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 "util/env.h" 37 #include "util/intel-tpebs.h" 38 #include "util/strbuf.h" 39 #include <signal.h> 40 #include <unistd.h> 41 #include <sched.h> 42 #include <stdlib.h> 43 44 #include "parse-events.h" 45 #include <subcmd/parse-options.h> 46 47 #include <fcntl.h> 48 #include <sys/ioctl.h> 49 #include <sys/mman.h> 50 #include <sys/prctl.h> 51 #include <sys/timerfd.h> 52 #include <sys/wait.h> 53 54 #include <linux/bitops.h> 55 #include <linux/hash.h> 56 #include <linux/log2.h> 57 #include <linux/err.h> 58 #include <linux/string.h> 59 #include <linux/time64.h> 60 #include <linux/zalloc.h> 61 #include <perf/evlist.h> 62 #include <perf/evsel.h> 63 #include <perf/cpumap.h> 64 #include <perf/mmap.h> 65 66 #include <internal/xyarray.h> 67 68 #ifdef LACKS_SIGQUEUE_PROTOTYPE 69 int sigqueue(pid_t pid, int sig, const union sigval value); 70 #endif 71 72 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 73 #define SID(e, x, y) xyarray__entry(e->core.sample_id, x, y) 74 75 void evlist__init(struct evlist *evlist, struct perf_cpu_map *cpus, 76 struct perf_thread_map *threads) 77 { 78 perf_evlist__init(&evlist->core); 79 perf_evlist__set_maps(&evlist->core, cpus, threads); 80 evlist->workload.pid = -1; 81 evlist->bkw_mmap_state = BKW_MMAP_NOTREADY; 82 evlist->ctl_fd.fd = -1; 83 evlist->ctl_fd.ack = -1; 84 evlist->ctl_fd.pos = -1; 85 evlist->nr_br_cntr = -1; 86 } 87 88 struct evlist *evlist__new(void) 89 { 90 struct evlist *evlist = zalloc(sizeof(*evlist)); 91 92 if (evlist != NULL) 93 evlist__init(evlist, NULL, NULL); 94 95 return evlist; 96 } 97 98 struct evlist *evlist__new_default(void) 99 { 100 struct evlist *evlist = evlist__new(); 101 bool can_profile_kernel; 102 int err; 103 104 if (!evlist) 105 return NULL; 106 107 can_profile_kernel = perf_event_paranoid_check(1); 108 err = parse_event(evlist, can_profile_kernel ? "cycles:P" : "cycles:Pu"); 109 if (err) { 110 evlist__delete(evlist); 111 return NULL; 112 } 113 114 if (evlist->core.nr_entries > 1) { 115 struct evsel *evsel; 116 117 evlist__for_each_entry(evlist, evsel) 118 evsel__set_sample_id(evsel, /*can_sample_identifier=*/false); 119 } 120 121 return evlist; 122 } 123 124 struct evlist *evlist__new_dummy(void) 125 { 126 struct evlist *evlist = evlist__new(); 127 128 if (evlist && evlist__add_dummy(evlist)) { 129 evlist__delete(evlist); 130 evlist = NULL; 131 } 132 133 return evlist; 134 } 135 136 /** 137 * evlist__set_id_pos - set the positions of event ids. 138 * @evlist: selected event list 139 * 140 * Events with compatible sample types all have the same id_pos 141 * and is_pos. For convenience, put a copy on evlist. 142 */ 143 void evlist__set_id_pos(struct evlist *evlist) 144 { 145 struct evsel *first = evlist__first(evlist); 146 147 evlist->id_pos = first->id_pos; 148 evlist->is_pos = first->is_pos; 149 } 150 151 static void evlist__update_id_pos(struct evlist *evlist) 152 { 153 struct evsel *evsel; 154 155 evlist__for_each_entry(evlist, evsel) 156 evsel__calc_id_pos(evsel); 157 158 evlist__set_id_pos(evlist); 159 } 160 161 static void evlist__purge(struct evlist *evlist) 162 { 163 struct evsel *pos, *n; 164 165 evlist__for_each_entry_safe(evlist, n, pos) { 166 list_del_init(&pos->core.node); 167 pos->evlist = NULL; 168 evsel__delete(pos); 169 } 170 171 evlist->core.nr_entries = 0; 172 } 173 174 void evlist__exit(struct evlist *evlist) 175 { 176 event_enable_timer__exit(&evlist->eet); 177 zfree(&evlist->mmap); 178 zfree(&evlist->overwrite_mmap); 179 perf_evlist__exit(&evlist->core); 180 } 181 182 void evlist__delete(struct evlist *evlist) 183 { 184 if (evlist == NULL) 185 return; 186 187 evlist__free_stats(evlist); 188 evlist__munmap(evlist); 189 evlist__close(evlist); 190 evlist__purge(evlist); 191 evlist__exit(evlist); 192 free(evlist); 193 } 194 195 void evlist__add(struct evlist *evlist, struct evsel *entry) 196 { 197 perf_evlist__add(&evlist->core, &entry->core); 198 entry->evlist = evlist; 199 entry->tracking = !entry->core.idx; 200 201 if (evlist->core.nr_entries == 1) 202 evlist__set_id_pos(evlist); 203 } 204 205 void evlist__remove(struct evlist *evlist, struct evsel *evsel) 206 { 207 evsel->evlist = NULL; 208 perf_evlist__remove(&evlist->core, &evsel->core); 209 } 210 211 void evlist__splice_list_tail(struct evlist *evlist, struct list_head *list) 212 { 213 while (!list_empty(list)) { 214 struct evsel *evsel, *temp, *leader = NULL; 215 216 __evlist__for_each_entry_safe(list, temp, evsel) { 217 list_del_init(&evsel->core.node); 218 evlist__add(evlist, evsel); 219 leader = evsel; 220 break; 221 } 222 223 __evlist__for_each_entry_safe(list, temp, evsel) { 224 if (evsel__has_leader(evsel, leader)) { 225 list_del_init(&evsel->core.node); 226 evlist__add(evlist, evsel); 227 } 228 } 229 } 230 } 231 232 int __evlist__set_tracepoints_handlers(struct evlist *evlist, 233 const struct evsel_str_handler *assocs, size_t nr_assocs) 234 { 235 size_t i; 236 int err; 237 238 for (i = 0; i < nr_assocs; i++) { 239 // Adding a handler for an event not in this evlist, just ignore it. 240 struct evsel *evsel = evlist__find_tracepoint_by_name(evlist, assocs[i].name); 241 if (evsel == NULL) 242 continue; 243 244 err = -EEXIST; 245 if (evsel->handler != NULL) 246 goto out; 247 evsel->handler = assocs[i].handler; 248 } 249 250 err = 0; 251 out: 252 return err; 253 } 254 255 static void evlist__set_leader(struct evlist *evlist) 256 { 257 perf_evlist__set_leader(&evlist->core); 258 } 259 260 static struct evsel *evlist__dummy_event(struct evlist *evlist) 261 { 262 struct perf_event_attr attr = { 263 .type = PERF_TYPE_SOFTWARE, 264 .config = PERF_COUNT_SW_DUMMY, 265 .size = sizeof(attr), /* to capture ABI version */ 266 /* Avoid frequency mode for dummy events to avoid associated timers. */ 267 .freq = 0, 268 .sample_period = 1, 269 }; 270 271 return evsel__new_idx(&attr, evlist->core.nr_entries); 272 } 273 274 int evlist__add_dummy(struct evlist *evlist) 275 { 276 struct evsel *evsel = evlist__dummy_event(evlist); 277 278 if (evsel == NULL) 279 return -ENOMEM; 280 281 evlist__add(evlist, evsel); 282 return 0; 283 } 284 285 struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide) 286 { 287 struct evsel *evsel = evlist__dummy_event(evlist); 288 289 if (!evsel) 290 return NULL; 291 292 evsel->core.attr.exclude_kernel = 1; 293 evsel->core.attr.exclude_guest = 1; 294 evsel->core.attr.exclude_hv = 1; 295 evsel->core.system_wide = system_wide; 296 evsel->no_aux_samples = true; 297 evsel->name = strdup("dummy:u"); 298 299 evlist__add(evlist, evsel); 300 return evsel; 301 } 302 303 #ifdef HAVE_LIBTRACEEVENT 304 struct evsel *evlist__add_sched_switch(struct evlist *evlist, bool system_wide) 305 { 306 struct evsel *evsel = evsel__newtp_idx("sched", "sched_switch", 0, 307 /*format=*/true); 308 309 if (IS_ERR(evsel)) 310 return evsel; 311 312 evsel__set_sample_bit(evsel, CPU); 313 evsel__set_sample_bit(evsel, TIME); 314 315 evsel->core.system_wide = system_wide; 316 evsel->no_aux_samples = true; 317 318 evlist__add(evlist, evsel); 319 return evsel; 320 } 321 #endif 322 323 struct evsel *evlist__find_tracepoint_by_name(struct evlist *evlist, const char *name) 324 { 325 struct evsel *evsel; 326 327 evlist__for_each_entry(evlist, evsel) { 328 if ((evsel->core.attr.type == PERF_TYPE_TRACEPOINT) && 329 (strcmp(evsel->name, name) == 0)) 330 return evsel; 331 } 332 333 return NULL; 334 } 335 336 #ifdef HAVE_LIBTRACEEVENT 337 int evlist__add_newtp(struct evlist *evlist, const char *sys, const char *name, void *handler) 338 { 339 struct evsel *evsel = evsel__newtp(sys, name); 340 341 if (IS_ERR(evsel)) 342 return -1; 343 344 evsel->handler = handler; 345 evlist__add(evlist, evsel); 346 return 0; 347 } 348 #endif 349 350 struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity) 351 { 352 struct evlist_cpu_iterator itr = { 353 .container = evlist, 354 .evsel = NULL, 355 .cpu_map_idx = 0, 356 .evlist_cpu_map_idx = 0, 357 .evlist_cpu_map_nr = perf_cpu_map__nr(evlist->core.all_cpus), 358 .cpu = (struct perf_cpu){ .cpu = -1}, 359 .affinity = affinity, 360 }; 361 362 if (evlist__empty(evlist)) { 363 /* Ensure the empty list doesn't iterate. */ 364 itr.evlist_cpu_map_idx = itr.evlist_cpu_map_nr; 365 } else { 366 itr.evsel = evlist__first(evlist); 367 if (itr.affinity) { 368 itr.cpu = perf_cpu_map__cpu(evlist->core.all_cpus, 0); 369 affinity__set(itr.affinity, itr.cpu.cpu); 370 itr.cpu_map_idx = perf_cpu_map__idx(itr.evsel->core.cpus, itr.cpu); 371 /* 372 * If this CPU isn't in the evsel's cpu map then advance 373 * through the list. 374 */ 375 if (itr.cpu_map_idx == -1) 376 evlist_cpu_iterator__next(&itr); 