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/debugfs.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 <unistd.h>
  19
  20 #include "parse-events.h"
  21 #include "parse-options.h"
  22
  23 #include <sys/mman.h>
  24
  25 #include <linux/bitops.h>
  26 #include <linux/hash.h>
  27
  28 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  29 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  30
  31 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  32                        struct thread_map *threads)
  33 {
  34         int i;
  35
  36         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  37                 INIT_HLIST_HEAD(&evlist->heads[i]);
  38         INIT_LIST_HEAD(&evlist->entries);
  39         perf_evlist__set_maps(evlist, cpus, threads);
  40         evlist->workload.pid = -1;
  41 }
  42
  43 struct perf_evlist *perf_evlist__new(void)
  44 {
  45         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  46
  47         if (evlist != NULL)
  48                 perf_evlist__init(evlist, NULL, NULL);
  49
  50         return evlist;
  51 }
  52
  53 struct perf_evlist *perf_evlist__new_default(void)
  54 {
  55         struct perf_evlist *evlist = perf_evlist__new();
  56
  57         if (evlist && perf_evlist__add_default(evlist)) {
  58                 perf_evlist__delete(evlist);
  59                 evlist = NULL;
  60         }
  61
  62         return evlist;
  63 }
  64
  65 /**
  66  * perf_evlist__set_id_pos - set the positions of event ids.
  67  * @evlist: selected event list
  68  *
  69  * Events with compatible sample types all have the same id_pos
  70  * and is_pos.  For convenience, put a copy on evlist.
  71  */
  72 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  73 {
  74         struct perf_evsel *first = perf_evlist__first(evlist);
  75
  76         evlist->id_pos = first->id_pos;
  77         evlist->is_pos = first->is_pos;
  78 }
  79
  80 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  81 {
  82         struct perf_evsel *evsel;
  83
  84         evlist__for_each(evlist, evsel)
  85                 perf_evsel__calc_id_pos(evsel);
  86
  87         perf_evlist__set_id_pos(evlist);
  88 }
  89
  90 static void perf_evlist__purge(struct perf_evlist *evlist)
  91 {
  92         struct perf_evsel *pos, *n;
  93
  94         evlist__for_each_safe(evlist, n, pos) {
  95                 list_del_init(&pos->node);
  96                 perf_evsel__delete(pos);
  97         }
  98
  99         evlist->nr_entries = 0;
 100 }
 101
 102 void perf_evlist__exit(struct perf_evlist *evlist)
 103 {
 104         zfree(&evlist->mmap);
 105         zfree(&evlist->pollfd);
 106 }
 107
 108 void perf_evlist__delete(struct perf_evlist *evlist)
 109 {
 110         perf_evlist__munmap(evlist);
 111         perf_evlist__close(evlist);
 112         cpu_map__delete(evlist->cpus);
 113         thread_map__delete(evlist->threads);
 114         evlist->cpus = NULL;
 115         evlist->threads = NULL;
 116         perf_evlist__purge(evlist);
 117         perf_evlist__exit(evlist);
 118         free(evlist);
 119 }
 120
 121 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 122 {
 123         list_add_tail(&entry->node, &evlist->entries);
 124         entry->idx = evlist->nr_entries;
 125         entry->tracking = !entry->idx;
 126
 127         if (!evlist->nr_entries++)
 128                 perf_evlist__set_id_pos(evlist);
 129 }
 130
 131 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 132                                    struct list_head *list,
 133                                    int nr_entries)
 134 {
 135         bool set_id_pos = !evlist->nr_entries;
 136
 137         list_splice_tail(list, &evlist->entries);
 138         evlist->nr_entries += nr_entries;
 139         if (set_id_pos)
 140                 perf_evlist__set_id_pos(evlist);
 141 }
 142
 143 void __perf_evlist__set_leader(struct list_head *list)
 144 {
 145         struct perf_evsel *evsel, *leader;
 146
 147         leader = list_entry(list->next, struct perf_evsel, node);
 148         evsel = list_entry(list->prev, struct perf_evsel, node);
 149
 150         leader->nr_members = evsel->idx - leader->idx + 1;
 151
 152         __evlist__for_each(list, evsel) {
 153                 evsel->leader = leader;
 154         }
 155 }
 156
 157 void perf_evlist__set_leader(struct perf_evlist *evlist)
 158 {
 159         if (evlist->nr_entries) {
 160                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 161                 __perf_evlist__set_leader(&evlist->entries);
 162         }
 163 }
 164
 165 int perf_evlist__add_default(struct perf_evlist *evlist)
 166 {
 167         struct perf_event_attr attr = {
 168                 .type = PERF_TYPE_HARDWARE,
 169                 .config = PERF_COUNT_HW_CPU_CYCLES,
 170         };
 171         struct perf_evsel *evsel;
 172
 173         event_attr_init(&attr);
 174
 175         evsel = perf_evsel__new(&attr);
 176         if (evsel == NULL)
 177                 goto error;
 178
 179         /* use strdup() because free(evsel) assumes name is allocated */
 180         evsel->name = strdup("cycles");
 181         if (!evsel->name)
 182                 goto error_free;
 183
 184         perf_evlist__add(evlist, evsel);
 185         return 0;
