tools/perf/util/evlist.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
   4  *
   5  * Parts came from builtin-{top,stat,record}.c, see those files for further
   6  * copyright notes.
   7  */
   8 #include <api/fs/fs.h>
   9 #include <errno.h>
  10 #include <inttypes.h>
  11 #include <poll.h>
  12 #include "cpumap.h"
  13 #include "util/mmap.h"
  14 #include "thread_map.h"
  15 #include "target.h"
  16 #include "evlist.h"
  17 #include "evsel.h"
  18 #include "debug.h"
  19 #include "units.h"
  20 #include <internal/lib.h> // page_size
  21 #include "affinity.h"
  22 #include "../perf.h"
  23 #include "asm/bug.h"
  24 #include "bpf-event.h"
  25 #include "util/string2.h"
  26 #include <signal.h>
  27 #include <unistd.h>
  28 #include <sched.h>
  29 #include <stdlib.h>
  30
  31 #include "parse-events.h"
  32 #include <subcmd/parse-options.h>
  33
  34 #include <fcntl.h>
  35 #include <sys/ioctl.h>
  36 #include <sys/mman.h>
  37
  38 #include <linux/bitops.h>
  39 #include <linux/hash.h>
  40 #include <linux/log2.h>
  41 #include <linux/err.h>
  42 #include <linux/string.h>
  43 #include <linux/zalloc.h>
  44 #include <perf/evlist.h>
  45 #include <perf/evsel.h>
  46 #include <perf/cpumap.h>
  47 #include <perf/mmap.h>
  48
  49 #include <internal/xyarray.h>
  50
  51 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  52 int sigqueue(pid_t pid, int sig, const union sigval value);
  53 #endif
  54
  55 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  56 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  57
  58 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  59                   struct perf_thread_map *threads)
  60 {
  61         perf_evlist__init(&evlist->core);
  62         perf_evlist__set_maps(&evlist->core, cpus, threads);
  63         evlist->workload.pid = -1;
  64         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  65 }
  66
  67 struct evlist *evlist__new(void)
  68 {
  69         struct evlist *evlist = zalloc(sizeof(*evlist));
  70
  71         if (evlist != NULL)
  72                 evlist__init(evlist, NULL, NULL);
  73
  74         return evlist;
  75 }
  76
  77 struct evlist *perf_evlist__new_default(void)
  78 {
  79         struct evlist *evlist = evlist__new();
  80
  81         if (evlist && perf_evlist__add_default(evlist)) {
  82                 evlist__delete(evlist);
  83                 evlist = NULL;
  84         }
  85
  86         return evlist;
  87 }
  88
  89 struct evlist *perf_evlist__new_dummy(void)
  90 {
  91         struct evlist *evlist = evlist__new();
  92
  93         if (evlist && perf_evlist__add_dummy(evlist)) {
  94                 evlist__delete(evlist);
  95                 evlist = NULL;
  96         }
  97
  98         return evlist;
  99 }
 100
 101 /**
 102  * perf_evlist__set_id_pos - set the positions of event ids.
 103  * @evlist: selected event list
 104  *
 105  * Events with compatible sample types all have the same id_pos
 106  * and is_pos.  For convenience, put a copy on evlist.
 107  */
 108 void perf_evlist__set_id_pos(struct evlist *evlist)
 109 {
 110         struct evsel *first = evlist__first(evlist);
 111
 112         evlist->id_pos = first->id_pos;
 113         evlist->is_pos = first->is_pos;
 114 }
 115
 116 static void perf_evlist__update_id_pos(struct evlist *evlist)
 117 {
 118         struct evsel *evsel;
 119
 120         evlist__for_each_entry(evlist, evsel)
 121                 perf_evsel__calc_id_pos(evsel);
 122
 123         perf_evlist__set_id_pos(evlist);
 124 }
 125
 126 static void evlist__purge(struct evlist *evlist)
 127 {
 128         struct evsel *pos, *n;
 129
 130         evlist__for_each_entry_safe(evlist, n, pos) {
 131                 list_del_init(&pos->core.node);
 132                 pos->evlist = NULL;
 133                 evsel__delete(pos);
 134         }
 135
 136         evlist->core.nr_entries = 0;
 137 }
 138
 139 void evlist__exit(struct evlist *evlist)
 140 {
 141         zfree(&evlist->mmap);
 142         zfree(&evlist->overwrite_mmap);
 143         perf_evlist__exit(&evlist->core);
 144 }
 145
 146 void evlist__delete(struct evlist *evlist)
 147 {
 148         if (evlist == NULL)
 149                 return;
 150
 151         evlist__munmap(evlist);
 152         evlist__close(evlist);
 153         evlist__purge(evlist);
 154         evlist__exit(evlist);
 155         free(evlist);
 156 }
 157
 158 void evlist__add(struct evlist *evlist, struct evsel *entry)
 159 {
 160         entry->evlist = evlist;
 161         entry->idx = evlist->core.nr_entries;
 162         entry->tracking = !entry->idx;
 163
 164         perf_evlist__add(&evlist->core, &entry->core);
 165
 166         if (evlist->core.nr_entries == 1)
 167                 perf_evlist__set_id_pos(evlist);
 168 }
 169
 170 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 171 {
 172         evsel->evlist = NULL;
 173         perf_evlist__remove(&evlist->core, &evsel->core);
 174 }
 175
 176 void perf_evlist__splice_list_tail(struct evlist *evlist,
 177                                    struct list_head *list)
 178 {
 179         struct evsel *evsel, *temp;
 180
 181         __evlist__for_each_entry_safe(list, temp, evsel) {
 182                 list_del_init(&evsel->core.node);
 183                 evlist__add(evlist, evsel);
 184         }
 185 }
 186
 187 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 188                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 189 {
 190         struct evsel *evsel;
 191         size_t i;
 192         int err;
 193
 194         for (i = 0; i < nr_assocs; i++) {
 195                 // Adding a handler for an event not in this evlist, just ignore it.
