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         fdarray__init(&evlist->pollfd, 64);
  41         evlist->workload.pid = -1;
  42 }
  43
  44 struct perf_evlist *perf_evlist__new(void)
  45 {
  46         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  47
  48         if (evlist != NULL)
  49                 perf_evlist__init(evlist, NULL, NULL);
  50
  51         return evlist;
  52 }
  53
  54 struct perf_evlist *perf_evlist__new_default(void)
  55 {
  56         struct perf_evlist *evlist = perf_evlist__new();
  57
  58         if (evlist && perf_evlist__add_default(evlist)) {
  59                 perf_evlist__delete(evlist);
  60                 evlist = NULL;
  61         }
  62
  63         return evlist;
  64 }
  65
  66 /**
  67  * perf_evlist__set_id_pos - set the positions of event ids.
  68  * @evlist: selected event list
  69  *
  70  * Events with compatible sample types all have the same id_pos
  71  * and is_pos.  For convenience, put a copy on evlist.
  72  */
  73 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  74 {
  75         struct perf_evsel *first = perf_evlist__first(evlist);
  76
  77         evlist->id_pos = first->id_pos;
  78         evlist->is_pos = first->is_pos;
  79 }
  80
  81 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  82 {
  83         struct perf_evsel *evsel;
  84
  85         evlist__for_each(evlist, evsel)
  86                 perf_evsel__calc_id_pos(evsel);
  87
  88         perf_evlist__set_id_pos(evlist);
  89 }
  90
  91 static void perf_evlist__purge(struct perf_evlist *evlist)
  92 {
  93         struct perf_evsel *pos, *n;
  94
  95         evlist__for_each_safe(evlist, n, pos) {
  96                 list_del_init(&pos->node);
  97                 perf_evsel__delete(pos);
  98         }
  99
 100         evlist->nr_entries = 0;
 101 }
 102
 103 void perf_evlist__exit(struct perf_evlist *evlist)
 104 {
 105         zfree(&evlist->mmap);
 106         fdarray__exit(&evlist->pollfd);
 107 }
 108
 109 void perf_evlist__delete(struct perf_evlist *evlist)
 110 {
 111         perf_evlist__munmap(evlist);
 112         perf_evlist__close(evlist);
 113         cpu_map__delete(evlist->cpus);
 114         thread_map__delete(evlist->threads);
 115         evlist->cpus = NULL;
 116         evlist->threads = NULL;
 117         perf_evlist__purge(evlist);
 118         perf_evlist__exit(evlist);
 119         free(evlist);
 120 }
 121
 122 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 123 {
 124         list_add_tail(&entry->node, &evlist->entries);
 125         entry->idx = evlist->nr_entries;
 126         entry->tracking = !entry->idx;
 127
 128         if (!evlist->nr_entries++)
 129                 perf_evlist__set_id_pos(evlist);
 130 }
 131
 132 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 133                                    struct list_head *list,
 134                                    int nr_entries)
 135 {
 136         bool set_id_pos = !evlist->nr_entries;
 137
 138         list_splice_tail(list, &evlist->entries);
 139         evlist->nr_entries += nr_entries;
 140         if (set_id_pos)
 141                 perf_evlist__set_id_pos(evlist);
 142 }
 143
 144 void __perf_evlist__set_leader(struct list_head *list)
 145 {
 146         struct perf_evsel *evsel, *leader;
 147
 148         leader = list_entry(list->next, struct perf_evsel, node);
 149         evsel = list_entry(list->prev, struct perf_evsel, node);
 150
 151         leader->nr_members = evsel->idx - leader->idx + 1;
 152
 153         __evlist__for_each(list, evsel) {
 154                 evsel->leader = leader;
 155         }
 156 }
 157
 158 void perf_evlist__set_leader(struct perf_evlist *evlist)
 159 {
 160         if (evlist->nr_entries) {
 161                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 162                 __perf_evlist__set_leader(&evlist->entries);
 163         }
 164 }
 165
 166 int perf_evlist__add_default(struct perf_evlist *evlist)
 167 {
 168         struct perf_event_attr attr = {
 169                 .type = PERF_TYPE_HARDWARE,
 170                 .config = PERF_COUNT_HW_CPU_CYCLES,
 171         };
 172         struct perf_evsel *evsel;
 173
 174         event_attr_init(&attr);
 175
 176         evsel = perf_evsel__new(&attr);
 177         if (evsel == NULL)
 178                 goto error;
 179
 180         /* use strdup() because free(evsel) assumes name is allocated */
 181         evsel->name = strdup("cycles");
 182         if (!evsel->name)
 183                 goto error_free;
 184
 185         perf_evlist__add(evlist, evsel);
 186         return 0;
 187 error_free:
 188         perf_evsel__delete(evsel);
 189 error:
 190         return -ENOMEM;
 191 }
 192
 193 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 194                                   struct perf_event_attr *attrs, size_t nr_attrs)
 195 {
 196         struct perf_evsel *evsel, *n;
 197         LIST_HEAD(head);
 198         size_t i;
 199
 200         for (i = 0; i < nr_attrs; i++) {
 201                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 202                 if (evsel == NULL)
 203                         goto out_delete_partial_list;
 204                 list_add_tail(&evsel->node, &head);
 205         }
 206
 207         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 208
 209         return 0;
 210
 211 out_delete_partial_list:
 212         __evlist__for_each_safe(&head, n, evsel)
 213                 perf_evsel__delete(evsel);
 214         return -1;
 215 }
 216
 217 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 218                                      struct perf_event_attr *attrs, size_t nr_attrs)
 219 {
 220         size_t i;
 221
 222         for (i = 0; i < nr_attrs; i++)
