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 /* 9 * Powerpc needs __SANE_USERSPACE_TYPES__ before <linux/types.h> to select 10 * 'int-ll64.h' and avoid compile warnings when printing __u64 with %llu. 11 */ 12 #define __SANE_USERSPACE_TYPES__ 13 14 #include "evsel.h" 15 16 #include <errno.h> 17 #include <inttypes.h> 18 #include <stdlib.h> 19 20 #include <dirent.h> 21 #include <linux/bitops.h> 22 #include <linux/compiler.h> 23 #include <linux/ctype.h> 24 #include <linux/err.h> 25 #include <linux/hw_breakpoint.h> 26 #include <linux/perf_event.h> 27 #include <linux/zalloc.h> 28 #include <sys/ioctl.h> 29 #include <sys/resource.h> 30 #include <sys/syscall.h> 31 #include <sys/types.h> 32 33 #include <api/fs/fs.h> 34 #include <api/fs/tracing_path.h> 35 #include <byteswap.h> 36 #include <internal/lib.h> 37 #include <internal/threadmap.h> 38 #include <internal/xyarray.h> 39 #include <perf/cpumap.h> 40 #include <perf/evsel.h> 41 42 #include "../perf-sys.h" 43 #include "asm/bug.h" 44 #include "bpf-filter.h" 45 #include "bpf_counter.h" 46 #include "callchain.h" 47 #include "cgroup.h" 48 #include "counts.h" 49 #include "debug.h" 50 #include "drm_pmu.h" 51 #include "dwarf-regs.h" 52 #include "env.h" 53 #include "event.h" 54 #include "evlist.h" 55 #include "evsel_config.h" 56 #include "evsel_fprintf.h" 57 #include "hashmap.h" 58 #include "hist.h" 59 #include "hwmon_pmu.h" 60 #include "intel-tpebs.h" 61 #include "memswap.h" 62 #include "off_cpu.h" 63 #include "parse-branch-options.h" 64 #include "perf_regs.h" 65 #include "pmu.h" 66 #include "pmus.h" 67 #include "record.h" 68 #include "rlimit.h" 69 #include "session.h" 70 #include "stat.h" 71 #include "string2.h" 72 #include "target.h" 73 #include "thread_map.h" 74 #include "time-utils.h" 75 #include "tool_pmu.h" 76 #include "tp_pmu.h" 77 #include "trace-event.h" 78 #include "util.h" 79 80 #ifdef HAVE_LIBTRACEEVENT 81 #include <event-parse.h> 82 #endif 83 84 struct perf_missing_features perf_missing_features; 85 86 static clockid_t clockid; 87 88 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused) 89 { 90 return 0; 91 } 92 93 static bool test_attr__enabled(void) 94 { 95 static bool test_attr__enabled; 96 static bool test_attr__enabled_tested; 97 98 if (!test_attr__enabled_tested) { 99 char *dir = getenv("PERF_TEST_ATTR"); 100 101 test_attr__enabled = (dir != NULL); 102 test_attr__enabled_tested = true; 103 } 104 return test_attr__enabled; 105 } 106 107 #define __WRITE_ASS(str, fmt, data) \ 108 do { \ 109 if (fprintf(file, #str "=%"fmt "\n", data) < 0) { \ 110 perror("test attr - failed to write event file"); \ 111 fclose(file); \ 112 return -1; \ 113 } \ 114 } while (0) 115 116 #define WRITE_ASS(field, fmt) __WRITE_ASS(field, fmt, attr->field) 117 118 static int store_event(struct perf_event_attr *attr, pid_t pid, struct perf_cpu cpu, 119 int fd, int group_fd, unsigned long flags) 120 { 121 FILE *file; 122 char path[PATH_MAX]; 123 char *dir = getenv("PERF_TEST_ATTR"); 124 125 snprintf(path, PATH_MAX, "%s/event-%d-%llu-%d", dir, 126 attr->type, attr->config, fd); 127 128 file = fopen(path, "w+"); 129 if (!file) { 130 perror("test attr - failed to open event file"); 131 return -1; 132 } 133 134 if (fprintf(file, "[event-%d-%llu-%d]\n", 135 attr->type, attr->config, fd) < 0) { 136 perror("test attr - failed to write event file"); 137 fclose(file); 138 return -1; 139 } 140 141 /* syscall arguments */ 142 __WRITE_ASS(fd, "d", fd); 143 __WRITE_ASS(group_fd, "d", group_fd); 144 __WRITE_ASS(cpu, "d", cpu.cpu); 145 __WRITE_ASS(pid, "d", pid); 146 __WRITE_ASS(flags, "lu", flags); 147 148 /* struct perf_event_attr */ 149 WRITE_ASS(type, PRIu32); 150 WRITE_ASS(size, PRIu32); 151 WRITE_ASS(config, "llu"); 152 WRITE_ASS(sample_period, "llu"); 153 WRITE_ASS(sample_type, "llu"); 154 WRITE_ASS(read_format, "llu"); 155 WRITE_ASS(disabled, "d"); 156 WRITE_ASS(inherit, "d"); 157 WRITE_ASS(pinned, "d"); 158 WRITE_ASS(exclusive, "d"); 159 WRITE_ASS(exclude_user, "d"); 160 WRITE_ASS(exclude_kernel, "d"); 161 WRITE_ASS(exclude_hv, "d"); 162 WRITE_ASS(exclude_idle, "d"); 163 WRITE_ASS(mmap, "d"); 164 WRITE_ASS(comm, "d"); 165 WRITE_ASS(freq, "d"); 166 WRITE_ASS(inherit_stat, "d"); 167 WRITE_ASS(enable_on_exec, "d"); 168 WRITE_ASS(task, "d"); 169 WRITE_ASS(watermark, "d"); 170 WRITE_ASS(precise_ip, "d"); 171 WRITE_ASS(mmap_data, "d"); 172 WRITE_ASS(sample_id_all, "d"); 173 WRITE_ASS(exclude_host, "d"); 174 WRITE_ASS(exclude_guest, "d"); 175 WRITE_ASS(exclude_callchain_kernel, "d"); 176 WRITE_ASS(exclude_callchain_user, "d"); 177 WRITE_ASS(mmap2, "d"); 178 WRITE_ASS(comm_exec, "d"); 179 WRITE_ASS(context_switch, "d"); 180 WRITE_ASS(write_backward, "d"); 181 WRITE_ASS(namespaces, "d"); 182 WRITE_ASS(use_clockid, "d"); 183 WRITE_ASS(wakeup_events, PRIu32); 184 WRITE_ASS(bp_type, PRIu32); 185 WRITE_ASS(config1, "llu"); 186 WRITE_ASS(config2, "llu"); 187 WRITE_ASS(branch_sample_type, "llu"); 188 WRITE_ASS(sample_regs_user, "llu"); 189 WRITE_ASS(sample_stack_user, PRIu32); 190 191 fclose(file); 192 return 0; 193 } 194 195 #undef __WRITE_ASS 196 #undef WRITE_ASS 197 198 static void test_attr__open(struct perf_event_attr *attr, pid_t pid, struct perf_cpu cpu, 199 int fd, int group_fd, unsigned long flags) 200 { 201 int errno_saved = errno; 202 203 if ((fd != -1) && store_event(attr, pid, cpu, fd, group_fd, flags)) { 204 pr_err("test attr FAILED"); 205 exit(128); 206 } 207 208 errno = errno_saved; 209 } 210 211 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused) 212 { 213 } 214 215 static struct { 216 size_t size; 217 int (*init)(struct evsel *evsel); 218 void (*fini)(struct evsel *evsel); 219 } perf_evsel__object = { 220 .size = sizeof(struct evsel), 221 .init = evsel__no_extra_init, 222 .fini = evsel__no_extra_fini, 223 }; 224 225 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel), 226 void (*fini)(struct evsel *evsel)) 227 { 228 229 if (object_size == 0) 230 goto set_methods; 231 232 if (perf_evsel__object.size > object_size) 233 return -EINVAL; 234 235 perf_evsel__object.size = object_size; 236 237 set_methods: 238 if (init != NULL) 239 perf_evsel__object.init = init; 240 241 if (fini != NULL) 242 perf_evsel__object.fini = fini; 243 244 return 0; 245 } 246 247 const char *evsel__pmu_name(const struct evsel *evsel) 248 { 249 struct perf_pmu *pmu = evsel__find_pmu(evsel); 250 251 if (pmu) 252 return pmu->name; 253 254 return event_type(evsel->core.attr.type); 255 } 256 257 enum evsel_probe_type { 258 PROBE__UNKNOWN = 0, 259 PROBE__NOPE = 1, 260 PROBE__KPROBE = 2, 261 PROBE__UPROBE = 3, 262 /* 263 * Ftrace-based dynamic probes (kprobes/uprobes/fprobes created via 264 * tracefs) report PMU "tracepoint", not "kprobe"/"uprobe". Detect 265 * them by the __probe_ip field that the kernel adds to all dynamic 266 * probe formats. 267 */ 268 PROBE__FTRACE = 4, 269 }; 270 271 static void evsel__resolve_probe_type(struct evsel *evsel) 272 { 273 const char *name = evsel__pmu_name(evsel); 274 275 if (!strcmp(name, "kprobe")) 276 evsel->probe_type = PROBE__KPROBE; 277 else if (!strcmp(name, "uprobe")) 278 evsel->probe_type = PROBE__UPROBE; 279 else if (!strcmp(name, "tracepoint") && evsel__field(evsel, "__probe_ip")) 280 evsel->probe_type = PROBE__FTRACE; 281 else 282 evsel->probe_type = PROBE__NOPE; 283 } 284 285 bool evsel__is_probe(struct evsel *evsel) 286 { 287 if (evsel->probe_type == PROBE__UNKNOWN) 288 evsel__resolve_probe_type(evsel); 289 290 return evsel->probe_type > PROBE__NOPE; 291 } 292 293 bool evsel__is_kprobe(struct evsel *evsel) 294 { 295 if (evsel->probe_type == PROBE__UNKNOWN) 296 evsel__resolve_probe_type(evsel); 297 298 return evsel->probe_type == PROBE__KPROBE; 299 } 300 301 bool evsel__is_uprobe(struct evsel *evsel) 302 { 303 if (evsel->probe_type == PROBE__UNKNOWN) 304 evsel__resolve_probe_type(evsel); 305 306 return evsel->probe_type == PROBE__UPROBE; 307 } 308 309 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 310 311 int __evsel__sample_size(u64 sample_type) 312 { 313 u64 mask = sample_type & PERF_SAMPLE_MASK; 314 int size = 0; 315 int i; 316 317 for (i = 0; i < 64; i++) { 318 if (mask & (1ULL << i)) 319 size++; 320 } 321 322 size *= sizeof(u64); 323 324 return size; 325 } 326 327 /** 328 * __perf_evsel__calc_id_pos - calculate id_pos. 329 * @sample_type: sample type 330 * 331 * This function returns the position of the event id (PERF_SAMPLE_ID or 332 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct 333 * perf_record_sample. 334 */ 335 static int __perf_evsel__calc_id_pos(u64 sample_type) 336 { 337 int idx = 0; 338 339 if (sample_type & PERF_SAMPLE_IDENTIFIER) 340 return 0; 341 342 if (!(sample_type & PERF_SAMPLE_ID)) 343 return -1; 344 345 if (sample_type & PERF_SAMPLE_IP) 346 idx += 1; 347 348 if (sample_type & PERF_SAMPLE_TID) 349 idx += 1; 350 351 if (sample_type & PERF_SAMPLE_TIME) 352 idx += 1; 353 354 if (sample_type & PERF_SAMPLE_ADDR) 355 idx += 1; 356 357 return idx; 358 } 359 360 /** 361 * __perf_evsel__calc_is_pos - calculate is_pos. 362 * @sample_type: sample type 363 * 364 * This function returns the position (counting backwards) of the event id 365 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if 366 * sample_id_all is used there is an id sample appended to non-sample events. 367 */ 368 static int __perf_evsel__calc_is_pos(u64 sample_type) 369 { 370 int idx = 1; 371 372 if (sample_type & PERF_SAMPLE_IDENTIFIER) 373 return 1; 374 375 if (!(sample_type & PERF_SAMPLE_ID)) 376 return -1; 377 378 if (sample_type & PERF_SAMPLE_CPU) 379 idx += 1; 380 381 if (sample_type & PERF_SAMPLE_STREAM_ID) 382 idx += 1; 383 384 return idx; 385 } 386 387 void evsel__calc_id_pos(struct evsel *evsel) 388 { 389 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type); 390 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type); 391 } 392 393 void __evsel__set_sample_bit(struct evsel *evsel, 394 enum perf_event_sample_format bit) 395 { 396 if (!(evsel->core.attr.sample_type & bit)) { 397 evsel->core.attr.sample_type |= bit; 398 evsel->sample_size += sizeof(u64); 399 evsel__calc_id_pos(evsel); 400 } 401 } 402 403 void __evsel__reset_sample_bit(struct evsel *evsel, 404 enum perf_event_sample_format bit) 405 { 406 if (evsel->core.attr.sample_type & bit) { 407 evsel->core.attr.sample_type &= ~bit; 408 evsel->sample_size -= sizeof(u64); 409 evsel__calc_id_pos(evsel); 410 } 411 } 412 413 void evsel__set_sample_id(struct evsel *evsel, 414 bool can_sample_identifier) 415 { 416 if (can_sample_identifier) { 417 evsel__reset_sample_bit(evsel, ID); 418 evsel__set_sample_bit(evsel, IDENTIFIER); 419 } else { 420 evsel__set_sample_bit(evsel, ID); 421 } 422 evsel->core.attr.read_format |= PERF_FORMAT_ID; 423 } 424 425 /** 426 * evsel__is_function_event - Return whether given evsel is a function 427 * trace event 428 * 429 * @evsel - evsel selector to be tested 430 * 431 * Return %true if event is function trace event 432 */ 433 bool evsel__is_function_event(struct evsel *evsel) 434 { 435 #define FUNCTION_EVENT "ftrace:function" 436 437 return evsel->name && 438 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT)); 439 440 #undef FUNCTION_EVENT 441 } 442 443 void evsel__init(struct evsel *evsel, 444 struct perf_event_attr *attr, int idx) 445 { 446 perf_evsel__init(&evsel->core, attr, idx); 447 evsel->tracking = !idx; 448 evsel->unit = strdup(""); 449 evsel->scale = 1.0; 450 evsel->max_events = ULONG_MAX; 451 evsel->evlist = NULL; 452 evsel->bpf_obj = NULL; 453 evsel->bpf_fd = -1; 454 INIT_LIST_HEAD(&evsel->config_terms); 455 INIT_LIST_HEAD(&evsel->bpf_counter_list); 456 INIT_LIST_HEAD(&evsel->bpf_filters); 457 perf_evsel__object.init(evsel); 458 evsel->sample_size = __evsel__sample_size(attr->sample_type); 459 evsel__calc_id_pos(evsel); 460 evsel->cmdline_group_boundary = false; 461 evsel->per_pkg_mask = NULL; 462 evsel->collect_stat = false; 463 evsel->group_pmu_name = NULL; 464 evsel->skippable = false; 465 evsel->supported = true; 466 evsel->alternate_hw_config = PERF_COUNT_HW_MAX; 467 evsel->script_output_type = -1; // FIXME: OUTPUT_TYPE_UNSET, see builtin-script.c 468 evsel->probe_type = PROBE__UNKNOWN; 469 } 470 471 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx) 472 { 473 struct evsel *evsel = zalloc(perf_evsel__object.size); 474 475 if (!evsel) 476 return NULL; 477 evsel__init(evsel, attr, idx); 478 479 if (evsel__is_bpf_output(evsel) && !attr->sample_type) { 480 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 481 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 482 evsel->core.attr.sample_period = 1; 483 } 484 485 if (evsel__is_clock(evsel)) { 486 free((char *)evsel->unit); 487 evsel->unit = strdup("msec"); 488 evsel->scale = 1e-6; 489 } 490 491 return evsel; 492 } 493 494 int copy_config_terms(struct list_head *dst, struct list_head *src) 495 { 496 struct evsel_config_term *pos, *tmp; 497 498 list_for_each_entry(pos, src, list) { 499 tmp = malloc(sizeof(*tmp)); 500 if (tmp == NULL) 501 return -ENOMEM; 502 503 *tmp = *pos; 504 if (tmp->free_str) { 505 tmp->val.str = strdup(pos->val.str); 506 if (tmp->val.str == NULL) { 507 free(tmp); 508 return -ENOMEM; 509 } 510 } 511 list_add_tail(&tmp->list, dst); 512 } 513 return 0; 514 } 515 516 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src) 517 { 518 return copy_config_terms(&dst->config_terms, &src->config_terms); 519 } 520 521 /** 522 * evsel__clone - create a new evsel copied from @orig 523 * @orig: original evsel 524 * 525 * The assumption is that @orig is not configured nor opened yet. 526 * So we only care about the attributes that can be set while it's parsed. 527 */ 528 struct evsel *evsel__clone(struct evsel *dest, struct evsel *orig) 529 { 530 struct evsel *evsel; 531 532 BUG_ON(orig->core.fd); 533 BUG_ON(orig->counts); 534 BUG_ON(orig->priv); 535 BUG_ON(orig->per_pkg_mask); 536 537 /* cannot handle BPF objects for now */ 538 if (orig->bpf_obj) 539 return NULL; 540 541 if (dest) 542 evsel = dest; 543 else 544 evsel = evsel__new(&orig->core.attr); 545 546 if (evsel == NULL) 547 return NULL; 548 549 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus); 550 evsel->core.pmu_cpus = perf_cpu_map__get(orig->core.pmu_cpus); 551 evsel->core.threads = perf_thread_map__get(orig->core.threads); 552 evsel->core.nr_members = orig->core.nr_members; 553 evsel->core.system_wide = orig->core.system_wide; 554 evsel->core.requires_cpu = orig->core.requires_cpu; 555 evsel->core.is_pmu_core = orig->core.is_pmu_core; 556 557 if (orig->name) { 558 evsel->name = strdup(orig->name); 559 if (evsel->name == NULL) 560 goto out_err; 561 } 562 if (orig->group_name) { 563 evsel->group_name = strdup(orig->group_name); 564 if (evsel->group_name == NULL) 565 goto out_err; 566 } 567 if (orig->group_pmu_name) { 568 evsel->group_pmu_name = strdup(orig->group_pmu_name); 569 if (evsel->group_pmu_name == NULL) 570 goto out_err; 571 } 572 if (orig->filter) { 573 evsel->filter = strdup(orig->filter); 574 if (evsel->filter == NULL) 575 goto out_err; 576 } 577 if (orig->metric_id) { 578 evsel->metric_id = strdup(orig->metric_id); 579 if (evsel->metric_id == NULL) 580 goto out_err; 581 } 582 evsel->cgrp = cgroup__get(orig->cgrp); 583 #ifdef HAVE_LIBTRACEEVENT 584 if (orig->tp_sys) { 585 evsel->tp_sys = strdup(orig->tp_sys); 586 if (evsel->tp_sys == NULL) 587 goto out_err; 588 } 589 if (orig->tp_name) { 590 evsel->tp_name = strdup(orig->tp_name); 591 if (evsel->tp_name == NULL) 592 goto out_err; 593 } 594 evsel->tp_format = orig->tp_format; 595 #endif 596 evsel->handler = orig->handler; 597 evsel->core.leader = orig->core.leader; 598 evsel->metric_leader = orig->metric_leader; 599 600 evsel->max_events = orig->max_events; 601 zfree(&evsel->unit); 602 if (orig->unit) { 603 evsel->unit = strdup(orig->unit); 604 if (evsel->unit == NULL) 605 goto out_err; 606 } 607 evsel->scale = orig->scale; 608 evsel->snapshot = orig->snapshot; 609 evsel->per_pkg = orig->per_pkg; 610 evsel->percore = orig->percore; 611 evsel->precise_max = orig->precise_max; 612 evsel->is_libpfm_event = orig->is_libpfm_event; 613 614 evsel->exclude_GH = orig->exclude_GH; 615 evsel->sample_read = orig->sample_read; 616 evsel->collect_stat = orig->collect_stat; 617 evsel->weak_group = orig->weak_group; 618 evsel->use_config_name = orig->use_config_name; 619 evsel->pmu = orig->pmu; 620 evsel->first_wildcard_match = orig->first_wildcard_match; 621 622 if (evsel__copy_config_terms(evsel, orig) < 0) 623 goto out_err; 624 625 evsel->alternate_hw_config = orig->alternate_hw_config; 626 627 return evsel; 628 629 out_err: 630 evsel__delete(evsel); 631 return NULL; 632 } 633 634 /* 635 * Returns pointer with encoded error via <linux/err.h> interface. 