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/fs/fs.h> 14 #include <api/fs/tracing_path.h> 15 #include <traceevent/event-parse.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 "callchain.h" 29 #include "cgroup.h" 30 #include "counts.h" 31 #include "event.h" 32 #include "evsel.h" 33 #include "util/env.h" 34 #include "util/evsel_config.h" 35 #include "util/evsel_fprintf.h" 36 #include "evlist.h" 37 #include <perf/cpumap.h> 38 #include "thread_map.h" 39 #include "target.h" 40 #include "perf_regs.h" 41 #include "record.h" 42 #include "debug.h" 43 #include "trace-event.h" 44 #include "stat.h" 45 #include "string2.h" 46 #include "memswap.h" 47 #include "util.h" 48 #include "../perf-sys.h" 49 #include "util/parse-branch-options.h" 50 #include <internal/xyarray.h> 51 #include <internal/lib.h> 52 53 #include <linux/ctype.h> 54 55 struct perf_missing_features perf_missing_features; 56 57 static clockid_t clockid; 58 59 static int evsel__no_extra_init(struct evsel *evsel __maybe_unused) 60 { 61 return 0; 62 } 63 64 void __weak test_attr__ready(void) { } 65 66 static void evsel__no_extra_fini(struct evsel *evsel __maybe_unused) 67 { 68 } 69 70 static struct { 71 size_t size; 72 int (*init)(struct evsel *evsel); 73 void (*fini)(struct evsel *evsel); 74 } perf_evsel__object = { 75 .size = sizeof(struct evsel), 76 .init = evsel__no_extra_init, 77 .fini = evsel__no_extra_fini, 78 }; 79 80 int evsel__object_config(size_t object_size, int (*init)(struct evsel *evsel), 81 void (*fini)(struct evsel *evsel)) 82 { 83 84 if (object_size == 0) 85 goto set_methods; 86 87 if (perf_evsel__object.size > object_size) 88 return -EINVAL; 89 90 perf_evsel__object.size = object_size; 91 92 set_methods: 93 if (init != NULL) 94 perf_evsel__object.init = init; 95 96 if (fini != NULL) 97 perf_evsel__object.fini = fini; 98 99 return 0; 100 } 101 102 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 103 104 int __evsel__sample_size(u64 sample_type) 105 { 106 u64 mask = sample_type & PERF_SAMPLE_MASK; 107 int size = 0; 108 int i; 109 110 for (i = 0; i < 64; i++) { 111 if (mask & (1ULL << i)) 112 size++; 113 } 114 115 size *= sizeof(u64); 116 117 return size; 118 } 119 120 /** 121 * __perf_evsel__calc_id_pos - calculate id_pos. 122 * @sample_type: sample type 123 * 124 * This function returns the position of the event id (PERF_SAMPLE_ID or 125 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct 126 * perf_record_sample. 127 */ 128 static int __perf_evsel__calc_id_pos(u64 sample_type) 129 { 130 int idx = 0; 131 132 if (sample_type & PERF_SAMPLE_IDENTIFIER) 133 return 0; 134 135 if (!(sample_type & PERF_SAMPLE_ID)) 136 return -1; 137 138 if (sample_type & PERF_SAMPLE_IP) 139 idx += 1; 140 141 if (sample_type & PERF_SAMPLE_TID) 142 idx += 1; 143 144 if (sample_type & PERF_SAMPLE_TIME) 145 idx += 1; 146 147 if (sample_type & PERF_SAMPLE_ADDR) 148 idx += 1; 149 150 return idx; 151 } 152 153 /** 154 * __perf_evsel__calc_is_pos - calculate is_pos. 155 * @sample_type: sample type 156 * 157 * This function returns the position (counting backwards) of the event id 158 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if 159 * sample_id_all is used there is an id sample appended to non-sample events. 160 */ 161 static int __perf_evsel__calc_is_pos(u64 sample_type) 162 { 163 int idx = 1; 164 165 if (sample_type & PERF_SAMPLE_IDENTIFIER) 166 return 1; 167 168 if (!(sample_type & PERF_SAMPLE_ID)) 169 return -1; 170 171 if (sample_type & PERF_SAMPLE_CPU) 172 idx += 1; 173 174 if (sample_type & PERF_SAMPLE_STREAM_ID) 175 idx += 1; 176 177 return idx; 178 } 179 180 void evsel__calc_id_pos(struct evsel *evsel) 181 { 182 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type); 183 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type); 184 } 185 186 void __evsel__set_sample_bit(struct evsel *evsel, 187 enum perf_event_sample_format bit) 188 { 189 if (!(evsel->core.attr.sample_type & bit)) { 190 evsel->core.attr.sample_type |= bit; 191 evsel->sample_size += sizeof(u64); 192 evsel__calc_id_pos(evsel); 193 } 194 } 195 196 void __evsel__reset_sample_bit(struct evsel *evsel, 197 enum perf_event_sample_format bit) 198 { 199 if (evsel->core.attr.sample_type & bit) { 200 evsel->core.attr.sample_type &= ~bit; 201 evsel->sample_size -= sizeof(u64); 202 evsel__calc_id_pos(evsel); 203 } 204 } 205 206 void evsel__set_sample_id(struct evsel *evsel, 207 bool can_sample_identifier) 208 { 209 if (can_sample_identifier) { 210 evsel__reset_sample_bit(evsel, ID); 211 evsel__set_sample_bit(evsel, IDENTIFIER); 212 } else { 213 evsel__set_sample_bit(evsel, ID); 214 } 215 evsel->core.attr.read_format |= PERF_FORMAT_ID; 216 } 217 218 /** 219 * evsel__is_function_event - Return whether given evsel is a function 220 * trace event 221 * 222 * @evsel - evsel selector to be tested 223 * 224 * Return %true if event is function trace event 225 */ 226 bool evsel__is_function_event(struct evsel *evsel) 227 { 228 #define FUNCTION_EVENT "ftrace:function" 229 230 return evsel->name && 231 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT)); 232 233 #undef FUNCTION_EVENT 234 } 235 236 void evsel__init(struct evsel *evsel, 237 struct perf_event_attr *attr, int idx) 238 { 239 perf_evsel__init(&evsel->core, attr); 240 evsel->idx = idx; 241 evsel->tracking = !idx; 242 evsel->leader = evsel; 243 evsel->unit = ""; 244 evsel->scale = 1.0; 245 evsel->max_events = ULONG_MAX; 246 evsel->evlist = NULL; 247 evsel->bpf_obj = NULL; 248 evsel->bpf_fd = -1; 249 INIT_LIST_HEAD(&evsel->config_terms); 250 perf_evsel__object.init(evsel); 251 evsel->sample_size = __evsel__sample_size(attr->sample_type); 252 evsel__calc_id_pos(evsel); 253 evsel->cmdline_group_boundary = false; 254 evsel->metric_expr = NULL; 255 evsel->metric_name = NULL; 256 evsel->metric_events = NULL; 257 evsel->per_pkg_mask = NULL; 258 evsel->collect_stat = false; 259 evsel->pmu_name = NULL; 260 } 261 262 struct evsel *evsel__new_idx(struct perf_event_attr *attr, int idx) 263 { 264 struct evsel *evsel = zalloc(perf_evsel__object.size); 265 266 if (!evsel) 267 return NULL; 268 evsel__init(evsel, attr, idx); 269 270 if (evsel__is_bpf_output(evsel)) { 271 evsel->core.attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 272 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 273 evsel->core.attr.sample_period = 1; 274 } 275 276 if (evsel__is_clock(evsel)) { 277 /* 278 * The evsel->unit points to static alias->unit 279 * so it's ok to use static string in here. 280 */ 281 static const char *unit = "msec"; 282 283 evsel->unit = unit; 284 evsel->scale = 1e-6; 285 } 286 287 return evsel; 288 } 289 290 static bool perf_event_can_profile_kernel(void) 291 { 292 return perf_event_paranoid_check(1); 293 } 294 295 struct evsel *evsel__new_cycles(bool precise) 296 { 297 struct perf_event_attr attr = { 298 .type = PERF_TYPE_HARDWARE, 299 .config = PERF_COUNT_HW_CPU_CYCLES, 300 .exclude_kernel = !perf_event_can_profile_kernel(), 301 }; 302 struct evsel *evsel; 303 304 event_attr_init(&attr); 305 306 if (!precise) 307 goto new_event; 308 309 /* 310 * Now let the usual logic to set up the perf_event_attr defaults 311 * to kick in when we return and before perf_evsel__open() is called. 312 */ 313 new_event: 314 evsel = evsel__new(&attr); 315 if (evsel == NULL) 316 goto out; 317 318 evsel->precise_max = true; 319 320 /* use asprintf() because free(evsel) assumes name is allocated */ 321 if (asprintf(&evsel->name, "cycles%s%s%.*s", 322 (attr.precise_ip || attr.exclude_kernel) ? ":" : "", 323 attr.exclude_kernel ? "u" : "", 324 attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0) 325 goto error_free; 326 out: 327 return evsel; 328 error_free: 329 evsel__delete(evsel); 330 evsel = NULL; 331 goto out; 332 } 333 334 /* 335 * Returns pointer with encoded error via <linux/err.h> interface. 