Pileus Git - ~andy/linux/blob - 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/debugfs.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 "parse-options.h"
  22
  23 #include <sys/mman.h>
  24
  25 #include <linux/bitops.h>
  26 #include <linux/hash.h>
  27
  28 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  29 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  30
  31 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  32                        struct thread_map *threads)
  33 {
  34         int i;
  35
  36         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  37                 INIT_HLIST_HEAD(&evlist->heads[i]);
  38         INIT_LIST_HEAD(&evlist->entries);
  39         perf_evlist__set_maps(evlist, cpus, threads);
  40         evlist->workload.pid = -1;
  41 }
  42
  43 struct perf_evlist *perf_evlist__new(void)
  44 {
  45         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  46
  47         if (evlist != NULL)
  48                 perf_evlist__init(evlist, NULL, NULL);
  49
  50         return evlist;
  51 }
  52
  53 struct perf_evlist *perf_evlist__new_default(void)
  54 {
  55         struct perf_evlist *evlist = perf_evlist__new();
  56
  57         if (evlist && perf_evlist__add_default(evlist)) {
  58                 perf_evlist__delete(evlist);
  59                 evlist = NULL;
  60         }
  61
  62         return evlist;
  63 }
  64
  65 /**
  66  * perf_evlist__set_id_pos - set the positions of event ids.
  67  * @evlist: selected event list
  68  *
  69  * Events with compatible sample types all have the same id_pos
  70  * and is_pos.  For convenience, put a copy on evlist.
  71  */
  72 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  73 {
  74         struct perf_evsel *first = perf_evlist__first(evlist);
  75
  76         evlist->id_pos = first->id_pos;
  77         evlist->is_pos = first->is_pos;
  78 }
  79
  80 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  81 {
  82         struct perf_evsel *evsel;
  83
  84         list_for_each_entry(evsel, &evlist->entries, node)
  85                 perf_evsel__calc_id_pos(evsel);
  86
  87         perf_evlist__set_id_pos(evlist);
  88 }
  89
  90 static void perf_evlist__purge(struct perf_evlist *evlist)
  91 {
  92         struct perf_evsel *pos, *n;
  93
  94         list_for_each_entry_safe(pos, n, &evlist->entries, node) {
  95                 list_del_init(&pos->node);
  96                 perf_evsel__delete(pos);
  97         }
  98
  99         evlist->nr_entries = 0;
 100 }
 101
 102 void perf_evlist__exit(struct perf_evlist *evlist)
 103 {
 104         zfree(&evlist->mmap);
 105         zfree(&evlist->pollfd);
 106 }
 107
 108 void perf_evlist__delete(struct perf_evlist *evlist)
 109 {
 110         cpu_map__delete(evlist->cpus);
 111         thread_map__delete(evlist->threads);
 112         evlist->cpus = NULL;
 113         evlist->threads = NULL;
 114         perf_evlist__purge(evlist);
 115         perf_evlist__exit(evlist);
 116         free(evlist);
 117 }
 118
 119 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 120 {
 121         list_add_tail(&entry->node, &evlist->entries);
 122         entry->idx = evlist->nr_entries;
 123
 124         if (!evlist->nr_entries++)
 125                 perf_evlist__set_id_pos(evlist);
 126 }
 127
 128 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 129                                    struct list_head *list,
 130                                    int nr_entries)
 131 {
 132         bool set_id_pos = !evlist->nr_entries;
 133
 134         list_splice_tail(list, &evlist->entries);
 135         evlist->nr_entries += nr_entries;
 136         if (set_id_pos)
 137                 perf_evlist__set_id_pos(evlist);
 138 }
 139
 140 void __perf_evlist__set_leader(struct list_head *list)
 141 {
 142         struct perf_evsel *evsel, *leader;
 143
 144         leader = list_entry(list->next, struct perf_evsel, node);
 145         evsel = list_entry(list->prev, struct perf_evsel, node);
 146
 147         leader->nr_members = evsel->idx - leader->idx + 1;
 148
 149         list_for_each_entry(evsel, list, node) {
 150                 evsel->leader = leader;
 151         }
 152 }
 153
 154 void perf_evlist__set_leader(struct perf_evlist *evlist)
 155 {
 156         if (evlist->nr_entries) {
 157                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 158                 __perf_evlist__set_leader(&evlist->entries);
 159         }
 160 }
 161
 162 int perf_evlist__add_default(struct perf_evlist *evlist)
 163 {
 164         struct perf_event_attr attr = {
 165                 .type = PERF_TYPE_HARDWARE,
 166                 .config = PERF_COUNT_HW_CPU_CYCLES,
 167         };
 168         struct perf_evsel *evsel;
 169
 170         event_attr_init(&attr);
 171
 172         evsel = perf_evsel__new(&attr);
 173         if (evsel == NULL)
 174                 goto error;
 175
 176         /* use strdup() because free(evsel) assumes name is allocated */
 177         evsel->name = strdup("cycles");
 178         if (!evsel->name)
 179                 goto error_free;
 180
 181         perf_evlist__add(evlist, evsel);
 182         return 0;
 183 error_free:
 184         perf_evsel__delete(evsel);
 185 error:
 186         return -ENOMEM;
 187 }
 188
 189 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 190                                   struct perf_event_attr *attrs, size_t nr_attrs)
 191 {
 192         struct perf_evsel *evsel, *n;
 193         LIST_HEAD(head);
 194         size_t i;
 195
 196         for (i = 0; i < nr_attrs; i++) {
