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_cgroup) { 1108 attr->cgroup = track && !perf_missing_features.cgroup; 1109 perf_evsel__set_sample_bit(evsel, CGROUP); 1110 } 1111 1112 if (opts->record_switch_events) 1113 attr->context_switch = track; 1114 1115 if (opts->sample_transaction) 1116 perf_evsel__set_sample_bit(evsel, TRANSACTION); 1117 1118 if (opts->running_time) { 1119 evsel->core.attr.read_format |= 1120 PERF_FORMAT_TOTAL_TIME_ENABLED | 1121 PERF_FORMAT_TOTAL_TIME_RUNNING; 1122 } 1123 1124 /* 1125 * XXX see the function comment above 1126 * 1127 * Disabling only independent events or group leaders, 1128 * keeping group members enabled. 1129 */ 1130 if (perf_evsel__is_group_leader(evsel)) 1131 attr->disabled = 1; 1132 1133 /* 1134 * Setting enable_on_exec for independent events and 1135 * group leaders for traced executed by perf. 1136 */ 1137 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) && 1138 !opts->initial_delay) 1139 attr->enable_on_exec = 1; 1140 1141 if (evsel->immediate) { 1142 attr->disabled = 0; 1143 attr->enable_on_exec = 0; 1144 } 1145 1146 clockid = opts->clockid; 1147 if (opts->use_clockid) { 1148 attr->use_clockid = 1; 1149 attr->clockid = opts->clockid; 1150 } 1151 1152 if (evsel->precise_max) 1153 attr->precise_ip = 3; 1154 1155 if (opts->all_user) { 1156 attr->exclude_kernel = 1; 1157 attr->exclude_user = 0; 1158 } 1159 1160 if (opts->all_kernel) { 1161 attr->exclude_kernel = 0; 1162 attr->exclude_user = 1; 1163 } 1164 1165 if (evsel->core.own_cpus || evsel->unit) 1166 evsel->core.attr.read_format |= PERF_FORMAT_ID; 1167 1168 /* 1169 * Apply event specific term settings, 1170 * it overloads any global configuration. 1171 */ 1172 apply_config_terms(evsel, opts, track); 1173 1174 evsel->ignore_missing_thread = opts->ignore_missing_thread; 1175 1176 /* The --period option takes the precedence. */ 1177 if (opts->period_set) { 1178 if (opts->period) 1179 perf_evsel__set_sample_bit(evsel, PERIOD); 1180 else 1181 perf_evsel__reset_sample_bit(evsel, PERIOD); 1182 } 1183 1184 /* 1185 * For initial_delay, a dummy event is added implicitly. 1186 * The software event will trigger -EOPNOTSUPP error out, 1187 * if BRANCH_STACK bit is set. 1188 */ 1189 if (opts->initial_delay && is_dummy_event(evsel)) 1190 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK); 1191 } 1192 1193 int perf_evsel__set_filter(struct evsel *evsel, const char *filter) 1194 { 1195 char *new_filter = strdup(filter); 1196 1197 if (new_filter != NULL) { 1198 free(evsel->filter); 1199 evsel->filter = new_filter; 1200 return 0; 1201 } 1202 1203 return -1; 1204 } 1205 1206 static int perf_evsel__append_filter(struct evsel *evsel, 1207 const char *fmt, const char *filter) 1208 { 1209 char *new_filter; 1210 1211 if (evsel->filter == NULL) 1212 return perf_evsel__set_filter(evsel, filter); 1213 1214 if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) { 1215 free(evsel->filter); 1216 evsel->filter = new_filter; 1217 return 0; 1218 } 1219 1220 return -1; 1221 } 1222 1223 int perf_evsel__append_tp_filter(struct evsel *evsel, const char *filter) 1224 { 1225 return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter); 1226 } 1227 1228 int perf_evsel__append_addr_filter(struct evsel *evsel, const char *filter) 1229 { 1230 return perf_evsel__append_filter(evsel, "%s,%s", filter); 1231 } 1232 1233 /* Caller has to clear disabled after going through all CPUs. */ 1234 int evsel__enable_cpu(struct evsel *evsel, int cpu) 1235 { 1236 return perf_evsel__enable_cpu(&evsel->core, cpu); 1237 } 1238 1239 int evsel__enable(struct evsel *evsel) 1240 { 1241 int err = perf_evsel__enable(&evsel->core); 1242 1243 if (!err) 1244 evsel->disabled = false; 1245 return err; 1246 } 1247 1248 /* Caller has to set disabled after going through all CPUs. */ 1249 int evsel__disable_cpu(struct evsel *evsel, int cpu) 1250 { 1251 return perf_evsel__disable_cpu(&evsel->core, cpu); 1252 } 1253 1254 int evsel__disable(struct evsel *evsel) 1255 { 1256 int err = perf_evsel__disable(&evsel->core); 1257 /* 1258 * We mark it disabled here so that tools that disable a event can 1259 * ignore events after they disable it. I.e. the ring buffer may have 1260 * already a few more events queued up before the kernel got the stop 1261 * request. 1262 */ 1263 if (!err) 1264 evsel->disabled = true; 1265 1266 return err; 1267 } 1268 1269 static void perf_evsel__free_config_terms(struct evsel *evsel) 1270 { 1271 struct perf_evsel_config_term *term, *h; 1272 1273 list_for_each_entry_safe(term, h, &evsel->config_terms, list) { 1274 list_del_init(&term->list); 1275 if (term->free_str) 1276 zfree(&term->val.str); 1277 free(term); 1278 } 1279 } 1280 1281 void perf_evsel__exit(struct evsel *evsel) 1282 { 1283 assert(list_empty(&evsel->core.node)); 1284 assert(evsel->evlist == NULL); 1285 perf_evsel__free_counts(evsel); 1286 perf_evsel__free_fd(&evsel->core); 1287 perf_evsel__free_id(&evsel->core); 1288 perf_evsel__free_config_terms(evsel); 1289 cgroup__put(evsel->cgrp); 1290 perf_cpu_map__put(evsel->core.cpus); 1291 perf_cpu_map__put(evsel->core.own_cpus); 1292 perf_thread_map__put(evsel->core.threads); 1293 zfree(&evsel->group_name); 1294 zfree(&evsel->name); 1295 zfree(&evsel->pmu_name); 1296 perf_evsel__object.fini(evsel); 1297 } 1298 1299 void evsel__delete(struct evsel *evsel) 1300 { 1301 perf_evsel__exit(evsel); 1302 free(evsel); 1303 } 1304 1305 