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 *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 *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  * 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 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 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         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                 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 evlist__splice_list_tail(struct evlist *evlist, struct list_head *list)
 181 {
 182         struct evsel *evsel, *temp;
 183
 184         __evlist__for_each_entry_safe(list, temp, evsel) {
 185                 list_del_init(&evsel->core.node);
 186                 evlist__add(evlist, evsel);
 187         }
 188 }
 189
 190 int __evlist__set_tracepoints_handlers(struct evlist *evlist,
 191                                        const struct evsel_str_handler *assocs, size_t nr_assocs)
 192 {
 193         size_t i;
 194         int err;
 195
 196         for (i = 0; i < nr_assocs; i++) {
 197                 // Adding a handler for an event not in this evlist, just ignore it.
 198                 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name);
 199                 if (evsel == NULL)
 200                         continue;
 201
 202                 err = -EEXIST;
 203                 if (evsel->handler != NULL)
 204                         goto out;
 205                 evsel->handler = assocs[i].handler;
 206         }
 207
 208         err = 0;
 209 out:
 210         return err;
 211 }
 212
 213 void __evlist__set_leader(struct list_head *list)
 214 {
 215         struct evsel *evsel, *leader;
 216
 217         leader = list_entry(list->next, struct evsel, core.node);
 218         evsel = list_entry(list->prev, struct evsel, core.node);
 219
 220         leader->core.nr_members = evsel->idx - leader->idx + 1;
 221
 222         __evlist__for_each_entry(list, evsel) {
 223                 evsel->leader = leader;
 224         }
 225 }
 226
 227 void evlist__set_leader(struct evlist *evlist)
 228 {
 229         if (evlist->core.nr_entries) {
 230                 evlist->nr_groups = evlist->core.nr_entries > 1 ? 1 : 0;
 231                 __evlist__set_leader(&evlist->core.entries);
 232         }
 233 }
 234
 235 int __evlist__add_default(struct evlist *evlist, bool precise)
 236 {
 237         struct evsel *evsel = evsel__new_cycles(precise);
 238
 239         if (evsel == NULL)
 240                 return -ENOMEM;
 241
 242         evlist__add(evlist, evsel);
 243         return 0;
 244 }
 245
 246 int evlist__add_dummy(struct evlist *evlist)
 247 {
 248         struct perf_event_attr attr = {
 249                 .type   = PERF_TYPE_SOFTWARE,
 250                 .config = PERF_COUNT_SW_DUMMY,
 251                 .size   = sizeof(attr), /* to capture ABI version */
 252         };
 253         struct evsel *evsel = evsel__new_idx(&attr, evlist->core.nr_entries);
 254
 255         if (evsel == NULL)
 256                 return -ENOMEM;
 257
 258         evlist__add(evlist, evsel);
 259         return 0;
 260 }
 261
 262 static int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 263 {
 264         struct evsel *evsel, *n;
 265         LIST_HEAD(head);
 266         size_t i;
 267
 268         for (i = 0; i < nr_attrs; i++) {
 269                 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i);
 270                 if (evsel == NULL)
 271                         goto out_delete_partial_list;
 272                 list_add_tail(&evsel->core.node, &head);
 273         }
 274
 275         evlist__splice_list_tail(evlist, &head);
 276
 277         return 0;
 278
 279 out_delete_partial_list:
 280         __evlist__for_each_entry_safe(&head, n, evsel)
 281                 evsel__delete(evsel);
 282         return -1;
 283 }
 284
 285 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
 286 {
 287         size_t i;
 288
 289         for (i = 0; i < nr_attrs; i++)
 290                 event_attr_init(attrs + i);
 291
 292         return evlist__add_attrs(evlist, attrs, nr_attrs);
 293 }
 294
 295 struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id)
 296 {
 297         struct evsel *evsel;
 298
 299         evlist__for_each_entry(evlist, evsel) {
 300                 if (evsel->core.attr.type   == PERF_TYPE_TRACEPOINT &&
 301                     (int)evsel->core.attr.config == id)
 302                         return evsel;
 303         }
 304
 305         return NULL;
 306 }
 307
 308 struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name)
 309 {
 310         struct evsel *evsel;
 311
 312         evlist__for_each_entry(evlist, evsel) {
 313                 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) &&
 314                     (strcmp(evsel->name, name) == 0))
 315                         return evsel;
 316         }
 317
 318         return NULL;
 319 }
 320
 321 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler)
 322 {
 323         struct evsel *evsel = evsel__newtp(sys, name);
 324
 325         if (IS_ERR(evsel))
 326                 return -1;
 327
 328         evsel->handler = handler;
 329         evlist__add(evlist, evsel);
 330         return 0;
 331 }
 332
 333 static int perf_evlist__nr_threads(struct evlist *evlist,
 334                                    struct evsel *evsel)
 335 {
 336         if (evsel->core.system_wide)
 337                 return 1;
 338         else
 339                 return perf_thread_map__nr(evlist->core.threads);
 340 }
 341
 342 void evlist__cpu_iter_start(struct evlist *evlist)
 343 {
 344         struct evsel *pos;
 345
 346         /*
 347          * Reset the per evsel cpu_iter. This is needed because
 348          * each evsel's cpumap may have a different index space,
 349          * and some operations need the index to modify
 350          * the FD xyarray (e.g. open, close)
 351          */
 352         evlist__for_each_entry(evlist, pos)
 353                 pos->cpu_iter = 0;
 354 }
 355
 356 bool evsel__cpu_iter_skip_no_inc(struct evsel *ev, int cpu)
 357 {
 358         if (ev->cpu_iter >= ev->core.cpus->nr)
 359                 return true;
 360         if (cpu >= 0 && ev->core.cpus->map[ev->cpu_iter] != cpu)
 361                 return true;
 362         return false;
 363 }
 364
 365 bool evsel__cpu_iter_skip(struct evsel *ev, int cpu)
 366 {
 367         if (!evsel__cpu_iter_skip_no_inc(ev, cpu)) {
 368                 ev->cpu_iter++;
 369                 return false;
 370         }
 371         return true;
 372 }
 373
 374 void evlist__disable(struct evlist *evlist)
 375 {
 376         struct evsel *pos;
 377         struct affinity affinity;
 378         int cpu, i, imm = 0;
 379         bool has_imm = false;
 380
 381         if (affinity__setup(&affinity) < 0)
 382                 return;
 383
 384         /* Disable 'immediate' events last */
 385         for (imm = 0; imm <= 1; imm++) {
 386                 evlist__for_each_cpu(evlist, i, cpu) {
 387                         affinity__set(&affinity, cpu);
 388
 389                         evlist__for_each_entry(evlist, pos) {
 390                                 if (evsel__cpu_iter_skip(pos, cpu))
 391                                         continue;
 392                                 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd)
 393                                         continue;
 394                                 if (pos->immediate)
