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