1 #define _FILE_OFFSET_BITS 64 2 3 #include <linux/kernel.h> 4 5 #include <byteswap.h> 6 #include <unistd.h> 7 #include <sys/types.h> 8 #include <sys/mman.h> 9 10 #include "evlist.h" 11 #include "evsel.h" 12 #include "session.h" 13 #include "tool.h" 14 #include "sort.h" 15 #include "util.h" 16 #include "cpumap.h" 17 18 static int perf_session__open(struct perf_session *self, bool force) 19 { 20 struct stat input_stat; 21 22 if (!strcmp(self->filename, "-")) { 23 self->fd_pipe = true; 24 self->fd = STDIN_FILENO; 25 26 if (perf_session__read_header(self, self->fd) < 0) 27 pr_err("incompatible file format"); 28 29 return 0; 30 } 31 32 self->fd = open(self->filename, O_RDONLY); 33 if (self->fd < 0) { 34 int err = errno; 35 36 pr_err("failed to open %s: %s", self->filename, strerror(err)); 37 if (err == ENOENT && !strcmp(self->filename, "perf.data")) 38 pr_err(" (try 'perf record' first)"); 39 pr_err("\n"); 40 return -errno; 41 } 42 43 if (fstat(self->fd, &input_stat) < 0) 44 goto out_close; 45 46 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) { 47 pr_err("file %s not owned by current user or root\n", 48 self->filename); 49 goto out_close; 50 } 51 52 if (!input_stat.st_size) { 53 pr_info("zero-sized file (%s), nothing to do!\n", 54 self->filename); 55 goto out_close; 56 } 57 58 if (perf_session__read_header(self, self->fd) < 0) { 59 pr_err("incompatible file format"); 60 goto out_close; 61 } 62 63 if (!perf_evlist__valid_sample_type(self->evlist)) { 64 pr_err("non matching sample_type"); 65 goto out_close; 66 } 67 68 if (!perf_evlist__valid_sample_id_all(self->evlist)) { 69 pr_err("non matching sample_id_all"); 70 goto out_close; 71 } 72 73 self->size = input_stat.st_size; 74 return 0; 75 76 out_close: 77 close(self->fd); 78 self->fd = -1; 79 return -1; 80 } 81 82 void perf_session__update_sample_type(struct perf_session *self) 83 { 84 self->sample_type = perf_evlist__sample_type(self->evlist); 85 self->sample_size = __perf_evsel__sample_size(self->sample_type); 86 self->sample_id_all = perf_evlist__sample_id_all(self->evlist); 87 self->id_hdr_size = perf_evlist__id_hdr_size(self->evlist); 88 self->host_machine.id_hdr_size = self->id_hdr_size; 89 } 90 91 int perf_session__create_kernel_maps(struct perf_session *self) 92 { 93 int ret = machine__create_kernel_maps(&self->host_machine); 94 95 if (ret >= 0) 96 ret = machines__create_guest_kernel_maps(&self->machines); 97 return ret; 98 } 99 100 static void perf_session__destroy_kernel_maps(struct perf_session *self) 101 { 102 machine__destroy_kernel_maps(&self->host_machine); 103 machines__destroy_guest_kernel_maps(&self->machines); 104 } 105 106 struct perf_session *perf_session__new(const char *filename, int mode, 107 bool force, bool repipe, 108 struct perf_tool *tool) 109 { 110 struct perf_session *self; 111 struct stat st; 112 size_t len; 113 114 if (!filename || !strlen(filename)) { 115 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 116 filename = "-"; 117 else 118 filename = "perf.data"; 119 } 120 121 len = strlen(filename); 122 self = zalloc(sizeof(*self) + len); 123 124 if (self == NULL) 125 goto out; 126 127 memcpy(self->filename, filename, len); 128 /* 129 * On 64bit we can mmap the data file in one go. No need for tiny mmap 130 * slices. On 32bit we use 32MB. 131 */ 132 #if BITS_PER_LONG == 64 133 self->mmap_window = ULLONG_MAX; 134 #else 135 self->mmap_window = 32 * 1024 * 1024ULL; 136 #endif 137 self->machines = RB_ROOT; 138 self->repipe = repipe; 139 INIT_LIST_HEAD(&self->ordered_samples.samples); 140 INIT_LIST_HEAD(&self->ordered_samples.sample_cache); 141 INIT_LIST_HEAD(&self->ordered_samples.to_free); 142 machine__init(&self->host_machine, "", HOST_KERNEL_ID); 143 144 if (mode == O_RDONLY) { 145 if (perf_session__open(self, force) < 0) 146 goto out_delete; 147 perf_session__update_sample_type(self); 148 } else if (mode == O_WRONLY) { 149 /* 150 * In O_RDONLY mode this will be performed when reading the 151 * kernel MMAP event, in perf_event__process_mmap(). 152 */ 153 if (perf_session__create_kernel_maps(self) < 0) 154 goto out_delete; 155 } 156 157 if (tool && tool->ordering_requires_timestamps && 158 tool->ordered_samples && !self->sample_id_all) { 159 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 160 tool->ordered_samples = false; 161 } 162 163 out: 164 return self; 165 out_delete: 166 perf_session__delete(self); 167 return NULL; 168 } 169 170 static void machine__delete_dead_threads(struct machine *machine) 171 { 172 struct thread *n, *t; 173 174 list_for_each_entry_safe(t, n, &machine->dead_threads, node) { 175 list_del(&t->node); 176 thread__delete(t); 177 } 178 } 179 180 static void perf_session__delete_dead_threads(struct perf_session *session) 181 { 182 machine__delete_dead_threads(&session->host_machine); 183 } 184 185 static void machine__delete_threads(struct machine *self) 186 { 187 struct rb_node *nd = rb_first(&self->threads); 188 189 while (nd) { 190 struct thread *t = rb_entry(nd, struct thread, rb_node); 