 395                                         has_imm = true;
 396                                 if (pos->immediate != imm)
 397                                         continue;
 398                                 evsel__disable_cpu(pos, pos->cpu_iter - 1);
 399                         }
 400                 }
 401                 if (!has_imm)
 402                         break;
 403         }
 404
 405         affinity__cleanup(&affinity);
 406         evlist__for_each_entry(evlist, pos) {
 407                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 408                         continue;
 409                 pos->disabled = true;
 410         }
 411
 412         evlist->enabled = false;
 413 }
 414
 415 void evlist__enable(struct evlist *evlist)
 416 {
 417         struct evsel *pos;
 418         struct affinity affinity;
 419         int cpu, i;
 420
 421         if (affinity__setup(&affinity) < 0)
 422                 return;
 423
 424         evlist__for_each_cpu(evlist, i, cpu) {
 425                 affinity__set(&affinity, cpu);
 426
 427                 evlist__for_each_entry(evlist, pos) {
 428                         if (evsel__cpu_iter_skip(pos, cpu))
 429                                 continue;
 430                         if (!evsel__is_group_leader(pos) || !pos->core.fd)
 431                                 continue;
 432                         evsel__enable_cpu(pos, pos->cpu_iter - 1);
 433                 }
 434         }
 435         affinity__cleanup(&affinity);
 436         evlist__for_each_entry(evlist, pos) {
 437                 if (!evsel__is_group_leader(pos) || !pos->core.fd)
 438                         continue;
 439                 pos->disabled = false;
 440         }
 441
 442         evlist->enabled = true;
 443 }
 444
 445 void evlist__toggle_enable(struct evlist *evlist)
 446 {
 447         (evlist->enabled ? evlist__disable : evlist__enable)(evlist);
 448 }
 449
 450 static int evlist__enable_event_cpu(struct evlist *evlist, struct evsel *evsel, int cpu)
 451 {
 452         int thread;
 453         int nr_threads = perf_evlist__nr_threads(evlist, evsel);
 454
 455         if (!evsel->core.fd)
 456                 return -EINVAL;
 457
 458         for (thread = 0; thread < nr_threads; thread++) {
 459                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 460                 if (err)
 461                         return err;
 462         }
 463         return 0;
 464 }
 465
 466 static int evlist__enable_event_thread(struct evlist *evlist, struct evsel *evsel, int thread)
 467 {
 468         int cpu;
 469         int nr_cpus = perf_cpu_map__nr(evlist->core.cpus);
 470
 471         if (!evsel->core.fd)
 472                 return -EINVAL;
 473
 474         for (cpu = 0; cpu < nr_cpus; cpu++) {
 475                 int err = ioctl(FD(evsel, cpu, thread), PERF_EVENT_IOC_ENABLE, 0);
 476                 if (err)
 477                         return err;
 478         }
 479         return 0;
 480 }
 481
 482 int evlist__enable_event_idx(struct evlist *evlist, struct evsel *evsel, int idx)
 483 {
 484         bool per_cpu_mmaps = !perf_cpu_map__empty(evlist->core.cpus);
 485
 486         if (per_cpu_mmaps)
 487                 return evlist__enable_event_cpu(evlist, evsel, idx);
 488
 489         return evlist__enable_event_thread(evlist, evsel, idx);
 490 }
 491
 492 int evlist__add_pollfd(struct evlist *evlist, int fd)
 493 {
 494         return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default);
 495 }
 496
 497 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask)
 498 {
 499         return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask);
 500 }
 501
 502 int evlist__poll(struct evlist *evlist, int timeout)
 503 {
 504         return perf_evlist__poll(&evlist->core, timeout);
 505 }
 506
 507 struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id)
 508 {
 509         struct hlist_head *head;
 510         struct perf_sample_id *sid;
 511         int hash;
 512
 513         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 514         head = &evlist->core.heads[hash];
 515
 516         hlist_for_each_entry(sid, head, node)
 517                 if (sid->id == id)
 518                         return sid;
 519
 520         return NULL;
 521 }
 522
 523 struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id)
 524 {
 525         struct perf_sample_id *sid;
 526
 527         if (evlist->core.nr_entries == 1 || !id)
 528                 return evlist__first(evlist);
 529
 530         sid = evlist__id2sid(evlist, id);
 531         if (sid)
 532                 return container_of(sid->evsel, struct evsel, core);
 533
 534         if (!evlist__sample_id_all(evlist))
 535                 return evlist__first(evlist);
 536
 537         return NULL;
 538 }
 539
 540 struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id)
 541 {
 542         struct perf_sample_id *sid;
 543
 544         if (!id)
 545                 return NULL;
 546
 547         sid = evlist__id2sid(evlist, id);
 548         if (sid)
 549                 return container_of(sid->evsel, struct evsel, core);
 550
 551         return NULL;
 552 }
 553
 554 static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id)
 555 {
 556         const __u64 *array = event->sample.array;
 557         ssize_t n;
 558
 559         n = (event->header.size - sizeof(event->header)) >> 3;
 560
 561         if (event->header.type == PERF_RECORD_SAMPLE) {
 562                 if (evlist->id_pos >= n)
 563                         return -1;
 564                 *id = array[evlist->id_pos];
 565         } else {
 566                 if (evlist->is_pos > n)
 567                         return -1;
 568                 n -= evlist->is_pos;
 569                 *id = array[n];
 570         }
 571         return 0;
 572 }
 573
 574 struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event)
 575 {
 576         struct evsel *first = evlist__first(evlist);
 577         struct hlist_head *head;
 578         struct perf_sample_id *sid;
 579         int hash;
 580         u64 id;
 581
 582         if (evlist->core.nr_entries == 1)
 583                 return first;
 584
 585         if (!first->core.attr.sample_id_all &&
 586             event->header.type != PERF_RECORD_SAMPLE)
 587                 return first;
 588
 589         if (evlist__event2id(evlist, event, &id))
 590                 return NULL;
 591
 592         /* Synthesized events have an id of zero */
 593         if (!id)
 594                 return first;
 595
 596         hash = hash_64(id, PERF_EVLIST__HLIST_BITS);
 597         head = &evlist->core.heads[hash];
 598
 599         hlist_for_each_entry(sid, head, node) {
 600                 if (sid->id == id)
 601                         return container_of(sid->evsel, struct evsel, core);
 602         }
 603         return NULL;
 604 }
 605
 606 static int evlist__set_paused(struct evlist *evlist, bool value)
 607 {
 608         int i;
 609
 610         if (!evlist->overwrite_mmap)
 611                 return 0;
 612
 613         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 614                 int fd = evlist->overwrite_mmap[i].core.fd;
 615                 int err;
 616
 617                 if (fd < 0)
 618                         continue;
 619                 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0);
 620                 if (err)
 621                         return err;
