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 && !md->auxtrace_mmap.base;
 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 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 729                                struct auxtrace_mmap_params *mp __maybe_unused,
 730                                void *userpg __maybe_unused,
 731                                int fd __maybe_unused)
 732 {
 733         return 0;
 734 }
 735
 736 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 737 {
 738 }
 739
 740 void __weak auxtrace_mmap_params__init(
 741                         struct auxtrace_mmap_params *mp __maybe_unused,
 742                         off_t auxtrace_offset __maybe_unused,
 743                         unsigned int auxtrace_pages __maybe_unused,
 744                         bool auxtrace_overwrite __maybe_unused)
 745 {
 746 }
 747
 748 void __weak auxtrace_mmap_params__set_idx(
 749                         struct auxtrace_mmap_params *mp __maybe_unused,
 750                         struct perf_evlist *evlist __maybe_unused,
 751                         int idx __maybe_unused,
 752                         bool per_cpu __maybe_unused)
 753 {
 754 }
 755
 756 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 757 {
 758         if (evlist->mmap[idx].base != NULL) {
 759                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 760                 evlist->mmap[idx].base = NULL;
 761                 evlist->mmap[idx].refcnt = 0;
 762         }
 763         auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
 764 }
 765
 766 void perf_evlist__munmap(struct perf_evlist *evlist)
 767 {
 768         int i;
 769
 770         if (evlist->mmap == NULL)
 771                 return;
 772
 773         for (i = 0; i < evlist->nr_mmaps; i++)
 774                 __perf_evlist__munmap(evlist, i);
 775
 776         zfree(&evlist->mmap);
 777 }
 778
 779 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 780 {
 781         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 782         if (cpu_map__empty(evlist->cpus))
 783                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 784         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 785         return evlist->mmap != NULL ? 0 : -ENOMEM;
 786 }
 787
 788 struct mmap_params {
 789         int prot;
 790         int mask;
 791         struct auxtrace_mmap_params auxtrace_mp;
 792 };
 793
 794 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 795                                struct mmap_params *mp, int fd)
 796 {
 797         /*
 798          * The last one will be done at perf_evlist__mmap_consume(), so that we
 799          * make sure we don't prevent tools from consuming every last event in
 800          * the ring buffer.
 801          *
 802          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 803          * anymore, but the last events for it are still in the ring buffer,
 804          * waiting to be consumed.
 805          *
 806          * Tools can chose to ignore this at their own discretion, but the
 807          * evlist layer can't just drop it when filtering events in
 808          * perf_evlist__filter_pollfd().
 809          */
 810         evlist->mmap[idx].refcnt = 2;
 811         evlist->mmap[idx].prev = 0;
 812         evlist->mmap[idx].mask = mp->mask;
 813         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 814                                       MAP_SHARED, fd, 0);
 815         if (evlist->mmap[idx].base == MAP_FAILED) {
 816                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 817                           errno);
 818                 evlist->mmap[idx].base = NULL;
 819                 return -1;
 820         }
 821
 822         if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
 823                                 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
 824                 return -1;
 825
 826         return 0;
 827 }
 828
 829 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 830                                        struct mmap_params *mp, int cpu,
 831                                        int thread, int *output)
 832 {
 833         struct perf_evsel *evsel;
 834
 835         evlist__for_each(evlist, evsel) {
 836                 int fd;
 837
 838                 if (evsel->system_wide && thread)
 839                         continue;
 840
 841                 fd = FD(evsel, cpu, thread);
 842
 843                 if (*output == -1) {
 844                         *output = fd;
 845                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 846                                 return -1;
 847                 } else {
 848                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 849                                 return -1;
 850
 851                         perf_evlist__mmap_get(evlist, idx);
 852                 }
 853
 854                 /*
 855                  * The system_wide flag causes a selected event to be opened
 856                  * always without a pid.  Consequently it will never get a
 857                  * POLLHUP, but it is used for tracking in combination with
 858                  * other events, so it should not need to be polled anyway.
 859                  * Therefore don't add it for polling.
