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 <lk/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         free(evlist->mmap);
 105         free(evlist->pollfd);
 106         evlist->mmap = NULL;
 107         evlist->pollfd = NULL;
 108 }
 109
 110 void perf_evlist__delete(struct perf_evlist *evlist)
 111 {
 112         perf_evlist__purge(evlist);
 113         perf_evlist__exit(evlist);
 114         free(evlist);
 115 }
 116
 117 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 118 {
 119         list_add_tail(&entry->node, &evlist->entries);
 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, 0);
 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(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;
 253
 254         evsel = perf_evsel__newtp(sys, name, evlist->nr_entries);
 255         if (evsel == NULL)
 256                 return -1;
 257
 258         evsel->handler.func = handler;
 259         perf_evlist__add(evlist, evsel);
 260         return 0;
 261 }
 262
 263 void perf_evlist__disable(struct perf_evlist *evlist)
 264 {
 265         int cpu, thread;
 266         struct perf_evsel *pos;
 267         int nr_cpus = cpu_map__nr(evlist->cpus);
 268         int nr_threads = thread_map__nr(evlist->threads);
 269
 270         for (cpu = 0; cpu < nr_cpus; cpu++) {
 271                 list_for_each_entry(pos, &evlist->entries, node) {
 272                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 273                                 continue;
 274                         for (thread = 0; thread < nr_threads; thread++)
 275                                 ioctl(FD(pos, cpu, thread),
 276                                       PERF_EVENT_IOC_DISABLE, 0);
 277                 }
 278         }
 279 }
 280
 281 void perf_evlist__enable(struct perf_evlist *evlist)
 282 {
 283         int cpu, thread;
 284         struct perf_evsel *pos;
 285         int nr_cpus = cpu_map__nr(evlist->cpus);
 286         int nr_threads = thread_map__nr(evlist->threads);
 287
 288         for (cpu = 0; cpu < nr_cpus; cpu++) {
 289                 list_for_each_entry(pos, &evlist->entries, node) {
 290                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 291                                 continue;
 292                         for (thread = 0; thread < nr_threads; thread++)
 293                                 ioctl(FD(pos, cpu, thread),
 294                                       PERF_EVENT_IOC_ENABLE, 0);
 295                 }
 296         }
 297 }
 298
 299 int perf_evlist__disable_event(struct perf_evlist *evlist,
 300                                struct perf_evsel *evsel)
 301 {
 302         int cpu, thread, err;
 303
 304         if (!evsel->fd)
 305                 return 0;
 306
 307         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 308                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 309                         err = ioctl(FD(evsel, cpu, thread),
 310                                     PERF_EVENT_IOC_DISABLE, 0);
 311                         if (err)
 312                                 return err;
 313                 }
 314         }
 315         return 0;
 316 }
 317
 318 int perf_evlist__enable_event(struct perf_evlist *evlist,
 319                               struct perf_evsel *evsel)
 320 {
 321         int cpu, thread, err;
 322
 323         if (!evsel->fd)
 324                 return -EINVAL;
 325
 326         for (cpu = 0; cpu < evlist->cpus->nr; cpu++) {
 327                 for (thread = 0; thread < evlist->threads->nr; thread++) {
 328                         err = ioctl(FD(evsel, cpu, thread),
 329                                     PERF_EVENT_IOC_ENABLE, 0);
 330                         if (err)
 331                                 return err;
 332                 }
 333         }
 334         return 0;
 335 }
 336
 337 static int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 338 {
 339         int nr_cpus = cpu_map__nr(evlist->cpus);
 340         int nr_threads = thread_map__nr(evlist->threads);
 341         int nfds = nr_cpus * nr_threads * evlist->nr_entries;
 342         evlist->pollfd = malloc(sizeof(struct pollfd) * nfds);
 343         return evlist->pollfd != NULL ? 0 : -ENOMEM;
 344 }
 345
 346 void perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 347 {
 348         fcntl(fd, F_SETFL, O_NONBLOCK);
 349         evlist->pollfd[evlist->nr_fds].fd = fd;
 350         evlist->pollfd[evlist->nr_fds].events = POLLIN;
 351         evlist->nr_fds++;
 352 }
 353
 354 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 355                                  struct perf_evsel *evsel,
 356                                  int cpu, int thread, u64 id)
 357 {
 358         int hash;
 359         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 360
 361         sid->id = id;
 362         sid->evsel = evsel;
 363         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 364         hlist_add_head(&sid->node, &evlist->heads[hash]);