 622         }
 623         return 0;
 624 }
 625
 626 static int evlist__pause(struct evlist *evlist)
 627 {
 628         return evlist__set_paused(evlist, true);
 629 }
 630
 631 static int evlist__resume(struct evlist *evlist)
 632 {
 633         return evlist__set_paused(evlist, false);
 634 }
 635
 636 static void evlist__munmap_nofree(struct evlist *evlist)
 637 {
 638         int i;
 639
 640         if (evlist->mmap)
 641                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 642                         perf_mmap__munmap(&evlist->mmap[i].core);
 643
 644         if (evlist->overwrite_mmap)
 645                 for (i = 0; i < evlist->core.nr_mmaps; i++)
 646                         perf_mmap__munmap(&evlist->overwrite_mmap[i].core);
 647 }
 648
 649 void evlist__munmap(struct evlist *evlist)
 650 {
 651         evlist__munmap_nofree(evlist);
 652         zfree(&evlist->mmap);
 653         zfree(&evlist->overwrite_mmap);
 654 }
 655
 656 static void perf_mmap__unmap_cb(struct perf_mmap *map)
 657 {
 658         struct mmap *m = container_of(map, struct mmap, core);
 659
 660         mmap__munmap(m);
 661 }
 662
 663 static struct mmap *evlist__alloc_mmap(struct evlist *evlist,
 664                                        bool overwrite)
 665 {
 666         int i;
 667         struct mmap *map;
 668
 669         map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap));
 670         if (!map)
 671                 return NULL;
 672
 673         for (i = 0; i < evlist->core.nr_mmaps; i++) {
 674                 struct perf_mmap *prev = i ? &map[i - 1].core : NULL;
 675
 676                 /*
 677                  * When the perf_mmap() call is made we grab one refcount, plus
 678                  * one extra to let perf_mmap__consume() get the last
 679                  * events after all real references (perf_mmap__get()) are
 680                  * dropped.
 681                  *
 682                  * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and
 683                  * thus does perf_mmap__get() on it.
 684                  */
 685                 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb);
 686         }
 687
 688         return map;
 689 }
 690
 691 static void
 692 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist,
 693                          struct perf_mmap_param *_mp,
 694                          int idx, bool per_cpu)
 695 {
 696         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 697         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 698
 699         auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, idx, per_cpu);
 700 }
 701
 702 static struct perf_mmap*
 703 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx)
 704 {
 705         struct evlist *evlist = container_of(_evlist, struct evlist, core);
 706         struct mmap *maps;
 707
 708         maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
 709
 710         if (!maps) {
 711                 maps = evlist__alloc_mmap(evlist, overwrite);
 712                 if (!maps)
 713                         return NULL;
 714
 715                 if (overwrite) {
 716                         evlist->overwrite_mmap = maps;
 717                         if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY)
 718                                 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING);
 719                 } else {
 720                         evlist->mmap = maps;
 721                 }
 722         }
 723
 724         return &maps[idx].core;
 725 }
 726
 727 static int
 728 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp,
 729                           int output, int cpu)
 730 {
 731         struct mmap *map = container_of(_map, struct mmap, core);
 732         struct mmap_params *mp = container_of(_mp, struct mmap_params, core);
 733
 734         return mmap__mmap(map, mp, output, cpu);
 735 }
 736
 737 unsigned long perf_event_mlock_kb_in_pages(void)
 738 {
 739         unsigned long pages;
 740         int max;
 741
 742         if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) {
 743                 /*
 744                  * Pick a once upon a time good value, i.e. things look
 745                  * strange since we can't read a sysctl value, but lets not
 746                  * die yet...
 747                  */
 748                 max = 512;
 749         } else {
 750                 max -= (page_size / 1024);
 751         }
 752
 753         pages = (max * 1024) / page_size;
 754         if (!is_power_of_2(pages))
 755                 pages = rounddown_pow_of_two(pages);
 756
 757         return pages;
 758 }
 759
 760 size_t evlist__mmap_size(unsigned long pages)
 761 {
 762         if (pages == UINT_MAX)
 763                 pages = perf_event_mlock_kb_in_pages();
 764         else if (!is_power_of_2(pages))
 765                 return 0;
 766
 767         return (pages + 1) * page_size;
 768 }
 769
 770 static long parse_pages_arg(const char *str, unsigned long min,
 771                             unsigned long max)
 772 {
 773         unsigned long pages, val;
 774         static struct parse_tag tags[] = {
 775                 { .tag  = 'B', .mult = 1       },
 776                 { .tag  = 'K', .mult = 1 << 10 },
 777                 { .tag  = 'M', .mult = 1 << 20 },
 778                 { .tag  = 'G', .mult = 1 << 30 },
 779                 { .tag  = 0 },
 780         };
 781
 782         if (str == NULL)
 783                 return -EINVAL;
 784
 785         val = parse_tag_value(str, tags);
 786         if (val != (unsigned long) -1) {
 787                 /* we got file size value */
 788                 pages = PERF_ALIGN(val, page_size) / page_size;
 789         } else {
 790                 /* we got pages count value */
 791                 char *eptr;
 792                 pages = strtoul(str, &eptr, 10);
 793                 if (*eptr != '\0')
 794                         return -EINVAL;
 795         }
 796
 797         if (pages == 0 && min == 0) {
 798                 /* leave number of pages at 0 */
 799         } else if (!is_power_of_2(pages)) {
 800                 char buf[100];
 801
 802                 /* round pages up to next power of 2 */
 803                 pages = roundup_pow_of_two(pages);
 804                 if (!pages)
 805                         return -EINVAL;
 806
 807                 unit_number__scnprintf(buf, sizeof(buf), pages * page_size);
 808                 pr_info("rounding mmap pages size to %s (%lu pages)\n",
 809                         buf, pages);
 810         }
 811
 812         if (pages > max)
 813                 return -EINVAL;
 814
 815         return pages;
 816 }
 817
 818 int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str)
 819 {
 820         unsigned long max = UINT_MAX;
 821         long pages;
 822
 823         if (max > SIZE_MAX / page_size)
 824                 max = SIZE_MAX / page_size;
 825
 826         pages = parse_pages_arg(str, 1, max);
 827         if (pages < 0) {
 828                 pr_err("Invalid argument for --mmap_pages/-m\n");
 829                 return -1;
 830         }
 831
 832         *mmap_pages = pages;
 833         return 0;
 834 }
 835
 836 int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused)
 837 {
 838         return __evlist__parse_mmap_pages(opt->value, str);
 839 }
 840
 841 /**
 842  * evlist__mmap_ex - Create mmaps to receive events.
