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 perf_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 perf_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 = perf_evsel__no_extra_init, 77 .fini = perf_evsel__no_extra_fini, 78 }; 79 80 int perf_evsel__object_config(size_t object_size, 81 int (*init)(struct evsel *evsel), 82 void (*fini)(struct evsel *evsel)) 83 { 84 85 if (object_size == 0) 86 goto set_methods; 87 88 if (perf_evsel__object.size > object_size) 89 return -EINVAL; 90 91 perf_evsel__object.size = object_size; 92 93 set_methods: 94 if (init != NULL) 95 perf_evsel__object.init = init; 96 97 if (fini != NULL) 98 perf_evsel__object.fini = fini; 99 100 return 0; 101 } 102 103 #define FD(e, x, y) (*(int *)xyarray__entry(e->core.fd, x, y)) 104 105 int __perf_evsel__sample_size(u64 sample_type) 106 { 107 u64 mask = sample_type & PERF_SAMPLE_MASK; 108 int size = 0; 109 int i; 110 111 for (i = 0; i < 64; i++) { 112 if (mask & (1ULL << i)) 113 size++; 114 } 115 116 size *= sizeof(u64); 117 118 return size; 119 } 120 121 /** 122 * __perf_evsel__calc_id_pos - calculate id_pos. 123 * @sample_type: sample type 124 * 125 * This function returns the position of the event id (PERF_SAMPLE_ID or 126 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct 127 * perf_record_sample. 128 */ 129 static int __perf_evsel__calc_id_pos(u64 sample_type) 130 { 131 int idx = 0; 132 133 if (sample_type & PERF_SAMPLE_IDENTIFIER) 134 return 0; 135 136 if (!(sample_type & PERF_SAMPLE_ID)) 137 return -1; 138 139 if (sample_type & PERF_SAMPLE_IP) 140 idx += 1; 141 142 if (sample_type & PERF_SAMPLE_TID) 143 idx += 1; 144 145 if (sample_type & PERF_SAMPLE_TIME) 146 idx += 1; 147 148 if (sample_type & PERF_SAMPLE_ADDR) 149 idx += 1; 150 151 return idx; 152 } 153 154 /** 155 * __perf_evsel__calc_is_pos - calculate is_pos. 156 * @sample_type: sample type 157 * 158 * This function returns the position (counting backwards) of the event id 159 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if 160 * sample_id_all is used there is an id sample appended to non-sample events. 161 */ 162 static int __perf_evsel__calc_is_pos(u64 sample_type) 163 { 164 int idx = 1; 165 166 if (sample_type & PERF_SAMPLE_IDENTIFIER) 167 return 1; 168 169 if (!(sample_type & PERF_SAMPLE_ID)) 170 return -1; 171 172 if (sample_type & PERF_SAMPLE_CPU) 173 idx += 1; 174 175 if (sample_type & PERF_SAMPLE_STREAM_ID) 176 idx += 1; 177 178 return idx; 179 } 180 181 void perf_evsel__calc_id_pos(struct evsel *evsel) 182 { 183 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->core.attr.sample_type); 184 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->core.attr.sample_type); 185 } 186 187 void __perf_evsel__set_sample_bit(struct evsel *evsel, 188 enum perf_event_sample_format bit) 189 { 190 if (!(evsel->core.attr.sample_type & bit)) { 191 evsel->core.attr.sample_type |= bit; 192 evsel->sample_size += sizeof(u64); 193 perf_evsel__calc_id_pos(evsel); 194 } 195 } 196 197 void __perf_evsel__reset_sample_bit(struct evsel *evsel, 198 enum perf_event_sample_format bit) 199 { 200 if (evsel->core.attr.sample_type & bit) { 201 evsel->core.attr.sample_type &= ~bit; 202 evsel->sample_size -= sizeof(u64); 203 perf_evsel__calc_id_pos(evsel); 204 } 205 } 206 207 void perf_evsel__set_sample_id(struct evsel *evsel, 208 bool can_sample_identifier) 209 { 210 if (can_sample_identifier) { 211 perf_evsel__reset_sample_bit(evsel, ID); 212 perf_evsel__set_sample_bit(evsel, IDENTIFIER); 213 } else { 214 perf_evsel__set_sample_bit(evsel, ID); 215 } 216 evsel->core.attr.read_format |= PERF_FORMAT_ID; 217 } 218 219 /** 220 * perf_evsel__is_function_event - Return whether given evsel is a function 221 * trace event 222 * 223 * @evsel - evsel selector to be tested 224 * 225 * Return %true if event is function trace event 226 */ 227 bool perf_evsel__is_function_event(struct evsel *evsel) 228 { 229 #define FUNCTION_EVENT "ftrace:function" 230 231 return evsel->name && 232 !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT)); 233 234 #undef FUNCTION_EVENT 235 } 236 237 void evsel__init(struct evsel *evsel, 238 struct perf_event_attr *attr, int idx) 239 { 240 perf_evsel__init(&evsel->core, attr); 241 evsel->idx = idx; 242 evsel->tracking = !idx; 243 evsel->leader = evsel; 244 evsel->unit = ""; 245 evsel->scale = 1.0; 246 evsel->max_events = ULONG_MAX; 247 evsel->evlist = NULL; 248 evsel->bpf_obj = NULL; 249 evsel->bpf_fd = -1; 250 INIT_LIST_HEAD(&evsel->config_terms); 251 perf_evsel__object.init(evsel); 252 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type); 253 perf_evsel__calc_id_pos(evsel); 254 evsel->cmdline_group_boundary = false; 255 evsel->metric_expr = NULL; 256 evsel->metric_name = NULL; 257 evsel->metric_events = NULL; 258 evsel->collect_stat = false; 259 evsel->pmu_name = NULL; 260 } 261 262 struct evsel *perf_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 (perf_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 (perf_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 *perf_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 *perf_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 *perf_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 *__perf_evsel__hw_name(u64 config) 389 { 390 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config]) 391 return perf_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 perf_evsel__hw_name(struct evsel *evsel, char *bf, size_t size) 433 { 434 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->core.attr.config)); 435 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 436 } 437 438 const char *perf_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 *__perf_evsel__sw_name(u64 config) 452 { 453 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config]) 454 return perf_evsel__sw_names[config]; 455 return "unknown-software"; 456 } 457 458 static int perf_evsel__sw_name(struct evsel *evsel, char *bf, size_t size) 459 { 460 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->core.attr.config)); 461 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 462 } 463 464 static int __perf_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 perf_evsel__bp_name(struct evsel *evsel, char *bf, size_t size) 483 { 484 struct perf_event_attr *attr = &evsel->core.attr; 485 int r = __perf_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 *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX] 490 [PERF_EVSEL__MAX_ALIASES] = { 491 { "L1-dcache", "l1-d", "l1d", "L1-data", }, 492 { "L1-icache", "l1-i", "l1i", "L1-instruction", }, 493 { "LLC", "L2", }, 494 { "dTLB", "d-tlb", "Data-TLB", }, 495 { "iTLB", "i-tlb", "Instruction-TLB", }, 496 { "branch", "branches", "bpu", "btb", "bpc", }, 497 { "node", }, 498 }; 499 500 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX] 501 [PERF_EVSEL__MAX_ALIASES] = { 502 { "load", "loads", "read", }, 503 { "store", "stores", "write", }, 504 { "prefetch", "prefetches", "speculative-read", "speculative-load", }, 505 }; 506 507 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX] 508 [PERF_EVSEL__MAX_ALIASES] = { 509 { "refs", "Reference", "ops", "access", }, 510 { "misses", "miss", }, 511 }; 512 513 #define C(x) PERF_COUNT_HW_CACHE_##x 514 #define CACHE_READ (1 << C(OP_READ)) 515 #define CACHE_WRITE (1 << C(OP_WRITE)) 516 #define CACHE_PREFETCH (1 << C(OP_PREFETCH)) 517 #define COP(x) (1 << x) 518 519 /* 520 * cache operartion stat 521 * L1I : Read and prefetch only 522 * ITLB and BPU : Read-only 523 */ 524 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = { 525 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 526 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH), 527 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 528 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 529 [C(ITLB)] = (CACHE_READ), 530 [C(BPU)] = (CACHE_READ), 531 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 532 }; 533 534 bool perf_evsel__is_cache_op_valid(u8 type, u8 op) 535 { 536 if (perf_evsel__hw_cache_stat[type] & COP(op)) 537 return true; /* valid */ 538 else 539 return false; /* invalid */ 540 } 541 542 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, 543 char *bf, size_t size) 544 { 545 if (result) { 546 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0], 547 perf_evsel__hw_cache_op[op][0], 548 perf_evsel__hw_cache_result[result][0]); 549 } 550 551 