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 <api/fs/fs.h>
  12 #include <poll.h>
  13 #include "cpumap.h"
  14 #include "thread_map.h"
  15 #include "target.h"
  16 #include "evlist.h"
  17 #include "evsel.h"
  18 #include "debug.h"
  19 #include <unistd.h>
  20
  21 #include "parse-events.h"
  22 #include "parse-options.h"
  23
  24 #include <sys/mman.h>
  25
  26 #include <linux/bitops.h>
  27 #include <linux/hash.h>
  28 #include <linux/log2.h>
  29
  30 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx);
  31 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx);
  32
  33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
  34 #define SID(e, x, y) xyarray__entry(e->sample_id, x, y)
  35
  36 void perf_evlist__init(struct perf_evlist *evlist, struct cpu_map *cpus,
  37                        struct thread_map *threads)
  38 {
  39         int i;
  40
  41         for (i = 0; i < PERF_EVLIST__HLIST_SIZE; ++i)
  42                 INIT_HLIST_HEAD(&evlist->heads[i]);
  43         INIT_LIST_HEAD(&evlist->entries);
  44         perf_evlist__set_maps(evlist, cpus, threads);
  45         fdarray__init(&evlist->pollfd, 64);
  46         evlist->workload.pid = -1;
  47 }
  48
  49 struct perf_evlist *perf_evlist__new(void)
  50 {
  51         struct perf_evlist *evlist = zalloc(sizeof(*evlist));
  52
  53         if (evlist != NULL)
  54                 perf_evlist__init(evlist, NULL, NULL);
  55
  56         return evlist;
  57 }
  58
  59 struct perf_evlist *perf_evlist__new_default(void)
  60 {
  61         struct perf_evlist *evlist = perf_evlist__new();
  62
  63         if (evlist && perf_evlist__add_default(evlist)) {
  64                 perf_evlist__delete(evlist);
  65                 evlist = NULL;
  66         }
  67
  68         return evlist;
  69 }
  70
  71 /**
  72  * perf_evlist__set_id_pos - set the positions of event ids.
  73  * @evlist: selected event list
  74  *
  75  * Events with compatible sample types all have the same id_pos
  76  * and is_pos.  For convenience, put a copy on evlist.
  77  */
  78 void perf_evlist__set_id_pos(struct perf_evlist *evlist)
  79 {
  80         struct perf_evsel *first = perf_evlist__first(evlist);
  81
  82         evlist->id_pos = first->id_pos;
  83         evlist->is_pos = first->is_pos;
  84 }
  85
  86 static void perf_evlist__update_id_pos(struct perf_evlist *evlist)
  87 {
  88         struct perf_evsel *evsel;
  89
  90         evlist__for_each(evlist, evsel)
  91                 perf_evsel__calc_id_pos(evsel);
  92
  93         perf_evlist__set_id_pos(evlist);
  94 }
  95
  96 static void perf_evlist__purge(struct perf_evlist *evlist)
  97 {
  98         struct perf_evsel *pos, *n;
  99
 100         evlist__for_each_safe(evlist, n, pos) {
 101                 list_del_init(&pos->node);
 102                 perf_evsel__delete(pos);
 103         }
 104
 105         evlist->nr_entries = 0;
 106 }
 107
 108 void perf_evlist__exit(struct perf_evlist *evlist)
 109 {
 110         zfree(&evlist->mmap);
 111         fdarray__exit(&evlist->pollfd);
 112 }
 113
 114 void perf_evlist__delete(struct perf_evlist *evlist)
 115 {
 116         perf_evlist__munmap(evlist);
 117         perf_evlist__close(evlist);
 118         cpu_map__delete(evlist->cpus);
 119         thread_map__delete(evlist->threads);
 120         evlist->cpus = NULL;
 121         evlist->threads = NULL;
 122         perf_evlist__purge(evlist);
 123         perf_evlist__exit(evlist);
 124         free(evlist);
 125 }
 126
 127 void perf_evlist__add(struct perf_evlist *evlist, struct perf_evsel *entry)
 128 {
 129         list_add_tail(&entry->node, &evlist->entries);
 130         entry->idx = evlist->nr_entries;
 131         entry->tracking = !entry->idx;
 132
 133         if (!evlist->nr_entries++)
 134                 perf_evlist__set_id_pos(evlist);
 135 }
 136
 137 void perf_evlist__splice_list_tail(struct perf_evlist *evlist,
 138                                    struct list_head *list,
 139                                    int nr_entries)
 140 {
 141         bool set_id_pos = !evlist->nr_entries;
 142
 143         list_splice_tail(list, &evlist->entries);
 144         evlist->nr_entries += nr_entries;
 145         if (set_id_pos)
 146                 perf_evlist__set_id_pos(evlist);
 147 }
 148
 149 void __perf_evlist__set_leader(struct list_head *list)
 150 {
 151         struct perf_evsel *evsel, *leader;
 152
 153         leader = list_entry(list->next, struct perf_evsel, node);
 154         evsel = list_entry(list->prev, struct perf_evsel, node);
 155
 156         leader->nr_members = evsel->idx - leader->idx + 1;
 157
 158         __evlist__for_each(list, evsel) {
 159                 evsel->leader = leader;
 160         }
 161 }
 162
 163 void perf_evlist__set_leader(struct perf_evlist *evlist)
 164 {
 165         if (evlist->nr_entries) {
 166                 evlist->nr_groups = evlist->nr_entries > 1 ? 1 : 0;
 167                 __perf_evlist__set_leader(&evlist->entries);
 168         }
 169 }
 170
 171 int perf_evlist__add_default(struct perf_evlist *evlist)
 172 {
 173         struct perf_event_attr attr = {
 174                 .type = PERF_TYPE_HARDWARE,
 175                 .config = PERF_COUNT_HW_CPU_CYCLES,
 176         };
 177         struct perf_evsel *evsel;
 178
 179         event_attr_init(&attr);
 180
 181         evsel = perf_evsel__new(&attr);
 182         if (evsel == NULL)
 183                 goto error;
 184
 185         /* use strdup() because free(evsel) assumes name is allocated */
 186         evsel->name = strdup("cycles");
 187         if (!evsel->name)
 188                 goto error_free;
 189
 190         perf_evlist__add(evlist, evsel);
 191         return 0;
 192 error_free:
 193         perf_evsel__delete(evsel);
 194 error:
 195         return -ENOMEM;
 196 }
 197
 198 static int perf_evlist__add_attrs(struct perf_evlist *evlist,
 199                                   struct perf_event_attr *attrs, size_t nr_attrs)
 200 {
 201         struct perf_evsel *evsel, *n;
 202         LIST_HEAD(head);
 203         size_t i;
 204
 205         for (i = 0; i < nr_attrs; i++) {
 206                 evsel = perf_evsel__new_idx(attrs + i, evlist->nr_entries + i);
 207                 if (evsel == NULL)
 208                         goto out_delete_partial_list;
 209                 list_add_tail(&evsel->node, &head);
 210         }
 211
 212         perf_evlist__splice_list_tail(evlist, &head, nr_attrs);
 213
 214         return 0;
 215
 216 out_delete_partial_list:
 217         __evlist__for_each_safe(&head, n, evsel)
 218                 perf_evsel__delete(evsel);
 219         return -1;
 220 }
 221
 222 int __perf_evlist__add_default_attrs(struct perf_evlist *evlist,
 223                                      struct perf_event_attr *attrs, size_t nr_attrs)
 224 {
 225         size_t i;
 226
 227         for (i = 0; i < nr_attrs; i++)
 228                 event_attr_init(attrs + i);
 229
 230         return perf_evlist__add_attrs(evlist, attrs, nr_attrs);
 231 }
 232
 233 struct perf_evsel *
 234 perf_evlist__find_tracepoint_by_id(struct perf_evlist *evlist, int id)
 235 {
 236         struct perf_evsel *evsel;
 237
 238         evlist__for_each(evlist, evsel) {