 186 error_free:
 187         perf_evsel__delete(evsel);
 188 error:
 189         return -ENOMEM;
 190 }
 191
 192 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 193                                   struct perf_event_attr *attrs, size_t nr_attrs)
 194 {
 195         struct perf_evsel *evsel, *n;
 196         LIST_HEAD(head);
 197         size_t i;
 198
 199         for (i = 0; i < nr_attrs; i++) {
 200                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 201                 if (evsel == NULL)
 202                         goto out_delete_partial_list;
 203                 list_add_tail(&evsel->node, &head);
 204         }
 205
 206         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 207
 208         return 0;
 209
 210 out_delete_partial_list:
 211         __evlist__for_each_safe(&head, n, evsel)
 212                 perf_evsel__delete(evsel);
 213         return -1;
 214 }
 215
 216 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 217                                      struct perf_event_attr *attrs, size_t nr_attrs)
 218 {
 219         size_t i;
 220
 221         for (i = 0; i < nr_attrs; i++)
 222                 event_attr_init(attrs + i);
 223
 224         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 225 }
 226
 227 struct perf_evsel *
 228 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 229 {
 230         struct perf_evsel *evsel;
 231
 232         evlist__for_each(evlist, evsel) {
 233                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 234                     (int)evsel->attr.config == id)
 235                         return evsel;
 236         }
 237
 238         return NULL;
 239 }
 240
 241 struct perf_evsel *
 242 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 243                                      const char *name)
 244 {
 245         struct perf_evsel *evsel;
 246
 247         evlist__for_each(evlist, evsel) {
 248                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 249                     (strcmp(evsel->name, name) == 0))
 250                         return evsel;
 251         }
 252
 253         return NULL;
 254 }
 255
 256 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 257                            const char *sys, const char *name, void *handler)
 258 {
 259         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 260
 261         if (evsel == NULL)
 262                 return -1;
 263
 264         evsel->handler = handler;
 265         perf_evlist__add(evlist, evsel);
 266         return 0;
 267 }
 268
 269 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 270                                    struct perf_evsel *evsel)
 271 {
 272         if (evsel->system_wide)
 273                 return 1;
 274         else
 275                 return thread_map__nr(evlist->threads);
 276 }
 277
 278 void perf_evlist__disable(struct perf_evlist *evlist)
 279 {
 280         int cpu, thread;
 281         struct perf_evsel *pos;
 282         int nr_cpus = cpu_map__nr(evlist->cpus);
 283         int nr_threads;
 284
 285         for (cpu = 0; cpu < nr_cpus; cpu++) {
 286                 evlist__for_each(evlist, pos) {
 287                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 288                                 continue;
 289                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 290                         for (thread = 0; thread < nr_threads; thread++)
 291                                 ioctl(FD(pos, cpu, thread),
 292                                       PERF_EVENT_IOC_DISABLE, 0);
 293                 }
 294         }
 295 }
 296
 297 void perf_evlist__enable(struct perf_evlist *evlist)
 298 {
 299         int cpu, thread;
 300         struct perf_evsel *pos;
 301         int nr_cpus = cpu_map__nr(evlist->cpus);
 302         int nr_threads;
 303
 304         for (cpu = 0; cpu < nr_cpus; cpu++) {
 305                 evlist__for_each(evlist, pos) {
 306                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 307                                 continue;
 308                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 309                         for (thread = 0; thread < nr_threads; thread++)
 310                                 ioctl(FD(pos, cpu, thread),
 311                                       PERF_EVENT_IOC_ENABLE, 0);
 312                 }
 313         }
 314 }
 315
 316 int perf_evlist__disable_event(struct perf_evlist *evlist,
 317                                struct perf_evsel *evsel)
 318 {
 319         int cpu, thread, err;
 320         int nr_cpus = cpu_map__nr(evlist->cpus);
 321         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 322
 323         if (!evsel->fd)
 324                 return 0;
 325
 326         for (cpu = 0; cpu < nr_cpus; cpu++) {
 327                 for (thread = 0; thread < nr_threads; thread++) {
 328                         err = ioctl(FD(evsel, cpu, thread),
 329                                     PERF_EVENT_IOC_DISABLE, 0);
 330                         if (err)
 331                                 return err;
 332                 }
 333         }
 334         return 0;
 335 }
 336
 337 int perf_evlist__enable_event(struct perf_evlist *evlist,
 338                               struct perf_evsel *evsel)
