tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <poll.h>
  12 #include "cpumap.h"
  13 #include "thread_map.h"
  14 #include "target.h"
  15 #include "evlist.h"
  16 #include "evsel.h"
  17 #include "debug.h"
  18 #include <unistd.h>
  19
  20 #include "parse-events.h"
  21 #include <subcmd/parse-options.h>
  22
  23 #include <sys/mman.h>
  24
  25 #include <linux/bitops.h>
  26 #include <linux/hash.h>
  27 #include <linux/log2.h>
  28 #include <linux/err.h>
  29
  30 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
  31 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
  32
  33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  34 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  35
  36 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  37                        struct thread_map *threads)
  38 {
  39         int i;
  40
  41         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  42                 INIT_HLIST_HEAD(&evlist->heads[i]);
  43         INIT_LIST_HEAD(&evlist->entries);
  44         perf_evlist__set_maps(evlist, cpus, threads);
  45         fdarray__init(&evlist->pollfd, 64);
  46         evlist->workload.pid = -1;
  47 }
  48
  49 struct perf_evlist *perf_evlist__new(void)
  50 {
  51         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  52
  53         if (evlist != NULL)
  54                 perf_evlist__init(evlist, NULL, NULL);
  55
  56         return evlist;
  57 }
  58
  59 struct perf_evlist *perf_evlist__new_default(void)
  60 {
  61         struct perf_evlist *evlist = perf_evlist__new();
  62
  63         if (evlist && perf_evlist__add_default(evlist)) {
  64                 perf_evlist__delete(evlist);
  65                 evlist = NULL;
  66         }
  67
  68         return evlist;
  69 }
  70
  71 struct perf_evlist *perf_evlist__new_dummy(void)
  72 {
  73         struct perf_evlist *evlist = perf_evlist__new();
  74
  75         if (evlist && perf_evlist__add_dummy(evlist)) {
  76                 perf_evlist__delete(evlist);
  77                 evlist = NULL;
  78         }
  79
  80         return evlist;
  81 }
  82
  83 /**
  84  * perf_evlist__set_id_pos - set the positions of event ids.
  85  * @evlist: selected event list
  86  *
  87  * Events with compatible sample types all have the same id_pos
  88  * and is_pos.  For convenience, put a copy on evlist.
  89  */
  90 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  91 {
  92         struct perf_evsel *first = perf_evlist__first(evlist);
  93
  94         evlist->id_pos = first->id_pos;
  95         evlist->is_pos = first->is_pos;
  96 }
  97
  98 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  99 {
 100         struct perf_evsel *evsel;
 101
 102         evlist__for_each(evlist, evsel)
 103                 perf_evsel__calc_id_pos(evsel);
 104
 105         perf_evlist__set_id_pos(evlist);
 106 }
 107
 108 static void perf_evlist__purge(struct perf_evlist *evlist)
 109 {
 110         struct perf_evsel *pos, *n;
 111
 112         evlist__for_each_safe(evlist, n, pos) {
 113                 list_del_init(&pos->node);
 114                 pos->evlist = NULL;
 115                 perf_evsel__delete(pos);
 116         }
 117
 118         evlist->nr_entries = 0;
 119 }
 120
 121 void perf_evlist__exit(struct perf_evlist *evlist)
 122 {
 123         zfree(&evlist->mmap);
 124         fdarray__exit(&evlist->pollfd);
 125 }
 126
 127 void perf_evlist__delete(struct perf_evlist *evlist)
 128 {
 129         perf_evlist__munmap(evlist);
 130         perf_evlist__close(evlist);
 131         cpu_map__put(evlist->cpus);
 132         thread_map__put(evlist->threads);
 133         evlist->cpus = NULL;
 134         evlist->threads = NULL;
 135         perf_evlist__purge(evlist);
 136         perf_evlist__exit(evlist);
 137         free(evlist);
 138 }
 139
 140 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 141                                           struct perf_evsel *evsel)
 142 {
 143         /*
 144          * We already have cpus for evsel (via PMU sysfs) so
 145          * keep it, if there's no target cpu list defined.
 146          */
 147         if (!evsel->own_cpus || evlist->has_user_cpus) {
 148                 cpu_map__put(evsel->cpus);
 149                 evsel->cpus = cpu_map__get(evlist->cpus);
 150         } else if (evsel->cpus != evsel->own_cpus) {
 151                 cpu_map__put(evsel->cpus);
 152                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 153         }
 154
 155         thread_map__put(evsel->threads);
 156         evsel->threads = thread_map__get(evlist->threads);
 157 }
 158
 159 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 160 {
 161         struct perf_evsel *evsel;
 162
 163         evlist__for_each(evlist, evsel)
 164                 __perf_evlist__propagate_maps(evlist, evsel);
 165 }
 166
 167 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 168 {
 169         entry->evlist = evlist;
 170         list_add_tail(&entry->node, &evlist->entries);
 171         entry->idx = evlist->nr_entries;
 172         entry->tracking = !entry->idx;
 173
 174         if (!evlist->nr_entries++)
 175                 perf_evlist__set_id_pos(evlist);
 176
 177         __perf_evlist__propagate_maps(evlist, entry);
 178 }
 179
 180 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 181 {
 182         evsel->evlist = NULL;
 183         list_del_init(&evsel->node);
 184         evlist->nr_entries -= 1;
 185 }
 186
 187 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 188                                    struct list_head *list)
 189 {
 190         struct perf_evsel *evsel, *temp;
 191
 192         __evlist__for_each_safe(list, temp, evsel) {
 193                 list_del_init(&evsel->node);
 194                 perf_evlist__add(evlist, evsel);
 195         }
 196 }
 197
 198 void __perf_evlist__set_leader(struct list_head *list)
 199 {
 200         struct perf_evsel *evsel, *leader;
 201
 202         leader = list_entry(list->next, struct perf_evsel, node);
 203         evsel = list_entry(list->prev, struct perf_evsel, node);
 204
 205         leader->nr_members = evsel->idx - leader->idx + 1;
 206
 207         __evlist__for_each(list, evsel) {
 208                 evsel->leader = leader;
 209         }
 210 }
 211
 212 void perf_evlist__set_leader(struct perf_evlist *evlist)
 213 {
 214         if (evlist->nr_entries) {
 215                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 216                 __perf_evlist__set_leader(&evlist->entries);
 217         }
 218 }
 219
 220 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 221 {
 222         attr->precise_ip = 3;
 223
 224         while (attr->precise_ip != 0) {
 225                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 226                 if (fd != -1) {
 227                         close(fd);
 228                         break;
 229                 }
 230                 --attr->precise_ip;
 231         }
 232 }
 233
 234 int perf_evlist__add_default(struct perf_evlist *evlist)
 235 {
 236         struct perf_event_attr attr = {
 237                 .type = PERF_TYPE_HARDWARE,
 238                 .config = PERF_COUNT_HW_CPU_CYCLES,
 239         };
 240         struct perf_evsel *evsel;
 241
 242         event_attr_init(&attr);
 243
 244         perf_event_attr__set_max_precise_ip(&attr);
