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,
1036                                        int thread, int *_output, int *_output_backward)
1037 {
1038         struct perf_evsel *evsel;
1039         int revent;
1040
1041         evlist__for_each_entry(evlist, evsel) {
1042                 struct perf_mmap *maps = evlist->mmap;
1043                 int *output = _output;
1044                 int fd;
1045
1046                 if (evsel->attr.write_backward) {
1047                         output = _output_backward;
1048                         maps = evlist->backward_mmap;
1049
1050                         if (!maps) {
1051                                 maps = perf_evlist__alloc_mmap(evlist);
1052                                 if (!maps)
1053                                         return -1;
1054                                 evlist->backward_mmap = maps;
1055                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1056                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1057                         }
1058                 }
1059
1060                 if (evsel->system_wide && thread)
1061                         continue;
1062
1063                 fd = FD(evsel, cpu, thread);
1064
1065                 if (*output == -1) {
1066                         *output = fd;
1067
1068                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1069                                 return -1;
1070                 } else {
1071                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1072                                 return -1;
1073
1074                         perf_mmap__get(&maps[idx]);
1075                 }
1076
1077                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1078
1079                 /*
1080                  * The system_wide flag causes a selected event to be opened
1081                  * always without a pid.  Consequently it will never get a
1082                  * POLLHUP, but it is used for tracking in combination with
1083                  * other events, so it should not need to be polled anyway.
1084                  * Therefore don't add it for polling.
1085                  */
1086                 if (!evsel->system_wide &&
1087                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1088                         perf_mmap__put(&maps[idx]);
1089                         return -1;
1090                 }
1091
1092                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1093                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1094                                                    fd) < 0)
1095                                 return -1;
1096                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1097                                                  thread);
1098                 }
1099         }
1100
1101         return 0;
1102 }
1103
1104 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1105                                      struct mmap_params *mp)
1106 {
1107         int cpu, thread;
1108         int nr_cpus = cpu_map__nr(evlist->cpus);
1109         int nr_threads = thread_map__nr(evlist->threads);
1110
1111         pr_debug2("perf event ring buffer mmapped per cpu\n");
1112         for (cpu = 0; cpu < nr_cpus; cpu++) {
1113                 int output = -1;
1114                 int output_backward = -1;
1115
1116                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1117                                               true);
1118
1119                 for (thread = 0; thread < nr_threads; thread++) {
1120                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1121                                                         thread, &output, &output_backward))
1122                                 goto out_unmap;
1123                 }
1124         }
1125
1126         return 0;
1127
1128 out_unmap:
1129         perf_evlist__munmap_nofree(evlist);
1130         return -1;
1131 }
1132
1133 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1134                                         struct mmap_params *mp)
1135 {
1136         int thread;
1137         int nr_threads = thread_map__nr(evlist->threads);
1138
1139         pr_debug2("perf event ring buffer mmapped per thread\n");
1140         for (thread = 0; thread < nr_threads; thread++) {
1141                 int output = -1;
1142                 int output_backward = -1;
1143
1144                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1145                                               false);
1146
1147                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1148                                                 &output, &output_backward))
1149                         goto out_unmap;
1150         }
1151
1152         return 0;
1153
1154 out_unmap:
1155         perf_evlist__munmap_nofree(evlist);
1156         return -1;
1157 }
1158
1159 unsigned long perf_event_mlock_kb_in_pages(void)
1160 {
1161         unsigned long pages;
1162         int max;
1163
1164         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1165                 /*
1166                  * Pick a once upon a time good value, i.e. things look
1167                  * strange since we can't read a sysctl value, but lets not
1168                  * die yet...