 223                 event_attr_init(attrs + i);
 224
 225         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 226 }
 227
 228 struct perf_evsel *
 229 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 230 {
 231         struct perf_evsel *evsel;
 232
 233         evlist__for_each(evlist, evsel) {
 234                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 235                     (int)evsel->attr.config == id)
 236                         return evsel;
 237         }
 238
 239         return NULL;
 240 }
 241
 242 struct perf_evsel *
 243 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 244                                      const char *name)
 245 {
 246         struct perf_evsel *evsel;
 247
 248         evlist__for_each(evlist, evsel) {
 249                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 250                     (strcmp(evsel->name, name) == 0))
 251                         return evsel;
 252         }
 253
 254         return NULL;
 255 }
 256
 257 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 258                            const char *sys, const char *name, void *handler)
 259 {
 260         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 261
 262         if (evsel == NULL)
 263                 return -1;
 264
 265         evsel->handler = handler;
 266         perf_evlist__add(evlist, evsel);
 267         return 0;
 268 }
 269
 270 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 271                                    struct perf_evsel *evsel)
 272 {
 273         if (evsel->system_wide)
 274                 return 1;
 275         else
 276                 return thread_map__nr(evlist->threads);
 277 }
 278
 279 void perf_evlist__disable(struct perf_evlist *evlist)
 280 {
 281         int cpu, thread;
 282         struct perf_evsel *pos;
 283         int nr_cpus = cpu_map__nr(evlist->cpus);
 284         int nr_threads;
 285
 286         for (cpu = 0; cpu < nr_cpus; cpu++) {
 287                 evlist__for_each(evlist, pos) {
 288                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 289                                 continue;
 290                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 291                         for (thread = 0; thread < nr_threads; thread++)
 292                                 ioctl(FD(pos, cpu, thread),
 293                                       PERF_EVENT_IOC_DISABLE, 0);
 294                 }
 295         }
 296 }
 297
 298 void perf_evlist__enable(struct perf_evlist *evlist)
 299 {
 300         int cpu, thread;
 301         struct perf_evsel *pos;
 302         int nr_cpus = cpu_map__nr(evlist->cpus);
 303         int nr_threads;
 304
 305         for (cpu = 0; cpu < nr_cpus; cpu++) {
 306                 evlist__for_each(evlist, pos) {
 307                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 308                                 continue;
 309                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 310                         for (thread = 0; thread < nr_threads; thread++)
 311                                 ioctl(FD(pos, cpu, thread),
 312                                       PERF_EVENT_IOC_ENABLE, 0);
 313                 }
 314         }
 315 }
 316
 317 int perf_evlist__disable_event(struct perf_evlist *evlist,
 318                                struct perf_evsel *evsel)
 319 {
 320         int cpu, thread, err;
 321         int nr_cpus = cpu_map__nr(evlist->cpus);
 322         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 323
 324         if (!evsel->fd)
 325                 return 0;
 326
 327         for (cpu = 0; cpu < nr_cpus; cpu++) {
 328                 for (thread = 0; thread < nr_threads; thread++) {
 329                         err = ioctl(FD(evsel, cpu, thread),
 330                                     PERF_EVENT_IOC_DISABLE, 0);
 331                         if (err)
 332                                 return err;
 333                 }
 334         }
 335         return 0;
 336 }
 337
 338 int perf_evlist__enable_event(struct perf_evlist *evlist,
 339                               struct perf_evsel *evsel)
 340 {
 341         int cpu, thread, err;
 342         int nr_cpus = cpu_map__nr(evlist->cpus);
 343         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 344
 345         if (!evsel->fd)
 346                 return -EINVAL;
 347
 348         for (cpu = 0; cpu < nr_cpus; cpu++) {
 349                 for (thread = 0; thread < nr_threads; thread++) {
 350                         err = ioctl(FD(evsel, cpu, thread),
 351                                     PERF_EVENT_IOC_ENABLE, 0);
 352                         if (err)
 353                                 return err;
 354                 }
 355         }
 356         return 0;
 357 }
 358
 359 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 360                                          struct perf_evsel *evsel, int cpu)
 361 {
 362         int thread, err;
 363         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 364
 365         if (!evsel->fd)
 366                 return -EINVAL;
 367
 368         for (thread = 0; thread < nr_threads; thread++) {
 369                 err = ioctl(FD(evsel, cpu, thread),
 370                             PERF_EVENT_IOC_ENABLE, 0);
 371                 if (err)
 372                         return err;
 373         }
 374         return 0;
 375 }
 376
 377 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 378                                             struct perf_evsel *evsel,
 379                                             int thread)
 380 {
 381         int cpu, err;
 382         int nr_cpus = cpu_map__nr(evlist->cpus);
 383
 384         if (!evsel->fd)