377 } 378 } 379 return itr; 380 } 381 382 void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr) 383 { 384 while (evlist_cpu_itr->evsel != evlist__last(evlist_cpu_itr->container)) { 385 evlist_cpu_itr->evsel = evsel__next(evlist_cpu_itr->evsel); 386 evlist_cpu_itr->cpu_map_idx = 387 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus, 388 evlist_cpu_itr->cpu); 389 if (evlist_cpu_itr->cpu_map_idx != -1) 390 return; 391 } 392 evlist_cpu_itr->evlist_cpu_map_idx++; 393 if (evlist_cpu_itr->evlist_cpu_map_idx < evlist_cpu_itr->evlist_cpu_map_nr) { 394 evlist_cpu_itr->evsel = evlist__first(evlist_cpu_itr->container); 395 evlist_cpu_itr->cpu = 396 perf_cpu_map__cpu(evlist_cpu_itr->container->core.all_cpus, 397 evlist_cpu_itr->evlist_cpu_map_idx); 398 if (evlist_cpu_itr->affinity) 399 affinity__set(evlist_cpu_itr->affinity, evlist_cpu_itr->cpu.cpu); 400 evlist_cpu_itr->cpu_map_idx = 401 perf_cpu_map__idx(evlist_cpu_itr->evsel->core.cpus, 402 evlist_cpu_itr->cpu); 403 /* 404 * If this CPU isn't in the evsel's cpu map then advance through 405 * the list. 406 */ 407 if (evlist_cpu_itr->cpu_map_idx == -1) 408 evlist_cpu_iterator__next(evlist_cpu_itr); 409 } 410 } 411 412 bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr) 413 { 414 return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr; 415 } 416 417 static int evsel__strcmp(struct evsel *pos, char *evsel_name) 418 { 419 if (!evsel_name) 420 return 0; 421 if (evsel__is_dummy_event(pos)) 422 return 1; 423 return !evsel__name_is(pos, evsel_name); 424 } 425 426 static int evlist__is_enabled(struct evlist *evlist) 427 { 428 struct evsel *pos; 429 430 evlist__for_each_entry(evlist, pos) { 431 if (!evsel__is_group_leader(pos) || !pos->core.fd) 432 continue; 433 /* If at least one event is enabled, evlist is enabled. */ 434 if (!pos->disabled) 435 return true; 436 } 437 return false; 438 } 439 440 static void __evlist__disable(struct evlist *evlist, char *evsel_name, bool excl_dummy) 441 { 442 struct evsel *pos; 443 struct evlist_cpu_iterator evlist_cpu_itr; 444 struct affinity saved_affinity, *affinity = NULL; 445 bool has_imm = false; 446 447 // See explanation in evlist__close() 448 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 449 if (affinity__setup(&saved_affinity) < 0) 450 return; 451 affinity = &saved_affinity; 452 } 453 454 /* Disable 'immediate' events last */ 455 for (int imm = 0; imm <= 1; imm++) { 456 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) { 457 pos = evlist_cpu_itr.evsel; 458 if (evsel__strcmp(pos, evsel_name)) 459 continue; 460 if (pos->disabled || !evsel__is_group_leader(pos) || !pos->core.fd) 461 continue; 462 if (excl_dummy && evsel__is_dummy_event(pos)) 463 continue; 464 if (pos->immediate) 465 has_imm = true; 466 if (pos->immediate != imm) 467 continue; 468 evsel__disable_cpu(pos, evlist_cpu_itr.cpu_map_idx); 469 } 470 if (!has_imm) 471 break; 472 } 473 474 affinity__cleanup(affinity); 475 evlist__for_each_entry(evlist, pos) { 476 if (evsel__strcmp(pos, evsel_name)) 477 continue; 478 if (!evsel__is_group_leader(pos) || !pos->core.fd) 479 continue; 480 if (excl_dummy && evsel__is_dummy_event(pos)) 481 continue; 482 pos->disabled = true; 483 } 484 485 /* 486 * If we disabled only single event, we need to check 487 * the enabled state of the evlist manually. 488 */ 489 if (evsel_name) 490 evlist->enabled = evlist__is_enabled(evlist); 491 else 492 evlist->enabled = false; 493 } 494 495 void evlist__disable(struct evlist *evlist) 496 { 497 __evlist__disable(evlist, NULL, false); 498 } 499 500 void evlist__disable_non_dummy(struct evlist *evlist) 501 { 502 __evlist__disable(evlist, NULL, true); 503 } 504 505 void evlist__disable_evsel(struct evlist *evlist, char *evsel_name) 506 { 507 __evlist__disable(evlist, evsel_name, false); 508 } 509 510 static void __evlist__enable(struct evlist *evlist, char *evsel_name, bool excl_dummy) 511 { 512 struct evsel *pos; 513 struct evlist_cpu_iterator evlist_cpu_itr; 514 struct affinity saved_affinity, *affinity = NULL; 515 516 // See explanation in evlist__close() 517 if (!cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 518 if (affinity__setup(&saved_affinity) < 0) 519 return; 520 affinity = &saved_affinity; 521 } 522 523 evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity) { 524 pos = evlist_cpu_itr.evsel; 525 if (evsel__strcmp(pos, evsel_name)) 526 continue; 527 if (!evsel__is_group_leader(pos) || !pos->core.fd) 528 continue; 529 if (excl_dummy && evsel__is_dummy_event(pos)) 530 continue; 531 evsel__enable_cpu(pos, evlist_cpu_itr.cpu_map_idx); 532 } 533 affinity__cleanup(affinity); 534 evlist__for_each_entry(evlist, pos) { 535 if (evsel__strcmp(pos, evsel_name)) 536 continue; 537 if (!evsel__is_group_leader(pos) || !pos->core.fd) 538 continue; 539 if (excl_dummy && evsel__is_dummy_event(pos)) 540 continue; 541 pos->disabled = false; 542 } 543 544 /* 545 * Even single event sets the 'enabled' for evlist, 546 * so the toggle can work properly and toggle to 547 * 'disabled' state. 548 */ 549 evlist->enabled = true; 550 } 551 552 void evlist__enable(struct evlist *evlist) 553 { 554 __evlist__enable(evlist, NULL, false); 555 } 556 557 void evlist__enable_non_dummy(struct evlist *evlist) 558 { 559 __evlist__enable(evlist, NULL, true); 560 } 561 562 void evlist__enable_evsel(struct evlist *evlist, char *evsel_name) 563 { 564 __evlist__enable(evlist, evsel_name, false); 565 } 566 567 void evlist__toggle_enable(struct evlist *evlist) 568 { 569 (evlist->enabled ? evlist__disable : evlist__enable)(evlist); 570 } 571 572 int evlist__add_pollfd(struct evlist *evlist, int fd) 573 { 574 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, fdarray_flag__default); 575 } 576 577 int evlist__filter_pollfd(struct evlist *evlist, short revents_and_mask) 578 { 579 return perf_evlist__filter_pollfd(&evlist->core, revents_and_mask); 580 } 581 582 #ifdef HAVE_EVENTFD_SUPPORT 583 int evlist__add_wakeup_eventfd(struct evlist *evlist, int fd) 584 { 585 return perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, 586 fdarray_flag__nonfilterable | 587 fdarray_flag__non_perf_event); 588 } 589 #endif 590 591 int evlist__poll(struct evlist *evlist, int timeout) 592 { 593 return perf_evlist__poll(&evlist->core, timeout); 594 } 595 596 struct perf_sample_id *evlist__id2sid(struct evlist *evlist, u64 id) 597 { 598 struct hlist_head *head; 599 struct perf_sample_id *sid; 600 int hash; 601 602 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 603 head = &evlist->core.heads[hash]; 604 605 hlist_for_each_entry(sid, head, node) 606 if (sid->id == id) 607 return sid; 608 609 return NULL; 610 } 611 612 struct evsel *evlist__id2evsel(struct evlist *evlist, u64 id) 613 { 614 struct perf_sample_id *sid; 615 616 if (evlist->core.nr_entries == 1 || !id) 617 return evlist__first(evlist); 618 619 sid = evlist__id2sid(evlist, id); 620 if (sid) 621 return container_of(sid->evsel, struct evsel, core); 622 623 if (!evlist__sample_id_all(evlist)) 624 return evlist__first(evlist); 625 626 return NULL; 627 } 628 629 struct evsel *evlist__id2evsel_strict(struct evlist *evlist, u64 id) 630 { 631 struct perf_sample_id *sid; 632 633 if (!id) 634 return NULL; 635 636 sid = evlist__id2sid(evlist, id); 637 if (sid) 638 return container_of(sid->evsel, struct evsel, core); 639 640 return NULL; 641 } 642 643 static int evlist__event2id(struct evlist *evlist, union perf_event *event, u64 *id) 644 { 645 const __u64 *array = event->sample.array; 646 ssize_t n; 647 648 n = (event->header.size - sizeof(event->header)) >> 3; 649 650 if (event->header.type == PERF_RECORD_SAMPLE) { 651 if (evlist->id_pos >= n) 652 return -1; 653 *id = array[evlist->id_pos]; 654 } else { 655 if (evlist->is_pos > n) 656 return -1; 657 n -= evlist->is_pos; 658 *id = array[n]; 659 } 660 return 0; 661 } 662 663 struct evsel *evlist__event2evsel(struct evlist *evlist, union perf_event *event) 664 { 665 struct evsel *first = evlist__first(evlist); 666 struct hlist_head *head; 667 struct perf_sample_id *sid; 668 int hash; 669 u64 id; 670 671 if (evlist->core.nr_entries == 1) 672 return first; 673 674 if (!first->core.attr.sample_id_all && 675 event->header.type != PERF_RECORD_SAMPLE) 676 return first; 677 678 if (evlist__event2id(evlist, event, &id)) 679 return NULL; 680 681 /* Synthesized events have an id of zero */ 682 if (!id) 683 return first; 684 685 hash = hash_64(id, PERF_EVLIST__HLIST_BITS); 686 head = &evlist->core.heads[hash]; 687 688 hlist_for_each_entry(sid, head, node) { 689 if (sid->id == id) 690 return container_of(sid->evsel, struct evsel, core); 691 } 692 return NULL; 693 } 694 695 static int evlist__set_paused(struct evlist *evlist, bool value) 696 { 697 int i; 698 699 if (!evlist->overwrite_mmap) 700 return 0; 701 702 for (i = 0; i < evlist->core.nr_mmaps; i++) { 703 int fd = evlist->overwrite_mmap[i].core.fd; 704 int err; 705 706 if (fd < 0) 707 continue; 708 err = ioctl(fd, PERF_EVENT_IOC_PAUSE_OUTPUT, value ? 