tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <poll.h>
  14 #include "cpumap.h"
  15 #include "thread_map.h"
  16 #include "target.h"
  17 #include "evlist.h"
  18 #include "evsel.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "asm/bug.h"
  22 #include <signal.h>
  23 #include <unistd.h>
  24
  25 #include "parse-events.h"
  26 #include <subcmd/parse-options.h>
  27
  28 #include <fcntl.h>
  29 #include <sys/ioctl.h>
  30 #include <sys/mman.h>
  31
  32 #include <linux/bitops.h>
  33 #include <linux/hash.h>
  34 #include <linux/log2.h>
  35 #include <linux/err.h>
  36
  37 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  38 int sigqueue(pid_t pid, int sig, const union sigval value);
  39 #endif
  40
  41 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  42 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  43
  44 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  45                        struct thread_map *threads)
  46 {
  47         int i;
  48
  49         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  50                 INIT_HLIST_HEAD(&evlist->heads[i]);
  51         INIT_LIST_HEAD(&evlist->entries);
  52         perf_evlist__set_maps(evlist, cpus, threads);
  53         fdarray__init(&evlist->pollfd, 64);
  54         evlist->workload.pid = -1;
  55         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  56 }
  57
  58 struct perf_evlist *perf_evlist__new(void)
  59 {
  60         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  61
  62         if (evlist != NULL)
  63                 perf_evlist__init(evlist, NULL, NULL);
  64
  65         return evlist;
  66 }
  67
  68 struct perf_evlist *perf_evlist__new_default(void)
  69 {
  70         struct perf_evlist *evlist = perf_evlist__new();
  71
  72         if (evlist && perf_evlist__add_default(evlist)) {
  73                 perf_evlist__delete(evlist);
  74                 evlist = NULL;
  75         }
  76
  77         return evlist;
  78 }
  79
  80 struct perf_evlist *perf_evlist__new_dummy(void)
  81 {
  82         struct perf_evlist *evlist = perf_evlist__new();
  83
  84         if (evlist && perf_evlist__add_dummy(evlist)) {
  85                 perf_evlist__delete(evlist);
  86                 evlist = NULL;
  87         }
  88
  89         return evlist;
  90 }
  91
  92 /**
  93  * perf_evlist__set_id_pos - set the positions of event ids.
  94  * @evlist: selected event list
  95  *
  96  * Events with compatible sample types all have the same id_pos
  97  * and is_pos.  For convenience, put a copy on evlist.
  98  */
  99 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
 100 {
 101         struct perf_evsel *first = perf_evlist__first(evlist);
 102
 103         evlist->id_pos = first->id_pos;
 104         evlist->is_pos = first->is_pos;
 105 }
 106
 107 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 108 {
 109         struct perf_evsel *evsel;
 110
 111         evlist__for_each_entry(evlist, evsel)
 112                 perf_evsel__calc_id_pos(evsel);
 113
 114         perf_evlist__set_id_pos(evlist);
 115 }
 116
 117 static void perf_evlist__purge(struct perf_evlist *evlist)
 118 {
 119         struct perf_evsel *pos, *n;
 120
 121         evlist__for_each_entry_safe(evlist, n, pos) {
 122                 list_del_init(&pos->node);
 123                 pos->evlist = NULL;
 124                 perf_evsel__delete(pos);
 125         }
 126
 127         evlist->nr_entries = 0;
 128 }
 129
 130 void perf_evlist__exit(struct perf_evlist *evlist)
 131 {
 132         zfree(&evlist->mmap);
 133         zfree(&evlist->overwrite_mmap);
 134         fdarray__exit(&evlist->pollfd);
 135 }
 136
 137 void perf_evlist__delete(struct perf_evlist *evlist)
 138 {
 139         if (evlist == NULL)
 140                 return;
 141
 142         perf_evlist__munmap(evlist);
 143         perf_evlist__close(evlist);
 144         cpu_map__put(evlist->cpus);
 145         thread_map__put(evlist->threads);
 146         evlist->cpus = NULL;
 147         evlist->threads = NULL;
 148         perf_evlist__purge(evlist);
 149         perf_evlist__exit(evlist);
 150         free(evlist);
 151 }
 152
 153 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 154                                           struct perf_evsel *evsel)
 155 {
 156         /*
 157          * We already have cpus for evsel (via PMU sysfs) so
 158          * keep it, if there's no target cpu list defined.
