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/pmu.h" 32 #include "util/sample.h" 33 #include "util/bpf-filter.h" 34 #include "util/stat.h" 35 #include "util/util.h" 36 #include <signal.h> 37 #include <unistd.h> 38 #include <sched.h> 39 #include <stdlib.h> 40 41 #include "parse-events.h" 42 #include <subcmd/parse-options.h> 43 44 #include <fcntl.h> 45 #include <sys/ioctl.h> 46 #include <sys/mman.h> 47 #include <sys/prctl.h> 48 #include <sys/timerfd.h> 49 50 #include <linux/bitops.h> 51 #include <linux/hash.h> 52 #include <linux/log2.h> 53 #include <linux/err.h> 54 #include <linux/string.h> 55 #include <linux/time64.h> 56 #include <linux/zalloc.h> 57 #include <perf/evlist.h> 58 #include <perf/evsel.h> 59 #include <perf/cpumap.h> 60 #include <perf/mmap.h> 61 62 #include <internal/xyarray.h> 63 64 #ifdef LACKS_SIGQUEUE_PROTOTYPE 65 int sigqueue(pid_t pid, int sig, const union sigval value); 66 #endif 67 68 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 69 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 70 71 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus, 72 struct perf_thread_map *threads) 73 { 74 perf_evlist__init(&evlist->core); 75 perf_evlist__set_maps(&evlist->core, cpus, threads); 76 evlist->workload.pid = -1; 77 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY; 78 evlist->ctl_fd.fd = -1; 79 evlist->ctl_fd.ack = -1; 80 evlist->ctl_fd.pos = -1; 81 } 82 83 struct evlist *evlist__new(void) 84 { 85 struct evlist *evlist = zalloc(sizeof(*evlist)); 86 87 if (evlist != NULL) 88 evlist__init(evlist, NULL, NULL); 89 90 return evlist; 91 } 92 93 struct evlist *evlist__new_default(void) 94 { 95 struct evlist *evlist = evlist__new(); 96 bool can_profile_kernel; 97 int err; 98 99 if (!evlist) 100 return NULL; 101 102 can_profile_kernel = perf_event_paranoid_check(1); 103 err = parse_event(evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu"); 104 if (err) { 105 evlist__delete(evlist); 106 return NULL; 107 } 108 109 if (evlist->core.nr_entries > 1) { 110 struct evsel *evsel; 111 112 evlist__for_each_entry(evlist, evsel) 113 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false); 114 } 115 116 return evlist; 117 } 118 119 struct evlist *evlist__new_dummy(void) 120 { 121 struct evlist *evlist = evlist__new(); 122 123 if (evlist && evlist__add_dummy(evlist)) { 124 evlist__delete(evlist); 125 evlist = NULL; 126 } 127 128 return evlist; 129 } 130 131 /** 132 * evlist__set_id_pos - set the positions of event ids. 133 * @evlist: selected event list 134 * 135 * Events with compatible sample types all have the same id_pos 136 * and is_pos. For convenience, put a copy on evlist. 137 */ 138 void evlist__set_id_pos(struct evlist *evlist) 139 { 140 struct evsel *first = evlist__first(evlist); 141 142 evlist->id_pos = first->id_pos; 143 evlist->is_pos = first->is_pos; 144 } 145 146 static void evlist__update_id_pos(struct evlist *evlist) 147 { 148 struct evsel *evsel; 149 150 evlist__for_each_entry(evlist, evsel) 151 evsel__calc_id_pos(evsel); 152 153 evlist__set_id_pos(evlist); 154 } 155 156 static void evlist__purge(struct evlist *evlist) 157 { 158 struct evsel *pos, *n; 159 160 evlist__for_each_entry_safe(evlist, n, pos) { 161 list_del_init(&pos->core.node); 162 pos->evlist = NULL; 163 evsel__delete(pos); 164 } 165 166 evlist->core.nr_entries = 0; 167 } 168 169 void evlist__exit(struct evlist *evlist) 170 { 171 event_enable_timer__exit(&evlist->eet); 172 zfree(&evlist->mmap); 173 zfree(&evlist->overwrite_mmap); 174 perf_evlist__exit(&evlist->core); 175 } 176 177 void evlist__delete(struct evlist *evlist) 178 { 179 if (evlist == NULL) 180 return; 181 182 evlist__free_stats(evlist); 183 evlist__munmap(evlist); 184 evlist__close(evlist); 185 evlist__purge(evlist); 186 evlist__exit(evlist); 187 free(evlist); 188 } 189 190 void evlist__add(struct evlist *evlist, struct evsel *entry) 191 { 192 perf_evlist__add(&evlist->core, &entry->core); 193 entry->evlist = evlist; 194 entry->tracking = !entry->core.idx; 195 196 if (evlist->core.nr_entries == 1) 197 evlist__set_id_pos(evlist); 198 } 199 200 void evlist__remove(struct evlist *evlist, struct evsel *evsel) 201 { 202 evsel->evlist = NULL; 203 perf_evlist__remove(&evlist->core, &evsel->core); 204 } 205 206 void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list) 207 { 208 while (!list_empty(list)) { 209 struct evsel *evsel, *temp, *leader = NULL; 210 211 __evlist__for_each_entry_safe(list, temp, evsel) { 212 list_del_init(&evsel->core.node); 213 evlist__add(evlist, evsel); 214 leader = evsel; 215 break; 216 } 217 218 __evlist__for_each_entry_safe(list, temp, evsel) { 219 if (evsel__has_leader(evsel, leader)) { 220 list_del_init(&evsel->core.node); 221 evlist__add(evlist, evsel); 222 } 223 } 224 } 225 } 226 227 int __evlist__set_tracepoints_handlers(struct evlist *evlist, 228 const struct evsel_str_handler *assocs, size_t nr_assocs) 229 { 230 size_t i; 231 int err; 232 233 for (i = 0; i < nr_assocs; i++) { 234 // Adding a handler for an event not in this evlist, just ignore it. 235 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name); 236 if (evsel == NULL) 237 continue; 238 239 err = -EEXIST; 240 if (evsel->handler != NULL) 241 goto out; 242 evsel->handler = assocs[i].handler; 243 } 244 245 err = 0; 246 out: 247 return err; 248 } 249 250 static void evlist__set_leader(struct evlist *evlist) 251 { 252 perf_evlist__set_leader(&evlist->core); 253 } 254 255 static struct evsel *evlist__dummy_event(struct evlist *evlist) 256 { 257 struct perf_event_attr attr = { 258 .type = PERF_TYPE_SOFTWARE, 259 .config = PERF_COUNT_SW_DUMMY, 260 .size = sizeof(attr), /* to capture ABI version */ 261 /* Avoid frequency mode for dummy events to avoid associated timers. */ 262 .freq = 0, 263 .sample_period = 1, 264 }; 265 266 return evsel__new_idx(&attr, evlist->core.nr_entries); 267 } 268 269 int evlist__add_dummy(struct evlist *evlist) 270 { 271 struct evsel *evsel = evlist__dummy_event(evlist); 272 273 if (evsel == NULL) 274 return -ENOMEM; 275 276 evlist__add(evlist, evsel); 277 return 0; 278 } 279 280 struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide) 281 { 282 struct evsel *evsel = evlist__dummy_event(evlist); 283 284 if (!evsel) 285 return NULL; 286 287 evsel->core.attr.exclude_kernel = 1; 288 evsel->core.attr.exclude_guest = 1; 289 evsel->core.attr.exclude_hv = 1; 290 evsel->core.system_wide = system_wide; 291 evsel->no_aux_samples = true; 292 evsel->name = strdup("dummy:u"); 293 294 evlist__add(evlist, evsel); 295 return evsel; 296 } 297 298 #ifdef HAVE_LIBTRACEEVENT 299 struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide) 300 { 301 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0, 302 /*format=*/true); 303 304 if (IS_ERR(evsel)) 305 return evsel; 306 307 evsel__set_sample_bit(evsel, CPU); 308 evsel__set_sample_bit(evsel, TIME); 309 310 evsel->core.system_wide = system_wide; 311 evsel->no_aux_samples = true; 312 313 evlist__add(evlist, evsel); 314 return evsel; 315 } 316 #endif 317 318 int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs) 319 { 320 struct evsel *evsel, *n; 321 LIST_HEAD(head); 322 size_t i; 323 324 for (i = 0; i < nr_attrs; i++) { 325 evsel = evsel__new_idx(attrs + i, evlist->core.nr_entries + i); 326 if (evsel == NULL) 327 goto out_delete_partial_list; 328 list_add_tail(&evsel->core.node, &head); 329 } 330 331 evlist__splice_list_tail(evlist, &head); 332 333 return 0; 334 335 out_delete_partial_list: 336 __evlist__for_each_entry_safe(&head, n, evsel) 337 evsel__delete(evsel); 338 return -1; 339 } 340 341 int __evlist__add_default_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs) 342 { 343 size_t i; 344 345 for (i = 0; i < nr_attrs; i++) 346 event_attr_init(attrs + i); 347 348 return evlist__add_attrs(evlist, attrs, nr_attrs); 349 } 350 351 __weak int arch_evlist__add_default_attrs(struct evlist *evlist, 352 struct perf_event_attr *attrs, 353 size_t nr_attrs) 354 { 355 if (!nr_attrs) 356 return 0; 357 358 return __evlist__add_default_attrs(evlist, attrs, nr_attrs); 359 } 360 361 struct evsel *evlist__find_tracepoint_by_id(struct evlist *evlist, int id) 362 { 363 struct evsel *evsel; 364 365 