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 *pattr)
 234 {
 235         struct perf_event_attr attr = {
 236                 .type           = PERF_TYPE_HARDWARE,
 237                 .config         = PERF_COUNT_HW_CPU_CYCLES,
 238                 .exclude_kernel = 1,
 239                 .precise_ip     = 3,
 240         };
 241
 242         event_attr_init(&attr);
 243
 244         /*
 245          * Unnamed union member, not supported as struct member named
 246          * initializer in older compilers such as gcc 4.4.7
 247          */
 248         attr.sample_period = 1;
 249
 250         while (attr.precise_ip != 0) {
 251                 int fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
 252                 if (fd != -1) {
 253                         close(fd);
 254                         break;
 255                 }
 256                 --attr.precise_ip;
 257         }
 258
 259         pattr->precise_ip = attr.precise_ip;
 260 }
 261
 262 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 263 {
 264         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 265
 266         if (evsel == NULL)
 267                 return -ENOMEM;
 268
 269         perf_evlist__add(evlist, evsel);
 270         return 0;
 271 }
 272
 273 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 274 {
 275         struct perf_event_attr attr = {
 276                 .type   = PERF_TYPE_SOFTWARE,
 277                 .config = PERF_COUNT_SW_DUMMY,
 278                 .size   = sizeof(attr), /* to capture ABI version */
 279         };
 280         struct perf_evsel *evsel = perf_evsel__new_idx(&attr, evlist->nr_entries);
 281
 282         if (evsel == NULL)
 283                 return -ENOMEM;
 284
 285         perf_evlist__add(evlist, evsel);
 286         return 0;
 287 }
 288
 289 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 290                                   struct perf_event_attr *attrs, size_t nr_attrs)
 291 {
 292         struct perf_evsel *evsel, *n;
 293         LIST_HEAD(head);
 294         size_t i;
 295
 296         for (i = 0; i < nr_attrs; i++) {
 297                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 298                 if (evsel == NULL)
 299                         goto out_delete_partial_list;
 300                 list_add_tail(&evsel->node, &head);
 301         }
 302
 303         perf_evlist__splice_list_tail(evlist, &head);
 304
 305         return 0;
 306
 307 out_delete_partial_list:
 308         __evlist__for_each_entry_safe(&head, n, evsel)
 309                 perf_evsel__delete(evsel);
 310         return -1;
 311 }
 312
 313 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 314                                      struct perf_event_attr *attrs, size_t nr_attrs)
 315 {
 316         size_t i;
 317
 318         for (i = 0; i < nr_attrs; i++)
 319                 event_attr_init(attrs + i);
 320
 321         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 322 }
 323
 324 struct perf_evsel *
 325 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 326 {
 327         struct perf_evsel *evsel;
 328
 329         evlist__for_each_entry(evlist, evsel) {
 330                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 331                     (int)evsel->attr.config == id)
 332                         return evsel;
 333         }
 334
 335         return NULL;
 336 }
 337
 338 struct perf_evsel *
 339 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 340                                      const char *name)
 341 {
 342         struct perf_evsel *evsel;
 343
 344         evlist__for_each_entry(evlist, evsel) {
 345                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 346                     (strcmp(evsel->name, name) == 0))
 347                         return evsel;
 348         }
 349
 350         return NULL;
 351 }
 352
 353 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 354                            const char *sys, const char *name, void *handler)
 355 {
 356         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 357
 358         if (IS_ERR(evsel))
 359                 return -1;
 360
 361         evsel->handler = handler;
 362         perf_evlist__add(evlist, evsel);
 363         return 0;
 364 }
 365
 366 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 367                                    struct perf_evsel *evsel)
 368 {
 369         if (evsel->system_wide)
 370                 return 1;
 371         else
 372                 return thread_map__nr(evlist->threads);
 373 }
 374
 375 void perf_evlist__disable(struct perf_evlist *evlist)
 376 {
 377         struct perf_evsel *pos;
 378
 379         evlist__for_each_entry(evlist, pos) {
 380                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->fd)
 381                         continue;
 382                 perf_evsel__disable(pos);
 383         }
 384
 385         evlist->enabled = false;
 386 }
 387
 388 void perf_evlist__enable(struct perf_evlist *evlist)
 389 {
 390         struct perf_evsel *pos;
 391
 392         evlist__for_each_entry(evlist, pos) {
 393                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 394                         continue;
 395                 perf_evsel__enable(pos);
 396         }
 397
 398         evlist->enabled = true;
 399 }
 400
 401 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 402 {
 403         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 404 }
 405
 406 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 407                                          struct perf_evsel *evsel, int cpu)
 408 {
 409         int thread;
 410         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 411
 412         if (!evsel->fd)
 413                 return -EINVAL;
 414
 415         for (thread = 0; thread < nr_threads; thread++) {
 416                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 417                 if (err)
 418                         return err;
 419         }
 420         return 0;
 421 }
 422
 423 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 424                                             struct perf_evsel *evsel,
 425                                             int thread)
 426 {
 427         int cpu;
 428         int nr_cpus = cpu_map__nr(evlist->cpus);
 429
 430         if (!evsel->fd)
 431                 return -EINVAL;
 432
 433         for (cpu = 0; cpu < nr_cpus; cpu++) {
 434                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 435                 if (err)
 436                         return err;
 437         }
 438         return 0;
 439 }
 440
 441 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 442                                   struct perf_evsel *evsel, int idx)
 443 {
 444         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 445
 446         if (per_cpu_mmaps)
 447                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 448         else
 449                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 450 }
 451
 452 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 453 {
 454         int nr_cpus = cpu_map__nr(evlist->cpus);
 455         int nr_threads = thread_map__nr(evlist->threads);
 456         int nfds = 0;
 457         struct perf_evsel *evsel;
 458
 459         evlist__for_each_entry(evlist, evsel) {
 460                 if (evsel->system_wide)
 461                         nfds += nr_cpus;
 462                 else
 463                         nfds += nr_cpus * nr_threads;
 464         }
 465
 466         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 467             fdarray__grow(&evlist->pollfd, nfds) < 0)
 468                 return -ENOMEM;
 469
 470         return 0;
 471 }
 472
 473 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 474                                      struct perf_mmap *map, short revent)
 475 {
 476         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 477         /*
 478          * Save the idx so that when we filter out fds POLLHUP'ed we can
 479          * close the associated evlist->mmap[] entry.
