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 = 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 = (union perf_event *) md->event_copy;
 554                 }
 555
 556                 old += size;
 557         }
 558
 559         md->prev = old;
 560
 561         return event;
 562 }
 563
 564 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 565 {
 566         if (!evlist->overwrite) {
 567                 struct perf_mmap *md = &evlist->mmap[idx];
 568                 unsigned int old = md->prev;
 569
 570                 perf_mmap__write_tail(md, old);
 571         }
 572 }
 573
 574 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 575 {
 576         if (evlist->mmap[idx].base != NULL) {
 577                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 578                 evlist->mmap[idx].base = NULL;
 579         }
 580 }
 581
 582 void perf_evlist__munmap(struct perf_evlist *evlist)
 583 {
 584         int i;
 585
 586         for (i = 0; i < evlist->nr_mmaps; i++)
 587                 __perf_evlist__munmap(evlist, i);
 588
 589         free(evlist->mmap);
 590         evlist->mmap = NULL;
 591 }
 592
 593 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 594 {
 595         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 596         if (cpu_map__empty(evlist->cpus))
 597                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 598         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 599         return evlist->mmap != NULL ? 0 : -ENOMEM;
 600 }
 601
 602 static int __perf_evlist__mmap(struct perf_evlist *evlist,
 603                                int idx, int prot, int mask, int fd)
 604 {
 605         evlist->mmap[idx].prev = 0;
 606         evlist->mmap[idx].mask = mask;
 607         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, prot,
 608                                       MAP_SHARED, fd, 0);
 609         if (evlist->mmap[idx].base == MAP_FAILED) {
 610                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 611                           errno);
 612                 evlist->mmap[idx].base = NULL;
 613                 return -1;
 614         }
 615
 616         perf_evlist__add_pollfd(evlist, fd);
 617         return 0;
 618 }
 619
 620 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 621                                        int prot, int mask, int cpu, int thread,
 622                                        int *output)
 623 {
 624         struct perf_evsel *evsel;
 625
 626         list_for_each_entry(evsel, &evlist->entries, node) {
 627                 int fd = FD(evsel, cpu, thread);
 628
 629                 if (*output == -1) {
 630                         *output = fd;
 631                         if (__perf_evlist__mmap(evlist, idx, prot, mask,
 632                                                 *output) < 0)
 633                                 return -1;
 634                 } else {
 635                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 636                                 return -1;
 637                 }
 638
 639                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 640                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 641                         return -1;
 642         }
 643
 644         return 0;
 645 }
 646
 647 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist, int prot,
 648                                      int mask)
 649 {
 650         int cpu, thread;
 651         int nr_cpus = cpu_map__nr(evlist->cpus);
 652         int nr_threads = thread_map__nr(evlist->threads);
 653
 654         pr_debug2("perf event ring buffer mmapped per cpu\n");
 655         for (cpu = 0; cpu < nr_cpus; cpu++) {
 656                 int output = -1;
 657
 658                 for (thread = 0; thread < nr_threads; thread++) {
 659                         if (perf_evlist__mmap_per_evsel(evlist, cpu, prot, mask,
 660                                                         cpu, thread, &output))
 661                                 goto out_unmap;
 662                 }
 663         }
 664
 665         return 0;
 666
 667 out_unmap:
 668         for (cpu = 0; cpu < nr_cpus; cpu++)
 669                 __perf_evlist__munmap(evlist, cpu);
 670         return -1;
 671 }
 672
 673 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist, int prot,