191 192 rb_erase(&t->rb_node, &self->threads); 193 nd = rb_next(nd); 194 thread__delete(t); 195 } 196 } 197 198 static void perf_session__delete_threads(struct perf_session *session) 199 { 200 machine__delete_threads(&session->host_machine); 201 } 202 203 void perf_session__delete(struct perf_session *self) 204 { 205 perf_session__destroy_kernel_maps(self); 206 perf_session__delete_dead_threads(self); 207 perf_session__delete_threads(self); 208 machine__exit(&self->host_machine); 209 close(self->fd); 210 free(self); 211 } 212 213 void machine__remove_thread(struct machine *self, struct thread *th) 214 { 215 self->last_match = NULL; 216 rb_erase(&th->rb_node, &self->threads); 217 /* 218 * We may have references to this thread, for instance in some hist_entry 219 * instances, so just move them to a separate list. 220 */ 221 list_add_tail(&th->node, &self->dead_threads); 222 } 223 224 static bool symbol__match_parent_regex(struct symbol *sym) 225 { 226 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0)) 227 return 1; 228 229 return 0; 230 } 231 232 int machine__resolve_callchain(struct machine *self, struct perf_evsel *evsel, 233 struct thread *thread, 234 struct ip_callchain *chain, 235 struct symbol **parent) 236 { 237 u8 cpumode = PERF_RECORD_MISC_USER; 238 unsigned int i; 239 int err; 240 241 callchain_cursor_reset(&evsel->hists.callchain_cursor); 242 243 for (i = 0; i < chain->nr; i++) { 244 u64 ip; 245 struct addr_location al; 246 247 if (callchain_param.order == ORDER_CALLEE) 248 ip = chain->ips[i]; 249 else 250 ip = chain->ips[chain->nr - i - 1]; 251 252 if (ip >= PERF_CONTEXT_MAX) { 253 switch (ip) { 254 case PERF_CONTEXT_HV: 255 cpumode = PERF_RECORD_MISC_HYPERVISOR; break; 256 case PERF_CONTEXT_KERNEL: 257 cpumode = PERF_RECORD_MISC_KERNEL; break; 258 case PERF_CONTEXT_USER: 259 cpumode = PERF_RECORD_MISC_USER; break; 260 default: 261 break; 262 } 263 continue; 264 } 265 266 al.filtered = false; 267 thread__find_addr_location(thread, self, cpumode, 268 MAP__FUNCTION, ip, &al, NULL); 269 if (al.sym != NULL) { 270 if (sort__has_parent && !*parent && 271 symbol__match_parent_regex(al.sym)) 272 *parent = al.sym; 273 if (!symbol_conf.use_callchain) 274 break; 275 } 276 277 err = callchain_cursor_append(&evsel->hists.callchain_cursor, 278 ip, al.map, al.sym); 279 if (err) 280 return err; 281 } 282 283 return 0; 284 } 285 286 static int process_event_synth_tracing_data_stub(union perf_event *event __used, 287 struct perf_session *session __used) 288 { 289 dump_printf(": unhandled!\n"); 290 return 0; 291 } 292 293 static int process_event_synth_attr_stub(union perf_event *event __used, 294 struct perf_evlist **pevlist __used) 295 { 296 dump_printf(": unhandled!\n"); 297 return 0; 298 } 299 300 static int process_event_sample_stub(struct perf_tool *tool __used, 301 union perf_event *event __used, 302 struct perf_sample *sample __used, 303 struct perf_evsel *evsel __used, 304 struct machine *machine __used) 305 { 306 dump_printf(": unhandled!\n"); 307 return 0; 308 } 309 310 static int process_event_stub(struct perf_tool *tool __used, 311 union perf_event *event __used, 312 struct perf_sample *sample __used, 313 struct machine *machine __used) 314 { 315 dump_printf(": unhandled!\n"); 316 return 0; 317 } 318 319 static int process_finished_round_stub(struct perf_tool *tool __used, 320 union perf_event *event __used, 321 struct perf_session *perf_session __used) 322 { 323 dump_printf(": unhandled!\n"); 324 return 0; 325 } 326 327 static int process_event_type_stub(struct perf_tool *tool __used, 328 union perf_event *event __used) 329 { 330 dump_printf(": unhandled!\n"); 331 return 0; 332 } 333 334 static int process_finished_round(struct perf_tool *tool, 335 union perf_event *event, 336 struct perf_session *session); 337 338 static void perf_tool__fill_defaults(struct perf_tool *tool) 339 { 340 if (tool->sample == NULL) 341 tool->sample = process_event_sample_stub; 342 if (tool->mmap == NULL) 343 tool->mmap = process_event_stub; 344 if (tool->comm == NULL) 345 tool->comm = process_event_stub; 346 if (tool->fork == NULL) 347 tool->fork = process_event_stub; 348 if (tool->exit == NULL) 349 tool->exit = process_event_stub; 350 if (tool->lost == NULL) 351 tool->lost = perf_event__process_lost; 352 if (tool->read == NULL) 353 tool->read = process_event_sample_stub; 354 if (tool->throttle == NULL) 355 tool->throttle = process_event_stub; 356 if (tool->unthrottle == NULL) 357 tool->unthrottle = process_event_stub; 358 if (tool->attr == NULL) 359 tool->attr = process_event_synth_attr_stub; 360 if (tool->event_type == NULL) 361 tool->event_type = process_event_type_stub; 362 if (tool->tracing_data == NULL) 363 tool->tracing_data = process_event_synth_tracing_data_stub; 364 if (tool->build_id == NULL) 365 tool->build_id = process_finished_round_stub; 366 if (tool->finished_round == NULL) { 367 if (tool->ordered_samples) 368 tool->finished_round = process_finished_round; 369 else 370 tool->finished_round = process_finished_round_stub; 371 } 372 } 373 374 void mem_bswap_64(void *src, int byte_size) 375 { 376 u64 *m = src; 377 378 while (byte_size > 0) { 379 *m = bswap_64(*m); 380 byte_size -= sizeof(u64); 381 ++m; 382 } 383 } 384 385 static void perf_event__all64_swap(union perf_event *event) 386 { 387 struct perf_event_header *hdr = &event->header; 388 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 389 } 390 391 static void perf_event__comm_swap(union perf_event *event) 392 { 393 event->comm.pid = bswap_32(event->comm.pid); 394 event->comm.tid = bswap_32(event->comm.tid); 395 } 396 397 static void perf_event__mmap_swap(union perf_event *event) 398 { 399 event->mmap.pid = bswap_32(event->mmap.pid); 400 event->mmap.tid = bswap_32(event->mmap.tid); 401 event->mmap.start = bswap_64(event->mmap.start); 402 event->mmap.len = bswap_64(event->mmap.len); 403 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 404 } 405 406 static void perf_event__task_swap(union perf_event *event) 407 { 408 event->fork.pid = bswap_32(event->fork.pid); 409 event->fork.tid = bswap_32(event->fork.tid); 410 event->fork.ppid = bswap_32(event->fork.ppid); 411 event->fork.ptid = bswap_32(event->fork.ptid); 412 event->fork.time = bswap_64(event->fork.time); 413 } 414 415 static void perf_event__read_swap(union perf_event *event) 416 { 417 event->read.pid = bswap_32(event->read.pid); 418 event->read.tid = bswap_32(event->read.tid); 419 event->read.value = bswap_64(event->read.value); 420 event->read.time_enabled = bswap_64(event->read.time_enabled); 421 event->read.time_running = bswap_64(event->read.time_running); 422 event->read.id = bswap_64(event->read.id); 423 } 424 425 /* exported for swapping attributes in file header */ 426 void perf_event__attr_swap(struct perf_event_attr *attr) 427 { 428 attr->type = bswap_32(attr->type); 429 attr->size = bswap_32(attr->size); 430 attr->config = bswap_64(attr->config); 431 attr->sample_period = bswap_64(attr->sample_period); 432 attr->sample_type = bswap_64(attr->sample_type); 433 attr->read_format = bswap_64(attr->read_format); 434 attr->wakeup_events = bswap_32(attr->wakeup_events); 435 attr->bp_type = bswap_32(attr->bp_type); 436 attr->bp_addr = bswap_64(attr->bp_addr); 437 attr->bp_len = bswap_64(attr->bp_len); 438 } 439 440 static void perf_event__hdr_attr_swap(union perf_event *event) 441 { 442 size_t size; 443 444 perf_event__attr_swap(&event->attr.attr); 445 446 size = event->header.size; 447 size -= (void *)&event->attr.id - (void *)event; 448 mem_bswap_64(event->attr.id, size); 449 } 450 451 static void perf_event__event_type_swap(union perf_event *event) 452 { 453 event->event_type.event_type.event_id = 454 bswap_64(event->event_type.event_type.event_id); 455 } 456 457 static void perf_event__tracing_data_swap(union perf_event *event) 458 { 459 event->tracing_data.size = bswap_32(event->tracing_data.size); 460 } 461 462 typedef void (*perf_event__swap_op)(union perf_event *event); 463 464 static perf_event__swap_op perf_event__swap_ops[] = { 465 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 466 [PERF_RECORD_COMM] = perf_event__comm_swap, 467 [PERF_RECORD_FORK] = perf_event__task_swap, 468 [PERF_RECORD_EXIT] = perf_event__task_swap, 469 [PERF_RECORD_LOST] = perf_event__all64_swap, 470 [PERF_RECORD_READ] = perf_event__read_swap, 471 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 472 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 473 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 474 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 475 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 476 [PERF_RECORD_HEADER_MAX] = NULL, 477 }; 478 479 struct sample_queue { 480 u64 timestamp; 481 u64 file_offset; 482 union perf_event *event; 483 struct list_head list; 484 }; 485 486 static void perf_session_free_sample_buffers(struct perf_session *session) 487 { 488 struct ordered_samples *os = &session->ordered_samples; 489 490 while (!list_empty(&os->to_free)) { 491 struct sample_queue *sq; 492 493 sq = list_entry(os->to_free.next, struct sample_queue, list); 494 list_del(&sq->list); 495 free(sq); 496 } 497 } 498 499 static int perf_session_deliver_event(struct perf_session *session, 500 union perf_event *event, 501 struct perf_sample *sample, 502 struct perf_tool *tool, 503 u64 file_offset); 504 505 static void flush_sample_queue(struct perf_session *s, 506 struct perf_tool *tool) 507 { 508 struct ordered_samples *os = &s->ordered_samples; 509 struct list_head *head = &os->samples; 510 struct sample_queue *tmp, *iter; 511 struct perf_sample sample; 512 u64 limit = os->next_flush; 513 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL; 514 unsigned idx = 0, progress_next = os->nr_samples / 16; 515 int ret; 516 517 if (!tool->ordered_samples || !limit) 518 return; 519 520 list_for_each_entry_safe(iter, tmp, head, list) { 521 if (iter->timestamp > limit) 522 break; 523 524 ret = perf_session__parse_sample(s, iter->event, &sample); 525 if (ret) 