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;
 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 (!refcount_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 (!refcount_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 (!refcount_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         refcount_inc(&map->refcnt);
 879 }
 880
 881 static void perf_mmap__put(struct perf_mmap *md)
 882 {
 883         BUG_ON(md->base && refcount_read(&md->refcnt) == 0);
 884
 885         if (refcount_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 (refcount_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                 refcount_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                 /*
 980                  * When the perf_mmap() call is made we grab one refcount, plus
 981                  * one extra to let perf_evlist__mmap_consume() get the last
 982                  * events after all real references (perf_mmap__get()) are
 983                  * dropped.
 984                  *
 985                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 986                  * thus does perf_mmap__get() on it.
 987                  */
 988                 refcount_set(&map[i].refcnt, 0);
 989         }
 990         return map;
 991 }
 992
 993 struct mmap_params {
 994         int prot;
 995         int mask;
 996         struct auxtrace_mmap_params auxtrace_mp;
 997 };
 998
 999 static int perf_mmap__mmap(struct perf_mmap *map,
1000                            struct mmap_params *mp, int fd)
1001 {
1002         /*
1003          * The last one will be done at perf_evlist__mmap_consume(), so that we
1004          * make sure we don't prevent tools from consuming every last event in
1005          * the ring buffer.
1006          *
1007          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
1008          * anymore, but the last events for it are still in the ring buffer,
1009          * waiting to be consumed.
1010          *
1011          * Tools can chose to ignore this at their own discretion, but the
1012          * evlist layer can't just drop it when filtering events in
1013          * perf_evlist__filter_pollfd().
1014          */
1015         refcount_set(&map->refcnt, 2);
1016         map->prev = 0;
1017         map->mask = mp->mask;
1018         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
1019                          MAP_SHARED, fd, 0);
1020         if (map->base == MAP_FAILED) {
1021                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1022                           errno);
1023                 map->base = NULL;
1024                 return -1;
1025         }
1026         map->fd = fd;
1027
1028         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
1029                                 &mp->auxtrace_mp, map->base, fd))
1030                 return -1;
1031
1032         return 0;
1033 }
1034
1035 static bool
1036 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
1037                          struct perf_evsel *evsel)
1038 {
1039         if (evsel->attr.write_backward)
1040                 return false;
1041         return true;
1042 }
1043
1044 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
1045                                        struct mmap_params *mp, int cpu_idx,
1046                                        int thread, int *_output, int *_output_backward)
1047 {
1048         struct perf_evsel *evsel;
1049         int revent;
1050         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
1051
1052         evlist__for_each_entry(evlist, evsel) {
1053                 struct perf_mmap *maps = evlist->mmap;
1054                 int *output = _output;
1055                 int fd;
1056                 int cpu;
1057
1058                 if (evsel->attr.write_backward) {
1059                         output = _output_backward;
1060                         maps = evlist->backward_mmap;
1061
1062                         if (!maps) {
1063                                 maps = perf_evlist__alloc_mmap(evlist);
1064                                 if (!maps)
1065                                         return -1;
1066                                 evlist->backward_mmap = maps;
1067                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1068                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1069                         }
1070                 }
1071
1072                 if (evsel->system_wide && thread)
1073                         continue;
1074
1075                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
1076                 if (cpu == -1)
1077                         continue;
1078
1079                 fd = FD(evsel, cpu, thread);
1080
1081                 if (*output == -1) {
1082                         *output = fd;
1083
1084                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1085                                 return -1;
1086                 } else {
1087                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1088                                 return -1;
1089
1090                         perf_mmap__get(&maps[idx]);
1091                 }
1092
1093                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1094
1095                 /*
1096                  * The system_wide flag causes a selected event to be opened
1097                  * always without a pid.  Consequently it will never get a
1098                  * POLLHUP, but it is used for tracking in combination with
1099                  * other events, so it should not need to be polled anyway.
1100                  * Therefore don't add it for polling.
