tools/perf/util/evlist.c

   1 /*
   2  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   3  *
   4  * Parts came from builtin-{top,stat,record}.c, see those files for further
   5  * copyright notes.
   6  *
   7  * Released under the GPL v2. (and only v2, not any later version)
   8  */
   9 #include "util.h"
  10 #include <api/fs/fs.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <poll.h>
  14 #include "cpumap.h"
  15 #include "thread_map.h"
  16 #include "target.h"
  17 #include "evlist.h"
  18 #include "evsel.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "asm/bug.h"
  22 #include <signal.h>
  23 #include <unistd.h>
  24
  25 #include "parse-events.h"
  26 #include <subcmd/parse-options.h>
  27
  28 #include <sys/ioctl.h>
  29 #include <sys/mman.h>
  30
  31 #include <linux/bitops.h>
  32 #include <linux/hash.h>
  33 #include <linux/log2.h>
  34 #include <linux/err.h>
  35
  36 static void perf_mmap__munmap(struct perf_mmap *map);
  37 static void perf_mmap__put(struct perf_mmap *map);
  38
  39 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  40 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  41
  42 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  43                        struct thread_map *threads)
  44 {
  45         int i;
  46
  47         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  48                 INIT_HLIST_HEAD(&evlist->heads[i]);
  49         INIT_LIST_HEAD(&evlist->entries);
  50         perf_evlist__set_maps(evlist, cpus, threads);
  51         fdarray__init(&evlist->pollfd, 64);
  52         evlist->workload.pid = -1;
  53         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  54 }
  55
  56 struct perf_evlist *perf_evlist__new(void)
  57 {
  58         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  59
  60         if (evlist != NULL)
  61                 perf_evlist__init(evlist, NULL, NULL);
  62
  63         return evlist;
  64 }
  65
  66 struct perf_evlist *perf_evlist__new_default(void)
  67 {
  68         struct perf_evlist *evlist = perf_evlist__new();
  69
  70         if (evlist && perf_evlist__add_default(evlist)) {
  71                 perf_evlist__delete(evlist);
  72                 evlist = NULL;
  73         }
  74
  75         return evlist;
  76 }
  77
  78 struct perf_evlist *perf_evlist__new_dummy(void)
  79 {
  80         struct perf_evlist *evlist = perf_evlist__new();
  81
  82         if (evlist && perf_evlist__add_dummy(evlist)) {
  83                 perf_evlist__delete(evlist);
  84                 evlist = NULL;
  85         }
  86
  87         return evlist;
  88 }
  89
  90 /**
  91  * perf_evlist__set_id_pos - set the positions of event ids.
  92  * @evlist: selected event list
  93  *
  94  * Events with compatible sample types all have the same id_pos
  95  * and is_pos.  For convenience, put a copy on evlist.
  96  */
  97 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  98 {
  99         struct perf_evsel *first = perf_evlist__first(evlist);
 100
 101         evlist->id_pos = first->id_pos;
 102         evlist->is_pos = first->is_pos;
 103 }
 104
 105 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
 106 {
 107         struct perf_evsel *evsel;
 108
 109         evlist__for_each_entry(evlist, evsel)
 110                 perf_evsel__calc_id_pos(evsel);
 111
 112         perf_evlist__set_id_pos(evlist);
 113 }
 114
 115 static void perf_evlist__purge(struct perf_evlist *evlist)
 116 {
 117         struct perf_evsel *pos, *n;
 118
 119         evlist__for_each_entry_safe(evlist, n, pos) {
 120                 list_del_init(&pos->node);
 121                 pos->evlist = NULL;
 122                 perf_evsel__delete(pos);
 123         }
 124
 125         evlist->nr_entries = 0;
 126 }
 127
 128 void perf_evlist__exit(struct perf_evlist *evlist)
 129 {
 130         zfree(&evlist->mmap);
 131         zfree(&evlist->backward_mmap);
 132         fdarray__exit(&evlist->pollfd);
 133 }
 134
 135 void perf_evlist__delete(struct perf_evlist *evlist)
 136 {
 137         if (evlist == NULL)
 138                 return;
 139
 140         perf_evlist__munmap(evlist);
 141         perf_evlist__close(evlist);
 142         cpu_map__put(evlist->cpus);
 143         thread_map__put(evlist->threads);
 144         evlist->cpus = NULL;
 145         evlist->threads = NULL;
 146         perf_evlist__purge(evlist);
 147         perf_evlist__exit(evlist);
 148         free(evlist);
 149 }
 150
 151 static void __perf_evlist__propagate_maps(struct perf_evlist *evlist,
 152                                           struct perf_evsel *evsel)
 153 {
 154         /*
 155          * We already have cpus for evsel (via PMU sysfs) so
 156          * keep it, if there's no target cpu list defined.
 157          */
 158         if (!evsel->own_cpus || evlist->has_user_cpus) {
 159                 cpu_map__put(evsel->cpus);
 160                 evsel->cpus = cpu_map__get(evlist->cpus);
 161         } else if (evsel->cpus != evsel->own_cpus) {
 162                 cpu_map__put(evsel->cpus);
 163                 evsel->cpus = cpu_map__get(evsel->own_cpus);
 164         }
 165
 166         thread_map__put(evsel->threads);
 167         evsel->threads = thread_map__get(evlist->threads);
 168 }
 169
 170 static void perf_evlist__propagate_maps(struct perf_evlist *evlist)
 171 {
 172         struct perf_evsel *evsel;
 173
 174         evlist__for_each_entry(evlist, evsel)
 175                 __perf_evlist__propagate_maps(evlist, evsel);
 176 }
 177
 178 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 179 {
 180         entry->evlist = evlist;
 181         list_add_tail(&entry->node, &evlist->entries);
 182         entry->idx = evlist->nr_entries;
 183         entry->tracking = !entry->idx;
 184
 185         if (!evlist->nr_entries++)
 186                 perf_evlist__set_id_pos(evlist);
 187
 188         __perf_evlist__propagate_maps(evlist, entry);
 189 }
 190
 191 void perf_evlist__remove(struct perf_evlist *evlist, struct perf_evsel *evsel)
 192 {
 193         evsel->evlist = NULL;
 194         list_del_init(&evsel->node);
 195         evlist->nr_entries -= 1;
 196 }
 197
 198 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 199                                    struct list_head *list)
 200 {
 201         struct perf_evsel *evsel, *temp;
 202
 203         __evlist__for_each_entry_safe(list, temp, evsel) {
 204                 list_del_init(&evsel->node);
 205                 perf_evlist__add(evlist, evsel);
 206         }
 207 }
 208
 209 void __perf_evlist__set_leader(struct list_head *list)
 210 {
 211         struct perf_evsel *evsel, *leader;
 212
 213         leader = list_entry(list->next, struct perf_evsel, node);
 214         evsel = list_entry(list->prev, struct perf_evsel, node);
 215
 216         leader->nr_members = evsel->idx - leader->idx + 1;
 217
 218         __evlist__for_each_entry(list, evsel) {
 219                 evsel->leader = leader;
 220         }
 221 }
 222
 223 void perf_evlist__set_leader(struct perf_evlist *evlist)
 224 {
 225         if (evlist->nr_entries) {
 226                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 227                 __perf_evlist__set_leader(&evlist->entries);
 228         }
 229 }
 230
 231 void perf_event_attr__set_max_precise_ip(struct perf_event_attr *attr)
 232 {
 233         attr->precise_ip = 3;
 234
 235         while (attr->precise_ip != 0) {
 236                 int fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 237                 if (fd != -1) {
 238                         close(fd);
 239                         break;
 240                 }
 241                 --attr->precise_ip;
 242         }
 243 }
 244
 245 int __perf_evlist__add_default(struct perf_evlist *evlist, bool precise)
 246 {
 247         struct perf_evsel *evsel = perf_evsel__new_cycles(precise);
 248
 249         if (evsel == NULL)
 250                 return -ENOMEM;
 251
 252         perf_evlist__add(evlist, evsel);
 253         return 0;
 254 }
 255
 256 int perf_evlist__add_dummy(struct perf_evlist *evlist)
 257 {
 258         struct perf_event_attr attr = {
 259                 .type   = PERF_TYPE_SOFTWARE,
 260                 .config = PERF_COUNT_SW_DUMMY,
 261                 .size   = sizeof(attr), /* to capture ABI version */
 262         };
 263         struct perf_evsel *evsel = perf_evsel__new(&attr);
 264
 265         if (evsel == NULL)
 266                 return -ENOMEM;
 267
 268         perf_evlist__add(evlist, evsel);
 269         return 0;
 270 }
 271
 272 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 273                                   struct perf_event_attr *attrs, size_t nr_attrs)