636 */ 637 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx, bool format) 638 { 639 struct perf_event_attr attr = { 640 .type = PERF_TYPE_TRACEPOINT, 641 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 642 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 643 }; 644 struct evsel *evsel = zalloc(perf_evsel__object.size); 645 int err = -ENOMEM, id = -1; 646 647 if (evsel == NULL) 648 goto out_err; 649 650 651 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0) 652 goto out_free; 653 654 #ifdef HAVE_LIBTRACEEVENT 655 evsel->tp_sys = strdup(sys); 656 if (!evsel->tp_sys) 657 goto out_free; 658 659 evsel->tp_name = strdup(name); 660 if (!evsel->tp_name) 661 goto out_free; 662 #endif 663 664 event_attr_init(&attr); 665 666 if (format) { 667 id = tp_pmu__id(sys, name); 668 if (id < 0) { 669 err = id; 670 goto out_free; 671 } 672 } 673 attr.config = (__u64)id; 674 attr.sample_period = 1; 675 evsel__init(evsel, &attr, idx); 676 return evsel; 677 678 out_free: 679 zfree(&evsel->name); 680 #ifdef HAVE_LIBTRACEEVENT 681 zfree(&evsel->tp_sys); 682 zfree(&evsel->tp_name); 683 #endif 684 free(evsel); 685 out_err: 686 return ERR_PTR(err); 687 } 688 689 #ifdef HAVE_LIBTRACEEVENT 690 struct tep_event *evsel__tp_format(struct evsel *evsel) 691 { 692 struct tep_event *tp_format = evsel->tp_format; 693 694 if (tp_format) 695 return tp_format; 696 697 if (evsel->core.attr.type != PERF_TYPE_TRACEPOINT) 698 return NULL; 699 700 if (!evsel->tp_sys) 701 tp_format = trace_event__tp_format_id(evsel->core.attr.config); 702 else 703 tp_format = trace_event__tp_format(evsel->tp_sys, evsel->tp_name); 704 705 if (IS_ERR(tp_format)) { 706 int err = -PTR_ERR(evsel->tp_format); 707 708 errno = err; 709 pr_err("Error getting tracepoint format '%s': %m\n", 710 evsel__name(evsel)); 711 return NULL; 712 } 713 evsel->tp_format = tp_format; 714 return evsel->tp_format; 715 } 716 #endif 717 718 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = { 719 "cycles", 720 "instructions", 721 "cache-references", 722 "cache-misses", 723 "branches", 724 "branch-misses", 725 "bus-cycles", 726 "stalled-cycles-frontend", 727 "stalled-cycles-backend", 728 "ref-cycles", 729 }; 730 731 char *evsel__bpf_counter_events; 732 733 bool evsel__match_bpf_counter_events(const char *name) 734 { 735 int name_len; 736 bool match; 737 char *ptr; 738 739 if (!evsel__bpf_counter_events) 740 return false; 741 742 ptr = strstr(evsel__bpf_counter_events, name); 743 name_len = strlen(name); 744 745 /* check name matches a full token in evsel__bpf_counter_events */ 746 match = (ptr != NULL) && 747 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) && 748 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0')); 749 750 return match; 751 } 752 753 static const char *__evsel__hw_name(u64 config) 754 { 755 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config]) 756 return evsel__hw_names[config]; 757 758 return "unknown-hardware"; 759 } 760 761 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size) 762 { 763 int colon = 0, r = 0; 764 struct perf_event_attr *attr = &evsel->core.attr; 765 766 #define MOD_PRINT(context, mod) do { \ 767 if (!attr->exclude_##context) { \ 768 if (!colon) colon = ++r; \ 769 r += scnprintf(bf + r, size - r, "%c", mod); \ 770 } } while(0) 771 772 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) { 773 MOD_PRINT(kernel, 'k'); 774 MOD_PRINT(user, 'u'); 775 MOD_PRINT(hv, 'h'); 776 } 777 778 if (attr->precise_ip) { 779 if (!colon) 780 colon = ++r; 781 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp"); 782 } 783 784 if (attr->exclude_host || attr->exclude_guest) { 785 MOD_PRINT(host, 'H'); 786 MOD_PRINT(guest, 'G'); 787 } 788 #undef MOD_PRINT 789 if (colon) 790 bf[colon - 1] = ':'; 791 return r; 792 } 793 794 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 795 { 796 return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config)); 797 } 798 799 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 800 { 801 int r = arch_evsel__hw_name(evsel, bf, size); 802 return r + evsel__add_modifiers(evsel, bf + r, size - r); 803 } 804 805 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = { 806 "cpu-clock", 807 "task-clock", 808 "page-faults", 809 "context-switches", 810 "cpu-migrations", 811 "minor-faults", 812 "major-faults", 813 "alignment-faults", 814 "emulation-faults", 815 "dummy", 816 }; 817 818 static const char *__evsel__sw_name(u64 config) 819 { 820 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config]) 821 return evsel__sw_names[config]; 822 return "unknown-software"; 823 } 824 825 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size) 826 { 827 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config)); 828 return r + evsel__add_modifiers(evsel, bf + r, size - r); 829 } 830 831 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type) 832 { 833 int r; 834 835 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr); 836 837 if (type & HW_BREAKPOINT_R) 838 r += scnprintf(bf + r, size - r, "r"); 839 840 if (type & HW_BREAKPOINT_W) 841 r += scnprintf(bf + r, size - r, "w"); 842 843 if (type & HW_BREAKPOINT_X) 844 r += scnprintf(bf + r, size - r, "x"); 845 846 return r; 847 } 848 849 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size) 850 { 851 struct perf_event_attr *attr = &evsel->core.attr; 852 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type); 853 return r + evsel__add_modifiers(evsel, bf + r, size - r); 854 } 855 856 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = { 857 { "L1-dcache", "l1-d", "l1d", "L1-data", }, 858 { "L1-icache", "l1-i", "l1i", "L1-instruction", }, 859 { "LLC", "L2", }, 860 { "dTLB", "d-tlb", "Data-TLB", }, 861 { "iTLB", "i-tlb", "Instruction-TLB", }, 862 { "branch", "branches", "bpu", "btb", "bpc", }, 863 { "node", }, 864 }; 865 866 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = { 867 { "load", "loads", "read", }, 868 { "store", "stores", "write", }, 869 { "prefetch", "prefetches", "speculative-read", "speculative-load", }, 870 }; 871 872 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = { 873 { "refs", "Reference", "ops", "access", }, 874 { "misses", "miss", }, 875 }; 876 877 #define C(x) PERF_COUNT_HW_CACHE_##x 878 #define CACHE_READ (1 << C(OP_READ)) 879 #define CACHE_WRITE (1 << C(OP_WRITE)) 880 #define CACHE_PREFETCH (1 << C(OP_PREFETCH)) 881 #define COP(x) (1 << x) 882 883 /* 884 * cache operation stat 885 * L1I : Read and prefetch only 886 * ITLB and BPU : Read-only 887 */ 888 static const unsigned long evsel__hw_cache_stat[C(MAX)] = { 889 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 890 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH), 891 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 892 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 893 [C(ITLB)] = (CACHE_READ), 894 [C(BPU)] = (CACHE_READ), 895 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 896 }; 897 898 bool evsel__is_cache_op_valid(u8 type, u8 op) 899 { 900 if (evsel__hw_cache_stat[type] & COP(op)) 901 return true; /* valid */ 902 else 903 return false; /* invalid */ 904 } 905 906 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size) 907 { 908 if (result) { 909 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0], 910 evsel__hw_cache_op[op][0], 911 evsel__hw_cache_result[result][0]); 912 } 913 914 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0], 915 evsel__hw_cache_op[op][1]); 916 } 917 918 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size) 919 { 920 u8 op, result, type = (config >> 0) & 0xff; 921 const char *err = "unknown-ext-hardware-cache-type"; 922 923 if (type >= PERF_COUNT_HW_CACHE_MAX) 924 goto out_err; 925 926 op = (config >> 8) & 0xff; 927 err = "unknown-ext-hardware-cache-op"; 928 if (op >= PERF_COUNT_HW_CACHE_OP_MAX) 929 goto out_err; 930 931 result = (config >> 16) & 0xff; 932 err = "unknown-ext-hardware-cache-result"; 933 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 934 goto out_err; 935 936 err = "invalid-cache"; 937 if (!evsel__is_cache_op_valid(type, op)) 938 goto out_err; 939 940 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size); 941 out_err: 942 return scnprintf(bf, size, "%s", err); 943 } 944 945 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size) 946 { 947 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size); 948 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret); 949 } 950 951 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size) 952 { 953 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config); 954 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret); 955 } 956 957 const char *evsel__name(struct evsel *evsel) 958 { 959 char bf[128]; 960 961 if (!evsel) 962 goto out_unknown; 963 964 if (evsel->name) 965 return evsel->name; 966 967 switch (evsel->core.attr.type) { 968 case PERF_TYPE_RAW: 969 evsel__raw_name(evsel, bf, sizeof(bf)); 970 break; 971 972 case PERF_TYPE_HARDWARE: 973 evsel__hw_name(evsel, bf, sizeof(bf)); 974 break; 975 976 case PERF_TYPE_HW_CACHE: 977 evsel__hw_cache_name(evsel, bf, sizeof(bf)); 978 break; 979 980 case PERF_TYPE_SOFTWARE: 981 evsel__sw_name(evsel, bf, sizeof(bf)); 982 break; 983 984 case PERF_TYPE_TRACEPOINT: 985 scnprintf(bf, sizeof(bf), "unknown tracepoint id=%#"PRIx64, 986 evsel->core.attr.config); 987 break; 988 989 case PERF_TYPE_BREAKPOINT: 990 evsel__bp_name(evsel, bf, sizeof(bf)); 991 break; 992 993 case PERF_PMU_TYPE_TOOL: 994 scnprintf(bf, sizeof(bf), "%s", evsel__tool_pmu_event_name(evsel)); 995 break; 996 997 default: 998 scnprintf(bf, sizeof(bf), "unknown event PMU=%d config=%#"PRIx64, 999 evsel->core.attr.type, evsel->core.attr.config); 1000 break; 1001 } 1002 1003 evsel->name = strdup(bf); 1004 1005 if (evsel->name) 1006 return evsel->name; 1007 out_unknown: 1008 return "unknown"; 1009 } 1010 1011 bool evsel__name_is(struct evsel *evsel, const char *name) 1012 { 1013 return !strcmp(evsel__name(evsel), name); 1014 } 1015 1016 const char *evsel__metric_id(const struct evsel *evsel) 1017 { 1018 if (evsel->metric_id) 1019 return evsel->metric_id; 1020 1021 if (evsel__is_tool(evsel)) 1022 return evsel__tool_pmu_event_name(evsel); 1023 1024 return "unknown"; 1025 } 1026 1027 const char *evsel__group_name(struct evsel *evsel) 1028 { 1029 return evsel->group_name ?: "anon group"; 1030 } 1031 1032 /* 1033 * Returns the group details for the specified leader, 1034 * with following rules. 1035 * 1036 * For record -e '{cycles,instructions}' 1037 * 'anon group { cycles:u, instructions:u }' 1038 * 1039 * For record -e 'cycles,instructions' and report --group 1040 * 'cycles:u, instructions:u' 1041 */ 1042 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size) 1043 { 1044 int ret = 0; 1045 bool first = true; 1046 struct evsel *pos; 1047 const char *group_name = evsel__group_name(evsel); 1048 1049 if (!evsel->forced_leader) 1050 ret = scnprintf(buf, size, "%s { ", group_name); 1051 1052 for_each_group_evsel(pos, evsel) { 1053 if (symbol_conf.skip_empty && 1054 evsel__hists(pos)->stats.nr_samples == 0) 1055 continue; 1056 1057 ret += scnprintf(buf + ret, size - ret, "%s%s", 1058 first ? "" : ", ", evsel__name(pos)); 1059 first = false; 1060 } 1061 1062 if (!evsel->forced_leader) 1063 ret += scnprintf(buf + ret, size - ret, " }"); 1064 1065 return ret; 1066 } 1067 1068 uint16_t evsel__e_machine(struct evsel *evsel, uint32_t *e_flags) 1069 { 1070 struct perf_session *session = evsel__session(evsel); 1071 1072 return perf_session__e_machine(session, e_flags); 1073 } 1074 1075 static void __evsel__config_callchain(struct evsel *evsel, const struct record_opts *opts, 1076 const struct callchain_param *param) 1077 { 1078 bool function = evsel__is_function_event(evsel); 1079 struct perf_event_attr *attr = &evsel->core.attr; 1080 1081 if (EM_HOST == EM_S390 && param->record_mode == CALLCHAIN_FP) { 1082 pr_warning_once( 1083 "Framepointer unwinding lacks kernel support. Use '--call-graph dwarf'\n"); 1084 } 1085 1086 evsel__set_sample_bit(evsel, CALLCHAIN); 1087 1088 attr->sample_max_stack = param->max_stack; 1089 1090 if (opts->kernel_callchains) 1091 attr->exclude_callchain_user = 1; 1092 if (opts->user_callchains) 1093 attr->exclude_callchain_kernel = 1; 1094 if (param->record_mode == CALLCHAIN_LBR) { 1095 if (!opts->branch_stack) { 1096 if (attr->exclude_user) { 1097 pr_warning("LBR callstack option is only available " 1098 "to get user callchain information. " 1099 "Falling back to framepointers.\n"); 1100 } else { 1101 evsel__set_sample_bit(evsel, BRANCH_STACK); 1102 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER | 1103 PERF_SAMPLE_BRANCH_CALL_STACK | 1104 PERF_SAMPLE_BRANCH_NO_CYCLES | 1105 PERF_SAMPLE_BRANCH_NO_FLAGS | 1106 PERF_SAMPLE_BRANCH_HW_INDEX; 1107 } 1108 } else 1109 pr_warning("Cannot use LBR callstack with branch stack. " 1110 "Falling back to framepointers.\n"); 1111 } 1112 1113 if (param->record_mode == CALLCHAIN_DWARF) { 1114 if (!function) { 1115 uint16_t e_machine = evsel__e_machine(evsel, /*e_flags=*/NULL); 1116 1117 evsel__set_sample_bit(evsel, REGS_USER); 1118 evsel__set_sample_bit(evsel, STACK_USER); 1119 if (opts->sample_user_regs && 1120 DWARF_MINIMAL_REGS(e_machine) != perf_user_reg_mask(EM_HOST)) { 1121 attr->sample_regs_user |= DWARF_MINIMAL_REGS(e_machine); 1122 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, " 1123 "specifying a subset with --user-regs may render DWARF unwinding unreliable, " 1124 "so the minimal registers set (IP, SP) is explicitly forced.\n"); 1125 } else { 1126 attr->sample_regs_user |= perf_user_reg_mask(EM_HOST); 1127 } 1128 attr->sample_stack_user = param->dump_size; 1129 attr->exclude_callchain_user = 1; 1130 } else { 1131 pr_info("Cannot use DWARF unwind for function trace event," 1132 " falling back to framepointers.\n"); 1133 } 1134 } 1135 1136 if (function) { 1137 pr_info("Disabling user space callchains for function trace event.\n"); 1138 attr->exclude_callchain_user = 1; 1139 } 1140 1141 if (param->defer && !attr->exclude_callchain_user) 1142 attr->defer_callchain = 1; 1143 } 1144 1145 void evsel__config_callchain(struct evsel *evsel, const struct record_opts *opts, 1146 const struct callchain_param *param) 1147 { 1148 if (param->enabled) 1149 return __evsel__config_callchain(evsel, opts, param); 1150 } 1151 1152 static void evsel__reset_callgraph(struct evsel *evsel, const struct callchain_param *param) 1153 { 1154 struct perf_event_attr *attr = &evsel->core.attr; 1155 1156 evsel__reset_sample_bit(evsel, CALLCHAIN); 1157 if (param->record_mode == CALLCHAIN_LBR) { 1158 evsel__reset_sample_bit(evsel, BRANCH_STACK); 1159 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER | 1160 PERF_SAMPLE_BRANCH_CALL_STACK | 1161 PERF_SAMPLE_BRANCH_HW_INDEX); 1162 } 1163 if (param->record_mode == CALLCHAIN_DWARF) { 1164 evsel__reset_sample_bit(evsel, REGS_USER); 1165 evsel__reset_sample_bit(evsel, STACK_USER); 1166 } 1167 } 1168 1169 static void evsel__apply_ratio_to_prev(struct evsel *evsel, 1170 struct perf_event_attr *attr, 1171 const struct record_opts *opts, 1172 const char *buf) 1173 { 1174 struct perf_event_attr *prev_attr = NULL; 1175 struct evsel *evsel_prev = NULL; 1176 u64 type = evsel->core.attr.sample_type; 1177 u64 prev_type = 0; 1178 double rtp; 1179 1180 rtp = strtod(buf, NULL); 1181 if (rtp <= 0) { 1182 pr_err("Invalid ratio-to-prev value %lf\n", rtp); 1183 return; 1184 } 1185 if (evsel == evsel__leader(evsel)) { 1186 pr_err("Invalid use of ratio-to-prev term without preceding element in group\n"); 1187 return; 1188 } 1189 if (!evsel->pmu->is_core) { 1190 pr_err("Event using ratio-to-prev term must have a core PMU\n"); 1191 return; 1192 } 1193 1194 evsel_prev = evsel__prev(evsel); 1195 if (!evsel_prev) { 1196 pr_err("Previous event does not exist.\n"); 1197 return; 1198 } 1199 1200 if (evsel_prev->pmu->type != evsel->pmu->type) { 1201 pr_err("Compared events (\"%s\", \"%s\") must have same PMU\n", 1202 evsel->name, evsel_prev->name); 1203 return; 1204 } 1205 1206 prev_attr = &evsel_prev->core.attr; 1207 prev_type = evsel_prev->core.attr.sample_type; 1208 1209 if (!(prev_type & PERF_SAMPLE_PERIOD)) { 1210 attr->sample_period = prev_attr->sample_period * rtp; 1211 attr->freq = 0; 1212 evsel__reset_sample_bit(evsel, PERIOD); 1213 } else if (!(type & PERF_SAMPLE_PERIOD)) { 1214 prev_attr->sample_period = attr->sample_period / rtp; 1215 prev_attr->freq = 0; 1216 evsel__reset_sample_bit(evsel_prev, PERIOD); 1217 } else { 1218 if (opts->user_interval != ULLONG_MAX) { 1219 prev_attr->sample_period = opts->user_interval; 1220 attr->sample_period = prev_attr->sample_period * rtp; 1221 prev_attr->freq = 0; 1222 attr->freq = 0; 1223 evsel__reset_sample_bit(evsel_prev, PERIOD); 1224 evsel__reset_sample_bit(evsel, PERIOD); 1225 } else { 1226 pr_err("Event period term or count (-c) must be set when using ratio-to-prev term.