336 */ 337 struct evsel *evsel__newtp_idx(const char *sys, const char *name, int idx) 338 { 339 struct evsel *evsel = zalloc(perf_evsel__object.size); 340 int err = -ENOMEM; 341 342 if (evsel == NULL) { 343 goto out_err; 344 } else { 345 struct perf_event_attr attr = { 346 .type = PERF_TYPE_TRACEPOINT, 347 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 348 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 349 }; 350 351 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0) 352 goto out_free; 353 354 evsel->tp_format = trace_event__tp_format(sys, name); 355 if (IS_ERR(evsel->tp_format)) { 356 err = PTR_ERR(evsel->tp_format); 357 goto out_free; 358 } 359 360 event_attr_init(&attr); 361 attr.config = evsel->tp_format->id; 362 attr.sample_period = 1; 363 evsel__init(evsel, &attr, idx); 364 } 365 366 return evsel; 367 368 out_free: 369 zfree(&evsel->name); 370 free(evsel); 371 out_err: 372 return ERR_PTR(err); 373 } 374 375 const char *evsel__hw_names[PERF_COUNT_HW_MAX] = { 376 "cycles", 377 "instructions", 378 "cache-references", 379 "cache-misses", 380 "branches", 381 "branch-misses", 382 "bus-cycles", 383 "stalled-cycles-frontend", 384 "stalled-cycles-backend", 385 "ref-cycles", 386 }; 387 388 static const char *__evsel__hw_name(u64 config) 389 { 390 if (config < PERF_COUNT_HW_MAX && evsel__hw_names[config]) 391 return evsel__hw_names[config]; 392 393 return "unknown-hardware"; 394 } 395 396 static int perf_evsel__add_modifiers(struct evsel *evsel, char *bf, size_t size) 397 { 398 int colon = 0, r = 0; 399 struct perf_event_attr *attr = &evsel->core.attr; 400 bool exclude_guest_default = false; 401 402 #define MOD_PRINT(context, mod) do { \ 403 if (!attr->exclude_##context) { \ 404 if (!colon) colon = ++r; \ 405 r += scnprintf(bf + r, size - r, "%c", mod); \ 406 } } while(0) 407 408 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) { 409 MOD_PRINT(kernel, 'k'); 410 MOD_PRINT(user, 'u'); 411 MOD_PRINT(hv, 'h'); 412 exclude_guest_default = true; 413 } 414 415 if (attr->precise_ip) { 416 if (!colon) 417 colon = ++r; 418 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp"); 419 exclude_guest_default = true; 420 } 421 422 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) { 423 MOD_PRINT(host, 'H'); 424 MOD_PRINT(guest, 'G'); 425 } 426 #undef MOD_PRINT 427 if (colon) 428 bf[colon - 1] = ':'; 429 return r; 430 } 431 432 static int evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 433 { 434 int r = scnprintf(bf, size, "%s", __evsel__hw_name(evsel->core.attr.config)); 435 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 436 } 437 438 const char *evsel__sw_names[PERF_COUNT_SW_MAX] = { 439 "cpu-clock", 440 "task-clock", 441 "page-faults", 442 "context-switches", 443 "cpu-migrations", 444 "minor-faults", 445 "major-faults", 446 "alignment-faults", 447 "emulation-faults", 448 "dummy", 449 }; 450 451 static const char *__evsel__sw_name(u64 config) 452 { 453 if (config < PERF_COUNT_SW_MAX && evsel__sw_names[config]) 454 return evsel__sw_names[config]; 455 return "unknown-software"; 456 } 457 458 static int evsel__sw_name(struct evsel *evsel, char *bf, size_t size) 459 { 460 int r = scnprintf(bf, size, "%s", __evsel__sw_name(evsel->core.attr.config)); 461 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 462 } 463 464 static int __evsel__bp_name(char *bf, size_t size, u64 addr, u64 type) 465 { 466 int r; 467 468 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr); 469 470 if (type & HW_BREAKPOINT_R) 471 r += scnprintf(bf + r, size - r, "r"); 472 473 if (type & HW_BREAKPOINT_W) 474 r += scnprintf(bf + r, size - r, "w"); 475 476 if (type & HW_BREAKPOINT_X) 477 r += scnprintf(bf + r, size - r, "x"); 478 479 return r; 480 } 481 482 static int evsel__bp_name(struct evsel *evsel, char *bf, size_t size) 483 { 484 struct perf_event_attr *attr = &evsel->core.attr; 485 int r = __evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type); 486 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 487 } 488 489 const char *evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX][EVSEL__MAX_ALIASES] = { 490 { "L1-dcache", "l1-d", "l1d", "L1-data", }, 491 { "L1-icache", "l1-i", "l1i", "L1-instruction", }, 492 { "LLC", "L2", }, 493 { "dTLB", "d-tlb", "Data-TLB", }, 494 { "iTLB", "i-tlb", "Instruction-TLB", }, 495 { "branch", "branches", "bpu", "btb", "bpc", }, 496 { "node", }, 497 }; 498 499 const char *evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX][EVSEL__MAX_ALIASES] = { 500 { "load", "loads", "read", }, 501 { "store", "stores", "write", }, 502 { "prefetch", "prefetches", "speculative-read", "speculative-load", }, 503 }; 504 505 const char *evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX][EVSEL__MAX_ALIASES] = { 506 { "refs", "Reference", "ops", "access", }, 507 { "misses", "miss", }, 508 }; 509 510 #define C(x) PERF_COUNT_HW_CACHE_##x 511 #define CACHE_READ (1 << C(OP_READ)) 512 #define CACHE_WRITE (1 << C(OP_WRITE)) 513 #define CACHE_PREFETCH (1 << C(OP_PREFETCH)) 514 #define COP(x) (1 << x) 515 516 /* 517 * cache operartion stat 518 * L1I : Read and prefetch only 519 * ITLB and BPU : Read-only 520 */ 521 static unsigned long evsel__hw_cache_stat[C(MAX)] = { 522 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 523 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH), 524 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 525 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 526 [C(ITLB)] = (CACHE_READ), 527 [C(BPU)] = (CACHE_READ), 528 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 529 }; 530 531 bool evsel__is_cache_op_valid(u8 type, u8 op) 532 { 533 if (evsel__hw_cache_stat[type] & COP(op)) 534 return true; /* valid */ 535 else 536 return false; /* invalid */ 537 } 538 539 int __evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, char *bf, size_t size) 540 { 541 if (result) { 542 return scnprintf(bf, size, "%s-%s-%s", evsel__hw_cache[type][0], 543 evsel__hw_cache_op[op][0], 544 evsel__hw_cache_result[result][0]); 545 } 546 547 return scnprintf(bf, size, "%s-%s", evsel__hw_cache[type][0], 548 evsel__hw_cache_op[op][1]); 549 } 550 551 static int __evsel__hw_cache_name(u64 config, char *bf, size_t size) 552 { 553 u8 op, result, type = (config >> 0) & 0xff; 554 const char *err = "unknown-ext-hardware-cache-type"; 555 556 if (type >= PERF_COUNT_HW_CACHE_MAX) 557 goto out_err; 558 559 op = (config >> 8) & 0xff; 560 err = "unknown-ext-hardware-cache-op"; 561 if (op >= PERF_COUNT_HW_CACHE_OP_MAX) 562 goto out_err; 563 564 result = (config >> 16) & 0xff; 565 err = "unknown-ext-hardware-cache-result"; 566 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 567 goto out_err; 568 569 err = "invalid-cache"; 570 if (!evsel__is_cache_op_valid(type, op)) 571 goto out_err; 572 573 return __evsel__hw_cache_type_op_res_name(type, op, result, bf, size); 574 out_err: 575 return scnprintf(bf, size, "%s", err); 576 } 577 578 static int evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size) 579 { 580 int ret = __evsel__hw_cache_name(evsel->core.attr.config, bf, size); 581 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 582 } 583 584 static int evsel__raw_name(struct evsel *evsel, char *bf, size_t size) 585 { 586 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config); 587 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 588 } 589 590 static int evsel__tool_name(char *bf, size_t size) 591 { 592 int ret = scnprintf(bf, size, "duration_time"); 593 return ret; 594 } 595 596 const char *evsel__name(struct evsel *evsel) 597 { 598 char bf[128]; 599 600 if (!evsel) 601 goto out_unknown; 602 603 if (evsel->name) 604 return evsel->name; 605 606 switch (evsel->core.attr.type) { 607 case PERF_TYPE_RAW: 608 evsel__raw_name(evsel, bf, sizeof(bf)); 609 break; 610 611 case PERF_TYPE_HARDWARE: 612 evsel__hw_name(evsel, bf, sizeof(bf)); 613 break; 614 615 case PERF_TYPE_HW_CACHE: 616 evsel__hw_cache_name(evsel, bf, sizeof(bf)); 617 break; 618 619 case PERF_TYPE_SOFTWARE: 620 if (evsel->tool_event) 621 evsel__tool_name(bf, sizeof(bf)); 622 else 623 evsel__sw_name(evsel, bf, sizeof(bf)); 624 break; 625 626 case PERF_TYPE_TRACEPOINT: 627 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint"); 628 break; 629 630 case PERF_TYPE_BREAKPOINT: 631 evsel__bp_name(evsel, bf, sizeof(bf)); 632 break; 633 634 default: 635 scnprintf(bf, sizeof(bf), "unknown attr type: %d", 636 evsel->core.attr.type); 637 break; 638 } 639 640 evsel->name = strdup(bf); 641 642 if (evsel->name) 643 return evsel->name; 644 out_unknown: 645 return "unknown"; 646 } 647 648 const char *evsel__group_name(struct evsel *evsel) 649 { 650 return evsel->group_name ?: "anon group"; 651 } 652 653 /* 654 * Returns the group details for the specified leader, 655 * with following rules. 656 * 657 * For record -e '{cycles,instructions}' 658 * 'anon group { cycles:u, instructions:u }' 659 * 660 * For record -e 'cycles,instructions' and report --group 661 * 'cycles:u, instructions:u' 662 */ 663 int evsel__group_desc(struct evsel *evsel, char *buf, size_t size) 664 { 665 int ret = 0; 666 struct evsel *pos; 667 const char *group_name = evsel__group_name(evsel); 668 669 if (!evsel->forced_leader) 670 ret = scnprintf(buf, size, "%s { ", group_name); 671 672 ret += scnprintf(buf + ret, size - ret, "%s", evsel__name(evsel)); 673 674 for_each_group_member(pos, evsel) 675 ret += scnprintf(buf + ret, size - ret, ", %s", evsel__name(pos)); 676 677 if (!evsel->forced_leader) 678 ret += scnprintf(buf + ret, size - ret, " }"); 679 680 return ret; 681 } 682 683 static void __evsel__config_callchain(struct evsel *evsel, struct record_opts *opts, 684 struct callchain_param *param) 685 { 686 bool function = evsel__is_function_event(evsel); 687 struct perf_event_attr *attr = &evsel->core.attr; 688 689 evsel__set_sample_bit(evsel, CALLCHAIN); 690 691 attr->sample_max_stack = param->max_stack; 692 693 if (opts->kernel_callchains) 694 attr->exclude_callchain_user = 1; 695 if (opts->user_callchains) 696 attr->exclude_callchain_kernel = 1; 697 if (param->record_mode == CALLCHAIN_LBR) { 698 if (!opts->branch_stack) { 699 if (attr->exclude_user) { 700 pr_warning("LBR callstack option is only available " 701 "to get user callchain information. " 702 "Falling back to framepointers.