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 "asm/bug.h"
  21 #include <unistd.h>
  22
  23 #include "parse-events.h"
  24 #include <subcmd/parse-options.h>
  25
  26 #include <sys/mman.h>
  27
  28 #include <linux/bitops.h>
  29 #include <linux/hash.h>
  30 #include <linux/log2.h>
  31 #include <linux/err.h>
  32
  33 static void perf_mmap__munmap(struct perf_mmap *map);
  34 static void perf_mmap__put(struct perf_mmap *map);
  35
  36 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  37 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  38
  39 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  40                        struct thread_map *threads)
  41 {
  42         int i;
  43
  44         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  45                 INIT_HLIST_HEAD(&evlist->heads[i]);
  46         INIT_LIST_HEAD(&evlist->entries);
  47         perf_evlist__set_maps(evlist, cpus, threads);
  48         fdarray__init(&evlist->pollfd, 64);
  49         evlist->workload.pid = -1;
  50         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  51 }
  52
  53 struct perf_evlist *perf_evlist__new(void)
  54 {
  55         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  56
  57         if (evlist != NULL)
  58                 perf_evlist__init(evlist, NULL, NULL);
  59
  60         return evlist;
  61 }
  62
  63 struct perf_evlist *perf_evlist__new_default(void)
  64 {
  65         struct perf_evlist *evlist = perf_evlist__new();
  66
  67         if (evlist && perf_evlist__add_default(evlist)) {
  68                 perf_evlist__delete(evlist);
  69                 evlist = NULL;
  70         }
  71
  72         return evlist;
  73 }
  74
  75 struct perf_evlist *perf_evlist__new_dummy(void)
  76 {
  77         struct perf_evlist *evlist = perf_evlist__new();
  78
  79         if (evlist && perf_evlist__add_dummy(evlist)) {
  80                 perf_evlist__delete(evlist);
  81                 evlist = NULL;
  82         }
  83
  84         return evlist;
  85 }
  86
  87 /**
  88  * perf_evlist__set_id_pos - set the positions of event ids.
  89  * @evlist: selected event list
  90  *
  91  * Events with compatible sample types all have the same id_pos
  92  * and is_pos.  For convenience, put a copy on evlist.
  93  */
  94 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  95 {
  96         struct perf_evsel *first = perf_evlist__first(evlist);
  97
  98         evlist->id_pos = first->id_pos;
  99         evlist->is_pos = first->is_pos;
 100 }
 101
 102 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 103 {
 104         struct perf_evsel *evsel;
 105
 106         evlist__for_each_entry(evlist, evsel)
 107                 perf_evsel__calc_id_pos(evsel);
 108
 109         perf_evlist__set_id_pos(evlist);
 110 }
 111
 112 static void perf_evlist__purge(struct perf_evlist *evlist)
 113 {
 114         struct perf_evsel *pos, *n;
 115
 116         evlist__for_each_entry_safe(evlist, n, pos) {
 117                 list_del_init(&pos->node);
 118                 pos->evlist = NULL;
 119                 perf_evsel__delete(pos);
 120         }
 121
 122         evlist->nr_entries = 0;
 123 }
 124
 125 void perf_evlist__exit(struct perf_evlist *evlist)
 126 {
 127         zfree(&evlist->mmap);
 128         zfree(&evlist->backward_mmap);
 129         fdarray__exit(&evlist->pollfd);
 130 }
 131
 132 void perf_evlist__delete(struct perf_evlist *evlist)
 133 {
 134         if (evlist == NULL)
 135                 return;
 136
 137         perf_evlist__munmap(evlist);
 138         perf_evlist__close(evlist);
 139         cpu_map__put(evlist->cpus);
 140         thread_map__put(evlist->threads);
 141         evlist->cpus = NULL;
 142         evlist->threads = NULL;
 143         perf_evlist__purge(evlist);
 144         perf_evlist__exit(evlist);
 145         free(evlist);
 146 }
 147
 148 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 149                                           struct perf_evsel *evsel)
 150 {
 151         /*
 152          * We already have cpus for evsel (via PMU sysfs) so
 153          * keep it, if there's no target cpu list defined.
 154          */
 155         if (!evsel->own_cpus || evlist->has_user_cpus) {
 156                 cpu_map__put(evsel->cpus);
 157                 evsel->cpus = cpu_map__get(evlist->cpus);
 158         } else if (evsel->cpus != evsel->own_cpus) {
 159                 cpu_map__put(evsel->cpus);
 160                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 161         }
 162
 163         thread_map__put(evsel->threads);
 164         evsel->threads = thread_map__get(evlist->threads);
 165 }
 166
 167 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 168 {
 169         struct perf_evsel *evsel;
 170
 171         evlist__for_each_entry(evlist, evsel)
 172                 __perf_evlist__propagate_maps(evlist, evsel);
 173 }
 174
 175 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 176 {
 177         entry->evlist = evlist;
 178         list_add_tail(&entry->node, &evlist->entries);
 179         entry->idx = evlist->nr_entries;
 180         entry->tracking = !entry->idx;
 181
 182         if (!evlist->nr_entries++)
 183                 perf_evlist__set_id_pos(evlist);
 184
 185         __perf_evlist__propagate_maps(evlist, entry);
 186 }
 187
 188 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 189 {
 190         evsel->evlist = NULL;
 191         list_del_init(&evsel->node);
 192         evlist->nr_entries -= 1;
 193 }
 194
 195 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 196                                    struct list_head *list)
 197 {
 198         struct perf_evsel *evsel, *temp;
 199
 200         __evlist__for_each_entry_safe(list, temp, evsel) {
 201                 list_del_init(&evsel->node);
 202                 perf_evlist__add(evlist, evsel);
 203         }
 204 }
 205
 206 void __perf_evlist__set_leader(struct list_head *list)
 207 {
 208         struct perf_evsel *evsel, *leader;
 209
 210         leader = list_entry(list->next, struct perf_evsel, node);
 211         evsel = list_entry(list->prev, struct perf_evsel, node);
 212
 213         leader->nr_members = evsel->idx - leader->idx + 1;
 214
 215         __evlist__for_each_entry(list, evsel) {
 216                 evsel->leader = leader;
 217         }
 218 }
 219
 220 void perf_evlist__set_leader(struct perf_evlist *evlist)
 221 {
 222         if (evlist->nr_entries) {
 223                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 224                 __perf_evlist__set_leader(&evlist->entries);
 225         }
 226 }
 227
 228 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 229 {
 230         attr->precise_ip = 3;
 231
 232         while (attr->precise_ip != 0) {
 233                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 234                 if (fd != -1) {
 235                         close(fd);
 236                         break;
 237                 }
 238                 --attr->precise_ip;
 239         }
 240 }
 241
 242 int perf_evlist__add_default(struct perf_evlist *evlist)
 243 {
 244         struct perf_evsel *evsel = perf_evsel__new_cycles();
 245
 246         if (evsel == NULL)
 247                 return -ENOMEM;
 248
 249         perf_evlist__add(evlist, evsel);
 250         return 0;
 251 }
 252
 253 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 254 {
 255         struct perf_event_attr attr = {
 256                 .type   = PERF_TYPE_SOFTWARE,
 257                 .config = PERF_COUNT_SW_DUMMY,
 258                 .size   = sizeof(attr), /* to capture ABI version */
 259         };
 260         struct perf_evsel *evsel = perf_evsel__new(&attr);
 261
 262         if (evsel == NULL)
 263                 return -ENOMEM;
 264
 265         perf_evlist__add(evlist, evsel);
 266         return 0;
 267 }
 268
 269 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 270                                   struct perf_event_attr *attrs, size_t nr_attrs)
 271 {
 272         struct perf_evsel *evsel, *n;
 273         LIST_HEAD(head);
 274         size_t i;
 275
 276         for (i = 0; i < nr_attrs; i++) {
 277                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 278                 if (evsel == NULL)
 279                         goto out_delete_partial_list;
