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 "util/perf_api_probe.h"
  27 #include <signal.h>
  28 #include <unistd.h>
  29 #include <sched.h>
  30 #include <stdlib.h>
  31
  32 #include "parse-events.h"
  33 #include <subcmd/parse-options.h>
  34
  35 #include <fcntl.h>
  36 #include <sys/ioctl.h>
  37 #include <sys/mman.h>
  38
  39 #include <linux/bitops.h>
  40 #include <linux/hash.h>
  41 #include <linux/log2.h>
  42 #include <linux/err.h>
  43 #include <linux/string.h>
  44 #include <linux/zalloc.h>
  45 #include <perf/evlist.h>
  46 #include <perf/evsel.h>
  47 #include <perf/cpumap.h>
  48 #include <perf/mmap.h>
  49
  50 #include <internal/xyarray.h>
  51
  52 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  53 int sigqueue(pid_t pid, int sig, const union sigval value);
  54 #endif
  55
  56 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  57 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  58
  59 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  60                   struct perf_thread_map *threads)
  61 {
  62         perf_evlist__init(&evlist->core);
  63         perf_evlist__set_maps(&evlist->core, cpus, threads);
  64         evlist->workload.pid = -1;
  65         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  66         evlist->ctl_fd.fd = -1;
  67         evlist->ctl_fd.ack = -1;
  68         evlist->ctl_fd.pos = -1;
  69 }
  70
  71 struct evlist *evlist__new(void)
  72 {
  73         struct evlist *evlist = zalloc(sizeof(*evlist));
  74
  75         if (evlist != NULL)
  76                 evlist__init(evlist, NULL, NULL);
  77
  78         return evlist;
  79 }
  80
  81 struct evlist *perf_evlist__new_default(void)
  82 {
  83         struct evlist *evlist = evlist__new();
  84
  85         if (evlist && evlist__add_default(evlist)) {
  86                 evlist__delete(evlist);
  87                 evlist = NULL;
  88         }
  89
  90         return evlist;
  91 }
  92
  93 struct evlist *perf_evlist__new_dummy(void)
  94 {
  95         struct evlist *evlist = evlist__new();
  96
  97         if (evlist && evlist__add_dummy(evlist)) {
  98                 evlist__delete(evlist);
  99                 evlist = NULL;
 100         }
 101
 102         return evlist;
 103 }
 104
 105 /**
 106  * perf_evlist__set_id_pos - set the positions of event ids.
 107  * @evlist: selected event list
 108  *
 109  * Events with compatible sample types all have the same id_pos
 110  * and is_pos.  For convenience, put a copy on evlist.
 111  */
 112 void perf_evlist__set_id_pos(struct evlist *evlist)
 113 {
 114         struct evsel *first = evlist__first(evlist);
 115
 116         evlist->id_pos = first->id_pos;
 117         evlist->is_pos = first->is_pos;
 118 }
 119
 120 static void perf_evlist__update_id_pos(struct evlist *evlist)
 121 {
 122         struct evsel *evsel;
 123
 124         evlist__for_each_entry(evlist, evsel)
 125                 evsel__calc_id_pos(evsel);
 126
 127         perf_evlist__set_id_pos(evlist);
 128 }
 129
 130 static void evlist__purge(struct evlist *evlist)
 131 {
 132         struct evsel *pos, *n;
 133
 134         evlist__for_each_entry_safe(evlist, n, pos) {
 135                 list_del_init(&pos->core.node);
 136                 pos->evlist = NULL;
 137                 evsel__delete(pos);
 138         }
 139
 140         evlist->core.nr_entries = 0;
 141 }
 142
 143 void evlist__exit(struct evlist *evlist)
 144 {
 145         zfree(&evlist->mmap);
 146         zfree(&evlist->overwrite_mmap);
 147         perf_evlist__exit(&evlist->core);
 148 }
 149
 150 void evlist__delete(struct evlist *evlist)
 151 {
 152         if (evlist == NULL)
 153                 return;
 154
 155         evlist__munmap(evlist);
 156         evlist__close(evlist);
 157         evlist__purge(evlist);
 158         evlist__exit(evlist);
 159         free(evlist);
 160 }
 161
 162 void evlist__add(struct evlist *evlist, struct evsel *entry)
 163 {
 164         entry->evlist = evlist;
 165         entry->idx = evlist->core.nr_entries;
 166         entry->tracking = !entry->idx;
 167
 168         perf_evlist__add(&evlist->core, &entry->core);
 169
 170         if (evlist->core.nr_entries == 1)
 171                 perf_evlist__set_id_pos(evlist);
 172 }
 173
 174 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 175 {
 176         evsel->evlist = NULL;
 177         perf_evlist__remove(&evlist->core, &evsel->core);
 178 }
 179
 180 void perf_evlist__splice_list_tail(struct evlist *evlist,
 181                                    struct list_head *list)
 182 {
 183         struct evsel *evsel, *temp;
 184
 185         __evlist__for_each_entry_safe(list, temp, evsel) {
 186                 list_del_init(&evsel->core.node);
 187                 evlist__add(evlist, evsel);
 188         }
 189 }
 190
 191 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 192                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 193 {
 194         struct evsel *evsel;
 195         size_t i;
 196         int err;
 197
 198         for (i = 0; i < nr_assocs; i++) {
 199                 // Adding a handler for an event not in this evlist, just ignore it.