1 : 0); 709 if (err) 710 return err; 711 } 712 return 0; 713 } 714 715 static int evlist__pause(struct evlist *evlist) 716 { 717 return evlist__set_paused(evlist, true); 718 } 719 720 static int evlist__resume(struct evlist *evlist) 721 { 722 return evlist__set_paused(evlist, false); 723 } 724 725 static void evlist__munmap_nofree(struct evlist *evlist) 726 { 727 int i; 728 729 if (evlist->mmap) 730 for (i = 0; i < evlist->core.nr_mmaps; i++) 731 perf_mmap__munmap(&evlist->mmap[i].core); 732 733 if (evlist->overwrite_mmap) 734 for (i = 0; i < evlist->core.nr_mmaps; i++) 735 perf_mmap__munmap(&evlist->overwrite_mmap[i].core); 736 } 737 738 void evlist__munmap(struct evlist *evlist) 739 { 740 evlist__munmap_nofree(evlist); 741 zfree(&evlist->mmap); 742 zfree(&evlist->overwrite_mmap); 743 } 744 745 static void perf_mmap__unmap_cb(struct perf_mmap *map) 746 { 747 struct mmap *m = container_of(map, struct mmap, core); 748 749 mmap__munmap(m); 750 } 751 752 static struct mmap *evlist__alloc_mmap(struct evlist *evlist, 753 bool overwrite) 754 { 755 int i; 756 struct mmap *map; 757 758 map = zalloc(evlist->core.nr_mmaps * sizeof(struct mmap)); 759 if (!map) 760 return NULL; 761 762 for (i = 0; i < evlist->core.nr_mmaps; i++) { 763 struct perf_mmap *prev = i ? &map[i - 1].core : NULL; 764 765 /* 766 * When the perf_mmap() call is made we grab one refcount, plus 767 * one extra to let perf_mmap__consume() get the last 768 * events after all real references (perf_mmap__get()) are 769 * dropped. 770 * 771 * Each PERF_EVENT_IOC_SET_OUTPUT points to this mmap and 772 * thus does perf_mmap__get() on it. 773 */ 774 perf_mmap__init(&map[i].core, prev, overwrite, perf_mmap__unmap_cb); 775 } 776 777 return map; 778 } 779 780 static void 781 perf_evlist__mmap_cb_idx(struct perf_evlist *_evlist, 782 struct perf_evsel *_evsel, 783 struct perf_mmap_param *_mp, 784 int idx) 785 { 786 struct evlist *evlist = container_of(_evlist, struct evlist, core); 787 struct mmap_params *mp = container_of(_mp, struct mmap_params, core); 788 struct evsel *evsel = container_of(_evsel, struct evsel, core); 789 790 auxtrace_mmap_params__set_idx(&mp->auxtrace_mp, evlist, evsel, idx); 791 } 792 793 static struct perf_mmap* 794 perf_evlist__mmap_cb_get(struct perf_evlist *_evlist, bool overwrite, int idx) 795 { 796 struct evlist *evlist = container_of(_evlist, struct evlist, core); 797 struct mmap *maps; 798 799 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap; 800 801 if (!maps) { 802 maps = evlist__alloc_mmap(evlist, overwrite); 803 if (!maps) 804 return NULL; 805 806 if (overwrite) { 807 evlist->overwrite_mmap = maps; 808 if (evlist->bkw_mmap_state == BKW_MMAP_NOTREADY) 809 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_RUNNING); 810 } else { 811 evlist->mmap = maps; 812 } 813 } 814 815 return &maps[idx].core; 816 } 817 818 static int 819 perf_evlist__mmap_cb_mmap(struct perf_mmap *_map, struct perf_mmap_param *_mp, 820 int output, struct perf_cpu cpu) 821 { 822 struct mmap *map = container_of(_map, struct mmap, core); 823 struct mmap_params *mp = container_of(_mp, struct mmap_params, core); 824 825 return mmap__mmap(map, mp, output, cpu); 826 } 827 828 unsigned long perf_event_mlock_kb_in_pages(void) 829 { 830 unsigned long pages; 831 int max; 832 833 if (sysctl__read_int("kernel/perf_event_mlock_kb", &max) < 0) { 834 /* 835 * Pick a once upon a time good value, i.e. things look 836 * strange since we can't read a sysctl value, but lets not 837 * die yet... 838 */ 839 max = 512; 840 } else { 841 max -= (page_size / 1024); 842 } 843 844 pages = (max * 1024) / page_size; 845 if (!is_power_of_2(pages)) 846 pages = rounddown_pow_of_two(pages); 847 848 return pages; 849 } 850 851 size_t evlist__mmap_size(unsigned long pages) 852 { 853 if (pages == UINT_MAX) 854 pages = perf_event_mlock_kb_in_pages(); 855 else if (!is_power_of_2(pages)) 856 return 0; 857 858 return (pages + 1) * page_size; 859 } 860 861 static long parse_pages_arg(const char *str, unsigned long min, 862 unsigned long max) 863 { 864 unsigned long pages, val; 865 static struct parse_tag tags[] = { 866 { .tag = 'B', .mult = 1 }, 867 { .tag = 'K', .mult = 1 << 10 }, 868 { .tag = 'M', .mult = 1 << 20 }, 869 { .tag = 'G', .mult = 1 << 30 }, 870 { .tag = 0 }, 871 }; 872 873 if (str == NULL) 874 return -EINVAL; 875 876 val = parse_tag_value(str, tags); 877 if (val != (unsigned long) -1) { 878 /* we got file size value */ 879 pages = PERF_ALIGN(val, page_size) / page_size; 880 } else { 881 /* we got pages count value */ 882 char *eptr; 883 pages = strtoul(str, &eptr, 10); 884 if (*eptr != '\0') 885 return -EINVAL; 886 } 887 888 if (pages == 0 && min == 0) { 889 /* leave number of pages at 0 */ 890 } else if (!is_power_of_2(pages)) { 891 char buf[100]; 892 893 /* round pages up to next power of 2 */ 894 pages = roundup_pow_of_two(pages); 895 if (!pages) 896 return -EINVAL; 897 898 unit_number__scnprintf(buf, sizeof(buf), pages * page_size); 899 pr_info("rounding mmap pages size to %s (%lu pages)\n", 900 buf, pages); 901 } 902 903 if (pages > max) 904 return -EINVAL; 905 906 return pages; 907 } 908 909 int __evlist__parse_mmap_pages(unsigned int *mmap_pages, const char *str) 910 { 911 unsigned long max = UINT_MAX; 912 long pages; 913 914 if (max > SIZE_MAX / page_size) 915 max = SIZE_MAX / page_size; 916 917 pages = parse_pages_arg(str, 1, max); 918 if (pages < 0) { 919 pr_err("Invalid argument for --mmap_pages/-m\n"); 920 return -1; 921 } 922 923 *mmap_pages = pages; 924 return 0; 925 } 926 927 int evlist__parse_mmap_pages(const struct option *opt, const char *str, int unset __maybe_unused) 928 { 929 return __evlist__parse_mmap_pages(opt->value, str); 930 } 931 932 /** 933 * evlist__mmap_ex - Create mmaps to receive events. 934 * @evlist: list of events 935 * @pages: map length in pages 936 * @overwrite: overwrite older events? 937 * @auxtrace_pages - auxtrace map length in pages 938 * @auxtrace_overwrite - overwrite older auxtrace data? 939 * 940 * If @overwrite is %false the user needs to signal event consumption using 941 * perf_mmap__write_tail(). Using evlist__mmap_read() does this 942 * automatically. 943 * 944 * Similarly, if @auxtrace_overwrite is %false the user needs to signal data 945 * consumption using auxtrace_mmap__write_tail(). 946 * 947 * Return: %0 on success, negative error code otherwise. 948 */ 949 int evlist__mmap_ex(struct evlist *evlist, unsigned int pages, 950 unsigned int auxtrace_pages, 951 bool auxtrace_overwrite, int nr_cblocks, int affinity, int flush, 952 int comp_level) 953 { 954 /* 955 * Delay setting mp.prot: set it before calling perf_mmap__mmap. 956 * Its value is decided by evsel's write_backward. 957 * So &mp should not be passed through const pointer. 958 */ 959 struct mmap_params mp = { 960 .nr_cblocks = nr_cblocks, 961 .affinity = affinity, 962 .flush = flush, 963 .comp_level = comp_level 964 }; 965 struct perf_evlist_mmap_ops ops = { 966 .idx = perf_evlist__mmap_cb_idx, 967 .get = perf_evlist__mmap_cb_get, 968 .mmap = perf_evlist__mmap_cb_mmap, 969 }; 970 971 evlist->core.mmap_len = evlist__mmap_size(pages); 972 pr_debug("mmap size %zuB\n", evlist->core.mmap_len); 973 974 auxtrace_mmap_params__init(&mp.auxtrace_mp, evlist->core.mmap_len, 975 auxtrace_pages, auxtrace_overwrite); 976 977 return perf_evlist__mmap_ops(&evlist->core, &ops, &mp.core); 978 } 979 980 int evlist__mmap(struct evlist *evlist, unsigned int pages) 981 { 982 return evlist__mmap_ex(evlist, pages, 0, false, 0, PERF_AFFINITY_SYS, 1, 0); 983 } 984 985 int evlist__create_maps(struct evlist *evlist, struct target *target) 986 { 987 bool all_threads = (target->per_thread && target->system_wide); 988 struct perf_cpu_map *cpus; 989 struct perf_thread_map *threads; 990 991 /* 992 * If specify '-a' and '--per-thread' to perf record, perf record 993 * will override '--per-thread'. target->per_thread = false and 994 * target->system_wide = true. 995 * 996 * If specify '--per-thread' only to perf record, 997 * target->per_thread = true and target->system_wide = false. 998 * 999 * So target->per_thread && target->system_wide is false. 1000 * For perf record, thread_map__new_str doesn't call 1001 * thread_map__new_all_cpus. That will keep perf record's 1002 * current behavior. 