 280                 list_add_tail(&evsel->node, &head);
 281         }
 282
 283         perf_evlist__splice_list_tail(evlist, &head);
 284
 285         return 0;
 286
 287 out_delete_partial_list:
 288         __evlist__for_each_entry_safe(&head, n, evsel)
 289                 perf_evsel__delete(evsel);
 290         return -1;
 291 }
 292
 293 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 294                                      struct perf_event_attr *attrs, size_t nr_attrs)
 295 {
 296         size_t i;
 297
 298         for (i = 0; i < nr_attrs; i++)
 299                 event_attr_init(attrs + i);
 300
 301         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 302 }
 303
 304 struct perf_evsel *
 305 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 306 {
 307         struct perf_evsel *evsel;
 308
 309         evlist__for_each_entry(evlist, evsel) {
 310                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 311                     (int)evsel->attr.config == id)
 312                         return evsel;
 313         }
 314
 315         return NULL;
 316 }
 317
 318 struct perf_evsel *
 319 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 320                                      const char *name)
 321 {
 322         struct perf_evsel *evsel;
 323
 324         evlist__for_each_entry(evlist, evsel) {
 325                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 326                     (strcmp(evsel->name, name) == 0))
 327                         return evsel;
 328         }
 329
 330         return NULL;
 331 }
 332
 333 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 334                            const char *sys, const char *name, void *handler)
 335 {
 336         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 337
 338         if (IS_ERR(evsel))
 339                 return -1;
 340
 341         evsel->handler = handler;
 342         perf_evlist__add(evlist, evsel);
 343         return 0;
 344 }
 345
 346 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 347                                    struct perf_evsel *evsel)
 348 {
 349         if (evsel->system_wide)
 350                 return 1;
 351         else
 352                 return thread_map__nr(evlist->threads);
 353 }
 354
 355 void perf_evlist__disable(struct perf_evlist *evlist)
 356 {
 357         struct perf_evsel *pos;
 358
 359         evlist__for_each_entry(evlist, pos) {
 360                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 361                         continue;
 362                 perf_evsel__disable(pos);
 363         }
 364
 365         evlist->enabled = false;
 366 }
 367
 368 void perf_evlist__enable(struct perf_evlist *evlist)
 369 {
 370         struct perf_evsel *pos;
 371
 372         evlist__for_each_entry(evlist, pos) {
 373                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 374                         continue;
 375                 perf_evsel__enable(pos);
 376         }
 377
 378         evlist->enabled = true;
 379 }
 380
 381 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 382 {
 383         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 384 }
 385
 386 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 387                                          struct perf_evsel *evsel, int cpu)
 388 {
 389         int thread;
 390         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 391
 392         if (!evsel->fd)
 393                 return -EINVAL;
 394
 395         for (thread = 0; thread < nr_threads; thread++) {
 396                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 397                 if (err)
 398                         return err;
 399         }
 400         return 0;
 401 }
 402
 403 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 404                                             struct perf_evsel *evsel,
 405                                             int thread)
 406 {
 407         int cpu;
 408         int nr_cpus = cpu_map__nr(evlist->cpus);
 409
 410         if (!evsel->fd)
 411                 return -EINVAL;
 412
 413         for (cpu = 0; cpu < nr_cpus; cpu++) {
 414                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 415                 if (err)
 416                         return err;
 417         }
 418         return 0;
 419 }
 420
 421 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 422                                   struct perf_evsel *evsel, int idx)
 423 {
 424         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 425
 426         if (per_cpu_mmaps)
 427                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 428         else
 429                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 430 }
 431
 432 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 433 {
 434         int nr_cpus = cpu_map__nr(evlist->cpus);
 435         int nr_threads = thread_map__nr(evlist->threads);
 436         int nfds = 0;
 437         struct perf_evsel *evsel;
 438
 439         evlist__for_each_entry(evlist, evsel) {
 440                 if (evsel->system_wide)
 441                         nfds += nr_cpus;
 442                 else
 443                         nfds += nr_cpus * nr_threads;
 444         }
 445
 446         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 447             fdarray__grow(&evlist->pollfd, nfds) < 0)
 448                 return -ENOMEM;
 449
 450         return 0;
 451 }
 452
 453 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 454                                      struct perf_mmap *map, short revent)
 455 {
 456         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 457         /*
 458          * Save the idx so that when we filter out fds POLLHUP'ed we can
 459          * close the associated evlist->mmap[] entry.
 460          */
 461         if (pos >= 0) {
 462                 evlist->pollfd.priv[pos].ptr = map;
 463
 464                 fcntl(fd, F_SETFL, O_NONBLOCK);
 465         }
 466
 467         return pos;
 468 }
 469
 470 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 471 {
 472         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 473 }
 474
 475 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 476                                          void *arg __maybe_unused)
 477 {
 478         struct perf_mmap *map = fda->priv[fd].ptr;
 479
 480         if (map)
 481                 perf_mmap__put(map);
 482 }
 483
 484 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 485 {
 486         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 487                                perf_evlist__munmap_filtered, NULL);
 488 }
 489
 490 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 491 {
 492         return fdarray__poll(&evlist->pollfd, timeout);
 493 }
 494
 495 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 496                                  struct perf_evsel *evsel,
 497                                  int cpu, int thread, u64 id)
 498 {
 499         int hash;
 500         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 501
 502         sid->id = id;
 503         sid->evsel = evsel;
 504         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 505         hlist_add_head(&sid->node, &evlist->heads[hash]);
 506 }
 507
 508 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 509                          int cpu, int thread, u64 id)
 510 {
 511         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 512         evsel->id[evsel->ids++] = id;
 513 }
 514
 515 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 516                            struct perf_evsel *evsel,
 517                            int cpu, int thread, int fd)
 518 {
 519         u64 read_data[4] = { 0, };
 520         int id_idx = 1; /* The first entry is the counter value */
 521         u64 id;
 522         int ret;
 523
 524         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 525         if (!ret)
 526                 goto add;
 527
 528         if (errno != ENOTTY)
 529                 return -1;
 530
 531         /* Legacy way to get event id.. All hail to old kernels! */
 532
 533         /*
 534          * This way does not work with group format read, so bail
 535          * out in that case.
