_TYPE_SOFTWARE,
 273                 .config = PERF_COUNT_SW_DUMMY,
 274                 .size   = sizeof(attr), /* to capture ABI version */
 275         };
 276         struct perf_evsel *evsel = perf_evsel__new(&attr);
 277
 278         if (evsel == NULL)
 279                 return -ENOMEM;
 280
 281         perf_evlist__add(evlist, evsel);
 282         return 0;
 283 }
 284
 285 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 286                                   struct perf_event_attr *attrs, size_t nr_attrs)
 287 {
 288         struct perf_evsel *evsel, *n;
 289         LIST_HEAD(head);
 290         size_t i;
 291
 292         for (i = 0; i < nr_attrs; i++) {
 293                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 294                 if (evsel == NULL)
 295                         goto out_delete_partial_list;
 296                 list_add_tail(&evsel->node, &head);
 297         }
 298
 299         perf_evlist__splice_list_tail(evlist, &head);
 300
 301         return 0;
 302
 303 out_delete_partial_list:
 304         __evlist__for_each_entry_safe(&head, n, evsel)
 305                 perf_evsel__delete(evsel);
 306         return -1;
 307 }
 308
 309 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 310                                      struct perf_event_attr *attrs, size_t nr_attrs)
 311 {
 312         size_t i;
 313
 314         for (i = 0; i < nr_attrs; i++)
 315                 event_attr_init(attrs + i);
 316
 317         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 318 }
 319
 320 struct perf_evsel *
 321 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 322 {
 323         struct perf_evsel *evsel;
 324
 325         evlist__for_each_entry(evlist, evsel) {
 326                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 327                     (int)evsel->attr.config == id)
 328                         return evsel;
 329         }
 330
 331         return NULL;
 332 }
 333
 334 struct perf_evsel *
 335 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 336                                      const char *name)
 337 {
 338         struct perf_evsel *evsel;
 339
 340         evlist__for_each_entry(evlist, evsel) {
 341                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 342                     (strcmp(evsel->name, name) == 0))
 343                         return evsel;
 344         }
 345
 346         return NULL;
 347 }
 348
 349 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 350                            const char *sys, const char *name, void *handler)
 351 {
 352         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 353
 354         if (IS_ERR(evsel))
 355                 return -1;
 356
 357         evsel->handler = handler;
 358         perf_evlist__add(evlist, evsel);
 359         return 0;
 360 }
 361
 362 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 363                                    struct perf_evsel *evsel)
 364 {
 365         if (evsel->system_wide)
 366                 return 1;
 367         else
 368                 return thread_map__nr(evlist->threads);
 369 }
 370
 371 void perf_evlist__disable(struct perf_evlist *evlist)
 372 {
 373         struct perf_evsel *pos;
 374
 375         evlist__for_each_entry(evlist, pos) {
 376                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 377                         continue;
 378                 perf_evsel__disable(pos);
 379         }
 380
 381         evlist->enabled = false;
 382 }
 383
 384 void perf_evlist__enable(struct perf_evlist *evlist)
 385 {
 386         struct perf_evsel *pos;
 387
 388         evlist__for_each_entry(evlist, pos) {
 389                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 390                         continue;
 391                 perf_evsel__enable(pos);
 392         }
 393
 394         evlist->enabled = true;
 395 }
 396
 397 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 398 {
 399         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 400 }
 401
 402 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 403                                          struct perf_evsel *evsel, int cpu)
 404 {
 405         int thread, err;
 406         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 407
 408         if (!evsel->fd)
 409                 return -EINVAL;
 410
 411         for (thread = 0; thread < nr_threads; thread++) {
 412                 err = ioctl(FD(evsel, cpu, thread),
 413                             PERF_EVENT_IOC_ENABLE, 0);
 414                 if (err)
 415                         return err;
 416         }
 417         return 0;
 418 }
 419
 420 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 421                                             struct perf_evsel *evsel,
 422                                             int thread)
 423 {
 424         int cpu, err;
 425         int nr_cpus = cpu_map__nr(evlist->cpus);
 426
 427         if (!evsel->fd)
 428                 return -EINVAL;
 429
 430         for (cpu = 0; cpu < nr_cpus; cpu++) {
 431                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 432                 if (err)
 433                         return err;
 434         }
 435         return 0;
 436 }
 437
 438 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 439                                   struct perf_evsel *evsel, int idx)
 440 {
 441         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 442
 443         if (per_cpu_mmaps)
 444                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 445         else
 446                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 447 }
 448
 449 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 450 {
 451         int nr_cpus = cpu_map__nr(evlist->cpus);
 452         int nr_threads = thread_map__nr(evlist->threads);
 453         int nfds = 0;
 454         struct perf_evsel *evsel;
 455
 456         evlist__for_each_entry(evlist, evsel) {
 457                 if (evsel->system_wide)
 458                         nfds += nr_cpus;
 459                 else
 460                         nfds += nr_cpus * nr_threads;
 461         }
 462
 463         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 464             fdarray__grow(&evlist->pollfd, nfds) < 0)
 465                 return -ENOMEM;
 466
 467         return 0;
 468 }
 469
 470 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 471                                      struct perf_mmap *map, short revent)
 472 {
 473         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 474         /*
 475          * Save the idx so that when we filter out fds POLLHUP'ed we can
 476          * close the associated evlist->mmap[] entry.