 239                 if (evsel->attr.type   == PERF_TYPE_TRACEPOINT &&
 240                     (int)evsel->attr.config == id)
 241                         return evsel;
 242         }
 243
 244         return NULL;
 245 }
 246
 247 struct perf_evsel *
 248 perf_evlist__find_tracepoint_by_name(struct perf_evlist *evlist,
 249                                      const char *name)
 250 {
 251         struct perf_evsel *evsel;
 252
 253         evlist__for_each(evlist, evsel) {
 254                 if ((evsel->attr.type == PERF_TYPE_TRACEPOINT) &&
 255                     (strcmp(evsel->name, name) == 0))
 256                         return evsel;
 257         }
 258
 259         return NULL;
 260 }
 261
 262 int perf_evlist__add_newtp(struct perf_evlist *evlist,
 263                            const char *sys, const char *name, void *handler)
 264 {
 265         struct perf_evsel *evsel = perf_evsel__newtp(sys, name);
 266
 267         if (evsel == NULL)
 268                 return -1;
 269
 270         evsel->handler = handler;
 271         perf_evlist__add(evlist, evsel);
 272         return 0;
 273 }
 274
 275 static int perf_evlist__nr_threads(struct perf_evlist *evlist,
 276                                    struct perf_evsel *evsel)
 277 {
 278         if (evsel->system_wide)
 279                 return 1;
 280         else
 281                 return thread_map__nr(evlist->threads);
 282 }
 283
 284 void perf_evlist__disable(struct perf_evlist *evlist)
 285 {
 286         int cpu, thread;
 287         struct perf_evsel *pos;
 288         int nr_cpus = cpu_map__nr(evlist->cpus);
 289         int nr_threads;
 290
 291         for (cpu = 0; cpu < nr_cpus; cpu++) {
 292                 evlist__for_each(evlist, pos) {
 293                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 294                                 continue;
 295                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 296                         for (thread = 0; thread < nr_threads; thread++)
 297                                 ioctl(FD(pos, cpu, thread),
 298                                       PERF_EVENT_IOC_DISABLE, 0);
 299                 }
 300         }
 301 }
 302
 303 void perf_evlist__enable(struct perf_evlist *evlist)
 304 {
 305         int cpu, thread;
 306         struct perf_evsel *pos;
 307         int nr_cpus = cpu_map__nr(evlist->cpus);
 308         int nr_threads;
 309
 310         for (cpu = 0; cpu < nr_cpus; cpu++) {
 311                 evlist__for_each(evlist, pos) {
 312                         if (!perf_evsel__is_group_leader(pos) || !pos->fd)
 313                                 continue;
 314                         nr_threads = perf_evlist__nr_threads(evlist, pos);
 315                         for (thread = 0; thread < nr_threads; thread++)
 316                                 ioctl(FD(pos, cpu, thread),
 317                                       PERF_EVENT_IOC_ENABLE, 0);
 318                 }
 319         }
 320 }
 321
 322 int perf_evlist__disable_event(struct perf_evlist *evlist,
 323                                struct perf_evsel *evsel)
 324 {
 325         int cpu, thread, err;
 326         int nr_cpus = cpu_map__nr(evlist->cpus);
 327         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 328
 329         if (!evsel->fd)
 330                 return 0;
 331
 332         for (cpu = 0; cpu < nr_cpus; cpu++) {
 333                 for (thread = 0; thread < nr_threads; thread++) {
 334                         err = ioctl(FD(evsel, cpu, thread),
 335                                     PERF_EVENT_IOC_DISABLE, 0);
 336                         if (err)
 337                                 return err;
 338                 }
 339         }
 340         return 0;
 341 }
 342
 343 int perf_evlist__enable_event(struct perf_evlist *evlist,
 344                               struct perf_evsel *evsel)
 345 {
 346         int cpu, thread, err;
 347         int nr_cpus = cpu_map__nr(evlist->cpus);
 348         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 349
 350         if (!evsel->fd)
 351                 return -EINVAL;
 352
 353         for (cpu = 0; cpu < nr_cpus; cpu++) {
 354                 for (thread = 0; thread < nr_threads; thread++) {
 355                         err = ioctl(FD(evsel, cpu, thread),
 356                                     PERF_EVENT_IOC_ENABLE, 0);
 357                         if (err)
 358                                 return err;
 359                 }
 360         }
 361         return 0;
 362 }
 363
 364 static int perf_evlist__enable_event_cpu(struct perf_evlist *evlist,
 365                                          struct perf_evsel *evsel, int cpu)
 366 {
 367         int thread, err;
 368         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 369
 370         if (!evsel->fd)
 371                 return -EINVAL;
 372
 373         for (thread = 0; thread < nr_threads; thread++) {
 374                 err = ioctl(FD(evsel, cpu, thread),
 375                             PERF_EVENT_IOC_ENABLE, 0);
 376                 if (err)
 377                         return err;
 378         }
 379         return 0;
 380 }
 381
 382 static int perf_evlist__enable_event_thread(struct perf_evlist *evlist,
 383                                             struct perf_evsel *evsel,
 384                                             int thread)
 385 {
 386         int cpu, err;
 387         int nr_cpus = cpu_map__nr(evlist->cpus);
 388
 389         if (!evsel->fd)
 390                 return -EINVAL;
 391
 392         for (cpu = 0; cpu < nr_cpus; cpu++) {
 393                 err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 394                 if (err)
 395                         return err;
 396         }
 397         return 0;
 398 }
 399
 400 int perf_evlist__enable_event_idx(struct perf_evlist *evlist,
 401                                   struct perf_evsel *evsel, int idx)
 402 {
 403         bool per_cpu_mmaps = !cpu_map__empty(evlist->cpus);
 404
 405         if (per_cpu_mmaps)
 406                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 407         else
 408                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 409 }
 410
 411 int perf_evlist__alloc_pollfd(struct perf_evlist *evlist)
 412 {
 413         int nr_cpus = cpu_map__nr(evlist->cpus);
 414         int nr_threads = thread_map__nr(evlist->threads);
 415         int nfds = 0;
 416         struct perf_evsel *evsel;
 417
 418         evlist__for_each(evlist, evsel) {
 419                 if (evsel->system_wide)
 420                         nfds += nr_cpus;
 421                 else
 422                         nfds += nr_cpus * nr_threads;
 423         }
 424
 425         if (fdarray__available_entries(&evlist->pollfd) < nfds &&
 426             fdarray__grow(&evlist->pollfd, nfds) < 0)
 427                 return -ENOMEM;
 428
 429         return 0;
 430 }
 431
 432 static int __perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd, int idx)
 433 {
 434         int pos = fdarray__add(&evlist->pollfd, fd, POLLIN | POLLERR | POLLHUP);
 435         /*
 436          * Save the idx so that when we filter out fds POLLHUP'ed we can
 437          * close the associated evlist->mmap[] entry.
