tools/perf/util/evlist.c

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