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0], 552 perf_evsel__hw_cache_op[op][1]); 553 } 554 555 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size) 556 { 557 u8 op, result, type = (config >> 0) & 0xff; 558 const char *err = "unknown-ext-hardware-cache-type"; 559 560 if (type >= PERF_COUNT_HW_CACHE_MAX) 561 goto out_err; 562 563 op = (config >> 8) & 0xff; 564 err = "unknown-ext-hardware-cache-op"; 565 if (op >= PERF_COUNT_HW_CACHE_OP_MAX) 566 goto out_err; 567 568 result = (config >> 16) & 0xff; 569 err = "unknown-ext-hardware-cache-result"; 570 if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX) 571 goto out_err; 572 573 err = "invalid-cache"; 574 if (!perf_evsel__is_cache_op_valid(type, op)) 575 goto out_err; 576 577 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size); 578 out_err: 579 return scnprintf(bf, size, "%s", err); 580 } 581 582 static int perf_evsel__hw_cache_name(struct evsel *evsel, char *bf, size_t size) 583 { 584 int ret = __perf_evsel__hw_cache_name(evsel->core.attr.config, bf, size); 585 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 586 } 587 588 static int perf_evsel__raw_name(struct evsel *evsel, char *bf, size_t size) 589 { 590 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->core.attr.config); 591 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 592 } 593 594 static int perf_evsel__tool_name(char *bf, size_t size) 595 { 596 int ret = scnprintf(bf, size, "duration_time"); 597 return ret; 598 } 599 600 const char *perf_evsel__name(struct evsel *evsel) 601 { 602 char bf[128]; 603 604 if (!evsel) 605 goto out_unknown; 606 607 if (evsel->name) 608 return evsel->name; 609 610 switch (evsel->core.attr.type) { 611 case PERF_TYPE_RAW: 612 perf_evsel__raw_name(evsel, bf, sizeof(bf)); 613 break; 614 615 case PERF_TYPE_HARDWARE: 616 perf_evsel__hw_name(evsel, bf, sizeof(bf)); 617 break; 618 619 case PERF_TYPE_HW_CACHE: 620 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf)); 621 break; 622 623 case PERF_TYPE_SOFTWARE: 624 if (evsel->tool_event) 625 perf_evsel__tool_name(bf, sizeof(bf)); 626 else 627 perf_evsel__sw_name(evsel, bf, sizeof(bf)); 628 break; 629 630 case PERF_TYPE_TRACEPOINT: 631 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint"); 632 break; 633 634 case PERF_TYPE_BREAKPOINT: 635 perf_evsel__bp_name(evsel, bf, sizeof(bf)); 636 break; 637 638 default: 639 scnprintf(bf, sizeof(bf), "unknown attr type: %d", 640 evsel->core.attr.type); 641 break; 642 } 643 644 evsel->name = strdup(bf); 645 646 if (evsel->name) 647 return evsel->name; 648 out_unknown: 649 return "unknown"; 650 } 651 652 const char *perf_evsel__group_name(struct evsel *evsel) 653 { 654 return evsel->group_name ?: "anon group"; 655 } 656 657 /* 658 * Returns the group details for the specified leader, 659 * with following rules. 660 * 661 * For record -e '{cycles,instructions}' 662 * 'anon group { cycles:u, instructions:u }' 663 * 664 * For record -e 'cycles,instructions' and report --group 665 * 'cycles:u, instructions:u' 666 */ 667 int perf_evsel__group_desc(struct evsel *evsel, char *buf, size_t size) 668 { 669 int ret = 0; 670 struct evsel *pos; 671 const char *group_name = perf_evsel__group_name(evsel); 672 673 if (!evsel->forced_leader) 674 ret = scnprintf(buf, size, "%s { ", group_name); 675 676 ret += scnprintf(buf + ret, size - ret, "%s", 677 perf_evsel__name(evsel)); 678 679 for_each_group_member(pos, evsel) 680 ret += scnprintf(buf + ret, size - ret, ", %s", 681 perf_evsel__name(pos)); 682 683 if (!evsel->forced_leader) 684 ret += scnprintf(buf + ret, size - ret, " }"); 685 686 return ret; 687 } 688 689 static void __perf_evsel__config_callchain(struct evsel *evsel, 690 struct record_opts *opts, 691 struct callchain_param *param) 692 { 693 bool function = perf_evsel__is_function_event(evsel); 694 struct perf_event_attr *attr = &evsel->core.attr; 695 696 perf_evsel__set_sample_bit(evsel, CALLCHAIN); 697 698 attr->sample_max_stack = param->max_stack; 699 700 if (opts->kernel_callchains) 701 attr->exclude_callchain_user = 1; 702 if (opts->user_callchains) 703 attr->exclude_callchain_kernel = 1; 704 if (param->record_mode == CALLCHAIN_LBR) { 705 if (!opts->branch_stack) { 706 if (attr->exclude_user) { 707 pr_warning("LBR callstack option is only available " 708 "to get user callchain information. " 709 "Falling back to framepointers.\n"); 710 } else { 711 perf_evsel__set_sample_bit(evsel, BRANCH_STACK); 712 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER | 713 PERF_SAMPLE_BRANCH_CALL_STACK | 714 PERF_SAMPLE_BRANCH_NO_CYCLES | 715 PERF_SAMPLE_BRANCH_NO_FLAGS | 716 PERF_SAMPLE_BRANCH_HW_INDEX; 717 } 718 } else 719 pr_warning("Cannot use LBR callstack with branch stack. " 720 "Falling back to framepointers.\n"); 721 } 722 723 if (param->record_mode == CALLCHAIN_DWARF) { 724 if (!function) { 725 perf_evsel__set_sample_bit(evsel, REGS_USER); 726 perf_evsel__set_sample_bit(evsel, STACK_USER); 727 if (opts->sample_user_regs && DWARF_MINIMAL_REGS != PERF_REGS_MASK) { 728 attr->sample_regs_user |= DWARF_MINIMAL_REGS; 729 pr_warning("WARNING: The use of --call-graph=dwarf may require all the user registers, " 730 "specifying a subset with --user-regs may render DWARF unwinding unreliable, " 731 "so the minimal registers set (IP, SP) is explicitly forced.\n"); 732 } else { 733 attr->sample_regs_user |= PERF_REGS_MASK; 734 } 735 attr->sample_stack_user = param->dump_size; 736 attr->exclude_callchain_user = 1; 737 } else { 738 pr_info("Cannot use DWARF unwind for function trace event," 739 " falling back to framepointers.\n"); 740 } 741 } 742 743 if (function) { 744 pr_info("Disabling user space callchains for function trace event.\n"); 745 attr->exclude_callchain_user = 1; 746 } 747 } 748 749 void perf_evsel__config_callchain(struct evsel *evsel, 750 struct record_opts *opts, 751 struct callchain_param *param) 752 { 753 if (param->enabled) 754 return __perf_evsel__config_callchain(evsel, opts, param); 755 } 756 757 static void 758 perf_evsel__reset_callgraph(struct evsel *evsel, 759 struct callchain_param *param) 760 { 761 struct perf_event_attr *attr = &evsel->core.attr; 762 763 perf_evsel__reset_sample_bit(evsel, CALLCHAIN); 764 if (param->record_mode == CALLCHAIN_LBR) { 765 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK); 766 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER | 767 PERF_SAMPLE_BRANCH_CALL_STACK | 768 PERF_SAMPLE_BRANCH_HW_INDEX); 769 } 770 if (param->record_mode == CALLCHAIN_DWARF) { 771 perf_evsel__reset_sample_bit(evsel, REGS_USER); 772 perf_evsel__reset_sample_bit(evsel, STACK_USER); 773 } 774 } 775 776 static void apply_config_terms(struct evsel *evsel, 777 struct record_opts *opts, bool track) 778 { 779 struct perf_evsel_config_term *term; 780 struct list_head *config_terms = &evsel->config_terms; 781 struct perf_event_attr *attr = &evsel->core.attr; 782 /* callgraph default */ 783 struct callchain_param param = { 784 .record_mode = callchain_param.record_mode, 785 }; 786 u32 dump_size = 0; 787 int max_stack = 0; 788 const char *callgraph_buf = NULL; 789 790 list_for_each_entry(term, config_terms, list) { 791 switch (term->type) { 792 case PERF_EVSEL__CONFIG_TERM_PERIOD: 793 if (!(term->weak && opts->user_interval != ULLONG_MAX)) { 794 attr->sample_period = term->val.period; 795 attr->freq = 0; 796 perf_evsel__reset_sample_bit(evsel, PERIOD); 797 } 798 break; 799 case PERF_EVSEL__CONFIG_TERM_FREQ: 800 if (!(term->weak && opts->user_freq != UINT_MAX)) { 801 attr->sample_freq = term->val.freq; 802 attr->freq = 1; 803 perf_evsel__set_sample_bit(evsel, PERIOD); 804 } 805 break; 806 case PERF_EVSEL__CONFIG_TERM_TIME: 807 if (term->val.time) 808 perf_evsel__set_sample_bit(evsel, TIME); 809 else 810 perf_evsel__reset_sample_bit(evsel, TIME); 811 break; 812 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH: 813 callgraph_buf = term->val.str; 814 break; 815 case PERF_EVSEL__CONFIG_TERM_BRANCH: 816 if (term->val.str && strcmp(term->val.str, "no")) { 817 perf_evsel__set_sample_bit(evsel, BRANCH_STACK); 818 parse_branch_str(term->val.str, 819 &attr->branch_sample_type); 820 } else 821 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK); 822 break; 823 case PERF_EVSEL__CONFIG_TERM_STACK_USER: 824 dump_size = term->val.stack_user; 825 break; 826 case PERF_EVSEL__CONFIG_TERM_MAX_STACK: 827 max_stack = term->val.max_stack; 828 break; 829 case PERF_EVSEL__CONFIG_TERM_MAX_EVENTS: 830 evsel->max_events = term->val.max_events; 831 break; 832 case PERF_EVSEL__CONFIG_TERM_INHERIT: 833 /* 834 * attr->inherit should has already been set by 835 * perf_evsel__config. If user explicitly set 836 * inherit using config terms, override global 837 * opt->no_inherit setting. 