 860                  */
 861                 if (!evsel->system_wide &&
 862                     __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
 863                         perf_evlist__mmap_put(evlist, idx);
 864                         return -1;
 865                 }
 866
 867                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 868                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 869                                                    fd) < 0)
 870                                 return -1;
 871                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 872                                                  thread);
 873                 }
 874         }
 875
 876         return 0;
 877 }
 878
 879 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 880                                      struct mmap_params *mp)
 881 {
 882         int cpu, thread;
 883         int nr_cpus = cpu_map__nr(evlist->cpus);
 884         int nr_threads = thread_map__nr(evlist->threads);
 885
 886         pr_debug2("perf event ring buffer mmapped per cpu\n");
 887         for (cpu = 0; cpu < nr_cpus; cpu++) {
 888                 int output = -1;
 889
 890                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 891                                               true);
 892
 893                 for (thread = 0; thread < nr_threads; thread++) {
 894                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 895                                                         thread, &output))
 896                                 goto out_unmap;
 897                 }
 898         }
 899
 900         return 0;
 901
 902 out_unmap:
 903         for (cpu = 0; cpu < nr_cpus; cpu++)
 904                 __perf_evlist__munmap(evlist, cpu);
 905         return -1;
 906 }
 907
 908 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 909                                         struct mmap_params *mp)
 910 {
 911         int thread;
 912         int nr_threads = thread_map__nr(evlist->threads);
 913
 914         pr_debug2("perf event ring buffer mmapped per thread\n");
 915         for (thread = 0; thread < nr_threads; thread++) {
 916                 int output = -1;
 917
 918                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 919                                               false);
 920
 921                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 922                                                 &output))
 923                         goto out_unmap;
 924         }
 925
 926         return 0;
 927
 928 out_unmap:
 929         for (thread = 0; thread < nr_threads; thread++)
 930                 __perf_evlist__munmap(evlist, thread);
 931         return -1;
 932 }
 933
 934 static size_t perf_evlist__mmap_size(unsigned long pages)
 935 {
 936         if (pages == UINT_MAX) {
 937                 int max;
 938
 939                 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 940                         /*
 941                          * Pick a once upon a time good value, i.e. things look
 942                          * strange since we can't read a sysctl value, but lets not
 943                          * die yet...
 944                          */
 945                         max = 512;
 946                 } else {
 947                         max -= (page_size / 1024);
 948                 }
 949
 950                 pages = (max * 1024) / page_size;
 951                 if (!is_power_of_2(pages))
 952                         pages = rounddown_pow_of_two(pages);
 953         } else if (!is_power_of_2(pages))
 954                 return 0;
 955
 956         return (pages + 1) * page_size;
 957 }
 958
 959 static long parse_pages_arg(const char *str, unsigned long min,
 960                             unsigned long max)
 961 {
 962         unsigned long pages, val;
 963         static struct parse_tag tags[] = {
 964                 { .tag  = 'B', .mult = 1       },
 965                 { .tag  = 'K', .mult = 1 << 10 },
 966                 { .tag  = 'M', .mult = 1 << 20 },
 967                 { .tag  = 'G', .mult = 1 << 30 },
 968                 { .tag  = 0 },
 969         };
 970
 971         if (str == NULL)
 972                 return -EINVAL;
 973
 974         val = parse_tag_value(str, tags);
 975         if (val != (unsigned long) -1) {
 976                 /* we got file size value */
 977                 pages = PERF_ALIGN(val, page_size) / page_size;
 978         } else {
 979                 /* we got pages count value */
 980                 char *eptr;
 981                 pages = strtoul(str, &eptr, 10);
 982                 if (*eptr != '\0')
 983                         return -EINVAL;
 984         }
 985
 986         if (pages == 0 && min == 0) {
 987                 /* leave number of pages at 0 */
 988         } else if (!is_power_of_2(pages)) {
 989                 /* round pages up to next power of 2 */
 990                 pages = roundup_pow_of_two(pages);
 991                 if (!pages)
 992                         return -EINVAL;
 993                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 994                         pages * page_size, pages);
 995         }
 996
 997         if (pages > max)
 998                 return -EINVAL;
 999
1000         return pages;
1001 }
1002
1003 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1004 {
1005         unsigned long max = UINT_MAX;
1006         long pages;
1007
1008         if (max > SIZE_MAX / page_size)
1009                 max = SIZE_MAX / page_size;
1010
1011         pages = parse_pages_arg(str, 1, max);
1012         if (pages < 0) {
1013                 pr_err("Invalid argument for --mmap_pages/-m\n");
1014                 return -1;
1015         }
1016
1017         *mmap_pages = pages;
1018         return 0;
1019 }
1020
1021 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1022                                   int unset __maybe_unused)
1023 {
1024         return __perf_evlist__parse_mmap_pages(opt->value, str);
1025 }
1026
1027 /**
1028  * perf_evlist__mmap_ex - Create mmaps to receive events.
