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 <errno.h>
  12 #include <inttypes.h>
  13 #include <poll.h>
  14 #include "cpumap.h"
  15 #include "thread_map.h"
  16 #include "target.h"
  17 #include "evlist.h"
  18 #include "evsel.h"
  19 #include "debug.h"
  20 #include "asm/bug.h"
  21 #include <signal.h>
  22 #include <unistd.h>
  23
  24 #include "parse-events.h"
  25 #include <subcmd/parse-options.h>
  26
  27 #include <sys/mman.h>
  28
  29 #include <linux/bitops.h>
  30 #include <linux/hash.h>
  31 #include <linux/log2.h>
  32 #include <linux/err.h>
  33
  34 static void perf_mmap__munmap(struct perf_mmap *map);
  35 static void perf_mmap__put(struct perf_mmap *map);
  36
  37 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  38 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  39
  40 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  41                        struct thread_map *threads)
  42 {
  43         int i;
  44
  45         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  46                 INIT_HLIST_HEAD(&evlist->heads[i]);
  47         INIT_LIST_HEAD(&evlist->entries);
  48         perf_evlist__set_maps(evlist, cpus, threads);
  49         fdarray__init(&evlist->pollfd, 64);
  50         evlist->workload.pid = -1;
  51         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  52 }
  53
  54 struct perf_evlist *perf_evlist__new(void)
  55 {
  56         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  57
  58         if (evlist != NULL)
  59                 perf_evlist__init(evlist, NULL, NULL);
  60
  61         return evlist;
  62 }
  63
  64 struct perf_evlist *perf_evlist__new_default(void)
  65 {
  66         struct perf_evlist *evlist = perf_evlist__new();
  67
  68         if (evlist && perf_evlist__add_default(evlist)) {
  69                 perf_evlist__delete(evlist);
  70                 evlist = NULL;
  71         }
  72
  73         return evlist;
  74 }
  75
  76 struct perf_evlist *perf_evlist__new_dummy(void)
  77 {
  78         struct perf_evlist *evlist = perf_evlist__new();
  79
  80         if (evlist && perf_evlist__add_dummy(evlist)) {
  81                 perf_evlist__delete(evlist);
  82                 evlist = NULL;
  83         }
  84
  85         return evlist;
  86 }
  87
  88 /**
  89  * perf_evlist__set_id_pos - set the positions of event ids.
  90  * @evlist: selected event list
  91  *
  92  * Events with compatible sample types all have the same id_pos
  93  * and is_pos.  For convenience, put a copy on evlist.
  94  */
  95 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  96 {
  97         struct perf_evsel *first = perf_evlist__first(evlist);
  98
  99         evlist->id_pos = first->id_pos;
 100         evlist->is_pos = first->is_pos;
 101 }
 102
 103 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 104 {
 105         struct perf_evsel *evsel;
 106
 107         evlist__for_each_entry(evlist, evsel)
 108                 perf_evsel__calc_id_pos(evsel);
 109
 110         perf_evlist__set_id_pos(evlist);
 111 }
 112
 113 static void perf_evlist__purge(struct perf_evlist *evlist)
 114 {
 115         struct perf_evsel *pos, *n;
 116
 117         evlist__for_each_entry_safe(evlist, n, pos) {
 118                 list_del_init(&pos->node);
 119                 pos->evlist = NULL;
 120                 perf_evsel__delete(pos);
 121         }
 122
 123         evlist->nr_entries = 0;
 124 }
 125
 126 void perf_evlist__exit(struct perf_evlist *evlist)
 127 {
 128         zfree(&evlist->mmap);
 129         zfree(&evlist->backward_mmap);
 130         fdarray__exit(&evlist->pollfd);
 131 }
 132
 133 void perf_evlist__delete(struct perf_evlist *evlist)
 134 {
 135         if (evlist == NULL)
 136                 return;
 137
 138         perf_evlist__munmap(evlist);
 139         perf_evlist__close(evlist);
 140         cpu_map__put(evlist->cpus);
 141         thread_map__put(evlist->threads);
 142         evlist->cpus = NULL;
 143         evlist->threads = NULL;
 144         perf_evlist__purge(evlist);
 145         perf_evlist__exit(evlist);
 146         free(evlist);
 147 }
 148
 149 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 150                                           struct perf_evsel *evsel)
 151 {
 152         /*
 153          * We already have cpus for evsel (via PMU sysfs) so
 154          * keep it, if there's no target cpu list defined.
 155          */
 156         if (!evsel->own_cpus || evlist->has_user_cpus) {
 157                 cpu_map__put(evsel->cpus);
 158                 evsel->cpus = cpu_map__get(evlist->cpus);
 159         } else if (evsel->cpus != evsel->own_cpus) {
 160                 cpu_map__put(evsel->cpus);
 161                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 162         }
 163
 164         thread_map__put(evsel->threads);
 165         evsel->threads = thread_map__get(evlist->threads);
 166 }
 167
 168 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 169 {
 170         struct perf_evsel *evsel;
 171
 172         evlist__for_each_entry(evlist, evsel)
 173                 __perf_evlist__propagate_maps(evlist, evsel);
 174 }
 175
 176 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 177 {
 178         entry->evlist = evlist;
 179         list_add_tail(&entry->node, &evlist->entries);
 180         entry->idx = evlist->nr_entries;
 181         entry->tracking = !entry->idx;
 182
 183         if (!evlist->nr_entries++)
 184                 perf_evlist__set_id_pos(evlist);
 185
 186         __perf_evlist__propagate_maps(evlist, entry);
 187 }
 188
 189 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 190 {
 191         evsel->evlist = NULL;
 192         list_del_init(&evsel->node);
 193         evlist->nr_entries -= 1;
 194 }
 195
 196 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 197                                    struct list_head *list)
 198 {
 199         struct perf_evsel *evsel, *temp;
 200
 201         __evlist__for_each_entry_safe(list, temp, evsel) {
 202                 list_del_init(&evsel->node);
 203                 perf_evlist__add(evlist, evsel);
 204         }
 205 }
 206
 207 void __perf_evlist__set_leader(struct list_head *list)
 208 {
 209         struct perf_evsel *evsel, *leader;
 210
 211         leader = list_entry(list->next, struct perf_evsel, node);
 212         evsel = list_entry(list->prev, struct perf_evsel, node);
 213
 214         leader->nr_members = evsel->idx - leader->idx + 1;
 215
 216         __evlist__for_each_entry(list, evsel) {
 217                 evsel->leader = leader;
 218         }
 219 }
 220
 221 void perf_evlist__set_leader(struct perf_evlist *evlist)
 222 {
 223         if (evlist->nr_entries) {
 224                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 225                 __perf_evlist__set_leader(&evlist->entries);
 226         }
 227 }
 228
 229 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 230 {
 231         attr->precise_ip = 3;
 232
 233         while (attr->precise_ip != 0) {
 234                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 235                 if (fd != -1) {
 236                         close(fd);
 237                         break;
 238                 }
 239                 --attr->precise_ip;
 240         }
 241 }
 242
 243 int perf_evlist__add_default(struct perf_evlist *evlist)
 244 {
 245         struct perf_evsel *evsel = perf_evsel__new_cycles();
 246
 247         if (evsel == NULL)
 248                 return -ENOMEM;
 249
 250         perf_evlist__add(evlist, evsel);
 251         return 0;
 252 }
 253
 254 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 255 {
 256         struct perf_event_attr attr = {
 257                 .type   = PERF_TYPE_SOFTWARE,
 258                 .config = PERF_COUNT_SW_DUMMY,
 259                 .size   = sizeof(attr), /* to capture ABI version */
 260         };
 261         struct perf_evsel *evsel = perf_evsel__new(&attr);
 262
 263         if (evsel == NULL)
 264                 return -ENOMEM;
 265
 266         perf_evlist__add(evlist, evsel);
 267         return 0;
 268 }
 269
 270 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 271                                   struct perf_event_attr *attrs, size_t nr_attrs)
 272 {
 273         struct perf_evsel *evsel, *n;
 274         LIST_HEAD(head);
 275         size_t i;
 276
 277         for (i = 0; i < nr_attrs; i++) {