 843  * @evlist: list of events
 844  * @pages: map length in pages
 845  * @overwrite: overwrite older events?
 846  * @auxtrace_pages - auxtrace map length in pages
 847  * @auxtrace_overwrite - overwrite older auxtrace data?
 848  *
 849  * If @overwrite is %false the user needs to signal event consumption using
 850  * perf_mmap__write_tail().  Using evlist__mmap_read() does this
 851  * automatically.
 852  *
 853  * Similarly, if @auxtrace_overwrite is %false the user needs to signal data
 854  * consumption using auxtrace_mmap__write_tail().
 855  *
 856  * Return: %0 on success, negative error code otherwise.
 857  */
 858 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages,
 859                          unsigned int auxtrace_pages,
 860                          bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush,
 861                          int comp_level)
 862 {
 863         /*
 864          * Delay setting mp.prot: set it before calling perf_mmap__mmap.
 865          * Its value is decided by evsel's write_backward.
 866          * So &mp should not be passed through const pointer.
 867          */
 868         struct mmap_params mp = {
 869                 .nr_cblocks     = nr_cblocks,
 870                 .affinity       = affinity,
 871                 .flush          = flush,
 872                 .comp_level     = comp_level
 873         };
 874         struct perf_evlist_mmap_ops ops = {
 875                 .idx  = perf_evlist__mmap_cb_idx,
 876                 .get  = perf_evlist__mmap_cb_get,
 877                 .mmap = perf_evlist__mmap_cb_mmap,
 878         };
 879
 880         evlist->core.mmap_len = evlist__mmap_size(pages);
 881         pr_debug("mmap size %zuB\n", evlist->core.mmap_len);
 882
 883         auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len,
 884                                    auxtrace_pages, auxtrace_overwrite);
 885
 886         return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core);
 887 }
 888
 889 int evlist__mmap(struct evlist *evlist, unsigned int pages)
 890 {
 891         return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0);
 892 }
 893
 894 int evlist__create_maps(struct evlist *evlist, struct target *target)
 895 {
 896         bool all_threads = (target->per_thread && target->system_wide);
 897         struct perf_cpu_map *cpus;
 898         struct perf_thread_map *threads;
 899
 900         /*
 901          * If specify '-a' and '--per-thread' to perf record, perf record
 902          * will override '--per-thread'. target->per_thread = false and
 903          * target->system_wide = true.
 904          *
 905          * If specify '--per-thread' only to perf record,
 906          * target->per_thread = true and target->system_wide = false.
 907          *
 908          * So target->per_thread && target->system_wide is false.
 909          * For perf record, thread_map__new_str doesn't call
 910          * thread_map__new_all_cpus. That will keep perf record's
 911          * current behavior.
 912          *
 913          * For perf stat, it allows the case that target->per_thread and
 914          * target->system_wide are all true. It means to collect system-wide
 915          * per-thread data. thread_map__new_str will call
 916          * thread_map__new_all_cpus to enumerate all threads.
 917          */
 918         threads = thread_map__new_str(target->pid, target->tid, target->uid,
 919                                       all_threads);
 920
 921         if (!threads)
 922                 return -1;
 923
 924         if (target__uses_dummy_map(target))
 925                 cpus = perf_cpu_map__dummy_new();
 926         else
 927                 cpus = perf_cpu_map__new(target->cpu_list);
 928
 929         if (!cpus)
 930                 goto out_delete_threads;
 931
 932         evlist->core.has_user_cpus = !!target->cpu_list;
 933
 934         perf_evlist__set_maps(&evlist->core, cpus, threads);
 935
 936         /* as evlist now has references, put count here */
 937         perf_cpu_map__put(cpus);
 938         perf_thread_map__put(threads);
 939
 940         return 0;
 941
 942 out_delete_threads:
 943         perf_thread_map__put(threads);
 944         return -1;
 945 }
 946
 947 int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 948 {
 949         struct evsel *evsel;
 950         int err = 0;
 951
 952         evlist__for_each_entry(evlist, evsel) {
 953                 if (evsel->filter == NULL)
 954                         continue;
 955
 956                 /*
 957                  * filters only work for tracepoint event, which doesn't have cpu limit.
 958                  * So evlist and evsel should always be same.
 959                  */
 960                 err = perf_evsel__apply_filter(&evsel->core, evsel->filter);
 961                 if (err) {
 962                         *err_evsel = evsel;
 963                         break;
 964                 }
 965         }
 966
 967         return err;
 968 }
 969
 970 int evlist__set_tp_filter(struct evlist *evlist, const char *filter)
 971 {
 972         struct evsel *evsel;
 973         int err = 0;
 974
 975         if (filter == NULL)
 976                 return -1;
 977
 978         evlist__for_each_entry(evlist, evsel) {
 979                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
 980                         continue;
 981
 982                 err = evsel__set_filter(evsel, filter);
 983                 if (err)
 984                         break;
 985         }
 986
 987         return err;
 988 }
 989
 990 int evlist__append_tp_filter(struct evlist *evlist, const char *filter)
 991 {
 992         struct evsel *evsel;
 993         int err = 0;
 994
 995         if (filter == NULL)
 996                 return -1;
 997
 998         evlist__for_each_entry(evlist, evsel) {
 999                 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT)
1000                         continue;
1001
1002                 err = evsel__append_tp_filter(evsel, filter);
1003                 if (err)
1004                         break;
1005         }
1006
1007         return err;
1008 }
1009
1010 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids)
1011 {
1012         char *filter;
1013         size_t i;
1014
1015         for (i = 0; i < npids; ++i) {
1016                 if (i == 0) {
1017                         if (asprintf(&filter, "common_pid != %d", pids[i]) < 0)
1018                                 return NULL;
1019                 } else {
1020                         char *tmp;
1021
1022                         if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0)
1023                                 goto out_free;