1003 * 1004 * For perf stat, it allows the case that target->per_thread and 1005 * target->system_wide are all true. It means to collect system-wide 1006 * per-thread data. thread_map__new_str will call 1007 * thread_map__new_all_cpus to enumerate all threads. 1008 */ 1009 threads = thread_map__new_str(target->pid, target->tid, all_threads); 1010 1011 if (!threads) 1012 return -1; 1013 1014 if (target__uses_dummy_map(target) && !evlist__has_bpf_output(evlist)) 1015 cpus = perf_cpu_map__new_any_cpu(); 1016 else 1017 cpus = perf_cpu_map__new(target->cpu_list); 1018 1019 if (!cpus) 1020 goto out_delete_threads; 1021 1022 evlist->core.has_user_cpus = !!target->cpu_list; 1023 1024 perf_evlist__set_maps(&evlist->core, cpus, threads); 1025 1026 /* as evlist now has references, put count here */ 1027 perf_cpu_map__put(cpus); 1028 perf_thread_map__put(threads); 1029 1030 return 0; 1031 1032 out_delete_threads: 1033 perf_thread_map__put(threads); 1034 return -1; 1035 } 1036 1037 int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel, 1038 struct target *target) 1039 { 1040 struct evsel *evsel; 1041 int err = 0; 1042 1043 evlist__for_each_entry(evlist, evsel) { 1044 /* 1045 * filters only work for tracepoint event, which doesn't have cpu limit. 1046 * So evlist and evsel should always be same. 1047 */ 1048 if (evsel->filter) { 1049 err = perf_evsel__apply_filter(&evsel->core, evsel->filter); 1050 if (err) { 1051 *err_evsel = evsel; 1052 break; 1053 } 1054 } 1055 1056 /* 1057 * non-tracepoint events can have BPF filters. 1058 */ 1059 if (!list_empty(&evsel->bpf_filters)) { 1060 err = perf_bpf_filter__prepare(evsel, target); 1061 if (err) { 1062 *err_evsel = evsel; 1063 break; 1064 } 1065 } 1066 } 1067 1068 return err; 1069 } 1070 1071 int evlist__set_tp_filter(struct evlist *evlist, const char *filter) 1072 { 1073 struct evsel *evsel; 1074 int err = 0; 1075 1076 if (filter == NULL) 1077 return -1; 1078 1079 evlist__for_each_entry(evlist, evsel) { 1080 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT) 1081 continue; 1082 1083 err = evsel__set_filter(evsel, filter); 1084 if (err) 1085 break; 1086 } 1087 1088 return err; 1089 } 1090 1091 int evlist__append_tp_filter(struct evlist *evlist, const char *filter) 1092 { 1093 struct evsel *evsel; 1094 int err = 0; 1095 1096 if (filter == NULL) 1097 return -1; 1098 1099 evlist__for_each_entry(evlist, evsel) { 1100 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT) 1101 continue; 1102 1103 err = evsel__append_tp_filter(evsel, filter); 1104 if (err) 1105 break; 1106 } 1107 1108 return err; 1109 } 1110 1111 char *asprintf__tp_filter_pids(size_t npids, pid_t *pids) 1112 { 1113 char *filter; 1114 size_t i; 1115 1116 for (i = 0; i < npids; ++i) { 1117 if (i == 0) { 1118 if (asprintf(&filter, "common_pid != %d", pids[i]) < 0) 1119 return NULL; 1120 } else { 1121 char *tmp; 1122 1123 if (asprintf(&tmp, "%s && common_pid != %d", filter, pids[i]) < 0) 1124 goto out_free; 1125 1126 free(filter); 1127 filter = tmp; 1128 } 1129 } 1130 1131 return filter; 1132 out_free: 1133 free(filter); 1134 return NULL; 1135 } 1136 1137 int evlist__set_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids) 1138 { 1139 char *filter = asprintf__tp_filter_pids(npids, pids); 1140 int ret = evlist__set_tp_filter(evlist, filter); 1141 1142 free(filter); 1143 return ret; 1144 } 1145 1146 int evlist__append_tp_filter_pids(struct evlist *evlist, size_t npids, pid_t *pids) 1147 { 1148 char *filter = asprintf__tp_filter_pids(npids, pids); 1149 int ret = evlist__append_tp_filter(evlist, filter); 1150 1151 free(filter); 1152 return ret; 1153 } 1154 1155 int evlist__append_tp_filter_pid(struct evlist *evlist, pid_t pid) 1156 { 1157 return evlist__append_tp_filter_pids(evlist, 1, &pid); 1158 } 1159 1160 bool evlist__valid_sample_type(struct evlist *evlist) 1161 { 1162 struct evsel *pos; 1163 1164 if (evlist->core.nr_entries == 1) 1165 return true; 1166 1167 if (evlist->id_pos < 0 || evlist->is_pos < 0) 1168 return false; 1169 1170 evlist__for_each_entry(evlist, pos) { 1171 if (pos->id_pos != evlist->id_pos || 1172 pos->is_pos != evlist->is_pos) 1173 return false; 1174 } 1175 1176 return true; 1177 } 1178 1179 u64 __evlist__combined_sample_type(struct evlist *evlist) 1180 { 1181 struct evsel *evsel; 1182 1183 if (evlist->combined_sample_type) 1184 return evlist->combined_sample_type; 1185 1186 evlist__for_each_entry(evlist, evsel) 1187 evlist->combined_sample_type |= evsel->core.attr.sample_type; 1188 1189 return evlist->combined_sample_type; 1190 } 1191 1192 u64 evlist__combined_sample_type(struct evlist *evlist) 1193 { 1194 evlist->combined_sample_type = 0; 1195 return __evlist__combined_sample_type(evlist); 1196 } 1197 1198 u64 evlist__combined_branch_type(struct evlist *evlist) 1199 { 1200 struct evsel *evsel; 1201 u64 branch_type = 0; 1202 1203 evlist__for_each_entry(evlist, evsel) 1204 branch_type |= evsel->core.attr.branch_sample_type; 1205 return branch_type; 1206 } 1207 1208 static struct evsel * 1209 evlist__find_dup_event_from_prev(struct evlist *evlist, struct evsel *event) 1210 { 1211 struct evsel *pos; 1212 1213 evlist__for_each_entry(evlist, pos) { 1214 if (event == pos) 1215 break; 1216 if ((pos->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) && 1217 !strcmp(pos->name, event->name)) 1218 return pos; 1219 } 1220 return NULL; 1221 } 1222 1223 #define MAX_NR_ABBR_NAME (26 * 11) 1224 1225 /* 1226 * The abbr name is from A to Z9. If the number of event 1227 * which requires the branch counter > MAX_NR_ABBR_NAME, 1228 * return NA. 1229 */ 1230 static void evlist__new_abbr_name(char *name) 1231 { 1232 static int idx; 1233 int i = idx / 26; 1234 1235 if (idx >= MAX_NR_ABBR_NAME) { 1236 name[0] = 'N'; 1237 name[1] = 'A'; 1238 name[2] = '\0'; 1239 return; 1240 } 1241 1242 name[0] = 'A' + (idx % 26); 1243 1244 if (!i) 1245 name[1] = '\0'; 1246 else { 1247 name[1] = '0' + i - 1; 1248 name[2] = '\0'; 1249 } 1250 1251 idx++; 1252 } 1253 1254 void evlist__update_br_cntr(struct evlist *evlist) 1255 { 1256 struct evsel *evsel, *dup; 1257 int i = 0; 1258 1259 evlist__for_each_entry(evlist, evsel) { 1260 if (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) { 1261 evsel->br_cntr_idx = i++; 1262 evsel__leader(evsel)->br_cntr_nr++; 1263 1264 dup = evlist__find_dup_event_from_prev(evlist, evsel); 1265 if (dup) 1266 memcpy(evsel->abbr_name, dup->abbr_name, 3 * sizeof(char)); 1267 else 1268 evlist__new_abbr_name(evsel->abbr_name); 1269 } 1270 } 1271 evlist->nr_br_cntr = i; 1272 } 1273 1274 bool evlist__valid_read_format(struct evlist *evlist) 1275 { 1276 struct evsel *first = evlist__first(evlist), *pos = first; 1277 u64 read_format = first->core.attr.read_format; 1278 u64 sample_type = first->core.attr.sample_type; 1279 1280 evlist__for_each_entry(evlist, pos) { 1281 if (read_format != pos->core.attr.read_format) { 1282 pr_debug("Read format differs %#" PRIx64 " vs %#" PRIx64 "\n", 1283 read_format, (u64)pos->core.attr.read_format); 1284 } 1285 } 1286 1287 /* PERF_SAMPLE_READ implies PERF_FORMAT_ID. */ 1288 if ((sample_type & PERF_SAMPLE_READ) && 1289 !(read_format & PERF_FORMAT_ID)) { 1290 return false; 1291 } 1292 1293 return true; 1294 } 1295 1296 u16 evlist__id_hdr_size(struct evlist *evlist) 1297 { 1298 struct evsel *first = evlist__first(evlist); 1299 1300 return first->core.attr.sample_id_all ? evsel__id_hdr_size(first) : 0; 1301 } 1302 1303 bool evlist__valid_sample_id_all(struct evlist *evlist) 1304 { 1305 struct evsel *first = evlist__first(evlist), *pos = first; 1306 1307 evlist__for_each_entry_continue(evlist, pos) { 1308 if (first->core.attr.sample_id_all != pos->core.attr.sample_id_all) 1309 return false; 1310 } 1311 1312 return true; 1313 } 1314 1315 bool evlist__sample_id_all(struct evlist *evlist) 1316 { 1317 struct evsel *first = evlist__first(evlist); 1318 return first->core.attr.sample_id_all; 1319 } 1320 1321 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel) 1322 { 1323 evlist->selected = evsel; 1324 } 1325 1326 void evlist__close(struct evlist *evlist) 1327 { 1328 struct evsel *evsel; 1329 struct evlist_cpu_iterator evlist_cpu_itr; 1330 struct affinity affinity; 1331 1332 /* 1333 * With perf record core.user_requested_cpus is usually NULL. 1334 * Use the old method to handle this for now. 1335 */ 1336 if (!evlist->core.user_requested_cpus || 1337 cpu_map__is_dummy(evlist->core.user_requested_cpus)) { 1338 evlist__for_each_entry_reverse(evlist, evsel) 1339 evsel__close(evsel); 1340 return; 1341 } 1342 1343 if (affinity__setup(&affinity) < 0) 1344 return; 1345 1346 evlist__for_each_cpu(evlist_cpu_itr, evlist, &affinity) { 1347 perf_evsel__close_cpu(&evlist_cpu_itr.evsel->core, 