 438          */
 439         if (pos >= 0) {
 440                 evlist->pollfd.priv[pos].idx = idx;
 441
 442                 fcntl(fd, F_SETFL, O_NONBLOCK);
 443         }
 444
 445         return pos;
 446 }
 447
 448 int perf_evlist__add_pollfd(struct perf_evlist *evlist, int fd)
 449 {
 450         return __perf_evlist__add_pollfd(evlist, fd, -1);
 451 }
 452
 453 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd)
 454 {
 455         struct perf_evlist *evlist = container_of(fda, struct perf_evlist, pollfd);
 456
 457         perf_evlist__mmap_put(evlist, fda->priv[fd].idx);
 458 }
 459
 460 int perf_evlist__filter_pollfd(struct perf_evlist *evlist, short revents_and_mask)
 461 {
 462         return fdarray__filter(&evlist->pollfd, revents_and_mask,
 463                                perf_evlist__munmap_filtered);
 464 }
 465
 466 int perf_evlist__poll(struct perf_evlist *evlist, int timeout)
 467 {
 468         return fdarray__poll(&evlist->pollfd, timeout);
 469 }
 470
 471 static void perf_evlist__id_hash(struct perf_evlist *evlist,
 472                                  struct perf_evsel *evsel,
 473                                  int cpu, int thread, u64 id)
 474 {
 475         int hash;
 476         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 477
 478         sid->id = id;
 479         sid->evsel = evsel;
 480         hash = hash_64(sid->id, PERF_EVLIST__HLIST_BITS);
 481         hlist_add_head(&sid->node, &evlist->heads[hash]);
 482 }
 483
 484 void perf_evlist__id_add(struct perf_evlist *evlist, struct perf_evsel *evsel,
 485                          int cpu, int thread, u64 id)
 486 {
 487         perf_evlist__id_hash(evlist, evsel, cpu, thread, id);
 488         evsel->id[evsel->ids++] = id;
 489 }
 490
 491 static int perf_evlist__id_add_fd(struct perf_evlist *evlist,
 492                                   struct perf_evsel *evsel,
 493                                   int cpu, int thread, int fd)
 494 {
 495         u64 read_data[4] = { 0, };
 496         int id_idx = 1; /* The first entry is the counter value */
 497         u64 id;
 498         int ret;
 499
 500         ret = ioctl(fd, PERF_EVENT_IOC_ID, &id);
 501         if (!ret)
 502                 goto add;
 503
 504         if (errno != ENOTTY)
 505                 return -1;
 506
 507         /* Legacy way to get event id.. All hail to old kernels! */
 508
 509         /*
 510          * This way does not work with group format read, so bail
 511          * out in that case.