 159          */
 160         if (!evsel->own_cpus || evlist->has_user_cpus) {
 161                 cpu_map__put(evsel->cpus);
 162                 evsel->cpus = cpu_map__get(evlist->cpus);
 163         } else if (evsel->cpus != evsel->own_cpus) {
 164                 cpu_map__put(evsel->cpus);
 165                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 166         }
 167
 168         thread_map__put(evsel->threads);
 169         evsel->threads = thread_map__get(evlist->threads);
 170 }
 171
 172 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 173 {
 174         struct perf_evsel *evsel;
 175
 176         evlist__for_each_entry(evlist, evsel)
 177                 __perf_evlist__propagate_maps(evlist, evsel);
 178 }
 179
 180 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 181 {
 182         entry->evlist = evlist;
 183         list_add_tail(&entry->node, &evlist->entries);
 184         entry->idx = evlist->nr_entries;
 185         entry->tracking = !entry->idx;
 186
 187         if (!evlist->nr_entries++)
 188                 perf_evlist__set_id_pos(evlist);
 189
 190         __perf_evlist__propagate_maps(evlist, entry);
 191 }
 192
 193 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 194 {
 195         evsel->evlist = NULL;
 196         list_del_init(&evsel->node);
 197         evlist->nr_entries -= 1;
 198 }
 199
 200 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 201                                    struct list_head *list)
 202 {
 203         struct perf_evsel *evsel, *temp;
 204
 205         __evlist__for_each_entry_safe(list, temp, evsel) {
 206                 list_del_init(&evsel->node);
 207                 perf_evlist__add(evlist, evsel);
 208         }
 209 }
 210
 211 void __perf_evlist__set_leader(struct list_head *list)
 212 {
 213         struct perf_evsel *evsel, *leader;
 214
 215         leader = list_entry(list->next, struct perf_evsel, node);
 216         evsel = list_entry(list->prev, struct perf_evsel, node);
 217
 218         leader->nr_members = evsel->idx - leader->idx + 1;
 219
 220         __evlist__for_each_entry(list, evsel) {
 221                 evsel->leader = leader;
 222         }
 223 }
 224
 225 void perf_evlist__set_leader(struct perf_evlist *evlist)
 226 {
 227         if (evlist->nr_entries) {
 228                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 229                 __perf_evlist__set_leader(&evlist->entries);
 230         }
 231 }
 232
 233 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 234 {
 235         attr->precise_ip = 3;
 236
 237         while (attr->precise_ip != 0) {
 238                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 239                 if (fd != -1) {
 240                         close(fd);
 241                         break;
 242                 }
 243                 --attr->precise_ip;
 244         }
 245 }
 246
 247 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 248 {
 249         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 250
 251         if (evsel == NULL)
 252                 return -ENOMEM;
 253
 254         perf_evlist__add(evlist, evsel);
 255         return 0;
 256 }
 257
 258 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 259 {
 260         struct perf_event_attr attr = {
 261                 .type   = PERF_TYPE_SOFTWARE,
 262                 .config = PERF_COUNT_SW_DUMMY,
 263                 .size   = sizeof(attr), /* to capture ABI version */
 264         };
 265         struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
 266
 267         if (evsel == NULL)
 268                 return -ENOMEM;
 269
 270         perf_evlist__add(evlist, evsel);
 271         return 0;
 272 }
 273
 274 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 275                                   struct perf_event_attr *attrs, size_t nr_attrs)
 276 {
 277         struct perf_evsel *evsel, *n;
 278         LIST_HEAD(head);
 279         size_t i;
 280
 281         for (i = 0; i < nr_attrs; i++) {
 282                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 283                 if (evsel == NULL)
 284                         goto out_delete_partial_list;
 285                 list_add_tail(&evsel->node, &head);
 286         }
 287
 288         perf_evlist__splice_list_tail(evlist, &head);
 289
 290         return 0;
 291
 292 out_delete_partial_list:
 293         __evlist__for_each_entry_safe(&head, n, evsel)
 294                 perf_evsel__delete(evsel);
 295         return -1;
 296 }
 297
 298 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 299                                      struct perf_event_attr *attrs, size_t nr_attrs)
 300 {
 301         size_t i;
 302
 303         for (i = 0; i < nr_attrs; i++)
 304                 event_attr_init(attrs + i);
 305
 306         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 307 }
 308
 309 struct perf_evsel *
 310 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 311 {
 312         struct perf_evsel *evsel;
 313
 314         evlist__for_each_entry(evlist, evsel) {
 315                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 316                     (int)evsel->attr.config == id)
 317                         return evsel;
 318         }
 319
 320         return NULL;
 321 }
 322
 323 struct perf_evsel *
 324 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 325                                      const char *name)
 326 {
 327         struct perf_evsel *evsel;
 328
 329         evlist__for_each_entry(evlist, evsel) {
 330                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 331                     (strcmp(evsel->name, name) == 0))
 332                         return evsel;
 333         }
 334
 335         return NULL;
 336 }
 337
 338 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 339                            const char *sys, const char *name, void *handler)
 340 {
 341         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 342
 343         if (IS_ERR(evsel))
 344                 return -1;
 345
 346         evsel->handler = handler;
 347         perf_evlist__add(evlist, evsel);
 348         return 0;
 349 }
 350
 351 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 352                                    struct perf_evsel *evsel)
 353 {
 354         if (evsel->system_wide)
 355                 return 1;
 356         else
 357                 return thread_map__nr(evlist->threads);
 358 }
 359
 360 void perf_evlist__disable(struct perf_evlist *evlist)
 361 {
 362         struct perf_evsel *pos;
 363
 364         evlist__for_each_entry(evlist, pos) {
 365                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 366                         continue;
 367                 perf_evsel__disable(pos);
 368         }
 369
 370         evlist->enabled = false;
 371 }
 372
 373 void perf_evlist__enable(struct perf_evlist *evlist)
 374 {
 375         struct perf_evsel *pos;
 376
 377         evlist__for_each_entry(evlist, pos) {
 378                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 379                         continue;
 380                 perf_evsel__enable(pos);
 381         }
 382
 383         evlist->enabled = true;
 384 }
 385
 386 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 387 {
 388         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 389 }
 390
 391 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 392                                          struct perf_evsel *evsel, int cpu)
 393 {
 394         int thread;
 395         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 396
 397         if (!evsel->fd)
 398                 return -EINVAL;
 399
 400         for (thread = 0; thread < nr_threads; thread++) {
 401                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 402                 if (err)
 403                         return err;
 404         }
 405         return 0;
 406 }
 407
 408 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 409                                             struct perf_evsel *evsel,
 410                                             int thread)
 411 {
 412         int cpu;
 413         int nr_cpus = cpu_map__nr(evlist->cpus);
 414
 415         if (!evsel->fd)
 416                 return -EINVAL;
 417
 418         for (cpu = 0; cpu < nr_cpus; cpu++) {
 419                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 420                 if (err)
 421                         return err;
 422         }
 423         return 0;
 424 }
 425
 426 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 427                                   struct perf_evsel *evsel, int idx)
 428 {
 429         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 430
 431         if (per_cpu_mmaps)
 432                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 433         else
 434                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 435 }
 436
 437 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 438 {
 439         int nr_cpus = cpu_map__nr(evlist->cpus);
 440         int nr_threads = thread_map__nr(evlist->threads);
 441         int nfds = 0;
 442         struct perf_evsel *evsel;
 443
 444         evlist__for_each_entry(evlist, evsel) {
 445                 if (evsel->system_wide)
 446                         nfds += nr_cpus;
 447                 else
 448                         nfds += nr_cpus * nr_threads;
 449         }
 450
 451         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 452             fdarray__grow(&evlist->pollfd, nfds) < 0)
 453                 return -ENOMEM;
 454
 455         return 0;
 456 }
 457
 458 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 459                                      struct perf_mmap *map, short revent)
 460 {
 461         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 462         /*
 463          * Save the idx so that when we filter out fds POLLHUP'ed we can
 464          * close the associated evlist->mmap[] entry.