 245
 246         evsel = perf_evsel__new(&attr);
 247         if (evsel == NULL)
 248                 goto error;
 249
 250         /* use asprintf() because free(evsel) assumes name is allocated */
 251         if (asprintf(&evsel->name, "cycles%.*s",
 252                      attr.precise_ip ? attr.precise_ip + 1 : 0, ":ppp") < 0)
 253                 goto error_free;
 254
 255         perf_evlist__add(evlist, evsel);
 256         return 0;
 257 error_free:
 258         perf_evsel__delete(evsel);
 259 error:
 260         return -ENOMEM;
 261 }
 262
 263 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 264 {
 265         struct perf_event_attr attr = {
 266                 .type   = PERF_TYPE_SOFTWARE,
 267                 .config = PERF_COUNT_SW_DUMMY,
 268                 .size   = sizeof(attr), /* to capture ABI version */
 269         };
 270         struct perf_evsel *evsel = perf_evsel__new(&attr);
 271
 272         if (evsel == NULL)
 273                 return -ENOMEM;
 274
 275         perf_evlist__add(evlist, evsel);
 276         return 0;
 277 }
 278
 279 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 280                                   struct perf_event_attr *attrs, size_t nr_attrs)
 281 {
 282         struct perf_evsel *evsel, *n;
 283         LIST_HEAD(head);
 284         size_t i;
 285
 286         for (i = 0; i < nr_attrs; i++) {
 287                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 288                 if (evsel == NULL)
 289                         goto out_delete_partial_list;
 290                 list_add_tail(&evsel->node, &head);
 291         }
 292
 293         perf_evlist__splice_list_tail(evlist, &head);
 294
 295         return 0;
 296
 297 out_delete_partial_list:
 298         __evlist__for_each_safe(&head, n, evsel)
 299                 perf_evsel__delete(evsel);
 300         return -1;
 301 }
 302
 303 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 304                                      struct perf_event_attr *attrs, size_t nr_attrs)
 305 {
 306         size_t i;
 307
 308         for (i = 0; i < nr_attrs; i++)
 309                 event_attr_init(attrs + i);
 310
 311         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 312 }
 313
 314 struct perf_evsel *
 315 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 316 {
 317         struct perf_evsel *evsel;
 318
 319         evlist__for_each(evlist, evsel) {
 320                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 321                     (int)evsel->attr.config == id)
 322                         return evsel;
 323         }
 324
 325         return NULL;
 326 }
 327
 328 struct perf_evsel *
 329 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 330                                      const char *name)
 331 {
 332         struct perf_evsel *evsel;
 333
 334         evlist__for_each(evlist, evsel) {
 335                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 336                     (strcmp(evsel->name, name) == 0))
 337                         return evsel;
 338         }
 339
 340         return NULL;
 341 }
 342
 343 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 344                            const char *sys, const char *name, void *handler)
 345 {
 346         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 347
 348         if (IS_ERR(evsel))
 349                 return -1;
 350
 351         evsel->handler = handler;
 352         perf_evlist__add(evlist, evsel);
 353         return 0;
 354 }
 355
 356 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 357                                    struct perf_evsel *evsel)
 358 {
 359         if (evsel->system_wide)
 360                 return 1;
 361         else
 362                 return thread_map__nr(evlist->threads);
 363 }
 364
 365 void perf_evlist__disable(struct perf_evlist *evlist)
 366 {
 367         struct perf_evsel *pos;
 368
 369         evlist__for_each(evlist, pos) {
 370                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 371                         continue;
 372                 perf_evsel__disable(pos);
 373         }
 374
 375         evlist->enabled = false;
 376 }
 377
 378 void perf_evlist__enable(struct perf_evlist *evlist)
 379 {
 380         struct perf_evsel *pos;
 381
 382         evlist__for_each(evlist, pos) {
 383                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 384                         continue;
 385                 perf_evsel__enable(pos);
 386         }
 387
 388         evlist->enabled = true;
 389 }
 390
 391 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 392 {
 393         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 394 }
 395
 396 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 397                                          struct perf_evsel *evsel, int cpu)
 398 {
 399         int thread, err;
 400         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 401
 402         if (!evsel->fd)
 403                 return -EINVAL;
 404
 405         for (thread = 0; thread < nr_threads; thread++) {
 406                 err = ioctl(FD(evsel, cpu, thread),
 407                             PERF_EVENT_IOC_ENABLE, 0);
 408                 if (err)
 409                         return err;
 410         }
 411         return 0;
 412 }
 413
 414 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 415                                             struct perf_evsel *evsel,
 416                                             int thread)
 417 {
 418         int cpu, err;
 419         int nr_cpus = cpu_map__nr(evlist->cpus);
 420
 421         if (!evsel->fd)
 422                 return -EINVAL;
 423
 424         for (cpu = 0; cpu < nr_cpus; cpu++) {
 425                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 426                 if (err)
 427                         return err;
 428         }
 429         return 0;
 430 }
 431
 432 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 433                                   struct perf_evsel *evsel, int idx)
 434 {
 435         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 436
 437         if (per_cpu_mmaps)
 438                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 439         else
 440                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 441 }
 442
 443 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 444 {
 445         int nr_cpus = cpu_map__nr(evlist->cpus);
 446         int nr_threads = thread_map__nr(evlist->threads);
 447         int nfds = 0;
 448         struct perf_evsel *evsel;
 449
 450         evlist__for_each(evlist, evsel) {
 451                 if (evsel->system_wide)
 452                         nfds += nr_cpus;
 453                 else
 454                         nfds += nr_cpus * nr_threads;
 455         }
 456
 457         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 458             fdarray__grow(&evlist->pollfd, nfds) < 0)
 459                 return -ENOMEM;
 460
 461         return 0;
 462 }
 463
 464 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
 465 {
 466         int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
 467         /*
 468          * Save the idx so that when we filter out fds POLLHUP'ed we can
 469          * close the associated evlist->mmap[] entry.