 536          */
 537         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 538                 return -1;
 539
 540         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 541             read(fd, &read_data, sizeof(read_data)) == -1)
 542                 return -1;
 543
 544         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 545                 ++id_idx;
 546         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 547                 ++id_idx;
 548
 549         id = read_data[id_idx];
 550
 551  add:
 552         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 553         return 0;
 554 }
 555
 556 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 557                                      struct perf_evsel *evsel, int idx, int cpu,
 558                                      int thread)
 559 {
 560         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 561         sid->idx = idx;
 562         if (evlist->cpus && cpu >= 0)
 563                 sid->cpu = evlist->cpus->map[cpu];
 564         else
 565                 sid->cpu = -1;
 566         if (!evsel->system_wide && evlist->threads && thread >= 0)
 567                 sid->tid = thread_map__pid(evlist->threads, thread);
 568         else
 569                 sid->tid = -1;
 570 }
 571
 572 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 573 {
 574         struct hlist_head *head;
 575         struct perf_sample_id *sid;
 576         int hash;
 577
 578         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 579         head = &evlist->heads[hash];
 580
 581         hlist_for_each_entry(sid, head, node)
 582                 if (sid->id == id)
 583                         return sid;
 584
 585         return NULL;
 586 }
 587
 588 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 589 {
 590         struct perf_sample_id *sid;
 591
 592         if (evlist->nr_entries == 1 || !id)
 593                 return perf_evlist__first(evlist);
 594
 595         sid = perf_evlist__id2sid(evlist, id);
 596         if (sid)
 597                 return sid->evsel;
 598
 599         if (!perf_evlist__sample_id_all(evlist))
 600                 return perf_evlist__first(evlist);
 601
 602         return NULL;
 603 }
 604
 605 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 606                                                 u64 id)
 607 {
 608         struct perf_sample_id *sid;
 609
 610         if (!id)
 611                 return NULL;
 612
 613         sid = perf_evlist__id2sid(evlist, id);
 614         if (sid)
 615                 return sid->evsel;
 616
 617         return NULL;
 618 }
 619
 620 static int perf_evlist__event2id(struct perf_evlist *evlist,
 621                                  union perf_event *event, u64 *id)
 622 {
 623         const u64 *array = event->sample.array;
 624         ssize_t n;
 625
 626         n = (event->header.size - sizeof(event->header)) >> 3;
 627
 628         if (event->header.type == PERF_RECORD_SAMPLE) {
 629                 if (evlist->id_pos >= n)
 630                         return -1;
 631                 *id = array[evlist->id_pos];
 632         } else {
 633                 if (evlist->is_pos > n)
 634                         return -1;
 635                 n -= evlist->is_pos;
 636                 *id = array[n];
 637         }
 638         return 0;
 639 }
 640
 641 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 642                                             union perf_event *event)
 643 {
 644         struct perf_evsel *first = perf_evlist__first(evlist);
 645         struct hlist_head *head;
 646         struct perf_sample_id *sid;
 647         int hash;
 648         u64 id;
 649
 650         if (evlist->nr_entries == 1)
 651                 return first;
 652
 653         if (!first->attr.sample_id_all &&
 654             event->header.type != PERF_RECORD_SAMPLE)
 655                 return first;
 656
 657         if (perf_evlist__event2id(evlist, event, &id))
 658                 return NULL;
 659
 660         /* Synthesized events have an id of zero */
 661         if (!id)
 662                 return first;
 663
 664         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 665         head = &evlist->heads[hash];
 666
 667         hlist_for_each_entry(sid, head, node) {
 668                 if (sid->id == id)
 669                         return sid->evsel;
 670         }
 671         return NULL;
 672 }
 673
 674 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 675 {
 676         int i;
 677
 678         if (!evlist->backward_mmap)
 679                 return 0;
 680
 681         for (i = 0; i < evlist->nr_mmaps; i++) {
 682                 int fd = evlist->backward_mmap[i].fd;
 683                 int err;
 684
 685                 if (fd < 0)
 686                         continue;
 687                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 688                 if (err)
 689                         return err;
 690         }
 691         return 0;
 692 }
 693
 694 static int perf_evlist__pause(struct perf_evlist *evlist)
 695 {
 696         return perf_evlist__set_paused(evlist, true);
 697 }
 698
 699 static int perf_evlist__resume(struct perf_evlist *evlist)
 700 {
 701         return perf_evlist__set_paused(evlist, false);
 702 }
 703
 704 /* When check_messup is true, 'end' must points to a good entry */
 705 static union perf_event *
 706 perf_mmap__read(struct perf_mmap *md, bool check_messup, u64 start,
 707                 u64 end, u64 *prev)
 708 {
 709         unsigned char *data = md->base + page_size;
 710         union perf_event *event = NULL;
 711         int diff = end - start;
 712
 713         if (check_messup) {
 714                 /*
 715                  * If we're further behind than half the buffer, there's a chance
 716                  * the writer will bite our tail and mess up the samples under us.
 717                  *
 718                  * If we somehow ended up ahead of the 'end', we got messed up.
 719                  *
 720                  * In either case, truncate and restart at 'end'.
 721                  */
 722                 if (diff > md->mask / 2 || diff < 0) {
 723                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 724
 725                         /*
 726                          * 'end' points to a known good entry, start there.
 727                          */
 728                         start = end;
 729                         diff = 0;
 730                 }
 731         }
 732
 733         if (diff >= (int)sizeof(event->header)) {
 734                 size_t size;
 735
 736                 event = (union perf_event *)&data[start & md->mask];
 737                 size = event->header.size;
 738
 739                 if (size < sizeof(event->header) || diff < (int)size) {
 740                         event = NULL;
 741                         goto broken_event;
 742                 }
 743
 744                 /*
 745                  * Event straddles the mmap boundary -- header should always
 746                  * be inside due to u64 alignment of output.
 747                  */
 748                 if ((start & md->mask) + size != ((start + size) & md->mask)) {
 749                         unsigned int offset = start;
 750                         unsigned int len = min(sizeof(*event), size), cpy;
 751                         void *dst = md->event_copy;
 752
 753                         do {
 754                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 755                                 memcpy(dst, &data[offset & md->mask], cpy);
 756                                 offset += cpy;
 757                                 dst += cpy;
 758                                 len -= cpy;
 759                         } while (len);
 760
 761                         event = (union perf_event *) md->event_copy;
 762                 }
 763
 764                 start += size;
 765         }
 766
 767 broken_event:
 768         if (prev)
 769                 *prev = start;
 770
 771         return event;
 772 }
 773
 774 union perf_event *perf_mmap__read_forward(struct perf_mmap *md, bool check_messup)
 775 {
 776         u64 head;
 777         u64 old = md->prev;
 778
 779         /*
 780          * Check if event was unmapped due to a POLLHUP/POLLERR.
 781          */
 782         if (!refcount_read(&md->refcnt))
 783                 return NULL;
 784
 785         head = perf_mmap__read_head(md);
 786
 787         return perf_mmap__read(md, check_messup, old, head, &md->prev);
 788 }
 789
 790 union perf_event *
 791 perf_mmap__read_backward(struct perf_mmap *md)
 792 {
 793         u64 head, end;
 794         u64 start = md->prev;
 795
 796         /*
 797          * Check if event was unmapped due to a POLLHUP/POLLERR.