\n"); 703 } else { 704 evsel__set_sample_bit(evsel, BRANCH_STACK); 705 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER | 706 PERF_SAMPLE_BRANCH_CALL_STACK | 707 PERF_SAMPLE_BRANCH_NO_CYCLES | 708 PERF_SAMPLE_BRANCH_NO_FLAGS | 709 PERF_SAMPLE_BRANCH_HW_INDEX; 710 } 711 } else 712 pr_warning("Cannot use LBR callstack with branch stack. " 713 "Falling back to framepointers.\n"); 714 } 715 716 if (param->record_mode == CALLCHAIN_DWARF) { 717 if (!function) { 718 evsel__set_sample_bit(evsel, REGS_USER); 719 evsel__set_sample_bit(evsel, STACK_USER); 720 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) { 721 attr->sample_regs_user |= DWARF_MINIMAL_REGS; 722 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, " 723 "specifying a subset with --user-regs may render DWARF unwinding unreliable, " 724 "so the minimal registers set (IP, SP) is explicitly forced.\n"); 725 } else { 726 attr->sample_regs_user |= PERF_REGS_MASK; 727 } 728 attr->sample_stack_user = param->dump_size; 729 attr->exclude_callchain_user = 1; 730 } else { 731 pr_info("Cannot use DWARF unwind for function trace event," 732 " falling back to framepointers.\n"); 733 } 734 } 735 736 if (function) { 737 pr_info("Disabling user space callchains for function trace event.\n"); 738 attr->exclude_callchain_user = 1; 739 } 740 } 741 742 void evsel__config_callchain(struct evsel *evsel, struct record_opts *opts, 743 struct callchain_param *param) 744 { 745 if (param->enabled) 746 return __evsel__config_callchain(evsel, opts, param); 747 } 748 749 static void 750 perf_evsel__reset_callgraph(struct evsel *evsel, 751 struct callchain_param *param) 752 { 753 struct perf_event_attr *attr = &evsel->core.attr; 754 755 evsel__reset_sample_bit(evsel, CALLCHAIN); 756 if (param->record_mode == CALLCHAIN_LBR) { 757 evsel__reset_sample_bit(evsel, BRANCH_STACK); 758 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER | 759 PERF_SAMPLE_BRANCH_CALL_STACK | 760 PERF_SAMPLE_BRANCH_HW_INDEX); 761 } 762 if (param->record_mode == CALLCHAIN_DWARF) { 763 evsel__reset_sample_bit(evsel, REGS_USER); 764 evsel__reset_sample_bit(evsel, STACK_USER); 765 } 766 } 767 768 static void evsel__apply_config_terms(struct evsel *evsel, 769 struct record_opts *opts, bool track) 770 { 771 struct evsel_config_term *term; 772 struct list_head *config_terms = &evsel->config_terms; 773 struct perf_event_attr *attr = &evsel->core.attr; 774 /* callgraph default */ 775 struct callchain_param param = { 776 .record_mode = callchain_param.record_mode, 777 }; 778 u32 dump_size = 0; 779 int max_stack = 0; 780 const char *callgraph_buf = NULL; 781 782 list_for_each_entry(term, config_terms, list) { 783 switch (term->type) { 784 case EVSEL__CONFIG_TERM_PERIOD: 785 if (!(term->weak && opts->user_interval != ULLONG_MAX)) { 786 attr->sample_period = term->val.period; 787 attr->freq = 0; 788 evsel__reset_sample_bit(evsel, PERIOD); 789 } 790 break; 791 case EVSEL__CONFIG_TERM_FREQ: 792 if (!(term->weak && opts->user_freq != UINT_MAX)) { 793 attr->sample_freq = term->val.freq; 794 attr->freq = 1; 795 evsel__set_sample_bit(evsel, PERIOD); 796 } 797 break; 798 case EVSEL__CONFIG_TERM_TIME: 799 if (term->val.time) 800 evsel__set_sample_bit(evsel, TIME); 801 else 802 evsel__reset_sample_bit(evsel, TIME); 803 break; 804 case EVSEL__CONFIG_TERM_CALLGRAPH: 805 callgraph_buf = term->val.str; 806 break; 807 case EVSEL__CONFIG_TERM_BRANCH: 808 if (term->val.str && strcmp(term->val.str, "no")) { 809 evsel__set_sample_bit(evsel, BRANCH_STACK); 810 parse_branch_str(term->val.str, 811 &attr->branch_sample_type); 812 } else 813 evsel__reset_sample_bit(evsel, BRANCH_STACK); 814 break; 815 case EVSEL__CONFIG_TERM_STACK_USER: 816 dump_size = term->val.stack_user; 817 break; 818 case EVSEL__CONFIG_TERM_MAX_STACK: 819 max_stack = term->val.max_stack; 820 break; 821 case EVSEL__CONFIG_TERM_MAX_EVENTS: 822 evsel->max_events = term->val.max_events; 823 break; 824 case EVSEL__CONFIG_TERM_INHERIT: 825 /* 826 * attr->inherit should has already been set by 827 * evsel__config. If user explicitly set 828 * inherit using config terms, override global 829 * opt->no_inherit setting. 830 */ 831 attr->inherit = term->val.inherit ? 1 : 0; 832 break; 833 case EVSEL__CONFIG_TERM_OVERWRITE: 834 attr->write_backward = term->val.overwrite ? 1 : 0; 835 break; 836 case EVSEL__CONFIG_TERM_DRV_CFG: 837 break; 838 case EVSEL__CONFIG_TERM_PERCORE: 839 break; 840 case EVSEL__CONFIG_TERM_AUX_OUTPUT: 841 attr->aux_output = term->val.aux_output ? 1 : 0; 842 break; 843 case EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE: 844 /* Already applied by auxtrace */ 845 break; 846 case EVSEL__CONFIG_TERM_CFG_CHG: 847 break; 848 default: 849 break; 850 } 851 } 852 853 /* User explicitly set per-event callgraph, clear the old setting and reset. */ 854 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) { 855 bool sample_address = false; 856 857 if (max_stack) { 858 param.max_stack = max_stack; 859 if (callgraph_buf == NULL) 860 callgraph_buf = "fp"; 861 } 862 863 /* parse callgraph parameters */ 864 if (callgraph_buf != NULL) { 865 if (!strcmp(callgraph_buf, "no")) { 866 param.enabled = false; 867 param.record_mode = CALLCHAIN_NONE; 868 } else { 869 param.enabled = true; 870 if (parse_callchain_record(callgraph_buf, ¶m)) { 871 pr_err("per-event callgraph setting for %s failed. " 872 "Apply callgraph global setting for it\n", 873 evsel->name); 874 return; 875 } 876 if (param.record_mode == CALLCHAIN_DWARF) 877 sample_address = true; 878 } 879 } 880 if (dump_size > 0) { 881 dump_size = round_up(dump_size, sizeof(u64)); 882 param.dump_size = dump_size; 883 } 884 885 /* If global callgraph set, clear it */ 886 if (callchain_param.enabled) 887 perf_evsel__reset_callgraph(evsel, &callchain_param); 888 889 /* set perf-event callgraph */ 890 if (param.enabled) { 891 if (sample_address) { 892 evsel__set_sample_bit(evsel, ADDR); 893 evsel__set_sample_bit(evsel, DATA_SRC); 894 evsel->core.attr.mmap_data = track; 895 } 896 evsel__config_callchain(evsel, opts, ¶m); 897 } 898 } 899 } 900 901 struct evsel_config_term *__evsel__get_config_term(struct evsel *evsel, enum evsel_term_type type) 902 { 903 struct evsel_config_term *term, *found_term = NULL; 904 905 list_for_each_entry(term, &evsel->config_terms, list) { 906 if (term->type == type) 907 found_term = term; 908 } 909 910 return found_term; 911 } 912 913 /* 914 * The enable_on_exec/disabled value strategy: 915 * 916 * 1) For any type of traced program: 917 * - all independent events and group leaders are disabled 918 * - all group members are enabled 919 * 920 * Group members are ruled by group leaders. They need to 921 * be enabled, because the group scheduling relies on that. 922 * 923 * 2) For traced programs executed by perf: 924 * - all independent events and group leaders have 925 * enable_on_exec set 926 * - we don't specifically enable or disable any event during 927 * the record command 928 * 929 * Independent events and group leaders are initially disabled 930 * and get enabled by exec. Group members are ruled by group 931 * leaders as stated in 1). 932 * 933 * 3) For traced programs attached by perf (pid/tid): 934 * - we specifically enable or disable all events during 935 * the record command 936 * 937 * When attaching events to already running traced we 938 * enable/disable events specifically, as there's no 939 * initial traced exec call. 940 */ 941 void evsel__config(struct evsel *evsel, struct record_opts *opts, 942 struct callchain_param *callchain) 943 { 944 struct evsel *leader = evsel->leader; 945 struct perf_event_attr *attr = &evsel->core.attr; 946 int track = evsel->tracking; 947 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread; 948 949 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1; 950 attr->inherit = !opts->no_inherit; 951 attr->write_backward = opts->overwrite ? 1 : 0; 952 953 evsel__set_sample_bit(evsel, IP); 954 evsel__set_sample_bit(evsel, TID); 955 956 if (evsel->sample_read) { 957 evsel__set_sample_bit(evsel, READ); 958 959 /* 960 * We need ID even in case of single event, because 961 * PERF_SAMPLE_READ process ID specific data. 962 */ 963 evsel__set_sample_id(evsel, false); 964 965 /* 966 * Apply group format only if we belong to group 967 * with more than one members. 968 */ 969 if (leader->core.nr_members > 1) { 970 attr->read_format |= PERF_FORMAT_GROUP; 971 attr->inherit = 0; 972 } 973 } 974 975 /* 976 * We default some events to have a default interval. But keep 977 * it a weak assumption overridable by the user. 978 */ 979 if (!attr->sample_period || (opts->user_freq != UINT_MAX || 980 opts->user_interval != ULLONG_MAX)) { 981 if (opts->freq) { 982 evsel__set_sample_bit(evsel, PERIOD); 983 attr->freq = 1; 984 attr->sample_freq = opts->freq; 985 } else { 986 attr->sample_period = opts->default_interval; 987 } 988 } 989 990 if (opts->no_samples) 991 attr->sample_freq = 0; 992 993 if (opts->inherit_stat) { 994 evsel->core.attr.read_format |= 995 PERF_FORMAT_TOTAL_TIME_ENABLED | 996 PERF_FORMAT_TOTAL_TIME_RUNNING | 997 PERF_FORMAT_ID; 998 attr->inherit_stat = 1; 999 } 1000 1001 if (opts->sample_address) { 1002 evsel__set_sample_bit(evsel, ADDR); 1003 attr->mmap_data = track; 1004 } 1005 1006 /* 1007 * We don't allow user space callchains for function trace 1008 * event, due to issues with page faults while tracing page 1009 * fault handler and its overall trickiness nature. 