1169                  */
1170                 max = 512;
1171         } else {
1172                 max -= (page_size / 1024);
1173         }
1174
1175         pages = (max * 1024) / page_size;
1176         if (!is_power_of_2(pages))
1177                 pages = rounddown_pow_of_two(pages);
1178
1179         return pages;
1180 }
1181
1182 static size_t perf_evlist__mmap_size(unsigned long pages)
1183 {
1184         if (pages == UINT_MAX)
1185                 pages = perf_event_mlock_kb_in_pages();
1186         else if (!is_power_of_2(pages))
1187                 return 0;
1188
1189         return (pages + 1) * page_size;
1190 }
1191
1192 static long parse_pages_arg(const char *str, unsigned long min,
1193                             unsigned long max)
1194 {
1195         unsigned long pages, val;
1196         static struct parse_tag tags[] = {
1197                 { .tag  = 'B', .mult = 1       },
1198                 { .tag  = 'K', .mult = 1 << 10 },
1199                 { .tag  = 'M', .mult = 1 << 20 },
1200                 { .tag  = 'G', .mult = 1 << 30 },
1201                 { .tag  = 0 },
1202         };
1203
1204         if (str == NULL)
1205                 return -EINVAL;
1206
1207         val = parse_tag_value(str, tags);
1208         if (val != (unsigned long) -1) {
1209                 /* we got file size value */
1210                 pages = PERF_ALIGN(val, page_size) / page_size;
1211         } else {
1212                 /* we got pages count value */
1213                 char *eptr;
1214                 pages = strtoul(str, &eptr, 10);
1215                 if (*eptr != '\0')
1216                         return -EINVAL;
1217         }
1218
1219         if (pages == 0 && min == 0) {
1220                 /* leave number of pages at 0 */
1221         } else if (!is_power_of_2(pages)) {
1222                 /* round pages up to next power of 2 */
1223                 pages = roundup_pow_of_two(pages);
1224                 if (!pages)
1225                         return -EINVAL;
1226                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1227                         pages * page_size, pages);
1228         }
1229
1230         if (pages > max)
1231                 return -EINVAL;
1232
1233         return pages;
1234 }
1235
1236 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1237 {
1238         unsigned long max = UINT_MAX;
1239         long pages;
1240
1241         if (max > SIZE_MAX / page_size)
1242                 max = SIZE_MAX / page_size;
1243
1244         pages = parse_pages_arg(str, 1, max);
1245         if (pages < 0) {
1246                 pr_err("Invalid argument for --mmap_pages/-m\n");
1247                 return -1;
1248         }
1249
1250         *mmap_pages = pages;
1251         return 0;
1252 }
1253
1254 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1255                                   int unset __maybe_unused)
1256 {
1257         return __perf_evlist__parse_mmap_pages(opt->value, str);
1258 }
1259
1260 /**
1261  * perf_evlist__mmap_ex - Create mmaps to receive events.
1262  * @evlist: list of events
1263  * @pages: map length in pages
1264  * @overwrite: overwrite older events?
1265  * @auxtrace_pages - auxtrace map length in pages
1266  * @auxtrace_overwrite - overwrite older auxtrace data?
1267  *
1268  * If @overwrite is %false the user needs to signal event consumption using
1269  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1270  * automatically.
1271  *
1272  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1273  * consumption using auxtrace_mmap__write_tail().
1274  *
1275  * Return: %0 on success, negative error code otherwise.