838 */ 839 attr->inherit = term->val.inherit ? 1 : 0; 840 break; 841 case PERF_EVSEL__CONFIG_TERM_OVERWRITE: 842 attr->write_backward = term->val.overwrite ? 1 : 0; 843 break; 844 case PERF_EVSEL__CONFIG_TERM_DRV_CFG: 845 break; 846 case PERF_EVSEL__CONFIG_TERM_PERCORE: 847 break; 848 case PERF_EVSEL__CONFIG_TERM_AUX_OUTPUT: 849 attr->aux_output = term->val.aux_output ? 1 : 0; 850 break; 851 case PERF_EVSEL__CONFIG_TERM_AUX_SAMPLE_SIZE: 852 /* Already applied by auxtrace */ 853 break; 854 case PERF_EVSEL__CONFIG_TERM_CFG_CHG: 855 break; 856 default: 857 break; 858 } 859 } 860 861 /* User explicitly set per-event callgraph, clear the old setting and reset. */ 862 if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) { 863 bool sample_address = false; 864 865 if (max_stack) { 866 param.max_stack = max_stack; 867 if (callgraph_buf == NULL) 868 callgraph_buf = "fp"; 869 } 870 871 /* parse callgraph parameters */ 872 if (callgraph_buf != NULL) { 873 if (!strcmp(callgraph_buf, "no")) { 874 param.enabled = false; 875 param.record_mode = CALLCHAIN_NONE; 876 } else { 877 param.enabled = true; 878 if (parse_callchain_record(callgraph_buf, ¶m)) { 879 pr_err("per-event callgraph setting for %s failed. " 880 "Apply callgraph global setting for it\n", 881 evsel->name); 882 return; 883 } 884 if (param.record_mode == CALLCHAIN_DWARF) 885 sample_address = true; 886 } 887 } 888 if (dump_size > 0) { 889 dump_size = round_up(dump_size, sizeof(u64)); 890 param.dump_size = dump_size; 891 } 892 893 /* If global callgraph set, clear it */ 894 if (callchain_param.enabled) 895 perf_evsel__reset_callgraph(evsel, &callchain_param); 896 897 /* set perf-event callgraph */ 898 if (param.enabled) { 899 if (sample_address) { 900 perf_evsel__set_sample_bit(evsel, ADDR); 901 perf_evsel__set_sample_bit(evsel, DATA_SRC); 902 evsel->core.attr.mmap_data = track; 903 } 904 perf_evsel__config_callchain(evsel, opts, ¶m); 905 } 906 } 907 } 908 909 static bool is_dummy_event(struct evsel *evsel) 910 { 911 return (evsel->core.attr.type == PERF_TYPE_SOFTWARE) && 912 (evsel->core.attr.config == PERF_COUNT_SW_DUMMY); 913 } 914 915 struct perf_evsel_config_term *__perf_evsel__get_config_term(struct evsel *evsel, 916 enum evsel_term_type type) 917 { 918 struct perf_evsel_config_term *term, *found_term = NULL; 919 920 list_for_each_entry(term, &evsel->config_terms, list) { 921 if (term->type == type) 922 found_term = term; 923 } 924 925 return found_term; 926 } 927 928 /* 929 * The enable_on_exec/disabled value strategy: 930 * 931 * 1) For any type of traced program: 932 * - all independent events and group leaders are disabled 933 * - all group members are enabled 934 * 935 * Group members are ruled by group leaders. They need to 936 * be enabled, because the group scheduling relies on that. 937 * 938 * 2) For traced programs executed by perf: 939 * - all independent events and group leaders have 940 * enable_on_exec set 941 * - we don't specifically enable or disable any event during 942 * the record command 943 * 944 * Independent events and group leaders are initially disabled 945 * and get enabled by exec. Group members are ruled by group 946 * leaders as stated in 1). 947 * 948 * 3) For traced programs attached by perf (pid/tid): 949 * - we specifically enable or disable all events during 950 * the record command 951 * 952 * When attaching events to already running traced we 953 * enable/disable events specifically, as there's no 954 * initial traced exec call. 955 */ 956 void perf_evsel__config(struct evsel *evsel, struct record_opts *opts, 957 struct callchain_param *callchain) 958 { 959 struct evsel *leader = evsel->leader; 960 struct perf_event_attr *attr = &evsel->core.attr; 961 int track = evsel->tracking; 962 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread; 963 964 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1; 965 attr->inherit = !opts->no_inherit; 966 attr->write_backward = opts->overwrite ? 1 : 0; 967 968 perf_evsel__set_sample_bit(evsel, IP); 969 perf_evsel__set_sample_bit(evsel, TID); 970 971 if (evsel->sample_read) { 972 perf_evsel__set_sample_bit(evsel, READ); 973 974 /* 975 * We need ID even in case of single event, because 976 * PERF_SAMPLE_READ process ID specific data. 977 */ 978 perf_evsel__set_sample_id(evsel, false); 979 980 /* 981 * Apply group format only if we belong to group 982 * with more than one members. 983 */ 984 if (leader->core.nr_members > 1) { 985 attr->read_format |= PERF_FORMAT_GROUP; 986 attr->inherit = 0; 987 } 988 } 989 990 /* 991 * We default some events to have a default interval. But keep 992 * it a weak assumption overridable by the user. 993 */ 994 if (!attr->sample_period || (opts->user_freq != UINT_MAX || 995 opts->user_interval != ULLONG_MAX)) { 996 if (opts->freq) { 997 perf_evsel__set_sample_bit(evsel, PERIOD); 998 attr->freq = 1; 999 attr->sample_freq = opts->freq; 1000 } else { 1001 attr->sample_period = opts->default_interval; 1002 } 1003 } 1004 1005 /* 1006 * Disable sampling for all group members other 1007 * than leader in case leader 'leads' the sampling. 1008 */ 1009 if ((leader != evsel) && leader->sample_read) { 1010 attr->freq = 0; 1011 attr->sample_freq = 0; 1012 attr->sample_period = 0; 1013 attr->write_backward = 0; 1014 1015 /* 1016 * We don't get sample for slave events, we make them 1017 * when delivering group leader sample. Set the slave 1018 * event to follow the master sample_type to ease up 1019 * report. 1020 */ 1021 attr->sample_type = leader->core.attr.sample_type; 1022 } 1023 1024 if (opts->no_samples) 1025 attr->sample_freq = 0; 1026 1027 if (opts->inherit_stat) { 1028 evsel->core.attr.read_format |= 1029 PERF_FORMAT_TOTAL_TIME_ENABLED | 1030 PERF_FORMAT_TOTAL_TIME_RUNNING | 1031 PERF_FORMAT_ID; 1032 attr->inherit_stat = 1; 1033 } 1034 1035 if (opts->sample_address) { 1036 perf_evsel__set_sample_bit(evsel, ADDR); 1037 attr->mmap_data = track; 1038 } 1039 1040 /* 1041 * We don't allow user space callchains for function trace 1042 * event, due to issues with page faults while tracing page 1043 * fault handler and its overall trickiness nature. 1044 */ 1045 if (perf_evsel__is_function_event(evsel)) 1046 evsel->core.attr.exclude_callchain_user = 1; 1047 1048 if (callchain && callchain->enabled && !evsel->no_aux_samples) 1049 perf_evsel__config_callchain(evsel, opts, callchain); 1050 1051 if (opts->sample_intr_regs) { 1052 attr->sample_regs_intr = opts->sample_intr_regs; 1053 perf_evsel__set_sample_bit(evsel, REGS_INTR); 1054 } 1055 1056 if (opts->sample_user_regs) { 1057 attr->sample_regs_user |= opts->sample_user_regs; 1058 perf_evsel__set_sample_bit(evsel, REGS_USER); 1059 } 1060 1061 if (target__has_cpu(&opts->target) || opts->sample_cpu) 1062 perf_evsel__set_sample_bit(evsel, CPU); 1063 1064 /* 1065 * When the user explicitly disabled time don't force it here. 1066 */ 1067 if (opts->sample_time && 1068 (!perf_missing_features.sample_id_all && 1069 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu || 1070 opts->sample_time_set))) 1071 perf_evsel__set_sample_bit(evsel, TIME); 1072 1073 if (opts->raw_samples && !evsel->no_aux_samples) { 1074 perf_evsel__set_sample_bit(evsel, TIME); 1075 perf_evsel__set_sample_bit(evsel, RAW); 1076 perf_evsel__set_sample_bit(evsel, CPU); 1077 } 1078 1079 if (opts->sample_address) 1080 perf_evsel__set_sample_bit(evsel, DATA_SRC); 1081 1082 if (opts->sample_phys_addr) 1083 perf_evsel__set_sample_bit(evsel, PHYS_ADDR); 1084 1085 if (opts->no_buffering) { 1086 attr->watermark = 0; 1087 attr->wakeup_events = 1; 1088 } 1089 if (opts->branch_stack && !evsel->no_aux_samples) { 1090 perf_evsel__set_sample_bit(evsel, BRANCH_STACK); 1091 attr->branch_sample_type = opts->branch_stack; 1092 } 1093 1094 if (opts->sample_weight) 1095 perf_evsel__set_sample_bit(evsel, WEIGHT); 1096 1097 attr->task = track; 1098 attr->mmap = track; 1099 attr->mmap2 = track && !perf_missing_features.mmap2; 1100 attr->comm = track; 1101 attr->ksymbol = track && !perf_missing_features.ksymbol; 1102 attr->bpf_event = track && !opts->no_bpf_event && !perf_missing_features.bpf; 1103 1104 if (opts->record_namespaces) 1105 attr->namespaces = track; 1106 1107 if (opts->record_switch_events) 1108 attr->context_switch = track; 1109 1110 if (opts->sample_transaction) 1111 perf_evsel__set_sample_bit(evsel, TRANSACTION); 1112 1113 if (opts->running_time) { 1114 evsel->core.attr.read_format |= 1115 PERF_FORMAT_TOTAL_TIME_ENABLED | 1116 PERF_FORMAT_TOTAL_TIME_RUNNING; 1117 } 1118 1119 /* 1120 * XXX see the function comment above 1121 * 1122 * Disabling only independent events or group leaders, 1123 * keeping group members enabled. 