1 /* 2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com> 3 * 4 * Parts came from builtin-{top,stat,record}.c, see those files for further 5 * copyright notes. 6 * 7 * Released under the GPL v2. (and only v2, not any later version) 8 */ 9 10 #include <byteswap.h> 11 #include <linux/bitops.h> 12 #include <api/fs/debugfs.h> 13 #include <traceevent/event-parse.h> 14 #include <linux/hw_breakpoint.h> 15 #include <linux/perf_event.h> 16 #include <sys/resource.h> 17 #include "asm/bug.h" 18 #include "callchain.h" 19 #include "cgroup.h" 20 #include "evsel.h" 21 #include "evlist.h" 22 #include "util.h" 23 #include "cpumap.h" 24 #include "thread_map.h" 25 #include "target.h" 26 #include "perf_regs.h" 27 #include "debug.h" 28 #include "trace-event.h" 29 30 static struct { 31 bool sample_id_all; 32 bool exclude_guest; 33 bool mmap2; 34 bool cloexec; 35 } perf_missing_features; 36 37 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused) 38 { 39 return 0; 40 } 41 42 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused) 43 { 44 } 45 46 static struct { 47 size_t size; 48 int (*init)(struct perf_evsel *evsel); 49 void (*fini)(struct perf_evsel *evsel); 50 } perf_evsel__object = { 51 .size = sizeof(struct perf_evsel), 52 .init = perf_evsel__no_extra_init, 53 .fini = perf_evsel__no_extra_fini, 54 }; 55 56 int perf_evsel__object_config(size_t object_size, 57 int (*init)(struct perf_evsel *evsel), 58 void (*fini)(struct perf_evsel *evsel)) 59 { 60 61 if (object_size == 0) 62 goto set_methods; 63 64 if (perf_evsel__object.size > object_size) 65 return -EINVAL; 66 67 perf_evsel__object.size = object_size; 68 69 set_methods: 70 if (init != NULL) 71 perf_evsel__object.init = init; 72 73 if (fini != NULL) 74 perf_evsel__object.fini = fini; 75 76 return 0; 77 } 78 79 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y)) 80 81 int __perf_evsel__sample_size(u64 sample_type) 82 { 83 u64 mask = sample_type & PERF_SAMPLE_MASK; 84 int size = 0; 85 int i; 86 87 for (i = 0; i < 64; i++) { 88 if (mask & (1ULL << i)) 89 size++; 90 } 91 92 size *= sizeof(u64); 93 94 return size; 95 } 96 97 /** 98 * __perf_evsel__calc_id_pos - calculate id_pos. 99 * @sample_type: sample type 100 * 101 * This function returns the position of the event id (PERF_SAMPLE_ID or 102 * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct 103 * sample_event. 104 */ 105 static int __perf_evsel__calc_id_pos(u64 sample_type) 106 { 107 int idx = 0; 108 109 if (sample_type & PERF_SAMPLE_IDENTIFIER) 110 return 0; 111 112 if (!(sample_type & PERF_SAMPLE_ID)) 113 return -1; 114 115 if (sample_type & PERF_SAMPLE_IP) 116 idx += 1; 117 118 if (sample_type & PERF_SAMPLE_TID) 119 idx += 1; 120 121 if (sample_type & PERF_SAMPLE_TIME) 122 idx += 1; 123 124 if (sample_type & PERF_SAMPLE_ADDR) 125 idx += 1; 126 127 return idx; 128 } 129 130 /** 131 * __perf_evsel__calc_is_pos - calculate is_pos. 132 * @sample_type: sample type 133 * 134 * This function returns the position (counting backwards) of the event id 135 * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if 136 * sample_id_all is used there is an id sample appended to non-sample events. 137 */ 138 static int __perf_evsel__calc_is_pos(u64 sample_type) 139 { 140 int idx = 1; 141 142 if (sample_type & PERF_SAMPLE_IDENTIFIER) 143 return 1; 144 145 if (!(sample_type & PERF_SAMPLE_ID)) 146 return -1; 147 148 if (sample_type & PERF_SAMPLE_CPU) 149 idx += 1; 150 151 if (sample_type & PERF_SAMPLE_STREAM_ID) 152 idx += 1; 153 154 return idx; 155 } 156 157 void perf_evsel__calc_id_pos(struct perf_evsel *evsel) 158 { 159 evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type); 160 evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type); 161 } 162 163 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel, 164 enum perf_event_sample_format bit) 165 { 166 if (!(evsel->attr.sample_type & bit)) { 167 evsel->attr.sample_type |= bit; 168 evsel->sample_size += sizeof(u64); 169 perf_evsel__calc_id_pos(evsel); 170 } 171 } 172 173 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel, 174 enum perf_event_sample_format bit) 175 { 176 if (evsel->attr.sample_type & bit) { 177 evsel->attr.sample_type &= ~bit; 178 evsel->sample_size -= sizeof(u64); 179 perf_evsel__calc_id_pos(evsel); 180 } 181 } 182 183 void perf_evsel__set_sample_id(struct perf_evsel *evsel, 184 bool can_sample_identifier) 185 { 186 if (can_sample_identifier) { 187 perf_evsel__reset_sample_bit(evsel, ID); 188 perf_evsel__set_sample_bit(evsel, IDENTIFIER); 189 } else { 190 perf_evsel__set_sample_bit(evsel, ID); 191 } 192 evsel->attr.read_format |= PERF_FORMAT_ID; 193 } 194 195 void perf_evsel__init(struct perf_evsel *evsel, 196 struct perf_event_attr *attr, int idx) 197 { 198 evsel->idx = idx; 199 evsel->tracking = !idx; 200 evsel->attr = *attr; 201 evsel->leader = evsel; 202 evsel->unit = ""; 203 evsel->scale = 1.0; 204 INIT_LIST_HEAD(&evsel->node); 205 perf_evsel__object.init(evsel); 206 evsel->sample_size = __perf_evsel__sample_size(attr->sample_type); 207 perf_evsel__calc_id_pos(evsel); 208 } 209 210 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx) 211 { 212 struct perf_evsel *evsel = zalloc(perf_evsel__object.size); 213 214 if (evsel != NULL) 215 perf_evsel__init(evsel, attr, idx); 216 217 return evsel; 218 } 219 220 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx) 221 { 222 struct perf_evsel *evsel = zalloc(perf_evsel__object.size); 223 224 if (evsel != NULL) { 225 struct perf_event_attr attr = { 226 .type = PERF_TYPE_TRACEPOINT, 227 .sample_type = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME | 228 PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD), 229 }; 230 231 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0) 232 goto out_free; 233 234 evsel->tp_format = trace_event__tp_format(sys, name); 235 if (evsel->tp_format == NULL) 236 goto out_free; 237 238 event_attr_init(&attr); 239 attr.config = evsel->tp_format->id; 240 attr.sample_period = 1; 241 perf_evsel__init(evsel, &attr, idx); 242 } 243 244 return evsel; 245 246 out_free: 247 zfree(&evsel->name); 248 free(evsel); 249 return NULL; 250 } 251 252 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = { 253 "cycles", 254 "instructions", 255 "cache-references", 256 "cache-misses", 257 "branches", 258 "branch-misses", 259 "bus-cycles", 260 "stalled-cycles-frontend", 261 "stalled-cycles-backend", 262 "ref-cycles", 263 }; 264 265 static const char *__perf_evsel__hw_name(u64 config) 266 { 267 if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config]) 268 return perf_evsel__hw_names[config]; 269 270 return "unknown-hardware"; 271 } 272 273 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size) 274 { 275 int colon = 0, r = 0; 276 struct perf_event_attr *attr = &evsel->attr; 277 bool exclude_guest_default = false; 278 279 #define MOD_PRINT(context, mod) do { \ 280 if (!attr->exclude_##context) { \ 281 if (!colon) colon = ++r; \ 282 r += scnprintf(bf + r, size - r, "%c", mod); \ 283 } } while(0) 284 285 