1276  */
1277 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1278                          bool overwrite, unsigned int auxtrace_pages,
1279                          bool auxtrace_overwrite)
1280 {
1281         struct perf_evsel *evsel;
1282         const struct cpu_map *cpus = evlist->cpus;
1283         const struct thread_map *threads = evlist->threads;
1284         struct mmap_params mp = {
1285                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1286         };
1287
1288         if (!evlist->mmap)
1289                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1290         if (!evlist->mmap)
1291                 return -ENOMEM;
1292
1293         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1294                 return -ENOMEM;
1295
1296         evlist->overwrite = overwrite;
1297         evlist->mmap_len = perf_evlist__mmap_size(pages);
1298         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1299         mp.mask = evlist->mmap_len - page_size - 1;
1300
1301         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1302                                    auxtrace_pages, auxtrace_overwrite);
1303
1304         evlist__for_each_entry(evlist, evsel) {
1305                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1306                     evsel->sample_id == NULL &&
1307                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1308                         return -ENOMEM;
1309         }
1310
1311         if (cpu_map__empty(cpus))
1312                 return perf_evlist__mmap_per_thread(evlist, &mp);
1313
1314         return perf_evlist__mmap_per_cpu(evlist, &mp);
1315 }
1316
1317 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1318                       bool overwrite)
1319 {
1320         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1321 }
1322
1323 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1324 {
1325         struct cpu_map *cpus;
1326         struct thread_map *threads;
1327
1328         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1329
1330         if (!threads)
1331                 return -1;
1332
1333         if (target__uses_dummy_map(target))
1334                 cpus = cpu_map__dummy_new();
1335         else
1336                 cpus = cpu_map__new(target->cpu_list);
1337
1338         if (!cpus)
1339                 goto out_delete_threads;
1340
1341         evlist->has_user_cpus = !!target->cpu_list;
1342
1343         perf_evlist__set_maps(evlist, cpus, threads);
1344
1345         return 0;
1346
1347 out_delete_threads:
1348         thread_map__put(threads);
1349         return -1;
1350 }
1351
1352 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1353                            struct thread_map *threads)
1354 {
1355         /*
1356          * Allow for the possibility that one or another of the maps isn't being
1357          * changed i.e. don't put it.  Note we are assuming the maps that are
1358          * being applied are brand new and evlist is taking ownership of the
1359          * original reference count of 1.  If that is not the case it is up to
1360          * the caller to increase the reference count.
1361          */
1362         if (cpus != evlist->cpus) {
1363                 cpu_map__put(evlist->cpus);
1364                 evlist->cpus = cpu_map__get(cpus);
1365         }
1366
1367         if (threads != evlist->threads) {
1368                 thread_map__put(evlist->threads);
1369                 evlist->threads = thread_map__get(threads);
1370         }
1371
1372         perf_evlist__propagate_maps(evlist);
1373 }
1374
1375 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1376                                    enum perf_event_sample_format bit)
1377 {
1378         struct perf_evsel *evsel;
1379
1380         evlist__for_each_entry(evlist, evsel)
1381                 __perf_evsel__set_sample_bit(evsel, bit);
1382 }
1383
1384 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1385                                      enum perf_event_sample_format bit)
1386 {
1387         struct perf_evsel *evsel;
1388
1389         evlist__for_each_entry(evlist, evsel)
1390                 __perf_evsel__reset_sample_bit(evsel, bit);
1391 }
1392
1393 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1394 {
1395         struct perf_evsel *evsel;
1396         int err = 0;
1397         const int ncpus = cpu_map__nr(evlist->cpus),
1398                   nthreads = thread_map__nr(evlist->threads);
1399
1400         evlist__for_each_entry(evlist, evsel) {
1401                 if (evsel->filter == NULL)
1402                         continue;
1403
1404                 /*
1405                  * filters only work for tracepoint event, which doesn't have cpu limit.
1406                  * So evlist and evsel should always be same.