 674                                         int mask)
 675 {
 676         int thread;
 677         int nr_threads = thread_map__nr(evlist->threads);
 678
 679         pr_debug2("perf event ring buffer mmapped per thread\n");
 680         for (thread = 0; thread < nr_threads; thread++) {
 681                 int output = -1;
 682
 683                 if (perf_evlist__mmap_per_evsel(evlist, thread, prot, mask, 0,
 684                                                 thread, &output))
 685                         goto out_unmap;
 686         }
 687
 688         return 0;
 689
 690 out_unmap:
 691         for (thread = 0; thread < nr_threads; thread++)
 692                 __perf_evlist__munmap(evlist, thread);
 693         return -1;
 694 }
 695
 696 static size_t perf_evlist__mmap_size(unsigned long pages)
 697 {
 698         /* 512 kiB: default amount of unprivileged mlocked memory */
 699         if (pages == UINT_MAX)
 700                 pages = (512 * 1024) / page_size;
 701         else if (!is_power_of_2(pages))
 702                 return 0;
 703
 704         return (pages + 1) * page_size;
 705 }
 706
 707 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 708                                   int unset __maybe_unused)
 709 {
 710         unsigned int *mmap_pages = opt->value;
 711         unsigned long pages, val;
 712         size_t size;
 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         val = parse_tag_value(str, tags);
 722         if (val != (unsigned long) -1) {
 723                 /* we got file size value */
 724                 pages = PERF_ALIGN(val, page_size) / page_size;
 725                 if (pages < (1UL << 31) && !is_power_of_2(pages)) {
 726                         pages = next_pow2(pages);
 727                         pr_info("rounding mmap pages size to %lu (%lu pages)\n",
 728                                 pages * page_size, pages);
 729                 }
 730         } else {
 731                 /* we got pages count value */
 732                 char *eptr;
 733                 pages = strtoul(str, &eptr, 10);
 734                 if (*eptr != '\0') {
 735                         pr_err("failed to parse --mmap_pages/-m value\n");
 736                         return -1;
 737                 }
 738         }
 739
 740         if (pages > UINT_MAX || pages > SIZE_MAX / page_size) {
 741                 pr_err("--mmap_pages/-m value too big\n");
 742                 return -1;
 743         }
 744
 745         size = perf_evlist__mmap_size(pages);
 746         if (!size) {
 747                 pr_err("--mmap_pages/-m value must be a power of two.");
 748                 return -1;
 749         }
 750
 751         *mmap_pages = pages;
 752         return 0;
 753 }
 754
 755 /**
 756  * perf_evlist__mmap - Create mmaps to receive events.
 757  * @evlist: list of events
 758  * @pages: map length in pages
 759  * @overwrite: overwrite older events?
 760  *
 761  * If @overwrite is %false the user needs to signal event consumption using
 762  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 763  * automatically.
 764  *
 765  * Return: %0 on success, negative error code otherwise.
 766  */
 767 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 768                       bool overwrite)
 769 {
 770         struct perf_evsel *evsel;
 771         const struct cpu_map *cpus = evlist->cpus;
 772         const struct thread_map *threads = evlist->threads;
 773         int prot = PROT_READ | (overwrite ? 0 : PROT_WRITE), mask;
 774
 775         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 776                 return -ENOMEM;
 777
 778         if (evlist->pollfd == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 779                 return -ENOMEM;
 780
 781         evlist->overwrite = overwrite;
 782         evlist->mmap_len = perf_evlist__mmap_size(pages);
 783         pr_debug("mmap size %zuB\n", evlist->mmap_len);
 784         mask = evlist->mmap_len - page_size - 1;
 785
 786         list_for_each_entry(evsel, &evlist->entries, node) {
 787                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 788                     evsel->sample_id == NULL &&
 789                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
 790                         return -ENOMEM;
 791         }
 792
 793         if (cpu_map__empty(cpus))