 365 }
 366
 367 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 368                          int cpu, int thread, u64 id)
 369 {
 370         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 371         evsel->id[evsel->ids++] = id;
 372 }
 373
 374 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 375                                   struct perf_evsel *evsel,
 376                                   int cpu, int thread, int fd)
 377 {
 378         u64 read_data[4] = { 0, };
 379         int id_idx = 1; /* The first entry is the counter value */
 380         u64 id;
 381         int ret;
 382
 383         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 384         if (!ret)
 385                 goto add;
 386
 387         if (errno != ENOTTY)
 388                 return -1;
 389
 390         /* Legacy way to get event id.. All hail to old kernels! */
 391
 392         /*
 393          * This way does not work with group format read, so bail
 394          * out in that case.
 395          */
 396         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 397                 return -1;
 398
 399         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 400             read(fd, &read_data, sizeof(read_data)) == -1)
 401                 return -1;
 402
 403         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 404                 ++id_idx;
 405         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 406                 ++id_idx;
 407
 408         id = read_data[id_idx];
 409
 410  add:
 411         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 412         return 0;
 413 }
 414
 415 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 416 {
 417         struct hlist_head *head;
 418         struct perf_sample_id *sid;
 419         int hash;
 420
 421         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 422         head = &evlist->heads[hash];
 423
 424         hlist_for_each_entry(sid, head, node)
 425                 if (sid->id == id)
 426                         return sid;
 427
 428         return NULL;
 429 }
 430
 431 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 432 {
 433         struct perf_sample_id *sid;
 434
 435         if (evlist->nr_entries == 1)
 436                 return perf_evlist__first(evlist);
 437
 438         sid = perf_evlist__id2sid(evlist, id);
 439         if (sid)
 440                 return sid->evsel;
 441
 442         if (!perf_evlist__sample_id_all(evlist))
 443                 return perf_evlist__first(evlist);
 444
 445         return NULL;
 446 }
 447
 448 static int perf_evlist__event2id(struct perf_evlist *evlist,
 449                                  union perf_event *event, u64 *id)
 450 {
 451         const u64 *array = event->sample.array;
 452         ssize_t n;
 453
 454         n = (event->header.size - sizeof(event->header)) >> 3;
 455
 456         if (event->header.type == PERF_RECORD_SAMPLE) {
 457                 if (evlist->id_pos >= n)
 458                         return -1;
 459                 *id = array[evlist->id_pos];
 460         } else {
 461                 if (evlist->is_pos > n)
 462                         return -1;
 463                 n -= evlist->is_pos;
 464                 *id = array[n];
 465         }
 466         return 0;
 467 }
 468
 469 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 470                                                    union perf_event *event)
 471 {
 472         struct perf_evsel *first = perf_evlist__first(evlist);
 473         struct hlist_head *head;
 474         struct perf_sample_id *sid;
 475         int hash;
 476         u64 id;
 477
 478         if (evlist->nr_entries == 1)
 479                 return first;
 480
 481         if (!first->attr.sample_id_all &&
 482             event->header.type != PERF_RECORD_SAMPLE)
 483                 return first;
 484
 485         if (perf_evlist__event2id(evlist, event, &id))
 486                 return NULL;
 487
 488         /* Synthesized events have an id of zero */
 489         if (!id)
 490                 return first;
 491
 492         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 493         head = &evlist->heads[hash];
 494
 495         hlist_for_each_entry(sid, head, node) {
 496                 if (sid->id == id)
 497                         return sid->evsel;
 498         }
 499         return NULL;
 500 }
 501
 502 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 503 {
 504         struct perf_mmap *md = &evlist->mmap[idx];
 505         unsigned int head = perf_mmap__read_head(md);
 506         unsigned int old = md->prev;
 507         unsigned char *data = md->base + page_size;
 508         union perf_event *event = NULL;
 509
 510         if (evlist->overwrite) {
 511                 /*
 512                  * If we're further behind than half the buffer, there's a chance
 513                  * the writer will bite our tail and mess up the samples under us.