 477          */
 478         if (pos >= 0) {
 479                 evlist->pollfd.priv[pos].ptr = map;
 480
 481                 fcntl(fd, F_SETFL, O_NONBLOCK);
 482         }
 483
 484         return pos;
 485 }
 486
 487 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 488 {
 489         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 490 }
 491
 492 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 493                                          void *arg __maybe_unused)
 494 {
 495         struct perf_mmap *map = fda->priv[fd].ptr;
 496
 497         if (map)
 498                 perf_mmap__put(map);
 499 }
 500
 501 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 502 {
 503         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 504                                perf_evlist__munmap_filtered, NULL);
 505 }
 506
 507 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 508 {
 509         return fdarray__poll(&evlist->pollfd, timeout);
 510 }
 511
 512 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 513                                  struct perf_evsel *evsel,
 514                                  int cpu, int thread, u64 id)
 515 {
 516         int hash;
 517         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 518
 519         sid->id = id;
 520         sid->evsel = evsel;
 521         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 522         hlist_add_head(&sid->node, &evlist->heads[hash]);
 523 }
 524
 525 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 526                          int cpu, int thread, u64 id)
 527 {
 528         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 529         evsel->id[evsel->ids++] = id;
 530 }
 531
 532 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 533                            struct perf_evsel *evsel,
 534                            int cpu, int thread, int fd)
 535 {
 536         u64 read_data[4] = { 0, };
 537         int id_idx = 1; /* The first entry is the counter value */
 538         u64 id;
 539         int ret;
 540
 541         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 542         if (!ret)
 543                 goto add;
 544
 545         if (errno != ENOTTY)
 546                 return -1;
 547
 548         /* Legacy way to get event id.. All hail to old kernels! */
 549
 550         /*
 551          * This way does not work with group format read, so bail
 552          * out in that case.
 553          */
 554         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 555                 return -1;
 556
 557         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 558             read(fd, &read_data, sizeof(read_data)) == -1)
 559                 return -1;
 560
 561         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 562                 ++id_idx;
 563         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 564                 ++id_idx;
 565
 566         id = read_data[id_idx];
 567
 568  add:
 569         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 570         return 0;
 571 }
 572
 573 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 574                                      struct perf_evsel *evsel, int idx, int cpu,
 575                                      int thread)
 576 {
 577         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 578         sid->idx = idx;
 579         if (evlist->cpus && cpu >= 0)
 580                 sid->cpu = evlist->cpus->map[cpu];
 581         else
 582                 sid->cpu = -1;
 583         if (!evsel->system_wide && evlist->threads && thread >= 0)
 584                 sid->tid = thread_map__pid(evlist->threads, thread);
 585         else
 586                 sid->tid = -1;
 587 }
 588
 589 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 590 {
 591         struct hlist_head *head;
 592         struct perf_sample_id *sid;
 593         int hash;
 594
 595         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 596         head = &evlist->heads[hash];
 597
 598         hlist_for_each_entry(sid, head, node)
 599                 if (sid->id == id)
 600                         return sid;
 601
 602         return NULL;
 603 }
 604
 605 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 606 {
 607         struct perf_sample_id *sid;
 608
 609         if (evlist->nr_entries == 1 || !id)
 610                 return perf_evlist__first(evlist);
 611
 612         sid = perf_evlist__id2sid(evlist, id);
 613         if (sid)
 614                 return sid->evsel;
 615
 616         if (!perf_evlist__sample_id_all(evlist))
 617                 return perf_evlist__first(evlist);
 618
 619         return NULL;
 620 }
 621
 622 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 623                                                 u64 id)
 624 {
 625         struct perf_sample_id *sid;
 626
 627         if (!id)
 628                 return NULL;
 629
 630         sid = perf_evlist__id2sid(evlist, id);
 631         if (sid)
 632                 return sid->evsel;
 633
 634         return NULL;
 635 }
 636
 637 static int perf_evlist__event2id(struct perf_evlist *evlist,
 638                                  union perf_event *event, u64 *id)
 639 {
 640         const u64 *array = event->sample.array;
 641         ssize_t n;
 642
 643         n = (event->header.size - sizeof(event->header)) >> 3;
 644
 645         if (event->header.type == PERF_RECORD_SAMPLE) {
 646                 if (evlist->id_pos >= n)
 647                         return -1;
 648                 *id = array[evlist->id_pos];
 649         } else {
 650                 if (evlist->is_pos > n)
 651                         return -1;
 652                 n -= evlist->is_pos;
 653                 *id = array[n];
 654         }
 655         return 0;
 656 }
 657
 658 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 659                                             union perf_event *event)
 660 {
 661         struct perf_evsel *first = perf_evlist__first(evlist);
 662         struct hlist_head *head;
 663         struct perf_sample_id *sid;
 664         int hash;
 665         u64 id;
 666
 667         if (evlist->nr_entries == 1)
 668                 return first;
 669
 670         if (!first->attr.sample_id_all &&
 671             event->header.type != PERF_RECORD_SAMPLE)
 672                 return first;
 673
 674         if (perf_evlist__event2id(evlist, event, &id))
 675                 return NULL;
 676
 677         /* Synthesized events have an id of zero */
 678         if (!id)
 679                 return first;
 680
 681         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 682         head = &evlist->heads[hash];
 683
 684         hlist_for_each_entry(sid, head, node) {
 685                 if (sid->id == id)
 686                         return sid->evsel;
 687         }
 688         return NULL;
 689 }
 690
 691 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 692 {
 693         int i;
 694
 695         if (!evlist->backward_mmap)
 696                 return 0;
 697
 698         for (i = 0; i < evlist->nr_mmaps; i++) {
 699                 int fd = evlist->backward_mmap[i].fd;
 700                 int err;
 701
 702                 if (fd < 0)
 703                         continue;
 704                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 705                 if (err)
 706                         return err;
 707         }
 708         return 0;
 709 }
 710
 711 static int perf_evlist__pause(struct perf_evlist *evlist)
 712 {
 713         return perf_evlist__set_paused(evlist, true);
 714 }
 715
 716 static int perf_evlist__resume(struct perf_evlist *evlist)
 717 {
 718         return perf_evlist__set_paused(evlist, false);
 719 }
 720
 721 /* When check_messup is true, 'end' must points to a good entry */
 722 static union perf_event *
 723 perf_mmap__read(struct perf_mmap *md, bool check_messup, u64 start,
 724                 u64 end, u64 *prev)
 725 {
 726         unsigned char *data = md->base + page_size;
 727         union perf_event *event = NULL;
 728         int diff = end - start;
 729
 730         if (check_messup) {
 731                 /*
 732                  * If we're further behind than half the buffer, there's a chance
 733                  * the writer will bite our tail and mess up the samples under us.
 734                  *
 735                  * If we somehow ended up ahead of the 'end', we got messed up.
 736                  *
 737                  * In either case, truncate and restart at 'end'.