1101                  */
1102                 if (!evsel->system_wide &&
1103                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1104                         perf_mmap__put(&maps[idx]);
1105                         return -1;
1106                 }
1107
1108                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1109                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1110                                                    fd) < 0)
1111                                 return -1;
1112                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1113                                                  thread);
1114                 }
1115         }
1116
1117         return 0;
1118 }
1119
1120 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1121                                      struct mmap_params *mp)
1122 {
1123         int cpu, thread;
1124         int nr_cpus = cpu_map__nr(evlist->cpus);
1125         int nr_threads = thread_map__nr(evlist->threads);
1126
1127         pr_debug2("perf event ring buffer mmapped per cpu\n");
1128         for (cpu = 0; cpu < nr_cpus; cpu++) {
1129                 int output = -1;
1130                 int output_backward = -1;
1131
1132                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1133                                               true);
1134
1135                 for (thread = 0; thread < nr_threads; thread++) {
1136                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1137                                                         thread, &output, &output_backward))
1138                                 goto out_unmap;
1139                 }
1140         }
1141
1142         return 0;
1143
1144 out_unmap:
1145         perf_evlist__munmap_nofree(evlist);
1146         return -1;
1147 }
1148
1149 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1150                                         struct mmap_params *mp)
1151 {
1152         int thread;
1153         int nr_threads = thread_map__nr(evlist->threads);
1154
1155         pr_debug2("perf event ring buffer mmapped per thread\n");
1156         for (thread = 0; thread < nr_threads; thread++) {
1157                 int output = -1;
1158                 int output_backward = -1;
1159
1160                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1161                                               false);
1162
1163                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1164                                                 &output, &output_backward))
1165                         goto out_unmap;
1166         }
1167
1168         return 0;
1169
1170 out_unmap:
1171         perf_evlist__munmap_nofree(evlist);
1172         return -1;
1173 }
1174
1175 unsigned long perf_event_mlock_kb_in_pages(void)
1176 {
1177         unsigned long pages;
1178         int max;
1179
1180         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1181                 /*
1182                  * Pick a once upon a time good value, i.e. things look
1183                  * strange since we can't read a sysctl value, but lets not
1184                  * die yet...
1185                  */
1186                 max = 512;
1187         } else {
1188                 max -= (page_size / 1024);
1189         }
1190
1191         pages = (max * 1024) / page_size;
1192         if (!is_power_of_2(pages))
1193                 pages = rounddown_pow_of_two(pages);
1194
1195         return pages;
1196 }
1197
1198 size_t perf_evlist__mmap_size(unsigned long pages)
1199 {
1200         if (pages == UINT_MAX)
1201                 pages = perf_event_mlock_kb_in_pages();
1202         else if (!is_power_of_2(pages))
1203                 return 0;
1204
1205         return (pages + 1) * page_size;
1206 }
1207
1208 static long parse_pages_arg(const char *str, unsigned long min,
1209                             unsigned long max)
1210 {
1211         unsigned long pages, val;
1212         static struct parse_tag tags[] = {
1213                 { .tag  = 'B', .mult = 1       },
1214                 { .tag  = 'K', .mult = 1 << 10 },
1215                 { .tag  = 'M', .mult = 1 << 20 },
1216                 { .tag  = 'G', .mult = 1 << 30 },
1217                 { .tag  = 0 },
1218         };
1219
1220         if (str == NULL)
1221                 return -EINVAL;
1222
1223         val = parse_tag_value(str, tags);
1224         if (val != (unsigned long) -1) {
1225                 /* we got file size value */
1226                 pages = PERF_ALIGN(val, page_size) / page_size;
1227         } else {
1228                 /* we got pages count value */
1229                 char *eptr;
1230                 pages = strtoul(str, &eptr, 10);
1231                 if (*eptr != '\0')
1232                         return -EINVAL;
1233         }
1234
1235         if (pages == 0 && min == 0) {
1236                 /* leave number of pages at 0 */
1237         } else if (!is_power_of_2(pages)) {
1238                 char buf[100];
1239
1240                 /* round pages up to next power of 2 */
1241                 pages = roundup_pow_of_two(pages);
1242                 if (!pages)
1243                         return -EINVAL;
1244
1245                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1246                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1247                         buf, pages);
1248         }
1249
1250         if (pages > max)
1251                 return -EINVAL;
1252
1253         return pages;
1254 }
1255
1256 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1257 {
1258         unsigned long max = UINT_MAX;
1259         long pages;
1260
1261         if (max > SIZE_MAX / page_size)
1262                 max = SIZE_MAX / page_size;
1263
1264         pages = parse_pages_arg(str, 1, max);
1265         if (pages < 0) {
1266                 pr_err("Invalid argument for --mmap_pages/-m\n");
1267                 return -1;
1268         }
1269
1270         *mmap_pages = pages;
1271         return 0;
1272 }
1273
1274 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1275                                   int unset __maybe_unused)
1276 {
1277         return __perf_evlist__parse_mmap_pages(opt->value, str);
1278 }
1279
1280 /**
1281  * perf_evlist__mmap_ex - Create mmaps to receive events.