 480          */
 481         if (pos >= 0) {
 482                 evlist->pollfd.priv[pos].ptr = map;
 483
 484                 fcntl(fd, F_SETFL, O_NONBLOCK);
 485         }
 486
 487         return pos;
 488 }
 489
 490 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 491 {
 492         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 493 }
 494
 495 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 496                                          void *arg __maybe_unused)
 497 {
 498         struct perf_mmap *map = fda->priv[fd].ptr;
 499
 500         if (map)
 501                 perf_mmap__put(map);
 502 }
 503
 504 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 505 {
 506         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 507                                perf_evlist__munmap_filtered, NULL);
 508 }
 509
 510 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 511 {
 512         return fdarray__poll(&evlist->pollfd, timeout);
 513 }
 514
 515 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 516                                  struct perf_evsel *evsel,
 517                                  int cpu, int thread, u64 id)
 518 {
 519         int hash;
 520         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 521
 522         sid->id = id;
 523         sid->evsel = evsel;
 524         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 525         hlist_add_head(&sid->node, &evlist->heads[hash]);
 526 }
 527
 528 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 529                          int cpu, int thread, u64 id)
 530 {
 531         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 532         evsel->id[evsel->ids++] = id;
 533 }
 534
 535 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 536                            struct perf_evsel *evsel,
 537                            int cpu, int thread, int fd)
 538 {
 539         u64 read_data[4] = { 0, };
 540         int id_idx = 1; /* The first entry is the counter value */
 541         u64 id;
 542         int ret;
 543
 544         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 545         if (!ret)
 546                 goto add;
 547
 548         if (errno != ENOTTY)
 549                 return -1;
 550
 551         /* Legacy way to get event id.. All hail to old kernels! */
 552
 553         /*
 554          * This way does not work with group format read, so bail
 555          * out in that case.
 556          */
 557         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 558                 return -1;
 559
 560         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 561             read(fd, &read_data, sizeof(read_data)) == -1)
 562                 return -1;
 563
 564         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 565                 ++id_idx;
 566         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 567                 ++id_idx;
 568
 569         id = read_data[id_idx];
 570
 571  add:
 572         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 573         return 0;
 574 }
 575
 576 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 577                                      struct perf_evsel *evsel, int idx, int cpu,
 578                                      int thread)
 579 {
 580         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 581         sid->idx = idx;
 582         if (evlist->cpus && cpu >= 0)
 583                 sid->cpu = evlist->cpus->map[cpu];
 584         else
 585                 sid->cpu = -1;
 586         if (!evsel->system_wide && evlist->threads && thread >= 0)
 587                 sid->tid = thread_map__pid(evlist->threads, thread);
 588         else
 589                 sid->tid = -1;
 590 }
 591
 592 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 593 {
 594         struct hlist_head *head;
 595         struct perf_sample_id *sid;
 596         int hash;
 597
 598         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 599         head = &evlist->heads[hash];
 600
 601         hlist_for_each_entry(sid, head, node)
 602                 if (sid->id == id)
 603                         return sid;
 604
 605         return NULL;
 606 }
 607
 608 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 609 {
 610         struct perf_sample_id *sid;
 611
 612         if (evlist->nr_entries == 1 || !id)
 613                 return perf_evlist__first(evlist);
 614
 615         sid = perf_evlist__id2sid(evlist, id);
 616         if (sid)
 617                 return sid->evsel;
 618
 619         if (!perf_evlist__sample_id_all(evlist))
 620                 return perf_evlist__first(evlist);
 621
 622         return NULL;
 623 }
 624
 625 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 626                                                 u64 id)
 627 {
 628         struct perf_sample_id *sid;
 629
 630         if (!id)
 631                 return NULL;
 632
 633         sid = perf_evlist__id2sid(evlist, id);
 634         if (sid)
 635                 return sid->evsel;
 636
 637         return NULL;
 638 }
 639
 640 static int perf_evlist__event2id(struct perf_evlist *evlist,
 641                                  union perf_event *event, u64 *id)
 642 {
 643         const u64 *array = event->sample.array;
 644         ssize_t n;
 645
 646         n = (event->header.size - sizeof(event->header)) >> 3;
 647
 648         if (event->header.type == PERF_RECORD_SAMPLE) {
 649                 if (evlist->id_pos >= n)
 650                         return -1;
 651                 *id = array[evlist->id_pos];
 652         } else {
 653                 if (evlist->is_pos > n)
 654                         return -1;
 655                 n -= evlist->is_pos;
 656                 *id = array[n];
 657         }
 658         return 0;
 659 }
 660
 661 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 662                                             union perf_event *event)
 663 {
 664         struct perf_evsel *first = perf_evlist__first(evlist);
 665         struct hlist_head *head;
 666         struct perf_sample_id *sid;
 667         int hash;
 668         u64 id;
 669
 670         if (evlist->nr_entries == 1)
 671                 return first;
 672
 673         if (!first->attr.sample_id_all &&