 278                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 279                 if (evsel == NULL)
 280                         goto out_delete_partial_list;
 281                 list_add_tail(&evsel->node, &head);
 282         }
 283
 284         perf_evlist__splice_list_tail(evlist, &head);
 285
 286         return 0;
 287
 288 out_delete_partial_list:
 289         __evlist__for_each_entry_safe(&head, n, evsel)
 290                 perf_evsel__delete(evsel);
 291         return -1;
 292 }
 293
 294 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 295                                      struct perf_event_attr *attrs, size_t nr_attrs)
 296 {
 297         size_t i;
 298
 299         for (i = 0; i < nr_attrs; i++)
 300                 event_attr_init(attrs + i);
 301
 302         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 303 }
 304
 305 struct perf_evsel *
 306 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 307 {
 308         struct perf_evsel *evsel;
 309
 310         evlist__for_each_entry(evlist, evsel) {
 311                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 312                     (int)evsel->attr.config == id)
 313                         return evsel;
 314         }
 315
 316         return NULL;
 317 }
 318
 319 struct perf_evsel *
 320 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 321                                      const char *name)
 322 {
 323         struct perf_evsel *evsel;
 324
 325         evlist__for_each_entry(evlist, evsel) {
 326                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 327                     (strcmp(evsel->name, name) == 0))
 328                         return evsel;
 329         }
 330
 331         return NULL;
 332 }
 333
 334 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 335                            const char *sys, const char *name, void *handler)
 336 {
 337         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 338
 339         if (IS_ERR(evsel))
 340                 return -1;
 341
 342         evsel->handler = handler;
 343         perf_evlist__add(evlist, evsel);
 344         return 0;
 345 }
 346
 347 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 348                                    struct perf_evsel *evsel)
 349 {
 350         if (evsel->system_wide)
 351                 return 1;
 352         else
 353                 return thread_map__nr(evlist->threads);
 354 }
 355
 356 void perf_evlist__disable(struct perf_evlist *evlist)
 357 {
 358         struct perf_evsel *pos;
 359
 360         evlist__for_each_entry(evlist, pos) {
 361                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 362                         continue;
 363                 perf_evsel__disable(pos);
 364         }
 365
 366         evlist->enabled = false;
 367 }
 368
 369 void perf_evlist__enable(struct perf_evlist *evlist)
 370 {
 371         struct perf_evsel *pos;
 372
 373         evlist__for_each_entry(evlist, pos) {
 374                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 375                         continue;
 376                 perf_evsel__enable(pos);
 377         }
 378
 379         evlist->enabled = true;
 380 }
 381
 382 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 383 {
 384         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 385 }
 386
 387 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 388                                          struct perf_evsel *evsel, int cpu)
 389 {
 390         int thread;
 391         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 392
 393         if (!evsel->fd)
 394                 return -EINVAL;
 395
 396         for (thread = 0; thread < nr_threads; thread++) {
 397                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 398                 if (err)
 399                         return err;
 400         }
 401         return 0;
 402 }
 403
 404 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 405                                             struct perf_evsel *evsel,
 406                                             int thread)
 407 {
 408         int cpu;
 409         int nr_cpus = cpu_map__nr(evlist->cpus);
 410
 411         if (!evsel->fd)
 412                 return -EINVAL;
 413
 414         for (cpu = 0; cpu < nr_cpus; cpu++) {
 415                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 416                 if (err)
 417                         return err;
 418         }
 419         return 0;
 420 }
 421
 422 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 423                                   struct perf_evsel *evsel, int idx)
 424 {
 425         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 426
 427         if (per_cpu_mmaps)
 428                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 429         else
 430                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 431 }
 432
 433 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 434 {
 435         int nr_cpus = cpu_map__nr(evlist->cpus);
 436         int nr_threads = thread_map__nr(evlist->threads);
 437         int nfds = 0;
 438         struct perf_evsel *evsel;
 439
 440         evlist__for_each_entry(evlist, evsel) {
 441                 if (evsel->system_wide)
 442                         nfds += nr_cpus;
 443                 else
 444                         nfds += nr_cpus * nr_threads;
 445         }
 446
 447         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 448             fdarray__grow(&evlist->pollfd, nfds) < 0)
 449                 return -ENOMEM;
 450
 451         return 0;
 452 }
 453
 454 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 455                                      struct perf_mmap *map, short revent)
 456 {
 457         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 458         /*
 459          * Save the idx so that when we filter out fds POLLHUP'ed we can
 460          * close the associated evlist->mmap[] entry.
 461          */
 462         if (pos >= 0) {
 463                 evlist->pollfd.priv[pos].ptr = map;
 464
 465                 fcntl(fd, F_SETFL, O_NONBLOCK);
 466         }
 467
 468         return pos;
 469 }
 470
 471 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 472 {
 473         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 474 }
 475
 476 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 477                                          void *arg __maybe_unused)
 478 {
 479         struct perf_mmap *map = fda->priv[fd].ptr;
 480
 481         if (map)
 482                 perf_mmap__put(map);
 483 }
 484
 485 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 486 {
 487         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 488                                perf_evlist__munmap_filtered, NULL);
 489 }
 490
 491 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 492 {
 493         return fdarray__poll(&evlist->pollfd, timeout);
 494 }
 495
 496 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 497                                  struct perf_evsel *evsel,
 498                                  int cpu, int thread, u64 id)
 499 {
 500         int hash;
 501         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 502
 503         sid->id = id;
 504         sid->evsel = evsel;
 505         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 506         hlist_add_head(&sid->node, &evlist->heads[hash]);
 507 }
 508
 509 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 510                          int cpu, int thread, u64 id)
 511 {
 512         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 513         evsel->id[evsel->ids++] = id;
 514 }
 515
 516 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 517                            struct perf_evsel *evsel,
 518                            int cpu, int thread, int fd)
 519 {
 520         u64 read_data[4] = { 0, };
 521         int id_idx = 1; /* The first entry is the counter value */
 522         u64 id;
 523         int ret;
 524
 525         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 526         if (!ret)
 527                 goto add;
 528
 529         if (errno != ENOTTY)
 530                 return -1;
 531
 532         /* Legacy way to get event id.. All hail to old kernels! */
 533
 534         /*
 535          * This way does not work with group format read, so bail
 536          * out in that case.