 465          */
 466         if (pos >= 0) {
 467                 evlist->pollfd.priv[pos].ptr = map;
 468
 469                 fcntl(fd, F_SETFL, O_NONBLOCK);
 470         }
 471
 472         return pos;
 473 }
 474
 475 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 476 {
 477         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 478 }
 479
 480 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 481                                          void *arg __maybe_unused)
 482 {
 483         struct perf_mmap *map = fda->priv[fd].ptr;
 484
 485         if (map)
 486                 perf_mmap__put(map);
 487 }
 488
 489 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 490 {
 491         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 492                                perf_evlist__munmap_filtered, NULL);
 493 }
 494
 495 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 496 {
 497         return fdarray__poll(&evlist->pollfd, timeout);
 498 }
 499
 500 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 501                                  struct perf_evsel *evsel,
 502                                  int cpu, int thread, u64 id)
 503 {
 504         int hash;
 505         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 506
 507         sid->id = id;
 508         sid->evsel = evsel;
 509         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 510         hlist_add_head(&sid->node, &evlist->heads[hash]);
 511 }
 512
 513 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 514                          int cpu, int thread, u64 id)
 515 {
 516         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 517         evsel->id[evsel->ids++] = id;
 518 }
 519
 520 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 521                            struct perf_evsel *evsel,
 522                            int cpu, int thread, int fd)
 523 {
 524         u64 read_data[4] = { 0, };
 525         int id_idx = 1; /* The first entry is the counter value */
 526         u64 id;
 527         int ret;
 528
 529         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 530         if (!ret)
 531                 goto add;
 532
 533         if (errno != ENOTTY)
 534                 return -1;
 535
 536         /* Legacy way to get event id.. All hail to old kernels! */
 537
 538         /*
 539          * This way does not work with group format read, so bail
 540          * out in that case.
 541          */
 542         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 543                 return -1;
 544
 545         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 546             read(fd, &read_data, sizeof(read_data)) == -1)
 547                 return -1;
 548
 549         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 550                 ++id_idx;
 551         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 552                 ++id_idx;
 553
 554         id = read_data[id_idx];
 555
 556  add:
 557         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 558         return 0;
 559 }
 560
 561 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 562                                      struct perf_evsel *evsel, int idx, int cpu,
 563                                      int thread)
 564 {
 565         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 566         sid->idx = idx;
 567         if (evlist->cpus && cpu >= 0)
 568                 sid->cpu = evlist->cpus->map[cpu];
 569         else
 570                 sid->cpu = -1;
 571         if (!evsel->system_wide && evlist->threads && thread >= 0)
 572                 sid->tid = thread_map__pid(evlist->threads, thread);
 573         else
 574                 sid->tid = -1;
 575 }
 576
 577 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 578 {
 579         struct hlist_head *head;
 580         struct perf_sample_id *sid;
 581         int hash;
 582
 583         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 584         head = &evlist->heads[hash];
 585
 586         hlist_for_each_entry(sid, head, node)
 587                 if (sid->id == id)
 588                         return sid;
 589
 590         return NULL;
 591 }
 592
 593 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 594 {
 595         struct perf_sample_id *sid;
 596
 597         if (evlist->nr_entries == 1 || !id)
 598                 return perf_evlist__first(evlist);
 599
 600         sid = perf_evlist__id2sid(evlist, id);
 601         if (sid)
 602                 return sid->evsel;
 603
 604         if (!perf_evlist__sample_id_all(evlist))
 605                 return perf_evlist__first(evlist);
 606
 607         return NULL;
 608 }
 609
 610 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 611                                                 u64 id)
 612 {
 613         struct perf_sample_id *sid;
 614
 615         if (!id)
 616                 return NULL;
 617
 618         sid = perf_evlist__id2sid(evlist, id);
 619         if (sid)
 620                 return sid->evsel;
 621
 622         return NULL;
 623 }
 624
 625 static int perf_evlist__event2id(struct perf_evlist *evlist,
 626                                  union perf_event *event, u64 *id)
 627 {
 628         const u64 *array = event->sample.array;
 629         ssize_t n;
 630
 631         n = (event->header.size - sizeof(event->header)) >> 3;
 632
 633         if (event->header.type == PERF_RECORD_SAMPLE) {
 634                 if (evlist->id_pos >= n)
 635                         return -1;
 636                 *id = array[evlist->id_pos];
 637         } else {
 638                 if (evlist->is_pos > n)
 639                         return -1;
 640                 n -= evlist->is_pos;
 641                 *id = array[n];
 642         }
 643         return 0;
 644 }
 645
 646 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 647                                             union perf_event *event)
 648 {
 649         struct perf_evsel *first = perf_evlist__first(evlist);
 650         struct hlist_head *head;
 651         struct perf_sample_id *sid;
 652         int hash;
 653         u64 id;
 654
 655         if (evlist->nr_entries == 1)
 656                 return first;
 657
 658         if (!first->attr.sample_id_all &&
 659             event->header.type != PERF_RECORD_SAMPLE)
 660                 return first;
 661
 662         if (perf_evlist__event2id(evlist, event, &id))
 663                 return NULL;
 664
 665         /* Synthesized events have an id of zero */
 666         if (!id)
 667                 return first;
 668
 669         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 670         head = &evlist->heads[hash];
 671
 672         hlist_for_each_entry(sid, head, node) {
 673                 if (sid->id == id)
 674                         return sid->evsel;
 675         }
 676         return NULL;
 677 }
 678
 679 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 680 {
 681         int i;
 682
 683         if (!evlist->overwrite_mmap)
 684                 return 0;
 685
 686         for (i = 0; i < evlist->nr_mmaps; i++) {
 687                 int fd = evlist->overwrite_mmap[i].fd;
 688                 int err;
 689
 690                 if (fd < 0)
 691                         continue;
 692                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 693                 if (err)
 694                         return err;
 695         }
 696         return 0;
 697 }
 698
 699 static int perf_evlist__pause(struct perf_evlist *evlist)
 700 {
 701         return perf_evlist__set_paused(evlist, true);
 702 }
 703
 704 static int perf_evlist__resume(struct perf_evlist *evlist)
 705 {
 706         return perf_evlist__set_paused(evlist, false);
 707 }
 708
 709 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 710 {
 711         int i;
 712
 713         if (evlist->mmap)
 714                 for (i = 0; i < evlist->nr_mmaps; i++)
 715                         perf_mmap__munmap(&evlist->mmap[i]);
 716
 717         if (evlist->overwrite_mmap)
 718                 for (i = 0; i < evlist->nr_mmaps; i++)
 719                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 720 }
 721
 722 void perf_evlist__munmap(struct perf_evlist *evlist)
 723 {
 724         perf_evlist__munmap_nofree(evlist);
 725         zfree(&evlist->mmap);
 726         zfree(&evlist->overwrite_mmap);
 727 }
 728
 729 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
 730                                                  bool overwrite)
 731 {
 732         int i;
 733         struct perf_mmap *map;
 734
 735         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 736         if (cpu_map__empty(evlist->cpus))
 737                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 738         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 739         if (!map)
 740                 return NULL;
 741
 742         for (i = 0; i < evlist->nr_mmaps; i++) {
 743                 map[i].fd = -1;
 744                 map[i].overwrite = overwrite;
 745                 /*
 746                  * When the perf_mmap() call is made we grab one refcount, plus
 747                  * one extra to let perf_mmap__consume() get the last
 748                  * events after all real references (perf_mmap__get()) are
 749                  * dropped.