 794                 return perf_evlist__mmap_per_thread(evlist, prot, mask);
 795
 796         return perf_evlist__mmap_per_cpu(evlist, prot, mask);
 797 }
 798
 799 int perf_evlist__create_maps(struct perf_evlist *evlist,
 800                              struct perf_target *target)
 801 {
 802         evlist->threads = thread_map__new_str(target->pid, target->tid,
 803                                               target->uid);
 804
 805         if (evlist->threads == NULL)
 806                 return -1;
 807
 808         if (perf_target__has_task(target))
 809                 evlist->cpus = cpu_map__dummy_new();
 810         else if (!perf_target__has_cpu(target) && !target->uses_mmap)
 811                 evlist->cpus = cpu_map__dummy_new();
 812         else
 813                 evlist->cpus = cpu_map__new(target->cpu_list);
 814
 815         if (evlist->cpus == NULL)
 816                 goto out_delete_threads;
 817
 818         return 0;
 819
 820 out_delete_threads:
 821         thread_map__delete(evlist->threads);
 822         return -1;
 823 }
 824
 825 void perf_evlist__delete_maps(struct perf_evlist *evlist)
 826 {
 827         cpu_map__delete(evlist->cpus);
 828         thread_map__delete(evlist->threads);
 829         evlist->cpus    = NULL;
 830         evlist->threads = NULL;
 831 }
 832
 833 int perf_evlist__apply_filters(struct perf_evlist *evlist)
 834 {
 835         struct perf_evsel *evsel;
 836         int err = 0;
 837         const int ncpus = cpu_map__nr(evlist->cpus),
 838                   nthreads = thread_map__nr(evlist->threads);
 839
 840         list_for_each_entry(evsel, &evlist->entries, node) {
 841                 if (evsel->filter == NULL)
 842                         continue;
 843
 844                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
 845                 if (err)
 846                         break;
 847         }
 848
 849         return err;
 850 }
 851
 852 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 853 {
 854         struct perf_evsel *evsel;
 855         int err = 0;
 856         const int ncpus = cpu_map__nr(evlist->cpus),
 857                   nthreads = thread_map__nr(evlist->threads);
 858
 859         list_for_each_entry(evsel, &evlist->entries, node) {
 860                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
 861                 if (err)
 862                         break;
 863         }
 864
 865         return err;
 866 }
 867
 868 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 869 {
 870         struct perf_evsel *pos;
 871
 872         if (evlist->nr_entries == 1)
 873                 return true;
 874
 875         if (evlist->id_pos < 0 || evlist->is_pos < 0)
 876                 return false;
 877
 878         list_for_each_entry(pos, &evlist->entries, node) {
 879                 if (pos->id_pos != evlist->id_pos ||
 880                     pos->is_pos != evlist->is_pos)
 881                         return false;
 882         }
 883
 884         return true;
 885 }
 886
 887 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 888 {
 889         struct perf_evsel *evsel;
 890
 891         if (evlist->combined_sample_type)
 892                 return evlist->combined_sample_type;
 893
 894         list_for_each_entry(evsel, &evlist->entries, node)
 895                 evlist->combined_sample_type |= evsel->attr.sample_type;
 896
 897         return evlist->combined_sample_type;
 898 }
 899
 900 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 901 {
 902         evlist->combined_sample_type = 0;
 903         return __perf_evlist__combined_sample_type(evlist);
 904 }
 905
 906 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
 907 {
 908         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 909         u64 read_format = first->attr.read_format;
 910         u64 sample_type = first->attr.sample_type;
 911
 912         list_for_each_entry_continue(pos, &evlist->entries, node) {
 913                 if (read_format != pos->attr.read_format)
 914                         return false;
 915         }
 916
 917         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
 918         if ((sample_type & PERF_SAMPLE_READ) &&
 919             !(read_format & PERF_FORMAT_ID)) {
 920                 return false;
 921         }
 922
 923         return true;
 924 }
 925