evlist__for_each_entry(evlist, evsel) { 366 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT && 367 (int)evsel->core.attr.config == id) 368 return evsel; 369 } 370 371 return NULL; 372 } 373 374 struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name) 375 { 376 struct evsel *evsel; 377 378 evlist__for_each_entry(evlist, evsel) { 379 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) && 380 (strcmp(evsel->name, name) == 0)) 381 return evsel; 382 } 383 384 return NULL; 385 } 386 387 #ifdef HAVE_LIBTRACEEVENT 388 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler) 389 { 390 struct evsel *evsel = evsel__newtp(sys, name); 391 392 if (IS_ERR(evsel)) 393 return -1; 394 395 evsel->handler = handler; 396 evlist__add(evlist, evsel); 397 return 0; 398 } 399 #endif 400 401 struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity) 402 { 403 struct evlist_cpu_iterator itr = { 404 .container = evlist, 405 .evsel = NULL, 406 .cpu_map_idx = 0, 407 .evlist_cpu_map_idx = 0, 408 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus), 409 .cpu = (struct perf_cpu){ .cpu = -1}, 410 .affinity = affinity, 411 }; 412 413 if (evlist__empty(evlist)) { 414 /* Ensure the empty list doesn't iterate. */ 415 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr; 416 } else { 417 itr.evsel = evlist__first(evlist); 418 if (itr.affinity) { 419 itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0); 420 affinity__set(itr.affinity, itr.cpu.cpu); 421 itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu); 422 /* 423 * If this CPU isn't in the evsel's cpu map then advance 424 * through the list. 425 */ 426 if (itr.cpu_map_idx == -1) 427 evlist_cpu_iterator__next(&itr); 428 } 429 } 430 return itr; 431 } 432 433 void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr) 434 { 435 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) { 436 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel); 437 evlist_cpu_itr->cpu_map_idx = 438 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus, 439 evlist_cpu_itr->cpu); 440 if (evlist_cpu_itr->cpu_map_idx != -1) 441 return; 442 } 443 evlist_cpu_itr->evlist_cpu_map_idx++; 444 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) { 445 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container); 446 evlist_cpu_itr->cpu = 447 perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus, 448 evlist_cpu_itr->evlist_cpu_map_idx); 449 if (evlist_cpu_itr->affinity) 450 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu); 451 evlist_cpu_itr->cpu_map_idx = 452 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus, 453 evlist_cpu_itr->cpu); 454 /* 455 * If this CPU isn't in the evsel's cpu map then advance through 456 * the list. 457 */ 458 if (evlist_cpu_itr->cpu_map_idx == -1) 459 evlist_cpu_iterator__next(evlist_cpu_itr); 460 } 461 } 462 463 bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr) 464 { 465 return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr; 466 } 467 468 static int evsel__strcmp(struct evsel *pos, char *evsel_name) 469 { 470 if (!evsel_name) 471 return 0; 472 if (evsel__is_dummy_event(pos)) 473 return 1; 474 return !evsel__name_is(pos, evsel_name); 475 } 476 477 static int evlist__is_enabled(struct evlist *evlist) 478 { 479 struct evsel *pos; 480 481 evlist__for_each_entry(evlist, pos) { 482 if (!evsel__is_group_leader(pos) || !pos->core.fd) 483 continue; 484 /* If at least one event is enabled, evlist is enabled. */ 485 if (!pos->disabled) 486 return true; 487 } 488 return false; 489 } 490 491 static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy) 492 { 493 struct evsel *pos; 494 struct evlist_cpu_iterator evlist_cpu_itr; 495 struct affinity saved_affinity, *affinity = NULL; 496 bool has_imm = false; 497 498 // See explanation in evlist__close() 499 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 500 if (affinity__setup(&saved_affinity) < 0) 501 return; 502 affinity = &saved_affinity; 503 } 504 505 /* Disable 'immediate' events last */ 506 for (int imm = 0; imm <= 1; imm++) { 507 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) { 508 pos = evlist_cpu_itr.evsel; 509 if (evsel__strcmp(pos, evsel_name)) 510 continue; 511 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd) 512 continue; 513 if (excl_dummy && evsel__is_dummy_event(pos)) 514 continue; 515 if (pos->immediate) 516 has_imm = true; 517 if (pos->immediate != imm) 518 continue; 519 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx); 520 } 521 if (!has_imm) 522 break; 523 } 524 525 affinity__cleanup(affinity); 526 evlist__for_each_entry(evlist, pos) { 527 if (evsel__strcmp(pos, evsel_name)) 528 continue; 529 if (!evsel__is_group_leader(pos) || !pos->core.fd) 530 continue; 531 if (excl_dummy && evsel__is_dummy_event(pos)) 532 continue; 533 pos->disabled = true; 534 } 535 536 /* 537 * If we disabled only single event, we need to check 538 * the enabled state of the evlist manually. 539 */ 540 if (evsel_name) 541 evlist->enabled = evlist__is_enabled(evlist); 542 else 543 evlist->enabled = false; 544 } 545 546 void evlist__disable(struct evlist *evlist) 547 { 548 __evlist__disable(evlist, NULL, false); 549 } 550 551 void evlist__disable_non_dummy(struct evlist *evlist) 552 { 553 __evlist__disable(evlist, NULL, true); 554 } 555 556 void evlist__disable_evsel(struct evlist *evlist, char *evsel_name) 557 { 558 __evlist__disable(evlist, evsel_name, false); 559 } 560 561 static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy) 562 { 563 struct evsel *pos; 564 struct evlist_cpu_iterator evlist_cpu_itr; 565 struct affinity saved_affinity, *affinity = NULL; 566 567 // See explanation in evlist__close() 568 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 569 if (affinity__setup(&saved_affinity) < 0) 570 return; 571 affinity = &saved_affinity; 572 } 573 574 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) { 575 pos = evlist_cpu_itr.evsel; 576 if (evsel__strcmp(pos, evsel_name)) 577 continue; 578 if (!evsel__is_group_leader(pos) || !pos->core.fd) 579 continue; 580 if (excl_dummy && evsel__is_dummy_event(pos)) 581 continue; 582 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx); 583 } 584 affinity__cleanup(affinity); 585 evlist__for_each_entry(evlist, pos) { 586 if (evsel__strcmp(pos, evsel_name)) 587 continue; 588 if (!evsel__is_group_leader(pos) || !pos->core.fd) 589 continue; 590 if (excl_dummy && evsel__is_dummy_event(pos)) 591 continue; 592 pos->disabled = false; 593 } 594 595 /* 596 * Even single event sets the 'enabled' for evlist, 597 * so the toggle can work properly and toggle to 598 * 'disabled' state. 599 */ 600 evlist->enabled = true; 601 } 602 603 void evlist__enable(struct evlist *evlist) 604 { 605 __evlist__enable(evlist, NULL, false); 606 } 607 608 void evlist__enable_non_dummy(struct evlist *evlist) 609 { 610 __evlist__enable(evlist, NULL, true); 611 } 612 613 void evlist__enable_evsel(struct evlist *evlist, char *evsel_name) 614 { 615 __evlist__enable(evlist, evsel_name, false); 616 } 617 618 void evlist__toggle_enable(struct evlist *evlist) 619 { 620 (evlist->enabled ? evlist__disable : evlist__enable)(evlist); 621 } 622 623 int evlist__add_pollfd(struct evlist *evlist, int fd) 624 { 625 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default); 626 } 627 628 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask) 629 { 630 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask); 631 } 632 633 #ifdef HAVE_EVENTFD_SUPPORT 634 int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd) 635 { 636 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, 637 fdarray_flag__nonfilterable | 638 fdarray_flag__non_perf_event); 639 } 640 #endif 641 642 int evlist__poll(struct evlist *evlist, int timeout) 643 { 644 return perf_evlist__poll(&evlist->core, timeout); 645 } 646 647 struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id) 648 { 649 struct hlist_head *head; 650 struct perf_sample_id *sid; 651 int hash; 652 653 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 654 head = &evlist->core.heads[hash]; 655 656 hlist_for_each_entry(sid, head, node) 657 if (sid->id == id) 658 return sid; 659 660 return NULL; 661 } 662 663 struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id) 