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 "record.h"
  19 #include "debug.h"
  20 #include "units.h"
  21 #include "bpf_counter.h"
  22 #include <internal/lib.h> // page_size
  23 #include "affinity.h"
  24 #include "../perf.h"
  25 #include "asm/bug.h"
  26 #include "bpf-event.h"
  27 #include "util/event.h"
  28 #include "util/string2.h"
  29 #include "util/perf_api_probe.h"
  30 #include "util/evsel_fprintf.h"
  31 #include "util/evlist-hybrid.h"
  32 #include "util/pmu.h"
  33 #include "util/sample.h"
  34 #include <signal.h>
  35 #include <unistd.h>
  36 #include <sched.h>
  37 #include <stdlib.h>
  38
  39 #include "parse-events.h"
  40 #include <subcmd/parse-options.h>
  41
  42 #include <fcntl.h>
  43 #include <sys/ioctl.h>
  44 #include <sys/mman.h>
  45 #include <sys/prctl.h>
  46 #include <sys/timerfd.h>
  47
  48 #include <linux/bitops.h>
  49 #include <linux/hash.h>
  50 #include <linux/log2.h>
  51 #include <linux/err.h>
  52 #include <linux/string.h>
  53 #include <linux/time64.h>
  54 #include <linux/zalloc.h>
  55 #include <perf/evlist.h>
  56 #include <perf/evsel.h>
  57 #include <perf/cpumap.h>
  58 #include <perf/mmap.h>
  59
  60 #include <internal/xyarray.h>
  61
  62 #ifdef LACKS_SIGQUEUE_PROTOTYPE
  63 int sigqueue(pid_t pid, int sig, const union sigval value);
  64 #endif
  65
  66 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y))
  67 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y)
  68
  69 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus,
  70                   struct perf_thread_map *threads)
  71 {
  72         perf_evlist__init(&evlist->core);
  73         perf_evlist__set_maps(&evlist->core, cpus, threads);
  74         evlist->workload.pid = -1;
  75         evlist->bkw_mmap_state = BKW_MMAP_NOTREADY;
  76         evlist->ctl_fd.fd = -1;
  77         evlist->ctl_fd.ack = -1;
  78         evlist->ctl_fd.pos = -1;
  79 }
  80
  81 struct evlist *evlist__new(void)
  82 {
  83         struct evlist *evlist = zalloc(sizeof(*evlist));
  84
  85         if (evlist != NULL)
  86                 evlist__init(evlist, NULL, NULL);
  87
  88         return evlist;
  89 }
  90
  91 struct evlist *evlist__new_default(void)
  92 {
  93         struct evlist *evlist = evlist__new();
  94
  95         if (evlist && evlist__add_default(evlist)) {
  96                 evlist__delete(evlist);
  97                 evlist = NULL;
  98         }
  99
 100         return evlist;
 101 }
 102
 103 struct evlist *evlist__new_dummy(void)
 104 {
 105         struct evlist *evlist = evlist__new();
 106
 107         if (evlist && evlist__add_dummy(evlist)) {
 108                 evlist__delete(evlist);
 109                 evlist = NULL;
 110         }
 111
 112         return evlist;
 113 }
 114
 115 /**
 116  * evlist__set_id_pos - set the positions of event ids.
 117  * @evlist: selected event list
 118  *
 119  * Events with compatible sample types all have the same id_pos
 120  * and is_pos.  For convenience, put a copy on evlist.
 121  */
 122 void evlist__set_id_pos(struct evlist *evlist)
 123 {
 124         struct evsel *first = evlist__first(evlist);
 125
 126         evlist->id_pos = first->id_pos;
 127         evlist->is_pos = first->is_pos;
 128 }
 129
 130 static void evlist__update_id_pos(struct evlist *evlist)
 131 {
 132         struct evsel *evsel;
 133
 134         evlist__for_each_entry(evlist, evsel)
 135                 evsel__calc_id_pos(evsel);
 136
 137         evlist__set_id_pos(evlist);
 138 }
 139
 140 static void evlist__purge(struct evlist *evlist)
 141 {
 142         struct evsel *pos, *n;
 143
 144         evlist__for_each_entry_safe(evlist, n, pos) {
 145                 list_del_init(&pos->core.node);
 146                 pos->evlist = NULL;
 147                 evsel__delete(pos);
 148         }
 149
 150         evlist->core.nr_entries = 0;
 151 }
 152
 153 void evlist__exit(struct evlist *evlist)
 154 {
 155         event_enable_timer__exit(&evlist->eet);
 156         zfree(&evlist->mmap);
 157         zfree(&evlist->overwrite_mmap);
 158         perf_evlist__exit(&evlist->core);
 159 }
 160
 161 void evlist__delete(struct evlist *evlist)
 162 {
 163         if (evlist == NULL)
 164                 return;
 165
 166         evlist__munmap(evlist);
 167         evlist__close(evlist);
 168         evlist__purge(evlist);
 169         evlist__exit(evlist);
 170         free(evlist);
 171 }
 172
 173 void evlist__add(struct evlist *evlist, struct evsel *entry)
 174 {
 175         perf_evlist__add(&evlist->core, &entry->core);
 176         entry->evlist = evlist;
 177         entry->tracking = !entry->core.idx;
 178
 179         if (evlist->core.nr_entries == 1)
 180                 evlist__set_id_pos(evlist);
 181 }
 182
 183 void evlist__remove(struct evlist *evlist, struct evsel *evsel)
 184 {
 185         evsel->evlist = NULL;
 186         perf_evlist__remove(&evlist->core, &evsel->core);
 187 }
 188
 189 void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
 190 {
 191         while (!list_empty(list)) {
 192                 struct evsel *evsel, *temp, *leader = NULL;
 193
 194                 __evlist__for_each_entry_safe(list, temp, evsel) {
 195                         list_del_init(&evsel->core.node);
 196                         evlist__add(evlist, evsel);
 197                         leader = evsel;
 198                         break;
 199                 }
 200
 201                 __evlist__for_each_entry_safe(list, temp, evsel) {
 202                         if (evsel__has_leader(evsel, leader)) {
 203                                 list_del_init(&evsel->core.node);
 204                                 evlist__add(evlist, evsel);
 205                         }
 206                 }
 207         }
 208 }
 209
 210 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 211                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 212 {
 213         size_t i;
 214         int err;
 215
 216         for (i = 0; i < nr_assocs; i++) {
 217                 // Adding a handler for an event not in this evlist, just ignore it.
