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 "../perf.h"
  22 #include "asm/bug.h"
  23 #include "bpf-event.h"
  24 #include <signal.h>
  25 #include <unistd.h>
  26 #include <sched.h>
  27 #include <stdlib.h>
  28
  29 #include "parse-events.h"
  30 #include <subcmd/parse-options.h>
  31
  32 #include <fcntl.h>
  33 #include <sys/ioctl.h>
  34 #include <sys/mman.h>
  35
  36 #include <linux/bitops.h>
  37 #include <linux/hash.h>
  38 #include <linux/log2.h>
  39 #include <linux/err.h>
  40 #include <linux/string.h>
  41 #include <linux/zalloc.h>
  42 #include <perf/evlist.h>
  43 #include <perf/evsel.h>
  44 #include <perf/cpumap.h>
  45
  46 #include <internal/xyarray.h>
  47
  48 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  49 int sigqueue(pid_t pid, int sig, const union sigval value);
  50 #endif
  51
  52 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  53 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  54
  55 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  56                   struct perf_thread_map *threads)
  57 {
  58         perf_evlist__init(&evlist->core);
  59         perf_evlist__set_maps(&evlist->core, cpus, threads);
  60         fdarray__init(&evlist->core.pollfd, 64);
  61         evlist->workload.pid = -1;
  62         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  63 }
  64
  65 struct evlist *evlist__new(void)
  66 {
  67         struct evlist *evlist = zalloc(sizeof(*evlist));
  68
  69         if (evlist != NULL)
  70                 evlist__init(evlist, NULL, NULL);
  71
  72         return evlist;
  73 }
  74
  75 struct evlist *perf_evlist__new_default(void)
  76 {
  77         struct evlist *evlist = evlist__new();
  78
  79         if (evlist && perf_evlist__add_default(evlist)) {
  80                 evlist__delete(evlist);
  81                 evlist = NULL;
  82         }
  83
  84         return evlist;
  85 }
  86
  87 struct evlist *perf_evlist__new_dummy(void)
  88 {
  89         struct evlist *evlist = evlist__new();
  90
  91         if (evlist && perf_evlist__add_dummy(evlist)) {
  92                 evlist__delete(evlist);
  93                 evlist = NULL;
  94         }
  95
  96         return evlist;
  97 }
  98
  99 /**
 100  * perf_evlist__set_id_pos - set the positions of event ids.
 101  * @evlist: selected event list
 102  *
 103  * Events with compatible sample types all have the same id_pos
 104  * and is_pos.  For convenience, put a copy on evlist.
 105  */
 106 void perf_evlist__set_id_pos(struct evlist *evlist)
 107 {
 108         struct evsel *first = evlist__first(evlist);
 109
 110         evlist->id_pos = first->id_pos;
 111         evlist->is_pos = first->is_pos;
 112 }
 113
 114 static void perf_evlist__update_id_pos(struct evlist *evlist)
 115 {
 116         struct evsel *evsel;
 117
 118         evlist__for_each_entry(evlist, evsel)
 119                 perf_evsel__calc_id_pos(evsel);
 120
 121         perf_evlist__set_id_pos(evlist);
 122 }
 123
 124 static void evlist__purge(struct evlist *evlist)
 125 {
 126         struct evsel *pos, *n;
 127
 128         evlist__for_each_entry_safe(evlist, n, pos) {
 129                 list_del_init(&pos->core.node);
 130                 pos->evlist = NULL;
 131                 evsel__delete(pos);
 132         }
 133
 134         evlist->core.nr_entries = 0;
 135 }
 136
 137 void evlist__exit(struct evlist *evlist)
 138 {
 139         zfree(&evlist->mmap);
 140         zfree(&evlist->overwrite_mmap);
 141         fdarray__exit(&evlist->core.pollfd);
 142 }
 143
 144 void evlist__delete(struct evlist *evlist)
 145 {
 146         if (evlist == NULL)
 147                 return;
 148
 149         evlist__munmap(evlist);
 150         evlist__close(evlist);
 151         perf_cpu_map__put(evlist->core.cpus);
 152         perf_thread_map__put(evlist->core.threads);
 153         evlist->core.cpus = NULL;
 154         evlist->core.threads = NULL;
 155         evlist__purge(evlist);
 156         evlist__exit(evlist);
 157         free(evlist);
 158 }
 159
 160 void evlist__add(struct evlist *evlist, struct evsel *entry)
 161 {
 162         entry->evlist = evlist;
 163         entry->idx = evlist->core.nr_entries;
 164         entry->tracking = !entry->idx;
 165
 166         perf_evlist__add(&evlist->core, &entry->core);
 167
 168         if (evlist->core.nr_entries == 1)
 169                 perf_evlist__set_id_pos(evlist);
 170 }
 171
 172 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 173 {
 174         evsel->evlist = NULL;
 175         perf_evlist__remove(&evlist->core, &evsel->core);
 176 }
 177
 178 void perf_evlist__splice_list_tail(struct evlist *evlist,
 179                                    struct list_head *list)
 180 {
 181         struct evsel *evsel, *temp;
 182
 183         __evlist__for_each_entry_safe(list, temp, evsel) {
 184                 list_del_init(&evsel->core.node);
 185                 evlist__add(evlist, evsel);
 186         }
 187 }
 188
 189 void __perf_evlist__set_leader(struct list_head *list)
 190 {
 191         struct evsel *evsel, *leader;
 192
 193         leader = list_entry(list->next, struct evsel, core.node);
 194         evsel = list_entry(list->prev, struct evsel, core.node);
 195
 196         leader->core.nr_members = evsel->idx - leader->idx + 1;
 197
 198         __evlist__for_each_entry(list, evsel) {
 199                 evsel->leader = leader;
 200         }
 201 }
 202
 203 void perf_evlist__set_leader(struct evlist *evlist)
 204 {
 205         if (evlist->core.nr_entries) {
 206                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 207                 __perf_evlist__set_leader(&evlist->core.entries);
 208         }
 209 }
 210
 211 int __perf_evlist__add_default(struct evlist *evlist, bool precise)
 212 {
 213         struct evsel *evsel = perf_evsel__new_cycles(precise);
 214
 215         if (evsel == NULL)
 216                 return -ENOMEM;
 217
 218         evlist__add(evlist, evsel);
 219         return 0;
 220 }
 221
 222 int perf_evlist__add_dummy(struct evlist *evlist)
 223 {
 224         struct perf_event_attr attr = {
 225                 .type   = PERF_TYPE_SOFTWARE,
 226                 .config = PERF_COUNT_SW_DUMMY,
 227                 .size   = sizeof(attr), /* to capture ABI version */
 228         };
 229         struct evsel *evsel = perf_evsel__new_idx(&attr, evlist->core.nr_entries);
 230
 231         if (evsel == NULL)
 232                 return -ENOMEM;
 233
 234         evlist__add(evlist, evsel);
 235         return 0;
 236 }
 237
 238 static int evlist__add_attrs(struct evlist *evlist,
 239                                   struct perf_event_attr *attrs, size_t nr_attrs)
 240 {
 241         struct evsel *evsel, *n;
 242         LIST_HEAD(head);
 243         size_t i;
 244
 245         for (i = 0; i < nr_attrs; i++) {
 246                 evsel = perf_evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 247                 if (evsel == NULL)
 248                         goto out_delete_partial_list;
 249                 list_add_tail(&evsel->core.node, &head);
 250         }
 251
 252         perf_evlist__splice_list_tail(evlist, &head);
 253
 254         return 0;
 255
 256 out_delete_partial_list:
 257         __evlist__for_each_entry_safe(&head, n, evsel)
 258                 evsel__delete(evsel);
 259         return -1;
 260 }
 261
 262 int __perf_evlist__add_default_attrs(struct evlist *evlist,