 512          */
 513         if (perf_evlist__read_format(evlist) & PERF_FORMAT_GROUP)
 514                 return -1;
 515
 516         if (!(evsel->attr.read_format & PERF_FORMAT_ID) ||
 517             read(fd, &read_data, sizeof(read_data)) == -1)
 518                 return -1;
 519
 520         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 521                 ++id_idx;
 522         if (evsel->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 523                 ++id_idx;
 524
 525         id = read_data[id_idx];
 526
 527  add:
 528         perf_evlist__id_add(evlist, evsel, cpu, thread, id);
 529         return 0;
 530 }
 531
 532 static void perf_evlist__set_sid_idx(struct perf_evlist *evlist,
 533                                      struct perf_evsel *evsel, int idx, int cpu,
 534                                      int thread)
 535 {
 536         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 537         sid->idx = idx;
 538         if (evlist->cpus && cpu >= 0)
 539                 sid->cpu = evlist->cpus->map[cpu];
 540         else
 541                 sid->cpu = -1;
 542         if (!evsel->system_wide && evlist->threads && thread >= 0)
 543                 sid->tid = evlist->threads->map[thread];
 544         else
 545                 sid->tid = -1;
 546 }
 547
 548 struct perf_sample_id *perf_evlist__id2sid(struct perf_evlist *evlist, u64 id)
 549 {
 550         struct hlist_head *head;
 551         struct perf_sample_id *sid;
 552         int hash;
 553
 554         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 555         head = &evlist->heads[hash];
 556
 557         hlist_for_each_entry(sid, head, node)
 558                 if (sid->id == id)
 559                         return sid;
 560
 561         return NULL;
 562 }
 563
 564 struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id)
 565 {
 566         struct perf_sample_id *sid;
 567
 568         if (evlist->nr_entries == 1)
 569                 return perf_evlist__first(evlist);
 570
 571         sid = perf_evlist__id2sid(evlist, id);
 572         if (sid)
 573                 return sid->evsel;
 574
 575         if (!perf_evlist__sample_id_all(evlist))
 576                 return perf_evlist__first(evlist);
 577
 578         return NULL;
 579 }
 580
 581 static int perf_evlist__event2id(struct perf_evlist *evlist,
 582                                  union perf_event *event, u64 *id)
 583 {
 584         const u64 *array = event->sample.array;
 585         ssize_t n;
 586
 587         n = (event->header.size - sizeof(event->header)) >> 3;
 588
 589         if (event->header.type == PERF_RECORD_SAMPLE) {
 590                 if (evlist->id_pos >= n)
 591                         return -1;
 592                 *id = array[evlist->id_pos];
 593         } else {
 594                 if (evlist->is_pos > n)
 595                         return -1;
 596                 n -= evlist->is_pos;
 597                 *id = array[n];
 598         }
 599         return 0;
 600 }
 601
 602 static struct perf_evsel *perf_evlist__event2evsel(struct perf_evlist *evlist,
 603                                                    union perf_event *event)
 604 {
 605         struct perf_evsel *first = perf_evlist__first(evlist);
 606         struct hlist_head *head;
 607         struct perf_sample_id *sid;
 608         int hash;
 609         u64 id;
 610
 611         if (evlist->nr_entries == 1)
 612                 return first;
 613
 614         if (!first->attr.sample_id_all &&
 615             event->header.type != PERF_RECORD_SAMPLE)
 616                 return first;
 617
 618         if (perf_evlist__event2id(evlist, event, &id))
 619                 return NULL;
 620
 621         /* Synthesized events have an id of zero */
 622         if (!id)
 623                 return first;
 624
 625         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 626         head = &evlist->heads[hash];
 627
 628         hlist_for_each_entry(sid, head, node) {
 629                 if (sid->id == id)
 630                         return sid->evsel;
 631         }
 632         return NULL;
 633 }
 634
 635 union perf_event *perf_evlist__mmap_read(struct perf_evlist *evlist, int idx)
 636 {
 637         struct perf_mmap *md = &evlist->mmap[idx];
 638         unsigned int head = perf_mmap__read_head(md);
 639         unsigned int old = md->prev;
 640         unsigned char *data = md->base + page_size;
 641         union perf_event *event = NULL;
 642
 643         if (evlist->overwrite) {
 644                 /*
 645                  * If we're further behind than half the buffer, there's a chance
 646                  * the writer will bite our tail and mess up the samples under us.
 647                  *
 648                  * If we somehow ended up ahead of the head, we got messed up.
 649                  *
 650                  * In either case, truncate and restart at head.
 651                  */
 652                 int diff = head - old;
 653                 if (diff > md->mask / 2 || diff < 0) {
 654                         fprintf(stderr, "WARNING: failed to keep up with mmap data.\n");
 655
 656                         /*
 657                          * head points to a known good entry, start there.
 658                          */
 659                         old = head;
 660                 }
 661         }
 662
 663         if (old != head) {
 664                 size_t size;
 665
 666                 event = (union perf_event *)&data[old & md->mask];
 667                 size = event->header.size;
 668
 669                 /*
 670                  * Event straddles the mmap boundary -- header should always
 671                  * be inside due to u64 alignment of output.
 672                  */
 673                 if ((old & md->mask) + size != ((old + size) & md->mask)) {
 674                         unsigned int offset = old;
 675                         unsigned int len = min(sizeof(*event), size), cpy;
 676                         void *dst = md->event_copy;
 677
 678                         do {
 679                                 cpy = min(md->mask + 1 - (offset & md->mask), len);
 680                                 memcpy(dst, &data[offset & md->mask], cpy);
 681                                 offset += cpy;
 682                                 dst += cpy;
 683                                 len -= cpy;
 684                         } while (len);
 685
 686                         event = (union perf_event *) md->event_copy;
 687                 }
 688
 689                 old += size;
 690         }
 691
 692         md->prev = old;
 693
 694         return event;
 695 }
 696
 697 static bool perf_mmap__empty(struct perf_mmap *md)
 698 {
 699         return perf_mmap__read_head(md) != md->prev;
 700 }
 701
 702 static void perf_evlist__mmap_get(struct perf_evlist *evlist, int idx)
 703 {
 704         ++evlist->mmap[idx].refcnt;
 705 }
 706
 707 static void perf_evlist__mmap_put(struct perf_evlist *evlist, int idx)
 708 {
 709         BUG_ON(evlist->mmap[idx].refcnt == 0);
 710
 711         if (--evlist->mmap[idx].refcnt == 0)
 712                 __perf_evlist__munmap(evlist, idx);
 713 }
 714
 715 void perf_evlist__mmap_consume(struct perf_evlist *evlist, int idx)
 716 {
 717         struct perf_mmap *md = &evlist->mmap[idx];
 718
 719         if (!evlist->overwrite) {
 720                 unsigned int old = md->prev;
 721
 722                 perf_mmap__write_tail(md, old);
 723         }
 724
 725         if (md->refcnt == 1 && perf_mmap__empty(md))
 726                 perf_evlist__mmap_put(evlist, idx);
 727 }
 728
 729 static void __perf_evlist__munmap(struct perf_evlist *evlist, int idx)
 730 {
 731         if (evlist->mmap[idx].base != NULL) {
 732                 munmap(evlist->mmap[idx].base, evlist->mmap_len);
 733                 evlist->mmap[idx].base = NULL;
 734                 evlist->mmap[idx].refcnt = 0;
 735         }
 736 }
 737
 738 void perf_evlist__munmap(struct perf_evlist *evlist)
 739 {
 740         int i;
 741
 742         if (evlist->mmap == NULL)
 743                 return;
 744
 745         for (i = 0; i < evlist->nr_mmaps; i++)
 746                 __perf_evlist__munmap(evlist, i);
 747
 748         zfree(&evlist->mmap);
 749 }
 750
 751 static int perf_evlist__alloc_mmap(struct perf_evlist *evlist)
 752 {
 753         evlist->nr_mmaps = cpu_map__nr(evlist->cpus);
 754         if (cpu_map__empty(evlist->cpus))
 755                 evlist->nr_mmaps = thread_map__nr(evlist->threads);
 756         evlist->mmap = zalloc(evlist->nr_mmaps * sizeof(struct perf_mmap));
 757         return evlist->mmap != NULL ? 0 : -ENOMEM;
 758 }
 759
 760 struct mmap_params {
 761         int prot;
 762         int mask;
 763 };
 764
 765 static int __perf_evlist__mmap(struct perf_evlist *evlist, int idx,
 766                                struct mmap_params *mp, int fd)
 767 {
 768         /*
 769          * The last one will be done at perf_evlist__mmap_consume(), so that we
 770          * make sure we don't prevent tools from consuming every last event in
 771          * the ring buffer.