 798          */
 799         if (!refcount_read(&md->refcnt))
 800                 return NULL;
 801
 802         head = perf_mmap__read_head(md);
 803         if (!head)
 804                 return NULL;
 805
 806         /*
 807          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 808          * it each time when kernel writes to it, so in fact 'head' is
 809          * negative. 'end' pointer is made manually by adding the size of
 810          * the ring buffer to 'head' pointer, means the validate data can
 811          * read is the whole ring buffer. If 'end' is positive, the ring
 812          * buffer has not fully filled, so we must adjust 'end' to 0.
 813          *
 814          * However, since both 'head' and 'end' is unsigned, we can't
 815          * simply compare 'end' against 0. Here we compare '-head' and
 816          * the size of the ring buffer, where -head is the number of bytes
 817          * kernel write to the ring buffer.
 818          */
 819         if (-head < (u64)(md->mask + 1))
 820                 end = 0;
 821         else
 822                 end = head + md->mask + 1;
 823
 824         return perf_mmap__read(md, false, start, end, &md->prev);
 825 }
 826
 827 union perf_event *perf_evlist__mmap_read_forward(struct perf_evlist *evlist, int idx)
 828 {
 829         struct perf_mmap *md = &evlist->mmap[idx];
 830
 831         /*
 832          * Check messup is required for forward overwritable ring buffer:
 833          * memory pointed by md->prev can be overwritten in this case.
 834          * No need for read-write ring buffer: kernel stop outputting when
 835          * it hit md->prev (perf_mmap__consume()).
 836          */
 837         return perf_mmap__read_forward(md, evlist->overwrite);
 838 }
 839
 840 union perf_event *perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
 841 {
 842         struct perf_mmap *md = &evlist->mmap[idx];
 843
 844         /*
 845          * No need to check messup for backward ring buffer:
 846          * We can always read arbitrary long data from a backward
 847          * ring buffer unless we forget to pause it before reading.
 848          */
 849         return perf_mmap__read_backward(md);
 850 }
 851
 852 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 853 {
 854         return perf_evlist__mmap_read_forward(evlist, idx);
 855 }
 856
 857 void perf_mmap__read_catchup(struct perf_mmap *md)
 858 {
 859         u64 head;
 860
 861         if (!refcount_read(&md->refcnt))
 862                 return;
 863
 864         head = perf_mmap__read_head(md);
 865         md->prev = head;
 866 }
 867
 868 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 869 {
 870         perf_mmap__read_catchup(&evlist->mmap[idx]);
 871 }
 872
 873 static bool perf_mmap__empty(struct perf_mmap *md)
 874 {
 875         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 876 }
 877
 878 static void perf_mmap__get(struct perf_mmap *map)
 879 {
 880         refcount_inc(&map->refcnt);
 881 }
 882
 883 static void perf_mmap__put(struct perf_mmap *md)
 884 {
 885         BUG_ON(md->base && refcount_read(&md->refcnt) == 0);
 886
 887         if (refcount_dec_and_test(&md->refcnt))
 888                 perf_mmap__munmap(md);
 889 }
 890
 891 void perf_mmap__consume(struct perf_mmap *md, bool overwrite)
 892 {
 893         if (!overwrite) {
 894                 u64 old = md->prev;
 895
 896                 perf_mmap__write_tail(md, old);
 897         }
 898
 899         if (refcount_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 900                 perf_mmap__put(md);
 901 }
 902
 903 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 904 {
 905         perf_mmap__consume(&evlist->mmap[idx], evlist->overwrite);
 906 }
 907
 908 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 909                                struct auxtrace_mmap_params *mp __maybe_unused,
 910                                void *userpg __maybe_unused,
 911                                int fd __maybe_unused)
 912 {
 913         return 0;
 914 }
 915
 916 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 917 {
 918 }
 919
 920 void __weak auxtrace_mmap_params__init(
 921                         struct auxtrace_mmap_params *mp __maybe_unused,
 922                         off_t auxtrace_offset __maybe_unused,
 923                         unsigned int auxtrace_pages __maybe_unused,
 924                         bool auxtrace_overwrite __maybe_unused)
 925 {
 926 }
 927
 928 void __weak auxtrace_mmap_params__set_idx(
 929                         struct auxtrace_mmap_params *mp __maybe_unused,
 930                         struct perf_evlist *evlist __maybe_unused,
 931                         int idx __maybe_unused,
 932                         bool per_cpu __maybe_unused)
 933 {
 934 }
 935
 936 static void perf_mmap__munmap(struct perf_mmap *map)
 937 {
 938         if (map->base != NULL) {
 939                 munmap(map->base, perf_mmap__mmap_len(map));
 940                 map->base = NULL;
 941                 map->fd = -1;
 942                 refcount_set(&map->refcnt, 0);
 943         }
 944         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 945 }
 946
 947 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 948 {
 949         int i;
 950
 951         if (evlist->mmap)
 952                 for (i = 0; i < evlist->nr_mmaps; i++)
 953                         perf_mmap__munmap(&evlist->mmap[i]);
 954
 955         if (evlist->backward_mmap)
 956                 for (i = 0; i < evlist->nr_mmaps; i++)
 957                         perf_mmap__munmap(&evlist->backward_mmap[i]);
 958 }
 959
 960 void perf_evlist__munmap(struct perf_evlist *evlist)
 961 {
 962         perf_evlist__munmap_nofree(evlist);
 963         zfree(&evlist->mmap);
 964         zfree(&evlist->backward_mmap);
 965 }
 966
 967 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 968 {
 969         int i;
 970         struct perf_mmap *map;
 971
 972         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 973         if (cpu_map__empty(evlist->cpus))
 974                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 975         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 976         if (!map)
 977                 return NULL;
 978
 979         for (i = 0; i < evlist->nr_mmaps; i++) {
 980                 map[i].fd = -1;
 981                 /*
 982                  * When the perf_mmap() call is made we grab one refcount, plus
 983                  * one extra to let perf_evlist__mmap_consume() get the last
 984                  * events after all real references (perf_mmap__get()) are
 985                  * dropped.
 986                  *
 987                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 988                  * thus does perf_mmap__get() on it.
 989                  */
 990                 refcount_set(&map[i].refcnt, 0);
 991         }
 992         return map;
 993 }
 994
 995 struct mmap_params {
 996         int prot;
 997         int mask;
 998         struct auxtrace_mmap_params auxtrace_mp;
 999 };
1000
1001 static int perf_mmap__mmap(struct perf_mmap *map,
1002                            struct mmap_params *mp, int fd)
1003 {
1004         /*
1005          * The last one will be done at perf_evlist__mmap_consume(), so that we
1006          * make sure we don't prevent tools from consuming every last event in
1007          * the ring buffer.
1008          *
1009          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
1010          * anymore, but the last events for it are still in the ring buffer,
1011          * waiting to be consumed.
1012          *
1013          * Tools can chose to ignore this at their own discretion, but the
1014          * evlist layer can't just drop it when filtering events in
1015          * perf_evlist__filter_pollfd().