 926 u64 perf_evlist__read_format(struct perf_evlist *evlist)
 927 {
 928         struct perf_evsel *first = perf_evlist__first(evlist);
 929         return first->attr.read_format;
 930 }
 931
 932 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
 933 {
 934         struct perf_evsel *first = perf_evlist__first(evlist);
 935         struct perf_sample *data;
 936         u64 sample_type;
 937         u16 size = 0;
 938
 939         if (!first->attr.sample_id_all)
 940                 goto out;
 941
 942         sample_type = first->attr.sample_type;
 943
 944         if (sample_type & PERF_SAMPLE_TID)
 945                 size += sizeof(data->tid) * 2;
 946
 947        if (sample_type & PERF_SAMPLE_TIME)
 948                 size += sizeof(data->time);
 949
 950         if (sample_type & PERF_SAMPLE_ID)
 951                 size += sizeof(data->id);
 952
 953         if (sample_type & PERF_SAMPLE_STREAM_ID)
 954                 size += sizeof(data->stream_id);
 955
 956         if (sample_type & PERF_SAMPLE_CPU)
 957                 size += sizeof(data->cpu) * 2;
 958
 959         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 960                 size += sizeof(data->id);
 961 out:
 962         return size;
 963 }
 964
 965 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
 966 {
 967         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
 968
 969         list_for_each_entry_continue(pos, &evlist->entries, node) {
 970                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
 971                         return false;
 972         }
 973
 974         return true;
 975 }
 976
 977 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
 978 {
 979         struct perf_evsel *first = perf_evlist__first(evlist);
 980         return first->attr.sample_id_all;
 981 }
 982
 983 void perf_evlist__set_selected(struct perf_evlist *evlist,
 984                                struct perf_evsel *evsel)
 985 {
 986         evlist->selected = evsel;
 987 }
 988
 989 void perf_evlist__close(struct perf_evlist *evlist)
 990 {
 991         struct perf_evsel *evsel;
 992         int ncpus = cpu_map__nr(evlist->cpus);
 993         int nthreads = thread_map__nr(evlist->threads);
 994
 995         list_for_each_entry_reverse(evsel, &evlist->entries, node)
 996                 perf_evsel__close(evsel, ncpus, nthreads);
 997 }
 998
 999 int perf_evlist__open(struct perf_evlist *evlist)
1000 {
1001         struct perf_evsel *evsel;
1002         int err;
1003
1004         perf_evlist__update_id_pos(evlist);
1005
1006         list_for_each_entry(evsel, &evlist->entries, node) {
1007                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1008                 if (err < 0)
1009                         goto out_err;
1010         }
1011
1012         return 0;
1013 out_err:
1014         perf_evlist__close(evlist);
1015         errno = -err;
1016         return err;
1017 }
1018
1019 int perf_evlist__prepare_workload(struct perf_evlist *evlist,
1020                                   struct perf_target *target,
1021                                   const char *argv[], bool pipe_output,
1022                                   bool want_signal)
1023 {
1024         int child_ready_pipe[2], go_pipe[2];
1025         char bf;
1026
1027         if (pipe(child_ready_pipe) < 0) {
1028                 perror("failed to create 'ready' pipe");
1029                 return -1;
1030         }
1031
1032         if (pipe(go_pipe) < 0) {
1033                 perror("failed to create 'go' pipe");
1034                 goto out_close_ready_pipe;
1035         }
1036
1037         evlist->workload.pid = fork();
1038         if (evlist->workload.pid < 0) {
1039                 perror("failed to fork");
1040                 goto out_close_pipes;
1041         }
1042
1043         if (!evlist->workload.pid) {
1044                 if (pipe_output)
1045                         dup2(2, 1);
1046
1047                 signal(SIGTERM, SIG_DFL);
1048
1049                 close(child_ready_pipe[0]);
1050                 close(go_pipe[1]);
1051                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1052
1053                 /*
1054                  * Tell the parent we're ready to go
1055                  */
1056                 close(child_ready_pipe[1]);
1057
1058                 /*
1059                  * Wait until the parent tells us to go.