 537          */
 538         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 539                 return -1;
 540
 541         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 542             read(fd, &read_data, sizeof(read_data)) == -1)
 543                 return -1;
 544
 545         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 546                 ++id_idx;
 547         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 548                 ++id_idx;
 549
 550         id = read_data[id_idx];
 551
 552  add:
 553         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 554         return 0;
 555 }
 556
 557 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 558                                      struct perf_evsel *evsel, int idx, int cpu,
 559                                      int thread)
 560 {
 561         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 562         sid->idx = idx;
 563         if (evlist->cpus && cpu >= 0)
 564                 sid->cpu = evlist->cpus->map[cpu];
 565         else
 566                 sid->cpu = -1;
 567         if (!evsel->system_wide && evlist->threads && thread >= 0)
 568                 sid->tid = thread_map__pid(evlist->threads, thread);
 569         else
 570                 sid->tid = -1;
 571 }
 572
 573 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 574 {
 575         struct hlist_head *head;
 576         struct perf_sample_id *sid;
 577         int hash;
 578
 579         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 580         head = &evlist->heads[hash];
 581
 582         hlist_for_each_entry(sid, head, node)
 583                 if (sid->id == id)
 584                         return sid;
 585
 586         return NULL;
 587 }
 588
 589 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 590 {
 591         struct perf_sample_id *sid;
 592
 593         if (evlist->nr_entries == 1 || !id)
 594                 return perf_evlist__first(evlist);
 595
 596         sid = perf_evlist__id2sid(evlist, id);
 597         if (sid)
 598                 return sid->evsel;
 599
 600         if (!perf_evlist__sample_id_all(evlist))
 601                 return perf_evlist__first(evlist);
 602
 603         return NULL;
 604 }
 605
 606 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 607                                                 u64 id)
 608 {
 609         struct perf_sample_id *sid;
 610
 611         if (!id)
 612                 return NULL;
 613
 614         sid = perf_evlist__id2sid(evlist, id);
 615         if (sid)
 616                 return sid->evsel;
 617
 618         return NULL;
 619 }
 620
 621 static int perf_evlist__event2id(struct perf_evlist *evlist,
 622                                  union perf_event *event, u64 *id)
 623 {
 624         const u64 *array = event->sample.array;
 625         ssize_t n;
 626
 627         n = (event->header.size - sizeof(event->header)) >> 3;
 628
 629         if (event->header.type == PERF_RECORD_SAMPLE) {
 630                 if (evlist->id_pos >= n)
 631                         return -1;
 632                 *id = array[evlist->id_pos];
 633         } else {
 634                 if (evlist->is_pos > n)
 635                         return -1;
 636                 n -= evlist->is_pos;
 637                 *id = array[n];
 638         }
 639         return 0;
 640 }
 641
 642 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 643                                             union perf_event *event)
 644 {
 645         struct perf_evsel *first = perf_evlist__first(evlist);
 646         struct hlist_head *head;
 647         struct perf_sample_id *sid;
 648         int hash;
 649         u64 id;
 650
 651         if (evlist->nr_entries == 1)
 652                 return first;
 653
 654         if (!first->attr.sample_id_all &&
 655             event->header.type != PERF_RECORD_SAMPLE)
 656                 return first;
 657
 658         if (perf_evlist__event2id(evlist, event, &id))
 659                 return NULL;
 660
 661         /* Synthesized events have an id of zero */
 662         if (!id)
 663                 return first;
 664
 665         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 666         head = &evlist->heads[hash];
 667
 668         hlist_for_each_entry(sid, head, node) {
 669                 if (sid->id == id)
 670                         return sid->evsel;
 671         }
 672         return NULL;
 673 }
 674
 675 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 676 {
 677         int i;
 678
 679         if (!evlist->backward_mmap)
 680                 return 0;
 681
 682         for (i = 0; i < evlist->nr_mmaps; i++) {
 683                 int fd = evlist->backward_mmap[i].fd;
 684                 int err;
 685
 686                 if (fd < 0)
 687                         continue;
 688                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 689                 if (err)
 690                         return err;
 691         }
 692         return 0;
 693 }
 694
 695 static int perf_evlist__pause(struct perf_evlist *evlist)
 696 {
 697         return perf_evlist__set_paused(evlist, true);
 698 }
 699
 700 static int perf_evlist__resume(struct perf_evlist *evlist)
 701 {
 702         return perf_evlist__set_paused(evlist, false);
 703 }
 704
 705 /* When check_messup is true, 'end' must points to a good entry */
 706 static union perf_event *
 707 perf_mmap__read(struct perf_mmap *md, bool check_messup, u64 start,
 708                 u64 end, u64 *prev)
 709 {
 710         unsigned char *data = md->base + page_size;
 711         union perf_event *event = NULL;
 712         int diff = end - start;
 713
 714         if (check_messup) {
 715                 /*
 716                  * If we're further behind than half the buffer, there's a chance
 717                  * the writer will bite our tail and mess up the samples under us.
 718                  *
 719                  * If we somehow ended up ahead of the 'end', we got messed up.
 720                  *
 721                  * In either case, truncate and restart at 'end'.
 722                  */
 723                 if (diff > md->mask / 2 || diff < 0) {
 724                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 725
 726                         /*
 727                          * 'end' points to a known good entry, start there.
 728                          */
 729                         start = end;
 730                         diff = 0;
 731                 }
 732         }
 733
 734         if (diff >= (int)sizeof(event->header)) {
 735                 size_t size;
 736
 737                 event = (union perf_event *)&data[start & md->mask];
 738                 size = event->header.size;
 739
 740                 if (size < sizeof(event->header) || diff < (int)size) {
 741                         event = NULL;
 742                         goto broken_event;
 743                 }
 744
 745                 /*
 746                  * Event straddles the mmap boundary -- header should always
 747                  * be inside due to u64 alignment of output.
 748                  */
 749                 if ((start & md->mask) + size != ((start + size) & md->mask)) {
 750                         unsigned int offset = start;
 751                         unsigned int len = min(sizeof(*event), size), cpy;
 752                         void *dst = md->event_copy;
 753
 754                         do {
 755                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 756                                 memcpy(dst, &data[offset & md->mask], cpy);
 757                                 offset += cpy;
 758                                 dst += cpy;
 759                                 len -= cpy;
 760                         } while (len);
 761
 762                         event = (union perf_event *) md->event_copy;
 763                 }
 764
 765                 start += size;
 766         }
 767
 768 broken_event:
 769         if (prev)
 770                 *prev = start;
 771
 772         return event;
 773 }
 774
 775 union perf_event *perf_mmap__read_forward(struct perf_mmap *md, bool check_messup)
 776 {
 777         u64 head;
 778         u64 old = md->prev;
 779
 780         /*
 781          * Check if event was unmapped due to a POLLHUP/POLLERR.