 197                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 198                 if (evsel == NULL)
 199                         goto out_delete_partial_list;
 200                 list_add_tail(&evsel->node, &head);
 201         }
 202
 203         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 204
 205         return 0;
 206
 207 out_delete_partial_list:
 208         list_for_each_entry_safe(evsel, n, &head, node)
 209                 perf_evsel__delete(evsel);
 210         return -1;
 211 }
 212
 213 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 214                                      struct perf_event_attr *attrs, size_t nr_attrs)
 215 {
 216         size_t i;
 217
 218         for (i = 0; i < nr_attrs; i++)
 219                 event_attr_init(attrs + i);
 220
 221         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 222 }
 223
 224 struct perf_evsel *
 225 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 226 {
 227         struct perf_evsel *evsel;
 228
 229         list_for_each_entry(evsel, &evlist->entries, node) {
 230                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 231                     (int)evsel->attr.config == id)
 232                         return evsel;
 233         }
 234
 235         return NULL;
 236 }
 237
 238 struct perf_evsel *
 239 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 240                                      const char *name)
 241 {
 242         struct perf_evsel *evsel;
 243
 244         list_for_each_entry(evsel, &evlist->entries, node) {
 245                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 246                     (strcmp(evsel->name, name) == 0))
 247                         return evsel;
 248         }
 249
 250         return NULL;
 251 }
 252
 253 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 254                            const char *sys, const char *name, void *handler)
 255 {
 256         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 257
 258         if (evsel == NULL)
 259                 return -1;
 260
 261         evsel->handler = handler;
 262         perf_evlist__add(evlist, evsel);
 263         return 0;
 264 }
 265
 266 void perf_evlist__disable(struct perf_evlist *evlist)
 267 {
 268         int cpu, thread;
 269         struct perf_evsel *pos;
 270         int nr_cpus = cpu_map__nr(evlist->cpus);
 271         int nr_threads = thread_map__nr(evlist->threads);
 272
 273         for (cpu = 0; cpu < nr_cpus; cpu++) {
 274                 list_for_each_entry(pos, &evlist->entries, node) {
 275                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 276                                 continue;
 277                         for (thread = 0; thread < nr_threads; thread++)
 278                                 ioctl(FD(pos, cpu, thread),
 279                                       PERF_EVENT_IOC_DISABLE, 0);
 280                 }
 281         }
 282 }
 283
 284 void perf_evlist__enable(struct perf_evlist *evlist)
 285 {
 286         int cpu, thread;
 287         struct perf_evsel *pos;
 288         int nr_cpus = cpu_map__nr(evlist->cpus);
 289         int nr_threads = thread_map__nr(evlist->threads);
 290
 291         for (cpu = 0; cpu < nr_cpus; cpu++) {
 292                 list_for_each_entry(pos, &evlist->entries, node) {
 293                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 294                                 continue;
 295                         for (thread = 0; thread < nr_threads; thread++)
 296                                 ioctl(FD(pos, cpu, thread),
 297                                       PERF_EVENT_IOC_ENABLE, 0);
 298                 }
 299         }
 300 }
 301
 302 int perf_evlist__disable_event(struct perf_evlist *evlist,
 303                                struct perf_evsel *evsel)
 304 {
 305         int cpu, thread, err;
 306
 307         if (!evsel->fd)
 308                 return 0;
 309
 310         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 311                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 312                         err = ioctl(FD(evsel, cpu, thread),
 313                                     PERF_EVENT_IOC_DISABLE, 0);
 314                         if (err)
 315                                 return err;
 316                 }
 317         }
 318         return 0;
 319 }
 320
 321 int perf_evlist__enable_event(struct perf_evlist *evlist,
 322                               struct perf_evsel *evsel)
 323 {
 324         int cpu, thread, err;
 325
 326         if (!evsel->fd)
 327                 return -EINVAL;
 328
 329         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 330                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 331                         err = ioctl(FD(evsel, cpu, thread),
 332                                     PERF_EVENT_IOC_ENABLE, 0);
 333                         if (err)
 334                                 return err;
 335                 }
 336         }
 337         return 0;
 338 }
 339
 340 static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 341 {
 342         int nr_cpus = cpu_map__nr(evlist->cpus);
 343         int nr_threads = thread_map__nr(evlist->threads);
 344         int nfds = nr_cpus * nr_threads * evlist->nr_entries;
 345         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 346         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 347 }
 348
 349 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 350 {
 351         fcntl(fd, F_SETFL, O_NONBLOCK);