 470          */
 471         if (pos >= 0) {
 472                 evlist->pollfd.priv[pos].idx = idx;
 473
 474                 fcntl(fd, F_SETFL, O_NONBLOCK);
 475         }
 476
 477         return pos;
 478 }
 479
 480 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 481 {
 482         return __perf_evlist__add_pollfd(evlist, fd, -1);
 483 }
 484
 485 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
 486 {
 487         struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
 488
 489         perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
 490 }
 491
 492 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 493 {
 494         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 495                                perf_evlist__munmap_filtered);
 496 }
 497
 498 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 499 {
 500         return fdarray__poll(&evlist->pollfd, timeout);
 501 }
 502
 503 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 504                                  struct perf_evsel *evsel,
 505                                  int cpu, int thread, u64 id)
 506 {
 507         int hash;
 508         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 509
 510         sid->id = id;
 511         sid->evsel = evsel;
 512         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 513         hlist_add_head(&sid->node, &evlist->heads[hash]);
 514 }
 515
 516 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 517                          int cpu, int thread, u64 id)
 518 {
 519         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 520         evsel->id[evsel->ids++] = id;
 521 }
 522
 523 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 524                            struct perf_evsel *evsel,
 525                            int cpu, int thread, int fd)
 526 {
 527         u64 read_data[4] = { 0, };
 528         int id_idx = 1; /* The first entry is the counter value */
 529         u64 id;
 530         int ret;
 531
 532         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 533         if (!ret)
 534                 goto add;
 535
 536         if (errno != ENOTTY)
 537                 return -1;
 538
 539         /* Legacy way to get event id.. All hail to old kernels! */
 540
 541         /*
 542          * This way does not work with group format read, so bail
 543          * out in that case.
 544          */
 545         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 546                 return -1;
 547
 548         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 549             read(fd, &read_data, sizeof(read_data)) == -1)
 550                 return -1;
 551
 552         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 553                 ++id_idx;
 554         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 555                 ++id_idx;
 556
 557         id = read_data[id_idx];
 558
 559  add:
 560         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 561         return 0;
 562 }
 563
 564 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 565                                      struct perf_evsel *evsel, int idx, int cpu,
 566                                      int thread)
 567 {
 568         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 569         sid->idx = idx;
 570         if (evlist->cpus && cpu >= 0)
 571                 sid->cpu = evlist->cpus->map[cpu];
 572         else
 573                 sid->cpu = -1;
 574         if (!evsel->system_wide && evlist->threads && thread >= 0)
 575                 sid->tid = thread_map__pid(evlist->threads, thread);
 576         else
 577                 sid->tid = -1;
 578 }
 579
 580 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 581 {
 582         struct hlist_head *head;
 583         struct perf_sample_id *sid;
 584         int hash;
 585
 586         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 587         head = &evlist->heads[hash];
 588
 589         hlist_for_each_entry(sid, head, node)
 590                 if (sid->id == id)
 591                         return sid;
 592
 593         return NULL;
 594 }
 595
 596 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 597 {
 598         struct perf_sample_id *sid;
 599
 600         if (evlist->nr_entries == 1 || !id)
 601                 return perf_evlist__first(evlist);
 602
 603         sid = perf_evlist__id2sid(evlist, id);
 604         if (sid)
 605                 return sid->evsel;
 606
 607         if (!perf_evlist__sample_id_all(evlist))
 608                 return perf_evlist__first(evlist);
 609
 610         return NULL;
 611 }
 612
 613 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 614                                                 u64 id)
 615 {
 616         struct perf_sample_id *sid;
 617
 618         if (!id)
 619                 return NULL;
 620
 621         sid = perf_evlist__id2sid(evlist, id);
 622         if (sid)
 623                 return sid->evsel;
 624
 625         return NULL;
 626 }
 627
 628 static int perf_evlist__event2id(struct perf_evlist *evlist,
 629                                  union perf_event *event, u64 *id)
 630 {
 631         const u64 *array = event->sample.array;
 632         ssize_t n;
 633
 634         n = (event->header.size - sizeof(event->header)) >> 3;
 635
 636         if (event->header.type == PERF_RECORD_SAMPLE) {
 637                 if (evlist->id_pos >= n)
 638                         return -1;
 639                 *id = array[evlist->id_pos];
 640         } else {
 641                 if (evlist->is_pos > n)
 642                         return -1;
 643                 n -= evlist->is_pos;
 644                 *id = array[n];
 645         }
 646         return 0;
 647 }
 648
 649 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 650                                                    union perf_event *event)
 651 {
 652         struct perf_evsel *first = perf_evlist__first(evlist);
 653         struct hlist_head *head;
 654         struct perf_sample_id *sid;
 655         int hash;
 656         u64 id;
 657
 658         if (evlist->nr_entries == 1)
 659                 return first;
 660
 661         if (!first->attr.sample_id_all &&
 662             event->header.type != PERF_RECORD_SAMPLE)
 663                 return first;
 664
 665         if (perf_evlist__event2id(evlist, event, &id))
 666                 return NULL;
 667
 668         /* Synthesized events have an id of zero */
 669         if (!id)
 670                 return first;
 671
 672         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 673         head = &evlist->heads[hash];
 674
 675         hlist_for_each_entry(sid, head, node) {
 676                 if (sid->id == id)
 677                         return sid->evsel;
 678         }
 679         return NULL;
 680 }
 681
 682 /* When check_messup is true, 'end' must points to a good entry */
 683 static union perf_event *
 684 perf_mmap__read(struct perf_mmap *md, bool check_messup, u64 start,
 685                 u64 end, u64 *prev)
 686 {
 687         unsigned char *data = md->base + page_size;
 688         union perf_event *event = NULL;
 689         int diff = end - start;
 690
 691         if (check_messup) {
 692                 /*
 693                  * If we're further behind than half the buffer, there's a chance
 694                  * the writer will bite our tail and mess up the samples under us.
 695                  *
 696                  * If we somehow ended up ahead of the 'end', we got messed up.
 697                  *
 698                  * In either case, truncate and restart at 'end'.
 699                  */
 700                 if (diff > md->mask / 2 || diff < 0) {
 701                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 702
 703                         /*
 704                          * 'end' points to a known good entry, start there.