 200                 evsel = perf_evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 201                 if (evsel == NULL)
 202                         continue;
 203
 204                 err = -EEXIST;
 205                 if (evsel->handler != NULL)
 206                         goto out;
 207                 evsel->handler = assocs[i].handler;
 208         }
 209
 210         err = 0;
 211 out:
 212         return err;
 213 }
 214
 215 void __perf_evlist__set_leader(struct list_head *list)
 216 {
 217         struct evsel *evsel, *leader;
 218
 219         leader = list_entry(list->next, struct evsel, core.node);
 220         evsel = list_entry(list->prev, struct evsel, core.node);
 221
 222         leader->core.nr_members = evsel->idx - leader->idx + 1;
 223
 224         __evlist__for_each_entry(list, evsel) {
 225                 evsel->leader = leader;
 226         }
 227 }
 228
 229 void perf_evlist__set_leader(struct evlist *evlist)
 230 {
 231         if (evlist->core.nr_entries) {
 232                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 233                 __perf_evlist__set_leader(&evlist->core.entries);
 234         }
 235 }
 236
 237 int __evlist__add_default(struct evlist *evlist, bool precise)
 238 {
 239         struct evsel *evsel = evsel__new_cycles(precise);
 240
 241         if (evsel == NULL)
 242                 return -ENOMEM;
 243
 244         evlist__add(evlist, evsel);
 245         return 0;
 246 }
 247
 248 int evlist__add_dummy(struct evlist *evlist)
 249 {
 250         struct perf_event_attr attr = {
 251                 .type   = PERF_TYPE_SOFTWARE,
 252                 .config = PERF_COUNT_SW_DUMMY,
 253                 .size   = sizeof(attr), /* to capture ABI version */
 254         };
 255         struct evsel *evsel = evsel__new_idx(&attr, evlist->core.nr_entries);
 256
 257         if (evsel == NULL)
 258                 return -ENOMEM;
 259
 260         evlist__add(evlist, evsel);
 261         return 0;
 262 }
 263
 264 static int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 265 {
 266         struct evsel *evsel, *n;
 267         LIST_HEAD(head);
 268         size_t i;
 269
 270         for (i = 0; i < nr_attrs; i++) {
 271                 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 272                 if (evsel == NULL)
 273                         goto out_delete_partial_list;
 274                 list_add_tail(&evsel->core.node, &head);
 275         }
 276
 277         perf_evlist__splice_list_tail(evlist, &head);
 278
 279         return 0;
 280
 281 out_delete_partial_list:
 282         __evlist__for_each_entry_safe(&head, n, evsel)
 283                 evsel__delete(evsel);
 284         return -1;
 285 }
 286
 287 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 288 {
 289         size_t i;
 290
 291         for (i = 0; i < nr_attrs; i++)
 292                 event_attr_init(attrs + i);
 293
 294         return evlist__add_attrs(evlist, attrs, nr_attrs);
 295 }
 296
 297 struct evsel *
 298 perf_evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 299 {
 300         struct evsel *evsel;
 301
 302         evlist__for_each_entry(evlist, evsel) {
 303                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 304                     (int)evsel->core.attr.config == id)
 305                         return evsel;
 306         }
 307
 308         return NULL;
 309 }
 310
 311 struct evsel *
 312 perf_evlist__find_tracepoint_by_name(struct evlist *evlist,
 313                                      const char *name)
 314 {
 315         struct evsel *evsel;
 316
 317         evlist__for_each_entry(evlist, evsel) {
 318                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 319                     (strcmp(evsel->name, name) == 0))
 320                         return evsel;
 321         }
 322
 323         return NULL;
 324 }
 325
 326 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
 327 {
 328         struct evsel *evsel = evsel__newtp(sys, name);
 329
 330         if (IS_ERR(evsel))
 331                 return -1;
 332
 333         evsel->handler = handler;
 334         evlist__add(evlist, evsel);
 335         return 0;
 336 }
 337
 338 static int perf_evlist__nr_threads(struct evlist *evlist,
 339                                    struct evsel *evsel)
 340 {
 341         if (evsel->core.system_wide)
 342                 return 1;
 343         else
 344                 return perf_thread_map__nr(evlist->core.threads);
 345 }
 346
 347 void evlist__cpu_iter_start(struct evlist *evlist)
 348 {
 349         struct evsel *pos;
 350
 351         /*
 352          * Reset the per evsel cpu_iter. This is needed because
 353          * each evsel's cpumap may have a different index space,
 354          * and some operations need the index to modify
 355          * the FD xyarray (e.g. open, close)
 356          */
 357         evlist__for_each_entry(evlist, pos)
 358                 pos->cpu_iter = 0;
 359 }
 360
 361 bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
 362 {
 363         if (ev->cpu_iter >= ev->core.cpus->nr)
 364                 return true;
 365         if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
 366                 return true;
 367         return false;
 368 }
 369
 370 bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
 371 {
 372         if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
 373                 ev->cpu_iter++;
 374                 return false;
 375         }
 376         return true;
 377 }
 378
 379 void evlist__disable(struct evlist *evlist)
 380 {
 381         struct evsel *pos;
 382         struct affinity affinity;
 383         int cpu, i, imm = 0;
 384         bool has_imm = false;
 385
 386         if (affinity__setup(&affinity) < 0)
 387                 return;
 388
 389         /* Disable 'immediate' events last */
 390         for (imm = 0; imm <= 1; imm++) {
 391                 evlist__for_each_cpu(evlist, i, cpu) {
 392                         affinity__set(&affinity, cpu);
 393
 394                         evlist__for_each_entry(evlist, pos) {
 395                                 if (evsel__cpu_iter_skip(pos, cpu))
 396                                         continue;
 397                                 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
 398                                         continue;
 399                                 if (pos->immediate)
 400                                         has_imm = true;
 401                                 if (pos->immediate != imm)
 402                                         continue;
 403                                 evsel__disable_cpu(pos, pos->cpu_iter - 1);
 404                         }
 405                 }
 406                 if (!has_imm)
 407                         break;
 408         }
 409
 410         affinity__cleanup(&affinity);
 411         evlist__for_each_entry(evlist, pos) {
 412                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 413                         continue;
 414                 pos->disabled = true;
 415         }
 416
 417         evlist->enabled = false;
 418 }
 419
 420 void evlist__enable(struct evlist *evlist)
 421 {
 422         struct evsel *pos;
 423         struct affinity affinity;
 424         int cpu, i;
 425
 426         if (affinity__setup(&affinity) < 0)
 427                 return;
 428
 429         evlist__for_each_cpu(evlist, i, cpu) {
 430                 affinity__set(&affinity, cpu);
 431
 432                 evlist__for_each_entry(evlist, pos) {
 433                         if (evsel__cpu_iter_skip(pos, cpu))
 434                                 continue;
 435                         if (!evsel__is_group_leader(pos) || !pos->core.fd)
 436                                 continue;
 437                         evsel__enable_cpu(pos, pos->cpu_iter - 1);
 438                 }
 439         }
 440         affinity__cleanup(&affinity);
 441         evlist__for_each_entry(evlist, pos) {
 442                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 443                         continue;
 444                 pos->disabled = false;
 445         }
 446
 447         evlist->enabled = true;
 448 }
 449
 450 void perf_evlist__toggle_enable(struct evlist *evlist)
 451 {
 452         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 453 }
 454
 455 static int perf_evlist__enable_event_cpu(struct evlist *evlist,
 456                                          struct evsel *evsel, int cpu)
 457 {
 458         int thread;
 459         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 460
 461         if (!evsel->core.fd)
 462                 return -EINVAL;
 463
 464         for (thread = 0; thread < nr_threads; thread++) {
 465                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 466                 if (err)
 467                         return err;
 468         }
 469         return 0;
 470 }
 471
 472 static int perf_evlist__enable_event_thread(struct evlist *evlist,
 473                                             struct evsel *evsel,
 474                                             int thread)
 475 {
 476         int cpu;
 477         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 478