1124 */ 1125 if (perf_evsel__is_group_leader(evsel)) 1126 attr->disabled = 1; 1127 1128 /* 1129 * Setting enable_on_exec for independent events and 1130 * group leaders for traced executed by perf. 1131 */ 1132 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) && 1133 !opts->initial_delay) 1134 attr->enable_on_exec = 1; 1135 1136 if (evsel->immediate) { 1137 attr->disabled = 0; 1138 attr->enable_on_exec = 0; 1139 } 1140 1141 clockid = opts->clockid; 1142 if (opts->use_clockid) { 1143 attr->use_clockid = 1; 1144 attr->clockid = opts->clockid; 1145 } 1146 1147 if (evsel->precise_max) 1148 attr->precise_ip = 3; 1149 1150 if (opts->all_user) { 1151 attr->exclude_kernel = 1; 1152 attr->exclude_user = 0; 1153 } 1154 1155 if (opts->all_kernel) { 1156 attr->exclude_kernel = 0; 1157 attr->exclude_user = 1; 1158 } 1159 1160 if (evsel->core.own_cpus || evsel->unit) 1161 evsel->core.attr.read_format |= PERF_FORMAT_ID; 1162 1163 /* 1164 * Apply event specific term settings, 1165 * it overloads any global configuration. 1166 */ 1167 apply_config_terms(evsel, opts, track); 1168 1169 evsel->ignore_missing_thread = opts->ignore_missing_thread; 1170 1171 /* The --period option takes the precedence. */ 1172 if (opts->period_set) { 1173 if (opts->period) 1174 perf_evsel__set_sample_bit(evsel, PERIOD); 1175 else 1176 perf_evsel__reset_sample_bit(evsel, PERIOD); 1177 } 1178 1179 /* 1180 * For initial_delay, a dummy event is added implicitly. 1181 * The software event will trigger -EOPNOTSUPP error out, 1182 * if BRANCH_STACK bit is set. 1183 */ 1184 if (opts->initial_delay && is_dummy_event(evsel)) 1185 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK); 1186 } 1187 1188 int perf_evsel__set_filter(struct evsel *evsel, const char *filter) 1189 { 1190 char *new_filter = strdup(filter); 1191 1192 if (new_filter != NULL) { 1193 free(evsel->filter); 1194 evsel->filter = new_filter; 1195 return 0; 1196 } 1197 1198 return -1; 1199 } 1200 1201 static int perf_evsel__append_filter(struct evsel *evsel, 1202 const char *fmt, const char *filter) 1203 { 1204 char *new_filter; 1205 1206 if (evsel->filter == NULL) 1207 return perf_evsel__set_filter(evsel, filter); 1208 1209 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) { 1210 free(evsel->filter); 1211 evsel->filter = new_filter; 1212 return 0; 1213 } 1214 1215 return -1; 1216 } 1217 1218 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter) 1219 { 1220 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter); 1221 } 1222 1223 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter) 1224 { 1225 return perf_evsel__append_filter(evsel, "%s,%s", filter); 1226 } 1227 1228 /* Caller has to clear disabled after going through all CPUs. */ 1229 int evsel__enable_cpu(struct evsel *evsel, int cpu) 1230 { 1231 return perf_evsel__enable_cpu(&evsel->core, cpu); 1232 } 1233 1234 int evsel__enable(struct evsel *evsel) 1235 { 1236 int err = perf_evsel__enable(&evsel->core); 1237 1238 if (!err) 1239 evsel->disabled = false; 1240 return err; 1241 } 1242 1243 /* Caller has to set disabled after going through all CPUs. */ 1244 int evsel__disable_cpu(struct evsel *evsel, int cpu) 1245 { 1246 return perf_evsel__disable_cpu(&evsel->core, cpu); 1247 } 1248 1249 int evsel__disable(struct evsel *evsel) 1250 { 1251 int err = perf_evsel__disable(&evsel->core); 1252 /* 1253 * We mark it disabled here so that tools that disable a event can 1254 * ignore events after they disable it. I.e. the ring buffer may have 1255 * already a few more events queued up before the kernel got the stop 1256 * request. 1257 */ 1258 if (!err) 1259 evsel->disabled = true; 1260 1261 return err; 1262 } 1263 1264 static void perf_evsel__free_config_terms(struct evsel *evsel) 1265 { 1266 struct perf_evsel_config_term *term, *h; 1267 1268 list_for_each_entry_safe(term, h, &evsel->config_terms, list) { 1269 list_del_init(&term->list); 1270 if (term->free_str) 1271 zfree(&term->val.str); 1272 free(term); 1273 } 1274 } 1275 1276 void perf_evsel__exit(struct evsel *evsel) 1277 { 1278 assert(list_empty(&evsel->core.node)); 1279 assert(evsel->evlist == NULL); 1280 perf_evsel__free_counts(evsel); 1281 perf_evsel__free_fd(&evsel->core); 1282 perf_evsel__free_id(&evsel->core); 1283 perf_evsel__free_config_terms(evsel); 1284 cgroup__put(evsel->cgrp); 1285 perf_cpu_map__put(evsel->core.cpus); 1286 perf_cpu_map__put(evsel->core.own_cpus); 1287 perf_thread_map__put(evsel->core.threads); 1288 zfree(&evsel->group_name); 1289 zfree(&evsel->name); 1290 perf_evsel__object.fini(evsel); 1291 } 1292 1293 void evsel__delete(struct evsel *evsel) 1294 { 1295 perf_evsel__exit(evsel); 1296 free(evsel); 1297 } 1298 1299 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread, 1300 struct perf_counts_values *count) 1301 { 1302 struct perf_counts_values tmp; 1303 1304 if (!evsel->prev_raw_counts) 1305 return; 1306 1307 if (cpu == -1) { 1308 tmp = evsel->prev_raw_counts->aggr; 1309 evsel->prev_raw_counts->aggr = *count; 1310 } else { 1311 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread); 1312 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count; 1313 } 1314 1315 count->val = count->val - tmp.val; 1316 count->ena = count->ena - tmp.ena; 1317 count->run = count->run - tmp.run; 1318 } 1319 1320 void perf_counts_values__scale(struct perf_counts_values *count, 1321 bool scale, s8 *pscaled) 1322 { 1323 s8 scaled = 0; 1324 1325 if (scale) { 1326 if (count->run == 0) { 1327 scaled = -1; 1328 count->val = 0; 1329 } else if (count->run < count->ena) { 1330 scaled = 1; 1331 count->val = (u64)((double) count->val * count->ena / count->run); 1332 } 1333 } 1334 1335 if (pscaled) 1336 *pscaled = scaled; 1337 } 1338 1339 static int 1340 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread) 1341 { 1342 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread); 1343 1344 return perf_evsel__read(&evsel->core, cpu, thread, count); 1345 } 1346 1347 static void 1348 perf_evsel__set_count(struct evsel *counter, int cpu, int thread, 1349 u64 val, u64 ena, u64 run) 1350 { 1351 struct perf_counts_values *count; 1352 1353 count = perf_counts(counter->counts, cpu, thread); 1354 1355 count->val = val; 1356 count->ena = ena; 1357 count->run = run; 1358 1359 perf_counts__set_loaded(counter->counts, cpu, thread, true); 1360 } 1361 1362 static int 1363 perf_evsel__process_group_data(struct evsel *leader, 1364 int cpu, int thread, u64 *data) 1365 { 1366 u64 read_format = leader->core.attr.read_format; 1367 struct sample_read_value *v; 1368 u64 nr, ena = 0, run = 0, i; 1369 1370 nr = *data++; 1371 1372 if (nr != (u64) leader->core.nr_members) 1373 return -EINVAL; 1374 1375 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1376 ena = *data++; 1377 1378 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1379 run = *data++; 1380 1381 v = (struct sample_read_value *) data; 1382 1383 perf_evsel__set_count(leader, cpu, thread, 1384 v[0].value, ena, run); 1385 1386 for (i = 1; i < nr; i++) { 1387 struct evsel *counter; 1388 1389 counter = perf_evlist__id2evsel(leader->evlist, v[i].id); 1390 if (!counter) 1391 return -EINVAL; 1392 1393 perf_evsel__set_count(counter, cpu, thread, 1394 v[i].value, ena, run); 1395 } 1396 1397 return 0; 1398 } 1399 1400 static int 1401 perf_evsel__read_group(struct evsel *leader, int cpu, int thread) 1402 { 1403 struct perf_stat_evsel *ps = leader->stats; 1404 u64 read_format = leader->core.attr.read_format; 1405 int size = perf_evsel__read_size(&leader->core); 1406 u64 *data = ps->group_data; 1407 1408 if (!