 514                  *
 515                  * If we somehow ended up ahead of the head, we got messed up.
 516                  *
 517                  * In either case, truncate and restart at head.
 518                  */
 519                 int diff = head - old;
 520                 if (diff > md->mask / 2 || diff < 0) {
 521                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 522
 523                         /*
 524                          * head points to a known good entry, start there.
 525                          */
 526                         old = head;
 527                 }
 528         }
 529
 530         if (old != head) {
 531                 size_t size;
 532
 533                 event = (union perf_event *)&data[old & md->mask];
 534                 size = event->header.size;
 535
 536                 /*
 537                  * Event straddles the mmap boundary -- header should always
 538                  * be inside due to u64 alignment of output.
 539                  */
 540                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 541                         unsigned int offset = old;
 542                         unsigned int len = min(sizeof(*event), size), cpy;
 543                         void *dst = &md->event_copy;
 544
 545                         do {
 546                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 547                                 memcpy(dst, &data[offset & md->mask], cpy);
 548                                 offset += cpy;
 549                                 dst += cpy;
 550                                 len -= cpy;
 551                         } while (len);
 552
 553                         event = &md->event_copy;
 554                 }
 555
 556                 old += size;
 557         }
 558
 559         md->prev = old;
 560
 561         if (!evlist->overwrite)
 562                 perf_mmap__write_tail(md, old);
 563
 564         return event;
 565 }
 566
 567 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 568 {
 569         if (evlist->mmap[idx].base != NULL) {
 570                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 571                 evlist->mmap[idx].base = NULL;
 572         }
 573 }
 574
 575 void perf_evlist__munmap(struct perf_evlist *evlist)
 576 {
 577         int i;
 578
 579         for (i = 0; i < evlist->nr_mmaps; i++)
 580                 __perf_evlist__munmap(evlist, i);
 581
 582         free(evlist->mmap);
 583         evlist->mmap = NULL;
 584 }
 585
 586 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 587 {
 588         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 589         if (cpu_map__empty(evlist->cpus))
 590                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 591         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 592         return evlist->mmap != NULL ? 0 : -ENOMEM;
 593 }
 594
 595 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 596                                int idx, int prot, int mask, int fd)
 597 {
 598         evlist->mmap[idx].prev = 0;
 599         evlist->mmap[idx].mask = mask;
 600         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 601                                       MAP_SHARED, fd, 0);
 602         if (evlist->mmap[idx].base == MAP_FAILED) {
 603                 evlist->mmap[idx].base = NULL;
 604                 return -1;
 605         }
 606
 607         perf_evlist__add_pollfd(evlist, fd);
 608         return 0;
 609 }
 610
 611 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot, int mask)
 612 {
 613         struct perf_evsel *evsel;
 614         int cpu, thread;
 615         int nr_cpus = cpu_map__nr(evlist->cpus);
 616         int nr_threads = thread_map__nr(evlist->threads);
 617
 618         pr_debug2("perf event ring buffer mmapped per cpu\n");
 619         for (cpu = 0; cpu < nr_cpus; cpu++) {
 620                 int output = -1;
 621
 622                 for (thread = 0; thread < nr_threads; thread++) {
 623                         list_for_each_entry(evsel, &evlist->entries, node) {
 624                                 int fd = FD(evsel, cpu, thread);
 625
 626                                 if (output == -1) {
 627                                         output = fd;
 628                                         if (__perf_evlist__mmap(evlist, cpu,
 629                                                                 prot, mask, output) < 0)
 630                                                 goto out_unmap;