 674             event->header.type != PERF_RECORD_SAMPLE)
 675                 return first;
 676
 677         if (perf_evlist__event2id(evlist, event, &id))
 678                 return NULL;
 679
 680         /* Synthesized events have an id of zero */
 681         if (!id)
 682                 return first;
 683
 684         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 685         head = &evlist->heads[hash];
 686
 687         hlist_for_each_entry(sid, head, node) {
 688                 if (sid->id == id)
 689                         return sid->evsel;
 690         }
 691         return NULL;
 692 }
 693
 694 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 695 {
 696         int i;
 697
 698         if (!evlist->overwrite_mmap)
 699                 return 0;
 700
 701         for (i = 0; i < evlist->nr_mmaps; i++) {
 702                 int fd = evlist->overwrite_mmap[i].fd;
 703                 int err;
 704
 705                 if (fd < 0)
 706                         continue;
 707                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 708                 if (err)
 709                         return err;
 710         }
 711         return 0;
 712 }
 713
 714 static int perf_evlist__pause(struct perf_evlist *evlist)
 715 {
 716         return perf_evlist__set_paused(evlist, true);
 717 }
 718
 719 static int perf_evlist__resume(struct perf_evlist *evlist)
 720 {
 721         return perf_evlist__set_paused(evlist, false);
 722 }
 723
 724 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 725 {
 726         int i;
 727
 728         if (evlist->mmap)
 729                 for (i = 0; i < evlist->nr_mmaps; i++)
 730                         perf_mmap__munmap(&evlist->mmap[i]);
 731
 732         if (evlist->overwrite_mmap)
 733                 for (i = 0; i < evlist->nr_mmaps; i++)
 734                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 735 }
 736
 737 void perf_evlist__munmap(struct perf_evlist *evlist)
 738 {
 739         perf_evlist__munmap_nofree(evlist);
 740         zfree(&evlist->mmap);
 741         zfree(&evlist->overwrite_mmap);
 742 }
 743
 744 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist,
 745                                                  bool overwrite)
 746 {
 747         int i;
 748         struct perf_mmap *map;
 749
 750         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 751         if (cpu_map__empty(evlist->cpus))
 752                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 753         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 754         if (!map)
 755                 return NULL;
 756
 757         for (i = 0; i < evlist->nr_mmaps; i++) {
 758                 map[i].fd = -1;
 759                 map[i].overwrite = overwrite;
 760                 /*
 761                  * When the perf_mmap() call is made we grab one refcount, plus
 762                  * one extra to let perf_mmap__consume() get the last
 763                  * events after all real references (perf_mmap__get()) are
 764                  * dropped.
 765                  *
 766                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 767                  * thus does perf_mmap__get() on it.
 768                  */
 769                 refcount_set(&map[i].refcnt, 0);
 770         }
 771         return map;
 772 }
 773
 774 static bool
 775 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
 776                          struct perf_evsel *evsel)
 777 {
 778         if (evsel->attr.write_backward)
 779                 return false;
 780         return true;
 781 }
 782
 783 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 784                                        struct mmap_params *mp, int cpu_idx,
 785                                        int thread, int *_output, int *_output_overwrite)
 786 {
 787         struct perf_evsel *evsel;
 788         int revent;
 789         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
 790
 791         evlist__for_each_entry(evlist, evsel) {
 792                 struct perf_mmap *maps = evlist->mmap;
 793                 int *output = _output;
 794                 int fd;
 795                 int cpu;
 796
 797                 mp->prot = PROT_READ | PROT_WRITE;
 798                 if (evsel->attr.write_backward) {
 799                         output = _output_overwrite;
 800                         maps = evlist->overwrite_mmap;
 801
 802                         if (!maps) {
 803                                 maps = perf_evlist__alloc_mmap(evlist, true);
 804                                 if (!maps)
 805                                         return -1;
 806                                 evlist->overwrite_mmap = maps;
 807                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 808                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 809                         }
 810                         mp->prot &= ~PROT_WRITE;
 811                 }
 812
 813                 if (evsel->system_wide && thread)
 814                         continue;
 815
 816                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
 817                 if (cpu == -1)
 818                         continue;
 819
 820                 fd = FD(evsel, cpu, thread);
 821
 822                 if (*output == -1) {
 823                         *output = fd;
 824
 825                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 826                                 return -1;
 827                 } else {
 828                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 829                                 return -1;
 830
 831                         perf_mmap__get(&maps[idx]);
 832                 }
 833
 834                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 835
 836                 /*
 837                  * The system_wide flag causes a selected event to be opened
 838                  * always without a pid.  Consequently it will never get a
 839                  * POLLHUP, but it is used for tracking in combination with
 840                  * other events, so it should not need to be polled anyway.
 841                  * Therefore don't add it for polling.
