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 #include <byteswap.h> 10 #include <errno.h> 11 #include <inttypes.h> 12 #include <linux/bitops.h> 13 #include <api/io.h> 14 #include <api/fs/fs.h> 15 #include <api/fs/tracing_path.h> 16 #include <linux/hw_breakpoint.h> 17 #include <linux/perf_event.h> 18 #include <linux/compiler.h> 19 #include <linux/err.h> 20 #include <linux/zalloc.h> 21 #include <sys/ioctl.h> 22 #include <sys/resource.h> 23 #include <sys/types.h> 24 #include <dirent.h> 25 #include <stdlib.h> 26 #include <perf/evsel.h> 27 #include "asm/bug.h" 28 #include "bpf_counter.h" 29 #include "callchain.h" 30 #include "cgroup.h" 31 #include "counts.h" 32 #include "event.h" 33 #include "evsel.h" 34 #include "time-utils.h" 35 #include "util/env.h" 36 #include "util/evsel_config.h" 37 #include "util/evsel_fprintf.h" 38 #include "evlist.h" 39 #include <perf/cpumap.h> 40 #include "thread_map.h" 41 #include "target.h" 42 #include "perf_regs.h" 43 #include "record.h" 44 #include "debug.h" 45 #include "trace-event.h" 46 #include "stat.h" 47 #include "string2.h" 48 #include "memswap.h" 49 #include "util.h" 50 #include "util/hashmap.h" 51 #include "off_cpu.h" 52 #include "pmu.h" 53 #include "pmus.h" 54 #include "rlimit.h" 55 #include "../perf-sys.h" 56 #include "util/parse-branch-options.h" 57 #include "util/bpf-filter.h" 58 #include "util/hist.h" 59 #include <internal/xyarray.h> 60 #include <internal/lib.h> 61 #include <internal/threadmap.h> 62 #include "util/intel-tpebs.h" 63 64 #include <linux/ctype.h> 65 66 #ifdef HAVE_LIBTRACEEVENT 67 #include <traceevent/event-parse.h> 68 #endif 69 70 struct perf_missing_features perf_missing_features; 71 72 static clockid_t clockid; 73 74 static const char *const perf_tool_event__tool_names[PERF_TOOL_MAX] = { 75 NULL, 76 "duration_time", 77 "user_time", 78 "system_time", 79 }; 80 81 const char *perf_tool_event__to_str(enum perf_tool_event ev) 82 { 83 if (ev > PERF_TOOL_NONE && ev < PERF_TOOL_MAX) 84 return perf_tool_event__tool_names[ev]; 85 86 return NULL; 87 } 88 89 enum perf_tool_event perf_tool_event__from_str(const char *str) 90 { 91 int i; 92 93 perf_tool_event__for_each_event(i) { 94 if (!strcmp(str, perf_tool_event__tool_names[i])) 95 return i; 96 } 97 return PERF_TOOL_NONE; 98 } 99 100 101 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused) 102 { 103 return 0; 104 } 105 106 void __weak test_attr__ready(void) { } 107 108 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused) 109 { 110 } 111 112 static struct { 113 size_t size; 114 int (*init)(struct evsel *evsel); 115 void (*fini)(struct evsel *evsel); 116 } perf_evsel__object = { 117 .size = sizeof(struct evsel), 118 .init = evsel__no_extra_init, 119 .fini = evsel__no_extra_fini, 120 }; 121 122 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel), 123 void (*fini)(struct evsel *evsel)) 124 { 125 126 if (object_size == 0) 127 goto set_methods; 128 129 if (perf_evsel__object.size > object_size) 130 return -EINVAL; 131 132 perf_evsel__object.size = object_size; 133 134 set_methods: 135 if (init != NULL) 136 perf_evsel__object.init = init; 137 138 if (fini != NULL) 139 perf_evsel__object.fini = fini; 140 141 return 0; 142 } 143 144 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 145 146 int __evsel__sample_size(u64 sample_type) 147 { 148 u64 mask = sample_type & PERF_SAMPLE_MASK; 149 int size = 0; 150 int i; 151 152 for (i = 0; i < 64; i++) { 153 if (mask & (1ULL << i)) 154 size++; 155 } 156 157 size *= sizeof(u64); 158 159 return size; 160 } 161 162 /** 163 * __perf_evsel__calc_id_pos - calculate id_pos. 164 * @sample_type: sample type 165 * 166 * This function returns the position of the event id (PERF_SAMPLE_ID or 167 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct 168 * perf_record_sample. 169 */ 170 static int __perf_evsel__calc_id_pos(u64 sample_type) 171 { 172 int idx = 0; 173 174 if (sample_type & PERF_SAMPLE_IDENTIFIER) 175 return 0; 176 177 if (!(sample_type & PERF_SAMPLE_ID)) 178 return -1; 179 180 if (sample_type & PERF_SAMPLE_IP) 181 idx += 1; 182 183 if (sample_type & PERF_SAMPLE_TID) 184 idx += 1; 185 186 if (sample_type & PERF_SAMPLE_TIME) 187 idx += 1; 188 189 if (sample_type & PERF_SAMPLE_ADDR) 190 idx += 1; 191 192 return idx; 193 } 194 195 /** 196 * __perf_evsel__calc_is_pos - calculate is_pos. 197 * @sample_type: sample type 198 * 199 * This function returns the position (counting backwards) of the event id 200 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if 201 * sample_id_all is used there is an id sample appended to non-sample events. 202 */ 203 static int __perf_evsel__calc_is_pos(u64 sample_type) 204 { 205 int idx = 1; 206 207 if (sample_type & PERF_SAMPLE_IDENTIFIER) 208 return 1; 209 210 if (!(sample_type & PERF_SAMPLE_ID)) 211 return -1; 212 213 if (sample_type & PERF_SAMPLE_CPU) 214 idx += 1; 215 216 if (sample_type & PERF_SAMPLE_STREAM_ID) 217 idx += 1; 218 219 return idx; 220 } 221 222 void evsel__calc_id_pos(struct evsel *evsel) 223 { 224 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type); 225 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type); 226 } 227 228 void __evsel__set_sample_bit(struct evsel *evsel, 229 enum perf_event_sample_format bit) 230 { 231 if (!(evsel->core.attr.sample_type & bit)) { 232 evsel->core.attr.sample_type |= bit; 233 evsel->sample_size += sizeof(u64); 234 evsel__calc_id_pos(evsel); 235 } 236 } 237 238 void __evsel__reset_sample_bit(struct evsel *evsel, 239 enum perf_event_sample_format bit) 240 { 241 if (evsel->core.attr.sample_type & bit) { 242 evsel->core.attr.sample_type &= ~bit; 243 evsel->sample_size -= sizeof(u64); 244 evsel__calc_id_pos(evsel); 245 } 246 } 247 248 void evsel__set_sample_id(struct evsel *evsel, 249 bool can_sample_identifier) 250 { 251 if (can_sample_identifier) { 252 evsel__reset_sample_bit(evsel, ID); 253 evsel__set_sample_bit(evsel, IDENTIFIER); 254 } else { 255 evsel__set_sample_bit(evsel, ID); 256 } 257 evsel->core.attr.read_format |= PERF_FORMAT_ID; 258 } 259 260 /** 261 * evsel__is_function_event - Return whether given evsel is a function 262 * trace event 263 * 264 * @evsel - evsel selector to be tested 265 * 266 * Return %true if event is function trace event 267 */ 268 bool evsel__is_function_event(struct evsel *evsel) 269 { 270 #define FUNCTION_EVENT "ftrace:function" 271 272 return evsel->name && 273 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT)); 274 275 #undef FUNCTION_EVENT 276 } 277 278 void evsel__init(struct evsel *evsel, 279 struct perf_event_attr *attr, int idx) 280 { 281 perf_evsel__init(&evsel->core, attr, idx); 282 evsel->tracking = !idx; 283 evsel->unit = strdup(""); 284 evsel->scale = 1.0; 285 evsel->max_events = ULONG_MAX; 286 evsel->evlist = NULL; 287 evsel->bpf_obj = NULL; 288 evsel->bpf_fd = -1; 289 INIT_LIST_HEAD(&evsel->config_terms); 290 INIT_LIST_HEAD(&evsel->bpf_counter_list); 291 INIT_LIST_HEAD(&evsel->bpf_filters); 292 perf_evsel__object.init(evsel); 293 evsel->sample_size = __evsel__sample_size(attr->sample_type); 294 evsel__calc_id_pos(evsel); 295 evsel->cmdline_group_boundary = false; 296 evsel->metric_events = NULL; 297 evsel->per_pkg_mask = NULL; 298 evsel->collect_stat = false; 299 evsel->group_pmu_name = NULL; 300 evsel->skippable = false; 301 evsel->alternate_hw_config = PERF_COUNT_HW_MAX; 302 } 303 304 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx) 305 { 306 struct evsel *evsel = zalloc(perf_evsel__object.size); 307 308 if (!evsel) 309 return NULL; 310 evsel__init(evsel, attr, idx); 311 312 if (evsel__is_bpf_output(evsel) && !attr->sample_type) { 313 evsel->core.attr.sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 314 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 315 evsel->core.attr.sample_period = 1; 316 } 317 318 if (evsel__is_clock(evsel)) { 319 free((char *)evsel->unit); 320 evsel->unit = strdup("msec"); 321 evsel->scale = 1e-6; 322 } 323 324 return evsel; 325 } 326 327 int copy_config_terms(struct list_head *dst, struct list_head *src) 328 { 329 struct evsel_config_term *pos, *tmp; 330 331 list_for_each_entry(pos, src, list) { 332 tmp = malloc(sizeof(*tmp)); 333 if (tmp == NULL) 334 return -ENOMEM; 335 336 *tmp = *pos; 337 if (tmp->free_str) { 338 tmp->val.str = strdup(pos->val.str); 339 if (tmp->val.str == NULL) { 340 free(tmp); 341 return -ENOMEM; 342 } 343 } 344 list_add_tail(&tmp->list, dst); 345 } 346 return 0; 347 } 348 349 static int evsel__copy_config_terms(struct evsel *dst, struct evsel *src) 350 { 351 return copy_config_terms(&dst->config_terms, &src->config_terms); 352 } 353 354 /** 355 * evsel__clone - create a new evsel copied from @orig 356 * @orig: original evsel 357 * 358 * The assumption is that @orig is not configured nor opened yet. 359 * So we only care about the attributes that can be set while it's parsed. 360 */ 361 struct evsel *evsel__clone(struct evsel *orig) 362 { 363 struct evsel *evsel; 364 365 BUG_ON(orig->core.fd); 366 BUG_ON(orig->counts); 367 BUG_ON(orig->priv); 368 BUG_ON(orig->per_pkg_mask); 369 370 /* cannot handle BPF objects for now */ 371 if (orig->bpf_obj) 372 return NULL; 373 374 evsel = evsel__new(&orig->core.attr); 375 if (evsel == NULL) 376 return NULL; 377 378 evsel->core.cpus = perf_cpu_map__get(orig->core.cpus); 379 evsel->core.own_cpus = perf_cpu_map__get(orig->core.own_cpus); 380 evsel->core.threads = perf_thread_map__get(orig->core.threads); 381 evsel->core.nr_members = orig->core.nr_members; 382 evsel->core.system_wide = orig->core.system_wide; 383 evsel->core.requires_cpu = orig->core.requires_cpu; 384 evsel->core.is_pmu_core = orig->core.is_pmu_core; 385 386 if (orig->name) { 387 evsel->name = strdup(orig->name); 388 if (evsel->name == NULL) 389 goto out_err; 390 } 391 if (orig->group_name) { 392 evsel->group_name = strdup(orig->group_name); 393 if (evsel->group_name == NULL) 394 goto out_err; 395 } 396 if (orig->group_pmu_name) { 397 evsel->group_pmu_name = strdup(orig->group_pmu_name); 398 if (evsel->group_pmu_name == NULL) 399 goto out_err; 400 } 401 if (orig->filter) { 402 evsel->filter = strdup(orig->filter); 403 if (evsel->filter == NULL) 404 goto out_err; 405 } 406 if (orig->metric_id) { 407 evsel->metric_id = strdup(orig->metric_id); 408 if (evsel->metric_id == NULL) 409 goto out_err; 410 } 411 evsel->cgrp = cgroup__get(orig->cgrp); 412 #ifdef HAVE_LIBTRACEEVENT 413 evsel->tp_format = orig->tp_format; 414 #endif 415 evsel->handler = orig->handler; 416 evsel->core.leader = orig->core.leader; 417 418 evsel->max_events = orig->max_events; 419 evsel->tool_event = orig->tool_event; 420 free((char *)evsel->unit); 421 evsel->unit = strdup(orig->unit); 422 if (evsel->unit == NULL) 423 goto out_err; 424 425 evsel->scale = orig->scale; 426 evsel->snapshot = orig->snapshot; 427 evsel->per_pkg = orig->per_pkg; 428 evsel->percore = orig->percore; 429 evsel->precise_max = orig->precise_max; 430 evsel->is_libpfm_event = orig->is_libpfm_event; 431 432 evsel->exclude_GH = orig->exclude_GH; 433 evsel->sample_read = orig->sample_read; 434 evsel->auto_merge_stats = orig->auto_merge_stats; 435 evsel->collect_stat = orig->collect_stat; 436 evsel->weak_group = orig->weak_group; 437 evsel->use_config_name = orig->use_config_name; 438 evsel->pmu = orig->pmu; 439 440 if (evsel__copy_config_terms(evsel, orig) < 0) 441 goto out_err; 442 443 evsel->alternate_hw_config = orig->alternate_hw_config; 444 445 return evsel; 446 447 out_err: 448 evsel__delete(evsel); 449 return NULL; 450 } 451 452 /* 453 * Returns pointer with encoded error via <linux/err.h> interface. 