1407                  */
1408                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1409                 if (err) {
1410                         *err_evsel = evsel;
1411                         break;
1412                 }
1413         }
1414
1415         return err;
1416 }
1417
1418 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1419 {
1420         struct perf_evsel *evsel;
1421         int err = 0;
1422
1423         evlist__for_each_entry(evlist, evsel) {
1424                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1425                         continue;
1426
1427                 err = perf_evsel__set_filter(evsel, filter);
1428                 if (err)
1429                         break;
1430         }
1431
1432         return err;
1433 }
1434
1435 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1436 {
1437         char *filter;
1438         int ret = -1;
1439         size_t i;
1440
1441         for (i = 0; i < npids; ++i) {
1442                 if (i == 0) {
1443                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1444                                 return -1;
1445                 } else {
1446                         char *tmp;
1447
1448                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1449                                 goto out_free;
1450
1451                         free(filter);
1452                         filter = tmp;
1453                 }
1454         }
1455
1456         ret = perf_evlist__set_filter(evlist, filter);
1457 out_free:
1458         free(filter);
1459         return ret;
1460 }
1461
1462 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1463 {
1464         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1465 }
1466
1467 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1468 {
1469         struct perf_evsel *pos;
1470
1471         if (evlist->nr_entries == 1)
1472                 return true;
1473
1474         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1475                 return false;
1476
1477         evlist__for_each_entry(evlist, pos) {
1478                 if (pos->id_pos != evlist->id_pos ||
1479                     pos->is_pos != evlist->is_pos)
1480                         return false;
1481         }
1482
1483         return true;
1484 }
1485
1486 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1487 {
1488         struct perf_evsel *evsel;
1489
1490         if (evlist->combined_sample_type)
1491                 return evlist->combined_sample_type;
1492
1493         evlist__for_each_entry(evlist, evsel)
1494                 evlist->combined_sample_type |= evsel->attr.sample_type;
1495
1496         return evlist->combined_sample_type;
1497 }
1498
1499 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1500 {
1501         evlist->combined_sample_type = 0;
1502         return __perf_evlist__combined_sample_type(evlist);
1503 }
1504
1505 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1506 {
1507         struct perf_evsel *evsel;
1508         u64 branch_type = 0;
1509
1510         evlist__for_each_entry(evlist, evsel)
1511                 branch_type |= evsel->attr.branch_sample_type;
1512         return branch_type;
1513 }
1514
1515 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1516 {
1517         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1518         u64 read_format = first->attr.read_format;
1519         u64 sample_type = first->attr.sample_type;
1520
1521         evlist__for_each_entry(evlist, pos) {
1522                 if (read_format != pos->attr.read_format)
1523                         return false;
1524         }
1525
1526         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1527         if ((sample_type & PERF_SAMPLE_READ) &&
1528             !(read_format & PERF_FORMAT_ID)) {
1529                 return false;
1530         }
1531
1532         return true;
1533 }
1534
1535 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1536 {
1537         struct perf_evsel *first = perf_evlist__first(evlist);
1538         return first->attr.read_format;
1539 }
1540
1541 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1542 {
1543         struct perf_evsel *first = perf_evlist__first(evlist);
1544         struct perf_sample *data;
1545         u64 sample_type;
1546         u16 size = 0;
1547
1548         if (!first->attr.sample_id_all)
1549                 goto out;
1550
1551         sample_type = first->attr.sample_type;
1552
1553         if (sample_type & PERF_SAMPLE_TID)
1554                 size += sizeof(data->tid) * 2;
1555
1556        if (sample_type & PERF_SAMPLE_TIME)
1557                 size += sizeof(data->time);
1558
1559         if (sample_type & PERF_SAMPLE_ID)
1560                 size += sizeof(data->id);
1561
1562         if (sample_type & PERF_SAMPLE_STREAM_ID)
1563                 size += sizeof(data->stream_id);
1564
1565         if (sample_type & PERF_SAMPLE_CPU)
1566                 size += sizeof(data->cpu) * 2;
1567
1568         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1569                 size += sizeof(data->id);
1570 out:
1571         return size;
1572 }
1573
1574 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1575 {
1576         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1577
1578         evlist__for_each_entry_continue(evlist, pos) {
1579                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1580                         return false;
1581         }
1582
1583         return true;
1584 }
1585
1586 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1587 {
1588         struct perf_evsel *first = perf_evlist__first(evlist);
1589         return first->attr.sample_id_all;
1590 }
1591
1592 void perf_evlist__set_selected(struct perf_evlist *evlist,
1593                                struct perf_evsel *evsel)
1594 {
1595         evlist->selected = evsel;
1596 }
1597
1598 void perf_evlist__close(struct perf_evlist *evlist)
1599 {
1600         struct perf_evsel *evsel;
1601         int ncpus = cpu_map__nr(evlist->cpus);
1602         int nthreads = thread_map__nr(evlist->threads);
1603         int n;
1604
1605         evlist__for_each_entry_reverse(evlist, evsel) {
1606                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1607                 perf_evsel__close(evsel, n, nthreads);
1608         }
1609 }
1610
1611 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1612 {
1613         struct cpu_map    *cpus;
1614         struct thread_map *threads;
1615         int err = -ENOMEM;
1616
1617         /*
1618          * Try reading /sys/devices/system/cpu/online to get
1619          * an all cpus map.
