tools/perf/util/evlist.c

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