(read_format & PERF_FORMAT_ID)) 1409 return -EINVAL; 1410 1411 if (!perf_evsel__is_group_leader(leader)) 1412 return -EINVAL; 1413 1414 if (!data) { 1415 data = zalloc(size); 1416 if (!data) 1417 return -ENOMEM; 1418 1419 ps->group_data = data; 1420 } 1421 1422 if (FD(leader, cpu, thread) < 0) 1423 return -EINVAL; 1424 1425 if (readn(FD(leader, cpu, thread), data, size) <= 0) 1426 return -errno; 1427 1428 return perf_evsel__process_group_data(leader, cpu, thread, data); 1429 } 1430 1431 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread) 1432 { 1433 u64 read_format = evsel->core.attr.read_format; 1434 1435 if (read_format & PERF_FORMAT_GROUP) 1436 return perf_evsel__read_group(evsel, cpu, thread); 1437 else 1438 return perf_evsel__read_one(evsel, cpu, thread); 1439 } 1440 1441 int __perf_evsel__read_on_cpu(struct evsel *evsel, 1442 int cpu, int thread, bool scale) 1443 { 1444 struct perf_counts_values count; 1445 size_t nv = scale ? 3 : 1; 1446 1447 if (FD(evsel, cpu, thread) < 0) 1448 return -EINVAL; 1449 1450 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0) 1451 return -ENOMEM; 1452 1453 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0) 1454 return -errno; 1455 1456 perf_evsel__compute_deltas(evsel, cpu, thread, &count); 1457 perf_counts_values__scale(&count, scale, NULL); 1458 *perf_counts(evsel->counts, cpu, thread) = count; 1459 return 0; 1460 } 1461 1462 static int get_group_fd(struct evsel *evsel, int cpu, int thread) 1463 { 1464 struct evsel *leader = evsel->leader; 1465 int fd; 1466 1467 if (perf_evsel__is_group_leader(evsel)) 1468 return -1; 1469 1470 /* 1471 * Leader must be already processed/open, 1472 * if not it's a bug. 1473 */ 1474 BUG_ON(!leader->core.fd); 1475 1476 fd = FD(leader, cpu, thread); 1477 BUG_ON(fd == -1); 1478 1479 return fd; 1480 } 1481 1482 static void perf_evsel__remove_fd(struct evsel *pos, 1483 int nr_cpus, int nr_threads, 1484 int thread_idx) 1485 { 1486 for (int cpu = 0; cpu < nr_cpus; cpu++) 1487 for (int thread = thread_idx; thread < nr_threads - 1; thread++) 1488 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1); 1489 } 1490 1491 static int update_fds(struct evsel *evsel, 1492 int nr_cpus, int cpu_idx, 1493 int nr_threads, int thread_idx) 1494 { 1495 struct evsel *pos; 1496 1497 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads) 1498 return -EINVAL; 1499 1500 evlist__for_each_entry(evsel->evlist, pos) { 1501 nr_cpus = pos != evsel ? nr_cpus : cpu_idx; 1502 1503 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx); 1504 1505 /* 1506 * Since fds for next evsel has not been created, 1507 * there is no need to iterate whole event list. 1508 */ 1509 if (pos == evsel) 1510 break; 1511 } 1512 return 0; 1513 } 1514 1515 static bool ignore_missing_thread(struct evsel *evsel, 1516 int nr_cpus, int cpu, 1517 struct perf_thread_map *threads, 1518 int thread, int err) 1519 { 1520 pid_t ignore_pid = perf_thread_map__pid(threads, thread); 1521 1522 if (!evsel->ignore_missing_thread) 1523 return false; 1524 1525 /* The system wide setup does not work with threads. */ 1526 if (evsel->core.system_wide) 1527 return false; 1528 1529 /* The -ESRCH is perf event syscall errno for pid's not found. */ 1530 if (err != -ESRCH) 1531 return false; 1532 1533 /* If there's only one thread, let it fail. */ 1534 if (threads->nr == 1) 1535 return false; 1536 1537 /* 1538 * We should remove fd for missing_thread first 1539 * because thread_map__remove() will decrease threads->nr. 1540 */ 1541 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread)) 1542 return false; 1543 1544 if (thread_map__remove(threads, thread)) 1545 return false; 1546 1547 pr_warning("WARNING: Ignored open failure for pid %d\n", 1548 ignore_pid); 1549 return true; 1550 } 1551 1552 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val, 1553 void *priv __maybe_unused) 1554 { 1555 return fprintf(fp, " %-32s %s\n", name, val); 1556 } 1557 1558 static void display_attr(struct perf_event_attr *attr) 1559 { 1560 if (verbose >= 2 || debug_peo_args) { 1561 fprintf(stderr, "%.60s\n", graph_dotted_line); 1562 fprintf(stderr, "perf_event_attr:\n"); 1563 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL); 1564 fprintf(stderr, "%.60s\n", graph_dotted_line); 1565 } 1566 } 1567 1568 static int perf_event_open(struct evsel *evsel, 1569 pid_t pid, int cpu, int group_fd, 1570 unsigned long flags) 1571 { 1572 int precise_ip = evsel->core.attr.precise_ip; 1573 int fd; 1574 1575 while (1) { 1576 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx", 1577 pid, cpu, group_fd, flags); 1578 1579 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags); 1580 if (fd >= 0) 1581 break; 1582 1583 /* Do not try less precise if not requested. */ 1584 if (!evsel->precise_max) 1585 break; 1586 1587 /* 1588 * We tried all the precise_ip values, and it's 1589 * still failing, so leave it to standard fallback. 1590 */ 1591 if (!evsel->core.attr.precise_ip) { 1592 evsel->core.attr.precise_ip = precise_ip; 1593 break; 1594 } 1595 1596 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP); 1597 evsel->core.attr.precise_ip--; 1598 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip); 1599 display_attr(&evsel->core.attr); 1600 } 1601 1602 return fd; 1603 } 1604 1605 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, 1606 struct perf_thread_map *threads, 1607 int start_cpu, int end_cpu) 1608 { 1609 int cpu, thread, nthreads; 1610 unsigned long flags = PERF_FLAG_FD_CLOEXEC; 1611 int pid = -1, err, old_errno; 1612 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE; 1613 1614 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) || 1615 (perf_missing_features.aux_output && evsel->core.attr.aux_output)) 1616 return -EINVAL; 1617 1618 if (cpus == NULL) { 1619 static struct perf_cpu_map *empty_cpu_map; 1620 1621 if (empty_cpu_map == NULL) { 1622 empty_cpu_map = perf_cpu_map__dummy_new(); 1623 if (empty_cpu_map == NULL) 1624 return -ENOMEM; 1625 } 1626 1627 cpus = empty_cpu_map; 1628 } 1629 1630 if (threads == NULL) { 1631 static struct perf_thread_map *empty_thread_map; 1632 1633 if (empty_thread_map == NULL) { 1634 empty_thread_map = thread_map__new_by_tid(-1); 1635 if (empty_thread_map == NULL) 1636 return -ENOMEM; 1637 } 1638 1639 threads = empty_thread_map; 1640 } 1641 1642 if (evsel->core.system_wide) 1643 nthreads = 1; 1644 else 1645 nthreads = threads->nr; 1646 1647 if (evsel->core.fd == NULL && 1648 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0) 1649 return -ENOMEM; 1650 1651 if (evsel->cgrp) { 1652 flags |= PERF_FLAG_PID_CGROUP; 1653 pid = evsel->cgrp->fd; 1654 } 1655 1656 fallback_missing_features: 1657 if (perf_missing_features.clockid_wrong) 1658 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */ 1659 if (perf_missing_features.clockid) { 1660 evsel->core.attr.use_clockid = 0; 1661 evsel->core.attr.clockid = 0; 1662 } 1663 if (perf_missing_features.cloexec) 1664 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 1665 if (perf_missing_features.mmap2) 1666 evsel->core.attr.mmap2 = 0; 1667 if (perf_missing_features.exclude_guest) 1668 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0; 1669 if (perf_missing_features.lbr_flags) 1670 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS | 1671 PERF_SAMPLE_BRANCH_NO_CYCLES); 1672 if (perf_missing_features.group_read && evsel->core.attr.inherit) 1673 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID); 1674 if (perf_missing_features.ksymbol) 1675 evsel->core.attr.ksymbol = 0; 1676 if (perf_missing_features.bpf) 1677 evsel->core.attr.bpf_event = 0; 1678 if (perf_missing_features.branch_hw_idx) 1679 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX; 1680 retry_sample_id: 1681 if (perf_missing_features.sample_id_all) 1682 evsel->core.attr.sample_id_all = 0; 1683 1684 display_attr(&evsel->core.attr); 1685 1686 for (cpu = start_cpu; cpu < end_cpu; cpu++) { 1687 1688 for (thread = 0; thread < nthreads; thread++) { 1689 int fd, group_fd; 1690 1691 if (!evsel->cgrp && !evsel->core.system_wide) 1692 pid = perf_thread_map__pid(threads, thread); 1693 1694 group_fd = get_group_fd(evsel, cpu, thread); 1695 retry_open: 1696 test_attr__ready(); 1697 1698 fd = perf_event_open(evsel, pid, cpus->map[cpu], 1699 group_fd, flags); 1700 1701 FD(evsel, cpu, thread) = fd; 1702 1703 if (fd < 0) { 1704 err = -errno; 1705 1706 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) { 1707 /* 1708 * We just removed 1 thread, so take a step 1709 * back on thread index and lower the upper 1710 * nthreads limit. 