 385                 return -EINVAL;
 386
 387         for (cpu = 0; cpu < nr_cpus; cpu++) {
 388                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 389                 if (err)
 390                         return err;
 391         }
 392         return 0;
 393 }
 394
 395 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 396                                   struct perf_evsel *evsel, int idx)
 397 {
 398         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 399
 400         if (per_cpu_mmaps)
 401                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 402         else
 403                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 404 }
 405
 406 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 407 {
 408         int nr_cpus = cpu_map__nr(evlist->cpus);
 409         int nr_threads = thread_map__nr(evlist->threads);
 410         int nfds = 0;
 411         struct perf_evsel *evsel;
 412
 413         list_for_each_entry(evsel, &evlist->entries, node) {
 414                 if (evsel->system_wide)
 415                         nfds += nr_cpus;
 416                 else
 417                         nfds += nr_cpus * nr_threads;
 418         }
 419
 420         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 421             fdarray__grow(&evlist->pollfd, nfds) < 0)
 422                 return -ENOMEM;
 423
 424         return 0;
 425 }
 426
 427 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 428 {
 429         fcntl(fd, F_SETFL, O_NONBLOCK);
 430
 431         return fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
 432 }
 433
 434 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 435 {
 436         return fdarray__filter(&evlist->pollfd, revents_and_mask);
 437 }
 438
 439 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 440 {
 441         return fdarray__poll(&evlist->pollfd, timeout);
 442 }
 443
 444 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 445                                  struct perf_evsel *evsel,
 446                                  int cpu, int thread, u64 id)
 447 {
 448         int hash;
 449         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 450
 451         sid->id = id;
 452         sid->evsel = evsel;
 453         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 454         hlist_add_head(&sid->node, &evlist->heads[hash]);
 455 }
 456
 457 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 458                          int cpu, int thread, u64 id)
 459 {
 460         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 461         evsel->id[evsel->ids++] = id;
 462 }
 463
 464 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 465                                   struct perf_evsel *evsel,
 466                                   int cpu, int thread, int fd)
 467 {
 468         u64 read_data[4] = { 0, };
 469         int id_idx = 1; /* The first entry is the counter value */
 470         u64 id;
 471         int ret;
 472
 473         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 474         if (!ret)
 475                 goto add;
 476
 477         if (errno != ENOTTY)
 478                 return -1;
 479
 480         /* Legacy way to get event id.. All hail to old kernels! */
 481
 482         /*
 483          * This way does not work with group format read, so bail
 484          * out in that case.
 485          */
 486         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 487                 return -1;
 488
 489         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 490             read(fd, &read_data, sizeof(read_data)) == -1)
 491                 return -1;
 492
 493         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 494                 ++id_idx;
 495         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 496                 ++id_idx;
 497
 498         id = read_data[id_idx];
 499
 500  add:
 501         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 502         return 0;
 503 }
 504
 505 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 506 {
 507         struct hlist_head *head;
 508         struct perf_sample_id *sid;
 509         int hash;
 510
 511         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 512         head = &evlist->heads[hash];
 513
 514         hlist_for_each_entry(sid, head, node)
 515                 if (sid->id == id)
 516                         return sid;
 517
 518         return NULL;
 519 }
 520
 521 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 522 {
 523         struct perf_sample_id *sid;
 524
 525         if (evlist->nr_entries == 1)
 526                 return perf_evlist__first(evlist);
 527
 528         sid = perf_evlist__id2sid(evlist, id);
 529         if (sid)
 530                 return sid->evsel;
 531
 532         if (!perf_evlist__sample_id_all(evlist))
 533                 return perf_evlist__first(evlist);
 534
 535         return NULL;
 536 }
 537
 538 static int perf_evlist__event2id(struct perf_evlist *evlist,
 539                                  union perf_event *event, u64 *id)
 540 {
 541         const u64 *array = event->sample.array;
 542         ssize_t n;
 543
 544         n = (event->header.size - sizeof(event->header)) >> 3;
 545
 546         if (event->header.type == PERF_RECORD_SAMPLE) {
 547                 if (evlist->id_pos >= n)
 548                         return -1;
 549                 *id = array[evlist->id_pos];
 550         } else {
 551                 if (evlist->is_pos > n)
 552                         return -1;
 553                 n -= evlist->is_pos;
 554                 *id = array[n];
 555         }
 556         return 0;
 557 }
 558