1282  * @evlist: list of events
1283  * @pages: map length in pages
1284  * @overwrite: overwrite older events?
1285  * @auxtrace_pages - auxtrace map length in pages
1286  * @auxtrace_overwrite - overwrite older auxtrace data?
1287  *
1288  * If @overwrite is %false the user needs to signal event consumption using
1289  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1290  * automatically.
1291  *
1292  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1293  * consumption using auxtrace_mmap__write_tail().
1294  *
1295  * Return: %0 on success, negative error code otherwise.
1296  */
1297 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1298                          bool overwrite, unsigned int auxtrace_pages,
1299                          bool auxtrace_overwrite)
1300 {
1301         struct perf_evsel *evsel;
1302         const struct cpu_map *cpus = evlist->cpus;
1303         const struct thread_map *threads = evlist->threads;
1304         struct mmap_params mp = {
1305                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1306         };
1307
1308         if (!evlist->mmap)
1309                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1310         if (!evlist->mmap)
1311                 return -ENOMEM;
1312
1313         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1314                 return -ENOMEM;
1315
1316         evlist->overwrite = overwrite;
1317         evlist->mmap_len = perf_evlist__mmap_size(pages);
1318         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1319         mp.mask = evlist->mmap_len - page_size - 1;
1320
1321         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1322                                    auxtrace_pages, auxtrace_overwrite);
1323
1324         evlist__for_each_entry(evlist, evsel) {
1325                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1326                     evsel->sample_id == NULL &&
1327                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1328                         return -ENOMEM;
1329         }
1330
1331         if (cpu_map__empty(cpus))
1332                 return perf_evlist__mmap_per_thread(evlist, &mp);
1333
1334         return perf_evlist__mmap_per_cpu(evlist, &mp);
1335 }
1336
1337 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1338                       bool overwrite)
1339 {
1340         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1341 }
1342
1343 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1344 {
1345         struct cpu_map *cpus;
1346         struct thread_map *threads;
1347
1348         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1349
1350         if (!threads)
1351                 return -1;
1352
1353         if (target__uses_dummy_map(target))
1354                 cpus = cpu_map__dummy_new();
1355         else
1356                 cpus = cpu_map__new(target->cpu_list);
1357
1358         if (!cpus)
1359                 goto out_delete_threads;
1360
1361         evlist->has_user_cpus = !!target->cpu_list;
1362
1363         perf_evlist__set_maps(evlist, cpus, threads);
1364
1365         return 0;
1366
1367 out_delete_threads:
1368         thread_map__put(threads);
1369         return -1;
1370 }
1371
1372 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1373                            struct thread_map *threads)
1374 {
1375         /*
1376          * Allow for the possibility that one or another of the maps isn't being
1377          * changed i.e. don't put it.  Note we are assuming the maps that are
1378          * being applied are brand new and evlist is taking ownership of the
1379          * original reference count of 1.  If that is not the case it is up to
1380          * the caller to increase the reference count.
1381          */
1382         if (cpus != evlist->cpus) {
1383                 cpu_map__put(evlist->cpus);
1384                 evlist->cpus = cpu_map__get(cpus);
1385         }
1386
1387         if (threads != evlist->threads) {
1388                 thread_map__put(evlist->threads);
1389                 evlist->threads = thread_map__get(threads);
1390         }
1391
1392         perf_evlist__propagate_maps(evlist);
1393 }
1394
1395 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1396                                    enum perf_event_sample_format bit)
1397 {
1398         struct perf_evsel *evsel;
1399
1400         evlist__for_each_entry(evlist, evsel)
1401                 __perf_evsel__set_sample_bit(evsel, bit);
1402 }
1403
1404 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1405                                      enum perf_event_sample_format bit)
1406 {
1407         struct perf_evsel *evsel;
1408
1409         evlist__for_each_entry(evlist, evsel)
1410                 __perf_evsel__reset_sample_bit(evsel, bit);
1411 }
1412
1413 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1414 {
1415         struct perf_evsel *evsel;
1416         int err = 0;
1417         const int ncpus = cpu_map__nr(evlist->cpus),
1418                   nthreads = thread_map__nr(evlist->threads);
1419
1420         evlist__for_each_entry(evlist, evsel) {
1421                 if (evsel->filter == NULL)
1422                         continue;
1423
1424                 /*
1425                  * filters only work for tracepoint event, which doesn't have cpu limit.