if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) { 286 MOD_PRINT(kernel, 'k'); 287 MOD_PRINT(user, 'u'); 288 MOD_PRINT(hv, 'h'); 289 exclude_guest_default = true; 290 } 291 292 if (attr->precise_ip) { 293 if (!colon) 294 colon = ++r; 295 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp"); 296 exclude_guest_default = true; 297 } 298 299 if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) { 300 MOD_PRINT(host, 'H'); 301 MOD_PRINT(guest, 'G'); 302 } 303 #undef MOD_PRINT 304 if (colon) 305 bf[colon - 1] = ':'; 306 return r; 307 } 308 309 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size) 310 { 311 int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config)); 312 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 313 } 314 315 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = { 316 "cpu-clock", 317 "task-clock", 318 "page-faults", 319 "context-switches", 320 "cpu-migrations", 321 "minor-faults", 322 "major-faults", 323 "alignment-faults", 324 "emulation-faults", 325 "dummy", 326 }; 327 328 static const char *__perf_evsel__sw_name(u64 config) 329 { 330 if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config]) 331 return perf_evsel__sw_names[config]; 332 return "unknown-software"; 333 } 334 335 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size) 336 { 337 int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config)); 338 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 339 } 340 341 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type) 342 { 343 int r; 344 345 r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr); 346 347 if (type & HW_BREAKPOINT_R) 348 r += scnprintf(bf + r, size - r, "r"); 349 350 if (type & HW_BREAKPOINT_W) 351 r += scnprintf(bf + r, size - r, "w"); 352 353 if (type & HW_BREAKPOINT_X) 354 r += scnprintf(bf + r, size - r, "x"); 355 356 return r; 357 } 358 359 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size) 360 { 361 struct perf_event_attr *attr = &evsel->attr; 362 int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type); 363 return r + perf_evsel__add_modifiers(evsel, bf + r, size - r); 364 } 365 366 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX] 367 [PERF_EVSEL__MAX_ALIASES] = { 368 { "L1-dcache", "l1-d", "l1d", "L1-data", }, 369 { "L1-icache", "l1-i", "l1i", "L1-instruction", }, 370 { "LLC", "L2", }, 371 { "dTLB", "d-tlb", "Data-TLB", }, 372 { "iTLB", "i-tlb", "Instruction-TLB", }, 373 { "branch", "branches", "bpu", "btb", "bpc", }, 374 { "node", }, 375 }; 376 377 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX] 378 [PERF_EVSEL__MAX_ALIASES] = { 379 { "load", "loads", "read", }, 380 { "store", "stores", "write", }, 381 { "prefetch", "prefetches", "speculative-read", "speculative-load", }, 382 }; 383 384 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX] 385 [PERF_EVSEL__MAX_ALIASES] = { 386 { "refs", "Reference", "ops", "access", }, 387 { "misses", "miss", }, 388 }; 389 390 #define C(x) PERF_COUNT_HW_CACHE_##x 391 #define CACHE_READ (1 << C(OP_READ)) 392 #define CACHE_WRITE (1 << C(OP_WRITE)) 393 #define CACHE_PREFETCH (1 << C(OP_PREFETCH)) 394 #define COP(x) (1 << x) 395 396 /* 397 * cache operartion stat 398 * L1I : Read and prefetch only 399 * ITLB and BPU : Read-only 400 */ 401 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = { 402 [C(L1D)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 403 [C(L1I)] = (CACHE_READ | CACHE_PREFETCH), 404 [C(LL)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 405 [C(DTLB)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 406 [C(ITLB)] = (CACHE_READ), 407 [C(BPU)] = (CACHE_READ), 408 [C(NODE)] = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH), 409 }; 410 411 bool perf_evsel__is_cache_op_valid(u8 type, u8 op) 412 { 413 if (perf_evsel__hw_cache_stat[type] & COP(op)) 414 return true; /* valid */ 415 else 416 return false; /* invalid */ 417 } 418 419 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result, 420 char *bf, size_t size) 421 { 422 if (result) { 423 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0], 424 perf_evsel__hw_cache_op[op][0], 425 perf_evsel__hw_cache_result[result][0]); 426 } 427 428 return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0], 429 perf_evsel__hw_cache_op[op][1]); 430 } 431 432 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size) 433 { 434 u8 op, result, type = (config >> 0) & 0xff; 435 const char *err = "unknown-ext-hardware-cache-type"; 436 437 if (type > PERF_COUNT_HW_CACHE_MAX) 438 goto out_err; 439 440 op = (config >> 8) & 0xff; 441 err = "unknown-ext-hardware-cache-op"; 442 if (op > PERF_COUNT_HW_CACHE_OP_MAX) 443 goto out_err; 444 445 result = (config >> 16) & 0xff; 446 err = "unknown-ext-hardware-cache-result"; 447 if (result > PERF_COUNT_HW_CACHE_RESULT_MAX) 448 goto out_err; 449 450 err = "invalid-cache"; 451 if (!perf_evsel__is_cache_op_valid(type, op)) 452 goto out_err; 453 454 return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size); 455 out_err: 456 return scnprintf(bf, size, "%s", err); 457 } 458 459 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size) 460 { 461 int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size); 462 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 463 } 464 465 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size) 466 { 467 int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config); 468 return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret); 469 } 470 471 const char *perf_evsel__name(struct perf_evsel *evsel) 472 { 473 char bf[128]; 474 475 if (evsel->name) 476 return evsel->name; 477 478 switch (evsel->attr.type) { 479 case PERF_TYPE_RAW: 480 perf_evsel__raw_name(evsel, bf, sizeof(bf)); 481 break; 482 483 case PERF_TYPE_HARDWARE: 484 perf_evsel__hw_name(evsel, bf, sizeof(bf)); 485 break; 486 487 case PERF_TYPE_HW_CACHE: 488 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf)); 489 break; 490 491 case PERF_TYPE_SOFTWARE: 492 perf_evsel__sw_name(evsel, bf, sizeof(bf)); 493 break; 494 495 case PERF_TYPE_TRACEPOINT: 496 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint"); 497 break; 498 499 case PERF_TYPE_BREAKPOINT: 500 perf_evsel__bp_name(evsel, bf, sizeof(bf)); 501 break; 502 503 default: 504 scnprintf(bf, sizeof(bf), "unknown attr type: %d", 505 evsel->attr.type); 506 break; 507 } 508 509 evsel->name = strdup(bf); 510 511 return evsel->name ?: "unknown"; 512 } 513 514 const char *perf_evsel__group_name(struct perf_evsel *evsel) 515 { 516 return evsel->group_name ?: "anon group"; 517 } 518 519 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size) 520 { 521 int ret; 522 struct perf_evsel *pos; 523 const char *group_name = perf_evsel__group_name(evsel); 524 525 ret = scnprintf(buf, size, "%s", group_name); 526 527 ret += scnprintf(buf + ret, size - ret, " { %s", 528 perf_evsel__name(evsel)); 529 530 for_each_group_member(pos, evsel) 531 ret += scnprintf(buf + ret, size - ret, ", %s", 532 perf_evsel__name(pos)); 533 534 ret += scnprintf(buf + ret, size - ret, " }"); 535 536 return ret; 537 } 538 539 static void 540 perf_evsel__config_callgraph(struct perf_evsel *evsel) 541 { 542 bool function = perf_evsel__is_function_event(evsel); 543 struct perf_event_attr *attr = &evsel->attr; 544 545 perf_evsel__set_sample_bit(evsel, CALLCHAIN); 546 547 if (callchain_param.record_mode == CALLCHAIN_DWARF) { 548 if (!function) { 549 perf_evsel__set_sample_bit(evsel, REGS_USER); 550 perf_evsel__set_sample_bit(evsel, STACK_USER); 551 attr->sample_regs_user = PERF_REGS_MASK; 552 attr->sample_stack_user = callchain_param.dump_size; 553 attr->exclude_callchain_user = 1; 554 } else { 555 pr_info("Cannot use DWARF unwind for function trace event," 556 " falling back to framepointers.