1348 evlist_cpu_itr.cpu_map_idx); 1349 } 1350 1351 affinity__cleanup(&affinity); 1352 evlist__for_each_entry_reverse(evlist, evsel) { 1353 perf_evsel__free_fd(&evsel->core); 1354 perf_evsel__free_id(&evsel->core); 1355 } 1356 perf_evlist__reset_id_hash(&evlist->core); 1357 } 1358 1359 static int evlist__create_syswide_maps(struct evlist *evlist) 1360 { 1361 struct perf_cpu_map *cpus; 1362 struct perf_thread_map *threads; 1363 1364 /* 1365 * Try reading /sys/devices/system/cpu/online to get 1366 * an all cpus map. 1367 * 1368 * FIXME: -ENOMEM is the best we can do here, the cpu_map 1369 * code needs an overhaul to properly forward the 1370 * error, and we may not want to do that fallback to a 1371 * default cpu identity map :-\ 1372 */ 1373 cpus = perf_cpu_map__new_online_cpus(); 1374 if (!cpus) 1375 return -ENOMEM; 1376 1377 threads = perf_thread_map__new_dummy(); 1378 if (!threads) { 1379 perf_cpu_map__put(cpus); 1380 return -ENOMEM; 1381 } 1382 1383 perf_evlist__set_maps(&evlist->core, cpus, threads); 1384 perf_thread_map__put(threads); 1385 perf_cpu_map__put(cpus); 1386 return 0; 1387 } 1388 1389 int evlist__open(struct evlist *evlist) 1390 { 1391 struct evsel *evsel; 1392 int err; 1393 1394 /* 1395 * Default: one fd per CPU, all threads, aka systemwide 1396 * as sys_perf_event_open(cpu = -1, thread = -1) is EINVAL 1397 */ 1398 if (evlist->core.threads == NULL && evlist->core.user_requested_cpus == NULL) { 1399 err = evlist__create_syswide_maps(evlist); 1400 if (err < 0) 1401 goto out_err; 1402 } 1403 1404 evlist__update_id_pos(evlist); 1405 1406 evlist__for_each_entry(evlist, evsel) { 1407 err = evsel__open(evsel, evsel->core.cpus, evsel->core.threads); 1408 if (err < 0) 1409 goto out_err; 1410 } 1411 1412 return 0; 1413 out_err: 1414 evlist__close(evlist); 1415 errno = -err; 1416 return err; 1417 } 1418 1419 int evlist__prepare_workload(struct evlist *evlist, struct target *target, const char *argv[], 1420 bool pipe_output, void (*exec_error)(int signo, siginfo_t *info, void *ucontext)) 1421 { 1422 int child_ready_pipe[2], go_pipe[2]; 1423 char bf; 1424 1425 evlist->workload.cork_fd = -1; 1426 1427 if (pipe(child_ready_pipe) < 0) { 1428 perror("failed to create 'ready' pipe"); 1429 return -1; 1430 } 1431 1432 if (pipe(go_pipe) < 0) { 1433 perror("failed to create 'go' pipe"); 1434 goto out_close_ready_pipe; 1435 } 1436 1437 evlist->workload.pid = fork(); 1438 if (evlist->workload.pid < 0) { 1439 perror("failed to fork"); 1440 goto out_close_pipes; 1441 } 1442 1443 if (!evlist->workload.pid) { 1444 int ret; 1445 1446 if (pipe_output) 1447 dup2(2, 1); 1448 1449 signal(SIGTERM, SIG_DFL); 1450 1451 close(child_ready_pipe[0]); 1452 close(go_pipe[1]); 1453 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC); 1454 1455 /* 1456 * Change the name of this process not to confuse --exclude-perf users 1457 * that sees 'perf' in the window up to the execvp() and thinks that 1458 * perf samples are not being excluded. 1459 */ 1460 prctl(PR_SET_NAME, "perf-exec"); 1461 1462 /* 1463 * Tell the parent we're ready to go 1464 */ 1465 close(child_ready_pipe[1]); 1466 1467 /* 1468 * Wait until the parent tells us to go. 1469 */ 1470 ret = read(go_pipe[0], &bf, 1); 1471 /* 1472 * The parent will ask for the execvp() to be performed by 1473 * writing exactly one byte, in workload.cork_fd, usually via 1474 * evlist__start_workload(). 1475 * 1476 * For cancelling the workload without actually running it, 1477 * the parent will just close workload.cork_fd, without writing 1478 * anything, i.e. read will return zero and we just exit() 1479 * here (See evlist__cancel_workload()). 1480 */ 1481 if (ret != 1) { 1482 if (ret == -1) 1483 perror("unable to read pipe"); 1484 exit(ret); 1485 } 1486 1487 execvp(argv[0], (char **)argv); 1488 1489 if (exec_error) { 1490 union sigval val; 1491 1492 val.sival_int = errno; 1493 if (sigqueue(getppid(), SIGUSR1, val)) 1494 perror(argv[0]); 1495 } else 1496 perror(argv[0]); 1497 exit(-1); 1498 } 1499 1500 if (exec_error) { 1501 struct sigaction act = { 1502 .sa_flags = SA_SIGINFO, 1503 .sa_sigaction = exec_error, 1504 }; 1505 sigaction(SIGUSR1, &act, NULL); 1506 } 1507 1508 if (target__none(target)) { 1509 if (evlist->core.threads == NULL) { 1510 fprintf(stderr, "FATAL: evlist->threads need to be set at this point (%s:%d).\n", 1511 __func__, __LINE__); 1512 goto out_close_pipes; 1513 } 1514 perf_thread_map__set_pid(evlist->core.threads, 0, evlist->workload.pid); 1515 } 1516 1517 close(child_ready_pipe[1]); 1518 close(go_pipe[0]); 1519 /* 1520 * wait for child to settle 1521 */ 1522 if (read(child_ready_pipe[0], &bf, 1) == -1) { 1523 perror("unable to read pipe"); 1524 goto out_close_pipes; 1525 } 1526 1527 fcntl(go_pipe[1], F_SETFD, FD_CLOEXEC); 1528 evlist->workload.cork_fd = go_pipe[1]; 1529 close(child_ready_pipe[0]); 1530 return 0; 1531 1532 out_close_pipes: 1533 close(go_pipe[0]); 1534 close(go_pipe[1]); 1535 out_close_ready_pipe: 1536 close(child_ready_pipe[0]); 1537 close(child_ready_pipe[1]); 1538 return -1; 1539 } 1540 1541 int evlist__start_workload(struct evlist *evlist) 1542 { 1543 if (evlist->workload.cork_fd >= 0) { 1544 char bf = 0; 1545 int ret; 1546 /* 1547 * Remove the cork, let it rip! 1548 */ 1549 ret = write(evlist->workload.cork_fd, &bf, 1); 1550 if (ret < 0) 1551 perror("unable to write to pipe"); 1552 1553 close(evlist->workload.cork_fd); 1554 evlist->workload.cork_fd = -1; 1555 return ret; 1556 } 1557 1558 return 0; 1559 } 1560 1561 void evlist__cancel_workload(struct evlist *evlist) 1562 { 1563 int status; 1564 1565 if (evlist->workload.cork_fd >= 0) { 1566 close(evlist->workload.cork_fd); 1567 evlist->workload.cork_fd = -1; 1568 waitpid(evlist->workload.pid, &status, WNOHANG); 1569 } 1570 } 1571 1572 int evlist__parse_sample(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1573 { 1574 struct evsel *evsel = evlist__event2evsel(evlist, event); 1575 int ret; 1576 1577 if (!evsel) 1578 return -EFAULT; 1579 ret = evsel__parse_sample(evsel, event, sample); 1580 if (ret) 1581 return ret; 1582 if (perf_guest && sample->id) { 1583 struct perf_sample_id *sid = evlist__id2sid(evlist, sample->id); 1584 1585 if (sid) { 1586 sample->machine_pid = sid->machine_pid; 1587 sample->vcpu = sid->vcpu.cpu; 1588 } 1589 } 1590 return 0; 1591 } 1592 1593 int evlist__parse_sample_timestamp(struct evlist *evlist, union perf_event *event, u64 *timestamp) 1594 { 1595 struct evsel *evsel = evlist__event2evsel(evlist, event); 1596 1597 if (!evsel) 1598 return -EFAULT; 1599 return evsel__parse_sample_timestamp(evsel, event, timestamp); 1600 } 1601 1602 int evlist__strerror_open(struct evlist *evlist, int err, char *buf, size_t size) 1603 { 1604 int printed, value; 1605 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf)); 1606 1607 switch (err) { 1608 case EACCES: 1609 case EPERM: 1610 printed = scnprintf(buf, size, 1611 "Error:\t%s.\n" 1612 "Hint:\tCheck /proc/sys/kernel/perf_event_paranoid setting.", emsg); 1613 1614 value = perf_event_paranoid(); 1615 1616 printed += scnprintf(buf + printed, size - printed, "\nHint:\t"); 1617 1618 if (value >= 2) { 1619 printed += scnprintf(buf + printed, size - printed, 1620 "For your workloads it needs to be <= 1\nHint:\t"); 1621 } 1622 printed += scnprintf(buf + printed, size - printed, 1623 "For system wide tracing it needs to be set to -1.\n"); 1624 1625 printed += scnprintf(buf + printed, size - printed, 1626 "Hint:\tTry: 'sudo sh -c \"echo -1 > /proc/sys/kernel/perf_event_paranoid\"'\n" 1627 "Hint:\tThe current value is %d.", value); 1628 break; 1629 case EINVAL: { 1630 struct evsel *first = evlist__first(evlist); 1631 int max_freq; 1632 1633 if (sysctl__read_int("kernel/perf_event_max_sample_rate", &max_freq) < 0) 1634 goto out_default; 1635 1636 if (first->core.attr.sample_freq < (u64)max_freq) 1637 goto out_default; 1638 1639 printed = scnprintf(buf, size, 1640 "Error:\t%s.\n" 1641 "Hint:\tCheck /proc/sys/kernel/perf_event_max_sample_rate.\n" 1642 "Hint:\tThe current value is %d and %" PRIu64 " is being requested.", 1643 emsg, max_freq, first->core.attr.sample_freq); 1644 break; 1645 } 1646 default: 1647 out_default: 1648 scnprintf(buf, size, "%s", emsg); 1649 break; 1650 } 1651 1652 return 0; 1653 } 1654 1655 int evlist__strerror_mmap(struct evlist *evlist, int err, char *buf, size_t size) 1656 { 1657 char sbuf[STRERR_BUFSIZE], *emsg = str_error_r(err, sbuf, sizeof(sbuf)); 1658 int pages_attempted = evlist->core.mmap_len / 1024, pages_max_per_user, printed = 0; 1659 1660 switch (err) { 1661 case EPERM: 1662 sysctl__read_int("kernel/perf_event_mlock_kb", &pages_max_per_user); 1663 printed += scnprintf(buf + printed, size - printed, 1664 "Error:\t%s.