 750                  *
 751                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 752                  * thus does perf_mmap__get() on it.
 753                  */
 754                 refcount_set(&map[i].refcnt, 0);
 755         }
 756         return map;
 757 }
 758
 759 static bool
 760 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 761                          struct perf_evsel *evsel)
 762 {
 763         if (evsel->attr.write_backward)
 764                 return false;
 765         return true;
 766 }
 767
 768 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 769                                        struct mmap_params *mp, int cpu_idx,
 770                                        int thread, int *_output, int *_output_overwrite)
 771 {
 772         struct perf_evsel *evsel;
 773         int revent;
 774         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 775
 776         evlist__for_each_entry(evlist, evsel) {
 777                 struct perf_mmap *maps = evlist->mmap;
 778                 int *output = _output;
 779                 int fd;
 780                 int cpu;
 781
 782                 mp->prot = PROT_READ | PROT_WRITE;
 783                 if (evsel->attr.write_backward) {
 784                         output = _output_overwrite;
 785                         maps = evlist->overwrite_mmap;
 786
 787                         if (!maps) {
 788                                 maps = perf_evlist__alloc_mmap(evlist, true);
 789                                 if (!maps)
 790                                         return -1;
 791                                 evlist->overwrite_mmap = maps;
 792                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 793                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 794                         }
 795                         mp->prot &= ~PROT_WRITE;
 796                 }
 797
 798                 if (evsel->system_wide && thread)
 799                         continue;
 800
 801                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 802                 if (cpu == -1)
 803                         continue;
 804
 805                 fd = FD(evsel, cpu, thread);
 806
 807                 if (*output == -1) {
 808                         *output = fd;
 809
 810                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 811                                 return -1;
 812                 } else {
 813                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 814                                 return -1;
 815
 816                         perf_mmap__get(&maps[idx]);
 817                 }
 818
 819                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 820
 821                 /*
 822                  * The system_wide flag causes a selected event to be opened
 823                  * always without a pid.  Consequently it will never get a
 824                  * POLLHUP, but it is used for tracking in combination with
 825                  * other events, so it should not need to be polled anyway.
 826                  * Therefore don't add it for polling.
 827                  */
 828                 if (!evsel->system_wide &&
 829                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 830                         perf_mmap__put(&maps[idx]);
 831                         return -1;
 832                 }
 833
 834                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 835                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 836                                                    fd) < 0)
 837                                 return -1;
 838                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 839                                                  thread);
 840                 }
 841         }
 842
 843         return 0;
 844 }
 845
 846 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 847                                      struct mmap_params *mp)
 848 {
 849         int cpu, thread;
 850         int nr_cpus = cpu_map__nr(evlist->cpus);
 851         int nr_threads = thread_map__nr(evlist->threads);
 852
 853         pr_debug2("perf event ring buffer mmapped per cpu\n");
 854         for (cpu = 0; cpu < nr_cpus; cpu++) {
 855                 int output = -1;
 856                 int output_overwrite = -1;
 857
 858                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 859                                               true);
 860
 861                 for (thread = 0; thread < nr_threads; thread++) {
 862                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 863                                                         thread, &output, &output_overwrite))
 864                                 goto out_unmap;
 865                 }
 866         }
 867
 868         return 0;
 869
 870 out_unmap:
 871         perf_evlist__munmap_nofree(evlist);
 872         return -1;
 873 }
 874
 875 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 876                                         struct mmap_params *mp)
 877 {
 878         int thread;
 879         int nr_threads = thread_map__nr(evlist->threads);
 880
 881         pr_debug2("perf event ring buffer mmapped per thread\n");
 882         for (thread = 0; thread < nr_threads; thread++) {
 883                 int output = -1;
 884                 int output_overwrite = -1;
 885
 886                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 887                                               false);
 888
 889                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 890                                                 &output, &output_overwrite))
 891                         goto out_unmap;
 892         }
 893
 894         return 0;
 895
 896 out_unmap:
 897         perf_evlist__munmap_nofree(evlist);
 898         return -1;
 899 }
 900
 901 unsigned long perf_event_mlock_kb_in_pages(void)
 902 {
 903         unsigned long pages;
 904         int max;
 905
 906         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 907                 /*
 908                  * Pick a once upon a time good value, i.e. things look
 909                  * strange since we can't read a sysctl value, but lets not
 910                  * die yet...