\n"); 557 } 558 } 559 560 if (function) { 561 pr_info("Disabling user space callchains for function trace event.\n"); 562 attr->exclude_callchain_user = 1; 563 } 564 } 565 566 /* 567 * The enable_on_exec/disabled value strategy: 568 * 569 * 1) For any type of traced program: 570 * - all independent events and group leaders are disabled 571 * - all group members are enabled 572 * 573 * Group members are ruled by group leaders. They need to 574 * be enabled, because the group scheduling relies on that. 575 * 576 * 2) For traced programs executed by perf: 577 * - all independent events and group leaders have 578 * enable_on_exec set 579 * - we don't specifically enable or disable any event during 580 * the record command 581 * 582 * Independent events and group leaders are initially disabled 583 * and get enabled by exec. Group members are ruled by group 584 * leaders as stated in 1). 585 * 586 * 3) For traced programs attached by perf (pid/tid): 587 * - we specifically enable or disable all events during 588 * the record command 589 * 590 * When attaching events to already running traced we 591 * enable/disable events specifically, as there's no 592 * initial traced exec call. 593 */ 594 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts) 595 { 596 struct perf_evsel *leader = evsel->leader; 597 struct perf_event_attr *attr = &evsel->attr; 598 int track = evsel->tracking; 599 bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread; 600 601 attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1; 602 attr->inherit = !opts->no_inherit; 603 604 perf_evsel__set_sample_bit(evsel, IP); 605 perf_evsel__set_sample_bit(evsel, TID); 606 607 if (evsel->sample_read) { 608 perf_evsel__set_sample_bit(evsel, READ); 609 610 /* 611 * We need ID even in case of single event, because 612 * PERF_SAMPLE_READ process ID specific data. 613 */ 614 perf_evsel__set_sample_id(evsel, false); 615 616 /* 617 * Apply group format only if we belong to group 618 * with more than one members. 619 */ 620 if (leader->nr_members > 1) { 621 attr->read_format |= PERF_FORMAT_GROUP; 622 attr->inherit = 0; 623 } 624 } 625 626 /* 627 * We default some events to have a default interval. But keep 628 * it a weak assumption overridable by the user. 629 */ 630 if (!attr->sample_period || (opts->user_freq != UINT_MAX || 631 opts->user_interval != ULLONG_MAX)) { 632 if (opts->freq) { 633 perf_evsel__set_sample_bit(evsel, PERIOD); 634 attr->freq = 1; 635 attr->sample_freq = opts->freq; 636 } else { 637 attr->sample_period = opts->default_interval; 638 } 639 } 640 641 /* 642 * Disable sampling for all group members other 643 * than leader in case leader 'leads' the sampling. 644 */ 645 if ((leader != evsel) && leader->sample_read) { 646 attr->sample_freq = 0; 647 attr->sample_period = 0; 648 } 649 650 if (opts->no_samples) 651 attr->sample_freq = 0; 652 653 if (opts->inherit_stat) 654 attr->inherit_stat = 1; 655 656 if (opts->sample_address) { 657 perf_evsel__set_sample_bit(evsel, ADDR); 658 attr->mmap_data = track; 659 } 660 661 if (callchain_param.enabled && !evsel->no_aux_samples) 662 perf_evsel__config_callgraph(evsel); 663 664 if (target__has_cpu(&opts->target)) 665 perf_evsel__set_sample_bit(evsel, CPU); 666 667 if (opts->period) 668 perf_evsel__set_sample_bit(evsel, PERIOD); 669 670 /* 671 * When the user explicitely disabled time don't force it here. 672 */ 673 if (opts->sample_time && 674 (!perf_missing_features.sample_id_all && 675 (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu))) 676 perf_evsel__set_sample_bit(evsel, TIME); 677 678 if (opts->raw_samples && !evsel->no_aux_samples) { 679 perf_evsel__set_sample_bit(evsel, TIME); 680 perf_evsel__set_sample_bit(evsel, RAW); 681 perf_evsel__set_sample_bit(evsel, CPU); 682 } 683 684 if (opts->sample_address) 685 perf_evsel__set_sample_bit(evsel, DATA_SRC); 686 687 if (opts->no_buffering) { 688 attr->watermark = 0; 689 attr->wakeup_events = 1; 690 } 691 if (opts->branch_stack && !evsel->no_aux_samples) { 692 perf_evsel__set_sample_bit(evsel, BRANCH_STACK); 693 attr->branch_sample_type = opts->branch_stack; 694 } 695 696 if (opts->sample_weight) 697 perf_evsel__set_sample_bit(evsel, WEIGHT); 698 699 attr->mmap = track; 700 attr->mmap2 = track && !perf_missing_features.mmap2; 701 attr->comm = track; 702 703 if (opts->sample_transaction) 704 perf_evsel__set_sample_bit(evsel, TRANSACTION); 705 706 /* 707 * XXX see the function comment above 708 * 709 * Disabling only independent events or group leaders, 710 * keeping group members enabled. 711 */ 712 if (perf_evsel__is_group_leader(evsel)) 713 attr->disabled = 1; 714 715 /* 716 * Setting enable_on_exec for independent events and 717 * group leaders for traced executed by perf. 718 */ 719 if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) && 720 !opts->initial_delay) 721 attr->enable_on_exec = 1; 722 723 if (evsel->immediate) { 724 attr->disabled = 0; 725 attr->enable_on_exec = 0; 726 } 727 } 728 729 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads) 730 { 731 int cpu, thread; 732 733 if (evsel->system_wide) 734 nthreads = 1; 735 736 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int)); 737 738 if (evsel->fd) { 739 for (cpu = 0; cpu < ncpus; cpu++) { 740 for (thread = 0; thread < nthreads; thread++) { 741 FD(evsel, cpu, thread) = -1; 742 } 743 } 744 } 745 746 return evsel->fd != NULL ? 0 : -ENOMEM; 747 } 748 749 static int perf_evsel__run_ioctl(struct perf_evsel *evsel, int ncpus, int nthreads, 750 int ioc, void *arg) 751 { 752 int cpu, thread; 753 754 if (evsel->system_wide) 755 nthreads = 1; 756 757 for (cpu = 0; cpu < ncpus; cpu++) { 758 for (thread = 0; thread < nthreads; thread++) { 759 int fd = FD(evsel, cpu, thread), 760 err = ioctl(fd, ioc, arg); 761 762 if (err) 763 return err; 764 } 765 } 766 767 return 0; 768 } 769 770 int perf_evsel__set_filter(struct perf_evsel *evsel, int ncpus, int nthreads, 771 const char *filter) 772 { 773 return perf_evsel__run_ioctl(evsel, ncpus, nthreads, 774 PERF_EVENT_IOC_SET_FILTER, 775 (void *)filter); 776 } 777 778 int perf_evsel__enable(struct perf_evsel *evsel, int ncpus, int nthreads) 779 { 780 return perf_evsel__run_ioctl(evsel, ncpus, nthreads, 781 PERF_EVENT_IOC_ENABLE, 782 0); 783 } 784 785 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads) 786 { 787 if (evsel->system_wide) 788 nthreads = 1; 789 790 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id)); 791 if (evsel->sample_id == NULL) 792 return -ENOMEM; 793 794 evsel->id = zalloc(ncpus * nthreads * sizeof(u64)); 795 if (evsel->id == NULL) { 796 xyarray__delete(evsel->sample_id); 797 evsel->sample_id = NULL; 798 return -ENOMEM; 799 } 800 801 return 0; 802 } 803 804 void perf_evsel__reset_counts(struct perf_evsel *evsel, int ncpus) 805 { 806 memset(evsel->counts, 0, (sizeof(*evsel->counts) + 807 (ncpus * sizeof(struct perf_counts_values)))); 808 } 809 810 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus) 811 { 812 evsel->counts = zalloc((sizeof(*evsel->counts) + 813 (ncpus * sizeof(struct perf_counts_values)))); 814 return evsel->counts != NULL ? 