\n" 1665 "Hint:\tCheck /proc/sys/kernel/perf_event_mlock_kb (%d kB) setting.\n" 1666 "Hint:\tTried using %zd kB.\n", 1667 emsg, pages_max_per_user, pages_attempted); 1668 1669 if (pages_attempted >= pages_max_per_user) { 1670 printed += scnprintf(buf + printed, size - printed, 1671 "Hint:\tTry 'sudo sh -c \"echo %d > /proc/sys/kernel/perf_event_mlock_kb\"', or\n", 1672 pages_max_per_user + pages_attempted); 1673 } 1674 1675 printed += scnprintf(buf + printed, size - printed, 1676 "Hint:\tTry using a smaller -m/--mmap-pages value."); 1677 break; 1678 default: 1679 scnprintf(buf, size, "%s", emsg); 1680 break; 1681 } 1682 1683 return 0; 1684 } 1685 1686 void evlist__to_front(struct evlist *evlist, struct evsel *move_evsel) 1687 { 1688 struct evsel *evsel, *n; 1689 LIST_HEAD(move); 1690 1691 if (move_evsel == evlist__first(evlist)) 1692 return; 1693 1694 evlist__for_each_entry_safe(evlist, n, evsel) { 1695 if (evsel__leader(evsel) == evsel__leader(move_evsel)) 1696 list_move_tail(&evsel->core.node, &move); 1697 } 1698 1699 list_splice(&move, &evlist->core.entries); 1700 } 1701 1702 struct evsel *evlist__get_tracking_event(struct evlist *evlist) 1703 { 1704 struct evsel *evsel; 1705 1706 evlist__for_each_entry(evlist, evsel) { 1707 if (evsel->tracking) 1708 return evsel; 1709 } 1710 1711 return evlist__first(evlist); 1712 } 1713 1714 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel) 1715 { 1716 struct evsel *evsel; 1717 1718 if (tracking_evsel->tracking) 1719 return; 1720 1721 evlist__for_each_entry(evlist, evsel) { 1722 if (evsel != tracking_evsel) 1723 evsel->tracking = false; 1724 } 1725 1726 tracking_evsel->tracking = true; 1727 } 1728 1729 struct evsel *evlist__findnew_tracking_event(struct evlist *evlist, bool system_wide) 1730 { 1731 struct evsel *evsel; 1732 1733 evsel = evlist__get_tracking_event(evlist); 1734 if (!evsel__is_dummy_event(evsel)) { 1735 evsel = evlist__add_aux_dummy(evlist, system_wide); 1736 if (!evsel) 1737 return NULL; 1738 1739 evlist__set_tracking_event(evlist, evsel); 1740 } else if (system_wide) { 1741 perf_evlist__go_system_wide(&evlist->core, &evsel->core); 1742 } 1743 1744 return evsel; 1745 } 1746 1747 struct evsel *evlist__find_evsel_by_str(struct evlist *evlist, const char *str) 1748 { 1749 struct evsel *evsel; 1750 1751 evlist__for_each_entry(evlist, evsel) { 1752 if (!evsel->name) 1753 continue; 1754 if (evsel__name_is(evsel, str)) 1755 return evsel; 1756 } 1757 1758 return NULL; 1759 } 1760 1761 void evlist__toggle_bkw_mmap(struct evlist *evlist, enum bkw_mmap_state state) 1762 { 1763 enum bkw_mmap_state old_state = evlist->bkw_mmap_state; 1764 enum action { 1765 NONE, 1766 PAUSE, 1767 RESUME, 1768 } action = NONE; 1769 1770 if (!evlist->overwrite_mmap) 1771 return; 1772 1773 switch (old_state) { 1774 case BKW_MMAP_NOTREADY: { 1775 if (state != BKW_MMAP_RUNNING) 1776 goto state_err; 1777 break; 1778 } 1779 case BKW_MMAP_RUNNING: { 1780 if (state != BKW_MMAP_DATA_PENDING) 1781 goto state_err; 1782 action = PAUSE; 1783 break; 1784 } 1785 case BKW_MMAP_DATA_PENDING: { 1786 if (state != BKW_MMAP_EMPTY) 1787 goto state_err; 1788 break; 1789 } 1790 case BKW_MMAP_EMPTY: { 1791 if (state != BKW_MMAP_RUNNING) 1792 goto state_err; 1793 action = RESUME; 1794 break; 1795 } 1796 default: 1797 WARN_ONCE(1, "Shouldn't get there\n"); 1798 } 1799 1800 evlist->bkw_mmap_state = state; 1801 1802 switch (action) { 1803 case PAUSE: 1804 evlist__pause(evlist); 1805 break; 1806 case RESUME: 1807 evlist__resume(evlist); 1808 break; 1809 case NONE: 1810 default: 1811 break; 1812 } 1813 1814 state_err: 1815 return; 1816 } 1817 1818 bool evlist__exclude_kernel(struct evlist *evlist) 1819 { 1820 struct evsel *evsel; 1821 1822 evlist__for_each_entry(evlist, evsel) { 1823 if (!evsel->core.attr.exclude_kernel) 1824 return false; 1825 } 1826 1827 return true; 1828 } 1829 1830 /* 1831 * Events in data file are not collect in groups, but we still want 1832 * the group display. Set the artificial group and set the leader's 1833 * forced_leader flag to notify the display code. 1834 */ 1835 void evlist__force_leader(struct evlist *evlist) 1836 { 1837 if (evlist__nr_groups(evlist) == 0) { 1838 struct evsel *leader = evlist__first(evlist); 1839 1840 evlist__set_leader(evlist); 1841 leader->forced_leader = true; 1842 } 1843 } 1844 1845 struct evsel *evlist__reset_weak_group(struct evlist *evsel_list, struct evsel *evsel, bool close) 1846 { 1847 struct evsel *c2, *leader; 1848 bool is_open = true; 1849 1850 leader = evsel__leader(evsel); 1851 1852 pr_debug("Weak group for %s/%d failed\n", 1853 leader->name, leader->core.nr_members); 1854 1855 /* 1856 * for_each_group_member doesn't work here because it doesn't 1857 * include the first entry. 1858 */ 1859 evlist__for_each_entry(evsel_list, c2) { 1860 if (c2 == evsel) 1861 is_open = false; 1862 if (evsel__has_leader(c2, leader)) { 1863 if (is_open && close) 1864 perf_evsel__close(&c2->core); 1865 /* 1866 * We want to close all members of the group and reopen 1867 * them. Some events, like Intel topdown, require being 1868 * in a group and so keep these in the group. 1869 */ 1870 evsel__remove_from_group(c2, leader); 1871 1872 /* 1873 * Set this for all former members of the group 1874 * to indicate they get reopened. 1875 */ 1876 c2->reset_group = true; 1877 } 1878 } 1879 /* Reset the leader count if all entries were removed. */ 1880 if (leader->core.nr_members == 1) 1881 leader->core.nr_members = 0; 1882 return leader; 1883 } 1884 1885 static int evlist__parse_control_fifo(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close) 1886 { 1887 char *s, *p; 1888 int ret = 0, fd; 1889 1890 if (strncmp(str, "fifo:", 5)) 1891 return -EINVAL; 1892 1893 str += 5; 1894 if (!*str || *str == ',') 1895 return -EINVAL; 1896 1897 s = strdup(str); 1898 if (!s) 1899 return -ENOMEM; 1900 1901 p = strchr(s, ','); 1902 if (p) 1903 *p = '\0'; 1904 1905 /* 1906 * O_RDWR avoids POLLHUPs which is necessary to allow the other 1907 * end of a FIFO to be repeatedly opened and closed. 1908 */ 1909 fd = open(s, O_RDWR | O_NONBLOCK | O_CLOEXEC); 1910 if (fd < 0) { 1911 pr_err("Failed to open '%s'\n", s); 1912 ret = -errno; 1913 goto out_free; 1914 } 1915 *ctl_fd = fd; 1916 *ctl_fd_close = true; 1917 1918 if (p && *++p) { 1919 /* O_RDWR | O_NONBLOCK means the other end need not be open */ 1920 fd = open(p, O_RDWR | O_NONBLOCK | O_CLOEXEC); 1921 if (fd < 0) { 1922 pr_err("Failed to open '%s'\n", p); 1923 ret = -errno; 1924 goto out_free; 1925 } 1926 *ctl_fd_ack = fd; 1927 } 1928 1929 out_free: 1930 free(s); 1931 return ret; 1932 } 1933 1934 int evlist__parse_control(const char *str, int *ctl_fd, int *ctl_fd_ack, bool *ctl_fd_close) 1935 { 1936 char *comma = NULL, *endptr = NULL; 1937 1938 *ctl_fd_close = false; 1939 1940 if (strncmp(str, "fd:", 3)) 1941 return evlist__parse_control_fifo(str, ctl_fd, ctl_fd_ack, ctl_fd_close); 1942 1943 *ctl_fd = strtoul(&str[3], &endptr, 0); 1944 if (endptr == &str[3]) 1945 return -EINVAL; 1946 1947 comma = strchr(str, ','); 1948 if (comma) { 1949 if (endptr != comma) 1950 return -EINVAL; 1951 1952 *ctl_fd_ack = strtoul(comma + 1, &endptr, 0); 1953 if (endptr == comma + 1 || *endptr != '\0') 1954 return -EINVAL; 1955 } 1956 1957 return 0; 1958 } 1959 1960 void evlist__close_control(int ctl_fd, int ctl_fd_ack, bool *ctl_fd_close) 1961 { 1962 if (*ctl_fd_close) { 1963 *ctl_fd_close = false; 1964 close(ctl_fd); 1965 if (ctl_fd_ack >= 0) 1966 close(ctl_fd_ack); 1967 } 1968 } 1969 1970 int evlist__initialize_ctlfd(struct evlist *evlist, int fd, int ack) 1971 { 1972 if (fd == -1) { 1973 pr_debug("Control descriptor is not initialized\n"); 1974 return 0; 1975 } 1976 1977 evlist->ctl_fd.pos = perf_evlist__add_pollfd(&evlist->core, fd, NULL, POLLIN, 1978 fdarray_flag__nonfilterable | 1979 fdarray_flag__non_perf_event); 1980 if (evlist->ctl_fd.pos < 0) { 1981 evlist->ctl_fd.pos = -1; 1982 pr_err("Failed to add ctl fd entry: %m\n"); 1983 return -1; 1984 } 1985 1986 evlist->ctl_fd.fd = fd; 1987 evlist->ctl_fd.ack = ack; 1988 1989 return 0; 1990 } 1991 1992 bool evlist__ctlfd_initialized(struct evlist *evlist) 1993 { 1994 return evlist->ctl_fd.pos >= 0; 1995 } 1996 1997 int evlist__finalize_ctlfd(struct evlist *evlist) 1998 { 1999 struct pollfd *entries = evlist->core.pollfd.entries; 2000 2001 if (!evlist__ctlfd_initialized(evlist)) 2002 return 0; 2003 2004 entries[evlist->ctl_fd.pos].fd = -1; 2005 entries[evlist->ctl_fd.pos].events = 0; 2006 entries[evlist->ctl_fd.pos].revents = 0; 2007 2008 evlist->ctl_fd.pos = -1; 2009 evlist->ctl_fd.ack = -1; 2010 evlist->ctl_fd.fd = -1; 