 352         evlist->pollfd[evlist->nr_fds].fd = fd;
 353         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 354         evlist->nr_fds++;
 355 }
 356
 357 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 358                                  struct perf_evsel *evsel,
 359                                  int cpu, int thread, u64 id)
 360 {
 361         int hash;
 362         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 363
 364         sid->id = id;
 365         sid->evsel = evsel;
 366         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 367         hlist_add_head(&sid->node, &evlist->heads[hash]);
 368 }
 369
 370 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 371                          int cpu, int thread, u64 id)
 372 {
 373         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 374         evsel->id[evsel->ids++] = id;
 375 }
 376
 377 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 378                                   struct perf_evsel *evsel,
 379                                   int cpu, int thread, int fd)
 380 {
 381         u64 read_data[4] = { 0, };
 382         int id_idx = 1; /* The first entry is the counter value */
 383         u64 id;
 384         int ret;
 385
 386         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 387         if (!ret)
 388                 goto add;
 389
 390         if (errno != ENOTTY)
 391                 return -1;
 392
 393         /* Legacy way to get event id.. All hail to old kernels! */
 394
 395         /*
 396          * This way does not work with group format read, so bail
 397          * out in that case.
 398          */
 399         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 400                 return -1;
 401
 402         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 403             read(fd, &read_data, sizeof(read_data)) == -1)
 404                 return -1;
 405
 406         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 407                 ++id_idx;
 408         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 409                 ++id_idx;
 410
 411         id = read_data[id_idx];
 412
 413  add:
 414         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 415         return 0;
 416 }
 417
 418 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 419 {
 420         struct hlist_head *head;
 421         struct perf_sample_id *sid;
 422         int hash;
 423
 424         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 425         head = &evlist->heads[hash];
 426
 427         hlist_for_each_entry(sid, head, node)
 428                 if (sid->id == id)
 429                         return sid;
 430
 431         return NULL;
 432 }
 433
 434 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 435 {
 436         struct perf_sample_id *sid;
 437
 438         if (evlist->nr_entries == 1)
 439                 return perf_evlist__first(evlist);
 440
 441         sid = perf_evlist__id2sid(evlist, id);
 442         if (sid)
 443                 return sid->evsel;
 444
 445         if (!perf_evlist__sample_id_all(evlist))
 446                 return perf_evlist__first(evlist);
 447
 448         return NULL;
 449 }
 450
 451 static int perf_evlist__event2id(struct perf_evlist *evlist,
 452                                  union perf_event *event, u64 *id)
 453 {
 454         const u64 *array = event->sample.array;
 455         ssize_t n;
 456
 457         n = (event->header.size - sizeof(event->header)) >> 3;
 458
 459         if (event->header.type == PERF_RECORD_SAMPLE) {
 460                 if (evlist->id_pos >= n)
 461                         return -1;
 462                 *id = array[evlist->id_pos];
 463         } else {
 464                 if (evlist->is_pos > n)
 465                         return -1;
 466                 n -= evlist->is_pos;
 467                 *id = array[n];
 468         }
 469         return 0;
 470 }
 471
 472 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 473                                                    union perf_event *event)
 474 {
 475         struct perf_evsel *first = perf_evlist__first(evlist);
 476         struct hlist_head *head;
 477         struct perf_sample_id *sid;
 478         int hash;
 479         u64 id;
 480
 481         if (evlist->nr_entries == 1)
 482                 return first;
 483
 484         if (!first->attr.sample_id_all &&
 485             event->header.type != PERF_RECORD_SAMPLE)
 486                 return first;
 487
 488         if (perf_evlist__event2id(evlist, event, &id))
 489                 return NULL;
 490
 491         /* Synthesized events have an id of zero */
 492         if (!id)
 493                 return first;
 494
 495         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 496         head = &evlist->heads[hash];
 497
 498         hlist_for_each_entry(sid, head, node) {
 499                 if (sid->id == id)
 500                         return sid->evsel;
 501         }
 502         return NULL;
 503 }
 504
 505 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 506 {
 507         struct perf_mmap *md = &evlist->mmap[idx];
 508         unsigned int head = perf_mmap__read_head(md);
 509         unsigned int old = md->prev;
 510         unsigned char *data = md->base + page_size;
 511         union perf_event *event = NULL;
 512
 513         if (evlist->overwrite) {
 514                 /*
 515                  * If we're further behind than half the buffer, there's a chance
 516                  * the writer will bite our tail and mess up the samples under us.