1024
1025                         free(filter);
1026                         filter = tmp;
1027                 }
1028         }
1029
1030         return filter;
1031 out_free:
1032         free(filter);
1033         return NULL;
1034 }
1035
1036 int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1037 {
1038         char *filter = asprintf__tp_filter_pids(npids, pids);
1039         int ret = evlist__set_tp_filter(evlist, filter);
1040
1041         free(filter);
1042         return ret;
1043 }
1044
1045 int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid)
1046 {
1047         return evlist__set_tp_filter_pids(evlist, 1, &pid);
1048 }
1049
1050 int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids)
1051 {
1052         char *filter = asprintf__tp_filter_pids(npids, pids);
1053         int ret = evlist__append_tp_filter(evlist, filter);
1054
1055         free(filter);
1056         return ret;
1057 }
1058
1059 int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid)
1060 {
1061         return evlist__append_tp_filter_pids(evlist, 1, &pid);
1062 }
1063
1064 bool evlist__valid_sample_type(struct evlist *evlist)
1065 {
1066         struct evsel *pos;
1067
1068         if (evlist->core.nr_entries == 1)
1069                 return true;
1070
1071         if (evlist->id_pos < 0 || evlist->is_pos < 0)
1072                 return false;
1073
1074         evlist__for_each_entry(evlist, pos) {
1075                 if (pos->id_pos != evlist->id_pos ||
1076                     pos->is_pos != evlist->is_pos)
1077                         return false;
1078         }
1079
1080         return true;
1081 }
1082
1083 u64 __evlist__combined_sample_type(struct evlist *evlist)
1084 {
1085         struct evsel *evsel;
1086
1087         if (evlist->combined_sample_type)
1088                 return evlist->combined_sample_type;
1089
1090         evlist__for_each_entry(evlist, evsel)
1091                 evlist->combined_sample_type |= evsel->core.attr.sample_type;
1092
1093         return evlist->combined_sample_type;
1094 }
1095
1096 u64 evlist__combined_sample_type(struct evlist *evlist)
1097 {
1098         evlist->combined_sample_type = 0;
1099         return __evlist__combined_sample_type(evlist);
1100 }
1101
1102 u64 evlist__combined_branch_type(struct evlist *evlist)
1103 {
1104         struct evsel *evsel;
1105         u64 branch_type = 0;
1106
1107         evlist__for_each_entry(evlist, evsel)
1108                 branch_type |= evsel->core.attr.branch_sample_type;
1109         return branch_type;
1110 }
1111
1112 bool evlist__valid_read_format(struct evlist *evlist)
1113 {
1114         struct evsel *first = evlist__first(evlist), *pos = first;
1115         u64 read_format = first->core.attr.read_format;
1116         u64 sample_type = first->core.attr.sample_type;
1117
1118         evlist__for_each_entry(evlist, pos) {
1119                 if (read_format != pos->core.attr.read_format) {
1120                         pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n",
1121                                  read_format, (u64)pos->core.attr.read_format);
1122                 }
1123         }
1124
1125         /* PERF_SAMPLE_READ imples PERF_FORMAT_ID. */
1126         if ((sample_type & PERF_SAMPLE_READ) &&
1127             !(read_format & PERF_FORMAT_ID)) {
1128                 return false;
1129         }
1130
1131         return true;
1132 }
1133
1134 u16 perf_evlist__id_hdr_size(struct evlist *evlist)
1135 {
1136         struct evsel *first = evlist__first(evlist);
1137         struct perf_sample *data;
1138         u64 sample_type;
1139         u16 size = 0;
1140
1141         if (!first->core.attr.sample_id_all)
1142                 goto out;
1143
1144         sample_type = first->core.attr.sample_type;
1145
1146         if (sample_type & PERF_SAMPLE_TID)
1147                 size += sizeof(data->tid) * 2;
1148
1149        if (sample_type & PERF_SAMPLE_TIME)
1150                 size += sizeof(data->time);
1151
1152         if (sample_type & PERF_SAMPLE_ID)
1153                 size += sizeof(data->id);
1154
1155         if (sample_type & PERF_SAMPLE_STREAM_ID)
1156                 size += sizeof(data->stream_id);
1157
1158         if (sample_type & PERF_SAMPLE_CPU)
1159                 size += sizeof(data->cpu) * 2;
1160
1161         if (sample_type & PERF_SAMPLE_IDENTIFIER)
1162                 size += sizeof(data->id);
1163 out:
1164         return size;
1165 }
1166
1167 bool evlist__valid_sample_id_all(struct evlist *evlist)
1168 {
1169         struct evsel *first = evlist__first(evlist), *pos = first;
1170
1171         evlist__for_each_entry_continue(evlist, pos) {
1172                 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all)
1173                         return false;
1174         }
1175
1176         return true;
1177 }
1178
1179 bool evlist__sample_id_all(struct evlist *evlist)
1180 {
1181         struct evsel *first = evlist__first(evlist);
1182         return first->core.attr.sample_id_all;
1183 }
1184
1185 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel)
1186 {
1187         evlist->selected = evsel;
1188 }
1189
1190 void evlist__close(struct evlist *evlist)
1191 {
1192         struct evsel *evsel;
1193         struct affinity affinity;
1194         int cpu, i;
1195
1196         /*
1197          * With perf record core.cpus is usually NULL.
1198          * Use the old method to handle this for now.
1199          */
1200         if (!evlist->core.cpus) {
1201                 evlist__for_each_entry_reverse(evlist, evsel)
1202                         evsel__close(evsel);
1203                 return;
1204         }
1205
1206         if (affinity__setup(&affinity) < 0)
1207                 return;
1208         evlist__for_each_cpu(evlist, i, cpu) {
1209                 affinity__set(&affinity, cpu);
1210
1211                 evlist__for_each_entry_reverse(evlist, evsel) {
1212                         if (evsel__cpu_iter_skip(evsel, cpu))
1213                             continue;
1214                         perf_evsel__close_cpu(&evsel->core, evsel->cpu_iter - 1);
1215                 }
1216         }
1217         affinity__cleanup(&affinity);
1218         evlist__for_each_entry_reverse(evlist, evsel) {
1219                 perf_evsel__free_fd(&evsel->core);
1220                 perf_evsel__free_id(&evsel->core);
1221         }
1222 }
1223
1224 static int evlist__create_syswide_maps(struct evlist *evlist)
1225 {
1226         struct perf_cpu_map *cpus;
1227         struct perf_thread_map *threads;
1228         int err = -ENOMEM;
1229
1230         /*
1231          * Try reading /sys/devices/system/cpu/online to get
1232          * an all cpus map.