0 : -ENOMEM; 815 } 816 817 static void perf_evsel__free_fd(struct perf_evsel *evsel) 818 { 819 xyarray__delete(evsel->fd); 820 evsel->fd = NULL; 821 } 822 823 static void perf_evsel__free_id(struct perf_evsel *evsel) 824 { 825 xyarray__delete(evsel->sample_id); 826 evsel->sample_id = NULL; 827 zfree(&evsel->id); 828 } 829 830 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads) 831 { 832 int cpu, thread; 833 834 if (evsel->system_wide) 835 nthreads = 1; 836 837 for (cpu = 0; cpu < ncpus; cpu++) 838 for (thread = 0; thread < nthreads; ++thread) { 839 close(FD(evsel, cpu, thread)); 840 FD(evsel, cpu, thread) = -1; 841 } 842 } 843 844 void perf_evsel__free_counts(struct perf_evsel *evsel) 845 { 846 zfree(&evsel->counts); 847 } 848 849 void perf_evsel__exit(struct perf_evsel *evsel) 850 { 851 assert(list_empty(&evsel->node)); 852 perf_evsel__free_fd(evsel); 853 perf_evsel__free_id(evsel); 854 close_cgroup(evsel->cgrp); 855 zfree(&evsel->group_name); 856 if (evsel->tp_format) 857 pevent_free_format(evsel->tp_format); 858 zfree(&evsel->name); 859 perf_evsel__object.fini(evsel); 860 } 861 862 void perf_evsel__delete(struct perf_evsel *evsel) 863 { 864 perf_evsel__exit(evsel); 865 free(evsel); 866 } 867 868 static inline void compute_deltas(struct perf_evsel *evsel, 869 int cpu, 870 struct perf_counts_values *count) 871 { 872 struct perf_counts_values tmp; 873 874 if (!evsel->prev_raw_counts) 875 return; 876 877 if (cpu == -1) { 878 tmp = evsel->prev_raw_counts->aggr; 879 evsel->prev_raw_counts->aggr = *count; 880 } else { 881 tmp = evsel->prev_raw_counts->cpu[cpu]; 882 evsel->prev_raw_counts->cpu[cpu] = *count; 883 } 884 885 count->val = count->val - tmp.val; 886 count->ena = count->ena - tmp.ena; 887 count->run = count->run - tmp.run; 888 } 889 890 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel, 891 int cpu, int thread, bool scale) 892 { 893 struct perf_counts_values count; 894 size_t nv = scale ? 3 : 1; 895 896 if (FD(evsel, cpu, thread) < 0) 897 return -EINVAL; 898 899 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0) 900 return -ENOMEM; 901 902 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0) 903 return -errno; 904 905 compute_deltas(evsel, cpu, &count); 906 907 if (scale) { 908 if (count.run == 0) 909 count.val = 0; 910 else if (count.run < count.ena) 911 count.val = (u64)((double)count.val * count.ena / count.run + 0.5); 912 } else 913 count.ena = count.run = 0; 914 915 evsel->counts->cpu[cpu] = count; 916 return 0; 917 } 918 919 int __perf_evsel__read(struct perf_evsel *evsel, 920 int ncpus, int nthreads, bool scale) 921 { 922 size_t nv = scale ? 3 : 1; 923 int cpu, thread; 924 struct perf_counts_values *aggr = &evsel->counts->aggr, count; 925 926 if (evsel->system_wide) 927 nthreads = 1; 928 929 aggr->val = aggr->ena = aggr->run = 0; 930 931 for (cpu = 0; cpu < ncpus; cpu++) { 932 for (thread = 0; thread < nthreads; thread++) { 933 if (FD(evsel, cpu, thread) < 0) 934 continue; 935 936 if (readn(FD(evsel, cpu, thread), 937 &count, nv * sizeof(u64)) < 0) 938 return -errno; 939 940 aggr->val += count.val; 941 if (scale) { 942 aggr->ena += count.ena; 943 aggr->run += count.run; 944 } 945 } 946 } 947 948 compute_deltas(evsel, -1, aggr); 949 950 evsel->counts->scaled = 0; 951 if (scale) { 952 if (aggr->run == 0) { 953 evsel->counts->scaled = -1; 954 aggr->val = 0; 955 return 0; 956 } 957 958 if (aggr->run < aggr->ena) { 959 evsel->counts->scaled = 1; 960 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5); 961 } 962 } else 963 aggr->ena = aggr->run = 0; 964 965 return 0; 966 } 967 968 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread) 969 { 970 struct perf_evsel *leader = evsel->leader; 971 int fd; 972 973 if (perf_evsel__is_group_leader(evsel)) 974 return -1; 975 976 /* 977 * Leader must be already processed/open, 978 * if not it's a bug. 979 */ 980 BUG_ON(!leader->fd); 981 982 fd = FD(leader, cpu, thread); 983 BUG_ON(fd == -1); 984 985 return fd; 986 } 987 988 #define __PRINT_ATTR(fmt, cast, field) \ 989 fprintf(fp, " %-19s "fmt"\n", #field, cast attr->field) 990 991 #define PRINT_ATTR_U32(field) __PRINT_ATTR("%u" , , field) 992 #define PRINT_ATTR_X32(field) __PRINT_ATTR("%#x", , field) 993 #define PRINT_ATTR_U64(field) __PRINT_ATTR("%" PRIu64, (uint64_t), field) 994 #define PRINT_ATTR_X64(field) __PRINT_ATTR("%#"PRIx64, (uint64_t), field) 995 996 #define PRINT_ATTR2N(name1, field1, name2, field2) \ 997 fprintf(fp, " %-19s %u %-19s %u\n", \ 998 name1, attr->field1, name2, attr->field2) 999 1000 #define PRINT_ATTR2(field1, field2) \ 1001 PRINT_ATTR2N(#field1, field1, #field2, field2) 1002 1003 static size_t perf_event_attr__fprintf(struct perf_event_attr *attr, FILE *fp) 1004 { 1005 size_t ret = 0; 1006 1007 ret += fprintf(fp, "%.60s\n", graph_dotted_line); 1008 ret += fprintf(fp, "perf_event_attr:\n"); 1009 1010 ret += PRINT_ATTR_U32(type); 1011 ret += PRINT_ATTR_U32(size); 1012 ret += PRINT_ATTR_X64(config); 1013 ret += PRINT_ATTR_U64(sample_period); 1014 ret += PRINT_ATTR_U64(sample_freq); 1015 ret += PRINT_ATTR_X64(sample_type); 1016 ret += PRINT_ATTR_X64(read_format); 1017 1018 ret += PRINT_ATTR2(disabled, inherit); 1019 ret += PRINT_ATTR2(pinned, exclusive); 1020 ret += PRINT_ATTR2(exclude_user, exclude_kernel); 1021 ret += PRINT_ATTR2(exclude_hv, exclude_idle); 1022 ret += PRINT_ATTR2(mmap, comm); 1023 ret += PRINT_ATTR2(mmap2, comm_exec); 1024 ret += PRINT_ATTR2(freq, inherit_stat); 1025 ret += PRINT_ATTR2(enable_on_exec, task); 1026 ret += PRINT_ATTR2(watermark, precise_ip); 1027 ret += PRINT_ATTR2(mmap_data, sample_id_all); 1028 ret += PRINT_ATTR2(exclude_host, exclude_guest); 1029 ret += PRINT_ATTR2N("excl.callchain_kern", exclude_callchain_kernel, 1030 "excl.callchain_user", exclude_callchain_user); 1031 1032 ret += PRINT_ATTR_U32(wakeup_events); 1033 ret += PRINT_ATTR_U32(wakeup_watermark); 1034 ret += PRINT_ATTR_X32(bp_type); 1035 ret += PRINT_ATTR_X64(bp_addr); 1036 ret += PRINT_ATTR_X64(config1); 1037 ret += PRINT_ATTR_U64(bp_len); 1038 ret += PRINT_ATTR_X64(config2); 1039 ret += PRINT_ATTR_X64(branch_sample_type); 1040 ret += PRINT_ATTR_X64(sample_regs_user); 1041 ret += PRINT_ATTR_U32(sample_stack_user); 1042 1043 ret += fprintf(fp, "%.60s\n", graph_dotted_line); 1044 