 263                                      struct perf_event_attr *attrs, size_t nr_attrs)
 264 {
 265         size_t i;
 266
 267         for (i = 0; i < nr_attrs; i++)
 268                 event_attr_init(attrs + i);
 269
 270         return evlist__add_attrs(evlist, attrs, nr_attrs);
 271 }
 272
 273 struct evsel *
 274 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 275 {
 276         struct evsel *evsel;
 277
 278         evlist__for_each_entry(evlist, evsel) {
 279                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 280                     (int)evsel->core.attr.config == id)
 281                         return evsel;
 282         }
 283
 284         return NULL;
 285 }
 286
 287 struct evsel *
 288 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
 289                                      const char *name)
 290 {
 291         struct evsel *evsel;
 292
 293         evlist__for_each_entry(evlist, evsel) {
 294                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 295                     (strcmp(evsel->name, name) == 0))
 296                         return evsel;
 297         }
 298
 299         return NULL;
 300 }
 301
 302 int perf_evlist__add_newtp(struct evlist *evlist,
 303                            const char *sys, const char *name, void *handler)
 304 {
 305         struct evsel *evsel = perf_evsel__newtp(sys, name);
 306
 307         if (IS_ERR(evsel))
 308                 return -1;
 309
 310         evsel->handler = handler;
 311         evlist__add(evlist, evsel);
 312         return 0;
 313 }
 314
 315 static int perf_evlist__nr_threads(struct evlist *evlist,
 316                                    struct evsel *evsel)
 317 {
 318         if (evsel->core.system_wide)
 319                 return 1;
 320         else
 321                 return perf_thread_map__nr(evlist->core.threads);
 322 }
 323
 324 void evlist__disable(struct evlist *evlist)
 325 {
 326         struct evsel *pos;
 327
 328         evlist__for_each_entry(evlist, pos) {
 329                 if (pos->disabled || !perf_evsel__is_group_leader(pos) || !pos->core.fd)
 330                         continue;
 331                 evsel__disable(pos);
 332         }
 333
 334         evlist->enabled = false;
 335 }
 336
 337 void evlist__enable(struct evlist *evlist)
 338 {
 339         struct evsel *pos;
 340
 341         evlist__for_each_entry(evlist, pos) {
 342                 if (!perf_evsel__is_group_leader(pos) || !pos->core.fd)
 343                         continue;
 344                 evsel__enable(pos);
 345         }
 346
 347         evlist->enabled = true;
 348 }
 349
 350 void perf_evlist__toggle_enable(struct evlist *evlist)
 351 {
 352         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 353 }
 354
 355 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
 356                                          struct evsel *evsel, int cpu)
 357 {
 358         int thread;
 359         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 360
 361         if (!evsel->core.fd)
 362                 return -EINVAL;
 363
 364         for (thread = 0; thread < nr_threads; thread++) {
 365                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 366                 if (err)
 367                         return err;
 368         }
 369         return 0;
 370 }
 371
 372 static int perf_evlist__enable_event_thread(struct evlist *evlist,
 373                                             struct evsel *evsel,
 374                                             int thread)
 375 {
 376         int cpu;
 377         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 378
 379         if (!evsel->core.fd)
 380                 return -EINVAL;
 381
 382         for (cpu = 0; cpu < nr_cpus; cpu++) {
 383                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 384                 if (err)
 385                         return err;
 386         }
 387         return 0;
 388 }
 389
 390 int perf_evlist__enable_event_idx(struct evlist *evlist,
 391                                   struct evsel *evsel, int idx)
 392 {
 393         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 394
 395         if (per_cpu_mmaps)
 396                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 397         else
 398                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 399 }
 400
 401 int evlist__add_pollfd(struct evlist *evlist, int fd)
 402 {
 403         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN);
 404 }
 405
 406 static void perf_evlist__munmap_filtered(struct fdarray *fda, int fd,
 407                                          void *arg __maybe_unused)
 408 {
 409         struct mmap *map = fda->priv[fd].ptr;
 410
 411         if (map)
 412                 perf_mmap__put(map);
 413 }
 414
 415 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 416 {
 417         return fdarray__filter(&evlist->core.pollfd, revents_and_mask,
 418                                perf_evlist__munmap_filtered, NULL);
 419 }
 420
 421 int evlist__poll(struct evlist *evlist, int timeout)
 422 {
 423         return perf_evlist__poll(&evlist->core, timeout);
 424 }
 425
 426 static void perf_evlist__set_sid_idx(struct evlist *evlist,
 427                                      struct evsel *evsel, int idx, int cpu,
 428                                      int thread)
 429 {
 430         struct perf_sample_id *sid = SID(evsel, cpu, thread);
 431         sid->idx = idx;
 432         if (evlist->core.cpus && cpu >= 0)
 433                 sid->cpu = evlist->core.cpus->map[cpu];
 434         else
 435                 sid->cpu = -1;
 436         if (!evsel->core.system_wide && evlist->core.threads && thread >= 0)
 437                 sid->tid = perf_thread_map__pid(evlist->core.threads, thread);
 438         else
 439                 sid->tid = -1;
 440 }
 441
 442 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
 443 {
 444         struct hlist_head *head;
 445         struct perf_sample_id *sid;
 446         int hash;
 447
 448         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 449         head = &evlist->core.heads[hash];
 450
 451         hlist_for_each_entry(sid, head, node)
 452                 if (sid->id == id)
 453                         return sid;
 454
 455         return NULL;
 456 }
 457
 458 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
 459 {
 460         struct perf_sample_id *sid;
 461
 462         if (evlist->core.nr_entries == 1 || !id)
 463                 return evlist__first(evlist);
 464
 465         sid = perf_evlist__id2sid(evlist, id);
 466         if (sid)
 467                 return container_of(sid->evsel, struct evsel, core);
 468
 469         if (!perf_evlist__sample_id_all(evlist))
 470                 return evlist__first(evlist);
 471
 472         return NULL;
 473 }
 474
 475 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
 476                                                 u64 id)
 477 {
 478         struct perf_sample_id *sid;
 479
 480         if (!id)
 481                 return NULL;
 482
 483         sid = perf_evlist__id2sid(evlist, id);
 484         if (sid)
 485                 return container_of(sid->evsel, struct evsel, core);
 486
 487         return NULL;
 488 }
 489
 490 static int perf_evlist__event2id(struct evlist *evlist,
 491                                  union perf_event *event, u64 *id)
 492 {