 772          *
 773          * I.e. we can get the POLLHUP meaning that the fd doesn't exist
 774          * anymore, but the last events for it are still in the ring buffer,
 775          * waiting to be consumed.
 776          *
 777          * Tools can chose to ignore this at their own discretion, but the
 778          * evlist layer can't just drop it when filtering events in
 779          * perf_evlist__filter_pollfd().
 780          */
 781         evlist->mmap[idx].refcnt = 2;
 782         evlist->mmap[idx].prev = 0;
 783         evlist->mmap[idx].mask = mp->mask;
 784         evlist->mmap[idx].base = mmap(NULL, evlist->mmap_len, mp->prot,
 785                                       MAP_SHARED, fd, 0);
 786         if (evlist->mmap[idx].base == MAP_FAILED) {
 787                 pr_debug2("failed to mmap perf event ring buffer, error %d\n",
 788                           errno);
 789                 evlist->mmap[idx].base = NULL;
 790                 return -1;
 791         }
 792
 793         return 0;
 794 }
 795
 796 static int perf_evlist__mmap_per_evsel(struct perf_evlist *evlist, int idx,
 797                                        struct mmap_params *mp, int cpu,
 798                                        int thread, int *output)
 799 {
 800         struct perf_evsel *evsel;
 801
 802         evlist__for_each(evlist, evsel) {
 803                 int fd;
 804
 805                 if (evsel->system_wide && thread)
 806                         continue;
 807
 808                 fd = FD(evsel, cpu, thread);
 809
 810                 if (*output == -1) {
 811                         *output = fd;
 812                         if (__perf_evlist__mmap(evlist, idx, mp, *output) < 0)
 813                                 return -1;
 814                 } else {
 815                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 816                                 return -1;
 817
 818                         perf_evlist__mmap_get(evlist, idx);
 819                 }
 820
 821                 /*
 822                  * The system_wide flag causes a selected event to be opened
 823                  * always without a pid.  Consequently it will never get a
 824                  * POLLHUP, but it is used for tracking in combination with
 825                  * other events, so it should not need to be polled anyway.
 826                  * Therefore don't add it for polling.
 827                  */
 828                 if (!evsel->system_wide &&
 829                     __perf_evlist__add_pollfd(evlist, fd, idx) < 0) {
 830                         perf_evlist__mmap_put(evlist, idx);
 831                         return -1;
 832                 }
 833
 834                 if (evsel->attr.read_format & PERF_FORMAT_ID) {
 835                         if (perf_evlist__id_add_fd(evlist, evsel, cpu, thread,
 836                                                    fd) < 0)
 837                                 return -1;
 838                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 839                                                  thread);
 840                 }
 841         }
 842
 843         return 0;
 844 }
 845
 846 static int perf_evlist__mmap_per_cpu(struct perf_evlist *evlist,
 847                                      struct mmap_params *mp)
 848 {
 849         int cpu, thread;
 850         int nr_cpus = cpu_map__nr(evlist->cpus);
 851         int nr_threads = thread_map__nr(evlist->threads);
 852
 853         pr_debug2("perf event ring buffer mmapped per cpu\n");
 854         for (cpu = 0; cpu < nr_cpus; cpu++) {
 855                 int output = -1;
 856
 857                 for (thread = 0; thread < nr_threads; thread++) {
 858                         if (perf_evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 859                                                         thread, &output))
 860                                 goto out_unmap;
 861                 }
 862         }
 863
 864         return 0;
 865
 866 out_unmap:
 867         for (cpu = 0; cpu < nr_cpus; cpu++)
 868                 __perf_evlist__munmap(evlist, cpu);
 869         return -1;
 870 }
 871
 872 static int perf_evlist__mmap_per_thread(struct perf_evlist *evlist,
 873                                         struct mmap_params *mp)
 874 {
 875         int thread;
 876         int nr_threads = thread_map__nr(evlist->threads);
 877
 878         pr_debug2("perf event ring buffer mmapped per thread\n");
 879         for (thread = 0; thread < nr_threads; thread++) {
 880                 int output = -1;
 881
 882                 if (perf_evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 883                                                 &output))
 884                         goto out_unmap;
 885         }
 886
 887         return 0;
 888
 889 out_unmap:
 890         for (thread = 0; thread < nr_threads; thread++)
 891                 __perf_evlist__munmap(evlist, thread);
 892         return -1;
 893 }
 894
 895 static size_t perf_evlist__mmap_size(unsigned long pages)
 896 {
 897         if (pages == UINT_MAX) {
 898                 int max;
 899
 900                 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 901                         /*
 902                          * Pick a once upon a time good value, i.e. things look
 903                          * strange since we can't read a sysctl value, but lets not
 904                          * die yet...
 905                          */
 906                         max = 512;
 907                 } else {
 908                         max -= (page_size / 1024);
 909                 }
 910
 911                 pages = (max * 1024) / page_size;
 912                 if (!is_power_of_2(pages))
 913                         pages = rounddown_pow_of_two(pages);
 914         } else if (!is_power_of_2(pages))
 915                 return 0;
 916
 917         return (pages + 1) * page_size;
 918 }
 919
 920 static long parse_pages_arg(const char *str, unsigned long min,
 921                             unsigned long max)
 922 {
 923         unsigned long pages, val;
 924         static struct parse_tag tags[] = {
 925                 { .tag  = 'B', .mult = 1       },
 926                 { .tag  = 'K', .mult = 1 << 10 },
 927                 { .tag  = 'M', .mult = 1 << 20 },
 928                 { .tag  = 'G', .mult = 1 << 30 },
 929                 { .tag  = 0 },
 930         };
 931
 932         if (str == NULL)
 933                 return -EINVAL;
 934
 935         val = parse_tag_value(str, tags);
 936         if (val != (unsigned long) -1) {
 937                 /* we got file size value */
 938                 pages = PERF_ALIGN(val, page_size) / page_size;
 939         } else {
 940                 /* we got pages count value */
 941                 char *eptr;
 942                 pages = strtoul(str, &eptr, 10);
 943                 if (*eptr != '\0')
 944                         return -EINVAL;
 945         }
 946
 947         if (pages == 0 && min == 0) {
 948                 /* leave number of pages at 0 */
 949         } else if (!is_power_of_2(pages)) {
 950                 /* round pages up to next power of 2 */
 951                 pages = roundup_pow_of_two(pages);
 952                 if (!pages)
 953                         return -EINVAL;
 954                 pr_info("rounding mmap pages size to %lu bytes (%lu pages)\n",
 955                         pages * page_size, pages);
 956         }
 957
 958         if (pages > max)
 959                 return -EINVAL;
 960
 961         return pages;
 962 }
 963
 964 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 965                                   int unset __maybe_unused)
 966 {
 967         unsigned int *mmap_pages = opt->value;
 968         unsigned long max = UINT_MAX;
 969         long pages;
 970
 971         if (max > SIZE_MAX / page_size)
 972                 max = SIZE_MAX / page_size;
 973
 974         pages = parse_pages_arg(str, 1, max);
 975         if (pages < 0) {
 976                 pr_err("Invalid argument for --mmap_pages/-m\n");
 977                 return -1;
 978         }
 979
 980         *mmap_pages = pages;
 981         return 0;
 982 }
 983
 984 /**
 985  * perf_evlist__mmap - Create mmaps to receive events.