1045 return ret; 1046 } 1047 1048 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus, 1049 struct thread_map *threads) 1050 { 1051 int cpu, thread, nthreads; 1052 unsigned long flags = PERF_FLAG_FD_CLOEXEC; 1053 int pid = -1, err; 1054 enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE; 1055 1056 if (evsel->system_wide) 1057 nthreads = 1; 1058 else 1059 nthreads = threads->nr; 1060 1061 if (evsel->fd == NULL && 1062 perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0) 1063 return -ENOMEM; 1064 1065 if (evsel->cgrp) { 1066 flags |= PERF_FLAG_PID_CGROUP; 1067 pid = evsel->cgrp->fd; 1068 } 1069 1070 fallback_missing_features: 1071 if (perf_missing_features.cloexec) 1072 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC; 1073 if (perf_missing_features.mmap2) 1074 evsel->attr.mmap2 = 0; 1075 if (perf_missing_features.exclude_guest) 1076 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0; 1077 retry_sample_id: 1078 if (perf_missing_features.sample_id_all) 1079 evsel->attr.sample_id_all = 0; 1080 1081 if (verbose >= 2) 1082 perf_event_attr__fprintf(&evsel->attr, stderr); 1083 1084 for (cpu = 0; cpu < cpus->nr; cpu++) { 1085 1086 for (thread = 0; thread < nthreads; thread++) { 1087 int group_fd; 1088 1089 if (!evsel->cgrp && !evsel->system_wide) 1090 pid = threads->map[thread]; 1091 1092 group_fd = get_group_fd(evsel, cpu, thread); 1093 retry_open: 1094 pr_debug2("sys_perf_event_open: pid %d cpu %d group_fd %d flags %#lx\n", 1095 pid, cpus->map[cpu], group_fd, flags); 1096 1097 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr, 1098 pid, 1099 cpus->map[cpu], 1100 group_fd, flags); 1101 if (FD(evsel, cpu, thread) < 0) { 1102 err = -errno; 1103 pr_debug2("sys_perf_event_open failed, error %d\n", 1104 err); 1105 goto try_fallback; 1106 } 1107 set_rlimit = NO_CHANGE; 1108 } 1109 } 1110 1111 return 0; 1112 1113 try_fallback: 1114 /* 1115 * perf stat needs between 5 and 22 fds per CPU. When we run out 1116 * of them try to increase the limits. 1117 */ 1118 if (err == -EMFILE && set_rlimit < INCREASED_MAX) { 1119 struct rlimit l; 1120 int old_errno = errno; 1121 1122 if (getrlimit(RLIMIT_NOFILE, &l) == 0) { 1123 if (set_rlimit == NO_CHANGE) 1124 l.rlim_cur = l.rlim_max; 1125 else { 1126 l.rlim_cur = l.rlim_max + 1000; 1127 l.rlim_max = l.rlim_cur; 1128 } 1129 if (setrlimit(RLIMIT_NOFILE, &l) == 0) { 1130 set_rlimit++; 1131 errno = old_errno; 1132 goto retry_open; 1133 } 1134 } 1135 errno = old_errno; 1136 } 1137 1138 if (err != -EINVAL || cpu > 0 || thread > 0) 1139 goto out_close; 1140 1141 if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) { 1142 perf_missing_features.cloexec = true; 1143 goto fallback_missing_features; 1144 } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) { 1145 perf_missing_features.mmap2 = true; 1146 goto fallback_missing_features; 1147 } else if (!perf_missing_features.exclude_guest && 1148 (evsel->attr.exclude_guest || evsel->attr.exclude_host)) { 1149 perf_missing_features.exclude_guest = true; 1150 goto fallback_missing_features; 1151 } else if (!perf_missing_features.sample_id_all) { 1152 perf_missing_features.sample_id_all = true; 1153 goto retry_sample_id; 1154 } 1155 1156 out_close: 1157 do { 1158 while (--thread >= 0) { 1159 close(FD(evsel, cpu, thread)); 1160 FD(evsel, cpu, thread) = -1; 1161 } 1162 thread = nthreads; 1163 } while (--cpu >= 0); 1164 return err; 1165 } 1166 1167 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads) 1168 { 1169 if (evsel->fd == NULL) 1170 return; 1171 1172 perf_evsel__close_fd(evsel, ncpus, nthreads); 1173 perf_evsel__free_fd(evsel); 1174 } 1175 1176 static struct { 1177 struct cpu_map map; 1178 int cpus[1]; 1179 } empty_cpu_map = { 1180 .map.nr = 1, 1181 .cpus = { -1, }, 1182 }; 1183 1184 static struct { 1185 struct thread_map map; 1186 int threads[1]; 1187 } empty_thread_map = { 1188 .map.nr = 1, 1189 .threads = { -1, }, 1190 }; 1191 1192 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus, 1193 struct thread_map *threads) 1194 { 1195 if (cpus == NULL) { 1196 /* Work around old compiler warnings about strict aliasing */ 1197 cpus = &empty_cpu_map.map; 1198 } 1199 1200 if (threads == NULL) 1201 threads = &empty_thread_map.map; 1202 1203 return __perf_evsel__open(evsel, cpus, threads); 1204 } 1205 1206 int perf_evsel__open_per_cpu(struct perf_evsel *evsel, 1207 struct cpu_map *cpus) 1208 { 1209 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map); 1210 } 1211 1212 int perf_evsel__open_per_thread(struct perf_evsel *evsel, 1213 struct thread_map *threads) 1214 { 1215 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads); 1216 } 1217 1218 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel, 1219 const union perf_event *event, 1220 struct perf_sample *sample) 1221 { 1222 u64 type = evsel->attr.sample_type; 1223 const u64 *array = event->sample.array; 1224 bool swapped = evsel->needs_swap; 1225 union u64_swap u; 1226 1227 array += ((event->header.size - 1228 sizeof(event->header)) / sizeof(u64)) - 1; 1229 1230 if (type & PERF_SAMPLE_IDENTIFIER) { 1231 sample->id = *array; 1232 array--; 1233 } 1234 1235 if (type & PERF_SAMPLE_CPU) { 1236 u.val64 = *array; 1237 if (swapped) { 1238 /* undo swap of u64, then swap on individual u32s */ 1239 u.val64 = bswap_64(u.val64); 1240 u.val32[0] = bswap_32(u.val32[0]); 1241 } 1242 1243 sample->cpu = u.val32[0]; 1244 array--; 1245 } 1246 1247 if (type & PERF_SAMPLE_STREAM_ID) { 1248 sample->stream_id = *array; 1249 array--; 1250 } 1251 1252 if (type & PERF_SAMPLE_ID) { 1253 sample->id = *array; 1254 array--; 1255 } 1256 1257 if (type & PERF_SAMPLE_TIME) { 1258 sample->time = *array; 1259 array--; 1260 } 1261 1262 if (type & PERF_SAMPLE_TID) { 1263 u.val64 = *array; 1264 if (swapped) { 1265 /* undo swap of u64, then swap on individual u32s */ 1266 u.val64 = bswap_64(u.val64); 1267 u.val32[0] = bswap_32(u.val32[0]); 1268 u.val32[1] = bswap_32(u.val32[1]); 1269 } 1270 1271 sample->pid = u.val32[0]; 1272 sample->tid = u.val32[1]; 1273 array--; 1274 } 1275 1276 return 0; 1277 } 1278 1279 static inline bool overflow(const void *endp, u16 max_size, const void *offset, 1280 u64 size) 1281 { 1282 return size > max_size || offset + size > endp; 1283 } 1284 1285 #define OVERFLOW_CHECK(offset, size, max_size) \ 1286 do { \ 1287 if (overflow(endp, (max_size), (offset), (size))) \ 1288 return -EFAULT; \ 1289 } while (0) 1290 1291 #define OVERFLOW_CHECK_u64(offset) \ 1292 OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64)) 1293 1294 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event, 1295 struct perf_sample *data) 1296 { 1297 u64 type = evsel->attr.sample_type; 1298 bool swapped = evsel->needs_swap; 1299 const u64 *array; 1300 u16 max_size = event->header.size; 1301 const void *endp = (void *)event + max_size; 1302 u64 sz; 1303 1304 /* 1305 * used for cross-endian analysis. See git commit 65014ab3 1306 * for why this goofiness is needed. 