 559 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 560                                                    union perf_event *event)
 561 {
 562         struct perf_evsel *first = perf_evlist__first(evlist);
 563         struct hlist_head *head;
 564         struct perf_sample_id *sid;
 565         int hash;
 566         u64 id;
 567
 568         if (evlist->nr_entries == 1)
 569                 return first;
 570
 571         if (!first->attr.sample_id_all &&
 572             event->header.type != PERF_RECORD_SAMPLE)
 573                 return first;
 574
 575         if (perf_evlist__event2id(evlist, event, &id))
 576                 return NULL;
 577
 578         /* Synthesized events have an id of zero */
 579         if (!id)
 580                 return first;
 581
 582         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 583         head = &evlist->heads[hash];
 584
 585         hlist_for_each_entry(sid, head, node) {
 586                 if (sid->id == id)
 587                         return sid->evsel;
 588         }
 589         return NULL;
 590 }
 591
 592 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 593 {
 594         struct perf_mmap *md = &evlist->mmap[idx];
 595         unsigned int head = perf_mmap__read_head(md);
 596         unsigned int old = md->prev;
 597         unsigned char *data = md->base + page_size;
 598         union perf_event *event = NULL;
 599
 600         if (evlist->overwrite) {
 601                 /*
 602                  * If we're further behind than half the buffer, there's a chance
 603                  * the writer will bite our tail and mess up the samples under us.
 604                  *
 605                  * If we somehow ended up ahead of the head, we got messed up.
 606                  *
 607                  * In either case, truncate and restart at head.
 608                  */
 609                 int diff = head - old;
 610                 if (diff > md->mask / 2 || diff < 0) {
 611                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 612
 613                         /*
 614                          * head points to a known good entry, start there.
 615                          */
 616                         old = head;
 617                 }
 618         }
 619
 620         if (old != head) {
 621                 size_t size;
 622
 623                 event = (union perf_event *)&data[old & md->mask];
 624                 size = event->header.size;
 625
 626                 /*
 627                  * Event straddles the mmap boundary -- header should always
 628                  * be inside due to u64 alignment of output.
 629                  */
 630                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 631                         unsigned int offset = old;
 632                         unsigned int len = min(sizeof(*event), size), cpy;
 633                         void *dst = md->event_copy;
 634
 635                         do {
 636                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 637                                 memcpy(dst, &data[offset & md->mask], cpy);
 638                                 offset += cpy;
 639                                 dst += cpy;
 640                                 len -= cpy;
 641                         } while (len);
 642
 643                         event = (union perf_event *) md->event_copy;
 644                 }
 645
 646                 old += size;
 647         }
 648
 649         md->prev = old;
 650
 651         return event;
 652 }
 653
 654 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 655 {
 656         if (!evlist->overwrite) {
 657                 struct perf_mmap *md = &evlist->mmap[idx];
 658                 unsigned int old = md->prev;
 659
 660                 perf_mmap__write_tail(md, old);
 661         }
 662 }
 663
 664 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 665 {
 666         if (evlist->mmap[idx].base != NULL) {
 667                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 668                 evlist->mmap[idx].base = NULL;
 669         }
 670 }
 671
 672 void perf_evlist__munmap(struct perf_evlist *evlist)
 673 {
 674         int i;
 675
 676         if (evlist->mmap == NULL)
 677                 return;
 678
 679         for (i = 0; i < evlist->nr_mmaps; i++)
 680                 __perf_evlist__munmap(evlist, i);
 681
 682         zfree(&evlist->mmap);
 683 }
 684
 685 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 686 {
 687         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 688         if (cpu_map__empty(evlist->cpus))
 689                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 690         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 691         return evlist->mmap != NULL ? 0 : -ENOMEM;
 692 }
 693
 694 struct mmap_params {
 695         int prot;
 696         int mask;
 697 };
 698
 699 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 700                                struct mmap_params *mp, int fd)
 701 {
 702         evlist->mmap[idx].prev = 0;
 703         evlist->mmap[idx].mask = mp->mask;
 704         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 705                                       MAP_SHARED, fd, 0);
 706         if (evlist->mmap[idx].base == MAP_FAILED) {
 707                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 708                           errno);
 709                 evlist->mmap[idx].base = NULL;
 710                 return -1;
 711         }
 712
 713         return 0;
 714 }
 715
 716 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 717                                        struct mmap_params *mp, int cpu,
 718                                        int thread, int *output)
 719 {