 986  * @evlist: list of events
 987  * @pages: map length in pages
 988  * @overwrite: overwrite older events?
 989  *
 990  * If @overwrite is %false the user needs to signal event consumption using
 991  * perf_mmap__write_tail().  Using perf_evlist__mmap_read() does this
 992  * automatically.
 993  *
 994  * Return: %0 on success, negative error code otherwise.
 995  */
 996 int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,
 997                       bool overwrite)
 998 {
 999         struct perf_evsel *evsel;
1000         const struct cpu_map *cpus = evlist->cpus;
1001         const struct thread_map *threads = evlist->threads;
1002         struct mmap_params mp = {
1003                 .prot = PROT_READ | (overwrite ? 0 : PROT_WRITE),
1004         };
1005
1006         if (evlist->mmap == NULL && perf_evlist__alloc_mmap(evlist) < 0)
1007                 return -ENOMEM;
1008
1009         if (evlist->pollfd.entries == NULL && perf_evlist__alloc_pollfd(evlist) < 0)
1010                 return -ENOMEM;
1011
1012         evlist->overwrite = overwrite;
1013         evlist->mmap_len = perf_evlist__mmap_size(pages);
1014         pr_debug("mmap size %zuB\n", evlist->mmap_len);
1015         mp.mask = evlist->mmap_len - page_size - 1;
1016
1017         evlist__for_each(evlist, evsel) {
1018                 if ((evsel->attr.read_format & PERF_FORMAT_ID) &&
1019                     evsel->sample_id == NULL &&
1020                     perf_evsel__alloc_id(evsel, cpu_map__nr(cpus), threads->nr) < 0)
1021                         return -ENOMEM;
1022         }
1023
1024         if (cpu_map__empty(cpus))
1025                 return perf_evlist__mmap_per_thread(evlist, &mp);
1026
1027         return perf_evlist__mmap_per_cpu(evlist, &mp);
1028 }
1029
1030 int perf_evlist__create_maps(struct perf_evlist *evlist, struct target *target)
1031 {
1032         evlist->threads = thread_map__new_str(target->pid, target->tid,
1033                                               target->uid);
1034
1035         if (evlist->threads == NULL)
1036                 return -1;
1037
1038         if (target__uses_dummy_map(target))
1039                 evlist->cpus = cpu_map__dummy_new();
1040         else
1041                 evlist->cpus = cpu_map__new(target->cpu_list);
1042
1043         if (evlist->cpus == NULL)
1044                 goto out_delete_threads;
1045
1046         return 0;
1047
1048 out_delete_threads:
1049         thread_map__delete(evlist->threads);
1050         evlist->threads = NULL;
1051         return -1;
1052 }
1053
1054 int perf_evlist__apply_filters(struct perf_evlist *evlist)
1055 {
1056         struct perf_evsel *evsel;
1057         int err = 0;
1058         const int ncpus = cpu_map__nr(evlist->cpus),
1059                   nthreads = thread_map__nr(evlist->threads);
1060
1061         evlist__for_each(evlist, evsel) {
1062                 if (evsel->filter == NULL)
1063                         continue;
1064
1065                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, evsel->filter);
1066                 if (err)
1067                         break;
1068         }
1069
1070         return err;
1071 }
1072
1073 int perf_evlist__set_filter(struct perf_evlist *evlist, const char *filter)
1074 {
1075         struct perf_evsel *evsel;
1076         int err = 0;
1077         const int ncpus = cpu_map__nr(evlist->cpus),
1078                   nthreads = thread_map__nr(evlist->threads);
1079
1080         evlist__for_each(evlist, evsel) {
1081                 err = perf_evsel__set_filter(evsel, ncpus, nthreads, filter);
1082                 if (err)
1083                         break;
1084         }
1085
1086         return err;
1087 }
1088
1089 bool perf_evlist__valid_sample_type(struct perf_evlist *evlist)
1090 {
1091         struct perf_evsel *pos;
1092
1093         if (evlist->nr_entries == 1)
1094                 return true;
1095
1096         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1097                 return false;
1098
1099         evlist__for_each(evlist, pos) {
1100                 if (pos->id_pos != evlist->id_pos ||
1101                     pos->is_pos != evlist->is_pos)
1102                         return false;
1103         }
1104
1105         return true;
1106 }
1107
1108 u64 __perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1109 {
1110         struct perf_evsel *evsel;
1111
1112         if (evlist->combined_sample_type)
1113                 return evlist->combined_sample_type;
1114
1115         evlist__for_each(evlist, evsel)
1116                 evlist->combined_sample_type |= evsel->attr.sample_type;
1117
1118         return evlist->combined_sample_type;
1119 }
1120
1121 u64 perf_evlist__combined_sample_type(struct perf_evlist *evlist)
1122 {
1123         evlist->combined_sample_type = 0;
1124         return __perf_evlist__combined_sample_type(evlist);
1125 }
1126
1127 bool perf_evlist__valid_read_format(struct perf_evlist *evlist)
1128 {
1129         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1130         u64 read_format = first->attr.read_format;
1131         u64 sample_type = first->attr.sample_type;
1132
1133         evlist__for_each(evlist, pos) {
1134                 if (read_format != pos->attr.read_format)
1135                         return false;
1136         }
1137
1138         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1139         if ((sample_type & PERF_SAMPLE_READ) &&
1140             !(read_format & PERF_FORMAT_ID)) {
1141                 return false;
1142         }
1143
1144         return true;
1145 }
1146
1147 u64 perf_evlist__read_format(struct perf_evlist *evlist)
1148 {
1149         struct perf_evsel *first = perf_evlist__first(evlist);
1150         return first->attr.read_format;
1151 }
1152
1153 u16 perf_evlist__id_hdr_size(struct perf_evlist *evlist)
1154 {
1155         struct perf_evsel *first = perf_evlist__first(evlist);
1156         struct perf_sample *data;
1157         u64 sample_type;
1158         u16 size = 0;
1159
1160         if (!first->attr.sample_id_all)