1426                  * So evlist and evsel should always be same.
1427                  */
1428                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1429                 if (err) {
1430                         *err_evsel = evsel;
1431                         break;
1432                 }
1433         }
1434
1435         return err;
1436 }
1437
1438 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1439 {
1440         struct perf_evsel *evsel;
1441         int err = 0;
1442
1443         evlist__for_each_entry(evlist, evsel) {
1444                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1445                         continue;
1446
1447                 err = perf_evsel__set_filter(evsel, filter);
1448                 if (err)
1449                         break;
1450         }
1451
1452         return err;
1453 }
1454
1455 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1456 {
1457         char *filter;
1458         int ret = -1;
1459         size_t i;
1460
1461         for (i = 0; i < npids; ++i) {
1462                 if (i == 0) {
1463                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1464                                 return -1;
1465                 } else {
1466                         char *tmp;
1467
1468                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1469                                 goto out_free;
1470
1471                         free(filter);
1472                         filter = tmp;
1473                 }
1474         }
1475
1476         ret = perf_evlist__set_filter(evlist, filter);
1477 out_free:
1478         free(filter);
1479         return ret;
1480 }
1481
1482 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1483 {
1484         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1485 }
1486
1487 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1488 {
1489         struct perf_evsel *pos;
1490
1491         if (evlist->nr_entries == 1)
1492                 return true;
1493
1494         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1495                 return false;
1496
1497         evlist__for_each_entry(evlist, pos) {
1498                 if (pos->id_pos != evlist->id_pos ||
1499                     pos->is_pos != evlist->is_pos)
1500                         return false;
1501         }
1502
1503         return true;
1504 }
1505
1506 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1507 {
1508         struct perf_evsel *evsel;
1509
1510         if (evlist->combined_sample_type)
1511                 return evlist->combined_sample_type;
1512
1513         evlist__for_each_entry(evlist, evsel)
1514                 evlist->combined_sample_type |= evsel->attr.sample_type;
1515
1516         return evlist->combined_sample_type;
1517 }
1518
1519 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1520 {
1521         evlist->combined_sample_type = 0;
1522         return __perf_evlist__combined_sample_type(evlist);
1523 }
1524
1525 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1526 {
1527         struct perf_evsel *evsel;
1528         u64 branch_type = 0;
1529
1530         evlist__for_each_entry(evlist, evsel)
1531                 branch_type |= evsel->attr.branch_sample_type;
1532         return branch_type;
1533 }
1534
1535 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1536 {
1537         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1538         u64 read_format = first->attr.read_format;
1539         u64 sample_type = first->attr.sample_type;
1540
1541         evlist__for_each_entry(evlist, pos) {
1542                 if (read_format != pos->attr.read_format)
1543                         return false;
1544         }
1545
1546         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1547         if ((sample_type & PERF_SAMPLE_READ) &&
1548             !(read_format & PERF_FORMAT_ID)) {
1549                 return false;
1550         }
1551
1552         return true;
1553 }
1554
1555 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1556 {
1557         struct perf_evsel *first = perf_evlist__first(evlist);
1558         return first->attr.read_format;
1559 }
1560
1561 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1562 {
1563         struct perf_evsel *first = perf_evlist__first(evlist);
1564         struct perf_sample *data;
1565         u64 sample_type;
1566         u16 size = 0;
1567
1568         if (!first->attr.sample_id_all)
1569                 goto out;
1570
1571         sample_type = first->attr.sample_type;
1572
1573         if (sample_type & PERF_SAMPLE_TID)
1574                 size += sizeof(data->tid) * 2;
1575
1576        if (sample_type & PERF_SAMPLE_TIME)
1577                 size += sizeof(data->time);
1578
1579         if (sample_type & PERF_SAMPLE_ID)
1580                 size += sizeof(data->id);
1581
1582         if (sample_type & PERF_SAMPLE_STREAM_ID)