\n"); 1227 return; 1228 } 1229 } 1230 1231 arch_evsel__apply_ratio_to_prev(evsel, attr); 1232 } 1233 1234 static void evsel__apply_config_terms(struct evsel *evsel, 1235 const struct record_opts *opts, bool track) 1236 { 1237 struct evsel_config_term *term; 1238 struct list_head *config_terms = &evsel->config_terms; 1239 struct perf_event_attr *attr = &evsel->core.attr; 1240 /* callgraph default */ 1241 struct callchain_param param = { 1242 .record_mode = callchain_param.record_mode, 1243 }; 1244 u32 dump_size = 0; 1245 int max_stack = 0; 1246 const char *callgraph_buf = NULL; 1247 const char *rtp_buf = NULL; 1248 1249 list_for_each_entry(term, config_terms, list) { 1250 switch (term->type) { 1251 case EVSEL__CONFIG_TERM_PERIOD: 1252 if (!(term->weak && opts->user_interval != ULLONG_MAX)) { 1253 attr->sample_period = term->val.period; 1254 attr->freq = 0; 1255 evsel__reset_sample_bit(evsel, PERIOD); 1256 } 1257 break; 1258 case EVSEL__CONFIG_TERM_FREQ: 1259 if (!(term->weak && opts->user_freq != UINT_MAX)) { 1260 attr->sample_freq = term->val.freq; 1261 attr->freq = 1; 1262 evsel__set_sample_bit(evsel, PERIOD); 1263 } 1264 break; 1265 case EVSEL__CONFIG_TERM_TIME: 1266 if (term->val.time) 1267 evsel__set_sample_bit(evsel, TIME); 1268 else 1269 evsel__reset_sample_bit(evsel, TIME); 1270 break; 1271 case EVSEL__CONFIG_TERM_CALLGRAPH: 1272 callgraph_buf = term->val.str; 1273 break; 1274 case EVSEL__CONFIG_TERM_BRANCH: 1275 if (term->val.str && strcmp(term->val.str, "no")) { 1276 evsel__set_sample_bit(evsel, BRANCH_STACK); 1277 parse_branch_str(term->val.str, 1278 &attr->branch_sample_type); 1279 } else 1280 evsel__reset_sample_bit(evsel, BRANCH_STACK); 1281 break; 1282 case EVSEL__CONFIG_TERM_STACK_USER: 1283 dump_size = term->val.stack_user; 1284 break; 1285 case EVSEL__CONFIG_TERM_MAX_STACK: 1286 max_stack = term->val.max_stack; 1287 break; 1288 case EVSEL__CONFIG_TERM_MAX_EVENTS: 1289 evsel->max_events = term->val.max_events; 1290 break; 1291 case EVSEL__CONFIG_TERM_INHERIT: 1292 /* 1293 * attr->inherit should has already been set by 1294 * evsel__config. If user explicitly set 1295 * inherit using config terms, override global 1296 * opt->no_inherit setting. 1297 */ 1298 attr->inherit = term->val.inherit ? 1 : 0; 1299 break; 1300 case EVSEL__CONFIG_TERM_OVERWRITE: 1301 attr->write_backward = term->val.overwrite ? 1 : 0; 1302 break; 1303 case EVSEL__CONFIG_TERM_DRV_CFG: 1304 break; 1305 case EVSEL__CONFIG_TERM_PERCORE: 1306 break; 1307 case EVSEL__CONFIG_TERM_AUX_OUTPUT: 1308 attr->aux_output = term->val.aux_output ? 1 : 0; 1309 break; 1310 case EVSEL__CONFIG_TERM_AUX_ACTION: 1311 /* Already applied by auxtrace */ 1312 break; 1313 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE: 1314 /* Already applied by auxtrace */ 1315 break; 1316 case EVSEL__CONFIG_TERM_USR_CHG_CONFIG: 1317 case EVSEL__CONFIG_TERM_USR_CHG_CONFIG1: 1318 case EVSEL__CONFIG_TERM_USR_CHG_CONFIG2: 1319 case EVSEL__CONFIG_TERM_USR_CHG_CONFIG3: 1320 case EVSEL__CONFIG_TERM_USR_CHG_CONFIG4: 1321 break; 1322 case EVSEL__CONFIG_TERM_RATIO_TO_PREV: 1323 rtp_buf = term->val.str; 1324 break; 1325 default: 1326 break; 1327 } 1328 } 1329 1330 /* User explicitly set per-event callgraph, clear the old setting and reset. */ 1331 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) { 1332 bool sample_address = false; 1333 1334 if (max_stack) { 1335 param.max_stack = max_stack; 1336 if (callgraph_buf == NULL) 1337 callgraph_buf = "fp"; 1338 } 1339 1340 /* parse callgraph parameters */ 1341 if (callgraph_buf != NULL) { 1342 if (!strcmp(callgraph_buf, "no")) { 1343 param.enabled = false; 1344 param.record_mode = CALLCHAIN_NONE; 1345 } else { 1346 param.enabled = true; 1347 if (parse_callchain_record(callgraph_buf, ¶m)) { 1348 pr_err("per-event callgraph setting for %s failed. " 1349 "Apply callgraph global setting for it\n", 1350 evsel->name); 1351 return; 1352 } 1353 if (param.record_mode == CALLCHAIN_DWARF) 1354 sample_address = true; 1355 } 1356 } 1357 if (dump_size > 0) { 1358 dump_size = round_up(dump_size, sizeof(u64)); 1359 param.dump_size = dump_size; 1360 } 1361 1362 /* If global callgraph set, clear it */ 1363 if (callchain_param.enabled) 1364 evsel__reset_callgraph(evsel, &callchain_param); 1365 1366 /* set perf-event callgraph */ 1367 if (param.enabled) { 1368 if (sample_address) { 1369 evsel__set_sample_bit(evsel, ADDR); 1370 evsel__set_sample_bit(evsel, DATA_SRC); 1371 evsel->core.attr.mmap_data = track; 1372 } 1373 evsel__config_callchain(evsel, opts, ¶m); 1374 } 1375 } 1376 if (rtp_buf) 1377 evsel__apply_ratio_to_prev(evsel, attr, opts, rtp_buf); 1378 } 1379 1380 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type) 1381 { 1382 struct evsel_config_term *term, *found_term = NULL; 1383 1384 list_for_each_entry(term, &evsel->config_terms, list) { 1385 if (term->type == type) 1386 found_term = term; 1387 } 1388 1389 return found_term; 1390 } 1391 1392 /* 1393 * Set @config_name to @val as long as the user hasn't already set or cleared it 1394 * by passing a config term on the command line. 1395 * 1396 * @val is the value to put into the bits specified by @config_name rather than 1397 * the bit pattern. It is shifted into position by this function, so to set 1398 * something to true, pass 1 for val rather than a pre shifted value. 1399 */ 1400 void evsel__set_config_if_unset(struct evsel *evsel, const char *config_name, 1401 u64 val) 1402 { 1403 u64 user_bits = 0; 1404 struct evsel_config_term *term = evsel__get_config_term(evsel, 1405 USR_CHG_CONFIG); 1406 struct perf_pmu_format *format = pmu_find_format(&evsel->pmu->format, 1407 config_name); 1408 int fbit; 1409 __u64 *vp; 1410 1411 if (!format) 1412 return; 1413 1414 switch (format->value) { 1415 case PERF_PMU_FORMAT_VALUE_CONFIG: 1416 term = evsel__get_config_term(evsel, USR_CHG_CONFIG); 1417 vp = &evsel->core.attr.config; 1418 break; 1419 case PERF_PMU_FORMAT_VALUE_CONFIG1: 1420 term = evsel__get_config_term(evsel, USR_CHG_CONFIG1); 1421 vp = &evsel->core.attr.config1; 1422 break; 1423 case PERF_PMU_FORMAT_VALUE_CONFIG2: 1424 term = evsel__get_config_term(evsel, USR_CHG_CONFIG2); 1425 vp = &evsel->core.attr.config2; 1426 break; 1427 case PERF_PMU_FORMAT_VALUE_CONFIG3: 1428 term = evsel__get_config_term(evsel, USR_CHG_CONFIG3); 1429 vp = &evsel->core.attr.config3; 1430 break; 1431 case PERF_PMU_FORMAT_VALUE_CONFIG4: 1432 term = evsel__get_config_term(evsel, USR_CHG_CONFIG4); 1433 vp = &evsel->core.attr.config4; 1434 break; 1435 default: 1436 pr_err("Unknown format value: %d\n", format->value); 1437 return; 1438 } 1439 1440 if (!format) 1441 return; 1442 1443 if (term) 1444 user_bits = term->val.cfg_chg; 1445 1446 /* Do nothing if the user changed the value */ 1447 for_each_set_bit(fbit, format->bits, PERF_PMU_FORMAT_BITS) 1448 if ((1ULL << fbit) & user_bits) 1449 return; 1450 1451 /* Otherwise replace it */ 1452 perf_pmu__format_pack(format->bits, val, vp, /*zero=*/true); 1453 } 1454 1455 bool evsel__config_exists(const struct evsel *evsel, const char *config_name) 1456 { 1457 struct perf_pmu_format *format = pmu_find_format(&evsel->pmu->format, config_name); 1458 1459 return format && !bitmap_empty(format->bits, PERF_PMU_FORMAT_BITS); 1460 } 1461 1462 int evsel__get_config_val(const struct evsel *evsel, const char *config_name, 1463 u64 *val) 1464 { 1465 struct perf_pmu_format *format = pmu_find_format(&evsel->pmu->format, config_name); 1466 1467 if (!format || bitmap_empty(format->bits, PERF_PMU_FORMAT_BITS)) { 1468 pr_err("Unknown/empty format name: %s\n", config_name); 1469 *val = 0; 1470 return -EINVAL; 1471 } 1472 1473 switch (format->value) { 1474 case PERF_PMU_FORMAT_VALUE_CONFIG: 1475 *val = perf_pmu__format_unpack(format->bits, 1476 evsel->core.attr.config); 1477 return 0; 1478 case PERF_PMU_FORMAT_VALUE_CONFIG1: 1479 *val = perf_pmu__format_unpack(format->bits, 1480 evsel->core.attr.config1); 1481 return 0; 1482 case PERF_PMU_FORMAT_VALUE_CONFIG2: 1483 *val = perf_pmu__format_unpack(format->bits, 1484 evsel->core.attr.config2); 1485 return 0; 1486 case PERF_PMU_FORMAT_VALUE_CONFIG3: 1487 *val = perf_pmu__format_unpack(format->bits, 1488 evsel->core.attr.config3); 1489 return 0; 1490 case PERF_PMU_FORMAT_VALUE_CONFIG4: 1491 *val = perf_pmu__format_unpack(format->bits, 1492 evsel->core.attr.config4); 1493 return 0; 1494 default: 1495 pr_err("Unknown format value: %d\n", format->value); 1496 *val = 0; 1497 return -EINVAL; 1498 } 1499 } 1500 1501 void __weak arch_evsel__set_sample_weight(struct evsel *evsel) 1502 { 1503 evsel__set_sample_bit(evsel, WEIGHT); 1504 } 1505 1506 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused, 1507 struct perf_event_attr *attr __maybe_unused) 1508 { 1509 } 1510 1511 void __weak arch_evsel__apply_ratio_to_prev(struct evsel *evsel __maybe_unused, 1512 struct perf_event_attr *attr __maybe_unused) 1513 { 1514 } 1515 1516 static void evsel__set_default_freq_period(const struct record_opts *opts, 1517 struct perf_event_attr *attr) 1518 { 1519 if (opts->freq) { 1520 attr->freq = 1; 1521 attr->sample_freq = opts->freq; 1522 } else { 1523 attr->sample_period = opts->default_interval; 1524 } 1525 } 1526 1527 bool evsel__is_offcpu_event(struct evsel *evsel) 1528 { 1529 return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT) && 1530 evsel->core.attr.sample_type & PERF_SAMPLE_RAW; 1531 } 1532 1533 /* 1534 * The enable_on_exec/disabled value strategy: 1535 * 1536 * 1) For any type of traced program: 1537 * - all independent events and group leaders are disabled 1538 * - all group members are enabled 1539 * 1540 * Group members are ruled by group leaders. They need to 1541 * be enabled, because the group scheduling relies on that. 1542 * 1543 * 2) For traced programs executed by perf: 1544 * - all independent events and group leaders have 1545 * enable_on_exec set 1546 * - we don't specifically enable or disable any event during 1547 * the record command 1548 * 1549 * Independent events and group leaders are initially disabled 1550 * and get enabled by exec. Group members are ruled by group 1551 * leaders as stated in 1). 1552 * 1553 * 3) For traced programs attached by perf (pid/tid): 1554 * - we specifically enable or disable all events during 1555 * the record command 1556 * 1557 * When attaching events to already running traced we 1558 * enable/disable events specifically, as there's no 1559 * initial traced exec call. 1560 */ 1561 void evsel__config(struct evsel *evsel, const struct record_opts *opts, 1562 const struct callchain_param *callchain) 1563 { 1564 struct evsel *leader = evsel__leader(evsel); 1565 struct perf_event_attr *attr = &evsel->core.attr; 1566 int track = evsel->tracking; 1567 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread; 1568 1569 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1; 1570 attr->inherit = target__has_cpu(&opts->target) ? 0 : !opts->no_inherit; 1571 attr->write_backward = opts->overwrite ? 1 : 0; 1572 attr->read_format = PERF_FORMAT_LOST; 1573 1574 evsel__set_sample_bit(evsel, IP); 1575 evsel__set_sample_bit(evsel, TID); 1576 1577 if (evsel->sample_read) { 1578 evsel__set_sample_bit(evsel, READ); 1579 1580 /* 1581 * We need ID even in case of single event, because 1582 * PERF_SAMPLE_READ process ID specific data. 1583 */ 1584 evsel__set_sample_id(evsel, false); 1585 1586 /* 1587 * Apply group format only if we belong to group 1588 * with more than one members. 1589 */ 1590 if (leader->core.nr_members > 1) { 1591 attr->read_format |= PERF_FORMAT_GROUP; 1592 } 1593 1594 /* 1595 * Inherit + SAMPLE_READ requires SAMPLE_TID in the read_format 1596 */ 1597 if (attr->inherit) { 1598 evsel__set_sample_bit(evsel, TID); 1599 evsel->core.attr.read_format |= 1600 PERF_FORMAT_ID; 1601 } 1602 } 1603 1604 /* 1605 * We default some events to have a default interval. But keep 1606 * it a weak assumption overridable by the user. 1607 */ 1608 if ((evsel->is_libpfm_event && !attr->sample_period) || 1609 (!evsel->is_libpfm_event && (!attr->sample_period || 1610 opts->user_freq != UINT_MAX || 1611 opts->user_interval != ULLONG_MAX))) 1612 evsel__set_default_freq_period(opts, attr); 1613 1614 /* 1615 * If attr->freq was set (here or earlier), ask for period 1616 * to be sampled. 1617 */ 1618 if (attr->freq) 1619 evsel__set_sample_bit(evsel, PERIOD); 1620 1621 if (opts->no_samples) 1622 attr->sample_freq = 0; 1623 1624 if (opts->inherit_stat) { 1625 evsel->core.attr.read_format |= 1626 PERF_FORMAT_TOTAL_TIME_ENABLED | 1627 PERF_FORMAT_TOTAL_TIME_RUNNING | 1628 PERF_FORMAT_ID; 1629 attr->inherit_stat = 1; 1630 } 1631 1632 if (opts->sample_address) 1633 evsel__set_sample_bit(evsel, ADDR); 1634 1635 if (opts->record_data_mmap) 1636 attr->mmap_data = track; 1637 1638 /* 1639 * We don't allow user space callchains for function trace 1640 * event, due to issues with page faults while tracing page 1641 * fault handler and its overall trickiness nature. 1642 */ 1643 if (evsel__is_function_event(evsel)) 1644 evsel->core.attr.exclude_callchain_user = 1; 1645 1646 if (callchain && callchain->enabled && !evsel->no_aux_samples) 1647 evsel__config_callchain(evsel, opts, callchain); 1648 1649 if (opts->sample_intr_regs && !evsel->no_aux_samples && 1650 !evsel__is_dummy_event(evsel)) { 1651 attr->sample_regs_intr = opts->sample_intr_regs; 1652 evsel__set_sample_bit(evsel, REGS_INTR); 1653 } 1654 1655 if (opts->sample_user_regs && !evsel->no_aux_samples && 1656 !evsel__is_dummy_event(evsel)) { 1657 attr->sample_regs_user |= opts->sample_user_regs; 1658 evsel__set_sample_bit(evsel, REGS_USER); 1659 } 1660 1661 if (target__has_cpu(&opts->target) || opts->sample_cpu) 1662 evsel__set_sample_bit(evsel, CPU); 1663 1664 /* 1665 * When the user explicitly disabled time don't force it here. 1666 */ 1667 if (opts->sample_time && 1668 (!perf_missing_features.sample_id_all && 1669 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu || 1670 opts->sample_time_set))) 1671 evsel__set_sample_bit(evsel, TIME); 1672 1673 if (opts->raw_samples && !evsel->no_aux_samples) { 1674 evsel__set_sample_bit(evsel, TIME); 1675 evsel__set_sample_bit(evsel, RAW); 1676 evsel__set_sample_bit(evsel, CPU); 1677 } 1678 1679 if (opts->sample_data_src) 1680 evsel__set_sample_bit(evsel, DATA_SRC); 1681 1682 if (opts->sample_phys_addr) 1683 evsel__set_sample_bit(evsel, PHYS_ADDR); 1684 1685 if (opts->no_buffering) { 1686 attr->watermark = 0; 1687 attr->wakeup_events = 1; 1688 } 1689 if (opts->branch_stack && !evsel->no_aux_samples) { 1690 evsel__set_sample_bit(evsel, BRANCH_STACK); 1691 attr->branch_sample_type = opts->branch_stack; 1692 } 1693 1694 if (opts->sample_weight || evsel->retire_lat) { 1695 arch_evsel__set_sample_weight(evsel); 1696 evsel->retire_lat = false; 1697 } 1698 attr->task = track; 1699 attr->mmap = track; 1700 attr->mmap2 = track && !perf_missing_features.mmap2; 1701 attr->comm = track; 1702 attr->build_id = track && opts->build_id; 1703 attr->defer_output = track && callchain && callchain->defer; 1704 1705 /* 1706 * ksymbol is tracked separately with text poke because it needs to be 1707 * system wide and enabled immediately. 1708 */ 1709 if (!opts->text_poke) 1710 attr->ksymbol = track && !perf_missing_features.ksymbol; 1711 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf; 1712 1713 if (opts->record_namespaces) 1714 attr->namespaces = track; 1715 1716 if (opts->record_cgroup) { 1717 attr->cgroup = track && !perf_missing_features.cgroup; 1718 evsel__set_sample_bit(evsel, CGROUP); 1719 } 1720 1721 if (opts->sample_data_page_size) 1722 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE); 1723 1724 if (opts->sample_code_page_size) 1725 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE); 1726 1727 if (opts->record_switch_events) 1728 attr->context_switch = track; 1729 1730 if (opts->sample_transaction) 1731 evsel__set_sample_bit(evsel, TRANSACTION); 1732 1733 if (opts->running_time) { 1734 evsel->core.attr.read_format |= 1735 PERF_FORMAT_TOTAL_TIME_ENABLED | 1736 PERF_FORMAT_TOTAL_TIME_RUNNING; 1737 } 1738 1739 /* 1740 * XXX see the function comment above 1741 * 1742 * Disabling only independent events or group leaders, 1743 * keeping group members enabled. 