664 { 665 struct perf_sample_id *sid; 666 667 if (evlist->core.nr_entries == 1 || !id) 668 return evlist__first(evlist); 669 670 sid = evlist__id2sid(evlist, id); 671 if (sid) 672 return container_of(sid->evsel, struct evsel, core); 673 674 if (!evlist__sample_id_all(evlist)) 675 return evlist__first(evlist); 676 677 return NULL; 678 } 679 680 struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id) 681 { 682 struct perf_sample_id *sid; 683 684 if (!id) 685 return NULL; 686 687 sid = evlist__id2sid(evlist, id); 688 if (sid) 689 return container_of(sid->evsel, struct evsel, core); 690 691 return NULL; 692 } 693 694 static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id) 695 { 696 const __u64 *array = event->sample.array; 697 ssize_t n; 698 699 n = (event->header.size - sizeof(event->header)) >> 3; 700 701 if (event->header.type == PERF_RECORD_SAMPLE) { 702 if (evlist->id_pos >= n) 703 return -1; 704 *id = array[evlist->id_pos]; 705 } else { 706 if (evlist->is_pos > n) 707 return -1; 708 n -= evlist->is_pos; 709 *id = array[n]; 710 } 711 return 0; 712 } 713 714 struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event) 715 { 716 struct evsel *first = evlist__first(evlist); 717 struct hlist_head *head; 718 struct perf_sample_id *sid; 719 int hash; 720 u64 id; 721 722 if (evlist->core.nr_entries == 1) 723 return first; 724 725 if (!first->core.attr.sample_id_all && 726 event->header.type != PERF_RECORD_SAMPLE) 727 return first; 728 729 if (evlist__event2id(evlist, event, &id)) 730 return NULL; 731 732 /* Synthesized events have an id of zero */ 733 if (!id) 734 return first; 735 736 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 737 head = &evlist->core.heads[hash]; 738 739 hlist_for_each_entry(sid, head, node) { 740 if (sid->id == id) 741 return container_of(sid->evsel, struct evsel, core); 742 } 743 return NULL; 744 } 745 746 static int evlist__set_paused(struct evlist *evlist, bool value) 747 { 748 int i; 749 750 if (!evlist->overwrite_mmap) 751 return 0; 752 753 for (i = 0; i < evlist->core.nr_mmaps; i++) { 754 int fd = evlist->overwrite_mmap[i].core.fd; 755 int err; 756 757 if (fd < 0) 758 continue; 759 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0); 760 if (err) 761 return err; 762 } 763 return 0; 764 } 765 766 static int evlist__pause(struct evlist *evlist) 767 { 768 return evlist__set_paused(evlist, true); 769 } 770 771 static int evlist__resume(struct evlist *evlist) 772 { 773 return evlist__set_paused(evlist, false); 774 } 775 776 static void evlist__munmap_nofree(struct evlist *evlist) 777 { 778 int i; 779 780 if (evlist->mmap) 781 for (i = 0; i < evlist->core.nr_mmaps; i++) 782 perf_mmap__munmap(&evlist->mmap[i].core); 783 784 if (evlist->overwrite_mmap) 785 for (i = 0; i < evlist->core.nr_mmaps; i++) 786 perf_mmap__munmap(&evlist->overwrite_mmap[i].core); 787 } 788 789 void evlist__munmap(struct evlist *evlist) 790 { 791 evlist__munmap_nofree(evlist); 792 zfree(&evlist->mmap); 793 zfree(&evlist->overwrite_mmap); 794 } 795 796 static void perf_mmap__unmap_cb(struct perf_mmap *map) 797 { 798 struct mmap *m = container_of(map, struct mmap, core); 799 800 mmap__munmap(m); 801 } 802 803 static struct mmap *evlist__alloc_mmap(struct evlist *evlist, 804 bool overwrite) 805 { 806 int i; 807 struct mmap *map; 808 809 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap)); 810 if (!map) 811 return NULL; 812 813 for (i = 0; i < evlist->core.nr_mmaps; i++) { 814 struct perf_mmap *prev = i ? &map[i - 1].core : NULL; 815 816 /* 817 * When the perf_mmap() call is made we grab one refcount, plus 818 * one extra to let perf_mmap__consume() get the last 819 * events after all real references (perf_mmap__get()) are 820 * dropped. 821 * 822 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and 823 * thus does perf_mmap__get() on it. 824 */ 825 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb); 826 } 827 828 return map; 829 } 830 831 static void 832 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist, 833 struct perf_evsel *_evsel, 834 struct perf_mmap_param *_mp, 835 int idx) 836 { 837 struct evlist *evlist = container_of(_evlist, struct evlist, core); 838 struct mmap_params *mp = container_of(_mp, struct mmap_params, core); 839 struct evsel *evsel = container_of(_evsel, struct evsel, core); 840 841 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx); 842 } 843 844 static struct perf_mmap* 845 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx) 846 { 847 struct evlist *evlist = container_of(_evlist, struct evlist, core); 848 struct mmap *maps; 849 850 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap; 851 852 if (!maps) { 853 maps = evlist__alloc_mmap(evlist, overwrite); 854 if (!maps) 855 return NULL; 856 857 if (overwrite) { 858 evlist->overwrite_mmap = maps; 859 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY) 860 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING); 861 } else { 862 evlist->mmap = maps; 863 } 864 } 865 866 return &maps[idx].core; 867 } 868 869 static int 870 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp, 871 int output, struct perf_cpu cpu) 872 { 873 struct mmap *map = container_of(_map, struct mmap, core); 874 struct mmap_params *mp = container_of(_mp, struct mmap_params, core); 875 876 return mmap__mmap(map, mp, output, cpu); 877 } 878 879 unsigned long perf_event_mlock_kb_in_pages(void) 880 { 881 unsigned long pages; 882 int max; 883 884 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) { 885 /* 886 * Pick a once upon a time good value, i.e. things look 887 * strange since we can't read a sysctl value, but lets not 888 * die yet... 889 */ 890 max = 512; 891 } else { 892 max -= (page_size / 1024); 893 } 894 895 pages = (max * 1024) / page_size; 896 if (!is_power_of_2(pages)) 897 pages = rounddown_pow_of_two(pages); 898 899 return pages; 900 } 901 902 size_t evlist__mmap_size(unsigned long pages) 903 { 904 if (pages == UINT_MAX) 905 pages = perf_event_mlock_kb_in_pages(); 906 else if (!is_power_of_2(pages)) 907 return 0; 908 909 return (pages + 1) * page_size; 910 } 911 912 static long parse_pages_arg(const char *str, unsigned long min, 913 unsigned long max) 914 { 915 unsigned long pages, val; 916 static struct parse_tag tags[] = { 917 { .tag = 'B', .mult = 1 }, 918 { .tag = 'K', .mult = 1 << 10 }, 919 { .tag = 'M', .mult = 1 << 20 }, 920 { .tag = 'G', .mult = 1 << 30 }, 921 { .tag = 0 }, 922 }; 923 924 if (str == NULL) 925 return -EINVAL; 926 927 val = parse_tag_value(str, tags); 928 if (val != (unsigned long) -1) { 929 /* we got file size value */ 930 pages = PERF_ALIGN(val, page_size) / page_size; 931 } else { 932 /* we got pages count value */ 933 char *eptr; 934 pages = strtoul(str, &eptr, 10); 935 if (*eptr != '\0') 936 return -EINVAL; 937 } 938 939 if (pages == 0 && min == 0) { 940 /* leave number of pages at 0 */ 941 } else if (!is_power_of_2(pages)) { 942 char buf[100]; 943 944 /* round pages up to next power of 2 */ 945 pages = roundup_pow_of_two(pages); 946 if (!pages) 947 return -EINVAL; 948 949 unit_number__scnprintf(buf, sizeof(buf), pages * page_size); 950 pr_info("rounding mmap pages size to %s (%lu pages)\n", 951 buf, pages); 952 } 953 954 if (pages > max) 955 return -EINVAL; 956 957 return pages; 958 } 959 960 int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str) 961 { 962 unsigned long max = UINT_MAX; 963 long pages; 964 965 if (max > SIZE_MAX / page_size) 966 max = SIZE_MAX / page_size; 967 968 pages = parse_pages_arg(str, 1, max); 969 if (pages < 0) { 970 pr_err("Invalid argument for --mmap_pages/-m\n"); 971 return -1; 972 } 973 974 *mmap_pages = pages; 975 return 0; 976 } 977 978 int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused) 979 { 980 return __evlist__parse_mmap_pages(opt->value, str); 981 } 982 983 /** 984 * evlist__mmap_ex - Create mmaps to receive events. 985 * @evlist: list of events 986 * @pages: map length in pages 987 * @overwrite: overwrite older events? 988 * @auxtrace_pages - auxtrace map length in pages 989 * @auxtrace_overwrite - overwrite older auxtrace data? 990 * 991 * If @overwrite is %false the user needs to signal event consumption using 992 * perf_mmap__write_tail(). Using evlist__mmap_read() does this 993 * automatically. 