1233          *
1234          * FIXME: -ENOMEM is the best we can do here, the cpu_map
1235          * code needs an overhaul to properly forward the
1236          * error, and we may not want to do that fallback to a
1237          * default cpu identity map :-\
1238          */
1239         cpus = perf_cpu_map__new(NULL);
1240         if (!cpus)
1241                 goto out;
1242
1243         threads = perf_thread_map__new_dummy();
1244         if (!threads)
1245                 goto out_put;
1246
1247         perf_evlist__set_maps(&evlist->core, cpus, threads);
1248
1249         perf_thread_map__put(threads);
1250 out_put:
1251         perf_cpu_map__put(cpus);
1252 out:
1253         return err;
1254 }
1255
1256 int evlist__open(struct evlist *evlist)
1257 {
1258         struct evsel *evsel;
1259         int err;
1260
1261         /*
1262          * Default: one fd per CPU, all threads, aka systemwide
1263          * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL
1264          */
1265         if (evlist->core.threads == NULL && evlist->core.cpus == NULL) {
1266                 err = evlist__create_syswide_maps(evlist);
1267                 if (err < 0)
1268                         goto out_err;
1269         }
1270
1271         evlist__update_id_pos(evlist);
1272
1273         evlist__for_each_entry(evlist, evsel) {
1274                 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads);
1275                 if (err < 0)
1276                         goto out_err;
1277         }
1278
1279         return 0;
1280 out_err:
1281         evlist__close(evlist);
1282         errno = -err;
1283         return err;
1284 }
1285
1286 int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[],
1287                              bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext))
1288 {
1289         int child_ready_pipe[2], go_pipe[2];
1290         char bf;
1291
1292         if (pipe(child_ready_pipe) < 0) {
1293                 perror("failed to create 'ready' pipe");
1294                 return -1;
1295         }
1296
1297         if (pipe(go_pipe) < 0) {
1298                 perror("failed to create 'go' pipe");
1299                 goto out_close_ready_pipe;
1300         }
1301
1302         evlist->workload.pid = fork();
1303         if (evlist->workload.pid < 0) {
1304                 perror("failed to fork");
1305                 goto out_close_pipes;
1306         }
1307
1308         if (!evlist->workload.pid) {
1309                 int ret;
1310
1311                 if (pipe_output)
1312                         dup2(2, 1);
1313
1314                 signal(SIGTERM, SIG_DFL);
1315
1316                 close(child_ready_pipe[0]);
1317                 close(go_pipe[1]);
1318                 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
1319
1320                 /*
1321                  * Tell the parent we're ready to go
1322                  */
1323                 close(child_ready_pipe[1]);
1324
1325                 /*
1326                  * Wait until the parent tells us to go.
1327                  */
1328                 ret = read(go_pipe[0], &bf, 1);
1329                 /*
1330                  * The parent will ask for the execvp() to be performed by
1331                  * writing exactly one byte, in workload.cork_fd, usually via
1332                  * evlist__start_workload().
1333                  *
1334                  * For cancelling the workload without actually running it,
1335                  * the parent will just close workload.cork_fd, without writing
1336                  * anything, i.e. read will return zero and we just exit()
1337                  * here.
1338                  */
1339                 if (ret != 1) {
1340                         if (ret == -1)
1341                                 perror("unable to read pipe");
1342                         exit(ret);
1343                 }
1344
1345                 execvp(argv[0], (char **)argv);
1346
1347                 if (exec_error) {
1348                         union sigval val;
1349
1350                         val.sival_int = errno;
1351                         if (sigqueue(getppid(), SIGUSR1, val))
1352                                 perror(argv[0]);
1353                 } else
1354                         perror(argv[0]);
1355                 exit(-1);
1356         }
1357
1358         if (exec_error) {
1359                 struct sigaction act = {
1360                         .sa_flags     = SA_SIGINFO,
1361                         .sa_sigaction = exec_error,
1362                 };
1363                 sigaction(SIGUSR1, &act, NULL);
1364         }
1365
1366         if (target__none(target)) {
1367                 if (evlist->core.threads == NULL) {
1368                         fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n",
1369                                 __func__, __LINE__);
1370                         goto out_close_pipes;
1371                 }
1372                 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid);
1373         }
1374
1375         close(child_ready_pipe[1]);
1376         close(go_pipe[0]);
1377         /*
1378          * wait for child to settle
1379          */
1380         if (read(child_ready_pipe[0], &bf, 1) == -1) {
1381                 perror("unable to read pipe");
1382                 goto out_close_pipes;
1383         }
1384
1385         fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC);
1386         evlist->workload.cork_fd = go_pipe[1];
1387         close(child_ready_pipe[0]);
1388         return 0;
1389
1390 out_close_pipes:
1391         close(go_pipe[0]);
1392         close(go_pipe[1]);
1393 out_close_ready_pipe:
1394         close(child_ready_pipe[0]);
1395         close(child_ready_pipe[1]);
1396         return -1;
1397 }
1398
1399 int evlist__start_workload(struct evlist *evlist)
1400 {
1401         if (evlist->workload.cork_fd > 0) {
1402                 char bf = 0;
1403                 int ret;
1404                 /*
1405                  * Remove the cork, let it rip!
1406                  */
1407                 ret = write(evlist->workload.cork_fd, &bf, 1);
1408                 if (ret < 0)
1409                         perror("unable to write to pipe");
1410
1411                 close(evlist->workload.cork_fd);
1412                 return ret;
1413         }
1414
1415         return 0;
1416 }
1417
1418 int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1419 {
1420         struct evsel *evsel = evlist__event2evsel(evlist, event);
1421
1422         if (!evsel)
1423                 return -EFAULT;
1424         return evsel__parse_sample(evsel, event, sample);
1425 }
1426
1427 int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp)
1428 {
1429         struct evsel *evsel = evlist__event2evsel(evlist, event);
1430
1431         if (!evsel)
1432                 return -EFAULT;
1433         return evsel__parse_sample_timestamp(evsel, event, timestamp);
1434 }
1435
1436 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size)
1437 {
1438         int printed, value;
1439         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1440
1441         switch (err) {
1442         case EACCES:
1443         case EPERM:
1444                 printed = scnprintf(buf, size,
1445                                     "Error:\t%s.\n"
1446                                     "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg);
1447
1448                 value = perf_event_paranoid();
1449
1450                 printed += scnprintf(buf + printed, size - printed, "\nHint:\t");
1451
1452                 if (value >= 2) {
1453                         printed += scnprintf(buf + printed, size - printed,
1454                                              "For your workloads it needs to be <= 1\nHint:\t");
1455                 }
1456                 printed += scnprintf(buf + printed, size - printed,
1457                                      "For system wide tracing it needs to be set to -1.\n");
1458
1459                 printed += scnprintf(buf + printed, size - printed,
1460                                     "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n"
1461                                     "Hint:\tThe current value is %d.", value);
1462                 break;
1463         case EINVAL: {
1464                 struct evsel *first = evlist__first(evlist);
1465                 int max_freq;
1466