 782          */
 783         if (!refcount_read(&md->refcnt))
 784                 return NULL;
 785
 786         head = perf_mmap__read_head(md);
 787
 788         return perf_mmap__read(md, check_messup, old, head, &md->prev);
 789 }
 790
 791 union perf_event *
 792 perf_mmap__read_backward(struct perf_mmap *md)
 793 {
 794         u64 head, end;
 795         u64 start = md->prev;
 796
 797         /*
 798          * Check if event was unmapped due to a POLLHUP/POLLERR.
 799          */
 800         if (!refcount_read(&md->refcnt))
 801                 return NULL;
 802
 803         head = perf_mmap__read_head(md);
 804         if (!head)
 805                 return NULL;
 806
 807         /*
 808          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 809          * it each time when kernel writes to it, so in fact 'head' is
 810          * negative. 'end' pointer is made manually by adding the size of
 811          * the ring buffer to 'head' pointer, means the validate data can
 812          * read is the whole ring buffer. If 'end' is positive, the ring
 813          * buffer has not fully filled, so we must adjust 'end' to 0.
 814          *
 815          * However, since both 'head' and 'end' is unsigned, we can't
 816          * simply compare 'end' against 0. Here we compare '-head' and
 817          * the size of the ring buffer, where -head is the number of bytes
 818          * kernel write to the ring buffer.
 819          */
 820         if (-head < (u64)(md->mask + 1))
 821                 end = 0;
 822         else
 823                 end = head + md->mask + 1;
 824
 825         return perf_mmap__read(md, false, start, end, &md->prev);
 826 }
 827
 828 union perf_event *perf_evlist__mmap_read_forward(struct perf_evlist *evlist, int idx)
 829 {
 830         struct perf_mmap *md = &evlist->mmap[idx];
 831
 832         /*
 833          * Check messup is required for forward overwritable ring buffer:
 834          * memory pointed by md->prev can be overwritten in this case.
 835          * No need for read-write ring buffer: kernel stop outputting when
 836          * it hit md->prev (perf_mmap__consume()).
 837          */
 838         return perf_mmap__read_forward(md, evlist->overwrite);
 839 }
 840
 841 union perf_event *perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
 842 {
 843         struct perf_mmap *md = &evlist->mmap[idx];
 844
 845         /*
 846          * No need to check messup for backward ring buffer:
 847          * We can always read arbitrary long data from a backward
 848          * ring buffer unless we forget to pause it before reading.
 849          */
 850         return perf_mmap__read_backward(md);
 851 }
 852
 853 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 854 {
 855         return perf_evlist__mmap_read_forward(evlist, idx);
 856 }
 857
 858 void perf_mmap__read_catchup(struct perf_mmap *md)
 859 {
 860         u64 head;
 861
 862         if (!refcount_read(&md->refcnt))
 863                 return;
 864
 865         head = perf_mmap__read_head(md);
 866         md->prev = head;
 867 }
 868
 869 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 870 {
 871         perf_mmap__read_catchup(&evlist->mmap[idx]);
 872 }
 873
 874 static bool perf_mmap__empty(struct perf_mmap *md)
 875 {
 876         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 877 }
 878
 879 static void perf_mmap__get(struct perf_mmap *map)
 880 {
 881         refcount_inc(&map->refcnt);
 882 }
 883
 884 static void perf_mmap__put(struct perf_mmap *md)
 885 {
 886         BUG_ON(md->base && refcount_read(&md->refcnt) == 0);
 887
 888         if (refcount_dec_and_test(&md->refcnt))
 889                 perf_mmap__munmap(md);
 890 }
 891
 892 void perf_mmap__consume(struct perf_mmap *md, bool overwrite)
 893 {
 894         if (!overwrite) {
 895                 u64 old = md->prev;
 896
 897                 perf_mmap__write_tail(md, old);
 898         }
 899
 900         if (refcount_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 901                 perf_mmap__put(md);
 902 }
 903
 904 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 905 {
 906         perf_mmap__consume(&evlist->mmap[idx], evlist->overwrite);
 907 }
 908
 909 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 910                                struct auxtrace_mmap_params *mp __maybe_unused,
 911                                void *userpg __maybe_unused,
 912                                int fd __maybe_unused)
 913 {
 914         return 0;
 915 }
 916
 917 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 918 {
 919 }
 920
 921 void __weak auxtrace_mmap_params__init(
 922                         struct auxtrace_mmap_params *mp __maybe_unused,
 923                         off_t auxtrace_offset __maybe_unused,
 924                         unsigned int auxtrace_pages __maybe_unused,
 925                         bool auxtrace_overwrite __maybe_unused)
 926 {
 927 }
 928
 929 void __weak auxtrace_mmap_params__set_idx(
 930                         struct auxtrace_mmap_params *mp __maybe_unused,
 931                         struct perf_evlist *evlist __maybe_unused,
 932                         int idx __maybe_unused,
 933                         bool per_cpu __maybe_unused)
 934 {
 935 }
 936
 937 static void perf_mmap__munmap(struct perf_mmap *map)
 938 {
 939         if (map->base != NULL) {
 940                 munmap(map->base, perf_mmap__mmap_len(map));
 941                 map->base = NULL;
 942                 map->fd = -1;
 943                 refcount_set(&map->refcnt, 0);
 944         }
 945         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 946 }
 947
 948 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 949 {
 950         int i;
 951
 952         if (evlist->mmap)
 953                 for (i = 0; i < evlist->nr_mmaps; i++)
 954                         perf_mmap__munmap(&evlist->mmap[i]);
 955
 956         if (evlist->backward_mmap)
 957                 for (i = 0; i < evlist->nr_mmaps; i++)
 958                         perf_mmap__munmap(&evlist->backward_mmap[i]);
 959 }
 960
 961 void perf_evlist__munmap(struct perf_evlist *evlist)
 962 {
 963         perf_evlist__munmap_nofree(evlist);
 964         zfree(&evlist->mmap);
 965         zfree(&evlist->backward_mmap);
 966 }
 967
 968 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 969 {
 970         int i;
 971         struct perf_mmap *map;
 972
 973         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 974         if (cpu_map__empty(evlist->cpus))
 975                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 976         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 977         if (!map)
 978                 return NULL;
 979
 980         for (i = 0; i < evlist->nr_mmaps; i++) {
 981                 map[i].fd = -1;
 982                 /*
 983                  * When the perf_mmap() call is made we grab one refcount, plus
 984                  * one extra to let perf_evlist__mmap_consume() get the last
 985                  * events after all real references (perf_mmap__get()) are
 986                  * dropped.
 987                  *
 988                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 989                  * thus does perf_mmap__get() on it.
 990                  */
 991                 refcount_set(&map[i].refcnt, 0);
 992         }
 993         return map;
 994 }
 995
 996 struct mmap_params {
 997         int prot;
 998         int mask;
 999         struct auxtrace_mmap_params auxtrace_mp;
1000 };
1001
1002 static int perf_mmap__mmap(struct perf_mmap *map,
1003                            struct mmap_params *mp, int fd)
1004 {
1005         /*
1006          * The last one will be done at perf_evlist__mmap_consume(), so that we
1007          * make sure we don't prevent tools from consuming every last event in
1008          * the ring buffer.