1583                 size += sizeof(data->stream_id);
1584
1585         if (sample_type & PERF_SAMPLE_CPU)
1586                 size += sizeof(data->cpu) * 2;
1587
1588         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1589                 size += sizeof(data->id);
1590 out:
1591         return size;
1592 }
1593
1594 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1595 {
1596         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1597
1598         evlist__for_each_entry_continue(evlist, pos) {
1599                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1600                         return false;
1601         }
1602
1603         return true;
1604 }
1605
1606 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1607 {
1608         struct perf_evsel *first = perf_evlist__first(evlist);
1609         return first->attr.sample_id_all;
1610 }
1611
1612 void perf_evlist__set_selected(struct perf_evlist *evlist,
1613                                struct perf_evsel *evsel)
1614 {
1615         evlist->selected = evsel;
1616 }
1617
1618 void perf_evlist__close(struct perf_evlist *evlist)
1619 {
1620         struct perf_evsel *evsel;
1621         int ncpus = cpu_map__nr(evlist->cpus);
1622         int nthreads = thread_map__nr(evlist->threads);
1623
1624         evlist__for_each_entry_reverse(evlist, evsel) {
1625                 int n = evsel->cpus ? evsel->cpus->nr : ncpus;
1626                 perf_evsel__close(evsel, n, nthreads);
1627         }
1628 }
1629
1630 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1631 {
1632         struct cpu_map    *cpus;
1633         struct thread_map *threads;
1634         int err = -ENOMEM;
1635
1636         /*
1637          * Try reading /sys/devices/system/cpu/online to get
1638          * an all cpus map.
1639          *
1640          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1641          * code needs an overhaul to properly forward the
1642          * error, and we may not want to do that fallback to a
1643          * default cpu identity map :-\
1644          */
1645         cpus = cpu_map__new(NULL);
1646         if (!cpus)
1647                 goto out;
1648
1649         threads = thread_map__new_dummy();
1650         if (!threads)
1651                 goto out_put;
1652
1653         perf_evlist__set_maps(evlist, cpus, threads);
1654 out:
1655         return err;
1656 out_put:
1657         cpu_map__put(cpus);
1658         goto out;
1659 }
1660
1661 int perf_evlist__open(struct perf_evlist *evlist)
1662 {
1663         struct perf_evsel *evsel;
1664         int err;
1665
1666         /*
1667          * Default: one fd per CPU, all threads, aka systemwide
1668          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1669          */
1670         if (evlist->threads == NULL && evlist->cpus == NULL) {
1671                 err = perf_evlist__create_syswide_maps(evlist);
1672                 if (err < 0)
1673                         goto out_err;
1674         }
1675
1676         perf_evlist__update_id_pos(evlist);
1677
1678         evlist__for_each_entry(evlist, evsel) {
1679                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1680                 if (err < 0)
1681                         goto out_err;
1682         }
1683
1684         return 0;
1685 out_err:
1686         perf_evlist__close(evlist);
1687         errno = -err;
1688         return err;
1689 }
1690
1691 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1692                                   const char *argv[], bool pipe_output,
1693                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1694 {
1695         int child_ready_pipe[2], go_pipe[2];
1696         char bf;
1697
1698         if (pipe(child_ready_pipe) < 0) {
1699                 perror("failed to create 'ready' pipe");
1700                 return -1;
1701         }
1702
1703         if (pipe(go_pipe) < 0) {
1704                 perror("failed to create 'go' pipe");
1705                 goto out_close_ready_pipe;
1706         }
1707
1708         evlist->workload.pid = fork();
1709         if (evlist->workload.pid < 0) {
1710                 perror("failed to fork");
1711                 goto out_close_pipes;
1712         }
1713
1714         if (!evlist->workload.pid) {
1715                 int ret;
1716
1717                 if (pipe_output)
1718                         dup2(2, 1);
1719
1720                 signal(SIGTERM, SIG_DFL);
1721
1722                 close(child_ready_pipe[0]);
1723                 close(go_pipe[1]);
1724                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1725
1726                 /*
1727                  * Tell the parent we're ready to go
1728                  */
1729                 close(child_ready_pipe[1]);
1730
1731                 /*
1732                  * Wait until the parent tells us to go.