 493         const __u64 *array = event->sample.array;
 494         ssize_t n;
 495
 496         n = (event->header.size - sizeof(event->header)) >> 3;
 497
 498         if (event->header.type == PERF_RECORD_SAMPLE) {
 499                 if (evlist->id_pos >= n)
 500                         return -1;
 501                 *id = array[evlist->id_pos];
 502         } else {
 503                 if (evlist->is_pos > n)
 504                         return -1;
 505                 n -= evlist->is_pos;
 506                 *id = array[n];
 507         }
 508         return 0;
 509 }
 510
 511 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
 512                                             union perf_event *event)
 513 {
 514         struct evsel *first = evlist__first(evlist);
 515         struct hlist_head *head;
 516         struct perf_sample_id *sid;
 517         int hash;
 518         u64 id;
 519
 520         if (evlist->core.nr_entries == 1)
 521                 return first;
 522
 523         if (!first->core.attr.sample_id_all &&
 524             event->header.type != PERF_RECORD_SAMPLE)
 525                 return first;
 526
 527         if (perf_evlist__event2id(evlist, event, &id))
 528                 return NULL;
 529
 530         /* Synthesized events have an id of zero */
 531         if (!id)
 532                 return first;
 533
 534         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 535         head = &evlist->core.heads[hash];
 536
 537         hlist_for_each_entry(sid, head, node) {
 538                 if (sid->id == id)
 539                         return container_of(sid->evsel, struct evsel, core);
 540         }
 541         return NULL;
 542 }
 543
 544 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
 545 {
 546         int i;
 547
 548         if (!evlist->overwrite_mmap)
 549                 return 0;
 550
 551         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 552                 int fd = evlist->overwrite_mmap[i].core.fd;
 553                 int err;
 554
 555                 if (fd < 0)
 556                         continue;
 557                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 558                 if (err)
 559                         return err;
 560         }
 561         return 0;
 562 }
 563
 564 static int perf_evlist__pause(struct evlist *evlist)
 565 {
 566         return perf_evlist__set_paused(evlist, true);
 567 }
 568
 569 static int perf_evlist__resume(struct evlist *evlist)
 570 {
 571         return perf_evlist__set_paused(evlist, false);
 572 }
 573
 574 static void evlist__munmap_nofree(struct evlist *evlist)
 575 {
 576         int i;
 577
 578         if (evlist->mmap)
 579                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 580                         perf_mmap__munmap(&evlist->mmap[i]);
 581
 582         if (evlist->overwrite_mmap)
 583                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 584                         perf_mmap__munmap(&evlist->overwrite_mmap[i]);
 585 }
 586
 587 void evlist__munmap(struct evlist *evlist)
 588 {
 589         evlist__munmap_nofree(evlist);
 590         zfree(&evlist->mmap);
 591         zfree(&evlist->overwrite_mmap);
 592 }
 593
 594 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 595                                        bool overwrite)
 596 {
 597         int i;
 598         struct mmap *map;
 599
 600         evlist->core.nr_mmaps = perf_cpu_map__nr(evlist->core.cpus);
 601         if (perf_cpu_map__empty(evlist->core.cpus))
 602                 evlist->core.nr_mmaps = perf_thread_map__nr(evlist->core.threads);
 603         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 604         if (!map)
 605                 return NULL;
 606
 607         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 608                 map[i].core.fd = -1;
 609                 map[i].core.overwrite = overwrite;
 610                 /*
 611                  * When the perf_mmap() call is made we grab one refcount, plus
 612                  * one extra to let perf_mmap__consume() get the last
 613                  * events after all real references (perf_mmap__get()) are
 614                  * dropped.
 615                  *
 616                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 617                  * thus does perf_mmap__get() on it.
 618                  */
 619                 refcount_set(&map[i].core.refcnt, 0);
 620         }
 621         return map;
 622 }
 623
 624 static bool
 625 perf_evlist__should_poll(struct evlist *evlist __maybe_unused,
 626                          struct evsel *evsel)
 627 {
 628         if (evsel->core.attr.write_backward)
 629                 return false;
 630         return true;
 631 }
 632
 633 static int evlist__mmap_per_evsel(struct evlist *evlist, int idx,
 634                                        struct mmap_params *mp, int cpu_idx,
 635                                        int thread, int *_output, int *_output_overwrite)
 636 {
 637         struct evsel *evsel;
 638         int revent;
 639         int evlist_cpu = cpu_map__cpu(evlist->core.cpus, cpu_idx);
 640
 641         evlist__for_each_entry(evlist, evsel) {
 642                 struct mmap *maps = evlist->mmap;
 643                 int *output = _output;
 644                 int fd;
 645                 int cpu;
 646
 647                 mp->prot = PROT_READ | PROT_WRITE;
 648                 if (evsel->core.attr.write_backward) {
 649                         output = _output_overwrite;
 650                         maps = evlist->overwrite_mmap;
 651
 652                         if (!maps) {
 653                                 maps = evlist__alloc_mmap(evlist, true);
 654                                 if (!maps)
 655                                         return -1;
 656                                 evlist->overwrite_mmap = maps;
 657                                 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 658                                         perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 659                         }
 660                         mp->prot &= ~PROT_WRITE;
 661                 }
 662
 663                 if (evsel->core.system_wide && thread)
 664                         continue;
 665
 666                 cpu = perf_cpu_map__idx(evsel->core.cpus, evlist_cpu);
 667                 if (cpu == -1)
 668                         continue;
 669
 670                 fd = FD(evsel, cpu, thread);
 671
 672                 if (*output == -1) {
 673                         *output = fd;
 674
 675                         if (perf_mmap__mmap(&maps[idx], mp, *output, evlist_cpu) < 0)
 676                                 return -1;
 677                 } else {
 678                         if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
 679                                 return -1;
 680
 681                         perf_mmap__get(&maps[idx]);
 682                 }
 683
 684                 revent = perf_evlist__should_poll(evlist, evsel) ? POLLIN : 0;
 685
 686                 /*
 687                  * The system_wide flag causes a selected event to be opened
 688                  * always without a pid.  Consequently it will never get a
 689                  * POLLHUP, but it is used for tracking in combination with
 690                  * other events, so it should not need to be polled anyway.