1009          *
1010          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
1011          * anymore, but the last events for it are still in the ring buffer,
1012          * waiting to be consumed.
1013          *
1014          * Tools can chose to ignore this at their own discretion, but the
1015          * evlist layer can't just drop it when filtering events in
1016          * perf_evlist__filter_pollfd().
1017          */
1018         refcount_set(&map->refcnt, 2);
1019         map->prev = 0;
1020         map->mask = mp->mask;
1021         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
1022                          MAP_SHARED, fd, 0);
1023         if (map->base == MAP_FAILED) {
1024                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1025                           errno);
1026                 map->base = NULL;
1027                 return -1;
1028         }
1029         map->fd = fd;
1030
1031         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
1032                                 &mp->auxtrace_mp, map->base, fd))
1033                 return -1;
1034
1035         return 0;
1036 }
1037
1038 static bool
1039 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
1040                          struct perf_evsel *evsel)
1041 {
1042         if (evsel->attr.write_backward)
1043                 return false;
1044         return true;
1045 }
1046
1047 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
1048                                        struct mmap_params *mp, int cpu_idx,
1049                                        int thread, int *_output, int *_output_backward)
1050 {
1051         struct perf_evsel *evsel;
1052         int revent;
1053         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
1054
1055         evlist__for_each_entry(evlist, evsel) {
1056                 struct perf_mmap *maps = evlist->mmap;
1057                 int *output = _output;
1058                 int fd;
1059                 int cpu;
1060
1061                 if (evsel->attr.write_backward) {
1062                         output = _output_backward;
1063                         maps = evlist->backward_mmap;
1064
1065                         if (!maps) {
1066                                 maps = perf_evlist__alloc_mmap(evlist);
1067                                 if (!maps)
1068                                         return -1;
1069                                 evlist->backward_mmap = maps;
1070                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1071                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1072                         }
1073                 }
1074
1075                 if (evsel->system_wide && thread)
1076                         continue;
1077
1078                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
1079                 if (cpu == -1)
1080                         continue;
1081
1082                 fd = FD(evsel, cpu, thread);
1083
1084                 if (*output == -1) {
1085                         *output = fd;
1086
1087                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1088                                 return -1;
1089                 } else {
1090                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1091                                 return -1;
1092
1093                         perf_mmap__get(&maps[idx]);
1094                 }
1095
1096                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1097
1098                 /*
1099                  * The system_wide flag causes a selected event to be opened
1100                  * always without a pid.  Consequently it will never get a
1101                  * POLLHUP, but it is used for tracking in combination with
1102                  * other events, so it should not need to be polled anyway.
1103                  * Therefore don't add it for polling.
1104                  */
1105                 if (!evsel->system_wide &&
1106                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1107                         perf_mmap__put(&maps[idx]);
1108                         return -1;
1109                 }
1110
1111                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1112                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1113                                                    fd) < 0)
1114                                 return -1;
1115                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1116                                                  thread);
1117                 }
1118         }
1119
1120         return 0;
1121 }
1122
1123 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1124                                      struct mmap_params *mp)
1125 {
1126         int cpu, thread;
1127         int nr_cpus = cpu_map__nr(evlist->cpus);
1128         int nr_threads = thread_map__nr(evlist->threads);
1129
1130         pr_debug2("perf event ring buffer mmapped per cpu\n");
1131         for (cpu = 0; cpu < nr_cpus; cpu++) {
1132                 int output = -1;
1133                 int output_backward = -1;
1134
1135                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1136                                               true);
1137
1138                 for (thread = 0; thread < nr_threads; thread++) {
1139                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1140                                                         thread, &output, &output_backward))
1141                                 goto out_unmap;
1142                 }
1143         }
1144
1145         return 0;
1146
1147 out_unmap:
1148         perf_evlist__munmap_nofree(evlist);
1149         return -1;
1150 }
1151
1152 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1153                                         struct mmap_params *mp)
1154 {
1155         int thread;
1156         int nr_threads = thread_map__nr(evlist->threads);
1157
1158         pr_debug2("perf event ring buffer mmapped per thread\n");
1159         for (thread = 0; thread < nr_threads; thread++) {
1160                 int output = -1;
1161                 int output_backward = -1;
1162
1163                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1164                                               false);
1165
1166                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1167                                                 &output, &output_backward))
1168                         goto out_unmap;
1169         }
1170
1171         return 0;
1172
1173 out_unmap:
1174         perf_evlist__munmap_nofree(evlist);
1175         return -1;
1176 }
1177
1178 unsigned long perf_event_mlock_kb_in_pages(void)
1179 {
1180         unsigned long pages;
1181         int max;
1182
1183         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1184                 /*
1185                  * Pick a once upon a time good value, i.e. things look
1186                  * strange since we can't read a sysctl value, but lets not
1187                  * die yet...
1188                  */
1189                 max = 512;
1190         } else {
1191                 max -= (page_size / 1024);
1192         }
1193
1194         pages = (max * 1024) / page_size;
1195         if (!is_power_of_2(pages))
1196                 pages = rounddown_pow_of_two(pages);
1197
1198         return pages;
1199 }
1200
1201 size_t perf_evlist__mmap_size(unsigned long pages)
1202 {
1203         if (pages == UINT_MAX)
1204                 pages = perf_event_mlock_kb_in_pages();
1205         else if (!is_power_of_2(pages))
1206                 return 0;
1207
1208         return (pages + 1) * page_size;
1209 }
1210
1211 static long parse_pages_arg(const char *str, unsigned long min,
1212                             unsigned long max)
1213 {
1214         unsigned long pages, val;
1215         static struct parse_tag tags[] = {
1216                 { .tag  = 'B', .mult = 1       },
1217                 { .tag  = 'K', .mult = 1 << 10 },
1218                 { .tag  = 'M', .mult = 1 << 20 },
1219                 { .tag  = 'G', .mult = 1 << 30 },
1220                 { .tag  = 0 },
1221         };
1222
1223         if (str == NULL)
1224                 return -EINVAL;
1225
1226         val = parse_tag_value(str, tags);
1227         if (val != (unsigned long) -1) {
1228                 /* we got file size value */
1229                 pages = PERF_ALIGN(val, page_size) / page_size;
1230         } else {
1231                 /* we got pages count value */
1232                 char *eptr;
1233                 pages = strtoul(str, &eptr, 10);
1234                 if (*eptr != '\0')
1235                         return -EINVAL;
1236         }
1237
1238         if (pages == 0 && min == 0) {
1239                 /* leave number of pages at 0 */
1240         } else if (!is_power_of_2(pages)) {
1241                 char buf[100];
1242
1243                 /* round pages up to next power of 2 */
1244                 pages = roundup_pow_of_two(pages);
1245                 if (!pages)
1246                         return -EINVAL;
1247
1248                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1249                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1250                         buf, pages);
1251         }
1252
1253         if (pages > max)
1254                 return -EINVAL;
1255
1256         return pages;
1257 }
1258
1259 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1260 {
1261         unsigned long max = UINT_MAX;
1262         long pages;
1263
1264         if (max > SIZE_MAX / page_size)
1265                 max = SIZE_MAX / page_size;
1266
1267         pages = parse_pages_arg(str, 1, max);
1268         if (pages < 0) {
1269                 pr_err("Invalid argument for --mmap_pages/-m\n");
1270                 return -1;
1271         }
1272
1273         *mmap_pages = pages;
1274         return 0;
1275 }
1276
1277 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1278                                   int unset __maybe_unused)
1279 {
1280         return __perf_evlist__parse_mmap_pages(opt->value, str);
1281 }
1282
1283 /**
1284  * perf_evlist__mmap_ex - Create mmaps to receive events.