1029  * @evlist: list of events
1030  * @pages: map length in pages
1031  * @overwrite: overwrite older events?
1032  * @auxtrace_pages - auxtrace map length in pages
1033  * @auxtrace_overwrite - overwrite older auxtrace data?
1034  *
1035  * If @overwrite is %false the user needs to signal event consumption using
1036  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1037  * automatically.
1038  *
1039  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1040  * consumption using auxtrace_mmap__write_tail().
1041  *
1042  * Return: %0 on success, negative error code otherwise.
1043  */
1044 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1045                          bool overwrite, unsigned int auxtrace_pages,
1046                          bool auxtrace_overwrite)
1047 {
1048         struct perf_evsel *evsel;
1049         const struct cpu_map *cpus = evlist->cpus;
1050         const struct thread_map *threads = evlist->threads;
1051         struct mmap_params mp = {
1052                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1053         };
1054
1055         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1056                 return -ENOMEM;
1057
1058         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1059                 return -ENOMEM;
1060
1061         evlist->overwrite = overwrite;
1062         evlist->mmap_len = perf_evlist__mmap_size(pages);
1063         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1064         mp.mask = evlist->mmap_len - page_size - 1;
1065
1066         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1067                                    auxtrace_pages, auxtrace_overwrite);
1068
1069         evlist__for_each(evlist, evsel) {
1070                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1071                     evsel->sample_id == NULL &&
1072                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1073                         return -ENOMEM;
1074         }
1075
1076         if (cpu_map__empty(cpus))
1077                 return perf_evlist__mmap_per_thread(evlist, &mp);
1078
1079         return perf_evlist__mmap_per_cpu(evlist, &mp);
1080 }
1081
1082 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1083                       bool overwrite)
1084 {
1085         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1086 }
1087
1088 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1089 {
1090         evlist->threads = thread_map__new_str(target->pid, target->tid,
1091                                               target->uid);
1092
1093         if (evlist->threads == NULL)
1094                 return -1;
1095
1096         if (target__uses_dummy_map(target))
1097                 evlist->cpus = cpu_map__dummy_new();
1098         else
1099                 evlist->cpus = cpu_map__new(target->cpu_list);
1100
1101         if (evlist->cpus == NULL)
1102                 goto out_delete_threads;
1103
1104         return 0;
1105
1106 out_delete_threads:
1107         thread_map__delete(evlist->threads);
1108         evlist->threads = NULL;
1109         return -1;
1110 }
1111
1112 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1113 {
1114         struct perf_evsel *evsel;
1115         int err = 0;
1116         const int ncpus = cpu_map__nr(evlist->cpus),
1117                   nthreads = thread_map__nr(evlist->threads);
1118
1119         evlist__for_each(evlist, evsel) {
1120                 if (evsel->filter == NULL)
1121                         continue;
1122
1123                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1124                 if (err) {
1125                         *err_evsel = evsel;
1126                         break;
1127                 }
1128         }
1129
1130         return err;
1131 }
1132
1133 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1134 {
1135         struct perf_evsel *evsel;
1136         int err = 0;
1137         const int ncpus = cpu_map__nr(evlist->cpus),
1138                   nthreads = thread_map__nr(evlist->threads);
1139
1140         evlist__for_each(evlist, evsel) {
1141                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1142                 if (err)
1143                         break;
1144         }
1145
1146         return err;
1147 }
1148
1149 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1150 {
1151         char *filter;
1152         int ret = -1;
1153         size_t i;
1154
1155         for (i = 0; i < npids; ++i) {
1156                 if (i == 0) {
1157                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1158                                 return -1;
1159                 } else {
1160                         char *tmp;
1161
1162                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1163                                 goto out_free;
1164
1165                         free(filter);