1161                 goto out;
1162
1163         sample_type = first->attr.sample_type;
1164
1165         if (sample_type & PERF_SAMPLE_TID)
1166                 size += sizeof(data->tid) * 2;
1167
1168        if (sample_type & PERF_SAMPLE_TIME)
1169                 size += sizeof(data->time);
1170
1171         if (sample_type & PERF_SAMPLE_ID)
1172                 size += sizeof(data->id);
1173
1174         if (sample_type & PERF_SAMPLE_STREAM_ID)
1175                 size += sizeof(data->stream_id);
1176
1177         if (sample_type & PERF_SAMPLE_CPU)
1178                 size += sizeof(data->cpu) * 2;
1179
1180         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1181                 size += sizeof(data->id);
1182 out:
1183         return size;
1184 }
1185
1186 bool perf_evlist__valid_sample_id_all(struct perf_evlist *evlist)
1187 {
1188         struct perf_evsel *first = perf_evlist__first(evlist), *pos = first;
1189
1190         evlist__for_each_continue(evlist, pos) {
1191                 if (first->attr.sample_id_all != pos->attr.sample_id_all)
1192                         return false;
1193         }
1194
1195         return true;
1196 }
1197
1198 bool perf_evlist__sample_id_all(struct perf_evlist *evlist)
1199 {
1200         struct perf_evsel *first = perf_evlist__first(evlist);
1201         return first->attr.sample_id_all;
1202 }
1203
1204 void perf_evlist__set_selected(struct perf_evlist *evlist,
1205                                struct perf_evsel *evsel)
1206 {
1207         evlist->selected = evsel;
1208 }
1209
1210 void perf_evlist__close(struct perf_evlist *evlist)
1211 {
1212         struct perf_evsel *evsel;
1213         int ncpus = cpu_map__nr(evlist->cpus);
1214         int nthreads = thread_map__nr(evlist->threads);
1215         int n;
1216
1217         evlist__for_each_reverse(evlist, evsel) {
1218                 n = evsel->cpus ? evsel->cpus->nr : ncpus;
1219                 perf_evsel__close(evsel, n, nthreads);
1220         }
1221 }
1222
1223 static int perf_evlist__create_syswide_maps(struct perf_evlist *evlist)
1224 {
1225         int err = -ENOMEM;
1226
1227         /*
1228          * Try reading /sys/devices/system/cpu/online to get
1229          * an all cpus map.
1230          *
1231          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1232          * code needs an overhaul to properly forward the
1233          * error, and we may not want to do that fallback to a
1234          * default cpu identity map :-\
1235          */
1236         evlist->cpus = cpu_map__new(NULL);
1237         if (evlist->cpus == NULL)
1238                 goto out;
1239
1240         evlist->threads = thread_map__new_dummy();
1241         if (evlist->threads == NULL)
1242                 goto out_free_cpus;
1243
1244         err = 0;
1245 out:
1246         return err;
1247 out_free_cpus:
1248         cpu_map__delete(evlist->cpus);
1249         evlist->cpus = NULL;
1250         goto out;
1251 }
1252
1253 int perf_evlist__open(struct perf_evlist *evlist)
1254 {
1255         struct perf_evsel *evsel;
1256         int err;
1257
1258         /*
1259          * Default: one fd per CPU, all threads, aka systemwide
1260          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1261          */
1262         if (evlist->threads == NULL && evlist->cpus == NULL) {
1263                 err = perf_evlist__create_syswide_maps(evlist);
1264                 if (err < 0)
1265                         goto out_err;
1266         }
1267
1268         perf_evlist__update_id_pos(evlist);
1269
1270         evlist__for_each(evlist, evsel) {
1271                 err = perf_evsel__open(evsel, evlist->cpus, evlist->threads);
1272                 if (err < 0)
1273                         goto out_err;
1274         }
1275
1276         return 0;
1277 out_err:
1278         perf_evlist__close(evlist);
1279         errno = -err;
1280         return err;
1281 }
1282
1283 int perf_evlist__prepare_workload(struct perf_evlist *evlist, struct target *target,
1284                                   const char *argv[], bool pipe_output,
1285                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1286 {
1287         int child_ready_pipe[2], go_pipe[2];
1288         char bf;
1289
1290         if (pipe(child_ready_pipe) < 0) {
1291                 perror("failed to create 'ready' pipe");
1292                 return -1;
1293         }
1294
1295         if (pipe(go_pipe) < 0) {
1296                 perror("failed to create 'go' pipe");
1297                 goto out_close_ready_pipe;
1298         }
1299
1300         evlist->workload.pid = fork();
1301         if (evlist->workload.pid < 0) {
1302                 perror("failed to fork");
1303                 goto out_close_pipes;
1304         }
1305
1306         if (!evlist->workload.pid) {
1307                 int ret;
1308
1309                 if (pipe_output)
1310                         dup2(2, 1);
1311
1312                 signal(SIGTERM, SIG_DFL);
1313
1314                 close(child_ready_pipe[0]);
1315                 close(go_pipe[1]);
1316                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1317
1318                 /*
1319                  * Tell the parent we're ready to go
1320                  */
1321                 close(child_ready_pipe[1]);
1322
1323                 /*
1324                  * Wait until the parent tells us to go.
1325                  */
1326                 ret = read(go_pipe[0], &bf, 1);
1327                 /*
1328                  * The parent will ask for the execvp() to be performed by
1329                  * writing exactly one byte, in workload.cork_fd, usually via
1330                  * perf_evlist__start_workload().
1331                  *
1332                  * For cancelling the workload without actuallin running it,
1333                  * the parent will just close workload.cork_fd, without writing
1334                  * anything, i.e. read will return zero and we just exit()
1335                  * here.