454 */ 455 #ifdef HAVE_LIBTRACEEVENT 456 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx, bool format) 457 { 458 struct evsel *evsel = zalloc(perf_evsel__object.size); 459 int err = -ENOMEM; 460 461 if (evsel == NULL) { 462 goto out_err; 463 } else { 464 struct perf_event_attr attr = { 465 .type = PERF_TYPE_TRACEPOINT, 466 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 467 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 468 }; 469 470 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0) 471 goto out_free; 472 473 event_attr_init(&attr); 474 475 if (format) { 476 evsel->tp_format = trace_event__tp_format(sys, name); 477 if (IS_ERR(evsel->tp_format)) { 478 err = PTR_ERR(evsel->tp_format); 479 goto out_free; 480 } 481 attr.config = evsel->tp_format->id; 482 } else { 483 attr.config = (__u64) -1; 484 } 485 486 487 attr.sample_period = 1; 488 evsel__init(evsel, &attr, idx); 489 } 490 491 return evsel; 492 493 out_free: 494 zfree(&evsel->name); 495 free(evsel); 496 out_err: 497 return ERR_PTR(err); 498 } 499 #endif 500 501 const char *const evsel__hw_names[PERF_COUNT_HW_MAX] = { 502 "cycles", 503 "instructions", 504 "cache-references", 505 "cache-misses", 506 "branches", 507 "branch-misses", 508 "bus-cycles", 509 "stalled-cycles-frontend", 510 "stalled-cycles-backend", 511 "ref-cycles", 512 }; 513 514 char *evsel__bpf_counter_events; 515 516 bool evsel__match_bpf_counter_events(const char *name) 517 { 518 int name_len; 519 bool match; 520 char *ptr; 521 522 if (!evsel__bpf_counter_events) 523 return false; 524 525 ptr = strstr(evsel__bpf_counter_events, name); 526 name_len = strlen(name); 527 528 /* check name matches a full token in evsel__bpf_counter_events */ 529 match = (ptr != NULL) && 530 ((ptr == evsel__bpf_counter_events) || (*(ptr - 1) == ',')) && 531 ((*(ptr + name_len) == ',') || (*(ptr + name_len) == '\0')); 532 533 return match; 534 } 535 536 static const char *__evsel__hw_name(u64 config) 537 { 538 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config]) 539 return evsel__hw_names[config]; 540 541 return "unknown-hardware"; 542 } 543 544 static int evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size) 545 { 546 int colon = 0, r = 0; 547 struct perf_event_attr *attr = &evsel->core.attr; 548 bool exclude_guest_default = false; 549 550 #define MOD_PRINT(context, mod) do { \ 551 if (!attr->exclude_##context) { \ 552 if (!colon) colon = ++r; \ 553 r += scnprintf(bf + r, size - r, "%c", mod); \ 554 } } while(0) 555 556 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) { 557 MOD_PRINT(kernel, 'k'); 558 MOD_PRINT(user, 'u'); 559 MOD_PRINT(hv, 'h'); 560 exclude_guest_default = true; 561 } 562 563 if (attr->precise_ip) { 564 if (!colon) 565 colon = ++r; 566 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp"); 567 exclude_guest_default = true; 568 } 569 570 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) { 571 MOD_PRINT(host, 'H'); 572 MOD_PRINT(guest, 'G'); 573 } 574 #undef MOD_PRINT 575 if (colon) 576 bf[colon - 1] = ':'; 577 return r; 578 } 579 580 int __weak arch_evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 581 { 582 return scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config)); 583 } 584 585 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 586 { 587 int r = arch_evsel__hw_name(evsel, bf, size); 588 return r + evsel__add_modifiers(evsel, bf + r, size - r); 589 } 590 591 const char *const evsel__sw_names[PERF_COUNT_SW_MAX] = { 592 "cpu-clock", 593 "task-clock", 594 "page-faults", 595 "context-switches", 596 "cpu-migrations", 597 "minor-faults", 598 "major-faults", 599 "alignment-faults", 600 "emulation-faults", 601 "dummy", 602 }; 603 604 static const char *__evsel__sw_name(u64 config) 605 { 606 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config]) 607 return evsel__sw_names[config]; 608 return "unknown-software"; 609 } 610 611 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size) 612 { 613 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config)); 614 return r + evsel__add_modifiers(evsel, bf + r, size - r); 615 } 616 617 static int evsel__tool_name(enum perf_tool_event ev, char *bf, size_t size) 618 { 619 return scnprintf(bf, size, "%s", perf_tool_event__to_str(ev)); 620 } 621 622 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type) 623 { 624 int r; 625 626 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr); 627 628 if (type & HW_BREAKPOINT_R) 629 r += scnprintf(bf + r, size - r, "r"); 630 631 if (type & HW_BREAKPOINT_W) 632 r += scnprintf(bf + r, size - r, "w"); 633 634 if (type & HW_BREAKPOINT_X) 635 r += scnprintf(bf + r, size - r, "x"); 636 637 return r; 638 } 639 640 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size) 641 { 642 struct perf_event_attr *attr = &evsel->core.attr; 643 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type); 644 return r + evsel__add_modifiers(evsel, bf + r, size - r); 645 } 646 647 const char *const evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = { 648 { "L1-dcache", "l1-d", "l1d", "L1-data", }, 649 { "L1-icache", "l1-i", "l1i", "L1-instruction", }, 650 { "LLC", "L2", }, 651 { "dTLB", "d-tlb", "Data-TLB", }, 652 { "iTLB", "i-tlb", "Instruction-TLB", }, 653 { "branch", "branches", "bpu", "btb", "bpc", }, 654 { "node", }, 655 }; 656 657 const char *const evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = { 658 { "load", "loads", "read", }, 659 { "store", "stores", "write", }, 660 { "prefetch", "prefetches", "speculative-read", "speculative-load", }, 661 }; 662 663 const char *const evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = { 664 { "refs", "Reference", "ops", "access", }, 665 { "misses", "miss", }, 666 }; 667 668 #define C(x) PERF_COUNT_HW_CACHE_##x 669 #define CACHE_READ (1 << C(OP_READ)) 670 #define CACHE_WRITE (1 << C(OP_WRITE)) 671 #define CACHE_PREFETCH (1 << C(OP_PREFETCH)) 672 #define COP(x) (1 << x) 673 674 /* 675 * cache operation stat 676 * L1I : Read and prefetch only 677 * ITLB and BPU : Read-only 678 */ 679 static const unsigned long evsel__hw_cache_stat[C(MAX)] = { 680 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 681 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH), 682 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 683 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 684 [C(ITLB)] = (CACHE_READ), 685 [C(BPU)] = (CACHE_READ), 686 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 687 }; 688 689 bool evsel__is_cache_op_valid(u8 type, u8 op) 690 { 691 if (evsel__hw_cache_stat[type] & COP(op)) 692 return true; /* valid */ 693 else 694 return false; /* invalid */ 695 } 696 697 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size) 698 { 699 if (result) { 700 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0], 701 evsel__hw_cache_op[op][0], 702 evsel__hw_cache_result[result][0]); 703 } 704 705 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0], 706 evsel__hw_cache_op[op][1]); 707 } 708 709 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size) 710 { 711 u8 op, result, type = (config >> 0) & 0xff; 712 const char *err = "unknown-ext-hardware-cache-type"; 713 714 if (type >= PERF_COUNT_HW_CACHE_MAX) 715 goto out_err; 716 717 op = (config >> 8) & 0xff; 718 err = "unknown-ext-hardware-cache-op"; 719 if (op >= PERF_COUNT_HW_CACHE_OP_MAX) 720 goto out_err; 721 722 result = (config >> 16) & 0xff; 723 err = "unknown-ext-hardware-cache-result"; 724 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 725 goto out_err; 726 727 err = "invalid-cache"; 728 if (!evsel__is_cache_op_valid(type, op)) 729 goto out_err; 730 731 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size); 732 out_err: 733 return scnprintf(bf, size, "%s", err); 734 } 735 736 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size) 737 { 738 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size); 739 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret); 740 } 741 742 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size) 743 { 744 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config); 745 return ret + evsel__add_modifiers(evsel, bf + ret, size - ret); 746 } 747 748 const char *evsel__name(struct evsel *evsel) 749 { 750 char bf[128]; 751 752 if (!evsel) 753 goto out_unknown; 754 755 if (evsel->name) 756 return evsel->name; 757 758 switch (evsel->core.attr.type) { 759 case PERF_TYPE_RAW: 760 evsel__raw_name(evsel, bf, sizeof(bf)); 761 break; 762 763 case PERF_TYPE_HARDWARE: 764 evsel__hw_name(evsel, bf, sizeof(bf)); 765 break; 766 767 case PERF_TYPE_HW_CACHE: 768 evsel__hw_cache_name(evsel, bf, sizeof(bf)); 769 break; 770 771 case PERF_TYPE_SOFTWARE: 772 if (evsel__is_tool(evsel)) 773 evsel__tool_name(evsel__tool_event(evsel), bf, sizeof(bf)); 774 else 775 evsel__sw_name(evsel, bf, sizeof(bf)); 776 break; 777 778 case PERF_TYPE_TRACEPOINT: 779 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint"); 780 break; 781 782 case PERF_TYPE_BREAKPOINT: 783 evsel__bp_name(evsel, bf, sizeof(bf)); 784 break; 785 786 default: 787 scnprintf(bf, sizeof(bf), "unknown attr type: %d", 788 evsel->core.attr.type); 789 break; 790 } 791 792 evsel->name = strdup(bf); 793 794 if (evsel->name) 795 return evsel->name; 796 out_unknown: 797 return "unknown"; 798 } 799 800 bool evsel__name_is(struct evsel *evsel, const char *name) 801 { 802 return !strcmp(evsel__name(evsel), name); 803 } 804 805 const char *evsel__metric_id(const struct evsel *evsel) 806 { 807 if (evsel->metric_id) 808 return evsel->metric_id; 809 810 if (evsel__is_tool(evsel)) 811 return perf_tool_event__to_str(evsel__tool_event(evsel)); 812 813 return "unknown"; 814 } 815 816 const char *evsel__group_name(struct evsel *evsel) 817 { 818 return evsel->group_name ?: "anon group"; 819 } 820 821 /* 822 * Returns the group details for the specified leader, 823 * with following rules. 824 * 825 * For record -e '{cycles,instructions}' 826 * 'anon group { cycles:u, instructions:u }' 827 * 828 * For record -e 'cycles,instructions' and report --group 829 * 'cycles:u, instructions:u' 830 */ 831 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size) 832 { 833 int ret = 0; 834 bool first = true; 835 struct evsel *pos; 836 const char *group_name = evsel__group_name(evsel); 837 838 if (!evsel->forced_leader) 839 ret = scnprintf(buf, size, "%s { ", group_name); 840 841 for_each_group_evsel(pos, evsel) { 842 if (symbol_conf.skip_empty && 843 evsel__hists(pos)->stats.nr_samples == 0) 844 continue; 845 846 ret += scnprintf(buf + ret, size - ret, "%s%s", 847 first ? "" : ", ", evsel__name(pos)); 848 first = false; 849 } 850 851 if (!evsel->forced_leader) 852 ret += scnprintf(buf + ret, size - ret, " }"); 853 854 return ret; 855 } 856 857 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts, 858 struct callchain_param *param) 859 { 860 bool function = evsel__is_function_event(evsel); 861 struct perf_event_attr *attr = &evsel->core.attr; 862 863 evsel__set_sample_bit(evsel, CALLCHAIN); 864 865 attr->sample_max_stack = param->max_stack; 866 867 if (opts->kernel_callchains) 868 attr->exclude_callchain_user = 1; 869 if (opts->user_callchains) 870 attr->exclude_callchain_kernel = 1; 871 if (param->record_mode == CALLCHAIN_LBR) { 872 if (!opts->branch_stack) { 873 if (attr->exclude_user) { 874 pr_warning("LBR callstack option is only available " 875 "to get user callchain information. " 876 "Falling back to framepointers.\n"); 877 } else { 878 evsel__set_sample_bit(evsel, BRANCH_STACK); 879 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER | 880 PERF_SAMPLE_BRANCH_CALL_STACK | 881 PERF_SAMPLE_BRANCH_NO_CYCLES | 882 PERF_SAMPLE_BRANCH_NO_FLAGS | 883 PERF_SAMPLE_BRANCH_HW_INDEX; 884 } 885 } else 886 pr_warning("Cannot use LBR callstack with branch stack. " 887 "Falling back to framepointers.\n"); 888 } 889 890 if (param->record_mode == CALLCHAIN_DWARF) { 891 if (!function) { 892 const char *arch = perf_env__arch(evsel__env(evsel)); 893 894 evsel__set_sample_bit(evsel, REGS_USER); 895 evsel__set_sample_bit(evsel, STACK_USER); 896 if (opts->sample_user_regs && 897 DWARF_MINIMAL_REGS(arch) != arch__user_reg_mask()) { 898 attr->sample_regs_user |= DWARF_MINIMAL_REGS(arch); 899 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, " 900 "specifying a subset with --user-regs may render DWARF unwinding unreliable, " 901 "so the minimal registers set (IP, SP) is explicitly forced.\n"); 902 } else { 903 attr->sample_regs_user |= arch__user_reg_mask(); 904 } 905 attr->sample_stack_user = param->dump_size; 906 attr->exclude_callchain_user = 1; 907 } else { 908 pr_info("Cannot use DWARF unwind for function trace event," 909 " falling back to framepointers.\n"); 910 } 911 } 912 913 if (function) { 914 pr_info("Disabling user space callchains for function trace event.\n"); 915 attr->exclude_callchain_user = 1; 916 } 917 } 918 919 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts, 920 struct callchain_param *param) 921 { 922 if (param->enabled) 923 return __evsel__config_callchain(evsel, opts, param); 924 } 925 926 static void evsel__reset_callgraph(struct evsel *evsel, struct callchain_param *param) 927 { 928 struct perf_event_attr *attr = &evsel->core.attr; 929 930 evsel__reset_sample_bit(evsel, CALLCHAIN); 931 if (param->record_mode == CALLCHAIN_LBR) { 932 evsel__reset_sample_bit(evsel, BRANCH_STACK); 933 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER | 934 PERF_SAMPLE_BRANCH_CALL_STACK | 935 PERF_SAMPLE_BRANCH_HW_INDEX); 936 } 937 if (param->record_mode == CALLCHAIN_DWARF) { 938 evsel__reset_sample_bit(evsel, REGS_USER); 939 evsel__reset_sample_bit(evsel, STACK_USER); 940 } 941 } 942 943 static void evsel__apply_config_terms(struct evsel *evsel, 944 struct record_opts *opts, bool track) 945 { 946 struct evsel_config_term *term; 947 struct list_head *config_terms = &evsel->config_terms; 948 struct perf_event_attr *attr = &evsel->core.attr; 949 /* callgraph default */ 950 struct callchain_param param = { 951 .record_mode = callchain_param.record_mode, 952 }; 953 u32 dump_size = 0; 954 int max_stack = 0; 955 const char *callgraph_buf = NULL; 956 957 list_for_each_entry(term, config_terms, list) { 958 switch (term->type) { 959 case EVSEL__CONFIG_TERM_PERIOD: 960 if (!