 479         if (!evsel->core.fd)
 480                 return -EINVAL;
 481
 482         for (cpu = 0; cpu < nr_cpus; cpu++) {
 483                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 484                 if (err)
 485                         return err;
 486         }
 487         return 0;
 488 }
 489
 490 int perf_evlist__enable_event_idx(struct evlist *evlist,
 491                                   struct evsel *evsel, int idx)
 492 {
 493         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 494
 495         if (per_cpu_mmaps)
 496                 return perf_evlist__enable_event_cpu(evlist, evsel, idx);
 497         else
 498                 return perf_evlist__enable_event_thread(evlist, evsel, idx);
 499 }
 500
 501 int evlist__add_pollfd(struct evlist *evlist, int fd)
 502 {
 503         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
 504 }
 505
 506 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 507 {
 508         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 509 }
 510
 511 int evlist__poll(struct evlist *evlist, int timeout)
 512 {
 513         return perf_evlist__poll(&evlist->core, timeout);
 514 }
 515
 516 struct perf_sample_id *perf_evlist__id2sid(struct evlist *evlist, u64 id)
 517 {
 518         struct hlist_head *head;
 519         struct perf_sample_id *sid;
 520         int hash;
 521
 522         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 523         head = &evlist->core.heads[hash];
 524
 525         hlist_for_each_entry(sid, head, node)
 526                 if (sid->id == id)
 527                         return sid;
 528
 529         return NULL;
 530 }
 531
 532 struct evsel *perf_evlist__id2evsel(struct evlist *evlist, u64 id)
 533 {
 534         struct perf_sample_id *sid;
 535
 536         if (evlist->core.nr_entries == 1 || !id)
 537                 return evlist__first(evlist);
 538
 539         sid = perf_evlist__id2sid(evlist, id);
 540         if (sid)
 541                 return container_of(sid->evsel, struct evsel, core);
 542
 543         if (!evlist__sample_id_all(evlist))
 544                 return evlist__first(evlist);
 545
 546         return NULL;
 547 }
 548
 549 struct evsel *perf_evlist__id2evsel_strict(struct evlist *evlist,
 550                                                 u64 id)
 551 {
 552         struct perf_sample_id *sid;
 553
 554         if (!id)
 555                 return NULL;
 556
 557         sid = perf_evlist__id2sid(evlist, id);
 558         if (sid)
 559                 return container_of(sid->evsel, struct evsel, core);
 560
 561         return NULL;
 562 }
 563
 564 static int perf_evlist__event2id(struct evlist *evlist,
 565                                  union perf_event *event, u64 *id)
 566 {
 567         const __u64 *array = event->sample.array;
 568         ssize_t n;
 569
 570         n = (event->header.size - sizeof(event->header)) >> 3;
 571
 572         if (event->header.type == PERF_RECORD_SAMPLE) {
 573                 if (evlist->id_pos >= n)
 574                         return -1;
 575                 *id = array[evlist->id_pos];
 576         } else {
 577                 if (evlist->is_pos > n)
 578                         return -1;
 579                 n -= evlist->is_pos;
 580                 *id = array[n];
 581         }
 582         return 0;
 583 }
 584
 585 struct evsel *perf_evlist__event2evsel(struct evlist *evlist,
 586                                             union perf_event *event)
 587 {
 588         struct evsel *first = evlist__first(evlist);
 589         struct hlist_head *head;
 590         struct perf_sample_id *sid;
 591         int hash;
 592         u64 id;
 593
 594         if (evlist->core.nr_entries == 1)
 595                 return first;
 596
 597         if (!first->core.attr.sample_id_all &&
 598             event->header.type != PERF_RECORD_SAMPLE)
 599                 return first;
 600
 601         if (perf_evlist__event2id(evlist, event, &id))
 602                 return NULL;
 603
 604         /* Synthesized events have an id of zero */
 605         if (!id)
 606                 return first;
 607
 608         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 609         head = &evlist->core.heads[hash];
 610
 611         hlist_for_each_entry(sid, head, node) {
 612                 if (sid->id == id)
 613                         return container_of(sid->evsel, struct evsel, core);
 614         }
 615         return NULL;
 616 }
 617
 618 static int perf_evlist__set_paused(struct evlist *evlist, bool value)
 619 {
 620         int i;
 621
 622         if (!evlist->overwrite_mmap)
 623                 return 0;
 624
 625         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 626                 int fd = evlist->overwrite_mmap[i].core.fd;
 627                 int err;
 628
 629                 if (fd < 0)
 630                         continue;
 631                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 632                 if (err)
 633                         return err;
 634         }
 635         return 0;
 636 }
 637
 638 static int perf_evlist__pause(struct evlist *evlist)
 639 {
 640         return perf_evlist__set_paused(evlist, true);
 641 }
 642
 643 static int perf_evlist__resume(struct evlist *evlist)
 644 {
 645         return perf_evlist__set_paused(evlist, false);
 646 }
 647
 648 static void evlist__munmap_nofree(struct evlist *evlist)
 649 {
 650         int i;
 651
 652         if (evlist->mmap)
 653                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 654                         perf_mmap__munmap(&evlist->mmap[i].core);
 655
 656         if (evlist->overwrite_mmap)
 657                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 658                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 659 }
 660
 661 void evlist__munmap(struct evlist *evlist)
 662 {
 663         evlist__munmap_nofree(evlist);
 664         zfree(&evlist->mmap);
 665         zfree(&evlist->overwrite_mmap);
 666 }
 667
 668 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 669 {
 670         struct mmap *m = container_of(map, struct mmap, core);
 671
 672         mmap__munmap(m);
 673 }
 674
 675 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 676                                        bool overwrite)
 677 {
 678         int i;
 679         struct mmap *map;
 680
 681         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 682         if (!map)
 683                 return NULL;
 684
 685         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 686                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 687
 688                 /*
 689                  * When the perf_mmap() call is made we grab one refcount, plus
 690                  * one extra to let perf_mmap__consume() get the last
 691                  * events after all real references (perf_mmap__get()) are
 692                  * dropped.
 693                  *
 694                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 695                  * thus does perf_mmap__get() on it.
 696                  */
 697                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 698         }
 699
 700         return map;
 701 }
 702
 703 static void
 704 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 705                          struct perf_mmap_param *_mp,
 706                          int idx, bool per_cpu)
 707 {
 708         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 709         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 710
 711         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
 712 }
 713
 714 static struct perf_mmap*
 715 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 716 {
 717         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 718         struct mmap *maps;
 719
 720         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 721
 722         if (!maps) {
 723                 maps = evlist__alloc_mmap(evlist, overwrite);
 724                 if (!maps)
 725                         return NULL;
 726
 727                 if (overwrite) {
 728                         evlist->overwrite_mmap = maps;
 729                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 730                                 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 731                 } else {
 732                         evlist->mmap = maps;
 733                 }
 734         }
 735
 736         return &maps[idx].core;
 737 }
 738
 739 static int
 740 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 741                           int output, int cpu)
 742 {
 743         struct mmap *map = container_of(_map, struct mmap, core);
 744         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 745
 746         return mmap__mmap(map, mp, output, cpu);
 747 }
 748
 749 unsigned long perf_event_mlock_kb_in_pages(void)
 750 {
 751         unsigned long pages;
 752         int max;
 753
 754         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 755                 /*
 756                  * Pick a once upon a time good value, i.e. things look
 757                  * strange since we can't read a sysctl value, but lets not
 758                  * die yet...