 218                 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 219                 if (evsel == NULL)
 220                         continue;
 221
 222                 err = -EEXIST;
 223                 if (evsel->handler != NULL)
 224                         goto out;
 225                 evsel->handler = assocs[i].handler;
 226         }
 227
 228         err = 0;
 229 out:
 230         return err;
 231 }
 232
 233 static void evlist__set_leader(struct evlist *evlist)
 234 {
 235         perf_evlist__set_leader(&evlist->core);
 236 }
 237
 238 int __evlist__add_default(struct evlist *evlist, bool precise)
 239 {
 240         struct evsel *evsel;
 241
 242         evsel = evsel__new_cycles(precise, PERF_TYPE_HARDWARE,
 243                                   PERF_COUNT_HW_CPU_CYCLES);
 244         if (evsel == NULL)
 245                 return -ENOMEM;
 246
 247         evlist__add(evlist, evsel);
 248         return 0;
 249 }
 250
 251 static struct evsel *evlist__dummy_event(struct evlist *evlist)
 252 {
 253         struct perf_event_attr attr = {
 254                 .type   = PERF_TYPE_SOFTWARE,
 255                 .config = PERF_COUNT_SW_DUMMY,
 256                 .size   = sizeof(attr), /* to capture ABI version */
 257         };
 258
 259         return evsel__new_idx(&attr, evlist->core.nr_entries);
 260 }
 261
 262 int evlist__add_dummy(struct evlist *evlist)
 263 {
 264         struct evsel *evsel = evlist__dummy_event(evlist);
 265
 266         if (evsel == NULL)
 267                 return -ENOMEM;
 268
 269         evlist__add(evlist, evsel);
 270         return 0;
 271 }
 272
 273 struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
 274 {
 275         struct evsel *evsel = evlist__dummy_event(evlist);
 276
 277         if (!evsel)
 278                 return NULL;
 279
 280         evsel->core.attr.exclude_kernel = 1;
 281         evsel->core.attr.exclude_guest = 1;
 282         evsel->core.attr.exclude_hv = 1;
 283         evsel->core.attr.freq = 0;
 284         evsel->core.attr.sample_period = 1;
 285         evsel->core.system_wide = system_wide;
 286         evsel->no_aux_samples = true;
 287         evsel->name = strdup("dummy:u");
 288
 289         evlist__add(evlist, evsel);
 290         return evsel;
 291 }
 292
 293 #ifdef HAVE_LIBTRACEEVENT
 294 struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide)
 295 {
 296         struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0);
 297
 298         if (IS_ERR(evsel))
 299                 return evsel;
 300
 301         evsel__set_sample_bit(evsel, CPU);
 302         evsel__set_sample_bit(evsel, TIME);
 303
 304         evsel->core.system_wide = system_wide;
 305         evsel->no_aux_samples = true;
 306
 307         evlist__add(evlist, evsel);
 308         return evsel;
 309 }
 310 #endif
 311
 312 int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 313 {
 314         struct evsel *evsel, *n;
 315         LIST_HEAD(head);
 316         size_t i;
 317
 318         for (i = 0; i < nr_attrs; i++) {
 319                 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 320                 if (evsel == NULL)
 321                         goto out_delete_partial_list;
 322                 list_add_tail(&evsel->core.node, &head);
 323         }
 324
 325         evlist__splice_list_tail(evlist, &head);
 326
 327         return 0;
 328
 329 out_delete_partial_list:
 330         __evlist__for_each_entry_safe(&head, n, evsel)
 331                 evsel__delete(evsel);
 332         return -1;
 333 }
 334
 335 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 336 {
 337         size_t i;
 338
 339         for (i = 0; i < nr_attrs; i++)
 340                 event_attr_init(attrs + i);
 341
 342         return evlist__add_attrs(evlist, attrs, nr_attrs);
 343 }
 344
 345 __weak int arch_evlist__add_default_attrs(struct evlist *evlist,
 346                                           struct perf_event_attr *attrs,
 347                                           size_t nr_attrs)
 348 {
 349         if (!nr_attrs)
 350                 return 0;
 351
 352         return __evlist__add_default_attrs(evlist, attrs, nr_attrs);
 353 }
 354
 355 struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 356 {
 357         struct evsel *evsel;
 358
 359         evlist__for_each_entry(evlist, evsel) {
 360                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 361                     (int)evsel->core.attr.config == id)
 362                         return evsel;
 363         }
 364
 365         return NULL;
 366 }
 367
 368 struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
 369 {
 370         struct evsel *evsel;
 371
 372         evlist__for_each_entry(evlist, evsel) {
 373                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 374                     (strcmp(evsel->name, name) == 0))
 375                         return evsel;
 376         }
 377
 378         return NULL;
 379 }
 380
 381 #ifdef HAVE_LIBTRACEEVENT
 382 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
 383 {
 384         struct evsel *evsel = evsel__newtp(sys, name);
 385
 386         if (IS_ERR(evsel))
 387                 return -1;
 388
 389         evsel->handler = handler;
 390         evlist__add(evlist, evsel);
 391         return 0;
 392 }
 393 #endif
 394
 395 struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity)
 396 {
 397         struct evlist_cpu_iterator itr = {
 398                 .container = evlist,
 399                 .evsel = NULL,
 400                 .cpu_map_idx = 0,
 401                 .evlist_cpu_map_idx = 0,
 402                 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus),
 403                 .cpu = (struct perf_cpu){ .cpu = -1},
 404                 .affinity = affinity,
 405         };
 406
 407         if (evlist__empty(evlist)) {
 408                 /* Ensure the empty list doesn't iterate. */
 409                 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr;
 410         } else {
 411                 itr.evsel = evlist__first(evlist);
 412                 if (itr.affinity) {
 413                         itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0);
 414                         affinity__set(itr.affinity, itr.cpu.cpu);
 415                         itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu);
 416                         /*
 417                          * If this CPU isn't in the evsel's cpu map then advance
 418                          * through the list.
 419                          */
 420                         if (itr.cpu_map_idx == -1)
 421                                 evlist_cpu_iterator__next(&itr);
 422                 }
 423         }
 424         return itr;
 425 }
 426
 427 void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr)
 428 {
 429         while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) {
 430                 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel);
 431                 evlist_cpu_itr->cpu_map_idx =
 432                         perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 433                                           evlist_cpu_itr->cpu);
 434                 if (evlist_cpu_itr->cpu_map_idx != -1)
 435                         return;
 436         }
 437         evlist_cpu_itr->evlist_cpu_map_idx++;
 438         if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) {
 439                 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container);
 440                 evlist_cpu_itr->cpu =
 441                         perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus,
 442                                           evlist_cpu_itr->evlist_cpu_map_idx);
 443                 if (evlist_cpu_itr->affinity)
 444                         affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu);
 445                 evlist_cpu_itr->cpu_map_idx =
 446                         perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus,
 447                                           evlist_cpu_itr->cpu);
 448                 /*
 449                  * If this CPU isn't in the evsel's cpu map then advance through
 450                  * the list.
 451                  */
 452                 if (evlist_cpu_itr->cpu_map_idx == -1)
 453                         evlist_cpu_iterator__next(evlist_cpu_itr);
 454         }
 455 }
 456
 457 bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
 458 {
 459         return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
 460 }
 461
 462 static int evsel__strcmp(struct evsel *pos, char *evsel_name)
 463 {
 464         if (!evsel_name)
 465                 return 0;
 466         if (evsel__is_dummy_event(pos))
 467                 return 1;
 468         return strcmp(pos->name, evsel_name);
 469 }
 470
 471 static int evlist__is_enabled(struct evlist *evlist)
 472 {
 473         struct evsel *pos;
 474
 475         evlist__for_each_entry(evlist, pos) {
 476                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 477                         continue;
 478                 /* If at least one event is enabled, evlist is enabled. */
 479                 if (!pos->disabled)
 480                         return true;
 481         }
 482         return false;
 483 }
 484
 485 static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 486 {
 487         struct evsel *pos;
 488         struct evlist_cpu_iterator evlist_cpu_itr;
 489         struct affinity saved_affinity, *affinity = NULL;
 490         bool has_imm = false;
 491
 492         // See explanation in evlist__close()
 493         if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 494                 if (affinity__setup(&saved_affinity) < 0)
 495                         return;
 496                 affinity = &saved_affinity;
 497         }
 498
 499         /* Disable 'immediate' events last */
 500         for (int imm = 0; imm <= 1; imm++) {
 501                 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 502                         pos = evlist_cpu_itr.evsel;
 503                         if (evsel__strcmp(pos, evsel_name))
 504                                 continue;
 505                         if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
 506                                 continue;
 507                         if (excl_dummy && evsel__is_dummy_event(pos))
 508                                 continue;
 509                         if (pos->immediate)
 510                                 has_imm = true;
 511                         if (pos->immediate != imm)
 512                                 continue;
 513                         evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 514                 }
 515                 if (!has_imm)
 516                         break;
 517         }
 518
 519         affinity__cleanup(affinity);
 520         evlist__for_each_entry(evlist, pos) {
 521                 if (evsel__strcmp(pos, evsel_name))
 522                         continue;
 523                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 524                         continue;
 525                 if (excl_dummy && evsel__is_dummy_event(pos))
 526                         continue;
 527                 pos->disabled = true;
 528         }
 529
 530         /*
 531          * If we disabled only single event, we need to check
 532          * the enabled state of the evlist manually.
 533          */
 534         if (evsel_name)
 535                 evlist->enabled = evlist__is_enabled(evlist);
 536         else
 537                 evlist->enabled = false;
 538 }
 539
 540 void evlist__disable(struct evlist *evlist)
 541 {
 542         __evlist__disable(evlist, NULL, false);
 543 }
 544
 545 void evlist__disable_non_dummy(struct evlist *evlist)
 546 {
 547         __evlist__disable(evlist, NULL, true);
 548 }
 549
 550 void evlist__disable_evsel(struct evlist *evlist, char *evsel_name)
 551 {
 552         __evlist__disable(evlist, evsel_name, false);
 553 }
 554
 555 static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy)
 556 {
 557         struct evsel *pos;
 558         struct evlist_cpu_iterator evlist_cpu_itr;
 559         struct affinity saved_affinity, *affinity = NULL;
 560
 561         // See explanation in evlist__close()
 562         if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
 563                 if (affinity__setup(&saved_affinity) < 0)
 564                         return;
 565                 affinity = &saved_affinity;
 566         }
 567
 568         evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) {
 569                 pos = evlist_cpu_itr.evsel;
 570                 if (evsel__strcmp(pos, evsel_name))
 571                         continue;
 572                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 573                         continue;
 574                 if (excl_dummy && evsel__is_dummy_event(pos))
 575                         continue;
 576                 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx);
 577         }
 578         affinity__cleanup(affinity);
 579         evlist__for_each_entry(evlist, pos) {
 580                 if (evsel__strcmp(pos, evsel_name))
 581                         continue;
 582                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 583                         continue;
 584                 if (excl_dummy && evsel__is_dummy_event(pos))
 585                         continue;
 586                 pos->disabled = false;
 587         }
 588
 589         /*
 590          * Even single event sets the 'enabled' for evlist,
 591          * so the toggle can work properly and toggle to
 592          * 'disabled' state.