 339 {
 340         int cpu, thread, err;
 341         int nr_cpus = cpu_map__nr(evlist->cpus);
 342         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 343
 344         if (!evsel->fd)
 345                 return -EINVAL;
 346
 347         for (cpu = 0; cpu < nr_cpus; cpu++) {
 348                 for (thread = 0; thread < nr_threads; thread++) {
 349                         err = ioctl(FD(evsel, cpu, thread),
 350                                     PERF_EVENT_IOC_ENABLE, 0);
 351                         if (err)
 352                                 return err;
 353                 }
 354         }
 355         return 0;
 356 }
 357
 358 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 359                                          struct perf_evsel *evsel, int cpu)
 360 {
 361         int thread, err;
 362         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 363
 364         if (!evsel->fd)
 365                 return -EINVAL;
 366
 367         for (thread = 0; thread < nr_threads; thread++) {
 368                 err = ioctl(FD(evsel, cpu, thread),
 369                             PERF_EVENT_IOC_ENABLE, 0);
 370                 if (err)
 371                         return err;
 372         }
 373         return 0;
 374 }
 375
 376 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 377                                             struct perf_evsel *evsel,
 378                                             int thread)
 379 {
 380         int cpu, err;
 381         int nr_cpus = cpu_map__nr(evlist->cpus);
 382
 383         if (!evsel->fd)
 384                 return -EINVAL;
 385
 386         for (cpu = 0; cpu < nr_cpus; cpu++) {
 387                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 388                 if (err)
 389                         return err;
 390         }
 391         return 0;
 392 }
 393
 394 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 395                                   struct perf_evsel *evsel, int idx)
 396 {
 397         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 398
 399         if (per_cpu_mmaps)
 400                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 401         else
 402                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 403 }
 404
 405 static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 406 {
 407         int nr_cpus = cpu_map__nr(evlist->cpus);
 408         int nr_threads = thread_map__nr(evlist->threads);
 409         int nfds = 0;
 410         struct perf_evsel *evsel;
 411
 412         list_for_each_entry(evsel, &evlist->entries, node) {
 413                 if (evsel->system_wide)
 414                         nfds += nr_cpus;
 415                 else
 416                         nfds += nr_cpus * nr_threads;
 417         }
 418
 419         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 420         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 421 }
 422
 423 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 424 {
 425         fcntl(fd, F_SETFL, O_NONBLOCK);
 426         evlist->pollfd[evlist->nr_fds].fd = fd;
 427         evlist->pollfd[evlist->nr_fds].events = POLLIN | POLLERR | POLLHUP;
 428         evlist->nr_fds++;
 429 }
 430
 431 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 432 {
 433         int fd, nr_fds = 0;
 434
 435         if (evlist->nr_fds == 0)
 436                 return 0;
 437
 438         for (fd = 0; fd < evlist->nr_fds; ++fd) {
 439                 if (evlist->pollfd[fd].revents & revents_and_mask)
 440                         continue;
 441
 442                 if (fd != nr_fds)
 443                         evlist->pollfd[nr_fds] = evlist->pollfd[fd];
 444
 445                 ++nr_fds;
 446         }
 447
 448         evlist->nr_fds = nr_fds;
 449         return nr_fds;
 450 }
 451
 452 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 453                                  struct perf_evsel *evsel,
 454                                  int cpu, int thread, u64 id)
 455 {
 456         int hash;
 457         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 458
 459         sid->id = id;
 460         sid->evsel = evsel;
 461         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 462         hlist_add_head(&sid->node, &evlist->heads[hash]);
 463 }
 464
 465 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 466                          int cpu, int thread, u64 id)
 467 {
 468         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 469         evsel->id[evsel->ids++] = id;
 470 }
 471
 472 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 473                                   struct perf_evsel *evsel,
 474                                   int cpu, int thread, int fd)
 475 {
 476         u64 read_data[4] = { 0, };
 477         int id_idx = 1; /* The first entry is the counter value */
 478         u64 id;
 479         int ret;
 480
 481         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 482         if (!ret)
 483                 goto add;
 484
 485         if (errno != ENOTTY)
 486                 return -1;
 487
 488         /* Legacy way to get event id.. All hail to old kernels! */
 489
 490         /*
 491          * This way does not work with group format read, so bail
 492          * out in that case.