 720         struct perf_evsel *evsel;
 721
 722         evlist__for_each(evlist, evsel) {
 723                 int fd;
 724
 725                 if (evsel->system_wide && thread)
 726                         continue;
 727
 728                 fd = FD(evsel, cpu, thread);
 729
 730                 if (*output == -1) {
 731                         *output = fd;
 732                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 733                                 return -1;
 734                 } else {
 735                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 736                                 return -1;
 737                 }
 738
 739                 if (perf_evlist__add_pollfd(evlist, fd) < 0)
 740                         return -1;
 741
 742                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 743                     perf_evlist__id_add_fd(evlist, evsel, cpu, thread, fd) < 0)
 744                         return -1;
 745         }
 746
 747         return 0;
 748 }
 749
 750 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 751                                      struct mmap_params *mp)
 752 {
 753         int cpu, thread;
 754         int nr_cpus = cpu_map__nr(evlist->cpus);
 755         int nr_threads = thread_map__nr(evlist->threads);
 756
 757         pr_debug2("perf event ring buffer mmapped per cpu\n");
 758         for (cpu = 0; cpu < nr_cpus; cpu++) {
 759                 int output = -1;
 760
 761                 for (thread = 0; thread < nr_threads; thread++) {
 762                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 763                                                         thread, &output))
 764                                 goto out_unmap;
 765                 }
 766         }
 767
 768         return 0;
 769
 770 out_unmap:
 771         for (cpu = 0; cpu < nr_cpus; cpu++)
 772                 __perf_evlist__munmap(evlist, cpu);
 773         return -1;
 774 }
 775
 776 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 777                                         struct mmap_params *mp)
 778 {
 779         int thread;
 780         int nr_threads = thread_map__nr(evlist->threads);
 781
 782         pr_debug2("perf event ring buffer mmapped per thread\n");
 783         for (thread = 0; thread < nr_threads; thread++) {
 784                 int output = -1;
 785
 786                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 787                                                 &output))
 788                         goto out_unmap;
 789         }
 790
 791         return 0;
 792
 793 out_unmap:
 794         for (thread = 0; thread < nr_threads; thread++)
 795                 __perf_evlist__munmap(evlist, thread);
 796         return -1;
 797 }
 798
 799 static size_t perf_evlist__mmap_size(unsigned long pages)
 800 {
 801         /* 512 kiB: default amount of unprivileged mlocked memory */
 802         if (pages == UINT_MAX)
 803                 pages = (512 * 1024) / page_size;
 804         else if (!is_power_of_2(pages))
 805                 return 0;
 806
 807         return (pages + 1) * page_size;
 808 }
 809
 810 static long parse_pages_arg(const char *str, unsigned long min,
 811                             unsigned long max)
 812 {
 813         unsigned long pages, val;
 814         static struct parse_tag tags[] = {
 815                 { .tag  = 'B', .mult = 1       },
 816                 { .tag  = 'K', .mult = 1 << 10 },
 817                 { .tag  = 'M', .mult = 1 << 20 },
 818                 { .tag  = 'G', .mult = 1 << 30 },
 819                 { .tag  = 0 },
 820         };
 821
 822         if (str == NULL)
 823                 return -EINVAL;
 824
 825         val = parse_tag_value(str, tags);
 826         if (val != (unsigned long) -1) {
 827                 /* we got file size value */
 828                 pages = PERF_ALIGN(val, page_size) / page_size;
 829         } else {
 830                 /* we got pages count value */
 831                 char *eptr;
 832                 pages = strtoul(str, &eptr, 10);
 833                 if (*eptr != '\0')
 834                         return -EINVAL;
 835         }
 836
 837         if (pages == 0 && min == 0) {
 838                 /* leave number of pages at 0 */
 839         } else if (!is_power_of_2(pages)) {
 840                 /* round pages up to next power of 2 */
 841                 pages = next_pow2_l(pages);
 842                 if (!pages)
 843                         return -EINVAL;
 844                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 845                         pages * page_size, pages);
 846         }
 847
 848         if (pages > max)
 849                 return -EINVAL;
 850
 851         return pages;
 852 }
 853
 854 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 855                                   int unset __maybe_unused)
 856 {
 857         unsigned int *mmap_pages = opt->value;
 858         unsigned long max = UINT_MAX;
 859         long pages;
 860
 861         if (max > SIZE_MAX / page_size)
 862                 max = SIZE_MAX / page_size;
 863
 864         pages = parse_pages_arg(str, 1, max);
 865         if (pages < 0) {
 866                 pr_err("Invalid argument for --mmap_pages/-m\n");
 867                 return -1;
 868         }
 869
 870         *mmap_pages = pages;
 871         return 0;
 872 }
 873
 874 /**
 875  * perf_evlist__mmap - Create mmaps to receive events.
 876  * @evlist: list of events
 877  * @pages: map length in pages
 878  * @overwrite: overwrite older events?
 879  *
 880  * If @overwrite is %false the user needs to signal event consumption using
 881  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 882  * automatically.