 738                  */
 739                 if (diff > md->mask / 2 || diff < 0) {
 740                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 741
 742                         /*
 743                          * 'end' points to a known good entry, start there.
 744                          */
 745                         start = end;
 746                         diff = 0;
 747                 }
 748         }
 749
 750         if (diff >= (int)sizeof(event->header)) {
 751                 size_t size;
 752
 753                 event = (union perf_event *)&data[start & md->mask];
 754                 size = event->header.size;
 755
 756                 if (size < sizeof(event->header) || diff < (int)size) {
 757                         event = NULL;
 758                         goto broken_event;
 759                 }
 760
 761                 /*
 762                  * Event straddles the mmap boundary -- header should always
 763                  * be inside due to u64 alignment of output.
 764                  */
 765                 if ((start & md->mask) + size != ((start + size) & md->mask)) {
 766                         unsigned int offset = start;
 767                         unsigned int len = min(sizeof(*event), size), cpy;
 768                         void *dst = md->event_copy;
 769
 770                         do {
 771                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 772                                 memcpy(dst, &data[offset & md->mask], cpy);
 773                                 offset += cpy;
 774                                 dst += cpy;
 775                                 len -= cpy;
 776                         } while (len);
 777
 778                         event = (union perf_event *) md->event_copy;
 779                 }
 780
 781                 start += size;
 782         }
 783
 784 broken_event:
 785         if (prev)
 786                 *prev = start;
 787
 788         return event;
 789 }
 790
 791 union perf_event *perf_mmap__read_forward(struct perf_mmap *md, bool check_messup)
 792 {
 793         u64 head;
 794         u64 old = md->prev;
 795
 796         /*
 797          * Check if event was unmapped due to a POLLHUP/POLLERR.
 798          */
 799         if (!atomic_read(&md->refcnt))
 800                 return NULL;
 801
 802         head = perf_mmap__read_head(md);
 803
 804         return perf_mmap__read(md, check_messup, old, head, &md->prev);
 805 }
 806
 807 union perf_event *
 808 perf_mmap__read_backward(struct perf_mmap *md)
 809 {
 810         u64 head, end;
 811         u64 start = md->prev;
 812
 813         /*
 814          * Check if event was unmapped due to a POLLHUP/POLLERR.
 815          */
 816         if (!atomic_read(&md->refcnt))
 817                 return NULL;
 818
 819         head = perf_mmap__read_head(md);
 820         if (!head)
 821                 return NULL;
 822
 823         /*
 824          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 825          * it each time when kernel writes to it, so in fact 'head' is
 826          * negative. 'end' pointer is made manually by adding the size of
 827          * the ring buffer to 'head' pointer, means the validate data can
 828          * read is the whole ring buffer. If 'end' is positive, the ring
 829          * buffer has not fully filled, so we must adjust 'end' to 0.
 830          *
 831          * However, since both 'head' and 'end' is unsigned, we can't
 832          * simply compare 'end' against 0. Here we compare '-head' and
 833          * the size of the ring buffer, where -head is the number of bytes
 834          * kernel write to the ring buffer.
 835          */
 836         if (-head < (u64)(md->mask + 1))
 837                 end = 0;
 838         else
 839                 end = head + md->mask + 1;
 840
 841         return perf_mmap__read(md, false, start, end, &md->prev);
 842 }
 843
 844 union perf_event *perf_evlist__mmap_read_forward(struct perf_evlist *evlist, int idx)
 845 {
 846         struct perf_mmap *md = &evlist->mmap[idx];
 847
 848         /*
 849          * Check messup is required for forward overwritable ring buffer:
 850          * memory pointed by md->prev can be overwritten in this case.
 851          * No need for read-write ring buffer: kernel stop outputting when
 852          * it hit md->prev (perf_mmap__consume()).
 853          */
 854         return perf_mmap__read_forward(md, evlist->overwrite);
 855 }
 856
 857 union perf_event *perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
 858 {
 859         struct perf_mmap *md = &evlist->mmap[idx];
 860
 861         /*
 862          * No need to check messup for backward ring buffer:
 863          * We can always read arbitrary long data from a backward
 864          * ring buffer unless we forget to pause it before reading.
 865          */
 866         return perf_mmap__read_backward(md);
 867 }
 868
 869 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 870 {
 871         return perf_evlist__mmap_read_forward(evlist, idx);
 872 }
 873
 874 void perf_mmap__read_catchup(struct perf_mmap *md)
 875 {
 876         u64 head;
 877
 878         if (!atomic_read(&md->refcnt))
 879                 return;
 880
 881         head = perf_mmap__read_head(md);
 882         md->prev = head;
 883 }
 884
 885 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 886 {
 887         perf_mmap__read_catchup(&evlist->mmap[idx]);
 888 }
 889
 890 static bool perf_mmap__empty(struct perf_mmap *md)
 891 {
 892         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 893 }
 894
 895 static void perf_mmap__get(struct perf_mmap *map)
 896 {
 897         atomic_inc(&map->refcnt);
 898 }
 899
 900 static void perf_mmap__put(struct perf_mmap *md)
 901 {
 902         BUG_ON(md->base && atomic_read(&md->refcnt) == 0);
 903
 904         if (atomic_dec_and_test(&md->refcnt))
 905                 perf_mmap__munmap(md);
 906 }
 907
 908 void perf_mmap__consume(struct perf_mmap *md, bool overwrite)
 909 {
 910         if (!overwrite) {
 911                 u64 old = md->prev;
 912
 913                 perf_mmap__write_tail(md, old);
 914         }
 915
 916         if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 917                 perf_mmap__put(md);
 918 }
 919
 920 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 921 {
 922         perf_mmap__consume(&evlist->mmap[idx], evlist->overwrite);
 923 }
 924
 925 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 926                                struct auxtrace_mmap_params *mp __maybe_unused,
 927                                void *userpg __maybe_unused,
 928                                int fd __maybe_unused)
 929 {
 930         return 0;
 931 }
 932
 933 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 934 {
 935 }
 936
 937 void __weak auxtrace_mmap_params__init(
 938                         struct auxtrace_mmap_params *mp __maybe_unused,
 939                         off_t auxtrace_offset __maybe_unused,
 940                         unsigned int auxtrace_pages __maybe_unused,
 941                         bool auxtrace_overwrite __maybe_unused)
 942 {
 943 }
 944
 945 void __weak auxtrace_mmap_params__set_idx(
 946                         struct auxtrace_mmap_params *mp __maybe_unused,
 947                         struct perf_evlist *evlist __maybe_unused,
 948                         int idx __maybe_unused,
 949                         bool per_cpu __maybe_unused)
 950 {
 951 }
 952
 953 static void perf_mmap__munmap(struct perf_mmap *map)
 954 {
 955         if (map->base != NULL) {
 956                 munmap(map->base, perf_mmap__mmap_len(map));
 957                 map->base = NULL;
 958                 map->fd = -1;
 959                 atomic_set(&map->refcnt, 0);
 960         }
 961         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 962 }
 963
 964 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 965 {
 966         int i;
 967
 968         if (evlist->mmap)
 969                 for (i = 0; i < evlist->nr_mmaps; i++)
 970                         perf_mmap__munmap(&evlist->mmap[i]);
 971
 972         if (evlist->backward_mmap)