void perf_evsel__compute_deltas(struct evsel *evsel, int cpu, int thread, 1306 struct perf_counts_values *count) 1307 { 1308 struct perf_counts_values tmp; 1309 1310 if (!evsel->prev_raw_counts) 1311 return; 1312 1313 if (cpu == -1) { 1314 tmp = evsel->prev_raw_counts->aggr; 1315 evsel->prev_raw_counts->aggr = *count; 1316 } else { 1317 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread); 1318 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count; 1319 } 1320 1321 count->val = count->val - tmp.val; 1322 count->ena = count->ena - tmp.ena; 1323 count->run = count->run - tmp.run; 1324 } 1325 1326 void perf_counts_values__scale(struct perf_counts_values *count, 1327 bool scale, s8 *pscaled) 1328 { 1329 s8 scaled = 0; 1330 1331 if (scale) { 1332 if (count->run == 0) { 1333 scaled = -1; 1334 count->val = 0; 1335 } else if (count->run < count->ena) { 1336 scaled = 1; 1337 count->val = (u64)((double) count->val * count->ena / count->run); 1338 } 1339 } 1340 1341 if (pscaled) 1342 *pscaled = scaled; 1343 } 1344 1345 static int 1346 perf_evsel__read_one(struct evsel *evsel, int cpu, int thread) 1347 { 1348 struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread); 1349 1350 return perf_evsel__read(&evsel->core, cpu, thread, count); 1351 } 1352 1353 static void 1354 perf_evsel__set_count(struct evsel *counter, int cpu, int thread, 1355 u64 val, u64 ena, u64 run) 1356 { 1357 struct perf_counts_values *count; 1358 1359 count = perf_counts(counter->counts, cpu, thread); 1360 1361 count->val = val; 1362 count->ena = ena; 1363 count->run = run; 1364 1365 perf_counts__set_loaded(counter->counts, cpu, thread, true); 1366 } 1367 1368 static int 1369 perf_evsel__process_group_data(struct evsel *leader, 1370 int cpu, int thread, u64 *data) 1371 { 1372 u64 read_format = leader->core.attr.read_format; 1373 struct sample_read_value *v; 1374 u64 nr, ena = 0, run = 0, i; 1375 1376 nr = *data++; 1377 1378 if (nr != (u64) leader->core.nr_members) 1379 return -EINVAL; 1380 1381 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1382 ena = *data++; 1383 1384 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1385 run = *data++; 1386 1387 v = (struct sample_read_value *) data; 1388 1389 perf_evsel__set_count(leader, cpu, thread, 1390 v[0].value, ena, run); 1391 1392 for (i = 1; i < nr; i++) { 1393 struct evsel *counter; 1394 1395 counter = perf_evlist__id2evsel(leader->evlist, v[i].id); 1396 if (!counter) 1397 return -EINVAL; 1398 1399 perf_evsel__set_count(counter, cpu, thread, 1400 v[i].value, ena, run); 1401 } 1402 1403 return 0; 1404 } 1405 1406 static int 1407 perf_evsel__read_group(struct evsel *leader, int cpu, int thread) 1408 { 1409 struct perf_stat_evsel *ps = leader->stats; 1410 u64 read_format = leader->core.attr.read_format; 1411 int size = perf_evsel__read_size(&leader->core); 1412 u64 *data = ps->group_data; 1413 1414 if (!(read_format & PERF_FORMAT_ID)) 1415 return -EINVAL; 1416 1417 if (!perf_evsel__is_group_leader(leader)) 1418 return -EINVAL; 1419 1420 if (!data) { 1421 data = zalloc(size); 1422 if (!data) 1423 return -ENOMEM; 1424 1425 ps->group_data = data; 1426 } 1427 1428 if (FD(leader, cpu, thread) < 0) 1429 return -EINVAL; 1430 1431 if (readn(FD(leader, cpu, thread), data, size) <= 0) 1432 return -errno; 1433 1434 return perf_evsel__process_group_data(leader, cpu, thread, data); 1435 } 1436 1437 int perf_evsel__read_counter(struct evsel *evsel, int cpu, int thread) 1438 { 1439 u64 read_format = evsel->core.attr.read_format; 1440 1441 if (read_format & PERF_FORMAT_GROUP) 1442 return perf_evsel__read_group(evsel, cpu, thread); 1443 else 1444 return perf_evsel__read_one(evsel, cpu, thread); 1445 } 1446 1447 int __perf_evsel__read_on_cpu(struct evsel *evsel, 1448 int cpu, int thread, bool scale) 1449 { 1450 struct perf_counts_values count; 1451 size_t nv = scale ? 3 : 1; 1452 1453 if (FD(evsel, cpu, thread) < 0) 1454 return -EINVAL; 1455 1456 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0) 1457 return -ENOMEM; 1458 1459 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0) 1460 return -errno; 1461 1462 perf_evsel__compute_deltas(evsel, cpu, thread, &count); 1463 perf_counts_values__scale(&count, scale, NULL); 1464 *perf_counts(evsel->counts, cpu, thread) = count; 1465 return 0; 1466 } 1467 1468 static int get_group_fd(struct evsel *evsel, int cpu, int thread) 1469 { 1470 struct evsel *leader = evsel->leader; 1471 int fd; 1472 1473 if (perf_evsel__is_group_leader(evsel)) 1474 return -1; 1475 1476 /* 1477 * Leader must be already processed/open, 1478 * if not it's a bug. 1479 */ 1480 BUG_ON(!leader->core.fd); 1481 1482 fd = FD(leader, cpu, thread); 1483 BUG_ON(fd == -1); 1484 1485 return fd; 1486 } 1487 1488 static void perf_evsel__remove_fd(struct evsel *pos, 1489 int nr_cpus, int nr_threads, 1490 int thread_idx) 1491 { 1492 for (int cpu = 0; cpu < nr_cpus; cpu++) 1493 for (int thread = thread_idx; thread < nr_threads - 1; thread++) 1494 FD(pos, cpu, thread) = FD(pos, cpu, thread + 1); 1495 } 1496 1497 static int update_fds(struct evsel *evsel, 1498 int nr_cpus, int cpu_idx, 1499 int nr_threads, int thread_idx) 1500 { 1501 struct evsel *pos; 1502 1503 if (cpu_idx >= nr_cpus || thread_idx >= nr_threads) 1504 return -EINVAL; 1505 1506 evlist__for_each_entry(evsel->evlist, pos) { 1507 nr_cpus = pos != evsel ? nr_cpus : cpu_idx; 1508 1509 perf_evsel__remove_fd(pos, nr_cpus, nr_threads, thread_idx); 