 196                 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 197                 if (evsel == NULL)
 198                         continue;
 199
 200                 err = -EEXIST;
 201                 if (evsel->handler != NULL)
 202                         goto out;
 203                 evsel->handler = assocs[i].handler;
 204         }
 205
 206         err = 0;
 207 out:
 208         return err;
 209 }
 210
 211 void __perf_evlist__set_leader(struct list_head *list)
 212 {
 213         struct evsel *evsel, *leader;
 214
 215         leader = list_entry(list->next, struct evsel, core.node);
 216         evsel = list_entry(list->prev, struct evsel, core.node);
 217
 218         leader->core.nr_members = evsel->idx - leader->idx + 1;
 219
 220         __evlist__for_each_entry(list, evsel) {
 221                 evsel->leader = leader;
 222         }
 223 }
 224
 225 void perf_evlist__set_leader(struct evlist *evlist)
 226 {
 227         if (evlist->core.nr_entries) {
 228                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 229                 __perf_evlist__set_leader(&evlist->core.entries);
 230         }
 231 }
 232
 233 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
 234 {
 235         struct evsel *evsel = perf_evsel__new_cycles(precise);
 236
 237         if (evsel == NULL)
 238                 return -ENOMEM;
 239
 240         evlist__add(evlist, evsel);
 241         return 0;
 242 }
 243
 244 int perf_evlist__add_dummy(struct evlist *evlist)
 245 {
 246         struct perf_event_attr attr = {
 247                 .type   = PERF_TYPE_SOFTWARE,
 248                 .config = PERF_COUNT_SW_DUMMY,
 249                 .size   = sizeof(attr), /* to capture ABI version */
 250         };
 251         struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
 252
 253         if (evsel == NULL)
 254                 return -ENOMEM;
 255
 256         evlist__add(evlist, evsel);
 257         return 0;
 258 }
 259
 260 static int evlist__add_attrs(struct evlist *evlist,
 261                                   struct perf_event_attr *attrs, size_t nr_attrs)
 262 {
 263         struct evsel *evsel, *n;
 264         LIST_HEAD(head);
 265         size_t i;
 266
 267         for (i = 0; i < nr_attrs; i++) {
 268                 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 269                 if (evsel == NULL)
 270                         goto out_delete_partial_list;
 271                 list_add_tail(&evsel->core.node, &head);
 272         }
 273
 274         perf_evlist__splice_list_tail(evlist, &head);
 275
 276         return 0;
 277
 278 out_delete_partial_list:
 279         __evlist__for_each_entry_safe(&head, n, evsel)
 280                 evsel__delete(evsel);
 281         return -1;
 282 }
 283
 284 int __perf_evlist__add_default_attrs(struct evlist *evlist,
 285                                      struct perf_event_attr *attrs, size_t nr_attrs)
 286 {
 287         size_t i;
 288
 289         for (i = 0; i < nr_attrs; i++)
 290                 event_attr_init(attrs + i);
 291
 292         return evlist__add_attrs(evlist, attrs, nr_attrs);
 293 }
 294
 295 struct evsel *
 296 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 297 {
 298         struct evsel *evsel;
 299
 300         evlist__for_each_entry(evlist, evsel) {
 301                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 302                     (int)evsel->core.attr.config == id)
 303                         return evsel;
 304         }
 305
 306         return NULL;
 307 }
 308
 309 struct evsel *
 310 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
 311                                      const char *name)
 312 {
 313         struct evsel *evsel;
 314
 315         evlist__for_each_entry(evlist, evsel) {
 316                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 317                     (strcmp(evsel->name, name) == 0))
 318                         return evsel;
 319         }
 320
 321         return NULL;
 322 }
 323
 324 int perf_evlist__add_newtp(struct evlist *evlist,
 325                            const char *sys, const char *name, void *handler)
 326 {
 327         struct evsel *evsel = perf_evsel__newtp(sys, name);
 328
 329         if (IS_ERR(evsel))
 330                 return -1;
 331
 332         evsel->handler = handler;
 333         evlist__add(evlist, evsel);
 334         return 0;
 335 }
 336
 337 static int perf_evlist__nr_threads(struct evlist *evlist,
 338                                    struct evsel *evsel)
 339 {
 340         if (evsel->core.system_wide)
 341                 return 1;
 342         else
 343                 return perf_thread_map__nr(evlist->core.threads);
 344 }
 345
 346 void evlist__cpu_iter_start(struct evlist *evlist)
 347 {
 348         struct evsel *pos;
 349
 350         /*
 351          * Reset the per evsel cpu_iter. This is needed because
 352          * each evsel's cpumap may have a different index space,
 353          * and some operations need the index to modify
 354          * the FD xyarray (e.g. open, close)
 355          */
 356         evlist__for_each_entry(evlist, pos)
 357                 pos->cpu_iter = 0;
 358 }
 359
 360 bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
 361 {
 362         if (ev->cpu_iter >= ev->core.cpus->nr)
 363                 return true;
 364         if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
 365                 return true;
 366         return false;
 367 }
 368
 369 bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
 370 {
 371         if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
 372                 ev->cpu_iter++;
 373                 return false;
 374         }
 375         return true;
 376 }
 377
 378 void evlist__disable(struct evlist *evlist)
 379 {
 380         struct evsel *pos;
 381         struct affinity affinity;
 382         int cpu, i;
 383
 384         if (affinity__setup(&affinity) < 0)
 385                 return;
 386
 387         evlist__for_each_cpu(evlist, i, cpu) {
 388                 affinity__set(&affinity, cpu);
 389
 390                 evlist__for_each_entry(evlist, pos) {
 391                         if (evsel__cpu_iter_skip(pos, cpu))
 392                                 continue;
 393                         if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
 394                                 continue;
 395                         evsel__disable_cpu(pos, pos->cpu_iter - 1);
 396                 }
 397         }
 398         affinity__cleanup(&affinity);
 399         evlist__for_each_entry(evlist, pos) {
 400                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 401                         continue;
 402                 pos->disabled = true;
 403         }
 404
 405         evlist->enabled = false;
 406 }
 407
 408 void evlist__enable(struct evlist *evlist)
 409 {
 410         struct evsel *pos;
 411         struct affinity affinity;
 412         int cpu, i;
 413
 414         if (affinity__setup(&affinity) < 0)
 415                 return;
 416
 417         evlist__for_each_cpu(evlist, i, cpu) {
 418                 affinity__set(&affinity, cpu);
 419
 420                 evlist__for_each_entry(evlist, pos) {
 421                         if (evsel__cpu_iter_skip(pos, cpu))
 422                                 continue;
 423                         if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 424                                 continue;
 425                         evsel__enable_cpu(pos, pos->cpu_iter - 1);
 426                 }
 427         }
 428         affinity__cleanup(&affinity);
 429         evlist__for_each_entry(evlist, pos) {
 430                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 431                         continue;
 432                 pos->disabled = false;
 433         }
 434
 435         evlist->enabled = true;
 436 }