1166                         filter = tmp;
1167                 }
1168         }
1169
1170         ret = perf_evlist__set_filter(evlist, filter);
1171 out_free:
1172         free(filter);
1173         return ret;
1174 }
1175
1176 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1177 {
1178         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1179 }
1180
1181 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1182 {
1183         struct perf_evsel *pos;
1184
1185         if (evlist->nr_entries == 1)
1186                 return true;
1187
1188         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1189                 return false;
1190
1191         evlist__for_each(evlist, pos) {
1192                 if (pos->id_pos != evlist->id_pos ||
1193                     pos->is_pos != evlist->is_pos)
1194                         return false;
1195         }
1196
1197         return true;
1198 }
1199
1200 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1201 {
1202         struct perf_evsel *evsel;
1203
1204         if (evlist->combined_sample_type)
1205                 return evlist->combined_sample_type;
1206
1207         evlist__for_each(evlist, evsel)
1208                 evlist->combined_sample_type |= evsel->attr.sample_type;
1209
1210         return evlist->combined_sample_type;
1211 }
1212
1213 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1214 {
1215         evlist->combined_sample_type = 0;
1216         return __perf_evlist__combined_sample_type(evlist);
1217 }
1218
1219 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1220 {
1221         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1222         u64 read_format = first->attr.read_format;
1223         u64 sample_type = first->attr.sample_type;
1224
1225         evlist__for_each(evlist, pos) {
1226                 if (read_format != pos->attr.read_format)
1227                         return false;
1228         }
1229
1230         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1231         if ((sample_type & PERF_SAMPLE_READ) &&
1232             !(read_format & PERF_FORMAT_ID)) {
1233                 return false;
1234         }
1235
1236         return true;
1237 }
1238
1239 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1240 {
1241         struct perf_evsel *first = perf_evlist__first(evlist);
1242         return first->attr.read_format;
1243 }
1244
1245 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1246 {
1247         struct perf_evsel *first = perf_evlist__first(evlist);
1248         struct perf_sample *data;
1249         u64 sample_type;
1250         u16 size = 0;
1251
1252         if (!first->attr.sample_id_all)
1253                 goto out;
1254
1255         sample_type = first->attr.sample_type;
1256
1257         if (sample_type & PERF_SAMPLE_TID)
1258                 size += sizeof(data->tid) * 2;
1259
1260        if (sample_type & PERF_SAMPLE_TIME)
1261                 size += sizeof(data->time);
1262
1263         if (sample_type & PERF_SAMPLE_ID)
1264                 size += sizeof(data->id);
1265
1266         if (sample_type & PERF_SAMPLE_STREAM_ID)
1267                 size += sizeof(data->stream_id);
1268
1269         if (sample_type & PERF_SAMPLE_CPU)
1270                 size += sizeof(data->cpu) * 2;
1271
1272         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1273                 size += sizeof(data->id);
1274 out:
1275         return size;
1276 }
1277
1278 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1279 {
1280         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1281
1282         evlist__for_each_continue(evlist, pos) {
1283                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1284                         return false;
1285         }
1286
1287         return true;
1288 }
1289
1290 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1291 {
1292         struct perf_evsel *first = perf_evlist__first(evlist);
1293         return first->attr.sample_id_all;
1294 }
1295
1296 void perf_evlist__set_selected(struct perf_evlist *evlist,
1297                                struct perf_evsel *evsel)
1298 {
1299         evlist->selected = evsel;
1300 }
1301
1302 void perf_evlist__close(struct perf_evlist *evlist)
1303 {
1304         struct perf_evsel *evsel;
1305         int ncpus = cpu_map__nr(evlist->cpus);
1306         int nthreads = thread_map__nr(evlist->threads);
1307         int n;
1308
1309         evlist__for_each_reverse(evlist, evsel) {
1310                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1311                 perf_evsel__close(evsel, n, nthreads);
1312         }
1313 }
1314
1315 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1316 {
1317         int err = -ENOMEM;
1318
1319         /*
1320          * Try reading /sys/devices/system/cpu/online to get
1321          * an all cpus map.