1016          */
1017         refcount_set(&map->refcnt, 2);
1018         map->prev = 0;
1019         map->mask = mp->mask;
1020         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
1021                          MAP_SHARED, fd, 0);
1022         if (map->base == MAP_FAILED) {
1023                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1024                           errno);
1025                 map->base = NULL;
1026                 return -1;
1027         }
1028         map->fd = fd;
1029
1030         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
1031                                 &mp->auxtrace_mp, map->base, fd))
1032                 return -1;
1033
1034         return 0;
1035 }
1036
1037 static bool
1038 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
1039                          struct perf_evsel *evsel)
1040 {
1041         if (evsel->attr.write_backward)
1042                 return false;
1043         return true;
1044 }
1045
1046 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
1047                                        struct mmap_params *mp, int cpu_idx,
1048                                        int thread, int *_output, int *_output_backward)
1049 {
1050         struct perf_evsel *evsel;
1051         int revent;
1052         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
1053
1054         evlist__for_each_entry(evlist, evsel) {
1055                 struct perf_mmap *maps = evlist->mmap;
1056                 int *output = _output;
1057                 int fd;
1058                 int cpu;
1059
1060                 if (evsel->attr.write_backward) {
1061                         output = _output_backward;
1062                         maps = evlist->backward_mmap;
1063
1064                         if (!maps) {
1065                                 maps = perf_evlist__alloc_mmap(evlist);
1066                                 if (!maps)
1067                                         return -1;
1068                                 evlist->backward_mmap = maps;
1069                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1070                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1071                         }
1072                 }
1073
1074                 if (evsel->system_wide && thread)
1075                         continue;
1076
1077                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
1078                 if (cpu == -1)
1079                         continue;
1080
1081                 fd = FD(evsel, cpu, thread);
1082
1083                 if (*output == -1) {
1084                         *output = fd;
1085
1086                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1087                                 return -1;
1088                 } else {
1089                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1090                                 return -1;
1091
1092                         perf_mmap__get(&maps[idx]);
1093                 }
1094
1095                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1096
1097                 /*
1098                  * The system_wide flag causes a selected event to be opened
1099                  * always without a pid.  Consequently it will never get a
1100                  * POLLHUP, but it is used for tracking in combination with
1101                  * other events, so it should not need to be polled anyway.
1102                  * Therefore don't add it for polling.
1103                  */
1104                 if (!evsel->system_wide &&
1105                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1106                         perf_mmap__put(&maps[idx]);
1107                         return -1;
1108                 }
1109
1110                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1111                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1112                                                    fd) < 0)
1113                                 return -1;
1114                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1115                                                  thread);
1116                 }
1117         }
1118
1119         return 0;
1120 }
1121
1122 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1123                                      struct mmap_params *mp)
1124 {
1125         int cpu, thread;
1126         int nr_cpus = cpu_map__nr(evlist->cpus);
1127         int nr_threads = thread_map__nr(evlist->threads);
1128
1129         pr_debug2("perf event ring buffer mmapped per cpu\n");
1130         for (cpu = 0; cpu < nr_cpus; cpu++) {
1131                 int output = -1;
1132                 int output_backward = -1;
1133
1134                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1135                                               true);
1136
1137                 for (thread = 0; thread < nr_threads; thread++) {
1138                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1139                                                         thread, &output, &output_backward))
1140                                 goto out_unmap;
1141                 }
1142         }
1143
1144         return 0;
1145
1146 out_unmap:
1147         perf_evlist__munmap_nofree(evlist);
1148         return -1;
1149 }
1150
1151 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1152                                         struct mmap_params *mp)
1153 {
1154         int thread;
1155         int nr_threads = thread_map__nr(evlist->threads);
1156
1157         pr_debug2("perf event ring buffer mmapped per thread\n");
1158         for (thread = 0; thread < nr_threads; thread++) {
1159                 int output = -1;
1160                 int output_backward = -1;
1161
1162                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1163                                               false);
1164
1165                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1166                                                 &output, &output_backward))
1167                         goto out_unmap;
1168         }
1169
1170         return 0;
1171
1172 out_unmap:
1173         perf_evlist__munmap_nofree(evlist);
1174         return -1;
1175 }
1176
1177 unsigned long perf_event_mlock_kb_in_pages(void)
1178 {
1179         unsigned long pages;
1180         int max;
1181
1182         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1183                 /*
1184                  * Pick a once upon a time good value, i.e. things look
1185                  * strange since we can't read a sysctl value, but lets not
1186                  * die yet...
1187                  */
1188                 max = 512;
1189         } else {
1190                 max -= (page_size / 1024);
1191         }
1192
1193         pages = (max * 1024) / page_size;
1194         if (!is_power_of_2(pages))
1195                 pages = rounddown_pow_of_two(pages);
1196
1197         return pages;
1198 }
1199
1200 size_t perf_evlist__mmap_size(unsigned long pages)
1201 {
1202         if (pages == UINT_MAX)
1203                 pages = perf_event_mlock_kb_in_pages();
1204         else if (!is_power_of_2(pages))
1205                 return 0;
1206
1207         return (pages + 1) * page_size;
1208 }
1209
1210 static long parse_pages_arg(const char *str, unsigned long min,
1211                             unsigned long max)
1212 {
1213         unsigned long pages, val;
1214         static struct parse_tag tags[] = {
1215                 { .tag  = 'B', .mult = 1       },
1216                 { .tag  = 'K', .mult = 1 << 10 },
1217                 { .tag  = 'M', .mult = 1 << 20 },
1218                 { .tag  = 'G', .mult = 1 << 30 },
1219                 { .tag  = 0 },
1220         };
1221
1222         if (str == NULL)
1223                 return -EINVAL;
1224
1225         val = parse_tag_value(str, tags);
1226         if (val != (unsigned long) -1) {
1227                 /* we got file size value */
1228                 pages = PERF_ALIGN(val, page_size) / page_size;
1229         } else {
1230                 /* we got pages count value */
1231                 char *eptr;
1232                 pages = strtoul(str, &eptr, 10);
1233                 if (*eptr != '\0')
1234                         return -EINVAL;
1235         }
1236
1237         if (pages == 0 && min == 0) {
1238                 /* leave number of pages at 0 */
1239         } else if (!is_power_of_2(pages)) {
1240                 char buf[100];
1241
1242                 /* round pages up to next power of 2 */
1243                 pages = roundup_pow_of_two(pages);
1244                 if (!pages)
1245                         return -EINVAL;
1246
1247                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1248                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1249                         buf, pages);
1250         }
1251
1252         if (pages > max)
1253                 return -EINVAL;
1254
1255         return pages;
1256 }
1257
1258 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1259 {
1260         unsigned long max = UINT_MAX;
1261         long pages;
1262
1263         if (max > SIZE_MAX / page_size)
1264                 max = SIZE_MAX / page_size;
1265
1266         pages = parse_pages_arg(str, 1, max);
1267         if (pages < 0) {
1268                 pr_err("Invalid argument for --mmap_pages/-m\n");
1269                 return -1;
1270         }
1271
1272         *mmap_pages = pages;
1273         return 0;
1274 }
1275
1276 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1277                                   int unset __maybe_unused)
1278 {
1279         return __perf_evlist__parse_mmap_pages(opt->value, str);
1280 }
1281
1282 /**
1283  * perf_evlist__mmap_ex - Create mmaps to receive events.