1010 */ 1011 if (evsel__is_function_event(evsel)) 1012 evsel->core.attr.exclude_callchain_user = 1; 1013 1014 if (callchain && callchain->enabled && !evsel->no_aux_samples) 1015 evsel__config_callchain(evsel, opts, callchain); 1016 1017 if (opts->sample_intr_regs && !evsel->no_aux_samples) { 1018 attr->sample_regs_intr = opts->sample_intr_regs; 1019 evsel__set_sample_bit(evsel, REGS_INTR); 1020 } 1021 1022 if (opts->sample_user_regs && !evsel->no_aux_samples) { 1023 attr->sample_regs_user |= opts->sample_user_regs; 1024 evsel__set_sample_bit(evsel, REGS_USER); 1025 } 1026 1027 if (target__has_cpu(&opts->target) || opts->sample_cpu) 1028 evsel__set_sample_bit(evsel, CPU); 1029 1030 /* 1031 * When the user explicitly disabled time don't force it here. 1032 */ 1033 if (opts->sample_time && 1034 (!perf_missing_features.sample_id_all && 1035 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu || 1036 opts->sample_time_set))) 1037 evsel__set_sample_bit(evsel, TIME); 1038 1039 if (opts->raw_samples && !evsel->no_aux_samples) { 1040 evsel__set_sample_bit(evsel, TIME); 1041 evsel__set_sample_bit(evsel, RAW); 1042 evsel__set_sample_bit(evsel, CPU); 1043 } 1044 1045 if (opts->sample_address) 1046 evsel__set_sample_bit(evsel, DATA_SRC); 1047 1048 if (opts->sample_phys_addr) 1049 evsel__set_sample_bit(evsel, PHYS_ADDR); 1050 1051 if (opts->no_buffering) { 1052 attr->watermark = 0; 1053 attr->wakeup_events = 1; 1054 } 1055 if (opts->branch_stack && !evsel->no_aux_samples) { 1056 evsel__set_sample_bit(evsel, BRANCH_STACK); 1057 attr->branch_sample_type = opts->branch_stack; 1058 } 1059 1060 if (opts->sample_weight) 1061 evsel__set_sample_bit(evsel, WEIGHT); 1062 1063 attr->task = track; 1064 attr->mmap = track; 1065 attr->mmap2 = track && !perf_missing_features.mmap2; 1066 attr->comm = track; 1067 /* 1068 * ksymbol is tracked separately with text poke because it needs to be 1069 * system wide and enabled immediately. 1070 */ 1071 if (!opts->text_poke) 1072 attr->ksymbol = track && !perf_missing_features.ksymbol; 1073 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf; 1074 1075 if (opts->record_namespaces) 1076 attr->namespaces = track; 1077 1078 if (opts->record_cgroup) { 1079 attr->cgroup = track && !perf_missing_features.cgroup; 1080 evsel__set_sample_bit(evsel, CGROUP); 1081 } 1082 1083 if (opts->record_switch_events) 1084 attr->context_switch = track; 1085 1086 if (opts->sample_transaction) 1087 evsel__set_sample_bit(evsel, TRANSACTION); 1088 1089 if (opts->running_time) { 1090 evsel->core.attr.read_format |= 1091 PERF_FORMAT_TOTAL_TIME_ENABLED | 1092 PERF_FORMAT_TOTAL_TIME_RUNNING; 1093 } 1094 1095 /* 1096 * XXX see the function comment above 1097 * 1098 * Disabling only independent events or group leaders, 1099 * keeping group members enabled. 1100 */ 1101 if (evsel__is_group_leader(evsel)) 1102 attr->disabled = 1; 1103 1104 /* 1105 * Setting enable_on_exec for independent events and 1106 * group leaders for traced executed by perf. 1107 */ 1108 if (target__none(&opts->target) && evsel__is_group_leader(evsel) && 1109 !opts->initial_delay) 1110 attr->enable_on_exec = 1; 1111 1112 if (evsel->immediate) { 1113 attr->disabled = 0; 1114 attr->enable_on_exec = 0; 1115 } 1116 1117 clockid = opts->clockid; 1118 if (opts->use_clockid) { 1119 attr->use_clockid = 1; 1120 attr->clockid = opts->clockid; 1121 } 1122 1123 if (evsel->precise_max) 1124 attr->precise_ip = 3; 1125 1126 if (opts->all_user) { 1127 attr->exclude_kernel = 1; 1128 attr->exclude_user = 0; 1129 } 1130 1131 if (opts->all_kernel) { 1132 attr->exclude_kernel = 0; 1133 attr->exclude_user = 1; 1134 } 1135 1136 if (evsel->core.own_cpus || evsel->unit) 1137 evsel->core.attr.read_format |= PERF_FORMAT_ID; 1138 1139 /* 1140 * Apply event specific term settings, 1141 * it overloads any global configuration. 1142 */ 1143 evsel__apply_config_terms(evsel, opts, track); 1144 1145 evsel->ignore_missing_thread = opts->ignore_missing_thread; 1146 1147 /* The --period option takes the precedence. */ 1148 if (opts->period_set) { 1149 if (opts->period) 1150 evsel__set_sample_bit(evsel, PERIOD); 1151 else 1152 evsel__reset_sample_bit(evsel, PERIOD); 1153 } 1154 1155 /* 1156 * A dummy event never triggers any actual counter and therefore 1157 * cannot be used with branch_stack. 1158 * 1159 * For initial_delay, a dummy event is added implicitly. 1160 * The software event will trigger -EOPNOTSUPP error out, 1161 * if BRANCH_STACK bit is set. 1162 */ 1163 if (evsel__is_dummy_event(evsel)) 1164 evsel__reset_sample_bit(evsel, BRANCH_STACK); 1165 } 1166 1167 int evsel__set_filter(struct evsel *evsel, const char *filter) 1168 { 1169 char *new_filter = strdup(filter); 1170 1171 if (new_filter != NULL) { 1172 free(evsel->filter); 1173 evsel->filter = new_filter; 1174 return 0; 1175 } 1176 1177 return -1; 1178 } 1179 1180 static int evsel__append_filter(struct evsel *evsel, const char *fmt, const char *filter) 1181 { 1182 char *new_filter; 1183 1184 if (evsel->filter == NULL) 1185 return evsel__set_filter(evsel, filter); 1186 1187 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) { 1188 free(evsel->filter); 1189 evsel->filter = new_filter; 1190 return 0; 1191 } 1192 1193 return -1; 1194 } 1195 1196 int evsel__append_tp_filter(struct evsel *evsel, const char *filter) 1197 { 1198 return evsel__append_filter(evsel, "(%s) && (%s)", filter); 1199 } 1200 1201 int evsel__append_addr_filter(struct evsel *evsel, const char *filter) 1202 { 1203 return evsel__append_filter(evsel, "%s,%s", filter); 1204 } 1205 1206 /* Caller has to clear disabled after going through all CPUs. */ 1207 int evsel__enable_cpu(struct evsel *evsel, int cpu) 1208 { 1209 return perf_evsel__enable_cpu(&evsel->core, cpu); 1210 } 1211 1212 int evsel__enable(struct evsel *evsel) 1213 { 1214 int err = perf_evsel__enable(&evsel->core); 1215 1216 if (!err) 1217 evsel->disabled = false; 1218 return err; 1219 } 1220 1221 /* Caller has to set disabled after going through all CPUs. */ 1222 int evsel__disable_cpu(struct evsel *evsel, int cpu) 1223 { 1224 return perf_evsel__disable_cpu(&evsel->core, cpu); 1225 } 1226 1227 int evsel__disable(struct evsel *evsel) 1228 { 1229 int err = perf_evsel__disable(&evsel->core); 1230 /* 1231 * We mark it disabled here so that tools that disable a event can 1232 * ignore events after they disable it. I.e. the ring buffer may have 1233 * already a few more events queued up before the kernel got the stop 1234 * request. 1235 */ 1236 if (!err) 1237 evsel->disabled = true; 1238 1239 return err; 1240 } 1241 1242 static void evsel__free_config_terms(struct evsel *evsel) 1243 { 1244 struct evsel_config_term *term, *h; 1245 1246 list_for_each_entry_safe(term, h, &evsel->config_terms, list) { 1247 list_del_init(&term->list); 1248 if (term->free_str) 1249 zfree(&term->val.str); 1250 free(term); 1251 } 1252 } 1253 1254 void evsel__exit(struct evsel *evsel) 1255 { 1256 assert(list_empty(&evsel->core.node)); 1257 assert(evsel->evlist == NULL); 1258 evsel__free_counts(evsel); 1259 perf_evsel__free_fd(&evsel->core); 1260 perf_evsel__free_id(&evsel->core); 1261 evsel__free_config_terms(evsel); 1262 cgroup__put(evsel->cgrp); 1263 perf_cpu_map__put(evsel->core.cpus); 1264 perf_cpu_map__put(evsel->core.own_cpus); 1265 perf_thread_map__put(evsel->core.threads); 1266 zfree(&evsel->group_name); 1267 zfree(&evsel->name); 1268 zfree(&evsel->pmu_name); 1269 zfree(&evsel->per_pkg_mask); 1270 zfree(&evsel->metric_events); 1271 perf_evsel__object.fini(evsel); 1272 } 1273 1274 void evsel__delete(struct evsel *evsel) 1275 { 1276 evsel__exit(evsel); 1277 free(evsel); 1278 } 1279 1280 void evsel__compute_deltas(struct evsel *evsel, int cpu, int thread, 1281 struct perf_counts_values *count) 1282 { 1283 struct perf_counts_values tmp; 1284 1285 if (!evsel->prev_raw_counts) 1286 return; 1287 1288 if (cpu == -1) { 1289 tmp = evsel->prev_raw_counts->aggr; 1290 evsel->prev_raw_counts->aggr = *count; 1291 } else { 1292 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread); 1293 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count; 1294 } 1295 1296 count->val = count->val - tmp.val; 1297 count->ena = count->ena - tmp.ena; 1298 count->run = count->run - tmp.run; 1299 } 1300 1301 void perf_counts_values__scale(struct perf_counts_values *count, 1302 bool scale, s8 *pscaled) 1303 { 1304 s8 scaled = 0; 1305 1306 if (scale) { 1307 if (count->run == 0) { 1308 scaled = -1; 1309 count->val = 0; 1310 } else if (count->run < count->ena) { 1311 scaled = 1; 1312 count->val = (u64)((double) count->val * count->ena / count->run); 1313 } 1314 } 1315 1316 if (pscaled) 1317 *pscaled = scaled; 1318 } 1319 1320 static int evsel__read_one(struct evsel *evsel, int cpu, int thread) 1321 { 1322 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread); 1323 1324 return perf_evsel__read(&evsel->core, cpu, thread, count); 1325 } 1326 1327 static void 1328 