1285  * @evlist: list of events
1286  * @pages: map length in pages
1287  * @overwrite: overwrite older events?
1288  * @auxtrace_pages - auxtrace map length in pages
1289  * @auxtrace_overwrite - overwrite older auxtrace data?
1290  *
1291  * If @overwrite is %false the user needs to signal event consumption using
1292  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1293  * automatically.
1294  *
1295  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1296  * consumption using auxtrace_mmap__write_tail().
1297  *
1298  * Return: %0 on success, negative error code otherwise.
1299  */
1300 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1301                          bool overwrite, unsigned int auxtrace_pages,
1302                          bool auxtrace_overwrite)
1303 {
1304         struct perf_evsel *evsel;
1305         const struct cpu_map *cpus = evlist->cpus;
1306         const struct thread_map *threads = evlist->threads;
1307         struct mmap_params mp = {
1308                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1309         };
1310
1311         if (!evlist->mmap)
1312                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1313         if (!evlist->mmap)
1314                 return -ENOMEM;
1315
1316         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1317                 return -ENOMEM;
1318
1319         evlist->overwrite = overwrite;
1320         evlist->mmap_len = perf_evlist__mmap_size(pages);
1321         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1322         mp.mask = evlist->mmap_len - page_size - 1;
1323
1324         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1325                                    auxtrace_pages, auxtrace_overwrite);
1326
1327         evlist__for_each_entry(evlist, evsel) {
1328                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1329                     evsel->sample_id == NULL &&
1330                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1331                         return -ENOMEM;
1332         }
1333
1334         if (cpu_map__empty(cpus))
1335                 return perf_evlist__mmap_per_thread(evlist, &mp);
1336
1337         return perf_evlist__mmap_per_cpu(evlist, &mp);
1338 }
1339
1340 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1341                       bool overwrite)
1342 {
1343         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1344 }
1345
1346 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1347 {
1348         struct cpu_map *cpus;
1349         struct thread_map *threads;
1350
1351         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1352
1353         if (!threads)
1354                 return -1;
1355
1356         if (target__uses_dummy_map(target))
1357                 cpus = cpu_map__dummy_new();
1358         else
1359                 cpus = cpu_map__new(target->cpu_list);
1360
1361         if (!cpus)
1362                 goto out_delete_threads;
1363
1364         evlist->has_user_cpus = !!target->cpu_list;
1365
1366         perf_evlist__set_maps(evlist, cpus, threads);
1367
1368         return 0;
1369
1370 out_delete_threads:
1371         thread_map__put(threads);
1372         return -1;
1373 }
1374
1375 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1376                            struct thread_map *threads)
1377 {
1378         /*
1379          * Allow for the possibility that one or another of the maps isn't being
1380          * changed i.e. don't put it.  Note we are assuming the maps that are
1381          * being applied are brand new and evlist is taking ownership of the
1382          * original reference count of 1.  If that is not the case it is up to
1383          * the caller to increase the reference count.
1384          */
1385         if (cpus != evlist->cpus) {
1386                 cpu_map__put(evlist->cpus);
1387                 evlist->cpus = cpu_map__get(cpus);
1388         }
1389
1390         if (threads != evlist->threads) {
1391                 thread_map__put(evlist->threads);
1392                 evlist->threads = thread_map__get(threads);
1393         }
1394
1395         perf_evlist__propagate_maps(evlist);
1396 }
1397
1398 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1399                                    enum perf_event_sample_format bit)
1400 {
1401         struct perf_evsel *evsel;
1402
1403         evlist__for_each_entry(evlist, evsel)
1404                 __perf_evsel__set_sample_bit(evsel, bit);
1405 }
1406
1407 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1408                                      enum perf_event_sample_format bit)
1409 {
1410         struct perf_evsel *evsel;
1411
1412         evlist__for_each_entry(evlist, evsel)
1413                 __perf_evsel__reset_sample_bit(evsel, bit);
1414 }
1415
1416 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1417 {
1418         struct perf_evsel *evsel;
1419         int err = 0;
1420         const int ncpus = cpu_map__nr(evlist->cpus),
1421                   nthreads = thread_map__nr(evlist->threads);
1422
1423         evlist__for_each_entry(evlist, evsel) {
1424                 if (evsel->filter == NULL)
1425                         continue;
1426
1427                 /*
1428                  * filters only work for tracepoint event, which doesn't have cpu limit.
1429                  * So evlist and evsel should always be same.
1430                  */
1431                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1432                 if (err) {
1433                         *err_evsel = evsel;
1434                         break;
1435                 }
1436         }
1437
1438         return err;
1439 }
1440
1441 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1442 {
1443         struct perf_evsel *evsel;
1444         int err = 0;
1445
1446         evlist__for_each_entry(evlist, evsel) {
1447                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1448                         continue;
1449
1450                 err = perf_evsel__set_filter(evsel, filter);
1451                 if (err)
1452                         break;
1453         }
1454
1455         return err;
1456 }
1457
1458 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1459 {
1460         char *filter;
1461         int ret = -1;
1462         size_t i;
1463
1464         for (i = 0; i < npids; ++i) {
1465                 if (i == 0) {
1466                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1467                                 return -1;
1468                 } else {
1469                         char *tmp;
1470
1471                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1472                                 goto out_free;
1473
1474                         free(filter);
1475                         filter = tmp;
1476                 }
1477         }
1478
1479         ret = perf_evlist__set_filter(evlist, filter);
1480 out_free:
1481         free(filter);
1482         return ret;
1483 }
1484
1485 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1486 {
1487         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1488 }
1489
1490 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1491 {
1492         struct perf_evsel *pos;
1493
1494         if (evlist->nr_entries == 1)
1495                 return true;
1496
1497         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1498                 return false;
1499
1500         evlist__for_each_entry(evlist, pos) {
1501                 if (pos->id_pos != evlist->id_pos ||
1502                     pos->is_pos != evlist->is_pos)
1503                         return false;
1504         }
1505
1506         return true;
1507 }
1508
1509 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1510 {
1511         struct perf_evsel *evsel;
1512
1513         if (evlist->combined_sample_type)