1284  * @evlist: list of events
1285  * @pages: map length in pages
1286  * @overwrite: overwrite older events?
1287  * @auxtrace_pages - auxtrace map length in pages
1288  * @auxtrace_overwrite - overwrite older auxtrace data?
1289  *
1290  * If @overwrite is %false the user needs to signal event consumption using
1291  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1292  * automatically.
1293  *
1294  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1295  * consumption using auxtrace_mmap__write_tail().
1296  *
1297  * Return: %0 on success, negative error code otherwise.
1298  */
1299 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1300                          bool overwrite, unsigned int auxtrace_pages,
1301                          bool auxtrace_overwrite)
1302 {
1303         struct perf_evsel *evsel;
1304         const struct cpu_map *cpus = evlist->cpus;
1305         const struct thread_map *threads = evlist->threads;
1306         struct mmap_params mp = {
1307                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1308         };
1309
1310         if (!evlist->mmap)
1311                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1312         if (!evlist->mmap)
1313                 return -ENOMEM;
1314
1315         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1316                 return -ENOMEM;
1317
1318         evlist->overwrite = overwrite;
1319         evlist->mmap_len = perf_evlist__mmap_size(pages);
1320         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1321         mp.mask = evlist->mmap_len - page_size - 1;
1322
1323         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1324                                    auxtrace_pages, auxtrace_overwrite);
1325
1326         evlist__for_each_entry(evlist, evsel) {
1327                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1328                     evsel->sample_id == NULL &&
1329                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1330                         return -ENOMEM;
1331         }
1332
1333         if (cpu_map__empty(cpus))
1334                 return perf_evlist__mmap_per_thread(evlist, &mp);
1335
1336         return perf_evlist__mmap_per_cpu(evlist, &mp);
1337 }
1338
1339 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1340                       bool overwrite)
1341 {
1342         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1343 }
1344
1345 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1346 {
1347         struct cpu_map *cpus;
1348         struct thread_map *threads;
1349
1350         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1351
1352         if (!threads)
1353                 return -1;
1354
1355         if (target__uses_dummy_map(target))
1356                 cpus = cpu_map__dummy_new();
1357         else
1358                 cpus = cpu_map__new(target->cpu_list);
1359
1360         if (!cpus)
1361                 goto out_delete_threads;
1362
1363         evlist->has_user_cpus = !!target->cpu_list;
1364
1365         perf_evlist__set_maps(evlist, cpus, threads);
1366
1367         return 0;
1368
1369 out_delete_threads:
1370         thread_map__put(threads);
1371         return -1;
1372 }
1373
1374 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1375                            struct thread_map *threads)
1376 {
1377         /*
1378          * Allow for the possibility that one or another of the maps isn't being
1379          * changed i.e. don't put it.  Note we are assuming the maps that are
1380          * being applied are brand new and evlist is taking ownership of the
1381          * original reference count of 1.  If that is not the case it is up to
1382          * the caller to increase the reference count.
1383          */
1384         if (cpus != evlist->cpus) {
1385                 cpu_map__put(evlist->cpus);
1386                 evlist->cpus = cpu_map__get(cpus);
1387         }
1388
1389         if (threads != evlist->threads) {
1390                 thread_map__put(evlist->threads);
1391                 evlist->threads = thread_map__get(threads);
1392         }
1393
1394         perf_evlist__propagate_maps(evlist);
1395 }
1396
1397 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1398                                    enum perf_event_sample_format bit)
1399 {
1400         struct perf_evsel *evsel;
1401
1402         evlist__for_each_entry(evlist, evsel)
1403                 __perf_evsel__set_sample_bit(evsel, bit);
1404 }
1405
1406 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1407                                      enum perf_event_sample_format bit)
1408 {
1409         struct perf_evsel *evsel;
1410
1411         evlist__for_each_entry(evlist, evsel)
1412                 __perf_evsel__reset_sample_bit(evsel, bit);
1413 }
1414
1415 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1416 {
1417         struct perf_evsel *evsel;
1418         int err = 0;
1419         const int ncpus = cpu_map__nr(evlist->cpus),
1420                   nthreads = thread_map__nr(evlist->threads);
1421
1422         evlist__for_each_entry(evlist, evsel) {
1423                 if (evsel->filter == NULL)
1424                         continue;
1425
1426                 /*
1427                  * filters only work for tracepoint event, which doesn't have cpu limit.
1428                  * So evlist and evsel should always be same.
1429                  */
1430                 err = perf_evsel__apply_filter(evsel, ncpus, nthreads, evsel->filter);
1431                 if (err) {
1432                         *err_evsel = evsel;
1433                         break;
1434                 }
1435         }
1436
1437         return err;
1438 }
1439
1440 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1441 {
1442         struct perf_evsel *evsel;
1443         int err = 0;
1444
1445         evlist__for_each_entry(evlist, evsel) {
1446                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1447                         continue;
1448
1449                 err = perf_evsel__set_filter(evsel, filter);
1450                 if (err)
1451                         break;
1452         }
1453
1454         return err;
1455 }
1456
1457 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1458 {
1459         char *filter;
1460         int ret = -1;
1461         size_t i;
1462
1463         for (i = 0; i < npids; ++i) {
1464                 if (i == 0) {
1465                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1466                                 return -1;
1467                 } else {
1468                         char *tmp;
1469
1470                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1471                                 goto out_free;
1472
1473                         free(filter);
1474                         filter = tmp;
1475                 }
1476         }
1477
1478         ret = perf_evlist__set_filter(evlist, filter);
1479 out_free:
1480         free(filter);
1481         return ret;
1482 }
1483
1484 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1485 {
1486         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1487 }
1488
1489 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1490 {
1491         struct perf_evsel *pos;
1492
1493         if (evlist->nr_entries == 1)
1494                 return true;
1495
1496         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1497                 return false;
1498
1499         evlist__for_each_entry(evlist, pos) {
1500                 if (pos->id_pos != evlist->id_pos ||
1501                     pos->is_pos != evlist->is_pos)
1502                         return false;
1503         }
1504
1505         return true;
1506 }
1507
1508 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1509 {
1510         struct perf_evsel *evsel;
1511