 842                  */
 843                 if (!evsel->system_wide &&
 844                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
 845                         perf_mmap__put(&maps[idx]);
 846                         return -1;
 847                 }
 848
 849                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 850                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 851                                                    fd) < 0)
 852                                 return -1;
 853                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 854                                                  thread);
 855                 }
 856         }
 857
 858         return 0;
 859 }
 860
 861 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 862                                      struct mmap_params *mp)
 863 {
 864         int cpu, thread;
 865         int nr_cpus = cpu_map__nr(evlist->cpus);
 866         int nr_threads = thread_map__nr(evlist->threads);
 867
 868         pr_debug2("perf event ring buffer mmapped per cpu\n");
 869         for (cpu = 0; cpu < nr_cpus; cpu++) {
 870                 int output = -1;
 871                 int output_overwrite = -1;
 872
 873                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 874                                               true);
 875
 876                 for (thread = 0; thread < nr_threads; thread++) {
 877                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 878                                                         thread, &output, &output_overwrite))
 879                                 goto out_unmap;
 880                 }
 881         }
 882
 883         return 0;
 884
 885 out_unmap:
 886         perf_evlist__munmap_nofree(evlist);
 887         return -1;
 888 }
 889
 890 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 891                                         struct mmap_params *mp)
 892 {
 893         int thread;
 894         int nr_threads = thread_map__nr(evlist->threads);
 895
 896         pr_debug2("perf event ring buffer mmapped per thread\n");
 897         for (thread = 0; thread < nr_threads; thread++) {
 898                 int output = -1;
 899                 int output_overwrite = -1;
 900
 901                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 902                                               false);
 903
 904                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 905                                                 &output, &output_overwrite))
 906                         goto out_unmap;
 907         }
 908
 909         return 0;
 910
 911 out_unmap:
 912         perf_evlist__munmap_nofree(evlist);
 913         return -1;
 914 }
 915
 916 unsigned long perf_event_mlock_kb_in_pages(void)
 917 {
 918         unsigned long pages;
 919         int max;
 920
 921         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 922                 /*
 923                  * Pick a once upon a time good value, i.e. things look
 924                  * strange since we can't read a sysctl value, but lets not
 925                  * die yet...
 926                  */
 927                 max = 512;
 928         } else {
 929                 max -= (page_size / 1024);
 930         }
 931
 932         pages = (max * 1024) / page_size;
 933         if (!is_power_of_2(pages))
 934                 pages = rounddown_pow_of_two(pages);
 935
 936         return pages;
 937 }
 938
 939 size_t perf_evlist__mmap_size(unsigned long pages)
 940 {
 941         if (pages == UINT_MAX)
 942                 pages = perf_event_mlock_kb_in_pages();
 943         else if (!is_power_of_2(pages))
 944                 return 0;
 945
 946         return (pages + 1) * page_size;
 947 }
 948
 949 static long parse_pages_arg(const char *str, unsigned long min,
 950                             unsigned long max)
 951 {
 952         unsigned long pages, val;
 953         static struct parse_tag tags[] = {
 954                 { .tag  = 'B', .mult = 1       },
 955                 { .tag  = 'K', .mult = 1 << 10 },
 956                 { .tag  = 'M', .mult = 1 << 20 },
 957                 { .tag  = 'G', .mult = 1 << 30 },
 958                 { .tag  = 0 },
 959         };
 960
 961         if (str == NULL)
 962                 return -EINVAL;
 963
 964         val = parse_tag_value(str, tags);
 965         if (val != (unsigned long) -1) {
 966                 /* we got file size value */
 967                 pages = PERF_ALIGN(val, page_size) / page_size;
 968         } else {
 969                 /* we got pages count value */
 970                 char *eptr;
 971                 pages = strtoul(str, &eptr, 10);
 972                 if (*eptr != '\0')
 973                         return -EINVAL;
 974         }
 975
 976         if (pages == 0 && min == 0) {
 977                 /* leave number of pages at 0 */
 978         } else if (!is_power_of_2(pages)) {
 979                 char buf[100];
 980
 981                 /* round pages up to next power of 2 */
 982                 pages = roundup_pow_of_two(pages);
 983                 if (!pages)
 984                         return -EINVAL;
 985
 986                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 987                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 988                         buf, pages);
 989         }
 990
 991         if (pages > max)
 992                 return -EINVAL;
 993
 994         return pages;
 995 }
 996
 997 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 998 {
 999         unsigned long max = UINT_MAX;
1000         long pages;
1001
1002         if (max > SIZE_MAX / page_size)
1003                 max = SIZE_MAX / page_size;
1004
1005         pages = parse_pages_arg(str, 1, max);
1006         if (pages < 0) {
1007                 pr_err("Invalid argument for --mmap_pages/-m\n");
1008                 return -1;
1009         }
1010
1011         *mmap_pages = pages;
1012         return 0;
1013 }
1014
1015 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1016                                   int unset __maybe_unused)
1017 {
1018         return __perf_evlist__parse_mmap_pages(opt->value, str);
1019 }
1020
1021 /**
1022  * perf_evlist__mmap_ex - Create mmaps to receive events.
1023  * @evlist: list of events
1024  * @pages: map length in pages
1025  * @overwrite: overwrite older events?
1026  * @auxtrace_pages - auxtrace map length in pages
1027  * @auxtrace_overwrite - overwrite older auxtrace data?
1028  *
1029  * If @overwrite is %false the user needs to signal event consumption using
1030  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1031  * automatically.
1032  *
1033  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1034  * consumption using auxtrace_mmap__write_tail().
1035  *
1036  * Return: %0 on success, negative error code otherwise.
1037  */
1038 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1039                          unsigned int auxtrace_pages,
1040                          bool auxtrace_overwrite, int nr_cblocks, int affinity)
1041 {
1042         struct perf_evsel *evsel;
1043         const struct cpu_map *cpus = evlist->cpus;
1044         const struct thread_map *threads = evlist->threads;
1045         /*
1046          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1047          * Its value is decided by evsel's write_backward.
1048          * So &mp should not be passed through const pointer.