 631                                 } else {
 632                                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 633                                                 goto out_unmap;
 634                                 }
 635
 636                                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 637                                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 638                                         goto out_unmap;
 639                         }
 640                 }
 641         }
 642
 643         return 0;
 644
 645 out_unmap:
 646         for (cpu = 0; cpu < nr_cpus; cpu++)
 647                 __perf_evlist__munmap(evlist, cpu);
 648         return -1;
 649 }
 650
 651 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot, int mask)
 652 {
 653         struct perf_evsel *evsel;
 654         int thread;
 655         int nr_threads = thread_map__nr(evlist->threads);
 656
 657         pr_debug2("perf event ring buffer mmapped per thread\n");
 658         for (thread = 0; thread < nr_threads; thread++) {
 659                 int output = -1;
 660
 661                 list_for_each_entry(evsel, &evlist->entries, node) {
 662                         int fd = FD(evsel, 0, thread);
 663
 664                         if (output == -1) {
 665                                 output = fd;
 666                                 if (__perf_evlist__mmap(evlist, thread,
 667                                                         prot, mask, output) < 0)
 668                                         goto out_unmap;
 669                         } else {
 670                                 if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, output) != 0)
 671                                         goto out_unmap;
 672                         }
 673
 674                         if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 675                             perf_evlist__id_add_fd(evlist, evsel, 0, thread, fd) < 0)
 676                                 goto out_unmap;
 677                 }
 678         }
 679
 680         return 0;
 681
 682 out_unmap:
 683         for (thread = 0; thread < nr_threads; thread++)
 684                 __perf_evlist__munmap(evlist, thread);
 685         return -1;
 686 }
 687
 688 static size_t perf_evlist__mmap_size(unsigned long pages)
 689 {
 690         /* 512 kiB: default amount of unprivileged mlocked memory */
 691         if (pages == UINT_MAX)
 692                 pages = (512 * 1024) / page_size;
 693         else if (!is_power_of_2(pages))
 694                 return 0;
 695
 696         return (pages + 1) * page_size;
 697 }
 698
 699 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 700                                   int unset __maybe_unused)
 701 {
 702         unsigned int pages, val, *mmap_pages = opt->value;
 703         size_t size;
 704         static struct parse_tag tags[] = {
 705                 { .tag  = 'B', .mult = 1       },
 706                 { .tag  = 'K', .mult = 1 << 10 },
 707                 { .tag  = 'M', .mult = 1 << 20 },
 708                 { .tag  = 'G', .mult = 1 << 30 },
 709                 { .tag  = 0 },
 710         };
 711
 712         val = parse_tag_value(str, tags);
 713         if (val != (unsigned int) -1) {
 714                 /* we got file size value */
 715                 pages = PERF_ALIGN(val, page_size) / page_size;
 716                 if (!is_power_of_2(pages)) {
 717                         pages = next_pow2(pages);
 718                         pr_info("rounding mmap pages size to %u (%u pages)\n",
 719                                 pages * page_size, pages);
 720                 }
 721         } else {
 722                 /* we got pages count value */
 723                 char *eptr;
 724                 pages = strtoul(str, &eptr, 10);
 725                 if (*eptr != '\0') {
 726                         pr_err("failed to parse --mmap_pages/-m value\n");
 727                         return -1;
 728                 }
 729         }
 730
 731         size = perf_evlist__mmap_size(pages);
 732         if (!size) {
 733                 pr_err("--mmap_pages/-m value must be a power of two.");
 734                 return -1;
 735         }
 736
 737         *mmap_pages = pages;
 738         return 0;
 739 }
 740
 741 /** perf_evlist__mmap - Create per cpu maps to receive events
 742  *
 743  * @evlist - list of events
 744  * @pages - map length in pages
 745  * @overwrite - overwrite older events?