1336                  */
1337                 if (ret != 1) {
1338                         if (ret == -1)
1339                                 perror("unable to read pipe");
1340                         exit(ret);
1341                 }
1342
1343                 execvp(argv[0], (char **)argv);
1344
1345                 if (exec_error) {
1346                         union sigval val;
1347
1348                         val.sival_int = errno;
1349                         if (sigqueue(getppid(), SIGUSR1, val))
1350                                 perror(argv[0]);
1351                 } else
1352                         perror(argv[0]);
1353                 exit(-1);
1354         }
1355
1356         if (exec_error) {
1357                 struct sigaction act = {
1358                         .sa_flags     = SA_SIGINFO,
1359                         .sa_sigaction = exec_error,
1360                 };
1361                 sigaction(SIGUSR1, &act, NULL);
1362         }
1363
1364         if (target__none(target)) {
1365                 if (evlist->threads == NULL) {
1366                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1367                                 __func__, __LINE__);
1368                         goto out_close_pipes;
1369                 }
1370                 evlist->threads->map[0] = evlist->workload.pid;
1371         }
1372
1373         close(child_ready_pipe[1]);
1374         close(go_pipe[0]);
1375         /*
1376          * wait for child to settle
1377          */
1378         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1379                 perror("unable to read pipe");
1380                 goto out_close_pipes;
1381         }
1382
1383         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1384         evlist->workload.cork_fd = go_pipe[1];
1385         close(child_ready_pipe[0]);
1386         return 0;
1387
1388 out_close_pipes:
1389         close(go_pipe[0]);
1390         close(go_pipe[1]);
1391 out_close_ready_pipe:
1392         close(child_ready_pipe[0]);
1393         close(child_ready_pipe[1]);
1394         return -1;
1395 }
1396
1397 int perf_evlist__start_workload(struct perf_evlist *evlist)
1398 {
1399         if (evlist->workload.cork_fd > 0) {
1400                 char bf = 0;
1401                 int ret;
1402                 /*
1403                  * Remove the cork, let it rip!
1404                  */
1405                 ret = write(evlist->workload.cork_fd, &bf, 1);
1406                 if (ret < 0)
1407                         perror("enable to write to pipe");
1408
1409                 close(evlist->workload.cork_fd);
1410                 return ret;
1411         }
1412
1413         return 0;
1414 }
1415
1416 int perf_evlist__parse_sample(struct perf_evlist *evlist, union perf_event *event,
1417                               struct perf_sample *sample)
1418 {
1419         struct perf_evsel *evsel = perf_evlist__event2evsel(evlist, event);
1420
1421         if (!evsel)
1422                 return -EFAULT;
1423         return perf_evsel__parse_sample(evsel, event, sample);
1424 }
1425
1426 size_t perf_evlist__fprintf(struct perf_evlist *evlist, FILE *fp)
1427 {
1428         struct perf_evsel *evsel;
1429         size_t printed = 0;
1430
1431         evlist__for_each(evlist, evsel) {
1432                 printed += fprintf(fp, "%s%s", evsel->idx ? ", " : "",
1433                                    perf_evsel__name(evsel));
1434         }
1435
1436         return printed + fprintf(fp, "\n");
1437 }
1438
1439 int perf_evlist__strerror_open(struct perf_evlist *evlist __maybe_unused,
1440                                int err, char *buf, size_t size)
1441 {
1442         int printed, value;
1443         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1444
1445         switch (err) {
1446         case EACCES:
1447         case EPERM:
1448                 printed = scnprintf(buf, size,
1449                                     "Error:\t%s.\n"
1450                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1451
1452                 value = perf_event_paranoid();
1453
1454                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1455
1456                 if (value >= 2) {
1457                         printed += scnprintf(buf + printed, size - printed,
1458                                              "For your workloads it needs to be <= 1\nHint:\t");
1459                 }
1460                 printed += scnprintf(buf + printed, size - printed,
1461                                      "For system wide tracing it needs to be set to -1.\n");
1462
1463                 printed += scnprintf(buf + printed, size - printed,
1464                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1465                                     "Hint:\tThe current value is %d.", value);
1466                 break;
1467         default:
1468                 scnprintf(buf, size, "%s", emsg);
1469                 break;
1470         }
1471
1472         return 0;
1473 }
1474
1475 int perf_evlist__strerror_mmap(struct perf_evlist *evlist, int err, char *buf, size_t size)
1476 {
1477         char sbuf[STRERR_BUFSIZE], *emsg = strerror_r(err, sbuf, sizeof(sbuf));
1478         int pages_attempted = evlist->mmap_len / 1024, pages_max_per_user, printed = 0;
1479
1480         switch (err) {
1481         case EPERM:
1482                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1483                 printed += scnprintf(buf + printed, size - printed,
1484                                      "Error:\t%s.\n"
1485                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1486                                      "Hint:\tTried using %zd kB.\n",
1487                                      emsg, pages_max_per_user, pages_attempted);
1488
1489                 if (pages_attempted >= pages_max_per_user) {
1490                         printed += scnprintf(buf + printed, size - printed,
1491                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1492                                              pages_max_per_user + pages_attempted);
1493                 }
1494
1495                 printed += scnprintf(buf + printed, size - printed,
1496                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1497                 break;
1498         default:
1499                 scnprintf(buf, size, "%s", emsg);
1500                 break;
1501         }
1502
1503         return 0;
1504 }
1505
1506 void perf_evlist__to_front(struct perf_evlist *evlist,
1507                            struct perf_evsel *move_evsel)
1508 {
1509         struct perf_evsel *evsel, *n;
1510         LIST_HEAD(move);
1511
1512         if (move_evsel == perf_evlist__first(evlist))
1513                 return;
1514
1515         evlist__for_each_safe(evlist, n, evsel) {
1516                 if (evsel->leader == move_evsel->leader)
1517                         list_move_tail(&evsel->node, &move);
1518         }
1519
1520         list_splice(&move, &evlist->entries);
1521 }
1522
1523 void perf_evlist__set_tracking_event(struct perf_evlist *evlist,
1524                                      struct perf_evsel *tracking_evsel)
1525 {
1526         struct perf_evsel *evsel;
1527
1528         if (tracking_evsel->tracking)
1529                 return;
1530
1531         evlist__for_each(evlist, evsel) {
1532                 if (evsel != tracking_evsel)
1533                         evsel->tracking = false;
1534         }
1535
1536         tracking_evsel->tracking = true;
1537 }