1322          *
1323          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1324          * code needs an overhaul to properly forward the
1325          * error, and we may not want to do that fallback to a
1326          * default cpu identity map :-\
1327          */
1328         evlist->cpus = cpu_map__new(NULL);
1329         if (evlist->cpus == NULL)
1330                 goto out;
1331
1332         evlist->threads = thread_map__new_dummy();
1333         if (evlist->threads == NULL)
1334                 goto out_free_cpus;
1335
1336         err = 0;
1337 out:
1338         return err;
1339 out_free_cpus:
1340         cpu_map__delete(evlist->cpus);
1341         evlist->cpus = NULL;
1342         goto out;
1343 }
1344
1345 int perf_evlist__open(struct perf_evlist *evlist)
1346 {
1347         struct perf_evsel *evsel;
1348         int err;
1349
1350         /*
1351          * Default: one fd per CPU, all threads, aka systemwide
1352          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1353          */
1354         if (evlist->threads == NULL && evlist->cpus == NULL) {
1355                 err = perf_evlist__create_syswide_maps(evlist);
1356                 if (err < 0)
1357                         goto out_err;
1358         }
1359
1360         perf_evlist__update_id_pos(evlist);
1361
1362         evlist__for_each(evlist, evsel) {
1363                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1364                 if (err < 0)
1365                         goto out_err;
1366         }
1367
1368         return 0;
1369 out_err:
1370         perf_evlist__close(evlist);
1371         errno = -err;
1372         return err;
1373 }
1374
1375 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1376                                   const char *argv[], bool pipe_output,
1377                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1378 {
1379         int child_ready_pipe[2], go_pipe[2];
1380         char bf;
1381
1382         if (pipe(child_ready_pipe) < 0) {
1383                 perror("failed to create 'ready' pipe");
1384                 return -1;
1385         }
1386
1387         if (pipe(go_pipe) < 0) {
1388                 perror("failed to create 'go' pipe");
1389                 goto out_close_ready_pipe;
1390         }
1391
1392         evlist->workload.pid = fork();
1393         if (evlist->workload.pid < 0) {
1394                 perror("failed to fork");
1395                 goto out_close_pipes;
1396         }
1397
1398         if (!evlist->workload.pid) {
1399                 int ret;
1400
1401                 if (pipe_output)
1402                         dup2(2, 1);
1403
1404                 signal(SIGTERM, SIG_DFL);
1405
1406                 close(child_ready_pipe[0]);
1407                 close(go_pipe[1]);
1408                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1409
1410                 /*
1411                  * Tell the parent we're ready to go
1412                  */
1413                 close(child_ready_pipe[1]);
1414
1415                 /*
1416                  * Wait until the parent tells us to go.
1417                  */
1418                 ret = read(go_pipe[0], &bf, 1);
1419                 /*
1420                  * The parent will ask for the execvp() to be performed by
1421                  * writing exactly one byte, in workload.cork_fd, usually via
1422                  * perf_evlist__start_workload().
1423                  *
1424                  * For cancelling the workload without actually running it,
1425                  * the parent will just close workload.cork_fd, without writing
1426                  * anything, i.e. read will return zero and we just exit()
1427                  * here.
1428                  */
1429                 if (ret != 1) {
1430                         if (ret == -1)
1431                                 perror("unable to read pipe");
1432                         exit(ret);
1433                 }
1434
1435                 execvp(argv[0], (char **)argv);
1436
1437                 if (exec_error) {
1438                         union sigval val;
1439
1440                         val.sival_int = errno;
1441                         if (sigqueue(getppid(), SIGUSR1, val))
1442                                 perror(argv[0]);
1443                 } else
1444                         perror(argv[0]);
1445                 exit(-1);
1446         }
1447
1448         if (exec_error) {
1449                 struct sigaction act = {
1450                         .sa_flags     = SA_SIGINFO,
1451                         .sa_sigaction = exec_error,
1452                 };
1453                 sigaction(SIGUSR1, &act, NULL);
1454         }
1455
1456         if (target__none(target)) {
1457                 if (evlist->threads == NULL) {
1458                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1459                                 __func__, __LINE__);
1460                         goto out_close_pipes;
1461                 }
1462                 evlist->threads->map[0] = evlist->workload.pid;
1463         }
1464
1465         close(child_ready_pipe[1]);
1466         close(go_pipe[0]);
1467         /*
1468          * wait for child to settle
1469          */
1470         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1471                 perror("unable to read pipe");
1472                 goto out_close_pipes;
1473         }
1474
1475         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1476         evlist->workload.cork_fd = go_pipe[1];
1477         close(child_ready_pipe[0]);
1478         return 0;
1479
1480 out_close_pipes:
1481         close(go_pipe[0]);
1482         close(go_pipe[1]);
1483 out_close_ready_pipe:
1484         close(child_ready_pipe[0]);
1485         close(child_ready_pipe[1]);
1486         return -1;
1487 }
1488
1489 int perf_evlist__start_workload(struct perf_evlist *evlist)
1490 {
1491         if (evlist->workload.cork_fd > 0) {
1492                 char bf = 0;
1493                 int ret;
1494                 /*
1495                  * Remove the cork, let it rip!