1620          *
1621          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1622          * code needs an overhaul to properly forward the
1623          * error, and we may not want to do that fallback to a
1624          * default cpu identity map :-\
1625          */
1626         cpus = cpu_map__new(NULL);
1627         if (!cpus)
1628                 goto out;
1629
1630         threads = thread_map__new_dummy();
1631         if (!threads)
1632                 goto out_put;
1633
1634         perf_evlist__set_maps(evlist, cpus, threads);
1635 out:
1636         return err;
1637 out_put:
1638         cpu_map__put(cpus);
1639         goto out;
1640 }
1641
1642 int perf_evlist__open(struct perf_evlist *evlist)
1643 {
1644         struct perf_evsel *evsel;
1645         int err;
1646
1647         /*
1648          * Default: one fd per CPU, all threads, aka systemwide
1649          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1650          */
1651         if (evlist->threads == NULL && evlist->cpus == NULL) {
1652                 err = perf_evlist__create_syswide_maps(evlist);
1653                 if (err < 0)
1654                         goto out_err;
1655         }
1656
1657         perf_evlist__update_id_pos(evlist);
1658
1659         evlist__for_each_entry(evlist, evsel) {
1660                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1661                 if (err < 0)
1662                         goto out_err;
1663         }
1664
1665         return 0;
1666 out_err:
1667         perf_evlist__close(evlist);
1668         errno = -err;
1669         return err;
1670 }
1671
1672 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1673                                   const char *argv[], bool pipe_output,
1674                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1675 {
1676         int child_ready_pipe[2], go_pipe[2];
1677         char bf;
1678
1679         if (pipe(child_ready_pipe) < 0) {
1680                 perror("failed to create 'ready' pipe");
1681                 return -1;
1682         }
1683
1684         if (pipe(go_pipe) < 0) {
1685                 perror("failed to create 'go' pipe");
1686                 goto out_close_ready_pipe;
1687         }
1688
1689         evlist->workload.pid = fork();
1690         if (evlist->workload.pid < 0) {
1691                 perror("failed to fork");
1692                 goto out_close_pipes;
1693         }
1694
1695         if (!evlist->workload.pid) {
1696                 int ret;
1697
1698                 if (pipe_output)
1699                         dup2(2, 1);
1700
1701                 signal(SIGTERM, SIG_DFL);
1702
1703                 close(child_ready_pipe[0]);
1704                 close(go_pipe[1]);
1705                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1706
1707                 /*
1708                  * Tell the parent we're ready to go
1709                  */
1710                 close(child_ready_pipe[1]);
1711
1712                 /*
1713                  * Wait until the parent tells us to go.
1714                  */
1715                 ret = read(go_pipe[0], &bf, 1);
1716                 /*
1717                  * The parent will ask for the execvp() to be performed by
1718                  * writing exactly one byte, in workload.cork_fd, usually via
1719                  * perf_evlist__start_workload().
1720                  *
1721                  * For cancelling the workload without actually running it,
1722                  * the parent will just close workload.cork_fd, without writing
1723                  * anything, i.e. read will return zero and we just exit()
1724                  * here.