 437
 438 void perf_evlist__toggle_enable(struct evlist *evlist)
 439 {
 440         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 441 }
 442
 443 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
 444                                          struct evsel *evsel, int cpu)
 445 {
 446         int thread;
 447         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 448
 449         if (!evsel->core.fd)
 450                 return -EINVAL;
 451
 452         for (thread = 0; thread < nr_threads; thread++) {
 453                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 454                 if (err)
 455                         return err;
 456         }
 457         return 0;
 458 }
 459
 460 static int perf_evlist__enable_event_thread(struct evlist *evlist,
 461                                             struct evsel *evsel,
 462                                             int thread)
 463 {
 464         int cpu;
 465         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 466
 467         if (!evsel->core.fd)
 468                 return -EINVAL;
 469
 470         for (cpu = 0; cpu < nr_cpus; cpu++) {
 471                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 472                 if (err)
 473                         return err;
 474         }
 475         return 0;
 476 }
 477
 478 int perf_evlist__enable_event_idx(struct evlist *evlist,
 479                                   struct evsel *evsel, int idx)
 480 {
 481         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 482
 483         if (per_cpu_mmaps)
 484                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 485         else
 486                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 487 }
 488
 489 int evlist__add_pollfd(struct evlist *evlist, int fd)
 490 {
 491         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
 492 }
 493
 494 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 495 {
 496         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 497 }
 498
 499 int evlist__poll(struct evlist *evlist, int timeout)
 500 {
 501         return perf_evlist__poll(&evlist->core, timeout);
 502 }
 503
 504 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
 505 {
 506         struct hlist_head *head;
 507         struct perf_sample_id *sid;
 508         int hash;
 509
 510         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 511         head = &evlist->core.heads[hash];
 512
 513         hlist_for_each_entry(sid, head, node)
 514                 if (sid->id == id)
 515                         return sid;
 516
 517         return NULL;
 518 }
 519
 520 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
 521 {
 522         struct perf_sample_id *sid;
 523
 524         if (evlist->core.nr_entries == 1 || !id)
 525                 return evlist__first(evlist);
 526
 527         sid = perf_evlist__id2sid(evlist, id);
 528         if (sid)
 529                 return container_of(sid->evsel, struct evsel, core);
 530
 531         if (!perf_evlist__sample_id_all(evlist))
 532                 return evlist__first(evlist);
 533
 534         return NULL;
 535 }
 536
 537 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
 538                                                 u64 id)
 539 {
 540         struct perf_sample_id *sid;
 541
 542         if (!id)
 543                 return NULL;
 544
 545         sid = perf_evlist__id2sid(evlist, id);
 546         if (sid)
 547                 return container_of(sid->evsel, struct evsel, core);
 548
 549         return NULL;
 550 }
 551
 552 static int perf_evlist__event2id(struct evlist *evlist,
 553                                  union perf_event *event, u64 *id)
 554 {
 555         const __u64 *array = event->sample.array;
 556         ssize_t n;
 557
 558         n = (event->header.size - sizeof(event->header)) >> 3;
 559
 560         if (event->header.type == PERF_RECORD_SAMPLE) {
 561                 if (evlist->id_pos >= n)
 562                         return -1;
 563                 *id = array[evlist->id_pos];
 564         } else {
 565                 if (evlist->is_pos > n)
 566                         return -1;
 567                 n -= evlist->is_pos;
 568                 *id = array[n];
 569         }
 570         return 0;
 571 }
 572
 573 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
 574                                             union perf_event *event)
 575 {
 576         struct evsel *first = evlist__first(evlist);
 577         struct hlist_head *head;
 578         struct perf_sample_id *sid;
 579         int hash;
 580         u64 id;
 581
 582         if (evlist->core.nr_entries == 1)
 583                 return first;
 584
 585         if (!first->core.attr.sample_id_all &&
 586             event->header.type != PERF_RECORD_SAMPLE)
 587                 return first;
 588
 589         if (perf_evlist__event2id(evlist, event, &id))
 590                 return NULL;
 591
 592         /* Synthesized events have an id of zero */
 593         if (!id)
 594                 return first;
 595
 596         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 597         head = &evlist->core.heads[hash];
 598
 599         hlist_for_each_entry(sid, head, node) {
 600                 if (sid->id == id)
 601                         return container_of(sid->evsel, struct evsel, core);
 602         }
 603         return NULL;
 604 }
 605
 606 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
 607 {
 608         int i;
 609
 610         if (!evlist->overwrite_mmap)
 611                 return 0;
 612
 613         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 614                 int fd = evlist->overwrite_mmap[i].core.fd;
 615                 int err;
 616
 617                 if (fd < 0)
 618                         continue;
 619                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 620                 if (err)
 621                         return err;
 622         }
 623         return 0;
 624 }
 625
 626 static int perf_evlist__pause(struct evlist *evlist)
 627 {
 628         return perf_evlist__set_paused(evlist, true);
 629 }
 630
 631 static int perf_evlist__resume(struct evlist *evlist)
 632 {
 633         return perf_evlist__set_paused(evlist, false);
 634 }
 635
 636 static void evlist__munmap_nofree(struct evlist *evlist)
 637 {
 638         int i;
 639
 640         if (evlist->mmap)
 641                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 642                         perf_mmap__munmap(&evlist->mmap[i].core);
 643
 644         if (evlist->overwrite_mmap)
 645                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 646                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 647 }
 648
 649 void evlist__munmap(struct evlist *evlist)
 650 {
 651         evlist__munmap_nofree(evlist);
 652         zfree(&evlist->mmap);
 653         zfree(&evlist->overwrite_mmap);
 654 }
 655
 656 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 657 {
 658         struct mmap *m = container_of(map, struct mmap, core);
 659
 660         mmap__munmap(m);
 661 }
 662
 663 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 664                                        bool overwrite)
 665 {
 666         int i;
 667         struct mmap *map;
 668
 669         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 670         if (!map)
 671                 return NULL;
 672
 673         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 674                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 675
 676                 /*
 677                  * When the perf_mmap() call is made we grab one refcount, plus
 678                  * one extra to let perf_mmap__consume() get the last
 679                  * events after all real references (perf_mmap__get()) are
 680                  * dropped.