 274 {
 275         struct perf_evsel *evsel, *n;
 276         LIST_HEAD(head);
 277         size_t i;
 278
 279         for (i = 0; i < nr_attrs; i++) {
 280                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 281                 if (evsel == NULL)
 282                         goto out_delete_partial_list;
 283                 list_add_tail(&evsel->node, &head);
 284         }
 285
 286         perf_evlist__splice_list_tail(evlist, &head);
 287
 288         return 0;
 289
 290 out_delete_partial_list:
 291         __evlist__for_each_entry_safe(&head, n, evsel)
 292                 perf_evsel__delete(evsel);
 293         return -1;
 294 }
 295
 296 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 297                                      struct perf_event_attr *attrs, size_t nr_attrs)
 298 {
 299         size_t i;
 300
 301         for (i = 0; i < nr_attrs; i++)
 302                 event_attr_init(attrs + i);
 303
 304         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 305 }
 306
 307 struct perf_evsel *
 308 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 309 {
 310         struct perf_evsel *evsel;
 311
 312         evlist__for_each_entry(evlist, evsel) {
 313                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 314                     (int)evsel->attr.config == id)
 315                         return evsel;
 316         }
 317
 318         return NULL;
 319 }
 320
 321 struct perf_evsel *
 322 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 323                                      const char *name)
 324 {
 325         struct perf_evsel *evsel;
 326
 327         evlist__for_each_entry(evlist, evsel) {
 328                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 329                     (strcmp(evsel->name, name) == 0))
 330                         return evsel;
 331         }
 332
 333         return NULL;
 334 }
 335
 336 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 337                            const char *sys, const char *name, void *handler)
 338 {
 339         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 340
 341         if (IS_ERR(evsel))
 342                 return -1;
 343
 344         evsel->handler = handler;
 345         perf_evlist__add(evlist, evsel);
 346         return 0;
 347 }
 348
 349 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 350                                    struct perf_evsel *evsel)
 351 {
 352         if (evsel->system_wide)
 353                 return 1;
 354         else
 355                 return thread_map__nr(evlist->threads);
 356 }
 357
 358 void perf_evlist__disable(struct perf_evlist *evlist)
 359 {
 360         struct perf_evsel *pos;
 361
 362         evlist__for_each_entry(evlist, pos) {
 363                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 364                         continue;
 365                 perf_evsel__disable(pos);
 366         }
 367
 368         evlist->enabled = false;
 369 }
 370
 371 void perf_evlist__enable(struct perf_evlist *evlist)
 372 {
 373         struct perf_evsel *pos;
 374
 375         evlist__for_each_entry(evlist, pos) {
 376                 if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 377                         continue;
 378                 perf_evsel__enable(pos);
 379         }
 380
 381         evlist->enabled = true;
 382 }
 383
 384 void perf_evlist__toggle_enable(struct perf_evlist *evlist)
 385 {
 386         (evlist->enabled ? perf_evlist__disable : perf_evlist__enable)(evlist);
 387 }
 388
 389 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 390                                          struct perf_evsel *evsel, int cpu)
 391 {
 392         int thread;
 393         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 394
 395         if (!evsel->fd)
 396                 return -EINVAL;
 397
 398         for (thread = 0; thread < nr_threads; thread++) {
 399                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 400                 if (err)
 401                         return err;
 402         }
 403         return 0;
 404 }
 405
 406 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 407                                             struct perf_evsel *evsel,
 408                                             int thread)
 409 {
 410         int cpu;
 411         int nr_cpus = cpu_map__nr(evlist->cpus);
 412
 413         if (!evsel->fd)
 414                 return -EINVAL;
 415
 416         for (cpu = 0; cpu < nr_cpus; cpu++) {
 417                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 418                 if (err)
 419                         return err;
 420         }
 421         return 0;
 422 }
 423
 424 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 425                                   struct perf_evsel *evsel, int idx)
 426 {
 427         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 428
 429         if (per_cpu_mmaps)
 430                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 431         else
 432                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 433 }
 434
 435 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 436 {
 437         int nr_cpus = cpu_map__nr(evlist->cpus);
 438         int nr_threads = thread_map__nr(evlist->threads);
 439         int nfds = 0;
 440         struct perf_evsel *evsel;
 441
 442         evlist__for_each_entry(evlist, evsel) {
 443                 if (evsel->system_wide)
 444                         nfds += nr_cpus;
 445                 else
 446                         nfds += nr_cpus * nr_threads;
 447         }
 448
 449         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 450             fdarray__grow(&evlist->pollfd, nfds) < 0)
 451                 return -ENOMEM;
 452
 453         return 0;
 454 }
 455
 456 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd,
 457                                      struct perf_mmap *map, short revent)
 458 {
 459         int pos = fdarray__add(&evlist->pollfd, fd, revent | POLLERR | POLLHUP);
 460         /*
 461          * Save the idx so that when we filter out fds POLLHUP'ed we can
 462          * close the associated evlist->mmap[] entry.
 463          */
 464         if (pos >= 0) {
 465                 evlist->pollfd.priv[pos].ptr = map;
 466
 467                 fcntl(fd, F_SETFL, O_NONBLOCK);
 468         }
 469
 470         return pos;
 471 }
 472
 473 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 474 {
 475         return __perf_evlist__add_pollfd(evlist, fd, NULL, POLLIN);
 476 }
 477
 478 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 479                                          void *arg __maybe_unused)
 480 {
 481         struct perf_mmap *map = fda->priv[fd].ptr;
 482
 483         if (map)
 484                 perf_mmap__put(map);
 485 }
 486
 487 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 488 {
 489         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 490                                perf_evlist__munmap_filtered, NULL);
 491 }
 492
 493 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 494 {
 495         return fdarray__poll(&evlist->pollfd, timeout);
 496 }
 497
 498 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 499                                  struct perf_evsel *evsel,
 500                                  int cpu, int thread, u64 id)
 501 {
 502         int hash;
 503         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 504
 505         sid->id = id;
 506         sid->evsel = evsel;
 507         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 508         hlist_add_head(&sid->node, &evlist->heads[hash]);
 509 }
 510
 511 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 512                          int cpu, int thread, u64 id)
 513 {
 514         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 515         evsel->id[evsel->ids++] = id;
 516 }
 517
 518 int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 519                            struct perf_evsel *evsel,
 520                            int cpu, int thread, int fd)
 521 {
 522         u64 read_data[4] = { 0, };
 523         int id_idx = 1; /* The first entry is the counter value */
 524         u64 id;
 525         int ret;
 526
 527         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 528         if (!ret)
 529                 goto add;
 530
 531         if (errno != ENOTTY)
 532                 return -1;
 533
 534         /* Legacy way to get event id.. All hail to old kernels! */
 535
 536         /*
 537          * This way does not work with group format read, so bail
 538          * out in that case.