 705                          */
 706                         start = end;
 707                         diff = 0;
 708                 }
 709         }
 710
 711         if (diff >= (int)sizeof(event->header)) {
 712                 size_t size;
 713
 714                 event = (union perf_event *)&data[start & md->mask];
 715                 size = event->header.size;
 716
 717                 if (size < sizeof(event->header) || diff < (int)size) {
 718                         event = NULL;
 719                         goto broken_event;
 720                 }
 721
 722                 /*
 723                  * Event straddles the mmap boundary -- header should always
 724                  * be inside due to u64 alignment of output.
 725                  */
 726                 if ((start & md->mask) + size != ((start + size) & md->mask)) {
 727                         unsigned int offset = start;
 728                         unsigned int len = min(sizeof(*event), size), cpy;
 729                         void *dst = md->event_copy;
 730
 731                         do {
 732                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 733                                 memcpy(dst, &data[offset & md->mask], cpy);
 734                                 offset += cpy;
 735                                 dst += cpy;
 736                                 len -= cpy;
 737                         } while (len);
 738
 739                         event = (union perf_event *) md->event_copy;
 740                 }
 741
 742                 start += size;
 743         }
 744
 745 broken_event:
 746         if (prev)
 747                 *prev = start;
 748
 749         return event;
 750 }
 751
 752 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 753 {
 754         struct perf_mmap *md = &evlist->mmap[idx];
 755         u64 head;
 756         u64 old = md->prev;
 757
 758         /*
 759          * Check if event was unmapped due to a POLLHUP/POLLERR.
 760          */
 761         if (!atomic_read(&md->refcnt))
 762                 return NULL;
 763
 764         head = perf_mmap__read_head(md);
 765
 766         return perf_mmap__read(md, evlist->overwrite, old, head, &md->prev);
 767 }
 768
 769 union perf_event *
 770 perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
 771 {
 772         struct perf_mmap *md = &evlist->mmap[idx];
 773         u64 head, end;
 774         u64 start = md->prev;
 775
 776         /*
 777          * Check if event was unmapped due to a POLLHUP/POLLERR.
 778          */
 779         if (!atomic_read(&md->refcnt))
 780                 return NULL;
 781
 782         head = perf_mmap__read_head(md);
 783         if (!head)
 784                 return NULL;
 785
 786         /*
 787          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 788          * it each time when kernel writes to it, so in fact 'head' is
 789          * negative. 'end' pointer is made manually by adding the size of
 790          * the ring buffer to 'head' pointer, means the validate data can
 791          * read is the whole ring buffer. If 'end' is positive, the ring
 792          * buffer has not fully filled, so we must adjust 'end' to 0.
 793          *
 794          * However, since both 'head' and 'end' is unsigned, we can't
 795          * simply compare 'end' against 0. Here we compare '-head' and
 796          * the size of the ring buffer, where -head is the number of bytes
 797          * kernel write to the ring buffer.
 798          */
 799         if (-head < (u64)(md->mask + 1))
 800                 end = 0;
 801         else
 802                 end = head + md->mask + 1;
 803
 804         return perf_mmap__read(md, false, start, end, &md->prev);
 805 }
 806
 807 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 808 {
 809         struct perf_mmap *md = &evlist->mmap[idx];
 810         u64 head;
 811
 812         if (!atomic_read(&md->refcnt))
 813                 return;
 814
 815         head = perf_mmap__read_head(md);
 816         md->prev = head;
 817 }
 818
 819 static bool perf_mmap__empty(struct perf_mmap *md)
 820 {
 821         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 822 }
 823
 824 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
 825 {
 826         atomic_inc(&evlist->mmap[idx].refcnt);
 827 }
 828
 829 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
 830 {
 831         BUG_ON(atomic_read(&evlist->mmap[idx].refcnt) == 0);
 832
 833         if (atomic_dec_and_test(&evlist->mmap[idx].refcnt))
 834                 __perf_evlist__munmap(evlist, idx);
 835 }
 836
 837 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 838 {
 839         struct perf_mmap *md = &evlist->mmap[idx];
 840
 841         if (!evlist->overwrite) {
 842                 u64 old = md->prev;
 843
 844                 perf_mmap__write_tail(md, old);
 845         }
 846
 847         if (atomic_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 848                 perf_evlist__mmap_put(evlist, idx);
 849 }
 850
 851 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 852                                struct auxtrace_mmap_params *mp __maybe_unused,
 853                                void *userpg __maybe_unused,
 854                                int fd __maybe_unused)
 855 {
 856         return 0;
 857 }
 858
 859 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 860 {
 861 }
 862
 863 void __weak auxtrace_mmap_params__init(
 864                         struct auxtrace_mmap_params *mp __maybe_unused,
 865                         off_t auxtrace_offset __maybe_unused,
 866                         unsigned int auxtrace_pages __maybe_unused,
 867                         bool auxtrace_overwrite __maybe_unused)
 868 {
 869 }
 870
 871 void __weak auxtrace_mmap_params__set_idx(
 872                         struct auxtrace_mmap_params *mp __maybe_unused,
 873                         struct perf_evlist *evlist __maybe_unused,
 874                         int idx __maybe_unused,
 875                         bool per_cpu __maybe_unused)
 876 {
 877 }
 878
 879 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 880 {
 881         if (evlist->mmap[idx].base != NULL) {
 882                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 883                 evlist->mmap[idx].base = NULL;
 884                 atomic_set(&evlist->mmap[idx].refcnt, 0);
 885         }
 886         auxtrace_mmap__munmap(&evlist->mmap[idx].auxtrace_mmap);
 887 }
 888
 889 void perf_evlist__munmap(struct perf_evlist *evlist)
 890 {
 891         int i;
 892
 893         if (evlist->mmap == NULL)
 894                 return;
 895
 896         for (i = 0; i < evlist->nr_mmaps; i++)
 897                 __perf_evlist__munmap(evlist, i);
 898
 899         zfree(&evlist->mmap);
 900 }
 901
 902 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 903 {
 904         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 905         if (cpu_map__empty(evlist->cpus))
 906                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 907         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 908         return evlist->mmap != NULL ? 0 : -ENOMEM;
 909 }
 910
 911 struct mmap_params {
 912         int prot;
 913         int mask;
 914         struct auxtrace_mmap_params auxtrace_mp;
 915 };
 916
 917 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 918                                struct mmap_params *mp, int fd)
 919 {
 920         /*
 921          * The last one will be done at perf_evlist__mmap_consume(), so that we
 922          * make sure we don't prevent tools from consuming every last event in
 923          * the ring buffer.
 924          *
 925          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 926          * anymore, but the last events for it are still in the ring buffer,
 927          * waiting to be consumed.