 517                  *
 518                  * If we somehow ended up ahead of the head, we got messed up.
 519                  *
 520                  * In either case, truncate and restart at head.
 521                  */
 522                 int diff = head - old;
 523                 if (diff > md->mask / 2 || diff < 0) {
 524                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 525
 526                         /*
 527                          * head points to a known good entry, start there.
 528                          */
 529                         old = head;
 530                 }
 531         }
 532
 533         if (old != head) {
 534                 size_t size;
 535
 536                 event = (union perf_event *)&data[old & md->mask];
 537                 size = event->header.size;
 538
 539                 /*
 540                  * Event straddles the mmap boundary -- header should always
 541                  * be inside due to u64 alignment of output.
 542                  */
 543                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 544                         unsigned int offset = old;
 545                         unsigned int len = min(sizeof(*event), size), cpy;
 546                         void *dst = md->event_copy;
 547
 548                         do {
 549                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 550                                 memcpy(dst, &data[offset & md->mask], cpy);
 551                                 offset += cpy;
 552                                 dst += cpy;
 553                                 len -= cpy;
 554                         } while (len);
 555
 556                         event = (union perf_event *) md->event_copy;
 557                 }
 558
 559                 old += size;
 560         }
 561
 562         md->prev = old;
 563
 564         return event;
 565 }
 566
 567 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 568 {
 569         if (!evlist->overwrite) {
 570                 struct perf_mmap *md = &evlist->mmap[idx];
 571                 unsigned int old = md->prev;
 572
 573                 perf_mmap__write_tail(md, old);
 574         }
 575 }
 576
 577 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 578 {
 579         if (evlist->mmap[idx].base != NULL) {
 580                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 581                 evlist->mmap[idx].base = NULL;
 582         }
 583 }
 584
 585 void perf_evlist__munmap(struct perf_evlist *evlist)
 586 {
 587         int i;
 588
 589         for (i = 0; i < evlist->nr_mmaps; i++)
 590                 __perf_evlist__munmap(evlist, i);
 591
 592         zfree(&evlist->mmap);
 593 }
 594
 595 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 596 {
 597         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 598         if (cpu_map__empty(evlist->cpus))
 599                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 600         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 601         return evlist->mmap != NULL ? 0 : -ENOMEM;
 602 }
 603
 604 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 605                                int idx, int prot, int mask, int fd)
 606 {
 607         evlist->mmap[idx].prev = 0;
 608         evlist->mmap[idx].mask = mask;
 609         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 610                                       MAP_SHARED, fd, 0);
 611         if (evlist->mmap[idx].base == MAP_FAILED) {
 612                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 613                           errno);
 614                 evlist->mmap[idx].base = NULL;
 615                 return -1;
 616         }
 617
 618         perf_evlist__add_pollfd(evlist, fd);
 619         return 0;
 620 }
 621
 622 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 623                                        int prot, int mask, int cpu, int thread,
 624                                        int *output)
 625 {
 626         struct perf_evsel *evsel;
 627
 628         list_for_each_entry(evsel, &evlist->entries, node) {
 629                 int fd = FD(evsel, cpu, thread);
 630
 631                 if (*output == -1) {
 632                         *output = fd;
 633                         if (__perf_evlist__mmap(evlist, idx, prot, mask,
 634                                                 *output) < 0)
 635                                 return -1;
 636                 } else {
 637                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 638                                 return -1;
 639                 }
 640