 691                  * Therefore don't add it for polling.
 692                  */
 693                 if (!evsel->core.system_wide &&
 694                      perf_evlist__add_pollfd(&evlist->core, fd, &maps[idx], revent) < 0) {
 695                         perf_mmap__put(&maps[idx]);
 696                         return -1;
 697                 }
 698
 699                 if (evsel->core.attr.read_format & PERF_FORMAT_ID) {
 700                         if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, cpu, thread,
 701                                                    fd) < 0)
 702                                 return -1;
 703                         perf_evlist__set_sid_idx(evlist, evsel, idx, cpu,
 704                                                  thread);
 705                 }
 706         }
 707
 708         return 0;
 709 }
 710
 711 static int evlist__mmap_per_cpu(struct evlist *evlist,
 712                                      struct mmap_params *mp)
 713 {
 714         int cpu, thread;
 715         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 716         int nr_threads = perf_thread_map__nr(evlist->core.threads);
 717
 718         pr_debug2("perf event ring buffer mmapped per cpu\n");
 719         for (cpu = 0; cpu < nr_cpus; cpu++) {
 720                 int output = -1;
 721                 int output_overwrite = -1;
 722
 723                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, cpu,
 724                                               true);
 725
 726                 for (thread = 0; thread < nr_threads; thread++) {
 727                         if (evlist__mmap_per_evsel(evlist, cpu, mp, cpu,
 728                                                         thread, &output, &output_overwrite))
 729                                 goto out_unmap;
 730                 }
 731         }
 732
 733         return 0;
 734
 735 out_unmap:
 736         evlist__munmap_nofree(evlist);
 737         return -1;
 738 }
 739
 740 static int evlist__mmap_per_thread(struct evlist *evlist,
 741                                         struct mmap_params *mp)
 742 {
 743         int thread;
 744         int nr_threads = perf_thread_map__nr(evlist->core.threads);
 745
 746         pr_debug2("perf event ring buffer mmapped per thread\n");
 747         for (thread = 0; thread < nr_threads; thread++) {
 748                 int output = -1;
 749                 int output_overwrite = -1;
 750
 751                 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, thread,
 752                                               false);
 753
 754                 if (evlist__mmap_per_evsel(evlist, thread, mp, 0, thread,
 755                                                 &output, &output_overwrite))
 756                         goto out_unmap;
 757         }
 758
 759         return 0;
 760
 761 out_unmap:
 762         evlist__munmap_nofree(evlist);
 763         return -1;
 764 }
 765
 766 unsigned long perf_event_mlock_kb_in_pages(void)
 767 {
 768         unsigned long pages;
 769         int max;
 770
 771         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 772                 /*
 773                  * Pick a once upon a time good value, i.e. things look
 774                  * strange since we can't read a sysctl value, but lets not
 775                  * die yet...
 776                  */
 777                 max = 512;
 778         } else {
 779                 max -= (page_size / 1024);
 780         }
 781
 782         pages = (max * 1024) / page_size;
 783         if (!is_power_of_2(pages))
 784                 pages = rounddown_pow_of_two(pages);
 785
 786         return pages;
 787 }
 788
 789 size_t evlist__mmap_size(unsigned long pages)
 790 {
 791         if (pages == UINT_MAX)
 792                 pages = perf_event_mlock_kb_in_pages();
 793         else if (!is_power_of_2(pages))
 794                 return 0;
 795
 796         return (pages + 1) * page_size;
 797 }
 798
 799 static long parse_pages_arg(const char *str, unsigned long min,
 800                             unsigned long max)
 801 {
 802         unsigned long pages, val;
 803         static struct parse_tag tags[] = {
 804                 { .tag  = 'B', .mult = 1       },
 805                 { .tag  = 'K', .mult = 1 << 10 },
 806                 { .tag  = 'M', .mult = 1 << 20 },
 807                 { .tag  = 'G', .mult = 1 << 30 },
 808                 { .tag  = 0 },
 809         };
 810
 811         if (str == NULL)
 812                 return -EINVAL;
 813
 814         val = parse_tag_value(str, tags);
 815         if (val != (unsigned long) -1) {
 816                 /* we got file size value */
 817                 pages = PERF_ALIGN(val, page_size) / page_size;
 818         } else {
 819                 /* we got pages count value */
 820                 char *eptr;
 821                 pages = strtoul(str, &eptr, 10);
 822                 if (*eptr != '\0')
 823                         return -EINVAL;
 824         }
 825
 826         if (pages == 0 && min == 0) {
 827                 /* leave number of pages at 0 */
 828         } else if (!is_power_of_2(pages)) {
 829                 char buf[100];
 830
 831                 /* round pages up to next power of 2 */
 832                 pages = roundup_pow_of_two(pages);
 833                 if (!pages)
 834                         return -EINVAL;
 835
 836                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 837                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 838                         buf, pages);
 839         }
 840
 841         if (pages > max)
 842                 return -EINVAL;
 843
 844         return pages;
 845 }
 846
 847 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 848 {
 849         unsigned long max = UINT_MAX;
 850         long pages;
 851
 852         if (max > SIZE_MAX / page_size)
 853                 max = SIZE_MAX / page_size;
 854
 855         pages = parse_pages_arg(str, 1, max);
 856         if (pages < 0) {
 857                 pr_err("Invalid argument for --mmap_pages/-m\n");
 858                 return -1;
 859         }
 860
 861         *mmap_pages = pages;
 862         return 0;
 863 }
 864
 865 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 866                                   int unset __maybe_unused)
 867 {
 868         return __perf_evlist__parse_mmap_pages(opt->value, str);
 869 }
 870
 871 /**
 872  * evlist__mmap_ex - Create mmaps to receive events.
 873  * @evlist: list of events
 874  * @pages: map length in pages
 875  * @overwrite: overwrite older events?
 876  * @auxtrace_pages - auxtrace map length in pages
 877  * @auxtrace_overwrite - overwrite older auxtrace data?
 878  *
 879  * If @overwrite is %false the user needs to signal event consumption using
 880  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 881  * automatically.
 882  *
 883  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 884  * consumption using auxtrace_mmap__write_tail().
 885  *
 886  * Return: %0 on success, negative error code otherwise.