(term->weak && opts->user_interval != ULLONG_MAX)) { 961 attr->sample_period = term->val.period; 962 attr->freq = 0; 963 evsel__reset_sample_bit(evsel, PERIOD); 964 } 965 break; 966 case EVSEL__CONFIG_TERM_FREQ: 967 if (!(term->weak && opts->user_freq != UINT_MAX)) { 968 attr->sample_freq = term->val.freq; 969 attr->freq = 1; 970 evsel__set_sample_bit(evsel, PERIOD); 971 } 972 break; 973 case EVSEL__CONFIG_TERM_TIME: 974 if (term->val.time) 975 evsel__set_sample_bit(evsel, TIME); 976 else 977 evsel__reset_sample_bit(evsel, TIME); 978 break; 979 case EVSEL__CONFIG_TERM_CALLGRAPH: 980 callgraph_buf = term->val.str; 981 break; 982 case EVSEL__CONFIG_TERM_BRANCH: 983 if (term->val.str && strcmp(term->val.str, "no")) { 984 evsel__set_sample_bit(evsel, BRANCH_STACK); 985 parse_branch_str(term->val.str, 986 &attr->branch_sample_type); 987 } else 988 evsel__reset_sample_bit(evsel, BRANCH_STACK); 989 break; 990 case EVSEL__CONFIG_TERM_STACK_USER: 991 dump_size = term->val.stack_user; 992 break; 993 case EVSEL__CONFIG_TERM_MAX_STACK: 994 max_stack = term->val.max_stack; 995 break; 996 case EVSEL__CONFIG_TERM_MAX_EVENTS: 997 evsel->max_events = term->val.max_events; 998 break; 999 case EVSEL__CONFIG_TERM_INHERIT: 1000 /* 1001 * attr->inherit should has already been set by 1002 * evsel__config. If user explicitly set 1003 * inherit using config terms, override global 1004 * opt->no_inherit setting. 1005 */ 1006 attr->inherit = term->val.inherit ? 1 : 0; 1007 break; 1008 case EVSEL__CONFIG_TERM_OVERWRITE: 1009 attr->write_backward = term->val.overwrite ? 1 : 0; 1010 break; 1011 case EVSEL__CONFIG_TERM_DRV_CFG: 1012 break; 1013 case EVSEL__CONFIG_TERM_PERCORE: 1014 break; 1015 case EVSEL__CONFIG_TERM_AUX_OUTPUT: 1016 attr->aux_output = term->val.aux_output ? 1 : 0; 1017 break; 1018 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE: 1019 /* Already applied by auxtrace */ 1020 break; 1021 case EVSEL__CONFIG_TERM_CFG_CHG: 1022 break; 1023 default: 1024 break; 1025 } 1026 } 1027 1028 /* User explicitly set per-event callgraph, clear the old setting and reset. */ 1029 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) { 1030 bool sample_address = false; 1031 1032 if (max_stack) { 1033 param.max_stack = max_stack; 1034 if (callgraph_buf == NULL) 1035 callgraph_buf = "fp"; 1036 } 1037 1038 /* parse callgraph parameters */ 1039 if (callgraph_buf != NULL) { 1040 if (!strcmp(callgraph_buf, "no")) { 1041 param.enabled = false; 1042 param.record_mode = CALLCHAIN_NONE; 1043 } else { 1044 param.enabled = true; 1045 if (parse_callchain_record(callgraph_buf, ¶m)) { 1046 pr_err("per-event callgraph setting for %s failed. " 1047 "Apply callgraph global setting for it\n", 1048 evsel->name); 1049 return; 1050 } 1051 if (param.record_mode == CALLCHAIN_DWARF) 1052 sample_address = true; 1053 } 1054 } 1055 if (dump_size > 0) { 1056 dump_size = round_up(dump_size, sizeof(u64)); 1057 param.dump_size = dump_size; 1058 } 1059 1060 /* If global callgraph set, clear it */ 1061 if (callchain_param.enabled) 1062 evsel__reset_callgraph(evsel, &callchain_param); 1063 1064 /* set perf-event callgraph */ 1065 if (param.enabled) { 1066 if (sample_address) { 1067 evsel__set_sample_bit(evsel, ADDR); 1068 evsel__set_sample_bit(evsel, DATA_SRC); 1069 evsel->core.attr.mmap_data = track; 1070 } 1071 evsel__config_callchain(evsel, opts, ¶m); 1072 } 1073 } 1074 } 1075 1076 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type) 1077 { 1078 struct evsel_config_term *term, *found_term = NULL; 1079 1080 list_for_each_entry(term, &evsel->config_terms, list) { 1081 if (term->type == type) 1082 found_term = term; 1083 } 1084 1085 return found_term; 1086 } 1087 1088 void __weak arch_evsel__set_sample_weight(struct evsel *evsel) 1089 { 1090 evsel__set_sample_bit(evsel, WEIGHT); 1091 } 1092 1093 void __weak arch__post_evsel_config(struct evsel *evsel __maybe_unused, 1094 struct perf_event_attr *attr __maybe_unused) 1095 { 1096 } 1097 1098 static void evsel__set_default_freq_period(struct record_opts *opts, 1099 struct perf_event_attr *attr) 1100 { 1101 if (opts->freq) { 1102 attr->freq = 1; 1103 attr->sample_freq = opts->freq; 1104 } else { 1105 attr->sample_period = opts->default_interval; 1106 } 1107 } 1108 1109 static bool evsel__is_offcpu_event(struct evsel *evsel) 1110 { 1111 return evsel__is_bpf_output(evsel) && evsel__name_is(evsel, OFFCPU_EVENT); 1112 } 1113 1114 /* 1115 * The enable_on_exec/disabled value strategy: 1116 * 1117 * 1) For any type of traced program: 1118 * - all independent events and group leaders are disabled 1119 * - all group members are enabled 1120 * 1121 * Group members are ruled by group leaders. They need to 1122 * be enabled, because the group scheduling relies on that. 1123 * 1124 * 2) For traced programs executed by perf: 1125 * - all independent events and group leaders have 1126 * enable_on_exec set 1127 * - we don't specifically enable or disable any event during 1128 * the record command 1129 * 1130 * Independent events and group leaders are initially disabled 1131 * and get enabled by exec. Group members are ruled by group 1132 * leaders as stated in 1). 1133 * 1134 * 3) For traced programs attached by perf (pid/tid): 1135 * - we specifically enable or disable all events during 1136 * the record command 1137 * 1138 * When attaching events to already running traced we 1139 * enable/disable events specifically, as there's no 1140 * initial traced exec call. 1141 */ 1142 void evsel__config(struct evsel *evsel, struct record_opts *opts, 1143 struct callchain_param *callchain) 1144 { 1145 struct evsel *leader = evsel__leader(evsel); 1146 struct perf_event_attr *attr = &evsel->core.attr; 1147 int track = evsel->tracking; 1148 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread; 1149 1150 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1; 1151 attr->inherit = !opts->no_inherit; 1152 attr->write_backward = opts->overwrite ? 1 : 0; 1153 attr->read_format = PERF_FORMAT_LOST; 1154 1155 evsel__set_sample_bit(evsel, IP); 1156 evsel__set_sample_bit(evsel, TID); 1157 1158 if (evsel->sample_read) { 1159 evsel__set_sample_bit(evsel, READ); 1160 1161 /* 1162 * We need ID even in case of single event, because 1163 * PERF_SAMPLE_READ process ID specific data. 1164 */ 1165 evsel__set_sample_id(evsel, false); 1166 1167 /* 1168 * Apply group format only if we belong to group 1169 * with more than one members. 1170 */ 1171 if (leader->core.nr_members > 1) { 1172 attr->read_format |= PERF_FORMAT_GROUP; 1173 attr->inherit = 0; 1174 } 1175 } 1176 1177 /* 1178 * We default some events to have a default interval. But keep 1179 * it a weak assumption overridable by the user. 1180 */ 1181 if ((evsel->is_libpfm_event && !attr->sample_period) || 1182 (!evsel->is_libpfm_event && (!attr->sample_period || 1183 opts->user_freq != UINT_MAX || 1184 opts->user_interval != ULLONG_MAX))) 1185 evsel__set_default_freq_period(opts, attr); 1186 1187 /* 1188 * If attr->freq was set (here or earlier), ask for period 1189 * to be sampled. 1190 */ 1191 if (attr->freq) 1192 evsel__set_sample_bit(evsel, PERIOD); 1193 1194 if (opts->no_samples) 1195 attr->sample_freq = 0; 1196 1197 if (opts->inherit_stat) { 1198 evsel->core.attr.read_format |= 1199 PERF_FORMAT_TOTAL_TIME_ENABLED | 1200 PERF_FORMAT_TOTAL_TIME_RUNNING | 1201 PERF_FORMAT_ID; 1202 attr->inherit_stat = 1; 1203 } 1204 1205 if (opts->sample_address) { 1206 evsel__set_sample_bit(evsel, ADDR); 1207 attr->mmap_data = track; 1208 } 1209 1210 /* 1211 * We don't allow user space callchains for function trace 1212 * event, due to issues with page faults while tracing page 1213 * fault handler and its overall trickiness nature. 1214 */ 1215 if (evsel__is_function_event(evsel)) 1216 evsel->core.attr.exclude_callchain_user = 1; 1217 1218 if (callchain && callchain->enabled && !evsel->no_aux_samples) 1219 evsel__config_callchain(evsel, opts, callchain); 1220 1221 if (opts->sample_intr_regs && !evsel->no_aux_samples && 1222 !evsel__is_dummy_event(evsel)) { 1223 attr->sample_regs_intr = opts->sample_intr_regs; 1224 evsel__set_sample_bit(evsel, REGS_INTR); 1225 } 1226 1227 if (opts->sample_user_regs && !evsel->no_aux_samples && 1228 !evsel__is_dummy_event(evsel)) { 1229 attr->sample_regs_user |= opts->sample_user_regs; 1230 evsel__set_sample_bit(evsel, REGS_USER); 1231 } 1232 1233 if (target__has_cpu(&opts->target) || opts->sample_cpu) 1234 evsel__set_sample_bit(evsel, CPU); 1235 1236 /* 1237 * When the user explicitly disabled time don't force it here. 1238 */ 1239 if (opts->sample_time && 1240 (!perf_missing_features.sample_id_all && 1241 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu || 1242 opts->sample_time_set))) 1243 evsel__set_sample_bit(evsel, TIME); 1244 1245 if (opts->raw_samples && !evsel->no_aux_samples) { 1246 evsel__set_sample_bit(evsel, TIME); 1247 evsel__set_sample_bit(evsel, RAW); 1248 evsel__set_sample_bit(evsel, CPU); 1249 } 1250 1251 if (opts->sample_address) 1252 evsel__set_sample_bit(evsel, DATA_SRC); 1253 1254 if (opts->sample_phys_addr) 1255 evsel__set_sample_bit(evsel, PHYS_ADDR); 1256 1257 if (opts->no_buffering) { 1258 attr->watermark = 0; 1259 attr->wakeup_events = 1; 1260 } 1261 if (opts->branch_stack && !evsel->no_aux_samples) { 1262 evsel__set_sample_bit(evsel, BRANCH_STACK); 1263 attr->branch_sample_type = opts->branch_stack; 1264 } 1265 1266 if (opts->sample_weight) 1267 arch_evsel__set_sample_weight(evsel); 1268 1269 attr->task = track; 1270 attr->mmap = track; 1271 attr->mmap2 = track && !perf_missing_features.mmap2; 1272 attr->comm = track; 1273 attr->build_id = track && opts->build_id; 1274 1275 /* 1276 * ksymbol is tracked separately with text poke because it needs to be 1277 * system wide and enabled immediately. 1278 */ 1279 if (!opts->text_poke) 1280 attr->ksymbol = track && !perf_missing_features.ksymbol; 1281 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf; 1282 1283 if (opts->record_namespaces) 1284 attr->namespaces = track; 1285 1286 if (opts->record_cgroup) { 1287 attr->cgroup = track && !perf_missing_features.cgroup; 1288 evsel__set_sample_bit(evsel, CGROUP); 1289 } 1290 1291 if (opts->sample_data_page_size) 1292 evsel__set_sample_bit(evsel, DATA_PAGE_SIZE); 1293 1294 if (opts->sample_code_page_size) 1295 evsel__set_sample_bit(evsel, CODE_PAGE_SIZE); 1296 1297 if (opts->record_switch_events) 1298 attr->context_switch = track; 1299 1300 if (opts->sample_transaction) 1301 evsel__set_sample_bit(evsel, TRANSACTION); 1302 1303 if (opts->running_time) { 1304 evsel->core.attr.read_format |= 1305 PERF_FORMAT_TOTAL_TIME_ENABLED | 1306 PERF_FORMAT_TOTAL_TIME_RUNNING; 1307 } 1308 1309 /* 1310 * XXX see the function comment above 1311 * 1312 * Disabling only independent events or group leaders, 1313 * keeping group members enabled. 1314 */ 1315 if (evsel__is_group_leader(evsel)) 1316 attr->disabled = 1; 1317 1318 /* 1319 * Setting enable_on_exec for independent events and 1320 * group leaders for traced executed by perf. 1321 */ 1322 if (target__none(&opts->target) && evsel__is_group_leader(evsel) && 1323 !opts->target.initial_delay) 1324 attr->enable_on_exec = 1; 1325 1326 if (evsel->immediate) { 1327 attr->disabled = 0; 1328 attr->enable_on_exec = 0; 1329 } 1330 1331 clockid = opts->clockid; 1332 if (opts->use_clockid) { 1333 attr->use_clockid = 1; 1334 attr->clockid = opts->clockid; 1335 } 1336 1337 if (evsel->precise_max) 1338 attr->precise_ip = 3; 1339 1340 if (opts->all_user) { 1341 attr->exclude_kernel = 1; 1342 attr->exclude_user = 0; 1343 } 1344 1345 if (opts->all_kernel) { 1346 attr->exclude_kernel = 0; 1347 attr->exclude_user = 1; 1348 } 1349 1350 if (evsel->core.own_cpus || evsel->unit) 1351 evsel->core.attr.read_format |= PERF_FORMAT_ID; 1352 1353 /* 1354 * Apply event specific term settings, 1355 * it overloads any global configuration. 1356 */ 1357 evsel__apply_config_terms(evsel, opts, track); 1358 1359 evsel->ignore_missing_thread = opts->ignore_missing_thread; 1360 1361 /* The --period option takes the precedence. */ 1362 if (opts->period_set) { 1363 if (opts->period) 1364 evsel__set_sample_bit(evsel, PERIOD); 1365 else 1366 evsel__reset_sample_bit(evsel, PERIOD); 1367 } 1368 1369 /* 1370 * A dummy event never triggers any actual counter and therefore 1371 * cannot be used with branch_stack. 