 928          *
 929          * Tools can chose to ignore this at their own discretion, but the
 930          * evlist layer can't just drop it when filtering events in
 931          * perf_evlist__filter_pollfd().
 932          */
 933         atomic_set(&evlist->mmap[idx].refcnt, 2);
 934         evlist->mmap[idx].prev = 0;
 935         evlist->mmap[idx].mask = mp->mask;
 936         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 937                                       MAP_SHARED, fd, 0);
 938         if (evlist->mmap[idx].base == MAP_FAILED) {
 939                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 940                           errno);
 941                 evlist->mmap[idx].base = NULL;
 942                 return -1;
 943         }
 944
 945         if (auxtrace_mmap__mmap(&evlist->mmap[idx].auxtrace_mmap,
 946                                 &mp->auxtrace_mp, evlist->mmap[idx].base, fd))
 947                 return -1;
 948
 949         return 0;
 950 }
 951
 952 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 953                                        struct mmap_params *mp, int cpu,
 954                                        int thread, int *output)
 955 {
 956         struct perf_evsel *evsel;
 957
 958         evlist__for_each(evlist, evsel) {
 959                 int fd;
 960
 961                 if (evsel->system_wide && thread)
 962                         continue;
 963
 964                 fd = FD(evsel, cpu, thread);
 965
 966                 if (*output == -1) {
 967                         *output = fd;
 968                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 969                                 return -1;
 970                 } else {
 971                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 972                                 return -1;
 973
 974                         perf_evlist__mmap_get(evlist, idx);
 975                 }
 976
 977                 /*
 978                  * The system_wide flag causes a selected event to be opened
 979                  * always without a pid.  Consequently it will never get a
 980                  * POLLHUP, but it is used for tracking in combination with
 981                  * other events, so it should not need to be polled anyway.
 982                  * Therefore don't add it for polling.
 983                  */
 984                 if (!evsel->system_wide &&
 985                     __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
 986                         perf_evlist__mmap_put(evlist, idx);
 987                         return -1;
 988                 }
 989
 990                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 991                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 992                                                    fd) < 0)
 993                                 return -1;
 994                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 995                                                  thread);
 996                 }
 997         }
 998
 999         return 0;
1000 }
1001
1002 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1003                                      struct mmap_params *mp)
1004 {
1005         int cpu, thread;
1006         int nr_cpus = cpu_map__nr(evlist->cpus);
1007         int nr_threads = thread_map__nr(evlist->threads);
1008
1009         pr_debug2("perf event ring buffer mmapped per cpu\n");
1010         for (cpu = 0; cpu < nr_cpus; cpu++) {
1011                 int output = -1;
1012
1013                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1014                                               true);
1015
1016                 for (thread = 0; thread < nr_threads; thread++) {
1017                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1018                                                         thread, &output))
1019                                 goto out_unmap;
1020                 }
1021         }
1022
1023         return 0;
1024
1025 out_unmap:
1026         for (cpu = 0; cpu < nr_cpus; cpu++)
1027                 __perf_evlist__munmap(evlist, cpu);
1028         return -1;
1029 }
1030
1031 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1032                                         struct mmap_params *mp)
1033 {
1034         int thread;
1035         int nr_threads = thread_map__nr(evlist->threads);
1036
1037         pr_debug2("perf event ring buffer mmapped per thread\n");
1038         for (thread = 0; thread < nr_threads; thread++) {
1039                 int output = -1;
1040
1041                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1042                                               false);
1043
1044                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1045                                                 &output))
1046                         goto out_unmap;
1047         }
1048
1049         return 0;
1050
1051 out_unmap:
1052         for (thread = 0; thread < nr_threads; thread++)
1053                 __perf_evlist__munmap(evlist, thread);
1054         return -1;
1055 }
1056
1057 unsigned long perf_event_mlock_kb_in_pages(void)
1058 {
1059         unsigned long pages;
1060         int max;
1061
1062         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1063                 /*
1064                  * Pick a once upon a time good value, i.e. things look
1065                  * strange since we can't read a sysctl value, but lets not
1066                  * die yet...
1067                  */
1068                 max = 512;
1069         } else {
1070                 max -= (page_size / 1024);
1071         }
1072
1073         pages = (max * 1024) / page_size;
1074         if (!is_power_of_2(pages))
1075                 pages = rounddown_pow_of_two(pages);
1076
1077         return pages;
1078 }
1079
1080 static size_t perf_evlist__mmap_size(unsigned long pages)
1081 {
1082         if (pages == UINT_MAX)
1083                 pages = perf_event_mlock_kb_in_pages();
1084         else if (!is_power_of_2(pages))
1085                 return 0;
1086
1087         return (pages + 1) * page_size;
1088 }
1089
1090 static long parse_pages_arg(const char *str, unsigned long min,
1091                             unsigned long max)
1092 {
1093         unsigned long pages, val;
1094         static struct parse_tag tags[] = {
1095                 { .tag  = 'B', .mult = 1       },
1096                 { .tag  = 'K', .mult = 1 << 10 },
1097                 { .tag  = 'M', .mult = 1 << 20 },
1098                 { .tag  = 'G', .mult = 1 << 30 },
1099                 { .tag  = 0 },
1100         };
1101
1102         if (str == NULL)
1103                 return -EINVAL;
1104
1105         val = parse_tag_value(str, tags);
1106         if (val != (unsigned long) -1) {
1107                 /* we got file size value */
1108                 pages = PERF_ALIGN(val, page_size) / page_size;
1109         } else {
1110                 /* we got pages count value */
1111                 char *eptr;
1112                 pages = strtoul(str, &eptr, 10);
1113                 if (*eptr != '\0')
1114                         return -EINVAL;
1115         }
1116
1117         if (pages == 0 && min == 0) {
1118                 /* leave number of pages at 0 */
1119         } else if (!is_power_of_2(pages)) {
1120                 /* round pages up to next power of 2 */
1121                 pages = roundup_pow_of_two(pages);
1122                 if (!pages)
1123                         return -EINVAL;
1124                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
1125                         pages * page_size, pages);
1126         }
1127
1128         if (pages > max)
1129                 return -EINVAL;
1130
1131         return pages;
1132 }
1133
1134 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1135 {
1136         unsigned long max = UINT_MAX;
1137         long pages;
1138
1139         if (max > SIZE_MAX / page_size)
1140                 max = SIZE_MAX / page_size;
1141
1142         pages = parse_pages_arg(str, 1, max);
1143         if (pages < 0) {
1144                 pr_err("Invalid argument for --mmap_pages/-m\n");
1145                 return -1;
1146         }
1147
1148         *mmap_pages = pages;
1149         return 0;
1150 }
1151
1152 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1153                                   int unset __maybe_unused)
1154 {
1155         return __perf_evlist__parse_mmap_pages(opt->value, str);
1156 }
1157
1158 /**
1159  * perf_evlist__mmap_ex - Create mmaps to receive events.
