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