 593          */
 594         evlist->enabled = true;
 595 }
 596
 597 void evlist__enable(struct evlist *evlist)
 598 {
 599         __evlist__enable(evlist, NULL, false);
 600 }
 601
 602 void evlist__enable_non_dummy(struct evlist *evlist)
 603 {
 604         __evlist__enable(evlist, NULL, true);
 605 }
 606
 607 void evlist__enable_evsel(struct evlist *evlist, char *evsel_name)
 608 {
 609         __evlist__enable(evlist, evsel_name, false);
 610 }
 611
 612 void evlist__toggle_enable(struct evlist *evlist)
 613 {
 614         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 615 }
 616
 617 int evlist__add_pollfd(struct evlist *evlist, int fd)
 618 {
 619         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
 620 }
 621
 622 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 623 {
 624         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 625 }
 626
 627 #ifdef HAVE_EVENTFD_SUPPORT
 628 int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd)
 629 {
 630         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
 631                                        fdarray_flag__nonfilterable |
 632                                        fdarray_flag__non_perf_event);
 633 }
 634 #endif
 635
 636 int evlist__poll(struct evlist *evlist, int timeout)
 637 {
 638         return perf_evlist__poll(&evlist->core, timeout);
 639 }
 640
 641 struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
 642 {
 643         struct hlist_head *head;
 644         struct perf_sample_id *sid;
 645         int hash;
 646
 647         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 648         head = &evlist->core.heads[hash];
 649
 650         hlist_for_each_entry(sid, head, node)
 651                 if (sid->id == id)
 652                         return sid;
 653
 654         return NULL;
 655 }
 656
 657 struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
 658 {
 659         struct perf_sample_id *sid;
 660
 661         if (evlist->core.nr_entries == 1 || !id)
 662                 return evlist__first(evlist);
 663
 664         sid = evlist__id2sid(evlist, id);
 665         if (sid)
 666                 return container_of(sid->evsel, struct evsel, core);
 667
 668         if (!evlist__sample_id_all(evlist))
 669                 return evlist__first(evlist);
 670
 671         return NULL;
 672 }
 673
 674 struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
 675 {
 676         struct perf_sample_id *sid;
 677
 678         if (!id)
 679                 return NULL;
 680
 681         sid = evlist__id2sid(evlist, id);
 682         if (sid)
 683                 return container_of(sid->evsel, struct evsel, core);
 684
 685         return NULL;
 686 }
 687
 688 static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
 689 {
 690         const __u64 *array = event->sample.array;
 691         ssize_t n;
 692
 693         n = (event->header.size - sizeof(event->header)) >> 3;
 694
 695         if (event->header.type == PERF_RECORD_SAMPLE) {
 696                 if (evlist->id_pos >= n)
 697                         return -1;
 698                 *id = array[evlist->id_pos];
 699         } else {
 700                 if (evlist->is_pos > n)
 701                         return -1;
 702                 n -= evlist->is_pos;
 703                 *id = array[n];
 704         }
 705         return 0;
 706 }
 707
 708 struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
 709 {
 710         struct evsel *first = evlist__first(evlist);
 711         struct hlist_head *head;
 712         struct perf_sample_id *sid;
 713         int hash;
 714         u64 id;
 715
 716         if (evlist->core.nr_entries == 1)
 717                 return first;
 718
 719         if (!first->core.attr.sample_id_all &&
 720             event->header.type != PERF_RECORD_SAMPLE)
 721                 return first;
 722
 723         if (evlist__event2id(evlist, event, &id))
 724                 return NULL;
 725
 726         /* Synthesized events have an id of zero */
 727         if (!id)
 728                 return first;
 729
 730         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 731         head = &evlist->core.heads[hash];
 732
 733         hlist_for_each_entry(sid, head, node) {
 734                 if (sid->id == id)
 735                         return container_of(sid->evsel, struct evsel, core);
 736         }
 737         return NULL;
 738 }
 739
 740 static int evlist__set_paused(struct evlist *evlist, bool value)
 741 {
 742         int i;
 743
 744         if (!evlist->overwrite_mmap)
 745                 return 0;
 746
 747         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 748                 int fd = evlist->overwrite_mmap[i].core.fd;
 749                 int err;
 750
 751                 if (fd < 0)
 752                         continue;
 753                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 754                 if (err)
 755                         return err;
 756         }
 757         return 0;
 758 }
 759
 760 static int evlist__pause(struct evlist *evlist)
 761 {
 762         return evlist__set_paused(evlist, true);
 763 }
 764
 765 static int evlist__resume(struct evlist *evlist)
 766 {
 767         return evlist__set_paused(evlist, false);
 768 }
 769
 770 static void evlist__munmap_nofree(struct evlist *evlist)
 771 {
 772         int i;
 773
 774         if (evlist->mmap)
 775                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 776                         perf_mmap__munmap(&evlist->mmap[i].core);
 777
 778         if (evlist->overwrite_mmap)
 779                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 780                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 781 }
 782
 783 void evlist__munmap(struct evlist *evlist)
 784 {
 785         evlist__munmap_nofree(evlist);
 786         zfree(&evlist->mmap);
 787         zfree(&evlist->overwrite_mmap);
 788 }
 789
 790 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 791 {
 792         struct mmap *m = container_of(map, struct mmap, core);
 793
 794         mmap__munmap(m);
 795 }
 796
 797 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 798                                        bool overwrite)
 799 {
 800         int i;
 801         struct mmap *map;
 802
 803         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 804         if (!map)
 805                 return NULL;
 806
 807         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 808                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 809
 810                 /*
 811                  * When the perf_mmap() call is made we grab one refcount, plus
 812                  * one extra to let perf_mmap__consume() get the last
 813                  * events after all real references (perf_mmap__get()) are
 814                  * dropped.
 815                  *
 816                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 817                  * thus does perf_mmap__get() on it.
 818                  */
 819                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 820         }
 821
 822         return map;
 823 }
 824
 825 static void
 826 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 827                          struct perf_evsel *_evsel,
 828                          struct perf_mmap_param *_mp,
 829                          int idx)
 830 {
 831         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 832         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 833         struct evsel *evsel = container_of(_evsel, struct evsel, core);
 834
 835         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx);
 836 }
 837
 838 static struct perf_mmap*
 839 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 840 {
 841         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 842         struct mmap *maps;
 843
 844         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 845
 846         if (!maps) {
 847                 maps = evlist__alloc_mmap(evlist, overwrite);
 848                 if (!maps)
 849                         return NULL;
 850
 851                 if (overwrite) {
 852                         evlist->overwrite_mmap = maps;
 853                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 854                                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 855                 } else {
 856                         evlist->mmap = maps;
 857                 }
 858         }
 859
 860         return &maps[idx].core;
 861 }
 862
 863 static int
 864 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 865                           int output, struct perf_cpu cpu)
 866 {
 867         struct mmap *map = container_of(_map, struct mmap, core);
 868         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 869
 870         return mmap__mmap(map, mp, output, cpu);
 871 }
 872
 873 unsigned long perf_event_mlock_kb_in_pages(void)
 874 {
 875         unsigned long pages;
 876         int max;
 877
 878         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 879                 /*
 880                  * Pick a once upon a time good value, i.e. things look
 881                  * strange since we can't read a sysctl value, but lets not
 882                  * die yet...