1733                  */
1734                 ret = read(go_pipe[0], &bf, 1);
1735                 /*
1736                  * The parent will ask for the execvp() to be performed by
1737                  * writing exactly one byte, in workload.cork_fd, usually via
1738                  * perf_evlist__start_workload().
1739                  *
1740                  * For cancelling the workload without actually running it,
1741                  * the parent will just close workload.cork_fd, without writing
1742                  * anything, i.e. read will return zero and we just exit()
1743                  * here.
1744                  */
1745                 if (ret != 1) {
1746                         if (ret == -1)
1747                                 perror("unable to read pipe");
1748                         exit(ret);
1749                 }
1750
1751                 execvp(argv[0], (char **)argv);
1752
1753                 if (exec_error) {
1754                         union sigval val;
1755
1756                         val.sival_int = errno;
1757                         if (sigqueue(getppid(), SIGUSR1, val))
1758                                 perror(argv[0]);
1759                 } else
1760                         perror(argv[0]);
1761                 exit(-1);
1762         }
1763
1764         if (exec_error) {
1765                 struct sigaction act = {
1766                         .sa_flags     = SA_SIGINFO,
1767                         .sa_sigaction = exec_error,
1768                 };
1769                 sigaction(SIGUSR1, &act, NULL);
1770         }
1771
1772         if (target__none(target)) {
1773                 if (evlist->threads == NULL) {
1774                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1775                                 __func__, __LINE__);
1776                         goto out_close_pipes;
1777                 }
1778                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1779         }
1780
1781         close(child_ready_pipe[1]);
1782         close(go_pipe[0]);
1783         /*
1784          * wait for child to settle
1785          */
1786         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1787                 perror("unable to read pipe");
1788                 goto out_close_pipes;
1789         }
1790
1791         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1792         evlist->workload.cork_fd = go_pipe[1];
1793         close(child_ready_pipe[0]);
1794         return 0;
1795
1796 out_close_pipes:
1797         close(go_pipe[0]);
1798         close(go_pipe[1]);
1799 out_close_ready_pipe:
1800         close(child_ready_pipe[0]);
1801         close(child_ready_pipe[1]);
1802         return -1;
1803 }
1804
1805 int perf_evlist__start_workload(struct perf_evlist *evlist)
1806 {
1807         if (evlist->workload.cork_fd > 0) {
1808                 char bf = 0;
1809                 int ret;
1810                 /*
1811                  * Remove the cork, let it rip!
1812                  */
1813                 ret = write(evlist->workload.cork_fd, &bf, 1);
1814                 if (ret < 0)
1815                         perror("unable to write to pipe");
1816
1817                 close(evlist->workload.cork_fd);
1818                 return ret;
1819         }
1820
1821         return 0;
1822 }
1823
1824 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1825                               struct perf_sample *sample)
1826 {
1827         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1828
1829         if (!evsel)
1830                 return -EFAULT;
1831         return perf_evsel__parse_sample(evsel, event, sample);
1832 }
1833
1834 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1835 {
1836         struct perf_evsel *evsel;
1837         size_t printed = 0;
1838
1839         evlist__for_each_entry(evlist, evsel) {
1840                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1841                                    perf_evsel__name(evsel));
1842         }
1843
1844         return printed + fprintf(fp, "\n");
1845 }
1846
1847 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1848                                int err, char *buf, size_t size)
1849 {
1850         int printed, value;
1851         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1852
1853         switch (err) {
1854         case EACCES:
1855         case EPERM:
1856                 printed = scnprintf(buf, size,
1857                                     "Error:\t%s.\n"
1858                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1859
1860                 value = perf_event_paranoid();
1861
1862                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1863
1864                 if (value >= 2) {
1865                         printed += scnprintf(buf + printed, size - printed,
1866                                              "For your workloads it needs to be <= 1\nHint:\t");
1867                 }
1868                 printed += scnprintf(buf + printed, size - printed,
1869                                      "For system wide tracing it needs to be set to -1.\n");
1870
1871                 printed += scnprintf(buf + printed, size - printed,
1872                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1873                                     "Hint:\tThe current value is %d.", value);
1874                 break;
1875         case EINVAL: {
1876                 struct perf_evsel *first = perf_evlist__first(evlist);