 759                  */
 760                 max = 512;
 761         } else {
 762                 max -= (page_size / 1024);
 763         }
 764
 765         pages = (max * 1024) / page_size;
 766         if (!is_power_of_2(pages))
 767                 pages = rounddown_pow_of_two(pages);
 768
 769         return pages;
 770 }
 771
 772 size_t evlist__mmap_size(unsigned long pages)
 773 {
 774         if (pages == UINT_MAX)
 775                 pages = perf_event_mlock_kb_in_pages();
 776         else if (!is_power_of_2(pages))
 777                 return 0;
 778
 779         return (pages + 1) * page_size;
 780 }
 781
 782 static long parse_pages_arg(const char *str, unsigned long min,
 783                             unsigned long max)
 784 {
 785         unsigned long pages, val;
 786         static struct parse_tag tags[] = {
 787                 { .tag  = 'B', .mult = 1       },
 788                 { .tag  = 'K', .mult = 1 << 10 },
 789                 { .tag  = 'M', .mult = 1 << 20 },
 790                 { .tag  = 'G', .mult = 1 << 30 },
 791                 { .tag  = 0 },
 792         };
 793
 794         if (str == NULL)
 795                 return -EINVAL;
 796
 797         val = parse_tag_value(str, tags);
 798         if (val != (unsigned long) -1) {
 799                 /* we got file size value */
 800                 pages = PERF_ALIGN(val, page_size) / page_size;
 801         } else {
 802                 /* we got pages count value */
 803                 char *eptr;
 804                 pages = strtoul(str, &eptr, 10);
 805                 if (*eptr != '\0')
 806                         return -EINVAL;
 807         }
 808
 809         if (pages == 0 && min == 0) {
 810                 /* leave number of pages at 0 */
 811         } else if (!is_power_of_2(pages)) {
 812                 char buf[100];
 813
 814                 /* round pages up to next power of 2 */
 815                 pages = roundup_pow_of_two(pages);
 816                 if (!pages)
 817                         return -EINVAL;
 818
 819                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 820                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 821                         buf, pages);
 822         }
 823
 824         if (pages > max)
 825                 return -EINVAL;
 826
 827         return pages;
 828 }
 829
 830 int __perf_evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 831 {
 832         unsigned long max = UINT_MAX;
 833         long pages;
 834
 835         if (max > SIZE_MAX / page_size)
 836                 max = SIZE_MAX / page_size;
 837
 838         pages = parse_pages_arg(str, 1, max);
 839         if (pages < 0) {
 840                 pr_err("Invalid argument for --mmap_pages/-m\n");
 841                 return -1;
 842         }
 843
 844         *mmap_pages = pages;
 845         return 0;
 846 }
 847
 848 int perf_evlist__parse_mmap_pages(const struct option *opt, const char *str,
 849                                   int unset __maybe_unused)
 850 {
 851         return __perf_evlist__parse_mmap_pages(opt->value, str);
 852 }
 853
 854 /**
 855  * evlist__mmap_ex - Create mmaps to receive events.
 856  * @evlist: list of events
 857  * @pages: map length in pages
 858  * @overwrite: overwrite older events?
 859  * @auxtrace_pages - auxtrace map length in pages
 860  * @auxtrace_overwrite - overwrite older auxtrace data?
 861  *
 862  * If @overwrite is %false the user needs to signal event consumption using
 863  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 864  * automatically.
 865  *
 866  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 867  * consumption using auxtrace_mmap__write_tail().
 868  *
 869  * Return: %0 on success, negative error code otherwise.
 870  */
 871 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 872                          unsigned int auxtrace_pages,
 873                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 874                          int comp_level)
 875 {
 876         /*
 877          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 878          * Its value is decided by evsel's write_backward.
 879          * So &mp should not be passed through const pointer.
 880          */
 881         struct mmap_params mp = {
 882                 .nr_cblocks     = nr_cblocks,
 883                 .affinity       = affinity,
 884                 .flush          = flush,
 885                 .comp_level     = comp_level
 886         };
 887         struct perf_evlist_mmap_ops ops = {
 888                 .idx  = perf_evlist__mmap_cb_idx,
 889                 .get  = perf_evlist__mmap_cb_get,
 890                 .mmap = perf_evlist__mmap_cb_mmap,
 891         };
 892
 893         evlist->core.mmap_len = evlist__mmap_size(pages);
 894         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 895
 896         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 897                                    auxtrace_pages, auxtrace_overwrite);
 898
 899         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 900 }
 901
 902 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 903 {
 904         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 905 }
 906
 907 int perf_evlist__create_maps(struct evlist *evlist, struct target *target)
 908 {
 909         bool all_threads = (target->per_thread && target->system_wide);
 910         struct perf_cpu_map *cpus;
 911         struct perf_thread_map *threads;
 912
 913         /*
 914          * If specify '-a' and '--per-thread' to perf record, perf record
 915          * will override '--per-thread'. target->per_thread = false and
 916          * target->system_wide = true.
 917          *
 918          * If specify '--per-thread' only to perf record,
 919          * target->per_thread = true and target->system_wide = false.
 920          *
 921          * So target->per_thread && target->system_wide is false.
 922          * For perf record, thread_map__new_str doesn't call
 923          * thread_map__new_all_cpus. That will keep perf record's
 924          * current behavior.
 925          *
 926          * For perf stat, it allows the case that target->per_thread and
 927          * target->system_wide are all true. It means to collect system-wide
 928          * per-thread data. thread_map__new_str will call
 929          * thread_map__new_all_cpus to enumerate all threads.
 930          */
 931         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 932                                       all_threads);
 933
 934         if (!threads)
 935                 return -1;
 936
 937         if (target__uses_dummy_map(target))
 938                 cpus = perf_cpu_map__dummy_new();
 939         else
 940                 cpus = perf_cpu_map__new(target->cpu_list);
 941
 942         if (!cpus)
 943                 goto out_delete_threads;
 944
 945         evlist->core.has_user_cpus = !!target->cpu_list;
 946
 947         perf_evlist__set_maps(&evlist->core, cpus, threads);
 948
 949         return 0;
 950
 951 out_delete_threads:
 952         perf_thread_map__put(threads);
 953         return -1;
 954 }
 955
 956 void __perf_evlist__set_sample_bit(struct evlist *evlist,
 957                                    enum perf_event_sample_format bit)
 958 {
 959         struct evsel *evsel;
 960
 961         evlist__for_each_entry(evlist, evsel)
 962                 __evsel__set_sample_bit(evsel, bit);
 963 }
 964
 965 void __perf_evlist__reset_sample_bit(struct evlist *evlist,
 966                                      enum perf_event_sample_format bit)
 967 {
 968         struct evsel *evsel;
 969
 970         evlist__for_each_entry(evlist, evsel)
 971                 __evsel__reset_sample_bit(evsel, bit);
 972 }
 973
 974 int perf_evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 975 {
 976         struct evsel *evsel;
 977         int err = 0;
 978
 979         evlist__for_each_entry(evlist, evsel) {
 980                 if (evsel->filter == NULL)
 981                         continue;
 982
 983                 /*
 984                  * filters only work for tracepoint event, which doesn't have cpu limit.