 883  *
 884  * Return: %0 on success, negative error code otherwise.
 885  */
 886 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 887                       bool overwrite)
 888 {
 889         struct perf_evsel *evsel;
 890         const struct cpu_map *cpus = evlist->cpus;
 891         const struct thread_map *threads = evlist->threads;
 892         struct mmap_params mp = {
 893                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
 894         };
 895
 896         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
 897                 return -ENOMEM;
 898
 899         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
 900                 return -ENOMEM;
 901
 902         evlist->overwrite = overwrite;
 903         evlist->mmap_len = perf_evlist__mmap_size(pages);
 904         pr_debug("mmap size %zuB\n", evlist->mmap_len);
 905         mp.mask = evlist->mmap_len - page_size - 1;
 906
 907         evlist__for_each(evlist, evsel) {
 908                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
 909                     evsel->sample_id == NULL &&
 910                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
 911                         return -ENOMEM;
 912         }
 913
 914         if (cpu_map__empty(cpus))
 915                 return perf_evlist__mmap_per_thread(evlist, &mp);
 916
 917         return perf_evlist__mmap_per_cpu(evlist, &mp);
 918 }
 919
 920 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
 921 {
 922         evlist->threads = thread_map__new_str(target->pid, target->tid,
 923                                               target->uid);
 924
 925         if (evlist->threads == NULL)
 926                 return -1;
 927
 928         if (target__uses_dummy_map(target))
 929                 evlist->cpus = cpu_map__dummy_new();
 930         else
 931                 evlist->cpus = cpu_map__new(target->cpu_list);
 932
 933         if (evlist->cpus == NULL)
 934                 goto out_delete_threads;
 935
 936         return 0;
 937
 938 out_delete_threads:
 939         thread_map__delete(evlist->threads);
 940         return -1;
 941 }
 942
 943 int perf_evlist__apply_filters(struct perf_evlist *evlist)
 944 {
 945         struct perf_evsel *evsel;
 946         int err = 0;
 947         const int ncpus = cpu_map__nr(evlist->cpus),
 948                   nthreads = thread_map__nr(evlist->threads);
 949
 950         evlist__for_each(evlist, evsel) {
 951                 if (evsel->filter == NULL)
 952                         continue;
 953
 954                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
 955                 if (err)
 956                         break;
 957         }
 958
 959         return err;
 960 }
 961
 962 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
 963 {
 964         struct perf_evsel *evsel;
 965         int err = 0;
 966         const int ncpus = cpu_map__nr(evlist->cpus),
 967                   nthreads = thread_map__nr(evlist->threads);
 968
 969         evlist__for_each(evlist, evsel) {
 970                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
 971                 if (err)
 972                         break;
 973         }
 974
 975         return err;
 976 }
 977
 978 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
 979 {
 980         struct perf_evsel *pos;
 981
 982         if (evlist->nr_entries == 1)
 983                 return true;
 984
 985         if (evlist->id_pos < 0 || evlist->is_pos < 0)
 986                 return false;
 987
 988         evlist__for_each(evlist, pos) {
 989                 if (pos->id_pos != evlist->id_pos ||
 990                     pos->is_pos != evlist->is_pos)
 991                         return false;
 992         }
 993
 994         return true;
 995 }
 996
 997 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
 998 {
 999         struct perf_evsel *evsel;
1000
1001         if (evlist->combined_sample_type)
1002                 return evlist->combined_sample_type;
1003
1004         evlist__for_each(evlist, evsel)
1005                 evlist->combined_sample_type |= evsel->attr.sample_type;
1006
1007         return evlist->combined_sample_type;
1008 }
1009
1010 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1011 {
1012         evlist->combined_sample_type = 0;
1013         return __perf_evlist__combined_sample_type(evlist);
1014 }
1015
1016 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1017 {
1018         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1019         u64 read_format = first->attr.read_format;
1020         u64 sample_type = first->attr.sample_type;
1021
1022         evlist__for_each(evlist, pos) {
1023                 if (read_format != pos->attr.read_format)
1024                         return false;
1025         }
1026
1027         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1028         if ((sample_type & PERF_SAMPLE_READ) &&
1029             !(read_format & PERF_FORMAT_ID)) {
1030                 return false;
1031         }
1032
1033         return true;
1034 }
1035
1036 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1037 {
1038         struct perf_evsel *first = perf_evlist__first(evlist);
1039         return first->attr.read_format;
1040 }
1041
1042 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1043 {
1044         struct perf_evsel *first = perf_evlist__first(evlist);
1045         struct perf_sample *data;
1046         u64 sample_type;
1047         u16 size = 0;
1048
1049         if (!first->attr.sample_id_all)
1050                 goto out;
1051
1052         sample_type = first->attr.sample_type;
1053
1054         if (sample_type & PERF_SAMPLE_TID)
1055                 size += sizeof(data->tid) * 2;
1056
1057        if (sample_type & PERF_SAMPLE_TIME)
1058                 size += sizeof(data->time);
1059
1060         if (sample_type & PERF_SAMPLE_ID)
1061                 size += sizeof(data->id);
1062
1063         if (sample_type & PERF_SAMPLE_STREAM_ID)
1064                 size += sizeof(data->stream_id);
1065
1066         if (sample_type & PERF_SAMPLE_CPU)