perf_evsel__set_count(struct evsel *counter, int cpu, int thread, 1329 u64 val, u64 ena, u64 run) 1330 { 1331 struct perf_counts_values *count; 1332 1333 count = perf_counts(counter->counts, cpu, thread); 1334 1335 count->val = val; 1336 count->ena = ena; 1337 count->run = run; 1338 1339 perf_counts__set_loaded(counter->counts, cpu, thread, true); 1340 } 1341 1342 static int 1343 perf_evsel__process_group_data(struct evsel *leader, 1344 int cpu, int thread, u64 *data) 1345 { 1346 u64 read_format = leader->core.attr.read_format; 1347 struct sample_read_value *v; 1348 u64 nr, ena = 0, run = 0, i; 1349 1350 nr = *data++; 1351 1352 if (nr != (u64) leader->core.nr_members) 1353 return -EINVAL; 1354 1355 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1356 ena = *data++; 1357 1358 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1359 run = *data++; 1360 1361 v = (struct sample_read_value *) data; 1362 1363 perf_evsel__set_count(leader, cpu, thread, 1364 v[0].value, ena, run); 1365 1366 for (i = 1; i < nr; i++) { 1367 struct evsel *counter; 1368 1369 counter = perf_evlist__id2evsel(leader->evlist, v[i].id); 1370 if (!counter) 1371 return -EINVAL; 1372 1373 perf_evsel__set_count(counter, cpu, thread, 1374 v[i].value, ena, run); 1375 } 1376 1377 return 0; 1378 } 1379 1380 static int evsel__read_group(struct evsel *leader, int cpu, int thread) 1381 { 1382 struct perf_stat_evsel *ps = leader->stats; 1383 u64 read_format = leader->core.attr.read_format; 1384 int size = perf_evsel__read_size(&leader->core); 1385 u64 *data = ps->group_data; 1386 1387 if (!(read_format & PERF_FORMAT_ID)) 1388 return -EINVAL; 1389 1390 if (!evsel__is_group_leader(leader)) 1391 return -EINVAL; 1392 1393 if (!data) { 1394 data = zalloc(size); 1395 if (!data) 1396 return -ENOMEM; 1397 1398 ps->group_data = data; 1399 } 1400 1401 if (FD(leader, cpu, thread) < 0) 1402 return -EINVAL; 1403 1404 if (readn(FD(leader, cpu, thread), data, size) <= 0) 1405 return -errno; 1406 1407 return perf_evsel__process_group_data(leader, cpu, thread, data); 1408 } 1409 1410 int evsel__read_counter(struct evsel *evsel, int cpu, int thread) 1411 { 1412 u64 read_format = evsel->core.attr.read_format; 1413 1414 if (read_format & PERF_FORMAT_GROUP) 1415 return evsel__read_group(evsel, cpu, thread); 1416 1417 return evsel__read_one(evsel, cpu, thread); 1418 } 1419 1420 int __evsel__read_on_cpu(struct evsel *evsel, int cpu, int thread, bool scale) 1421 { 1422 struct perf_counts_values count; 1423 size_t nv = scale ? 3 : 1; 1424 1425 if (FD(evsel, cpu, thread) < 0) 1426 return -EINVAL; 1427 1428 if (evsel->counts == NULL && evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0) 1429 return -ENOMEM; 1430 1431 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0) 1432 return -errno; 1433 1434 evsel__compute_deltas(evsel, cpu, thread, &count); 1435 perf_counts_values__scale(&count, scale, NULL); 1436 *perf_counts(evsel->counts, cpu, thread) = count; 1437 return 0; 1438 } 1439 1440 static int get_group_fd(struct evsel *evsel, int cpu, int thread) 1441 { 1442 struct evsel *leader = evsel->leader; 1443 int fd; 1444 1445 if (evsel__is_group_leader(evsel)) 1446 return -1; 1447 1448 /* 1449 * Leader must be already processed/open, 1450 * if not it's a bug. 1451 */ 1452 BUG_ON(!leader->core.fd); 1453 1454 fd = FD(leader, cpu, thread); 1455 BUG_ON(fd == -1); 1456 1457 return fd; 1458 } 1459 1460 static void perf_evsel__remove_fd(struct evsel *pos, 1461 int nr_cpus, int nr_threads, 1462 int thread_idx) 1463 { 1464 for (int cpu = 0; cpu < nr_cpus; cpu++) 1465 for (int thread = thread_idx; thread < nr_threads - 1; thread++) 1466 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1); 1467 } 1468 1469 static int update_fds(struct evsel *evsel, 1470 int nr_cpus, int cpu_idx, 1471 int nr_threads, int thread_idx) 1472 { 1473 struct evsel *pos; 1474 1475 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads) 1476 return -EINVAL; 1477 1478 evlist__for_each_entry(evsel->evlist, pos) { 1479 nr_cpus = pos != evsel ? nr_cpus : cpu_idx; 1480 1481 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx); 1482 1483 /* 1484 * Since fds for next evsel has not been created, 1485 * there is no need to iterate whole event list. 1486 */ 1487 if (pos == evsel) 1488 break; 1489 } 1490 return 0; 1491 } 1492 1493 static bool ignore_missing_thread(struct evsel *evsel, 1494 int nr_cpus, int cpu, 1495 struct perf_thread_map *threads, 1496 int thread, int err) 1497 { 1498 pid_t ignore_pid = perf_thread_map__pid(threads, thread); 1499 1500 if (!evsel->ignore_missing_thread) 1501 return false; 1502 1503 /* The system wide setup does not work with threads. */ 1504 if (evsel->core.system_wide) 1505 return false; 1506 1507 /* The -ESRCH is perf event syscall errno for pid's not found. */ 1508 if (err != -ESRCH) 1509 return false; 1510 1511 /* If there's only one thread, let it fail. */ 1512 if (threads->nr == 1) 1513 return false; 1514 1515 /* 1516 * We should remove fd for missing_thread first 1517 * because thread_map__remove() will decrease threads->nr. 1518 */ 1519 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread)) 1520 return false; 1521 1522 if (thread_map__remove(threads, thread)) 1523 return false; 1524 1525 pr_warning("WARNING: Ignored open failure for pid %d\n", 1526 ignore_pid); 1527 return true; 1528 } 1529 1530 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val, 1531 void *priv __maybe_unused) 1532 { 1533 return fprintf(fp, " %-32s %s\n", name, val); 1534 } 1535 1536 static void display_attr(struct perf_event_attr *attr) 1537 { 1538 if (verbose >= 2 || debug_peo_args) { 1539 fprintf(stderr, "%.60s\n", graph_dotted_line); 1540 fprintf(stderr, "perf_event_attr:\n"); 1541 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL); 1542 fprintf(stderr, "%.60s\n", graph_dotted_line); 1543 } 1544 } 1545 1546 static int perf_event_open(struct evsel *evsel, 1547 pid_t pid, int cpu, int group_fd, 1548 unsigned long flags) 1549 { 1550 int precise_ip = evsel->core.attr.precise_ip; 1551 int fd; 1552 1553 while (1) { 1554 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx", 1555 pid, cpu, group_fd, flags); 1556 1557 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags); 1558 if (fd >= 0) 1559 break; 1560 1561 /* Do not try less precise if not requested. */ 1562 if (!evsel->precise_max) 1563 break; 1564 1565 /* 1566 * We tried all the precise_ip values, and it's 1567 * still failing, so leave it to standard fallback. 1568 */ 1569 if (!evsel->core.attr.precise_ip) { 1570 evsel->core.attr.precise_ip = precise_ip; 1571 break; 1572 } 1573 1574 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP); 1575 evsel->core.attr.precise_ip--; 1576 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip); 1577 display_attr(&evsel->core.attr); 1578 } 1579 1580 return fd; 1581 } 1582 1583 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, 1584 struct perf_thread_map *threads, 1585 int start_cpu, int end_cpu) 1586 { 1587 int cpu, thread, nthreads; 1588 unsigned long flags = PERF_FLAG_FD_CLOEXEC; 1589 int pid = -1, err, old_errno; 1590 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE; 1591 1592 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) || 1593 (perf_missing_features.aux_output && evsel->core.attr.aux_output)) 1594 return -EINVAL; 1595 1596 if (cpus == NULL) { 1597 static struct perf_cpu_map *empty_cpu_map; 1598 1599 if (empty_cpu_map == NULL) { 1600 empty_cpu_map = perf_cpu_map__dummy_new(); 1601 if (empty_cpu_map == NULL) 1602 return -ENOMEM; 1603 } 1604 1605 cpus = empty_cpu_map; 1606 } 1607 1608 if (threads == NULL) { 1609 static struct perf_thread_map *empty_thread_map; 1610 1611 if (empty_thread_map == NULL) { 1612 empty_thread_map = thread_map__new_by_tid(-1); 1613 if (empty_thread_map == NULL) 1614 return -ENOMEM; 1615 } 1616 1617 threads = empty_thread_map; 1618 } 1619 1620 if (evsel->core.system_wide) 1621 nthreads = 1; 1622 else 1623 nthreads = threads->nr; 1624 1625 if (evsel->core.fd == NULL && 1626 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0) 1627 return -ENOMEM; 1628 1629 if (evsel->cgrp) { 1630 flags |= PERF_FLAG_PID_CGROUP; 1631 pid = evsel->cgrp->fd; 1632 } 1633 1634 fallback_missing_features: 1635 if (perf_missing_features.clockid_wrong) 1636 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */ 1637 if (perf_missing_features.clockid) { 1638 evsel->core.attr.use_clockid = 0; 1639 evsel->core.attr.clockid = 0; 1640 } 1641 if (perf_missing_features.cloexec) 1642 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 1643 if (perf_missing_features.mmap2) 1644 evsel->core.attr.mmap2 = 0; 1645 if (perf_missing_features.exclude_guest) 1646 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0; 1647 if (perf_missing_features.lbr_flags) 1648 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS | 1649 PERF_SAMPLE_BRANCH_NO_CYCLES); 1650 if (perf_missing_features.group_read && evsel->core.attr.inherit) 1651 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID); 1652 if (perf_missing_features.ksymbol) 1653 evsel->core.attr.ksymbol = 0; 1654 if (perf_missing_features.bpf) 1655 evsel->core.attr.bpf_event = 0; 1656 if (perf_missing_features.branch_hw_idx) 1657 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX; 1658 retry_sample_id: 1659 if (perf_missing_features.sample_id_all) 1660 evsel->core.attr.sample_id_all = 0; 1661 1662 display_attr(&evsel->core.attr); 1663 1664 for (cpu = start_cpu; cpu < end_cpu; cpu++) { 1665 1666 for (thread = 0; thread < nthreads; thread++) { 1667 int fd, group_fd; 1668 1669 if (!evsel->cgrp && !evsel->core.system_wide) 1670 pid = perf_thread_map__pid(threads, thread); 1671 1672 group_fd = get_group_fd(evsel, cpu, thread); 1673 retry_open: 1674 test_attr__ready(); 1675 1676 fd = perf_event_open(evsel, pid, cpus->map[cpu], 1677 group_fd, flags); 1678 1679 FD(evsel, cpu, thread) = fd; 1680 1681 if (fd < 0) { 1682 err = -errno; 1683 1684 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) { 1685 /* 1686 * We just removed 1 thread, so take a step 1687 * back on thread index and lower the upper 1688 * nthreads limit. 