1060                  */
1061                 if (read(go_pipe[0], &bf, 1) == -1)
1062                         perror("unable to read pipe");
1063
1064                 execvp(argv[0], (char **)argv);
1065
1066                 perror(argv[0]);
1067                 if (want_signal)
1068                         kill(getppid(), SIGUSR1);
1069                 exit(-1);
1070         }
1071
1072         if (perf_target__none(target))
1073                 evlist->threads->map[0] = evlist->workload.pid;
1074
1075         close(child_ready_pipe[1]);
1076         close(go_pipe[0]);
1077         /*
1078          * wait for child to settle
1079          */
1080         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1081                 perror("unable to read pipe");
1082                 goto out_close_pipes;
1083         }
1084
1085         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1086         evlist->workload.cork_fd = go_pipe[1];
1087         close(child_ready_pipe[0]);
1088         return 0;
1089
1090 out_close_pipes:
1091         close(go_pipe[0]);
1092         close(go_pipe[1]);
1093 out_close_ready_pipe:
1094         close(child_ready_pipe[0]);
1095         close(child_ready_pipe[1]);
1096         return -1;
1097 }
1098
1099 int perf_evlist__start_workload(struct perf_evlist *evlist)
1100 {
1101         if (evlist->workload.cork_fd > 0) {
1102                 char bf = 0;
1103                 int ret;
1104                 /*
1105                  * Remove the cork, let it rip!
1106                  */
1107                 ret = write(evlist->workload.cork_fd, &bf, 1);
1108                 if (ret < 0)
1109                         perror("enable to write to pipe");
1110
1111                 close(evlist->workload.cork_fd);
1112                 return ret;
1113         }
1114
1115         return 0;
1116 }
1117
1118 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1119                               struct perf_sample *sample)
1120 {
1121         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1122
1123         if (!evsel)
1124                 return -EFAULT;
1125         return perf_evsel__parse_sample(evsel, event, sample);
1126 }
1127
1128 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1129 {
1130         struct perf_evsel *evsel;
1131         size_t printed = 0;
1132
1133         list_for_each_entry(evsel, &evlist->entries, node) {
1134                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1135                                    perf_evsel__name(evsel));
1136         }
1137
1138         return printed + fprintf(fp, "\n");;
1139 }
1140
1141 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1142                              int err, char *buf, size_t size)
1143 {
1144         char sbuf[128];
1145
1146         switch (err) {
1147         case ENOENT:
1148                 scnprintf(buf, size, "%s",
1149                           "Error:\tUnable to find debugfs\n"
1150                           "Hint:\tWas your kernel was compiled with debugfs support?\n"
1151                           "Hint:\tIs the debugfs filesystem mounted?\n"
1152                           "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1153                 break;
1154         case EACCES:
1155                 scnprintf(buf, size,
1156                           "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1157                           "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1158                           debugfs_mountpoint, debugfs_mountpoint);
1159                 break;
1160         default:
1161                 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1162                 break;
1163         }
1164
1165         return 0;
1166 }
1167
1168 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1169                                int err, char *buf, size_t size)
1170 {
1171         int printed, value;
1172         char sbuf[128], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1173
1174         switch (err) {
1175         case EACCES:
1176         case EPERM:
1177                 printed = scnprintf(buf, size,
1178                                     "Error:\t%s.\n"
1179                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1180
1181                 if (filename__read_int("/proc/sys/kernel/perf_event_paranoid", &value))
1182                         break;
1183
1184                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1185
1186                 if (value >= 2) {
1187                         printed += scnprintf(buf + printed, size - printed,
1188                                              "For your workloads it needs to be <= 1\nHint:\t");
1189                 }
1190                 printed += scnprintf(buf + printed, size - printed,
1191                                      "For system wide tracing it needs to be set to -1");
1192
1193                 printed += scnprintf(buf + printed, size - printed,
1194                                     ".\nHint:\tThe current value is %d.", value);
1195                 break;
1196         default:
1197                 scnprintf(buf, size, "%s", emsg);
1198                 break;
1199         }
1200
1201         return 0;
1202 }