 493          */
 494         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 495                 return -1;
 496
 497         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 498             read(fd, &read_data, sizeof(read_data)) == -1)
 499                 return -1;
 500
 501         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 502                 ++id_idx;
 503         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 504                 ++id_idx;
 505
 506         id = read_data[id_idx];
 507
 508  add:
 509         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 510         return 0;
 511 }
 512
 513 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 514 {
 515         struct hlist_head *head;
 516         struct perf_sample_id *sid;
 517         int hash;
 518
 519         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 520         head = &evlist->heads[hash];
 521
 522         hlist_for_each_entry(sid, head, node)
 523                 if (sid->id == id)
 524                         return sid;
 525
 526         return NULL;
 527 }
 528
 529 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 530 {
 531         struct perf_sample_id *sid;
 532
 533         if (evlist->nr_entries == 1)
 534                 return perf_evlist__first(evlist);
 535
 536         sid = perf_evlist__id2sid(evlist, id);
 537         if (sid)
 538                 return sid->evsel;
 539
 540         if (!perf_evlist__sample_id_all(evlist))
 541                 return perf_evlist__first(evlist);
 542
 543         return NULL;
 544 }
 545
 546 static int perf_evlist__event2id(struct perf_evlist *evlist,
 547                                  union perf_event *event, u64 *id)
 548 {
 549         const u64 *array = event->sample.array;
 550         ssize_t n;
 551
 552         n = (event->header.size - sizeof(event->header)) >> 3;
 553
 554         if (event->header.type == PERF_RECORD_SAMPLE) {
 555                 if (evlist->id_pos >= n)
 556                         return -1;
 557                 *id = array[evlist->id_pos];
 558         } else {
 559                 if (evlist->is_pos > n)
 560                         return -1;
 561                 n -= evlist->is_pos;
 562                 *id = array[n];
 563         }
 564         return 0;
 565 }
 566
 567 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 568                                                    union perf_event *event)
 569 {
 570         struct perf_evsel *first = perf_evlist__first(evlist);
 571         struct hlist_head *head;
 572         struct perf_sample_id *sid;
 573         int hash;
 574         u64 id;
 575
 576         if (evlist->nr_entries == 1)
 577                 return first;
 578
 579         if (!first->attr.sample_id_all &&
 580             event->header.type != PERF_RECORD_SAMPLE)
 581                 return first;
 582
 583         if (perf_evlist__event2id(evlist, event, &id))
 584                 return NULL;
 585
 586         /* Synthesized events have an id of zero */
 587         if (!id)
 588                 return first;
 589
 590         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 591         head = &evlist->heads[hash];
 592
 593         hlist_for_each_entry(sid, head, node) {
 594                 if (sid->id == id)
 595                         return sid->evsel;
 596         }
 597         return NULL;
 598 }
 599
 600 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 601 {
 602         struct perf_mmap *md = &evlist->mmap[idx];
 603         unsigned int head = perf_mmap__read_head(md);
 604         unsigned int old = md->prev;
 605         unsigned char *data = md->base + page_size;
 606         union perf_event *event = NULL;
 607
 608         if (evlist->overwrite) {
 609                 /*
 610                  * If we're further behind than half the buffer, there's a chance
 611                  * the writer will bite our tail and mess up the samples under us.
 612                  *
 613                  * If we somehow ended up ahead of the head, we got messed up.
 614                  *
 615                  * In either case, truncate and restart at head.
 616                  */
 617                 int diff = head - old;
 618                 if (diff > md->mask / 2 || diff < 0) {
 619                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 620
 621                         /*
 622                          * head points to a known good entry, start there.
 623                          */
 624                         old = head;
 625                 }
 626         }
 627
 628         if (old != head) {
 629                 size_t size;
 630
 631                 event = (union perf_event *)&data[old & md->mask];
 632                 size = event->header.size;
 633
 634                 /*
 635                  * Event straddles the mmap boundary -- header should always
 636                  * be inside due to u64 alignment of output.
 637                  */
 638                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 639                         unsigned int offset = old;
 640                         unsigned int len = min(sizeof(*event), size), cpy;
 641                         void *dst = md->event_copy;
 642
 643                         do {
 644                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 645                                 memcpy(dst, &data[offset & md->mask], cpy);
 646                                 offset += cpy;
 647                                 dst += cpy;
 648                                 len -= cpy;
 649                         } while (len);
 650
 651                         event = (union perf_event *) md->event_copy;
 652                 }
 653
 654                 old += size;
 655         }
 656
 657         md->prev = old;
 658
 659         return event;
 660 }
 661
 662 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 663 {
 664         if (!evlist->overwrite) {
 665                 struct perf_mmap *md = &evlist->mmap[idx];
 666                 unsigned int old = md->prev;
 667
 668                 perf_mmap__write_tail(md, old);
 669         }
 670 }
 671
 672 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 673 {
 674         if (evlist->mmap[idx].base != NULL) {
 675                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 676                 evlist->mmap[idx].base = NULL;
 677         }
 678 }
 679
 680 void perf_evlist__munmap(struct perf_evlist *evlist)
 681 {
 682         int i;
 683
 684         if (evlist->mmap == NULL)
 685                 return;
 686
 687         for (i = 0; i < evlist->nr_mmaps; i++)
 688                 __perf_evlist__munmap(evlist, i);
 689
 690         zfree(&evlist->mmap);