1160  * @evlist: list of events
1161  * @pages: map length in pages
1162  * @overwrite: overwrite older events?
1163  * @auxtrace_pages - auxtrace map length in pages
1164  * @auxtrace_overwrite - overwrite older auxtrace data?
1165  *
1166  * If @overwrite is %false the user needs to signal event consumption using
1167  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1168  * automatically.
1169  *
1170  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1171  * consumption using auxtrace_mmap__write_tail().
1172  *
1173  * Return: %0 on success, negative error code otherwise.
1174  */
1175 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1176                          bool overwrite, unsigned int auxtrace_pages,
1177                          bool auxtrace_overwrite)
1178 {
1179         struct perf_evsel *evsel;
1180         const struct cpu_map *cpus = evlist->cpus;
1181         const struct thread_map *threads = evlist->threads;
1182         struct mmap_params mp = {
1183                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1184         };
1185
1186         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1187                 return -ENOMEM;
1188
1189         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1190                 return -ENOMEM;
1191
1192         evlist->overwrite = overwrite;
1193         evlist->mmap_len = perf_evlist__mmap_size(pages);
1194         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1195         mp.mask = evlist->mmap_len - page_size - 1;
1196
1197         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1198                                    auxtrace_pages, auxtrace_overwrite);
1199
1200         evlist__for_each(evlist, evsel) {
1201                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1202                     evsel->sample_id == NULL &&
1203                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1204                         return -ENOMEM;
1205         }
1206
1207         if (cpu_map__empty(cpus))
1208                 return perf_evlist__mmap_per_thread(evlist, &mp);
1209
1210         return perf_evlist__mmap_per_cpu(evlist, &mp);
1211 }
1212
1213 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1214                       bool overwrite)
1215 {
1216         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1217 }
1218
1219 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1220 {
1221         struct cpu_map *cpus;
1222         struct thread_map *threads;
1223
1224         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1225
1226         if (!threads)
1227                 return -1;
1228
1229         if (target__uses_dummy_map(target))
1230                 cpus = cpu_map__dummy_new();
1231         else
1232                 cpus = cpu_map__new(target->cpu_list);
1233
1234         if (!cpus)
1235                 goto out_delete_threads;
1236
1237         evlist->has_user_cpus = !!target->cpu_list;
1238
1239         perf_evlist__set_maps(evlist, cpus, threads);
1240
1241         return 0;
1242
1243 out_delete_threads:
1244         thread_map__put(threads);
1245         return -1;
1246 }
1247
1248 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1249                            struct thread_map *threads)
1250 {
1251         /*
1252          * Allow for the possibility that one or another of the maps isn't being
1253          * changed i.e. don't put it.  Note we are assuming the maps that are
1254          * being applied are brand new and evlist is taking ownership of the
1255          * original reference count of 1.  If that is not the case it is up to
1256          * the caller to increase the reference count.
1257          */
1258         if (cpus != evlist->cpus) {
1259                 cpu_map__put(evlist->cpus);
1260                 evlist->cpus = cpu_map__get(cpus);
1261         }
1262
1263         if (threads != evlist->threads) {
1264                 thread_map__put(evlist->threads);
1265                 evlist->threads = thread_map__get(threads);
1266         }
1267
1268         perf_evlist__propagate_maps(evlist);
1269 }
1270
1271 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1272                                    enum perf_event_sample_format bit)
1273 {
1274         struct perf_evsel *evsel;
1275
1276         evlist__for_each(evlist, evsel)
1277                 __perf_evsel__set_sample_bit(evsel, bit);
1278 }
1279
1280 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1281                                      enum perf_event_sample_format bit)
1282 {
1283         struct perf_evsel *evsel;
1284
1285         evlist__for_each(evlist, evsel)
1286                 __perf_evsel__reset_sample_bit(evsel, bit);
1287 }
1288
1289 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1290 {
1291         struct perf_evsel *evsel;
1292         int err = 0;
1293         const int ncpus = cpu_map__nr(evlist->cpus),
1294                   nthreads = thread_map__nr(evlist->threads);
1295
1296         evlist__for_each(evlist, evsel) {
1297                 if (evsel->filter == NULL)
1298                         continue;
1299
1300                 /*
1301                  * filters only work for tracepoint event, which doesn't have cpu limit.
1302                  * So evlist and evsel should always be same.
