kernel/tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <poll.h>
  12 #include "cpumap.h"
  13 #include "thread_map.h"
  14 #include "target.h"
  15 #include "evlist.h"
  16 #include "evsel.h"
  17 #include "debug.h"
  18 #include <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 #include <linux/log2.h>
  28
  29 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
  30 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
  31
  32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  33 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  34
  35 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  36                        struct thread_map *threads)
  37 {
  38         int i;
  39
  40         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  41                 INIT_HLIST_HEAD(&evlist->heads[i]);
  42         INIT_LIST_HEAD(&evlist->entries);
  43         perf_evlist__set_maps(evlist, cpus, threads);
  44         fdarray__init(&evlist->pollfd, 64);
  45         evlist->workload.pid = -1;
  46 }
  47
  48 struct perf_evlist *perf_evlist__new(void)
  49 {
  50         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  51
  52         if (evlist != NULL)
  53                 perf_evlist__init(evlist, NULL, NULL);
  54
  55         return evlist;
  56 }
  57
  58 struct perf_evlist *perf_evlist__new_default(void)
  59 {
  60         struct perf_evlist *evlist = perf_evlist__new();
  61
  62         if (evlist && perf_evlist__add_default(evlist)) {
  63                 perf_evlist__delete(evlist);
  64                 evlist = NULL;
  65         }
  66
  67         return evlist;
  68 }
  69
  70 /**
  71  * perf_evlist__set_id_pos - set the positions of event ids.
  72  * @evlist: selected event list
  73  *
  74  * Events with compatible sample types all have the same id_pos
  75  * and is_pos.  For convenience, put a copy on evlist.
  76  */
  77 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  78 {
  79         struct perf_evsel *first = perf_evlist__first(evlist);
  80
  81         evlist->id_pos = first->id_pos;
  82         evlist->is_pos = first->is_pos;
  83 }
  84
  85 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  86 {
  87         struct perf_evsel *evsel;
  88
  89         evlist__for_each(evlist, evsel)
  90                 perf_evsel__calc_id_pos(evsel);
  91
  92         perf_evlist__set_id_pos(evlist);
  93 }
  94
  95 static void perf_evlist__purge(struct perf_evlist *evlist)
  96 {
  97         struct perf_evsel *pos, *n;
  98
  99         evlist__for_each_safe(evlist, n, pos) {
 100                 list_del_init(&pos->node);
 101                 perf_evsel__delete(pos);
 102         }
 103
 104         evlist->nr_entries = 0;
 105 }
 106
 107 void perf_evlist__exit(struct perf_evlist *evlist)
 108 {
 109         zfree(&evlist->mmap);
 110         fdarray__exit(&evlist->pollfd);
 111 }
 112
 113 void perf_evlist__delete(struct perf_evlist *evlist)
 114 {
 115         perf_evlist__munmap(evlist);
 116         perf_evlist__close(evlist);
 117         cpu_map__delete(evlist->cpus);
 118         thread_map__delete(evlist->threads);
 119         evlist->cpus = NULL;
 120         evlist->threads = NULL;
 121         perf_evlist__purge(evlist);
 122         perf_evlist__exit(evlist);
 123         free(evlist);
 124 }
 125
 126 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 127 {
 128         list_add_tail(&entry->node, &evlist->entries);
 129         entry->idx = evlist->nr_entries;
 130         entry->tracking = !entry->idx;
 131
 132         if (!evlist->nr_entries++)
 133                 perf_evlist__set_id_pos(evlist);
 134 }
 135
 136 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 137                                    struct list_head *list,
 138                                    int nr_entries)
 139 {
 140         bool set_id_pos = !evlist->nr_entries;
 141
 142         list_splice_tail(list, &evlist->entries);
 143         evlist->nr_entries += nr_entries;
 144         if (set_id_pos)
 145                 perf_evlist__set_id_pos(evlist);
 146 }
 147
 148 void __perf_evlist__set_leader(struct list_head *list)
 149 {
 150         struct perf_evsel *evsel, *leader;
 151
 152         leader = list_entry(list->next, struct perf_evsel, node);
 153         evsel = list_entry(list->prev, struct perf_evsel, node);
 154
 155         leader->nr_members = evsel->idx - leader->idx + 1;
 156
 157         __evlist__for_each(list, evsel) {
 158                 evsel->leader = leader;
 159         }
 160 }
 161
 162 void perf_evlist__set_leader(struct perf_evlist *evlist)
 163 {
 164         if (evlist->nr_entries) {
 165                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 166                 __perf_evlist__set_leader(&evlist->entries);
 167         }
 168 }
 169
 170 int perf_evlist__add_default(struct perf_evlist *evlist)
 171 {
 172         struct perf_event_attr attr = {
 173                 .type = PERF_TYPE_HARDWARE,
 174                 .config = PERF_COUNT_HW_CPU_CYCLES,
 175         };
 176         struct perf_evsel *evsel;
 177
 178         event_attr_init(&attr);
 179
 180         evsel = perf_evsel__new(&attr);
 181         if (evsel == NULL)
 182                 goto error;
 183
 184         /* use strdup() because free(evsel) assumes name is allocated */
 185         evsel->name = strdup("cycles");
 186         if (!evsel->name)
 187                 goto error_free;
 188
 189         perf_evlist__add(evlist, evsel);
 190         return 0;
 191 error_free:
 192         perf_evsel__delete(evsel);
 193 error:
 194         return -ENOMEM;
 195 }
 196
 197 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 198                                   struct perf_event_attr *attrs, size_t nr_attrs)
 199 {
 200         struct perf_evsel *evsel, *n;
 201         LIST_HEAD(head);
 202         size_t i;
 203
 204         for (i = 0; i < nr_attrs; i++) {
 205                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 206                 if (evsel == NULL)
 207                         goto out_delete_partial_list;
 208                 list_add_tail(&evsel->node, &head);
 209         }
 210
 211         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 212
 213         return 0;
 214
 215 out_delete_partial_list:
 216         __evlist__for_each_safe(&head, n, evsel)
 217                 perf_evsel__delete(evsel);
 218         return -1;
 219 }
 220
 221 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 222                                      struct perf_event_attr *attrs, size_t nr_attrs)
 223 {
 224         size_t i;
 225
 226         for (i = 0; i < nr_attrs; i++)
 227                 event_attr_init(attrs + i);
 228
 229         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 230 }
 231
 232 struct perf_evsel *
 233 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 234 {
 235         struct perf_evsel *evsel;
 236
 237         evlist__for_each(evlist, evsel) {
 238                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 239                     (int)evsel->attr.config == id)