 539          */
 540         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 541                 return -1;
 542
 543         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 544             read(fd, &read_data, sizeof(read_data)) == -1)
 545                 return -1;
 546
 547         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 548                 ++id_idx;
 549         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 550                 ++id_idx;
 551
 552         id = read_data[id_idx];
 553
 554  add:
 555         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 556         return 0;
 557 }
 558
 559 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 560                                      struct perf_evsel *evsel, int idx, int cpu,
 561                                      int thread)
 562 {
 563         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 564         sid->idx = idx;
 565         if (evlist->cpus && cpu >= 0)
 566                 sid->cpu = evlist->cpus->map[cpu];
 567         else
 568                 sid->cpu = -1;
 569         if (!evsel->system_wide && evlist->threads && thread >= 0)
 570                 sid->tid = thread_map__pid(evlist->threads, thread);
 571         else
 572                 sid->tid = -1;
 573 }
 574
 575 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 576 {
 577         struct hlist_head *head;
 578         struct perf_sample_id *sid;
 579         int hash;
 580
 581         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 582         head = &evlist->heads[hash];
 583
 584         hlist_for_each_entry(sid, head, node)
 585                 if (sid->id == id)
 586                         return sid;
 587
 588         return NULL;
 589 }
 590
 591 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 592 {
 593         struct perf_sample_id *sid;
 594
 595         if (evlist->nr_entries == 1 || !id)
 596                 return perf_evlist__first(evlist);
 597
 598         sid = perf_evlist__id2sid(evlist, id);
 599         if (sid)
 600                 return sid->evsel;
 601
 602         if (!perf_evlist__sample_id_all(evlist))
 603                 return perf_evlist__first(evlist);
 604
 605         return NULL;
 606 }
 607
 608 struct perf_evsel *perf_evlist__id2evsel_strict(struct perf_evlist *evlist,
 609                                                 u64 id)
 610 {
 611         struct perf_sample_id *sid;
 612
 613         if (!id)
 614                 return NULL;
 615
 616         sid = perf_evlist__id2sid(evlist, id);
 617         if (sid)
 618                 return sid->evsel;
 619
 620         return NULL;
 621 }
 622
 623 static int perf_evlist__event2id(struct perf_evlist *evlist,
 624                                  union perf_event *event, u64 *id)
 625 {
 626         const u64 *array = event->sample.array;
 627         ssize_t n;
 628
 629         n = (event->header.size - sizeof(event->header)) >> 3;
 630
 631         if (event->header.type == PERF_RECORD_SAMPLE) {
 632                 if (evlist->id_pos >= n)
 633                         return -1;
 634                 *id = array[evlist->id_pos];
 635         } else {
 636                 if (evlist->is_pos > n)
 637                         return -1;
 638                 n -= evlist->is_pos;
 639                 *id = array[n];
 640         }
 641         return 0;
 642 }
 643
 644 struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 645                                             union perf_event *event)
 646 {
 647         struct perf_evsel *first = perf_evlist__first(evlist);
 648         struct hlist_head *head;
 649         struct perf_sample_id *sid;
 650         int hash;
 651         u64 id;
 652
 653         if (evlist->nr_entries == 1)
 654                 return first;
 655
 656         if (!first->attr.sample_id_all &&
 657             event->header.type != PERF_RECORD_SAMPLE)
 658                 return first;
 659
 660         if (perf_evlist__event2id(evlist, event, &id))
 661                 return NULL;
 662
 663         /* Synthesized events have an id of zero */
 664         if (!id)
 665                 return first;
 666
 667         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 668         head = &evlist->heads[hash];
 669
 670         hlist_for_each_entry(sid, head, node) {
 671                 if (sid->id == id)
 672                         return sid->evsel;
 673         }
 674         return NULL;
 675 }
 676
 677 static int perf_evlist__set_paused(struct perf_evlist *evlist, bool value)
 678 {
 679         int i;
 680
 681         if (!evlist->backward_mmap)
 682                 return 0;
 683
 684         for (i = 0; i < evlist->nr_mmaps; i++) {
 685                 int fd = evlist->backward_mmap[i].fd;
 686                 int err;
 687
 688                 if (fd < 0)
 689                         continue;
 690                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 691                 if (err)
 692                         return err;
 693         }
 694         return 0;
 695 }
 696
 697 static int perf_evlist__pause(struct perf_evlist *evlist)
 698 {
 699         return perf_evlist__set_paused(evlist, true);
 700 }
 701
 702 static int perf_evlist__resume(struct perf_evlist *evlist)
 703 {
 704         return perf_evlist__set_paused(evlist, false);
 705 }
 706
 707 /* When check_messup is true, 'end' must points to a good entry */
 708 static union perf_event *
 709 perf_mmap__read(struct perf_mmap *md, bool check_messup, u64 start,
 710                 u64 end, u64 *prev)
 711 {
 712         unsigned char *data = md->base + page_size;
 713         union perf_event *event = NULL;
 714         int diff = end - start;
 715
 716         if (check_messup) {
 717                 /*
 718                  * If we're further behind than half the buffer, there's a chance
 719                  * the writer will bite our tail and mess up the samples under us.
 720                  *
 721                  * If we somehow ended up ahead of the 'end', we got messed up.
 722                  *
 723                  * In either case, truncate and restart at 'end'.
 724                  */
 725                 if (diff > md->mask / 2 || diff < 0) {
 726                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 727
 728                         /*
 729                          * 'end' points to a known good entry, start there.
 730                          */
 731                         start = end;
 732                         diff = 0;
 733                 }
 734         }
 735
 736         if (diff >= (int)sizeof(event->header)) {
 737                 size_t size;
 738
 739                 event = (union perf_event *)&data[start & md->mask];
 740                 size = event->header.size;
 741
 742                 if (size < sizeof(event->header) || diff < (int)size) {
 743                         event = NULL;
 744                         goto broken_event;
 745                 }
 746
 747                 /*
 748                  * Event straddles the mmap boundary -- header should always
 749                  * be inside due to u64 alignment of output.
 750                  */
 751                 if ((start & md->mask) + size != ((start + size) & md->mask)) {
 752                         unsigned int offset = start;
 753                         unsigned int len = min(sizeof(*event), size), cpy;
 754                         void *dst = md->event_copy;
 755
 756                         do {
 757                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 758                                 memcpy(dst, &data[offset & md->mask], cpy);
 759                                 offset += cpy;
 760                                 dst += cpy;
 761                                 len -= cpy;
 762                         } while (len);
 763
 764                         event = (union perf_event *) md->event_copy;
 765                 }
 766
 767                 start += size;
 768         }
 769
 770 broken_event:
 771         if (prev)
 772                 *prev = start;
 773
 774         return event;
 775 }
 776
 777 union perf_event *perf_mmap__read_forward(struct perf_mmap *md, bool check_messup)
 778 {
 779         u64 head;
 780         u64 old = md->prev;
 781
 782         /*
 783          * Check if event was unmapped due to a POLLHUP/POLLERR.
 784          */
 785         if (!refcount_read(&md->refcnt))
 786                 return NULL;
 787
 788         head = perf_mmap__read_head(md);
 789
 790         return perf_mmap__read(md, check_messup, old, head, &md->prev);
 791 }
 792
 793 union perf_event *
 794 perf_mmap__read_backward(struct perf_mmap *md)
 795 {
 796         u64 head, end;
 797         u64 start = md->prev;
 798
 799         /*
 800          * Check if event was unmapped due to a POLLHUP/POLLERR.