1303                  */
1304                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1305                 if (err) {
1306                         *err_evsel = evsel;
1307                         break;
1308                 }
1309         }
1310
1311         return err;
1312 }
1313
1314 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1315 {
1316         struct perf_evsel *evsel;
1317         int err = 0;
1318
1319         evlist__for_each(evlist, evsel) {
1320                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1321                         continue;
1322
1323                 err = perf_evsel__set_filter(evsel, filter);
1324                 if (err)
1325                         break;
1326         }
1327
1328         return err;
1329 }
1330
1331 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1332 {
1333         char *filter;
1334         int ret = -1;
1335         size_t i;
1336
1337         for (i = 0; i < npids; ++i) {
1338                 if (i == 0) {
1339                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1340                                 return -1;
1341                 } else {
1342                         char *tmp;
1343
1344                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1345                                 goto out_free;
1346
1347                         free(filter);
1348                         filter = tmp;
1349                 }
1350         }
1351
1352         ret = perf_evlist__set_filter(evlist, filter);
1353 out_free:
1354         free(filter);
1355         return ret;
1356 }
1357
1358 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1359 {
1360         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1361 }
1362
1363 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1364 {
1365         struct perf_evsel *pos;
1366
1367         if (evlist->nr_entries == 1)
1368                 return true;
1369
1370         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1371                 return false;
1372
1373         evlist__for_each(evlist, pos) {
1374                 if (pos->id_pos != evlist->id_pos ||
1375                     pos->is_pos != evlist->is_pos)
1376                         return false;
1377         }
1378
1379         return true;
1380 }
1381
1382 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1383 {
1384         struct perf_evsel *evsel;
1385
1386         if (evlist->combined_sample_type)
1387                 return evlist->combined_sample_type;
1388
1389         evlist__for_each(evlist, evsel)
1390                 evlist->combined_sample_type |= evsel->attr.sample_type;
1391
1392         return evlist->combined_sample_type;
1393 }
1394
1395 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1396 {
1397         evlist->combined_sample_type = 0;
1398         return __perf_evlist__combined_sample_type(evlist);
1399 }
1400
1401 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1402 {
1403         struct perf_evsel *evsel;
1404         u64 branch_type = 0;
1405
1406         evlist__for_each(evlist, evsel)
1407                 branch_type |= evsel->attr.branch_sample_type;
1408         return branch_type;
1409 }
1410
1411 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1412 {
1413         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1414         u64 read_format = first->attr.read_format;
1415         u64 sample_type = first->attr.sample_type;
1416
1417         evlist__for_each(evlist, pos) {
1418                 if (read_format != pos->attr.read_format)
1419                         return false;
1420         }
1421
1422         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1423         if ((sample_type & PERF_SAMPLE_READ) &&
1424             !(read_format & PERF_FORMAT_ID)) {
1425                 return false;
1426         }
1427
1428         return true;
1429 }
1430
1431 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1432 {
1433         struct perf_evsel *first = perf_evlist__first(evlist);
1434         return first->attr.read_format;
1435 }
1436
1437 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1438 {
1439         struct perf_evsel *first = perf_evlist__first(evlist);
1440         struct perf_sample *data;
1441         u64 sample_type;
1442         u16 size = 0;
1443
1444         if (!first->attr.sample_id_all)
1445                 goto out;
1446
1447         sample_type = first->attr.sample_type;
1448
1449         if (sample_type & PERF_SAMPLE_TID)
1450                 size += sizeof(data->tid) * 2;
1451
1452        if (sample_type & PERF_SAMPLE_TIME)
1453                 size += sizeof(data->time);
1454
1455         if (sample_type & PERF_SAMPLE_ID)
1456                 size += sizeof(data->id);
1457
1458         if (sample_type & PERF_SAMPLE_STREAM_ID)
1459                 size += sizeof(data->stream_id);
1460
1461         if (sample_type & PERF_SAMPLE_CPU)
1462                 size += sizeof(data->cpu) * 2;
1463
1464         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1465                 size += sizeof(data->id);
1466 out:
1467         return size;
1468 }
1469
1470 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1471 {
1472         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1473
1474         evlist__for_each_continue(evlist, pos) {
1475                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1476                         return false;
1477         }
1478
1479         return true;
1480 }
1481
1482 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1483 {
1484         struct perf_evsel *first = perf_evlist__first(evlist);
1485         return first->attr.sample_id_all;
1486 }
1487
1488 void perf_evlist__set_selected(struct perf_evlist *evlist,
1489                                struct perf_evsel *evsel)
1490 {
1491         evlist->selected = evsel;
1492 }
1493
1494 void perf_evlist__close(struct perf_evlist *evlist)
1495 {
1496         struct perf_evsel *evsel;
1497         int ncpus = cpu_map__nr(evlist->cpus);
1498         int nthreads = thread_map__nr(evlist->threads);
1499         int n;
1500
1501         evlist__for_each_reverse(evlist, evsel) {
1502                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1503                 perf_evsel__close(evsel, n, nthreads);
1504         }
1505 }
1506
1507 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1508 {
1509         struct cpu_map    *cpus;
1510         struct thread_map *threads;
1511         int err = -ENOMEM;
1512
1513         /*
1514          * Try reading /sys/devices/system/cpu/online to get
1515          * an all cpus map.
1516          *
1517          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1518          * code needs an overhaul to properly forward the
1519          * error, and we may not want to do that fallback to a
1520          * default cpu identity map :-\
1521          */
1522         cpus = cpu_map__new(NULL);
1523         if (!cpus)
1524                 goto out;
1525
1526         threads = thread_map__new_dummy();
1527         if (!threads)
1528                 goto out_put;
1529
1530         perf_evlist__set_maps(evlist, cpus, threads);
1531 out:
1532         return err;
1533 out_put:
1534         cpu_map__put(cpus);
1535         goto out;
1536 }
1537
1538 int perf_evlist__open(struct perf_evlist *evlist)
1539 {
1540         struct perf_evsel *evsel;
1541         int err;
1542
1543         /*
1544          * Default: one fd per CPU, all threads, aka systemwide
1545          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1546          */
1547         if (evlist->threads == NULL && evlist->cpus == NULL) {
1548                 err = perf_evlist__create_syswide_maps(evlist);
1549                 if (err < 0)
1550                         goto out_err;
1551         }
1552
1553         perf_evlist__update_id_pos(evlist);
1554
1555         evlist__for_each(evlist, evsel) {
1556                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1557                 if (err < 0)
1558                         goto out_err;
1559         }
1560
1561         return 0;
1562 out_err:
1563         perf_evlist__close(evlist);
1564         errno = -err;
1565         return err;
1566 }
1567
1568 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1569                                   const char *argv[], bool pipe_output,
1570                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1571 {
1572         int child_ready_pipe[2], go_pipe[2];
1573         char bf;
1574
1575         if (pipe(child_ready_pipe) < 0) {
1576                 perror("failed to create 'ready' pipe");
1577                 return -1;
1578         }
1579
1580         if (pipe(go_pipe) < 0) {
1581                 perror("failed to create 'go' pipe");
1582                 goto out_close_ready_pipe;
1583         }
1584
1585         evlist->workload.pid = fork();
1586         if (evlist->workload.pid < 0) {
1587                 perror("failed to fork");
1588                 goto out_close_pipes;
1589         }
1590
1591         if (!evlist->workload.pid) {
1592                 int ret;
1593
1594                 if (pipe_output)
1595                         dup2(2, 1);
1596
1597                 signal(SIGTERM, SIG_DFL);
1598
1599                 close(child_ready_pipe[0]);
1600                 close(go_pipe[1]);
1601                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1602
1603                 /*
1604                  * Tell the parent we're ready to go
1605                  */
1606                 close(child_ready_pipe[1]);
1607
1608                 /*
1609                  * Wait until the parent tells us to go.
1610                  */
1611                 ret = read(go_pipe[0], &bf, 1);
1612                 /*
1613                  * The parent will ask for the execvp() to be performed by
1614                  * writing exactly one byte, in workload.cork_fd, usually via
1615                  * perf_evlist__start_workload().
1616                  *
1617                  * For cancelling the workload without actually running it,
1618                  * the parent will just close workload.cork_fd, without writing
1619                  * anything, i.e. read will return zero and we just exit()
1620                  * here.