1510 1511 /* 1512 * Since fds for next evsel has not been created, 1513 * there is no need to iterate whole event list. 1514 */ 1515 if (pos == evsel) 1516 break; 1517 } 1518 return 0; 1519 } 1520 1521 static bool ignore_missing_thread(struct evsel *evsel, 1522 int nr_cpus, int cpu, 1523 struct perf_thread_map *threads, 1524 int thread, int err) 1525 { 1526 pid_t ignore_pid = perf_thread_map__pid(threads, thread); 1527 1528 if (!evsel->ignore_missing_thread) 1529 return false; 1530 1531 /* The system wide setup does not work with threads. */ 1532 if (evsel->core.system_wide) 1533 return false; 1534 1535 /* The -ESRCH is perf event syscall errno for pid's not found. */ 1536 if (err != -ESRCH) 1537 return false; 1538 1539 /* If there's only one thread, let it fail. */ 1540 if (threads->nr == 1) 1541 return false; 1542 1543 /* 1544 * We should remove fd for missing_thread first 1545 * because thread_map__remove() will decrease threads->nr. 1546 */ 1547 if (update_fds(evsel, nr_cpus, cpu, threads->nr, thread)) 1548 return false; 1549 1550 if (thread_map__remove(threads, thread)) 1551 return false; 1552 1553 pr_warning("WARNING: Ignored open failure for pid %d\n", 1554 ignore_pid); 1555 return true; 1556 } 1557 1558 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val, 1559 void *priv __maybe_unused) 1560 { 1561 return fprintf(fp, " %-32s %s\n", name, val); 1562 } 1563 1564 static void display_attr(struct perf_event_attr *attr) 1565 { 1566 if (verbose >= 2 || debug_peo_args) { 1567 fprintf(stderr, "%.60s\n", graph_dotted_line); 1568 fprintf(stderr, "perf_event_attr:\n"); 1569 perf_event_attr__fprintf(stderr, attr, __open_attr__fprintf, NULL); 1570 fprintf(stderr, "%.60s\n", graph_dotted_line); 1571 } 1572 } 1573 1574 static int perf_event_open(struct evsel *evsel, 1575 pid_t pid, int cpu, int group_fd, 1576 unsigned long flags) 1577 { 1578 int precise_ip = evsel->core.attr.precise_ip; 1579 int fd; 1580 1581 while (1) { 1582 pr_debug2_peo("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx", 1583 pid, cpu, group_fd, flags); 1584 1585 fd = sys_perf_event_open(&evsel->core.attr, pid, cpu, group_fd, flags); 1586 if (fd >= 0) 1587 break; 1588 1589 /* Do not try less precise if not requested. */ 1590 if (!evsel->precise_max) 1591 break; 1592 1593 /* 1594 * We tried all the precise_ip values, and it's 1595 * still failing, so leave it to standard fallback. 1596 */ 1597 if (!evsel->core.attr.precise_ip) { 1598 evsel->core.attr.precise_ip = precise_ip; 1599 break; 1600 } 1601 1602 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", -ENOTSUP); 1603 evsel->core.attr.precise_ip--; 1604 pr_debug2_peo("decreasing precise_ip by one (%d)\n", evsel->core.attr.precise_ip); 1605 display_attr(&evsel->core.attr); 1606 } 1607 1608 return fd; 1609 } 1610 1611 static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus, 1612 struct perf_thread_map *threads, 1613 int start_cpu, int end_cpu) 1614 { 1615 int cpu, thread, nthreads; 1616 unsigned long flags = PERF_FLAG_FD_CLOEXEC; 1617 int pid = -1, err, old_errno; 1618 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE; 1619 1620 if ((perf_missing_features.write_backward && evsel->core.attr.write_backward) || 1621 (perf_missing_features.aux_output && evsel->core.attr.aux_output)) 1622 return -EINVAL; 1623 1624 if (cpus == NULL) { 1625 static struct perf_cpu_map *empty_cpu_map; 1626 1627 if (empty_cpu_map == NULL) { 1628 empty_cpu_map = perf_cpu_map__dummy_new(); 1629 if (empty_cpu_map == NULL) 1630 return -ENOMEM; 1631 } 1632 1633 cpus = empty_cpu_map; 1634 } 1635 1636 if (threads == NULL) { 1637 static struct perf_thread_map *empty_thread_map; 1638 1639 if (empty_thread_map == NULL) { 1640 empty_thread_map = thread_map__new_by_tid(-1); 1641 if (empty_thread_map == NULL) 1642 return -ENOMEM; 1643 } 1644 1645 threads = empty_thread_map; 1646 } 1647 1648 if (evsel->core.system_wide) 1649 nthreads = 1; 1650 else 1651 nthreads = threads->nr; 1652 1653 if (evsel->core.fd == NULL && 1654 perf_evsel__alloc_fd(&evsel->core, cpus->nr, nthreads) < 0) 1655 return -ENOMEM; 1656 1657 if (evsel->cgrp) { 1658 flags |= PERF_FLAG_PID_CGROUP; 1659 pid = evsel->cgrp->fd; 1660 } 1661 1662 fallback_missing_features: 1663 if (perf_missing_features.clockid_wrong) 1664 evsel->core.attr.clockid = CLOCK_MONOTONIC; /* should always work */ 1665 if (perf_missing_features.clockid) { 1666 evsel->core.attr.use_clockid = 0; 1667 evsel->core.attr.clockid = 0; 1668 } 1669 if (perf_missing_features.cloexec) 1670 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 1671 if (perf_missing_features.mmap2) 1672 evsel->core.attr.mmap2 = 0; 1673 if (perf_missing_features.exclude_guest) 1674 evsel->core.attr.exclude_guest = evsel->core.attr.exclude_host = 0; 1675 if (perf_missing_features.lbr_flags) 1676 evsel->core.attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS | 1677 PERF_SAMPLE_BRANCH_NO_CYCLES); 1678 if (perf_missing_features.group_read && evsel->core.attr.inherit) 1679 evsel->core.attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID); 1680 if (perf_missing_features.ksymbol) 1681 evsel->core.attr.ksymbol = 0; 1682 if (perf_missing_features.bpf) 1683 evsel->core.attr.bpf_event = 0; 1684 if (perf_missing_features.branch_hw_idx) 1685 evsel->core.attr.branch_sample_type &= ~PERF_SAMPLE_BRANCH_HW_INDEX; 1686 retry_sample_id: 1687 if (perf_missing_features.sample_id_all) 1688 evsel->core.attr.sample_id_all = 0; 1689 1690 display_attr(&evsel->core.attr); 1691 1692 for (cpu = start_cpu; cpu < end_cpu; cpu++) { 1693 1694 for (thread = 0; thread < nthreads; thread++) { 1695 int fd, group_fd; 1696 1697 if (!evsel->cgrp && !evsel->core.system_wide) 1698 pid = perf_thread_map__pid(threads, thread); 1699 1700 group_fd = get_group_fd(evsel, cpu, thread); 1701 retry_open: 1702 test_attr__ready(); 1703 1704 fd = perf_event_open(evsel, pid, cpus->map[cpu], 1705 group_fd, flags); 1706 1707 FD(evsel, cpu, thread) = fd; 1708 1709 if (fd < 0) { 1710 err = -errno; 1711 1712 if (ignore_missing_thread(evsel, cpus->nr, cpu, threads, thread, err)) { 1713 /* 1714 * We just removed 1 thread, so take a step 1715 * back on thread index and lower the upper 1716 * nthreads limit. 