 911                  */
 912                 max = 512;
 913         } else {
 914                 max -= (page_size / 1024);
 915         }
 916
 917         pages = (max * 1024) / page_size;
 918         if (!is_power_of_2(pages))
 919                 pages = rounddown_pow_of_two(pages);
 920
 921         return pages;
 922 }
 923
 924 size_t perf_evlist__mmap_size(unsigned long pages)
 925 {
 926         if (pages == UINT_MAX)
 927                 pages = perf_event_mlock_kb_in_pages();
 928         else if (!is_power_of_2(pages))
 929                 return 0;
 930
 931         return (pages + 1) * page_size;
 932 }
 933
 934 static long parse_pages_arg(const char *str, unsigned long min,
 935                             unsigned long max)
 936 {
 937         unsigned long pages, val;
 938         static struct parse_tag tags[] = {
 939                 { .tag  = 'B', .mult = 1       },
 940                 { .tag  = 'K', .mult = 1 << 10 },
 941                 { .tag  = 'M', .mult = 1 << 20 },
 942                 { .tag  = 'G', .mult = 1 << 30 },
 943                 { .tag  = 0 },
 944         };
 945
 946         if (str == NULL)
 947                 return -EINVAL;
 948
 949         val = parse_tag_value(str, tags);
 950         if (val != (unsigned long) -1) {
 951                 /* we got file size value */
 952                 pages = PERF_ALIGN(val, page_size) / page_size;
 953         } else {
 954                 /* we got pages count value */
 955                 char *eptr;
 956                 pages = strtoul(str, &eptr, 10);
 957                 if (*eptr != '\0')
 958                         return -EINVAL;
 959         }
 960
 961         if (pages == 0 && min == 0) {
 962                 /* leave number of pages at 0 */
 963         } else if (!is_power_of_2(pages)) {
 964                 char buf[100];
 965
 966                 /* round pages up to next power of 2 */
 967                 pages = roundup_pow_of_two(pages);
 968                 if (!pages)
 969                         return -EINVAL;
 970
 971                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 972                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 973                         buf, pages);
 974         }
 975
 976         if (pages > max)
 977                 return -EINVAL;
 978
 979         return pages;
 980 }
 981
 982 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 983 {
 984         unsigned long max = UINT_MAX;
 985         long pages;
 986
 987         if (max > SIZE_MAX / page_size)
 988                 max = SIZE_MAX / page_size;
 989
 990         pages = parse_pages_arg(str, 1, max);
 991         if (pages < 0) {
 992                 pr_err("Invalid argument for --mmap_pages/-m\n");
 993                 return -1;
 994         }
 995
 996         *mmap_pages = pages;
 997         return 0;
 998 }
 999
1000 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1001                                   int unset __maybe_unused)
1002 {
1003         return __perf_evlist__parse_mmap_pages(opt->value, str);
1004 }
1005
1006 /**
1007  * perf_evlist__mmap_ex - Create mmaps to receive events.
1008  * @evlist: list of events
1009  * @pages: map length in pages
1010  * @overwrite: overwrite older events?
1011  * @auxtrace_pages - auxtrace map length in pages
1012  * @auxtrace_overwrite - overwrite older auxtrace data?
1013  *
1014  * If @overwrite is %false the user needs to signal event consumption using
1015  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1016  * automatically.
1017  *
1018  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1019  * consumption using auxtrace_mmap__write_tail().
1020  *
1021  * Return: %0 on success, negative error code otherwise.
1022  */
1023 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1024                          unsigned int auxtrace_pages,
1025                          bool auxtrace_overwrite)
1026 {
1027         struct perf_evsel *evsel;
1028         const struct cpu_map *cpus = evlist->cpus;
1029         const struct thread_map *threads = evlist->threads;
1030         /*
1031          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1032          * Its value is decided by evsel's write_backward.
1033          * So &mp should not be passed through const pointer.
1034          */
1035         struct mmap_params mp;
1036
1037         if (!evlist->mmap)
1038                 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1039         if (!evlist->mmap)
1040                 return -ENOMEM;
1041
1042         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1043                 return -ENOMEM;
1044
1045         evlist->mmap_len = perf_evlist__mmap_size(pages);
1046         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1047         mp.mask = evlist->mmap_len - page_size - 1;
1048
1049         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1050                                    auxtrace_pages, auxtrace_overwrite);
1051
1052         evlist__for_each_entry(evlist, evsel) {
1053                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1054                     evsel->sample_id == NULL &&
1055                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1056                         return -ENOMEM;
1057         }
1058
1059         if (cpu_map__empty(cpus))
1060                 return perf_evlist__mmap_per_thread(evlist, &mp);
1061
1062         return perf_evlist__mmap_per_cpu(evlist, &mp);
1063 }
1064
1065 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1066 {
1067         return perf_evlist__mmap_ex(evlist, pages, 0, false);
1068 }
1069
1070 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1071 {
1072         bool all_threads = (target->per_thread && target->system_wide);
1073         struct cpu_map *cpus;
1074         struct thread_map *threads;
1075
1076         /*
1077          * If specify '-a' and '--per-thread' to perf record, perf record
1078          * will override '--per-thread'. target->per_thread = false and
1079          * target->system_wide = true.
1080          *
1081          * If specify '--per-thread' only to perf record,
1082          * target->per_thread = true and target->system_wide = false.
1083          *
1084          * So target->per_thread && target->system_wide is false.
1085          * For perf record, thread_map__new_str doesn't call
1086          * thread_map__new_all_cpus. That will keep perf record's
1087          * current behavior.
1088          *
1089          * For perf stat, it allows the case that target->per_thread and
1090          * target->system_wide are all true. It means to collect system-wide
1091          * per-thread data. thread_map__new_str will call
1092          * thread_map__new_all_cpus to enumerate all threads.
1093          */
1094         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1095                                       all_threads);
1096
1097         if (!threads)
1098                 return -1;
1099
1100         if (target__uses_dummy_map(target))
1101                 cpus = cpu_map__dummy_new();
1102         else
1103                 cpus = cpu_map__new(target->cpu_list);
1104
1105         if (!cpus)
1106                 goto out_delete_threads;
1107
1108         evlist->has_user_cpus = !!target->cpu_list;
1109
1110         perf_evlist__set_maps(evlist, cpus, threads);
1111
1112         return 0;
1113
1114 out_delete_threads:
1115         thread_map__put(threads);
1116         return -1;
1117 }
1118
1119 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1120                            struct thread_map *threads)
1121 {
1122         /*
1123          * Allow for the possibility that one or another of the maps isn't being
1124          * changed i.e. don't put it.  Note we are assuming the maps that are
1125          * being applied are brand new and evlist is taking ownership of the
1126          * original reference count of 1.  If that is not the case it is up to
1127          * the caller to increase the reference count.