 691 }
 692
 693 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 694 {
 695         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 696         if (cpu_map__empty(evlist->cpus))
 697                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 698         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 699         return evlist->mmap != NULL ? 0 : -ENOMEM;
 700 }
 701
 702 struct mmap_params {
 703         int prot;
 704         int mask;
 705 };
 706
 707 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 708                                struct mmap_params *mp, int fd)
 709 {
 710         evlist->mmap[idx].prev = 0;
 711         evlist->mmap[idx].mask = mp->mask;
 712         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 713                                       MAP_SHARED, fd, 0);
 714         if (evlist->mmap[idx].base == MAP_FAILED) {
 715                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 716                           errno);
 717                 evlist->mmap[idx].base = NULL;
 718                 return -1;
 719         }
 720
 721         perf_evlist__add_pollfd(evlist, fd);
 722         return 0;
 723 }
 724
 725 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 726                                        struct mmap_params *mp, int cpu,
 727                                        int thread, int *output)
 728 {
 729         struct perf_evsel *evsel;
 730
 731         evlist__for_each(evlist, evsel) {
 732                 int fd;
 733
 734                 if (evsel->system_wide && thread)
 735                         continue;
 736
 737                 fd = FD(evsel, cpu, thread);
 738
 739                 if (*output == -1) {
 740                         *output = fd;
 741                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 742                                 return -1;
 743                 } else {
 744                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 745                                 return -1;
 746                 }
 747
 748                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 749                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 750                         return -1;
 751         }
 752
 753         return 0;
 754 }
 755
 756 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 757                                      struct mmap_params *mp)
 758 {
 759         int cpu, thread;
 760         int nr_cpus = cpu_map__nr(evlist->cpus);
 761         int nr_threads = thread_map__nr(evlist->threads);
 762
 763         pr_debug2("perf event ring buffer mmapped per cpu\n");
 764         for (cpu = 0; cpu < nr_cpus; cpu++) {
 765                 int output = -1;
 766
 767                 for (thread = 0; thread < nr_threads; thread++) {
 768                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 769                                                         thread, &output))
 770                                 goto out_unmap;
 771                 }
 772         }
 773
 774         return 0;
 775
 776 out_unmap:
 777         for (cpu = 0; cpu < nr_cpus; cpu++)
 778                 __perf_evlist__munmap(evlist, cpu);
 779         return -1;
 780 }
 781
 782 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 783                                         struct mmap_params *mp)
 784 {
 785         int thread;
 786         int nr_threads = thread_map__nr(evlist->threads);
 787
 788         pr_debug2("perf event ring buffer mmapped per thread\n");
 789         for (thread = 0; thread < nr_threads; thread++) {
 790                 int output = -1;
 791
 792                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 793                                                 &output))
 794                         goto out_unmap;
 795         }
 796
 797         return 0;
 798
 799 out_unmap:
 800         for (thread = 0; thread < nr_threads; thread++)
 801                 __perf_evlist__munmap(evlist, thread);
 802         return -1;
 803 }
 804
 805 static size_t perf_evlist__mmap_size(unsigned long pages)
 806 {
 807         /* 512 kiB: default amount of unprivileged mlocked memory */
 808         if (pages == UINT_MAX)
 809                 pages = (512 * 1024) / page_size;
 810         else if (!is_power_of_2(pages))
 811                 return 0;
 812
 813         return (pages + 1) * page_size;
 814 }
 815
 816 static long parse_pages_arg(const char *str, unsigned long min,
 817                             unsigned long max)
 818 {
 819         unsigned long pages, val;
 820         static struct parse_tag tags[] = {
 821                 { .tag  = 'B', .mult = 1       },
 822                 { .tag  = 'K', .mult = 1 << 10 },
 823                 { .tag  = 'M', .mult = 1 << 20 },
 824                 { .tag  = 'G', .mult = 1 << 30 },
 825                 { .tag  = 0 },
 826         };
 827
 828         if (str == NULL)
 829                 return -EINVAL;
 830
 831         val = parse_tag_value(str, tags);
 832         if (val != (unsigned long) -1) {
 833                 /* we got file size value */
 834                 pages = PERF_ALIGN(val, page_size) / page_size;
 835         } else {
 836                 /* we got pages count value */
 837                 char *eptr;
 838                 pages = strtoul(str, &eptr, 10);
 839                 if (*eptr != '\0')
 840                         return -EINVAL;
 841         }
 842
 843         if (pages == 0 && min == 0) {
 844                 /* leave number of pages at 0 */
 845         } else if (!is_power_of_2(pages)) {
 846                 /* round pages up to next power of 2 */
 847                 pages = next_pow2_l(pages);
 848                 if (!pages)
 849                         return -EINVAL;
 850                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 851                         pages * page_size, pages);
 852         }
 853
 854         if (pages > max)
 855                 return -EINVAL;
 856
 857         return pages;
 858 }
 859
 860 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 861                                   int unset __maybe_unused)
 862 {
 863         unsigned int *mmap_pages = opt->value;
 864         unsigned long max = UINT_MAX;
 865         long pages;
 866
 867         if (max > SIZE_MAX / page_size)
 868                 max = SIZE_MAX / page_size;
 869
 870         pages = parse_pages_arg(str, 1, max);
 871         if (pages < 0) {
 872                 pr_err("Invalid argument for --mmap_pages/-m\n");
 873                 return -1;
 874         }
 875
 876         *mmap_pages = pages;
 877         return 0;
 878 }
 879
 880 /**
 881  * perf_evlist__mmap - Create mmaps to receive events.