 801          */
 802         if (!refcount_read(&md->refcnt))
 803                 return NULL;
 804
 805         head = perf_mmap__read_head(md);
 806         if (!head)
 807                 return NULL;
 808
 809         /*
 810          * 'head' pointer starts from 0. Kernel minus sizeof(record) form
 811          * it each time when kernel writes to it, so in fact 'head' is
 812          * negative. 'end' pointer is made manually by adding the size of
 813          * the ring buffer to 'head' pointer, means the validate data can
 814          * read is the whole ring buffer. If 'end' is positive, the ring
 815          * buffer has not fully filled, so we must adjust 'end' to 0.
 816          *
 817          * However, since both 'head' and 'end' is unsigned, we can't
 818          * simply compare 'end' against 0. Here we compare '-head' and
 819          * the size of the ring buffer, where -head is the number of bytes
 820          * kernel write to the ring buffer.
 821          */
 822         if (-head < (u64)(md->mask + 1))
 823                 end = 0;
 824         else
 825                 end = head + md->mask + 1;
 826
 827         return perf_mmap__read(md, false, start, end, &md->prev);
 828 }
 829
 830 union perf_event *perf_evlist__mmap_read_forward(struct perf_evlist *evlist, int idx)
 831 {
 832         struct perf_mmap *md = &evlist->mmap[idx];
 833
 834         /*
 835          * Check messup is required for forward overwritable ring buffer:
 836          * memory pointed by md->prev can be overwritten in this case.
 837          * No need for read-write ring buffer: kernel stop outputting when
 838          * it hit md->prev (perf_mmap__consume()).
 839          */
 840         return perf_mmap__read_forward(md, evlist->overwrite);
 841 }
 842
 843 union perf_event *perf_evlist__mmap_read_backward(struct perf_evlist *evlist, int idx)
 844 {
 845         struct perf_mmap *md = &evlist->mmap[idx];
 846
 847         /*
 848          * No need to check messup for backward ring buffer:
 849          * We can always read arbitrary long data from a backward
 850          * ring buffer unless we forget to pause it before reading.
 851          */
 852         return perf_mmap__read_backward(md);
 853 }
 854
 855 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 856 {
 857         return perf_evlist__mmap_read_forward(evlist, idx);
 858 }
 859
 860 void perf_mmap__read_catchup(struct perf_mmap *md)
 861 {
 862         u64 head;
 863
 864         if (!refcount_read(&md->refcnt))
 865                 return;
 866
 867         head = perf_mmap__read_head(md);
 868         md->prev = head;
 869 }
 870
 871 void perf_evlist__mmap_read_catchup(struct perf_evlist *evlist, int idx)
 872 {
 873         perf_mmap__read_catchup(&evlist->mmap[idx]);
 874 }
 875
 876 static bool perf_mmap__empty(struct perf_mmap *md)
 877 {
 878         return perf_mmap__read_head(md) == md->prev && !md->auxtrace_mmap.base;
 879 }
 880
 881 static void perf_mmap__get(struct perf_mmap *map)
 882 {
 883         refcount_inc(&map->refcnt);
 884 }
 885
 886 static void perf_mmap__put(struct perf_mmap *md)
 887 {
 888         BUG_ON(md->base && refcount_read(&md->refcnt) == 0);
 889
 890         if (refcount_dec_and_test(&md->refcnt))
 891                 perf_mmap__munmap(md);
 892 }
 893
 894 void perf_mmap__consume(struct perf_mmap *md, bool overwrite)
 895 {
 896         if (!overwrite) {
 897                 u64 old = md->prev;
 898
 899                 perf_mmap__write_tail(md, old);
 900         }
 901
 902         if (refcount_read(&md->refcnt) == 1 && perf_mmap__empty(md))
 903                 perf_mmap__put(md);
 904 }
 905
 906 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 907 {
 908         perf_mmap__consume(&evlist->mmap[idx], evlist->overwrite);
 909 }
 910
 911 int __weak auxtrace_mmap__mmap(struct auxtrace_mmap *mm __maybe_unused,
 912                                struct auxtrace_mmap_params *mp __maybe_unused,
 913                                void *userpg __maybe_unused,
 914                                int fd __maybe_unused)
 915 {
 916         return 0;
 917 }
 918
 919 void __weak auxtrace_mmap__munmap(struct auxtrace_mmap *mm __maybe_unused)
 920 {
 921 }
 922
 923 void __weak auxtrace_mmap_params__init(
 924                         struct auxtrace_mmap_params *mp __maybe_unused,
 925                         off_t auxtrace_offset __maybe_unused,
 926                         unsigned int auxtrace_pages __maybe_unused,
 927                         bool auxtrace_overwrite __maybe_unused)
 928 {
 929 }
 930
 931 void __weak auxtrace_mmap_params__set_idx(
 932                         struct auxtrace_mmap_params *mp __maybe_unused,
 933                         struct perf_evlist *evlist __maybe_unused,
 934                         int idx __maybe_unused,
 935                         bool per_cpu __maybe_unused)
 936 {
 937 }
 938
 939 static void perf_mmap__munmap(struct perf_mmap *map)
 940 {
 941         if (map->base != NULL) {
 942                 munmap(map->base, perf_mmap__mmap_len(map));
 943                 map->base = NULL;
 944                 map->fd = -1;
 945                 refcount_set(&map->refcnt, 0);
 946         }
 947         auxtrace_mmap__munmap(&map->auxtrace_mmap);
 948 }
 949
 950 static void perf_evlist__munmap_nofree(struct perf_evlist *evlist)
 951 {
 952         int i;
 953
 954         if (evlist->mmap)
 955                 for (i = 0; i < evlist->nr_mmaps; i++)
 956                         perf_mmap__munmap(&evlist->mmap[i]);
 957
 958         if (evlist->backward_mmap)
 959                 for (i = 0; i < evlist->nr_mmaps; i++)
 960                         perf_mmap__munmap(&evlist->backward_mmap[i]);
 961 }
 962
 963 void perf_evlist__munmap(struct perf_evlist *evlist)
 964 {
 965         perf_evlist__munmap_nofree(evlist);
 966         zfree(&evlist->mmap);
 967         zfree(&evlist->backward_mmap);
 968 }
 969
 970 static struct perf_mmap *perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 971 {
 972         int i;
 973         struct perf_mmap *map;
 974
 975         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 976         if (cpu_map__empty(evlist->cpus))
 977                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 978         map = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 979         if (!map)
 980                 return NULL;
 981
 982         for (i = 0; i < evlist->nr_mmaps; i++) {
 983                 map[i].fd = -1;
 984                 /*
 985                  * When the perf_mmap() call is made we grab one refcount, plus
 986                  * one extra to let perf_evlist__mmap_consume() get the last
 987                  * events after all real references (perf_mmap__get()) are
 988                  * dropped.
 989                  *
 990                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 991                  * thus does perf_mmap__get() on it.
 992                  */
 993                 refcount_set(&map[i].refcnt, 0);
 994         }
 995         return map;
 996 }
 997
 998 struct mmap_params {
 999         int prot;
1000         int mask;
1001         struct auxtrace_mmap_params auxtrace_mp;
1002 };
1003
1004 static int perf_mmap__mmap(struct perf_mmap *map,
1005                            struct mmap_params *mp, int fd)
1006 {
1007         /*
1008          * The last one will be done at perf_evlist__mmap_consume(), so that we
1009          * make sure we don't prevent tools from consuming every last event in
1010          * the ring buffer.