1307 */ 1308 union u64_swap u; 1309 1310 memset(data, 0, sizeof(*data)); 1311 data->cpu = data->pid = data->tid = -1; 1312 data->stream_id = data->id = data->time = -1ULL; 1313 data->period = evsel->attr.sample_period; 1314 data->weight = 0; 1315 1316 if (event->header.type != PERF_RECORD_SAMPLE) { 1317 if (!evsel->attr.sample_id_all) 1318 return 0; 1319 return perf_evsel__parse_id_sample(evsel, event, data); 1320 } 1321 1322 array = event->sample.array; 1323 1324 /* 1325 * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes 1326 * up to PERF_SAMPLE_PERIOD. After that overflow() must be used to 1327 * check the format does not go past the end of the event. 1328 */ 1329 if (evsel->sample_size + sizeof(event->header) > event->header.size) 1330 return -EFAULT; 1331 1332 data->id = -1ULL; 1333 if (type & PERF_SAMPLE_IDENTIFIER) { 1334 data->id = *array; 1335 array++; 1336 } 1337 1338 if (type & PERF_SAMPLE_IP) { 1339 data->ip = *array; 1340 array++; 1341 } 1342 1343 if (type & PERF_SAMPLE_TID) { 1344 u.val64 = *array; 1345 if (swapped) { 1346 /* undo swap of u64, then swap on individual u32s */ 1347 u.val64 = bswap_64(u.val64); 1348 u.val32[0] = bswap_32(u.val32[0]); 1349 u.val32[1] = bswap_32(u.val32[1]); 1350 } 1351 1352 data->pid = u.val32[0]; 1353 data->tid = u.val32[1]; 1354 array++; 1355 } 1356 1357 if (type & PERF_SAMPLE_TIME) { 1358 data->time = *array; 1359 array++; 1360 } 1361 1362 data->addr = 0; 1363 if (type & PERF_SAMPLE_ADDR) { 1364 data->addr = *array; 1365 array++; 1366 } 1367 1368 if (type & PERF_SAMPLE_ID) { 1369 data->id = *array; 1370 array++; 1371 } 1372 1373 if (type & PERF_SAMPLE_STREAM_ID) { 1374 data->stream_id = *array; 1375 array++; 1376 } 1377 1378 if (type & PERF_SAMPLE_CPU) { 1379 1380 u.val64 = *array; 1381 if (swapped) { 1382 /* undo swap of u64, then swap on individual u32s */ 1383 u.val64 = bswap_64(u.val64); 1384 u.val32[0] = bswap_32(u.val32[0]); 1385 } 1386 1387 data->cpu = u.val32[0]; 1388 array++; 1389 } 1390 1391 if (type & PERF_SAMPLE_PERIOD) { 1392 data->period = *array; 1393 array++; 1394 } 1395 1396 if (type & PERF_SAMPLE_READ) { 1397 u64 read_format = evsel->attr.read_format; 1398 1399 OVERFLOW_CHECK_u64(array); 1400 if (read_format & PERF_FORMAT_GROUP) 1401 data->read.group.nr = *array; 1402 else 1403 data->read.one.value = *array; 1404 1405 array++; 1406 1407 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1408 OVERFLOW_CHECK_u64(array); 1409 data->read.time_enabled = *array; 1410 array++; 1411 } 1412 1413 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1414 OVERFLOW_CHECK_u64(array); 1415 data->read.time_running = *array; 1416 array++; 1417 } 1418 1419 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1420 if (read_format & PERF_FORMAT_GROUP) { 1421 const u64 max_group_nr = UINT64_MAX / 1422 sizeof(struct sample_read_value); 1423 1424 if (data->read.group.nr > max_group_nr) 1425 return -EFAULT; 1426 sz = data->read.group.nr * 1427 sizeof(struct sample_read_value); 1428 OVERFLOW_CHECK(array, sz, max_size); 1429 data->read.group.values = 1430 (struct sample_read_value *)array; 1431 array = (void *)array + sz; 1432 } else { 1433 OVERFLOW_CHECK_u64(array); 1434 data->read.one.id = *array; 1435 array++; 1436 } 1437 } 1438 1439 if (type & PERF_SAMPLE_CALLCHAIN) { 1440 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64); 1441 1442 OVERFLOW_CHECK_u64(array); 1443 data->callchain = (struct ip_callchain *)array++; 1444 if (data->callchain->nr > max_callchain_nr) 1445 return -EFAULT; 1446 sz = data->callchain->nr * sizeof(u64); 1447 OVERFLOW_CHECK(array, sz, max_size); 1448 array = (void *)array + sz; 1449 } 1450 1451 if (type & PERF_SAMPLE_RAW) { 1452 OVERFLOW_CHECK_u64(array); 1453 u.val64 = *array; 1454 if (WARN_ONCE(swapped, 1455 "Endianness of raw data not corrected!\n")) { 1456 /* undo swap of u64, then swap on individual u32s */ 1457 u.val64 = bswap_64(u.val64); 1458 u.val32[0] = bswap_32(u.val32[0]); 1459 u.val32[1] = bswap_32(u.val32[1]); 1460 } 1461 data->raw_size = u.val32[0]; 1462 array = (void *)array + sizeof(u32); 1463 1464 OVERFLOW_CHECK(array, data->raw_size, max_size); 1465 data->raw_data = (void *)array; 1466 array = (void *)array + data->raw_size; 1467 } 1468 1469 if (type & PERF_SAMPLE_BRANCH_STACK) { 1470 const u64 max_branch_nr = UINT64_MAX / 1471 sizeof(struct branch_entry); 1472 1473 OVERFLOW_CHECK_u64(array); 1474 data->branch_stack = (struct branch_stack *)array++; 1475 1476 if (data->branch_stack->nr > max_branch_nr) 1477 return -EFAULT; 1478 sz = data->branch_stack->nr * sizeof(struct branch_entry); 1479 OVERFLOW_CHECK(array, sz, max_size); 1480 array = (void *)array + sz; 1481 } 1482 1483 if (type & PERF_SAMPLE_REGS_USER) { 1484 OVERFLOW_CHECK_u64(array); 1485 data->user_regs.abi = *array; 1486 array++; 1487 1488 if (data->user_regs.abi) { 1489 u64 mask = evsel->attr.sample_regs_user; 1490 1491 sz = hweight_long(mask) * sizeof(u64); 1492 OVERFLOW_CHECK(array, sz, max_size); 1493 data->user_regs.mask = mask; 1494 data->user_regs.regs = (u64 *)array; 1495 array = (void *)array + sz; 1496 } 1497 } 1498 1499 if (type & PERF_SAMPLE_STACK_USER) { 1500 OVERFLOW_CHECK_u64(array); 1501 sz = *array++; 1502 1503 data->user_stack.offset = ((char *)(array - 1) 1504 - (char *) event); 1505 1506 if (!sz) { 1507 data->user_stack.size = 0; 1508 } else { 1509 OVERFLOW_CHECK(array, sz, max_size); 1510 data->user_stack.data = (char *)array; 1511 array = (void *)array + sz; 1512 OVERFLOW_CHECK_u64(array); 1513 data->user_stack.size = *array++; 1514 if (WARN_ONCE(data->user_stack.size > sz, 1515 "user stack dump failure\n")) 1516 return -EFAULT; 1517 } 1518 } 1519 1520 data->weight = 0; 1521 if (type & PERF_SAMPLE_WEIGHT) { 1522 OVERFLOW_CHECK_u64(array); 1523 data->weight = *array; 1524 array++; 1525 } 1526 1527 data->data_src = PERF_MEM_DATA_SRC_NONE; 1528 if (type & PERF_SAMPLE_DATA_SRC) { 1529 OVERFLOW_CHECK_u64(array); 1530 data->data_src = *array; 1531 array++; 1532 } 1533 1534 data->transaction = 0; 1535 if (type & PERF_SAMPLE_TRANSACTION) { 1536 OVERFLOW_CHECK_u64(array); 1537 data->transaction = *array; 1538 array++; 1539 } 1540 1541 return 0; 1542 } 1543 1544 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, 1545 u64 read_format) 1546 { 1547 size_t sz, result = sizeof(struct sample_event); 1548 1549 if (type & PERF_SAMPLE_IDENTIFIER) 1550 result += sizeof(u64); 1551 1552 if (type & PERF_SAMPLE_IP) 1553 result += sizeof(u64); 1554 1555 if (type & PERF_SAMPLE_TID) 1556 result += sizeof(u64); 1557 1558 if (type & PERF_SAMPLE_TIME) 1559 result += sizeof(u64); 1560 1561 if (type & PERF_SAMPLE_ADDR) 1562 result += sizeof(u64); 1563 1564 if (type & PERF_SAMPLE_ID) 1565 result += sizeof(u64); 1566 1567 if (type & PERF_SAMPLE_STREAM_ID) 1568 result += sizeof(u64); 1569 1570 if (type & PERF_SAMPLE_CPU) 1571 result += sizeof(u64); 1572 1573 if (type & PERF_SAMPLE_PERIOD) 1574 result += sizeof(u64); 1575 1576 if (type & PERF_SAMPLE_READ) { 1577 result += sizeof(u64); 1578 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1579 result += sizeof(u64); 1580 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1581 