1744 */ 1745 if (evsel__is_group_leader(evsel)) 1746 attr->disabled = 1; 1747 1748 /* 1749 * Setting enable_on_exec for independent events and 1750 * group leaders for traced executed by perf. 1751 */ 1752 if (target__none(&opts->target) && evsel__is_group_leader(evsel) && 1753 !opts->target.initial_delay) 1754 attr->enable_on_exec = 1; 1755 1756 if (evsel->immediate) { 1757 attr->disabled = 0; 1758 attr->enable_on_exec = 0; 1759 } 1760 1761 clockid = opts->clockid; 1762 if (opts->use_clockid) { 1763 attr->use_clockid = 1; 1764 attr->clockid = opts->clockid; 1765 } 1766 1767 if (evsel->precise_max) 1768 attr->precise_ip = 3; 1769 1770 if (opts->all_user) { 1771 attr->exclude_kernel = 1; 1772 attr->exclude_user = 0; 1773 } 1774 1775 if (opts->all_kernel) { 1776 attr->exclude_kernel = 0; 1777 attr->exclude_user = 1; 1778 } 1779 1780 if (evsel->core.pmu_cpus || evsel->unit) 1781 evsel->core.attr.read_format |= PERF_FORMAT_ID; 1782 1783 /* 1784 * Apply event specific term settings, 1785 * it overloads any global configuration. 1786 */ 1787 evsel__apply_config_terms(evsel, opts, track); 1788 1789 evsel->ignore_missing_thread = opts->ignore_missing_thread; 1790 1791 /* The --period option takes the precedence. */ 1792 if (opts->period_set) { 1793 if (opts->period) 1794 evsel__set_sample_bit(evsel, PERIOD); 1795 else 1796 evsel__reset_sample_bit(evsel, PERIOD); 1797 } 1798 1799 /* 1800 * A dummy event never triggers any actual counter and therefore 1801 * cannot be used with branch_stack. 1802 * 1803 * For initial_delay, a dummy event is added implicitly. 1804 * The software event will trigger -EOPNOTSUPP error out, 1805 * if BRANCH_STACK bit is set. 1806 */ 1807 if (evsel__is_dummy_event(evsel)) 1808 evsel__reset_sample_bit(evsel, BRANCH_STACK); 1809 1810 if (evsel__is_offcpu_event(evsel)) { 1811 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES; 1812 attr->inherit = 0; 1813 } 1814 1815 arch__post_evsel_config(evsel, attr); 1816 } 1817 1818 int evsel__set_filter(struct evsel *evsel, const char *filter) 1819 { 1820 char *new_filter = strdup(filter); 1821 1822 if (new_filter != NULL) { 1823 free(evsel->filter); 1824 evsel->filter = new_filter; 1825 return 0; 1826 } 1827 1828 return -1; 1829 } 1830 1831 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter) 1832 { 1833 char *new_filter; 1834 1835 if (evsel->filter == NULL) 1836 return evsel__set_filter(evsel, filter); 1837 1838 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) { 1839 free(evsel->filter); 1840 evsel->filter = new_filter; 1841 return 0; 1842 } 1843 1844 return -1; 1845 } 1846 1847 int evsel__append_tp_filter(struct evsel *evsel, const char *filter) 1848 { 1849 return evsel__append_filter(evsel, "(%s) && (%s)", filter); 1850 } 1851 1852 int evsel__append_addr_filter(struct evsel *evsel, const char *filter) 1853 { 1854 return evsel__append_filter(evsel, "%s,%s", filter); 1855 } 1856 1857 /* Caller has to clear disabled after going through all CPUs. */ 1858 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx) 1859 { 1860 int err; 1861 1862 if (evsel__is_tool(evsel)) 1863 err = evsel__tool_pmu_enable_cpu(evsel, cpu_map_idx); 1864 else 1865 err = perf_evsel__enable_cpu(&evsel->core, cpu_map_idx); 1866 1867 if (!err && evsel__is_group_leader(evsel)) { 1868 struct evsel *member; 1869 1870 for_each_group_member(member, evsel) { 1871 if (evsel__is_non_perf_event_open_pmu(evsel) || 1872 evsel__is_non_perf_event_open_pmu(member)) { 1873 /* 1874 * In a mixed PMU group, userspace PMUs are not 1875 * grouped in the kernel (opened with group_fd = -1) 1876 * and are skipped by the kernel when enabling the 1877 * group leader. We must manually enable them in 1878 * userspace. 1879 */ 1880 int mem_err = evsel__enable_cpu(member, cpu_map_idx); 1881 1882 if (mem_err) 1883 return mem_err; 1884 } 1885 } 1886 } 1887 return err; 1888 } 1889 1890 int evsel__enable(struct evsel *evsel) 1891 { 1892 int err; 1893 1894 if (evsel__is_tool(evsel)) 1895 err = evsel__tool_pmu_enable(evsel); 1896 else 1897 err = perf_evsel__enable(&evsel->core); 1898 1899 if (!err) 1900 evsel->disabled = false; 1901 1902 if (!err && evsel__is_group_leader(evsel)) { 1903 struct evsel *member; 1904 1905 for_each_group_member(member, evsel) { 1906 if (evsel__is_non_perf_event_open_pmu(evsel) || 1907 evsel__is_non_perf_event_open_pmu(member)) { 1908 /* 1909 * In a mixed PMU group, userspace PMUs are not 1910 * grouped in the kernel (opened with group_fd = -1) 1911 * and are skipped by the kernel when enabling the 1912 * group leader. We must manually enable them in 1913 * userspace. 1914 */ 1915 int mem_err = evsel__enable(member); 1916 1917 if (mem_err) 1918 return mem_err; 1919 } 1920 member->disabled = false; 1921 } 1922 } 1923 1924 return err; 1925 } 1926 1927 /* Caller has to set disabled after going through all CPUs. */ 1928 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx) 1929 { 1930 int err; 1931 1932 if (evsel__is_tool(evsel)) 1933 err = evsel__tool_pmu_disable_cpu(evsel, cpu_map_idx); 1934 else 1935 err = perf_evsel__disable_cpu(&evsel->core, cpu_map_idx); 1936 1937 if (!err && evsel__is_group_leader(evsel)) { 1938 struct evsel *member; 1939 1940 for_each_group_member(member, evsel) { 1941 if (evsel__is_non_perf_event_open_pmu(evsel) || 1942 evsel__is_non_perf_event_open_pmu(member)) { 1943 /* 1944 * In a mixed PMU group, userspace PMUs are not 1945 * grouped in the kernel and are skipped by the 1946 * kernel when disabling the group leader. We must 1947 * manually disable them in userspace. 1948 */ 1949 int mem_err = evsel__disable_cpu(member, cpu_map_idx); 1950 1951 if (mem_err) 1952 return mem_err; 1953 } 1954 } 1955 } 1956 return err; 1957 } 1958 1959 int evsel__disable(struct evsel *evsel) 1960 { 1961 int err; 1962 1963 if (evsel__is_tool(evsel)) 1964 err = evsel__tool_pmu_disable(evsel); 1965 else 1966 err = perf_evsel__disable(&evsel->core); 1967 1968 /* 1969 * We mark it disabled here so that tools that disable a event can 1970 * ignore events after they disable it. I.e. the ring buffer may have 1971 * already a few more events queued up before the kernel got the stop 1972 * request. 1973 */ 1974 if (!err) 1975 evsel->disabled = true; 1976 1977 if (!err && evsel__is_group_leader(evsel)) { 1978 struct evsel *member; 1979 1980 for_each_group_member(member, evsel) { 1981 if (evsel__is_non_perf_event_open_pmu(evsel) || 1982 evsel__is_non_perf_event_open_pmu(member)) { 1983 /* 1984 * In a mixed PMU group, userspace PMUs are not 1985 * grouped in the kernel and are skipped by the 1986 * kernel when disabling the group leader. We must 1987 * manually disable them in userspace. 1988 */ 1989 int mem_err = evsel__disable(member); 1990 1991 if (mem_err) 1992 return mem_err; 1993 } 1994 member->disabled = true; 1995 } 1996 } 1997 1998 return err; 1999 } 2000 2001 void free_config_terms(struct list_head *config_terms) 2002 { 2003 struct evsel_config_term *term, *h; 2004 2005 list_for_each_entry_safe(term, h, config_terms, list) { 2006 list_del_init(&term->list); 2007 if (term->free_str) 2008 zfree(&term->val.str); 2009 free(term); 2010 } 2011 } 2012 2013 static void evsel__free_config_terms(struct evsel *evsel) 2014 { 2015 free_config_terms(&evsel->config_terms); 2016 } 2017 2018 static void (*evsel__priv_destructor)(void *priv); 2019 2020 void evsel__set_priv_destructor(void (*destructor)(void *priv)) 2021 { 2022 assert(evsel__priv_destructor == NULL); 2023 2024 evsel__priv_destructor = destructor; 2025 } 2026 2027 void evsel__exit(struct evsel *evsel) 2028 { 2029 assert(list_empty(&evsel->core.node)); 2030 assert(evsel->evlist == NULL); 2031 if (evsel__is_retire_lat(evsel)) 2032 evsel__tpebs_close(evsel); 2033 bpf_counter__destroy(evsel); 2034 perf_bpf_filter__destroy(evsel); 2035 evsel__free_counts(evsel); 2036 perf_evsel__free_fd(&evsel->core); 2037 perf_evsel__free_id(&evsel->core); 2038 evsel__free_config_terms(evsel); 2039 cgroup__put(evsel->cgrp); 2040 perf_evsel__exit(&evsel->core); 2041 zfree(&evsel->group_name); 2042 zfree(&evsel->name); 2043 #ifdef HAVE_LIBTRACEEVENT 2044 zfree(&evsel->tp_sys); 2045 zfree(&evsel->tp_name); 2046 #endif 2047 zfree(&evsel->filter); 2048 zfree(&evsel->group_pmu_name); 2049 zfree(&evsel->unit); 2050 zfree(&evsel->metric_id); 2051 evsel__zero_per_pkg(evsel); 2052 hashmap__free(evsel->per_pkg_mask); 2053 evsel->per_pkg_mask = NULL; 2054 if (evsel__priv_destructor) 2055 evsel__priv_destructor(evsel->priv); 2056 perf_evsel__object.fini(evsel); 2057 if (evsel__tool_event(evsel) == TOOL_PMU__EVENT_SYSTEM_TIME || 2058 evsel__tool_event(evsel) == TOOL_PMU__EVENT_USER_TIME) { 2059 xyarray__delete(evsel->process_time.start_times); 2060 xyarray__delete(evsel->process_time.accumulated_times); 2061 } 2062 } 2063 2064 void evsel__delete(struct evsel *evsel) 2065 { 2066 if (!evsel) 2067 return; 2068 2069 evsel__exit(evsel); 2070 free(evsel); 2071 } 2072 2073 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread, 2074 struct perf_counts_values *count) 2075 { 2076 struct perf_counts_values tmp; 2077 2078 if (!evsel->prev_raw_counts) 2079 return; 2080 2081 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread); 2082 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count; 2083 2084 count->val = count->val - tmp.val; 2085 count->ena = count->ena - tmp.ena; 2086 count->run = count->run - tmp.run; 2087 } 2088 2089 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread) 2090 { 2091 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread); 2092 2093 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count); 2094 } 2095 2096 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread, 2097 u64 val, u64 ena, u64 run, u64 lost) 2098 { 2099 struct perf_counts_values *count; 2100 2101 count = perf_counts(counter->counts, cpu_map_idx, thread); 2102 2103 if (evsel__is_retire_lat(counter)) { 2104 evsel__tpebs_read(counter, cpu_map_idx, thread); 2105 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true); 2106 return; 2107 } 2108 2109 count->val = val; 2110 count->ena = ena; 2111 count->run = run; 2112 count->lost = lost; 2113 2114 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true); 2115 } 2116 2117 static bool evsel__group_has_tpebs(struct evsel *leader) 2118 { 2119 struct evsel *evsel; 2120 2121 for_each_group_evsel(evsel, leader) { 2122 if (evsel__is_retire_lat(evsel)) 2123 return true; 2124 } 2125 return false; 2126 } 2127 2128 static u64 evsel__group_read_nr_members(struct evsel *leader) 2129 { 2130 u64 nr = leader->core.nr_members; 2131 struct evsel *evsel; 2132 2133 for_each_group_evsel(evsel, leader) { 2134 if (evsel__is_retire_lat(evsel)) 2135 nr--; 2136 } 2137 return nr; 2138 } 2139 2140 static u64 evsel__group_read_size(struct evsel *leader) 2141 { 2142 u64 read_format = leader->core.attr.read_format; 2143 int entry = sizeof(u64); /* value */ 2144 int size = 0; 2145 int nr = 1; 2146 2147 if (!evsel__group_has_tpebs(leader)) 2148 return perf_evsel__read_size(&leader->core); 2149 2150 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 2151 size += sizeof(u64); 2152 2153 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 2154 size += sizeof(u64); 2155 2156 if (read_format & PERF_FORMAT_ID) 2157 entry += sizeof(u64); 2158 2159 if (read_format & PERF_FORMAT_LOST) 2160 entry += sizeof(u64); 2161 2162 if (read_format & PERF_FORMAT_GROUP) { 2163 nr = evsel__group_read_nr_members(leader); 2164 size += sizeof(u64); 2165 } 2166 2167 size += entry * nr; 2168 return size; 2169 } 2170 2171 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data) 2172 { 2173 u64 read_format = leader->core.attr.read_format; 2174 struct sample_read_value *v; 2175 u64 nr, ena = 0, run = 0, lost = 0; 2176 2177 nr = *data++; 2178 2179 if (nr != evsel__group_read_nr_members(leader)) 2180 return -EINVAL; 2181 2182 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 2183 ena = *data++; 2184 2185 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 2186 run = *data++; 2187 2188 v = (void *)data; 2189 sample_read_group__for_each(v, nr, read_format) { 2190 struct evsel *counter; 2191 2192 counter = evlist__id2evsel(leader->evlist, v->id); 2193 if (!counter) 2194 return -EINVAL; 2195 2196 if (read_format & PERF_FORMAT_LOST) 2197 lost = v->lost; 2198 2199 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost); 2200 } 2201 2202 return 0; 2203 } 2204 2205 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread) 2206 { 2207 struct perf_stat_evsel *ps = leader->stats; 2208 u64 read_format = leader->core.attr.read_format; 2209 int size = evsel__group_read_size(leader); 2210 u64 *data = ps->group_data; 2211 2212 if (!(read_format & PERF_FORMAT_ID)) 2213 return -EINVAL; 2214 2215 if (!evsel__is_group_leader(leader)) 2216 return -EINVAL; 2217 2218 if (!data) { 2219 data = zalloc(size); 2220 if (!data) 2221 return -ENOMEM; 2222 2223 ps->group_data = data; 2224 } 2225 2226 if (FD(leader, cpu_map_idx, thread) < 0) 2227 return -EINVAL; 2228 2229 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0) 2230 return -errno; 2231 2232 return evsel__process_group_data(leader, cpu_map_idx, thread, data); 2233 } 2234 2235 bool __evsel__match(const struct evsel *evsel, u32 type, u64 config) 2236 { 2237 2238 u32 e_type = evsel->core.attr.type; 2239 u64 e_config = evsel->core.attr.config; 2240 2241 if (e_type == type && e_config == config) 2242 return true; 2243 if (type != PERF_TYPE_HARDWARE && type != PERF_TYPE_HW_CACHE) 2244 return false; 2245 if ((e_type == PERF_TYPE_HARDWARE || e_type == PERF_TYPE_HW_CACHE) && 2246 perf_pmus__supports_extended_type()) 2247 e_config &= PERF_HW_EVENT_MASK; 2248 if (e_type == type && e_config == config) 2249 return true; 2250 if (type == PERF_TYPE_HARDWARE && evsel->pmu && evsel->pmu->is_core && 2251 evsel->alternate_hw_config == config) 2252 return true; 2253 return false; 2254 } 2255 2256 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread) 2257 { 2258 if (evsel__is_tool(evsel)) 2259 return evsel__tool_pmu_read(evsel, cpu_map_idx, thread); 2260 2261 if (evsel__is_hwmon(evsel)) 2262 return evsel__hwmon_pmu_read(evsel, cpu_map_idx, thread); 2263 2264 if (evsel__is_drm(evsel)) 2265 return evsel__drm_pmu_read(evsel, cpu_map_idx, thread); 2266 2267 if (evsel__is_retire_lat(evsel)) 2268 return evsel__tpebs_read(evsel, cpu_map_idx, thread); 2269 2270 if (evsel->core.attr.read_format & PERF_FORMAT_GROUP) 2271 return evsel__read_group(evsel, cpu_map_idx, thread); 2272 2273 return evsel__read_one(evsel, cpu_map_idx, thread); 2274 } 2275 2276 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale) 2277 { 2278 struct perf_counts_values count; 2279 size_t nv = scale ? 3 : 1; 2280 2281 if (FD(evsel, cpu_map_idx, thread) < 0) 2282 return -EINVAL; 2283 2284 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0) 2285 return -ENOMEM; 2286 2287 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0) 2288 return -errno; 2289 2290 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count); 2291 perf_counts_values__scale(&count, scale, NULL); 2292 *perf_counts(evsel->counts, cpu_map_idx, thread) = count; 2293 return 0; 2294 } 2295 2296 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other, 2297 int cpu_map_idx) 2298 { 2299 struct perf_cpu cpu; 2300 2301 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx); 2302 return perf_cpu_map__idx(other->core.cpus, cpu); 2303 } 2304 2305 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx) 2306 { 2307 struct evsel *leader = evsel__leader(evsel); 2308 2309 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) || 2310 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) { 2311 return evsel__match_other_cpu(evsel, leader, cpu_map_idx); 2312 } 2313 2314 return cpu_map_idx; 2315 } 2316 2317 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread) 2318 { 2319 struct evsel *leader = evsel__leader(evsel); 2320 int fd; 2321 2322 if (!evsel->supported || evsel__is_group_leader(evsel)) 2323 return -1; 2324 2325 /* 2326 * Leader must be already processed/open, 2327 * if not it's a bug. 2328 */ 2329 BUG_ON(!leader->core.fd); 2330 2331 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx); 2332 if (cpu_map_idx == -1) 2333 return -1; 2334 2335 fd = FD(leader, cpu_map_idx, thread); 2336 BUG_ON(fd == -1 && leader->supported); 2337 2338 /* 2339 * When the leader has been skipped, return -2 to distinguish from no 2340 * group leader case. 2341 */ 2342 return fd == -1 ? -2 : fd; 2343 } 2344 2345 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx) 2346 { 2347 for (int cpu = 0; cpu < nr_cpus; cpu++) 2348 for (int thread = thread_idx; thread < nr_threads - 1; thread++) 2349 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1); 2350 } 2351 2352 static int update_fds(struct evsel *evsel, 2353 int nr_cpus, int cpu_map_idx, 2354 int nr_threads, int thread_idx) 2355 { 2356 struct evsel *pos; 2357 2358 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads) 2359 return -EINVAL; 2360 2361 evlist__for_each_entry(evsel->evlist, pos) { 2362 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx; 2363 2364 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx); 2365 2366 /* 2367 * Since fds for next evsel has not been created, 2368 * there is no need to iterate whole event list. 2369 */ 2370 if (pos == evsel) 2371 break; 2372 } 2373 return 0; 2374 } 2375 2376 static bool evsel__ignore_missing_thread(struct evsel *evsel, 2377 int nr_cpus, int cpu_map_idx, 2378 struct perf_thread_map *threads, 2379 int thread, int err) 2380 { 2381 pid_t ignore_pid = perf_thread_map__pid(threads, thread); 2382 2383 if (!evsel->ignore_missing_thread) 2384 return false; 2385 2386 /* The system wide setup does not work with threads. */ 2387 if (evsel->core.system_wide) 2388 return false; 2389 2390 /* The -ESRCH is perf event syscall errno for pid's not found. */ 2391 if (err != -ESRCH) 2392 return false; 2393 2394 /* If there's only one thread, let it fail. */ 2395 if (threads->nr == 1) 2396 return false; 2397 2398 /* 2399 * We should remove fd for missing_thread first 2400 * because thread_map__remove() will decrease threads->nr. 2401 */ 2402 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread)) 2403 return false; 2404 2405 if (thread_map__remove(threads, thread)) 2406 return false; 2407 2408 pr_warning("WARNING: Ignored open failure for pid %d\n", 2409 ignore_pid); 2410 return true; 2411 } 2412 2413 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val, 2414 void *priv __maybe_unused) 2415 { 2416 return fprintf(fp, " %-32s %s\n", name, val); 2417 } 2418 2419 static void display_attr(struct perf_event_attr *attr) 2420 { 2421 if (verbose >= 2 || debug_peo_args) { 2422 fprintf(stderr, "%.60s\n", graph_dotted_line); 2423 fprintf(stderr, "perf_event_attr:\n"); 2424 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL); 2425 fprintf(stderr, "%.60s\n", graph_dotted_line); 2426 } 2427 } 2428 2429 bool evsel__precise_ip_fallback(struct evsel *evsel) 2430 { 2431 /* Do not try less precise if not requested. */ 2432 if (!evsel->precise_max) 2433 return false; 2434 2435 /* 2436 * We tried all the precise_ip values, and it's 2437 * still failing, so leave it to standard fallback. 