 883                  */
 884                 max = 512;
 885         } else {
 886                 max -= (page_size / 1024);
 887         }
 888
 889         pages = (max * 1024) / page_size;
 890         if (!is_power_of_2(pages))
 891                 pages = rounddown_pow_of_two(pages);
 892
 893         return pages;
 894 }
 895
 896 size_t evlist__mmap_size(unsigned long pages)
 897 {
 898         if (pages == UINT_MAX)
 899                 pages = perf_event_mlock_kb_in_pages();
 900         else if (!is_power_of_2(pages))
 901                 return 0;
 902
 903         return (pages + 1) * page_size;
 904 }
 905
 906 static long parse_pages_arg(const char *str, unsigned long min,
 907                             unsigned long max)
 908 {
 909         unsigned long pages, val;
 910         static struct parse_tag tags[] = {
 911                 { .tag  = 'B', .mult = 1       },
 912                 { .tag  = 'K', .mult = 1 << 10 },
 913                 { .tag  = 'M', .mult = 1 << 20 },
 914                 { .tag  = 'G', .mult = 1 << 30 },
 915                 { .tag  = 0 },
 916         };
 917
 918         if (str == NULL)
 919                 return -EINVAL;
 920
 921         val = parse_tag_value(str, tags);
 922         if (val != (unsigned long) -1) {
 923                 /* we got file size value */
 924                 pages = PERF_ALIGN(val, page_size) / page_size;
 925         } else {
 926                 /* we got pages count value */
 927                 char *eptr;
 928                 pages = strtoul(str, &eptr, 10);
 929                 if (*eptr != '\0')
 930                         return -EINVAL;
 931         }
 932
 933         if (pages == 0 && min == 0) {
 934                 /* leave number of pages at 0 */
 935         } else if (!is_power_of_2(pages)) {
 936                 char buf[100];
 937
 938                 /* round pages up to next power of 2 */
 939                 pages = roundup_pow_of_two(pages);
 940                 if (!pages)
 941                         return -EINVAL;
 942
 943                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 944                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 945                         buf, pages);
 946         }
 947
 948         if (pages > max)
 949                 return -EINVAL;
 950
 951         return pages;
 952 }
 953
 954 int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 955 {
 956         unsigned long max = UINT_MAX;
 957         long pages;
 958
 959         if (max > SIZE_MAX / page_size)
 960                 max = SIZE_MAX / page_size;
 961
 962         pages = parse_pages_arg(str, 1, max);
 963         if (pages < 0) {
 964                 pr_err("Invalid argument for --mmap_pages/-m\n");
 965                 return -1;
 966         }
 967
 968         *mmap_pages = pages;
 969         return 0;
 970 }
 971
 972 int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
 973 {
 974         return __evlist__parse_mmap_pages(opt->value, str);
 975 }
 976
 977 /**
 978  * evlist__mmap_ex - Create mmaps to receive events.
 979  * @evlist: list of events
 980  * @pages: map length in pages
 981  * @overwrite: overwrite older events?
 982  * @auxtrace_pages - auxtrace map length in pages
 983  * @auxtrace_overwrite - overwrite older auxtrace data?
 984  *
 985  * If @overwrite is %false the user needs to signal event consumption using
 986  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 987  * automatically.
 988  *
 989  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 990  * consumption using auxtrace_mmap__write_tail().
 991  *
 992  * Return: %0 on success, negative error code otherwise.
 993  */
 994 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 995                          unsigned int auxtrace_pages,
 996                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 997                          int comp_level)
 998 {
 999         /*
1000          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
1001          * Its value is decided by evsel's write_backward.
1002          * So &mp should not be passed through const pointer.
1003          */
1004         struct mmap_params mp = {
1005                 .nr_cblocks     = nr_cblocks,
1006                 .affinity       = affinity,
1007                 .flush          = flush,
1008                 .comp_level     = comp_level
1009         };
1010         struct perf_evlist_mmap_ops ops = {
1011                 .idx  = perf_evlist__mmap_cb_idx,
1012                 .get  = perf_evlist__mmap_cb_get,
1013                 .mmap = perf_evlist__mmap_cb_mmap,
1014         };
1015
1016         evlist->core.mmap_len = evlist__mmap_size(pages);
1017         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
1018
1019         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
1020                                    auxtrace_pages, auxtrace_overwrite);
1021
1022         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
1023 }
1024
1025 int evlist__mmap(struct evlist *evlist, unsigned int pages)
1026 {
1027         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
1028 }
1029
1030 int evlist__create_maps(struct evlist *evlist, struct target *target)
1031 {
1032         bool all_threads = (target->per_thread && target->system_wide);
1033         struct perf_cpu_map *cpus;
1034         struct perf_thread_map *threads;
1035
1036         /*
1037          * If specify '-a' and '--per-thread' to perf record, perf record
1038          * will override '--per-thread'. target->per_thread = false and
1039          * target->system_wide = true.
1040          *
1041          * If specify '--per-thread' only to perf record,
1042          * target->per_thread = true and target->system_wide = false.
1043          *
1044          * So target->per_thread && target->system_wide is false.
1045          * For perf record, thread_map__new_str doesn't call
1046          * thread_map__new_all_cpus. That will keep perf record's
1047          * current behavior.
1048          *
1049          * For perf stat, it allows the case that target->per_thread and
1050          * target->system_wide are all true. It means to collect system-wide
1051          * per-thread data. thread_map__new_str will call
1052          * thread_map__new_all_cpus to enumerate all threads.
1053          */
1054         threads = thread_map__new_str(target->pid, target->tid, target->uid,
1055                                       all_threads);
1056
1057         if (!threads)
1058                 return -1;
1059
1060         if (target__uses_dummy_map(target))
1061                 cpus = perf_cpu_map__dummy_new();
1062         else
1063                 cpus = perf_cpu_map__new(target->cpu_list);
1064
1065         if (!cpus)
1066                 goto out_delete_threads;
1067
1068         evlist->core.has_user_cpus = !!target->cpu_list && !target->hybrid;
1069
1070         perf_evlist__set_maps(&evlist->core, cpus, threads);
1071
1072         /* as evlist now has references, put count here */
1073         perf_cpu_map__put(cpus);
1074         perf_thread_map__put(threads);
1075
1076         return 0;
1077
1078 out_delete_threads:
1079         perf_thread_map__put(threads);
1080         return -1;
1081 }
1082
1083 int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
1084 {
1085         struct evsel *evsel;
1086         int err = 0;
1087
1088         evlist__for_each_entry(evlist, evsel) {
1089                 if (evsel->filter == NULL)
1090                         continue;
1091
1092                 /*
1093                  * filters only work for tracepoint event, which doesn't have cpu limit.
1094                  * So evlist and evsel should always be same.
1095                  */
1096                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
1097                 if (err) {
1098                         *err_evsel = evsel;
1099                         break;
1100                 }
1101         }
1102
1103         return err;
1104 }
1105
1106 int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
1107 {
1108         struct evsel *evsel;
1109         int err = 0;
1110
1111         if (filter == NULL)
1112                 return -1;
1113
1114         evlist__for_each_entry(evlist, evsel) {
1115                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1116                         continue;
1117
1118                 err = evsel__set_filter(evsel, filter);
1119                 if (err)
1120                         break;
1121         }
1122
1123         return err;
1124 }
1125
1126 int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
1127 {
1128         struct evsel *evsel;
1129         int err = 0;
1130
1131         if (filter == NULL)
1132                 return -1;
1133
1134         evlist__for_each_entry(evlist, evsel) {
1135                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1136                         continue;
1137
1138                 err = evsel__append_tp_filter(evsel, filter);
1139                 if (err)
1140                         break;
1141         }
1142
1143         return err;
1144 }
1145
1146 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1147 {
1148         char *filter;
1149         size_t i;
1150
1151         for (i = 0; i < npids; ++i) {
1152                 if (i == 0) {
1153                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1154                                 return NULL;
1155                 } else {
1156                         char *tmp;
1157
1158                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1159                                 goto out_free;
1160
1161                         free(filter);
1162                         filter = tmp;
1163                 }
1164         }
1165
1166         return filter;
1167 out_free:
1168         free(filter);
1169         return NULL;
1170 }
1171
1172 int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1173 {
1174         char *filter = asprintf__tp_filter_pids(npids, pids);
1175         int ret = evlist__set_tp_filter(evlist, filter);
1176
1177         free(filter);
1178         return ret;
1179 }
1180
1181 int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1182 {
1183         return evlist__set_tp_filter_pids(evlist, 1, &pid);
1184 }
1185
1186 int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1187 {
1188         char *filter = asprintf__tp_filter_pids(npids, pids);
1189         int ret = evlist__append_tp_filter(evlist, filter);
1190
1191         free(filter);
1192         return ret;
1193 }
1194
1195 int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1196 {
1197         return evlist__append_tp_filter_pids(evlist, 1, &pid);
1198 }
1199
1200 bool evlist__valid_sample_type(struct evlist *evlist)
1201 {
1202         struct evsel *pos;
1203
1204         if (evlist->core.nr_entries == 1)
1205                 return true;
1206
1207         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1208                 return false;
1209
1210         evlist__for_each_entry(evlist, pos) {
1211                 if (pos->id_pos != evlist->id_pos ||
1212                     pos->is_pos != evlist->is_pos)
1213                         return false;
1214         }
1215
1216         return true;
1217 }
1218
1219 u64 __evlist__combined_sample_type(struct evlist *evlist)
1220 {
1221         struct evsel *evsel;
1222
1223         if (evlist->combined_sample_type)
1224                 return evlist->combined_sample_type;
1225
1226         evlist__for_each_entry(evlist, evsel)
1227                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1228
1229         return evlist->combined_sample_type;
1230 }
1231
1232 u64 evlist__combined_sample_type(struct evlist *evlist)
1233 {
1234         evlist->combined_sample_type = 0;
1235         return __evlist__combined_sample_type(evlist);
1236 }
1237
1238 u64 evlist__combined_branch_type(struct evlist *evlist)
1239 {
1240         struct evsel *evsel;
1241         u64 branch_type = 0;
1242
1243         evlist__for_each_entry(evlist, evsel)
1244                 branch_type |= evsel->core.attr.branch_sample_type;
1245         return branch_type;
1246 }
1247
1248 bool evlist__valid_read_format(struct evlist *evlist)
1249 {
1250         struct evsel *first = evlist__first(evlist), *pos = first;
1251         u64 read_format = first->core.attr.read_format;
1252         u64 sample_type = first->core.attr.sample_type;
1253
1254         evlist__for_each_entry(evlist, pos) {
1255                 if (read_format != pos->core.attr.read_format) {
1256                         pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1257                                  read_format, (u64)pos->core.attr.read_format);
1258                 }
1259         }
1260
1261         /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */
1262         if ((sample_type & PERF_SAMPLE_READ) &&
1263             !(read_format & PERF_FORMAT_ID)) {
1264                 return false;
1265         }
1266
1267         return true;
1268 }
1269
1270 u16 evlist__id_hdr_size(struct evlist *evlist)
1271 {
1272         struct evsel *first = evlist__first(evlist);
1273
1274         return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0;
1275 }
1276
1277 bool evlist__valid_sample_id_all(struct evlist *evlist)
1278 {
1279         struct evsel *first = evlist__first(evlist), *pos = first;
1280
1281         evlist__for_each_entry_continue(evlist, pos) {
1282                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1283                         return false;
1284         }
1285
1286         return true;
1287 }
1288
1289 bool evlist__sample_id_all(struct evlist *evlist)
1290 {
1291         struct evsel *first = evlist__first(evlist);
1292         return first->core.attr.sample_id_all;
1293 }
1294
1295 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1296 {
1297         evlist->selected = evsel;
1298 }
1299
1300 void evlist__close(struct evlist *evlist)
1301 {
1302         struct evsel *evsel;
1303         struct evlist_cpu_iterator evlist_cpu_itr;
1304         struct affinity affinity;
1305
1306         /*
1307          * With perf record core.user_requested_cpus is usually NULL.