1514                 return evlist->combined_sample_type;
1515
1516         evlist__for_each_entry(evlist, evsel)
1517                 evlist->combined_sample_type |= evsel->attr.sample_type;
1518
1519         return evlist->combined_sample_type;
1520 }
1521
1522 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1523 {
1524         evlist->combined_sample_type = 0;
1525         return __perf_evlist__combined_sample_type(evlist);
1526 }
1527
1528 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1529 {
1530         struct perf_evsel *evsel;
1531         u64 branch_type = 0;
1532
1533         evlist__for_each_entry(evlist, evsel)
1534                 branch_type |= evsel->attr.branch_sample_type;
1535         return branch_type;
1536 }
1537
1538 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1539 {
1540         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1541         u64 read_format = first->attr.read_format;
1542         u64 sample_type = first->attr.sample_type;
1543
1544         evlist__for_each_entry(evlist, pos) {
1545                 if (read_format != pos->attr.read_format)
1546                         return false;
1547         }
1548
1549         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1550         if ((sample_type & PERF_SAMPLE_READ) &&
1551             !(read_format & PERF_FORMAT_ID)) {
1552                 return false;
1553         }
1554
1555         return true;
1556 }
1557
1558 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1559 {
1560         struct perf_evsel *first = perf_evlist__first(evlist);
1561         return first->attr.read_format;
1562 }
1563
1564 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1565 {
1566         struct perf_evsel *first = perf_evlist__first(evlist);
1567         struct perf_sample *data;
1568         u64 sample_type;
1569         u16 size = 0;
1570
1571         if (!first->attr.sample_id_all)
1572                 goto out;
1573
1574         sample_type = first->attr.sample_type;
1575
1576         if (sample_type & PERF_SAMPLE_TID)
1577                 size += sizeof(data->tid) * 2;
1578
1579        if (sample_type & PERF_SAMPLE_TIME)
1580                 size += sizeof(data->time);
1581
1582         if (sample_type & PERF_SAMPLE_ID)
1583                 size += sizeof(data->id);
1584
1585         if (sample_type & PERF_SAMPLE_STREAM_ID)
1586                 size += sizeof(data->stream_id);
1587
1588         if (sample_type & PERF_SAMPLE_CPU)
1589                 size += sizeof(data->cpu) * 2;
1590
1591         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1592                 size += sizeof(data->id);
1593 out:
1594         return size;
1595 }
1596
1597 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1598 {
1599         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1600
1601         evlist__for_each_entry_continue(evlist, pos) {
1602                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1603                         return false;
1604         }
1605
1606         return true;
1607 }
1608
1609 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1610 {
1611         struct perf_evsel *first = perf_evlist__first(evlist);
1612         return first->attr.sample_id_all;
1613 }
1614
1615 void perf_evlist__set_selected(struct perf_evlist *evlist,
1616                                struct perf_evsel *evsel)
1617 {
1618         evlist->selected = evsel;
1619 }
1620
1621 void perf_evlist__close(struct perf_evlist *evlist)
1622 {
1623         struct perf_evsel *evsel;
1624         int ncpus = cpu_map__nr(evlist->cpus);
1625         int nthreads = thread_map__nr(evlist->threads);
1626
1627         evlist__for_each_entry_reverse(evlist, evsel) {
1628                 int n = evsel->cpus ? evsel->cpus->nr : ncpus;
1629                 perf_evsel__close(evsel, n, nthreads);
1630         }
1631 }
1632
1633 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1634 {
1635         struct cpu_map    *cpus;
1636         struct thread_map *threads;
1637         int err = -ENOMEM;
1638
1639         /*
1640          * Try reading /sys/devices/system/cpu/online to get
1641          * an all cpus map.
1642          *
1643          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1644          * code needs an overhaul to properly forward the
1645          * error, and we may not want to do that fallback to a
1646          * default cpu identity map :-\
1647          */
1648         cpus = cpu_map__new(NULL);
1649         if (!cpus)
1650                 goto out;
1651
1652         threads = thread_map__new_dummy();
1653         if (!threads)
1654                 goto out_put;
1655
1656         perf_evlist__set_maps(evlist, cpus, threads);
1657 out:
1658         return err;
1659 out_put:
1660         cpu_map__put(cpus);
1661         goto out;
1662 }
1663
1664 int perf_evlist__open(struct perf_evlist *evlist)
1665 {
1666         struct perf_evsel *evsel;
1667         int err;
1668
1669         /*
1670          * Default: one fd per CPU, all threads, aka systemwide
1671          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1672          */
1673         if (evlist->threads == NULL && evlist->cpus == NULL) {
1674                 err = perf_evlist__create_syswide_maps(evlist);
1675                 if (err < 0)
1676                         goto out_err;
1677         }
1678
1679         perf_evlist__update_id_pos(evlist);
1680
1681         evlist__for_each_entry(evlist, evsel) {
1682                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1683                 if (err < 0)
1684                         goto out_err;
1685         }
1686
1687         return 0;
1688 out_err:
1689         perf_evlist__close(evlist);
1690         errno = -err;
1691         return err;
1692 }
1693
1694 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1695                                   const char *argv[], bool pipe_output,
1696                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1697 {
1698         int child_ready_pipe[2], go_pipe[2];
1699         char bf;
1700
1701         if (pipe(child_ready_pipe) < 0) {
1702                 perror("failed to create 'ready' pipe");
1703                 return -1;
1704         }
1705
1706         if (pipe(go_pipe) < 0) {
1707                 perror("failed to create 'go' pipe");
1708                 goto out_close_ready_pipe;
1709         }
1710
1711         evlist->workload.pid = fork();
1712         if (evlist->workload.pid < 0) {
1713                 perror("failed to fork");
1714                 goto out_close_pipes;
1715         }
1716
1717         if (!evlist->workload.pid) {
1718                 int ret;
1719
1720                 if (pipe_output)
1721                         dup2(2, 1);
1722
1723                 signal(SIGTERM, SIG_DFL);
1724
1725                 close(child_ready_pipe[0]);
1726                 close(go_pipe[1]);
1727                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1728
1729                 /*
1730                  * Tell the parent we're ready to go
1731                  */
1732                 close(child_ready_pipe[1]);
1733
1734                 /*
1735                  * Wait until the parent tells us to go.
1736                  */
1737                 ret = read(go_pipe[0], &bf, 1);
1738                 /*
1739                  * The parent will ask for the execvp() to be performed by
1740                  * writing exactly one byte, in workload.cork_fd, usually via
1741                  * perf_evlist__start_workload().
1742                  *
1743                  * For cancelling the workload without actually running it,
1744                  * the parent will just close workload.cork_fd, without writing
1745                  * anything, i.e. read will return zero and we just exit()
1746                  * here.