result += sizeof(u64); 1582 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1583 if (read_format & PERF_FORMAT_GROUP) { 1584 sz = sample->read.group.nr * 1585 sizeof(struct sample_read_value); 1586 result += sz; 1587 } else { 1588 result += sizeof(u64); 1589 } 1590 } 1591 1592 if (type & PERF_SAMPLE_CALLCHAIN) { 1593 sz = (sample->callchain->nr + 1) * sizeof(u64); 1594 result += sz; 1595 } 1596 1597 if (type & PERF_SAMPLE_RAW) { 1598 result += sizeof(u32); 1599 result += sample->raw_size; 1600 } 1601 1602 if (type & PERF_SAMPLE_BRANCH_STACK) { 1603 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1604 sz += sizeof(u64); 1605 result += sz; 1606 } 1607 1608 if (type & PERF_SAMPLE_REGS_USER) { 1609 if (sample->user_regs.abi) { 1610 result += sizeof(u64); 1611 sz = hweight_long(sample->user_regs.mask) * sizeof(u64); 1612 result += sz; 1613 } else { 1614 result += sizeof(u64); 1615 } 1616 } 1617 1618 if (type & PERF_SAMPLE_STACK_USER) { 1619 sz = sample->user_stack.size; 1620 result += sizeof(u64); 1621 if (sz) { 1622 result += sz; 1623 result += sizeof(u64); 1624 } 1625 } 1626 1627 if (type & PERF_SAMPLE_WEIGHT) 1628 result += sizeof(u64); 1629 1630 if (type & PERF_SAMPLE_DATA_SRC) 1631 result += sizeof(u64); 1632 1633 if (type & PERF_SAMPLE_TRANSACTION) 1634 result += sizeof(u64); 1635 1636 return result; 1637 } 1638 1639 int perf_event__synthesize_sample(union perf_event *event, u64 type, 1640 u64 read_format, 1641 const struct perf_sample *sample, 1642 bool swapped) 1643 { 1644 u64 *array; 1645 size_t sz; 1646 /* 1647 * used for cross-endian analysis. See git commit 65014ab3 1648 * for why this goofiness is needed. 1649 */ 1650 union u64_swap u; 1651 1652 array = event->sample.array; 1653 1654 if (type & PERF_SAMPLE_IDENTIFIER) { 1655 *array = sample->id; 1656 array++; 1657 } 1658 1659 if (type & PERF_SAMPLE_IP) { 1660 *array = sample->ip; 1661 array++; 1662 } 1663 1664 if (type & PERF_SAMPLE_TID) { 1665 u.val32[0] = sample->pid; 1666 u.val32[1] = sample->tid; 1667 if (swapped) { 1668 /* 1669 * Inverse of what is done in perf_evsel__parse_sample 1670 */ 1671 u.val32[0] = bswap_32(u.val32[0]); 1672 u.val32[1] = bswap_32(u.val32[1]); 1673 u.val64 = bswap_64(u.val64); 1674 } 1675 1676 *array = u.val64; 1677 array++; 1678 } 1679 1680 if (type & PERF_SAMPLE_TIME) { 1681 *array = sample->time; 1682 array++; 1683 } 1684 1685 if (type & PERF_SAMPLE_ADDR) { 1686 *array = sample->addr; 1687 array++; 1688 } 1689 1690 if (type & PERF_SAMPLE_ID) { 1691 *array = sample->id; 1692 array++; 1693 } 1694 1695 if (type & PERF_SAMPLE_STREAM_ID) { 1696 *array = sample->stream_id; 1697 array++; 1698 } 1699 1700 if (type & PERF_SAMPLE_CPU) { 1701 u.val32[0] = sample->cpu; 1702 if (swapped) { 1703 /* 1704 * Inverse of what is done in perf_evsel__parse_sample 1705 */ 1706 u.val32[0] = bswap_32(u.val32[0]); 1707 u.val64 = bswap_64(u.val64); 1708 } 1709 *array = u.val64; 1710 array++; 1711 } 1712 1713 if (type & PERF_SAMPLE_PERIOD) { 1714 *array = sample->period; 1715 array++; 1716 } 1717 1718 if (type & PERF_SAMPLE_READ) { 1719 if (read_format & PERF_FORMAT_GROUP) 1720 *array = sample->read.group.nr; 1721 else 1722 *array = sample->read.one.value; 1723 array++; 1724 1725 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1726 *array = sample->read.time_enabled; 1727 array++; 1728 } 1729 1730 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1731 *array = sample->read.time_running; 1732 array++; 1733 } 1734 1735 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1736 if (read_format & PERF_FORMAT_GROUP) { 1737 sz = sample->read.group.nr * 1738 sizeof(struct sample_read_value); 1739 memcpy(array, sample->read.group.values, sz); 1740 array = (void *)array + sz; 1741 } else { 1742 *array = sample->read.one.id; 1743 array++; 1744 } 1745 } 1746 1747 if (type & PERF_SAMPLE_CALLCHAIN) { 1748 sz = (sample->callchain->nr + 1) * sizeof(u64); 1749 memcpy(array, sample->callchain, sz); 1750 array = (void *)array + sz; 1751 } 1752 1753 if (type & PERF_SAMPLE_RAW) { 1754 u.val32[0] = sample->raw_size; 1755 if (WARN_ONCE(swapped, 1756 "Endianness of raw data not corrected!\n")) { 1757 /* 1758 * Inverse of what is done in perf_evsel__parse_sample 1759 */ 1760 u.val32[0] = bswap_32(u.val32[0]); 1761 u.val32[1] = bswap_32(u.val32[1]); 1762 u.val64 = bswap_64(u.val64); 1763 } 1764 *array = u.val64; 1765 array = (void *)array + sizeof(u32); 1766 1767 memcpy(array, sample->raw_data, sample->raw_size); 1768 array = (void *)array + sample->raw_size; 1769 } 1770 1771 if (type & PERF_SAMPLE_BRANCH_STACK) { 1772 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1773 sz += sizeof(u64); 1774 memcpy(array, sample->branch_stack, sz); 1775 array = (void *)array + sz; 1776 } 1777 1778 if (type & PERF_SAMPLE_REGS_USER) { 1779 if (sample->user_regs.abi) { 1780 *array++ = sample->user_regs.abi; 1781 sz = hweight_long(sample->user_regs.mask) * sizeof(u64); 1782 memcpy(array, sample->user_regs.regs, sz); 1783 array = (void *)array + sz; 1784 } else { 1785 *array++ = 0; 1786 } 1787 } 1788 1789 if (type & PERF_SAMPLE_STACK_USER) { 1790 sz = sample->user_stack.size; 1791 *array++ = sz; 1792 if (sz) { 1793 memcpy(array, sample->user_stack.data, sz); 1794 array = (void *)array + sz; 1795 *array++ = sz; 1796 } 1797 } 1798 1799 if (type & PERF_SAMPLE_WEIGHT) { 1800 *array = sample->weight; 1801 array++; 1802 } 1803 1804 if (type & PERF_SAMPLE_DATA_SRC) { 1805 *array = sample->data_src; 1806 array++; 1807 } 1808 1809 if (type & PERF_SAMPLE_TRANSACTION) { 1810 *array = sample->transaction; 1811 array++; 1812 } 1813 1814 return 0; 1815 } 1816 1817 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name) 1818 { 1819 return pevent_find_field(evsel->tp_format, name); 1820 } 1821 1822 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample, 1823 const char *name) 1824 { 1825 struct format_field *field = perf_evsel__field(evsel, name); 1826 int offset; 1827 1828 if (!field) 1829 return NULL; 1830 1831 offset = field->offset; 1832 1833 if (field->flags & FIELD_IS_DYNAMIC) { 1834 offset = *(int *)(sample->raw_data + field->offset); 1835 offset &= 0xffff; 1836 } 1837 1838 return sample->raw_data + offset; 1839 } 1840 1841 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample, 1842 const char *name) 1843 { 1844 struct format_field *field = perf_evsel__field(evsel, name); 1845 void *ptr; 1846 u64 value; 1847 1848 if (!field) 1849 return 0; 1850 1851 ptr = sample->raw_data + field->offset; 1852 1853 switch (field->size) { 1854 case 1: 1855 return *(u8 *)ptr; 1856 case 2: 1857 value = *(u16 *)ptr; 1858 break; 1859 case 4: 1860 value = *(u32 *)ptr; 1861 break; 1862 case 8: 1863 value = *(u64 *)ptr; 1864 break; 1865 default: 1866 return 0; 1867 } 1868 1869 if (!evsel->needs_swap) 1870 return value; 1871 1872 switch (field->size) { 1873 case 2: 1874 return bswap_16(value); 1875 case 4: 1876 return bswap_32(value); 1877 case 8: 1878 return bswap_64(value); 1879 default: 1880 return 0; 1881 } 1882 1883 return 0; 1884 } 1885 1886 static int comma_fprintf(FILE *fp, bool *first, const char *fmt, ...) 