1308          * Use the old method to handle this for now.
1309          */
1310         if (!evlist->core.user_requested_cpus ||
1311             cpu_map__is_dummy(evlist->core.user_requested_cpus)) {
1312                 evlist__for_each_entry_reverse(evlist, evsel)
1313                         evsel__close(evsel);
1314                 return;
1315         }
1316
1317         if (affinity__setup(&affinity) < 0)
1318                 return;
1319
1320         evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) {
1321                 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core,
1322                                       evlist_cpu_itr.cpu_map_idx);
1323         }
1324
1325         affinity__cleanup(&affinity);
1326         evlist__for_each_entry_reverse(evlist, evsel) {
1327                 perf_evsel__free_fd(&evsel->core);
1328                 perf_evsel__free_id(&evsel->core);
1329         }
1330         perf_evlist__reset_id_hash(&evlist->core);
1331 }
1332
1333 static int evlist__create_syswide_maps(struct evlist *evlist)
1334 {
1335         struct perf_cpu_map *cpus;
1336         struct perf_thread_map *threads;
1337
1338         /*
1339          * Try reading /sys/devices/system/cpu/online to get
1340          * an all cpus map.
1341          *
1342          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1343          * code needs an overhaul to properly forward the
1344          * error, and we may not want to do that fallback to a
1345          * default cpu identity map :-\
1346          */
1347         cpus = perf_cpu_map__new(NULL);
1348         if (!cpus)
1349                 goto out;
1350
1351         threads = perf_thread_map__new_dummy();
1352         if (!threads)
1353                 goto out_put;
1354
1355         perf_evlist__set_maps(&evlist->core, cpus, threads);
1356
1357         perf_thread_map__put(threads);
1358 out_put:
1359         perf_cpu_map__put(cpus);
1360 out:
1361         return -ENOMEM;
1362 }
1363
1364 int evlist__open(struct evlist *evlist)
1365 {
1366         struct evsel *evsel;
1367         int err;
1368
1369         /*
1370          * Default: one fd per CPU, all threads, aka systemwide
1371          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1372          */
1373         if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) {
1374                 err = evlist__create_syswide_maps(evlist);
1375                 if (err < 0)
1376                         goto out_err;
1377         }
1378
1379         evlist__update_id_pos(evlist);
1380
1381         evlist__for_each_entry(evlist, evsel) {
1382                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1383                 if (err < 0)
1384                         goto out_err;
1385         }
1386
1387         return 0;
1388 out_err:
1389         evlist__close(evlist);
1390         errno = -err;
1391         return err;
1392 }
1393
1394 int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1395                              bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1396 {
1397         int child_ready_pipe[2], go_pipe[2];
1398         char bf;
1399
1400         if (pipe(child_ready_pipe) < 0) {
1401                 perror("failed to create 'ready' pipe");
1402                 return -1;
1403         }
1404
1405         if (pipe(go_pipe) < 0) {
1406                 perror("failed to create 'go' pipe");
1407                 goto out_close_ready_pipe;
1408         }
1409
1410         evlist->workload.pid = fork();
1411         if (evlist->workload.pid < 0) {
1412                 perror("failed to fork");
1413                 goto out_close_pipes;
1414         }
1415
1416         if (!evlist->workload.pid) {
1417                 int ret;
1418
1419                 if (pipe_output)
1420                         dup2(2, 1);
1421
1422                 signal(SIGTERM, SIG_DFL);
1423
1424                 close(child_ready_pipe[0]);
1425                 close(go_pipe[1]);
1426                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1427
1428                 /*
1429                  * Change the name of this process not to confuse --exclude-perf users
1430                  * that sees 'perf' in the window up to the execvp() and thinks that
1431                  * perf samples are not being excluded.
1432                  */
1433                 prctl(PR_SET_NAME, "perf-exec");
1434
1435                 /*
1436                  * Tell the parent we're ready to go
1437                  */
1438                 close(child_ready_pipe[1]);
1439
1440                 /*
1441                  * Wait until the parent tells us to go.
1442                  */
1443                 ret = read(go_pipe[0], &bf, 1);
1444                 /*
1445                  * The parent will ask for the execvp() to be performed by
1446                  * writing exactly one byte, in workload.cork_fd, usually via
1447                  * evlist__start_workload().
1448                  *
1449                  * For cancelling the workload without actually running it,
1450                  * the parent will just close workload.cork_fd, without writing
1451                  * anything, i.e. read will return zero and we just exit()
1452                  * here.
1453                  */
1454                 if (ret != 1) {
1455                         if (ret == -1)
1456                                 perror("unable to read pipe");
1457                         exit(ret);
1458                 }
1459
1460                 execvp(argv[0], (char **)argv);
1461
1462                 if (exec_error) {
1463                         union sigval val;
1464
1465                         val.sival_int = errno;
1466                         if (sigqueue(getppid(), SIGUSR1, val))
1467                                 perror(argv[0]);
1468                 } else
1469                         perror(argv[0]);
1470                 exit(-1);
1471         }
1472
1473         if (exec_error) {
1474                 struct sigaction act = {
1475                         .sa_flags     = SA_SIGINFO,
1476                         .sa_sigaction = exec_error,
1477                 };
1478                 sigaction(SIGUSR1, &act, NULL);
1479         }
1480
1481         if (target__none(target)) {
1482                 if (evlist->core.threads == NULL) {
1483                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1484                                 __func__, __LINE__);
1485                         goto out_close_pipes;
1486                 }
1487                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1488         }
1489
1490         close(child_ready_pipe[1]);
1491         close(go_pipe[0]);
1492         /*
1493          * wait for child to settle
1494          */
1495         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1496                 perror("unable to read pipe");
1497                 goto out_close_pipes;
1498         }
1499
1500         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1501         evlist->workload.cork_fd = go_pipe[1];
1502         close(child_ready_pipe[0]);
1503         return 0;
1504
1505 out_close_pipes:
1506         close(go_pipe[0]);
1507         close(go_pipe[1]);
1508 out_close_ready_pipe:
1509         close(child_ready_pipe[0]);
1510         close(child_ready_pipe[1]);
1511         return -1;
1512 }
1513
1514 int evlist__start_workload(struct evlist *evlist)
1515 {
1516         if (evlist->workload.cork_fd > 0) {
1517                 char bf = 0;
1518                 int ret;
1519                 /*
1520                  * Remove the cork, let it rip!