1372 * 1373 * For initial_delay, a dummy event is added implicitly. 1374 * The software event will trigger -EOPNOTSUPP error out, 1375 * if BRANCH_STACK bit is set. 1376 */ 1377 if (evsel__is_dummy_event(evsel)) 1378 evsel__reset_sample_bit(evsel, BRANCH_STACK); 1379 1380 if (evsel__is_offcpu_event(evsel)) 1381 evsel->core.attr.sample_type &= OFFCPU_SAMPLE_TYPES; 1382 1383 arch__post_evsel_config(evsel, attr); 1384 } 1385 1386 int evsel__set_filter(struct evsel *evsel, const char *filter) 1387 { 1388 char *new_filter = strdup(filter); 1389 1390 if (new_filter != NULL) { 1391 free(evsel->filter); 1392 evsel->filter = new_filter; 1393 return 0; 1394 } 1395 1396 return -1; 1397 } 1398 1399 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter) 1400 { 1401 char *new_filter; 1402 1403 if (evsel->filter == NULL) 1404 return evsel__set_filter(evsel, filter); 1405 1406 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) { 1407 free(evsel->filter); 1408 evsel->filter = new_filter; 1409 return 0; 1410 } 1411 1412 return -1; 1413 } 1414 1415 int evsel__append_tp_filter(struct evsel *evsel, const char *filter) 1416 { 1417 return evsel__append_filter(evsel, "(%s) && (%s)", filter); 1418 } 1419 1420 int evsel__append_addr_filter(struct evsel *evsel, const char *filter) 1421 { 1422 return evsel__append_filter(evsel, "%s,%s", filter); 1423 } 1424 1425 /* Caller has to clear disabled after going through all CPUs. */ 1426 int evsel__enable_cpu(struct evsel *evsel, int cpu_map_idx) 1427 { 1428 return perf_evsel__enable_cpu(&evsel->core, cpu_map_idx); 1429 } 1430 1431 int evsel__enable(struct evsel *evsel) 1432 { 1433 int err = perf_evsel__enable(&evsel->core); 1434 1435 if (!err) 1436 evsel->disabled = false; 1437 return err; 1438 } 1439 1440 /* Caller has to set disabled after going through all CPUs. */ 1441 int evsel__disable_cpu(struct evsel *evsel, int cpu_map_idx) 1442 { 1443 return perf_evsel__disable_cpu(&evsel->core, cpu_map_idx); 1444 } 1445 1446 int evsel__disable(struct evsel *evsel) 1447 { 1448 int err = perf_evsel__disable(&evsel->core); 1449 /* 1450 * We mark it disabled here so that tools that disable a event can 1451 * ignore events after they disable it. I.e. the ring buffer may have 1452 * already a few more events queued up before the kernel got the stop 1453 * request. 1454 */ 1455 if (!err) 1456 evsel->disabled = true; 1457 1458 return err; 1459 } 1460 1461 void free_config_terms(struct list_head *config_terms) 1462 { 1463 struct evsel_config_term *term, *h; 1464 1465 list_for_each_entry_safe(term, h, config_terms, list) { 1466 list_del_init(&term->list); 1467 if (term->free_str) 1468 zfree(&term->val.str); 1469 free(term); 1470 } 1471 } 1472 1473 static void evsel__free_config_terms(struct evsel *evsel) 1474 { 1475 free_config_terms(&evsel->config_terms); 1476 } 1477 1478 void evsel__exit(struct evsel *evsel) 1479 { 1480 assert(list_empty(&evsel->core.node)); 1481 assert(evsel->evlist == NULL); 1482 bpf_counter__destroy(evsel); 1483 perf_bpf_filter__destroy(evsel); 1484 evsel__free_counts(evsel); 1485 perf_evsel__free_fd(&evsel->core); 1486 perf_evsel__free_id(&evsel->core); 1487 evsel__free_config_terms(evsel); 1488 cgroup__put(evsel->cgrp); 1489 perf_cpu_map__put(evsel->core.cpus); 1490 perf_cpu_map__put(evsel->core.own_cpus); 1491 perf_thread_map__put(evsel->core.threads); 1492 zfree(&evsel->group_name); 1493 zfree(&evsel->name); 1494 zfree(&evsel->filter); 1495 zfree(&evsel->group_pmu_name); 1496 zfree(&evsel->unit); 1497 zfree(&evsel->metric_id); 1498 evsel__zero_per_pkg(evsel); 1499 hashmap__free(evsel->per_pkg_mask); 1500 evsel->per_pkg_mask = NULL; 1501 zfree(&evsel->metric_events); 1502 perf_evsel__object.fini(evsel); 1503 if (evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME || 1504 evsel__tool_event(evsel) == PERF_TOOL_USER_TIME) 1505 xyarray__delete(evsel->start_times); 1506 } 1507 1508 void evsel__delete(struct evsel *evsel) 1509 { 1510 if (!evsel) 1511 return; 1512 1513 evsel__exit(evsel); 1514 free(evsel); 1515 } 1516 1517 void evsel__compute_deltas(struct evsel *evsel, int cpu_map_idx, int thread, 1518 struct perf_counts_values *count) 1519 { 1520 struct perf_counts_values tmp; 1521 1522 if (!evsel->prev_raw_counts) 1523 return; 1524 1525 tmp = *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread); 1526 *perf_counts(evsel->prev_raw_counts, cpu_map_idx, thread) = *count; 1527 1528 count->val = count->val - tmp.val; 1529 count->ena = count->ena - tmp.ena; 1530 count->run = count->run - tmp.run; 1531 } 1532 1533 static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread) 1534 { 1535 struct perf_counts_values *count = perf_counts(evsel->counts, cpu_map_idx, thread); 1536 1537 return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count); 1538 } 1539 1540 static int evsel__read_retire_lat(struct evsel *evsel, int cpu_map_idx, int thread) 1541 { 1542 return tpebs_set_evsel(evsel, cpu_map_idx, thread); 1543 } 1544 1545 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread, 1546 u64 val, u64 ena, u64 run, u64 lost) 1547 { 1548 struct perf_counts_values *count; 1549 1550 count = perf_counts(counter->counts, cpu_map_idx, thread); 1551 1552 if (counter->retire_lat) { 1553 evsel__read_retire_lat(counter, cpu_map_idx, thread); 1554 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true); 1555 return; 1556 } 1557 1558 count->val = val; 1559 count->ena = ena; 1560 count->run = run; 1561 count->lost = lost; 1562 1563 perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true); 1564 } 1565 1566 static bool evsel__group_has_tpebs(struct evsel *leader) 1567 { 1568 struct evsel *evsel; 1569 1570 for_each_group_evsel(evsel, leader) { 1571 if (evsel__is_retire_lat(evsel)) 1572 return true; 1573 } 1574 return false; 1575 } 1576 1577 static u64 evsel__group_read_nr_members(struct evsel *leader) 1578 { 1579 u64 nr = leader->core.nr_members; 1580 struct evsel *evsel; 1581 1582 for_each_group_evsel(evsel, leader) { 1583 if (evsel__is_retire_lat(evsel)) 1584 nr--; 1585 } 1586 return nr; 1587 } 1588 1589 static u64 evsel__group_read_size(struct evsel *leader) 1590 { 1591 u64 read_format = leader->core.attr.read_format; 1592 int entry = sizeof(u64); /* value */ 1593 int size = 0; 1594 int nr = 1; 1595 1596 if (!evsel__group_has_tpebs(leader)) 1597 return perf_evsel__read_size(&leader->core); 1598 1599 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1600 size += sizeof(u64); 1601 1602 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1603 size += sizeof(u64); 1604 1605 if (read_format & PERF_FORMAT_ID) 1606 entry += sizeof(u64); 1607 1608 if (read_format & PERF_FORMAT_LOST) 1609 entry += sizeof(u64); 1610 1611 if (read_format & PERF_FORMAT_GROUP) { 1612 nr = evsel__group_read_nr_members(leader); 1613 size += sizeof(u64); 1614 } 1615 1616 size += entry * nr; 1617 return size; 1618 } 1619 1620 static int evsel__process_group_data(struct evsel *leader, int cpu_map_idx, int thread, u64 *data) 1621 { 1622 u64 read_format = leader->core.attr.read_format; 1623 struct sample_read_value *v; 1624 u64 nr, ena = 0, run = 0, lost = 0; 1625 1626 nr = *data++; 1627 1628 if (nr != evsel__group_read_nr_members(leader)) 1629 return -EINVAL; 1630 1631 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1632 ena = *data++; 1633 1634 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1635 run = *data++; 1636 1637 v = (void *)data; 1638 sample_read_group__for_each(v, nr, read_format) { 1639 struct evsel *counter; 1640 1641 counter = evlist__id2evsel(leader->evlist, v->id); 1642 if (!counter) 1643 return -EINVAL; 1644 1645 if (read_format & PERF_FORMAT_LOST) 1646 lost = v->lost; 1647 1648 evsel__set_count(counter, cpu_map_idx, thread, v->value, ena, run, lost); 1649 } 1650 1651 return 0; 1652 } 1653 1654 static int evsel__read_group(struct evsel *leader, int cpu_map_idx, int thread) 1655 { 1656 struct perf_stat_evsel *ps = leader->stats; 1657 u64 read_format = leader->core.attr.read_format; 1658 int size = evsel__group_read_size(leader); 1659 u64 *data = ps->group_data; 1660 1661 if (!(read_format & PERF_FORMAT_ID)) 1662 return -EINVAL; 1663 1664 if (!evsel__is_group_leader(leader)) 1665 return -EINVAL; 1666 1667 if (!data) { 1668 data = zalloc(size); 1669 if (!data) 1670 return -ENOMEM; 1671 1672 ps->group_data = data; 1673 } 1674 1675 if (FD(leader, cpu_map_idx, thread) < 0) 1676 return -EINVAL; 1677 1678 if (readn(FD(leader, cpu_map_idx, thread), data, size) <= 0) 1679 return -errno; 1680 1681 return evsel__process_group_data(leader, cpu_map_idx, thread, data); 1682 } 1683 1684 static bool read_until_char(struct io *io, char e) 1685 { 1686 int c; 1687 1688 do { 1689 c = io__get_char(io); 1690 if (c == -1) 1691 return false; 1692 } while (c != e); 1693 return true; 1694 } 1695 1696 static int read_stat_field(int fd, struct perf_cpu cpu, int field, __u64 *val) 1697 { 1698 char buf[256]; 1699 struct io io; 1700 int i; 1701 1702 io__init(&io, fd, buf, sizeof(buf)); 1703 1704 /* Skip lines to relevant CPU. */ 1705 for (i = -1; i < cpu.cpu; i++) { 1706 if (!read_until_char(&io, '\n')) 1707 return -EINVAL; 1708 } 1709 /* Skip to "cpu". */ 1710 if (io__get_char(&io) != 'c') return -EINVAL; 1711 if (io__get_char(&io) != 'p') return -EINVAL; 1712 if (io__get_char(&io) != 'u') return -EINVAL; 1713 1714 /* Skip N of cpuN. */ 1715 if (!read_until_char(&io, ' ')) 1716 return -EINVAL; 1717 1718 i = 1; 1719 while (true) { 1720 if (io__get_dec(&io, val) != ' ') 1721 break; 1722 if (field == i) 1723 return 0; 1724 i++; 1725 } 1726 return -EINVAL; 1727 } 1728 1729 static int read_pid_stat_field(int fd, int field, __u64 *val) 1730 { 1731 char buf[256]; 1732 struct io io; 1733 int c, i; 1734 1735 io__init(&io, fd, buf, sizeof(buf)); 1736 if (io__get_dec(&io, val) != ' ') 1737 return -EINVAL; 1738 if (field == 1) 1739 return 0; 1740 1741 /* Skip comm. */ 1742 if (io__get_char(&io) != '(' || !read_until_char(&io, ')')) 1743 return -EINVAL; 1744 if (field == 2) 1745 return -EINVAL; /* String can't be returned. */ 1746 1747 /* Skip state */ 1748 if (io__get_char(&io) != ' ' || io__get_char(&io) == -1) 1749 return -EINVAL; 1750 if (field == 3) 1751 return -EINVAL; /* String can't be returned. */ 1752 1753 /* Loop over numeric fields*/ 1754 if (io__get_char(&io) != ' ') 1755 return -EINVAL; 1756 1757 i = 4; 1758 while (true) { 1759 c = io__get_dec(&io, val); 1760 if (c == -1) 1761 return -EINVAL; 1762 if (c == -2) { 1763 /* Assume a -ve was read */ 1764 c = io__get_dec(&io, val); 1765 *val *= -1; 1766 } 1767 if (c != ' ') 1768 return -EINVAL; 1769 if (field == i) 1770 return 0; 1771 i++; 1772 } 1773 return -EINVAL; 1774 } 1775 1776 static int evsel__read_tool(struct evsel *evsel, int cpu_map_idx, int thread) 1777 { 1778 __u64 *start_time, cur_time, delta_start; 1779 int fd, err = 0; 1780 struct perf_counts_values *count; 1781 bool adjust = false; 1782 1783 count = perf_counts(evsel->counts, cpu_map_idx, thread); 1784 1785 switch (evsel__tool_event(evsel)) { 1786 case PERF_TOOL_DURATION_TIME: 1787 /* 1788 * Pretend duration_time is only on the first CPU and thread, or 1789 * else aggregation will scale duration_time by the number of 1790 * CPUs/threads. 1791 */ 1792 start_time = &evsel->start_time; 1793 if (cpu_map_idx == 0 && thread == 0) 1794 cur_time = rdclock(); 1795 else 1796 cur_time = *start_time; 1797 break; 1798 case PERF_TOOL_USER_TIME: 1799 case PERF_TOOL_SYSTEM_TIME: { 1800 bool system = evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME; 1801 1802 start_time = xyarray__entry(evsel->start_times, cpu_map_idx, thread); 1803 fd = FD(evsel, cpu_map_idx, thread); 1804 lseek(fd, SEEK_SET, 0); 1805 if (evsel->pid_stat) { 1806 /* The event exists solely on 1 CPU. */ 1807 if (cpu_map_idx == 0) 1808 err = read_pid_stat_field(fd, system ? 15 : 14, &cur_time); 1809 else 1810 cur_time = 0; 1811 } else { 1812 /* The event is for all threads. */ 1813 if (thread == 0) { 1814 struct perf_cpu cpu = perf_cpu_map__cpu(evsel->core.cpus, 1815 cpu_map_idx); 1816 1817 err = read_stat_field(fd, cpu, system ? 