2011 2012 return 0; 2013 } 2014 2015 static int evlist__ctlfd_recv(struct evlist *evlist, enum evlist_ctl_cmd *cmd, 2016 char *cmd_data, size_t data_size) 2017 { 2018 int err; 2019 char c; 2020 size_t bytes_read = 0; 2021 2022 *cmd = EVLIST_CTL_CMD_UNSUPPORTED; 2023 memset(cmd_data, 0, data_size); 2024 data_size--; 2025 2026 do { 2027 err = read(evlist->ctl_fd.fd, &c, 1); 2028 if (err > 0) { 2029 if (c == '\n' || c == '\0') 2030 break; 2031 cmd_data[bytes_read++] = c; 2032 if (bytes_read == data_size) 2033 break; 2034 continue; 2035 } else if (err == -1) { 2036 if (errno == EINTR) 2037 continue; 2038 if (errno == EAGAIN || errno == EWOULDBLOCK) 2039 err = 0; 2040 else 2041 pr_err("Failed to read from ctlfd %d: %m\n", evlist->ctl_fd.fd); 2042 } 2043 break; 2044 } while (1); 2045 2046 pr_debug("Message from ctl_fd: \"%s%s\"\n", cmd_data, 2047 bytes_read == data_size ? "" : c == '\n' ? "\\n" : "\\0"); 2048 2049 if (bytes_read > 0) { 2050 if (!strncmp(cmd_data, EVLIST_CTL_CMD_ENABLE_TAG, 2051 (sizeof(EVLIST_CTL_CMD_ENABLE_TAG)-1))) { 2052 *cmd = EVLIST_CTL_CMD_ENABLE; 2053 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_DISABLE_TAG, 2054 (sizeof(EVLIST_CTL_CMD_DISABLE_TAG)-1))) { 2055 *cmd = EVLIST_CTL_CMD_DISABLE; 2056 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_SNAPSHOT_TAG, 2057 (sizeof(EVLIST_CTL_CMD_SNAPSHOT_TAG)-1))) { 2058 *cmd = EVLIST_CTL_CMD_SNAPSHOT; 2059 pr_debug("is snapshot\n"); 2060 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_EVLIST_TAG, 2061 (sizeof(EVLIST_CTL_CMD_EVLIST_TAG)-1))) { 2062 *cmd = EVLIST_CTL_CMD_EVLIST; 2063 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_STOP_TAG, 2064 (sizeof(EVLIST_CTL_CMD_STOP_TAG)-1))) { 2065 *cmd = EVLIST_CTL_CMD_STOP; 2066 } else if (!strncmp(cmd_data, EVLIST_CTL_CMD_PING_TAG, 2067 (sizeof(EVLIST_CTL_CMD_PING_TAG)-1))) { 2068 *cmd = EVLIST_CTL_CMD_PING; 2069 } 2070 } 2071 2072 return bytes_read ? (int)bytes_read : err; 2073 } 2074 2075 int evlist__ctlfd_ack(struct evlist *evlist) 2076 { 2077 int err; 2078 2079 if (evlist->ctl_fd.ack == -1) 2080 return 0; 2081 2082 err = write(evlist->ctl_fd.ack, EVLIST_CTL_CMD_ACK_TAG, 2083 sizeof(EVLIST_CTL_CMD_ACK_TAG)); 2084 if (err == -1) 2085 pr_err("failed to write to ctl_ack_fd %d: %m\n", evlist->ctl_fd.ack); 2086 2087 return err; 2088 } 2089 2090 static int get_cmd_arg(char *cmd_data, size_t cmd_size, char **arg) 2091 { 2092 char *data = cmd_data + cmd_size; 2093 2094 /* no argument */ 2095 if (!*data) 2096 return 0; 2097 2098 /* there's argument */ 2099 if (*data == ' ') { 2100 *arg = data + 1; 2101 return 1; 2102 } 2103 2104 /* malformed */ 2105 return -1; 2106 } 2107 2108 static int evlist__ctlfd_enable(struct evlist *evlist, char *cmd_data, bool enable) 2109 { 2110 struct evsel *evsel; 2111 char *name; 2112 int err; 2113 2114 err = get_cmd_arg(cmd_data, 2115 enable ? sizeof(EVLIST_CTL_CMD_ENABLE_TAG) - 1 : 2116 sizeof(EVLIST_CTL_CMD_DISABLE_TAG) - 1, 2117 &name); 2118 if (err < 0) { 2119 pr_info("failed: wrong command\n"); 2120 return -1; 2121 } 2122 2123 if (err) { 2124 evsel = evlist__find_evsel_by_str(evlist, name); 2125 if (evsel) { 2126 if (enable) 2127 evlist__enable_evsel(evlist, name); 2128 else 2129 evlist__disable_evsel(evlist, name); 2130 pr_info("Event %s %s\n", evsel->name, 2131 enable ? "enabled" : "disabled"); 2132 } else { 2133 pr_info("failed: can't find '%s' event\n", name); 2134 } 2135 } else { 2136 if (enable) { 2137 evlist__enable(evlist); 2138 pr_info(EVLIST_ENABLED_MSG); 2139 } else { 2140 evlist__disable(evlist); 2141 pr_info(EVLIST_DISABLED_MSG); 2142 } 2143 } 2144 2145 return 0; 2146 } 2147 2148 static int evlist__ctlfd_list(struct evlist *evlist, char *cmd_data) 2149 { 2150 struct perf_attr_details details = { .verbose = false, }; 2151 struct evsel *evsel; 2152 char *arg; 2153 int err; 2154 2155 err = get_cmd_arg(cmd_data, 2156 sizeof(EVLIST_CTL_CMD_EVLIST_TAG) - 1, 2157 &arg); 2158 if (err < 0) { 2159 pr_info("failed: wrong command\n"); 2160 return -1; 2161 } 2162 2163 if (err) { 2164 if (!strcmp(arg, "-v")) { 2165 details.verbose = true; 2166 } else if (!strcmp(arg, "-g")) { 2167 details.event_group = true; 2168 } else if (!strcmp(arg, "-F")) { 2169 details.freq = true; 2170 } else { 2171 pr_info("failed: wrong command\n"); 2172 return -1; 2173 } 2174 } 2175 2176 evlist__for_each_entry(evlist, evsel) 2177 evsel__fprintf(evsel, &details, stderr); 2178 2179 return 0; 2180 } 2181 2182 int evlist__ctlfd_process(struct evlist *evlist, enum evlist_ctl_cmd *cmd) 2183 { 2184 int err = 0; 2185 char cmd_data[EVLIST_CTL_CMD_MAX_LEN]; 2186 int ctlfd_pos = evlist->ctl_fd.pos; 2187 struct pollfd *entries = evlist->core.pollfd.entries; 2188 2189 if (!evlist__ctlfd_initialized(evlist) || !entries[ctlfd_pos].revents) 2190 return 0; 2191 2192 if (entries[ctlfd_pos].revents & POLLIN) { 2193 err = evlist__ctlfd_recv(evlist, cmd, cmd_data, 2194 EVLIST_CTL_CMD_MAX_LEN); 2195 if (err > 0) { 2196 switch (*cmd) { 2197 case EVLIST_CTL_CMD_ENABLE: 2198 case EVLIST_CTL_CMD_DISABLE: 2199 err = evlist__ctlfd_enable(evlist, cmd_data, 2200 *cmd == EVLIST_CTL_CMD_ENABLE); 2201 break; 2202 case EVLIST_CTL_CMD_EVLIST: 2203 err = evlist__ctlfd_list(evlist, cmd_data); 2204 break; 2205 case EVLIST_CTL_CMD_SNAPSHOT: 2206 case EVLIST_CTL_CMD_STOP: 2207 case EVLIST_CTL_CMD_PING: 2208 break; 2209 case EVLIST_CTL_CMD_ACK: 2210 case EVLIST_CTL_CMD_UNSUPPORTED: 2211 default: 2212 pr_debug("ctlfd: unsupported %d\n", *cmd); 2213 break; 2214 } 2215 if (!(*cmd == EVLIST_CTL_CMD_ACK || *cmd == EVLIST_CTL_CMD_UNSUPPORTED || 2216 *cmd == EVLIST_CTL_CMD_SNAPSHOT)) 2217 evlist__ctlfd_ack(evlist); 2218 } 2219 } 2220 2221 if (entries[ctlfd_pos].revents & (POLLHUP | POLLERR)) 2222 evlist__finalize_ctlfd(evlist); 2223 else 2224 entries[ctlfd_pos].revents = 0; 2225 2226 return err; 2227 } 2228 2229 /** 2230 * struct event_enable_time - perf record -D/--delay single time range. 2231 * @start: start of time range to enable events in milliseconds 2232 * @end: end of time range to enable events in milliseconds 2233 * 2234 * N.B. this structure is also accessed as an array of int. 2235 */ 2236 struct event_enable_time { 2237 int start; 2238 int end; 2239 }; 2240 2241 static int parse_event_enable_time(const char *str, struct event_enable_time *range, bool first) 2242 { 2243 const char *fmt = first ? "%u - %u %n" : " , %u - %u %n"; 2244 int ret, start, end, n; 2245 2246 ret = sscanf(str, fmt, &start, &end, &n); 2247 if (ret != 2 || end <= start) 2248 return -EINVAL; 2249 if (range) { 2250 range->start = start; 2251 range->end = end; 2252 } 2253 return n; 2254 } 2255 2256 static ssize_t parse_event_enable_times(const char *str, struct event_enable_time *range) 2257 { 2258 int incr = !!range; 2259 bool first = true; 2260 ssize_t ret, cnt; 2261 2262 for (cnt = 0; *str; cnt++) { 2263 ret = parse_event_enable_time(str, range, first); 2264 if (ret < 0) 2265 return ret; 2266 /* Check no overlap */ 2267 if (!first && range && range->start <= range[-1].end) 2268 return -EINVAL; 2269 str += ret; 2270 range += incr; 2271 first = false; 2272 } 2273 return cnt; 2274 } 2275 2276 /** 2277 * struct event_enable_timer - control structure for perf record -D/--delay. 2278 * @evlist: event list 2279 * @times: time ranges that events are enabled (N.B. this is also accessed as an 2280 * array of int) 2281 * @times_cnt: number of time ranges 2282 * @timerfd: timer file descriptor 2283 * @pollfd_pos: position in @evlist array of file descriptors to poll (fdarray) 2284 * @times_step: current position in (int *)@times)[], 2285 * refer event_enable_timer__process() 2286 * 2287 * Note, this structure is only used when there are time ranges, not when there 2288 * is only an initial delay. 