 681                  *
 682                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 683                  * thus does perf_mmap__get() on it.
 684                  */
 685                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 686         }
 687
 688         return map;
 689 }
 690
 691 static void
 692 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 693                          struct perf_mmap_param *_mp,
 694                          int idx, bool per_cpu)
 695 {
 696         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 697         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 698
 699         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
 700 }
 701
 702 static struct perf_mmap*
 703 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 704 {
 705         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 706         struct mmap *maps;
 707
 708         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 709
 710         if (!maps) {
 711                 maps = evlist__alloc_mmap(evlist, overwrite);
 712                 if (!maps)
 713                         return NULL;
 714
 715                 if (overwrite) {
 716                         evlist->overwrite_mmap = maps;
 717                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 718                                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 719                 } else {
 720                         evlist->mmap = maps;
 721                 }
 722         }
 723
 724         return &maps[idx].core;
 725 }
 726
 727 static int
 728 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 729                           int output, int cpu)
 730 {
 731         struct mmap *map = container_of(_map, struct mmap, core);
 732         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 733
 734         return mmap__mmap(map, mp, output, cpu);
 735 }
 736
 737 unsigned long perf_event_mlock_kb_in_pages(void)
 738 {
 739         unsigned long pages;
 740         int max;
 741
 742         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 743                 /*
 744                  * Pick a once upon a time good value, i.e. things look
 745                  * strange since we can't read a sysctl value, but lets not
 746                  * die yet...
 747                  */
 748                 max = 512;
 749         } else {
 750                 max -= (page_size / 1024);
 751         }
 752
 753         pages = (max * 1024) / page_size;
 754         if (!is_power_of_2(pages))
 755                 pages = rounddown_pow_of_two(pages);
 756
 757         return pages;
 758 }
 759
 760 size_t evlist__mmap_size(unsigned long pages)
 761 {
 762         if (pages == UINT_MAX)
 763                 pages = perf_event_mlock_kb_in_pages();
 764         else if (!is_power_of_2(pages))
 765                 return 0;
 766
 767         return (pages + 1) * page_size;
 768 }
 769
 770 static long parse_pages_arg(const char *str, unsigned long min,
 771                             unsigned long max)
 772 {
 773         unsigned long pages, val;
 774         static struct parse_tag tags[] = {
 775                 { .tag  = 'B', .mult = 1       },
 776                 { .tag  = 'K', .mult = 1 << 10 },
 777                 { .tag  = 'M', .mult = 1 << 20 },
 778                 { .tag  = 'G', .mult = 1 << 30 },
 779                 { .tag  = 0 },
 780         };
 781
 782         if (str == NULL)
 783                 return -EINVAL;
 784
 785         val = parse_tag_value(str, tags);
 786         if (val != (unsigned long) -1) {
 787                 /* we got file size value */
 788                 pages = PERF_ALIGN(val, page_size) / page_size;
 789         } else {
 790                 /* we got pages count value */
 791                 char *eptr;
 792                 pages = strtoul(str, &eptr, 10);
 793                 if (*eptr != '\0')
 794                         return -EINVAL;
 795         }
 796
 797         if (pages == 0 && min == 0) {
 798                 /* leave number of pages at 0 */
 799         } else if (!is_power_of_2(pages)) {
 800                 char buf[100];
 801
 802                 /* round pages up to next power of 2 */
 803                 pages = roundup_pow_of_two(pages);
 804                 if (!pages)
 805                         return -EINVAL;
 806
 807                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 808                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 809                         buf, pages);
 810         }
 811
 812         if (pages > max)
 813                 return -EINVAL;
 814
 815         return pages;
 816 }
 817
 818 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 819 {
 820         unsigned long max = UINT_MAX;
 821         long pages;
 822
 823         if (max > SIZE_MAX / page_size)
 824                 max = SIZE_MAX / page_size;
 825
 826         pages = parse_pages_arg(str, 1, max);
 827         if (pages < 0) {
 828                 pr_err("Invalid argument for --mmap_pages/-m\n");
 829                 return -1;
 830         }
 831
 832         *mmap_pages = pages;
 833         return 0;
 834 }
 835
 836 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 837                                   int unset __maybe_unused)
 838 {
 839         return __perf_evlist__parse_mmap_pages(opt->value, str);
 840 }
 841
 842 /**
 843  * evlist__mmap_ex - Create mmaps to receive events.
 844  * @evlist: list of events
 845  * @pages: map length in pages
 846  * @overwrite: overwrite older events?
 847  * @auxtrace_pages - auxtrace map length in pages
 848  * @auxtrace_overwrite - overwrite older auxtrace data?
 849  *
 850  * If @overwrite is %false the user needs to signal event consumption using
 851  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 852  * automatically.
 853  *
 854  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 855  * consumption using auxtrace_mmap__write_tail().
 856  *
 857  * Return: %0 on success, negative error code otherwise.
 858  */
 859 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 860                          unsigned int auxtrace_pages,
 861                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 862                          int comp_level)
 863 {
 864         /*
 865          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 866          * Its value is decided by evsel's write_backward.
 867          * So &mp should not be passed through const pointer.
 868          */
 869         struct mmap_params mp = {
 870                 .nr_cblocks     = nr_cblocks,
 871                 .affinity       = affinity,
 872                 .flush          = flush,
 873                 .comp_level     = comp_level
 874         };
 875         struct perf_evlist_mmap_ops ops = {
 876                 .idx  = perf_evlist__mmap_cb_idx,
 877                 .get  = perf_evlist__mmap_cb_get,
 878                 .mmap = perf_evlist__mmap_cb_mmap,
 879         };
 880
 881         evlist->core.mmap_len = evlist__mmap_size(pages);
 882         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 883
 884         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 885                                    auxtrace_pages, auxtrace_overwrite);
 886
 887         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 888 }
 889
 890 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 891 {
 892         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 893 }
 894
 895 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
 896 {
 897         bool all_threads = (target->per_thread && target->system_wide);
 898         struct perf_cpu_map *cpus;
 899         struct perf_thread_map *threads;
 900
 901         /*
 902          * If specify '-a' and '--per-thread' to perf record, perf record
 903          * will override '--per-thread'. target->per_thread = false and
 904          * target->system_wide = true.
 905          *
 906          * If specify '--per-thread' only to perf record,
 907          * target->per_thread = true and target->system_wide = false.
 908          *
 909          * So target->per_thread && target->system_wide is false.