1887 { 1888 va_list args; 1889 int ret = 0; 1890 1891 if (!*first) { 1892 ret += fprintf(fp, ","); 1893 } else { 1894 ret += fprintf(fp, ":"); 1895 *first = false; 1896 } 1897 1898 va_start(args, fmt); 1899 ret += vfprintf(fp, fmt, args); 1900 va_end(args); 1901 return ret; 1902 } 1903 1904 static int __if_fprintf(FILE *fp, bool *first, const char *field, u64 value) 1905 { 1906 if (value == 0) 1907 return 0; 1908 1909 return comma_fprintf(fp, first, " %s: %" PRIu64, field, value); 1910 } 1911 1912 #define if_print(field) printed += __if_fprintf(fp, &first, #field, evsel->attr.field) 1913 1914 struct bit_names { 1915 int bit; 1916 const char *name; 1917 }; 1918 1919 static int bits__fprintf(FILE *fp, const char *field, u64 value, 1920 struct bit_names *bits, bool *first) 1921 { 1922 int i = 0, printed = comma_fprintf(fp, first, " %s: ", field); 1923 bool first_bit = true; 1924 1925 do { 1926 if (value & bits[i].bit) { 1927 printed += fprintf(fp, "%s%s", first_bit ? "" : "|", bits[i].name); 1928 first_bit = false; 1929 } 1930 } while (bits[++i].name != NULL); 1931 1932 return printed; 1933 } 1934 1935 static int sample_type__fprintf(FILE *fp, bool *first, u64 value) 1936 { 1937 #define bit_name(n) { PERF_SAMPLE_##n, #n } 1938 struct bit_names bits[] = { 1939 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR), 1940 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU), 1941 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW), 1942 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER), 1943 bit_name(IDENTIFIER), 1944 { .name = NULL, } 1945 }; 1946 #undef bit_name 1947 return bits__fprintf(fp, "sample_type", value, bits, first); 1948 } 1949 1950 static int read_format__fprintf(FILE *fp, bool *first, u64 value) 1951 { 1952 #define bit_name(n) { PERF_FORMAT_##n, #n } 1953 struct bit_names bits[] = { 1954 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING), 1955 bit_name(ID), bit_name(GROUP), 1956 { .name = NULL, } 1957 }; 1958 #undef bit_name 1959 return bits__fprintf(fp, "read_format", value, bits, first); 1960 } 1961 1962 int perf_evsel__fprintf(struct perf_evsel *evsel, 1963 struct perf_attr_details *details, FILE *fp) 1964 { 1965 bool first = true; 1966 int printed = 0; 1967 1968 if (details->event_group) { 1969 struct perf_evsel *pos; 1970 1971 if (!perf_evsel__is_group_leader(evsel)) 1972 return 0; 1973 1974 if (evsel->nr_members > 1) 1975 printed += fprintf(fp, "%s{", evsel->group_name ?: ""); 1976 1977 printed += fprintf(fp, "%s", perf_evsel__name(evsel)); 1978 for_each_group_member(pos, evsel) 1979 printed += fprintf(fp, ",%s", perf_evsel__name(pos)); 1980 1981 if (evsel->nr_members > 1) 1982 printed += fprintf(fp, "}"); 1983 goto out; 1984 } 1985 1986 printed += fprintf(fp, "%s", perf_evsel__name(evsel)); 1987 1988 if (details->verbose || details->freq) { 1989 printed += comma_fprintf(fp, &first, " sample_freq=%" PRIu64, 1990 (u64)evsel->attr.sample_freq); 1991 } 1992 1993 if (details->verbose) { 1994 if_print(type); 1995 if_print(config); 1996 if_print(config1); 1997 if_print(config2); 1998 if_print(size); 1999 printed += sample_type__fprintf(fp, &first, evsel->attr.sample_type); 2000 if (evsel->attr.read_format) 2001 printed += read_format__fprintf(fp, &first, evsel->attr.read_format); 2002 if_print(disabled); 2003 if_print(inherit); 2004 if_print(pinned); 2005 if_print(exclusive); 2006 if_print(exclude_user); 2007 if_print(exclude_kernel); 2008 if_print(exclude_hv); 2009 if_print(exclude_idle); 2010 if_print(mmap); 2011 if_print(mmap2); 2012 if_print(comm); 2013 if_print(comm_exec); 2014 if_print(freq); 2015 if_print(inherit_stat); 2016 if_print(enable_on_exec); 2017 if_print(task); 2018 if_print(watermark); 2019 if_print(precise_ip); 2020 if_print(mmap_data); 2021 if_print(sample_id_all); 2022 if_print(exclude_host); 2023 if_print(exclude_guest); 2024 if_print(__reserved_1); 2025 if_print(wakeup_events); 2026 if_print(bp_type); 2027 if_print(branch_sample_type); 2028 } 2029 out: 2030 fputc('\n', fp); 2031 return ++printed; 2032 } 2033 2034 bool perf_evsel__fallback(struct perf_evsel *evsel, int err, 2035 char *msg, size_t msgsize) 2036 { 2037 if ((err == ENOENT || err == ENXIO || err == ENODEV) && 2038 evsel->attr.type == PERF_TYPE_HARDWARE && 2039 evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) { 2040 /* 2041 * If it's cycles then fall back to hrtimer based 2042 * cpu-clock-tick sw counter, which is always available even if 2043 * no PMU support. 2044 * 2045 * PPC returns ENXIO until 2.6.37 (behavior changed with commit 2046 * b0a873e). 2047 */ 2048 scnprintf(msg, msgsize, "%s", 2049 "The cycles event is not supported, trying to fall back to cpu-clock-ticks"); 2050 2051 evsel->attr.type = PERF_TYPE_SOFTWARE; 2052 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK; 2053 2054 zfree(&evsel->name); 2055 return true; 2056 } 2057 2058 return false; 2059 } 2060 2061 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target, 2062 int err, char *msg, size_t size) 2063 { 2064 char sbuf[STRERR_BUFSIZE]; 2065 2066 switch (err) { 2067 case EPERM: 2068 case EACCES: 2069 return scnprintf(msg, size, 2070 "You may not have permission to collect %sstats.\n" 2071 "Consider tweaking /proc/sys/kernel/perf_event_paranoid:\n" 2072 " -1 - Not paranoid at all\n" 2073 " 0 - Disallow raw tracepoint access for unpriv\n" 2074 " 1 - Disallow cpu events for unpriv\n" 2075 " 2 - Disallow kernel profiling for unpriv", 2076 target->system_wide ? "system-wide " : ""); 2077 case ENOENT: 2078 return scnprintf(msg, size, "The %s event is not supported.", 2079 perf_evsel__name(evsel)); 2080 case EMFILE: 2081 return scnprintf(msg, size, "%s", 2082 "Too many events are opened.\n" 2083 "Try again after reducing the number of events."); 2084 case ENODEV: 2085 if (target->cpu_list) 2086 return scnprintf(msg, size, "%s", 2087 "No such device - did you specify an out-of-range profile CPU?\n"); 2088 break; 2089 case EOPNOTSUPP: 2090 if (evsel->attr.precise_ip) 2091 return scnprintf(msg, size, "%s", 2092 "\'precise\' request may not be supported. Try removing 'p' modifier."); 2093 #if defined(__i386__) || defined(__x86_64__) 2094 if (evsel->attr.type == PERF_TYPE_HARDWARE) 2095 return scnprintf(msg, size, "%s", 2096 "No hardware sampling interrupt available.\n" 2097 "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it."); 2098 #endif 2099 break; 2100 case EBUSY: 2101 if (find_process("oprofiled")) 2102 return scnprintf(msg, size, 2103 "The PMU counters are busy/taken by another profiler.\n" 2104 "We found oprofile daemon running, please stop it and try again."); 2105 break; 2106 default: 2107 break; 2108 } 2109 2110 return scnprintf(msg, size, 2111 "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n" 2112 "/bin/dmesg may provide additional information.\n" 2113 "No CONFIG_PERF_EVENTS=y kernel support configured?\n", 2114 err, strerror_r(err, sbuf, sizeof(sbuf)), 2115 perf_evsel__name(evsel)); 2116 } 2117