2438 */ 2439 if (!evsel->core.attr.precise_ip) { 2440 evsel->core.attr.precise_ip = evsel->precise_ip_original; 2441 return false; 2442 } 2443 2444 if (!evsel->precise_ip_original) 2445 evsel->precise_ip_original = evsel->core.attr.precise_ip; 2446 2447 evsel->core.attr.precise_ip--; 2448 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip); 2449 display_attr(&evsel->core.attr); 2450 return true; 2451 } 2452 2453 static struct perf_cpu_map *empty_cpu_map; 2454 static struct perf_thread_map *empty_thread_map; 2455 2456 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus, 2457 struct perf_thread_map *threads) 2458 { 2459 int ret = 0; 2460 int nthreads = perf_thread_map__nr(threads); 2461 2462 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) || 2463 (perf_missing_features.aux_output && evsel->core.attr.aux_output)) 2464 return -EINVAL; 2465 2466 if (cpus == NULL) { 2467 if (empty_cpu_map == NULL) { 2468 empty_cpu_map = perf_cpu_map__new_any_cpu(); 2469 if (empty_cpu_map == NULL) 2470 return -ENOMEM; 2471 } 2472 2473 cpus = empty_cpu_map; 2474 } 2475 2476 if (threads == NULL) { 2477 if (empty_thread_map == NULL) { 2478 empty_thread_map = thread_map__new_by_tid(-1); 2479 if (empty_thread_map == NULL) 2480 return -ENOMEM; 2481 } 2482 2483 threads = empty_thread_map; 2484 } 2485 2486 if (evsel->core.fd == NULL && 2487 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0) 2488 return -ENOMEM; 2489 2490 if (evsel__is_tool(evsel)) 2491 ret = evsel__tool_pmu_prepare_open(evsel, cpus, nthreads); 2492 2493 evsel->open_flags = PERF_FLAG_FD_CLOEXEC; 2494 if (evsel->cgrp) 2495 evsel->open_flags |= PERF_FLAG_PID_CGROUP; 2496 2497 return ret; 2498 } 2499 2500 static void evsel__disable_missing_features(struct evsel *evsel) 2501 { 2502 if (perf_missing_features.defer_callchain && evsel->core.attr.defer_callchain) 2503 evsel->core.attr.defer_callchain = 0; 2504 if (perf_missing_features.defer_callchain && evsel->core.attr.defer_output) 2505 evsel->core.attr.defer_output = 0; 2506 if (perf_missing_features.inherit_sample_read && evsel->core.attr.inherit && 2507 (evsel->core.attr.sample_type & PERF_SAMPLE_READ)) 2508 evsel->core.attr.inherit = 0; 2509 if (perf_missing_features.branch_counters) 2510 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_COUNTERS; 2511 if (perf_missing_features.read_lost) 2512 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST; 2513 if (perf_missing_features.weight_struct) { 2514 evsel__set_sample_bit(evsel, WEIGHT); 2515 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT); 2516 } 2517 if (perf_missing_features.clockid_wrong) 2518 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */ 2519 if (perf_missing_features.clockid) { 2520 evsel->core.attr.use_clockid = 0; 2521 evsel->core.attr.clockid = 0; 2522 } 2523 if (perf_missing_features.cloexec) 2524 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 2525 if (perf_missing_features.mmap2) 2526 evsel->core.attr.mmap2 = 0; 2527 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest) 2528 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0; 2529 if (perf_missing_features.lbr_flags) 2530 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS | 2531 PERF_SAMPLE_BRANCH_NO_CYCLES); 2532 if (perf_missing_features.group_read && evsel->core.attr.inherit) 2533 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID); 2534 if (perf_missing_features.ksymbol) 2535 evsel->core.attr.ksymbol = 0; 2536 if (perf_missing_features.bpf) 2537 evsel->core.attr.bpf_event = 0; 2538 if (perf_missing_features.branch_hw_idx) 2539 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX; 2540 if (perf_missing_features.sample_id_all) 2541 evsel->core.attr.sample_id_all = 0; 2542 } 2543 2544 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus, 2545 struct perf_thread_map *threads) 2546 { 2547 int err; 2548 2549 err = __evsel__prepare_open(evsel, cpus, threads); 2550 if (err) 2551 return err; 2552 2553 evsel__disable_missing_features(evsel); 2554 2555 return err; 2556 } 2557 2558 static bool __has_attr_feature(struct perf_event_attr *attr, 2559 struct perf_cpu cpu, unsigned long flags) 2560 { 2561 int fd = syscall(SYS_perf_event_open, attr, /*pid=*/0, cpu.cpu, 2562 /*group_fd=*/-1, flags); 2563 close(fd); 2564 2565 if (fd < 0) { 2566 attr->exclude_kernel = 1; 2567 2568 fd = syscall(SYS_perf_event_open, attr, /*pid=*/0, cpu.cpu, 2569 /*group_fd=*/-1, flags); 2570 close(fd); 2571 } 2572 2573 if (fd < 0) { 2574 attr->exclude_hv = 1; 2575 2576 fd = syscall(SYS_perf_event_open, attr, /*pid=*/0, cpu.cpu, 2577 /*group_fd=*/-1, flags); 2578 close(fd); 2579 } 2580 2581 if (fd < 0) { 2582 attr->exclude_guest = 1; 2583 2584 fd = syscall(SYS_perf_event_open, attr, /*pid=*/0, cpu.cpu, 2585 /*group_fd=*/-1, flags); 2586 close(fd); 2587 } 2588 2589 attr->exclude_kernel = 0; 2590 attr->exclude_guest = 0; 2591 attr->exclude_hv = 0; 2592 2593 return fd >= 0; 2594 } 2595 2596 static bool has_attr_feature(struct perf_event_attr *attr, unsigned long flags) 2597 { 2598 struct perf_cpu cpu = {.cpu = -1}; 2599 2600 return __has_attr_feature(attr, cpu, flags); 2601 } 2602 2603 static void evsel__detect_missing_pmu_features(struct evsel *evsel) 2604 { 2605 struct perf_event_attr attr = { 2606 .type = evsel->core.attr.type, 2607 .config = evsel->core.attr.config, 2608 .disabled = 1, 2609 }; 2610 struct perf_pmu *pmu = evsel->pmu; 2611 int old_errno; 2612 2613 old_errno = errno; 2614 2615 if (pmu == NULL) 2616 pmu = evsel->pmu = evsel__find_pmu(evsel); 2617 2618 if (pmu == NULL || pmu->missing_features.checked) 2619 goto out; 2620 2621 /* 2622 * Must probe features in the order they were added to the 2623 * perf_event_attr interface. These are kernel core limitation but 2624 * specific to PMUs with branch stack. So we can detect with the given 2625 * hardware event and stop on the first one succeeded. 2626 */ 2627 2628 /* Please add new feature detection here. */ 2629 2630 attr.exclude_guest = 1; 2631 if (has_attr_feature(&attr, /*flags=*/0)) 2632 goto found; 2633 pmu->missing_features.exclude_guest = true; 2634 pr_debug2("switching off exclude_guest for PMU %s\n", pmu->name); 2635 2636 found: 2637 pmu->missing_features.checked = true; 2638 out: 2639 errno = old_errno; 2640 } 2641 2642 static void evsel__detect_missing_brstack_features(struct evsel *evsel) 2643 { 2644 static bool detection_done = false; 2645 struct perf_event_attr attr = { 2646 .type = evsel->core.attr.type, 2647 .config = evsel->core.attr.config, 2648 .disabled = 1, 2649 .sample_type = PERF_SAMPLE_BRANCH_STACK, 2650 .sample_period = 1000, 2651 }; 2652 int old_errno; 2653 2654 if (detection_done) 2655 return; 2656 2657 old_errno = errno; 2658 2659 /* 2660 * Must probe features in the order they were added to the 2661 * perf_event_attr interface. These are PMU specific limitation 2662 * so we can detect with the given hardware event and stop on the 2663 * first one succeeded. 2664 */ 2665 2666 /* Please add new feature detection here. */ 2667 2668 attr.branch_sample_type = PERF_SAMPLE_BRANCH_COUNTERS; 2669 if (has_attr_feature(&attr, /*flags=*/0)) 2670 goto found; 2671 perf_missing_features.branch_counters = true; 2672 pr_debug2("switching off branch counters support\n"); 2673 2674 attr.branch_sample_type = PERF_SAMPLE_BRANCH_HW_INDEX; 2675 if (has_attr_feature(&attr, /*flags=*/0)) 2676 goto found; 2677 perf_missing_features.branch_hw_idx = true; 2678 pr_debug2("switching off branch HW index support\n"); 2679 2680 attr.branch_sample_type = PERF_SAMPLE_BRANCH_NO_CYCLES | PERF_SAMPLE_BRANCH_NO_FLAGS; 2681 if (has_attr_feature(&attr, /*flags=*/0)) 2682 goto found; 2683 perf_missing_features.lbr_flags = true; 2684 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n"); 2685 2686 found: 2687 detection_done = true; 2688 errno = old_errno; 2689 } 2690 2691 static bool evsel__probe_aux_action(struct evsel *evsel, struct perf_cpu cpu) 2692 { 2693 struct perf_event_attr attr = evsel->core.attr; 2694 int old_errno = errno; 2695 2696 attr.disabled = 1; 2697 attr.aux_start_paused = 1; 2698 2699 if (__has_attr_feature(&attr, cpu, /*flags=*/0)) { 2700 errno = old_errno; 2701 return true; 2702 } 2703 2704 /* 2705 * EOPNOTSUPP means the kernel supports the feature but the PMU does 2706 * not, so keep that distinction if possible. 2707 */ 2708 if (errno != EOPNOTSUPP) 2709 errno = old_errno; 2710 2711 return false; 2712 } 2713 2714 static void evsel__detect_missing_aux_action_feature(struct evsel *evsel, struct perf_cpu cpu) 2715 { 2716 static bool detection_done; 2717 struct evsel *leader; 2718 2719 /* 2720 * Don't bother probing aux_action if it is not being used or has been 2721 * probed before. 2722 */ 2723 if (!evsel->core.attr.aux_action || detection_done) 2724 return; 2725 2726 detection_done = true; 2727 2728 /* 2729 * The leader is an AUX area event. If it has failed, assume the feature 2730 * is not supported. 2731 */ 2732 leader = evsel__leader(evsel); 2733 if (evsel == leader) { 2734 perf_missing_features.aux_action = true; 2735 return; 2736 } 2737 2738 /* 2739 * AUX area event with aux_action must have been opened successfully 2740 * already, so feature is supported. 2741 */ 2742 if (leader->core.attr.aux_action) 2743 return; 2744 2745 if (!evsel__probe_aux_action(leader, cpu)) 2746 perf_missing_features.aux_action = true; 2747 } 2748 2749 static bool evsel__detect_missing_features(struct evsel *evsel, struct perf_cpu cpu) 2750 { 2751 static bool detection_done = false; 2752 struct perf_event_attr attr = { 2753 .type = PERF_TYPE_SOFTWARE, 2754 .config = PERF_COUNT_SW_TASK_CLOCK, 2755 .disabled = 1, 2756 }; 2757 int old_errno; 2758 2759 evsel__detect_missing_aux_action_feature(evsel, cpu); 2760 2761 evsel__detect_missing_pmu_features(evsel); 2762 2763 if (evsel__has_br_stack(evsel)) 2764 evsel__detect_missing_brstack_features(evsel); 2765 2766 if (detection_done) 2767 goto check; 2768 2769 old_errno = errno; 2770 2771 /* 2772 * Must probe features in the order they were added to the 2773 * perf_event_attr interface. These are kernel core limitation 2774 * not PMU-specific so we can detect with a software event and 2775 * stop on the first one succeeded. 2776 */ 2777 2778 /* Please add new feature detection here. */ 2779 2780 attr.defer_callchain = true; 2781 if (has_attr_feature(&attr, /*flags=*/0)) 2782 goto found; 2783 perf_missing_features.defer_callchain = true; 2784 pr_debug2("switching off deferred callchain support\n"); 2785 attr.defer_callchain = false; 2786 2787 attr.inherit = true; 2788 attr.sample_type = PERF_SAMPLE_READ | PERF_SAMPLE_TID; 2789 if (has_attr_feature(&attr, /*flags=*/0)) 2790 goto found; 2791 perf_missing_features.inherit_sample_read = true; 2792 pr_debug2("Using PERF_SAMPLE_READ / :S modifier is not compatible with inherit, falling back to no-inherit.\n"); 2793 attr.inherit = false; 2794 attr.sample_type = 0; 2795 2796 attr.read_format = PERF_FORMAT_LOST; 2797 if (has_attr_feature(&attr, /*flags=*/0)) 2798 goto found; 2799 perf_missing_features.read_lost = true; 2800 pr_debug2("switching off PERF_FORMAT_LOST support\n"); 2801 attr.read_format = 0; 2802 2803 attr.sample_type = PERF_SAMPLE_WEIGHT_STRUCT; 2804 if (has_attr_feature(&attr, /*flags=*/0)) 2805 goto found; 2806 perf_missing_features.weight_struct = true; 2807 pr_debug2("switching off weight struct support\n"); 2808 attr.sample_type = 0; 2809 2810 attr.sample_type = PERF_SAMPLE_CODE_PAGE_SIZE; 2811 if (has_attr_feature(&attr, /*flags=*/0)) 2812 goto found; 2813 perf_missing_features.code_page_size = true; 2814 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support\n"); 2815 attr.sample_type = 0; 2816 2817 attr.sample_type = PERF_SAMPLE_DATA_PAGE_SIZE; 2818 if (has_attr_feature(&attr, /*flags=*/0)) 2819 goto found; 2820 perf_missing_features.data_page_size = true; 2821 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support\n"); 2822 attr.sample_type = 0; 2823 2824 attr.cgroup = 1; 2825 if (has_attr_feature(&attr, /*flags=*/0)) 2826 goto found; 2827 perf_missing_features.cgroup = true; 2828 pr_debug2_peo("Kernel has no cgroup sampling support\n"); 2829 attr.cgroup = 0; 2830 2831 attr.aux_output = 1; 2832 if (has_attr_feature(&attr, /*flags=*/0)) 2833 goto found; 2834 perf_missing_features.aux_output = true; 2835 pr_debug2_peo("Kernel has no attr.aux_output support\n"); 2836 attr.aux_output = 0; 2837 2838 attr.bpf_event = 1; 2839 if (has_attr_feature(&attr, /*flags=*/0)) 2840 goto found; 2841 perf_missing_features.bpf = true; 2842 pr_debug2_peo("switching off bpf_event\n"); 2843 attr.bpf_event = 0; 2844 2845 attr.ksymbol = 1; 2846 if (has_attr_feature(&attr, /*flags=*/0)) 2847 goto found; 2848 perf_missing_features.ksymbol = true; 2849 pr_debug2_peo("switching off ksymbol\n"); 2850 attr.ksymbol = 0; 2851 2852 attr.write_backward = 1; 2853 if (has_attr_feature(&attr, /*flags=*/0)) 2854 goto found; 2855 perf_missing_features.write_backward = true; 2856 pr_debug2_peo("switching off write_backward\n"); 2857 attr.write_backward = 0; 2858 2859 attr.use_clockid = 1; 2860 attr.clockid = CLOCK_MONOTONIC; 2861 if (has_attr_feature(&attr, /*flags=*/0)) 2862 goto found; 2863 perf_missing_features.clockid = true; 2864 pr_debug2_peo("switching off clockid\n"); 2865 attr.use_clockid = 0; 2866 attr.clockid = 0; 2867 2868 if (has_attr_feature(&attr, /*flags=*/PERF_FLAG_FD_CLOEXEC)) 2869 goto found; 2870 perf_missing_features.cloexec = true; 2871 pr_debug2_peo("switching off cloexec flag\n"); 2872 2873 attr.mmap2 = 1; 2874 if (has_attr_feature(&attr, /*flags=*/0)) 2875 goto found; 2876 perf_missing_features.mmap2 = true; 2877 pr_debug2_peo("switching off mmap2\n"); 2878 attr.mmap2 = 0; 2879 2880 /* set this unconditionally? */ 2881 perf_missing_features.sample_id_all = true; 2882 pr_debug2_peo("switching off sample_id_all\n"); 2883 2884 attr.inherit = 1; 2885 attr.read_format = PERF_FORMAT_GROUP; 2886 if (has_attr_feature(&attr, /*flags=*/0)) 2887 goto found; 2888 perf_missing_features.group_read = true; 2889 pr_debug2_peo("switching off group read\n"); 2890 attr.inherit = 0; 2891 attr.read_format = 0; 2892 2893 found: 2894 detection_done = true; 2895 errno = old_errno; 2896 2897 check: 2898 if ((evsel->core.attr.defer_callchain || evsel->core.attr.defer_output) && 2899 perf_missing_features.defer_callchain) 2900 return true; 2901 2902 if (evsel->core.attr.inherit && 2903 (evsel->core.attr.sample_type & PERF_SAMPLE_READ) && 2904 perf_missing_features.inherit_sample_read) 2905 return true; 2906 2907 if ((evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS) && 2908 perf_missing_features.branch_counters) 2909 return true; 2910 2911 if ((evsel->core.attr.read_format & PERF_FORMAT_LOST) && 2912 perf_missing_features.read_lost) 2913 return true; 2914 2915 if ((evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT) && 2916 