 240                         return evsel;
 241         }
 242
 243         return NULL;
 244 }
 245
 246 struct perf_evsel *
 247 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 248                                      const char *name)
 249 {
 250         struct perf_evsel *evsel;
 251
 252         evlist__for_each(evlist, evsel) {
 253                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 254                     (strcmp(evsel->name, name) == 0))
 255                         return evsel;
 256         }
 257
 258         return NULL;
 259 }
 260
 261 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 262                            const char *sys, const char *name, void *handler)
 263 {
 264         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 265
 266         if (evsel == NULL)
 267                 return -1;
 268
 269         evsel->handler = handler;
 270         perf_evlist__add(evlist, evsel);
 271         return 0;
 272 }
 273
 274 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 275                                    struct perf_evsel *evsel)
 276 {
 277         if (evsel->system_wide)
 278                 return 1;
 279         else
 280                 return thread_map__nr(evlist->threads);
 281 }
 282
 283 void perf_evlist__disable(struct perf_evlist *evlist)
 284 {
 285         int cpu, thread;
 286         struct perf_evsel *pos;
 287         int nr_cpus = cpu_map__nr(evlist->cpus);
 288         int nr_threads;
 289
 290         for (cpu = 0; cpu < nr_cpus; cpu++) {
 291                 evlist__for_each(evlist, pos) {
 292                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 293                                 continue;
 294                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 295                         for (thread = 0; thread < nr_threads; thread++)
 296                                 ioctl(FD(pos, cpu, thread),
 297                                       PERF_EVENT_IOC_DISABLE, 0);
 298                 }
 299         }
 300 }
 301
 302 void perf_evlist__enable(struct perf_evlist *evlist)
 303 {
 304         int cpu, thread;
 305         struct perf_evsel *pos;
 306         int nr_cpus = cpu_map__nr(evlist->cpus);
 307         int nr_threads;
 308
 309         for (cpu = 0; cpu < nr_cpus; cpu++) {
 310                 evlist__for_each(evlist, pos) {
 311                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 312                                 continue;
 313                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 314                         for (thread = 0; thread < nr_threads; thread++)
 315                                 ioctl(FD(pos, cpu, thread),
 316                                       PERF_EVENT_IOC_ENABLE, 0);
 317                 }
 318         }
 319 }
 320
 321 int perf_evlist__disable_event(struct perf_evlist *evlist,
 322                                struct perf_evsel *evsel)
 323 {
 324         int cpu, thread, err;
 325         int nr_cpus = cpu_map__nr(evlist->cpus);
 326         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 327
 328         if (!evsel->fd)
 329                 return 0;
 330
 331         for (cpu = 0; cpu < nr_cpus; cpu++) {
 332                 for (thread = 0; thread < nr_threads; thread++) {
 333                         err = ioctl(FD(evsel, cpu, thread),
 334                                     PERF_EVENT_IOC_DISABLE, 0);
 335                         if (err)
 336                                 return err;
 337                 }
 338         }
 339         return 0;
 340 }
 341
 342 int perf_evlist__enable_event(struct perf_evlist *evlist,
 343                               struct perf_evsel *evsel)
 344 {
 345         int cpu, thread, err;
 346         int nr_cpus = cpu_map__nr(evlist->cpus);
 347         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 348
 349         if (!evsel->fd)
 350                 return -EINVAL;
 351
 352         for (cpu = 0; cpu < nr_cpus; cpu++) {
 353                 for (thread = 0; thread < nr_threads; thread++) {
 354                         err = ioctl(FD(evsel, cpu, thread),
 355                                     PERF_EVENT_IOC_ENABLE, 0);
 356                         if (err)
 357                                 return err;
 358                 }
 359         }
 360         return 0;
 361 }
 362
 363 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 364                                          struct perf_evsel *evsel, int cpu)
 365 {
 366         int thread, err;
 367         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 368
 369         if (!evsel->fd)
 370                 return -EINVAL;
 371
 372         for (thread = 0; thread < nr_threads; thread++) {
 373                 err = ioctl(FD(evsel, cpu, thread),
 374                             PERF_EVENT_IOC_ENABLE, 0);
 375                 if (err)
 376                         return err;
 377         }
 378         return 0;
 379 }
 380
 381 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 382                                             struct perf_evsel *evsel,
 383                                             int thread)
 384 {
 385         int cpu, err;
 386         int nr_cpus = cpu_map__nr(evlist->cpus);
 387
 388         if (!evsel->fd)
 389                 return -EINVAL;
 390
 391         for (cpu = 0; cpu < nr_cpus; cpu++) {
 392                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 393                 if (err)
 394                         return err;
 395         }
 396         return 0;
 397 }
 398
 399 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 400                                   struct perf_evsel *evsel, int idx)
 401 {
 402         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 403
 404         if (per_cpu_mmaps)
 405                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 406         else
 407                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 408 }
 409
 410 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 411 {
 412         int nr_cpus = cpu_map__nr(evlist->cpus);
 413         int nr_threads = thread_map__nr(evlist->threads);
 414         int nfds = 0;
 415         struct perf_evsel *evsel;
 416
 417         evlist__for_each(evlist, evsel) {
 418                 if (evsel->system_wide)
 419                         nfds += nr_cpus;
 420                 else
 421                         nfds += nr_cpus * nr_threads;
 422         }
 423
 424         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 425             fdarray__grow(&evlist->pollfd, nfds) < 0)
 426                 return -ENOMEM;
 427
 428         return 0;
 429 }
 430
 431 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
 432 {
 433         int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
 434         /*
 435          * Save the idx so that when we filter out fds POLLHUP'ed we can
 436          * close the associated evlist->mmap[] entry.