1711 */ 1712 nthreads--; 1713 thread--; 1714 1715 /* ... and pretend like nothing have happened. */ 1716 err = 0; 1717 continue; 1718 } 1719 1720 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", 1721 err); 1722 goto try_fallback; 1723 } 1724 1725 pr_debug2_peo(" = %d\n", fd); 1726 1727 if (evsel->bpf_fd >= 0) { 1728 int evt_fd = fd; 1729 int bpf_fd = evsel->bpf_fd; 1730 1731 err = ioctl(evt_fd, 1732 PERF_EVENT_IOC_SET_BPF, 1733 bpf_fd); 1734 if (err && errno != EEXIST) { 1735 pr_err("failed to attach bpf fd %d: %s\n", 1736 bpf_fd, strerror(errno)); 1737 err = -EINVAL; 1738 goto out_close; 1739 } 1740 } 1741 1742 set_rlimit = NO_CHANGE; 1743 1744 /* 1745 * If we succeeded but had to kill clockid, fail and 1746 * have perf_evsel__open_strerror() print us a nice 1747 * error. 1748 */ 1749 if (perf_missing_features.clockid || 1750 perf_missing_features.clockid_wrong) { 1751 err = -EINVAL; 1752 goto out_close; 1753 } 1754 } 1755 } 1756 1757 return 0; 1758 1759 try_fallback: 1760 /* 1761 * perf stat needs between 5 and 22 fds per CPU. When we run out 1762 * of them try to increase the limits. 1763 */ 1764 if (err == -EMFILE && set_rlimit < INCREASED_MAX) { 1765 struct rlimit l; 1766 1767 old_errno = errno; 1768 if (getrlimit(RLIMIT_NOFILE, &l) == 0) { 1769 if (set_rlimit == NO_CHANGE) 1770 l.rlim_cur = l.rlim_max; 1771 else { 1772 l.rlim_cur = l.rlim_max + 1000; 1773 l.rlim_max = l.rlim_cur; 1774 } 1775 if (setrlimit(RLIMIT_NOFILE, &l) == 0) { 1776 set_rlimit++; 1777 errno = old_errno; 1778 goto retry_open; 1779 } 1780 } 1781 errno = old_errno; 1782 } 1783 1784 if (err != -EINVAL || cpu > 0 || thread > 0) 1785 goto out_close; 1786 1787 /* 1788 * Must probe features in the order they were added to the 1789 * perf_event_attr interface. 1790 */ 1791 if (!perf_missing_features.branch_hw_idx && 1792 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) { 1793 perf_missing_features.branch_hw_idx = true; 1794 pr_debug2("switching off branch HW index support\n"); 1795 goto fallback_missing_features; 1796 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) { 1797 perf_missing_features.aux_output = true; 1798 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n"); 1799 goto out_close; 1800 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) { 1801 perf_missing_features.bpf = true; 1802 pr_debug2_peo("switching off bpf_event\n"); 1803 goto fallback_missing_features; 1804 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) { 1805 perf_missing_features.ksymbol = true; 1806 pr_debug2_peo("switching off ksymbol\n"); 1807 goto fallback_missing_features; 1808 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) { 1809 perf_missing_features.write_backward = true; 1810 pr_debug2_peo("switching off write_backward\n"); 1811 goto out_close; 1812 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) { 1813 perf_missing_features.clockid_wrong = true; 1814 pr_debug2_peo("switching off clockid\n"); 1815 goto fallback_missing_features; 1816 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) { 1817 perf_missing_features.clockid = true; 1818 pr_debug2_peo("switching off use_clockid\n"); 1819 goto fallback_missing_features; 1820 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) { 1821 perf_missing_features.cloexec = true; 1822 pr_debug2_peo("switching off cloexec flag\n"); 1823 goto fallback_missing_features; 1824 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) { 1825 perf_missing_features.mmap2 = true; 1826 pr_debug2_peo("switching off mmap2\n"); 1827 goto fallback_missing_features; 1828 } else if (!perf_missing_features.exclude_guest && 1829 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) { 1830 perf_missing_features.exclude_guest = true; 1831 pr_debug2_peo("switching off exclude_guest, exclude_host\n"); 1832 goto fallback_missing_features; 1833 } else if (!perf_missing_features.sample_id_all) { 1834 perf_missing_features.sample_id_all = true; 1835 pr_debug2_peo("switching off sample_id_all\n"); 1836 goto retry_sample_id; 1837 } else if (!perf_missing_features.lbr_flags && 1838 (evsel->core.attr.branch_sample_type & 1839 (PERF_SAMPLE_BRANCH_NO_CYCLES | 1840 PERF_SAMPLE_BRANCH_NO_FLAGS))) { 1841 perf_missing_features.lbr_flags = true; 1842 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n"); 1843 goto fallback_missing_features; 1844 } else if (!perf_missing_features.group_read && 1845 evsel->core.attr.inherit && 1846 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) && 1847 perf_evsel__is_group_leader(evsel)) { 1848 perf_missing_features.group_read = true; 1849 pr_debug2_peo("switching off group read\n"); 1850 goto fallback_missing_features; 1851 } 1852 out_close: 1853 if (err) 1854 threads->err_thread = thread; 1855 1856 old_errno = errno; 1857 do { 1858 while (--thread >= 0) { 1859 if (FD(evsel, cpu, thread) >= 0) 1860 close(FD(evsel, cpu, thread)); 1861 FD(evsel, cpu, thread) = -1; 1862 } 1863 thread = nthreads; 1864 } while (--cpu >= 0); 1865 errno = old_errno; 1866 return err; 1867 } 1868 1869 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus, 1870 struct perf_thread_map *threads) 1871 { 1872 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1); 1873 } 1874 1875 void evsel__close(struct evsel *evsel) 1876 { 1877 perf_evsel__close(&evsel->core); 1878 perf_evsel__free_id(&evsel->core); 1879 } 1880 1881 int perf_evsel__open_per_cpu(struct evsel *evsel, 1882 struct perf_cpu_map *cpus, 1883 int cpu) 1884 { 1885 if (cpu == -1) 1886 return evsel__open_cpu(evsel, cpus, NULL, 0, 1887 cpus ? cpus->nr : 1); 1888 1889 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1); 1890 } 1891 1892 int perf_evsel__open_per_thread(struct evsel *evsel, 1893 struct perf_thread_map *threads) 1894 { 1895 return evsel__open(evsel, NULL, threads); 1896 } 1897 1898 static int perf_evsel__parse_id_sample(const struct evsel *evsel, 1899 const union perf_event *event, 1900 struct perf_sample *sample) 1901 { 1902 u64 type = evsel->core.attr.sample_type; 1903 const __u64 *array = event->sample.array; 1904 bool swapped = evsel->needs_swap; 1905 union u64_swap u; 1906 1907 array += ((event->header.size - 1908 sizeof(event->header)) / sizeof(u64)) - 1; 1909 1910 if (type & PERF_SAMPLE_IDENTIFIER) { 1911 sample->id = *array; 1912 array--; 1913 } 1914 1915 if (type & PERF_SAMPLE_CPU) { 1916 u.val64 = *array; 1917 if (swapped) { 1918 /* undo swap of u64, then swap on individual u32s */ 1919 u.val64 = bswap_64(u.val64); 1920 u.val32[0] = bswap_32(u.val32[0]); 1921 } 1922 1923 sample->cpu = u.val32[0]; 1924 array--; 1925 } 1926 1927 if (type & PERF_SAMPLE_STREAM_ID) { 1928 sample->stream_id = *array; 1929 array--; 1930 } 1931 1932 if (type & PERF_SAMPLE_ID) { 1933 sample->id = *array; 1934 array--; 1935 } 1936 1937 if (type & PERF_SAMPLE_TIME) { 1938 sample->time = *array; 1939 array--; 1940 } 1941 1942 if (type & PERF_SAMPLE_TID) { 1943 u.val64 = *array; 1944 if (swapped) { 1945 /* undo swap of u64, then swap on individual u32s */ 1946 u.val64 = bswap_64(u.val64); 1947 u.val32[0] = bswap_32(u.val32[0]); 1948 u.val32[1] = bswap_32(u.val32[1]); 1949 } 1950 1951 sample->pid = u.val32[0]; 1952 sample->tid = u.val32[1]; 1953 array--; 1954 } 1955 1956 return 0; 1957 } 1958 1959 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 1960 u64 size) 1961 { 1962 return size > max_size || offset + size > endp; 1963 } 1964 1965 #define OVERFLOW_CHECK(offset, size, max_size) \ 1966 do { \ 1967 if (overflow(endp, (max_size), (offset), (size))) \ 1968 return -EFAULT; \ 1969 } while (0) 1970 1971 #define OVERFLOW_CHECK_u64(offset) \ 1972 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 1973 1974 static int 1975 perf_event__check_size(union perf_event *event, unsigned int sample_size) 1976 { 1977 /* 1978 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 1979 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 1980 * check the format does not go past the end of the event. 1981 */ 1982 if (sample_size + sizeof(event->header) > event->header.size) 1983 return -EFAULT; 1984 1985 return 0; 1986 } 1987 1988 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event, 1989 struct perf_sample *data) 1990 { 1991 u64 type = evsel->core.attr.sample_type; 1992 bool swapped = evsel->needs_swap; 1993 const __u64 *array; 1994 u16 max_size = event->header.size; 1995 const void *endp = (void *)event + max_size; 1996 u64 sz; 1997 1998 /* 1999 * used for cross-endian analysis. See git commit 65014ab3 2000 * for why this goofiness is needed. 