 973                 for (i = 0; i < evlist->nr_mmaps; i++)
 974                         perf_mmap__munmap(&evlist->backward_mmap[i]);
 975 }
 976
 977 void perf_evlist__munmap(struct perf_evlist *evlist)
 978 {
 979         perf_evlist__munmap_nofree(evlist);
 980         zfree(&evlist->mmap);
 981         zfree(&evlist->backward_mmap);
 982 }
 983
 984 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 985 {
 986         int i;
 987         struct perf_mmap *map;
 988
 989         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 990         if (cpu_map__empty(evlist->cpus))
 991                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 992         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 993         if (!map)
 994                 return NULL;
 995
 996         for (i = 0; i < evlist->nr_mmaps; i++)
 997                 map[i].fd = -1;
 998         return map;
 999 }
1000
1001 struct mmap_params {
1002         int prot;
1003         int mask;
1004         struct auxtrace_mmap_params auxtrace_mp;
1005 };
1006
1007 static int perf_mmap__mmap(struct perf_mmap *map,
1008                            struct mmap_params *mp, int fd)
1009 {
1010         /*
1011          * The last one will be done at perf_evlist__mmap_consume(), so that we
1012          * make sure we don't prevent tools from consuming every last event in
1013          * the ring buffer.
1014          *
1015          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
1016          * anymore, but the last events for it are still in the ring buffer,
1017          * waiting to be consumed.
1018          *
1019          * Tools can chose to ignore this at their own discretion, but the
1020          * evlist layer can't just drop it when filtering events in
1021          * perf_evlist__filter_pollfd().
1022          */
1023         atomic_set(&map->refcnt, 2);
1024         map->prev = 0;
1025         map->mask = mp->mask;
1026         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
1027                          MAP_SHARED, fd, 0);
1028         if (map->base == MAP_FAILED) {
1029                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1030                           errno);
1031                 map->base = NULL;
1032                 return -1;
1033         }
1034         map->fd = fd;
1035
1036         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
1037                                 &mp->auxtrace_mp, map->base, fd))
1038                 return -1;
1039
1040         return 0;
1041 }
1042
1043 static bool
1044 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
1045                          struct perf_evsel *evsel)
1046 {
1047         if (evsel->attr.write_backward)
1048                 return false;
1049         return true;
1050 }
1051
1052 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
1053                                        struct mmap_params *mp, int cpu,
1054                                        int thread, int *_output, int *_output_backward)
1055 {
1056         struct perf_evsel *evsel;
1057         int revent;
1058
1059         evlist__for_each_entry(evlist, evsel) {
1060                 struct perf_mmap *maps = evlist->mmap;
1061                 int *output = _output;
1062                 int fd;
1063
1064                 if (evsel->attr.write_backward) {
1065                         output = _output_backward;
1066                         maps = evlist->backward_mmap;
1067
1068                         if (!maps) {
1069                                 maps = perf_evlist__alloc_mmap(evlist);
1070                                 if (!maps)
1071                                         return -1;
1072                                 evlist->backward_mmap = maps;
1073                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1074                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1075                         }
1076                 }
1077
1078                 if (evsel->system_wide && thread)
1079                         continue;
1080
1081                 fd = FD(evsel, cpu, thread);
1082
1083                 if (*output == -1) {
1084                         *output = fd;
1085
1086                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1087                                 return -1;
1088                 } else {
1089                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1090                                 return -1;
1091
1092                         perf_mmap__get(&maps[idx]);
1093                 }
1094
1095                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1096
1097                 /*
1098                  * The system_wide flag causes a selected event to be opened
1099                  * always without a pid.  Consequently it will never get a
1100                  * POLLHUP, but it is used for tracking in combination with
1101                  * other events, so it should not need to be polled anyway.
1102                  * Therefore don't add it for polling.
1103                  */
1104                 if (!evsel->system_wide &&
1105                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1106                         perf_mmap__put(&maps[idx]);
1107                         return -1;
1108                 }
1109
1110                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1111                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1112                                                    fd) < 0)
1113                                 return -1;
1114                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1115                                                  thread);
1116                 }
1117         }
1118
1119         return 0;
1120 }
1121
1122 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1123                                      struct mmap_params *mp)
1124 {
1125         int cpu, thread;
1126         int nr_cpus = cpu_map__nr(evlist->cpus);
1127         int nr_threads = thread_map__nr(evlist->threads);
1128
1129         pr_debug2("perf event ring buffer mmapped per cpu\n");
1130         for (cpu = 0; cpu < nr_cpus; cpu++) {
1131                 int output = -1;
1132                 int output_backward = -1;
1133
1134                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1135                                               true);
1136
1137                 for (thread = 0; thread < nr_threads; thread++) {
1138                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1139                                                         thread, &output, &output_backward))
1140                                 goto out_unmap;
1141                 }
1142         }
1143
1144         return 0;
1145
1146 out_unmap:
1147         perf_evlist__munmap_nofree(evlist);
1148         return -1;
1149 }
1150
1151 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1152                                         struct mmap_params *mp)
1153 {
1154         int thread;
1155         int nr_threads = thread_map__nr(evlist->threads);
1156
1157         pr_debug2("perf event ring buffer mmapped per thread\n");
1158         for (thread = 0; thread < nr_threads; thread++) {
1159                 int output = -1;
1160                 int output_backward = -1;
1161
1162                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1163                                               false);
1164
1165                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1166                                                 &output, &output_backward))
1167                         goto out_unmap;
1168         }
1169
1170         return 0;
1171
1172 out_unmap:
1173         perf_evlist__munmap_nofree(evlist);
1174         return -1;
1175 }
1176
1177 unsigned long perf_event_mlock_kb_in_pages(void)
1178 {
1179         unsigned long pages;
1180         int max;
1181
1182         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1183                 /*
1184                  * Pick a once upon a time good value, i.e. things look
1185                  * strange since we can't read a sysctl value, but lets not
1186                  * die yet...