 437          */
 438         if (pos >= 0) {
 439                 evlist->pollfd.priv[pos].idx = idx;
 440
 441                 fcntl(fd, F_SETFL, O_NONBLOCK);
 442         }
 443
 444         return pos;
 445 }
 446
 447 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 448 {
 449         return __perf_evlist__add_pollfd(evlist, fd, -1);
 450 }
 451
 452 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
 453 {
 454         struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
 455
 456         perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
 457 }
 458
 459 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 460 {
 461         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 462                                perf_evlist__munmap_filtered);
 463 }
 464
 465 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 466 {
 467         return fdarray__poll(&evlist->pollfd, timeout);
 468 }
 469
 470 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 471                                  struct perf_evsel *evsel,
 472                                  int cpu, int thread, u64 id)
 473 {
 474         int hash;
 475         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 476
 477         sid->id = id;
 478         sid->evsel = evsel;
 479         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 480         hlist_add_head(&sid->node, &evlist->heads[hash]);
 481 }
 482
 483 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 484                          int cpu, int thread, u64 id)
 485 {
 486         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 487         evsel->id[evsel->ids++] = id;
 488 }
 489
 490 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 491                                   struct perf_evsel *evsel,
 492                                   int cpu, int thread, int fd)
 493 {
 494         u64 read_data[4] = { 0, };
 495         int id_idx = 1; /* The first entry is the counter value */
 496         u64 id;
 497         int ret;
 498
 499         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 500         if (!ret)
 501                 goto add;
 502
 503         if (errno != ENOTTY)
 504                 return -1;
 505
 506         /* Legacy way to get event id.. All hail to old kernels! */
 507
 508         /*
 509          * This way does not work with group format read, so bail
 510          * out in that case.
 511          */
 512         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 513                 return -1;
 514
 515         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 516             read(fd, &read_data, sizeof(read_data)) == -1)
 517                 return -1;
 518
 519         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 520                 ++id_idx;
 521         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 522                 ++id_idx;
 523
 524         id = read_data[id_idx];
 525
 526  add:
 527         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 528         return 0;
 529 }
 530
 531 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 532                                      struct perf_evsel *evsel, int idx, int cpu,
 533                                      int thread)
 534 {
 535         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 536         sid->idx = idx;
 537         if (evlist->cpus && cpu >= 0)
 538                 sid->cpu = evlist->cpus->map[cpu];
 539         else
 540                 sid->cpu = -1;
 541         if (!evsel->system_wide && evlist->threads && thread >= 0)
 542                 sid->tid = evlist->threads->map[thread];
 543         else
 544                 sid->tid = -1;
 545 }
 546
 547 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 548 {
 549         struct hlist_head *head;
 550         struct perf_sample_id *sid;
 551         int hash;
 552
 553         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 554         head = &evlist->heads[hash];
 555
 556         hlist_for_each_entry(sid, head, node)
 557                 if (sid->id == id)
 558                         return sid;
 559
 560         return NULL;
 561 }
 562
 563 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 564 {
 565         struct perf_sample_id *sid;
 566
 567         if (evlist->nr_entries == 1)
 568                 return perf_evlist__first(evlist);
 569
 570         sid = perf_evlist__id2sid(evlist, id);
 571         if (sid)
 572                 return sid->evsel;
 573
 574         if (!perf_evlist__sample_id_all(evlist))
 575                 return perf_evlist__first(evlist);
 576
 577         return NULL;
 578 }
 579
 580 static int perf_evlist__event2id(struct perf_evlist *evlist,
 581                                  union perf_event *event, u64 *id)
 582 {
 583         const u64 *array = event->sample.array;
 584         ssize_t n;
 585
 586         n = (event->header.size - sizeof(event->header)) >> 3;
 587
 588         if (event->header.type == PERF_RECORD_SAMPLE) {
 589                 if (evlist->id_pos >= n)
 590                         return -1;
 591                 *id = array[evlist->id_pos];
 592         } else {
 593                 if (evlist->is_pos > n)
 594                         return -1;
 595                 n -= evlist->is_pos;
 596                 *id = array[n];
 597         }
 598         return 0;
 599 }
 600
 601 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 602                                                    union perf_event *event)
 603 {
 604         struct perf_evsel *first = perf_evlist__first(evlist);
 605         struct hlist_head *head;
 606         struct perf_sample_id *sid;
 607         int hash;
 608         u64 id;
 609
 610         if (evlist->nr_entries == 1)
 611                 return first;
 612
 613         if (!first->attr.sample_id_all &&
 614             event->header.type != PERF_RECORD_SAMPLE)
 615                 return first;
 616
 617         if (perf_evlist__event2id(evlist, event, &id))
 618                 return NULL;
 619
 620         /* Synthesized events have an id of zero */
 621         if (!id)
 622                 return first;
 623
 624         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 625         head = &evlist->heads[hash];
 626
 627         hlist_for_each_entry(sid, head, node) {
 628                 if (sid->id == id)
 629                         return sid->evsel;
 630         }
 631         return NULL;
 632 }
 633
 634 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 635 {
 636         struct perf_mmap *md = &evlist->mmap[idx];
 637         u64 head = perf_mmap__read_head(md);
 638         u64 old = md->prev;
 639         unsigned char *data = md->base + page_size;
 640         union perf_event *event = NULL;
 641
 642         if (evlist->overwrite) {
 643                 /*
 644                  * If we're further behind than half the buffer, there's a chance
 645                  * the writer will bite our tail and mess up the samples under us.