2001 */ 2002 union u64_swap u; 2003 2004 memset(data, 0, sizeof(*data)); 2005 data->cpu = data->pid = data->tid = -1; 2006 data->stream_id = data->id = data->time = -1ULL; 2007 data->period = evsel->core.attr.sample_period; 2008 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 2009 data->misc = event->header.misc; 2010 data->id = -1ULL; 2011 data->data_src = PERF_MEM_DATA_SRC_NONE; 2012 2013 if (event->header.type != PERF_RECORD_SAMPLE) { 2014 if (!evsel->core.attr.sample_id_all) 2015 return 0; 2016 return perf_evsel__parse_id_sample(evsel, event, data); 2017 } 2018 2019 array = event->sample.array; 2020 2021 if (perf_event__check_size(event, evsel->sample_size)) 2022 return -EFAULT; 2023 2024 if (type & PERF_SAMPLE_IDENTIFIER) { 2025 data->id = *array; 2026 array++; 2027 } 2028 2029 if (type & PERF_SAMPLE_IP) { 2030 data->ip = *array; 2031 array++; 2032 } 2033 2034 if (type & PERF_SAMPLE_TID) { 2035 u.val64 = *array; 2036 if (swapped) { 2037 /* undo swap of u64, then swap on individual u32s */ 2038 u.val64 = bswap_64(u.val64); 2039 u.val32[0] = bswap_32(u.val32[0]); 2040 u.val32[1] = bswap_32(u.val32[1]); 2041 } 2042 2043 data->pid = u.val32[0]; 2044 data->tid = u.val32[1]; 2045 array++; 2046 } 2047 2048 if (type & PERF_SAMPLE_TIME) { 2049 data->time = *array; 2050 array++; 2051 } 2052 2053 if (type & PERF_SAMPLE_ADDR) { 2054 data->addr = *array; 2055 array++; 2056 } 2057 2058 if (type & PERF_SAMPLE_ID) { 2059 data->id = *array; 2060 array++; 2061 } 2062 2063 if (type & PERF_SAMPLE_STREAM_ID) { 2064 data->stream_id = *array; 2065 array++; 2066 } 2067 2068 if (type & PERF_SAMPLE_CPU) { 2069 2070 u.val64 = *array; 2071 if (swapped) { 2072 /* undo swap of u64, then swap on individual u32s */ 2073 u.val64 = bswap_64(u.val64); 2074 u.val32[0] = bswap_32(u.val32[0]); 2075 } 2076 2077 data->cpu = u.val32[0]; 2078 array++; 2079 } 2080 2081 if (type & PERF_SAMPLE_PERIOD) { 2082 data->period = *array; 2083 array++; 2084 } 2085 2086 if (type & PERF_SAMPLE_READ) { 2087 u64 read_format = evsel->core.attr.read_format; 2088 2089 OVERFLOW_CHECK_u64(array); 2090 if (read_format & PERF_FORMAT_GROUP) 2091 data->read.group.nr = *array; 2092 else 2093 data->read.one.value = *array; 2094 2095 array++; 2096 2097 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 2098 OVERFLOW_CHECK_u64(array); 2099 data->read.time_enabled = *array; 2100 array++; 2101 } 2102 2103 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 2104 OVERFLOW_CHECK_u64(array); 2105 data->read.time_running = *array; 2106 array++; 2107 } 2108 2109 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 2110 if (read_format & PERF_FORMAT_GROUP) { 2111 const u64 max_group_nr = UINT64_MAX / 2112 sizeof(struct sample_read_value); 2113 2114 if (data->read.group.nr > max_group_nr) 2115 return -EFAULT; 2116 sz = data->read.group.nr * 2117 sizeof(struct sample_read_value); 2118 OVERFLOW_CHECK(array, sz, max_size); 2119 data->read.group.values = 2120 (struct sample_read_value *)array; 2121 array = (void *)array + sz; 2122 } else { 2123 OVERFLOW_CHECK_u64(array); 2124 data->read.one.id = *array; 2125 array++; 2126 } 2127 } 2128 2129 if (evsel__has_callchain(evsel)) { 2130 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 2131 2132 OVERFLOW_CHECK_u64(array); 2133 data->callchain = (struct ip_callchain *)array++; 2134 if (data->callchain->nr > max_callchain_nr) 2135 return -EFAULT; 2136 sz = data->callchain->nr * sizeof(u64); 2137 OVERFLOW_CHECK(array, sz, max_size); 2138 array = (void *)array + sz; 2139 } 2140 2141 if (type & PERF_SAMPLE_RAW) { 2142 OVERFLOW_CHECK_u64(array); 2143 u.val64 = *array; 2144 2145 /* 2146 * Undo swap of u64, then swap on individual u32s, 2147 * get the size of the raw area and undo all of the 2148 * swap. The pevent interface handles endianity by 2149 * itself. 2150 */ 2151 if (swapped) { 2152 u.val64 = bswap_64(u.val64); 2153 u.val32[0] = bswap_32(u.val32[0]); 2154 u.val32[1] = bswap_32(u.val32[1]); 2155 } 2156 data->raw_size = u.val32[0]; 2157 2158 /* 2159 * The raw data is aligned on 64bits including the 2160 * u32 size, so it's safe to use mem_bswap_64. 2161 */ 2162 if (swapped) 2163 mem_bswap_64((void *) array, data->raw_size); 2164 2165 array = (void *)array + sizeof(u32); 2166 2167 OVERFLOW_CHECK(array, data->raw_size, max_size); 2168 data->raw_data = (void *)array; 2169 array = (void *)array + data->raw_size; 2170 } 2171 2172 if (type & PERF_SAMPLE_BRANCH_STACK) { 2173 const u64 max_branch_nr = UINT64_MAX / 2174 sizeof(struct branch_entry); 2175 2176 OVERFLOW_CHECK_u64(array); 2177 data->branch_stack = (struct branch_stack *)array++; 2178 2179 if (data->branch_stack->nr > max_branch_nr) 2180 return -EFAULT; 2181 2182 sz = data->branch_stack->nr * sizeof(struct branch_entry); 2183 if (perf_evsel__has_branch_hw_idx(evsel)) 2184 sz += sizeof(u64); 2185 else 2186 data->no_hw_idx = true; 2187 OVERFLOW_CHECK(array, sz, max_size); 2188 array = (void *)array + sz; 2189 } 2190 2191 if (type & PERF_SAMPLE_REGS_USER) { 2192 OVERFLOW_CHECK_u64(array); 2193 data->user_regs.abi = *array; 2194 array++; 2195 2196 if (data->user_regs.abi) { 2197 u64 mask = evsel->core.attr.sample_regs_user; 2198 2199 sz = hweight64(mask) * sizeof(u64); 2200 OVERFLOW_CHECK(array, sz, max_size); 2201 data->user_regs.mask = mask; 2202 data->user_regs.regs = (u64 *)array; 2203 array = (void *)array + sz; 2204 } 2205 } 2206 2207 if (type & PERF_SAMPLE_STACK_USER) { 2208 OVERFLOW_CHECK_u64(array); 2209 sz = *array++; 2210 2211 data->user_stack.offset = ((char *)(array - 1) 2212 - (char *) event); 2213 2214 if (!sz) { 2215 data->user_stack.size = 0; 2216 } else { 2217 OVERFLOW_CHECK(array, sz, max_size); 2218 data->user_stack.data = (char *)array; 2219 array = (void *)array + sz; 2220 OVERFLOW_CHECK_u64(array); 2221 data->user_stack.size = *array++; 2222 if (WARN_ONCE(data->user_stack.size > sz, 2223 "user stack dump failure\n")) 2224 return -EFAULT; 2225 } 2226 } 2227 2228 if (type & PERF_SAMPLE_WEIGHT) { 2229 OVERFLOW_CHECK_u64(array); 2230 data->weight = *array; 2231 array++; 2232 } 2233 2234 if (type & PERF_SAMPLE_DATA_SRC) { 2235 OVERFLOW_CHECK_u64(array); 2236 data->data_src = *array; 2237 array++; 2238 } 2239 2240 if (type & PERF_SAMPLE_TRANSACTION) { 2241 OVERFLOW_CHECK_u64(array); 2242 data->transaction = *array; 2243 array++; 2244 } 2245 2246 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE; 2247 if (type & PERF_SAMPLE_REGS_INTR) { 2248 OVERFLOW_CHECK_u64(array); 2249 data->intr_regs.abi = *array; 2250 array++; 2251 2252 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) { 2253 u64 mask = evsel->core.attr.sample_regs_intr; 2254 2255 sz = hweight64(mask) * sizeof(u64); 2256 OVERFLOW_CHECK(array, sz, max_size); 2257 data->intr_regs.mask = mask; 2258 data->intr_regs.regs = (u64 *)array; 2259 array = (void *)array + sz; 2260 } 2261 } 2262 2263 data->phys_addr = 0; 2264 if (type & PERF_SAMPLE_PHYS_ADDR) { 2265 data->phys_addr = *array; 2266 array++; 2267 } 2268 2269 if (type & PERF_SAMPLE_AUX) { 2270 OVERFLOW_CHECK_u64(array); 2271 sz = *array++; 2272 2273 OVERFLOW_CHECK(array, sz, max_size); 2274 /* Undo swap of data */ 2275 if (swapped) 2276 mem_bswap_64((char *)array, sz); 2277 data->aux_sample.size = sz; 2278 data->aux_sample.data = (char *)array; 2279 array = (void *)array + sz; 2280 } 2281 2282 return 0; 2283 } 2284 2285 int perf_evsel__parse_sample_timestamp(struct evsel *evsel, 2286 union perf_event *event, 2287 u64 *timestamp) 2288 { 2289 u64 type = evsel->core.attr.sample_type; 2290 const __u64 *array; 2291 2292 if (!