1011          *
1012          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
1013          * anymore, but the last events for it are still in the ring buffer,
1014          * waiting to be consumed.
1015          *
1016          * Tools can chose to ignore this at their own discretion, but the
1017          * evlist layer can't just drop it when filtering events in
1018          * perf_evlist__filter_pollfd().
1019          */
1020         refcount_set(&map->refcnt, 2);
1021         map->prev = 0;
1022         map->mask = mp->mask;
1023         map->base = mmap(NULL, perf_mmap__mmap_len(map), mp->prot,
1024                          MAP_SHARED, fd, 0);
1025         if (map->base == MAP_FAILED) {
1026                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
1027                           errno);
1028                 map->base = NULL;
1029                 return -1;
1030         }
1031         map->fd = fd;
1032
1033         if (auxtrace_mmap__mmap(&map->auxtrace_mmap,
1034                                 &mp->auxtrace_mp, map->base, fd))
1035                 return -1;
1036
1037         return 0;
1038 }
1039
1040 static bool
1041 perf_evlist__should_poll(struct perf_evlist *evlist __maybe_unused,
1042                          struct perf_evsel *evsel)
1043 {
1044         if (evsel->attr.write_backward)
1045                 return false;
1046         return true;
1047 }
1048
1049 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
1050                                        struct mmap_params *mp, int cpu_idx,
1051                                        int thread, int *_output, int *_output_backward)
1052 {
1053         struct perf_evsel *evsel;
1054         int revent;
1055         int evlist_cpu = cpu_map__cpu(evlist->cpus, cpu_idx);
1056
1057         evlist__for_each_entry(evlist, evsel) {
1058                 struct perf_mmap *maps = evlist->mmap;
1059                 int *output = _output;
1060                 int fd;
1061                 int cpu;
1062
1063                 if (evsel->attr.write_backward) {
1064                         output = _output_backward;
1065                         maps = evlist->backward_mmap;
1066
1067                         if (!maps) {
1068                                 maps = perf_evlist__alloc_mmap(evlist);
1069                                 if (!maps)
1070                                         return -1;
1071                                 evlist->backward_mmap = maps;
1072                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
1073                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
1074                         }
1075                 }
1076
1077                 if (evsel->system_wide && thread)
1078                         continue;
1079
1080                 cpu = cpu_map__idx(evsel->cpus, evlist_cpu);
1081                 if (cpu == -1)
1082                         continue;
1083
1084                 fd = FD(evsel, cpu, thread);
1085
1086                 if (*output == -1) {
1087                         *output = fd;
1088
1089                         if (perf_mmap__mmap(&maps[idx], mp, *output)  < 0)
1090                                 return -1;
1091                 } else {
1092                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
1093                                 return -1;
1094
1095                         perf_mmap__get(&maps[idx]);
1096                 }
1097
1098                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
1099
1100                 /*
1101                  * The system_wide flag causes a selected event to be opened
1102                  * always without a pid.  Consequently it will never get a
1103                  * POLLHUP, but it is used for tracking in combination with
1104                  * other events, so it should not need to be polled anyway.
1105                  * Therefore don't add it for polling.
1106                  */
1107                 if (!evsel->system_wide &&
1108                     __perf_evlist__add_pollfd(evlist, fd, &maps[idx], revent) < 0) {
1109                         perf_mmap__put(&maps[idx]);
1110                         return -1;
1111                 }
1112
1113                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
1114                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
1115                                                    fd) < 0)
1116                                 return -1;
1117                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
1118                                                  thread);
1119                 }
1120         }
1121
1122         return 0;
1123 }
1124
1125 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
1126                                      struct mmap_params *mp)
1127 {
1128         int cpu, thread;
1129         int nr_cpus = cpu_map__nr(evlist->cpus);
1130         int nr_threads = thread_map__nr(evlist->threads);
1131
1132         pr_debug2("perf event ring buffer mmapped per cpu\n");
1133         for (cpu = 0; cpu < nr_cpus; cpu++) {
1134                 int output = -1;
1135                 int output_backward = -1;
1136
1137                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
1138                                               true);
1139
1140                 for (thread = 0; thread < nr_threads; thread++) {
1141                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
1142                                                         thread, &output, &output_backward))
1143                                 goto out_unmap;
1144                 }
1145         }
1146
1147         return 0;
1148
1149 out_unmap:
1150         perf_evlist__munmap_nofree(evlist);
1151         return -1;
1152 }
1153
1154 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
1155                                         struct mmap_params *mp)
1156 {
1157         int thread;
1158         int nr_threads = thread_map__nr(evlist->threads);
1159
1160         pr_debug2("perf event ring buffer mmapped per thread\n");
1161         for (thread = 0; thread < nr_threads; thread++) {
1162                 int output = -1;
1163                 int output_backward = -1;
1164
1165                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
1166                                               false);
1167
1168                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
1169                                                 &output, &output_backward))
1170                         goto out_unmap;
1171         }
1172
1173         return 0;
1174
1175 out_unmap:
1176         perf_evlist__munmap_nofree(evlist);
1177         return -1;
1178 }
1179
1180 unsigned long perf_event_mlock_kb_in_pages(void)
1181 {
1182         unsigned long pages;
1183         int max;
1184
1185         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
1186                 /*
1187                  * Pick a once upon a time good value, i.e. things look
1188                  * strange since we can't read a sysctl value, but lets not
1189                  * die yet...
1190                  */
1191                 max = 512;
1192         } else {
1193                 max -= (page_size / 1024);
1194         }
1195
1196         pages = (max * 1024) / page_size;
1197         if (!is_power_of_2(pages))
1198                 pages = rounddown_pow_of_two(pages);
1199
1200         return pages;
1201 }
1202
1203 size_t perf_evlist__mmap_size(unsigned long pages)
1204 {
1205         if (pages == UINT_MAX)
1206                 pages = perf_event_mlock_kb_in_pages();
1207         else if (!is_power_of_2(pages))
1208                 return 0;
1209
1210         return (pages + 1) * page_size;
1211 }
1212
1213 static long parse_pages_arg(const char *str, unsigned long min,
1214                             unsigned long max)
1215 {
1216         unsigned long pages, val;
1217         static struct parse_tag tags[] = {
1218                 { .tag  = 'B', .mult = 1       },
1219                 { .tag  = 'K', .mult = 1 << 10 },
1220                 { .tag  = 'M', .mult = 1 << 20 },
1221                 { .tag  = 'G', .mult = 1 << 30 },
1222                 { .tag  = 0 },
1223         };
1224
1225         if (str == NULL)
1226                 return -EINVAL;
1227
1228         val = parse_tag_value(str, tags);
1229         if (val != (unsigned long) -1) {
1230                 /* we got file size value */
1231                 pages = PERF_ALIGN(val, page_size) / page_size;
1232         } else {
1233                 /* we got pages count value */
1234                 char *eptr;
1235                 pages = strtoul(str, &eptr, 10);
1236                 if (*eptr != '\0')
1237                         return -EINVAL;
1238         }
1239
1240         if (pages == 0 && min == 0) {
1241                 /* leave number of pages at 0 */
1242         } else if (!is_power_of_2(pages)) {
1243                 char buf[100];
1244
1245                 /* round pages up to next power of 2 */
1246                 pages = roundup_pow_of_two(pages);
1247                 if (!pages)
1248                         return -EINVAL;
1249
1250                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
1251                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
1252                         buf, pages);
1253         }
1254
1255         if (pages > max)
1256                 return -EINVAL;
1257
1258         return pages;
1259 }
1260
1261 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
1262 {
1263         unsigned long max = UINT_MAX;
1264         long pages;
1265
1266         if (max > SIZE_MAX / page_size)
1267                 max = SIZE_MAX / page_size;
1268
1269         pages = parse_pages_arg(str, 1, max);
1270         if (pages < 0) {
1271                 pr_err("Invalid argument for --mmap_pages/-m\n");
1272                 return -1;
1273         }
1274
1275         *mmap_pages = pages;
1276         return 0;
1277 }
1278
1279 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
1280                                   int unset __maybe_unused)
1281 {
1282         return __perf_evlist__parse_mmap_pages(opt->value, str);
1283 }
1284
1285 /**
1286  * perf_evlist__mmap_ex - Create mmaps to receive events.