1128          */
1129         if (cpus != evlist->cpus) {
1130                 cpu_map__put(evlist->cpus);
1131                 evlist->cpus = cpu_map__get(cpus);
1132         }
1133
1134         if (threads != evlist->threads) {
1135                 thread_map__put(evlist->threads);
1136                 evlist->threads = thread_map__get(threads);
1137         }
1138
1139         perf_evlist__propagate_maps(evlist);
1140 }
1141
1142 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1143                                    enum perf_event_sample_format bit)
1144 {
1145         struct perf_evsel *evsel;
1146
1147         evlist__for_each_entry(evlist, evsel)
1148                 __perf_evsel__set_sample_bit(evsel, bit);
1149 }
1150
1151 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1152                                      enum perf_event_sample_format bit)
1153 {
1154         struct perf_evsel *evsel;
1155
1156         evlist__for_each_entry(evlist, evsel)
1157                 __perf_evsel__reset_sample_bit(evsel, bit);
1158 }
1159
1160 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1161 {
1162         struct perf_evsel *evsel;
1163         int err = 0;
1164
1165         evlist__for_each_entry(evlist, evsel) {
1166                 if (evsel->filter == NULL)
1167                         continue;
1168
1169                 /*
1170                  * filters only work for tracepoint event, which doesn't have cpu limit.
1171                  * So evlist and evsel should always be same.
1172                  */
1173                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1174                 if (err) {
1175                         *err_evsel = evsel;
1176                         break;
1177                 }
1178         }
1179
1180         return err;
1181 }
1182
1183 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1184 {
1185         struct perf_evsel *evsel;
1186         int err = 0;
1187
1188         evlist__for_each_entry(evlist, evsel) {
1189                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1190                         continue;
1191
1192                 err = perf_evsel__set_filter(evsel, filter);
1193                 if (err)
1194                         break;
1195         }
1196
1197         return err;
1198 }
1199
1200 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1201 {
1202         char *filter;
1203         int ret = -1;
1204         size_t i;
1205
1206         for (i = 0; i < npids; ++i) {
1207                 if (i == 0) {
1208                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1209                                 return -1;
1210                 } else {
1211                         char *tmp;
1212
1213                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1214                                 goto out_free;
1215
1216                         free(filter);
1217                         filter = tmp;
1218                 }
1219         }
1220
1221         ret = perf_evlist__set_filter(evlist, filter);
1222 out_free:
1223         free(filter);
1224         return ret;
1225 }
1226
1227 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1228 {
1229         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1230 }
1231
1232 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1233 {
1234         struct perf_evsel *pos;
1235
1236         if (evlist->nr_entries == 1)
1237                 return true;
1238
1239         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1240                 return false;
1241
1242         evlist__for_each_entry(evlist, pos) {
1243                 if (pos->id_pos != evlist->id_pos ||
1244                     pos->is_pos != evlist->is_pos)
1245                         return false;
1246         }
1247
1248         return true;
1249 }
1250
1251 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1252 {
1253         struct perf_evsel *evsel;
1254
1255         if (evlist->combined_sample_type)
1256                 return evlist->combined_sample_type;
1257
1258         evlist__for_each_entry(evlist, evsel)
1259                 evlist->combined_sample_type |= evsel->attr.sample_type;
1260
1261         return evlist->combined_sample_type;
1262 }
1263
1264 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1265 {
1266         evlist->combined_sample_type = 0;
1267         return __perf_evlist__combined_sample_type(evlist);
1268 }
1269
1270 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1271 {
1272         struct perf_evsel *evsel;
1273         u64 branch_type = 0;
1274
1275         evlist__for_each_entry(evlist, evsel)
1276                 branch_type |= evsel->attr.branch_sample_type;
1277         return branch_type;
1278 }
1279
1280 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1281 {
1282         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1283         u64 read_format = first->attr.read_format;
1284         u64 sample_type = first->attr.sample_type;
1285
1286         evlist__for_each_entry(evlist, pos) {
1287                 if (read_format != pos->attr.read_format)
1288                         return false;
1289         }
1290
1291         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1292         if ((sample_type & PERF_SAMPLE_READ) &&
1293             !(read_format & PERF_FORMAT_ID)) {
1294                 return false;
1295         }
1296
1297         return true;
1298 }
1299
1300 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1301 {
1302         struct perf_evsel *first = perf_evlist__first(evlist);
1303         return first->attr.read_format;
1304 }
1305
1306 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1307 {
1308         struct perf_evsel *first = perf_evlist__first(evlist);
1309         struct perf_sample *data;
1310         u64 sample_type;
1311         u16 size = 0;
1312
1313         if (!first->attr.sample_id_all)
1314                 goto out;
1315
1316         sample_type = first->attr.sample_type;
1317
1318         if (sample_type & PERF_SAMPLE_TID)
1319                 size += sizeof(data->tid) * 2;
1320
1321        if (sample_type & PERF_SAMPLE_TIME)
1322                 size += sizeof(data->time);
1323
1324         if (sample_type & PERF_SAMPLE_ID)
1325                 size += sizeof(data->id);
1326
1327         if (sample_type & PERF_SAMPLE_STREAM_ID)
1328                 size += sizeof(data->stream_id);
1329
1330         if (sample_type & PERF_SAMPLE_CPU)
1331                 size += sizeof(data->cpu) * 2;
1332
1333         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1334                 size += sizeof(data->id);
1335 out:
1336         return size;
1337 }
1338
1339 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1340 {
1341         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1342
1343         evlist__for_each_entry_continue(evlist, pos) {
1344                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1345                         return false;
1346         }
1347
1348         return true;
1349 }
1350
1351 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1352 {
1353         struct perf_evsel *first = perf_evlist__first(evlist);
1354         return first->attr.sample_id_all;
1355 }
1356
1357 void perf_evlist__set_selected(struct perf_evlist *evlist,
1358                                struct perf_evsel *evsel)
1359 {
1360         evlist->selected = evsel;
1361 }
1362
1363 void perf_evlist__close(struct perf_evlist *evlist)
1364 {
1365         struct perf_evsel *evsel;
1366
1367         evlist__for_each_entry_reverse(evlist, evsel)
1368                 perf_evsel__close(evsel);
1369 }
1370
1371 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1372 {
1373         struct cpu_map    *cpus;
1374         struct thread_map *threads;
1375         int err = -ENOMEM;
1376
1377         /*
1378          * Try reading /sys/devices/system/cpu/online to get
1379          * an all cpus map.