1187                  */
1188                 max = 512;
1189         } else {
1190                 max -= (page_size / 1024);
1191         }
1192
1193         pages = (max * 1024) / page_size;
1194         if (!is_power_of_2(pages))
1195                 pages = rounddown_pow_of_two(pages);
1196
1197         return pages;
1198 }
1199
1200 static size_t perf_evlist__mmap_size(unsigned long pages)
1201 {
1202         if (pages == UINT_MAX)
1203                 pages = perf_event_mlock_kb_in_pages();
1204         else if (!is_power_of_2(pages))
1205                 return 0;
1206
1207         return (pages + 1) * page_size;
1208 }
1209
1210 static long parse_pages_arg(const char *str, unsigned long min,
1211                             unsigned long max)
1212 {
1213         unsigned long pages, val;
1214         static struct parse_tag tags[] = {
1215                 { .tag  = 'B', .mult = 1       },
1216                 { .tag  = 'K', .mult = 1 << 10 },
1217                 { .tag  = 'M', .mult = 1 << 20 },
1218                 { .tag  = 'G', .mult = 1 << 30 },
1219                 { .tag  = 0 },
1220         };
1221
1222         if (str == NULL)
1223                 return -EINVAL;
1224
1225         val = parse_tag_value(str, tags);
1226         if (val != (unsigned long) -1) {
1227                 /* we got file size value */
1228                 pages = PERF_ALIGN(val, page_size) / page_size;
1229         } else {
1230                 /* we got pages count value */
1231                 char *eptr;
1232                 pages = strtoul(str, &eptr, 10);
1233                 if (*eptr != '\0')
1234                         return -EINVAL;
1235         }
1236
1237         if (pages == 0 && min == 0) {
1238                 /* leave number of pages at 0 */
1239         } else if (!is_power_of_2(pages)) {
1240                 /* round pages up to next power of 2 */
1241                 pages = roundup_pow_of_two(pages);
1242                 if (!pages)
1243                         return -EINVAL;
1244                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1245                         pages * page_size, pages);
1246         }
1247
1248         if (pages > max)
1249                 return -EINVAL;
1250
1251         return pages;
1252 }
1253
1254 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1255 {
1256         unsigned long max = UINT_MAX;
1257         long pages;
1258
1259         if (max > SIZE_MAX / page_size)
1260                 max = SIZE_MAX / page_size;
1261
1262         pages = parse_pages_arg(str, 1, max);
1263         if (pages < 0) {
1264                 pr_err("Invalid argument for --mmap_pages/-m\n");
1265                 return -1;
1266         }
1267
1268         *mmap_pages = pages;
1269         return 0;
1270 }
1271
1272 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1273                                   int unset __maybe_unused)
1274 {
1275         return __perf_evlist__parse_mmap_pages(opt->value, str);
1276 }
1277
1278 /**
1279  * perf_evlist__mmap_ex - Create mmaps to receive events.
1280  * @evlist: list of events
1281  * @pages: map length in pages
1282  * @overwrite: overwrite older events?
1283  * @auxtrace_pages - auxtrace map length in pages
1284  * @auxtrace_overwrite - overwrite older auxtrace data?
1285  *
1286  * If @overwrite is %false the user needs to signal event consumption using
1287  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1288  * automatically.
1289  *
1290  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1291  * consumption using auxtrace_mmap__write_tail().
1292  *
1293  * Return: %0 on success, negative error code otherwise.
1294  */
1295 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1296                          bool overwrite, unsigned int auxtrace_pages,
1297                          bool auxtrace_overwrite)
1298 {
1299         struct perf_evsel *evsel;
1300         const struct cpu_map *cpus = evlist->cpus;
1301         const struct thread_map *threads = evlist->threads;
1302         struct mmap_params mp = {
1303                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1304         };
1305
1306         if (!evlist->mmap)
1307                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1308         if (!evlist->mmap)
1309                 return -ENOMEM;
1310
1311         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1312                 return -ENOMEM;
1313
1314         evlist->overwrite = overwrite;
1315         evlist->mmap_len = perf_evlist__mmap_size(pages);
1316         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1317         mp.mask = evlist->mmap_len - page_size - 1;
1318
1319         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1320                                    auxtrace_pages, auxtrace_overwrite);
1321
1322         evlist__for_each_entry(evlist, evsel) {
1323                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1324                     evsel->sample_id == NULL &&
1325                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1326                         return -ENOMEM;
1327         }
1328
1329         if (cpu_map__empty(cpus))
1330                 return perf_evlist__mmap_per_thread(evlist, &mp);
1331
1332         return perf_evlist__mmap_per_cpu(evlist, &mp);
1333 }
1334
1335 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1336                       bool overwrite)
1337 {
1338         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1339 }
1340
1341 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1342 {
1343         struct cpu_map *cpus;
1344         struct thread_map *threads;
1345
1346         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1347
1348         if (!threads)
1349                 return -1;
1350
1351         if (target__uses_dummy_map(target))
1352                 cpus = cpu_map__dummy_new();
1353         else
1354                 cpus = cpu_map__new(target->cpu_list);
1355
1356         if (!cpus)
1357                 goto out_delete_threads;
1358
1359         evlist->has_user_cpus = !!target->cpu_list;
1360
1361         perf_evlist__set_maps(evlist, cpus, threads);
1362
1363         return 0;
1364
1365 out_delete_threads:
1366         thread_map__put(threads);
1367         return -1;
1368 }
1369
1370 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1371                            struct thread_map *threads)
1372 {
1373         /*
1374          * Allow for the possibility that one or another of the maps isn't being
1375          * changed i.e. don't put it.  Note we are assuming the maps that are
1376          * being applied are brand new and evlist is taking ownership of the
1377          * original reference count of 1.  If that is not the case it is up to
1378          * the caller to increase the reference count.