994 * 995 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data 996 * consumption using auxtrace_mmap__write_tail(). 997 * 998 * Return: %0 on success, negative error code otherwise. 999 */ 1000 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages, 1001 unsigned int auxtrace_pages, 1002 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush, 1003 int comp_level) 1004 { 1005 /* 1006 * Delay setting mp.prot: set it before calling perf_mmap__mmap. 1007 * Its value is decided by evsel's write_backward. 1008 * So &mp should not be passed through const pointer. 1009 */ 1010 struct mmap_params mp = { 1011 .nr_cblocks = nr_cblocks, 1012 .affinity = affinity, 1013 .flush = flush, 1014 .comp_level = comp_level 1015 }; 1016 struct perf_evlist_mmap_ops ops = { 1017 .idx = perf_evlist__mmap_cb_idx, 1018 .get = perf_evlist__mmap_cb_get, 1019 .mmap = perf_evlist__mmap_cb_mmap, 1020 }; 1021 1022 evlist->core.mmap_len = evlist__mmap_size(pages); 1023 pr_debug("mmap size %zuB\n", evlist->core.mmap_len); 1024 1025 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len, 1026 auxtrace_pages, auxtrace_overwrite); 1027 1028 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core); 1029 } 1030 1031 int evlist__mmap(struct evlist *evlist, unsigned int pages) 1032 { 1033 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0); 1034 } 1035 1036 int evlist__create_maps(struct evlist *evlist, struct target *target) 1037 { 1038 bool all_threads = (target->per_thread && target->system_wide); 1039 struct perf_cpu_map *cpus; 1040 struct perf_thread_map *threads; 1041 1042 /* 1043 * If specify '-a' and '--per-thread' to perf record, perf record 1044 * will override '--per-thread'. target->per_thread = false and 1045 * target->system_wide = true. 1046 * 1047 * If specify '--per-thread' only to perf record, 1048 * target->per_thread = true and target->system_wide = false. 1049 * 1050 * So target->per_thread && target->system_wide is false. 1051 * For perf record, thread_map__new_str doesn't call 1052 * thread_map__new_all_cpus. That will keep perf record's 1053 * current behavior. 1054 * 1055 * For perf stat, it allows the case that target->per_thread and 1056 * target->system_wide are all true. It means to collect system-wide 1057 * per-thread data. thread_map__new_str will call 1058 * thread_map__new_all_cpus to enumerate all threads. 1059 */ 1060 threads = thread_map__new_str(target->pid, target->tid, target->uid, 1061 all_threads); 1062 1063 if (!threads) 1064 return -1; 1065 1066 if (target__uses_dummy_map(target)) 1067 cpus = perf_cpu_map__new_any_cpu(); 1068 else 1069 cpus = perf_cpu_map__new(target->cpu_list); 1070 1071 if (!cpus) 1072 goto out_delete_threads; 1073 1074 evlist->core.has_user_cpus = !!target->cpu_list; 1075 1076 perf_evlist__set_maps(&evlist->core, cpus, threads); 1077 1078 /* as evlist now has references, put count here */ 1079 perf_cpu_map__put(cpus); 1080 perf_thread_map__put(threads); 1081 1082 return 0; 1083 1084 out_delete_threads: 1085 perf_thread_map__put(threads); 1086 return -1; 1087 } 1088 1089 int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel, 1090 struct target *target) 1091 { 1092 struct evsel *evsel; 1093 int err = 0; 1094 1095 evlist__for_each_entry(evlist, evsel) { 1096 /* 1097 * filters only work for tracepoint event, which doesn't have cpu limit. 1098 * So evlist and evsel should always be same. 1099 */ 1100 if (evsel->filter) { 1101 err = perf_evsel__apply_filter(&evsel->core, evsel->filter); 1102 if (err) { 1103 *err_evsel = evsel; 1104 break; 1105 } 1106 } 1107 1108 /* 1109 * non-tracepoint events can have BPF filters. 1110 */ 1111 if (!list_empty(&evsel->bpf_filters)) { 1112 err = perf_bpf_filter__prepare(evsel, target); 1113 if (err) { 1114 *err_evsel = evsel; 1115 break; 1116 } 1117 } 1118 } 1119 1120 return err; 1121 } 1122 1123 int evlist__set_tp_filter(struct evlist *evlist, const char *filter) 1124 { 1125 struct evsel *evsel; 1126 int err = 0; 1127 1128 if (filter == NULL) 1129 return -1; 1130 1131 evlist__for_each_entry(evlist, evsel) { 1132 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT) 1133 continue; 1134 1135 err = evsel__set_filter(evsel, filter); 1136 if (err) 1137 break; 1138 } 1139 1140 return err; 1141 } 1142 1143 int evlist__append_tp_filter(struct evlist *evlist, const char *filter) 1144 { 1145 struct evsel *evsel; 1146 int err = 0; 1147 1148 if (filter == NULL) 1149 return -1; 1150 1151 evlist__for_each_entry(evlist, evsel) { 1152 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT) 1153 continue; 1154 1155 err = evsel__append_tp_filter(evsel, filter); 1156 if (err) 1157 break; 1158 } 1159 1160 return err; 1161 } 1162 1163 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids) 1164 { 1165 char *filter; 1166 size_t i; 1167 1168 for (i = 0; i < npids; ++i) { 1169 if (i == 0) { 1170 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0) 1171 return NULL; 1172 } else { 1173 char *tmp; 1174 1175 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0) 1176 goto out_free; 1177 1178 free(filter); 1179 filter = tmp; 1180 } 1181 } 1182 1183 return filter; 1184 out_free: 1185 free(filter); 1186 return NULL; 1187 } 1188 1189 int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids) 1190 { 1191 char *filter = asprintf__tp_filter_pids(npids, pids); 1192 int ret = evlist__set_tp_filter(evlist, filter); 1193 1194 free(filter); 1195 return ret; 1196 } 1197 1198 int evlist__set_tp_filter_pid(struct evlist *evlist, pid_t pid) 1199 { 1200 return evlist__set_tp_filter_pids(evlist, 1, &pid); 1201 } 1202 1203 int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids) 1204 { 1205 char *filter = asprintf__tp_filter_pids(npids, pids); 1206 int ret = evlist__append_tp_filter(evlist, filter); 1207 1208 free(filter); 1209 return ret; 1210 } 1211 1212 int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid) 1213 { 1214 return evlist__append_tp_filter_pids(evlist, 1, &pid); 1215 } 1216 1217 bool evlist__valid_sample_type(struct evlist *evlist) 1218 { 1219 struct evsel *pos; 1220 1221 if (evlist->core.nr_entries == 1) 1222 return true; 1223 1224 if (evlist->id_pos < 0 || evlist->is_pos < 0) 1225 return false; 1226 1227 evlist__for_each_entry(evlist, pos) { 1228 if (pos->id_pos != evlist->id_pos || 1229 pos->is_pos != evlist->is_pos) 1230 return false; 1231 } 1232 1233 return true; 1234 } 1235 1236 u64 __evlist__combined_sample_type(struct evlist *evlist) 1237 { 1238 struct evsel *evsel; 1239 1240 if (evlist->combined_sample_type) 1241 return evlist->combined_sample_type; 1242 1243 evlist__for_each_entry(evlist, evsel) 1244 evlist->combined_sample_type |= evsel->core.attr.sample_type; 1245 1246 return evlist->combined_sample_type; 1247 } 1248 1249 u64 evlist__combined_sample_type(struct evlist *evlist) 1250 { 1251 evlist->combined_sample_type = 0; 1252 return __evlist__combined_sample_type(evlist); 1253 } 1254 1255 u64 evlist__combined_branch_type(struct evlist *evlist) 1256 { 1257 struct evsel *evsel; 1258 u64 branch_type = 0; 1259 1260 evlist__for_each_entry(evlist, evsel) 1261 branch_type |= evsel->core.attr.branch_sample_type; 1262 return branch_type; 1263 } 1264 1265 bool evlist__valid_read_format(struct evlist *evlist) 1266 { 1267 struct evsel *first = evlist__first(evlist), *pos = first; 1268 u64 read_format = first->core.attr.read_format; 1269 u64 sample_type = first->core.attr.sample_type; 1270 1271 evlist__for_each_entry(evlist, pos) { 1272 if (read_format != pos->core.attr.read_format) { 1273 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n", 1274 read_format, (u64)pos->core.attr.read_format); 1275 } 1276 } 1277 1278 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */ 1279 if ((sample_type & PERF_SAMPLE_READ) && 1280 !