 985                  * So evlist and evsel should always be same.
 986                  */
 987                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
 988                 if (err) {
 989                         *err_evsel = evsel;
 990                         break;
 991                 }
 992         }
 993
 994         return err;
 995 }
 996
 997 int perf_evlist__set_tp_filter(struct evlist *evlist, const char *filter)
 998 {
 999         struct evsel *evsel;
1000         int err = 0;
1001
1002         if (filter == NULL)
1003                 return -1;
1004
1005         evlist__for_each_entry(evlist, evsel) {
1006                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1007                         continue;
1008
1009                 err = evsel__set_filter(evsel, filter);
1010                 if (err)
1011                         break;
1012         }
1013
1014         return err;
1015 }
1016
1017 int perf_evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1018 {
1019         struct evsel *evsel;
1020         int err = 0;
1021
1022         if (filter == NULL)
1023                 return -1;
1024
1025         evlist__for_each_entry(evlist, evsel) {
1026                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1027                         continue;
1028
1029                 err = evsel__append_tp_filter(evsel, filter);
1030                 if (err)
1031                         break;
1032         }
1033
1034         return err;
1035 }
1036
1037 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1038 {
1039         char *filter;
1040         size_t i;
1041
1042         for (i = 0; i < npids; ++i) {
1043                 if (i == 0) {
1044                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1045                                 return NULL;
1046                 } else {
1047                         char *tmp;
1048
1049                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1050                                 goto out_free;
1051
1052                         free(filter);
1053                         filter = tmp;
1054                 }
1055         }
1056
1057         return filter;
1058 out_free:
1059         free(filter);
1060         return NULL;
1061 }
1062
1063 int perf_evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1064 {
1065         char *filter = asprintf__tp_filter_pids(npids, pids);
1066         int ret = perf_evlist__set_tp_filter(evlist, filter);
1067
1068         free(filter);
1069         return ret;
1070 }
1071
1072 int perf_evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1073 {
1074         return perf_evlist__set_tp_filter_pids(evlist, 1, &pid);
1075 }
1076
1077 int perf_evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1078 {
1079         char *filter = asprintf__tp_filter_pids(npids, pids);
1080         int ret = perf_evlist__append_tp_filter(evlist, filter);
1081
1082         free(filter);
1083         return ret;
1084 }
1085
1086 int perf_evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1087 {
1088         return perf_evlist__append_tp_filter_pids(evlist, 1, &pid);
1089 }
1090
1091 bool evlist__valid_sample_type(struct evlist *evlist)
1092 {
1093         struct evsel *pos;
1094
1095         if (evlist->core.nr_entries == 1)
1096                 return true;
1097
1098         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1099                 return false;
1100
1101         evlist__for_each_entry(evlist, pos) {
1102                 if (pos->id_pos != evlist->id_pos ||
1103                     pos->is_pos != evlist->is_pos)
1104                         return false;
1105         }
1106
1107         return true;
1108 }
1109
1110 u64 __evlist__combined_sample_type(struct evlist *evlist)
1111 {
1112         struct evsel *evsel;
1113
1114         if (evlist->combined_sample_type)
1115                 return evlist->combined_sample_type;
1116
1117         evlist__for_each_entry(evlist, evsel)
1118                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1119
1120         return evlist->combined_sample_type;
1121 }
1122
1123 u64 evlist__combined_sample_type(struct evlist *evlist)
1124 {
1125         evlist->combined_sample_type = 0;
1126         return __evlist__combined_sample_type(evlist);
1127 }
1128
1129 u64 evlist__combined_branch_type(struct evlist *evlist)
1130 {
1131         struct evsel *evsel;
1132         u64 branch_type = 0;
1133
1134         evlist__for_each_entry(evlist, evsel)
1135                 branch_type |= evsel->core.attr.branch_sample_type;
1136         return branch_type;
1137 }
1138
1139 bool perf_evlist__valid_read_format(struct evlist *evlist)
1140 {
1141         struct evsel *first = evlist__first(evlist), *pos = first;
1142         u64 read_format = first->core.attr.read_format;
1143         u64 sample_type = first->core.attr.sample_type;
1144
1145         evlist__for_each_entry(evlist, pos) {
1146                 if (read_format != pos->core.attr.read_format) {
1147                         pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1148                                  read_format, (u64)pos->core.attr.read_format);
1149                 }
1150         }
1151
1152         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1153         if ((sample_type & PERF_SAMPLE_READ) &&
1154             !(read_format & PERF_FORMAT_ID)) {
1155                 return false;
1156         }
1157
1158         return true;
1159 }
1160
1161 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1162 {
1163         struct evsel *first = evlist__first(evlist);
1164         struct perf_sample *data;
1165         u64 sample_type;
1166         u16 size = 0;
1167
1168         if (!first->core.attr.sample_id_all)
1169                 goto out;
1170
1171         sample_type = first->core.attr.sample_type;
1172
1173         if (sample_type & PERF_SAMPLE_TID)
1174                 size += sizeof(data->tid) * 2;
1175
1176        if (sample_type & PERF_SAMPLE_TIME)
1177                 size += sizeof(data->time);
1178
1179         if (sample_type & PERF_SAMPLE_ID)
1180                 size += sizeof(data->id);
1181
1182         if (sample_type & PERF_SAMPLE_STREAM_ID)
1183                 size += sizeof(data->stream_id);
1184
1185         if (sample_type & PERF_SAMPLE_CPU)
1186                 size += sizeof(data->cpu) * 2;
1187
1188         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1189                 size += sizeof(data->id);
1190 out:
1191         return size;
1192 }
1193
1194 bool evlist__valid_sample_id_all(struct evlist *evlist)
1195 {
1196         struct evsel *first = evlist__first(evlist), *pos = first;
1197
1198         evlist__for_each_entry_continue(evlist, pos) {
1199                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1200                         return false;
1201         }
1202
1203         return true;
1204 }
1205
1206 bool evlist__sample_id_all(struct evlist *evlist)
1207 {
1208         struct evsel *first = evlist__first(evlist);
1209         return first->core.attr.sample_id_all;
1210 }
1211
1212 void perf_evlist__set_selected(struct evlist *evlist,
1213                                struct evsel *evsel)
1214 {
1215         evlist->selected = evsel;
1216 }
1217
1218 void evlist__close(struct evlist *evlist)
1219 {
1220         struct evsel *evsel;
1221         struct affinity affinity;
1222         int cpu, i;
1223
1224         /*
1225          * With perf record core.cpus is usually NULL.