1877                 int max_freq;
1878
1879                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1880                         goto out_default;
1881
1882                 if (first->attr.sample_freq < (u64)max_freq)
1883                         goto out_default;
1884
1885                 printed = scnprintf(buf, size,
1886                                     "Error:\t%s.\n"
1887                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1888                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1889                                     emsg, max_freq, first->attr.sample_freq);
1890                 break;
1891         }
1892         default:
1893 out_default:
1894                 scnprintf(buf, size, "%s", emsg);
1895                 break;
1896         }
1897
1898         return 0;
1899 }
1900
1901 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1902 {
1903         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1904         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1905
1906         switch (err) {
1907         case EPERM:
1908                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1909                 printed += scnprintf(buf + printed, size - printed,
1910                                      "Error:\t%s.\n"
1911                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1912                                      "Hint:\tTried using %zd kB.\n",
1913                                      emsg, pages_max_per_user, pages_attempted);
1914
1915                 if (pages_attempted >= pages_max_per_user) {
1916                         printed += scnprintf(buf + printed, size - printed,
1917                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1918                                              pages_max_per_user + pages_attempted);
1919                 }
1920
1921                 printed += scnprintf(buf + printed, size - printed,
1922                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1923                 break;
1924         default:
1925                 scnprintf(buf, size, "%s", emsg);
1926                 break;
1927         }
1928
1929         return 0;
1930 }
1931
1932 void perf_evlist__to_front(struct perf_evlist *evlist,
1933                            struct perf_evsel *move_evsel)
1934 {
1935         struct perf_evsel *evsel, *n;
1936         LIST_HEAD(move);
1937
1938         if (move_evsel == perf_evlist__first(evlist))
1939                 return;
1940
1941         evlist__for_each_entry_safe(evlist, n, evsel) {
1942                 if (evsel->leader == move_evsel->leader)
1943                         list_move_tail(&evsel->node, &move);
1944         }
1945
1946         list_splice(&move, &evlist->entries);
1947 }
1948
1949 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1950                                      struct perf_evsel *tracking_evsel)
1951 {
1952         struct perf_evsel *evsel;
1953
1954         if (tracking_evsel->tracking)
1955                 return;
1956
1957         evlist__for_each_entry(evlist, evsel) {
1958                 if (evsel != tracking_evsel)
1959                         evsel->tracking = false;
1960         }
1961
1962         tracking_evsel->tracking = true;
1963 }
1964
1965 struct perf_evsel *
1966 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1967                                const char *str)
1968 {
1969         struct perf_evsel *evsel;
1970
1971         evlist__for_each_entry(evlist, evsel) {
1972                 if (!evsel->name)
1973                         continue;
1974                 if (strcmp(str, evsel->name) == 0)
1975                         return evsel;
1976         }
1977
1978         return NULL;
1979 }
1980
1981 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1982                                   enum bkw_mmap_state state)
1983 {
1984         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1985         enum action {
1986                 NONE,
1987                 PAUSE,
1988                 RESUME,
1989         } action = NONE;
1990
1991         if (!evlist->backward_mmap)
1992                 return;
1993
1994         switch (old_state) {
1995         case BKW_MMAP_NOTREADY: {
1996                 if (state != BKW_MMAP_RUNNING)
1997                         goto state_err;;
1998                 break;
1999         }
2000         case BKW_MMAP_RUNNING: {
2001                 if (state != BKW_MMAP_DATA_PENDING)
2002                         goto state_err;
2003                 action = PAUSE;
2004                 break;
2005         }
2006         case BKW_MMAP_DATA_PENDING: {
2007                 if (state != BKW_MMAP_EMPTY)
2008                         goto state_err;
2009                 break;
2010         }
2011         case BKW_MMAP_EMPTY: {
2012                 if (state != BKW_MMAP_RUNNING)
2013                         goto state_err;
2014                 action = RESUME;
2015                 break;
2016         }
2017         default:
2018                 WARN_ONCE(1, "Shouldn't get there\n");
2019         }
2020
2021         evlist->bkw_mmap_state = state;
2022
2023         switch (action) {
2024         case PAUSE:
2025                 perf_evlist__pause(evlist);
2026                 break;
2027         case RESUME:
2028                 perf_evlist__resume(evlist);
2029                 break;
2030         case NONE:
2031         default:
2032                 break;
2033         }
2034
2035 state_err:
2036         return;
2037 }