 641                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 642                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 643                         return -1;
 644         }
 645
 646         return 0;
 647 }
 648
 649 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot,
 650                                      int mask)
 651 {
 652         int cpu, thread;
 653         int nr_cpus = cpu_map__nr(evlist->cpus);
 654         int nr_threads = thread_map__nr(evlist->threads);
 655
 656         pr_debug2("perf event ring buffer mmapped per cpu\n");
 657         for (cpu = 0; cpu < nr_cpus; cpu++) {
 658                 int output = -1;
 659
 660                 for (thread = 0; thread < nr_threads; thread++) {
 661                         if (perf_evlist__mmap_per_evsel(evlist, cpu, prot, mask,
 662                                                         cpu, thread, &output))
 663                                 goto out_unmap;
 664                 }
 665         }
 666
 667         return 0;
 668
 669 out_unmap:
 670         for (cpu = 0; cpu < nr_cpus; cpu++)
 671                 __perf_evlist__munmap(evlist, cpu);
 672         return -1;
 673 }
 674
 675 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot,
 676                                         int mask)
 677 {
 678         int thread;
 679         int nr_threads = thread_map__nr(evlist->threads);
 680
 681         pr_debug2("perf event ring buffer mmapped per thread\n");
 682         for (thread = 0; thread < nr_threads; thread++) {
 683                 int output = -1;
 684
 685                 if (perf_evlist__mmap_per_evsel(evlist, thread, prot, mask, 0,
 686                                                 thread, &output))
 687                         goto out_unmap;
 688         }
 689
 690         return 0;
 691
 692 out_unmap:
 693         for (thread = 0; thread < nr_threads; thread++)
 694                 __perf_evlist__munmap(evlist, thread);
 695         return -1;
 696 }
 697
 698 static size_t perf_evlist__mmap_size(unsigned long pages)
 699 {
 700         /* 512 kiB: default amount of unprivileged mlocked memory */
 701         if (pages == UINT_MAX)
 702                 pages = (512 * 1024) / page_size;
 703         else if (!is_power_of_2(pages))
 704                 return 0;
 705
 706         return (pages + 1) * page_size;
 707 }
 708
 709 static long parse_pages_arg(const char *str, unsigned long min,
 710                             unsigned long max)
 711 {
 712         unsigned long pages, val;
 713         static struct parse_tag tags[] = {
 714                 { .tag  = 'B', .mult = 1       },
 715                 { .tag  = 'K', .mult = 1 << 10 },
 716                 { .tag  = 'M', .mult = 1 << 20 },
 717                 { .tag  = 'G', .mult = 1 << 30 },
 718                 { .tag  = 0 },
 719         };
 720
 721         if (str == NULL)
 722                 return -EINVAL;
 723
 724         val = parse_tag_value(str, tags);
 725         if (val != (unsigned long) -1) {
 726                 /* we got file size value */
 727                 pages = PERF_ALIGN(val, page_size) / page_size;
 728         } else {
 729                 /* we got pages count value */
 730                 char *eptr;
 731                 pages = strtoul(str, &eptr, 10);
 732                 if (*eptr != '\0')
 733                         return -EINVAL;
 734         }
 735
 736         if (pages == 0 && min == 0) {
 737                 /* leave number of pages at 0 */
 738         } else if (!is_power_of_2(pages)) {
 739                 /* round pages up to next power of 2 */
 740                 pages = next_pow2_l(pages);
 741                 if (!pages)
 742                         return -EINVAL;
 743                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 744                         pages * page_size, pages);
 745         }
 746
 747         if (pages > max)
 748                 return -EINVAL;
 749
 750         return pages;
 751 }
 752
 753 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 754                                   int unset __maybe_unused)
 755 {
 756         unsigned int *mmap_pages = opt->value;
 757         unsigned long max = UINT_MAX;
 758         long pages;
 759
 760         if (max > SIZE_MAX / page_size)
 761                 max = SIZE_MAX / page_size;
 762
 763         pages = parse_pages_arg(str, 1, max);
 764         if (pages < 0) {
 765                 pr_err("Invalid argument for --mmap_pages/-m\n");
 766                 return -1;
 767         }
 768
 769         *mmap_pages = pages;
 770         return 0;
 771 }
 772
 773 /**
 774  * perf_evlist__mmap - Create mmaps to receive events.
