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