 887  */
 888 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 889                          unsigned int auxtrace_pages,
 890                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 891                          int comp_level)
 892 {
 893         struct evsel *evsel;
 894         const struct perf_cpu_map *cpus = evlist->core.cpus;
 895         const struct perf_thread_map *threads = evlist->core.threads;
 896         /*
 897          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 898          * Its value is decided by evsel's write_backward.
 899          * So &mp should not be passed through const pointer.
 900          */
 901         struct mmap_params mp = { .nr_cblocks = nr_cblocks, .affinity = affinity, .flush = flush,
 902                                   .comp_level = comp_level };
 903
 904         if (!evlist->mmap)
 905                 evlist->mmap = evlist__alloc_mmap(evlist, false);
 906         if (!evlist->mmap)
 907                 return -ENOMEM;
 908
 909         if (evlist->core.pollfd.entries == NULL && perf_evlist__alloc_pollfd(&evlist->core) < 0)
 910                 return -ENOMEM;
 911
 912         evlist->core.mmap_len = evlist__mmap_size(pages);
 913         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 914         mp.mask = evlist->core.mmap_len - page_size - 1;
 915
 916         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 917                                    auxtrace_pages, auxtrace_overwrite);
 918
 919         evlist__for_each_entry(evlist, evsel) {
 920                 if ((evsel->core.attr.read_format & PERF_FORMAT_ID) &&
 921                     evsel->core.sample_id == NULL &&
 922                     perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr) < 0)
 923                         return -ENOMEM;
 924         }
 925
 926         if (perf_cpu_map__empty(cpus))
 927                 return evlist__mmap_per_thread(evlist, &mp);
 928
 929         return evlist__mmap_per_cpu(evlist, &mp);
 930 }
 931
 932 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 933 {
 934         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 935 }
 936
 937 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
 938 {
 939         bool all_threads = (target->per_thread && target->system_wide);
 940         struct perf_cpu_map *cpus;
 941         struct perf_thread_map *threads;
 942
 943         /*
 944          * If specify '-a' and '--per-thread' to perf record, perf record
 945          * will override '--per-thread'. target->per_thread = false and
 946          * target->system_wide = true.
 947          *
 948          * If specify '--per-thread' only to perf record,
 949          * target->per_thread = true and target->system_wide = false.
 950          *
 951          * So target->per_thread && target->system_wide is false.
 952          * For perf record, thread_map__new_str doesn't call
 953          * thread_map__new_all_cpus. That will keep perf record's
 954          * current behavior.
 955          *
 956          * For perf stat, it allows the case that target->per_thread and
 957          * target->system_wide are all true. It means to collect system-wide
 958          * per-thread data. thread_map__new_str will call
 959          * thread_map__new_all_cpus to enumerate all threads.
 960          */
 961         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 962                                       all_threads);
 963
 964         if (!threads)
 965                 return -1;
 966
 967         if (target__uses_dummy_map(target))
 968                 cpus = perf_cpu_map__dummy_new();
 969         else
 970                 cpus = perf_cpu_map__new(target->cpu_list);
 971
 972         if (!cpus)
 973                 goto out_delete_threads;
 974
 975         evlist->core.has_user_cpus = !!target->cpu_list;
 976
 977         perf_evlist__set_maps(&evlist->core, cpus, threads);
 978
 979         return 0;
 980
 981 out_delete_threads:
 982         perf_thread_map__put(threads);
 983         return -1;
 984 }
 985
 986 void __perf_evlist__set_sample_bit(struct evlist *evlist,
 987                                    enum perf_event_sample_format bit)
 988 {
 989         struct evsel *evsel;
 990
 991         evlist__for_each_entry(evlist, evsel)
 992                 __perf_evsel__set_sample_bit(evsel, bit);
 993 }
 994
 995 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
 996                                      enum perf_event_sample_format bit)
 997 {
 998         struct evsel *evsel;
 999
1000         evlist__for_each_entry(evlist, evsel)
1001                 __perf_evsel__reset_sample_bit(evsel, bit);
1002 }
1003
1004 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1005 {
1006         struct evsel *evsel;
1007         int err = 0;
1008
1009         evlist__for_each_entry(evlist, evsel) {
1010                 if (evsel->filter == NULL)
1011                         continue;
1012
1013                 /*
1014                  * filters only work for tracepoint event, which doesn't have cpu limit.
1015                  * So evlist and evsel should always be same.
1016                  */
1017                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1018                 if (err) {
1019                         *err_evsel = evsel;
1020                         break;
1021                 }
1022         }
1023
1024         return err;
1025 }
1026
1027 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1028 {
1029         struct evsel *evsel;
1030         int err = 0;
1031
1032         evlist__for_each_entry(evlist, evsel) {
1033                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1034                         continue;
1035
1036                 err = perf_evsel__set_filter(evsel, filter);
1037                 if (err)
1038                         break;
1039         }
1040
1041         return err;
1042 }
1043
1044 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1045 {
1046         char *filter;
1047         int ret = -1;
1048         size_t i;
1049
1050         for (i = 0; i < npids; ++i) {
1051                 if (i == 0) {
1052                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1053                                 return -1;
1054                 } else {
1055                         char *tmp;
1056
1057                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1058                                 goto out_free;
1059
1060                         free(filter);
1061                         filter = tmp;
1062                 }
1063         }
1064
1065         ret = perf_evlist__set_tp_filter(evlist, filter);
1066 out_free:
1067         free(filter);
1068         return ret;
1069 }
1070
1071 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1072 {
1073         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1074 }
1075
1076 bool perf_evlist__valid_sample_type(struct evlist *evlist)
1077 {
1078         struct evsel *pos;
1079
1080         if (evlist->core.nr_entries == 1)
1081                 return true;
1082
1083         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1084                 return false;
1085
1086         evlist__for_each_entry(evlist, pos) {
1087                 if (pos->id_pos != evlist->id_pos ||
1088                     pos->is_pos != evlist->is_pos)
1089                         return false;
1090         }
1091
1092         return true;
1093 }
1094
1095 u64 __perf_evlist__combined_sample_type(struct evlist *evlist)
1096 {