3 : 1, &cur_time); 1818 } else { 1819 cur_time = 0; 1820 } 1821 } 1822 adjust = true; 1823 break; 1824 } 1825 case PERF_TOOL_NONE: 1826 case PERF_TOOL_MAX: 1827 default: 1828 err = -EINVAL; 1829 } 1830 if (err) 1831 return err; 1832 1833 delta_start = cur_time - *start_time; 1834 if (adjust) { 1835 __u64 ticks_per_sec = sysconf(_SC_CLK_TCK); 1836 1837 delta_start *= 1000000000 / ticks_per_sec; 1838 } 1839 count->val = delta_start; 1840 count->ena = count->run = delta_start; 1841 count->lost = 0; 1842 return 0; 1843 } 1844 1845 bool __evsel__match(const struct evsel *evsel, u32 type, u64 config) 1846 { 1847 1848 u32 e_type = evsel->core.attr.type; 1849 u64 e_config = evsel->core.attr.config; 1850 1851 if (e_type != type) { 1852 return type == PERF_TYPE_HARDWARE && evsel->pmu && evsel->pmu->is_core && 1853 evsel->alternate_hw_config == config; 1854 } 1855 1856 if ((type == PERF_TYPE_HARDWARE || type == PERF_TYPE_HW_CACHE) && 1857 perf_pmus__supports_extended_type()) 1858 e_config &= PERF_HW_EVENT_MASK; 1859 1860 return e_config == config; 1861 } 1862 1863 int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread) 1864 { 1865 if (evsel__is_tool(evsel)) 1866 return evsel__read_tool(evsel, cpu_map_idx, thread); 1867 1868 if (evsel__is_retire_lat(evsel)) 1869 return evsel__read_retire_lat(evsel, cpu_map_idx, thread); 1870 1871 if (evsel->core.attr.read_format & PERF_FORMAT_GROUP) 1872 return evsel__read_group(evsel, cpu_map_idx, thread); 1873 1874 return evsel__read_one(evsel, cpu_map_idx, thread); 1875 } 1876 1877 int __evsel__read_on_cpu(struct evsel *evsel, int cpu_map_idx, int thread, bool scale) 1878 { 1879 struct perf_counts_values count; 1880 size_t nv = scale ? 3 : 1; 1881 1882 if (FD(evsel, cpu_map_idx, thread) < 0) 1883 return -EINVAL; 1884 1885 if (evsel->counts == NULL && evsel__alloc_counts(evsel) < 0) 1886 return -ENOMEM; 1887 1888 if (readn(FD(evsel, cpu_map_idx, thread), &count, nv * sizeof(u64)) <= 0) 1889 return -errno; 1890 1891 evsel__compute_deltas(evsel, cpu_map_idx, thread, &count); 1892 perf_counts_values__scale(&count, scale, NULL); 1893 *perf_counts(evsel->counts, cpu_map_idx, thread) = count; 1894 return 0; 1895 } 1896 1897 static int evsel__match_other_cpu(struct evsel *evsel, struct evsel *other, 1898 int cpu_map_idx) 1899 { 1900 struct perf_cpu cpu; 1901 1902 cpu = perf_cpu_map__cpu(evsel->core.cpus, cpu_map_idx); 1903 return perf_cpu_map__idx(other->core.cpus, cpu); 1904 } 1905 1906 static int evsel__hybrid_group_cpu_map_idx(struct evsel *evsel, int cpu_map_idx) 1907 { 1908 struct evsel *leader = evsel__leader(evsel); 1909 1910 if ((evsel__is_hybrid(evsel) && !evsel__is_hybrid(leader)) || 1911 (!evsel__is_hybrid(evsel) && evsel__is_hybrid(leader))) { 1912 return evsel__match_other_cpu(evsel, leader, cpu_map_idx); 1913 } 1914 1915 return cpu_map_idx; 1916 } 1917 1918 static int get_group_fd(struct evsel *evsel, int cpu_map_idx, int thread) 1919 { 1920 struct evsel *leader = evsel__leader(evsel); 1921 int fd; 1922 1923 if (evsel__is_group_leader(evsel)) 1924 return -1; 1925 1926 /* 1927 * Leader must be already processed/open, 1928 * if not it's a bug. 1929 */ 1930 BUG_ON(!leader->core.fd); 1931 1932 cpu_map_idx = evsel__hybrid_group_cpu_map_idx(evsel, cpu_map_idx); 1933 if (cpu_map_idx == -1) 1934 return -1; 1935 1936 fd = FD(leader, cpu_map_idx, thread); 1937 BUG_ON(fd == -1 && !leader->skippable); 1938 1939 /* 1940 * When the leader has been skipped, return -2 to distinguish from no 1941 * group leader case. 1942 */ 1943 return fd == -1 ? -2 : fd; 1944 } 1945 1946 static void evsel__remove_fd(struct evsel *pos, int nr_cpus, int nr_threads, int thread_idx) 1947 { 1948 for (int cpu = 0; cpu < nr_cpus; cpu++) 1949 for (int thread = thread_idx; thread < nr_threads - 1; thread++) 1950 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1); 1951 } 1952 1953 static int update_fds(struct evsel *evsel, 1954 int nr_cpus, int cpu_map_idx, 1955 int nr_threads, int thread_idx) 1956 { 1957 struct evsel *pos; 1958 1959 if (cpu_map_idx >= nr_cpus || thread_idx >= nr_threads) 1960 return -EINVAL; 1961 1962 evlist__for_each_entry(evsel->evlist, pos) { 1963 nr_cpus = pos != evsel ? nr_cpus : cpu_map_idx; 1964 1965 evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx); 1966 1967 /* 1968 * Since fds for next evsel has not been created, 1969 * there is no need to iterate whole event list. 1970 */ 1971 if (pos == evsel) 1972 break; 1973 } 1974 return 0; 1975 } 1976 1977 static bool evsel__ignore_missing_thread(struct evsel *evsel, 1978 int nr_cpus, int cpu_map_idx, 1979 struct perf_thread_map *threads, 1980 int thread, int err) 1981 { 1982 pid_t ignore_pid = perf_thread_map__pid(threads, thread); 1983 1984 if (!evsel->ignore_missing_thread) 1985 return false; 1986 1987 /* The system wide setup does not work with threads. */ 1988 if (evsel->core.system_wide) 1989 return false; 1990 1991 /* The -ESRCH is perf event syscall errno for pid's not found. */ 1992 if (err != -ESRCH) 1993 return false; 1994 1995 /* If there's only one thread, let it fail. */ 1996 if (threads->nr == 1) 1997 return false; 1998 1999 /* 2000 * We should remove fd for missing_thread first 2001 * because thread_map__remove() will decrease threads->nr. 2002 */ 2003 if (update_fds(evsel, nr_cpus, cpu_map_idx, threads->nr, thread)) 2004 return false; 2005 2006 if (thread_map__remove(threads, thread)) 2007 return false; 2008 2009 pr_warning("WARNING: Ignored open failure for pid %d\n", 2010 ignore_pid); 2011 return true; 2012 } 2013 2014 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val, 2015 void *priv __maybe_unused) 2016 { 2017 return fprintf(fp, " %-32s %s\n", name, val); 2018 } 2019 2020 static void display_attr(struct perf_event_attr *attr) 2021 { 2022 if (verbose >= 2 || debug_peo_args) { 2023 fprintf(stderr, "%.60s\n", graph_dotted_line); 2024 fprintf(stderr, "perf_event_attr:\n"); 2025 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL); 2026 fprintf(stderr, "%.60s\n", graph_dotted_line); 2027 } 2028 } 2029 2030 bool evsel__precise_ip_fallback(struct evsel *evsel) 2031 { 2032 /* Do not try less precise if not requested. */ 2033 if (!evsel->precise_max) 2034 return false; 2035 2036 /* 2037 * We tried all the precise_ip values, and it's 2038 * still failing, so leave it to standard fallback. 2039 */ 2040 if (!evsel->core.attr.precise_ip) { 2041 evsel->core.attr.precise_ip = evsel->precise_ip_original; 2042 return false; 2043 } 2044 2045 if (!evsel->precise_ip_original) 2046 evsel->precise_ip_original = evsel->core.attr.precise_ip; 2047 2048 evsel->core.attr.precise_ip--; 2049 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip); 2050 display_attr(&evsel->core.attr); 2051 return true; 2052 } 2053 2054 static struct perf_cpu_map *empty_cpu_map; 2055 static struct perf_thread_map *empty_thread_map; 2056 2057 static int __evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus, 2058 struct perf_thread_map *threads) 2059 { 2060 int nthreads = perf_thread_map__nr(threads); 2061 2062 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) || 2063 (perf_missing_features.aux_output && evsel->core.attr.aux_output)) 2064 return -EINVAL; 2065 2066 if (cpus == NULL) { 2067 if (empty_cpu_map == NULL) { 2068 empty_cpu_map = perf_cpu_map__new_any_cpu(); 2069 if (empty_cpu_map == NULL) 2070 return -ENOMEM; 2071 } 2072 2073 cpus = empty_cpu_map; 2074 } 2075 2076 if (threads == NULL) { 2077 if (empty_thread_map == NULL) { 2078 empty_thread_map = thread_map__new_by_tid(-1); 2079 if (empty_thread_map == NULL) 2080 return -ENOMEM; 2081 } 2082 2083 threads = empty_thread_map; 2084 } 2085 2086 if (evsel->core.fd == NULL && 2087 perf_evsel__alloc_fd(&evsel->core, perf_cpu_map__nr(cpus), nthreads) < 0) 2088 return -ENOMEM; 2089 2090 if ((evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME || 2091 evsel__tool_event(evsel) == PERF_TOOL_USER_TIME) && 2092 !evsel->start_times) { 2093 evsel->start_times = xyarray__new(perf_cpu_map__nr(cpus), nthreads, sizeof(__u64)); 2094 if (!evsel->start_times) 2095 return -ENOMEM; 2096 } 2097 2098 evsel->open_flags = PERF_FLAG_FD_CLOEXEC; 2099 if (evsel->cgrp) 2100 evsel->open_flags |= PERF_FLAG_PID_CGROUP; 2101 2102 return 0; 2103 } 2104 2105 static void evsel__disable_missing_features(struct evsel *evsel) 2106 { 2107 if (perf_missing_features.branch_counters) 2108 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_COUNTERS; 2109 if (perf_missing_features.read_lost) 2110 evsel->core.attr.read_format &= ~PERF_FORMAT_LOST; 2111 if (perf_missing_features.weight_struct) { 2112 evsel__set_sample_bit(evsel, WEIGHT); 2113 evsel__reset_sample_bit(evsel, WEIGHT_STRUCT); 2114 } 2115 if (perf_missing_features.clockid_wrong) 2116 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */ 2117 if (perf_missing_features.clockid) { 2118 evsel->core.attr.use_clockid = 0; 2119 evsel->core.attr.clockid = 0; 2120 } 2121 if (perf_missing_features.cloexec) 2122 evsel->open_flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 2123 if (perf_missing_features.mmap2) 2124 evsel->core.attr.mmap2 = 0; 2125 if (evsel->pmu && evsel->pmu->missing_features.exclude_guest) 2126 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0; 2127 if (perf_missing_features.lbr_flags) 2128 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS | 2129 PERF_SAMPLE_BRANCH_NO_CYCLES); 2130 if (perf_missing_features.group_read && evsel->core.attr.inherit) 2131 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID); 2132 if (perf_missing_features.ksymbol) 2133 evsel->core.attr.ksymbol = 0; 2134 if (perf_missing_features.bpf) 2135 evsel->core.attr.bpf_event = 0; 2136 if (perf_missing_features.branch_hw_idx) 2137 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX; 2138 if (perf_missing_features.sample_id_all) 2139 evsel->core.attr.sample_id_all = 0; 2140 } 2141 2142 int evsel__prepare_open(struct evsel *evsel, struct perf_cpu_map *cpus, 2143 struct perf_thread_map *threads) 2144 { 2145 int err; 2146 2147 err = __evsel__prepare_open(evsel, cpus, threads); 2148 if (err) 2149 return err; 2150 2151 evsel__disable_missing_features(evsel); 2152 2153 return err; 2154 } 2155 2156 bool evsel__detect_missing_features(struct evsel *evsel) 2157 { 2158 /* 2159 * Must probe features in the order they were added to the 2160 * perf_event_attr interface. 2161 */ 2162 if (!perf_missing_features.branch_counters && 2163 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_COUNTERS)) { 2164 perf_missing_features.branch_counters = true; 2165 pr_debug2("switching off branch counters support\n"); 2166 return true; 2167 } else if (!perf_missing_features.read_lost && 2168 (evsel->core.attr.read_format & PERF_FORMAT_LOST)) { 2169 perf_missing_features.read_lost = true; 2170 pr_debug2("switching off PERF_FORMAT_LOST support\n"); 2171 return true; 2172 } else if (!perf_missing_features.weight_struct && 2173 (evsel->core.attr.sample_type & PERF_SAMPLE_WEIGHT_STRUCT)) { 2174 perf_missing_features.weight_struct = true; 2175 pr_debug2("switching off weight struct support\n"); 2176 return true; 2177 } else if (!perf_missing_features.code_page_size && 2178 (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)) { 2179 perf_missing_features.code_page_size = true; 2180 pr_debug2_peo("Kernel has no PERF_SAMPLE_CODE_PAGE_SIZE support, bailing out\n"); 2181 return false; 2182 } else if (!perf_missing_features.data_page_size && 2183 (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)) { 2184 perf_missing_features.data_page_size = true; 2185 pr_debug2_peo("Kernel has no PERF_SAMPLE_DATA_PAGE_SIZE support, bailing out\n"); 2186 return false; 2187 } else if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) { 2188 perf_missing_features.cgroup = true; 2189 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n"); 2190 return false; 2191 } else if (!perf_missing_features.branch_hw_idx && 2192 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) { 2193 perf_missing_features.branch_hw_idx = true; 2194 pr_debug2("switching off branch HW index support\n"); 2195 return true; 2196 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) { 2197 perf_missing_features.aux_output = true; 2198 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n"); 2199 return false; 2200 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) { 2201 perf_missing_features.bpf = true; 2202 pr_debug2_peo("switching off bpf_event\n"); 2203 return true; 2204 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) { 2205 perf_missing_features.ksymbol = true; 2206 pr_debug2_peo("switching off ksymbol\n"); 2207 return true; 2208 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) { 2209 perf_missing_features.write_backward = true; 2210 pr_debug2_peo("switching off write_backward\n"); 2211 return false; 2212 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) { 2213 perf_missing_features.clockid_wrong = true; 2214 pr_debug2_peo("switching off clockid\n"); 2215 return true; 2216 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) { 2217 perf_missing_features.clockid = true; 2218 pr_debug2_peo("switching off use_clockid\n"); 2219 return true; 2220 } else if (!perf_missing_features.cloexec && (evsel->open_flags & PERF_FLAG_FD_CLOEXEC)) { 2221 