1521                  */
1522                 ret = write(evlist->workload.cork_fd, &bf, 1);
1523                 if (ret < 0)
1524                         perror("unable to write to pipe");
1525
1526                 close(evlist->workload.cork_fd);
1527                 return ret;
1528         }
1529
1530         return 0;
1531 }
1532
1533 int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1534 {
1535         struct evsel *evsel = evlist__event2evsel(evlist, event);
1536         int ret;
1537
1538         if (!evsel)
1539                 return -EFAULT;
1540         ret = evsel__parse_sample(evsel, event, sample);
1541         if (ret)
1542                 return ret;
1543         if (perf_guest && sample->id) {
1544                 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id);
1545
1546                 if (sid) {
1547                         sample->machine_pid = sid->machine_pid;
1548                         sample->vcpu = sid->vcpu.cpu;
1549                 }
1550         }
1551         return 0;
1552 }
1553
1554 int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1555 {
1556         struct evsel *evsel = evlist__event2evsel(evlist, event);
1557
1558         if (!evsel)
1559                 return -EFAULT;
1560         return evsel__parse_sample_timestamp(evsel, event, timestamp);
1561 }
1562
1563 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1564 {
1565         int printed, value;
1566         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1567
1568         switch (err) {
1569         case EACCES:
1570         case EPERM:
1571                 printed = scnprintf(buf, size,
1572                                     "Error:\t%s.\n"
1573                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1574
1575                 value = perf_event_paranoid();
1576
1577                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1578
1579                 if (value >= 2) {
1580                         printed += scnprintf(buf + printed, size - printed,
1581                                              "For your workloads it needs to be <= 1\nHint:\t");
1582                 }
1583                 printed += scnprintf(buf + printed, size - printed,
1584                                      "For system wide tracing it needs to be set to -1.\n");
1585
1586                 printed += scnprintf(buf + printed, size - printed,
1587                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1588                                     "Hint:\tThe current value is %d.", value);
1589                 break;
1590         case EINVAL: {
1591                 struct evsel *first = evlist__first(evlist);
1592                 int max_freq;
1593
1594                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1595                         goto out_default;
1596
1597                 if (first->core.attr.sample_freq < (u64)max_freq)
1598                         goto out_default;
1599
1600                 printed = scnprintf(buf, size,
1601                                     "Error:\t%s.\n"
1602                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1603                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1604                                     emsg, max_freq, first->core.attr.sample_freq);
1605                 break;
1606         }
1607         default:
1608 out_default:
1609                 scnprintf(buf, size, "%s", emsg);
1610                 break;
1611         }
1612
1613         return 0;
1614 }
1615
1616 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1617 {
1618         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1619         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1620
1621         switch (err) {
1622         case EPERM:
1623                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1624                 printed += scnprintf(buf + printed, size - printed,
1625                                      "Error:\t%s.\n"
1626                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1627                                      "Hint:\tTried using %zd kB.\n",
1628                                      emsg, pages_max_per_user, pages_attempted);
1629
1630                 if (pages_attempted >= pages_max_per_user) {
1631                         printed += scnprintf(buf + printed, size - printed,
1632                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1633                                              pages_max_per_user + pages_attempted);
1634                 }
1635
1636                 printed += scnprintf(buf + printed, size - printed,
1637                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1638                 break;
1639         default:
1640                 scnprintf(buf, size, "%s", emsg);
1641                 break;
1642         }
1643
1644         return 0;
1645 }
1646
1647 void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1648 {
1649         struct evsel *evsel, *n;
1650         LIST_HEAD(move);
1651
1652         if (move_evsel == evlist__first(evlist))
1653                 return;
1654
1655         evlist__for_each_entry_safe(evlist, n, evsel) {
1656                 if (evsel__leader(evsel) == evsel__leader(move_evsel))
1657                         list_move_tail(&evsel->core.node, &move);
1658         }
1659
1660         list_splice(&move, &evlist->core.entries);
1661 }
1662
1663 struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1664 {
1665         struct evsel *evsel;
1666
1667         evlist__for_each_entry(evlist, evsel) {
1668                 if (evsel->tracking)
1669                         return evsel;
1670         }
1671
1672         return evlist__first(evlist);
1673 }
1674
1675 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1676 {
1677         struct evsel *evsel;
1678
1679         if (tracking_evsel->tracking)
1680                 return;
1681
1682         evlist__for_each_entry(evlist, evsel) {
1683                 if (evsel != tracking_evsel)
1684                         evsel->tracking = false;
1685         }
1686
1687         tracking_evsel->tracking = true;
1688 }
1689
1690 struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1691 {
1692         struct evsel *evsel;
1693
1694         evlist__for_each_entry(evlist, evsel) {
1695                 if (!evsel->name)
1696                         continue;
1697                 if (strcmp(str, evsel->name) == 0)
1698                         return evsel;
1699         }
1700
1701         return NULL;
1702 }
1703
1704 void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1705 {
1706         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1707         enum action {
1708                 NONE,
1709                 PAUSE,
1710                 RESUME,
1711         } action = NONE;
1712
1713         if (!evlist->overwrite_mmap)
1714                 return;
1715
1716         switch (old_state) {
1717         case BKW_MMAP_NOTREADY: {
1718                 if (state != BKW_MMAP_RUNNING)
1719                         goto state_err;
1720                 break;
1721         }
1722         case BKW_MMAP_RUNNING: {
1723                 if (state != BKW_MMAP_DATA_PENDING)
1724                         goto state_err;
1725                 action = PAUSE;
1726                 break;
1727         }
1728         case BKW_MMAP_DATA_PENDING: {
1729                 if (state != BKW_MMAP_EMPTY)
1730                         goto state_err;
1731                 break;
1732         }
1733         case BKW_MMAP_EMPTY: {
1734                 if (state != BKW_MMAP_RUNNING)
1735                         goto state_err;
1736                 action = RESUME;
1737                 break;
1738         }
1739         default:
1740                 WARN_ONCE(1, "Shouldn't get there\n");
1741         }
1742
1743         evlist->bkw_mmap_state = state;
1744
1745         switch (action) {
1746         case PAUSE:
1747                 evlist__pause(evlist);
1748                 break;
1749         case RESUME:
1750                 evlist__resume(evlist);
1751                 break;
1752         case NONE:
1753         default:
1754                 break;
1755         }
1756
1757 state_err:
1758         return;
1759 }
1760
1761 bool evlist__exclude_kernel(struct evlist *evlist)
1762 {
1763         struct evsel *evsel;
1764
1765         evlist__for_each_entry(evlist, evsel) {
1766                 if (!evsel->core.attr.exclude_kernel)
1767                         return false;
1768         }
1769
1770         return true;
1771 }
1772
1773 /*
1774  * Events in data file are not collect in groups, but we still want
1775  * the group display. Set the artificial group and set the leader's
1776  * forced_leader flag to notify the display code.
1777  */
1778 void evlist__force_leader(struct evlist *evlist)
1779 {
1780         if (!evlist->core.nr_groups) {
1781                 struct evsel *leader = evlist__first(evlist);
1782
1783                 evlist__set_leader(evlist);
1784                 leader->forced_leader = true;
1785         }
1786 }
1787
1788 struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1789 {
1790         struct evsel *c2, *leader;
1791         bool is_open = true;
1792
1793         leader = evsel__leader(evsel);
1794
1795         pr_debug("Weak group for %s/%d failed\n",
1796                         leader->name, leader->core.nr_members);
1797
1798         /*
1799          * for_each_group_member doesn't work here because it doesn't
1800          * include the first entry.
1801          */
1802         evlist__for_each_entry(evsel_list, c2) {
1803                 if (c2 == evsel)
1804                         is_open = false;
1805                 if (evsel__has_leader(c2, leader)) {
1806                         if (is_open && close)
1807                                 perf_evsel__close(&c2->core);
1808                         /*
1809                          * We want to close all members of the group and reopen
1810                          * them. Some events, like Intel topdown, require being
1811                          * in a group and so keep these in the group.
1812                          */
1813                         evsel__remove_from_group(c2, leader);
1814
1815                         /*
1816                          * Set this for all former members of the group
1817                          * to indicate they get reopened.
1818                          */
1819                         c2->reset_group = true;
1820                 }
1821         }
1822         /* Reset the leader count if all entries were removed. */
1823         if (leader->core.nr_members == 1)
1824                 leader->core.nr_members = 0;
1825         return leader;
1826 }
1827
1828 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1829 {
1830         char *s, *p;
1831         int ret = 0, fd;
1832
1833         if (strncmp(str, "fifo:", 5))
1834                 return -EINVAL;
1835
1836         str += 5;
1837         if (!*str || *str == ',')
1838                 return -EINVAL;
1839
1840         s = strdup(str);
1841         if (!s)
1842                 return -ENOMEM;
1843
1844         p = strchr(s, ',');
1845         if (p)
1846                 *p = '\0';
1847
1848         /*
1849          * O_RDWR avoids POLLHUPs which is necessary to allow the other
1850          * end of a FIFO to be repeatedly opened and closed.