1496                  */
1497                 ret = write(evlist->workload.cork_fd, &bf, 1);
1498                 if (ret < 0)
1499                         perror("enable to write to pipe");
1500
1501                 close(evlist->workload.cork_fd);
1502                 return ret;
1503         }
1504
1505         return 0;
1506 }
1507
1508 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1509                               struct perf_sample *sample)
1510 {
1511         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1512
1513         if (!evsel)
1514                 return -EFAULT;
1515         return perf_evsel__parse_sample(evsel, event, sample);
1516 }
1517
1518 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1519 {
1520         struct perf_evsel *evsel;
1521         size_t printed = 0;
1522
1523         evlist__for_each(evlist, evsel) {
1524                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1525                                    perf_evsel__name(evsel));
1526         }
1527
1528         return printed + fprintf(fp, "\n");
1529 }
1530
1531 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1532                                int err, char *buf, size_t size)
1533 {
1534         int printed, value;
1535         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1536
1537         switch (err) {
1538         case EACCES:
1539         case EPERM:
1540                 printed = scnprintf(buf, size,
1541                                     "Error:\t%s.\n"
1542                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1543
1544                 value = perf_event_paranoid();
1545
1546                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1547
1548                 if (value >= 2) {
1549                         printed += scnprintf(buf + printed, size - printed,
1550                                              "For your workloads it needs to be <= 1\nHint:\t");
1551                 }
1552                 printed += scnprintf(buf + printed, size - printed,
1553                                      "For system wide tracing it needs to be set to -1.\n");
1554
1555                 printed += scnprintf(buf + printed, size - printed,
1556                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1557                                     "Hint:\tThe current value is %d.", value);
1558                 break;
1559         default:
1560                 scnprintf(buf, size, "%s", emsg);
1561                 break;
1562         }
1563
1564         return 0;
1565 }
1566
1567 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1568 {
1569         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1570         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1571
1572         switch (err) {
1573         case EPERM:
1574                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1575                 printed += scnprintf(buf + printed, size - printed,
1576                                      "Error:\t%s.\n"
1577                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1578                                      "Hint:\tTried using %zd kB.\n",
1579                                      emsg, pages_max_per_user, pages_attempted);
1580
1581                 if (pages_attempted >= pages_max_per_user) {
1582                         printed += scnprintf(buf + printed, size - printed,
1583                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1584                                              pages_max_per_user + pages_attempted);
1585                 }
1586
1587                 printed += scnprintf(buf + printed, size - printed,
1588                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1589                 break;
1590         default:
1591                 scnprintf(buf, size, "%s", emsg);
1592                 break;
1593         }
1594
1595         return 0;
1596 }
1597
1598 void perf_evlist__to_front(struct perf_evlist *evlist,
1599                            struct perf_evsel *move_evsel)
1600 {
1601         struct perf_evsel *evsel, *n;
1602         LIST_HEAD(move);
1603
1604         if (move_evsel == perf_evlist__first(evlist))
1605                 return;
1606
1607         evlist__for_each_safe(evlist, n, evsel) {
1608                 if (evsel->leader == move_evsel->leader)
1609                         list_move_tail(&evsel->node, &move);
1610         }
1611
1612         list_splice(&move, &evlist->entries);
1613 }
1614
1615 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1616                                      struct perf_evsel *tracking_evsel)
1617 {
1618         struct perf_evsel *evsel;
1619
1620         if (tracking_evsel->tracking)
1621                 return;
1622
1623         evlist__for_each(evlist, evsel) {
1624                 if (evsel != tracking_evsel)
1625                         evsel->tracking = false;
1626         }
1627
1628         tracking_evsel->tracking = true;
1629 }