 910          * For perf record, thread_map__new_str doesn't call
 911          * thread_map__new_all_cpus. That will keep perf record's
 912          * current behavior.
 913          *
 914          * For perf stat, it allows the case that target->per_thread and
 915          * target->system_wide are all true. It means to collect system-wide
 916          * per-thread data. thread_map__new_str will call
 917          * thread_map__new_all_cpus to enumerate all threads.
 918          */
 919         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 920                                       all_threads);
 921
 922         if (!threads)
 923                 return -1;
 924
 925         if (target__uses_dummy_map(target))
 926                 cpus = perf_cpu_map__dummy_new();
 927         else
 928                 cpus = perf_cpu_map__new(target->cpu_list);
 929
 930         if (!cpus)
 931                 goto out_delete_threads;
 932
 933         evlist->core.has_user_cpus = !!target->cpu_list;
 934
 935         perf_evlist__set_maps(&evlist->core, cpus, threads);
 936
 937         return 0;
 938
 939 out_delete_threads:
 940         perf_thread_map__put(threads);
 941         return -1;
 942 }
 943
 944 void __perf_evlist__set_sample_bit(struct evlist *evlist,
 945                                    enum perf_event_sample_format bit)
 946 {
 947         struct evsel *evsel;
 948
 949         evlist__for_each_entry(evlist, evsel)
 950                 __perf_evsel__set_sample_bit(evsel, bit);
 951 }
 952
 953 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
 954                                      enum perf_event_sample_format bit)
 955 {
 956         struct evsel *evsel;
 957
 958         evlist__for_each_entry(evlist, evsel)
 959                 __perf_evsel__reset_sample_bit(evsel, bit);
 960 }
 961
 962 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 963 {
 964         struct evsel *evsel;
 965         int err = 0;
 966
 967         evlist__for_each_entry(evlist, evsel) {
 968                 if (evsel->filter == NULL)
 969                         continue;
 970
 971                 /*
 972                  * filters only work for tracepoint event, which doesn't have cpu limit.
 973                  * So evlist and evsel should always be same.
 974                  */
 975                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
 976                 if (err) {
 977                         *err_evsel = evsel;
 978                         break;
 979                 }
 980         }
 981
 982         return err;
 983 }
 984
 985 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
 986 {
 987         struct evsel *evsel;
 988         int err = 0;
 989
 990         if (filter == NULL)
 991                 return -1;
 992
 993         evlist__for_each_entry(evlist, evsel) {
 994                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
 995                         continue;
 996
 997                 err = perf_evsel__set_filter(evsel, filter);
 998                 if (err)
 999                         break;
1000         }
1001
1002         return err;
1003 }
1004
1005 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1006 {
1007         struct evsel *evsel;
1008         int err = 0;
1009
1010         if (filter == NULL)
1011                 return -1;
1012
1013         evlist__for_each_entry(evlist, evsel) {
1014                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1015                         continue;
1016
1017                 err = perf_evsel__append_tp_filter(evsel, filter);
1018                 if (err)
1019                         break;
1020         }
1021
1022         return err;
1023 }
1024
1025 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1026 {
1027         char *filter;
1028         size_t i;
1029
1030         for (i = 0; i < npids; ++i) {
1031                 if (i == 0) {
1032                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1033                                 return NULL;
1034                 } else {
1035                         char *tmp;
1036
1037                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1038                                 goto out_free;
1039
1040                         free(filter);
1041                         filter = tmp;
1042                 }
1043         }
1044
1045         return filter;
1046 out_free:
1047         free(filter);
1048         return NULL;
1049 }
1050
1051 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1052 {
1053         char *filter = asprintf__tp_filter_pids(npids, pids);
1054         int ret = perf_evlist__set_tp_filter(evlist, filter);
1055
1056         free(filter);
1057         return ret;
1058 }
1059
1060 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1061 {
1062         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1063 }
1064
1065 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1066 {
1067         char *filter = asprintf__tp_filter_pids(npids, pids);
1068         int ret = perf_evlist__append_tp_filter(evlist, filter);
1069
1070         free(filter);
1071         return ret;
1072 }
1073
1074 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1075 {
1076         return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1077 }
1078
1079 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1080 {
1081         struct evsel *pos;
1082
1083         if (evlist->core.nr_entries == 1)
1084                 return true;
1085
1086         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1087                 return false;
1088
1089         evlist__for_each_entry(evlist, pos) {
1090                 if (pos->id_pos != evlist->id_pos ||
1091                     pos->is_pos != evlist->is_pos)
1092                         return false;
1093         }
1094
1095         return true;
1096 }
1097
1098 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1099 {
1100         struct evsel *evsel;
1101
1102         if (evlist->combined_sample_type)
1103                 return evlist->combined_sample_type;
1104
1105         evlist__for_each_entry(evlist, evsel)
1106                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1107
1108         return evlist->combined_sample_type;
1109 }
1110
1111 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1112 {
1113         evlist->combined_sample_type = 0;
1114         return __perf_evlist__combined_sample_type(evlist);
1115 }
1116
1117 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1118 {
1119         struct evsel *evsel;
1120         u64 branch_type = 0;
1121
1122         evlist__for_each_entry(evlist, evsel)
1123                 branch_type |= evsel->core.attr.branch_sample_type;
1124         return branch_type;
1125 }
1126
1127 bool perf_evlist__valid_read_format(struct evlist *evlist)
1128 {
1129         struct evsel *first = evlist__first(evlist), *pos = first;
1130         u64 read_format = first->core.attr.read_format;
1131         u64 sample_type = first->core.attr.sample_type;
1132
1133         evlist__for_each_entry(evlist, pos) {
1134                 if (read_format != pos->core.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 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1148 {
1149         struct evsel *first = evlist__first(evlist);
1150         struct perf_sample *data;
1151         u64 sample_type;
1152         u16 size = 0;
1153
1154         if (!first->core.attr.sample_id_all)
1155                 goto out;
1156
1157         sample_type = first->core.attr.sample_type;
1158
1159         if (sample_type & PERF_SAMPLE_TID)
1160                 size += sizeof(data->tid) * 2;
1161
1162        if (sample_type & PERF_SAMPLE_TIME)
1163                 size += sizeof(data->time);
1164
1165         if (sample_type & PERF_SAMPLE_ID)
1166                 size += sizeof(data->id);
1167
1168         if (sample_type & PERF_SAMPLE_STREAM_ID)
1169                 size += sizeof(data->stream_id);
1170
1171         if (sample_type & PERF_SAMPLE_CPU)
1172                 size += sizeof(data->cpu) * 2;
1173
1174         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1175                 size += sizeof(data->id);
1176 out:
1177         return size;
1178 }
1179
1180 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1181 {
1182         struct evsel *first = evlist__first(evlist), *pos = first;
1183
1184         evlist__for_each_entry_continue(evlist, pos) {
1185                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1186                         return false;
1187         }
1188
1189         return true;
1190 }
1191
1192 bool perf_evlist__sample_id_all(struct evlist *evlist)
1193 {
1194         struct evsel *first = evlist__first(evlist);
1195         return first->core.attr.sample_id_all;
1196 }
1197
1198 void perf_evlist__set_selected(struct evlist *evlist,
1199                                struct evsel *evsel)
1200 {
1201         evlist->selected = evsel;
1202 }
1203
1204 void evlist__close(struct evlist *evlist)
1205 {
1206         struct evsel *evsel;
1207         struct affinity affinity;
1208         int cpu, i;
1209
1210         /*
1211          * With perf record core.cpus is usually NULL.