 746  *
 747  * If overwrite is false the user needs to signal event consuption using:
 748  *
 749  *      struct perf_mmap *m = &evlist->mmap[cpu];
 750  *      unsigned int head = perf_mmap__read_head(m);
 751  *
 752  *      perf_mmap__write_tail(m, head)
 753  *
 754  * Using perf_evlist__read_on_cpu does this automatically.
 755  */
 756 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 757                       bool overwrite)
 758 {
 759         struct perf_evsel *evsel;
 760         const struct cpu_map *cpus = evlist->cpus;
 761         const struct thread_map *threads = evlist->threads;
 762         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
 763
 764         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 765                 return -ENOMEM;
 766
 767         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 768                 return -ENOMEM;
 769
 770         evlist->overwrite = overwrite;
 771         evlist->mmap_len = perf_evlist__mmap_size(pages);
 772         pr_debug("mmap size %luB\n", evlist->mmap_len);
 773         mask = evlist->mmap_len - page_size - 1;
 774
 775         list_for_each_entry(evsel, &evlist->entries, node) {
 776                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 777                     evsel->sample_id == NULL &&
 778                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
 779                         return -ENOMEM;
 780         }
 781
 782         if (cpu_map__empty(cpus))
 783                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 784
 785         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 786 }
 787
 788 int perf_evlist__create_maps(struct perf_evlist *evlist,
 789                              struct perf_target *target)
 790 {
 791         evlist->threads = thread_map__new_str(target->pid, target->tid,
 792                                               target->uid);
 793
 794         if (evlist->threads == NULL)
 795                 return -1;
 796
 797         if (perf_target__has_task(target))
 798                 evlist->cpus = cpu_map__dummy_new();
 799         else if (!perf_target__has_cpu(target) && !target->uses_mmap)
 800                 evlist->cpus = cpu_map__dummy_new();
 801         else
 802                 evlist->cpus = cpu_map__new(target->cpu_list);
 803
 804         if (evlist->cpus == NULL)
 805                 goto out_delete_threads;
 806
 807         return 0;
 808
 809 out_delete_threads:
 810         thread_map__delete(evlist->threads);
 811         return -1;
 812 }
 813
 814 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 815 {
 816         cpu_map__delete(evlist->cpus);
 817         thread_map__delete(evlist->threads);
 818         evlist->cpus    = NULL;
 819         evlist->threads = NULL;
 820 }
 821
 822 int perf_evlist__apply_filters(struct perf_evlist *evlist)
 823 {
 824         struct perf_evsel *evsel;
 825         int err = 0;
 826         const int ncpus = cpu_map__nr(evlist->cpus),
 827                   nthreads = thread_map__nr(evlist->threads);
 828
 829         list_for_each_entry(evsel, &evlist->entries, node) {
 830                 if (evsel->filter == NULL)
 831                         continue;
 832
 833                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
 834                 if (err)
 835                         break;
 836         }
 837
 838         return err;
 839 }
 840
 841 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 842 {
 843         struct perf_evsel *evsel;
 844         int err = 0;
 845         const int ncpus = cpu_map__nr(evlist->cpus),
 846                   nthreads = thread_map__nr(evlist->threads);
 847
 848         list_for_each_entry(evsel, &evlist->entries, node) {
 849                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
 850                 if (err)
 851                         break;
 852         }
 853
 854         return err;
 855 }
 856
 857 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 858 {
 859         struct perf_evsel *pos;
 860
 861         if (evlist->nr_entries == 1)
 862                 return true;
 863
 864         if (evlist->id_pos < 0 || evlist->is_pos < 0)
 865                 return false;
 866
 867         list_for_each_entry(pos, &evlist->entries, node) {