 646                  *
 647                  * If we somehow ended up ahead of the head, we got messed up.
 648                  *
 649                  * In either case, truncate and restart at head.
 650                  */
 651                 int diff = head - old;
 652                 if (diff > md->mask / 2 || diff < 0) {
 653                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 654
 655                         /*
 656                          * head points to a known good entry, start there.
 657                          */
 658                         old = head;
 659                 }
 660         }
 661
 662         if (old != head) {
 663                 size_t size;
 664
 665                 event = (union perf_event *)&data[old & md->mask];
 666                 size = event->header.size;
 667
 668                 /*
 669                  * Event straddles the mmap boundary -- header should always
 670                  * be inside due to u64 alignment of output.
 671                  */
 672                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 673                         unsigned int offset = old;
 674                         unsigned int len = min(sizeof(*event), size), cpy;
 675                         void *dst = md->event_copy;
 676
 677                         do {
 678                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 679                                 memcpy(dst, &data[offset & md->mask], cpy);
 680                                 offset += cpy;
 681                                 dst += cpy;
 682                                 len -= cpy;
 683                         } while (len);
 684
 685                         event = (union perf_event *) md->event_copy;
 686                 }
 687
 688                 old += size;
 689         }
 690
 691         md->prev = old;
 692
 693         return event;
 694 }
 695
 696 static bool perf_mmap__empty(struct perf_mmap *md)
 697 {
 698         return perf_mmap__read_head(md) == md->prev;
 699 }
 700
 701 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
 702 {
 703         ++evlist->mmap[idx].refcnt;
 704 }
 705
 706 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
 707 {
 708         BUG_ON(evlist->mmap[idx].refcnt == 0);
 709
 710         if (--evlist->mmap[idx].refcnt == 0)
 711                 __perf_evlist__munmap(evlist, idx);
 712 }
 713
 714 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 715 {
 716         struct perf_mmap *md = &evlist->mmap[idx];
 717
 718         if (!evlist->overwrite) {
 719                 u64 old = md->prev;
 720
 721                 perf_mmap__write_tail(md, old);
 722         }
 723
 724         if (md->refcnt == 1 && perf_mmap__empty(md))
 725                 perf_evlist__mmap_put(evlist, idx);
 726 }
 727
 728 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 729 {
 730         if (evlist->mmap[idx].base != NULL) {
 731                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 732                 evlist->mmap[idx].base = NULL;
 733                 evlist->mmap[idx].refcnt = 0;
 734         }
 735 }
 736
 737 void perf_evlist__munmap(struct perf_evlist *evlist)
 738 {
 739         int i;
 740
 741         if (evlist->mmap == NULL)
 742                 return;
 743
 744         for (i = 0; i < evlist->nr_mmaps; i++)
 745                 __perf_evlist__munmap(evlist, i);
 746
 747         zfree(&evlist->mmap);
 748 }
 749
 750 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 751 {
 752         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 753         if (cpu_map__empty(evlist->cpus))
 754                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 755         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 756         return evlist->mmap != NULL ? 0 : -ENOMEM;
 757 }
 758
 759 struct mmap_params {
 760         int prot;
 761         int mask;
 762 };
 763
 764 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 765                                struct mmap_params *mp, int fd)
 766 {
 767         /*
 768          * The last one will be done at perf_evlist__mmap_consume(), so that we
 769          * make sure we don't prevent tools from consuming every last event in
 770          * the ring buffer.
 771          *
 772          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 773          * anymore, but the last events for it are still in the ring buffer,
 774          * waiting to be consumed.
 775          *
 776          * Tools can chose to ignore this at their own discretion, but the
 777          * evlist layer can't just drop it when filtering events in
 778          * perf_evlist__filter_pollfd().