1212          * Use the old method to handle this for now.
1213          */
1214         if (!evlist->core.cpus) {
1215                 evlist__for_each_entry_reverse(evlist, evsel)
1216                         evsel__close(evsel);
1217                 return;
1218         }
1219
1220         if (affinity__setup(&affinity) < 0)
1221                 return;
1222         evlist__for_each_cpu(evlist, i, cpu) {
1223                 affinity__set(&affinity, cpu);
1224
1225                 evlist__for_each_entry_reverse(evlist, evsel) {
1226                         if (evsel__cpu_iter_skip(evsel, cpu))
1227                             continue;
1228                         perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1229                 }
1230         }
1231         affinity__cleanup(&affinity);
1232         evlist__for_each_entry_reverse(evlist, evsel) {
1233                 perf_evsel__free_fd(&evsel->core);
1234                 perf_evsel__free_id(&evsel->core);
1235         }
1236 }
1237
1238 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1239 {
1240         struct perf_cpu_map *cpus;
1241         struct perf_thread_map *threads;
1242         int err = -ENOMEM;
1243
1244         /*
1245          * Try reading /sys/devices/system/cpu/online to get
1246          * an all cpus map.
1247          *
1248          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1249          * code needs an overhaul to properly forward the
1250          * error, and we may not want to do that fallback to a
1251          * default cpu identity map :-\
1252          */
1253         cpus = perf_cpu_map__new(NULL);
1254         if (!cpus)
1255                 goto out;
1256
1257         threads = perf_thread_map__new_dummy();
1258         if (!threads)
1259                 goto out_put;
1260
1261         perf_evlist__set_maps(&evlist->core, cpus, threads);
1262 out:
1263         return err;
1264 out_put:
1265         perf_cpu_map__put(cpus);
1266         goto out;
1267 }
1268
1269 int evlist__open(struct evlist *evlist)
1270 {
1271         struct evsel *evsel;
1272         int err;
1273
1274         /*
1275          * Default: one fd per CPU, all threads, aka systemwide
1276          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1277          */
1278         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1279                 err = perf_evlist__create_syswide_maps(evlist);
1280                 if (err < 0)
1281                         goto out_err;
1282         }
1283
1284         perf_evlist__update_id_pos(evlist);
1285
1286         evlist__for_each_entry(evlist, evsel) {
1287                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1288                 if (err < 0)
1289                         goto out_err;
1290         }
1291
1292         return 0;
1293 out_err:
1294         evlist__close(evlist);
1295         errno = -err;
1296         return err;
1297 }
1298
1299 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1300                                   const char *argv[], bool pipe_output,
1301                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1302 {
1303         int child_ready_pipe[2], go_pipe[2];
1304         char bf;
1305
1306         if (pipe(child_ready_pipe) < 0) {
1307                 perror("failed to create 'ready' pipe");
1308                 return -1;
1309         }
1310
1311         if (pipe(go_pipe) < 0) {
1312                 perror("failed to create 'go' pipe");
1313                 goto out_close_ready_pipe;
1314         }
1315
1316         evlist->workload.pid = fork();
1317         if (evlist->workload.pid < 0) {
1318                 perror("failed to fork");
1319                 goto out_close_pipes;
1320         }
1321
1322         if (!evlist->workload.pid) {
1323                 int ret;
1324
1325                 if (pipe_output)
1326                         dup2(2, 1);
1327
1328                 signal(SIGTERM, SIG_DFL);
1329
1330                 close(child_ready_pipe[0]);
1331                 close(go_pipe[1]);
1332                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1333
1334                 /*
1335                  * Tell the parent we're ready to go
1336                  */
1337                 close(child_ready_pipe[1]);
1338
1339                 /*
1340                  * Wait until the parent tells us to go.
1341                  */
1342                 ret = read(go_pipe[0], &bf, 1);
1343                 /*
1344                  * The parent will ask for the execvp() to be performed by
1345                  * writing exactly one byte, in workload.cork_fd, usually via
1346                  * perf_evlist__start_workload().
1347                  *
1348                  * For cancelling the workload without actually running it,
1349                  * the parent will just close workload.cork_fd, without writing
1350                  * anything, i.e. read will return zero and we just exit()
1351                  * here.
1352                  */
1353                 if (ret != 1) {
1354                         if (ret == -1)
1355                                 perror("unable to read pipe");
1356                         exit(ret);
1357                 }
1358
1359                 execvp(argv[0], (char **)argv);
1360
1361                 if (exec_error) {
1362                         union sigval val;
1363
1364                         val.sival_int = errno;
1365                         if (sigqueue(getppid(), SIGUSR1, val))
1366                                 perror(argv[0]);
1367                 } else
1368                         perror(argv[0]);
1369                 exit(-1);
1370         }
1371
1372         if (exec_error) {
1373                 struct sigaction act = {
1374                         .sa_flags     = SA_SIGINFO,
1375                         .sa_sigaction = exec_error,
1376                 };
1377                 sigaction(SIGUSR1, &act, NULL);
1378         }
1379
1380         if (target__none(target)) {
1381                 if (evlist->core.threads == NULL) {
1382                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1383                                 __func__, __LINE__);
1384                         goto out_close_pipes;
1385                 }
1386                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1387         }
1388
1389         close(child_ready_pipe[1]);
1390         close(go_pipe[0]);
1391         /*
1392          * wait for child to settle
1393          */
1394         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1395                 perror("unable to read pipe");
1396                 goto out_close_pipes;
1397         }
1398
1399         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1400         evlist->workload.cork_fd = go_pipe[1];
1401         close(child_ready_pipe[0]);
1402         return 0;
1403
1404 out_close_pipes:
1405         close(go_pipe[0]);
1406         close(go_pipe[1]);
1407 out_close_ready_pipe:
1408         close(child_ready_pipe[0]);
1409         close(child_ready_pipe[1]);
1410         return -1;
1411 }
1412
1413 int perf_evlist__start_workload(struct evlist *evlist)
1414 {
1415         if (evlist->workload.cork_fd > 0) {
1416                 char bf = 0;
1417                 int ret;
1418                 /*
1419                  * Remove the cork, let it rip!