1717 */ 1718 nthreads--; 1719 thread--; 1720 1721 /* ... and pretend like nothing have happened. */ 1722 err = 0; 1723 continue; 1724 } 1725 1726 pr_debug2_peo("\nsys_perf_event_open failed, error %d\n", 1727 err); 1728 goto try_fallback; 1729 } 1730 1731 pr_debug2_peo(" = %d\n", fd); 1732 1733 if (evsel->bpf_fd >= 0) { 1734 int evt_fd = fd; 1735 int bpf_fd = evsel->bpf_fd; 1736 1737 err = ioctl(evt_fd, 1738 PERF_EVENT_IOC_SET_BPF, 1739 bpf_fd); 1740 if (err && errno != EEXIST) { 1741 pr_err("failed to attach bpf fd %d: %s\n", 1742 bpf_fd, strerror(errno)); 1743 err = -EINVAL; 1744 goto out_close; 1745 } 1746 } 1747 1748 set_rlimit = NO_CHANGE; 1749 1750 /* 1751 * If we succeeded but had to kill clockid, fail and 1752 * have perf_evsel__open_strerror() print us a nice 1753 * error. 1754 */ 1755 if (perf_missing_features.clockid || 1756 perf_missing_features.clockid_wrong) { 1757 err = -EINVAL; 1758 goto out_close; 1759 } 1760 } 1761 } 1762 1763 return 0; 1764 1765 try_fallback: 1766 /* 1767 * perf stat needs between 5 and 22 fds per CPU. When we run out 1768 * of them try to increase the limits. 1769 */ 1770 if (err == -EMFILE && set_rlimit < INCREASED_MAX) { 1771 struct rlimit l; 1772 1773 old_errno = errno; 1774 if (getrlimit(RLIMIT_NOFILE, &l) == 0) { 1775 if (set_rlimit == NO_CHANGE) 1776 l.rlim_cur = l.rlim_max; 1777 else { 1778 l.rlim_cur = l.rlim_max + 1000; 1779 l.rlim_max = l.rlim_cur; 1780 } 1781 if (setrlimit(RLIMIT_NOFILE, &l) == 0) { 1782 set_rlimit++; 1783 errno = old_errno; 1784 goto retry_open; 1785 } 1786 } 1787 errno = old_errno; 1788 } 1789 1790 if (err != -EINVAL || cpu > 0 || thread > 0) 1791 goto out_close; 1792 1793 /* 1794 * Must probe features in the order they were added to the 1795 * perf_event_attr interface. 1796 */ 1797 if (!perf_missing_features.cgroup && evsel->core.attr.cgroup) { 1798 perf_missing_features.cgroup = true; 1799 pr_debug2_peo("Kernel has no cgroup sampling support, bailing out\n"); 1800 goto out_close; 1801 } else if (!perf_missing_features.branch_hw_idx && 1802 (evsel->core.attr.branch_sample_type & PERF_SAMPLE_BRANCH_HW_INDEX)) { 1803 perf_missing_features.branch_hw_idx = true; 1804 pr_debug2("switching off branch HW index support\n"); 1805 goto fallback_missing_features; 1806 } else if (!perf_missing_features.aux_output && evsel->core.attr.aux_output) { 1807 perf_missing_features.aux_output = true; 1808 pr_debug2_peo("Kernel has no attr.aux_output support, bailing out\n"); 1809 goto out_close; 1810 } else if (!perf_missing_features.bpf && evsel->core.attr.bpf_event) { 1811 perf_missing_features.bpf = true; 1812 pr_debug2_peo("switching off bpf_event\n"); 1813 goto fallback_missing_features; 1814 } else if (!perf_missing_features.ksymbol && evsel->core.attr.ksymbol) { 1815 perf_missing_features.ksymbol = true; 1816 pr_debug2_peo("switching off ksymbol\n"); 1817 goto fallback_missing_features; 1818 } else if (!perf_missing_features.write_backward && evsel->core.attr.write_backward) { 1819 perf_missing_features.write_backward = true; 1820 pr_debug2_peo("switching off write_backward\n"); 1821 goto out_close; 1822 } else if (!perf_missing_features.clockid_wrong && evsel->core.attr.use_clockid) { 1823 perf_missing_features.clockid_wrong = true; 1824 pr_debug2_peo("switching off clockid\n"); 1825 goto fallback_missing_features; 1826 } else if (!perf_missing_features.clockid && evsel->core.attr.use_clockid) { 1827 perf_missing_features.clockid = true; 1828 pr_debug2_peo("switching off use_clockid\n"); 1829 goto fallback_missing_features; 1830 } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) { 1831 perf_missing_features.cloexec = true; 1832 pr_debug2_peo("switching off cloexec flag\n"); 1833 goto fallback_missing_features; 1834 } else if (!perf_missing_features.mmap2 && evsel->core.attr.mmap2) { 1835 perf_missing_features.mmap2 = true; 1836 pr_debug2_peo("switching off mmap2\n"); 1837 goto fallback_missing_features; 1838 } else if (!perf_missing_features.exclude_guest && 1839 (evsel->core.attr.exclude_guest || evsel->core.attr.exclude_host)) { 1840 perf_missing_features.exclude_guest = true; 1841 pr_debug2_peo("switching off exclude_guest, exclude_host\n"); 1842 goto fallback_missing_features; 1843 } else if (!perf_missing_features.sample_id_all) { 1844 perf_missing_features.sample_id_all = true; 1845 pr_debug2_peo("switching off sample_id_all\n"); 1846 goto retry_sample_id; 1847 } else if (!perf_missing_features.lbr_flags && 1848 (evsel->core.attr.branch_sample_type & 1849 (PERF_SAMPLE_BRANCH_NO_CYCLES | 1850 PERF_SAMPLE_BRANCH_NO_FLAGS))) { 1851 perf_missing_features.lbr_flags = true; 1852 pr_debug2_peo("switching off branch sample type no (cycles/flags)\n"); 1853 goto fallback_missing_features; 1854 } else if (!perf_missing_features.group_read && 1855 evsel->core.attr.inherit && 1856 (evsel->core.attr.read_format & PERF_FORMAT_GROUP) && 1857 perf_evsel__is_group_leader(evsel)) { 1858 perf_missing_features.group_read = true; 1859 pr_debug2_peo("switching off group read\n"); 1860 goto fallback_missing_features; 1861 } 1862 out_close: 1863 if (err) 1864 threads->err_thread = thread; 1865 1866 old_errno = errno; 1867 do { 1868 while (--thread >= 0) { 1869 if (FD(evsel, cpu, thread) >= 0) 1870 close(FD(evsel, cpu, thread)); 1871 FD(evsel, cpu, thread) = -1; 1872 } 1873 thread = nthreads; 1874 } while (--cpu >= 0); 1875 errno = old_errno; 1876 return err; 1877 } 1878 1879 int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus, 1880 struct perf_thread_map *threads) 1881 { 1882 return evsel__open_cpu(evsel, cpus, threads, 0, cpus ? cpus->nr : 1); 1883 } 1884 1885 void evsel__close(struct evsel *evsel) 1886 { 1887 perf_evsel__close(&evsel->core); 1888 perf_evsel__free_id(&evsel->core); 1889 } 1890 1891 int perf_evsel__open_per_cpu(struct evsel *evsel, 1892 struct perf_cpu_map *cpus, 1893 int cpu) 1894 { 1895 if (cpu == -1) 1896 return evsel__open_cpu(evsel, cpus, NULL, 0, 1897 cpus ? cpus->nr : 1); 1898 1899 return evsel__open_cpu(evsel, cpus, NULL, cpu, cpu + 1); 1900 } 1901 1902 int perf_evsel__open_per_thread(struct evsel *evsel, 1903 struct perf_thread_map *threads) 1904 { 1905 return evsel__open(evsel, NULL, threads); 1906 } 1907 1908 static int perf_evsel__parse_id_sample(const struct evsel *evsel, 1909 const union perf_event *event, 1910 struct perf_sample *sample) 1911 { 1912 u64 type = evsel->core.attr.sample_type; 1913 const __u64 *array = event->sample.array; 1914 bool swapped = evsel->needs_swap; 1915 union u64_swap u; 1916 1917 array += ((event->header.size - 1918 sizeof(event->header)) / sizeof(u64)) - 1; 1919 1920 if (type & PERF_SAMPLE_IDENTIFIER) { 1921 sample->id = *array; 1922 array--; 1923 } 1924 1925 if (type & PERF_SAMPLE_CPU) { 1926 u.val64 = *array; 1927 if (swapped) { 1928 /* undo swap of u64, then swap on individual u32s */ 1929 u.val64 = bswap_64(u.val64); 1930 u.val32[0] = bswap_32(u.val32[0]); 1931 } 1932 1933 sample->cpu = u.val32[0]; 1934 array--; 1935 } 1936 1937 if (type & PERF_SAMPLE_STREAM_ID) { 1938 sample->stream_id = *array; 1939 array--; 1940 } 1941 1942 if (type & PERF_SAMPLE_ID) { 1943 sample->id = *array; 1944 array--; 1945 } 1946 1947 if (type & PERF_SAMPLE_TIME) { 1948 sample->time = *array; 1949 array--; 1950 } 1951 1952 if (type & PERF_SAMPLE_TID) { 1953 u.val64 = *array; 1954 if (swapped) { 1955 /* undo swap of u64, then swap on individual u32s */ 1956 u.val64 = bswap_64(u.val64); 1957 u.val32[0] = bswap_32(u.val32[0]); 1958 u.val32[1] = bswap_32(u.val32[1]); 1959 } 1960 1961 sample->pid = u.val32[0]; 1962 sample->tid = u.val32[1]; 1963 array--; 1964 } 1965 1966 return 0; 1967 } 1968 1969 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 1970 u64 size) 1971 { 1972 return size > max_size || offset + size > endp; 1973 } 1974 1975 #define OVERFLOW_CHECK(offset, size, max_size) \ 1976 do { \ 1977 if (overflow(endp, (max_size), (offset), (size))) \ 1978 return -EFAULT; \ 1979 } while (0) 1980 1981 #define OVERFLOW_CHECK_u64(offset) \ 1982 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 1983 1984 static int 1985 perf_event__check_size(union perf_event *event, unsigned int sample_size) 1986 { 1987 /* 1988 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 1989 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 1990 * check the format does not go past the end of the event. 1991 */ 1992 if (sample_size + sizeof(event->header) > event->header.size) 1993 return -EFAULT; 1994 1995 return 0; 1996 } 1997 1998 int perf_evsel__parse_sample(struct evsel *evsel, union perf_event *event, 1999 struct perf_sample *data) 2000 { 2001 u64 type = evsel->core.attr.sample_type; 2002 bool swapped = evsel->needs_swap; 2003 const __u64 *array; 2004 u16 max_size = event->header.size; 2005 const void *endp = (void *)event + max_size; 2006 u64 sz; 2007 2008 /* 2009 * used for cross-endian analysis. See git commit 65014ab3 2010 * for why this goofiness is needed. 2011 */ 2012 union u64_swap u; 2013 2014 memset(data, 0, sizeof(*data)); 2015 data->cpu = data->pid = data->tid = -1; 2016 data->stream_id = data->id = data->time = -1ULL; 2017 data->period = evsel->core.attr.sample_period; 2018 data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 2019 data->misc = event->header.misc; 2020 data->id = -1ULL; 2021 data->data_src = PERF_MEM_DATA_SRC_NONE; 2022 2023 if (event->header.type != PERF_RECORD_SAMPLE) { 2024 if (!evsel->core.attr.sample_id_all) 2025 return 0; 2026 return perf_evsel__parse_id_sample(evsel, event, data); 2027 } 2028 2029 array = event->sample.array; 2030 2031 if (perf_event__check_size(event, evsel->sample_size)) 2032 return -EFAULT; 2033 2034 if (type & PERF_SAMPLE_IDENTIFIER) { 2035 data->id = *array; 2036 array++; 2037 } 2038 2039 if (type & PERF_SAMPLE_IP) { 2040 data->ip = *array; 2041 array++; 2042 } 2043 2044 if (type & PERF_SAMPLE_TID) { 2045 u.val64 = *array; 2046 if (swapped) { 2047 /* undo swap of u64, then swap on individual u32s */ 2048 u.val64 = bswap_64(u.val64); 2049 u.val32[0] = bswap_32(u.val32[0]); 2050 