 779          */
 780         evlist->mmap[idx].refcnt = 2;
 781         evlist->mmap[idx].prev = 0;
 782         evlist->mmap[idx].mask = mp->mask;
 783         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 784                                       MAP_SHARED, fd, 0);
 785         if (evlist->mmap[idx].base == MAP_FAILED) {
 786                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 787                           errno);
 788                 evlist->mmap[idx].base = NULL;
 789                 return -1;
 790         }
 791
 792         return 0;
 793 }
 794
 795 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 796                                        struct mmap_params *mp, int cpu,
 797                                        int thread, int *output)
 798 {
 799         struct perf_evsel *evsel;
 800
 801         evlist__for_each(evlist, evsel) {
 802                 int fd;
 803
 804                 if (evsel->system_wide && thread)
 805                         continue;
 806
 807                 fd = FD(evsel, cpu, thread);
 808
 809                 if (*output == -1) {
 810                         *output = fd;
 811                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 812                                 return -1;
 813                 } else {
 814                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 815                                 return -1;
 816
 817                         perf_evlist__mmap_get(evlist, idx);
 818                 }
 819
 820                 /*
 821                  * The system_wide flag causes a selected event to be opened
 822                  * always without a pid.  Consequently it will never get a
 823                  * POLLHUP, but it is used for tracking in combination with
 824                  * other events, so it should not need to be polled anyway.
 825                  * Therefore don't add it for polling.
 826                  */
 827                 if (!evsel->system_wide &&
 828                     __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
 829                         perf_evlist__mmap_put(evlist, idx);
 830                         return -1;
 831                 }
 832
 833                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 834                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 835                                                    fd) < 0)
 836                                 return -1;
 837                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 838                                                  thread);
 839                 }
 840         }
 841
 842         return 0;
 843 }
 844
 845 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 846                                      struct mmap_params *mp)
 847 {
 848         int cpu, thread;
 849         int nr_cpus = cpu_map__nr(evlist->cpus);
 850         int nr_threads = thread_map__nr(evlist->threads);
 851
 852         pr_debug2("perf event ring buffer mmapped per cpu\n");
 853         for (cpu = 0; cpu < nr_cpus; cpu++) {
 854                 int output = -1;
 855
 856                 for (thread = 0; thread < nr_threads; thread++) {
 857                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 858                                                         thread, &output))
 859                                 goto out_unmap;
 860                 }
 861         }
 862
 863         return 0;
 864
 865 out_unmap:
 866         for (cpu = 0; cpu < nr_cpus; cpu++)
 867                 __perf_evlist__munmap(evlist, cpu);
 868         return -1;
 869 }
 870
 871 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 872                                         struct mmap_params *mp)
 873 {
 874         int thread;
 875         int nr_threads = thread_map__nr(evlist->threads);
 876
 877         pr_debug2("perf event ring buffer mmapped per thread\n");
 878         for (thread = 0; thread < nr_threads; thread++) {
 879                 int output = -1;
 880
 881                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 882                                                 &output))
 883                         goto out_unmap;
 884         }
 885
 886         return 0;
 887
 888 out_unmap:
 889         for (thread = 0; thread < nr_threads; thread++)
 890                 __perf_evlist__munmap(evlist, thread);
 891         return -1;
 892 }
 893
 894 static size_t perf_evlist__mmap_size(unsigned long pages)
 895 {
 896         if (pages == UINT_MAX) {
 897                 int max;
 898
 899                 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 900                         /*
 901                          * Pick a once upon a time good value, i.e. things look
 902                          * strange since we can't read a sysctl value, but lets not
 903                          * die yet...
 904                          */
 905                         max = 512;
 906                 } else {
 907                         max -= (page_size / 1024);
 908                 }
 909
 910                 pages = (max * 1024) / page_size;
 911                 if (!is_power_of_2(pages))
 912                         pages = rounddown_pow_of_two(pages);
 913         } else if (!is_power_of_2(pages))
 914                 return 0;
 915
 916         return (pages + 1) * page_size;
 917 }
 918
 919 static long parse_pages_arg(const char *str, unsigned long min,
 920                             unsigned long max)
 921 {
 922         unsigned long pages, val;
 923         static struct parse_tag tags[] = {
 924                 { .tag  = 'B', .mult = 1       },
 925                 { .tag  = 'K', .mult = 1 << 10 },
 926                 { .tag  = 'M', .mult = 1 << 20 },
 927                 { .tag  = 'G', .mult = 1 << 30 },
 928                 { .tag  = 0 },
 929         };
 930
 931         if (str == NULL)
 932                 return -EINVAL;
 933
 934         val = parse_tag_value(str, tags);
 935         if (val != (unsigned long) -1) {
 936                 /* we got file size value */
 937                 pages = PERF_ALIGN(val, page_size) / page_size;
 938         } else {
 939                 /* we got pages count value */
 940                 char *eptr;
 941                 pages = strtoul(str, &eptr, 10);
 942                 if (*eptr != '\0')
 943                         return -EINVAL;
 944         }
 945
 946         if (pages == 0 && min == 0) {
 947                 /* leave number of pages at 0 */
 948         } else if (!is_power_of_2(pages)) {
 949                 /* round pages up to next power of 2 */
 950                 pages = roundup_pow_of_two(pages);
 951                 if (!pages)
 952                         return -EINVAL;
 953                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 954                         pages * page_size, pages);
 955         }
 956
 957         if (pages > max)
 958                 return -EINVAL;
 959
 960         return pages;
 961 }
 962
 963 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 964                                   int unset __maybe_unused)
 965 {
 966         unsigned int *mmap_pages = opt->value;
 967         unsigned long max = UINT_MAX;
 968         long pages;
 969
 970         if (max > SIZE_MAX / page_size)
 971                 max = SIZE_MAX / page_size;
 972
 973         pages = parse_pages_arg(str, 1, max);
 974         if (pages < 0) {
 975                 pr_err("Invalid argument for --mmap_pages/-m\n");
 976                 return -1;
 977         }
 978
 979         *mmap_pages = pages;
 980         return 0;
 981 }
 982
 983 /**
 984  * perf_evlist__mmap - Create mmaps to receive events.