(read_format & PERF_FORMAT_ID)) { 1281 return false; 1282 } 1283 1284 return true; 1285 } 1286 1287 u16 evlist__id_hdr_size(struct evlist *evlist) 1288 { 1289 struct evsel *first = evlist__first(evlist); 1290 1291 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0; 1292 } 1293 1294 bool evlist__valid_sample_id_all(struct evlist *evlist) 1295 { 1296 struct evsel *first = evlist__first(evlist), *pos = first; 1297 1298 evlist__for_each_entry_continue(evlist, pos) { 1299 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all) 1300 return false; 1301 } 1302 1303 return true; 1304 } 1305 1306 bool evlist__sample_id_all(struct evlist *evlist) 1307 { 1308 struct evsel *first = evlist__first(evlist); 1309 return first->core.attr.sample_id_all; 1310 } 1311 1312 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel) 1313 { 1314 evlist->selected = evsel; 1315 } 1316 1317 void evlist__close(struct evlist *evlist) 1318 { 1319 struct evsel *evsel; 1320 struct evlist_cpu_iterator evlist_cpu_itr; 1321 struct affinity affinity; 1322 1323 /* 1324 * With perf record core.user_requested_cpus is usually NULL. 1325 * Use the old method to handle this for now. 1326 */ 1327 if (!evlist->core.user_requested_cpus || 1328 cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 1329 evlist__for_each_entry_reverse(evlist, evsel) 1330 evsel__close(evsel); 1331 return; 1332 } 1333 1334 if (affinity__setup(&affinity) < 0) 1335 return; 1336 1337 evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) { 1338 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core, 1339 evlist_cpu_itr.cpu_map_idx); 1340 } 1341 1342 affinity__cleanup(&affinity); 1343 evlist__for_each_entry_reverse(evlist, evsel) { 1344 perf_evsel__free_fd(&evsel->core); 1345 perf_evsel__free_id(&evsel->core); 1346 } 1347 perf_evlist__reset_id_hash(&evlist->core); 1348 } 1349 1350 static int evlist__create_syswide_maps(struct evlist *evlist) 1351 { 1352 struct perf_cpu_map *cpus; 1353 struct perf_thread_map *threads; 1354 1355 /* 1356 * Try reading /sys/devices/system/cpu/online to get 1357 * an all cpus map. 1358 * 1359 * FIXME: -ENOMEM is the best we can do here, the cpu_map 1360 * code needs an overhaul to properly forward the 1361 * error, and we may not want to do that fallback to a 1362 * default cpu identity map :-\ 1363 */ 1364 cpus = perf_cpu_map__new_online_cpus(); 1365 if (!cpus) 1366 goto out; 1367 1368 threads = perf_thread_map__new_dummy(); 1369 if (!threads) 1370 goto out_put; 1371 1372 perf_evlist__set_maps(&evlist->core, cpus, threads); 1373 1374 perf_thread_map__put(threads); 1375 out_put: 1376 perf_cpu_map__put(cpus); 1377 out: 1378 return -ENOMEM; 1379 } 1380 1381 int evlist__open(struct evlist *evlist) 1382 { 1383 struct evsel *evsel; 1384 int err; 1385 1386 /* 1387 * Default: one fd per CPU, all threads, aka systemwide 1388 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL 1389 */ 1390 if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) { 1391 err = evlist__create_syswide_maps(evlist); 1392 if (err < 0) 1393 goto out_err; 1394 } 1395 1396 evlist__update_id_pos(evlist); 1397 1398 evlist__for_each_entry(evlist, evsel) { 1399 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads); 1400 if (err < 0) 1401 goto out_err; 1402 } 1403 1404 return 0; 1405 out_err: 1406 evlist__close(evlist); 1407 errno = -err; 1408 return err; 1409 } 1410 1411 int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[], 1412 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext)) 1413 { 1414 int child_ready_pipe[2], go_pipe[2]; 1415 char bf; 1416 1417 if (pipe(child_ready_pipe) < 0) { 1418 perror("failed to create 'ready' pipe"); 1419 return -1; 1420 } 1421 1422 if (pipe(go_pipe) < 0) { 1423 perror("failed to create 'go' pipe"); 1424 goto out_close_ready_pipe; 1425 } 1426 1427 evlist->workload.pid = fork(); 1428 if (evlist->workload.pid < 0) { 1429 perror("failed to fork"); 1430 goto out_close_pipes; 1431 } 1432 1433 if (!evlist->workload.pid) { 1434 int ret; 1435 1436 if (pipe_output) 1437 dup2(2, 1); 1438 1439 signal(SIGTERM, SIG_DFL); 1440 1441 close(child_ready_pipe[0]); 1442 close(go_pipe[1]); 1443 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 1444 1445 /* 1446 * Change the name of this process not to confuse --exclude-perf users 1447 * that sees 'perf' in the window up to the execvp() and thinks that 1448 * perf samples are not being excluded. 1449 */ 1450 prctl(PR_SET_NAME, "perf-exec"); 1451 1452 /* 1453 * Tell the parent we're ready to go 1454 */ 1455 close(child_ready_pipe[1]); 1456 1457 /* 1458 * Wait until the parent tells us to go. 1459 */ 1460 ret = read(go_pipe[0], &bf, 1); 1461 /* 1462 * The parent will ask for the execvp() to be performed by 1463 * writing exactly one byte, in workload.cork_fd, usually via 1464 * evlist__start_workload(). 1465 * 1466 * For cancelling the workload without actually running it, 1467 * the parent will just close workload.cork_fd, without writing 1468 * anything, i.e. read will return zero and we just exit() 1469 * here. 1470 */ 1471 if (ret != 1) { 1472 if (ret == -1) 1473 perror("unable to read pipe"); 1474 exit(ret); 1475 } 1476 1477 execvp(argv[0], (char **)argv); 1478 1479 if (exec_error) { 1480 union sigval val; 1481 1482 val.sival_int = errno; 1483 if (sigqueue(getppid(), SIGUSR1, val)) 1484 perror(argv[0]); 1485 } else 1486 perror(argv[0]); 1487 exit(-1); 1488 } 1489 1490 if (exec_error) { 1491 struct sigaction act = { 1492 .sa_flags = SA_SIGINFO, 1493 .sa_sigaction = exec_error, 1494 }; 1495 sigaction(SIGUSR1, &act, NULL); 1496 } 1497 1498 if (target__none(target)) { 1499 if (evlist->core.threads == NULL) { 1500 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n", 1501 __func__, __LINE__); 1502 goto out_close_pipes; 1503 } 1504 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid); 1505 } 1506 1507 close(child_ready_pipe[1]); 1508 close(go_pipe[0]); 1509 /* 1510 * wait for child to settle 1511 */ 1512 if (read(child_ready_pipe[0], &bf, 1) == -1) { 1513 perror("unable to read pipe"); 1514 goto out_close_pipes; 1515 } 1516 1517 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC); 1518 evlist->workload.cork_fd = go_pipe[1]; 1519 close(child_ready_pipe[0]); 1520 return 0; 1521 1522 out_close_pipes: 1523 close(go_pipe[0]); 1524 close(go_pipe[1]); 1525 out_close_ready_pipe: 1526 close(child_ready_pipe[0]); 1527 close(child_ready_pipe[1]); 1528 return -1; 1529 } 1530 1531 int evlist__start_workload(struct evlist *evlist) 1532 { 1533 if (evlist->workload.cork_fd > 0) { 1534 char bf = 0; 1535 int ret; 1536 /* 1537 * Remove the cork, let it rip! 1538 */ 1539 ret = write(evlist->workload.cork_fd, &bf, 1); 1540 if (ret < 0) 1541 perror("unable to write to pipe"); 1542 1543 close(evlist->workload.cork_fd); 1544 return ret; 1545 } 1546 1547 return 0; 1548 } 1549 1550 int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1551 { 1552 struct evsel *evsel = evlist__event2evsel(evlist, event); 1553 int ret; 1554 1555 if (!evsel) 1556 return -EFAULT; 1557 ret = evsel__parse_sample(evsel, event, sample); 1558 if (ret) 1559 return ret; 1560 if (perf_guest && sample->id) { 1561 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id); 1562 1563 if (sid) { 1564 sample->machine_pid = sid->machine_pid; 1565 sample->vcpu = sid->vcpu.cpu; 1566 } 1567 } 1568 return 0; 1569 } 1570 1571 int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp) 1572 { 1573 struct evsel *evsel = evlist__event2evsel(evlist, event); 1574 1575 if (!evsel) 1576 return -EFAULT; 1577 return evsel__parse_sample_timestamp(evsel, event, timestamp); 1578 } 1579 1580 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size) 1581 { 1582 int printed, value; 1583 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf)); 1584 1585 switch (err) { 1586 case EACCES: 1587 case EPERM: 1588 printed = scnprintf(buf, size, 1589 "Error:\t%s.\n" 1590 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg); 1591 1592 value = perf_event_paranoid(); 1593 1594 printed += scnprintf(buf + printed, size - printed, "\nHint:\t"); 1595 1596 if (value >= 2) { 1597 printed += scnprintf(buf + printed, size - printed, 1598 "For your workloads it needs to be <= 1\nHint:\t"); 1599 } 1600 printed += scnprintf(buf + printed, size - printed, 1601 "For system wide tracing it needs to be set to -1.\n"); 1602 1603 printed += scnprintf(buf + printed, size - printed, 1604 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n" 1605 "Hint:\tThe current value is %d.", value); 1606 break; 1607 case EINVAL: { 1608 struct evsel *first = evlist__first(evlist); 1609 int max_freq; 1610 1611 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0) 1612 goto out_default; 1613 1614 if (first->core.attr.sample_freq < (u64)max_freq) 1615 goto out_default; 1616 1617 printed = scnprintf(buf, size, 1618 "Error:\t%s.