perf_missing_features.weight_struct) 2917 return true; 2918 2919 if (evsel->core.attr.use_clockid && evsel->core.attr.clockid != CLOCK_MONOTONIC && 2920 !perf_missing_features.clockid) { 2921 perf_missing_features.clockid_wrong = true; 2922 return true; 2923 } 2924 2925 if (evsel->core.attr.use_clockid && perf_missing_features.clockid) 2926 return true; 2927 2928 if ((evsel->open_flags & PERF_FLAG_FD_CLOEXEC) && 2929 perf_missing_features.cloexec) 2930 return true; 2931 2932 if (evsel->core.attr.mmap2 && perf_missing_features.mmap2) 2933 return true; 2934 2935 if ((evsel->core.attr.branch_sample_type & (PERF_SAMPLE_BRANCH_NO_FLAGS | 2936 PERF_SAMPLE_BRANCH_NO_CYCLES)) && 2937 perf_missing_features.lbr_flags) 2938 return true; 2939 2940 if (evsel->core.attr.inherit && (evsel->core.attr.read_format & PERF_FORMAT_GROUP) && 2941 perf_missing_features.group_read) 2942 return true; 2943 2944 if (evsel->core.attr.ksymbol && perf_missing_features.ksymbol) 2945 return true; 2946 2947 if (evsel->core.attr.bpf_event && perf_missing_features.bpf) 2948 return true; 2949 2950 if ((evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX) && 2951 perf_missing_features.branch_hw_idx) 2952 return true; 2953 2954 if (evsel->core.attr.sample_id_all && perf_missing_features.sample_id_all) 2955 return true; 2956 2957 return false; 2958 } 2959 2960 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, 2961 struct perf_thread_map *threads, 2962 int start_cpu_map_idx, int end_cpu_map_idx) 2963 { 2964 int idx, thread, nthreads; 2965 int pid = -1, err, old_errno; 2966 enum rlimit_action set_rlimit = NO_CHANGE; 2967 struct perf_cpu cpu; 2968 2969 if (evsel__is_retire_lat(evsel)) { 2970 err = evsel__tpebs_open(evsel); 2971 goto out; 2972 } 2973 2974 err = __evsel__prepare_open(evsel, cpus, threads); 2975 if (err) 2976 goto out; 2977 2978 if (cpus == NULL) 2979 cpus = empty_cpu_map; 2980 2981 if (threads == NULL) 2982 threads = empty_thread_map; 2983 2984 nthreads = perf_thread_map__nr(threads); 2985 2986 if (evsel->cgrp) 2987 pid = evsel->cgrp->fd; 2988 2989 fallback_missing_features: 2990 evsel__disable_missing_features(evsel); 2991 2992 pr_debug3("Opening: %s\n", evsel__name(evsel)); 2993 display_attr(&evsel->core.attr); 2994 2995 if (evsel__is_tool(evsel)) { 2996 err = evsel__tool_pmu_open(evsel, threads, 2997 start_cpu_map_idx, 2998 end_cpu_map_idx); 2999 goto out; 3000 } 3001 if (evsel__is_hwmon(evsel)) { 3002 err = evsel__hwmon_pmu_open(evsel, threads, 3003 start_cpu_map_idx, 3004 end_cpu_map_idx); 3005 goto out; 3006 } 3007 if (evsel__is_drm(evsel)) { 3008 err = evsel__drm_pmu_open(evsel, threads, 3009 start_cpu_map_idx, 3010 end_cpu_map_idx); 3011 goto out; 3012 } 3013 3014 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) { 3015 cpu = perf_cpu_map__cpu(cpus, idx); 3016 3017 for (thread = 0; thread < nthreads; thread++) { 3018 int fd, group_fd; 3019 retry_open: 3020 if (thread >= nthreads) 3021 break; 3022 3023 if (!evsel->cgrp && !evsel->core.system_wide) 3024 pid = perf_thread_map__pid(threads, thread); 3025 3026 group_fd = get_group_fd(evsel, idx, thread); 3027 3028 if (group_fd == -2) { 3029 pr_debug("broken group leader for %s\n", evsel->name); 3030 err = -EINVAL; 3031 goto out_close; 3032 } 3033 3034 /* Debug message used by test scripts */ 3035 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx", 3036 pid, cpu.cpu, group_fd, evsel->open_flags); 3037 3038 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu.cpu, 3039 group_fd, evsel->open_flags); 3040 3041 FD(evsel, idx, thread) = fd; 3042 3043 if (fd < 0) { 3044 err = -errno; 3045 3046 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", 3047 err); 3048 goto try_fallback; 3049 } 3050 3051 bpf_counter__install_pe(evsel, idx, fd); 3052 3053 if (unlikely(test_attr__enabled())) { 3054 test_attr__open(&evsel->core.attr, pid, cpu, 3055 fd, group_fd, evsel->open_flags); 3056 } 3057 3058 /* Debug message used by test scripts */ 3059 pr_debug2_peo(" = %d\n", fd); 3060 3061 if (evsel->bpf_fd >= 0) { 3062 int evt_fd = fd; 3063 int bpf_fd = evsel->bpf_fd; 3064 3065 err = ioctl(evt_fd, 3066 PERF_EVENT_IOC_SET_BPF, 3067 bpf_fd); 3068 if (err && errno != EEXIST) { 3069 pr_err("failed to attach bpf fd %d: %m\n", 3070 bpf_fd); 3071 err = -EINVAL; 3072 goto out_close; 3073 } 3074 } 3075 3076 set_rlimit = NO_CHANGE; 3077 3078 /* 3079 * If we succeeded but had to kill clockid, fail and 3080 * have evsel__open_strerror() print us a nice error. 3081 */ 3082 if (perf_missing_features.clockid || 3083 perf_missing_features.clockid_wrong) { 3084 err = -EINVAL; 3085 goto out_close; 3086 } 3087 } 3088 } 3089 3090 err = 0; 3091 goto out; 3092 3093 try_fallback: 3094 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus), 3095 idx, threads, thread, err)) { 3096 /* We just removed 1 thread, so lower the upper nthreads limit. */ 3097 nthreads--; 3098 3099 /* ... and pretend like nothing have happened. */ 3100 err = 0; 3101 goto retry_open; 3102 } 3103 /* 3104 * perf stat needs between 5 and 22 fds per CPU. When we run out 3105 * of them try to increase the limits. 3106 */ 3107 if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit)) 3108 goto retry_open; 3109 3110 if (err == -EINVAL && evsel__detect_missing_features(evsel, cpu)) 3111 goto fallback_missing_features; 3112 3113 if (evsel__precise_ip_fallback(evsel)) 3114 goto retry_open; 3115 3116 out_close: 3117 if (err) 3118 threads->err_thread = thread; 3119 3120 old_errno = errno; 3121 do { 3122 while (--thread >= 0) { 3123 if (FD(evsel, idx, thread) >= 0) 3124 close(FD(evsel, idx, thread)); 3125 FD(evsel, idx, thread) = -1; 3126 } 3127 thread = nthreads; 3128 } while (--idx >= 0); 3129 errno = old_errno; 3130 out: 3131 if (err) 3132 evsel->supported = false; 3133 return err; 3134 } 3135 3136 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus, 3137 struct perf_thread_map *threads) 3138 { 3139 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus)); 3140 } 3141 3142 void evsel__close(struct evsel *evsel) 3143 { 3144 if (evsel__is_retire_lat(evsel)) 3145 evsel__tpebs_close(evsel); 3146 perf_evsel__close(&evsel->core); 3147 perf_evsel__free_id(&evsel->core); 3148 } 3149 3150 int evsel__open_per_cpu_and_thread(struct evsel *evsel, 3151 struct perf_cpu_map *cpus, int cpu_map_idx, 3152 struct perf_thread_map *threads) 3153 { 3154 if (cpu_map_idx == -1) 3155 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus)); 3156 3157 return evsel__open_cpu(evsel, cpus, threads, cpu_map_idx, cpu_map_idx + 1); 3158 } 3159 3160 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx) 3161 { 3162 struct perf_thread_map *threads = thread_map__new_by_tid(-1); 3163 int ret = evsel__open_per_cpu_and_thread(evsel, cpus, cpu_map_idx, threads); 3164 3165 perf_thread_map__put(threads); 3166 return ret; 3167 } 3168 3169 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads) 3170 { 3171 struct perf_cpu_map *cpus = perf_cpu_map__new_any_cpu(); 3172 int ret = evsel__open_per_cpu_and_thread(evsel, cpus, -1, threads); 3173 3174 perf_cpu_map__put(cpus); 3175 return ret; 3176 } 3177 3178 static int perf_evsel__parse_id_sample(const union perf_event *event, 3179 struct perf_sample *sample) 3180 { 3181 const struct evsel *evsel = sample->evsel; 3182 u64 type = evsel->core.attr.sample_type; 3183 const __u64 *array = event->sample.array; 3184 bool swapped = evsel->needs_swap; 3185 union u64_swap u; 3186 int i = ((event->header.size - sizeof(event->header)) / sizeof(u64)) - 1; 3187 3188 if (type & PERF_SAMPLE_IDENTIFIER) { 3189 if (i < 0) 3190 return -EFAULT; 3191 3192 sample->id = array[i--]; 3193 } 3194 3195 if (type & PERF_SAMPLE_CPU) { 3196 if (i < 0) 3197 return -EFAULT; 3198 3199 u.val64 = array[i--]; 3200 if (swapped) { 3201 /* undo swap of u64, then swap on individual u32s */ 3202 u.val64 = bswap_64(u.val64); 3203 u.val32[0] = bswap_32(u.val32[0]); 3204 } 3205 sample->cpu = u.val32[0]; 3206 } 3207 3208 if (type & PERF_SAMPLE_STREAM_ID) { 3209 if (i < 0) 3210 return -EFAULT; 3211 3212 sample->stream_id = array[i--]; 3213 } 3214 3215 if (type & PERF_SAMPLE_ID) { 3216 if (i < 0) 3217 return -EFAULT; 3218 3219 sample->id = array[i--]; 3220 } 3221 3222 if (type & PERF_SAMPLE_TIME) { 3223 if (i < 0) 3224 return -EFAULT; 3225 3226 sample->time = array[i--]; 3227 } 3228 3229 if (type & PERF_SAMPLE_TID) { 3230 if (i < 0) 3231 return -EFAULT; 3232 3233 u.val64 = array[i--]; 3234 if (swapped) { 3235 /* undo swap of u64, then swap on individual u32s */ 3236 u.val64 = bswap_64(u.val64); 3237 u.val32[0] = bswap_32(u.val32[0]); 3238 u.val32[1] = bswap_32(u.val32[1]); 3239 } 3240 3241 sample->pid = u.val32[0]; 3242 sample->tid = u.val32[1]; 3243 } 3244 3245 return 0; 3246 } 3247 3248 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 3249 u64 size) 3250 { 3251 return size > max_size || offset + size > endp; 3252 } 3253 3254 #define OVERFLOW_CHECK(offset, size, max_size) \ 3255 do { \ 3256 if (overflow(endp, (max_size), (offset), (size))) \ 3257 goto out_efault; \ 3258 } while (0) 3259 3260 #define OVERFLOW_CHECK_u64(offset) \ 3261 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 3262 3263 static int 3264 perf_event__check_size(union perf_event *event, unsigned int sample_size) 3265 { 3266 /* 3267 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 3268 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 3269 * check the format does not go past the end of the event. 3270 */ 3271 if (sample_size + sizeof(event->header) > event->header.size) 3272 return -EFAULT; 3273 3274 return 0; 3275 } 3276 3277 static void perf_parse_sample_weight(struct perf_sample *data, const __u64 *array, u64 type) 3278 { 3279 union perf_sample_weight weight; 3280 3281 weight.full = *array; 3282 if (type & PERF_SAMPLE_WEIGHT_STRUCT) { 3283 data->weight = weight.var1_dw; 3284 data->ins_lat = weight.var2_w; 3285 data->weight3 = weight.var3_w; 3286 } else { 3287 data->weight = weight.full; 3288 } 3289 } 3290 3291 u64 evsel__bitfield_swap_branch_flags(u64 value) 3292 { 3293 u64 new_val = 0; 3294 3295 /* 3296 * branch_flags 3297 * union { 3298 * u64 values; 3299 * struct { 3300 * mispred:1 //target mispredicted 3301 * predicted:1 //target predicted 3302 * in_tx:1 //in transaction 3303 * abort:1 //transaction abort 3304 * cycles:16 //cycle count to last branch 3305 * type:4 //branch type 3306 * spec:2 //branch speculation info 3307 * new_type:4 //additional branch type 3308 * priv:3 //privilege level 3309 * reserved:31 3310 * } 3311 * } 3312 * 3313 * Avoid bswap64() the entire branch_flag.value, 3314 * as it has variable bit-field sizes. Instead the 3315 * macro takes the bit-field position/size, 3316 * swaps it based on the host endianness. 3317 */ 3318 if (host_is_bigendian()) { 3319 new_val = bitfield_swap(value, 0, 1); 3320 new_val |= bitfield_swap(value, 1, 1); 3321 new_val |= bitfield_swap(value, 2, 1); 3322 new_val |= bitfield_swap(value, 3, 1); 3323 new_val |= bitfield_swap(value, 4, 16); 3324 new_val |= bitfield_swap(value, 20, 4); 3325 new_val |= bitfield_swap(value, 24, 2); 3326 new_val |= bitfield_swap(value, 26, 4); 3327 new_val |= bitfield_swap(value, 30, 3); 3328 new_val |= bitfield_swap(value, 33, 31); 3329 } else { 3330 new_val = bitfield_swap(value, 63, 1); 3331 new_val |= bitfield_swap(value, 62, 1); 3332 new_val |= bitfield_swap(value, 61, 1); 3333 new_val |= bitfield_swap(value, 60, 1); 3334 new_val |= bitfield_swap(value, 44, 16); 3335 new_val |= bitfield_swap(value, 40, 4); 3336 new_val |= bitfield_swap(value, 38, 2); 3337 new_val |= bitfield_swap(value, 34, 4); 3338 new_val |= bitfield_swap(value, 31, 3); 3339 new_val |= bitfield_swap(value, 0, 31); 3340 } 3341 3342 return new_val; 3343 } 3344 3345 static inline bool evsel__has_branch_counters(const struct evsel *evsel) 3346 { 3347 struct evsel *leader = evsel__leader(evsel); 3348 3349 /* The branch counters feature only supports group */ 3350 if (!leader || !evsel->evlist) 3351 return false; 3352 3353 if (evsel->evlist->nr_br_cntr < 0) 3354 evlist__update_br_cntr(evsel->evlist); 3355 3356 if (leader->br_cntr_nr > 0) 3357 return true; 3358 3359 return false; 3360 } 3361 3362 static int __set_offcpu_sample(struct perf_sample *data) 3363 { 3364 u64 *array = data->raw_data; 3365 u32 max_size = data->raw_size, *p32; 3366 const void *endp = (void *)array + max_size; 3367 3368 if (array == NULL) 3369 return -EFAULT; 3370 3371 OVERFLOW_CHECK_u64(array); 3372 p32 = (void *)array++; 3373 data->pid = p32[0]; 3374 data->tid = p32[1]; 3375 3376 OVERFLOW_CHECK_u64(array); 3377 data->period = *array++; 3378 3379 OVERFLOW_CHECK_u64(array); 3380 data->callchain = (struct ip_callchain *)array++; 3381 OVERFLOW_CHECK(array, data->callchain->nr * sizeof(u64), max_size); 3382 data->ip = data->callchain->ips[1]; 3383 array += data->callchain->nr; 3384 3385 OVERFLOW_CHECK_u64(array); 3386 data->cgroup = *array; 3387 3388 return 0; 3389 out_efault: 3390 return -EFAULT; 3391 } 3392 3393 int __evsel__parse_sample(struct evsel *evsel, union perf_event *event, 3394 struct perf_sample *data, bool needs_swap) 3395 { 3396 u64 type = evsel->core.attr.sample_type; 3397 bool swapped = needs_swap; 3398 const __u64 *array; 3399 u16 max_size = event->header.size; 3400 const void *endp = (void *)event + max_size; 3401 u64 sz; 3402 3403 /* 3404 * used for cross-endian analysis. See git commit 65014ab3 3405 * for why this goofiness is needed. 3406 */ 3407 union u64_swap u; 3408 3409 perf_sample__init(data, /*all=*/true); 3410 data->evsel = evsel; 3411 data->cpu = data->pid = data->tid = -1; 3412 data->stream_id = data->id = data->time = -1ULL; 3413 data->period = evsel->core.attr.sample_period; 3414 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 3415 data->misc = event->header.misc; 3416 data->data_src = PERF_MEM_DATA_SRC_NONE; 3417 data->vcpu = -1; 3418 3419 if (event->header.type == PERF_RECORD_CALLCHAIN_DEFERRED) { 3420 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 3421 3422 data->callchain = (struct ip_callchain *)&event->callchain_deferred.nr; 3423 if (data->callchain->nr > max_callchain_nr) 3424 goto out_efault; 3425 3426 data->deferred_cookie = event->callchain_deferred.cookie; 3427 3428 if (evsel->core.attr.sample_id_all) { 3429 if (perf_evsel__parse_id_sample(event, data)) 3430 goto out_efault; 3431 } 3432 return 0; 3433 } 3434 3435 if (event->header.type != PERF_RECORD_SAMPLE) { 3436 if (evsel->core.attr.sample_id_all) { 3437 if (perf_evsel__parse_id_sample(event, data)) 3438 goto out_efault; 3439 } 3440 return 0; 3441 } 3442 3443 array = event->sample.array; 3444 3445 if (perf_event__check_size(event, evsel->sample_size)) 3446 goto out_efault; 3447 3448 if (type & PERF_SAMPLE_IDENTIFIER) { 3449 data->id = *array; 3450 array++; 3451 } 3452 3453 if (type & PERF_SAMPLE_IP) { 3454 data->ip = *array; 3455 array++; 3456 } 3457 3458 if (type & PERF_SAMPLE_TID) { 3459 u.val64 = *array; 3460 if (swapped) { 3461 /* undo swap of u64, then swap on individual u32s */ 3462 u.val64 = bswap_64(u.val64); 3463 u.val32[0] = bswap_32(u.val32[0]); 3464 u.val32[1] = bswap_32(u.val32[1]); 3465 } 3466 3467 data->pid = u.val32[0]; 3468 data->tid = u.val32[1]; 3469 array++; 3470 } 3471 3472 if (type & PERF_SAMPLE_TIME) { 3473 data->time = *array; 3474 array++; 3475 } 3476 3477 if (type & PERF_SAMPLE_ADDR) { 3478 data->addr = *array; 3479 array++; 3480 } 3481 3482 if (type & PERF_SAMPLE_ID) { 3483 data->id = *array; 3484 array++; 3485 } 3486 3487 if (type & PERF_SAMPLE_STREAM_ID) { 3488 data->stream_id = *array; 3489 array++; 3490 } 3491 3492 if (type & PERF_SAMPLE_CPU) { 3493 3494 u.val64 = *array; 3495 if (swapped) { 3496 /* undo swap of u64, then swap on individual u32s */ 3497 u.val64 = bswap_64(u.val64); 3498 u.val32[0] = bswap_32(u.val32[0]); 3499 } 3500 3501 data->cpu = u.val32[0]; 3502 array++; 3503 } 3504 3505 if (type & PERF_SAMPLE_PERIOD) { 3506 data->period = *array; 3507 array++; 3508 } 3509 3510 if (type & PERF_SAMPLE_READ) { 3511 u64 read_format = evsel->core.attr.read_format; 3512 3513 OVERFLOW_CHECK_u64(array); 3514 if (read_format & PERF_FORMAT_GROUP) 3515 data->read.group.nr = *array; 3516 else 3517 data->read.one.value = *array; 3518 3519 array++; 3520 3521 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 3522 OVERFLOW_CHECK_u64(array); 3523 data->read.time_enabled = *array; 3524 array++; 3525 } 3526 3527 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 3528 OVERFLOW_CHECK_u64(array); 3529 data->read.time_running = *array; 3530 array++; 3531 } 3532 3533 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 3534 if (read_format & PERF_FORMAT_GROUP) { 3535 const u64 max_group_nr = UINT64_MAX / 3536 sizeof(struct sample_read_value); 3537 3538 if (data->read.group.nr > max_group_nr) 3539 goto out_efault; 3540 3541 sz = data->read.group.nr * sample_read_value_size(read_format); 3542 OVERFLOW_CHECK(array, sz, max_size); 3543 data->read.group.values = 3544 (struct sample_read_value *)array; 3545 array = (void *)array + sz; 3546 } else { 3547 OVERFLOW_CHECK_u64(array); 3548 data->read.one.id = *array; 3549 array++; 3550 3551 if (read_format & PERF_FORMAT_LOST) { 3552 OVERFLOW_CHECK_u64(array); 3553 data->read.one.lost = *array; 3554 array++; 3555 } 3556 } 3557 } 3558 3559 if (type & PERF_SAMPLE_CALLCHAIN) { 3560 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 3561 u64 callchain_nr; 3562 3563 OVERFLOW_CHECK_u64(array); 3564 data->callchain = (struct ip_callchain *)array++; 3565 callchain_nr = data->callchain->nr; 3566 if (callchain_nr > max_callchain_nr) 3567 goto out_efault; 3568 sz = callchain_nr * sizeof(u64); 3569 /* 3570 * Save the cookie for the deferred user callchain. The last 2 3571 * entries in the callchain should be the context marker and the 3572 * cookie. The cookie will be used to match PERF_RECORD_ 3573 * CALLCHAIN_DEFERRED later. 