1287  * @evlist: list of events
1288  * @pages: map length in pages
1289  * @overwrite: overwrite older events?
1290  * @auxtrace_pages - auxtrace map length in pages
1291  * @auxtrace_overwrite - overwrite older auxtrace data?
1292  *
1293  * If @overwrite is %false the user needs to signal event consumption using
1294  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
1295  * automatically.
1296  *
1297  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
1298  * consumption using auxtrace_mmap__write_tail().
1299  *
1300  * Return: %0 on success, negative error code otherwise.
1301  */
1302 int perf_evlist__mmap_ex(struct perf_evlist *evlist, unsigned int pages,
1303                          bool overwrite, unsigned int auxtrace_pages,
1304                          bool auxtrace_overwrite)
1305 {
1306         struct perf_evsel *evsel;
1307         const struct cpu_map *cpus = evlist->cpus;
1308         const struct thread_map *threads = evlist->threads;
1309         struct mmap_params mp = {
1310                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1311         };
1312
1313         if (!evlist->mmap)
1314                 evlist->mmap = perf_evlist__alloc_mmap(evlist);
1315         if (!evlist->mmap)
1316                 return -ENOMEM;
1317
1318         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1319                 return -ENOMEM;
1320
1321         evlist->overwrite = overwrite;
1322         evlist->mmap_len = perf_evlist__mmap_size(pages);
1323         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1324         mp.mask = evlist->mmap_len - page_size - 1;
1325
1326         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->mmap_len,
1327                                    auxtrace_pages, auxtrace_overwrite);
1328
1329         evlist__for_each_entry(evlist, evsel) {
1330                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1331                     evsel->sample_id == NULL &&
1332                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1333                         return -ENOMEM;
1334         }
1335
1336         if (cpu_map__empty(cpus))
1337                 return perf_evlist__mmap_per_thread(evlist, &mp);
1338
1339         return perf_evlist__mmap_per_cpu(evlist, &mp);
1340 }
1341
1342 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
1343                       bool overwrite)
1344 {
1345         return perf_evlist__mmap_ex(evlist, pages, overwrite, 0, false);
1346 }
1347
1348 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1349 {
1350         struct cpu_map *cpus;
1351         struct thread_map *threads;
1352
1353         threads = thread_map__new_str(target->pid, target->tid, target->uid);
1354
1355         if (!threads)
1356                 return -1;
1357
1358         if (target__uses_dummy_map(target))
1359                 cpus = cpu_map__dummy_new();
1360         else
1361                 cpus = cpu_map__new(target->cpu_list);
1362
1363         if (!cpus)
1364                 goto out_delete_threads;
1365
1366         evlist->has_user_cpus = !!target->cpu_list;
1367
1368         perf_evlist__set_maps(evlist, cpus, threads);
1369
1370         return 0;
1371
1372 out_delete_threads:
1373         thread_map__put(threads);
1374         return -1;
1375 }
1376
1377 void perf_evlist__set_maps(struct perf_evlist *evlist, struct cpu_map *cpus,
1378                            struct thread_map *threads)
1379 {
1380         /*
1381          * Allow for the possibility that one or another of the maps isn't being
1382          * changed i.e. don't put it.  Note we are assuming the maps that are
1383          * being applied are brand new and evlist is taking ownership of the
1384          * original reference count of 1.  If that is not the case it is up to
1385          * the caller to increase the reference count.
1386          */
1387         if (cpus != evlist->cpus) {
1388                 cpu_map__put(evlist->cpus);
1389                 evlist->cpus = cpu_map__get(cpus);
1390         }
1391
1392         if (threads != evlist->threads) {
1393                 thread_map__put(evlist->threads);
1394                 evlist->threads = thread_map__get(threads);
1395         }
1396
1397         perf_evlist__propagate_maps(evlist);
1398 }
1399
1400 void __perf_evlist__set_sample_bit(struct perf_evlist *evlist,
1401                                    enum perf_event_sample_format bit)
1402 {
1403         struct perf_evsel *evsel;
1404
1405         evlist__for_each_entry(evlist, evsel)
1406                 __perf_evsel__set_sample_bit(evsel, bit);
1407 }
1408
1409 void __perf_evlist__reset_sample_bit(struct perf_evlist *evlist,
1410                                      enum perf_event_sample_format bit)
1411 {
1412         struct perf_evsel *evsel;
1413
1414         evlist__for_each_entry(evlist, evsel)
1415                 __perf_evsel__reset_sample_bit(evsel, bit);
1416 }
1417
1418 int perf_evlist__apply_filters(struct perf_evlist *evlist, struct perf_evsel **err_evsel)
1419 {
1420         struct perf_evsel *evsel;
1421         int err = 0;
1422
1423         evlist__for_each_entry(evlist, evsel) {
1424                 if (evsel->filter == NULL)
1425                         continue;
1426
1427                 /*
1428                  * filters only work for tracepoint event, which doesn't have cpu limit.
1429                  * So evlist and evsel should always be same.