1689 */ 1690 nthreads--; 1691 thread--; 1692 1693 /* ... and pretend like nothing have happened. */ 1694 err = 0; 1695 continue; 1696 } 1697 1698 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", 1699 err); 1700 goto try_fallback; 1701 } 1702 1703 pr_debug2_peo(" = %d\n", fd); 1704 1705 if (evsel->bpf_fd >= 0) { 1706 int evt_fd = fd; 1707 int bpf_fd = evsel->bpf_fd; 1708 1709 err = ioctl(evt_fd, 1710 PERF_EVENT_IOC_SET_BPF, 1711 bpf_fd); 1712 if (err && errno != EEXIST) { 1713 pr_err("failed to attach bpf fd %d: %s\n", 1714 bpf_fd, strerror(errno)); 1715 err = -EINVAL; 1716 goto out_close; 1717 } 1718 } 1719 1720 set_rlimit = NO_CHANGE; 1721 1722 /* 1723 * If we succeeded but had to kill clockid, fail and 1724 * have evsel__open_strerror() print us a nice error. 1725 */ 1726 if (perf_missing_features.clockid || 1727 perf_missing_features.clockid_wrong) { 1728 err = -EINVAL; 1729 goto out_close; 1730 } 1731 } 1732 } 1733 1734 return 0; 1735 1736 try_fallback: 1737 /* 1738 * perf stat needs between 5 and 22 fds per CPU. When we run out 1739 * of them try to increase the limits. 1740 */ 1741 if (err == -EMFILE && set_rlimit < INCREASED_MAX) { 1742 struct rlimit l; 1743 1744 old_errno = errno; 1745 if (getrlimit(RLIMIT_NOFILE, &l) == 0) { 1746 if (set_rlimit == NO_CHANGE) 1747 l.rlim_cur = l.rlim_max; 1748 else { 1749 l.rlim_cur = l.rlim_max + 1000; 1750 l.rlim_max = l.rlim_cur; 1751 } 1752 if (setrlimit(RLIMIT_NOFILE, &l) == 0) { 1753 set_rlimit++; 1754 errno = old_errno; 1755 goto retry_open; 1756 } 1757 } 1758 errno = old_errno; 1759 } 1760 1761 if (err != -EINVAL || cpu > 0 || thread > 0) 1762 goto out_close; 1763 1764 /* 1765 * Must probe features in the order they were added to the 1766 * perf_event_attr interface. 1767 */ 1768 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) { 1769 perf_missing_features.cgroup = true; 1770 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n"); 1771 goto out_close; 1772 } else if (!perf_missing_features.branch_hw_idx && 1773 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) { 1774 perf_missing_features.branch_hw_idx = true; 1775 pr_debug2("switching off branch HW index support\n"); 1776 goto fallback_missing_features; 1777 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) { 1778 perf_missing_features.aux_output = true; 1779 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n"); 1780 goto out_close; 1781 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) { 1782 perf_missing_features.bpf = true; 1783 pr_debug2_peo("switching off bpf_event\n"); 1784 goto fallback_missing_features; 1785 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) { 1786 perf_missing_features.ksymbol = true; 1787 pr_debug2_peo("switching off ksymbol\n"); 1788 goto fallback_missing_features; 1789 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) { 1790 perf_missing_features.write_backward = true; 1791 pr_debug2_peo("switching off write_backward\n"); 1792 goto out_close; 1793 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) { 1794 perf_missing_features.clockid_wrong = true; 1795 pr_debug2_peo("switching off clockid\n"); 1796 goto fallback_missing_features; 1797 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) { 1798 perf_missing_features.clockid = true; 1799 pr_debug2_peo("switching off use_clockid\n"); 1800 goto fallback_missing_features; 1801 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) { 1802 perf_missing_features.cloexec = true; 1803 pr_debug2_peo("switching off cloexec flag\n"); 1804 goto fallback_missing_features; 1805 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) { 1806 perf_missing_features.mmap2 = true; 1807 pr_debug2_peo("switching off mmap2\n"); 1808 goto fallback_missing_features; 1809 } else if (!perf_missing_features.exclude_guest && 1810 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) { 1811 perf_missing_features.exclude_guest = true; 1812 pr_debug2_peo("switching off exclude_guest, exclude_host\n"); 1813 goto fallback_missing_features; 1814 } else if (!perf_missing_features.sample_id_all) { 1815 perf_missing_features.sample_id_all = true; 1816 pr_debug2_peo("switching off sample_id_all\n"); 1817 goto retry_sample_id; 1818 } else if (!perf_missing_features.lbr_flags && 1819 (evsel->core.attr.branch_sample_type & 1820 (PERF_SAMPLE_BRANCH_NO_CYCLES | 1821 PERF_SAMPLE_BRANCH_NO_FLAGS))) { 1822 perf_missing_features.lbr_flags = true; 1823 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n"); 1824 goto fallback_missing_features; 1825 } else if (!perf_missing_features.group_read && 1826 evsel->core.attr.inherit && 1827 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) && 1828 evsel__is_group_leader(evsel)) { 1829 perf_missing_features.group_read = true; 1830 pr_debug2_peo("switching off group read\n"); 1831 goto fallback_missing_features; 1832 } 1833 out_close: 1834 if (err) 1835 threads->err_thread = thread; 1836 1837 old_errno = errno; 1838 do { 1839 while (--thread >= 0) { 1840 if (FD(evsel, cpu, thread) >= 0) 1841 close(FD(evsel, cpu, thread)); 1842 FD(evsel, cpu, thread) = -1; 1843 } 1844 thread = nthreads; 1845 } while (--cpu >= 0); 1846 errno = old_errno; 1847 return err; 1848 } 1849 1850 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus, 1851 struct perf_thread_map *threads) 1852 { 1853 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1); 1854 } 1855 1856 void evsel__close(struct evsel *evsel) 1857 { 1858 perf_evsel__close(&evsel->core); 1859 perf_evsel__free_id(&evsel->core); 1860 } 1861 1862 int evsel__open_per_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, int cpu) 1863 { 1864 if (cpu == -1) 1865 return evsel__open_cpu(evsel, cpus, NULL, 0, 1866 cpus ? cpus->nr : 1); 1867 1868 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1); 1869 } 1870 1871 int evsel__open_per_thread(struct evsel *evsel, struct perf_thread_map *threads) 1872 { 1873 return evsel__open(evsel, NULL, threads); 1874 } 1875 1876 static int perf_evsel__parse_id_sample(const struct evsel *evsel, 1877 const union perf_event *event, 1878 struct perf_sample *sample) 1879 { 1880 u64 type = evsel->core.attr.sample_type; 1881 const __u64 *array = event->sample.array; 1882 bool swapped = evsel->needs_swap; 1883 union u64_swap u; 1884 1885 array += ((event->header.size - 1886 sizeof(event->header)) / sizeof(u64)) - 1; 1887 1888 if (type & PERF_SAMPLE_IDENTIFIER) { 1889 sample->id = *array; 1890 array--; 1891 } 1892 1893 if (type & PERF_SAMPLE_CPU) { 1894 u.val64 = *array; 1895 if (swapped) { 1896 /* undo swap of u64, then swap on individual u32s */ 1897 u.val64 = bswap_64(u.val64); 1898 u.val32[0] = bswap_32(u.val32[0]); 1899 } 1900 1901 sample->cpu = u.val32[0]; 1902 array--; 1903 } 1904 1905 if (type & PERF_SAMPLE_STREAM_ID) { 1906 sample->stream_id = *array; 1907 array--; 1908 } 1909 1910 if (type & PERF_SAMPLE_ID) { 1911 sample->id = *array; 1912 array--; 1913 } 1914 1915 if (type & PERF_SAMPLE_TIME) { 1916 sample->time = *array; 1917 array--; 1918 } 1919 1920 if (type & PERF_SAMPLE_TID) { 1921 u.val64 = *array; 1922 if (swapped) { 1923 /* undo swap of u64, then swap on individual u32s */ 1924 u.val64 = bswap_64(u.val64); 1925 u.val32[0] = bswap_32(u.val32[0]); 1926 u.val32[1] = bswap_32(u.val32[1]); 1927 } 1928 1929 sample->pid = u.val32[0]; 1930 sample->tid = u.val32[1]; 1931 array--; 1932 } 1933 1934 return 0; 1935 } 1936 1937 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 1938 u64 size) 1939 { 1940 return size > max_size || offset + size > endp; 1941 } 1942 1943 #define OVERFLOW_CHECK(offset, size, max_size) \ 1944 do { \ 1945 if (overflow(endp, (max_size), (offset), (size))) \ 1946 return -EFAULT; \ 1947 } while (0) 1948 1949 #define OVERFLOW_CHECK_u64(offset) \ 1950 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 1951 1952 static int 1953 perf_event__check_size(union perf_event *event, unsigned int sample_size) 1954 { 1955 /* 1956 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 1957 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 1958 * check the format does not go past the end of the event. 