perf_missing_features.cloexec = true; 2222 pr_debug2_peo("switching off cloexec flag\n"); 2223 return true; 2224 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) { 2225 perf_missing_features.mmap2 = true; 2226 pr_debug2_peo("switching off mmap2\n"); 2227 return true; 2228 } else if (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host) { 2229 if (evsel->pmu == NULL) 2230 evsel->pmu = evsel__find_pmu(evsel); 2231 2232 if (evsel->pmu) 2233 evsel->pmu->missing_features.exclude_guest = true; 2234 else { 2235 /* we cannot find PMU, disable attrs now */ 2236 evsel->core.attr.exclude_host = false; 2237 evsel->core.attr.exclude_guest = false; 2238 } 2239 2240 if (evsel->exclude_GH) { 2241 pr_debug2_peo("PMU has no exclude_host/guest support, bailing out\n"); 2242 return false; 2243 } 2244 if (!perf_missing_features.exclude_guest) { 2245 perf_missing_features.exclude_guest = true; 2246 pr_debug2_peo("switching off exclude_guest, exclude_host\n"); 2247 } 2248 return true; 2249 } else if (!perf_missing_features.sample_id_all) { 2250 perf_missing_features.sample_id_all = true; 2251 pr_debug2_peo("switching off sample_id_all\n"); 2252 return true; 2253 } else if (!perf_missing_features.lbr_flags && 2254 (evsel->core.attr.branch_sample_type & 2255 (PERF_SAMPLE_BRANCH_NO_CYCLES | 2256 PERF_SAMPLE_BRANCH_NO_FLAGS))) { 2257 perf_missing_features.lbr_flags = true; 2258 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n"); 2259 return true; 2260 } else if (!perf_missing_features.group_read && 2261 evsel->core.attr.inherit && 2262 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) && 2263 evsel__is_group_leader(evsel)) { 2264 perf_missing_features.group_read = true; 2265 pr_debug2_peo("switching off group read\n"); 2266 return true; 2267 } else { 2268 return false; 2269 } 2270 } 2271 2272 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, 2273 struct perf_thread_map *threads, 2274 int start_cpu_map_idx, int end_cpu_map_idx) 2275 { 2276 int idx, thread, nthreads; 2277 int pid = -1, err, old_errno; 2278 enum rlimit_action set_rlimit = NO_CHANGE; 2279 2280 if (evsel__tool_event(evsel) == PERF_TOOL_DURATION_TIME) { 2281 if (evsel->core.attr.sample_period) /* no sampling */ 2282 return -EINVAL; 2283 evsel->start_time = rdclock(); 2284 return 0; 2285 } 2286 2287 if (evsel__is_retire_lat(evsel)) 2288 return tpebs_start(evsel->evlist); 2289 2290 err = __evsel__prepare_open(evsel, cpus, threads); 2291 if (err) 2292 return err; 2293 2294 if (cpus == NULL) 2295 cpus = empty_cpu_map; 2296 2297 if (threads == NULL) 2298 threads = empty_thread_map; 2299 2300 nthreads = perf_thread_map__nr(threads); 2301 2302 if (evsel->cgrp) 2303 pid = evsel->cgrp->fd; 2304 2305 fallback_missing_features: 2306 evsel__disable_missing_features(evsel); 2307 2308 pr_debug3("Opening: %s\n", evsel__name(evsel)); 2309 display_attr(&evsel->core.attr); 2310 2311 for (idx = start_cpu_map_idx; idx < end_cpu_map_idx; idx++) { 2312 2313 for (thread = 0; thread < nthreads; thread++) { 2314 int fd, group_fd; 2315 retry_open: 2316 if (thread >= nthreads) 2317 break; 2318 2319 if (!evsel->cgrp && !evsel->core.system_wide) 2320 pid = perf_thread_map__pid(threads, thread); 2321 2322 if (evsel__tool_event(evsel) == PERF_TOOL_USER_TIME || 2323 evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME) { 2324 bool system = evsel__tool_event(evsel) == PERF_TOOL_SYSTEM_TIME; 2325 __u64 *start_time = NULL; 2326 2327 if (evsel->core.attr.sample_period) { 2328 /* no sampling */ 2329 err = -EINVAL; 2330 goto out_close; 2331 } 2332 if (pid > -1) { 2333 char buf[64]; 2334 2335 snprintf(buf, sizeof(buf), "/proc/%d/stat", pid); 2336 fd = open(buf, O_RDONLY); 2337 evsel->pid_stat = true; 2338 } else { 2339 fd = open("/proc/stat", O_RDONLY); 2340 } 2341 FD(evsel, idx, thread) = fd; 2342 if (fd < 0) { 2343 err = -errno; 2344 goto out_close; 2345 } 2346 start_time = xyarray__entry(evsel->start_times, idx, thread); 2347 if (pid > -1) { 2348 err = read_pid_stat_field(fd, system ? 15 : 14, 2349 start_time); 2350 } else { 2351 struct perf_cpu cpu; 2352 2353 cpu = perf_cpu_map__cpu(evsel->core.cpus, idx); 2354 err = read_stat_field(fd, cpu, system ? 3 : 1, 2355 start_time); 2356 } 2357 if (err) 2358 goto out_close; 2359 continue; 2360 } 2361 2362 group_fd = get_group_fd(evsel, idx, thread); 2363 2364 if (group_fd == -2) { 2365 pr_debug("broken group leader for %s\n", evsel->name); 2366 err = -EINVAL; 2367 goto out_close; 2368 } 2369 2370 test_attr__ready(); 2371 2372 /* Debug message used by test scripts */ 2373 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx", 2374 pid, perf_cpu_map__cpu(cpus, idx).cpu, group_fd, evsel->open_flags); 2375 2376 fd = sys_perf_event_open(&evsel->core.attr, pid, 2377 perf_cpu_map__cpu(cpus, idx).cpu, 2378 group_fd, evsel->open_flags); 2379 2380 FD(evsel, idx, thread) = fd; 2381 2382 if (fd < 0) { 2383 err = -errno; 2384 2385 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", 2386 err); 2387 goto try_fallback; 2388 } 2389 2390 bpf_counter__install_pe(evsel, idx, fd); 2391 2392 if (unlikely(test_attr__enabled)) { 2393 test_attr__open(&evsel->core.attr, pid, 2394 perf_cpu_map__cpu(cpus, idx), 2395 fd, group_fd, evsel->open_flags); 2396 } 2397 2398 /* Debug message used by test scripts */ 2399 pr_debug2_peo(" = %d\n", fd); 2400 2401 if (evsel->bpf_fd >= 0) { 2402 int evt_fd = fd; 2403 int bpf_fd = evsel->bpf_fd; 2404 2405 err = ioctl(evt_fd, 2406 PERF_EVENT_IOC_SET_BPF, 2407 bpf_fd); 2408 if (err && errno != EEXIST) { 2409 pr_err("failed to attach bpf fd %d: %s\n", 2410 bpf_fd, strerror(errno)); 2411 err = -EINVAL; 2412 goto out_close; 2413 } 2414 } 2415 2416 set_rlimit = NO_CHANGE; 2417 2418 /* 2419 * If we succeeded but had to kill clockid, fail and 2420 * have evsel__open_strerror() print us a nice error. 2421 */ 2422 if (perf_missing_features.clockid || 2423 perf_missing_features.clockid_wrong) { 2424 err = -EINVAL; 2425 goto out_close; 2426 } 2427 } 2428 } 2429 2430 return 0; 2431 2432 try_fallback: 2433 if (evsel__precise_ip_fallback(evsel)) 2434 goto retry_open; 2435 2436 if (evsel__ignore_missing_thread(evsel, perf_cpu_map__nr(cpus), 2437 idx, threads, thread, err)) { 2438 /* We just removed 1 thread, so lower the upper nthreads limit. */ 2439 nthreads--; 2440 2441 /* ... and pretend like nothing have happened. */ 2442 err = 0; 2443 goto retry_open; 2444 } 2445 /* 2446 * perf stat needs between 5 and 22 fds per CPU. When we run out 2447 * of them try to increase the limits. 2448 */ 2449 if (err == -EMFILE && rlimit__increase_nofile(&set_rlimit)) 2450 goto retry_open; 2451 2452 if (err != -EINVAL || idx > 0 || thread > 0) 2453 goto out_close; 2454 2455 if (evsel__detect_missing_features(evsel)) 2456 goto fallback_missing_features; 2457 out_close: 2458 if (err) 2459 threads->err_thread = thread; 2460 2461 old_errno = errno; 2462 do { 2463 while (--thread >= 0) { 2464 if (FD(evsel, idx, thread) >= 0) 2465 close(FD(evsel, idx, thread)); 2466 FD(evsel, idx, thread) = -1; 2467 } 2468 thread = nthreads; 2469 } while (--idx >= 0); 2470 errno = old_errno; 2471 return err; 2472 } 2473 2474 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus, 2475 struct perf_thread_map *threads) 2476 { 2477 return evsel__open_cpu(evsel, cpus, threads, 0, perf_cpu_map__nr(cpus)); 2478 } 2479 2480 void evsel__close(struct evsel *evsel) 2481 { 2482 if (evsel__is_retire_lat(evsel)) 2483 tpebs_delete(); 2484 perf_evsel__close(&evsel->core); 2485 perf_evsel__free_id(&evsel->core); 2486 } 2487 2488 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu_map_idx) 2489 { 2490 if (cpu_map_idx == -1) 2491 return evsel__open_cpu(evsel, cpus, NULL, 0, perf_cpu_map__nr(cpus)); 2492 2493 return evsel__open_cpu(evsel, cpus, NULL, cpu_map_idx, cpu_map_idx + 1); 2494 } 2495 2496 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads) 2497 { 2498 return evsel__open(evsel, NULL, threads); 2499 } 2500 2501 static int perf_evsel__parse_id_sample(const struct evsel *evsel, 2502 const union perf_event *event, 2503 struct perf_sample *sample) 2504 { 2505 u64 type = evsel->core.attr.sample_type; 2506 const __u64 *array = event->sample.array; 2507 bool swapped = evsel->needs_swap; 2508 union u64_swap u; 2509 2510 array += ((event->header.size - 2511 sizeof(event->header)) / sizeof(u64)) - 1; 2512 2513 if (type & PERF_SAMPLE_IDENTIFIER) { 2514 sample->id = *array; 2515 array--; 2516 } 2517 2518 if (type & PERF_SAMPLE_CPU) { 2519 u.val64 = *array; 2520 if (swapped) { 2521 /* undo swap of u64, then swap on individual u32s */ 2522 u.val64 = bswap_64(u.val64); 2523 u.val32[0] = bswap_32(u.val32[0]); 2524 } 2525 2526 sample->cpu = u.val32[0]; 2527 array--; 2528 } 2529 2530 if (type & PERF_SAMPLE_STREAM_ID) { 2531 sample->stream_id = *array; 2532 array--; 2533 } 2534 2535 if (type & PERF_SAMPLE_ID) { 2536 sample->id = *array; 2537 array--; 2538 } 2539 2540 if (type & PERF_SAMPLE_TIME) { 2541 sample->time = *array; 2542 array--; 2543 } 2544 2545 if (type & PERF_SAMPLE_TID) { 2546 u.val64 = *array; 2547 if (swapped) { 2548 /* undo swap of u64, then swap on individual u32s */ 2549 u.val64 = bswap_64(u.val64); 2550 u.val32[0] = bswap_32(u.val32[0]); 2551 u.val32[1] = bswap_32(u.val32[1]); 2552 } 2553 2554 sample->pid = u.val32[0]; 2555 sample->tid = u.val32[1]; 2556 array--; 2557 } 2558 2559 return 0; 2560 } 2561 2562 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 2563 u64 size) 2564 { 2565 return size > max_size || offset + size > endp; 2566 } 2567 2568 #define OVERFLOW_CHECK(offset, size, max_size) \ 2569 do { \ 2570 if (overflow(endp, (max_size), (offset), (size))) \ 2571 return -EFAULT; \ 2572 } while (0) 2573 2574 #define OVERFLOW_CHECK_u64(offset) \ 2575 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 2576 2577 static int 2578 perf_event__check_size(union perf_event *event, unsigned int sample_size) 2579 { 2580 /* 2581 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 2582 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 2583 * check the format does not go past the end of the event. 2584 */ 2585 if (sample_size + sizeof(event->header) > event->header.size) 2586 return -EFAULT; 2587 2588 return 0; 2589 } 2590 2591 void __weak arch_perf_parse_sample_weight(struct perf_sample *data, 2592 const __u64 *array, 2593 u64 type __maybe_unused) 2594 { 2595 data->weight = *array; 2596 } 2597 2598 u64 evsel__bitfield_swap_branch_flags(u64 value) 2599 { 2600 u64 new_val = 0; 2601 2602 /* 2603 * branch_flags 2604 * union { 2605 * u64 values; 2606 * struct { 2607 * mispred:1 //target mispredicted 2608 * predicted:1 //target predicted 2609 * in_tx:1 //in transaction 2610 * abort:1 //transaction abort 2611 * cycles:16 //cycle count to last branch 2612 * type:4 //branch type 2613 * spec:2 //branch speculation info 2614 * new_type:4 //additional branch type 2615 * priv:3 //privilege level 2616 * reserved:31 2617 * } 2618 * } 2619 * 2620 * Avoid bswap64() the entire branch_flag.value, 2621 * as it has variable bit-field sizes. Instead the 2622 * macro takes the bit-field position/size, 2623 * swaps it based on the host endianness. 2624 */ 2625 if (host_is_bigendian()) { 2626 new_val = bitfield_swap(value, 0, 1); 2627 new_val |= bitfield_swap(value, 1, 1); 2628 new_val |= bitfield_swap(value, 2, 1); 2629 new_val |= bitfield_swap(value, 3, 1); 2630 new_val |= bitfield_swap(value, 4, 16); 2631 new_val |= bitfield_swap(value, 20, 4); 2632 new_val |= bitfield_swap(value, 24, 2); 2633 new_val |= bitfield_swap(value, 26, 4); 2634 new_val |= bitfield_swap(value, 30, 3); 2635 new_val |= bitfield_swap(value, 33, 31); 2636 } else { 2637 new_val = bitfield_swap(value, 63, 1); 2638 new_val |= bitfield_swap(value, 62, 1); 2639 new_val |= bitfield_swap(value, 61, 1); 2640 new_val |= bitfield_swap(value, 60, 1); 2641 new_val |= bitfield_swap(value, 44, 16); 2642 new_val |= bitfield_swap(value, 40, 4); 2643 new_val |= bitfield_swap(value, 38, 2); 2644 new_val |= bitfield_swap(value, 34, 4); 2645 new_val |= bitfield_swap(value, 31, 3); 2646 new_val |= bitfield_swap(value, 0, 31); 2647 } 2648 2649 return new_val; 2650 } 2651 2652 static inline bool evsel__has_branch_counters(const struct evsel *evsel) 2653 { 2654 struct evsel *leader = evsel__leader(evsel); 2655 2656 /* The branch counters feature only supports group */ 2657 if (!leader || !evsel->evlist) 2658 return false; 2659 2660 if (evsel->evlist->nr_br_cntr < 0) 2661 evlist__update_br_cntr(evsel->evlist); 2662 2663 if (leader->br_cntr_nr > 0) 2664 return true; 2665 2666 return false; 2667 } 2668 2669 int evsel__parse_sample(struct evsel *evsel, union perf_event *event, 2670 struct perf_sample *data) 2671 { 2672 u64 type = evsel->core.attr.sample_type; 2673 bool swapped = evsel->needs_swap; 2674 const __u64 *array; 2675 u16 max_size = event->header.size; 2676 const void *endp = (void *)event + max_size; 2677 u64 sz; 2678 2679 /* 2680 * used for cross-endian analysis. See git commit 65014ab3 2681 * for why this goofiness is needed. 