3574 */ 3575 if (evsel->core.attr.defer_callchain && callchain_nr >= 2 && 3576 data->callchain->ips[callchain_nr - 2] == PERF_CONTEXT_USER_DEFERRED) { 3577 data->deferred_cookie = data->callchain->ips[callchain_nr - 1]; 3578 data->deferred_callchain = true; 3579 } 3580 OVERFLOW_CHECK(array, sz, max_size); 3581 array = (void *)array + sz; 3582 } 3583 3584 if (type & PERF_SAMPLE_RAW) { 3585 OVERFLOW_CHECK_u64(array); 3586 u.val64 = *array; 3587 3588 /* 3589 * Undo swap of u64, then swap on individual u32s, 3590 * get the size of the raw area and undo all of the 3591 * swap. The pevent interface handles endianness by 3592 * itself. 3593 */ 3594 if (swapped) { 3595 u.val64 = bswap_64(u.val64); 3596 u.val32[0] = bswap_32(u.val32[0]); 3597 u.val32[1] = bswap_32(u.val32[1]); 3598 } 3599 data->raw_size = u.val32[0]; 3600 3601 /* 3602 * The raw data is aligned on 64bits including the 3603 * u32 size, so it's safe to use mem_bswap_64. 3604 */ 3605 if (swapped) 3606 mem_bswap_64((void *) array, data->raw_size); 3607 3608 array = (void *)array + sizeof(u32); 3609 3610 OVERFLOW_CHECK(array, data->raw_size, max_size); 3611 data->raw_data = (void *)array; 3612 array = (void *)array + data->raw_size; 3613 } 3614 3615 if (type & PERF_SAMPLE_BRANCH_STACK) { 3616 const u64 max_branch_nr = UINT64_MAX / 3617 sizeof(struct branch_entry); 3618 struct branch_entry *e; 3619 unsigned int i; 3620 3621 OVERFLOW_CHECK_u64(array); 3622 data->branch_stack = (struct branch_stack *)array++; 3623 3624 if (data->branch_stack->nr > max_branch_nr) 3625 goto out_efault; 3626 3627 sz = data->branch_stack->nr * sizeof(struct branch_entry); 3628 if (evsel__has_branch_hw_idx(evsel)) { 3629 sz += sizeof(u64); 3630 e = &data->branch_stack->entries[0]; 3631 } else { 3632 data->no_hw_idx = true; 3633 /* 3634 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied, 3635 * only nr and entries[] will be output by kernel. 3636 */ 3637 e = (struct branch_entry *)&data->branch_stack->hw_idx; 3638 } 3639 3640 if (swapped) { 3641 /* 3642 * struct branch_flag does not have endian 3643 * specific bit field definition. And bswap 3644 * will not resolve the issue, since these 3645 * are bit fields. 3646 * 3647 * evsel__bitfield_swap_branch_flags() uses a 3648 * bitfield_swap macro to swap the bit position 3649 * based on the host endians. 3650 */ 3651 for (i = 0; i < data->branch_stack->nr; i++, e++) 3652 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value); 3653 } 3654 3655 OVERFLOW_CHECK(array, sz, max_size); 3656 array = (void *)array + sz; 3657 3658 if (evsel__has_branch_counters(evsel)) { 3659 data->branch_stack_cntr = (u64 *)array; 3660 sz = data->branch_stack->nr * sizeof(u64); 3661 3662 OVERFLOW_CHECK(array, sz, max_size); 3663 array = (void *)array + sz; 3664 } 3665 } 3666 3667 if (type & PERF_SAMPLE_REGS_USER) { 3668 struct regs_dump *regs = perf_sample__user_regs(data); 3669 3670 OVERFLOW_CHECK_u64(array); 3671 regs->abi = *array; 3672 array++; 3673 3674 if (regs->abi) { 3675 u64 mask = evsel->core.attr.sample_regs_user; 3676 3677 sz = hweight64(mask) * sizeof(u64); 3678 OVERFLOW_CHECK(array, sz, max_size); 3679 regs->mask = mask; 3680 regs->regs = (u64 *)array; 3681 array = (void *)array + sz; 3682 } 3683 } 3684 3685 if (type & PERF_SAMPLE_STACK_USER) { 3686 OVERFLOW_CHECK_u64(array); 3687 sz = *array++; 3688 3689 data->user_stack.offset = ((char *)(array - 1) 3690 - (char *) event); 3691 3692 if (!sz) { 3693 data->user_stack.size = 0; 3694 } else { 3695 OVERFLOW_CHECK(array, sz, max_size); 3696 data->user_stack.data = (char *)array; 3697 array = (void *)array + sz; 3698 OVERFLOW_CHECK_u64(array); 3699 data->user_stack.size = *array++; 3700 if (WARN_ONCE(data->user_stack.size > sz, 3701 "user stack dump failure\n")) 3702 goto out_efault; 3703 } 3704 } 3705 3706 if (type & PERF_SAMPLE_WEIGHT_TYPE) { 3707 OVERFLOW_CHECK_u64(array); 3708 perf_parse_sample_weight(data, array, type); 3709 array++; 3710 } 3711 3712 if (type & PERF_SAMPLE_DATA_SRC) { 3713 OVERFLOW_CHECK_u64(array); 3714 data->data_src = *array; 3715 array++; 3716 } 3717 3718 if (type & PERF_SAMPLE_TRANSACTION) { 3719 OVERFLOW_CHECK_u64(array); 3720 data->transaction = *array; 3721 array++; 3722 } 3723 3724 if (type & PERF_SAMPLE_REGS_INTR) { 3725 struct regs_dump *regs = perf_sample__intr_regs(data); 3726 3727 OVERFLOW_CHECK_u64(array); 3728 regs->abi = *array; 3729 array++; 3730 3731 if (regs->abi != PERF_SAMPLE_REGS_ABI_NONE) { 3732 u64 mask = evsel->core.attr.sample_regs_intr; 3733 3734 sz = hweight64(mask) * sizeof(u64); 3735 OVERFLOW_CHECK(array, sz, max_size); 3736 regs->mask = mask; 3737 regs->regs = (u64 *)array; 3738 array = (void *)array + sz; 3739 } 3740 } 3741 3742 data->phys_addr = 0; 3743 if (type & PERF_SAMPLE_PHYS_ADDR) { 3744 data->phys_addr = *array; 3745 array++; 3746 } 3747 3748 data->cgroup = 0; 3749 if (type & PERF_SAMPLE_CGROUP) { 3750 data->cgroup = *array; 3751 array++; 3752 } 3753 3754 data->data_page_size = 0; 3755 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) { 3756 data->data_page_size = *array; 3757 array++; 3758 } 3759 3760 data->code_page_size = 0; 3761 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) { 3762 data->code_page_size = *array; 3763 array++; 3764 } 3765 3766 if (type & PERF_SAMPLE_AUX) { 3767 OVERFLOW_CHECK_u64(array); 3768 sz = *array++; 3769 3770 OVERFLOW_CHECK(array, sz, max_size); 3771 /* Undo swap of data */ 3772 if (swapped) 3773 mem_bswap_64((char *)array, sz); 3774 data->aux_sample.size = sz; 3775 data->aux_sample.data = (char *)array; 3776 array = (void *)array + sz; 3777 } 3778 3779 if (evsel__is_offcpu_event(evsel)) { 3780 if (__set_offcpu_sample(data)) 3781 goto out_efault; 3782 } 3783 3784 return 0; 3785 out_efault: 3786 perf_sample__exit(data); 3787 return -EFAULT; 3788 } 3789 3790 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event, 3791 u64 *timestamp) 3792 { 3793 u64 type = evsel->core.attr.sample_type; 3794 const __u64 *array; 3795 3796 if (!(type & PERF_SAMPLE_TIME)) 3797 return -1; 3798 3799 if (event->header.type != PERF_RECORD_SAMPLE) { 3800 struct perf_sample data = { 3801 .evsel = evsel, 3802 .time = -1ULL, 3803 }; 3804 3805 if (!evsel->core.attr.sample_id_all) 3806 return -1; 3807 if (perf_evsel__parse_id_sample(event, &data)) 3808 return -1; 3809 3810 *timestamp = data.time; 3811 return 0; 3812 } 3813 3814 array = event->sample.array; 3815 3816 if (perf_event__check_size(event, evsel->sample_size)) 3817 return -EFAULT; 3818 3819 if (type & PERF_SAMPLE_IDENTIFIER) 3820 array++; 3821 3822 if (type & PERF_SAMPLE_IP) 3823 array++; 3824 3825 if (type & PERF_SAMPLE_TID) 3826 array++; 3827 3828 if (type & PERF_SAMPLE_TIME) 3829 *timestamp = *array; 3830 3831 return 0; 3832 } 3833 3834 u16 evsel__id_hdr_size(const struct evsel *evsel) 3835 { 3836 u64 sample_type = evsel->core.attr.sample_type; 3837 u16 size = 0; 3838 3839 if (sample_type & PERF_SAMPLE_TID) 3840 size += sizeof(u64); 3841 3842 if (sample_type & PERF_SAMPLE_TIME) 3843 size += sizeof(u64); 3844 3845 if (sample_type & PERF_SAMPLE_ID) 3846 size += sizeof(u64); 3847 3848 if (sample_type & PERF_SAMPLE_STREAM_ID) 3849 size += sizeof(u64); 3850 3851 if (sample_type & PERF_SAMPLE_CPU) 3852 size += sizeof(u64); 3853 3854 if (sample_type & PERF_SAMPLE_IDENTIFIER) 3855 size += sizeof(u64); 3856 3857 return size; 3858 } 3859 3860 #ifdef HAVE_LIBTRACEEVENT 3861 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name) 3862 { 3863 struct tep_event *tp_format = evsel__tp_format(evsel); 3864 3865 return tp_format ? tep_find_field(tp_format, name) : NULL; 3866 } 3867 3868 struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name) 3869 { 3870 struct tep_event *tp_format = evsel__tp_format(evsel); 3871 3872 return tp_format ? tep_find_common_field(tp_format, name) : NULL; 3873 } 3874 3875 static bool out_of_bounds(const struct tep_format_field *field, int offset, int size, u32 raw_size) 3876 { 3877 if (offset < 0) { 3878 pr_warning("Negative trace point field offset %d in %s\n", 3879 offset, field->name); 3880 return true; 3881 } 3882 if (size < 0) { 3883 pr_warning("Negative trace point field size %d in %s\n", 3884 size, field->name); 3885 return true; 3886 } 3887 if ((u32)offset + (u32)size > raw_size) { 3888 pr_warning("Out of bound tracepoint field (%s) offset %d size %d in %u\n", 3889 field->name, offset, size, raw_size); 3890 return true; 3891 } 3892 return false; 3893 } 3894 3895 void *perf_sample__rawptr(struct perf_sample *sample, const char *name) 3896 { 3897 struct tep_format_field *field = evsel__field(sample->evsel, name); 3898 int offset, size; 3899 3900 if (!field) 3901 return NULL; 3902 3903 offset = field->offset; 3904 size = field->size; 3905 if (field->flags & TEP_FIELD_IS_DYNAMIC) { 3906 int dynamic_data; 3907 3908 if (out_of_bounds(field, offset, 4, sample->raw_size)) 3909 return NULL; 3910 3911 dynamic_data = *(int *)(sample->raw_data + field->offset); 3912 3913 if (sample->evsel->needs_swap) 3914 dynamic_data = bswap_32(dynamic_data); 3915 3916 offset = dynamic_data & 0xffff; 3917 size = (dynamic_data >> 16) & 0xffff; 3918 3919 if (tep_field_is_relative(field->flags)) { 3920 /* 3921 * Newer kernel feature: Relative offsets (__rel_loc). 3922 * If the relative flag is set, the parsed offset is not 3923 * absolute from the start of the record. Instead, it is 3924 * relative to the *end* of the dynamic field descriptor 3925 * itself. 3926 */ 3927 offset += field->offset + field->size; 3928 } 3929 } 3930 if (out_of_bounds(field, offset, size, sample->raw_size)) 3931 return NULL; 3932 3933 return sample->raw_data + offset; 3934 } 3935 3936 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample, 3937 bool needs_swap) 3938 { 3939 u64 value; 3940 void *ptr = sample->raw_data + field->offset; 3941 3942 if (out_of_bounds(field, field->offset, field->size, sample->raw_size)) 3943 return 0; 3944 3945 switch (field->size) { 3946 case 1: 3947 return *(u8 *)ptr; 3948 case 2: 3949 value = *(u16 *)ptr; 3950 break; 3951 case 4: 3952 value = *(u32 *)ptr; 3953 break; 3954 case 8: 3955 memcpy(&value, ptr, sizeof(u64)); 3956 break; 3957 default: 3958 return 0; 3959 } 3960 3961 if (!needs_swap) 3962 return value; 3963 3964 switch (field->size) { 3965 case 2: 3966 return bswap_16(value); 3967 case 4: 3968 return bswap_32(value); 3969 case 8: 3970 return bswap_64(value); 3971 default: 3972 return 0; 3973 } 3974 3975 return 0; 3976 } 3977 3978 u64 perf_sample__intval(struct perf_sample *sample, const char *name) 3979 { 3980 struct tep_format_field *field = evsel__field(sample->evsel, name); 3981 3982 return field ? format_field__intval(field, sample, sample->evsel->needs_swap) : 0; 3983 } 3984 3985 u64 perf_sample__intval_common(struct perf_sample *sample, const char *name) 3986 { 3987 struct tep_format_field *field = evsel__common_field(sample->evsel, name); 3988 3989 return field ? format_field__intval(field, sample, sample->evsel->needs_swap) : 0; 3990 } 3991 3992 char perf_sample__taskstate(struct perf_sample *sample, const char *name) 3993 { 3994 static struct tep_format_field *prev_state_field; 3995 static const char *states; 3996 struct tep_format_field *field; 3997 unsigned long long val; 3998 unsigned int bit; 3999 char state = '?'; /* '?' denotes unknown task state */ 4000 4001 field = evsel__field(sample->evsel, name); 4002 4003 if (!field) 4004 return state; 4005 4006 if (!states || field != prev_state_field) { 4007 states = parse_task_states(field); 4008 if (!states) 4009 return state; 4010 prev_state_field = field; 4011 } 4012 4013 /* 4014 * Note since the kernel exposes TASK_REPORT_MAX to userspace 4015 * to denote the 'preempted' state, we might as welll report 4016 * 'R' for this case, which make senses to users as well. 4017 * 4018 * We can change this if we have a good reason in the future. 4019 */ 4020 val = perf_sample__intval(sample, name); 4021 bit = val ? ffs(val) : 0; 4022 state = (!bit || bit > strlen(states)) ? 'R' : states[bit-1]; 4023 return state; 4024 } 4025 #endif 4026 4027 bool evsel__fallback(struct evsel *evsel, struct target *target, int err, 4028 char *msg, size_t msgsize) 4029 { 4030 int paranoid; 4031 4032 if ((err == ENODEV || err == ENOENT || err == ENXIO) && 4033 evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) { 4034 /* 4035 * If it's the legacy hardware cycles event fails then fall back 4036 * to hrtimer based cpu-clock sw counter, which is always 4037 * available even if no PMU support. PPC returned ENXIO rather 4038 * than ENODEV or ENOENT until 2.6.37. 4039 */ 4040 evsel->pmu = perf_pmus__find_by_type(PERF_TYPE_SOFTWARE); 4041 assert(evsel->pmu); /* software is a "well-known" and can't fail PMU type. */ 4042 4043 /* Configure the event. */ 4044 evsel->core.attr.type = PERF_TYPE_SOFTWARE; 4045 evsel->core.attr.config = target__has_cpu(target) 4046 ? PERF_COUNT_SW_CPU_CLOCK 4047 : PERF_COUNT_SW_TASK_CLOCK; 4048 evsel->core.is_pmu_core = false; 4049 4050 /* Remove excludes for new event. */ 4051 if (evsel->fallenback_eacces) { 4052 evsel->core.attr.exclude_kernel = 0; 4053 evsel->core.attr.exclude_hv = 0; 4054 evsel->fallenback_eacces = false; 4055 } 4056 if (evsel->fallenback_eopnotsupp) { 4057 evsel->core.attr.exclude_guest = 0; 4058 evsel->fallenback_eopnotsupp = false; 4059 } 4060 4061 /* Name is recomputed by evsel__name. */ 4062 zfree(&evsel->name); 4063 4064 /* Log message. */ 4065 scnprintf(msg, msgsize, 4066 "The cycles event is not supported, trying to fall back to %s", 4067 evsel__name(evsel)); 4068 return true; 4069 } else if (err == EACCES && !evsel->core.attr.exclude_kernel && 4070 (paranoid = perf_event_paranoid()) > 1) { 4071 const char *name = evsel__name(evsel); 4072 char *new_name; 4073 const char *sep = ":"; 4074 4075 /* If event has exclude user then don't exclude kernel. */ 4076 if (evsel->core.attr.exclude_user) 4077 goto no_fallback; 4078 4079 /* Is there already the separator in the name. */ 4080 if (strchr(name, '/') || 4081 (strchr(name, ':') && !evsel->is_libpfm_event)) 4082 sep = ""; 4083 4084 if (asprintf(&new_name, "%s%su", name, sep) < 0) 4085 goto no_fallback; 4086 4087 free(evsel->name); 4088 evsel->name = new_name; 4089 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying " 4090 "to fall back to excluding kernel and hypervisor " 4091 " samples", paranoid); 4092 evsel->core.attr.exclude_kernel = 1; 4093 evsel->core.attr.exclude_hv = 1; 4094 evsel->fallenback_eacces = true; 4095 return true; 4096 } else if (err == EOPNOTSUPP && !evsel->core.attr.exclude_guest && 4097 !evsel->exclude_GH) { 4098 const char *name = evsel__name(evsel); 4099 char *new_name; 4100 const char *sep = ":"; 4101 4102 /* Is there already the separator in the name. */ 4103 if (strchr(name, '/') || 4104 (strchr(name, ':') && !evsel->is_libpfm_event)) 4105 sep = ""; 4106 4107 if (asprintf(&new_name, "%s%sH", name, sep) < 0) 4108 goto no_fallback; 4109 4110 free(evsel->name); 4111 evsel->name = new_name; 4112 /* Apple M1 requires exclude_guest */ 4113 scnprintf(msg, msgsize, "Trying to fall back to excluding guest samples"); 4114 evsel->core.attr.exclude_guest = 1; 4115 evsel->fallenback_eopnotsupp = true; 4116 return true; 4117 } 4118 no_fallback: 4119 scnprintf(msg, msgsize, "No fallback found for '%s' for error %d", 4120 evsel__name(evsel), err); 4121 return false; 4122 } 4123 4124 static bool find_process(const char *name) 4125 { 4126 size_t len = strlen(name); 4127 DIR *dir; 4128 struct dirent *d; 4129 int ret = -1; 4130 4131 dir = opendir(procfs__mountpoint()); 4132 if (!dir) 4133 return false; 4134 4135 /* Walk through the directory. */ 4136 while (ret && (d = readdir(dir)) != NULL) { 4137 char path[PATH_MAX]; 