u.val32[1] = bswap_32(u.val32[1]); 2051 } 2052 2053 data->pid = u.val32[0]; 2054 data->tid = u.val32[1]; 2055 array++; 2056 } 2057 2058 if (type & PERF_SAMPLE_TIME) { 2059 data->time = *array; 2060 array++; 2061 } 2062 2063 if (type & PERF_SAMPLE_ADDR) { 2064 data->addr = *array; 2065 array++; 2066 } 2067 2068 if (type & PERF_SAMPLE_ID) { 2069 data->id = *array; 2070 array++; 2071 } 2072 2073 if (type & PERF_SAMPLE_STREAM_ID) { 2074 data->stream_id = *array; 2075 array++; 2076 } 2077 2078 if (type & PERF_SAMPLE_CPU) { 2079 2080 u.val64 = *array; 2081 if (swapped) { 2082 /* undo swap of u64, then swap on individual u32s */ 2083 u.val64 = bswap_64(u.val64); 2084 u.val32[0] = bswap_32(u.val32[0]); 2085 } 2086 2087 data->cpu = u.val32[0]; 2088 array++; 2089 } 2090 2091 if (type & PERF_SAMPLE_PERIOD) { 2092 data->period = *array; 2093 array++; 2094 } 2095 2096 if (type & PERF_SAMPLE_READ) { 2097 u64 read_format = evsel->core.attr.read_format; 2098 2099 OVERFLOW_CHECK_u64(array); 2100 if (read_format & PERF_FORMAT_GROUP) 2101 data->read.group.nr = *array; 2102 else 2103 data->read.one.value = *array; 2104 2105 array++; 2106 2107 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 2108 OVERFLOW_CHECK_u64(array); 2109 data->read.time_enabled = *array; 2110 array++; 2111 } 2112 2113 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 2114 OVERFLOW_CHECK_u64(array); 2115 data->read.time_running = *array; 2116 array++; 2117 } 2118 2119 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 2120 if (read_format & PERF_FORMAT_GROUP) { 2121 const u64 max_group_nr = UINT64_MAX / 2122 sizeof(struct sample_read_value); 2123 2124 if (data->read.group.nr > max_group_nr) 2125 return -EFAULT; 2126 sz = data->read.group.nr * 2127 sizeof(struct sample_read_value); 2128 OVERFLOW_CHECK(array, sz, max_size); 2129 data->read.group.values = 2130 (struct sample_read_value *)array; 2131 array = (void *)array + sz; 2132 } else { 2133 OVERFLOW_CHECK_u64(array); 2134 data->read.one.id = *array; 2135 array++; 2136 } 2137 } 2138 2139 if (evsel__has_callchain(evsel)) { 2140 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 2141 2142 OVERFLOW_CHECK_u64(array); 2143 data->callchain = (struct ip_callchain *)array++; 2144 if (data->callchain->nr > max_callchain_nr) 2145 return -EFAULT; 2146 sz = data->callchain->nr * sizeof(u64); 2147 OVERFLOW_CHECK(array, sz, max_size); 2148 array = (void *)array + sz; 2149 } 2150 2151 if (type & PERF_SAMPLE_RAW) { 2152 OVERFLOW_CHECK_u64(array); 2153 u.val64 = *array; 2154 2155 /* 2156 * Undo swap of u64, then swap on individual u32s, 2157 * get the size of the raw area and undo all of the 2158 * swap. The pevent interface handles endianity by 2159 * itself. 2160 */ 2161 if (swapped) { 2162 u.val64 = bswap_64(u.val64); 2163 u.val32[0] = bswap_32(u.val32[0]); 2164 u.val32[1] = bswap_32(u.val32[1]); 2165 } 2166 data->raw_size = u.val32[0]; 2167 2168 /* 2169 * The raw data is aligned on 64bits including the 2170 * u32 size, so it's safe to use mem_bswap_64. 2171 */ 2172 if (swapped) 2173 mem_bswap_64((void *) array, data->raw_size); 2174 2175 array = (void *)array + sizeof(u32); 2176 2177 OVERFLOW_CHECK(array, data->raw_size, max_size); 2178 data->raw_data = (void *)array; 2179 array = (void *)array + data->raw_size; 2180 } 2181 2182 if (type & PERF_SAMPLE_BRANCH_STACK) { 2183 const u64 max_branch_nr = UINT64_MAX / 2184 sizeof(struct branch_entry); 2185 2186 OVERFLOW_CHECK_u64(array); 2187 data->branch_stack = (struct branch_stack *)array++; 2188 2189 if (data->branch_stack->nr > max_branch_nr) 2190 return -EFAULT; 2191 2192 sz = data->branch_stack->nr * sizeof(struct branch_entry); 2193 if (perf_evsel__has_branch_hw_idx(evsel)) 2194 sz += sizeof(u64); 2195 else 2196 data->no_hw_idx = true; 2197 OVERFLOW_CHECK(array, sz, max_size); 2198 array = (void *)array + sz; 2199 } 2200 2201 if (type & PERF_SAMPLE_REGS_USER) { 2202 OVERFLOW_CHECK_u64(array); 2203 data->user_regs.abi = *array; 2204 array++; 2205 2206 if (data->user_regs.abi) { 2207 u64 mask = evsel->core.attr.sample_regs_user; 2208 2209 sz = hweight64(mask) * sizeof(u64); 2210 OVERFLOW_CHECK(array, sz, max_size); 2211 data->user_regs.mask = mask; 2212 data->user_regs.regs = (u64 *)array; 2213 array = (void *)array + sz; 2214 } 2215 } 2216 2217 if (type & PERF_SAMPLE_STACK_USER) { 2218 OVERFLOW_CHECK_u64(array); 2219 sz = *array++; 2220 2221 data->user_stack.offset = ((char *)(array - 1) 2222 - (char *) event); 2223 2224 if (!sz) { 2225 data->user_stack.size = 0; 2226 } else { 2227 OVERFLOW_CHECK(array, sz, max_size); 2228 data->user_stack.data = (char *)array; 2229 array = (void *)array + sz; 2230 OVERFLOW_CHECK_u64(array); 2231 data->user_stack.size = *array++; 2232 if (WARN_ONCE(data->user_stack.size > sz, 2233 "user stack dump failure\n")) 2234 return -EFAULT; 2235 } 2236 } 2237 2238 if (type & PERF_SAMPLE_WEIGHT) { 2239 OVERFLOW_CHECK_u64(array); 2240 data->weight = *array; 2241 array++; 2242 } 2243 2244 if (type & PERF_SAMPLE_DATA_SRC) { 2245 OVERFLOW_CHECK_u64(array); 2246 data->data_src = *array; 2247 array++; 2248 } 2249 2250 if (type & PERF_SAMPLE_TRANSACTION) { 2251 OVERFLOW_CHECK_u64(array); 2252 data->transaction = *array; 2253 array++; 2254 } 2255 2256 data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE; 2257 if (type & PERF_SAMPLE_REGS_INTR) { 2258 OVERFLOW_CHECK_u64(array); 2259 data->intr_regs.abi = *array; 2260 array++; 2261 