526 pr_err("Can't parse sample, err = %d\n", ret); 527 else 528 perf_session_deliver_event(s, iter->event, &sample, tool, 529 iter->file_offset); 530 531 os->last_flush = iter->timestamp; 532 list_del(&iter->list); 533 list_add(&iter->list, &os->sample_cache); 534 if (++idx >= progress_next) { 535 progress_next += os->nr_samples / 16; 536 ui_progress__update(idx, os->nr_samples, 537 "Processing time ordered events..."); 538 } 539 } 540 541 if (list_empty(head)) { 542 os->last_sample = NULL; 543 } else if (last_ts <= limit) { 544 os->last_sample = 545 list_entry(head->prev, struct sample_queue, list); 546 } 547 548 os->nr_samples = 0; 549 } 550 551 /* 552 * When perf record finishes a pass on every buffers, it records this pseudo 553 * event. 554 * We record the max timestamp t found in the pass n. 555 * Assuming these timestamps are monotonic across cpus, we know that if 556 * a buffer still has events with timestamps below t, they will be all 557 * available and then read in the pass n + 1. 558 * Hence when we start to read the pass n + 2, we can safely flush every 559 * events with timestamps below t. 560 * 561 * ============ PASS n ================= 562 * CPU 0 | CPU 1 563 * | 564 * cnt1 timestamps | cnt2 timestamps 565 * 1 | 2 566 * 2 | 3 567 * - | 4 <--- max recorded 568 * 569 * ============ PASS n + 1 ============== 570 * CPU 0 | CPU 1 571 * | 572 * cnt1 timestamps | cnt2 timestamps 573 * 3 | 5 574 * 4 | 6 575 * 5 | 7 <---- max recorded 576 * 577 * Flush every events below timestamp 4 578 * 579 * ============ PASS n + 2 ============== 580 * CPU 0 | CPU 1 581 * | 582 * cnt1 timestamps | cnt2 timestamps 583 * 6 | 8 584 * 7 | 9 585 * - | 10 586 * 587 * Flush every events below timestamp 7 588 * etc... 589 */ 590 static int process_finished_round(struct perf_tool *tool, 591 union perf_event *event __used, 592 struct perf_session *session) 593 { 594 flush_sample_queue(session, tool); 595 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp; 596 597 return 0; 598 } 599 600 /* The queue is ordered by time */ 601 static void __queue_event(struct sample_queue *new, struct perf_session *s) 602 { 603 struct ordered_samples *os = &s->ordered_samples; 604 struct sample_queue *sample = os->last_sample; 605 u64 timestamp = new->timestamp; 606 struct list_head *p; 607 608 ++os->nr_samples; 609 os->last_sample = new; 610 611 if (!sample) { 612 list_add(&new->list, &os->samples); 613 os->max_timestamp = timestamp; 614 return; 615 } 616 617 /* 618 * last_sample might point to some random place in the list as it's 619 * the last queued event. We expect that the new event is close to 620 * this. 621 */ 622 if (sample->timestamp <= timestamp) { 623 while (sample->timestamp <= timestamp) { 624 p = sample->list.next; 625 if (p == &os->samples) { 626 list_add_tail(&new->list, &os->samples); 627 os->max_timestamp = timestamp; 628 return; 629 } 630 sample = list_entry(p, struct sample_queue, list); 631 } 632 list_add_tail(&new->list, &sample->list); 633 } else { 634 while (sample->timestamp > timestamp) { 635 p = sample->list.prev; 636 if (p == &os->samples) { 637 list_add(&new->list, &os->samples); 638 return; 639 } 640 sample = list_entry(p, struct sample_queue, list); 641 } 642 list_add(&new->list, &sample->list); 643 } 644 } 645 646 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue)) 647 648 static int perf_session_queue_event(struct perf_session *s, union perf_event *event, 649 struct perf_sample *sample, u64 file_offset) 650 { 651 struct ordered_samples *os = &s->ordered_samples; 652 struct list_head *sc = &os->sample_cache; 653 u64 timestamp = sample->time; 654 struct sample_queue *new; 655 656 if (!timestamp || timestamp == ~0ULL) 657 return -ETIME; 658 659 if (timestamp < s->ordered_samples.last_flush) { 660 printf("Warning: Timestamp below last timeslice flush\n"); 661 return -EINVAL; 662 } 663 664 if (!list_empty(sc)) { 665 new = list_entry(sc->next, struct sample_queue, list); 666 list_del(&new->list); 667 } else if (os->sample_buffer) { 668 new = os->sample_buffer + os->sample_buffer_idx; 669 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER) 670 os->sample_buffer = NULL; 671 } else { 672 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new)); 673 if (!os->sample_buffer) 674 return -ENOMEM; 675 list_add(&os->sample_buffer->list, &os->to_free); 676 os->sample_buffer_idx = 2; 677 new = os->sample_buffer + 1; 678 } 679 680 new->timestamp = timestamp; 681 new->file_offset = file_offset; 682 new->event = event; 683 684 __queue_event(new, s); 685 686 return 0; 687 } 688 689 static void callchain__printf(struct perf_sample *sample) 690 { 691 unsigned int i; 692 693 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr); 694 695 for (i = 0; i < sample->callchain->nr; i++) 696 printf("..... %2d: %016" PRIx64 "\n", 697 i, sample->callchain->ips[i]); 698 } 699 700 static void perf_session__print_tstamp(struct perf_session *session, 701 union perf_event *event, 702 struct perf_sample *sample) 703 { 704 if (event->header.type != PERF_RECORD_SAMPLE && 705 !session->sample_id_all) { 706 fputs("-1 -1 ", stdout); 707 return; 708 } 709 710 if ((session->sample_type & PERF_SAMPLE_CPU)) 711 printf("%u ", sample->cpu); 712 713 if (session->sample_type & PERF_SAMPLE_TIME) 714 printf("%" PRIu64 " ", sample->time); 715 } 716 717 static void dump_event(struct perf_session *session, union perf_event *event, 718 u64 file_offset, struct perf_sample *sample) 719 { 720 if (!dump_trace) 721 return; 722 723 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 724 file_offset, event->header.size, event->header.type); 725 726 trace_event(event); 727 728 if (sample) 729 perf_session__print_tstamp(session, event, sample); 730 731 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 732 event->header.size, perf_event__name(event->header.type)); 733 } 734 735 static void dump_sample(struct perf_session *session, union perf_event *event, 736 struct perf_sample *sample) 737 { 738 if (!dump_trace) 739 return; 740 741 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 742 event->header.misc, sample->pid, sample->tid, sample->ip, 743 sample->period, sample->addr); 744 745 if (session->sample_type & PERF_SAMPLE_CALLCHAIN) 746 callchain__printf(sample); 747 } 748 749 static struct machine * 750 perf_session__find_machine_for_cpumode(struct perf_session *session, 751 union perf_event *event) 752 { 753 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 754 755 if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) 756 return perf_session__find_machine(session, event->ip.pid); 757 758 return perf_session__find_host_machine(session); 759 } 760 761 static int perf_session_deliver_event(struct perf_session *session, 762 union perf_event *event, 763 struct perf_sample *sample, 764 struct perf_tool *tool, 765 u64 file_offset) 766 { 767 struct perf_evsel *evsel; 768 struct machine *machine; 769 770 dump_event(session, event, file_offset, sample); 771 772 evsel = perf_evlist__id2evsel(session->evlist, sample->id); 773 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) { 774 /* 775 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here 776 * because the tools right now may apply filters, discarding 777 * some of the samples. For consistency, in the future we 778 * should have something like nr_filtered_samples and remove 779 * the sample->period from total_sample_period, etc, KISS for 780 * now tho. 781 * 782 * Also testing against NULL allows us to handle files without 783 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the 784 * future probably it'll be a good idea to restrict event 785 * processing via perf_session to files with both set. 786 */ 787 hists__inc_nr_events(&evsel->hists, event->header.type); 788 } 789 790 machine = perf_session__find_machine_for_cpumode(session, event); 791 792 switch (event->header.type) { 793 case PERF_RECORD_SAMPLE: 794 dump_sample(session, event, sample); 795 if (evsel == NULL) { 796 ++session->hists.stats.nr_unknown_id; 797 return -1; 798 } 799 return tool->sample(tool, event, sample, evsel, machine); 800 case PERF_RECORD_MMAP: 801 return tool->mmap(tool, event, sample, machine); 802 case PERF_RECORD_COMM: 803 return tool->comm(tool, event, sample, machine); 804 case PERF_RECORD_FORK: 805 return tool->fork(tool, event, sample, machine); 806 case PERF_RECORD_EXIT: 807 return tool->exit(tool, event, sample, machine); 808 case PERF_RECORD_LOST: 809 if (tool->lost == perf_event__process_lost) 810 session->hists.stats.total_lost += event->lost.lost; 811 return tool->lost(tool, event, sample, machine); 812 case PERF_RECORD_READ: 813 return tool->read(tool, event, sample, evsel, machine); 814 case PERF_RECORD_THROTTLE: 815 return tool->throttle(tool, event, sample, machine); 816 case PERF_RECORD_UNTHROTTLE: 817 return tool->unthrottle(tool, event, sample, machine); 818 default: 819 ++session->hists.stats.nr_unknown_events; 820 return -1; 821 } 822 } 823 824 static int perf_session__preprocess_sample(struct perf_session *session, 825 union perf_event *event, struct perf_sample *sample) 826 { 827 if (event->header.type != PERF_RECORD_SAMPLE || 828 !(session->sample_type & PERF_SAMPLE_CALLCHAIN)) 829 return 0; 830 831 if (!ip_callchain__valid(sample->callchain, event)) { 832 pr_debug("call-chain problem with event, skipping it.