 985  * @evlist: list of events
 986  * @pages: map length in pages
 987  * @overwrite: overwrite older events?
 988  *
 989  * If @overwrite is %false the user needs to signal event consumption using
 990  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 991  * automatically.
 992  *
 993  * Return: %0 on success, negative error code otherwise.
 994  */
 995 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 996                       bool overwrite)
 997 {
 998         struct perf_evsel *evsel;
 999         const struct cpu_map *cpus = evlist->cpus;
1000         const struct thread_map *threads = evlist->threads;
1001         struct mmap_params mp = {
1002                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1003         };
1004
1005         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1006                 return -ENOMEM;
1007
1008         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1009                 return -ENOMEM;
1010
1011         evlist->overwrite = overwrite;
1012         evlist->mmap_len = perf_evlist__mmap_size(pages);
1013         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1014         mp.mask = evlist->mmap_len - page_size - 1;
1015
1016         evlist__for_each(evlist, evsel) {
1017                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1018                     evsel->sample_id == NULL &&
1019                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1020                         return -ENOMEM;
1021         }
1022
1023         if (cpu_map__empty(cpus))
1024                 return perf_evlist__mmap_per_thread(evlist, &mp);
1025
1026         return perf_evlist__mmap_per_cpu(evlist, &mp);
1027 }
1028
1029 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1030 {
1031         evlist->threads = thread_map__new_str(target->pid, target->tid,
1032                                               target->uid);
1033
1034         if (evlist->threads == NULL)
1035                 return -1;
1036
1037         if (target__uses_dummy_map(target))
1038                 evlist->cpus = cpu_map__dummy_new();
1039         else
1040                 evlist->cpus = cpu_map__new(target->cpu_list);
1041
1042         if (evlist->cpus == NULL)
1043                 goto out_delete_threads;
1044
1045         return 0;
1046
1047 out_delete_threads:
1048         thread_map__delete(evlist->threads);
1049         evlist->threads = NULL;
1050         return -1;
1051 }
1052
1053 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1054 {
1055         struct perf_evsel *evsel;
1056         int err = 0;
1057         const int ncpus = cpu_map__nr(evlist->cpus),
1058                   nthreads = thread_map__nr(evlist->threads);
1059
1060         evlist__for_each(evlist, evsel) {
1061                 if (evsel->filter == NULL)
1062                         continue;
1063
1064                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1065                 if (err) {
1066                         *err_evsel = evsel;
1067                         break;
1068                 }
1069         }
1070
1071         return err;
1072 }
1073
1074 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1075 {
1076         struct perf_evsel *evsel;
1077         int err = 0;
1078         const int ncpus = cpu_map__nr(evlist->cpus),
1079                   nthreads = thread_map__nr(evlist->threads);
1080
1081         evlist__for_each(evlist, evsel) {
1082                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1083                 if (err)
1084                         break;
1085         }
1086
1087         return err;
1088 }
1089
1090 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1091 {
1092         char *filter;
1093         int ret = -1;
1094         size_t i;
1095
1096         for (i = 0; i < npids; ++i) {
1097                 if (i == 0) {
1098                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1099                                 return -1;
1100                 } else {
1101                         char *tmp;
1102
1103                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1104                                 goto out_free;
1105
1106                         free(filter);
1107                         filter = tmp;
1108                 }
1109         }
1110
1111         ret = perf_evlist__set_filter(evlist, filter);
1112 out_free:
1113         free(filter);
1114         return ret;
1115 }
1116
1117 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1118 {
1119         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1120 }
1121
1122 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1123 {
1124         struct perf_evsel *pos;
1125
1126         if (evlist->nr_entries == 1)
1127                 return true;
1128
1129         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1130                 return false;
1131
1132         evlist__for_each(evlist, pos) {
1133                 if (pos->id_pos != evlist->id_pos ||
1134                     pos->is_pos != evlist->is_pos)
1135                         return false;
1136         }
1137
1138         return true;
1139 }
1140
1141 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1142 {
1143         struct perf_evsel *evsel;
1144
1145         if (evlist->combined_sample_type)
1146                 return evlist->combined_sample_type;
1147
1148         evlist__for_each(evlist, evsel)
1149                 evlist->combined_sample_type |= evsel->attr.sample_type;
1150
1151         return evlist->combined_sample_type;
1152 }
1153
1154 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1155 {
1156         evlist->combined_sample_type = 0;
1157         return __perf_evlist__combined_sample_type(evlist);
1158 }
1159
1160 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1161 {
1162         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1163         u64 read_format = first->attr.read_format;
1164         u64 sample_type = first->attr.sample_type;
1165
1166         evlist__for_each(evlist, pos) {
1167                 if (read_format != pos->attr.read_format)
1168                         return false;
1169         }
1170
1171         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1172         if ((sample_type & PERF_SAMPLE_READ) &&
1173             !(read_format & PERF_FORMAT_ID)) {
1174                 return false;
1175         }
1176
1177         return true;