1226          * Use the old method to handle this for now.
1227          */
1228         if (!evlist->core.cpus) {
1229                 evlist__for_each_entry_reverse(evlist, evsel)
1230                         evsel__close(evsel);
1231                 return;
1232         }
1233
1234         if (affinity__setup(&affinity) < 0)
1235                 return;
1236         evlist__for_each_cpu(evlist, i, cpu) {
1237                 affinity__set(&affinity, cpu);
1238
1239                 evlist__for_each_entry_reverse(evlist, evsel) {
1240                         if (evsel__cpu_iter_skip(evsel, cpu))
1241                             continue;
1242                         perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1243                 }
1244         }
1245         affinity__cleanup(&affinity);
1246         evlist__for_each_entry_reverse(evlist, evsel) {
1247                 perf_evsel__free_fd(&evsel->core);
1248                 perf_evsel__free_id(&evsel->core);
1249         }
1250 }
1251
1252 static int perf_evlist__create_syswide_maps(struct evlist *evlist)
1253 {
1254         struct perf_cpu_map *cpus;
1255         struct perf_thread_map *threads;
1256         int err = -ENOMEM;
1257
1258         /*
1259          * Try reading /sys/devices/system/cpu/online to get
1260          * an all cpus map.
1261          *
1262          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1263          * code needs an overhaul to properly forward the
1264          * error, and we may not want to do that fallback to a
1265          * default cpu identity map :-\
1266          */
1267         cpus = perf_cpu_map__new(NULL);
1268         if (!cpus)
1269                 goto out;
1270
1271         threads = perf_thread_map__new_dummy();
1272         if (!threads)
1273                 goto out_put;
1274
1275         perf_evlist__set_maps(&evlist->core, cpus, threads);
1276 out:
1277         return err;
1278 out_put:
1279         perf_cpu_map__put(cpus);
1280         goto out;
1281 }
1282
1283 int evlist__open(struct evlist *evlist)
1284 {
1285         struct evsel *evsel;
1286         int err;
1287
1288         /*
1289          * Default: one fd per CPU, all threads, aka systemwide
1290          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1291          */
1292         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1293                 err = perf_evlist__create_syswide_maps(evlist);
1294                 if (err < 0)
1295                         goto out_err;
1296         }
1297
1298         perf_evlist__update_id_pos(evlist);
1299
1300         evlist__for_each_entry(evlist, evsel) {
1301                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1302                 if (err < 0)
1303                         goto out_err;
1304         }
1305
1306         return 0;
1307 out_err:
1308         evlist__close(evlist);
1309         errno = -err;
1310         return err;
1311 }
1312
1313 int perf_evlist__prepare_workload(struct evlist *evlist, struct target *target,
1314                                   const char *argv[], bool pipe_output,
1315                                   void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1316 {
1317         int child_ready_pipe[2], go_pipe[2];
1318         char bf;
1319
1320         if (pipe(child_ready_pipe) < 0) {
1321                 perror("failed to create 'ready' pipe");
1322                 return -1;
1323         }
1324
1325         if (pipe(go_pipe) < 0) {
1326                 perror("failed to create 'go' pipe");
1327                 goto out_close_ready_pipe;
1328         }
1329
1330         evlist->workload.pid = fork();
1331         if (evlist->workload.pid < 0) {
1332                 perror("failed to fork");
1333                 goto out_close_pipes;
1334         }
1335
1336         if (!evlist->workload.pid) {
1337                 int ret;
1338
1339                 if (pipe_output)
1340                         dup2(2, 1);
1341
1342                 signal(SIGTERM, SIG_DFL);
1343
1344                 close(child_ready_pipe[0]);
1345                 close(go_pipe[1]);
1346                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1347
1348                 /*
1349                  * Tell the parent we're ready to go
1350                  */
1351                 close(child_ready_pipe[1]);
1352
1353                 /*
1354                  * Wait until the parent tells us to go.
1355                  */
1356                 ret = read(go_pipe[0], &bf, 1);
1357                 /*
1358                  * The parent will ask for the execvp() to be performed by
1359                  * writing exactly one byte, in workload.cork_fd, usually via
1360                  * perf_evlist__start_workload().
1361                  *
1362                  * For cancelling the workload without actually running it,
1363                  * the parent will just close workload.cork_fd, without writing
1364                  * anything, i.e. read will return zero and we just exit()
1365                  * here.
1366                  */
1367                 if (ret != 1) {
1368                         if (ret == -1)
1369                                 perror("unable to read pipe");
1370                         exit(ret);
1371                 }
1372
1373                 execvp(argv[0], (char **)argv);
1374
1375                 if (exec_error) {
1376                         union sigval val;
1377
1378                         val.sival_int = errno;
1379                         if (sigqueue(getppid(), SIGUSR1, val))
1380                                 perror(argv[0]);
1381                 } else
1382                         perror(argv[0]);
1383                 exit(-1);
1384         }
1385
1386         if (exec_error) {
1387                 struct sigaction act = {
1388                         .sa_flags     = SA_SIGINFO,
1389                         .sa_sigaction = exec_error,
1390                 };
1391                 sigaction(SIGUSR1, &act, NULL);
1392         }
1393
1394         if (target__none(target)) {
1395                 if (evlist->core.threads == NULL) {
1396                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1397                                 __func__, __LINE__);
1398                         goto out_close_pipes;
1399                 }
1400                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1401         }
1402
1403         close(child_ready_pipe[1]);
1404         close(go_pipe[0]);
1405         /*
1406          * wait for child to settle
1407          */
1408         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1409                 perror("unable to read pipe");
1410                 goto out_close_pipes;
1411         }
1412
1413         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1414         evlist->workload.cork_fd = go_pipe[1];
1415         close(child_ready_pipe[0]);
1416         return 0;
1417
1418 out_close_pipes:
1419         close(go_pipe[0]);
1420         close(go_pipe[1]);
1421 out_close_ready_pipe:
1422         close(child_ready_pipe[0]);
1423         close(child_ready_pipe[1]);
1424         return -1;
1425 }
1426
1427 int perf_evlist__start_workload(struct evlist *evlist)
1428 {
1429         if (evlist->workload.cork_fd > 0) {
1430                 char bf = 0;
1431                 int ret;
1432                 /*
1433                  * Remove the cork, let it rip!