1959 */ 1960 if (sample_size + sizeof(event->header) > event->header.size) 1961 return -EFAULT; 1962 1963 return 0; 1964 } 1965 1966 int evsel__parse_sample(struct evsel *evsel, union perf_event *event, 1967 struct perf_sample *data) 1968 { 1969 u64 type = evsel->core.attr.sample_type; 1970 bool swapped = evsel->needs_swap; 1971 const __u64 *array; 1972 u16 max_size = event->header.size; 1973 const void *endp = (void *)event + max_size; 1974 u64 sz; 1975 1976 /* 1977 * used for cross-endian analysis. See git commit 65014ab3 1978 * for why this goofiness is needed. 1979 */ 1980 union u64_swap u; 1981 1982 memset(data, 0, sizeof(*data)); 1983 data->cpu = data->pid = data->tid = -1; 1984 data->stream_id = data->id = data->time = -1ULL; 1985 data->period = evsel->core.attr.sample_period; 1986 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 1987 data->misc = event->header.misc; 1988 data->id = -1ULL; 1989 data->data_src = PERF_MEM_DATA_SRC_NONE; 1990 1991 if (event->header.type != PERF_RECORD_SAMPLE) { 1992 if (!evsel->core.attr.sample_id_all) 1993 return 0; 1994 return perf_evsel__parse_id_sample(evsel, event, data); 1995 } 1996 1997 array = event->sample.array; 1998 1999 if (perf_event__check_size(event, evsel->sample_size)) 2000 return -EFAULT; 2001 2002 if (type & PERF_SAMPLE_IDENTIFIER) { 2003 data->id = *array; 2004 array++; 2005 } 2006 2007 if (type & PERF_SAMPLE_IP) { 2008 data->ip = *array; 2009 array++; 2010 } 2011 2012 if (type & PERF_SAMPLE_TID) { 2013 u.val64 = *array; 2014 if (swapped) { 2015 /* undo swap of u64, then swap on individual u32s */ 2016 u.val64 = bswap_64(u.val64); 2017 u.val32[0] = bswap_32(u.val32[0]); 2018 u.val32[1] = bswap_32(u.val32[1]); 2019 } 2020 2021 data->pid = u.val32[0]; 2022 data->tid = u.val32[1]; 2023 array++; 2024 } 2025 2026 if (type & PERF_SAMPLE_TIME) { 2027 data->time = *array; 2028 array++; 2029 } 2030 2031 if (type & PERF_SAMPLE_ADDR) { 2032 data->addr = *array; 2033 array++; 2034 } 2035 2036 if (type & PERF_SAMPLE_ID) { 2037 data->id = *array; 2038 array++; 2039 } 2040 2041 if (type & PERF_SAMPLE_STREAM_ID) { 2042 data->stream_id = *array; 2043 array++; 2044 } 2045 2046 if (type & PERF_SAMPLE_CPU) { 2047 2048 u.val64 = *array; 2049 if (swapped) { 2050 /* undo swap of u64, then swap on individual u32s */ 2051 u.val64 = bswap_64(u.val64); 2052 u.val32[0] = bswap_32(u.val32[0]); 2053 } 2054 2055 data->cpu = u.val32[0]; 2056 array++; 2057 } 2058 2059 if (type & PERF_SAMPLE_PERIOD) { 2060 data->period = *array; 2061 array++; 2062 } 2063 2064 if (type & PERF_SAMPLE_READ) { 2065 u64 read_format = evsel->core.attr.read_format; 2066 2067 OVERFLOW_CHECK_u64(array); 2068 if (read_format & PERF_FORMAT_GROUP) 2069 data->read.group.nr = *array; 2070 else 2071 data->read.one.value = *array; 2072 2073 array++; 2074 2075 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 2076 OVERFLOW_CHECK_u64(array); 2077 data->read.time_enabled = *array; 2078 array++; 2079 } 2080 2081 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 2082 OVERFLOW_CHECK_u64(array); 2083 data->read.time_running = *array; 2084 array++; 2085 } 2086 2087 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 2088 if (read_format & PERF_FORMAT_GROUP) { 2089 const u64 max_group_nr = UINT64_MAX / 2090 sizeof(struct sample_read_value); 2091 2092 if (data->read.group.nr > max_group_nr) 2093 return -EFAULT; 2094 sz = data->read.group.nr * 2095 sizeof(struct sample_read_value); 2096 OVERFLOW_CHECK(array, sz, max_size); 2097 data->read.group.values = 2098 (struct sample_read_value *)array; 2099 array = (void *)array + sz; 2100 } else { 2101 OVERFLOW_CHECK_u64(array); 2102 data->read.one.id = *array; 2103 array++; 2104 } 2105 } 2106 2107 if (type & PERF_SAMPLE_CALLCHAIN) { 2108 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 2109 2110 OVERFLOW_CHECK_u64(array); 2111 data->callchain = (struct ip_callchain *)array++; 2112 if (data->callchain->nr > max_callchain_nr) 2113 return -EFAULT; 2114 sz = data->callchain->nr * sizeof(u64); 2115 OVERFLOW_CHECK(array, sz, max_size); 2116 array = (void *)array + sz; 2117 } 2118 2119 if (type & PERF_SAMPLE_RAW) { 2120 OVERFLOW_CHECK_u64(array); 2121 u.val64 = *array; 2122 2123 /* 2124 * Undo swap of u64, then swap on individual u32s, 2125 * get the size of the raw area and undo all of the 2126 * swap. The pevent interface handles endianity by 2127 * itself. 2128 */ 2129 if (swapped) { 2130 u.val64 = bswap_64(u.val64); 2131 u.val32[0] = bswap_32(u.val32[0]); 2132 u.val32[1] = bswap_32(u.val32[1]); 2133 } 2134 data->raw_size = u.val32[0]; 2135 2136 /* 2137 * The raw data is aligned on 64bits including the 2138 * u32 size, so it's safe to use mem_bswap_64. 2139 */ 2140 if (swapped) 2141 mem_bswap_64((void *) array, data->raw_size); 2142 2143 array = (void *)array + sizeof(u32); 2144 2145 OVERFLOW_CHECK(array, data->raw_size, max_size); 2146 data->raw_data = (void *)array; 2147 array = (void *)array + data->raw_size; 2148 } 2149 2150 if (type & PERF_SAMPLE_BRANCH_STACK) { 2151 const u64 max_branch_nr = UINT64_MAX / 2152 sizeof(struct branch_entry); 2153 2154 OVERFLOW_CHECK_u64(array); 2155 data->branch_stack = (struct branch_stack *)array++; 2156 2157 if (data->branch_stack->nr > max_branch_nr) 2158 return -EFAULT; 2159 2160 sz = data->branch_stack->nr * sizeof(struct branch_entry); 2161 if (evsel__has_branch_hw_idx(evsel)) 2162 sz += sizeof(u64); 2163 else 2164 data->no_hw_idx = true; 2165 OVERFLOW_CHECK(array, sz, max_size); 2166 array = (void *)array + sz; 2167 } 2168 2169 if (type & PERF_SAMPLE_REGS_USER) { 2170 OVERFLOW_CHECK_u64(array); 2171 data->user_regs.abi = *array; 2172 array++; 2173 2174 if (data->user_regs.abi) { 2175 u64 mask = evsel->core.attr.sample_regs_user; 2176 2177 sz = hweight64(mask) * sizeof(u64); 2178 OVERFLOW_CHECK(array, sz, max_size); 2179 data->user_regs.mask = mask; 2180 data->user_regs.regs = (u64 *)array; 2181 array = (void *)array + sz; 2182 } 2183 } 2184 2185 if (type & PERF_SAMPLE_STACK_USER) { 2186 OVERFLOW_CHECK_u64(array); 2187 sz = *array++; 2188 2189 data->user_stack.offset = ((char *)(array - 1) 2190 - (char *) event); 2191 2192 if (!sz) { 2193 data->user_stack.size = 0; 2194 } else { 2195 OVERFLOW_CHECK(array, sz, max_size); 2196 data->user_stack.data = (char *)array; 2197 array = (void *)array + sz; 2198 OVERFLOW_CHECK_u64(array); 2199 data->user_stack.size = *array++; 2200 if (WARN_ONCE(data->user_stack.size > sz, 2201 "user stack dump failure\n")) 2202 return -EFAULT; 2203 } 2204 } 2205 2206 if (type & PERF_SAMPLE_WEIGHT) { 2207 OVERFLOW_CHECK_u64(array); 2208 data->weight = *array; 2209 array++; 2210 } 2211 2212 if (type & PERF_SAMPLE_DATA_SRC) { 2213 OVERFLOW_CHECK_u64(array); 2214 data->data_src = *array; 2215 array++; 2216 } 2217 2218 if (type & PERF_SAMPLE_TRANSACTION) { 2219 OVERFLOW_CHECK_u64(array); 2220 data->transaction = *array; 2221 array++; 2222 } 2223 2224 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE; 2225 if (type & PERF_SAMPLE_REGS_INTR) { 2226 OVERFLOW_CHECK_u64(array); 2227 data->intr_regs.abi = *array; 2228 array++; 2229 2230 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) { 2231 u64 mask = evsel->core.attr.sample_regs_intr; 2232 2233 sz = hweight64(mask) * sizeof(u64); 2234 OVERFLOW_CHECK(array, sz, max_size); 2235 data->intr_regs.mask = mask; 2236 data->intr_regs.regs = (u64 *)array; 2237 array = (void *)array + sz; 2238 } 2239 } 2240 2241 data->phys_addr = 0; 2242 if (type & PERF_SAMPLE_PHYS_ADDR) { 2243 data->phys_addr = *array; 2244 array++; 2245 } 2246 2247 data->cgroup = 0; 2248 if (type & PERF_SAMPLE_CGROUP) { 2249 data->cgroup = *array; 2250 array++; 2251 } 2252 2253 if (type & PERF_SAMPLE_AUX) { 2254 OVERFLOW_CHECK_u64(array); 2255 sz = *array++; 2256 2257 OVERFLOW_CHECK(array, sz, max_size); 2258 /* Undo swap of data */ 2259 if (swapped) 2260 mem_bswap_64((char *)array, sz); 2261 data->aux_sample.size = sz; 2262 data->aux_sample.data = (char *)array; 2263 array = (void *)array + sz; 2264 } 2265 2266 return 0; 2267 } 2268 2269 int evsel__parse_sample_timestamp(struct evsel *evsel, union perf_event *event, 2270 u64 *timestamp) 2271 { 2272 u64 type = evsel->core.attr.sample_type; 2273 const __u64 *array; 2274 2275 if (!