\n"); 833 ++session->hists.stats.nr_invalid_chains; 834 session->hists.stats.total_invalid_chains += sample->period; 835 return -EINVAL; 836 } 837 return 0; 838 } 839 840 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event, 841 struct perf_tool *tool, u64 file_offset) 842 { 843 int err; 844 845 dump_event(session, event, file_offset, NULL); 846 847 /* These events are processed right away */ 848 switch (event->header.type) { 849 case PERF_RECORD_HEADER_ATTR: 850 err = tool->attr(event, &session->evlist); 851 if (err == 0) 852 perf_session__update_sample_type(session); 853 return err; 854 case PERF_RECORD_HEADER_EVENT_TYPE: 855 return tool->event_type(tool, event); 856 case PERF_RECORD_HEADER_TRACING_DATA: 857 /* setup for reading amidst mmap */ 858 lseek(session->fd, file_offset, SEEK_SET); 859 return tool->tracing_data(event, session); 860 case PERF_RECORD_HEADER_BUILD_ID: 861 return tool->build_id(tool, event, session); 862 case PERF_RECORD_FINISHED_ROUND: 863 return tool->finished_round(tool, event, session); 864 default: 865 return -EINVAL; 866 } 867 } 868 869 static int perf_session__process_event(struct perf_session *session, 870 union perf_event *event, 871 struct perf_tool *tool, 872 u64 file_offset) 873 { 874 struct perf_sample sample; 875 int ret; 876 877 if (session->header.needs_swap && 878 perf_event__swap_ops[event->header.type]) 879 perf_event__swap_ops[event->header.type](event); 880 881 if (event->header.type >= PERF_RECORD_HEADER_MAX) 882 return -EINVAL; 883 884 hists__inc_nr_events(&session->hists, event->header.type); 885 886 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 887 return perf_session__process_user_event(session, event, tool, file_offset); 888 889 /* 890 * For all kernel events we get the sample data 891 */ 892 ret = perf_session__parse_sample(session, event, &sample); 893 if (ret) 894 return ret; 895 896 /* Preprocess sample records - precheck callchains */ 897 if (perf_session__preprocess_sample(session, event, &sample)) 898 return 0; 899 900 if (tool->ordered_samples) { 901 ret = perf_session_queue_event(session, event, &sample, 902 file_offset); 903 if (ret != -ETIME) 904 return ret; 905 } 906 907 return perf_session_deliver_event(session, event, &sample, tool, 908 file_offset); 909 } 910 911 void perf_event_header__bswap(struct perf_event_header *self) 912 { 913 self->type = bswap_32(self->type); 914 self->misc = bswap_16(self->misc); 915 self->size = bswap_16(self->size); 916 } 917 918 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 919 { 920 return machine__findnew_thread(&session->host_machine, pid); 921 } 922 923 static struct thread *perf_session__register_idle_thread(struct perf_session *self) 924 { 925 struct thread *thread = perf_session__findnew(self, 0); 926 927 if (thread == NULL || thread__set_comm(thread, "swapper")) { 928 pr_err("problem inserting idle task.\n"); 929 thread = NULL; 930 } 931 932 return thread; 933 } 934 935 static void perf_session__warn_about_errors(const struct perf_session *session, 936 const struct perf_tool *tool) 937 { 938 if (tool->lost == perf_event__process_lost && 939 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) { 940 ui__warning("Processed %d events and lost %d chunks!\n\n" 941 "Check IO/CPU overload!\n\n", 942 session->hists.stats.nr_events[0], 943 session->hists.stats.nr_events[PERF_RECORD_LOST]); 944 } 945 946 if (session->hists.stats.nr_unknown_events != 0) { 947 ui__warning("Found %u unknown events!\n\n" 948 "Is this an older tool processing a perf.data " 949 "file generated by a more recent tool?\n\n" 950 "If that is not the case, consider " 951 "reporting to linux-kernel@vger.kernel.org.\n\n", 952 session->hists.stats.nr_unknown_events); 953 } 954 955 if (session->hists.stats.nr_unknown_id != 0) { 956 ui__warning("%u samples with id not present in the header\n", 957 session->hists.stats.nr_unknown_id); 958 } 959 960 if (session->hists.stats.nr_invalid_chains != 0) { 961 ui__warning("Found invalid callchains!\n\n" 962 "%u out of %u events were discarded for this reason.\n\n" 963 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 964 session->hists.stats.nr_invalid_chains, 965 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]); 966 } 967 } 968 969 #define session_done() (*(volatile int *)(&session_done)) 970 volatile int session_done; 971 972 static int __perf_session__process_pipe_events(struct perf_session *self, 973 struct perf_tool *tool) 974 { 975 union perf_event event; 976 uint32_t size; 977 int skip = 0; 978 u64 head; 979 int err; 980 void *p; 981 982 perf_tool__fill_defaults(tool); 983 984 head = 0; 985 more: 986 err = readn(self->fd, &event, sizeof(struct perf_event_header)); 987 if (err <= 0) { 988 if (err == 0) 989 goto done; 990 991 pr_err("failed to read event header\n"); 992 goto out_err; 993 } 994 995 if (self->header.needs_swap) 996 perf_event_header__bswap(&event.header); 997 998 size = event.header.size; 999 if (size == 0) 1000 size = 8; 1001 1002 p = &event; 1003 p += sizeof(struct perf_event_header); 1004 1005 if (size - sizeof(struct perf_event_header)) { 1006 err = readn(self->fd, p, size - sizeof(struct perf_event_header)); 1007 if (err <= 0) { 1008 if (err == 0) { 1009 pr_err("unexpected end of event stream\n"); 1010 goto done; 1011 } 1012 1013 pr_err("failed to read event data\n"); 1014 goto out_err; 1015 } 1016 } 1017 1018 if ((skip = perf_session__process_event(self, &event, tool, head)) < 0) { 1019 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 1020 head, event.header.size, event.header.type); 1021 /* 1022 * assume we lost track of the stream, check alignment, and 1023 * increment a single u64 in the hope to catch on again 'soon'. 