1379          */
1380         if (cpus != evlist->cpus) {
1381                 cpu_map__put(evlist->cpus);
1382                 evlist->cpus = cpu_map__get(cpus);
1383         }
1384
1385         if (threads != evlist->threads) {
1386                 thread_map__put(evlist->threads);
1387                 evlist->threads = thread_map__get(threads);
1388         }
1389
1390         perf_evlist__propagate_maps(evlist);
1391 }
1392
1393 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1394                                    enum perf_event_sample_format bit)
1395 {
1396         struct perf_evsel *evsel;
1397
1398         evlist__for_each_entry(evlist, evsel)
1399                 __perf_evsel__set_sample_bit(evsel, bit);
1400 }
1401
1402 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1403                                      enum perf_event_sample_format bit)
1404 {
1405         struct perf_evsel *evsel;
1406
1407         evlist__for_each_entry(evlist, evsel)
1408                 __perf_evsel__reset_sample_bit(evsel, bit);
1409 }
1410
1411 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1412 {
1413         struct perf_evsel *evsel;
1414         int err = 0;
1415         const int ncpus = cpu_map__nr(evlist->cpus),
1416                   nthreads = thread_map__nr(evlist->threads);
1417
1418         evlist__for_each_entry(evlist, evsel) {
1419                 if (evsel->filter == NULL)
1420                         continue;
1421
1422                 /*
1423                  * filters only work for tracepoint event, which doesn't have cpu limit.
1424                  * So evlist and evsel should always be same.
1425                  */
1426                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1427                 if (err) {
1428                         *err_evsel = evsel;
1429                         break;
1430                 }
1431         }
1432
1433         return err;
1434 }
1435
1436 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1437 {
1438         struct perf_evsel *evsel;
1439         int err = 0;
1440
1441         evlist__for_each_entry(evlist, evsel) {
1442                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1443                         continue;
1444
1445                 err = perf_evsel__set_filter(evsel, filter);
1446                 if (err)
1447                         break;
1448         }
1449
1450         return err;
1451 }
1452
1453 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1454 {
1455         char *filter;
1456         int ret = -1;
1457         size_t i;
1458
1459         for (i = 0; i < npids; ++i) {
1460                 if (i == 0) {
1461                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1462                                 return -1;
1463                 } else {
1464                         char *tmp;
1465
1466                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1467                                 goto out_free;
1468
1469                         free(filter);
1470                         filter = tmp;
1471                 }
1472         }
1473
1474         ret = perf_evlist__set_filter(evlist, filter);
1475 out_free:
1476         free(filter);
1477         return ret;
1478 }
1479
1480 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1481 {
1482         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1483 }
1484
1485 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1486 {
1487         struct perf_evsel *pos;
1488
1489         if (evlist->nr_entries == 1)
1490                 return true;
1491
1492         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1493                 return false;
1494
1495         evlist__for_each_entry(evlist, pos) {
1496                 if (pos->id_pos != evlist->id_pos ||
1497                     pos->is_pos != evlist->is_pos)
1498                         return false;
1499         }
1500
1501         return true;
1502 }
1503
1504 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1505 {
1506         struct perf_evsel *evsel;
1507
1508         if (evlist->combined_sample_type)
1509                 return evlist->combined_sample_type;
1510
1511         evlist__for_each_entry(evlist, evsel)
1512                 evlist->combined_sample_type |= evsel->attr.sample_type;
1513
1514         return evlist->combined_sample_type;
1515 }
1516
1517 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1518 {
1519         evlist->combined_sample_type = 0;
1520         return __perf_evlist__combined_sample_type(evlist);
1521 }
1522
1523 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1524 {
1525         struct perf_evsel *evsel;
1526         u64 branch_type = 0;
1527
1528         evlist__for_each_entry(evlist, evsel)
1529                 branch_type |= evsel->attr.branch_sample_type;
1530         return branch_type;
1531 }
1532
1533 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1534 {
1535         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1536         u64 read_format = first->attr.read_format;
1537         u64 sample_type = first->attr.sample_type;
1538
1539         evlist__for_each_entry(evlist, pos) {
1540                 if (read_format != pos->attr.read_format)
1541                         return false;
1542         }
1543
1544         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1545         if ((sample_type & PERF_SAMPLE_READ) &&
1546             !(read_format & PERF_FORMAT_ID)) {
1547                 return false;
1548         }
1549
1550         return true;
1551 }
1552
1553 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1554 {
1555         struct perf_evsel *first = perf_evlist__first(evlist);
1556         return first->attr.read_format;
1557 }
1558
1559 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1560 {
1561         struct perf_evsel *first = perf_evlist__first(evlist);
1562         struct perf_sample *data;
1563         u64 sample_type;
1564         u16 size = 0;
1565
1566         if (!first->attr.sample_id_all)
1567                 goto out;
1568
1569         sample_type = first->attr.sample_type;
1570
1571         if (sample_type & PERF_SAMPLE_TID)
1572                 size += sizeof(data->tid) * 2;
1573
1574        if (sample_type & PERF_SAMPLE_TIME)
1575                 size += sizeof(data->time);
1576
1577         if (sample_type & PERF_SAMPLE_ID)
1578                 size += sizeof(data->id);
1579
1580         if (sample_type & PERF_SAMPLE_STREAM_ID)
1581                 size += sizeof(data->stream_id);
1582
1583         if (sample_type & PERF_SAMPLE_CPU)
1584                 size += sizeof(data->cpu) * 2;
1585
1586         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1587                 size += sizeof(data->id);
1588 out:
1589         return size;
1590 }
1591
1592 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1593 {
1594         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1595
1596         evlist__for_each_entry_continue(evlist, pos) {
1597                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1598                         return false;
1599         }
1600
1601         return true;
1602 }
1603
1604 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1605 {
1606         struct perf_evsel *first = perf_evlist__first(evlist);
1607         return first->attr.sample_id_all;
1608 }
1609
1610 void perf_evlist__set_selected(struct perf_evlist *evlist,
1611                                struct perf_evsel *evsel)
1612 {
1613         evlist->selected = evsel;
1614 }
1615
1616 void perf_evlist__close(struct perf_evlist *evlist)
1617 {
1618         struct perf_evsel *evsel;
1619         int ncpus = cpu_map__nr(evlist->cpus);
1620         int nthreads = thread_map__nr(evlist->threads);
1621         int n;
1622
1623         evlist__for_each_entry_reverse(evlist, evsel) {
1624                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1625                 perf_evsel__close(evsel, n, nthreads);
1626         }
1627 }
1628
1629 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1630 {
1631         struct cpu_map    *cpus;
1632         struct thread_map *threads;
1633         int err = -ENOMEM;
1634
1635         /*
1636          * Try reading /sys/devices/system/cpu/online to get
1637          * an all cpus map.