\n" 1619 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n" 1620 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.", 1621 emsg, max_freq, first->core.attr.sample_freq); 1622 break; 1623 } 1624 default: 1625 out_default: 1626 scnprintf(buf, size, "%s", emsg); 1627 break; 1628 } 1629 1630 return 0; 1631 } 1632 1633 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size) 1634 { 1635 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf)); 1636 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0; 1637 1638 switch (err) { 1639 case EPERM: 1640 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user); 1641 printed += scnprintf(buf + printed, size - printed, 1642 "Error:\t%s.\n" 1643 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n" 1644 "Hint:\tTried using %zd kB.\n", 1645 emsg, pages_max_per_user, pages_attempted); 1646 1647 if (pages_attempted >= pages_max_per_user) { 1648 printed += scnprintf(buf + printed, size - printed, 1649 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n", 1650 pages_max_per_user + pages_attempted); 1651 } 1652 1653 printed += scnprintf(buf + printed, size - printed, 1654 "Hint:\tTry using a smaller -m/--mmap-pages value."); 1655 break; 1656 default: 1657 scnprintf(buf, size, "%s", emsg); 1658 break; 1659 } 1660 1661 return 0; 1662 } 1663 1664 void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel) 1665 { 1666 struct evsel *evsel, *n; 1667 LIST_HEAD(move); 1668 1669 if (move_evsel == evlist__first(evlist)) 1670 return; 1671 1672 evlist__for_each_entry_safe(evlist, n, evsel) { 1673 if (evsel__leader(evsel) == evsel__leader(move_evsel)) 1674 list_move_tail(&evsel->core.node, &move); 1675 } 1676 1677 list_splice(&move, &evlist->core.entries); 1678 } 1679 1680 struct evsel *evlist__get_tracking_event(struct evlist *evlist) 1681 { 1682 struct evsel *evsel; 1683 1684 evlist__for_each_entry(evlist, evsel) { 1685 if (evsel->tracking) 1686 return evsel; 1687 } 1688 1689 return evlist__first(evlist); 1690 } 1691 1692 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel) 1693 { 1694 struct evsel *evsel; 1695 1696 if (tracking_evsel->tracking) 1697 return; 1698 1699 evlist__for_each_entry(evlist, evsel) { 1700 if (evsel != tracking_evsel) 1701 evsel->tracking = false; 1702 } 1703 1704 tracking_evsel->tracking = true; 1705 } 1706 1707 struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide) 1708 { 1709 struct evsel *evsel; 1710 1711 evsel = evlist__get_tracking_event(evlist); 1712 if (!evsel__is_dummy_event(evsel)) { 1713 evsel = evlist__add_aux_dummy(evlist, system_wide); 1714 if (!evsel) 1715 return NULL; 1716 1717 evlist__set_tracking_event(evlist, evsel); 1718 } else if (system_wide) { 1719 perf_evlist__go_system_wide(&evlist->core, &evsel->core); 1720 } 1721 1722 return evsel; 1723 } 1724 1725 struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str) 1726 { 1727 struct evsel *evsel; 1728 1729 evlist__for_each_entry(evlist, evsel) { 1730 if (!evsel->name) 1731 continue; 1732 if (evsel__name_is(evsel, str)) 1733 return evsel; 1734 } 1735 1736 return NULL; 1737 } 1738 1739 void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state) 1740 { 1741 enum bkw_mmap_state old_state = evlist->bkw_mmap_state; 1742 enum action { 1743 NONE, 1744 PAUSE, 1745 RESUME, 1746 } action = NONE; 1747 1748 if (!evlist->overwrite_mmap) 1749 return; 1750 1751 switch (old_state) { 1752 case BKW_MMAP_NOTREADY: { 1753 if (state != BKW_MMAP_RUNNING) 1754 goto state_err; 1755 break; 1756 } 1757 case BKW_MMAP_RUNNING: { 1758 if (state != BKW_MMAP_DATA_PENDING) 1759 goto state_err; 1760 action = PAUSE; 1761 break; 1762 } 1763 case BKW_MMAP_DATA_PENDING: { 1764 if (state != BKW_MMAP_EMPTY) 1765 goto state_err; 1766 break; 1767 } 1768 case BKW_MMAP_EMPTY: { 1769 if (state != BKW_MMAP_RUNNING) 1770 goto state_err; 1771 action = RESUME; 1772 break; 1773 } 1774 default: 1775 WARN_ONCE(1, "Shouldn't get there\n"); 1776 } 1777 1778 evlist->bkw_mmap_state = state; 1779 1780 switch (action) { 1781 case PAUSE: 1782 evlist__pause(evlist); 1783 break; 1784 case RESUME: 1785 evlist__resume(evlist); 1786 break; 1787 case NONE: 1788 default: 1789 break; 1790 } 1791 1792 state_err: 1793 return; 1794 } 1795 1796 bool evlist__exclude_kernel(struct evlist *evlist) 1797 { 1798 struct evsel *evsel; 1799 1800 evlist__for_each_entry(evlist, evsel) { 1801 if (!evsel->core.attr.exclude_kernel) 1802 return false; 1803 } 1804 1805 return true; 1806 } 1807 1808 /* 1809 * Events in data file are not collect in groups, but we still want 1810 * the group display. Set the artificial group and set the leader's 1811 * forced_leader flag to notify the display code. 1812 */ 1813 void evlist__force_leader(struct evlist *evlist) 1814 { 1815 if (evlist__nr_groups(evlist) == 0) { 1816 struct evsel *leader = evlist__first(evlist); 1817 1818 evlist__set_leader(evlist); 1819 leader->forced_leader = true; 1820 } 1821 } 1822 1823 struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close) 1824 { 1825 struct evsel *c2, *leader; 1826 bool is_open = true; 1827 1828 leader = evsel__leader(evsel); 1829 1830 pr_debug("Weak group for %s/%d failed\n", 1831 leader->name, leader->core.nr_members); 1832 1833 /* 1834 * for_each_group_member doesn't work here because it doesn't 1835 * include the first entry. 1836 */ 1837 evlist__for_each_entry(evsel_list, c2) { 1838 if (c2 == evsel) 1839 is_open = false; 1840 if (evsel__has_leader(c2, leader)) { 1841 if (is_open && close) 1842 perf_evsel__close(&c2->core); 1843 /* 1844 * We want to close all members of the group and reopen 1845 * them. Some events, like Intel topdown, require being 1846 * in a group and so keep these in the group. 1847 */ 1848 evsel__remove_from_group(c2, leader); 1849 1850 /* 1851 * Set this for all former members of the group 1852 * to indicate they get reopened. 1853 */ 1854 c2->reset_group = true; 1855 } 1856 } 1857 /* Reset the leader count if all entries were removed. */ 1858 if (leader->core.nr_members == 1) 1859 leader->core.nr_members = 0; 1860 return leader; 1861 } 1862 1863 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close) 1864 { 1865 char *s, *p; 1866 int ret = 0, fd; 1867 1868 if (strncmp(str, "fifo:", 5)) 1869 return -EINVAL; 1870 1871 str += 5; 1872 if (!*str || *str == ',') 1873 return -EINVAL; 1874 1875 s = strdup(str); 1876 if (!s) 1877 return -ENOMEM; 1878 1879 p = strchr(s, ','); 1880 if (p) 1881 *p = '\0'; 1882 1883 /* 1884 * O_RDWR avoids POLLHUPs which is necessary to allow the other 1885 * end of a FIFO to be repeatedly opened and closed. 1886 */ 1887 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC); 1888 if (fd < 0) { 1889 pr_err("Failed to open '%s'\n", s); 1890 ret = -errno; 1891 goto out_free; 1892 } 1893 *ctl_fd = fd; 1894 *ctl_fd_close = true; 1895 1896 if (p && *++p) { 1897 /* O_RDWR | O_NONBLOCK means the other end need not be open */ 1898 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC); 1899 if (fd < 0) { 1900 pr_err("Failed to open '%s'\n", p); 1901 ret = -errno; 1902 goto out_free; 1903 } 1904 *ctl_fd_ack = fd; 1905 } 1906 1907 out_free: 1908 free(s); 1909 return ret; 1910 } 1911 1912 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close) 1913 { 1914 char *comma = NULL, *endptr = NULL; 1915 1916 *ctl_fd_close = false; 1917 1918 if (strncmp(str, "fd:", 3)) 1919 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close); 1920 1921 *ctl_fd = strtoul(&str[3], &endptr, 0); 1922 if (endptr == &str[3]) 1923 return -EINVAL; 1924 1925 comma = strchr(str, ','); 1926 if (comma) { 1927 if (endptr != comma) 1928 return -EINVAL; 1929 1930 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0); 1931 if (endptr == comma + 1 || *endptr != '\0') 1932 return -EINVAL; 1933 } 1934 1935 return 0; 1936 } 1937 1938 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close) 1939 { 1940 if (*ctl_fd_close) { 1941 *ctl_fd_close = false; 1942 close(ctl_fd); 1943 if (ctl_fd_ack >= 0) 1944 close(ctl_fd_ack); 