2682 */ 2683 union u64_swap u; 2684 2685 memset(data, 0, sizeof(*data)); 2686 data->cpu = data->pid = data->tid = -1; 2687 data->stream_id = data->id = data->time = -1ULL; 2688 data->period = evsel->core.attr.sample_period; 2689 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 2690 data->misc = event->header.misc; 2691 data->data_src = PERF_MEM_DATA_SRC_NONE; 2692 data->vcpu = -1; 2693 2694 if (event->header.type != PERF_RECORD_SAMPLE) { 2695 if (!evsel->core.attr.sample_id_all) 2696 return 0; 2697 return perf_evsel__parse_id_sample(evsel, event, data); 2698 } 2699 2700 array = event->sample.array; 2701 2702 if (perf_event__check_size(event, evsel->sample_size)) 2703 return -EFAULT; 2704 2705 if (type & PERF_SAMPLE_IDENTIFIER) { 2706 data->id = *array; 2707 array++; 2708 } 2709 2710 if (type & PERF_SAMPLE_IP) { 2711 data->ip = *array; 2712 array++; 2713 } 2714 2715 if (type & PERF_SAMPLE_TID) { 2716 u.val64 = *array; 2717 if (swapped) { 2718 /* undo swap of u64, then swap on individual u32s */ 2719 u.val64 = bswap_64(u.val64); 2720 u.val32[0] = bswap_32(u.val32[0]); 2721 u.val32[1] = bswap_32(u.val32[1]); 2722 } 2723 2724 data->pid = u.val32[0]; 2725 data->tid = u.val32[1]; 2726 array++; 2727 } 2728 2729 if (type & PERF_SAMPLE_TIME) { 2730 data->time = *array; 2731 array++; 2732 } 2733 2734 if (type & PERF_SAMPLE_ADDR) { 2735 data->addr = *array; 2736 array++; 2737 } 2738 2739 if (type & PERF_SAMPLE_ID) { 2740 data->id = *array; 2741 array++; 2742 } 2743 2744 if (type & PERF_SAMPLE_STREAM_ID) { 2745 data->stream_id = *array; 2746 array++; 2747 } 2748 2749 if (type & PERF_SAMPLE_CPU) { 2750 2751 u.val64 = *array; 2752 if (swapped) { 2753 /* undo swap of u64, then swap on individual u32s */ 2754 u.val64 = bswap_64(u.val64); 2755 u.val32[0] = bswap_32(u.val32[0]); 2756 } 2757 2758 data->cpu = u.val32[0]; 2759 array++; 2760 } 2761 2762 if (type & PERF_SAMPLE_PERIOD) { 2763 data->period = *array; 2764 array++; 2765 } 2766 2767 if (type & PERF_SAMPLE_READ) { 2768 u64 read_format = evsel->core.attr.read_format; 2769 2770 OVERFLOW_CHECK_u64(array); 2771 if (read_format & PERF_FORMAT_GROUP) 2772 data->read.group.nr = *array; 2773 else 2774 data->read.one.value = *array; 2775 2776 array++; 2777 2778 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 2779 OVERFLOW_CHECK_u64(array); 2780 data->read.time_enabled = *array; 2781 array++; 2782 } 2783 2784 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 2785 OVERFLOW_CHECK_u64(array); 2786 data->read.time_running = *array; 2787 array++; 2788 } 2789 2790 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 2791 if (read_format & PERF_FORMAT_GROUP) { 2792 const u64 max_group_nr = UINT64_MAX / 2793 sizeof(struct sample_read_value); 2794 2795 if (data->read.group.nr > max_group_nr) 2796 return -EFAULT; 2797 2798 sz = data->read.group.nr * sample_read_value_size(read_format); 2799 OVERFLOW_CHECK(array, sz, max_size); 2800 data->read.group.values = 2801 (struct sample_read_value *)array; 2802 array = (void *)array + sz; 2803 } else { 2804 OVERFLOW_CHECK_u64(array); 2805 data->read.one.id = *array; 2806 array++; 2807 2808 if (read_format & PERF_FORMAT_LOST) { 2809 OVERFLOW_CHECK_u64(array); 2810 data->read.one.lost = *array; 2811 array++; 2812 } 2813 } 2814 } 2815 2816 if (type & PERF_SAMPLE_CALLCHAIN) { 2817 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 2818 2819 OVERFLOW_CHECK_u64(array); 2820 data->callchain = (struct ip_callchain *)array++; 2821 if (data->callchain->nr > max_callchain_nr) 2822 return -EFAULT; 2823 sz = data->callchain->nr * sizeof(u64); 2824 OVERFLOW_CHECK(array, sz, max_size); 2825 array = (void *)array + sz; 2826 } 2827 2828 if (type & PERF_SAMPLE_RAW) { 2829 OVERFLOW_CHECK_u64(array); 2830 u.val64 = *array; 2831 2832 /* 2833 * Undo swap of u64, then swap on individual u32s, 2834 * get the size of the raw area and undo all of the 2835 * swap. The pevent interface handles endianness by 2836 * itself. 2837 */ 2838 if (swapped) { 2839 u.val64 = bswap_64(u.val64); 2840 u.val32[0] = bswap_32(u.val32[0]); 2841 u.val32[1] = bswap_32(u.val32[1]); 2842 } 2843 data->raw_size = u.val32[0]; 2844 2845 /* 2846 * The raw data is aligned on 64bits including the 2847 * u32 size, so it's safe to use mem_bswap_64. 2848 */ 2849 if (swapped) 2850 mem_bswap_64((void *) array, data->raw_size); 2851 2852 array = (void *)array + sizeof(u32); 2853 2854 OVERFLOW_CHECK(array, data->raw_size, max_size); 2855 data->raw_data = (void *)array; 2856 array = (void *)array + data->raw_size; 2857 } 2858 2859 if (type & PERF_SAMPLE_BRANCH_STACK) { 2860 const u64 max_branch_nr = UINT64_MAX / 2861 sizeof(struct branch_entry); 2862 struct branch_entry *e; 2863 unsigned int i; 2864 2865 OVERFLOW_CHECK_u64(array); 2866 data->branch_stack = (struct branch_stack *)array++; 2867 2868 if (data->branch_stack->nr > max_branch_nr) 2869 return -EFAULT; 2870 2871 sz = data->branch_stack->nr * sizeof(struct branch_entry); 2872 if (evsel__has_branch_hw_idx(evsel)) { 2873 sz += sizeof(u64); 2874 e = &data->branch_stack->entries[0]; 2875 } else { 2876 data->no_hw_idx = true; 2877 /* 2878 * if the PERF_SAMPLE_BRANCH_HW_INDEX is not applied, 2879 * only nr and entries[] will be output by kernel. 2880 */ 2881 e = (struct branch_entry *)&data->branch_stack->hw_idx; 2882 } 2883 2884 if (swapped) { 2885 /* 2886 * struct branch_flag does not have endian 2887 * specific bit field definition. And bswap 2888 * will not resolve the issue, since these 2889 * are bit fields. 2890 * 2891 * evsel__bitfield_swap_branch_flags() uses a 2892 * bitfield_swap macro to swap the bit position 2893 * based on the host endians. 2894 */ 2895 for (i = 0; i < data->branch_stack->nr; i++, e++) 2896 e->flags.value = evsel__bitfield_swap_branch_flags(e->flags.value); 2897 } 2898 2899 OVERFLOW_CHECK(array, sz, max_size); 2900 array = (void *)array + sz; 2901 2902 if (evsel__has_branch_counters(evsel)) { 2903 data->branch_stack_cntr = (u64 *)array; 2904 sz = data->branch_stack->nr * sizeof(u64); 2905 2906 OVERFLOW_CHECK(array, sz, max_size); 2907 array = (void *)array + sz; 2908 } 2909 } 2910 2911 if (type & PERF_SAMPLE_REGS_USER) { 2912 OVERFLOW_CHECK_u64(array); 2913 data->user_regs.abi = *array; 2914 array++; 2915 2916 if (data->user_regs.abi) { 2917 u64 mask = evsel->core.attr.sample_regs_user; 2918 2919 sz = hweight64(mask) * sizeof(u64); 2920 OVERFLOW_CHECK(array, sz, max_size); 2921 data->user_regs.mask = mask; 2922 data->user_regs.regs = (u64 *)array; 2923 array = (void *)array + sz; 2924 } 2925 } 2926 2927 if (type & PERF_SAMPLE_STACK_USER) { 2928 OVERFLOW_CHECK_u64(array); 2929 sz = *array++; 2930 2931 data->user_stack.offset = ((char *)(array - 1) 2932 - (char *) event); 2933 2934 if (!sz) { 2935 data->user_stack.size = 0; 2936 } else { 2937 OVERFLOW_CHECK(array, sz, max_size); 2938 data->user_stack.data = (char *)array; 2939 array = (void *)array + sz; 2940 OVERFLOW_CHECK_u64(array); 2941 data->user_stack.size = *array++; 2942 if (WARN_ONCE(data->user_stack.size > sz, 2943 "user stack dump failure\n")) 2944 return -EFAULT; 2945 } 2946 } 2947 2948 if (type & PERF_SAMPLE_WEIGHT_TYPE) { 2949 OVERFLOW_CHECK_u64(array); 2950 arch_perf_parse_sample_weight(data, array, type); 2951 array++; 2952 } 2953 2954 if (type & PERF_SAMPLE_DATA_SRC) { 2955 OVERFLOW_CHECK_u64(array); 2956 data->data_src = *array; 2957 array++; 2958 } 2959 2960 if (type & PERF_SAMPLE_TRANSACTION) { 2961 OVERFLOW_CHECK_u64(array); 2962 data->transaction = *array; 2963 array++; 2964 } 2965 2966 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE; 2967 if (type & PERF_SAMPLE_REGS_INTR) { 2968 OVERFLOW_CHECK_u64(array); 2969 data->intr_regs.abi = *array; 2970 array++; 2971 2972 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) { 2973 u64 mask = evsel->core.attr.sample_regs_intr; 2974 2975 sz = hweight64(mask) * sizeof(u64); 2976 OVERFLOW_CHECK(array, sz, max_size); 2977 data->intr_regs.mask = mask; 2978 data->intr_regs.regs = (u64 *)array; 2979 array = (void *)array + sz; 2980 } 2981 } 2982 2983 data->phys_addr = 0; 2984 if (type & PERF_SAMPLE_PHYS_ADDR) { 2985 data->phys_addr = *array; 2986 array++; 2987 } 2988 2989 data->cgroup = 0; 2990 if (type & PERF_SAMPLE_CGROUP) { 2991 data->cgroup = *array; 2992 array++; 2993 } 2994 2995 data->data_page_size = 0; 2996 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) { 2997 data->data_page_size = *array; 2998 array++; 2999 } 3000 3001 data->code_page_size = 0; 3002 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) { 3003 data->code_page_size = *array; 3004 array++; 3005 } 3006 3007 if (type & PERF_SAMPLE_AUX) { 3008 OVERFLOW_CHECK_u64(array); 3009 sz = *array++; 3010 3011 OVERFLOW_CHECK(array, sz, max_size); 3012 /* Undo swap of data */ 3013 if (swapped) 3014 mem_bswap_64((char *)array, sz); 3015 data->aux_sample.size = sz; 3016 data->aux_sample.data = (char *)array; 3017 array = (void *)array + sz; 3018 } 3019 3020 return 0; 3021 } 3022 3023 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event, 3024 u64 *timestamp) 3025 { 3026 u64 type = evsel->core.attr.sample_type; 3027 const __u64 *array; 3028 3029 if (!(type & PERF_SAMPLE_TIME)) 3030 return -1; 3031 3032 if (event->header.type != PERF_RECORD_SAMPLE) { 3033 struct perf_sample data = { 3034 .time = -1ULL, 3035 }; 3036 3037 if (!evsel->core.attr.sample_id_all) 3038 return -1; 3039 if (perf_evsel__parse_id_sample(evsel, event, &data)) 3040 return -1; 3041 3042 *timestamp = data.time; 3043 return 0; 3044 } 3045 3046 array = event->sample.array; 3047 3048 if (perf_event__check_size(event, evsel->sample_size)) 3049 return -EFAULT; 3050 3051 if (type & PERF_SAMPLE_IDENTIFIER) 3052 array++; 3053 3054 if (type & PERF_SAMPLE_IP) 3055 array++; 3056 3057 if (type & PERF_SAMPLE_TID) 3058 array++; 3059 3060 if (type & PERF_SAMPLE_TIME) 3061 *timestamp = *array; 3062 3063 return 0; 3064 } 3065 3066 u16 evsel__id_hdr_size(const struct evsel *evsel) 3067 { 3068 u64 sample_type = evsel->core.attr.sample_type; 3069 u16 size = 0; 3070 3071 if (sample_type & PERF_SAMPLE_TID) 3072 size += sizeof(u64); 3073 3074 if (sample_type & PERF_SAMPLE_TIME) 3075 size += sizeof(u64); 3076 3077 if (sample_type & PERF_SAMPLE_ID) 3078 size += sizeof(u64); 3079 3080 if (sample_type & PERF_SAMPLE_STREAM_ID) 3081 size += sizeof(u64); 3082 3083 if (sample_type & PERF_SAMPLE_CPU) 3084 size += sizeof(u64); 3085 3086 if (sample_type & PERF_SAMPLE_IDENTIFIER) 3087 size += sizeof(u64); 3088 3089 return size; 3090 } 3091 3092 #ifdef HAVE_LIBTRACEEVENT 3093 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name) 3094 { 3095 return tep_find_field(evsel->tp_format, name); 3096 } 3097 3098 struct tep_format_field *evsel__common_field(struct evsel *evsel, const char *name) 3099 { 3100 return tep_find_common_field(evsel->tp_format, name); 3101 } 3102 3103 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name) 3104 { 3105 struct tep_format_field *field = evsel__field(evsel, name); 3106 int offset; 3107 3108 if (!field) 3109 return NULL; 3110 3111 offset = field->offset; 3112 3113 if (field->flags & TEP_FIELD_IS_DYNAMIC) { 3114 offset = *(int *)(sample->raw_data + field->offset); 3115 offset &= 0xffff; 3116 if (tep_field_is_relative(field->flags)) 3117 offset += field->offset + field->size; 3118 } 3119 3120 return sample->raw_data + offset; 3121 } 3122 3123 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample, 3124 bool needs_swap) 3125 { 3126 u64 value; 3127 void *ptr = sample->raw_data + field->offset; 3128 3129 switch (field->size) { 3130 case 1: 3131 return *(u8 *)ptr; 3132 case 2: 3133 value = *(u16 *)ptr; 3134 break; 3135 case 4: 3136 value = *(u32 *)ptr; 3137 break; 3138 case 8: 3139 memcpy(&value, ptr, sizeof(u64)); 3140 break; 3141 default: 3142 return 0; 3143 } 3144 3145 if (!needs_swap) 3146 return value; 3147 3148 switch (field->size) { 3149 case 2: 3150 return bswap_16(value); 3151 case 4: 3152 return bswap_32(value); 3153 case 8: 3154 return bswap_64(value); 3155 default: 3156 return 0; 3157 } 3158 3159 return 0; 3160 } 3161 3162 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name) 3163 { 3164 struct tep_format_field *field = evsel__field(evsel, name); 3165 3166 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0; 3167 } 3168 3169 u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const char *name) 3170 { 3171 struct tep_format_field *field = evsel__common_field(evsel, name); 3172 3173 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0; 3174 } 3175 3176 char evsel__taskstate(struct evsel *evsel, struct perf_sample *sample, const char *name) 3177 { 3178 static struct tep_format_field *prev_state_field; 3179 static const char *states; 3180 struct tep_format_field *field; 3181 unsigned long long val; 3182 unsigned int bit; 3183 char state = '?'; /* '?' denotes unknown task state */ 3184 3185 field = evsel__field(evsel, name); 3186 3187 if (!field) 3188 return state; 3189 3190 if (!states || field != prev_state_field) { 3191 states = parse_task_states(field); 3192 if (!states) 3193 return state; 3194 prev_state_field = field; 3195 } 3196 3197 /* 3198 * Note since the kernel exposes TASK_REPORT_MAX to userspace 3199 * to denote the 'preempted' state, we might as welll report 3200 * 'R' for this case, which make senses to users as well. 3201 * 3202 * We can change this if we have a good reason in the future. 