1467                 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0)
1468                         goto out_default;
1469
1470                 if (first->core.attr.sample_freq < (u64)max_freq)
1471                         goto out_default;
1472
1473                 printed = scnprintf(buf, size,
1474                                     "Error:\t%s.\n"
1475                                     "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n"
1476                                     "Hint:\tThe current value is %d and %" PRIu64 " is being requested.",
1477                                     emsg, max_freq, first->core.attr.sample_freq);
1478                 break;
1479         }
1480         default:
1481 out_default:
1482                 scnprintf(buf, size, "%s", emsg);
1483                 break;
1484         }
1485
1486         return 0;
1487 }
1488
1489 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size)
1490 {
1491         char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf));
1492         int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0;
1493
1494         switch (err) {
1495         case EPERM:
1496                 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user);
1497                 printed += scnprintf(buf + printed, size - printed,
1498                                      "Error:\t%s.\n"
1499                                      "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n"
1500                                      "Hint:\tTried using %zd kB.\n",
1501                                      emsg, pages_max_per_user, pages_attempted);
1502
1503                 if (pages_attempted >= pages_max_per_user) {
1504                         printed += scnprintf(buf + printed, size - printed,
1505                                              "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n",
1506                                              pages_max_per_user + pages_attempted);
1507                 }
1508
1509                 printed += scnprintf(buf + printed, size - printed,
1510                                      "Hint:\tTry using a smaller -m/--mmap-pages value.");
1511                 break;
1512         default:
1513                 scnprintf(buf, size, "%s", emsg);
1514                 break;
1515         }
1516
1517         return 0;
1518 }
1519
1520 void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel)
1521 {
1522         struct evsel *evsel, *n;
1523         LIST_HEAD(move);
1524
1525         if (move_evsel == evlist__first(evlist))
1526                 return;
1527
1528         evlist__for_each_entry_safe(evlist, n, evsel) {
1529                 if (evsel->leader == move_evsel->leader)
1530                         list_move_tail(&evsel->core.node, &move);
1531         }
1532
1533         list_splice(&move, &evlist->core.entries);
1534 }
1535
1536 struct evsel *evlist__get_tracking_event(struct evlist *evlist)
1537 {
1538         struct evsel *evsel;
1539
1540         evlist__for_each_entry(evlist, evsel) {
1541                 if (evsel->tracking)
1542                         return evsel;
1543         }
1544
1545         return evlist__first(evlist);
1546 }
1547
1548 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel)
1549 {
1550         struct evsel *evsel;
1551
1552         if (tracking_evsel->tracking)
1553                 return;
1554
1555         evlist__for_each_entry(evlist, evsel) {
1556                 if (evsel != tracking_evsel)
1557                         evsel->tracking = false;
1558         }
1559
1560         tracking_evsel->tracking = true;
1561 }
1562
1563 struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str)
1564 {
1565         struct evsel *evsel;
1566
1567         evlist__for_each_entry(evlist, evsel) {
1568                 if (!evsel->name)
1569                         continue;
1570                 if (strcmp(str, evsel->name) == 0)
1571                         return evsel;
1572         }
1573
1574         return NULL;
1575 }
1576
1577 void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state)
1578 {
1579         enum bkw_mmap_state old_state = evlist->bkw_mmap_state;
1580         enum action {
1581                 NONE,
1582                 PAUSE,
1583                 RESUME,
1584         } action = NONE;
1585
1586         if (!evlist->overwrite_mmap)
1587                 return;
1588
1589         switch (old_state) {
1590         case BKW_MMAP_NOTREADY: {
1591                 if (state != BKW_MMAP_RUNNING)
1592                         goto state_err;
1593                 break;
1594         }
1595         case BKW_MMAP_RUNNING: {
1596                 if (state != BKW_MMAP_DATA_PENDING)
1597                         goto state_err;
1598                 action = PAUSE;
1599                 break;
1600         }
1601         case BKW_MMAP_DATA_PENDING: {
1602                 if (state != BKW_MMAP_EMPTY)
1603                         goto state_err;
1604                 break;
1605         }
1606         case BKW_MMAP_EMPTY: {
1607                 if (state != BKW_MMAP_RUNNING)
1608                         goto state_err;
1609                 action = RESUME;
1610                 break;
1611         }
1612         default:
1613                 WARN_ONCE(1, "Shouldn't get there\n");
1614         }
1615
1616         evlist->bkw_mmap_state = state;
1617
1618         switch (action) {
1619         case PAUSE:
1620                 evlist__pause(evlist);
1621                 break;
1622         case RESUME:
1623                 evlist__resume(evlist);
1624                 break;
1625         case NONE:
1626         default:
1627                 break;
1628         }
1629
1630 state_err:
1631         return;
1632 }
1633
1634 bool evlist__exclude_kernel(struct evlist *evlist)
1635 {
1636         struct evsel *evsel;
1637
1638         evlist__for_each_entry(evlist, evsel) {
1639                 if (!evsel->core.attr.exclude_kernel)
1640                         return false;
1641         }
1642
1643         return true;
1644 }
1645
1646 /*
1647  * Events in data file are not collect in groups, but we still want
1648  * the group display. Set the artificial group and set the leader's
1649  * forced_leader flag to notify the display code.
1650  */
1651 void evlist__force_leader(struct evlist *evlist)
1652 {
1653         if (!evlist->nr_groups) {
1654                 struct evsel *leader = evlist__first(evlist);
1655
1656                 evlist__set_leader(evlist);
1657                 leader->forced_leader = true;
1658         }
1659 }
1660
1661 struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close)
1662 {
1663         struct evsel *c2, *leader;
1664         bool is_open = true;
1665
1666         leader = evsel->leader;
1667         pr_debug("Weak group for %s/%d failed\n",
1668                         leader->name, leader->core.nr_members);
1669
1670         /*
1671          * for_each_group_member doesn't work here because it doesn't
1672          * include the first entry.
1673          */
1674         evlist__for_each_entry(evsel_list, c2) {
1675                 if (c2 == evsel)
1676                         is_open = false;
1677                 if (c2->leader == leader) {
1678                         if (is_open && close)
1679                                 perf_evsel__close(&c2->core);
1680                         c2->leader = c2;
1681                         c2->core.nr_members = 0;
1682                         /*
1683                          * Set this for all former members of the group
1684                          * to indicate they get reopened.
1685                          */
1686                         c2->reset_group = true;
1687                 }
1688         }
1689         return leader;
1690 }
1691
1692 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1693 {
1694         char *s, *p;
1695         int ret = 0, fd;
1696
1697         if (strncmp(str, "fifo:", 5))
1698                 return -EINVAL;
1699
1700         str += 5;
1701         if (!*str || *str == ',')
1702                 return -EINVAL;
1703
1704         s = strdup(str);
1705         if (!s)
1706                 return -ENOMEM;
1707
1708         p = strchr(s, ',');
1709         if (p)
1710                 *p = '\0';
1711
1712         /*
1713          * O_RDWR avoids POLLHUPs which is necessary to allow the other
1714          * end of a FIFO to be repeatedly opened and closed.