 775  * @evlist: list of events
 776  * @pages: map length in pages
 777  * @overwrite: overwrite older events?
 778  *
 779  * If @overwrite is %false the user needs to signal event consumption using
 780  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 781  * automatically.
 782  *
 783  * Return: %0 on success, negative error code otherwise.
 784  */
 785 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 786                       bool overwrite)
 787 {
 788         struct perf_evsel *evsel;
 789         const struct cpu_map *cpus = evlist->cpus;
 790         const struct thread_map *threads = evlist->threads;
 791         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
 792
 793         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 794                 return -ENOMEM;
 795
 796         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 797                 return -ENOMEM;
 798
 799         evlist->overwrite = overwrite;
 800         evlist->mmap_len = perf_evlist__mmap_size(pages);
 801         pr_debug("mmap size %zuB\n", evlist->mmap_len);
 802         mask = evlist->mmap_len - page_size - 1;
 803
 804         list_for_each_entry(evsel, &evlist->entries, node) {
 805                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 806                     evsel->sample_id == NULL &&
 807                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
 808                         return -ENOMEM;
 809         }
 810
 811         if (cpu_map__empty(cpus))
 812                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 813
 814         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 815 }
 816
 817 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
 818 {
 819         evlist->threads = thread_map__new_str(target->pid, target->tid,
 820                                               target->uid);
 821
 822         if (evlist->threads == NULL)
 823                 return -1;
 824
 825         if (target__uses_dummy_map(target))
 826                 evlist->cpus = cpu_map__dummy_new();
 827         else
 828                 evlist->cpus = cpu_map__new(target->cpu_list);
 829
 830         if (evlist->cpus == NULL)
 831                 goto out_delete_threads;
 832
 833         return 0;
 834
 835 out_delete_threads:
 836         thread_map__delete(evlist->threads);
 837         return -1;
 838 }
 839
 840 int perf_evlist__apply_filters(struct perf_evlist *evlist)
 841 {
 842         struct perf_evsel *evsel;
 843         int err = 0;
 844         const int ncpus = cpu_map__nr(evlist->cpus),
 845                   nthreads = thread_map__nr(evlist->threads);
 846
 847         list_for_each_entry(evsel, &evlist->entries, node) {
 848                 if (evsel->filter == NULL)
 849                         continue;
 850
 851                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
 852                 if (err)
 853                         break;
 854         }
 855
 856         return err;
 857 }
 858
 859 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 860 {
 861         struct perf_evsel *evsel;
 862         int err = 0;
 863         const int ncpus = cpu_map__nr(evlist->cpus),
 864                   nthreads = thread_map__nr(evlist->threads);
 865
 866         list_for_each_entry(evsel, &evlist->entries, node) {
 867                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
 868                 if (err)
 869                         break;
 870         }
 871
 872         return err;
 873 }
 874
 875 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 876 {
 877         struct perf_evsel *pos;
 878
 879         if (evlist->nr_entries == 1)
 880                 return true;
 881
 882         if (evlist->id_pos < 0 || evlist->is_pos < 0)
 883                 return false;
 884
 885         list_for_each_entry(pos, &evlist->entries, node) {
 886                 if (pos->id_pos != evlist->id_pos ||
 887                     pos->is_pos != evlist->is_pos)
 888                         return false;
 889         }
 890
 891         return true;
 892 }
 893
 894 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 895 {
 896         struct perf_evsel *evsel;
 897
 898         if (evlist->combined_sample_type)
 899                 return evlist->combined_sample_type;
 900
 901         list_for_each_entry(evsel, &evlist->entries, node)
 902                 evlist->combined_sample_type |= evsel->attr.sample_type;
 903
 904         return evlist->combined_sample_type;
 905 }
 906
 907 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 908 {
 909         evlist->combined_sample_type = 0;
 910         return __perf_evlist__combined_sample_type(evlist);
 911 }
 912
 913 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
 914 {
 915         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 916         u64 read_format = first->attr.read_format;