3203 */ 3204 val = evsel__intval(evsel, sample, name); 3205 bit = val ? ffs(val) : 0; 3206 state = (!bit || bit > strlen(states)) ? 'R' : states[bit-1]; 3207 return state; 3208 } 3209 #endif 3210 3211 bool evsel__fallback(struct evsel *evsel, struct target *target, int err, 3212 char *msg, size_t msgsize) 3213 { 3214 int paranoid; 3215 3216 if ((err == ENOENT || err == ENXIO || err == ENODEV) && 3217 evsel->core.attr.type == PERF_TYPE_HARDWARE && 3218 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) { 3219 /* 3220 * If it's cycles then fall back to hrtimer based cpu-clock sw 3221 * counter, which is always available even if no PMU support. 3222 * 3223 * PPC returns ENXIO until 2.6.37 (behavior changed with commit 3224 * b0a873e). 3225 */ 3226 evsel->core.attr.type = PERF_TYPE_SOFTWARE; 3227 evsel->core.attr.config = target__has_cpu(target) 3228 ? PERF_COUNT_SW_CPU_CLOCK 3229 : PERF_COUNT_SW_TASK_CLOCK; 3230 scnprintf(msg, msgsize, 3231 "The cycles event is not supported, trying to fall back to %s", 3232 target__has_cpu(target) ? "cpu-clock" : "task-clock"); 3233 3234 zfree(&evsel->name); 3235 return true; 3236 } else if (err == EACCES && !evsel->core.attr.exclude_kernel && 3237 (paranoid = perf_event_paranoid()) > 1) { 3238 const char *name = evsel__name(evsel); 3239 char *new_name; 3240 const char *sep = ":"; 3241 3242 /* If event has exclude user then don't exclude kernel. */ 3243 if (evsel->core.attr.exclude_user) 3244 return false; 3245 3246 /* Is there already the separator in the name. */ 3247 if (strchr(name, '/') || 3248 (strchr(name, ':') && !evsel->is_libpfm_event)) 3249 sep = ""; 3250 3251 if (asprintf(&new_name, "%s%su", name, sep) < 0) 3252 return false; 3253 3254 free(evsel->name); 3255 evsel->name = new_name; 3256 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying " 3257 "to fall back to excluding kernel and hypervisor " 3258 " samples", paranoid); 3259 evsel->core.attr.exclude_kernel = 1; 3260 evsel->core.attr.exclude_hv = 1; 3261 3262 return true; 3263 } 3264 3265 return false; 3266 } 3267 3268 static bool find_process(const char *name) 3269 { 3270 size_t len = strlen(name); 3271 DIR *dir; 3272 struct dirent *d; 3273 int ret = -1; 3274 3275 dir = opendir(procfs__mountpoint()); 3276 if (!dir) 3277 return false; 3278 3279 /* Walk through the directory. */ 3280 while (ret && (d = readdir(dir)) != NULL) { 3281 char path[PATH_MAX]; 3282 char *data; 3283 size_t size; 3284 3285 if ((d->d_type != DT_DIR) || 3286 !strcmp(".", d->d_name) || 3287 !strcmp("..", d->d_name)) 3288 continue; 3289 3290 scnprintf(path, sizeof(path), "%s/%s/comm", 3291 procfs__mountpoint(), d->d_name); 3292 3293 if (filename__read_str(path, &data, &size)) 3294 continue; 3295 3296 ret = strncmp(name, data, len); 3297 free(data); 3298 } 3299 3300 closedir(dir); 3301 return ret ? false : true; 3302 } 3303 3304 int __weak arch_evsel__open_strerror(struct evsel *evsel __maybe_unused, 3305 char *msg __maybe_unused, 3306 size_t size __maybe_unused) 3307 { 3308 return 0; 3309 } 3310 3311 int evsel__open_strerror(struct evsel *evsel, struct target *target, 3312 int err, char *msg, size_t size) 3313 { 3314 char sbuf[STRERR_BUFSIZE]; 3315 int printed = 0, enforced = 0; 3316 int ret; 3317 3318 switch (err) { 3319 case EPERM: 3320 case EACCES: 3321 printed += scnprintf(msg + printed, size - printed, 3322 "Access to performance monitoring and observability operations is limited.\n"); 3323 3324 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) { 3325 if (enforced) { 3326 printed += scnprintf(msg + printed, size - printed, 3327 "Enforced MAC policy settings (SELinux) can limit access to performance\n" 3328 "monitoring and observability operations. Inspect system audit records for\n" 3329 "more perf_event access control information and adjusting the policy.\n"); 3330 } 3331 } 3332 3333 if (err == EPERM) 3334 printed += scnprintf(msg, size, 3335 "No permission to enable %s event.\n\n", evsel__name(evsel)); 3336 3337 return scnprintf(msg + printed, size - printed, 3338 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n" 3339 "access to performance monitoring and observability operations for processes\n" 3340 "without CAP_PERFMON, CAP_SYS_PTRACE or CAP_SYS_ADMIN Linux capability.\n" 3341 "More information can be found at 'Perf events and tool security' document:\n" 3342 "https://www.kernel.org/doc/html/latest/admin-guide/perf-security.html\n" 3343 "perf_event_paranoid setting is %d:\n" 3344 " -1: Allow use of (almost) all events by all users\n" 3345 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n" 3346 ">= 0: Disallow raw and ftrace function tracepoint access\n" 3347 ">= 1: Disallow CPU event access\n" 3348 ">= 2: Disallow kernel profiling\n" 3349 "To make the adjusted perf_event_paranoid setting permanent preserve it\n" 3350 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)", 3351 perf_event_paranoid()); 3352 case ENOENT: 3353 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel)); 3354 case EMFILE: 3355 return scnprintf(msg, size, "%s", 3356 "Too many events are opened.\n" 3357 "Probably the maximum number of open file descriptors has been reached.\n" 3358 "Hint: Try again after reducing the number of events.\n" 3359 "Hint: Try increasing the limit with 'ulimit -n <limit>'"); 3360 case ENOMEM: 3361 if (evsel__has_callchain(evsel) && 3362 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0) 3363 return scnprintf(msg, size, 3364 "Not enough memory to setup event with callchain.\n" 3365 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n" 3366 "Hint: Current value: %d", sysctl__max_stack()); 3367 break; 3368 case ENODEV: 3369 if (target->cpu_list) 3370 return scnprintf(msg, size, "%s", 3371 "No such device - did you specify an out-of-range profile CPU?"); 3372 break; 3373 case EOPNOTSUPP: 3374 if (evsel->core.attr.sample_type & PERF_SAMPLE_BRANCH_STACK) 3375 return scnprintf(msg, size, 3376 "%s: PMU Hardware or event type doesn't support branch stack sampling.", 3377 evsel__name(evsel)); 3378 if (evsel->core.attr.aux_output) 3379 return scnprintf(msg, size, 3380 "%s: PMU Hardware doesn't support 'aux_output' feature", 3381 evsel__name(evsel)); 3382 if (evsel->core.attr.sample_period != 0) 3383 return scnprintf(msg, size, 3384 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'", 3385 evsel__name(evsel)); 3386 if (evsel->core.attr.precise_ip) 3387 return scnprintf(msg, size, "%s", 3388 "\'precise\' request may not be supported. Try removing 'p' modifier."); 3389 #if defined(__i386__) || defined(__x86_64__) 3390 if (evsel->core.attr.type == PERF_TYPE_HARDWARE) 3391 return scnprintf(msg, size, "%s", 3392 "No hardware sampling interrupt available.\n"); 3393 #endif 3394 break; 3395 case EBUSY: 3396 if (find_process("oprofiled")) 3397 return scnprintf(msg, size, 3398 "The PMU counters are busy/taken by another profiler.\n" 3399 "We found oprofile daemon running, please stop it and try again."); 3400 break; 3401 case EINVAL: 3402 if (evsel->core.attr.sample_type & PERF_SAMPLE_CODE_PAGE_SIZE && perf_missing_features.code_page_size) 3403 return scnprintf(msg, size, "Asking for the code page size isn't supported by this kernel."); 3404 if (evsel->core.attr.sample_type & PERF_SAMPLE_DATA_PAGE_SIZE && perf_missing_features.data_page_size) 3405 return scnprintf(msg, size, "Asking for the data page size isn't supported by this kernel."); 3406 if (evsel->core.attr.write_backward && perf_missing_features.write_backward) 3407 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel."); 3408 if (perf_missing_features.clockid) 3409 return scnprintf(msg, size, "clockid feature not supported."); 3410 if (perf_missing_features.clockid_wrong) 3411 return scnprintf(msg, size, "wrong clockid (%d).", clockid); 3412 if (perf_missing_features.aux_output) 3413 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel."); 3414 if (!target__has_cpu(target)) 3415 return scnprintf(msg, size, 3416 "Invalid event (%s) in per-thread mode, enable system wide with '-a'.", 3417 evsel__name(evsel)); 3418 3419 break; 3420 case ENODATA: 3421 return scnprintf(msg, size, "Cannot collect data source with the load latency event alone. " 3422 "Please add an auxiliary event in front of the load latency event."); 3423 default: 3424 break; 3425 } 3426 3427 ret = arch_evsel__open_strerror(evsel, msg, size); 3428 if (ret) 3429 return ret; 3430 3431 return scnprintf(msg, size, 3432 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n" 3433 "\"dmesg | grep -i perf\" may provide additional information.\n", 3434 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel)); 3435 } 3436 3437 struct perf_env *evsel__env(struct evsel *evsel) 3438 { 3439 if (evsel && evsel->evlist && evsel->evlist->env) 3440 return evsel->evlist->env; 3441 return &perf_env; 3442 } 3443 3444 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist) 3445 { 3446 int cpu_map_idx, thread; 3447 3448 if (evsel__is_retire_lat(evsel)) 3449 return 0; 3450 3451 for (cpu_map_idx = 0; cpu_map_idx < xyarray__max_x(evsel->core.fd); cpu_map_idx++) { 3452 for (thread = 0; thread < xyarray__max_y(evsel->core.fd); 3453 thread++) { 3454 int fd = FD(evsel, cpu_map_idx, thread); 3455 3456 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, 3457 cpu_map_idx, thread, fd) < 0) 3458 return -1; 3459 } 3460 } 3461 3462 return 0; 3463 } 3464 3465 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist) 3466 { 3467 struct perf_cpu_map *cpus = evsel->core.cpus; 3468 struct perf_thread_map *threads = evsel->core.threads; 3469 3470 if (perf_evsel__alloc_id(&evsel->core, perf_cpu_map__nr(cpus), threads->nr)) 3471 return -ENOMEM; 3472 3473 return store_evsel_ids(evsel, evlist); 3474 } 3475 3476 void evsel__zero_per_pkg(struct evsel *evsel) 3477 { 3478 struct hashmap_entry *cur; 3479 size_t bkt; 3480 3481 if (evsel->per_pkg_mask) { 3482 hashmap__for_each_entry(evsel->per_pkg_mask, cur, bkt) 3483 zfree(&cur->pkey); 3484 3485 hashmap__clear(evsel->per_pkg_mask); 3486 } 3487 } 3488 3489 /** 3490 * evsel__is_hybrid - does the evsel have a known PMU that is hybrid. Note, this 3491 * will be false on hybrid systems for hardware and legacy 3492 * cache events. 3493 */ 3494 bool evsel__is_hybrid(const struct evsel *evsel) 3495 { 3496 if (perf_pmus__num_core_pmus() == 1) 3497 return false; 3498 3499 return evsel->core.is_pmu_core; 3500 } 3501 3502 struct evsel *evsel__leader(const struct evsel *evsel) 3503 { 3504 return container_of(evsel->core.leader, struct evsel, core); 3505 } 3506 3507 bool evsel__has_leader(struct evsel *evsel, struct evsel *leader) 3508 { 3509 return evsel->core.leader == &leader->core; 3510 } 3511 3512 bool evsel__is_leader(struct evsel *evsel) 3513 { 3514 return evsel__has_leader(evsel, evsel); 3515 } 3516 3517 void evsel__set_leader(struct evsel *evsel, struct evsel *leader) 3518 { 3519 evsel->core.leader = &leader->core; 3520 } 3521 3522 int evsel__source_count(const struct evsel *evsel) 3523 { 3524 struct evsel *pos; 3525 int count = 0; 3526 3527 evlist__for_each_entry(evsel->evlist, pos) { 3528 if (pos->metric_leader == evsel) 3529 count++; 3530 } 3531 return count; 3532 } 3533 3534 bool __weak arch_evsel__must_be_in_group(const struct evsel *evsel __maybe_unused) 3535 { 3536 return false; 3537 } 3538 3539 /* 3540 * Remove an event from a given group (leader). 3541 * Some events, e.g., perf metrics Topdown events, 3542 * must always be grouped. Ignore the events. 3543 */ 3544 void evsel__remove_from_group(struct evsel *evsel, struct evsel *leader) 3545 { 3546 if (!arch_evsel__must_be_in_group(evsel) && evsel != leader) { 3547 evsel__set_leader(evsel, evsel); 3548 evsel->core.nr_members = 0; 3549 leader->core.nr_members--; 3550 } 3551 } 3552