1747                  */
1748                 if (ret != 1) {
1749                         if (ret == -1)
1750                                 perror("unable to read pipe");
1751                         exit(ret);
1752                 }
1753
1754                 execvp(argv[0], (char **)argv);
1755
1756                 if (exec_error) {
1757                         union sigval val;
1758
1759                         val.sival_int = errno;
1760                         if (sigqueue(getppid(), SIGUSR1, val))
1761                                 perror(argv[0]);
1762                 } else
1763                         perror(argv[0]);
1764                 exit(-1);
1765         }
1766
1767         if (exec_error) {
1768                 struct sigaction act = {
1769                         .sa_flags     = SA_SIGINFO,
1770                         .sa_sigaction = exec_error,
1771                 };
1772                 sigaction(SIGUSR1, &act, NULL);
1773         }
1774
1775         if (target__none(target)) {
1776                 if (evlist->threads == NULL) {
1777                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1778                                 __func__, __LINE__);
1779                         goto out_close_pipes;
1780                 }
1781                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1782         }
1783
1784         close(child_ready_pipe[1]);
1785         close(go_pipe[0]);
1786         /*
1787          * wait for child to settle
1788          */
1789         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1790                 perror("unable to read pipe");
1791                 goto out_close_pipes;
1792         }
1793
1794         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1795         evlist->workload.cork_fd = go_pipe[1];
1796         close(child_ready_pipe[0]);
1797         return 0;
1798
1799 out_close_pipes:
1800         close(go_pipe[0]);
1801         close(go_pipe[1]);
1802 out_close_ready_pipe:
1803         close(child_ready_pipe[0]);
1804         close(child_ready_pipe[1]);
1805         return -1;
1806 }
1807
1808 int perf_evlist__start_workload(struct perf_evlist *evlist)
1809 {
1810         if (evlist->workload.cork_fd > 0) {
1811                 char bf = 0;
1812                 int ret;
1813                 /*
1814                  * Remove the cork, let it rip!
1815                  */
1816                 ret = write(evlist->workload.cork_fd, &bf, 1);
1817                 if (ret < 0)
1818                         perror("unable to write to pipe");
1819
1820                 close(evlist->workload.cork_fd);
1821                 return ret;
1822         }
1823
1824         return 0;
1825 }
1826
1827 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1828                               struct perf_sample *sample)
1829 {
1830         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1831
1832         if (!evsel)
1833                 return -EFAULT;
1834         return perf_evsel__parse_sample(evsel, event, sample);
1835 }
1836
1837 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1838 {
1839         struct perf_evsel *evsel;
1840         size_t printed = 0;
1841
1842         evlist__for_each_entry(evlist, evsel) {
1843                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1844                                    perf_evsel__name(evsel));
1845         }
1846
1847         return printed + fprintf(fp, "\n");
1848 }
1849
1850 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1851                                int err, char *buf, size_t size)
1852 {
1853         int printed, value;
1854         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1855
1856         switch (err) {
1857         case EACCES:
1858         case EPERM:
1859                 printed = scnprintf(buf, size,
1860                                     "Error:\t%s.\n"
1861                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1862
1863                 value = perf_event_paranoid();
1864
1865                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1866
1867                 if (value >= 2) {
1868                         printed += scnprintf(buf + printed, size - printed,
1869                                              "For your workloads it needs to be <= 1\nHint:\t");
1870                 }
1871                 printed += scnprintf(buf + printed, size - printed,
1872                                      "For system wide tracing it needs to be set to -1.\n");
1873
1874                 printed += scnprintf(buf + printed, size - printed,
1875                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1876                                     "Hint:\tThe current value is %d.", value);
1877                 break;
1878         case EINVAL: {
1879                 struct perf_evsel *first = perf_evlist__first(evlist);
1880                 int max_freq;
1881
1882                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1883                         goto out_default;
1884
1885                 if (first->attr.sample_freq < (u64)max_freq)
1886                         goto out_default;
1887
1888                 printed = scnprintf(buf, size,
1889                                     "Error:\t%s.\n"
1890                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1891                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1892                                     emsg, max_freq, first->attr.sample_freq);
1893                 break;
1894         }
1895         default:
1896 out_default:
1897                 scnprintf(buf, size, "%s", emsg);
1898                 break;
1899         }
1900
1901         return 0;
1902 }
1903
1904 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1905 {
1906         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1907         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1908
1909         switch (err) {
1910         case EPERM:
1911                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1912                 printed += scnprintf(buf + printed, size - printed,
1913                                      "Error:\t%s.\n"
1914                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1915                                      "Hint:\tTried using %zd kB.\n",
1916                                      emsg, pages_max_per_user, pages_attempted);
1917
1918                 if (pages_attempted >= pages_max_per_user) {
1919                         printed += scnprintf(buf + printed, size - printed,
1920                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1921                                              pages_max_per_user + pages_attempted);
1922                 }
1923
1924                 printed += scnprintf(buf + printed, size - printed,
1925                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1926                 break;
1927         default:
1928                 scnprintf(buf, size, "%s", emsg);
1929                 break;
1930         }
1931
1932         return 0;
1933 }
1934
1935 void perf_evlist__to_front(struct perf_evlist *evlist,
1936                            struct perf_evsel *move_evsel)
1937 {
1938         struct perf_evsel *evsel, *n;
1939         LIST_HEAD(move);
1940
1941         if (move_evsel == perf_evlist__first(evlist))
1942                 return;
1943
1944         evlist__for_each_entry_safe(evlist, n, evsel) {
1945                 if (evsel->leader == move_evsel->leader)
1946                         list_move_tail(&evsel->node, &move);
1947         }
1948
1949         list_splice(&move, &evlist->entries);
1950 }
1951
1952 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1953                                      struct perf_evsel *tracking_evsel)
1954 {
1955         struct perf_evsel *evsel;
1956
1957         if (tracking_evsel->tracking)
1958                 return;
1959
1960         evlist__for_each_entry(evlist, evsel) {
1961                 if (evsel != tracking_evsel)
1962                         evsel->tracking = false;
1963         }
1964
1965         tracking_evsel->tracking = true;
1966 }
1967
1968 struct perf_evsel *
1969 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1970                                const char *str)
1971 {
1972         struct perf_evsel *evsel;
1973
1974         evlist__for_each_entry(evlist, evsel) {
1975                 if (!evsel->name)
1976                         continue;
1977                 if (strcmp(str, evsel->name) == 0)
1978                         return evsel;
1979         }
1980
1981         return NULL;
1982 }
1983
1984 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1985                                   enum bkw_mmap_state state)
1986 {
1987         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1988         enum action {
1989                 NONE,
1990                 PAUSE,
1991                 RESUME,
1992         } action = NONE;
1993
1994         if (!evlist->backward_mmap)
1995                 return;
1996
1997         switch (old_state) {
1998         case BKW_MMAP_NOTREADY: {
1999                 if (state != BKW_MMAP_RUNNING)
2000                         goto state_err;;
2001                 break;
2002         }
2003         case BKW_MMAP_RUNNING: {
2004                 if (state != BKW_MMAP_DATA_PENDING)
2005                         goto state_err;
2006                 action = PAUSE;
2007                 break;
2008         }
2009         case BKW_MMAP_DATA_PENDING: {
2010                 if (state != BKW_MMAP_EMPTY)
2011                         goto state_err;
2012                 break;
2013         }
2014         case BKW_MMAP_EMPTY: {
2015                 if (state != BKW_MMAP_RUNNING)
2016                         goto state_err;
2017                 action = RESUME;
2018                 break;
2019         }
2020         default:
2021                 WARN_ONCE(1, "Shouldn't get there\n");
2022         }
2023
2024         evlist->bkw_mmap_state = state;
2025
2026         switch (action) {
2027         case PAUSE:
2028                 perf_evlist__pause(evlist);
2029                 break;
2030         case RESUME:
2031                 perf_evlist__resume(evlist);
2032                 break;
2033         case NONE:
2034         default:
2035                 break;
2036         }
2037
2038 state_err:
2039         return;
2040 }