2262 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) { 2263 u64 mask = evsel->core.attr.sample_regs_intr; 2264 2265 sz = hweight64(mask) * sizeof(u64); 2266 OVERFLOW_CHECK(array, sz, max_size); 2267 data->intr_regs.mask = mask; 2268 data->intr_regs.regs = (u64 *)array; 2269 array = (void *)array + sz; 2270 } 2271 } 2272 2273 data->phys_addr = 0; 2274 if (type & PERF_SAMPLE_PHYS_ADDR) { 2275 data->phys_addr = *array; 2276 array++; 2277 } 2278 2279 data->cgroup = 0; 2280 if (type & PERF_SAMPLE_CGROUP) { 2281 data->cgroup = *array; 2282 array++; 2283 } 2284 2285 if (type & PERF_SAMPLE_AUX) { 2286 OVERFLOW_CHECK_u64(array); 2287 sz = *array++; 2288 2289 OVERFLOW_CHECK(array, sz, max_size); 2290 /* Undo swap of data */ 2291 if (swapped) 2292 mem_bswap_64((char *)array, sz); 2293 data->aux_sample.size = sz; 2294 data->aux_sample.data = (char *)array; 2295 array = (void *)array + sz; 2296 } 2297 2298 return 0; 2299 } 2300 2301 int perf_evsel__parse_sample_timestamp(struct evsel *evsel, 2302 union perf_event *event, 2303 u64 *timestamp) 2304 { 2305 u64 type = evsel->core.attr.sample_type; 2306 const __u64 *array; 2307 2308 if (!(type & PERF_SAMPLE_TIME)) 2309 return -1; 2310 2311 if (event->header.type != PERF_RECORD_SAMPLE) { 2312 struct perf_sample data = { 2313 .time = -1ULL, 2314 }; 2315 2316 if (!evsel->core.attr.sample_id_all) 2317 return -1; 2318 if (perf_evsel__parse_id_sample(evsel, event, &data)) 2319 return -1; 2320 2321 *timestamp = data.time; 2322 return 0; 2323 } 2324 2325 array = event->sample.array; 2326 2327 if (perf_event__check_size(event, evsel->sample_size)) 2328 return -EFAULT; 2329 2330 if (type & PERF_SAMPLE_IDENTIFIER) 2331 array++; 2332 2333 if (type & PERF_SAMPLE_IP) 2334 array++; 2335 2336 if (type & PERF_SAMPLE_TID) 2337 array++; 2338 2339 if (type & PERF_SAMPLE_TIME) 2340 *timestamp = *array; 2341 2342 return 0; 2343 } 2344 2345 struct tep_format_field *perf_evsel__field(struct evsel *evsel, const char *name) 2346 { 2347 return tep_find_field(evsel->tp_format, name); 2348 } 2349 2350 void *perf_evsel__rawptr(struct evsel *evsel, struct perf_sample *sample, 2351 const char *name) 2352 { 2353 struct tep_format_field *field = perf_evsel__field(evsel, name); 2354 int offset; 2355 2356 if (!field) 2357 return NULL; 2358 2359 offset = field->offset; 2360 2361 if (field->flags & TEP_FIELD_IS_DYNAMIC) { 2362 offset = *(int *)(sample->raw_data + field->offset); 2363 offset &= 0xffff; 2364 } 2365 2366 return sample->raw_data + offset; 2367 } 2368 2369 u64 format_field__intval(struct tep_format_field *field, struct perf_sample *sample, 2370 bool needs_swap) 2371 { 2372 u64 value; 2373 void *ptr = sample->raw_data + field->offset; 2374 2375 switch (field->size) { 2376 case 1: 2377 return *(u8 *)ptr; 2378 case 2: 2379 value = *(u16 *)ptr; 2380 break; 2381 case 4: 2382 value = *(u32 *)ptr; 2383 break; 2384 case 8: 2385 memcpy(&value, ptr, sizeof(u64)); 2386 break; 2387 default: 2388 return 0; 2389 } 2390 2391 if (!needs_swap) 2392 return value; 2393 2394 switch (field->size) { 2395 case 2: 2396 return bswap_16(value); 2397 case 4: 2398 return bswap_32(value); 2399 case 8: 2400 return bswap_64(value); 2401 default: 2402 return 0; 2403 } 2404 2405 return 0; 2406 } 2407 2408 u64 perf_evsel__intval(struct evsel *evsel, struct perf_sample *sample, 2409 const char *name) 2410 { 2411 struct tep_format_field *field = perf_evsel__field(evsel, name); 2412 2413 if (!field) 2414 return 0; 2415 2416 return field ? format_field__intval(field, sample, evsel->needs_swap) : 0; 2417 } 2418 2419 bool perf_evsel__fallback(struct evsel *evsel, int err, 2420 char *msg, size_t msgsize) 2421 { 2422 int paranoid; 2423 2424 if ((err == ENOENT || err == ENXIO || err == ENODEV) && 2425 evsel->core.attr.type == PERF_TYPE_HARDWARE && 2426 evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) { 2427 /* 2428 * If it's cycles then fall back to hrtimer based 2429 * cpu-clock-tick sw counter, which is always available even if 2430 * no PMU support. 2431 * 2432 * PPC returns ENXIO until 2.6.37 (behavior changed with commit 2433 * b0a873e). 2434 */ 2435 scnprintf(msg, msgsize, "%s", 2436 "The cycles event is not supported, trying to fall back to cpu-clock-ticks"); 2437 2438 evsel->core.attr.type = PERF_TYPE_SOFTWARE; 2439 evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK; 2440 2441 zfree(&evsel->name); 2442 return true; 2443 } else if (err == EACCES && !evsel->core.attr.exclude_kernel && 2444 (paranoid = perf_event_paranoid()) > 1) { 2445 const char *name = perf_evsel__name(evsel); 2446 char *new_name; 2447 const char *sep = ":"; 2448 2449 /* Is there already the separator in the name. */ 2450 if (strchr(name, '/') || 2451 strchr(name, ':')) 2452 sep = ""; 2453 2454 if (asprintf(&new_name, "%s%su", name, sep) < 0) 2455 return false; 2456 2457 if (evsel->name) 2458 free(evsel->name); 2459 evsel->name = new_name; 2460 scnprintf(msg, msgsize, "kernel.perf_event_paranoid=%d, trying " 2461 "to fall back to excluding kernel and hypervisor " 2462 " samples", paranoid); 2463 evsel->core.attr.exclude_kernel = 1; 2464 evsel->core.attr.exclude_hv = 1; 2465 2466 return true; 2467 } 2468 2469 return false; 2470 } 2471 2472 static bool find_process(const char *name) 2473 { 2474 size_t len = strlen(name); 2475 DIR *dir; 2476 struct dirent *d; 2477 int ret = -1; 2478 2479 dir = opendir(procfs__mountpoint()); 2480 if (!dir) 