1420                  */
1421                 ret = write(evlist->workload.cork_fd, &bf, 1);
1422                 if (ret < 0)
1423                         perror("unable to write to pipe");
1424
1425                 close(evlist->workload.cork_fd);
1426                 return ret;
1427         }
1428
1429         return 0;
1430 }
1431
1432 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1433                               struct perf_sample *sample)
1434 {
1435         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1436
1437         if (!evsel)
1438                 return -EFAULT;
1439         return perf_evsel__parse_sample(evsel, event, sample);
1440 }
1441
1442 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1443                                         union perf_event *event,
1444                                         u64 *timestamp)
1445 {
1446         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1447
1448         if (!evsel)
1449                 return -EFAULT;
1450         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1451 }
1452
1453 int perf_evlist__strerror_open(struct evlist *evlist,
1454                                int err, char *buf, size_t size)
1455 {
1456         int printed, value;
1457         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1458
1459         switch (err) {
1460         case EACCES:
1461         case EPERM:
1462                 printed = scnprintf(buf, size,
1463                                     "Error:\t%s.\n"
1464                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1465
1466                 value = perf_event_paranoid();
1467
1468                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1469
1470                 if (value >= 2) {
1471                         printed += scnprintf(buf + printed, size - printed,
1472                                              "For your workloads it needs to be <= 1\nHint:\t");
1473                 }
1474                 printed += scnprintf(buf + printed, size - printed,
1475                                      "For system wide tracing it needs to be set to -1.\n");
1476
1477                 printed += scnprintf(buf + printed, size - printed,
1478                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1479                                     "Hint:\tThe current value is %d.", value);
1480                 break;
1481         case EINVAL: {
1482                 struct evsel *first = evlist__first(evlist);
1483                 int max_freq;
1484
1485                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1486                         goto out_default;
1487
1488                 if (first->core.attr.sample_freq < (u64)max_freq)
1489                         goto out_default;
1490
1491                 printed = scnprintf(buf, size,
1492                                     "Error:\t%s.\n"
1493                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1494                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1495                                     emsg, max_freq, first->core.attr.sample_freq);
1496                 break;
1497         }
1498         default:
1499 out_default:
1500                 scnprintf(buf, size, "%s", emsg);
1501                 break;
1502         }
1503
1504         return 0;
1505 }
1506
1507 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1508 {
1509         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1510         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1511
1512         switch (err) {
1513         case EPERM:
1514                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1515                 printed += scnprintf(buf + printed, size - printed,
1516                                      "Error:\t%s.\n"
1517                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1518                                      "Hint:\tTried using %zd kB.\n",
1519                                      emsg, pages_max_per_user, pages_attempted);
1520
1521                 if (pages_attempted >= pages_max_per_user) {
1522                         printed += scnprintf(buf + printed, size - printed,
1523                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1524                                              pages_max_per_user + pages_attempted);
1525                 }
1526
1527                 printed += scnprintf(buf + printed, size - printed,
1528                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1529                 break;
1530         default:
1531                 scnprintf(buf, size, "%s", emsg);
1532                 break;
1533         }
1534
1535         return 0;
1536 }
1537
1538 void perf_evlist__to_front(struct evlist *evlist,
1539                            struct evsel *move_evsel)
1540 {
1541         struct evsel *evsel, *n;
1542         LIST_HEAD(move);
1543
1544         if (move_evsel == evlist__first(evlist))
1545                 return;
1546
1547         evlist__for_each_entry_safe(evlist, n, evsel) {
1548                 if (evsel->leader == move_evsel->leader)
1549                         list_move_tail(&evsel->core.node, &move);
1550         }
1551
1552         list_splice(&move, &evlist->core.entries);
1553 }
1554
1555 void perf_evlist__set_tracking_event(struct evlist *evlist,
1556                                      struct evsel *tracking_evsel)
1557 {
1558         struct evsel *evsel;
1559
1560         if (tracking_evsel->tracking)
1561                 return;
1562
1563         evlist__for_each_entry(evlist, evsel) {
1564                 if (evsel != tracking_evsel)
1565                         evsel->tracking = false;
1566         }
1567
1568         tracking_evsel->tracking = true;
1569 }
1570
1571 struct evsel *
1572 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1573                                const char *str)
1574 {
1575         struct evsel *evsel;
1576
1577         evlist__for_each_entry(evlist, evsel) {
1578                 if (!evsel->name)
1579                         continue;
1580                 if (strcmp(str, evsel->name) == 0)
1581                         return evsel;
1582         }
1583
1584         return NULL;
1585 }
1586
1587 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1588                                   enum bkw_mmap_state state)
1589 {
1590         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1591         enum action {
1592                 NONE,
1593                 PAUSE,
1594                 RESUME,
1595         } action = NONE;
1596
1597         if (!evlist->overwrite_mmap)
1598                 return;
1599
1600         switch (old_state) {
1601         case BKW_MMAP_NOTREADY: {
1602                 if (state != BKW_MMAP_RUNNING)
1603                         goto state_err;
1604                 break;
1605         }
1606         case BKW_MMAP_RUNNING: {
1607                 if (state != BKW_MMAP_DATA_PENDING)
1608                         goto state_err;
1609                 action = PAUSE;
1610                 break;
1611         }
1612         case BKW_MMAP_DATA_PENDING: {
1613                 if (state != BKW_MMAP_EMPTY)
1614                         goto state_err;
1615                 break;
1616         }
1617         case BKW_MMAP_EMPTY: {
1618                 if (state != BKW_MMAP_RUNNING)
1619                         goto state_err;
1620                 action = RESUME;
1621                 break;
1622         }
1623         default:
1624                 WARN_ONCE(1, "Shouldn't get there\n");
1625         }
1626
1627         evlist->bkw_mmap_state = state;
1628
1629         switch (action) {
1630         case PAUSE:
1631                 perf_evlist__pause(evlist);
1632                 break;
1633         case RESUME:
1634                 perf_evlist__resume(evlist);
1635                 break;
1636         case NONE:
1637         default:
1638                 break;
1639         }
1640
1641 state_err:
1642         return;
1643 }
1644
1645 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1646 {
1647         struct evsel *evsel;
1648
1649         evlist__for_each_entry(evlist, evsel) {
1650                 if (!evsel->core.attr.exclude_kernel)
1651                         return false;
1652         }
1653
1654         return true;
1655 }
1656
1657 /*
1658  * Events in data file are not collect in groups, but we still want
1659  * the group display. Set the artificial group and set the leader's
1660  * forced_leader flag to notify the display code.