 868                 if (pos->id_pos != evlist->id_pos ||
 869                     pos->is_pos != evlist->is_pos)
 870                         return false;
 871         }
 872
 873         return true;
 874 }
 875
 876 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 877 {
 878         struct perf_evsel *evsel;
 879
 880         if (evlist->combined_sample_type)
 881                 return evlist->combined_sample_type;
 882
 883         list_for_each_entry(evsel, &evlist->entries, node)
 884                 evlist->combined_sample_type |= evsel->attr.sample_type;
 885
 886         return evlist->combined_sample_type;
 887 }
 888
 889 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 890 {
 891         evlist->combined_sample_type = 0;
 892         return __perf_evlist__combined_sample_type(evlist);
 893 }
 894
 895 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
 896 {
 897         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 898         u64 read_format = first->attr.read_format;
 899         u64 sample_type = first->attr.sample_type;
 900
 901         list_for_each_entry_continue(pos, &evlist->entries, node) {
 902                 if (read_format != pos->attr.read_format)
 903                         return false;
 904         }
 905
 906         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
 907         if ((sample_type & PERF_SAMPLE_READ) &&
 908             !(read_format & PERF_FORMAT_ID)) {
 909                 return false;
 910         }
 911
 912         return true;
 913 }
 914
 915 u64 perf_evlist__read_format(struct perf_evlist *evlist)
 916 {
 917         struct perf_evsel *first = perf_evlist__first(evlist);
 918         return first->attr.read_format;
 919 }
 920
 921 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
 922 {
 923         struct perf_evsel *first = perf_evlist__first(evlist);
 924         struct perf_sample *data;
 925         u64 sample_type;
 926         u16 size = 0;
 927
 928         if (!first->attr.sample_id_all)
 929                 goto out;
 930
 931         sample_type = first->attr.sample_type;
 932
 933         if (sample_type & PERF_SAMPLE_TID)
 934                 size += sizeof(data->tid) * 2;
 935
 936        if (sample_type & PERF_SAMPLE_TIME)
 937                 size += sizeof(data->time);
 938
 939         if (sample_type & PERF_SAMPLE_ID)
 940                 size += sizeof(data->id);
 941
 942         if (sample_type & PERF_SAMPLE_STREAM_ID)
 943                 size += sizeof(data->stream_id);
 944
 945         if (sample_type & PERF_SAMPLE_CPU)
 946                 size += sizeof(data->cpu) * 2;
 947
 948         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 949                 size += sizeof(data->id);
 950 out:
 951         return size;
 952 }
 953
 954 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
 955 {
 956         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 957
 958         list_for_each_entry_continue(pos, &evlist->entries, node) {
 959                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
 960                         return false;
 961         }
 962
 963         return true;
 964 }
 965
 966 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
 967 {
 968         struct perf_evsel *first = perf_evlist__first(evlist);
 969         return first->attr.sample_id_all;
 970 }
 971
 972 void perf_evlist__set_selected(struct perf_evlist *evlist,
 973                                struct perf_evsel *evsel)
 974 {
 975         evlist->selected = evsel;
 976 }
 977
 978 void perf_evlist__close(struct perf_evlist *evlist)
 979 {
 980         struct perf_evsel *evsel;
 981         int ncpus = cpu_map__nr(evlist->cpus);
 982         int nthreads = thread_map__nr(evlist->threads);
 983
 984         list_for_each_entry_reverse(evsel, &evlist->entries, node)
 985                 perf_evsel__close(evsel, ncpus, nthreads);
 986 }
 987
 988 int perf_evlist__open(struct perf_evlist *evlist)
 989 {
 990         struct perf_evsel *evsel;
 991         int err;
 992
 993         perf_evlist__update_id_pos(evlist);
 994
 995         list_for_each_entry(evsel, &evlist->entries, node) {
 996                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
 997                 if (err < 0)