1024 */ 1025 if (unlikely(head & 7)) 1026 head &= ~7ULL; 1027 1028 size = 8; 1029 } 1030 1031 head += size; 1032 1033 if (skip > 0) 1034 head += skip; 1035 1036 if (!session_done()) 1037 goto more; 1038 done: 1039 err = 0; 1040 out_err: 1041 perf_session__warn_about_errors(self, tool); 1042 perf_session_free_sample_buffers(self); 1043 return err; 1044 } 1045 1046 static union perf_event * 1047 fetch_mmaped_event(struct perf_session *session, 1048 u64 head, size_t mmap_size, char *buf) 1049 { 1050 union perf_event *event; 1051 1052 /* 1053 * Ensure we have enough space remaining to read 1054 * the size of the event in the headers. 1055 */ 1056 if (head + sizeof(event->header) > mmap_size) 1057 return NULL; 1058 1059 event = (union perf_event *)(buf + head); 1060 1061 if (session->header.needs_swap) 1062 perf_event_header__bswap(&event->header); 1063 1064 if (head + event->header.size > mmap_size) 1065 return NULL; 1066 1067 return event; 1068 } 1069 1070 int __perf_session__process_events(struct perf_session *session, 1071 u64 data_offset, u64 data_size, 1072 u64 file_size, struct perf_tool *tool) 1073 { 1074 u64 head, page_offset, file_offset, file_pos, progress_next; 1075 int err, mmap_prot, mmap_flags, map_idx = 0; 1076 size_t page_size, mmap_size; 1077 char *buf, *mmaps[8]; 1078 union perf_event *event; 1079 uint32_t size; 1080 1081 perf_tool__fill_defaults(tool); 1082 1083 page_size = sysconf(_SC_PAGESIZE); 1084 1085 page_offset = page_size * (data_offset / page_size); 1086 file_offset = page_offset; 1087 head = data_offset - page_offset; 1088 1089 if (data_offset + data_size < file_size) 1090 file_size = data_offset + data_size; 1091 1092 progress_next = file_size / 16; 1093 1094 mmap_size = session->mmap_window; 1095 if (mmap_size > file_size) 1096 mmap_size = file_size; 1097 1098 memset(mmaps, 0, sizeof(mmaps)); 1099 1100 mmap_prot = PROT_READ; 1101 mmap_flags = MAP_SHARED; 1102 1103 if (session->header.needs_swap) { 1104 mmap_prot |= PROT_WRITE; 1105 mmap_flags = MAP_PRIVATE; 1106 } 1107 remap: 1108 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd, 1109 file_offset); 1110 if (buf == MAP_FAILED) { 1111 pr_err("failed to mmap file\n"); 1112 err = -errno; 1113 goto out_err; 1114 } 1115 mmaps[map_idx] = buf; 1116 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 1117 file_pos = file_offset + head; 1118 1119 more: 1120 event = fetch_mmaped_event(session, head, mmap_size, buf); 1121 if (!event) { 1122 if (mmaps[map_idx]) { 1123 munmap(mmaps[map_idx], mmap_size); 1124 mmaps[map_idx] = NULL; 1125 } 1126 1127 page_offset = page_size * (head / page_size); 1128 file_offset += page_offset; 1129 head -= page_offset; 1130 goto remap; 1131 } 1132 1133 size = event->header.size; 1134 1135 if (size == 0 || 1136 perf_session__process_event(session, event, tool, file_pos) < 0) { 1137 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 1138 file_offset + head, event->header.size, 1139 event->header.type); 1140 /* 1141 * assume we lost track of the stream, check alignment, and 1142 * increment a single u64 in the hope to catch on again 'soon'. 1143 */ 1144 if (unlikely(head & 7)) 1145 head &= ~7ULL; 1146 1147 size = 8; 1148 } 1149 1150 head += size; 1151 file_pos += size; 1152 1153 if (file_pos >= progress_next) { 1154 progress_next += file_size / 16; 1155 ui_progress__update(file_pos, file_size, 1156 "Processing events..."); 1157 } 1158 1159 if (file_pos < file_size) 1160 goto more; 1161 1162 err = 0; 1163 /* do the final flush for ordered samples */ 1164 session->ordered_samples.next_flush = ULLONG_MAX; 1165 flush_sample_queue(session, tool); 1166 out_err: 1167 perf_session__warn_about_errors(session, tool); 1168 perf_session_free_sample_buffers(session); 1169 return err; 1170 } 1171 1172 int perf_session__process_events(struct perf_session *self, 1173 struct perf_tool *tool) 1174 { 1175 int err; 1176 1177 if (perf_session__register_idle_thread(self) == NULL) 1178 return -ENOMEM; 1179 1180 if (!self->fd_pipe) 1181 err = __perf_session__process_events(self, 1182 self->header.data_offset, 1183 self->header.data_size, 1184 self->size, tool); 1185 else 1186 err = __perf_session__process_pipe_events(self, tool); 1187 1188 return err; 1189 } 1190 1191 bool perf_session__has_traces(struct perf_session *self, const char *msg) 1192 { 1193 if (!