 882  * @evlist: list of events
 883  * @pages: map length in pages
 884  * @overwrite: overwrite older events?
 885  *
 886  * If @overwrite is %false the user needs to signal event consumption using
 887  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 888  * automatically.
 889  *
 890  * Return: %0 on success, negative error code otherwise.
 891  */
 892 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 893                       bool overwrite)
 894 {
 895         struct perf_evsel *evsel;
 896         const struct cpu_map *cpus = evlist->cpus;
 897         const struct thread_map *threads = evlist->threads;
 898         struct mmap_params mp = {
 899                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
 900         };
 901
 902         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 903                 return -ENOMEM;
 904
 905         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 906                 return -ENOMEM;
 907
 908         evlist->overwrite = overwrite;
 909         evlist->mmap_len = perf_evlist__mmap_size(pages);
 910         pr_debug("mmap size %zuB\n", evlist->mmap_len);
 911         mp.mask = evlist->mmap_len - page_size - 1;
 912
 913         evlist__for_each(evlist, evsel) {
 914                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 915                     evsel->sample_id == NULL &&
 916                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
 917                         return -ENOMEM;
 918         }
 919
 920         if (cpu_map__empty(cpus))
 921                 return perf_evlist__mmap_per_thread(evlist, &mp);
 922
 923         return perf_evlist__mmap_per_cpu(evlist, &mp);
 924 }
 925
 926 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
 927 {
 928         evlist->threads = thread_map__new_str(target->pid, target->tid,
 929                                               target->uid);
 930
 931         if (evlist->threads == NULL)
 932                 return -1;
 933
 934         if (target__uses_dummy_map(target))
 935                 evlist->cpus = cpu_map__dummy_new();
 936         else
 937                 evlist->cpus = cpu_map__new(target->cpu_list);
 938
 939         if (evlist->cpus == NULL)
 940                 goto out_delete_threads;
 941
 942         return 0;
 943
 944 out_delete_threads:
 945         thread_map__delete(evlist->threads);
 946         return -1;
 947 }
 948
 949 int perf_evlist__apply_filters(struct perf_evlist *evlist)
 950 {
 951         struct perf_evsel *evsel;
 952         int err = 0;
 953         const int ncpus = cpu_map__nr(evlist->cpus),
 954                   nthreads = thread_map__nr(evlist->threads);
 955
 956         evlist__for_each(evlist, evsel) {
 957                 if (evsel->filter == NULL)
 958                         continue;
 959
 960                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
 961                 if (err)
 962                         break;
 963         }
 964
 965         return err;
 966 }
 967
 968 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 969 {
 970         struct perf_evsel *evsel;
 971         int err = 0;
 972         const int ncpus = cpu_map__nr(evlist->cpus),
 973                   nthreads = thread_map__nr(evlist->threads);
 974
 975         evlist__for_each(evlist, evsel) {
 976                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
 977                 if (err)
 978                         break;
 979         }
 980
 981         return err;
 982 }
 983
 984 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 985 {
 986         struct perf_evsel *pos;
 987
 988         if (evlist->nr_entries == 1)
 989                 return true;
 990
 991         if (evlist->id_pos < 0 || evlist->is_pos < 0)
 992                 return false;
 993
 994         evlist__for_each(evlist, pos) {
 995                 if (pos->id_pos != evlist->id_pos ||
 996                     pos->is_pos != evlist->is_pos)
 997                         return false;
 998         }
 999
1000         return true;
1001 }
1002
1003 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1004 {
1005         struct perf_evsel *evsel;
1006
1007         if (evlist->combined_sample_type)
1008                 return evlist->combined_sample_type;
1009
1010         evlist__for_each(evlist, evsel)
1011                 evlist->combined_sample_type |= evsel->attr.sample_type;
1012
1013         return evlist->combined_sample_type;
1014 }
1015
1016 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1017 {
1018         evlist->combined_sample_type = 0;
1019         return __perf_evlist__combined_sample_type(evlist);
1020 }
1021
1022 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1023 {
1024         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1025         u64 read_format = first->attr.read_format;
1026         u64 sample_type = first->attr.sample_type;
1027
1028         evlist__for_each(evlist, pos) {
1029                 if (read_format != pos->attr.read_format)
1030                         return false;
1031         }
1032
1033         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1034         if ((sample_type & PERF_SAMPLE_READ) &&