1380          *
1381          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1382          * code needs an overhaul to properly forward the
1383          * error, and we may not want to do that fallback to a
1384          * default cpu identity map :-\
1385          */
1386         cpus = cpu_map__new(NULL);
1387         if (!cpus)
1388                 goto out;
1389
1390         threads = thread_map__new_dummy();
1391         if (!threads)
1392                 goto out_put;
1393
1394         perf_evlist__set_maps(evlist, cpus, threads);
1395 out:
1396         return err;
1397 out_put:
1398         cpu_map__put(cpus);
1399         goto out;
1400 }
1401
1402 int perf_evlist__open(struct perf_evlist *evlist)
1403 {
1404         struct perf_evsel *evsel;
1405         int err;
1406
1407         /*
1408          * Default: one fd per CPU, all threads, aka systemwide
1409          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1410          */
1411         if (evlist->threads == NULL && evlist->cpus == NULL) {
1412                 err = perf_evlist__create_syswide_maps(evlist);
1413                 if (err < 0)
1414                         goto out_err;
1415         }
1416
1417         perf_evlist__update_id_pos(evlist);
1418
1419         evlist__for_each_entry(evlist, evsel) {
1420                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1421                 if (err < 0)
1422                         goto out_err;
1423         }
1424
1425         return 0;
1426 out_err:
1427         perf_evlist__close(evlist);
1428         errno = -err;
1429         return err;
1430 }
1431
1432 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1433                                   const char *argv[], bool pipe_output,
1434                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1435 {
1436         int child_ready_pipe[2], go_pipe[2];
1437         char bf;
1438
1439         if (pipe(child_ready_pipe) < 0) {
1440                 perror("failed to create 'ready' pipe");
1441                 return -1;
1442         }
1443
1444         if (pipe(go_pipe) < 0) {
1445                 perror("failed to create 'go' pipe");
1446                 goto out_close_ready_pipe;
1447         }
1448
1449         evlist->workload.pid = fork();
1450         if (evlist->workload.pid < 0) {
1451                 perror("failed to fork");
1452                 goto out_close_pipes;
1453         }
1454
1455         if (!evlist->workload.pid) {
1456                 int ret;
1457
1458                 if (pipe_output)
1459                         dup2(2, 1);
1460
1461                 signal(SIGTERM, SIG_DFL);
1462
1463                 close(child_ready_pipe[0]);
1464                 close(go_pipe[1]);
1465                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1466
1467                 /*
1468                  * Tell the parent we're ready to go
1469                  */
1470                 close(child_ready_pipe[1]);
1471
1472                 /*
1473                  * Wait until the parent tells us to go.
1474                  */
1475                 ret = read(go_pipe[0], &bf, 1);
1476                 /*
1477                  * The parent will ask for the execvp() to be performed by
1478                  * writing exactly one byte, in workload.cork_fd, usually via
1479                  * perf_evlist__start_workload().
1480                  *
1481                  * For cancelling the workload without actually running it,
1482                  * the parent will just close workload.cork_fd, without writing
1483                  * anything, i.e. read will return zero and we just exit()
1484                  * here.
1485                  */
1486                 if (ret != 1) {
1487                         if (ret == -1)
1488                                 perror("unable to read pipe");
1489                         exit(ret);
1490                 }
1491
1492                 execvp(argv[0], (char **)argv);
1493
1494                 if (exec_error) {
1495                         union sigval val;
1496
1497                         val.sival_int = errno;
1498                         if (sigqueue(getppid(), SIGUSR1, val))
1499                                 perror(argv[0]);
1500                 } else
1501                         perror(argv[0]);
1502                 exit(-1);
1503         }
1504
1505         if (exec_error) {
1506                 struct sigaction act = {
1507                         .sa_flags     = SA_SIGINFO,
1508                         .sa_sigaction = exec_error,
1509                 };
1510                 sigaction(SIGUSR1, &act, NULL);
1511         }
1512
1513         if (target__none(target)) {
1514                 if (evlist->threads == NULL) {
1515                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1516                                 __func__, __LINE__);
1517                         goto out_close_pipes;
1518                 }
1519                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1520         }
1521
1522         close(child_ready_pipe[1]);
1523         close(go_pipe[0]);
1524         /*
1525          * wait for child to settle
1526          */
1527         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1528                 perror("unable to read pipe");
1529                 goto out_close_pipes;
1530         }
1531
1532         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1533         evlist->workload.cork_fd = go_pipe[1];
1534         close(child_ready_pipe[0]);
1535         return 0;
1536
1537 out_close_pipes:
1538         close(go_pipe[0]);
1539         close(go_pipe[1]);
1540 out_close_ready_pipe:
1541         close(child_ready_pipe[0]);
1542         close(child_ready_pipe[1]);
1543         return -1;
1544 }
1545
1546 int perf_evlist__start_workload(struct perf_evlist *evlist)
1547 {
1548         if (evlist->workload.cork_fd > 0) {
1549                 char bf = 0;
1550                 int ret;
1551                 /*
1552                  * Remove the cork, let it rip!