1725                  */
1726                 if (ret != 1) {
1727                         if (ret == -1)
1728                                 perror("unable to read pipe");
1729                         exit(ret);
1730                 }
1731
1732                 execvp(argv[0], (char **)argv);
1733
1734                 if (exec_error) {
1735                         union sigval val;
1736
1737                         val.sival_int = errno;
1738                         if (sigqueue(getppid(), SIGUSR1, val))
1739                                 perror(argv[0]);
1740                 } else
1741                         perror(argv[0]);
1742                 exit(-1);
1743         }
1744
1745         if (exec_error) {
1746                 struct sigaction act = {
1747                         .sa_flags     = SA_SIGINFO,
1748                         .sa_sigaction = exec_error,
1749                 };
1750                 sigaction(SIGUSR1, &act, NULL);
1751         }
1752
1753         if (target__none(target)) {
1754                 if (evlist->threads == NULL) {
1755                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1756                                 __func__, __LINE__);
1757                         goto out_close_pipes;
1758                 }
1759                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1760         }
1761
1762         close(child_ready_pipe[1]);
1763         close(go_pipe[0]);
1764         /*
1765          * wait for child to settle
1766          */
1767         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1768                 perror("unable to read pipe");
1769                 goto out_close_pipes;
1770         }
1771
1772         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1773         evlist->workload.cork_fd = go_pipe[1];
1774         close(child_ready_pipe[0]);
1775         return 0;
1776
1777 out_close_pipes:
1778         close(go_pipe[0]);
1779         close(go_pipe[1]);
1780 out_close_ready_pipe:
1781         close(child_ready_pipe[0]);
1782         close(child_ready_pipe[1]);
1783         return -1;
1784 }
1785
1786 int perf_evlist__start_workload(struct perf_evlist *evlist)
1787 {
1788         if (evlist->workload.cork_fd > 0) {
1789                 char bf = 0;
1790                 int ret;
1791                 /*
1792                  * Remove the cork, let it rip!
1793                  */
1794                 ret = write(evlist->workload.cork_fd, &bf, 1);
1795                 if (ret < 0)
1796                         perror("enable to write to pipe");
1797
1798                 close(evlist->workload.cork_fd);
1799                 return ret;
1800         }
1801
1802         return 0;
1803 }
1804
1805 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1806                               struct perf_sample *sample)
1807 {
1808         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1809
1810         if (!evsel)
1811                 return -EFAULT;
1812         return perf_evsel__parse_sample(evsel, event, sample);
1813 }
1814
1815 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1816 {
1817         struct perf_evsel *evsel;
1818         size_t printed = 0;
1819
1820         evlist__for_each_entry(evlist, evsel) {
1821                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1822                                    perf_evsel__name(evsel));
1823         }
1824
1825         return printed + fprintf(fp, "\n");
1826 }
1827
1828 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1829                                int err, char *buf, size_t size)
1830 {
1831         int printed, value;
1832         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1833
1834         switch (err) {
1835         case EACCES:
1836         case EPERM:
1837                 printed = scnprintf(buf, size,
1838                                     "Error:\t%s.\n"
1839                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1840
1841                 value = perf_event_paranoid();
1842
1843                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1844
1845                 if (value >= 2) {
1846                         printed += scnprintf(buf + printed, size - printed,
1847                                              "For your workloads it needs to be <= 1\nHint:\t");
1848                 }
1849                 printed += scnprintf(buf + printed, size - printed,
1850                                      "For system wide tracing it needs to be set to -1.\n");
1851
1852                 printed += scnprintf(buf + printed, size - printed,
1853                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1854                                     "Hint:\tThe current value is %d.", value);
1855                 break;
1856         case EINVAL: {
1857                 struct perf_evsel *first = perf_evlist__first(evlist);
1858                 int max_freq;
1859
1860                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1861                         goto out_default;
1862
1863                 if (first->attr.sample_freq < (u64)max_freq)
1864                         goto out_default;
1865
1866                 printed = scnprintf(buf, size,