1851          */
1852         fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1853         if (fd < 0) {
1854                 pr_err("Failed to open '%s'\n", s);
1855                 ret = -errno;
1856                 goto out_free;
1857         }
1858         *ctl_fd = fd;
1859         *ctl_fd_close = true;
1860
1861         if (p && *++p) {
1862                 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1863                 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1864                 if (fd < 0) {
1865                         pr_err("Failed to open '%s'\n", p);
1866                         ret = -errno;
1867                         goto out_free;
1868                 }
1869                 *ctl_fd_ack = fd;
1870         }
1871
1872 out_free:
1873         free(s);
1874         return ret;
1875 }
1876
1877 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1878 {
1879         char *comma = NULL, *endptr = NULL;
1880
1881         *ctl_fd_close = false;
1882
1883         if (strncmp(str, "fd:", 3))
1884                 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1885
1886         *ctl_fd = strtoul(&str[3], &endptr, 0);
1887         if (endptr == &str[3])
1888                 return -EINVAL;
1889
1890         comma = strchr(str, ',');
1891         if (comma) {
1892                 if (endptr != comma)
1893                         return -EINVAL;
1894
1895                 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1896                 if (endptr == comma + 1 || *endptr != '\0')
1897                         return -EINVAL;
1898         }
1899
1900         return 0;
1901 }
1902
1903 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1904 {
1905         if (*ctl_fd_close) {
1906                 *ctl_fd_close = false;
1907                 close(ctl_fd);
1908                 if (ctl_fd_ack >= 0)
1909                         close(ctl_fd_ack);
1910         }
1911 }
1912
1913 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1914 {
1915         if (fd == -1) {
1916                 pr_debug("Control descriptor is not initialized\n");
1917                 return 0;
1918         }
1919
1920         evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1921                                                      fdarray_flag__nonfilterable |
1922                                                      fdarray_flag__non_perf_event);
1923         if (evlist->ctl_fd.pos < 0) {
1924                 evlist->ctl_fd.pos = -1;
1925                 pr_err("Failed to add ctl fd entry: %m\n");
1926                 return -1;
1927         }
1928
1929         evlist->ctl_fd.fd = fd;
1930         evlist->ctl_fd.ack = ack;
1931
1932         return 0;
1933 }
1934
1935 bool evlist__ctlfd_initialized(struct evlist *evlist)
1936 {
1937         return evlist->ctl_fd.pos >= 0;
1938 }
1939
1940 int evlist__finalize_ctlfd(struct evlist *evlist)
1941 {
1942         struct pollfd *entries = evlist->core.pollfd.entries;
1943
1944         if (!evlist__ctlfd_initialized(evlist))
1945                 return 0;
1946
1947         entries[evlist->ctl_fd.pos].fd = -1;
1948         entries[evlist->ctl_fd.pos].events = 0;
1949         entries[evlist->ctl_fd.pos].revents = 0;
1950
1951         evlist->ctl_fd.pos = -1;
1952         evlist->ctl_fd.ack = -1;
1953         evlist->ctl_fd.fd = -1;
1954
1955         return 0;
1956 }
1957
1958 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1959                               char *cmd_data, size_t data_size)
1960 {
1961         int err;
1962         char c;
1963         size_t bytes_read = 0;
1964
1965         *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1966         memset(cmd_data, 0, data_size);
1967         data_size--;
1968
1969         do {
1970                 err = read(evlist->ctl_fd.fd, &c, 1);
1971                 if (err > 0) {
1972                         if (c == '\n' || c == '\0')
1973                                 break;
1974                         cmd_data[bytes_read++] = c;
1975                         if (bytes_read == data_size)
1976                                 break;
1977                         continue;
1978                 } else if (err == -1) {
1979                         if (errno == EINTR)
1980                                 continue;
1981                         if (errno == EAGAIN || errno == EWOULDBLOCK)
1982                                 err = 0;
1983                         else
1984                                 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1985                 }
1986                 break;
1987         } while (1);
1988
1989         pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1990                  bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1991
1992         if (bytes_read > 0) {
1993                 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1994                              (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1995                         *cmd = EVLIST_CTL_CMD_ENABLE;
1996                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1997                                     (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1998                         *cmd = EVLIST_CTL_CMD_DISABLE;
1999                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
2000                                     (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
2001                         *cmd = EVLIST_CTL_CMD_SNAPSHOT;
2002                         pr_debug("is snapshot\n");
2003                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG,
2004                                     (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) {
2005                         *cmd = EVLIST_CTL_CMD_EVLIST;
2006                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG,
2007                                     (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) {
2008                         *cmd = EVLIST_CTL_CMD_STOP;
2009                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG,
2010                                     (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) {
2011                         *cmd = EVLIST_CTL_CMD_PING;
2012                 }
2013         }
2014
2015         return bytes_read ? (int)bytes_read : err;
2016 }
2017
2018 int evlist__ctlfd_ack(struct evlist *evlist)
2019 {
2020         int err;
2021
2022         if (evlist->ctl_fd.ack == -1)
2023                 return 0;
2024
2025         err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
2026                     sizeof(EVLIST_CTL_CMD_ACK_TAG));
2027         if (err == -1)
2028                 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
2029
2030         return err;
2031 }
2032
2033 static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg)
2034 {
2035         char *data = cmd_data + cmd_size;
2036
2037         /* no argument */
2038         if (!*data)
2039                 return 0;
2040
2041         /* there's argument */
2042         if (*data == ' ') {
2043                 *arg = data + 1;
2044                 return 1;
2045         }
2046
2047         /* malformed */
2048         return -1;
2049 }
2050
2051 static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable)
2052 {
2053         struct evsel *evsel;
2054         char *name;
2055         int err;
2056
2057         err = get_cmd_arg(cmd_data,
2058                           enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 :
2059                                    sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1,
2060                           &name);
2061         if (err < 0) {
2062                 pr_info("failed: wrong command\n");
2063                 return -1;
2064         }
2065
2066         if (err) {
2067                 evsel = evlist__find_evsel_by_str(evlist, name);
2068                 if (evsel) {
2069                         if (enable)
2070                                 evlist__enable_evsel(evlist, name);
2071                         else
2072                                 evlist__disable_evsel(evlist, name);
2073                         pr_info("Event %s %s\n", evsel->name,
2074                                 enable ? "enabled" : "disabled");
2075                 } else {
2076                         pr_info("failed: can't find '%s' event\n", name);
2077                 }
2078         } else {
2079                 if (enable) {
2080                         evlist__enable(evlist);
2081                         pr_info(EVLIST_ENABLED_MSG);
2082                 } else {
2083                         evlist__disable(evlist);
2084                         pr_info(EVLIST_DISABLED_MSG);
2085                 }
2086         }
2087
2088         return 0;
2089 }
2090
2091 static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data)
2092 {
2093         struct perf_attr_details details = { .verbose = false, };
2094         struct evsel *evsel;
2095         char *arg;
2096         int err;
2097
2098         err = get_cmd_arg(cmd_data,
2099                           sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1,
2100                           &arg);
2101         if (err < 0) {
2102                 pr_info("failed: wrong command\n");
2103                 return -1;
2104         }
2105
2106         if (err) {
2107                 if (!strcmp(arg, "-v")) {
2108                         details.verbose = true;
2109                 } else if (!strcmp(arg, "-g")) {
2110                         details.event_group = true;
2111                 } else if (!strcmp(arg, "-F")) {
2112                         details.freq = true;
2113                 } else {
2114                         pr_info("failed: wrong command\n");
2115                         return -1;
2116                 }
2117         }
2118
2119         evlist__for_each_entry(evlist, evsel)
2120                 evsel__fprintf(evsel, &details, stderr);
2121
2122         return 0;
2123 }
2124
2125 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
2126 {
2127         int err = 0;
2128         char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
2129         int ctlfd_pos = evlist->ctl_fd.pos;
2130         struct pollfd *entries = evlist->core.pollfd.entries;
2131
2132         if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
2133                 return 0;
2134
2135         if (entries[ctlfd_pos].revents & POLLIN) {
2136                 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
2137                                          EVLIST_CTL_CMD_MAX_LEN);
2138                 if (err > 0) {
2139                         switch (*cmd) {
2140                         case EVLIST_CTL_CMD_ENABLE:
2141                         case EVLIST_CTL_CMD_DISABLE:
2142                                 err = evlist__ctlfd_enable(evlist, cmd_data,
2143                                                            *cmd == EVLIST_CTL_CMD_ENABLE);
2144                                 break;
2145                         case EVLIST_CTL_CMD_EVLIST:
2146                                 err = evlist__ctlfd_list(evlist, cmd_data);
2147                                 break;
2148                         case EVLIST_CTL_CMD_SNAPSHOT:
2149                         case EVLIST_CTL_CMD_STOP:
2150                         case EVLIST_CTL_CMD_PING:
2151                                 break;
2152                         case EVLIST_CTL_CMD_ACK:
2153                         case EVLIST_CTL_CMD_UNSUPPORTED:
2154                         default:
2155                                 pr_debug("ctlfd: unsupported %d\n", *cmd);
2156                                 break;
2157                         }
2158                         if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
2159                               *cmd == EVLIST_CTL_CMD_SNAPSHOT))
2160                                 evlist__ctlfd_ack(evlist);
2161                 }
2162         }
2163
2164         if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
2165                 evlist__finalize_ctlfd(evlist);
2166         else
2167                 entries[ctlfd_pos].revents = 0;
2168
2169         return err;
2170 }
2171
2172 /**
2173  * struct event_enable_time - perf record -D/--delay single time range.