1097         struct evsel *evsel;
1098
1099         if (evlist->combined_sample_type)
1100                 return evlist->combined_sample_type;
1101
1102         evlist__for_each_entry(evlist, evsel)
1103                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1104
1105         return evlist->combined_sample_type;
1106 }
1107
1108 u64 perf_evlist__combined_sample_type(struct evlist *evlist)
1109 {
1110         evlist->combined_sample_type = 0;
1111         return __perf_evlist__combined_sample_type(evlist);
1112 }
1113
1114 u64 perf_evlist__combined_branch_type(struct evlist *evlist)
1115 {
1116         struct evsel *evsel;
1117         u64 branch_type = 0;
1118
1119         evlist__for_each_entry(evlist, evsel)
1120                 branch_type |= evsel->core.attr.branch_sample_type;
1121         return branch_type;
1122 }
1123
1124 bool perf_evlist__valid_read_format(struct evlist *evlist)
1125 {
1126         struct evsel *first = evlist__first(evlist), *pos = first;
1127         u64 read_format = first->core.attr.read_format;
1128         u64 sample_type = first->core.attr.sample_type;
1129
1130         evlist__for_each_entry(evlist, pos) {
1131                 if (read_format != pos->core.attr.read_format)
1132                         return false;
1133         }
1134
1135         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1136         if ((sample_type & PERF_SAMPLE_READ) &&
1137             !(read_format & PERF_FORMAT_ID)) {
1138                 return false;
1139         }
1140
1141         return true;
1142 }
1143
1144 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1145 {
1146         struct evsel *first = evlist__first(evlist);
1147         struct perf_sample *data;
1148         u64 sample_type;
1149         u16 size = 0;
1150
1151         if (!first->core.attr.sample_id_all)
1152                 goto out;
1153
1154         sample_type = first->core.attr.sample_type;
1155
1156         if (sample_type & PERF_SAMPLE_TID)
1157                 size += sizeof(data->tid) * 2;
1158
1159        if (sample_type & PERF_SAMPLE_TIME)
1160                 size += sizeof(data->time);
1161
1162         if (sample_type & PERF_SAMPLE_ID)
1163                 size += sizeof(data->id);
1164
1165         if (sample_type & PERF_SAMPLE_STREAM_ID)
1166                 size += sizeof(data->stream_id);
1167
1168         if (sample_type & PERF_SAMPLE_CPU)
1169                 size += sizeof(data->cpu) * 2;
1170
1171         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1172                 size += sizeof(data->id);
1173 out:
1174         return size;
1175 }
1176
1177 bool perf_evlist__valid_sample_id_all(struct evlist *evlist)
1178 {
1179         struct evsel *first = evlist__first(evlist), *pos = first;
1180
1181         evlist__for_each_entry_continue(evlist, pos) {
1182                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1183                         return false;
1184         }
1185
1186         return true;
1187 }
1188
1189 bool perf_evlist__sample_id_all(struct evlist *evlist)
1190 {
1191         struct evsel *first = evlist__first(evlist);
1192         return first->core.attr.sample_id_all;
1193 }
1194
1195 void perf_evlist__set_selected(struct evlist *evlist,
1196                                struct evsel *evsel)
1197 {
1198         evlist->selected = evsel;
1199 }
1200
1201 void evlist__close(struct evlist *evlist)
1202 {
1203         struct evsel *evsel;
1204
1205         evlist__for_each_entry_reverse(evlist, evsel)
1206                 evsel__close(evsel);
1207 }
1208
1209 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1210 {
1211         struct perf_cpu_map *cpus;
1212         struct perf_thread_map *threads;
1213         int err = -ENOMEM;
1214
1215         /*
1216          * Try reading /sys/devices/system/cpu/online to get
1217          * an all cpus map.
1218          *
1219          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1220          * code needs an overhaul to properly forward the
1221          * error, and we may not want to do that fallback to a
1222          * default cpu identity map :-\
1223          */
1224         cpus = perf_cpu_map__new(NULL);
1225         if (!cpus)
1226                 goto out;
1227
1228         threads = perf_thread_map__new_dummy();
1229         if (!threads)
1230                 goto out_put;
1231
1232         perf_evlist__set_maps(&evlist->core, cpus, threads);
1233 out:
1234         return err;
1235 out_put:
1236         perf_cpu_map__put(cpus);
1237         goto out;
1238 }
1239
1240 int evlist__open(struct evlist *evlist)
1241 {
1242         struct evsel *evsel;
1243         int err;
1244
1245         /*
1246          * Default: one fd per CPU, all threads, aka systemwide
1247          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1248          */
1249         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1250                 err = perf_evlist__create_syswide_maps(evlist);
1251                 if (err < 0)
1252                         goto out_err;
1253         }
1254
1255         perf_evlist__update_id_pos(evlist);
1256
1257         evlist__for_each_entry(evlist, evsel) {
1258                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1259                 if (err < 0)
1260                         goto out_err;
1261         }
1262
1263         return 0;
1264 out_err:
1265         evlist__close(evlist);
1266         errno = -err;
1267         return err;
1268 }
1269
1270 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1271                                   const char *argv[], bool pipe_output,
1272                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1273 {
1274         int child_ready_pipe[2], go_pipe[2];
1275         char bf;
1276
1277         if (pipe(child_ready_pipe) < 0) {
1278                 perror("failed to create 'ready' pipe");
1279                 return -1;
1280         }
1281
1282         if (pipe(go_pipe) < 0) {
1283                 perror("failed to create 'go' pipe");
1284                 goto out_close_ready_pipe;
1285         }
1286
1287         evlist->workload.pid = fork();
1288         if (evlist->workload.pid < 0) {
1289                 perror("failed to fork");
1290                 goto out_close_pipes;
1291         }
1292
1293         if (!evlist->workload.pid) {
1294                 int ret;
1295
1296                 if (pipe_output)
1297                         dup2(2, 1);
1298
1299                 signal(SIGTERM, SIG_DFL);
1300
1301                 close(child_ready_pipe[0]);
1302                 close(go_pipe[1]);
1303                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1304
1305                 /*
1306                  * Tell the parent we're ready to go
1307                  */
1308                 close(child_ready_pipe[1]);
1309
1310                 /*
1311                  * Wait until the parent tells us to go.
1312                  */
1313                 ret = read(go_pipe[0], &bf, 1);
1314                 /*
1315                  * The parent will ask for the execvp() to be performed by
1316                  * writing exactly one byte, in workload.cork_fd, usually via
1317                  * perf_evlist__start_workload().
1318                  *
1319                  * For cancelling the workload without actually running it,
1320                  * the parent will just close workload.cork_fd, without writing
1321                  * anything, i.e. read will return zero and we just exit()
1322                  * here.