 998                         goto out_err;
 999         }
1000
1001         return 0;
1002 out_err:
1003         perf_evlist__close(evlist);
1004         errno = -err;
1005         return err;
1006 }
1007
1008 int perf_evlist__prepare_workload(struct perf_evlist *evlist,
1009                                   struct perf_target *target,
1010                                   const char *argv[], bool pipe_output,
1011                                   bool want_signal)
1012 {
1013         int child_ready_pipe[2], go_pipe[2];
1014         char bf;
1015
1016         if (pipe(child_ready_pipe) < 0) {
1017                 perror("failed to create 'ready' pipe");
1018                 return -1;
1019         }
1020
1021         if (pipe(go_pipe) < 0) {
1022                 perror("failed to create 'go' pipe");
1023                 goto out_close_ready_pipe;
1024         }
1025
1026         evlist->workload.pid = fork();
1027         if (evlist->workload.pid < 0) {
1028                 perror("failed to fork");
1029                 goto out_close_pipes;
1030         }
1031
1032         if (!evlist->workload.pid) {
1033                 if (pipe_output)
1034                         dup2(2, 1);
1035
1036                 signal(SIGTERM, SIG_DFL);
1037
1038                 close(child_ready_pipe[0]);
1039                 close(go_pipe[1]);
1040                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1041
1042                 /*
1043                  * Tell the parent we're ready to go
1044                  */
1045                 close(child_ready_pipe[1]);
1046
1047                 /*
1048                  * Wait until the parent tells us to go.
1049                  */
1050                 if (read(go_pipe[0], &bf, 1) == -1)
1051                         perror("unable to read pipe");
1052
1053                 execvp(argv[0], (char **)argv);
1054
1055                 perror(argv[0]);
1056                 if (want_signal)
1057                         kill(getppid(), SIGUSR1);
1058                 exit(-1);
1059         }
1060
1061         if (perf_target__none(target))
1062                 evlist->threads->map[0] = evlist->workload.pid;
1063
1064         close(child_ready_pipe[1]);
1065         close(go_pipe[0]);
1066         /*
1067          * wait for child to settle
1068          */
1069         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1070                 perror("unable to read pipe");
1071                 goto out_close_pipes;
1072         }
1073
1074         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1075         evlist->workload.cork_fd = go_pipe[1];
1076         close(child_ready_pipe[0]);
1077         return 0;
1078
1079 out_close_pipes:
1080         close(go_pipe[0]);
1081         close(go_pipe[1]);
1082 out_close_ready_pipe:
1083         close(child_ready_pipe[0]);
1084         close(child_ready_pipe[1]);
1085         return -1;
1086 }
1087
1088 int perf_evlist__start_workload(struct perf_evlist *evlist)
1089 {
1090         if (evlist->workload.cork_fd > 0) {
1091                 char bf = 0;
1092                 int ret;
1093                 /*
1094                  * Remove the cork, let it rip!
1095                  */
1096                 ret = write(evlist->workload.cork_fd, &bf, 1);
1097                 if (ret < 0)
1098                         perror("enable to write to pipe");
1099
1100                 close(evlist->workload.cork_fd);
1101                 return ret;
1102         }
1103
1104         return 0;
1105 }
1106
1107 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1108                               struct perf_sample *sample)
1109 {
1110         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1111
1112         if (!evsel)
1113                 return -EFAULT;
1114         return perf_evsel__parse_sample(evsel, event, sample);
1115 }
1116
1117 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1118 {
1119         struct perf_evsel *evsel;
1120         size_t printed = 0;
1121
1122         list_for_each_entry(evsel, &evlist->entries, node) {
1123                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1124                                    perf_evsel__name(evsel));
1125         }
1126
1127         return printed + fprintf(fp, "\n");;
1128 }