1049          */
1050         struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity };
1051
1052         if (!evlist->mmap)
1053                 evlist->mmap = perf_evlist__alloc_mmap(evlist, false);
1054         if (!evlist->mmap)
1055                 return -ENOMEM;
1056
1057         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1058                 return -ENOMEM;
1059
1060         evlist->mmap_len = perf_evlist__mmap_size(pages);
1061         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1062         mp.mask = evlist->mmap_len - page_size - 1;
1063
1064         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1065                                    auxtrace_pages, auxtrace_overwrite);
1066
1067         evlist__for_each_entry(evlist, evsel) {
1068                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1069                     evsel->sample_id == NULL &&
1070                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1071                         return -ENOMEM;
1072         }
1073
1074         if (cpu_map__empty(cpus))
1075                 return perf_evlist__mmap_per_thread(evlist, &mp);
1076
1077         return perf_evlist__mmap_per_cpu(evlist, &mp);
1078 }
1079
1080 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages)
1081 {
1082         return perf_evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS);
1083 }
1084
1085 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1086 {
1087         bool all_threads = (target->per_thread && target->system_wide);
1088         struct cpu_map *cpus;
1089         struct thread_map *threads;
1090
1091         /*
1092          * If specify '-a' and '--per-thread' to perf record, perf record
1093          * will override '--per-thread'. target->per_thread = false and
1094          * target->system_wide = true.
1095          *
1096          * If specify '--per-thread' only to perf record,
1097          * target->per_thread = true and target->system_wide = false.
1098          *
1099          * So target->per_thread && target->system_wide is false.
1100          * For perf record, thread_map__new_str doesn't call
1101          * thread_map__new_all_cpus. That will keep perf record's
1102          * current behavior.
1103          *
1104          * For perf stat, it allows the case that target->per_thread and
1105          * target->system_wide are all true. It means to collect system-wide
1106          * per-thread data. thread_map__new_str will call
1107          * thread_map__new_all_cpus to enumerate all threads.
1108          */
1109         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1110                                       all_threads);
1111
1112         if (!threads)
1113                 return -1;
1114
1115         if (target__uses_dummy_map(target))
1116                 cpus = cpu_map__dummy_new();
1117         else
1118                 cpus = cpu_map__new(target->cpu_list);
1119
1120         if (!cpus)
1121                 goto out_delete_threads;
1122
1123         evlist->has_user_cpus = !!target->cpu_list;
1124
1125         perf_evlist__set_maps(evlist, cpus, threads);
1126
1127         return 0;
1128
1129 out_delete_threads:
1130         thread_map__put(threads);
1131         return -1;
1132 }
1133
1134 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1135                            struct thread_map *threads)
1136 {
1137         /*
1138          * Allow for the possibility that one or another of the maps isn't being
1139          * changed i.e. don't put it.  Note we are assuming the maps that are
1140          * being applied are brand new and evlist is taking ownership of the
1141          * original reference count of 1.  If that is not the case it is up to
1142          * the caller to increase the reference count.
1143          */
1144         if (cpus != evlist->cpus) {
1145                 cpu_map__put(evlist->cpus);
1146                 evlist->cpus = cpu_map__get(cpus);
1147         }
1148
1149         if (threads != evlist->threads) {
1150                 thread_map__put(evlist->threads);
1151                 evlist->threads = thread_map__get(threads);
1152         }
1153
1154         perf_evlist__propagate_maps(evlist);
1155 }
1156
1157 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1158                                    enum perf_event_sample_format bit)
1159 {
1160         struct perf_evsel *evsel;
1161
1162         evlist__for_each_entry(evlist, evsel)
1163                 __perf_evsel__set_sample_bit(evsel, bit);
1164 }
1165
1166 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1167                                      enum perf_event_sample_format bit)
1168 {
1169         struct perf_evsel *evsel;
1170
1171         evlist__for_each_entry(evlist, evsel)
1172                 __perf_evsel__reset_sample_bit(evsel, bit);
1173 }
1174
1175 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1176 {
1177         struct perf_evsel *evsel;
1178         int err = 0;
1179
1180         evlist__for_each_entry(evlist, evsel) {
1181                 if (evsel->filter == NULL)
1182                         continue;
1183
1184                 /*
1185                  * filters only work for tracepoint event, which doesn't have cpu limit.
1186                  * So evlist and evsel should always be same.
1187                  */
1188                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1189                 if (err) {
1190                         *err_evsel = evsel;
1191                         break;
1192                 }
1193         }
1194
1195         return err;
1196 }
1197
1198 int perf_evlist__set_tp_filter(struct perf_evlist *evlist, const char *filter)
1199 {
1200         struct perf_evsel *evsel;
1201         int err = 0;
1202
1203         evlist__for_each_entry(evlist, evsel) {
1204                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1205                         continue;
1206
1207                 err = perf_evsel__set_filter(evsel, filter);
1208                 if (err)
1209                         break;
1210         }
1211
1212         return err;
1213 }
1214
1215 int perf_evlist__set_tp_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1216 {
1217         char *filter;
1218         int ret = -1;
1219         size_t i;
1220
1221         for (i = 0; i < npids; ++i) {
1222                 if (i == 0) {
1223                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1224                                 return -1;
1225                 } else {
1226                         char *tmp;
1227
1228                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1229                                 goto out_free;
1230
1231                         free(filter);
1232                         filter = tmp;
1233                 }
1234         }
1235
1236         ret = perf_evlist__set_tp_filter(evlist, filter);
1237 out_free:
1238         free(filter);
1239         return ret;
1240 }
1241
1242 int perf_evlist__set_tp_filter_pid(struct perf_evlist *evlist, pid_t pid)
1243 {
1244         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1245 }
1246
1247 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1248 {
1249         struct perf_evsel *pos;
1250
1251         if (evlist->nr_entries == 1)
1252                 return true;
1253
1254         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1255                 return false;