1323                  */
1324                 if (ret != 1) {
1325                         if (ret == -1)
1326                                 perror("unable to read pipe");
1327                         exit(ret);
1328                 }
1329
1330                 execvp(argv[0], (char **)argv);
1331
1332                 if (exec_error) {
1333                         union sigval val;
1334
1335                         val.sival_int = errno;
1336                         if (sigqueue(getppid(), SIGUSR1, val))
1337                                 perror(argv[0]);
1338                 } else
1339                         perror(argv[0]);
1340                 exit(-1);
1341         }
1342
1343         if (exec_error) {
1344                 struct sigaction act = {
1345                         .sa_flags     = SA_SIGINFO,
1346                         .sa_sigaction = exec_error,
1347                 };
1348                 sigaction(SIGUSR1, &act, NULL);
1349         }
1350
1351         if (target__none(target)) {
1352                 if (evlist->core.threads == NULL) {
1353                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1354                                 __func__, __LINE__);
1355                         goto out_close_pipes;
1356                 }
1357                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1358         }
1359
1360         close(child_ready_pipe[1]);
1361         close(go_pipe[0]);
1362         /*
1363          * wait for child to settle
1364          */
1365         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1366                 perror("unable to read pipe");
1367                 goto out_close_pipes;
1368         }
1369
1370         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1371         evlist->workload.cork_fd = go_pipe[1];
1372         close(child_ready_pipe[0]);
1373         return 0;
1374
1375 out_close_pipes:
1376         close(go_pipe[0]);
1377         close(go_pipe[1]);
1378 out_close_ready_pipe:
1379         close(child_ready_pipe[0]);
1380         close(child_ready_pipe[1]);
1381         return -1;
1382 }
1383
1384 int perf_evlist__start_workload(struct evlist *evlist)
1385 {
1386         if (evlist->workload.cork_fd > 0) {
1387                 char bf = 0;
1388                 int ret;
1389                 /*
1390                  * Remove the cork, let it rip!
1391                  */
1392                 ret = write(evlist->workload.cork_fd, &bf, 1);
1393                 if (ret < 0)
1394                         perror("unable to write to pipe");
1395
1396                 close(evlist->workload.cork_fd);
1397                 return ret;
1398         }
1399
1400         return 0;
1401 }
1402
1403 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1404                               struct perf_sample *sample)
1405 {
1406         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1407
1408         if (!evsel)
1409                 return -EFAULT;
1410         return perf_evsel__parse_sample(evsel, event, sample);
1411 }
1412
1413 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1414                                         union perf_event *event,
1415                                         u64 *timestamp)
1416 {
1417         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1418
1419         if (!evsel)
1420                 return -EFAULT;
1421         return perf_evsel__parse_sample_timestamp(evsel, event, timestamp);
1422 }
1423
1424 int perf_evlist__strerror_open(struct evlist *evlist,
1425                                int err, char *buf, size_t size)
1426 {
1427         int printed, value;
1428         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1429
1430         switch (err) {
1431         case EACCES:
1432         case EPERM:
1433                 printed = scnprintf(buf, size,
1434                                     "Error:\t%s.\n"
1435                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1436
1437                 value = perf_event_paranoid();
1438
1439                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1440
1441                 if (value >= 2) {
1442                         printed += scnprintf(buf + printed, size - printed,
1443                                              "For your workloads it needs to be <= 1\nHint:\t");
1444                 }
1445                 printed += scnprintf(buf + printed, size - printed,
1446                                      "For system wide tracing it needs to be set to -1.\n");
1447
1448                 printed += scnprintf(buf + printed, size - printed,
1449                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1450                                     "Hint:\tThe current value is %d.", value);
1451                 break;
1452         case EINVAL: {
1453                 struct evsel *first = evlist__first(evlist);
1454                 int max_freq;
1455
1456                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1457                         goto out_default;
1458
1459                 if (first->core.attr.sample_freq < (u64)max_freq)
1460                         goto out_default;
1461
1462                 printed = scnprintf(buf, size,
1463                                     "Error:\t%s.\n"
1464                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1465                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1466                                     emsg, max_freq, first->core.attr.sample_freq);
1467                 break;
1468         }
1469         default:
1470 out_default:
1471                 scnprintf(buf, size, "%s", emsg);
1472                 break;
1473         }
1474
1475         return 0;
1476 }
1477
1478 int perf_evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1479 {
1480         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1481         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1482
1483         switch (err) {
1484         case EPERM:
1485                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1486                 printed += scnprintf(buf + printed, size - printed,
1487                                      "Error:\t%s.\n"
1488                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1489                                      "Hint:\tTried using %zd kB.\n",
1490                                      emsg, pages_max_per_user, pages_attempted);
1491
1492                 if (pages_attempted >= pages_max_per_user) {
1493                         printed += scnprintf(buf + printed, size - printed,
1494                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1495                                              pages_max_per_user + pages_attempted);
1496                 }
1497
1498                 printed += scnprintf(buf + printed, size - printed,
1499                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1500                 break;
1501         default:
1502                 scnprintf(buf, size, "%s", emsg);
1503                 break;
1504         }
1505
1506         return 0;
1507 }
1508
1509 void perf_evlist__to_front(struct evlist *evlist,
1510                            struct evsel *move_evsel)
1511 {
1512         struct evsel *evsel, *n;
1513         LIST_HEAD(move);
1514
1515         if (move_evsel == evlist__first(evlist))
1516                 return;
1517
1518         evlist__for_each_entry_safe(evlist, n, evsel) {
1519                 if (evsel->leader == move_evsel->leader)
1520                         list_move_tail(&evsel->core.node, &move);
1521         }
1522
1523         list_splice(&move, &evlist->core.entries);
1524 }
1525
1526 void perf_evlist__set_tracking_event(struct evlist *evlist,
1527                                      struct evsel *tracking_evsel)
1528 {
1529         struct evsel *evsel;
1530
1531         if (tracking_evsel->tracking)
1532                 return;
1533
1534         evlist__for_each_entry(evlist, evsel) {
1535                 if (evsel != tracking_evsel)
1536                         evsel->tracking = false;
1537         }
1538
1539         tracking_evsel->tracking = true;
1540 }
1541
1542 struct evsel *
1543 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1544                                const char *str)
1545 {
1546         struct evsel *evsel;
1547
1548         evlist__for_each_entry(evlist, evsel) {
1549                 if (!evsel->name)
1550                         continue;
1551                 if (strcmp(str, evsel->name) == 0)
1552                         return evsel;
1553         }
1554
1555         return NULL;
1556 }
1557
1558 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1559                                   enum bkw_mmap_state state)
1560 {
1561         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1562         enum action {
1563                 NONE,
1564                 PAUSE,
1565                 RESUME,
1566         } action = NONE;
1567
1568         if (!evlist->overwrite_mmap)
1569                 return;
1570
1571         switch (old_state) {
1572         case BKW_MMAP_NOTREADY: {
1573                 if (state != BKW_MMAP_RUNNING)
1574                         goto state_err;
1575                 break;
1576         }
1577         case BKW_MMAP_RUNNING: {
1578                 if (state != BKW_MMAP_DATA_PENDING)
1579                         goto state_err;
1580                 action = PAUSE;
1581                 break;
1582         }
1583         case BKW_MMAP_DATA_PENDING: {
1584                 if (state != BKW_MMAP_EMPTY)
1585                         goto state_err;
1586                 break;
1587         }
1588         case BKW_MMAP_EMPTY: {
1589                 if (state != BKW_MMAP_RUNNING)
1590                         goto state_err;
1591                 action = RESUME;
1592                 break;
1593         }
1594         default:
1595                 WARN_ONCE(1, "Shouldn't get there\n");
1596         }
1597
1598         evlist->bkw_mmap_state = state;
1599
1600         switch (action) {
1601         case PAUSE:
1602                 perf_evlist__pause(evlist);
1603                 break;
1604         case RESUME:
1605                 perf_evlist__resume(evlist);
1606                 break;
1607         case NONE:
1608         default:
1609                 break;
1610         }
1611
1612 state_err:
1613         return;
1614 }
1615
1616 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1617 {
1618         struct evsel *evsel;
1619
1620         evlist__for_each_entry(evlist, evsel) {
1621                 if (!evsel->core.attr.exclude_kernel)
1622                         return false;
1623         }
1624
1625         return true;
1626 }
1627
1628 /*
1629  * Events in data file are not collect in groups, but we still want
1630  * the group display. Set the artificial group and set the leader's
1631  * forced_leader flag to notify the display code.