1945 } 1946 } 1947 1948 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack) 1949 { 1950 if (fd == -1) { 1951 pr_debug("Control descriptor is not initialized\n"); 1952 return 0; 1953 } 1954 1955 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, 1956 fdarray_flag__nonfilterable | 1957 fdarray_flag__non_perf_event); 1958 if (evlist->ctl_fd.pos < 0) { 1959 evlist->ctl_fd.pos = -1; 1960 pr_err("Failed to add ctl fd entry: %m\n"); 1961 return -1; 1962 } 1963 1964 evlist->ctl_fd.fd = fd; 1965 evlist->ctl_fd.ack = ack; 1966 1967 return 0; 1968 } 1969 1970 bool evlist__ctlfd_initialized(struct evlist *evlist) 1971 { 1972 return evlist->ctl_fd.pos >= 0; 1973 } 1974 1975 int evlist__finalize_ctlfd(struct evlist *evlist) 1976 { 1977 struct pollfd *entries = evlist->core.pollfd.entries; 1978 1979 if (!evlist__ctlfd_initialized(evlist)) 1980 return 0; 1981 1982 entries[evlist->ctl_fd.pos].fd = -1; 1983 entries[evlist->ctl_fd.pos].events = 0; 1984 entries[evlist->ctl_fd.pos].revents = 0; 1985 1986 evlist->ctl_fd.pos = -1; 1987 evlist->ctl_fd.ack = -1; 1988 evlist->ctl_fd.fd = -1; 1989 1990 return 0; 1991 } 1992 1993 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd, 1994 char *cmd_data, size_t data_size) 1995 { 1996 int err; 1997 char c; 1998 size_t bytes_read = 0; 1999 2000 *cmd = EVLIST_CTL_CMD_UNSUPPORTED; 2001 memset(cmd_data, 0, data_size); 2002 data_size--; 2003 2004 do { 2005 err = read(evlist->ctl_fd.fd, &c, 1); 2006 if (err > 0) { 2007 if (c == '\n' || c == '\0') 2008 break; 2009 cmd_data[bytes_read++] = c; 2010 if (bytes_read == data_size) 2011 break; 2012 continue; 2013 } else if (err == -1) { 2014 if (errno == EINTR) 2015 continue; 2016 if (errno == EAGAIN || errno == EWOULDBLOCK) 2017 err = 0; 2018 else 2019 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd); 2020 } 2021 break; 2022 } while (1); 2023 2024 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data, 2025 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0"); 2026 2027 if (bytes_read > 0) { 2028 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG, 2029 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) { 2030 *cmd = EVLIST_CTL_CMD_ENABLE; 2031 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG, 2032 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) { 2033 *cmd = EVLIST_CTL_CMD_DISABLE; 2034 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG, 2035 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) { 2036 *cmd = EVLIST_CTL_CMD_SNAPSHOT; 2037 pr_debug("is snapshot\n"); 2038 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG, 2039 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) { 2040 *cmd = EVLIST_CTL_CMD_EVLIST; 2041 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG, 2042 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) { 2043 *cmd = EVLIST_CTL_CMD_STOP; 2044 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG, 2045 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) { 2046 *cmd = EVLIST_CTL_CMD_PING; 2047 } 2048 } 2049 2050 return bytes_read ? (int)bytes_read : err; 2051 } 2052 2053 int evlist__ctlfd_ack(struct evlist *evlist) 2054 { 2055 int err; 2056 2057 if (evlist->ctl_fd.ack == -1) 2058 return 0; 2059 2060 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG, 2061 sizeof(EVLIST_CTL_CMD_ACK_TAG)); 2062 if (err == -1) 2063 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack); 2064 2065 return err; 2066 } 2067 2068 static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg) 2069 { 2070 char *data = cmd_data + cmd_size; 2071 2072 /* no argument */ 2073 if (!*data) 2074 return 0; 2075 2076 /* there's argument */ 2077 if (*data == ' ') { 2078 *arg = data + 1; 2079 return 1; 2080 } 2081 2082 /* malformed */ 2083 return -1; 2084 } 2085 2086 static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable) 2087 { 2088 struct evsel *evsel; 2089 char *name; 2090 int err; 2091 2092 err = get_cmd_arg(cmd_data, 2093 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 : 2094 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1, 2095 &name); 2096 if (err < 0) { 2097 pr_info("failed: wrong command\n"); 2098 return -1; 2099 } 2100 2101 if (err) { 2102 evsel = evlist__find_evsel_by_str(evlist, name); 2103 if (evsel) { 2104 if (enable) 2105 evlist__enable_evsel(evlist, name); 2106 else 2107 evlist__disable_evsel(evlist, name); 2108 pr_info("Event %s %s\n", evsel->name, 2109 enable ? "enabled" : "disabled"); 2110 } else { 2111 pr_info("failed: can't find '%s' event\n", name); 2112 } 2113 } else { 2114 if (enable) { 2115 evlist__enable(evlist); 2116 pr_info(EVLIST_ENABLED_MSG); 2117 } else { 2118 evlist__disable(evlist); 2119 pr_info(EVLIST_DISABLED_MSG); 2120 } 2121 } 2122 2123 return 0; 2124 } 2125 2126 static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data) 2127 { 2128 struct perf_attr_details details = { .verbose = false, }; 2129 struct evsel *evsel; 2130 char *arg; 2131 int err; 2132 2133 err = get_cmd_arg(cmd_data, 2134 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1, 2135 &arg); 2136 if (err < 0) { 2137 pr_info("failed: wrong command\n"); 2138 return -1; 2139 } 2140 2141 if (err) { 2142 if (!strcmp(arg, "-v")) { 2143 details.verbose = true; 2144 } else if (!strcmp(arg, "-g")) { 2145 details.event_group = true; 2146 } else if (!strcmp(arg, "-F")) { 2147 details.freq = true; 2148 } else { 2149 pr_info("failed: wrong command\n"); 2150 return -1; 2151 } 2152 } 2153 2154 evlist__for_each_entry(evlist, evsel) 2155 evsel__fprintf(evsel, &details, stderr); 2156 2157 return 0; 2158 } 2159 2160 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd) 2161 { 2162 int err = 0; 2163 char cmd_data[EVLIST_CTL_CMD_MAX_LEN]; 2164 int ctlfd_pos = evlist->ctl_fd.pos; 2165 struct pollfd *entries = evlist->core.pollfd.entries; 2166 2167 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents) 2168 return 0; 2169 2170 if (entries[ctlfd_pos].revents & POLLIN) { 2171 err = evlist__ctlfd_recv(evlist, cmd, cmd_data, 2172 EVLIST_CTL_CMD_MAX_LEN); 2173 if (err > 0) { 2174 switch (*cmd) { 2175 case EVLIST_CTL_CMD_ENABLE: 2176 case EVLIST_CTL_CMD_DISABLE: 2177 err = evlist__ctlfd_enable(evlist, cmd_data, 2178 *cmd == EVLIST_CTL_CMD_ENABLE); 2179 break; 2180 case EVLIST_CTL_CMD_EVLIST: 2181 err = evlist__ctlfd_list(evlist, cmd_data); 2182 break; 2183 case EVLIST_CTL_CMD_SNAPSHOT: 2184 case EVLIST_CTL_CMD_STOP: 2185 case EVLIST_CTL_CMD_PING: 2186 break; 2187 case EVLIST_CTL_CMD_ACK: 2188 case EVLIST_CTL_CMD_UNSUPPORTED: 2189 default: 2190 pr_debug("ctlfd: unsupported %d\n", *cmd); 2191 break; 2192 } 2193 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED || 2194 *cmd == EVLIST_CTL_CMD_SNAPSHOT)) 2195 evlist__ctlfd_ack(evlist); 2196 } 2197 } 2198 2199 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR)) 2200 evlist__finalize_ctlfd(evlist); 2201 else 2202 entries[ctlfd_pos].revents = 0; 2203 2204 return err; 2205 } 2206 2207 /** 2208 * struct event_enable_time - perf record -D/--delay single time range. 2209 * @start: start of time range to enable events in milliseconds 2210 * @end: end of time range to enable events in milliseconds 2211 * 2212 * N.B. this structure is also accessed as an array of int. 2213 */ 2214 struct event_enable_time { 2215 int start; 2216 int end; 2217 }; 2218 2219 static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first) 2220 { 2221 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n"; 2222 int ret, start, end, n; 2223 2224 ret = sscanf(str, fmt, &start, &end, &n); 2225 if (ret != 2 || end <= start) 2226 return -EINVAL; 2227 if (range) { 2228 range->start = start; 2229 range->end = end; 2230 } 2231 return n; 2232 } 2233 2234 static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range) 2235 { 2236 int incr = !!range; 2237 bool first = true; 2238 ssize_t ret, cnt; 2239 2240 for (cnt = 0; *str; cnt++) { 2241 ret = parse_event_enable_time(str, range, first); 2242 if (ret < 0) 2243 return ret; 2244 /* Check no overlap */ 2245 if (!first && range && range->start <= range[-1].end) 2246 return -EINVAL; 2247 str += ret; 2248 range += incr; 2249 first = false; 2250 } 2251 return cnt; 2252 } 2253 2254 /** 2255 * struct event_enable_timer - control structure for perf record -D/--delay. 