1256
1257         evlist__for_each_entry(evlist, pos) {
1258                 if (pos->id_pos != evlist->id_pos ||
1259                     pos->is_pos != evlist->is_pos)
1260                         return false;
1261         }
1262
1263         return true;
1264 }
1265
1266 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1267 {
1268         struct perf_evsel *evsel;
1269
1270         if (evlist->combined_sample_type)
1271                 return evlist->combined_sample_type;
1272
1273         evlist__for_each_entry(evlist, evsel)
1274                 evlist->combined_sample_type |= evsel->attr.sample_type;
1275
1276         return evlist->combined_sample_type;
1277 }
1278
1279 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1280 {
1281         evlist->combined_sample_type = 0;
1282         return __perf_evlist__combined_sample_type(evlist);
1283 }
1284
1285 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1286 {
1287         struct perf_evsel *evsel;
1288         u64 branch_type = 0;
1289
1290         evlist__for_each_entry(evlist, evsel)
1291                 branch_type |= evsel->attr.branch_sample_type;
1292         return branch_type;
1293 }
1294
1295 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1296 {
1297         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1298         u64 read_format = first->attr.read_format;
1299         u64 sample_type = first->attr.sample_type;
1300
1301         evlist__for_each_entry(evlist, pos) {
1302                 if (read_format != pos->attr.read_format)
1303                         return false;
1304         }
1305
1306         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1307         if ((sample_type & PERF_SAMPLE_READ) &&
1308             !(read_format & PERF_FORMAT_ID)) {
1309                 return false;
1310         }
1311
1312         return true;
1313 }
1314
1315 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1316 {
1317         struct perf_evsel *first = perf_evlist__first(evlist);
1318         return first->attr.read_format;
1319 }
1320
1321 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1322 {
1323         struct perf_evsel *first = perf_evlist__first(evlist);
1324         struct perf_sample *data;
1325         u64 sample_type;
1326         u16 size = 0;
1327
1328         if (!first->attr.sample_id_all)
1329                 goto out;
1330
1331         sample_type = first->attr.sample_type;
1332
1333         if (sample_type & PERF_SAMPLE_TID)
1334                 size += sizeof(data->tid) * 2;
1335
1336        if (sample_type & PERF_SAMPLE_TIME)
1337                 size += sizeof(data->time);
1338
1339         if (sample_type & PERF_SAMPLE_ID)
1340                 size += sizeof(data->id);
1341
1342         if (sample_type & PERF_SAMPLE_STREAM_ID)
1343                 size += sizeof(data->stream_id);
1344
1345         if (sample_type & PERF_SAMPLE_CPU)
1346                 size += sizeof(data->cpu) * 2;
1347
1348         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1349                 size += sizeof(data->id);
1350 out:
1351         return size;
1352 }
1353
1354 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1355 {
1356         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1357
1358         evlist__for_each_entry_continue(evlist, pos) {
1359                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1360                         return false;
1361         }
1362
1363         return true;
1364 }
1365
1366 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1367 {
1368         struct perf_evsel *first = perf_evlist__first(evlist);
1369         return first->attr.sample_id_all;
1370 }
1371
1372 void perf_evlist__set_selected(struct perf_evlist *evlist,
1373                                struct perf_evsel *evsel)
1374 {
1375         evlist->selected = evsel;
1376 }
1377
1378 void perf_evlist__close(struct perf_evlist *evlist)
1379 {
1380         struct perf_evsel *evsel;
1381
1382         evlist__for_each_entry_reverse(evlist, evsel)
1383                 perf_evsel__close(evsel);
1384 }
1385
1386 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1387 {
1388         struct cpu_map    *cpus;
1389         struct thread_map *threads;
1390         int err = -ENOMEM;
1391
1392         /*
1393          * Try reading /sys/devices/system/cpu/online to get
1394          * an all cpus map.
1395          *
1396          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1397          * code needs an overhaul to properly forward the
1398          * error, and we may not want to do that fallback to a
1399          * default cpu identity map :-\
1400          */
1401         cpus = cpu_map__new(NULL);
1402         if (!cpus)
1403                 goto out;
1404
1405         threads = thread_map__new_dummy();
1406         if (!threads)
1407                 goto out_put;
1408
1409         perf_evlist__set_maps(evlist, cpus, threads);
1410 out:
1411         return err;
1412 out_put:
1413         cpu_map__put(cpus);
1414         goto out;
1415 }
1416
1417 int perf_evlist__open(struct perf_evlist *evlist)
1418 {
1419         struct perf_evsel *evsel;
1420         int err;
1421
1422         /*
1423          * Default: one fd per CPU, all threads, aka systemwide
1424          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1425          */
1426         if (evlist->threads == NULL && evlist->cpus == NULL) {
1427                 err = perf_evlist__create_syswide_maps(evlist);
1428                 if (err < 0)
1429                         goto out_err;
1430         }
1431
1432         perf_evlist__update_id_pos(evlist);
1433
1434         evlist__for_each_entry(evlist, evsel) {
1435                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1436                 if (err < 0)
1437                         goto out_err;
1438         }
1439
1440         return 0;
1441 out_err:
1442         perf_evlist__close(evlist);
1443         errno = -err;
1444         return err;
1445 }
1446
1447 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1448                                   const char *argv[], bool pipe_output,
1449                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1450 {
1451         int child_ready_pipe[2], go_pipe[2];
1452         char bf;
1453
1454         if (pipe(child_ready_pipe) < 0) {
1455                 perror("failed to create 'ready' pipe");
1456                 return -1;
1457         }
1458
1459         if (pipe(go_pipe) < 0) {
1460                 perror("failed to create 'go' pipe");
1461                 goto out_close_ready_pipe;
1462         }
1463
1464         evlist->workload.pid = fork();
1465         if (evlist->workload.pid < 0) {
1466                 perror("failed to fork");
1467                 goto out_close_pipes;
1468         }
1469
1470         if (!evlist->workload.pid) {
1471                 int ret;
1472
1473                 if (pipe_output)
1474                         dup2(2, 1);
1475
1476                 signal(SIGTERM, SIG_DFL);
1477
1478                 close(child_ready_pipe[0]);
1479                 close(go_pipe[1]);
1480                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1481
1482                 /*
1483                  * Tell the parent we're ready to go
1484                  */
1485                 close(child_ready_pipe[1]);
1486
1487                 /*
1488                  * Wait until the parent tells us to go.