2174  * @start: start of time range to enable events in milliseconds
2175  * @end: end of time range to enable events in milliseconds
2176  *
2177  * N.B. this structure is also accessed as an array of int.
2178  */
2179 struct event_enable_time {
2180         int     start;
2181         int     end;
2182 };
2183
2184 static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first)
2185 {
2186         const char *fmt = first ? "%u - %u %n" : " , %u - %u %n";
2187         int ret, start, end, n;
2188
2189         ret = sscanf(str, fmt, &start, &end, &n);
2190         if (ret != 2 || end <= start)
2191                 return -EINVAL;
2192         if (range) {
2193                 range->start = start;
2194                 range->end = end;
2195         }
2196         return n;
2197 }
2198
2199 static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range)
2200 {
2201         int incr = !!range;
2202         bool first = true;
2203         ssize_t ret, cnt;
2204
2205         for (cnt = 0; *str; cnt++) {
2206                 ret = parse_event_enable_time(str, range, first);
2207                 if (ret < 0)
2208                         return ret;
2209                 /* Check no overlap */
2210                 if (!first && range && range->start <= range[-1].end)
2211                         return -EINVAL;
2212                 str += ret;
2213                 range += incr;
2214                 first = false;
2215         }
2216         return cnt;
2217 }
2218
2219 /**
2220  * struct event_enable_timer - control structure for perf record -D/--delay.
2221  * @evlist: event list
2222  * @times: time ranges that events are enabled (N.B. this is also accessed as an
2223  *         array of int)
2224  * @times_cnt: number of time ranges
2225  * @timerfd: timer file descriptor
2226  * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray)
2227  * @times_step: current position in (int *)@times)[],
2228  *              refer event_enable_timer__process()
2229  *
2230  * Note, this structure is only used when there are time ranges, not when there
2231  * is only an initial delay.
2232  */
2233 struct event_enable_timer {
2234         struct evlist *evlist;
2235         struct event_enable_time *times;
2236         size_t  times_cnt;
2237         int     timerfd;
2238         int     pollfd_pos;
2239         size_t  times_step;
2240 };
2241
2242 static int str_to_delay(const char *str)
2243 {
2244         char *endptr;
2245         long d;
2246
2247         d = strtol(str, &endptr, 10);
2248         if (*endptr || d > INT_MAX || d < -1)
2249                 return 0;
2250         return d;
2251 }
2252
2253 int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts,
2254                                     const char *str, int unset)
2255 {
2256         enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event;
2257         struct event_enable_timer *eet;
2258         ssize_t times_cnt;
2259         ssize_t ret;
2260         int err;
2261
2262         if (unset)
2263                 return 0;
2264
2265         opts->initial_delay = str_to_delay(str);
2266         if (opts->initial_delay)
2267                 return 0;
2268
2269         ret = parse_event_enable_times(str, NULL);
2270         if (ret < 0)
2271                 return ret;
2272
2273         times_cnt = ret;
2274         if (times_cnt == 0)
2275                 return -EINVAL;
2276
2277         eet = zalloc(sizeof(*eet));
2278         if (!eet)
2279                 return -ENOMEM;
2280
2281         eet->times = calloc(times_cnt, sizeof(*eet->times));
2282         if (!eet->times) {
2283                 err = -ENOMEM;
2284                 goto free_eet;
2285         }
2286
2287         if (parse_event_enable_times(str, eet->times) != times_cnt) {
2288                 err = -EINVAL;
2289                 goto free_eet_times;
2290         }
2291
2292         eet->times_cnt = times_cnt;
2293
2294         eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC);
2295         if (eet->timerfd == -1) {
2296                 err = -errno;
2297                 pr_err("timerfd_create failed: %s\n", strerror(errno));
2298                 goto free_eet_times;
2299         }
2300
2301         eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags);
2302         if (eet->pollfd_pos < 0) {
2303                 err = eet->pollfd_pos;
2304                 goto close_timerfd;
2305         }
2306
2307         eet->evlist = evlist;
2308         evlist->eet = eet;
2309         opts->initial_delay = eet->times[0].start;
2310
2311         return 0;
2312
2313 close_timerfd:
2314         close(eet->timerfd);
2315 free_eet_times:
2316         free(eet->times);
2317 free_eet:
2318         free(eet);
2319         return err;
2320 }
2321
2322 static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms)
2323 {
2324         struct itimerspec its = {
2325                 .it_value.tv_sec = ms / MSEC_PER_SEC,
2326                 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC,
2327         };
2328         int err = 0;
2329
2330         if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) {
2331                 err = -errno;
2332                 pr_err("timerfd_settime failed: %s\n", strerror(errno));
2333         }
2334         return err;
2335 }
2336
2337 int event_enable_timer__start(struct event_enable_timer *eet)
2338 {
2339         int ms;
2340
2341         if (!eet)
2342                 return 0;
2343
2344         ms = eet->times[0].end - eet->times[0].start;
2345         eet->times_step = 1;
2346
2347         return event_enable_timer__set_timer(eet, ms);
2348 }
2349
2350 int event_enable_timer__process(struct event_enable_timer *eet)
2351 {
2352         struct pollfd *entries;
2353         short revents;
2354
2355         if (!eet)
2356                 return 0;
2357
2358         entries = eet->evlist->core.pollfd.entries;
2359         revents = entries[eet->pollfd_pos].revents;
2360         entries[eet->pollfd_pos].revents = 0;
2361
2362         if (revents & POLLIN) {
2363                 size_t step = eet->times_step;
2364                 size_t pos = step / 2;
2365
2366                 if (step & 1) {
2367                         evlist__disable_non_dummy(eet->evlist);
2368                         pr_info(EVLIST_DISABLED_MSG);
2369                         if (pos >= eet->times_cnt - 1) {
2370                                 /* Disarm timer */
2371                                 event_enable_timer__set_timer(eet, 0);
2372                                 return 1; /* Stop */
2373                         }
2374                 } else {
2375                         evlist__enable_non_dummy(eet->evlist);
2376                         pr_info(EVLIST_ENABLED_MSG);
2377                 }
2378
2379                 step += 1;
2380                 pos = step / 2;
2381
2382                 if (pos < eet->times_cnt) {
2383                         int *times = (int *)eet->times; /* Accessing 'times' as array of int */
2384                         int ms = times[step] - times[step - 1];
2385
2386                         eet->times_step = step;
2387                         return event_enable_timer__set_timer(eet, ms);
2388                 }
2389         }
2390
2391         return 0;
2392 }
2393
2394 void event_enable_timer__exit(struct event_enable_timer **ep)
2395 {
2396         if (!ep || !*ep)
2397                 return;
2398         free((*ep)->times);
2399         zfree(ep);
2400 }
2401
2402 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
2403 {
2404         struct evsel *evsel;
2405
2406         evlist__for_each_entry(evlist, evsel) {
2407                 if (evsel->core.idx == idx)
2408                         return evsel;
2409         }
2410         return NULL;
2411 }
2412
2413 int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf)
2414 {
2415         struct evsel *evsel;
2416         int printed = 0;
2417
2418         evlist__for_each_entry(evlist, evsel) {
2419                 if (evsel__is_dummy_event(evsel))
2420                         continue;
2421                 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) {
2422                         printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel));
2423                 } else {
2424                         printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : "");
2425                         break;
2426                 }
2427         }
2428
2429         return printed;
2430 }
2431
2432 void evlist__check_mem_load_aux(struct evlist *evlist)
2433 {
2434         struct evsel *leader, *evsel, *pos;
2435
2436         /*
2437          * For some platforms, the 'mem-loads' event is required to use
2438          * together with 'mem-loads-aux' within a group and 'mem-loads-aux'
2439          * must be the group leader. Now we disable this group before reporting
2440          * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry
2441          * any valid memory load information.
2442          */
2443         evlist__for_each_entry(evlist, evsel) {
2444                 leader = evsel__leader(evsel);
2445                 if (leader == evsel)
2446                         continue;
2447
2448                 if (leader->name && strstr(leader->name, "mem-loads-aux")) {
2449                         for_each_group_evsel(pos, leader) {
2450                                 evsel__set_leader(pos, pos);
2451                                 pos->core.nr_members = 0;
2452                         }
2453                 }
2454         }
2455 }