1430                  */
1431                 err = perf_evsel__apply_filter(evsel, evsel->filter);
1432                 if (err) {
1433                         *err_evsel = evsel;
1434                         break;
1435                 }
1436         }
1437
1438         return err;
1439 }
1440
1441 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1442 {
1443         struct perf_evsel *evsel;
1444         int err = 0;
1445
1446         evlist__for_each_entry(evlist, evsel) {
1447                 if (evsel->attr.type != PERF_TYPE_TRACEPOINT)
1448                         continue;
1449
1450                 err = perf_evsel__set_filter(evsel, filter);
1451                 if (err)
1452                         break;
1453         }
1454
1455         return err;
1456 }
1457
1458 int perf_evlist__set_filter_pids(struct perf_evlist *evlist, size_t npids, pid_t *pids)
1459 {
1460         char *filter;
1461         int ret = -1;
1462         size_t i;
1463
1464         for (i = 0; i < npids; ++i) {
1465                 if (i == 0) {
1466                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1467                                 return -1;
1468                 } else {
1469                         char *tmp;
1470
1471                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1472                                 goto out_free;
1473
1474                         free(filter);
1475                         filter = tmp;
1476                 }
1477         }
1478
1479         ret = perf_evlist__set_filter(evlist, filter);
1480 out_free:
1481         free(filter);
1482         return ret;
1483 }
1484
1485 int perf_evlist__set_filter_pid(struct perf_evlist *evlist, pid_t pid)
1486 {
1487         return perf_evlist__set_filter_pids(evlist, 1, &pid);
1488 }
1489
1490 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1491 {
1492         struct perf_evsel *pos;
1493
1494         if (evlist->nr_entries == 1)
1495                 return true;
1496
1497         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1498                 return false;
1499
1500         evlist__for_each_entry(evlist, pos) {
1501                 if (pos->id_pos != evlist->id_pos ||
1502                     pos->is_pos != evlist->is_pos)
1503                         return false;
1504         }
1505
1506         return true;
1507 }
1508
1509 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1510 {
1511         struct perf_evsel *evsel;
1512
1513         if (evlist->combined_sample_type)
1514                 return evlist->combined_sample_type;
1515
1516         evlist__for_each_entry(evlist, evsel)
1517                 evlist->combined_sample_type |= evsel->attr.sample_type;
1518
1519         return evlist->combined_sample_type;
1520 }
1521
1522 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1523 {
1524         evlist->combined_sample_type = 0;
1525         return __perf_evlist__combined_sample_type(evlist);
1526 }
1527
1528 u64 perf_evlist__combined_branch_type(struct perf_evlist *evlist)
1529 {
1530         struct perf_evsel *evsel;
1531         u64 branch_type = 0;
1532
1533         evlist__for_each_entry(evlist, evsel)
1534                 branch_type |= evsel->attr.branch_sample_type;
1535         return branch_type;
1536 }
1537
1538 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1539 {
1540         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1541         u64 read_format = first->attr.read_format;
1542         u64 sample_type = first->attr.sample_type;
1543
1544         evlist__for_each_entry(evlist, pos) {
1545                 if (read_format != pos->attr.read_format)
1546                         return false;
1547         }
1548
1549         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1550         if ((sample_type & PERF_SAMPLE_READ) &&
1551             !(read_format & PERF_FORMAT_ID)) {
1552                 return false;
1553         }
1554
1555         return true;
1556 }
1557
1558 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1559 {
1560         struct perf_evsel *first = perf_evlist__first(evlist);
1561         return first->attr.read_format;
1562 }
1563
1564 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1565 {
1566         struct perf_evsel *first = perf_evlist__first(evlist);
1567         struct perf_sample *data;
1568         u64 sample_type;
1569         u16 size = 0;
1570
1571         if (!first->attr.sample_id_all)
1572                 goto out;
1573
1574         sample_type = first->attr.sample_type;
1575
1576         if (sample_type & PERF_SAMPLE_TID)
1577                 size += sizeof(data->tid) * 2;
1578
1579        if (sample_type & PERF_SAMPLE_TIME)
1580                 size += sizeof(data->time);
1581
1582         if (sample_type & PERF_SAMPLE_ID)
1583                 size += sizeof(data->id);
1584
1585         if (sample_type & PERF_SAMPLE_STREAM_ID)
1586                 size += sizeof(data->stream_id);
1587
1588         if (sample_type & PERF_SAMPLE_CPU)
1589                 size += sizeof(data->cpu) * 2;
1590
1591         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1592                 size += sizeof(data->id);
1593 out:
1594         return size;
1595 }
1596
1597 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1598 {
1599         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1600
1601         evlist__for_each_entry_continue(evlist, pos) {
1602                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1603                         return false;
1604         }
1605
1606         return true;
1607 }
1608
1609 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1610 {
1611         struct perf_evsel *first = perf_evlist__first(evlist);
1612         return first->attr.sample_id_all;
1613 }
1614
1615 void perf_evlist__set_selected(struct perf_evlist *evlist,
1616                                struct perf_evsel *evsel)
1617 {
1618         evlist->selected = evsel;
1619 }
1620
1621 void perf_evlist__close(struct perf_evlist *evlist)
1622 {
1623         struct perf_evsel *evsel;
1624
1625         evlist__for_each_entry_reverse(evlist, evsel)
1626                 perf_evsel__close(evsel);
1627 }
1628
1629 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1630 {
1631         struct cpu_map    *cpus;
1632         struct thread_map *threads;
1633         int err = -ENOMEM;
1634
1635         /*
1636          * Try reading /sys/devices/system/cpu/online to get
1637          * an all cpus map.
1638          *
1639          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1640          * code needs an overhaul to properly forward the
1641          * error, and we may not want to do that fallback to a
1642          * default cpu identity map :-\
1643          */
1644         cpus = cpu_map__new(NULL);
1645         if (!cpus)
1646                 goto out;
1647
1648         threads = thread_map__new_dummy();
1649         if (!threads)
1650                 goto out_put;
1651
1652         perf_evlist__set_maps(evlist, cpus, threads);
1653 out:
1654         return err;
1655 out_put:
1656         cpu_map__put(cpus);
1657         goto out;
1658 }
1659
1660 int perf_evlist__open(struct perf_evlist *evlist)
1661 {
1662         struct perf_evsel *evsel;
1663         int err;
1664
1665         /*
1666          * Default: one fd per CPU, all threads, aka systemwide
1667          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1668          */
1669         if (evlist->threads == NULL && evlist->cpus == NULL) {
1670                 err = perf_evlist__create_syswide_maps(evlist);
1671                 if (err < 0)
1672                         goto out_err;
1673         }
1674
1675         perf_evlist__update_id_pos(evlist);
1676
1677         evlist__for_each_entry(evlist, evsel) {
1678                 err = perf_evsel__open(evsel, evsel->cpus, evsel->threads);
1679                 if (err < 0)
1680                         goto out_err;
1681         }
1682
1683         return 0;
1684 out_err:
1685         perf_evlist__close(evlist);
1686         errno = -err;
1687         return err;
1688 }
1689
1690 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1691                                   const char *argv[], bool pipe_output,
1692                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1693 {
1694         int child_ready_pipe[2], go_pipe[2];
1695         char bf;
1696
1697         if (pipe(child_ready_pipe) < 0) {
1698                 perror("failed to create 'ready' pipe");
1699                 return -1;
1700         }
1701
1702         if (pipe(go_pipe) < 0) {
1703                 perror("failed to create 'go' pipe");
1704                 goto out_close_ready_pipe;
1705         }
1706
1707         evlist->workload.pid = fork();
1708         if (evlist->workload.pid < 0) {
1709                 perror("failed to fork");
1710                 goto out_close_pipes;
1711         }
1712
1713         if (!evlist->workload.pid) {
1714                 int ret;
1715
1716                 if (pipe_output)
1717                         dup2(2, 1);
1718
1719                 signal(SIGTERM, SIG_DFL);
1720
1721                 close(child_ready_pipe[0]);
1722                 close(go_pipe[1]);
1723                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1724
1725                 /*
1726                  * Tell the parent we're ready to go
1727                  */
1728                 close(child_ready_pipe[1]);
1729
1730                 /*
1731                  * Wait until the parent tells us to go.
1732                  */
1733                 ret = read(go_pipe[0], &bf, 1);
1734                 /*
1735                  * The parent will ask for the execvp() to be performed by
1736                  * writing exactly one byte, in workload.cork_fd, usually via
1737                  * perf_evlist__start_workload().
1738                  *
1739                  * For cancelling the workload without actually running it,
1740                  * the parent will just close workload.cork_fd, without writing
1741                  * anything, i.e. read will return zero and we just exit()
1742                  * here.