(type & PERF_SAMPLE_TIME)) 2293 return -1; 2294 2295 if (event->header.type != PERF_RECORD_SAMPLE) { 2296 struct perf_sample data = { 2297 .time = -1ULL, 2298 }; 2299 2300 if (!evsel->core.attr.sample_id_all) 2301 return -1; 2302 if (perf_evsel__parse_id_sample(evsel, event, &data)) 2303 return -1; 2304 2305 *timestamp = data.time; 2306 return 0; 2307 } 2308 2309 array = event->sample.array; 2310 2311 if (perf_event__check_size(event, evsel->sample_size)) 2312 return -EFAULT; 2313 2314 if (type & PERF_SAMPLE_IDENTIFIER) 2315 array++; 2316 2317 if (type & PERF_SAMPLE_IP) 2318 array++; 2319 2320 if (type & PERF_SAMPLE_TID) 2321 array++; 2322 2323 if (type & PERF_SAMPLE_TIME) 2324 *timestamp = *array; 2325 2326 return 0; 2327 } 2328 2329 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name) 2330 { 2331 return tep_find_field(evsel->tp_format, name); 2332 } 2333 2334 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, 2335 const char *name) 2336 { 2337 struct tep_format_field *field = perf_evsel__field(evsel, name); 2338 int offset; 2339 2340 if (!field) 2341 return NULL; 2342 2343 offset = field->offset; 2344 2345 if (field->flags & TEP_FIELD_IS_DYNAMIC) { 2346 offset = *(int *)(sample->raw_data + field->offset); 2347 offset &= 0xffff; 2348 } 2349 2350 return sample->raw_data + offset; 2351 } 2352 2353 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample, 2354 bool needs_swap) 2355 { 2356 u64 value; 2357 void *ptr = sample->raw_data + field->offset; 2358 2359 switch (field->size) { 2360 case 1: 2361 return *(u8 *)ptr; 2362 case 2: 2363 value = *(u16 *)ptr; 2364 break; 2365 case 4: 2366 value = *(u32 *)ptr; 2367 break; 2368 case 8: 2369 memcpy(&value, ptr, sizeof(u64)); 2370 break; 2371 default: 2372 return 0; 2373 } 2374 2375 if (!needs_swap) 2376 return value; 2377 2378 switch (field->size) { 2379 case 2: 2380 return bswap_16(value); 2381 case 4: 2382 return bswap_32(value); 2383 case 8: 2384 return bswap_64(value); 2385 default: 2386 return 0; 2387 } 2388 2389 return 0; 2390 } 2391 2392 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample, 2393 const char *name) 2394 { 2395 struct tep_format_field *field = perf_evsel__field(evsel, name); 2396 2397 if (!field) 2398 return 0; 2399 2400 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0; 2401 } 2402 2403 bool perf_evsel__fallback(struct evsel *evsel, int err, 2404 char *msg, size_t msgsize) 2405 { 2406 int paranoid; 2407 2408 if ((err == ENOENT || err == ENXIO || err == ENODEV) && 2409 evsel->core.attr.type == PERF_TYPE_HARDWARE && 2410 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) { 2411 /* 2412 * If it's cycles then fall back to hrtimer based 2413 * cpu-clock-tick sw counter, which is always available even if 2414 * no PMU support. 2415 * 2416 * PPC returns ENXIO until 2.6.37 (behavior changed with commit 2417 * b0a873e). 2418 */ 2419 scnprintf(msg, msgsize, "%s", 2420 "The cycles event is not supported, trying to fall back to cpu-clock-ticks"); 2421 2422 evsel->core.attr.type = PERF_TYPE_SOFTWARE; 2423 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK; 2424 2425 zfree(&evsel->name); 2426 return true; 2427 } else if (err == EACCES && !evsel->core.attr.exclude_kernel && 2428 (paranoid = perf_event_paranoid()) > 1) { 2429 const char *name = perf_evsel__name(evsel); 2430 char *new_name; 2431 const char *sep = ":"; 2432 2433 /* Is there already the separator in the name. */ 2434 if (strchr(name, '/') || 2435 strchr(name, ':')) 2436 sep = ""; 2437 2438 if (asprintf(&new_name, "%s%su", name, sep) < 0) 2439 return false; 2440 2441 if (evsel->name) 2442 free(evsel->name); 2443 evsel->name = new_name; 2444 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying " 2445 "to fall back to excluding kernel and hypervisor " 2446 " samples", paranoid); 2447 evsel->core.attr.exclude_kernel = 1; 2448 evsel->core.attr.exclude_hv = 1; 2449 2450 return true; 2451 } 2452 2453 return false; 2454 } 2455 2456 static bool find_process(const char *name) 2457 { 2458 size_t len = strlen(name); 2459 DIR *dir; 2460 struct dirent *d; 2461 int ret = -1; 2462 2463 dir = opendir(procfs__mountpoint()); 2464 if (!dir) 2465 return false; 2466 2467 /* Walk through the directory. */ 2468 while (ret && (d = readdir(dir)) != NULL) { 2469 char path[PATH_MAX]; 2470 char *data; 2471 size_t size; 2472 2473 if ((d->d_type != DT_DIR) || 2474 !strcmp(".", d->d_name) || 2475 !strcmp("..", d->d_name)) 2476 continue; 2477 2478 scnprintf(path, sizeof(path), "%s/%s/comm", 2479 procfs__mountpoint(), d->d_name); 2480 2481 if (filename__read_str(path, &data, &size)) 2482 continue; 2483 2484 ret = strncmp(name, data, len); 2485 free(data); 2486 } 2487 2488 closedir(dir); 2489 return ret ? false : true; 2490 } 2491 2492 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target, 2493 int err, char *msg, size_t size) 2494 { 2495 char sbuf[STRERR_BUFSIZE]; 2496 int printed = 0; 2497 2498 switch (err) { 2499 case EPERM: 2500 case EACCES: 2501 if (err == EPERM) 2502 printed = scnprintf(msg, size, 2503 "No permission to enable %s event.\n\n", 2504 perf_evsel__name(evsel)); 2505 2506 return scnprintf(msg + printed, size - printed, 2507 "You may not have permission to collect %sstats.\n\n" 2508 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n" 2509 "which controls use of the performance events system by\n" 2510 "unprivileged users (without CAP_SYS_ADMIN).\n\n" 2511 "The current value is %d:\n\n" 2512 " -1: Allow use of (almost) all events by all users\n" 2513 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n" 2514 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n" 2515 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n" 2516 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n" 2517 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n" 2518 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n" 2519 " kernel.perf_event_paranoid = -1\n" , 2520 target->system_wide ? "system-wide " : "", 2521 perf_event_paranoid()); 2522 case ENOENT: 2523 return scnprintf(msg, size, "The %s event is not supported.", 2524 perf_evsel__name(evsel)); 2525 case EMFILE: 2526 return scnprintf(msg, size, "%s", 2527 "Too many events are opened.\n" 2528 "Probably the maximum number of open file descriptors has been reached.\n" 2529 "Hint: Try again after reducing the number of events.\n" 2530 "Hint: Try increasing the limit with 'ulimit -n <limit>'"); 2531 case ENOMEM: 2532 if (evsel__has_callchain(evsel) && 2533 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0) 2534 return scnprintf(msg, size, 2535 "Not enough memory to setup event with callchain.\n" 2536 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n" 2537 "Hint: Current value: %d", sysctl__max_stack()); 2538 break; 2539 case ENODEV: 2540 if (target->cpu_list) 2541 return scnprintf(msg, size, "%s", 2542 "No such device - did you specify an out-of-range profile CPU?"); 2543 break; 2544 case EOPNOTSUPP: 2545 if (evsel->core.attr.sample_period != 0) 2546 return scnprintf(msg, size, 2547 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'", 2548 perf_evsel__name(evsel)); 2549 if (evsel->core.attr.precise_ip) 2550 return scnprintf(msg, size, "%s", 2551 "\'precise\' request may not be supported. Try removing 'p' modifier."); 2552 #if defined(__i386__) || defined(__x86_64__) 2553 if (evsel->core.attr.type == PERF_TYPE_HARDWARE) 2554 return scnprintf(msg, size, "%s", 2555 "No hardware sampling interrupt available.\n"); 2556 #endif 2557 break; 2558 case EBUSY: 2559 if (find_process("oprofiled")) 2560 return scnprintf(msg, size, 2561 "The PMU counters are busy/taken by another profiler.\n" 2562 "We found oprofile daemon running, please stop it and try again."); 2563 break; 2564 case EINVAL: 2565 if (evsel->core.attr.write_backward && perf_missing_features.write_backward) 2566 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel."); 2567 if (perf_missing_features.clockid) 2568 return scnprintf(msg, size, "clockid feature not supported."); 2569 if (perf_missing_features.clockid_wrong) 2570 return scnprintf(msg, size, "wrong clockid (%d).", clockid); 2571 if (perf_missing_features.aux_output) 2572 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel."); 2573 break; 2574 default: 2575 break; 2576 } 2577 2578 return scnprintf(msg, size, 2579 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n" 2580 "/bin/dmesg | grep -i perf may provide additional information.\n", 2581 err, str_error_r(err, sbuf, sizeof(sbuf)), 2582 perf_evsel__name(evsel)); 2583 } 2584 2585 struct perf_env *perf_evsel__env(struct evsel *evsel) 2586 { 2587 if (evsel && evsel->evlist) 2588 return evsel->evlist->env; 2589 return &perf_env; 2590 } 2591 2592 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist) 2593 { 2594 int cpu, thread; 2595 2596 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) { 2597 for (thread = 0; thread < xyarray__max_y(evsel->core.fd); 2598 thread++) { 2599 int fd = FD(evsel, cpu, thread); 2600 2601 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, 2602 cpu, thread, fd) < 0) 2603 return -1; 2604 } 2605 } 2606 2607 return 0; 2608 } 2609 2610 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist) 2611 { 2612 struct perf_cpu_map *cpus = evsel->core.cpus; 2613 struct perf_thread_map *threads = evsel->core.threads; 2614 2615 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr)) 2616 return -ENOMEM; 2617 2618 return store_evsel_ids(evsel, evlist); 2619 } 2620