1512         if (evlist->combined_sample_type)
1513                 return evlist->combined_sample_type;
1514
1515         evlist__for_each_entry(evlist, evsel)
1516                 evlist->combined_sample_type |= evsel->attr.sample_type;
1517
1518         return evlist->combined_sample_type;
1519 }
1520
1521 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1522 {
1523         evlist->combined_sample_type = 0;
1524         return __perf_evlist__combined_sample_type(evlist);
1525 }
1526
1527 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1528 {
1529         struct perf_evsel *evsel;
1530         u64 branch_type = 0;
1531
1532         evlist__for_each_entry(evlist, evsel)
1533                 branch_type |= evsel->attr.branch_sample_type;
1534         return branch_type;
1535 }
1536
1537 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1538 {
1539         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1540         u64 read_format = first->attr.read_format;
1541         u64 sample_type = first->attr.sample_type;
1542
1543         evlist__for_each_entry(evlist, pos) {
1544                 if (read_format != pos->attr.read_format)
1545                         return false;
1546         }
1547
1548         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1549         if ((sample_type & PERF_SAMPLE_READ) &&
1550             !(read_format & PERF_FORMAT_ID)) {
1551                 return false;
1552         }
1553
1554         return true;
1555 }
1556
1557 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1558 {
1559         struct perf_evsel *first = perf_evlist__first(evlist);
1560         return first->attr.read_format;
1561 }
1562
1563 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1564 {
1565         struct perf_evsel *first = perf_evlist__first(evlist);
1566         struct perf_sample *data;
1567         u64 sample_type;
1568         u16 size = 0;
1569
1570         if (!first->attr.sample_id_all)
1571                 goto out;
1572
1573         sample_type = first->attr.sample_type;
1574
1575         if (sample_type & PERF_SAMPLE_TID)
1576                 size += sizeof(data->tid) * 2;
1577
1578        if (sample_type & PERF_SAMPLE_TIME)
1579                 size += sizeof(data->time);
1580
1581         if (sample_type & PERF_SAMPLE_ID)
1582                 size += sizeof(data->id);
1583
1584         if (sample_type & PERF_SAMPLE_STREAM_ID)
1585                 size += sizeof(data->stream_id);
1586
1587         if (sample_type & PERF_SAMPLE_CPU)
1588                 size += sizeof(data->cpu) * 2;
1589
1590         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1591                 size += sizeof(data->id);
1592 out:
1593         return size;
1594 }
1595
1596 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1597 {
1598         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1599
1600         evlist__for_each_entry_continue(evlist, pos) {
1601                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1602                         return false;
1603         }
1604
1605         return true;
1606 }
1607
1608 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1609 {
1610         struct perf_evsel *first = perf_evlist__first(evlist);
1611         return first->attr.sample_id_all;
1612 }
1613
1614 void perf_evlist__set_selected(struct perf_evlist *evlist,
1615                                struct perf_evsel *evsel)
1616 {
1617         evlist->selected = evsel;
1618 }
1619
1620 void perf_evlist__close(struct perf_evlist *evlist)
1621 {
1622         struct perf_evsel *evsel;
1623         int ncpus = cpu_map__nr(evlist->cpus);
1624         int nthreads = thread_map__nr(evlist->threads);
1625
1626         evlist__for_each_entry_reverse(evlist, evsel) {
1627                 int n = evsel->cpus ? evsel->cpus->nr : ncpus;
1628                 perf_evsel__close(evsel, n, nthreads);
1629         }
1630 }
1631
1632 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1633 {
1634         struct cpu_map    *cpus;
1635         struct thread_map *threads;
1636         int err = -ENOMEM;
1637
1638         /*
1639          * Try reading /sys/devices/system/cpu/online to get
1640          * an all cpus map.
1641          *
1642          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1643          * code needs an overhaul to properly forward the
1644          * error, and we may not want to do that fallback to a
1645          * default cpu identity map :-\
1646          */
1647         cpus = cpu_map__new(NULL);
1648         if (!cpus)
1649                 goto out;
1650
1651         threads = thread_map__new_dummy();
1652         if (!threads)
1653                 goto out_put;
1654
1655         perf_evlist__set_maps(evlist, cpus, threads);
1656 out:
1657         return err;
1658 out_put:
1659         cpu_map__put(cpus);
1660         goto out;
1661 }
1662
1663 int perf_evlist__open(struct perf_evlist *evlist)
1664 {
1665         struct perf_evsel *evsel;
1666         int err;
1667
1668         /*
1669          * Default: one fd per CPU, all threads, aka systemwide
1670          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1671          */
1672         if (evlist->threads == NULL && evlist->cpus == NULL) {
1673                 err = perf_evlist__create_syswide_maps(evlist);
1674                 if (err < 0)
1675                         goto out_err;
1676         }
1677
1678         perf_evlist__update_id_pos(evlist);
1679
1680         evlist__for_each_entry(evlist, evsel) {
1681                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1682                 if (err < 0)
1683                         goto out_err;
1684         }
1685
1686         return 0;
1687 out_err:
1688         perf_evlist__close(evlist);
1689         errno = -err;
1690         return err;
1691 }
1692
1693 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1694                                   const char *argv[], bool pipe_output,
1695                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1696 {
1697         int child_ready_pipe[2], go_pipe[2];
1698         char bf;
1699
1700         if (pipe(child_ready_pipe) < 0) {
1701                 perror("failed to create 'ready' pipe");
1702                 return -1;
1703         }
1704
1705         if (pipe(go_pipe) < 0) {
1706                 perror("failed to create 'go' pipe");
1707                 goto out_close_ready_pipe;
1708         }
1709
1710         evlist->workload.pid = fork();
1711         if (evlist->workload.pid < 0) {
1712                 perror("failed to fork");
1713                 goto out_close_pipes;
1714         }
1715
1716         if (!evlist->workload.pid) {
1717                 int ret;
1718
1719                 if (pipe_output)
1720                         dup2(2, 1);
1721
1722                 signal(SIGTERM, SIG_DFL);
1723
1724                 close(child_ready_pipe[0]);
1725                 close(go_pipe[1]);
1726                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1727
1728                 /*
1729                  * Tell the parent we're ready to go
1730                  */
1731                 close(child_ready_pipe[1]);
1732
1733                 /*
1734                  * Wait until the parent tells us to go.
1735                  */
1736                 ret = read(go_pipe[0], &bf, 1);
1737                 /*
1738                  * The parent will ask for the execvp() to be performed by
1739                  * writing exactly one byte, in workload.cork_fd, usually via
1740                  * perf_evlist__start_workload().
1741                  *
1742                  * For cancelling the workload without actually running it,
1743                  * the parent will just close workload.cork_fd, without writing
1744                  * anything, i.e. read will return zero and we just exit()
1745                  * here.