1661  */
1662 void perf_evlist__force_leader(struct evlist *evlist)
1663 {
1664         if (!evlist->nr_groups) {
1665                 struct evsel *leader = evlist__first(evlist);
1666
1667                 perf_evlist__set_leader(evlist);
1668                 leader->forced_leader = true;
1669         }
1670 }
1671
1672 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1673                                                  struct evsel *evsel,
1674                                                 bool close)
1675 {
1676         struct evsel *c2, *leader;
1677         bool is_open = true;
1678
1679         leader = evsel->leader;
1680         pr_debug("Weak group for %s/%d failed\n",
1681                         leader->name, leader->core.nr_members);
1682
1683         /*
1684          * for_each_group_member doesn't work here because it doesn't
1685          * include the first entry.
1686          */
1687         evlist__for_each_entry(evsel_list, c2) {
1688                 if (c2 == evsel)
1689                         is_open = false;
1690                 if (c2->leader == leader) {
1691                         if (is_open && close)
1692                                 perf_evsel__close(&c2->core);
1693                         c2->leader = c2;
1694                         c2->core.nr_members = 0;
1695                         /*
1696                          * Set this for all former members of the group
1697                          * to indicate they get reopened.
1698                          */
1699                         c2->reset_group = true;
1700                 }
1701         }
1702         return leader;
1703 }
1704
1705 int perf_evlist__add_sb_event(struct evlist **evlist,
1706                               struct perf_event_attr *attr,
1707                               perf_evsel__sb_cb_t cb,
1708                               void *data)
1709 {
1710         struct evsel *evsel;
1711         bool new_evlist = (*evlist) == NULL;
1712
1713         if (*evlist == NULL)
1714                 *evlist = evlist__new();
1715         if (*evlist == NULL)
1716                 return -1;
1717
1718         if (!attr->sample_id_all) {
1719                 pr_warning("enabling sample_id_all for all side band events\n");
1720                 attr->sample_id_all = 1;
1721         }
1722
1723         evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1724         if (!evsel)
1725                 goto out_err;
1726
1727         evsel->side_band.cb = cb;
1728         evsel->side_band.data = data;
1729         evlist__add(*evlist, evsel);
1730         return 0;
1731
1732 out_err:
1733         if (new_evlist) {
1734                 evlist__delete(*evlist);
1735                 *evlist = NULL;
1736         }
1737         return -1;
1738 }
1739
1740 static void *perf_evlist__poll_thread(void *arg)
1741 {
1742         struct evlist *evlist = arg;
1743         bool draining = false;
1744         int i, done = 0;
1745         /*
1746          * In order to read symbols from other namespaces perf to needs to call
1747          * setns(2).  This isn't permitted if the struct_fs has multiple users.
1748          * unshare(2) the fs so that we may continue to setns into namespaces
1749          * that we're observing when, for instance, reading the build-ids at
1750          * the end of a 'perf record' session.
1751          */
1752         unshare(CLONE_FS);
1753
1754         while (!done) {
1755                 bool got_data = false;
1756
1757                 if (evlist->thread.done)
1758                         draining = true;
1759
1760                 if (!draining)
1761                         evlist__poll(evlist, 1000);
1762
1763                 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1764                         struct mmap *map = &evlist->mmap[i];
1765                         union perf_event *event;
1766
1767                         if (perf_mmap__read_init(&map->core))
1768                                 continue;
1769                         while ((event = perf_mmap__read_event(&map->core)) != NULL) {
1770                                 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1771
1772                                 if (evsel && evsel->side_band.cb)
1773                                         evsel->side_band.cb(event, evsel->side_band.data);
1774                                 else
1775                                         pr_warning("cannot locate proper evsel for the side band event\n");
1776
1777                                 perf_mmap__consume(&map->core);
1778                                 got_data = true;
1779                         }
1780                         perf_mmap__read_done(&map->core);
1781                 }
1782
1783                 if (draining && !got_data)
1784                         break;
1785         }
1786         return NULL;
1787 }
1788
1789 int perf_evlist__start_sb_thread(struct evlist *evlist,
1790                                  struct target *target)
1791 {
1792         struct evsel *counter;
1793
1794         if (!evlist)
1795                 return 0;
1796
1797         if (perf_evlist__create_maps(evlist, target))
1798                 goto out_delete_evlist;
1799
1800         evlist__for_each_entry(evlist, counter) {
1801                 if (evsel__open(counter, evlist->core.cpus,
1802                                      evlist->core.threads) < 0)
1803                         goto out_delete_evlist;
1804         }
1805
1806         if (evlist__mmap(evlist, UINT_MAX))
1807                 goto out_delete_evlist;
1808
1809         evlist__for_each_entry(evlist, counter) {
1810                 if (evsel__enable(counter))
1811                         goto out_delete_evlist;
1812         }
1813
1814         evlist->thread.done = 0;
1815         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1816                 goto out_delete_evlist;
1817
1818         return 0;
1819
1820 out_delete_evlist:
1821         evlist__delete(evlist);
1822         evlist = NULL;
1823         return -1;
1824 }
1825
1826 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1827 {
1828         if (!evlist)
1829                 return;
1830         evlist->thread.done = 1;
1831         pthread_join(evlist->thread.th, NULL);
1832         evlist__delete(evlist);
1833 }