 917         u64 sample_type = first->attr.sample_type;
 918
 919         list_for_each_entry_continue(pos, &evlist->entries, node) {
 920                 if (read_format != pos->attr.read_format)
 921                         return false;
 922         }
 923
 924         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
 925         if ((sample_type & PERF_SAMPLE_READ) &&
 926             !(read_format & PERF_FORMAT_ID)) {
 927                 return false;
 928         }
 929
 930         return true;
 931 }
 932
 933 u64 perf_evlist__read_format(struct perf_evlist *evlist)
 934 {
 935         struct perf_evsel *first = perf_evlist__first(evlist);
 936         return first->attr.read_format;
 937 }
 938
 939 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
 940 {
 941         struct perf_evsel *first = perf_evlist__first(evlist);
 942         struct perf_sample *data;
 943         u64 sample_type;
 944         u16 size = 0;
 945
 946         if (!first->attr.sample_id_all)
 947                 goto out;
 948
 949         sample_type = first->attr.sample_type;
 950
 951         if (sample_type & PERF_SAMPLE_TID)
 952                 size += sizeof(data->tid) * 2;
 953
 954        if (sample_type & PERF_SAMPLE_TIME)
 955                 size += sizeof(data->time);
 956
 957         if (sample_type & PERF_SAMPLE_ID)
 958                 size += sizeof(data->id);
 959
 960         if (sample_type & PERF_SAMPLE_STREAM_ID)
 961                 size += sizeof(data->stream_id);
 962
 963         if (sample_type & PERF_SAMPLE_CPU)
 964                 size += sizeof(data->cpu) * 2;
 965
 966         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 967                 size += sizeof(data->id);
 968 out:
 969         return size;
 970 }
 971
 972 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
 973 {
 974         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 975
 976         list_for_each_entry_continue(pos, &evlist->entries, node) {
 977                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
 978                         return false;
 979         }
 980
 981         return true;
 982 }
 983
 984 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
 985 {
 986         struct perf_evsel *first = perf_evlist__first(evlist);
 987         return first->attr.sample_id_all;
 988 }
 989
 990 void perf_evlist__set_selected(struct perf_evlist *evlist,
 991                                struct perf_evsel *evsel)
 992 {
 993         evlist->selected = evsel;
 994 }
 995
 996 void perf_evlist__close(struct perf_evlist *evlist)
 997 {
 998         struct perf_evsel *evsel;
 999         int ncpus = cpu_map__nr(evlist->cpus);
1000         int nthreads = thread_map__nr(evlist->threads);
1001
1002         list_for_each_entry_reverse(evsel, &evlist->entries, node)
1003                 perf_evsel__close(evsel, ncpus, nthreads);
1004 }
1005
1006 int perf_evlist__open(struct perf_evlist *evlist)
1007 {
1008         struct perf_evsel *evsel;
1009         int err;
1010
1011         perf_evlist__update_id_pos(evlist);
1012
1013         list_for_each_entry(evsel, &evlist->entries, node) {
1014                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1015                 if (err < 0)
1016                         goto out_err;
1017         }
1018
1019         return 0;
1020 out_err:
1021         perf_evlist__close(evlist);
1022         errno = -err;
1023         return err;
1024 }
1025
1026 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1027                                   const char *argv[], bool pipe_output,
1028                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1029 {
1030         int child_ready_pipe[2], go_pipe[2];
1031         char bf;
1032
1033         if (pipe(child_ready_pipe) < 0) {
1034                 perror("failed to create 'ready' pipe");
1035                 return -1;
1036         }
1037
1038         if (pipe(go_pipe) < 0) {
1039                 perror("failed to create 'go' pipe");
1040                 goto out_close_ready_pipe;
1041         }
1042
1043         evlist->workload.pid = fork();
1044         if (evlist->workload.pid < 0) {
1045                 perror("failed to fork");
1046                 goto out_close_pipes;
1047         }
1048
1049         if (!evlist->workload.pid) {
1050                 if (pipe_output)
1051                         dup2(2, 1);
1052
1053                 signal(SIGTERM, SIG_DFL);
1054
1055                 close(child_ready_pipe[0]);
1056                 close(go_pipe[1]);
1057                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1058
1059                 /*
1060                  * Tell the parent we're ready to go
1061                  */
1062                 close(child_ready_pipe[1]);
1063
1064                 /*
1065                  * Wait until the parent tells us to go.