1489                  */
1490                 ret = read(go_pipe[0], &bf, 1);
1491                 /*
1492                  * The parent will ask for the execvp() to be performed by
1493                  * writing exactly one byte, in workload.cork_fd, usually via
1494                  * perf_evlist__start_workload().
1495                  *
1496                  * For cancelling the workload without actually running it,
1497                  * the parent will just close workload.cork_fd, without writing
1498                  * anything, i.e. read will return zero and we just exit()
1499                  * here.
1500                  */
1501                 if (ret != 1) {
1502                         if (ret == -1)
1503                                 perror("unable to read pipe");
1504                         exit(ret);
1505                 }
1506
1507                 execvp(argv[0], (char **)argv);
1508
1509                 if (exec_error) {
1510                         union sigval val;
1511
1512                         val.sival_int = errno;
1513                         if (sigqueue(getppid(), SIGUSR1, val))
1514                                 perror(argv[0]);
1515                 } else
1516                         perror(argv[0]);
1517                 exit(-1);
1518         }
1519
1520         if (exec_error) {
1521                 struct sigaction act = {
1522                         .sa_flags     = SA_SIGINFO,
1523                         .sa_sigaction = exec_error,
1524                 };
1525                 sigaction(SIGUSR1, &act, NULL);
1526         }
1527
1528         if (target__none(target)) {
1529                 if (evlist->threads == NULL) {
1530                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1531                                 __func__, __LINE__);
1532                         goto out_close_pipes;
1533                 }
1534                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1535         }
1536
1537         close(child_ready_pipe[1]);
1538         close(go_pipe[0]);
1539         /*
1540          * wait for child to settle
1541          */
1542         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1543                 perror("unable to read pipe");
1544                 goto out_close_pipes;
1545         }
1546
1547         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1548         evlist->workload.cork_fd = go_pipe[1];
1549         close(child_ready_pipe[0]);
1550         return 0;
1551
1552 out_close_pipes:
1553         close(go_pipe[0]);
1554         close(go_pipe[1]);
1555 out_close_ready_pipe:
1556         close(child_ready_pipe[0]);
1557         close(child_ready_pipe[1]);
1558         return -1;
1559 }
1560
1561 int perf_evlist__start_workload(struct perf_evlist *evlist)
1562 {
1563         if (evlist->workload.cork_fd > 0) {
1564                 char bf = 0;
1565                 int ret;
1566                 /*
1567                  * Remove the cork, let it rip!
1568                  */
1569                 ret = write(evlist->workload.cork_fd, &bf, 1);
1570                 if (ret < 0)
1571                         perror("unable to write to pipe");
1572
1573                 close(evlist->workload.cork_fd);
1574                 return ret;
1575         }
1576
1577         return 0;
1578 }
1579
1580 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1581                               struct perf_sample *sample)
1582 {
1583         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1584
1585         if (!evsel)
1586                 return -EFAULT;
1587         return perf_evsel__parse_sample(evsel, event, sample);
1588 }
1589
1590 int perf_evlist__parse_sample_timestamp(struct perf_evlist *evlist,
1591                                         union perf_event *event,
1592                                         u64 *timestamp)
1593 {
1594         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1595
1596         if (!evsel)
1597                 return -EFAULT;
1598         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1599 }
1600
1601 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1602 {
1603         struct perf_evsel *evsel;
1604         size_t printed = 0;
1605
1606         evlist__for_each_entry(evlist, evsel) {
1607                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1608                                    perf_evsel__name(evsel));
1609         }
1610
1611         return printed + fprintf(fp, "\n");
1612 }
1613
1614 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1615                                int err, char *buf, size_t size)
1616 {
1617         int printed, value;
1618         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1619
1620         switch (err) {
1621         case EACCES:
1622         case EPERM:
1623                 printed = scnprintf(buf, size,
1624                                     "Error:\t%s.\n"
1625                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1626
1627                 value = perf_event_paranoid();
1628
1629                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1630
1631                 if (value >= 2) {
1632                         printed += scnprintf(buf + printed, size - printed,
1633                                              "For your workloads it needs to be <= 1\nHint:\t");
1634                 }
1635                 printed += scnprintf(buf + printed, size - printed,
1636                                      "For system wide tracing it needs to be set to -1.\n");
1637
1638                 printed += scnprintf(buf + printed, size - printed,
1639                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1640                                     "Hint:\tThe current value is %d.", value);
1641                 break;
1642         case EINVAL: {
1643                 struct perf_evsel *first = perf_evlist__first(evlist);
1644                 int max_freq;
1645
1646                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1647                         goto out_default;
1648
1649                 if (first->attr.sample_freq < (u64)max_freq)
1650                         goto out_default;
1651
1652                 printed = scnprintf(buf, size,
1653                                     "Error:\t%s.\n"
1654                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1655                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1656                                     emsg, max_freq, first->attr.sample_freq);
1657                 break;
1658         }
1659         default:
1660 out_default:
1661                 scnprintf(buf, size, "%s", emsg);
