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