1638          *
1639          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1640          * code needs an overhaul to properly forward the
1641          * error, and we may not want to do that fallback to a
1642          * default cpu identity map :-\
1643          */
1644         cpus = cpu_map__new(NULL);
1645         if (!cpus)
1646                 goto out;
1647
1648         threads = thread_map__new_dummy();
1649         if (!threads)
1650                 goto out_put;
1651
1652         perf_evlist__set_maps(evlist, cpus, threads);
1653 out:
1654         return err;
1655 out_put:
1656         cpu_map__put(cpus);
1657         goto out;
1658 }
1659
1660 int perf_evlist__open(struct perf_evlist *evlist)
1661 {
1662         struct perf_evsel *evsel;
1663         int err;
1664
1665         /*
1666          * Default: one fd per CPU, all threads, aka systemwide
1667          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1668          */
1669         if (evlist->threads == NULL && evlist->cpus == NULL) {
1670                 err = perf_evlist__create_syswide_maps(evlist);
1671                 if (err < 0)
1672                         goto out_err;
1673         }
1674
1675         perf_evlist__update_id_pos(evlist);
1676
1677         evlist__for_each_entry(evlist, evsel) {
1678                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1679                 if (err < 0)
1680                         goto out_err;
1681         }
1682
1683         return 0;
1684 out_err:
1685         perf_evlist__close(evlist);
1686         errno = -err;
1687         return err;
1688 }
1689
1690 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1691                                   const char *argv[], bool pipe_output,
1692                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1693 {
1694         int child_ready_pipe[2], go_pipe[2];
1695         char bf;
1696
1697         if (pipe(child_ready_pipe) < 0) {
1698                 perror("failed to create 'ready' pipe");
1699                 return -1;
1700         }
1701
1702         if (pipe(go_pipe) < 0) {
1703                 perror("failed to create 'go' pipe");
1704                 goto out_close_ready_pipe;
1705         }
1706
1707         evlist->workload.pid = fork();
1708         if (evlist->workload.pid < 0) {
1709                 perror("failed to fork");
1710                 goto out_close_pipes;
1711         }
1712
1713         if (!evlist->workload.pid) {
1714                 int ret;
1715
1716                 if (pipe_output)
1717                         dup2(2, 1);
1718
1719                 signal(SIGTERM, SIG_DFL);
1720
1721                 close(child_ready_pipe[0]);
1722                 close(go_pipe[1]);
1723                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1724
1725                 /*
1726                  * Tell the parent we're ready to go
1727                  */
1728                 close(child_ready_pipe[1]);
1729
1730                 /*
1731                  * Wait until the parent tells us to go.
1732                  */
1733                 ret = read(go_pipe[0], &bf, 1);
1734                 /*
1735                  * The parent will ask for the execvp() to be performed by
1736                  * writing exactly one byte, in workload.cork_fd, usually via
1737                  * perf_evlist__start_workload().
1738                  *
1739                  * For cancelling the workload without actually running it,
1740                  * the parent will just close workload.cork_fd, without writing
1741                  * anything, i.e. read will return zero and we just exit()
1742                  * here.
1743                  */
1744                 if (ret != 1) {
1745                         if (ret == -1)
1746                                 perror("unable to read pipe");
1747                         exit(ret);
1748                 }
1749
1750                 execvp(argv[0], (char **)argv);
1751
1752                 if (exec_error) {
1753                         union sigval val;
1754
1755                         val.sival_int = errno;
1756                         if (sigqueue(getppid(), SIGUSR1, val))
1757                                 perror(argv[0]);
1758                 } else
1759                         perror(argv[0]);
1760                 exit(-1);
1761         }
1762
1763         if (exec_error) {
1764                 struct sigaction act = {
1765                         .sa_flags     = SA_SIGINFO,
1766                         .sa_sigaction = exec_error,
1767                 };
1768                 sigaction(SIGUSR1, &act, NULL);
1769         }
1770
1771         if (target__none(target)) {
1772                 if (evlist->threads == NULL) {
1773                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1774                                 __func__, __LINE__);
1775                         goto out_close_pipes;
1776                 }
1777                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1778         }
1779
1780         close(child_ready_pipe[1]);
1781         close(go_pipe[0]);
1782         /*
1783          * wait for child to settle
1784          */
1785         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1786                 perror("unable to read pipe");
1787                 goto out_close_pipes;
1788         }
1789
1790         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1791         evlist->workload.cork_fd = go_pipe[1];
1792         close(child_ready_pipe[0]);
1793         return 0;
1794
1795 out_close_pipes:
1796         close(go_pipe[0]);
1797         close(go_pipe[1]);
1798 out_close_ready_pipe:
1799         close(child_ready_pipe[0]);
1800         close(child_ready_pipe[1]);
1801         return -1;
1802 }
1803
1804 int perf_evlist__start_workload(struct perf_evlist *evlist)
1805 {
1806         if (evlist->workload.cork_fd > 0) {
1807                 char bf = 0;
1808                 int ret;
1809                 /*
1810                  * Remove the cork, let it rip!