1035             !(read_format & PERF_FORMAT_ID)) {
1036                 return false;
1037         }
1038
1039         return true;
1040 }
1041
1042 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1043 {
1044         struct perf_evsel *first = perf_evlist__first(evlist);
1045         return first->attr.read_format;
1046 }
1047
1048 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1049 {
1050         struct perf_evsel *first = perf_evlist__first(evlist);
1051         struct perf_sample *data;
1052         u64 sample_type;
1053         u16 size = 0;
1054
1055         if (!first->attr.sample_id_all)
1056                 goto out;
1057
1058         sample_type = first->attr.sample_type;
1059
1060         if (sample_type & PERF_SAMPLE_TID)
1061                 size += sizeof(data->tid) * 2;
1062
1063        if (sample_type & PERF_SAMPLE_TIME)
1064                 size += sizeof(data->time);
1065
1066         if (sample_type & PERF_SAMPLE_ID)
1067                 size += sizeof(data->id);
1068
1069         if (sample_type & PERF_SAMPLE_STREAM_ID)
1070                 size += sizeof(data->stream_id);
1071
1072         if (sample_type & PERF_SAMPLE_CPU)
1073                 size += sizeof(data->cpu) * 2;
1074
1075         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1076                 size += sizeof(data->id);
1077 out:
1078         return size;
1079 }
1080
1081 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1082 {
1083         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1084
1085         evlist__for_each_continue(evlist, pos) {
1086                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1087                         return false;
1088         }
1089
1090         return true;
1091 }
1092
1093 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1094 {
1095         struct perf_evsel *first = perf_evlist__first(evlist);
1096         return first->attr.sample_id_all;
1097 }
1098
1099 void perf_evlist__set_selected(struct perf_evlist *evlist,
1100                                struct perf_evsel *evsel)
1101 {
1102         evlist->selected = evsel;
1103 }
1104
1105 void perf_evlist__close(struct perf_evlist *evlist)
1106 {
1107         struct perf_evsel *evsel;
1108         int ncpus = cpu_map__nr(evlist->cpus);
1109         int nthreads = thread_map__nr(evlist->threads);
1110         int n;
1111
1112         evlist__for_each_reverse(evlist, evsel) {
1113                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1114                 perf_evsel__close(evsel, n, nthreads);
1115         }
1116 }
1117
1118 int perf_evlist__open(struct perf_evlist *evlist)
1119 {
1120         struct perf_evsel *evsel;
1121         int err;
1122
1123         perf_evlist__update_id_pos(evlist);
1124
1125         evlist__for_each(evlist, evsel) {
1126                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1127                 if (err < 0)
1128                         goto out_err;
1129         }
1130
1131         return 0;
1132 out_err:
1133         perf_evlist__close(evlist);
1134         errno = -err;
1135         return err;
1136 }
1137
1138 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1139                                   const char *argv[], bool pipe_output,
1140                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1141 {
1142         int child_ready_pipe[2], go_pipe[2];
1143         char bf;
1144
1145         if (pipe(child_ready_pipe) < 0) {
1146                 perror("failed to create 'ready' pipe");
1147                 return -1;
1148         }
1149
1150         if (pipe(go_pipe) < 0) {
1151                 perror("failed to create 'go' pipe");
1152                 goto out_close_ready_pipe;
1153         }
1154
1155         evlist->workload.pid = fork();
1156         if (evlist->workload.pid < 0) {
1157                 perror("failed to fork");
1158                 goto out_close_pipes;
1159         }
1160
1161         if (!evlist->workload.pid) {
1162                 int ret;
1163
1164                 if (pipe_output)
1165                         dup2(2, 1);
1166
1167                 signal(SIGTERM, SIG_DFL);
1168
1169                 close(child_ready_pipe[0]);
1170                 close(go_pipe[1]);
1171                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1172
1173                 /*
1174                  * Tell the parent we're ready to go
1175                  */
1176                 close(child_ready_pipe[1]);
1177
1178                 /*
1179                  * Wait until the parent tells us to go.
1180                  */
1181                 ret = read(go_pipe[0], &bf, 1);
1182                 /*
1183                  * The parent will ask for the execvp() to be performed by
1184                  * writing exactly one byte, in workload.cork_fd, usually via
1185                  * perf_evlist__start_workload().
1186                  *
1187                  * For cancelling the workload without actuallin running it,
1188                  * the parent will just close workload.cork_fd, without writing
1189                  * anything, i.e. read will return zero and we just exit()
1190                  * here.