1067                 size += sizeof(data->cpu) * 2;
1068
1069         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1070                 size += sizeof(data->id);
1071 out:
1072         return size;
1073 }
1074
1075 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1076 {
1077         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1078
1079         evlist__for_each_continue(evlist, pos) {
1080                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1081                         return false;
1082         }
1083
1084         return true;
1085 }
1086
1087 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1088 {
1089         struct perf_evsel *first = perf_evlist__first(evlist);
1090         return first->attr.sample_id_all;
1091 }
1092
1093 void perf_evlist__set_selected(struct perf_evlist *evlist,
1094                                struct perf_evsel *evsel)
1095 {
1096         evlist->selected = evsel;
1097 }
1098
1099 void perf_evlist__close(struct perf_evlist *evlist)
1100 {
1101         struct perf_evsel *evsel;
1102         int ncpus = cpu_map__nr(evlist->cpus);
1103         int nthreads = thread_map__nr(evlist->threads);
1104         int n;
1105
1106         evlist__for_each_reverse(evlist, evsel) {
1107                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1108                 perf_evsel__close(evsel, n, nthreads);
1109         }
1110 }
1111
1112 int perf_evlist__open(struct perf_evlist *evlist)
1113 {
1114         struct perf_evsel *evsel;
1115         int err;
1116
1117         perf_evlist__update_id_pos(evlist);
1118
1119         evlist__for_each(evlist, evsel) {
1120                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1121                 if (err < 0)
1122                         goto out_err;
1123         }
1124
1125         return 0;
1126 out_err:
1127         perf_evlist__close(evlist);
1128         errno = -err;
1129         return err;
1130 }
1131
1132 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1133                                   const char *argv[], bool pipe_output,
1134                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1135 {
1136         int child_ready_pipe[2], go_pipe[2];
1137         char bf;
1138
1139         if (pipe(child_ready_pipe) < 0) {
1140                 perror("failed to create 'ready' pipe");
1141                 return -1;
1142         }
1143
1144         if (pipe(go_pipe) < 0) {
1145                 perror("failed to create 'go' pipe");
1146                 goto out_close_ready_pipe;
1147         }
1148
1149         evlist->workload.pid = fork();
1150         if (evlist->workload.pid < 0) {
1151                 perror("failed to fork");
1152                 goto out_close_pipes;
1153         }
1154
1155         if (!evlist->workload.pid) {
1156                 int ret;
1157
1158                 if (pipe_output)
1159                         dup2(2, 1);
1160
1161                 signal(SIGTERM, SIG_DFL);
1162
1163                 close(child_ready_pipe[0]);
1164                 close(go_pipe[1]);
1165                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1166
1167                 /*
1168                  * Tell the parent we're ready to go
1169                  */
1170                 close(child_ready_pipe[1]);
1171
1172                 /*
1173                  * Wait until the parent tells us to go.
1174                  */
1175                 ret = read(go_pipe[0], &bf, 1);
1176                 /*
1177                  * The parent will ask for the execvp() to be performed by
1178                  * writing exactly one byte, in workload.cork_fd, usually via
1179                  * perf_evlist__start_workload().
1180                  *
1181                  * For cancelling the workload without actuallin running it,
1182                  * the parent will just close workload.cork_fd, without writing
1183                  * anything, i.e. read will return zero and we just exit()
1184                  * here.
1185                  */
1186                 if (ret != 1) {
1187                         if (ret == -1)
1188                                 perror("unable to read pipe");
1189                         exit(ret);
1190                 }
1191
1192                 execvp(argv[0], (char **)argv);
1193
1194                 if (exec_error) {
1195                         union sigval val;
1196
1197                         val.sival_int = errno;
1198                         if (sigqueue(getppid(), SIGUSR1, val))
1199                                 perror(argv[0]);
1200                 } else
1201                         perror(argv[0]);
1202                 exit(-1);
1203         }
1204
1205         if (exec_error) {
1206                 struct sigaction act = {
1207                         .sa_flags     = SA_SIGINFO,
1208                         .sa_sigaction = exec_error,
1209                 };
1210                 sigaction(SIGUSR1, &act, NULL);
1211         }
1212
1213         if (target__none(target))
1214                 evlist->threads->map[0] = evlist->workload.pid;
1215
1216         close(child_ready_pipe[1]);
1217         close(go_pipe[0]);
1218         /*
1219          * wait for child to settle
1220          */
1221         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1222                 perror("unable to read pipe");
1223                 goto out_close_pipes;
1224         }
1225
1226         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1227         evlist->workload.cork_fd = go_pipe[1];
1228         close(child_ready_pipe[0]);
1229         return 0;
1230
1231 out_close_pipes:
1232         close(go_pipe[0]);
1233         close(go_pipe[1]);
1234 out_close_ready_pipe:
1235         close(child_ready_pipe[0]);
1236         close(child_ready_pipe[1]);
1237         return -1;
1238 }
1239
1240 int perf_evlist__start_workload(struct perf_evlist *evlist)
1241 {
1242         if (evlist->workload.cork_fd > 0) {
1243                 char bf = 0;
1244                 int ret;
1245                 /*
1246                  * Remove the cork, let it rip!