1746                  */
1747                 if (ret != 1) {
1748                         if (ret == -1)
1749                                 perror("unable to read pipe");
1750                         exit(ret);
1751                 }
1752
1753                 execvp(argv[0], (char **)argv);
1754
1755                 if (exec_error) {
1756                         union sigval val;
1757
1758                         val.sival_int = errno;
1759                         if (sigqueue(getppid(), SIGUSR1, val))
1760                                 perror(argv[0]);
1761                 } else
1762                         perror(argv[0]);
1763                 exit(-1);
1764         }
1765
1766         if (exec_error) {
1767                 struct sigaction act = {
1768                         .sa_flags     = SA_SIGINFO,
1769                         .sa_sigaction = exec_error,
1770                 };
1771                 sigaction(SIGUSR1, &act, NULL);
1772         }
1773
1774         if (target__none(target)) {
1775                 if (evlist->threads == NULL) {
1776                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1777                                 __func__, __LINE__);
1778                         goto out_close_pipes;
1779                 }
1780                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1781         }
1782
1783         close(child_ready_pipe[1]);
1784         close(go_pipe[0]);
1785         /*
1786          * wait for child to settle
1787          */
1788         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1789                 perror("unable to read pipe");
1790                 goto out_close_pipes;
1791         }
1792
1793         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1794         evlist->workload.cork_fd = go_pipe[1];
1795         close(child_ready_pipe[0]);
1796         return 0;
1797
1798 out_close_pipes:
1799         close(go_pipe[0]);
1800         close(go_pipe[1]);
1801 out_close_ready_pipe:
1802         close(child_ready_pipe[0]);
1803         close(child_ready_pipe[1]);
1804         return -1;
1805 }
1806
1807 int perf_evlist__start_workload(struct perf_evlist *evlist)
1808 {
1809         if (evlist->workload.cork_fd > 0) {
1810                 char bf = 0;
1811                 int ret;
1812                 /*
1813                  * Remove the cork, let it rip!
1814                  */
1815                 ret = write(evlist->workload.cork_fd, &bf, 1);
1816                 if (ret < 0)
1817                         perror("unable to write to pipe");
1818
1819                 close(evlist->workload.cork_fd);
1820                 return ret;
1821         }
1822
1823         return 0;
1824 }
1825
1826 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1827                               struct perf_sample *sample)
1828 {
1829         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1830
1831         if (!evsel)
1832                 return -EFAULT;
1833         return perf_evsel__parse_sample(evsel, event, sample);
1834 }
1835
1836 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1837 {
1838         struct perf_evsel *evsel;
1839         size_t printed = 0;
1840
1841         evlist__for_each_entry(evlist, evsel) {
1842                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1843                                    perf_evsel__name(evsel));
1844         }
1845
1846         return printed + fprintf(fp, "\n");
1847 }
1848
1849 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1850                                int err, char *buf, size_t size)
1851 {
1852         int printed, value;
1853         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1854
1855         switch (err) {
1856         case EACCES:
1857         case EPERM:
1858                 printed = scnprintf(buf, size,
1859                                     "Error:\t%s.\n"
1860                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1861
1862                 value = perf_event_paranoid();
1863
1864                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1865
1866                 if (value >= 2) {
1867                         printed += scnprintf(buf + printed, size - printed,
1868                                              "For your workloads it needs to be <= 1\nHint:\t");
1869                 }
1870                 printed += scnprintf(buf + printed, size - printed,
1871                                      "For system wide tracing it needs to be set to -1.\n");
1872
1873                 printed += scnprintf(buf + printed, size - printed,
1874                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1875                                     "Hint:\tThe current value is %d.", value);
1876                 break;
1877         case EINVAL: {
1878                 struct perf_evsel *first = perf_evlist__first(evlist);
1879                 int max_freq;
1880
1881                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1882                         goto out_default;
1883
1884                 if (first->attr.sample_freq < (u64)max_freq)
1885                         goto out_default;
1886
1887                 printed = scnprintf(buf, size,
1888                                     "Error:\t%s.\n"
1889                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1890                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1891                                     emsg, max_freq, first->attr.sample_freq);
1892                 break;
1893         }
1894         default:
1895 out_default:
1896                 scnprintf(buf, size, "%s", emsg);
1897                 break;
1898         }
1899
1900         return 0;
1901 }
1902
1903 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1904 {
1905         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1906         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1907
1908         switch (err) {
1909         case EPERM:
1910                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1911                 printed += scnprintf(buf + printed, size - printed,
1912                                      "Error:\t%s.\n"
1913                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1914                                      "Hint:\tTried using %zd kB.\n",
1915                                      emsg, pages_max_per_user, pages_attempted);
1916
1917                 if (pages_attempted >= pages_max_per_user) {
1918                         printed += scnprintf(buf + printed, size - printed,
1919                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1920                                              pages_max_per_user + pages_attempted);
1921                 }
1922
1923                 printed += scnprintf(buf + printed, size - printed,
1924                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1925                 break;
1926         default:
1927                 scnprintf(buf, size, "%s", emsg);
1928                 break;
1929         }
1930
1931         return 0;
1932 }
1933
1934 void perf_evlist__to_front(struct perf_evlist *evlist,
1935                            struct perf_evsel *move_evsel)
1936 {
1937         struct perf_evsel *evsel, *n;
1938         LIST_HEAD(move);
1939
1940         if (move_evsel == perf_evlist__first(evlist))
1941                 return;
1942
1943         evlist__for_each_entry_safe(evlist, n, evsel) {
1944                 if (evsel->leader == move_evsel->leader)
1945                         list_move_tail(&evsel->node, &move);
1946         }
1947
1948         list_splice(&move, &evlist->entries);
1949 }
1950
1951 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1952                                      struct perf_evsel *tracking_evsel)
1953 {
1954         struct perf_evsel *evsel;
1955
1956         if (tracking_evsel->tracking)
1957                 return;
1958
1959         evlist__for_each_entry(evlist, evsel) {
1960                 if (evsel != tracking_evsel)
1961                         evsel->tracking = false;
1962         }
1963
1964         tracking_evsel->tracking = true;
1965 }
1966
1967 struct perf_evsel *
1968 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1969                                const char *str)
1970 {
1971         struct perf_evsel *evsel;
1972
1973         evlist__for_each_entry(evlist, evsel) {
1974                 if (!evsel->name)
1975                         continue;
1976                 if (strcmp(str, evsel->name) == 0)
1977                         return evsel;
1978         }
1979
1980         return NULL;
1981 }
1982
1983 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1984                                   enum bkw_mmap_state state)
1985 {
1986         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1987         enum action {
1988                 NONE,
1989                 PAUSE,
1990                 RESUME,
1991         } action = NONE;
1992
1993         if (!evlist->backward_mmap)
1994                 return;
1995
1996         switch (old_state) {
1997         case BKW_MMAP_NOTREADY: {
1998                 if (state != BKW_MMAP_RUNNING)
1999                         goto state_err;;
2000                 break;
2001         }
2002         case BKW_MMAP_RUNNING: {
2003                 if (state != BKW_MMAP_DATA_PENDING)
2004                         goto state_err;
2005                 action = PAUSE;
2006                 break;
2007         }
2008         case BKW_MMAP_DATA_PENDING: {
2009                 if (state != BKW_MMAP_EMPTY)
2010                         goto state_err;
2011                 break;
2012         }
2013         case BKW_MMAP_EMPTY: {
2014                 if (state != BKW_MMAP_RUNNING)
2015                         goto state_err;
2016                 action = RESUME;
2017                 break;
2018         }
2019         default:
2020                 WARN_ONCE(1, "Shouldn't get there\n");
2021         }
2022
2023         evlist->bkw_mmap_state = state;
2024
2025         switch (action) {
2026         case PAUSE:
2027                 perf_evlist__pause(evlist);
2028                 break;
2029         case RESUME:
2030                 perf_evlist__resume(evlist);
2031                 break;
2032         case NONE:
2033         default:
2034                 break;
2035         }
2036
2037 state_err:
2038         return;
2039 }