1178 }
1179
1180 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1181 {
1182         struct perf_evsel *first = perf_evlist__first(evlist);
1183         return first->attr.read_format;
1184 }
1185
1186 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1187 {
1188         struct perf_evsel *first = perf_evlist__first(evlist);
1189         struct perf_sample *data;
1190         u64 sample_type;
1191         u16 size = 0;
1192
1193         if (!first->attr.sample_id_all)
1194                 goto out;
1195
1196         sample_type = first->attr.sample_type;
1197
1198         if (sample_type & PERF_SAMPLE_TID)
1199                 size += sizeof(data->tid) * 2;
1200
1201        if (sample_type & PERF_SAMPLE_TIME)
1202                 size += sizeof(data->time);
1203
1204         if (sample_type & PERF_SAMPLE_ID)
1205                 size += sizeof(data->id);
1206
1207         if (sample_type & PERF_SAMPLE_STREAM_ID)
1208                 size += sizeof(data->stream_id);
1209
1210         if (sample_type & PERF_SAMPLE_CPU)
1211                 size += sizeof(data->cpu) * 2;
1212
1213         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1214                 size += sizeof(data->id);
1215 out:
1216         return size;
1217 }
1218
1219 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1220 {
1221         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1222
1223         evlist__for_each_continue(evlist, pos) {
1224                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1225                         return false;
1226         }
1227
1228         return true;
1229 }
1230
1231 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1232 {
1233         struct perf_evsel *first = perf_evlist__first(evlist);
1234         return first->attr.sample_id_all;
1235 }
1236
1237 void perf_evlist__set_selected(struct perf_evlist *evlist,
1238                                struct perf_evsel *evsel)
1239 {
1240         evlist->selected = evsel;
1241 }
1242
1243 void perf_evlist__close(struct perf_evlist *evlist)
1244 {
1245         struct perf_evsel *evsel;
1246         int ncpus = cpu_map__nr(evlist->cpus);
1247         int nthreads = thread_map__nr(evlist->threads);
1248         int n;
1249
1250         evlist__for_each_reverse(evlist, evsel) {
1251                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1252                 perf_evsel__close(evsel, n, nthreads);
1253         }
1254 }
1255
1256 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1257 {
1258         int err = -ENOMEM;
1259
1260         /*
1261          * Try reading /sys/devices/system/cpu/online to get
1262          * an all cpus map.
1263          *
1264          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1265          * code needs an overhaul to properly forward the
1266          * error, and we may not want to do that fallback to a
1267          * default cpu identity map :-\
1268          */
1269         evlist->cpus = cpu_map__new(NULL);
1270         if (evlist->cpus == NULL)
1271                 goto out;
1272
1273         evlist->threads = thread_map__new_dummy();
1274         if (evlist->threads == NULL)
1275                 goto out_free_cpus;
1276
1277         err = 0;
1278 out:
1279         return err;
1280 out_free_cpus:
1281         cpu_map__delete(evlist->cpus);
1282         evlist->cpus = NULL;
1283         goto out;
1284 }
1285
1286 int perf_evlist__open(struct perf_evlist *evlist)
1287 {
1288         struct perf_evsel *evsel;
1289         int err;
1290
1291         /*
1292          * Default: one fd per CPU, all threads, aka systemwide
1293          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1294          */
1295         if (evlist->threads == NULL && evlist->cpus == NULL) {
1296                 err = perf_evlist__create_syswide_maps(evlist);
1297                 if (err < 0)
1298                         goto out_err;
1299         }
1300
1301         perf_evlist__update_id_pos(evlist);
1302
1303         evlist__for_each(evlist, evsel) {
1304                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1305                 if (err < 0)
1306                         goto out_err;
1307         }
1308
1309         return 0;
1310 out_err:
1311         perf_evlist__close(evlist);
1312         errno = -err;
1313         return err;
1314 }
1315
1316 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1317                                   const char *argv[], bool pipe_output,
1318                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1319 {
1320         int child_ready_pipe[2], go_pipe[2];
1321         char bf;
1322
1323         if (pipe(child_ready_pipe) < 0) {
1324                 perror("failed to create 'ready' pipe");
1325                 return -1;
1326         }
1327
1328         if (pipe(go_pipe) < 0) {
1329                 perror("failed to create 'go' pipe");
1330                 goto out_close_ready_pipe;
1331         }
1332
1333         evlist->workload.pid = fork();
1334         if (evlist->workload.pid < 0) {
1335                 perror("failed to fork");
1336                 goto out_close_pipes;
1337         }
1338
1339         if (!evlist->workload.pid) {
1340                 int ret;
1341
1342                 if (pipe_output)
1343                         dup2(2, 1);
1344
1345                 signal(SIGTERM, SIG_DFL);
1346
1347                 close(child_ready_pipe[0]);
1348                 close(go_pipe[1]);
1349                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1350
1351                 /*
1352                  * Tell the parent we're ready to go
1353                  */
1354                 close(child_ready_pipe[1]);
1355
1356                 /*
1357                  * Wait until the parent tells us to go.
1358                  */
1359                 ret = read(go_pipe[0], &bf, 1);
1360                 /*
1361                  * The parent will ask for the execvp() to be performed by
1362                  * writing exactly one byte, in workload.cork_fd, usually via
1363                  * perf_evlist__start_workload().
1364                  *
1365                  * For cancelling the workload without actually running it,
1366                  * the parent will just close workload.cork_fd, without writing
1367                  * anything, i.e. read will return zero and we just exit()
1368                  * here.