1621                  */
1622                 if (ret != 1) {
1623                         if (ret == -1)
1624                                 perror("unable to read pipe");
1625                         exit(ret);
1626                 }
1627
1628                 execvp(argv[0], (char **)argv);
1629
1630                 if (exec_error) {
1631                         union sigval val;
1632
1633                         val.sival_int = errno;
1634                         if (sigqueue(getppid(), SIGUSR1, val))
1635                                 perror(argv[0]);
1636                 } else
1637                         perror(argv[0]);
1638                 exit(-1);
1639         }
1640
1641         if (exec_error) {
1642                 struct sigaction act = {
1643                         .sa_flags     = SA_SIGINFO,
1644                         .sa_sigaction = exec_error,
1645                 };
1646                 sigaction(SIGUSR1, &act, NULL);
1647         }
1648
1649         if (target__none(target)) {
1650                 if (evlist->threads == NULL) {
1651                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1652                                 __func__, __LINE__);
1653                         goto out_close_pipes;
1654                 }
1655                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1656         }
1657
1658         close(child_ready_pipe[1]);
1659         close(go_pipe[0]);
1660         /*
1661          * wait for child to settle
1662          */
1663         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1664                 perror("unable to read pipe");
1665                 goto out_close_pipes;
1666         }
1667
1668         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1669         evlist->workload.cork_fd = go_pipe[1];
1670         close(child_ready_pipe[0]);
1671         return 0;
1672
1673 out_close_pipes:
1674         close(go_pipe[0]);
1675         close(go_pipe[1]);
1676 out_close_ready_pipe:
1677         close(child_ready_pipe[0]);
1678         close(child_ready_pipe[1]);
1679         return -1;
1680 }
1681
1682 int perf_evlist__start_workload(struct perf_evlist *evlist)
1683 {
1684         if (evlist->workload.cork_fd > 0) {
1685                 char bf = 0;
1686                 int ret;
1687                 /*
1688                  * Remove the cork, let it rip!
1689                  */
1690                 ret = write(evlist->workload.cork_fd, &bf, 1);
1691                 if (ret < 0)
1692                         perror("enable to write to pipe");
1693
1694                 close(evlist->workload.cork_fd);
1695                 return ret;
1696         }
1697
1698         return 0;
1699 }
1700
1701 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1702                               struct perf_sample *sample)
1703 {
1704         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1705
1706         if (!evsel)
1707                 return -EFAULT;
1708         return perf_evsel__parse_sample(evsel, event, sample);
1709 }
1710
1711 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1712 {
1713         struct perf_evsel *evsel;
1714         size_t printed = 0;
1715
1716         evlist__for_each(evlist, evsel) {
1717                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1718                                    perf_evsel__name(evsel));
1719         }
1720
1721         return printed + fprintf(fp, "\n");
1722 }
1723
1724 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1725                                int err, char *buf, size_t size)
1726 {
1727         int printed, value;
1728         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1729
1730         switch (err) {
1731         case EACCES:
1732         case EPERM:
1733                 printed = scnprintf(buf, size,
1734                                     "Error:\t%s.\n"
1735                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1736
1737                 value = perf_event_paranoid();
1738
1739                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1740
1741                 if (value >= 2) {
1742                         printed += scnprintf(buf + printed, size - printed,
1743                                              "For your workloads it needs to be <= 1\nHint:\t");
1744                 }
1745                 printed += scnprintf(buf + printed, size - printed,
1746                                      "For system wide tracing it needs to be set to -1.\n");
1747
1748                 printed += scnprintf(buf + printed, size - printed,
1749                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1750                                     "Hint:\tThe current value is %d.", value);
1751                 break;
1752         case EINVAL: {
1753                 struct perf_evsel *first = perf_evlist__first(evlist);
1754                 int max_freq;
1755
1756                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1757                         goto out_default;
1758
1759                 if (first->attr.sample_freq < (u64)max_freq)
1760                         goto out_default;
1761
1762                 printed = scnprintf(buf, size,
1763                                     "Error:\t%s.\n"
1764                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1765                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1766                                     emsg, max_freq, first->attr.sample_freq);
1767                 break;
1768         }
1769         default:
1770 out_default:
1771                 scnprintf(buf, size, "%s", emsg);
1772                 break;
1773         }
1774
1775         return 0;
1776 }
1777
1778 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1779 {
1780         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1781         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1782
1783         switch (err) {
1784         case EPERM:
1785                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1786                 printed += scnprintf(buf + printed, size - printed,
1787                                      "Error:\t%s.\n"
1788                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1789                                      "Hint:\tTried using %zd kB.\n",
1790                                      emsg, pages_max_per_user, pages_attempted);
1791
1792                 if (pages_attempted >= pages_max_per_user) {
1793                         printed += scnprintf(buf + printed, size - printed,
1794                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1795                                              pages_max_per_user + pages_attempted);
1796                 }
1797
1798                 printed += scnprintf(buf + printed, size - printed,
1799                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1800                 break;
1801         default:
1802                 scnprintf(buf, size, "%s", emsg);
1803                 break;
1804         }
1805
1806         return 0;
1807 }
1808
1809 void perf_evlist__to_front(struct perf_evlist *evlist,
1810                            struct perf_evsel *move_evsel)
1811 {
1812         struct perf_evsel *evsel, *n;
1813         LIST_HEAD(move);
1814
1815         if (move_evsel == perf_evlist__first(evlist))
1816                 return;
1817
1818         evlist__for_each_safe(evlist, n, evsel) {
1819                 if (evsel->leader == move_evsel->leader)
1820                         list_move_tail(&evsel->node, &move);
1821         }
1822
1823         list_splice(&move, &evlist->entries);
1824 }
1825
1826 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1827                                      struct perf_evsel *tracking_evsel)
1828 {
1829         struct perf_evsel *evsel;
1830
1831         if (tracking_evsel->tracking)
1832                 return;
1833
1834         evlist__for_each(evlist, evsel) {
1835                 if (evsel != tracking_evsel)
1836                         evsel->tracking = false;
1837         }
1838
1839         tracking_evsel->tracking = true;
1840 }
1841
1842 struct perf_evsel *
1843 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1844                                const char *str)
1845 {
1846         struct perf_evsel *evsel;
1847
1848         evlist__for_each(evlist, evsel) {
1849                 if (!evsel->name)
1850                         continue;
1851                 if (strcmp(str, evsel->name) == 0)
1852                         return evsel;
1853         }
1854
1855         return NULL;
1856 }