2289 */ 2290 struct event_enable_timer { 2291 struct evlist *evlist; 2292 struct event_enable_time *times; 2293 size_t times_cnt; 2294 int timerfd; 2295 int pollfd_pos; 2296 size_t times_step; 2297 }; 2298 2299 static int str_to_delay(const char *str) 2300 { 2301 char *endptr; 2302 long d; 2303 2304 d = strtol(str, &endptr, 10); 2305 if (*endptr || d > INT_MAX || d < -1) 2306 return 0; 2307 return d; 2308 } 2309 2310 int evlist__parse_event_enable_time(struct evlist *evlist, struct record_opts *opts, 2311 const char *str, int unset) 2312 { 2313 enum fdarray_flags flags = fdarray_flag__nonfilterable | fdarray_flag__non_perf_event; 2314 struct event_enable_timer *eet; 2315 ssize_t times_cnt; 2316 ssize_t ret; 2317 int err; 2318 2319 if (unset) 2320 return 0; 2321 2322 opts->target.initial_delay = str_to_delay(str); 2323 if (opts->target.initial_delay) 2324 return 0; 2325 2326 ret = parse_event_enable_times(str, NULL); 2327 if (ret < 0) 2328 return ret; 2329 2330 times_cnt = ret; 2331 if (times_cnt == 0) 2332 return -EINVAL; 2333 2334 eet = zalloc(sizeof(*eet)); 2335 if (!eet) 2336 return -ENOMEM; 2337 2338 eet->times = calloc(times_cnt, sizeof(*eet->times)); 2339 if (!eet->times) { 2340 err = -ENOMEM; 2341 goto free_eet; 2342 } 2343 2344 if (parse_event_enable_times(str, eet->times) != times_cnt) { 2345 err = -EINVAL; 2346 goto free_eet_times; 2347 } 2348 2349 eet->times_cnt = times_cnt; 2350 2351 eet->timerfd = timerfd_create(CLOCK_MONOTONIC, TFD_CLOEXEC); 2352 if (eet->timerfd == -1) { 2353 err = -errno; 2354 pr_err("timerfd_create failed: %s\n", strerror(errno)); 2355 goto free_eet_times; 2356 } 2357 2358 eet->pollfd_pos = perf_evlist__add_pollfd(&evlist->core, eet->timerfd, NULL, POLLIN, flags); 2359 if (eet->pollfd_pos < 0) { 2360 err = eet->pollfd_pos; 2361 goto close_timerfd; 2362 } 2363 2364 eet->evlist = evlist; 2365 evlist->eet = eet; 2366 opts->target.initial_delay = eet->times[0].start; 2367 2368 return 0; 2369 2370 close_timerfd: 2371 close(eet->timerfd); 2372 free_eet_times: 2373 zfree(&eet->times); 2374 free_eet: 2375 free(eet); 2376 return err; 2377 } 2378 2379 static int event_enable_timer__set_timer(struct event_enable_timer *eet, int ms) 2380 { 2381 struct itimerspec its = { 2382 .it_value.tv_sec = ms / MSEC_PER_SEC, 2383 .it_value.tv_nsec = (ms % MSEC_PER_SEC) * NSEC_PER_MSEC, 2384 }; 2385 int err = 0; 2386 2387 if (timerfd_settime(eet->timerfd, 0, &its, NULL) < 0) { 2388 err = -errno; 2389 pr_err("timerfd_settime failed: %s\n", strerror(errno)); 2390 } 2391 return err; 2392 } 2393 2394 int event_enable_timer__start(struct event_enable_timer *eet) 2395 { 2396 int ms; 2397 2398 if (!eet) 2399 return 0; 2400 2401 ms = eet->times[0].end - eet->times[0].start; 2402 eet->times_step = 1; 2403 2404 return event_enable_timer__set_timer(eet, ms); 2405 } 2406 2407 int event_enable_timer__process(struct event_enable_timer *eet) 2408 { 2409 struct pollfd *entries; 2410 short revents; 2411 2412 if (!eet) 2413 return 0; 2414 2415 entries = eet->evlist->core.pollfd.entries; 2416 revents = entries[eet->pollfd_pos].revents; 2417 entries[eet->pollfd_pos].revents = 0; 2418 2419 if (revents & POLLIN) { 2420 size_t step = eet->times_step; 2421 size_t pos = step / 2; 2422 2423 if (step & 1) { 2424 evlist__disable_non_dummy(eet->evlist); 2425 pr_info(EVLIST_DISABLED_MSG); 2426 if (pos >= eet->times_cnt - 1) { 2427 /* Disarm timer */ 2428 event_enable_timer__set_timer(eet, 0); 2429 return 1; /* Stop */ 2430 } 2431 } else { 2432 evlist__enable_non_dummy(eet->evlist); 2433 pr_info(EVLIST_ENABLED_MSG); 2434 } 2435 2436 step += 1; 2437 pos = step / 2; 2438 2439 if (pos < eet->times_cnt) { 2440 int *times = (int *)eet->times; /* Accessing 'times' as array of int */ 2441 int ms = times[step] - times[step - 1]; 2442 2443 eet->times_step = step; 2444 return event_enable_timer__set_timer(eet, ms); 2445 } 2446 } 2447 2448 return 0; 2449 } 2450 2451 void event_enable_timer__exit(struct event_enable_timer **ep) 2452 { 2453 if (!ep || !*ep) 2454 return; 2455 zfree(&(*ep)->times); 2456 zfree(ep); 2457 } 2458 2459 struct evsel *evlist__find_evsel(struct evlist *evlist, int idx) 2460 { 2461 struct evsel *evsel; 2462 2463 evlist__for_each_entry(evlist, evsel) { 2464 if (evsel->core.idx == idx) 2465 return evsel; 2466 } 2467 return NULL; 2468 } 2469 2470 void evlist__format_evsels(struct evlist *evlist, struct strbuf *sb, size_t max_length) 2471 { 2472 struct evsel *evsel, *leader = NULL; 2473 bool first = true; 2474 2475 evlist__for_each_entry(evlist, evsel) { 2476 struct evsel *new_leader = evsel__leader(evsel); 2477 2478 if (evsel__is_dummy_event(evsel)) 2479 continue; 2480 2481 if (leader != new_leader && leader && leader->core.nr_members > 1) 2482 strbuf_addch(sb, '}'); 2483 2484 if (!first) 2485 strbuf_addch(sb, ','); 2486 2487 if (sb->len > max_length) { 2488 strbuf_addstr(sb, "..."); 2489 return; 2490 } 2491 if (leader != new_leader && new_leader->core.nr_members > 1) 2492 strbuf_addch(sb, '{'); 2493 2494 strbuf_addstr(sb, evsel__name(evsel)); 2495 first = false; 2496 leader = new_leader; 2497 } 2498 if (leader && leader->core.nr_members > 1) 2499 strbuf_addch(sb, '}'); 2500 } 2501 2502 void evlist__check_mem_load_aux(struct evlist *evlist) 2503 { 2504 struct evsel *leader, *evsel, *pos; 2505 2506 /* 2507 * For some platforms, the 'mem-loads' event is required to use 2508 * together with 'mem-loads-aux' within a group and 'mem-loads-aux' 2509 * must be the group leader. Now we disable this group before reporting 2510 * because 'mem-loads-aux' is just an auxiliary event. It doesn't carry 2511 * any valid memory load information. 2512 */ 2513 evlist__for_each_entry(evlist, evsel) { 2514 leader = evsel__leader(evsel); 2515 if (leader == evsel) 2516 continue; 2517 2518 if (leader->name && strstr(leader->name, "mem-loads-aux")) { 2519 for_each_group_evsel(pos, leader) { 2520 evsel__set_leader(pos, pos); 2521 pos->core.nr_members = 0; 2522 } 2523 } 2524 } 2525 } 2526 2527 /** 2528 * evlist__warn_user_requested_cpus() - Check each evsel against requested CPUs 2529 * and warn if the user CPU list is inapplicable for the event's PMU's 2530 * CPUs. Not core PMUs list a CPU in sysfs, but this may be overwritten by a 2531 * user requested CPU and so any online CPU is applicable. Core PMUs handle 2532 * events on the CPUs in their list and otherwise the event isn't supported. 2533 * @evlist: The list of events being checked. 2534 * @cpu_list: The user provided list of CPUs. 2535 */ 2536 void evlist__warn_user_requested_cpus(struct evlist *evlist, const char *cpu_list) 2537 { 2538 struct perf_cpu_map *user_requested_cpus; 2539 struct evsel *pos; 2540 2541 if (!cpu_list) 2542 return; 2543 2544 user_requested_cpus = perf_cpu_map__new(cpu_list); 2545 if (!user_requested_cpus) 2546 return; 2547 2548 evlist__for_each_entry(evlist, pos) { 2549 struct perf_cpu_map *intersect, *to_test, *online = cpu_map__online(); 2550 const struct perf_pmu *pmu = evsel__find_pmu(pos); 2551 2552 to_test = pmu && pmu->is_core ? pmu->cpus : online; 2553 intersect = perf_cpu_map__intersect(to_test, user_requested_cpus); 2554 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) { 2555 char buf[128]; 2556 2557 cpu_map__snprint(to_test, buf, sizeof(buf)); 2558 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n", 2559 cpu_list, pmu ? pmu->name : "cpu", buf, evsel__name(pos)); 2560 } 2561 perf_cpu_map__put(intersect); 2562 perf_cpu_map__put(online); 2563 } 2564 perf_cpu_map__put(user_requested_cpus); 2565 } 2566 2567 /* Should uniquify be disabled for the evlist? */ 2568 static bool evlist__disable_uniquify(const struct evlist *evlist) 2569 { 2570 struct evsel *counter; 2571 struct perf_pmu *last_pmu = NULL; 2572 bool first = true; 2573 2574 evlist__for_each_entry(evlist, counter) { 2575 /* If PMUs vary then uniquify can be useful. */ 2576 if (!first && counter->pmu != last_pmu) 2577 return false; 2578 first = false; 2579 if (counter->pmu) { 2580 /* Allow uniquify for uncore PMUs. */ 2581 if (!counter->pmu->is_core) 2582 return false; 2583 /* Keep hybrid event names uniquified for clarity. */ 2584 if (perf_pmus__num_core_pmus() > 1) 2585 return false; 2586 } 2587 last_pmu = counter->pmu; 2588 } 2589 return true; 2590 } 2591 2592 static bool evlist__set_needs_uniquify(struct evlist *evlist, const struct perf_stat_config *config) 2593 { 2594 struct evsel *counter; 2595 bool needs_uniquify = false; 2596 2597 if (evlist__disable_uniquify(evlist)) { 2598 evlist__for_each_entry(evlist, counter) 2599 counter->uniquified_name = true; 2600 return false; 2601 } 2602 2603 evlist__for_each_entry(evlist, counter) { 2604 if (evsel__set_needs_uniquify(counter, config)) 2605 needs_uniquify = true; 2606 } 2607 return needs_uniquify; 2608 } 2609 2610 void evlist__uniquify_evsel_names(struct evlist *evlist, const struct perf_stat_config *config) 2611 { 2612 if (evlist__set_needs_uniquify(evlist, config)) { 2613 struct evsel *pos; 2614 2615 evlist__for_each_entry(evlist, pos) 2616 evsel__uniquify_counter(pos); 2617 } 2618 } 2619 2620 bool evlist__has_bpf_output(struct evlist *evlist) 2621 { 2622 struct evsel *evsel; 2623 2624 evlist__for_each_entry(evlist, evsel) { 2625 if (evsel__is_bpf_output(evsel)) 2626 return true; 2627 } 2628 2629 return false; 2630 } 2631 2632 bool evlist__needs_bpf_sb_event(struct evlist *evlist) 2633 { 2634 struct evsel *evsel; 2635 2636 evlist__for_each_entry(evlist, evsel) { 2637 if (evsel__is_dummy_event(evsel)) 2638 continue; 2639 if (!evsel->core.attr.exclude_kernel) 2640 return true; 2641 } 2642 2643 return false; 2644 } 2645