1662                 break;
1663         }
1664
1665         return 0;
1666 }
1667
1668 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1669 {
1670         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1671         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1672
1673         switch (err) {
1674         case EPERM:
1675                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1676                 printed += scnprintf(buf + printed, size - printed,
1677                                      "Error:\t%s.\n"
1678                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1679                                      "Hint:\tTried using %zd kB.\n",
1680                                      emsg, pages_max_per_user, pages_attempted);
1681
1682                 if (pages_attempted >= pages_max_per_user) {
1683                         printed += scnprintf(buf + printed, size - printed,
1684                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1685                                              pages_max_per_user + pages_attempted);
1686                 }
1687
1688                 printed += scnprintf(buf + printed, size - printed,
1689                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1690                 break;
1691         default:
1692                 scnprintf(buf, size, "%s", emsg);
1693                 break;
1694         }
1695
1696         return 0;
1697 }
1698
1699 void perf_evlist__to_front(struct perf_evlist *evlist,
1700                            struct perf_evsel *move_evsel)
1701 {
1702         struct perf_evsel *evsel, *n;
1703         LIST_HEAD(move);
1704
1705         if (move_evsel == perf_evlist__first(evlist))
1706                 return;
1707
1708         evlist__for_each_entry_safe(evlist, n, evsel) {
1709                 if (evsel->leader == move_evsel->leader)
1710                         list_move_tail(&evsel->node, &move);
1711         }
1712
1713         list_splice(&move, &evlist->entries);
1714 }
1715
1716 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1717                                      struct perf_evsel *tracking_evsel)
1718 {
1719         struct perf_evsel *evsel;
1720
1721         if (tracking_evsel->tracking)
1722                 return;
1723
1724         evlist__for_each_entry(evlist, evsel) {
1725                 if (evsel != tracking_evsel)
1726                         evsel->tracking = false;
1727         }
1728
1729         tracking_evsel->tracking = true;
1730 }
1731
1732 struct perf_evsel *
1733 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1734                                const char *str)
1735 {
1736         struct perf_evsel *evsel;
1737
1738         evlist__for_each_entry(evlist, evsel) {
1739                 if (!evsel->name)
1740                         continue;
1741                 if (strcmp(str, evsel->name) == 0)
1742                         return evsel;
1743         }
1744
1745         return NULL;
1746 }
1747
1748 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1749                                   enum bkw_mmap_state state)
1750 {
1751         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1752         enum action {
1753                 NONE,
1754                 PAUSE,
1755                 RESUME,
1756         } action = NONE;
1757
1758         if (!evlist->overwrite_mmap)
1759                 return;
1760
1761         switch (old_state) {
1762         case BKW_MMAP_NOTREADY: {
1763                 if (state != BKW_MMAP_RUNNING)
1764                         goto state_err;
1765                 break;
1766         }
1767         case BKW_MMAP_RUNNING: {
1768                 if (state != BKW_MMAP_DATA_PENDING)
1769                         goto state_err;
1770                 action = PAUSE;
1771                 break;
1772         }
1773         case BKW_MMAP_DATA_PENDING: {
1774                 if (state != BKW_MMAP_EMPTY)
1775                         goto state_err;
1776                 break;
1777         }
1778         case BKW_MMAP_EMPTY: {
1779                 if (state != BKW_MMAP_RUNNING)
1780                         goto state_err;
1781                 action = RESUME;
1782                 break;
1783         }
1784         default:
1785                 WARN_ONCE(1, "Shouldn't get there\n");
1786         }
1787
1788         evlist->bkw_mmap_state = state;
1789
1790         switch (action) {
1791         case PAUSE:
1792                 perf_evlist__pause(evlist);
1793                 break;
1794         case RESUME:
1795                 perf_evlist__resume(evlist);
1796                 break;
1797         case NONE:
1798         default:
1799                 break;
1800         }
1801
1802 state_err:
1803         return;
1804 }
1805
1806 bool perf_evlist__exclude_kernel(struct perf_evlist *evlist)
1807 {
1808         struct perf_evsel *evsel;
1809
1810         evlist__for_each_entry(evlist, evsel) {
1811                 if (!evsel->attr.exclude_kernel)
1812                         return false;
1813         }
1814
1815         return true;
1816 }
1817
1818 /*
1819  * Events in data file are not collect in groups, but we still want
1820  * the group display. Set the artificial group and set the leader's
1821  * forced_leader flag to notify the display code.
1822  */
1823 void perf_evlist__force_leader(struct perf_evlist *evlist)
1824 {
1825         if (!evlist->nr_groups) {
1826                 struct perf_evsel *leader = perf_evlist__first(evlist);
1827
1828                 perf_evlist__set_leader(evlist);
1829                 leader->forced_leader = true;
1830         }
1831 }
1832
1833 struct perf_evsel *perf_evlist__reset_weak_group(struct perf_evlist *evsel_list,
1834                                                  struct perf_evsel *evsel)
1835 {
1836         struct perf_evsel *c2, *leader;
1837         bool is_open = true;
1838
1839         leader = evsel->leader;
1840         pr_debug("Weak group for %s/%d failed\n",
1841                         leader->name, leader->nr_members);
1842
1843         /*
1844          * for_each_group_member doesn't work here because it doesn't
1845          * include the first entry.
1846          */
1847         evlist__for_each_entry(evsel_list, c2) {
1848                 if (c2 == evsel)
1849                         is_open = false;
1850                 if (c2->leader == leader) {
1851                         if (is_open)
1852                                 perf_evsel__close(c2);
1853                         c2->leader = c2;
1854                         c2->nr_members = 0;
1855                 }
1856         }
1857         return leader;
1858 }