2256 * @evlist: event list 2257 * @times: time ranges that events are enabled (N.B. this is also accessed as an 2258 * array of int) 2259 * @times_cnt: number of time ranges 2260 * @timerfd: timer file descriptor 2261 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray) 2262 * @times_step: current position in (int *)@times)[], 2263 * refer event_enable_timer__process() 2264 * 2265 * Note, this structure is only used when there are time ranges, not when there 2266 * is only an initial delay. 2267 */ 2268 struct event_enable_timer { 2269 struct evlist *evlist; 2270 struct event_enable_time *times; 2271 size_t times_cnt; 2272 int timerfd; 2273 int pollfd_pos; 2274 size_t times_step; 2275 }; 2276 2277 static int str_to_delay(const char *str) 2278 { 2279 char *endptr; 2280 long d; 2281 2282 d = strtol(str, &endptr, 10); 2283 if (*endptr || d > INT_MAX || d < -1) 2284 return 0; 2285 return d; 2286 } 2287 2288 int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts, 2289 const char *str, int unset) 2290 { 2291 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event; 2292 struct event_enable_timer *eet; 2293 ssize_t times_cnt; 2294 ssize_t ret; 2295 int err; 2296 2297 if (unset) 2298 return 0; 2299 2300 opts->target.initial_delay = str_to_delay(str); 2301 if (opts->target.initial_delay) 2302 return 0; 2303 2304 ret = parse_event_enable_times(str, NULL); 2305 if (ret < 0) 2306 return ret; 2307 2308 times_cnt = ret; 2309 if (times_cnt == 0) 2310 return -EINVAL; 2311 2312 eet = zalloc(sizeof(*eet)); 2313 if (!eet) 2314 return -ENOMEM; 2315 2316 eet->times = calloc(times_cnt, sizeof(*eet->times)); 2317 if (!eet->times) { 2318 err = -ENOMEM; 2319 goto free_eet; 2320 } 2321 2322 if (parse_event_enable_times(str, eet->times) != times_cnt) { 2323 err = -EINVAL; 2324 goto free_eet_times; 2325 } 2326 2327 eet->times_cnt = times_cnt; 2328 2329 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC); 2330 if (eet->timerfd == -1) { 2331 err = -errno; 2332 pr_err("timerfd_create failed: %s\n", strerror(errno)); 2333 goto free_eet_times; 2334 } 2335 2336 eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags); 2337 if (eet->pollfd_pos < 0) { 2338 err = eet->pollfd_pos; 2339 goto close_timerfd; 2340 } 2341 2342 eet->evlist = evlist; 2343 evlist->eet = eet; 2344 opts->target.initial_delay = eet->times[0].start; 2345 2346 return 0; 2347 2348 close_timerfd: 2349 close(eet->timerfd); 2350 free_eet_times: 2351 zfree(&eet->times); 2352 free_eet: 2353 free(eet); 2354 return err; 2355 } 2356 2357 static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms) 2358 { 2359 struct itimerspec its = { 2360 .it_value.tv_sec = ms / MSEC_PER_SEC, 2361 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC, 2362 }; 2363 int err = 0; 2364 2365 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) { 2366 err = -errno; 2367 pr_err("timerfd_settime failed: %s\n", strerror(errno)); 2368 } 2369 return err; 2370 } 2371 2372 int event_enable_timer__start(struct event_enable_timer *eet) 2373 { 2374 int ms; 2375 2376 if (!eet) 2377 return 0; 2378 2379 ms = eet->times[0].end - eet->times[0].start; 2380 eet->times_step = 1; 2381 2382 return event_enable_timer__set_timer(eet, ms); 2383 } 2384 2385 int event_enable_timer__process(struct event_enable_timer *eet) 2386 { 2387 struct pollfd *entries; 2388 short revents; 2389 2390 if (!eet) 2391 return 0; 2392 2393 entries = eet->evlist->core.pollfd.entries; 2394 revents = entries[eet->pollfd_pos].revents; 2395 entries[eet->pollfd_pos].revents = 0; 2396 2397 if (revents & POLLIN) { 2398 size_t step = eet->times_step; 2399 size_t pos = step / 2; 2400 2401 if (step & 1) { 2402 evlist__disable_non_dummy(eet->evlist); 2403 pr_info(EVLIST_DISABLED_MSG); 2404 if (pos >= eet->times_cnt - 1) { 2405 /* Disarm timer */ 2406 event_enable_timer__set_timer(eet, 0); 2407 return 1; /* Stop */ 2408 } 2409 } else { 2410 evlist__enable_non_dummy(eet->evlist); 2411 pr_info(EVLIST_ENABLED_MSG); 2412 } 2413 2414 step += 1; 2415 pos = step / 2; 2416 2417 if (pos < eet->times_cnt) { 2418 int *times = (int *)eet->times; /* Accessing 'times' as array of int */ 2419 int ms = times[step] - times[step - 1]; 2420 2421 eet->times_step = step; 2422 return event_enable_timer__set_timer(eet, ms); 2423 } 2424 } 2425 2426 return 0; 2427 } 2428 2429 void event_enable_timer__exit(struct event_enable_timer **ep) 2430 { 2431 if (!ep || !*ep) 2432 return; 2433 zfree(&(*ep)->times); 2434 zfree(ep); 2435 } 2436 2437 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx) 2438 { 2439 struct evsel *evsel; 2440 2441 evlist__for_each_entry(evlist, evsel) { 2442 if (evsel->core.idx == idx) 2443 return evsel; 2444 } 2445 return NULL; 2446 } 2447 2448 int evlist__scnprintf_evsels(struct evlist *evlist, size_t size, char *bf) 2449 { 2450 struct evsel *evsel; 2451 int printed = 0; 2452 2453 evlist__for_each_entry(evlist, evsel) { 2454 if (evsel__is_dummy_event(evsel)) 2455 continue; 2456 if (size > (strlen(evsel__name(evsel)) + (printed ? 2 : 1))) { 2457 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "," : "", evsel__name(evsel)); 2458 } else { 2459 printed += scnprintf(bf + printed, size - printed, "%s...", printed ? "," : ""); 2460 break; 2461 } 2462 } 2463 2464 return printed; 2465 } 2466 2467 void evlist__check_mem_load_aux(struct evlist *evlist) 2468 { 2469 struct evsel *leader, *evsel, *pos; 2470 2471 /* 2472 * For some platforms, the 'mem-loads' event is required to use 2473 * together with 'mem-loads-aux' within a group and 'mem-loads-aux' 2474 * must be the group leader. Now we disable this group before reporting 2475 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry 2476 * any valid memory load information. 2477 */ 2478 evlist__for_each_entry(evlist, evsel) { 2479 leader = evsel__leader(evsel); 2480 if (leader == evsel) 2481 continue; 2482 2483 if (leader->name && strstr(leader->name, "mem-loads-aux")) { 2484 for_each_group_evsel(pos, leader) { 2485 evsel__set_leader(pos, pos); 2486 pos->core.nr_members = 0; 2487 } 2488 } 2489 } 2490 } 2491 2492 /** 2493 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs 2494 * and warn if the user CPU list is inapplicable for the event's PMU's 2495 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a 2496 * user requested CPU and so any online CPU is applicable. Core PMUs handle 2497 * events on the CPUs in their list and otherwise the event isn't supported. 2498 * @evlist: The list of events being checked. 2499 * @cpu_list: The user provided list of CPUs. 2500 */ 2501 void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list) 2502 { 2503 struct perf_cpu_map *user_requested_cpus; 2504 struct evsel *pos; 2505 2506 if (!cpu_list) 2507 return; 2508 2509 user_requested_cpus = perf_cpu_map__new(cpu_list); 2510 if (!user_requested_cpus) 2511 return; 2512 2513 evlist__for_each_entry(evlist, pos) { 2514 struct perf_cpu_map *intersect, *to_test; 2515 const struct perf_pmu *pmu = evsel__find_pmu(pos); 2516 2517 to_test = pmu && pmu->is_core ? pmu->cpus : cpu_map__online(); 2518 intersect = perf_cpu_map__intersect(to_test, user_requested_cpus); 2519 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) { 2520 char buf[128]; 2521 2522 cpu_map__snprint(to_test, buf, sizeof(buf)); 2523 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n", 2524 cpu_list, pmu ? pmu->name : "cpu", buf, evsel__name(pos)); 2525 } 2526 perf_cpu_map__put(intersect); 2527 } 2528 perf_cpu_map__put(user_requested_cpus); 2529 } 2530 2531 void evlist__uniquify_name(struct evlist *evlist) 2532 { 2533 char *new_name, empty_attributes[2] = ":", *attributes; 2534 struct evsel *pos; 2535 2536 if (perf_pmus__num_core_pmus() == 1) 2537 return; 2538 2539 evlist__for_each_entry(evlist, pos) { 2540 if (!evsel__is_hybrid(pos)) 2541 continue; 2542 2543 if (strchr(pos->name, '/')) 2544 continue; 2545 2546 attributes = strchr(pos->name, ':'); 2547 if (attributes) 2548 *attributes = '\0'; 2549 else 2550 attributes = empty_attributes; 2551 2552 if (asprintf(&new_name, "%s/%s/%s", pos->pmu_name, pos->name, attributes + 1)) { 2553 free(pos->name); 2554 pos->name = new_name; 2555 } else { 2556 *attributes = ':'; 2557 } 2558 } 2559 } 2560