(type & PERF_SAMPLE_TIME)) 2276 return -1; 2277 2278 if (event->header.type != PERF_RECORD_SAMPLE) { 2279 struct perf_sample data = { 2280 .time = -1ULL, 2281 }; 2282 2283 if (!evsel->core.attr.sample_id_all) 2284 return -1; 2285 if (perf_evsel__parse_id_sample(evsel, event, &data)) 2286 return -1; 2287 2288 *timestamp = data.time; 2289 return 0; 2290 } 2291 2292 array = event->sample.array; 2293 2294 if (perf_event__check_size(event, evsel->sample_size)) 2295 return -EFAULT; 2296 2297 if (type & PERF_SAMPLE_IDENTIFIER) 2298 array++; 2299 2300 if (type & PERF_SAMPLE_IP) 2301 array++; 2302 2303 if (type & PERF_SAMPLE_TID) 2304 array++; 2305 2306 if (type & PERF_SAMPLE_TIME) 2307 *timestamp = *array; 2308 2309 return 0; 2310 } 2311 2312 struct tep_format_field *evsel__field(struct evsel *evsel, const char *name) 2313 { 2314 return tep_find_field(evsel->tp_format, name); 2315 } 2316 2317 void *evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, const char *name) 2318 { 2319 struct tep_format_field *field = evsel__field(evsel, name); 2320 int offset; 2321 2322 if (!field) 2323 return NULL; 2324 2325 offset = field->offset; 2326 2327 if (field->flags & TEP_FIELD_IS_DYNAMIC) { 2328 offset = *(int *)(sample->raw_data + field->offset); 2329 offset &= 0xffff; 2330 } 2331 2332 return sample->raw_data + offset; 2333 } 2334 2335 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample, 2336 bool needs_swap) 2337 { 2338 u64 value; 2339 void *ptr = sample->raw_data + field->offset; 2340 2341 switch (field->size) { 2342 case 1: 2343 return *(u8 *)ptr; 2344 case 2: 2345 value = *(u16 *)ptr; 2346 break; 2347 case 4: 2348 value = *(u32 *)ptr; 2349 break; 2350 case 8: 2351 memcpy(&value, ptr, sizeof(u64)); 2352 break; 2353 default: 2354 return 0; 2355 } 2356 2357 if (!needs_swap) 2358 return value; 2359 2360 switch (field->size) { 2361 case 2: 2362 return bswap_16(value); 2363 case 4: 2364 return bswap_32(value); 2365 case 8: 2366 return bswap_64(value); 2367 default: 2368 return 0; 2369 } 2370 2371 return 0; 2372 } 2373 2374 u64 evsel__intval(struct evsel *evsel, struct perf_sample *sample, const char *name) 2375 { 2376 struct tep_format_field *field = evsel__field(evsel, name); 2377 2378 if (!field) 2379 return 0; 2380 2381 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0; 2382 } 2383 2384 bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize) 2385 { 2386 int paranoid; 2387 2388 if ((err == ENOENT || err == ENXIO || err == ENODEV) && 2389 evsel->core.attr.type == PERF_TYPE_HARDWARE && 2390 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) { 2391 /* 2392 * If it's cycles then fall back to hrtimer based 2393 * cpu-clock-tick sw counter, which is always available even if 2394 * no PMU support. 2395 * 2396 * PPC returns ENXIO until 2.6.37 (behavior changed with commit 2397 * b0a873e). 2398 */ 2399 scnprintf(msg, msgsize, "%s", 2400 "The cycles event is not supported, trying to fall back to cpu-clock-ticks"); 2401 2402 evsel->core.attr.type = PERF_TYPE_SOFTWARE; 2403 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK; 2404 2405 zfree(&evsel->name); 2406 return true; 2407 } else if (err == EACCES && !evsel->core.attr.exclude_kernel && 2408 (paranoid = perf_event_paranoid()) > 1) { 2409 const char *name = evsel__name(evsel); 2410 char *new_name; 2411 const char *sep = ":"; 2412 2413 /* If event has exclude user then don't exclude kernel. */ 2414 if (evsel->core.attr.exclude_user) 2415 return false; 2416 2417 /* Is there already the separator in the name. */ 2418 if (strchr(name, '/') || 2419 (strchr(name, ':') && !evsel->is_libpfm_event)) 2420 sep = ""; 2421 2422 if (asprintf(&new_name, "%s%su", name, sep) < 0) 2423 return false; 2424 2425 if (evsel->name) 2426 free(evsel->name); 2427 evsel->name = new_name; 2428 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying " 2429 "to fall back to excluding kernel and hypervisor " 2430 " samples", paranoid); 2431 evsel->core.attr.exclude_kernel = 1; 2432 evsel->core.attr.exclude_hv = 1; 2433 2434 return true; 2435 } 2436 2437 return false; 2438 } 2439 2440 static bool find_process(const char *name) 2441 { 2442 size_t len = strlen(name); 2443 DIR *dir; 2444 struct dirent *d; 2445 int ret = -1; 2446 2447 dir = opendir(procfs__mountpoint()); 2448 if (!dir) 2449 return false; 2450 2451 /* Walk through the directory. */ 2452 while (ret && (d = readdir(dir)) != NULL) { 2453 char path[PATH_MAX]; 2454 char *data; 2455 size_t size; 2456 2457 if ((d->d_type != DT_DIR) || 2458 !strcmp(".", d->d_name) || 2459 !strcmp("..", d->d_name)) 2460 continue; 2461 2462 scnprintf(path, sizeof(path), "%s/%s/comm", 2463 procfs__mountpoint(), d->d_name); 2464 2465 if (filename__read_str(path, &data, &size)) 2466 continue; 2467 2468 ret = strncmp(name, data, len); 2469 free(data); 2470 } 2471 2472 closedir(dir); 2473 return ret ? false : true; 2474 } 2475 2476 int evsel__open_strerror(struct evsel *evsel, struct target *target, 2477 int err, char *msg, size_t size) 2478 { 2479 char sbuf[STRERR_BUFSIZE]; 2480 int printed = 0, enforced = 0; 2481 2482 switch (err) { 2483 case EPERM: 2484 case EACCES: 2485 printed += scnprintf(msg + printed, size - printed, 2486 "Access to performance monitoring and observability operations is limited.\n"); 2487 2488 if (!sysfs__read_int("fs/selinux/enforce", &enforced)) { 2489 if (enforced) { 2490 printed += scnprintf(msg + printed, size - printed, 2491 "Enforced MAC policy settings (SELinux) can limit access to performance\n" 2492 "monitoring and observability operations. Inspect system audit records for\n" 2493 "more perf_event access control information and adjusting the policy.\n"); 2494 } 2495 } 2496 2497 if (err == EPERM) 2498 printed += scnprintf(msg, size, 2499 "No permission to enable %s event.\n\n", evsel__name(evsel)); 2500 2501 return scnprintf(msg + printed, size - printed, 2502 "Consider adjusting /proc/sys/kernel/perf_event_paranoid setting to open\n" 2503 "access to performance monitoring and observability operations for users\n" 2504 "without CAP_PERFMON or CAP_SYS_ADMIN Linux capability.\n" 2505 "perf_event_paranoid setting is %d:\n" 2506 " -1: Allow use of (almost) all events by all users\n" 2507 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n" 2508 ">= 0: Disallow raw and ftrace function tracepoint access\n" 2509 ">= 1: Disallow CPU event access\n" 2510 ">= 2: Disallow kernel profiling\n" 2511 "To make the adjusted perf_event_paranoid setting permanent preserve it\n" 2512 "in /etc/sysctl.conf (e.g. kernel.perf_event_paranoid = <setting>)", 2513 perf_event_paranoid()); 2514 case ENOENT: 2515 return scnprintf(msg, size, "The %s event is not supported.", evsel__name(evsel)); 2516 case EMFILE: 2517 return scnprintf(msg, size, "%s", 2518 "Too many events are opened.\n" 2519 "Probably the maximum number of open file descriptors has been reached.\n" 2520 "Hint: Try again after reducing the number of events.\n" 2521 "Hint: Try increasing the limit with 'ulimit -n <limit>'"); 2522 case ENOMEM: 2523 if (evsel__has_callchain(evsel) && 2524 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0) 2525 return scnprintf(msg, size, 2526 "Not enough memory to setup event with callchain.\n" 2527 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n" 2528 "Hint: Current value: %d", sysctl__max_stack()); 2529 break; 2530 case ENODEV: 2531 if (target->cpu_list) 2532 return scnprintf(msg, size, "%s", 2533 "No such device - did you specify an out-of-range profile CPU?"); 2534 break; 2535 case EOPNOTSUPP: 2536 if (evsel->core.attr.sample_period != 0) 2537 return scnprintf(msg, size, 2538 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'", 2539 evsel__name(evsel)); 2540 if (evsel->core.attr.precise_ip) 2541 return scnprintf(msg, size, "%s", 2542 "\'precise\' request may not be supported. Try removing 'p' modifier."); 2543 #if defined(__i386__) || defined(__x86_64__) 2544 if (evsel->core.attr.type == PERF_TYPE_HARDWARE) 2545 return scnprintf(msg, size, "%s", 2546 "No hardware sampling interrupt available.\n"); 2547 #endif 2548 break; 2549 case EBUSY: 2550 if (find_process("oprofiled")) 2551 return scnprintf(msg, size, 2552 "The PMU counters are busy/taken by another profiler.\n" 2553 "We found oprofile daemon running, please stop it and try again."); 2554 break; 2555 case EINVAL: 2556 if (evsel->core.attr.write_backward && perf_missing_features.write_backward) 2557 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel."); 2558 if (perf_missing_features.clockid) 2559 return scnprintf(msg, size, "clockid feature not supported."); 2560 if (perf_missing_features.clockid_wrong) 2561 return scnprintf(msg, size, "wrong clockid (%d).", clockid); 2562 if (perf_missing_features.aux_output) 2563 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel."); 2564 break; 2565 default: 2566 break; 2567 } 2568 2569 return scnprintf(msg, size, 2570 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n" 2571 "/bin/dmesg | grep -i perf may provide additional information.\n", 2572 err, str_error_r(err, sbuf, sizeof(sbuf)), evsel__name(evsel)); 2573 } 2574 2575 struct perf_env *evsel__env(struct evsel *evsel) 2576 { 2577 if (evsel && evsel->evlist) 2578 return evsel->evlist->env; 2579 return &perf_env; 2580 } 2581 2582 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist) 2583 { 2584 int cpu, thread; 2585 2586 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) { 2587 for (thread = 0; thread < xyarray__max_y(evsel->core.fd); 2588 thread++) { 2589 int fd = FD(evsel, cpu, thread); 2590 2591 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, 2592 cpu, thread, fd) < 0) 2593 return -1; 2594 } 2595 } 2596 2597 return 0; 2598 } 2599 2600 int evsel__store_ids(struct evsel *evsel, struct evlist *evlist) 2601 { 2602 struct perf_cpu_map *cpus = evsel->core.cpus; 2603 struct perf_thread_map *threads = evsel->core.threads; 2604 2605 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr)) 2606 return -ENOMEM; 2607 2608 return store_evsel_ids(evsel, evlist); 2609 } 2610