1715          */
1716         fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1717         if (fd < 0) {
1718                 pr_err("Failed to open '%s'\n", s);
1719                 ret = -errno;
1720                 goto out_free;
1721         }
1722         *ctl_fd = fd;
1723         *ctl_fd_close = true;
1724
1725         if (p && *++p) {
1726                 /* O_RDWR | O_NONBLOCK means the other end need not be open */
1727                 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC);
1728                 if (fd < 0) {
1729                         pr_err("Failed to open '%s'\n", p);
1730                         ret = -errno;
1731                         goto out_free;
1732                 }
1733                 *ctl_fd_ack = fd;
1734         }
1735
1736 out_free:
1737         free(s);
1738         return ret;
1739 }
1740
1741 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close)
1742 {
1743         char *comma = NULL, *endptr = NULL;
1744
1745         *ctl_fd_close = false;
1746
1747         if (strncmp(str, "fd:", 3))
1748                 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close);
1749
1750         *ctl_fd = strtoul(&str[3], &endptr, 0);
1751         if (endptr == &str[3])
1752                 return -EINVAL;
1753
1754         comma = strchr(str, ',');
1755         if (comma) {
1756                 if (endptr != comma)
1757                         return -EINVAL;
1758
1759                 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
1760                 if (endptr == comma + 1 || *endptr != '\0')
1761                         return -EINVAL;
1762         }
1763
1764         return 0;
1765 }
1766
1767 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close)
1768 {
1769         if (*ctl_fd_close) {
1770                 *ctl_fd_close = false;
1771                 close(ctl_fd);
1772                 if (ctl_fd_ack >= 0)
1773                         close(ctl_fd_ack);
1774         }
1775 }
1776
1777 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack)
1778 {
1779         if (fd == -1) {
1780                 pr_debug("Control descriptor is not initialized\n");
1781                 return 0;
1782         }
1783
1784         evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN,
1785                                                      fdarray_flag__nonfilterable);
1786         if (evlist->ctl_fd.pos < 0) {
1787                 evlist->ctl_fd.pos = -1;
1788                 pr_err("Failed to add ctl fd entry: %m\n");
1789                 return -1;
1790         }
1791
1792         evlist->ctl_fd.fd = fd;
1793         evlist->ctl_fd.ack = ack;
1794
1795         return 0;
1796 }
1797
1798 bool evlist__ctlfd_initialized(struct evlist *evlist)
1799 {
1800         return evlist->ctl_fd.pos >= 0;
1801 }
1802
1803 int evlist__finalize_ctlfd(struct evlist *evlist)
1804 {
1805         struct pollfd *entries = evlist->core.pollfd.entries;
1806
1807         if (!evlist__ctlfd_initialized(evlist))
1808                 return 0;
1809
1810         entries[evlist->ctl_fd.pos].fd = -1;
1811         entries[evlist->ctl_fd.pos].events = 0;
1812         entries[evlist->ctl_fd.pos].revents = 0;
1813
1814         evlist->ctl_fd.pos = -1;
1815         evlist->ctl_fd.ack = -1;
1816         evlist->ctl_fd.fd = -1;
1817
1818         return 0;
1819 }
1820
1821 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd,
1822                               char *cmd_data, size_t data_size)
1823 {
1824         int err;
1825         char c;
1826         size_t bytes_read = 0;
1827
1828         *cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1829         memset(cmd_data, 0, data_size);
1830         data_size--;
1831
1832         do {
1833                 err = read(evlist->ctl_fd.fd, &c, 1);
1834                 if (err > 0) {
1835                         if (c == '\n' || c == '\0')
1836                                 break;
1837                         cmd_data[bytes_read++] = c;
1838                         if (bytes_read == data_size)
1839                                 break;
1840                         continue;
1841                 } else if (err == -1) {
1842                         if (errno == EINTR)
1843                                 continue;
1844                         if (errno == EAGAIN || errno == EWOULDBLOCK)
1845                                 err = 0;
1846                         else
1847                                 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd);
1848                 }
1849                 break;
1850         } while (1);
1851
1852         pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data,
1853                  bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0");
1854
1855         if (bytes_read > 0) {
1856                 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG,
1857                              (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) {
1858                         *cmd = EVLIST_CTL_CMD_ENABLE;
1859                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG,
1860                                     (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) {
1861                         *cmd = EVLIST_CTL_CMD_DISABLE;
1862                 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG,
1863                                     (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) {
1864                         *cmd = EVLIST_CTL_CMD_SNAPSHOT;
1865                         pr_debug("is snapshot\n");
1866                 }
1867         }
1868
1869         return bytes_read ? (int)bytes_read : err;
1870 }
1871
1872 int evlist__ctlfd_ack(struct evlist *evlist)
1873 {
1874         int err;
1875
1876         if (evlist->ctl_fd.ack == -1)
1877                 return 0;
1878
1879         err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG,
1880                     sizeof(EVLIST_CTL_CMD_ACK_TAG));
1881         if (err == -1)
1882                 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack);
1883
1884         return err;
1885 }
1886
1887 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd)
1888 {
1889         int err = 0;
1890         char cmd_data[EVLIST_CTL_CMD_MAX_LEN];
1891         int ctlfd_pos = evlist->ctl_fd.pos;
1892         struct pollfd *entries = evlist->core.pollfd.entries;
1893
1894         if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents)
1895                 return 0;
1896
1897         if (entries[ctlfd_pos].revents & POLLIN) {
1898                 err = evlist__ctlfd_recv(evlist, cmd, cmd_data,
1899                                          EVLIST_CTL_CMD_MAX_LEN);
1900                 if (err > 0) {
1901                         switch (*cmd) {
1902                         case EVLIST_CTL_CMD_ENABLE:
1903                                 evlist__enable(evlist);
1904                                 break;
1905                         case EVLIST_CTL_CMD_DISABLE:
1906                                 evlist__disable(evlist);
1907                                 break;
1908                         case EVLIST_CTL_CMD_SNAPSHOT:
1909                                 break;
1910                         case EVLIST_CTL_CMD_ACK:
1911                         case EVLIST_CTL_CMD_UNSUPPORTED:
1912                         default:
1913                                 pr_debug("ctlfd: unsupported %d\n", *cmd);
1914                                 break;
1915                         }
1916                         if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED ||
1917                               *cmd == EVLIST_CTL_CMD_SNAPSHOT))
1918                                 evlist__ctlfd_ack(evlist);
1919                 }
1920         }
1921
1922         if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR))
1923                 evlist__finalize_ctlfd(evlist);
1924         else
1925                 entries[ctlfd_pos].revents = 0;
1926
1927         return err;
1928 }
1929
1930 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx)
1931 {
1932         struct evsel *evsel;
1933
1934         evlist__for_each_entry(evlist, evsel) {
1935                 if (evsel->idx == idx)
1936                         return evsel;
1937         }
1938         return NULL;
1939 }