1867                                     "Error:\t%s.\n"
1868                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1869                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1870                                     emsg, max_freq, first->attr.sample_freq);
1871                 break;
1872         }
1873         default:
1874 out_default:
1875                 scnprintf(buf, size, "%s", emsg);
1876                 break;
1877         }
1878
1879         return 0;
1880 }
1881
1882 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1883 {
1884         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1885         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1886
1887         switch (err) {
1888         case EPERM:
1889                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1890                 printed += scnprintf(buf + printed, size - printed,
1891                                      "Error:\t%s.\n"
1892                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1893                                      "Hint:\tTried using %zd kB.\n",
1894                                      emsg, pages_max_per_user, pages_attempted);
1895
1896                 if (pages_attempted >= pages_max_per_user) {
1897                         printed += scnprintf(buf + printed, size - printed,
1898                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1899                                              pages_max_per_user + pages_attempted);
1900                 }
1901
1902                 printed += scnprintf(buf + printed, size - printed,
1903                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1904                 break;
1905         default:
1906                 scnprintf(buf, size, "%s", emsg);
1907                 break;
1908         }
1909
1910         return 0;
1911 }
1912
1913 void perf_evlist__to_front(struct perf_evlist *evlist,
1914                            struct perf_evsel *move_evsel)
1915 {
1916         struct perf_evsel *evsel, *n;
1917         LIST_HEAD(move);
1918
1919         if (move_evsel == perf_evlist__first(evlist))
1920                 return;
1921
1922         evlist__for_each_entry_safe(evlist, n, evsel) {
1923                 if (evsel->leader == move_evsel->leader)
1924                         list_move_tail(&evsel->node, &move);
1925         }
1926
1927         list_splice(&move, &evlist->entries);
1928 }
1929
1930 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1931                                      struct perf_evsel *tracking_evsel)
1932 {
1933         struct perf_evsel *evsel;
1934
1935         if (tracking_evsel->tracking)
1936                 return;
1937
1938         evlist__for_each_entry(evlist, evsel) {
1939                 if (evsel != tracking_evsel)
1940                         evsel->tracking = false;
1941         }
1942
1943         tracking_evsel->tracking = true;
1944 }
1945
1946 struct perf_evsel *
1947 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1948                                const char *str)
1949 {
1950         struct perf_evsel *evsel;
1951
1952         evlist__for_each_entry(evlist, evsel) {
1953                 if (!evsel->name)
1954                         continue;
1955                 if (strcmp(str, evsel->name) == 0)
1956                         return evsel;
1957         }
1958
1959         return NULL;
1960 }
1961
1962 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1963                                   enum bkw_mmap_state state)
1964 {
1965         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1966         enum action {
1967                 NONE,
1968                 PAUSE,
1969                 RESUME,
1970         } action = NONE;
1971
1972         if (!evlist->backward_mmap)
1973                 return;
1974
1975         switch (old_state) {
1976         case BKW_MMAP_NOTREADY: {
1977                 if (state != BKW_MMAP_RUNNING)
1978                         goto state_err;;
1979                 break;
1980         }
1981         case BKW_MMAP_RUNNING: {
1982                 if (state != BKW_MMAP_DATA_PENDING)
1983                         goto state_err;
1984                 action = PAUSE;
1985                 break;
1986         }
1987         case BKW_MMAP_DATA_PENDING: {
1988                 if (state != BKW_MMAP_EMPTY)
1989                         goto state_err;
1990                 break;
1991         }
1992         case BKW_MMAP_EMPTY: {
1993                 if (state != BKW_MMAP_RUNNING)
1994                         goto state_err;
1995                 action = RESUME;
1996                 break;
1997         }
1998         default:
1999                 WARN_ONCE(1, "Shouldn't get there\n");
2000         }
2001
2002         evlist->bkw_mmap_state = state;
2003
2004         switch (action) {
2005         case PAUSE:
2006                 perf_evlist__pause(evlist);
2007                 break;
2008         case RESUME:
2009                 perf_evlist__resume(evlist);
2010                 break;
2011         case NONE:
2012         default:
2013                 break;
2014         }
2015
2016 state_err:
2017         return;
2018 }