1553                  */
1554                 ret = write(evlist->workload.cork_fd, &bf, 1);
1555                 if (ret < 0)
1556                         perror("unable to write to pipe");
1557
1558                 close(evlist->workload.cork_fd);
1559                 return ret;
1560         }
1561
1562         return 0;
1563 }
1564
1565 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1566                               struct perf_sample *sample)
1567 {
1568         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1569
1570         if (!evsel)
1571                 return -EFAULT;
1572         return perf_evsel__parse_sample(evsel, event, sample);
1573 }
1574
1575 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1576                                         union perf_event *event,
1577                                         u64 *timestamp)
1578 {
1579         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1580
1581         if (!evsel)
1582                 return -EFAULT;
1583         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1584 }
1585
1586 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1587 {
1588         struct perf_evsel *evsel;
1589         size_t printed = 0;
1590
1591         evlist__for_each_entry(evlist, evsel) {
1592                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1593                                    perf_evsel__name(evsel));
1594         }
1595
1596         return printed + fprintf(fp, "\n");
1597 }
1598
1599 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1600                                int err, char *buf, size_t size)
1601 {
1602         int printed, value;
1603         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1604
1605         switch (err) {
1606         case EACCES:
1607         case EPERM:
1608                 printed = scnprintf(buf, size,
1609                                     "Error:\t%s.\n"
1610                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1611
1612                 value = perf_event_paranoid();
1613
1614                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1615
1616                 if (value >= 2) {
1617                         printed += scnprintf(buf + printed, size - printed,
1618                                              "For your workloads it needs to be <= 1\nHint:\t");
1619                 }
1620                 printed += scnprintf(buf + printed, size - printed,
1621                                      "For system wide tracing it needs to be set to -1.\n");
1622
1623                 printed += scnprintf(buf + printed, size - printed,
1624                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1625                                     "Hint:\tThe current value is %d.", value);
1626                 break;
1627         case EINVAL: {
1628                 struct perf_evsel *first = perf_evlist__first(evlist);
1629                 int max_freq;
1630
1631                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1632                         goto out_default;
1633
1634                 if (first->attr.sample_freq < (u64)max_freq)
1635                         goto out_default;
1636
1637                 printed = scnprintf(buf, size,
1638                                     "Error:\t%s.\n"
1639                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1640                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1641                                     emsg, max_freq, first->attr.sample_freq);
1642                 break;
1643         }
1644         default:
1645 out_default:
1646                 scnprintf(buf, size, "%s", emsg);
1647                 break;
1648         }
1649
1650         return 0;
1651 }
1652
1653 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1654 {
1655         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1656         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1657
1658         switch (err) {
1659         case EPERM:
1660                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1661                 printed += scnprintf(buf + printed, size - printed,
1662                                      "Error:\t%s.\n"
1663                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1664                                      "Hint:\tTried using %zd kB.\n",
1665                                      emsg, pages_max_per_user, pages_attempted);
1666
1667                 if (pages_attempted >= pages_max_per_user) {
1668                         printed += scnprintf(buf + printed, size - printed,
1669                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1670                                              pages_max_per_user + pages_attempted);
1671                 }
1672
1673                 printed += scnprintf(buf + printed, size - printed,
1674                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1675                 break;
1676         default:
1677                 scnprintf(buf, size, "%s", emsg);
1678                 break;
1679         }
1680
1681         return 0;
1682 }
1683
1684 void perf_evlist__to_front(struct perf_evlist *evlist,
1685                            struct perf_evsel *move_evsel)
1686 {
1687         struct perf_evsel *evsel, *n;
1688         LIST_HEAD(move);
1689
1690         if (move_evsel == perf_evlist__first(evlist))
1691                 return;
1692
1693         evlist__for_each_entry_safe(evlist, n, evsel) {
1694                 if (evsel->leader == move_evsel->leader)
1695                         list_move_tail(&evsel->node, &move);
1696         }
1697
1698         list_splice(&move, &evlist->entries);
1699 }
1700
1701 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1702                                      struct perf_evsel *tracking_evsel)
1703 {
1704         struct perf_evsel *evsel;
1705
1706         if (tracking_evsel->tracking)
1707                 return;
1708
1709         evlist__for_each_entry(evlist, evsel) {
1710                 if (evsel != tracking_evsel)
1711                         evsel->tracking = false;
1712         }
1713
1714         tracking_evsel->tracking = true;
1715 }
1716
1717 struct perf_evsel *
1718 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1719                                const char *str)
1720 {
1721         struct perf_evsel *evsel;
1722
1723         evlist__for_each_entry(evlist, evsel) {
1724                 if (!evsel->name)
1725                         continue;
1726                 if (strcmp(str, evsel->name) == 0)
1727                         return evsel;
1728         }
1729
1730         return NULL;
1731 }
1732
1733 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1734                                   enum bkw_mmap_state state)
1735 {
1736         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1737         enum action {
1738                 NONE,
1739                 PAUSE,
1740                 RESUME,
1741         } action = NONE;
1742
1743         if (!evlist->overwrite_mmap)
1744                 return;
1745
1746         switch (old_state) {
1747         case BKW_MMAP_NOTREADY: {
1748                 if (state != BKW_MMAP_RUNNING)
1749                         goto state_err;
1750                 break;
1751         }
1752         case BKW_MMAP_RUNNING: {
1753                 if (state != BKW_MMAP_DATA_PENDING)
1754                         goto state_err;
1755                 action = PAUSE;
1756                 break;
1757         }
1758         case BKW_MMAP_DATA_PENDING: {
1759                 if (state != BKW_MMAP_EMPTY)
1760                         goto state_err;
1761                 break;
1762         }
1763         case BKW_MMAP_EMPTY: {
1764                 if (state != BKW_MMAP_RUNNING)
1765                         goto state_err;
1766                 action = RESUME;
1767                 break;
1768         }
1769         default:
1770                 WARN_ONCE(1, "Shouldn't get there\n");
1771         }
1772
1773         evlist->bkw_mmap_state = state;
1774
1775         switch (action) {
1776         case PAUSE:
1777                 perf_evlist__pause(evlist);
1778                 break;
1779         case RESUME:
1780                 perf_evlist__resume(evlist);
1781                 break;
1782         case NONE:
1783         default:
1784                 break;
1785         }
1786
1787 state_err:
1788         return;
1789 }
1790
1791 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1792 {
1793         struct perf_evsel *evsel;
1794
1795         evlist__for_each_entry(evlist, evsel) {
1796                 if (!evsel->attr.exclude_kernel)
1797                         return false;
1798         }
1799
1800         return true;
1801 }
1802
1803 /*
1804  * Events in data file are not collect in groups, but we still want
1805  * the group display. Set the artificial group and set the leader's
1806  * forced_leader flag to notify the display code.
1807  */
1808 void perf_evlist__force_leader(struct perf_evlist *evlist)
1809 {
1810         if (!evlist->nr_groups) {
1811                 struct perf_evsel *leader = perf_evlist__first(evlist);
1812
1813                 perf_evlist__set_leader(evlist);
1814                 leader->forced_leader = true;
1815         }
1816 }