(self->sample_type & PERF_SAMPLE_RAW)) { 1194 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 1195 return false; 1196 } 1197 1198 return true; 1199 } 1200 1201 int maps__set_kallsyms_ref_reloc_sym(struct map **maps, 1202 const char *symbol_name, u64 addr) 1203 { 1204 char *bracket; 1205 enum map_type i; 1206 struct ref_reloc_sym *ref; 1207 1208 ref = zalloc(sizeof(struct ref_reloc_sym)); 1209 if (ref == NULL) 1210 return -ENOMEM; 1211 1212 ref->name = strdup(symbol_name); 1213 if (ref->name == NULL) { 1214 free(ref); 1215 return -ENOMEM; 1216 } 1217 1218 bracket = strchr(ref->name, ']'); 1219 if (bracket) 1220 *bracket = '\0'; 1221 1222 ref->addr = addr; 1223 1224 for (i = 0; i < MAP__NR_TYPES; ++i) { 1225 struct kmap *kmap = map__kmap(maps[i]); 1226 kmap->ref_reloc_sym = ref; 1227 } 1228 1229 return 0; 1230 } 1231 1232 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp) 1233 { 1234 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) + 1235 __dsos__fprintf(&self->host_machine.user_dsos, fp) + 1236 machines__fprintf_dsos(&self->machines, fp); 1237 } 1238 1239 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp, 1240 bool with_hits) 1241 { 1242 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits); 1243 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits); 1244 } 1245 1246 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 1247 { 1248 struct perf_evsel *pos; 1249 size_t ret = fprintf(fp, "Aggregated stats:\n"); 1250 1251 ret += hists__fprintf_nr_events(&session->hists, fp); 1252 1253 list_for_each_entry(pos, &session->evlist->entries, node) { 1254 ret += fprintf(fp, "%s stats:\n", event_name(pos)); 1255 ret += hists__fprintf_nr_events(&pos->hists, fp); 1256 } 1257 1258 return ret; 1259 } 1260 1261 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 1262 { 1263 /* 1264 * FIXME: Here we have to actually print all the machines in this 1265 * session, not just the host... 1266 */ 1267 return machine__fprintf(&session->host_machine, fp); 1268 } 1269 1270 void perf_session__remove_thread(struct perf_session *session, 1271 struct thread *th) 1272 { 1273 /* 1274 * FIXME: This one makes no sense, we need to remove the thread from 1275 * the machine it belongs to, perf_session can have many machines, so 1276 * doing it always on ->host_machine is wrong. Fix when auditing all 1277 * the 'perf kvm' code. 1278 */ 1279 machine__remove_thread(&session->host_machine, th); 1280 } 1281 1282 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session, 1283 unsigned int type) 1284 { 1285 struct perf_evsel *pos; 1286 1287 list_for_each_entry(pos, &session->evlist->entries, node) { 1288 if (pos->attr.type == type) 1289 return pos; 1290 } 1291 return NULL; 1292 } 1293 1294 void perf_event__print_ip(union perf_event *event, struct perf_sample *sample, 1295 struct machine *machine, struct perf_evsel *evsel, 1296 int print_sym, int print_dso) 1297 { 1298 struct addr_location al; 1299 struct callchain_cursor *cursor = &evsel->hists.callchain_cursor; 1300 struct callchain_cursor_node *node; 1301 1302 if (perf_event__preprocess_sample(event, machine, &al, sample, 1303 NULL) < 0) { 1304 error("problem processing %d event, skipping it.\n", 1305 event->header.type); 1306 return; 1307 } 1308 1309 if (symbol_conf.use_callchain && sample->callchain) { 1310 1311 if (machine__resolve_callchain(machine, evsel, al.thread, 1312 sample->callchain, NULL) != 0) { 1313 if (verbose) 1314 error("Failed to resolve callchain. Skipping\n"); 1315 return; 1316 } 1317 callchain_cursor_commit(cursor); 1318 1319 while (1) { 1320 node = callchain_cursor_current(cursor); 1321 if (!node) 1322 break; 1323 1324 printf("\t%16" PRIx64, node->ip); 1325 if (print_sym) { 1326 printf(" "); 1327 symbol__fprintf_symname(node->sym, stdout); 1328 } 1329 if (print_dso) { 1330 printf(" ("); 1331 map__fprintf_dsoname(al.map, stdout); 1332 printf(")"); 1333 } 1334 printf("\n"); 1335 1336 callchain_cursor_advance(cursor); 1337 } 1338 1339 } else { 1340 printf("%16" PRIx64, sample->ip); 1341 if (print_sym) { 1342 printf(" "); 1343 symbol__fprintf_symname(al.sym, stdout); 1344 } 1345 1346 if (print_dso) { 1347 printf(" ("); 1348 map__fprintf_dsoname(al.map, stdout); 1349 printf(")"); 1350 } 1351 } 1352 } 1353 1354 int perf_session__cpu_bitmap(struct perf_session *session, 1355 const char *cpu_list, unsigned long *cpu_bitmap) 1356 { 1357 int i; 1358 struct cpu_map *map; 1359 1360 for (i = 0; i < PERF_TYPE_MAX; ++i) { 1361 struct perf_evsel *evsel; 1362 1363 evsel = perf_session__find_first_evtype(session, i); 1364 if (!evsel) 1365 continue; 1366 1367 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) { 1368 pr_err("File does not contain CPU events. " 1369 "Remove -c option to proceed.\n"); 1370 return -1; 1371 } 1372 } 1373 1374 map = cpu_map__new(cpu_list); 1375 if (map == NULL) { 1376 pr_err("Invalid cpu_list\n"); 1377 return -1; 1378 } 1379 1380 for (i = 0; i < map->nr; i++) { 1381 int cpu = map->map[i]; 1382 1383 if (cpu >= MAX_NR_CPUS) { 1384 pr_err("Requested CPU %d too large. " 1385 "Consider raising MAX_NR_CPUS\n", cpu); 1386 return -1; 1387 } 1388 1389 set_bit(cpu, cpu_bitmap); 1390 } 1391 1392 return 0; 1393 } 1394 1395 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 1396 bool full) 1397 { 1398 struct stat st; 1399 int ret; 1400 1401 if (session == NULL || fp == NULL) 1402 return; 1403 1404 ret = fstat(session->fd, &st); 1405 if (ret == -1) 1406 return; 1407 1408 fprintf(fp, "# ========\n"); 1409 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime)); 1410 perf_header__fprintf_info(session, fp, full); 1411 fprintf(fp, "# ========\n#\n"); 1412 } 1413