1434                  */
1435                 ret = write(evlist->workload.cork_fd, &bf, 1);
1436                 if (ret < 0)
1437                         perror("unable to write to pipe");
1438
1439                 close(evlist->workload.cork_fd);
1440                 return ret;
1441         }
1442
1443         return 0;
1444 }
1445
1446 int perf_evlist__parse_sample(struct evlist *evlist, union perf_event *event,
1447                               struct perf_sample *sample)
1448 {
1449         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1450
1451         if (!evsel)
1452                 return -EFAULT;
1453         return evsel__parse_sample(evsel, event, sample);
1454 }
1455
1456 int perf_evlist__parse_sample_timestamp(struct evlist *evlist,
1457                                         union perf_event *event,
1458                                         u64 *timestamp)
1459 {
1460         struct evsel *evsel = perf_evlist__event2evsel(evlist, event);
1461
1462         if (!evsel)
1463                 return -EFAULT;
1464         return evsel__parse_sample_timestamp(evsel, event, timestamp);
1465 }
1466
1467 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1468 {
1469         int printed, value;
1470         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1471
1472         switch (err) {
1473         case EACCES:
1474         case EPERM:
1475                 printed = scnprintf(buf, size,
1476                                     "Error:\t%s.\n"
1477                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1478
1479                 value = perf_event_paranoid();
1480
1481                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1482
1483                 if (value >= 2) {
1484                         printed += scnprintf(buf + printed, size - printed,
1485                                              "For your workloads it needs to be <= 1\nHint:\t");
1486                 }
1487                 printed += scnprintf(buf + printed, size - printed,
1488                                      "For system wide tracing it needs to be set to -1.\n");
1489
1490                 printed += scnprintf(buf + printed, size - printed,
1491                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1492                                     "Hint:\tThe current value is %d.", value);
1493                 break;
1494         case EINVAL: {
1495                 struct evsel *first = evlist__first(evlist);
1496                 int max_freq;
1497
1498                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1499                         goto out_default;
1500
1501                 if (first->core.attr.sample_freq < (u64)max_freq)
1502                         goto out_default;
1503
1504                 printed = scnprintf(buf, size,
1505                                     "Error:\t%s.\n"
1506                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1507                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1508                                     emsg, max_freq, first->core.attr.sample_freq);
1509                 break;
1510         }
1511         default:
1512 out_default:
1513                 scnprintf(buf, size, "%s", emsg);
1514                 break;
1515         }
1516
1517         return 0;
1518 }
1519
1520 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1521 {
1522         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1523         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1524
1525         switch (err) {
1526         case EPERM:
1527                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1528                 printed += scnprintf(buf + printed, size - printed,
1529                                      "Error:\t%s.\n"
1530                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1531                                      "Hint:\tTried using %zd kB.\n",
1532                                      emsg, pages_max_per_user, pages_attempted);
1533
1534                 if (pages_attempted >= pages_max_per_user) {
1535                         printed += scnprintf(buf + printed, size - printed,
1536                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1537                                              pages_max_per_user + pages_attempted);
1538                 }
1539
1540                 printed += scnprintf(buf + printed, size - printed,
1541                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1542                 break;
1543         default:
1544                 scnprintf(buf, size, "%s", emsg);
1545                 break;
1546         }
1547
1548         return 0;
1549 }
1550
1551 void perf_evlist__to_front(struct evlist *evlist,
1552                            struct evsel *move_evsel)
1553 {
1554         struct evsel *evsel, *n;
1555         LIST_HEAD(move);
1556
1557         if (move_evsel == evlist__first(evlist))
1558                 return;
1559
1560         evlist__for_each_entry_safe(evlist, n, evsel) {
1561                 if (evsel->leader == move_evsel->leader)
1562                         list_move_tail(&evsel->core.node, &move);
1563         }
1564
1565         list_splice(&move, &evlist->core.entries);
1566 }
1567
1568 struct evsel *perf_evlist__get_tracking_event(struct evlist *evlist)
1569 {
1570         struct evsel *evsel;
1571
1572         evlist__for_each_entry(evlist, evsel) {
1573                 if (evsel->tracking)
1574                         return evsel;
1575         }
1576
1577         return evlist__first(evlist);
1578 }
1579
1580 void perf_evlist__set_tracking_event(struct evlist *evlist,
1581                                      struct evsel *tracking_evsel)
1582 {
1583         struct evsel *evsel;
1584
1585         if (tracking_evsel->tracking)
1586                 return;
1587
1588         evlist__for_each_entry(evlist, evsel) {
1589                 if (evsel != tracking_evsel)
1590                         evsel->tracking = false;
1591         }
1592
1593         tracking_evsel->tracking = true;
1594 }
1595
1596 struct evsel *
1597 perf_evlist__find_evsel_by_str(struct evlist *evlist,
1598                                const char *str)
1599 {
1600         struct evsel *evsel;
1601
1602         evlist__for_each_entry(evlist, evsel) {
1603                 if (!evsel->name)
1604                         continue;
1605                 if (strcmp(str, evsel->name) == 0)
1606                         return evsel;
1607         }
1608
1609         return NULL;
1610 }
1611
1612 void perf_evlist__toggle_bkw_mmap(struct evlist *evlist,
1613                                   enum bkw_mmap_state state)
1614 {
1615         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1616         enum action {
1617                 NONE,
1618                 PAUSE,
1619                 RESUME,
1620         } action = NONE;
1621
1622         if (!evlist->overwrite_mmap)
1623                 return;
1624
1625         switch (old_state) {
1626         case BKW_MMAP_NOTREADY: {
1627                 if (state != BKW_MMAP_RUNNING)
1628                         goto state_err;
1629                 break;
1630         }
1631         case BKW_MMAP_RUNNING: {
1632                 if (state != BKW_MMAP_DATA_PENDING)
1633                         goto state_err;
1634                 action = PAUSE;
1635                 break;
1636         }
1637         case BKW_MMAP_DATA_PENDING: {
1638                 if (state != BKW_MMAP_EMPTY)
1639                         goto state_err;
1640                 break;
1641         }
1642         case BKW_MMAP_EMPTY: {
1643                 if (state != BKW_MMAP_RUNNING)
1644                         goto state_err;
1645                 action = RESUME;
1646                 break;
1647         }
1648         default:
1649                 WARN_ONCE(1, "Shouldn't get there\n");
1650         }
1651
1652         evlist->bkw_mmap_state = state;
1653
1654         switch (action) {
1655         case PAUSE:
1656                 perf_evlist__pause(evlist);
1657                 break;
1658         case RESUME:
1659                 perf_evlist__resume(evlist);
1660                 break;
1661         case NONE:
1662         default:
1663                 break;
1664         }
1665
1666 state_err:
1667         return;
1668 }
1669
1670 bool perf_evlist__exclude_kernel(struct evlist *evlist)
1671 {
1672         struct evsel *evsel;
1673
1674         evlist__for_each_entry(evlist, evsel) {
1675                 if (!evsel->core.attr.exclude_kernel)
1676                         return false;
1677         }
1678
1679         return true;
1680 }
1681
1682 /*
1683  * Events in data file are not collect in groups, but we still want
1684  * the group display. Set the artificial group and set the leader's
1685  * forced_leader flag to notify the display code.