1066                  */
1067                 if (read(go_pipe[0], &bf, 1) == -1)
1068                         perror("unable to read pipe");
1069
1070                 execvp(argv[0], (char **)argv);
1071
1072                 if (exec_error) {
1073                         union sigval val;
1074
1075                         val.sival_int = errno;
1076                         if (sigqueue(getppid(), SIGUSR1, val))
1077                                 perror(argv[0]);
1078                 } else
1079                         perror(argv[0]);
1080                 exit(-1);
1081         }
1082
1083         if (exec_error) {
1084                 struct sigaction act = {
1085                         .sa_flags     = SA_SIGINFO,
1086                         .sa_sigaction = exec_error,
1087                 };
1088                 sigaction(SIGUSR1, &act, NULL);
1089         }
1090
1091         if (target__none(target))
1092                 evlist->threads->map[0] = evlist->workload.pid;
1093
1094         close(child_ready_pipe[1]);
1095         close(go_pipe[0]);
1096         /*
1097          * wait for child to settle
1098          */
1099         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1100                 perror("unable to read pipe");
1101                 goto out_close_pipes;
1102         }
1103
1104         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1105         evlist->workload.cork_fd = go_pipe[1];
1106         close(child_ready_pipe[0]);
1107         return 0;
1108
1109 out_close_pipes:
1110         close(go_pipe[0]);
1111         close(go_pipe[1]);
1112 out_close_ready_pipe:
1113         close(child_ready_pipe[0]);
1114         close(child_ready_pipe[1]);
1115         return -1;
1116 }
1117
1118 int perf_evlist__start_workload(struct perf_evlist *evlist)
1119 {
1120         if (evlist->workload.cork_fd > 0) {
1121                 char bf = 0;
1122                 int ret;
1123                 /*
1124                  * Remove the cork, let it rip!
1125                  */
1126                 ret = write(evlist->workload.cork_fd, &bf, 1);
1127                 if (ret < 0)
1128                         perror("enable to write to pipe");
1129
1130                 close(evlist->workload.cork_fd);
1131                 return ret;
1132         }
1133
1134         return 0;
1135 }
1136
1137 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1138                               struct perf_sample *sample)
1139 {
1140         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1141
1142         if (!evsel)
1143                 return -EFAULT;
1144         return perf_evsel__parse_sample(evsel, event, sample);
1145 }
1146
1147 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1148 {
1149         struct perf_evsel *evsel;
1150         size_t printed = 0;
1151
1152         list_for_each_entry(evsel, &evlist->entries, node) {
1153                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1154                                    perf_evsel__name(evsel));
1155         }
1156
1157         return printed + fprintf(fp, "\n");
1158 }
1159
1160 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1161                              int err, char *buf, size_t size)
1162 {
1163         char sbuf[128];
1164
1165         switch (err) {
1166         case ENOENT:
1167                 scnprintf(buf, size, "%s",
1168                           "Error:\tUnable to find debugfs\n"
1169                           "Hint:\tWas your kernel was compiled with debugfs support?\n"
1170                           "Hint:\tIs the debugfs filesystem mounted?\n"
1171                           "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1172                 break;
1173         case EACCES:
1174                 scnprintf(buf, size,
1175                           "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1176                           "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1177                           debugfs_mountpoint, debugfs_mountpoint);
1178                 break;
1179         default:
1180                 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1181                 break;
1182         }
1183
1184         return 0;
1185 }
1186
1187 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1188                                int err, char *buf, size_t size)
1189 {
1190         int printed, value;
1191         char sbuf[128], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1192
1193         switch (err) {
1194         case EACCES:
1195         case EPERM:
1196                 printed = scnprintf(buf, size,
1197                                     "Error:\t%s.\n"
1198                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1199
1200                 value = perf_event_paranoid();
1201
1202                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1203
1204                 if (value >= 2) {
1205                         printed += scnprintf(buf + printed, size - printed,
1206                                              "For your workloads it needs to be <= 1\nHint:\t");
1207                 }
1208                 printed += scnprintf(buf + printed, size - printed,
1209                                      "For system wide tracing it needs to be set to -1");
1210
1211                 printed += scnprintf(buf + printed, size - printed,
1212                                     ".\nHint:\tThe current value is %d.", value);
1213                 break;
1214         default:
1215                 scnprintf(buf, size, "%s", emsg);
1216                 break;
1217         }
1218
1219         return 0;
1220 }
1221
1222 void perf_evlist__to_front(struct perf_evlist *evlist,
1223                            struct perf_evsel *move_evsel)
1224 {
1225         struct perf_evsel *evsel, *n;
1226         LIST_HEAD(move);
1227
1228         if (move_evsel == perf_evlist__first(evlist))
1229                 return;
1230
1231         list_for_each_entry_safe(evsel, n, &evlist->entries, node) {
1232                 if (evsel->leader == move_evsel->leader)
1233                         list_move_tail(&evsel->node, &move);
1234         }
1235
1236         list_splice(&move, &evlist->entries);
1237 }