1811                  */
1812                 ret = write(evlist->workload.cork_fd, &bf, 1);
1813                 if (ret < 0)
1814                         perror("enable to write to pipe");
1815
1816                 close(evlist->workload.cork_fd);
1817                 return ret;
1818         }
1819
1820         return 0;
1821 }
1822
1823 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1824                               struct perf_sample *sample)
1825 {
1826         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1827
1828         if (!evsel)
1829                 return -EFAULT;
1830         return perf_evsel__parse_sample(evsel, event, sample);
1831 }
1832
1833 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1834 {
1835         struct perf_evsel *evsel;
1836         size_t printed = 0;
1837
1838         evlist__for_each_entry(evlist, evsel) {
1839                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1840                                    perf_evsel__name(evsel));
1841         }
1842
1843         return printed + fprintf(fp, "\n");
1844 }
1845
1846 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1847                                int err, char *buf, size_t size)
1848 {
1849         int printed, value;
1850         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1851
1852         switch (err) {
1853         case EACCES:
1854         case EPERM:
1855                 printed = scnprintf(buf, size,
1856                                     "Error:\t%s.\n"
1857                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1858
1859                 value = perf_event_paranoid();
1860
1861                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1862
1863                 if (value >= 2) {
1864                         printed += scnprintf(buf + printed, size - printed,
1865                                              "For your workloads it needs to be <= 1\nHint:\t");
1866                 }
1867                 printed += scnprintf(buf + printed, size - printed,
1868                                      "For system wide tracing it needs to be set to -1.\n");
1869
1870                 printed += scnprintf(buf + printed, size - printed,
1871                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1872                                     "Hint:\tThe current value is %d.", value);
1873                 break;
1874         case EINVAL: {
1875                 struct perf_evsel *first = perf_evlist__first(evlist);
1876                 int max_freq;
1877
1878                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1879                         goto out_default;
1880
1881                 if (first->attr.sample_freq < (u64)max_freq)
1882                         goto out_default;
1883
1884                 printed = scnprintf(buf, size,
1885                                     "Error:\t%s.\n"
1886                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1887                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1888                                     emsg, max_freq, first->attr.sample_freq);
1889                 break;
1890         }
1891         default:
1892 out_default:
1893                 scnprintf(buf, size, "%s", emsg);
1894                 break;
1895         }
1896
1897         return 0;
1898 }
1899
1900 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1901 {
1902         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1903         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1904
1905         switch (err) {
1906         case EPERM:
1907                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1908                 printed += scnprintf(buf + printed, size - printed,
1909                                      "Error:\t%s.\n"
1910                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1911                                      "Hint:\tTried using %zd kB.\n",
1912                                      emsg, pages_max_per_user, pages_attempted);
1913
1914                 if (pages_attempted >= pages_max_per_user) {
1915                         printed += scnprintf(buf + printed, size - printed,
1916                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1917                                              pages_max_per_user + pages_attempted);
1918                 }
1919
1920                 printed += scnprintf(buf + printed, size - printed,
1921                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1922                 break;
1923         default:
1924                 scnprintf(buf, size, "%s", emsg);
1925                 break;
1926         }
1927
1928         return 0;
1929 }
1930
1931 void perf_evlist__to_front(struct perf_evlist *evlist,
1932                            struct perf_evsel *move_evsel)
1933 {
1934         struct perf_evsel *evsel, *n;
1935         LIST_HEAD(move);
1936
1937         if (move_evsel == perf_evlist__first(evlist))
1938                 return;
1939
1940         evlist__for_each_entry_safe(evlist, n, evsel) {
1941                 if (evsel->leader == move_evsel->leader)
1942                         list_move_tail(&evsel->node, &move);
1943         }
1944
1945         list_splice(&move, &evlist->entries);
1946 }
1947
1948 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1949                                      struct perf_evsel *tracking_evsel)
1950 {
1951         struct perf_evsel *evsel;
1952
1953         if (tracking_evsel->tracking)
1954                 return;
1955
1956         evlist__for_each_entry(evlist, evsel) {
1957                 if (evsel != tracking_evsel)
1958                         evsel->tracking = false;
1959         }
1960
1961         tracking_evsel->tracking = true;
1962 }
1963
1964 struct perf_evsel *
1965 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1966                                const char *str)
1967 {
1968         struct perf_evsel *evsel;
1969
1970         evlist__for_each_entry(evlist, evsel) {
1971                 if (!evsel->name)
1972                         continue;
1973                 if (strcmp(str, evsel->name) == 0)
1974                         return evsel;
1975         }
1976
1977         return NULL;
1978 }
1979
1980 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1981                                   enum bkw_mmap_state state)
1982 {
1983         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1984         enum action {
1985                 NONE,
1986                 PAUSE,
1987                 RESUME,
1988         } action = NONE;
1989
1990         if (!evlist->backward_mmap)
1991                 return;
1992
1993         switch (old_state) {
1994         case BKW_MMAP_NOTREADY: {
1995                 if (state != BKW_MMAP_RUNNING)
1996                         goto state_err;;
1997                 break;
1998         }
1999         case BKW_MMAP_RUNNING: {
2000                 if (state != BKW_MMAP_DATA_PENDING)
2001                         goto state_err;
2002                 action = PAUSE;
2003                 break;
2004         }
2005         case BKW_MMAP_DATA_PENDING: {
2006                 if (state != BKW_MMAP_EMPTY)
2007                         goto state_err;
2008                 break;
2009         }
2010         case BKW_MMAP_EMPTY: {
2011                 if (state != BKW_MMAP_RUNNING)
2012                         goto state_err;
2013                 action = RESUME;
2014                 break;
2015         }
2016         default:
2017                 WARN_ONCE(1, "Shouldn't get there\n");
2018         }
2019
2020         evlist->bkw_mmap_state = state;
2021
2022         switch (action) {
2023         case PAUSE:
2024                 perf_evlist__pause(evlist);
2025                 break;
2026         case RESUME:
2027                 perf_evlist__resume(evlist);
2028                 break;
2029         case NONE:
2030         default:
2031                 break;
2032         }
2033
2034 state_err:
2035         return;
2036 }