1247                  */
1248                 ret = write(evlist->workload.cork_fd, &bf, 1);
1249                 if (ret < 0)
1250                         perror("enable to write to pipe");
1251
1252                 close(evlist->workload.cork_fd);
1253                 return ret;
1254         }
1255
1256         return 0;
1257 }
1258
1259 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1260                               struct perf_sample *sample)
1261 {
1262         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1263
1264         if (!evsel)
1265                 return -EFAULT;
1266         return perf_evsel__parse_sample(evsel, event, sample);
1267 }
1268
1269 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1270 {
1271         struct perf_evsel *evsel;
1272         size_t printed = 0;
1273
1274         evlist__for_each(evlist, evsel) {
1275                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1276                                    perf_evsel__name(evsel));
1277         }
1278
1279         return printed + fprintf(fp, "\n");
1280 }
1281
1282 int perf_evlist__strerror_tp(struct perf_evlist *evlist __maybe_unused,
1283                              int err, char *buf, size_t size)
1284 {
1285         char sbuf[128];
1286
1287         switch (err) {
1288         case ENOENT:
1289                 scnprintf(buf, size, "%s",
1290                           "Error:\tUnable to find debugfs\n"
1291                           "Hint:\tWas your kernel was compiled with debugfs support?\n"
1292                           "Hint:\tIs the debugfs filesystem mounted?\n"
1293                           "Hint:\tTry 'sudo mount -t debugfs nodev /sys/kernel/debug'");
1294                 break;
1295         case EACCES:
1296                 scnprintf(buf, size,
1297                           "Error:\tNo permissions to read %s/tracing/events/raw_syscalls\n"
1298                           "Hint:\tTry 'sudo mount -o remount,mode=755 %s'\n",
1299                           debugfs_mountpoint, debugfs_mountpoint);
1300                 break;
1301         default:
1302                 scnprintf(buf, size, "%s", strerror_r(err, sbuf, sizeof(sbuf)));
1303                 break;
1304         }
1305
1306         return 0;
1307 }
1308
1309 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1310                                int err, char *buf, size_t size)
1311 {
1312         int printed, value;
1313         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1314
1315         switch (err) {
1316         case EACCES:
1317         case EPERM:
1318                 printed = scnprintf(buf, size,
1319                                     "Error:\t%s.\n"
1320                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1321
1322                 value = perf_event_paranoid();
1323
1324                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1325
1326                 if (value >= 2) {
1327                         printed += scnprintf(buf + printed, size - printed,
1328                                              "For your workloads it needs to be <= 1\nHint:\t");
1329                 }
1330                 printed += scnprintf(buf + printed, size - printed,
1331                                      "For system wide tracing it needs to be set to -1.\n");
1332
1333                 printed += scnprintf(buf + printed, size - printed,
1334                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1335                                     "Hint:\tThe current value is %d.", value);
1336                 break;
1337         default:
1338                 scnprintf(buf, size, "%s", emsg);
1339                 break;
1340         }
1341
1342         return 0;
1343 }
1344
1345 void perf_evlist__to_front(struct perf_evlist *evlist,
1346                            struct perf_evsel *move_evsel)
1347 {
1348         struct perf_evsel *evsel, *n;
1349         LIST_HEAD(move);
1350
1351         if (move_evsel == perf_evlist__first(evlist))
1352                 return;
1353
1354         evlist__for_each_safe(evlist, n, evsel) {
1355                 if (evsel->leader == move_evsel->leader)
1356                         list_move_tail(&evsel->node, &move);
1357         }
1358
1359         list_splice(&move, &evlist->entries);
1360 }
1361
1362 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1363                                      struct perf_evsel *tracking_evsel)
1364 {
1365         struct perf_evsel *evsel;
1366
1367         if (tracking_evsel->tracking)
1368                 return;
1369
1370         evlist__for_each(evlist, evsel) {
1371                 if (evsel != tracking_evsel)
1372                         evsel->tracking = false;
1373         }
1374
1375         tracking_evsel->tracking = true;
1376 }