1743                  */
1744                 if (ret != 1) {
1745                         if (ret == -1)
1746                                 perror("unable to read pipe");
1747                         exit(ret);
1748                 }
1749
1750                 execvp(argv[0], (char **)argv);
1751
1752                 if (exec_error) {
1753                         union sigval val;
1754
1755                         val.sival_int = errno;
1756                         if (sigqueue(getppid(), SIGUSR1, val))
1757                                 perror(argv[0]);
1758                 } else
1759                         perror(argv[0]);
1760                 exit(-1);
1761         }
1762
1763         if (exec_error) {
1764                 struct sigaction act = {
1765                         .sa_flags     = SA_SIGINFO,
1766                         .sa_sigaction = exec_error,
1767                 };
1768                 sigaction(SIGUSR1, &act, NULL);
1769         }
1770
1771         if (target__none(target)) {
1772                 if (evlist->threads == NULL) {
1773                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1774                                 __func__, __LINE__);
1775                         goto out_close_pipes;
1776                 }
1777                 thread_map__set_pid(evlist->threads, 0, evlist->workload.pid);
1778         }
1779
1780         close(child_ready_pipe[1]);
1781         close(go_pipe[0]);
1782         /*
1783          * wait for child to settle
1784          */
1785         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1786                 perror("unable to read pipe");
1787                 goto out_close_pipes;
1788         }
1789
1790         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1791         evlist->workload.cork_fd = go_pipe[1];
1792         close(child_ready_pipe[0]);
1793         return 0;
1794
1795 out_close_pipes:
1796         close(go_pipe[0]);
1797         close(go_pipe[1]);
1798 out_close_ready_pipe:
1799         close(child_ready_pipe[0]);
1800         close(child_ready_pipe[1]);
1801         return -1;
1802 }
1803
1804 int perf_evlist__start_workload(struct perf_evlist *evlist)
1805 {
1806         if (evlist->workload.cork_fd > 0) {
1807                 char bf = 0;
1808                 int ret;
1809                 /*
1810                  * Remove the cork, let it rip!
1811                  */
1812                 ret = write(evlist->workload.cork_fd, &bf, 1);
1813                 if (ret < 0)
1814                         perror("unable to write to pipe");
1815
1816                 close(evlist->workload.cork_fd);
1817                 return ret;
1818         }
1819
1820         return 0;
1821 }
1822
1823 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1824                               struct perf_sample *sample)
1825 {
1826         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1827
1828         if (!evsel)
1829                 return -EFAULT;
1830         return perf_evsel__parse_sample(evsel, event, sample);
1831 }
1832
1833 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1834 {
1835         struct perf_evsel *evsel;
1836         size_t printed = 0;
1837
1838         evlist__for_each_entry(evlist, evsel) {
1839                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1840                                    perf_evsel__name(evsel));
1841         }
1842
1843         return printed + fprintf(fp, "\n");
1844 }
1845
1846 int perf_evlist__strerror_open(struct perf_evlist *evlist,
1847                                int err, char *buf, size_t size)
1848 {
1849         int printed, value;
1850         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1851
1852         switch (err) {
1853         case EACCES:
1854         case EPERM:
1855                 printed = scnprintf(buf, size,
1856                                     "Error:\t%s.\n"
1857                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1858
1859                 value = perf_event_paranoid();
1860
1861                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1862
1863                 if (value >= 2) {
1864                         printed += scnprintf(buf + printed, size - printed,
1865                                              "For your workloads it needs to be <= 1\nHint:\t");
1866                 }
1867                 printed += scnprintf(buf + printed, size - printed,
1868                                      "For system wide tracing it needs to be set to -1.\n");
1869
1870                 printed += scnprintf(buf + printed, size - printed,
1871                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1872                                     "Hint:\tThe current value is %d.", value);
1873                 break;
1874         case EINVAL: {
1875                 struct perf_evsel *first = perf_evlist__first(evlist);
1876                 int max_freq;
1877
1878                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1879                         goto out_default;
1880
1881                 if (first->attr.sample_freq < (u64)max_freq)
1882                         goto out_default;
1883
1884                 printed = scnprintf(buf, size,
1885                                     "Error:\t%s.\n"
1886                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1887                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1888                                     emsg, max_freq, first->attr.sample_freq);
1889                 break;
1890         }
1891         default:
1892 out_default:
1893                 scnprintf(buf, size, "%s", emsg);
1894                 break;
1895         }
1896
1897         return 0;
1898 }
1899
1900 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1901 {
1902         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1903         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1904
1905         switch (err) {
1906         case EPERM:
1907                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1908                 printed += scnprintf(buf + printed, size - printed,
1909                                      "Error:\t%s.\n"
1910                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1911                                      "Hint:\tTried using %zd kB.\n",
1912                                      emsg, pages_max_per_user, pages_attempted);
1913
1914                 if (pages_attempted >= pages_max_per_user) {
1915                         printed += scnprintf(buf + printed, size - printed,
1916                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1917                                              pages_max_per_user + pages_attempted);
1918                 }
1919
1920                 printed += scnprintf(buf + printed, size - printed,
1921                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1922                 break;
1923         default:
1924                 scnprintf(buf, size, "%s", emsg);
1925                 break;
1926         }
1927
1928         return 0;
1929 }
1930
1931 void perf_evlist__to_front(struct perf_evlist *evlist,
1932                            struct perf_evsel *move_evsel)
1933 {
1934         struct perf_evsel *evsel, *n;
1935         LIST_HEAD(move);
1936
1937         if (move_evsel == perf_evlist__first(evlist))
1938                 return;
1939
1940         evlist__for_each_entry_safe(evlist, n, evsel) {
1941                 if (evsel->leader == move_evsel->leader)
1942                         list_move_tail(&evsel->node, &move);
1943         }
1944
1945         list_splice(&move, &evlist->entries);
1946 }
1947
1948 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1949                                      struct perf_evsel *tracking_evsel)
1950 {
1951         struct perf_evsel *evsel;
1952
1953         if (tracking_evsel->tracking)
1954                 return;
1955
1956         evlist__for_each_entry(evlist, evsel) {
1957                 if (evsel != tracking_evsel)
1958                         evsel->tracking = false;
1959         }
1960
1961         tracking_evsel->tracking = true;
1962 }
1963
1964 struct perf_evsel *
1965 perf_evlist__find_evsel_by_str(struct perf_evlist *evlist,
1966                                const char *str)
1967 {
1968         struct perf_evsel *evsel;
1969
1970         evlist__for_each_entry(evlist, evsel) {
1971                 if (!evsel->name)
1972                         continue;
1973                 if (strcmp(str, evsel->name) == 0)
1974                         return evsel;
1975         }
1976
1977         return NULL;
1978 }
1979
1980 void perf_evlist__toggle_bkw_mmap(struct perf_evlist *evlist,
1981                                   enum bkw_mmap_state state)
1982 {
1983         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1984         enum action {
1985                 NONE,
1986                 PAUSE,
1987                 RESUME,
1988         } action = NONE;
1989
1990         if (!evlist->backward_mmap)
1991                 return;
1992
1993         switch (old_state) {
1994         case BKW_MMAP_NOTREADY: {
1995                 if (state != BKW_MMAP_RUNNING)
1996                         goto state_err;;
1997                 break;
1998         }
1999         case BKW_MMAP_RUNNING: {
2000                 if (state != BKW_MMAP_DATA_PENDING)
2001                         goto state_err;
2002                 action = PAUSE;
2003                 break;
2004         }
2005         case BKW_MMAP_DATA_PENDING: {
2006                 if (state != BKW_MMAP_EMPTY)
2007                         goto state_err;
2008                 break;
2009         }
2010         case BKW_MMAP_EMPTY: {
2011                 if (state != BKW_MMAP_RUNNING)
2012                         goto state_err;
2013                 action = RESUME;
2014                 break;
2015         }
2016         default:
2017                 WARN_ONCE(1, "Shouldn't get there\n");
2018         }
2019
2020         evlist->bkw_mmap_state = state;
2021
2022         switch (action) {
2023         case PAUSE:
2024                 perf_evlist__pause(evlist);
2025                 break;
2026         case RESUME:
2027                 perf_evlist__resume(evlist);
2028                 break;
2029         case NONE:
2030         default:
2031                 break;
2032         }
2033
2034 state_err:
2035         return;
2036 }