1191                  */
1192                 if (ret != 1) {
1193                         if (ret == -1)
1194                                 perror("unable to read pipe");
1195                         exit(ret);
1196                 }
1197
1198                 execvp(argv[0], (char **)argv);
1199
1200                 if (exec_error) {
1201                         union sigval val;
1202
1203                         val.sival_int = errno;
1204                         if (sigqueue(getppid(), SIGUSR1, val))
1205                                 perror(argv[0]);
1206                 } else
1207                         perror(argv[0]);
1208                 exit(-1);
1209         }
1210
1211         if (exec_error) {
1212                 struct sigaction act = {
1213                         .sa_flags     = SA_SIGINFO,
1214                         .sa_sigaction = exec_error,
1215                 };
1216                 sigaction(SIGUSR1, &act, NULL);
1217         }
1218
1219         if (target__none(target))
1220                 evlist->threads->map[0] = evlist->workload.pid;
1221
1222         close(child_ready_pipe[1]);
1223         close(go_pipe[0]);
1224         /*
1225          * wait for child to settle
1226          */
1227         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1228                 perror("unable to read pipe");
1229                 goto out_close_pipes;
1230         }
1231
1232         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1233         evlist->workload.cork_fd = go_pipe[1];
1234         close(child_ready_pipe[0]);
1235         return 0;
1236
1237 out_close_pipes:
1238         close(go_pipe[0]);
1239         close(go_pipe[1]);
1240 out_close_ready_pipe:
1241         close(child_ready_pipe[0]);
1242         close(child_ready_pipe[1]);
1243         return -1;
1244 }
1245
1246 int perf_evlist__start_workload(struct perf_evlist *evlist)
1247 {
1248         if (evlist->workload.cork_fd > 0) {
1249                 char bf = 0;
1250                 int ret;
1251                 /*
1252                  * Remove the cork, let it rip!
1253                  */
1254                 ret = write(evlist->workload.cork_fd, &bf, 1);
1255                 if (ret < 0)
1256                         perror("enable to write to pipe");
1257
1258                 close(evlist->workload.cork_fd);
1259                 return ret;
1260         }
1261
1262         return 0;
1263 }
1264
1265 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1266                               struct perf_sample *sample)
1267 {
1268         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1269
1270         if (!evsel)
1271                 return -EFAULT;
1272         return perf_evsel__parse_sample(evsel, event, sample);
1273 }
1274
1275 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1276 {
1277         struct perf_evsel *evsel;
1278         size_t printed = 0;
1279
1280         evlist__for_each(evlist, evsel) {
1281                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1282                                    perf_evsel__name(evsel));
1283         }
1284
1285         return printed + fprintf(fp, "\n");
1286 }
1287
1288 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1289                              int err, char *buf, size_t size)
1290 {
1291         char sbuf[128];
1292
1293         switch (err) {
1294         case ENOENT:
1295                 scnprintf(buf, size, "%s",
1296                           "Error:\tUnable to find debugfs\n"
1297                           "Hint:\tWas your kernel was compiled with debugfs support?\n"
1298                           "Hint:\tIs the debugfs filesystem mounted?\n"
1299                           "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1300                 break;
1301         case EACCES:
1302                 scnprintf(buf, size,
1303                           "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1304                           "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1305                           debugfs_mountpoint, debugfs_mountpoint);
1306                 break;
1307         default:
1308                 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1309                 break;
1310         }
1311
1312         return 0;
1313 }
1314
1315 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1316                                int err, char *buf, size_t size)
1317 {
1318         int printed, value;
1319         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1320
1321         switch (err) {
1322         case EACCES:
1323         case EPERM:
1324                 printed = scnprintf(buf, size,
1325                                     "Error:\t%s.\n"
1326                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1327
1328                 value = perf_event_paranoid();
1329
1330                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1331
1332                 if (value >= 2) {
1333                         printed += scnprintf(buf + printed, size - printed,
1334                                              "For your workloads it needs to be <= 1\nHint:\t");
1335                 }
1336                 printed += scnprintf(buf + printed, size - printed,
1337                                      "For system wide tracing it needs to be set to -1.\n");
1338
1339                 printed += scnprintf(buf + printed, size - printed,
1340                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1341                                     "Hint:\tThe current value is %d.", value);
1342                 break;
1343         default:
1344                 scnprintf(buf, size, "%s", emsg);
1345                 break;
1346         }
1347
1348         return 0;
1349 }
1350
1351 void perf_evlist__to_front(struct perf_evlist *evlist,
1352                            struct perf_evsel *move_evsel)
1353 {
1354         struct perf_evsel *evsel, *n;
1355         LIST_HEAD(move);
1356
1357         if (move_evsel == perf_evlist__first(evlist))
1358                 return;
1359
1360         evlist__for_each_safe(evlist, n, evsel) {
1361                 if (evsel->leader == move_evsel->leader)
1362                         list_move_tail(&evsel->node, &move);
1363         }
1364
1365         list_splice(&move, &evlist->entries);
1366 }
1367
1368 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1369                                      struct perf_evsel *tracking_evsel)
1370 {
1371         struct perf_evsel *evsel;
1372
1373         if (tracking_evsel->tracking)
1374                 return;
1375
1376         evlist__for_each(evlist, evsel) {
1377                 if (evsel != tracking_evsel)
1378                         evsel->tracking = false;
1379         }
1380
1381         tracking_evsel->tracking = true;
1382 }