1632  */
1633 void perf_evlist__force_leader(struct evlist *evlist)
1634 {
1635         if (!evlist->nr_groups) {
1636                 struct evsel *leader = evlist__first(evlist);
1637
1638                 perf_evlist__set_leader(evlist);
1639                 leader->forced_leader = true;
1640         }
1641 }
1642
1643 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1644                                                  struct evsel *evsel)
1645 {
1646         struct evsel *c2, *leader;
1647         bool is_open = true;
1648
1649         leader = evsel->leader;
1650         pr_debug("Weak group for %s/%d failed\n",
1651                         leader->name, leader->core.nr_members);
1652
1653         /*
1654          * for_each_group_member doesn't work here because it doesn't
1655          * include the first entry.
1656          */
1657         evlist__for_each_entry(evsel_list, c2) {
1658                 if (c2 == evsel)
1659                         is_open = false;
1660                 if (c2->leader == leader) {
1661                         if (is_open)
1662                                 perf_evsel__close(&c2->core);
1663                         c2->leader = c2;
1664                         c2->core.nr_members = 0;
1665                 }
1666         }
1667         return leader;
1668 }
1669
1670 int perf_evlist__add_sb_event(struct evlist **evlist,
1671                               struct perf_event_attr *attr,
1672                               perf_evsel__sb_cb_t cb,
1673                               void *data)
1674 {
1675         struct evsel *evsel;
1676         bool new_evlist = (*evlist) == NULL;
1677
1678         if (*evlist == NULL)
1679                 *evlist = evlist__new();
1680         if (*evlist == NULL)
1681                 return -1;
1682
1683         if (!attr->sample_id_all) {
1684                 pr_warning("enabling sample_id_all for all side band events\n");
1685                 attr->sample_id_all = 1;
1686         }
1687
1688         evsel = perf_evsel__new_idx(attr, (*evlist)->core.nr_entries);
1689         if (!evsel)
1690                 goto out_err;
1691
1692         evsel->side_band.cb = cb;
1693         evsel->side_band.data = data;
1694         evlist__add(*evlist, evsel);
1695         return 0;
1696
1697 out_err:
1698         if (new_evlist) {
1699                 evlist__delete(*evlist);
1700                 *evlist = NULL;
1701         }
1702         return -1;
1703 }
1704
1705 static void *perf_evlist__poll_thread(void *arg)
1706 {
1707         struct evlist *evlist = arg;
1708         bool draining = false;
1709         int i, done = 0;
1710         /*
1711          * In order to read symbols from other namespaces perf to needs to call
1712          * setns(2).  This isn't permitted if the struct_fs has multiple users.
1713          * unshare(2) the fs so that we may continue to setns into namespaces
1714          * that we're observing when, for instance, reading the build-ids at
1715          * the end of a 'perf record' session.
1716          */
1717         unshare(CLONE_FS);
1718
1719         while (!done) {
1720                 bool got_data = false;
1721
1722                 if (evlist->thread.done)
1723                         draining = true;
1724
1725                 if (!draining)
1726                         evlist__poll(evlist, 1000);
1727
1728                 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1729                         struct mmap *map = &evlist->mmap[i];
1730                         union perf_event *event;
1731
1732                         if (perf_mmap__read_init(map))
1733                                 continue;
1734                         while ((event = perf_mmap__read_event(map)) != NULL) {
1735                                 struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1736
1737                                 if (evsel && evsel->side_band.cb)
1738                                         evsel->side_band.cb(event, evsel->side_band.data);
1739                                 else
1740                                         pr_warning("cannot locate proper evsel for the side band event\n");
1741
1742                                 perf_mmap__consume(map);
1743                                 got_data = true;
1744                         }
1745                         perf_mmap__read_done(map);
1746                 }
1747
1748                 if (draining && !got_data)
1749                         break;
1750         }
1751         return NULL;
1752 }
1753
1754 int perf_evlist__start_sb_thread(struct evlist *evlist,
1755                                  struct target *target)
1756 {
1757         struct evsel *counter;
1758
1759         if (!evlist)
1760                 return 0;
1761
1762         if (perf_evlist__create_maps(evlist, target))
1763                 goto out_delete_evlist;
1764
1765         evlist__for_each_entry(evlist, counter) {
1766                 if (evsel__open(counter, evlist->core.cpus,
1767                                      evlist->core.threads) < 0)
1768                         goto out_delete_evlist;
1769         }
1770
1771         if (evlist__mmap(evlist, UINT_MAX))
1772                 goto out_delete_evlist;
1773
1774         evlist__for_each_entry(evlist, counter) {
1775                 if (evsel__enable(counter))
1776                         goto out_delete_evlist;
1777         }
1778
1779         evlist->thread.done = 0;
1780         if (pthread_create(&evlist->thread.th, NULL, perf_evlist__poll_thread, evlist))
1781                 goto out_delete_evlist;
1782
1783         return 0;
1784
1785 out_delete_evlist:
1786         evlist__delete(evlist);
1787         evlist = NULL;
1788         return -1;
1789 }
1790
1791 void perf_evlist__stop_sb_thread(struct evlist *evlist)
1792 {
1793         if (!evlist)
1794                 return;
1795         evlist->thread.done = 1;
1796         pthread_join(evlist->thread.th, NULL);
1797         evlist__delete(evlist);
1798 }