2481 return false; 2482 2483 /* Walk through the directory. */ 2484 while (ret && (d = readdir(dir)) != NULL) { 2485 char path[PATH_MAX]; 2486 char *data; 2487 size_t size; 2488 2489 if ((d->d_type != DT_DIR) || 2490 !strcmp(".", d->d_name) || 2491 !strcmp("..", d->d_name)) 2492 continue; 2493 2494 scnprintf(path, sizeof(path), "%s/%s/comm", 2495 procfs__mountpoint(), d->d_name); 2496 2497 if (filename__read_str(path, &data, &size)) 2498 continue; 2499 2500 ret = strncmp(name, data, len); 2501 free(data); 2502 } 2503 2504 closedir(dir); 2505 return ret ? false : true; 2506 } 2507 2508 int perf_evsel__open_strerror(struct evsel *evsel, struct target *target, 2509 int err, char *msg, size_t size) 2510 { 2511 char sbuf[STRERR_BUFSIZE]; 2512 int printed = 0; 2513 2514 switch (err) { 2515 case EPERM: 2516 case EACCES: 2517 if (err == EPERM) 2518 printed = scnprintf(msg, size, 2519 "No permission to enable %s event.\n\n", 2520 perf_evsel__name(evsel)); 2521 2522 return scnprintf(msg + printed, size - printed, 2523 "You may not have permission to collect %sstats.\n\n" 2524 "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n" 2525 "which controls use of the performance events system by\n" 2526 "unprivileged users (without CAP_SYS_ADMIN).\n\n" 2527 "The current value is %d:\n\n" 2528 " -1: Allow use of (almost) all events by all users\n" 2529 " Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n" 2530 ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n" 2531 " Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n" 2532 ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n" 2533 ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n" 2534 "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n" 2535 " kernel.perf_event_paranoid = -1\n" , 2536 target->system_wide ? "system-wide " : "", 2537 perf_event_paranoid()); 2538 case ENOENT: 2539 return scnprintf(msg, size, "The %s event is not supported.", 2540 perf_evsel__name(evsel)); 2541 case EMFILE: 2542 return scnprintf(msg, size, "%s", 2543 "Too many events are opened.\n" 2544 "Probably the maximum number of open file descriptors has been reached.\n" 2545 "Hint: Try again after reducing the number of events.\n" 2546 "Hint: Try increasing the limit with 'ulimit -n <limit>'"); 2547 case ENOMEM: 2548 if (evsel__has_callchain(evsel) && 2549 access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0) 2550 return scnprintf(msg, size, 2551 "Not enough memory to setup event with callchain.\n" 2552 "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n" 2553 "Hint: Current value: %d", sysctl__max_stack()); 2554 break; 2555 case ENODEV: 2556 if (target->cpu_list) 2557 return scnprintf(msg, size, "%s", 2558 "No such device - did you specify an out-of-range profile CPU?"); 2559 break; 2560 case EOPNOTSUPP: 2561 if (evsel->core.attr.sample_period != 0) 2562 return scnprintf(msg, size, 2563 "%s: PMU Hardware doesn't support sampling/overflow-interrupts. Try 'perf stat'", 2564 perf_evsel__name(evsel)); 2565 if (evsel->core.attr.precise_ip) 2566 return scnprintf(msg, size, "%s", 2567 "\'precise\' request may not be supported. Try removing 'p' modifier."); 2568 #if defined(__i386__) || defined(__x86_64__) 2569 if (evsel->core.attr.type == PERF_TYPE_HARDWARE) 2570 return scnprintf(msg, size, "%s", 2571 "No hardware sampling interrupt available.\n"); 2572 #endif 2573 break; 2574 case EBUSY: 2575 if (find_process("oprofiled")) 2576 return scnprintf(msg, size, 2577 "The PMU counters are busy/taken by another profiler.\n" 2578 "We found oprofile daemon running, please stop it and try again."); 2579 break; 2580 case EINVAL: 2581 if (evsel->core.attr.write_backward && perf_missing_features.write_backward) 2582 return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel."); 2583 if (perf_missing_features.clockid) 2584 return scnprintf(msg, size, "clockid feature not supported."); 2585 if (perf_missing_features.clockid_wrong) 2586 return scnprintf(msg, size, "wrong clockid (%d).", clockid); 2587 if (perf_missing_features.aux_output) 2588 return scnprintf(msg, size, "The 'aux_output' feature is not supported, update the kernel."); 2589 break; 2590 default: 2591 break; 2592 } 2593 2594 return scnprintf(msg, size, 2595 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n" 2596 "/bin/dmesg | grep -i perf may provide additional information.\n", 2597 err, str_error_r(err, sbuf, sizeof(sbuf)), 2598 perf_evsel__name(evsel)); 2599 } 2600 2601 struct perf_env *perf_evsel__env(struct evsel *evsel) 2602 { 2603 if (evsel && evsel->evlist) 2604 return evsel->evlist->env; 2605 return &perf_env; 2606 } 2607 2608 static int store_evsel_ids(struct evsel *evsel, struct evlist *evlist) 2609 { 2610 int cpu, thread; 2611 2612 for (cpu = 0; cpu < xyarray__max_x(evsel->core.fd); cpu++) { 2613 for (thread = 0; thread < xyarray__max_y(evsel->core.fd); 2614 thread++) { 2615 int fd = FD(evsel, cpu, thread); 2616 2617 if (perf_evlist__id_add_fd(&evlist->core, &evsel->core, 2618 cpu, thread, fd) < 0) 2619 return -1; 2620 } 2621 } 2622 2623 return 0; 2624 } 2625 2626 int perf_evsel__store_ids(struct evsel *evsel, struct evlist *evlist) 2627 { 2628 struct perf_cpu_map *cpus = evsel->core.cpus; 2629 struct perf_thread_map *threads = evsel->core.threads; 2630 2631 if (perf_evsel__alloc_id(&evsel->core, cpus->nr, threads->nr)) 2632 return -ENOMEM; 2633 2634 return store_evsel_ids(evsel, evlist); 2635 } 2636