1369                  */
1370                 if (ret != 1) {
1371                         if (ret == -1)
1372                                 perror("unable to read pipe");
1373                         exit(ret);
1374                 }
1375
1376                 execvp(argv[0], (char **)argv);
1377
1378                 if (exec_error) {
1379                         union sigval val;
1380
1381                         val.sival_int = errno;
1382                         if (sigqueue(getppid(), SIGUSR1, val))
1383                                 perror(argv[0]);
1384                 } else
1385                         perror(argv[0]);
1386                 exit(-1);
1387         }
1388
1389         if (exec_error) {
1390                 struct sigaction act = {
1391                         .sa_flags     = SA_SIGINFO,
1392                         .sa_sigaction = exec_error,
1393                 };
1394                 sigaction(SIGUSR1, &act, NULL);
1395         }
1396
1397         if (target__none(target)) {
1398                 if (evlist->threads == NULL) {
1399                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1400                                 __func__, __LINE__);
1401                         goto out_close_pipes;
1402                 }
1403                 evlist->threads->map[0] = evlist->workload.pid;
1404         }
1405
1406         close(child_ready_pipe[1]);
1407         close(go_pipe[0]);
1408         /*
1409          * wait for child to settle
1410          */
1411         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1412                 perror("unable to read pipe");
1413                 goto out_close_pipes;
1414         }
1415
1416         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1417         evlist->workload.cork_fd = go_pipe[1];
1418         close(child_ready_pipe[0]);
1419         return 0;
1420
1421 out_close_pipes:
1422         close(go_pipe[0]);
1423         close(go_pipe[1]);
1424 out_close_ready_pipe:
1425         close(child_ready_pipe[0]);
1426         close(child_ready_pipe[1]);
1427         return -1;
1428 }
1429
1430 int perf_evlist__start_workload(struct perf_evlist *evlist)
1431 {
1432         if (evlist->workload.cork_fd > 0) {
1433                 char bf = 0;
1434                 int ret;
1435                 /*
1436                  * Remove the cork, let it rip!
1437                  */
1438                 ret = write(evlist->workload.cork_fd, &bf, 1);
1439                 if (ret < 0)
1440                         perror("enable to write to pipe");
1441
1442                 close(evlist->workload.cork_fd);
1443                 return ret;
1444         }
1445
1446         return 0;
1447 }
1448
1449 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1450                               struct perf_sample *sample)
1451 {
1452         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1453
1454         if (!evsel)
1455                 return -EFAULT;
1456         return perf_evsel__parse_sample(evsel, event, sample);
1457 }
1458
1459 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1460 {
1461         struct perf_evsel *evsel;
1462         size_t printed = 0;
1463
1464         evlist__for_each(evlist, evsel) {
1465                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1466                                    perf_evsel__name(evsel));
1467         }
1468
1469         return printed + fprintf(fp, "\n");
1470 }
1471
1472 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1473                                int err, char *buf, size_t size)
1474 {
1475         int printed, value;
1476         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1477
1478         switch (err) {
1479         case EACCES:
1480         case EPERM:
1481                 printed = scnprintf(buf, size,
1482                                     "Error:\t%s.\n"
1483                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1484
1485                 value = perf_event_paranoid();
1486
1487                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1488
1489                 if (value >= 2) {
1490                         printed += scnprintf(buf + printed, size - printed,
1491                                              "For your workloads it needs to be <= 1\nHint:\t");
1492                 }
1493                 printed += scnprintf(buf + printed, size - printed,
1494                                      "For system wide tracing it needs to be set to -1.\n");
1495
1496                 printed += scnprintf(buf + printed, size - printed,
1497                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1498                                     "Hint:\tThe current value is %d.", value);
1499                 break;
1500         default:
1501                 scnprintf(buf, size, "%s", emsg);
1502                 break;
1503         }
1504
1505         return 0;
1506 }
1507
1508 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1509 {
1510         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1511         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1512
1513         switch (err) {
1514         case EPERM:
1515                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1516                 printed += scnprintf(buf + printed, size - printed,
1517                                      "Error:\t%s.\n"
1518                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1519                                      "Hint:\tTried using %zd kB.\n",
1520                                      emsg, pages_max_per_user, pages_attempted);
1521
1522                 if (pages_attempted >= pages_max_per_user) {
1523                         printed += scnprintf(buf + printed, size - printed,
1524                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1525                                              pages_max_per_user + pages_attempted);
1526                 }
1527
1528                 printed += scnprintf(buf + printed, size - printed,
1529                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1530                 break;
1531         default:
1532                 scnprintf(buf, size, "%s", emsg);
1533                 break;
1534         }
1535
1536         return 0;
1537 }
1538
1539 void perf_evlist__to_front(struct perf_evlist *evlist,
1540                            struct perf_evsel *move_evsel)
1541 {
1542         struct perf_evsel *evsel, *n;
1543         LIST_HEAD(move);
1544
1545         if (move_evsel == perf_evlist__first(evlist))
1546                 return;
1547
1548         evlist__for_each_safe(evlist, n, evsel) {
1549                 if (evsel->leader == move_evsel->leader)
1550                         list_move_tail(&evsel->node, &move);
1551         }
1552
1553         list_splice(&move, &evlist->entries);
1554 }
1555
1556 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1557                                      struct perf_evsel *tracking_evsel)
1558 {
1559         struct perf_evsel *evsel;
1560
1561         if (tracking_evsel->tracking)
1562                 return;
1563
1564         evlist__for_each(evlist, evsel) {
1565                 if (evsel != tracking_evsel)
1566                         evsel->tracking = false;
1567         }
1568
1569         tracking_evsel->tracking = true;
1570 }