1686  */
1687 void perf_evlist__force_leader(struct evlist *evlist)
1688 {
1689         if (!evlist->nr_groups) {
1690                 struct evsel *leader = evlist__first(evlist);
1691
1692                 perf_evlist__set_leader(evlist);
1693                 leader->forced_leader = true;
1694         }
1695 }
1696
1697 struct evsel *perf_evlist__reset_weak_group(struct evlist *evsel_list,
1698                                                  struct evsel *evsel,
1699                                                 bool close)
1700 {
1701         struct evsel *c2, *leader;
1702         bool is_open = true;
1703
1704         leader = evsel->leader;
1705         pr_debug("Weak group for %s/%d failed\n",
1706                         leader->name, leader->core.nr_members);
1707
1708         /*
1709          * for_each_group_member doesn't work here because it doesn't
1710          * include the first entry.
1711          */
1712         evlist__for_each_entry(evsel_list, c2) {
1713                 if (c2 == evsel)
1714                         is_open = false;
1715                 if (c2->leader == leader) {
1716                         if (is_open && close)
1717                                 perf_evsel__close(&c2->core);
1718                         c2->leader = c2;
1719                         c2->core.nr_members = 0;
1720                         /*
1721                          * Set this for all former members of the group
1722                          * to indicate they get reopened.
1723                          */
1724                         c2->reset_group = true;
1725                 }
1726         }
1727         return leader;
1728 }
1729
1730 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack)
1731 {
1732         char *comma = NULL, *endptr = NULL;
1733
1734         if (strncmp(str, "fd:", 3))
1735                 return -EINVAL;
1736
1737         *ctl_fd = strtoul(&str[3], &endptr, 0);
1738         if (endptr == &str[3])
1739                 return -EINVAL;
1740
1741         comma = strchr(str, ',');
1742         if (comma) {
1743                 if (endptr != comma)
1744                         return -EINVAL;
1745
1746                 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1747                 if (endptr == comma + 1 || *endptr != '\0')
1748                         return -EINVAL;
1749         }
1750
1751         return 0;
1752 }
1753
1754 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1755 {
1756         if (fd == -1) {
1757                 pr_debug("Control descriptor is not initialized\n");
1758                 return 0;
1759         }
1760
1761         evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1762                                                      fdarray_flag__nonfilterable);
1763         if (evlist->ctl_fd.pos < 0) {
1764                 evlist->ctl_fd.pos = -1;
1765                 pr_err("Failed to add ctl fd entry: %m\n");
1766                 return -1;
1767         }
1768
1769         evlist->ctl_fd.fd = fd;
1770         evlist->ctl_fd.ack = ack;
1771
1772         return 0;
1773 }
1774
1775 bool evlist__ctlfd_initialized(struct evlist *evlist)
1776 {
1777         return evlist->ctl_fd.pos >= 0;
1778 }
1779
1780 int evlist__finalize_ctlfd(struct evlist *evlist)
1781 {
1782         struct pollfd *entries = evlist->core.pollfd.entries;
1783
1784         if (!evlist__ctlfd_initialized(evlist))
1785                 return 0;
1786
1787         entries[evlist->ctl_fd.pos].fd = -1;
1788         entries[evlist->ctl_fd.pos].events = 0;
1789         entries[evlist->ctl_fd.pos].revents = 0;
1790
1791         evlist->ctl_fd.pos = -1;
1792         evlist->ctl_fd.ack = -1;
1793         evlist->ctl_fd.fd = -1;
1794
1795         return 0;
1796 }
1797
1798 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1799                               char *cmd_data, size_t data_size)
1800 {
1801         int err;
1802         char c;
1803         size_t bytes_read = 0;
1804
1805         *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1806         memset(cmd_data, 0, data_size);
1807         data_size--;
1808
1809         do {
1810                 err = read(evlist->ctl_fd.fd, &c, 1);
1811                 if (err > 0) {
1812                         if (c == '\n' || c == '\0')
1813                                 break;
1814                         cmd_data[bytes_read++] = c;
1815                         if (bytes_read == data_size)
1816                                 break;
1817                         continue;
1818                 } else if (err == -1) {
1819                         if (errno == EINTR)
1820                                 continue;
1821                         if (errno == EAGAIN || errno == EWOULDBLOCK)
1822                                 err = 0;
1823                         else
1824                                 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1825                 }
1826                 break;
1827         } while (1);
1828
1829         pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1830                  bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1831
1832         if (bytes_read > 0) {
1833                 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1834                              (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1835                         *cmd = EVLIST_CTL_CMD_ENABLE;
1836                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1837                                     (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1838                         *cmd = EVLIST_CTL_CMD_DISABLE;
1839                 }
1840         }
1841
1842         return bytes_read ? (int)bytes_read : err;
1843 }
1844
1845 static int evlist__ctlfd_ack(struct evlist *evlist)
1846 {
1847         int err;
1848
1849         if (evlist->ctl_fd.ack == -1)
1850                 return 0;
1851
1852         err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
1853                     sizeof(EVLIST_CTL_CMD_ACK_TAG));
1854         if (err == -1)
1855                 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
1856
1857         return err;
1858 }
1859
1860 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
1861 {
1862         int err = 0;
1863         char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
1864         int ctlfd_pos = evlist->ctl_fd.pos;
1865         struct pollfd *entries = evlist->core.pollfd.entries;
1866
1867         if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
1868                 return 0;
1869
1870         if (entries[ctlfd_pos].revents & POLLIN) {
1871                 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
1872                                          EVLIST_CTL_CMD_MAX_LEN);
1873                 if (err > 0) {
1874                         switch (*cmd) {
1875                         case EVLIST_CTL_CMD_ENABLE:
1876                                 evlist__enable(evlist);
1877                                 break;
1878                         case EVLIST_CTL_CMD_DISABLE:
1879                                 evlist__disable(evlist);
1880                                 break;
1881                         case EVLIST_CTL_CMD_ACK:
1882                         case EVLIST_CTL_CMD_UNSUPPORTED:
1883                         default:
1884                                 pr_debug("ctlfd: unsupported %d\n", *cmd);
1885                                 break;
1886                         }
1887                         if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED))
1888                                 evlist__ctlfd_ack(evlist);
1889                 }
1890         }
1891
1892         if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
1893                 evlist__finalize_ctlfd(evlist);
1894         else
1895                 entries[ctlfd_pos].revents = 0;
1896
1897         return err;
1898 }