4138 char *data; 4139 size_t size; 4140 4141 if ((d->d_type != DT_DIR) || 4142 !strcmp(".", d->d_name) || 4143 !strcmp("..", d->d_name)) 4144 continue; 4145 4146 scnprintf(path, sizeof(path), "%s/%s/comm", 4147 procfs__mountpoint(), d->d_name); 4148 4149 if (filename__read_str(path, &data, &size)) 4150 continue; 4151 4152 ret = strncmp(name, data, len); 4153 free(data); 4154 } 4155 4156 closedir(dir); 4157 return ret ? false : true; 4158 } 4159 4160 static int dump_perf_event_processes(char *msg, size_t size) 4161 { 4162 DIR *proc_dir; 4163 struct dirent *proc_entry; 4164 int printed = 0; 4165 4166 proc_dir = opendir(procfs__mountpoint()); 4167 if (!proc_dir) 4168 return 0; 4169 4170 /* Walk through the /proc directory. */ 4171 while ((proc_entry = readdir(proc_dir)) != NULL) { 4172 char buf[256]; 4173 DIR *fd_dir; 4174 struct dirent *fd_entry; 4175 int fd_dir_fd; 4176 4177 if (proc_entry->d_type != DT_DIR || 4178 !isdigit(proc_entry->d_name[0]) || 4179 strlen(proc_entry->d_name) > sizeof(buf) - 4) 4180 continue; 4181 4182 scnprintf(buf, sizeof(buf), "%s/fd", proc_entry->d_name); 4183 fd_dir_fd = openat(dirfd(proc_dir), buf, O_DIRECTORY); 4184 if (fd_dir_fd == -1) 4185 continue; 4186 fd_dir = fdopendir(fd_dir_fd); 4187 if (!fd_dir) { 4188 close(fd_dir_fd); 4189 continue; 4190 } 4191 while ((fd_entry = readdir(fd_dir)) != NULL) { 4192 ssize_t link_size; 4193 4194 if (fd_entry->d_type != DT_LNK) 4195 continue; 4196 link_size = readlinkat(fd_dir_fd, fd_entry->d_name, buf, sizeof(buf)); 4197 if (link_size < 0) 4198 continue; 4199 /* Take care as readlink doesn't null terminate the string. */ 4200 if (!strncmp(buf, "anon_inode:[perf_event]", link_size)) { 4201 int cmdline_fd; 4202 ssize_t cmdline_size; 4203 4204 scnprintf(buf, sizeof(buf), "%s/cmdline", proc_entry->d_name); 4205 cmdline_fd = openat(dirfd(proc_dir), buf, O_RDONLY); 4206 if (cmdline_fd == -1) 4207 continue; 4208 cmdline_size = read(cmdline_fd, buf, sizeof(buf) - 1); 4209 close(cmdline_fd); 4210 if (cmdline_size < 0) 4211 continue; 4212 buf[cmdline_size] = '\0'; 4213 for (ssize_t i = 0; i < cmdline_size; i++) { 4214 if (buf[i] == '\0') 4215 buf[i] = ' '; 4216 } 4217 4218 if (printed == 0) 4219 printed += scnprintf(msg, size, "Possible processes:\n"); 4220 4221 printed += scnprintf(msg + printed, size - printed, 4222 "%s %s\n", proc_entry->d_name, buf); 4223 break; 4224 } 4225 } 4226 closedir(fd_dir); 4227 } 4228 closedir(proc_dir); 4229 return printed; 4230 } 4231 4232 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused, 4233 int err __maybe_unused, 4234 char *msg __maybe_unused, 4235 size_t size __maybe_unused) 4236 { 4237 return 0; 4238 } 4239 4240 int evsel__open_strerror(struct evsel *evsel, struct target *target, 4241 int err, char *msg, size_t size) 4242 { 4243 struct perf_pmu *pmu; 4244 int printed = 0, enforced = 0; 4245 int ret; 4246 4247 switch (err) { 4248 case EPERM: 4249 case EACCES: 4250 printed += scnprintf(msg + printed, size - printed, 4251 "Access to performance monitoring and observability operations is limited.\n"); 4252 4253 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) { 4254 if (enforced) { 4255 printed += scnprintf(msg + printed, size - printed, 4256 "Enforced MAC policy settings (SELinux) can limit access to performance\n" 4257 "monitoring and observability operations. Inspect system audit records for\n" 4258 "more perf_event access control information and adjusting the policy.\n"); 4259 } 4260 } 4261 4262 if (err == EPERM) 4263 printed += scnprintf(msg, size, 4264 "No permission to enable %s event.\n\n", evsel__name(evsel)); 4265 4266 return printed + scnprintf(msg + printed, size - printed, 4267 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n" 4268 "access to performance monitoring and observability operations for processes\n" 4269 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n" 4270 "More information can be found at 'Perf events and tool security' document:\n" 4271 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n" 4272 "perf_event_paranoid setting is %d:\n" 4273 " -1: Allow use of (almost) all events by all users\n" 4274 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n" 4275 ">= 0: Disallow raw and ftrace function tracepoint access\n" 4276 ">= 1: Disallow CPU event access\n" 4277 ">= 2: Disallow kernel profiling\n" 4278 "To make the adjusted perf_event_paranoid setting permanent preserve it\n" 4279 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)", 4280 perf_event_paranoid()); 4281 case ENOENT: 4282 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel)); 4283 case EMFILE: 4284 return scnprintf(msg, size, "%s", 4285 "Too many events are opened.\n" 4286 "Probably the maximum number of open file descriptors has been reached.\n" 4287 "Hint: Try again after reducing the number of events.\n" 4288 "Hint: Try increasing the limit with 'ulimit -n <limit>'"); 4289 case ENOMEM: 4290 if (evsel__has_callchain(evsel) && 4291 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0) 4292 return scnprintf(msg, size, 4293 "Not enough memory to setup event with callchain.\n" 4294 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n" 4295 "Hint: Current value: %d", sysctl__max_stack()); 4296 break; 4297 case ENODEV: 4298 if (target->cpu_list) 4299 return scnprintf(msg, size, "%s", 4300 "No such device - did you specify an out-of-range profile CPU?"); 4301 break; 4302 case EOPNOTSUPP: 4303 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK) 4304 return scnprintf(msg, size, 4305 "%s: PMU Hardware or event type doesn't support branch stack sampling.", 4306 evsel__name(evsel)); 4307 if (evsel->core.attr.aux_output) 4308 return scnprintf(msg, size, 4309 "%s: PMU Hardware doesn't support 'aux_output' feature", 4310 evsel__name(evsel)); 4311 if (evsel->core.attr.aux_action) 4312 return scnprintf(msg, size, 4313 "%s: PMU Hardware doesn't support 'aux_action' feature", 4314 evsel__name(evsel)); 4315 if (evsel->core.attr.sample_period != 0) 4316 return scnprintf(msg, size, 4317 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'", 4318 evsel__name(evsel)); 4319 if (evsel->core.attr.precise_ip) 4320 return scnprintf(msg, size, "%s", 4321 "\'precise\' request may not be supported. Try removing 'p' modifier."); 4322 #if defined(__i386__) || defined(__x86_64__) 4323 if (evsel->core.attr.type == PERF_TYPE_HARDWARE) 4324 return scnprintf(msg, size, "%s", 4325 "No hardware sampling interrupt available.\n"); 4326 #endif 4327 if (!target__has_cpu(target)) 4328 return scnprintf(msg, size, 4329 "Unsupported event (%s) in per-thread mode, enable system wide with '-a'.", 4330 evsel__name(evsel)); 4331 break; 4332 case EBUSY: 4333 if (find_process("oprofiled")) 4334 return scnprintf(msg, size, 4335 "The PMU counters are busy/taken by another profiler.\n" 4336 "We found oprofile daemon running, please stop it and try again."); 4337 printed += scnprintf( 4338 msg, size, 4339 "The PMU %s counters are busy and in use by another process.\n", 4340 evsel->pmu ? evsel->pmu->name : ""); 4341 return printed + dump_perf_event_processes(msg + printed, size - printed); 4342 break; 4343 case EINVAL: 4344 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size) 4345 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel."); 4346 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size) 4347 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel."); 4348 if (evsel->core.attr.write_backward && perf_missing_features.write_backward) 4349 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel."); 4350 if (perf_missing_features.clockid) 4351 return scnprintf(msg, size, "clockid feature not supported."); 4352 if (perf_missing_features.clockid_wrong) 4353 return scnprintf(msg, size, "wrong clockid (%d).", clockid); 4354 if (perf_missing_features.aux_action) 4355 return scnprintf(msg, size, "The 'aux_action' feature is not supported, update the kernel."); 4356 if (perf_missing_features.aux_output) 4357 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel."); 4358 pmu = evsel__find_pmu(evsel); 4359 if (!pmu->is_core && !target__has_cpu(target)) 4360 return scnprintf(msg, size, 4361 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.", 4362 evsel__name(evsel)); 4363 4364 break; 4365 case ENODATA: 4366 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. " 4367 "Please add an auxiliary event in front of the load latency event."); 4368 default: 4369 break; 4370 } 4371 4372 ret = arch_evsel__open_strerror(evsel, err, msg, size); 4373 if (ret) 4374 return ret; 4375 4376 errno = err; 4377 return scnprintf(msg, size, 4378 "The sys_perf_event_open() syscall failed for event (%s): %m\n" 4379 "\"dmesg | grep -i perf\" may provide additional information.\n", 4380 evsel__name(evsel)); 4381 } 4382 4383 struct perf_session *evsel__session(struct evsel *evsel) 4384 { 4385 return evsel && evsel->evlist ? evsel->evlist->session : NULL; 4386 } 4387 4388 struct perf_env *evsel__env(struct evsel *evsel) 4389 { 4390 struct perf_session *session = evsel__session(evsel); 4391 4392 return session ? perf_session__env(session) : NULL; 4393 } 4394 4395 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist) 4396 { 4397 int cpu_map_idx, thread; 4398 4399 if (evsel__is_retire_lat(evsel)) 4400 return 0; 4401 4402 if (perf_pmu__kind(evsel->pmu) != PERF_PMU_KIND_PE) 4403 return 0; 4404 4405 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) { 4406 for (thread = 0; thread < xyarray__max_y(evsel->core.fd); 4407 thread++) { 4408 int fd = FD(evsel, cpu_map_idx, thread); 4409 4410 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, 4411 cpu_map_idx, thread, fd) < 0) 4412 return -1; 4413 } 4414 } 4415 4416 return 0; 4417 } 4418 4419 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist) 4420 { 4421 struct perf_cpu_map *cpus = evsel->core.cpus; 4422 struct perf_thread_map *threads = evsel->core.threads; 4423 4424 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr)) 4425 return -ENOMEM; 4426 4427 return store_evsel_ids(evsel, evlist); 4428 } 4429 4430 void evsel__zero_per_pkg(struct evsel *evsel) 4431 { 4432 struct hashmap_entry *cur; 4433 size_t bkt; 4434 4435 if (evsel->per_pkg_mask) { 4436 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt) 4437 zfree(&cur->pkey); 4438 4439 hashmap__clear(evsel->per_pkg_mask); 4440 } 4441 } 4442 4443 /** 4444 * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this 4445 * will be false on hybrid systems for hardware and legacy 4446 * cache events. 4447 */ 4448 bool evsel__is_hybrid(const struct evsel *evsel) 4449 { 4450 if (!evsel->core.is_pmu_core) 4451 return false; 4452 4453 return perf_pmus__num_core_pmus() > 1; 4454 } 4455 4456 struct evsel *evsel__leader(const struct evsel *evsel) 4457 { 4458 if (evsel->core.leader == NULL) 4459 return NULL; 4460 return container_of(evsel->core.leader, struct evsel, core); 4461 } 4462 4463 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader) 4464 { 4465 return evsel->core.leader == &leader->core; 4466 } 4467 4468 bool evsel__is_leader(struct evsel *evsel) 4469 { 4470 return evsel__has_leader(evsel, evsel); 4471 } 4472 4473 void evsel__set_leader(struct evsel *evsel, struct evsel *leader) 4474 { 4475 evsel->core.leader = &leader->core; 4476 } 4477 4478 bool evsel__is_aux_event(const struct evsel *evsel) 4479 { 4480 struct perf_pmu *pmu; 4481 4482 if (evsel->needs_auxtrace_mmap) 4483 return true; 4484 4485 pmu = evsel__find_pmu(evsel); 4486 return pmu && pmu->auxtrace; 4487 } 4488 4489 int evsel__source_count(const struct evsel *evsel) 4490 { 4491 struct evsel *pos; 4492 int count = 0; 4493 4494 evlist__for_each_entry(evsel->evlist, pos) { 4495 if (pos->metric_leader == evsel) 4496 count++; 4497 } 4498 return count; 4499 } 4500 4501 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused) 4502 { 4503 return false; 4504 } 4505 4506 /* 4507 * Remove an event from a given group (leader). 4508 * Some events, e.g., perf metrics Topdown events, 4509 * must always be grouped. Ignore the events. 4510 */ 4511 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader) 4512 { 4513 if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) { 4514 evsel__set_leader(evsel, evsel); 4515 evsel->core.nr_members = 0; 4516 leader->core.nr_members--; 4517 } 4518 } 4519 4520 bool evsel__set_needs_uniquify(struct evsel *counter, const struct perf_stat_config *config) 4521 { 4522 struct evsel *evsel; 4523 4524 if (counter->needs_uniquify) { 4525 /* Already set. */ 4526 return true; 4527 } 4528 4529 if (counter->use_config_name || counter->is_libpfm_event) { 4530 /* Original name will be used. */ 4531 return false; 4532 } 4533 4534 if (!config->hybrid_merge && evsel__is_hybrid(counter)) { 4535 /* Unique hybrid counters necessary. */ 4536 counter->needs_uniquify = true; 4537 return true; 4538 } 4539 4540 if (counter->core.attr.type < PERF_TYPE_MAX && counter->core.attr.type != PERF_TYPE_RAW) { 4541 /* Legacy event, don't uniquify. */ 4542 return false; 4543 } 4544 4545 if (counter->pmu && counter->pmu->is_core && 4546 counter->alternate_hw_config != PERF_COUNT_HW_MAX) { 4547 /* A sysfs or json event replacing a legacy event, don't uniquify. */ 4548 return false; 4549 } 4550 4551 if (config->aggr_mode == AGGR_NONE) { 4552 /* Always unique with no aggregation. */ 4553 counter->needs_uniquify = true; 4554 return true; 4555 } 4556 4557 if (counter->first_wildcard_match != NULL) { 4558 /* 4559 * If stats are merged then only the first_wildcard_match is 4560 * displayed, there is no need to uniquify this evsel as the 4561 * name won't be shown. 4562 */ 4563 return false; 4564 } 4565 4566 /* 4567 * Do other non-merged events in the evlist have the same name? If so 4568 * uniquify is necessary. 4569 */ 4570 evlist__for_each_entry(counter->evlist, evsel) { 4571 if (evsel == counter || evsel->first_wildcard_match || evsel->pmu == counter->pmu) 4572 continue; 4573 4574 if (evsel__name_is(counter, evsel__name(evsel))) { 4575 counter->needs_uniquify = true; 4576 return true; 4577 } 4578 } 4579 return false; 4580 } 4581 4582 void evsel__uniquify_counter(struct evsel *counter) 4583 { 4584 const char *name, *pmu_name, *config; 4585 char *new_name; 4586 int len, ret; 4587 4588 /* No uniquification necessary. */ 4589 if (!counter->needs_uniquify) 4590 return; 4591 4592 /* The evsel was already uniquified. */ 4593 if (counter->uniquified_name) 4594 return; 4595 4596 /* Avoid checking to uniquify twice. */ 4597 counter->uniquified_name = true; 4598 4599 name = evsel__name(counter); 4600 config = strchr(name, '/'); 4601 pmu_name = counter->pmu->name; 4602 4603 /* Already prefixed by the PMU name? */ 4604 len = pmu_name_len_no_suffix(pmu_name); 4605 4606 if (!strncmp(name, pmu_name, len)) { 4607 /* 4608 * If the PMU name is there, then there is no sense in not 4609 * having a slash. Do this for robustness. 4610 */ 4611 if (config == NULL) 4612 config = name - 1; 4613 4614 ret = asprintf(&new_name, "%s/%s", pmu_name, config + 1); 4615 } else if (config) { 4616 len = config - name; 4617 if (config[1] == '/') { 4618 /* case: event// */ 4619 ret = asprintf(&new_name, "%s/%.*s/%s", pmu_name, len, name, config + 2); 4620 } else { 4621 /* case: event/.../ */ 4622 ret = asprintf(&new_name, "%s/%.*s,%s", pmu_name, len, name, config + 1); 4623 } 4624 } else { 4625 config = strchr(name, ':'); 4626 if (config) { 4627 /* case: event:.. */ 4628 len = config - name; 4629 4630 ret = asprintf(&new_name, "%s/%.*s/%s", pmu_name, len, name, config + 1); 4631 } else { 4632 /* case: event */ 4633 ret = asprintf(&new_name, "%s/%s/", pmu_name, name); 4634 } 4635 } 4636 if (ret > 0) { 4637 free(counter->name); 4638 counter->name = new_name; 4639 } else { 4640 /* ENOMEM from asprintf. */ 4641 counter->uniquified_name = false; 4642 } 4643 } 4644 4645 void evsel__warn_user_requested_cpus(struct evsel *evsel, struct perf_cpu_map *user_requested_cpus) 4646 { 4647 struct perf_cpu_map *intersect, *online = NULL; 4648 const struct perf_pmu *pmu = evsel__find_pmu(evsel); 4649 4650 if (pmu && pmu->is_core) { 4651 intersect = perf_cpu_map__intersect(pmu->cpus, user_requested_cpus); 4652 } else { 4653 online = cpu_map__online(); 4654 intersect = perf_cpu_map__intersect(online, user_requested_cpus); 4655 } 4656 if (!perf_cpu_map__equal(intersect, user_requested_cpus)) { 4657 char buf1[128]; 4658 char buf2[128]; 4659 4660 cpu_map__snprint(user_requested_cpus, buf1, sizeof(buf1)); 4661 cpu_map__snprint(online ?: pmu->cpus, buf2, sizeof(buf2)); 4662 pr_warning("WARNING: A requested CPU in '%s' is not supported by PMU '%s' (CPUs %s) for event '%s'\n", 4663 buf1, pmu ? pmu->name : "cpu", buf2, evsel__name(evsel)); 4664 } 4665 perf_cpu_map__put(intersect); 4666 perf_cpu_map__put(online); 4667 } 4668