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 static void perf_session__id_header_size(struct perf_session *session) 82 { 83 struct perf_sample *data; 84 u64 sample_type = session->sample_type; 85 u16 size = 0; 86 87 if (!session->sample_id_all) 88 goto out; 89 90 if (sample_type & PERF_SAMPLE_TID) 91 size += sizeof(data->tid) * 2; 92 93 if (sample_type & PERF_SAMPLE_TIME) 94 size += sizeof(data->time); 95 96 if (sample_type & PERF_SAMPLE_ID) 97 size += sizeof(data->id); 98 99 if (sample_type & PERF_SAMPLE_STREAM_ID) 100 size += sizeof(data->stream_id); 101 102 if (sample_type & PERF_SAMPLE_CPU) 103 size += sizeof(data->cpu) * 2; 104 out: 105 session->id_hdr_size = size; 106 } 107 108 void perf_session__update_sample_type(struct perf_session *self) 109 { 110 self->sample_type = perf_evlist__sample_type(self->evlist); 111 self->sample_size = __perf_evsel__sample_size(self->sample_type); 112 self->sample_id_all = perf_evlist__sample_id_all(self->evlist); 113 perf_session__id_header_size(self); 114 } 115 116 int perf_session__create_kernel_maps(struct perf_session *self) 117 { 118 int ret = machine__create_kernel_maps(&self->host_machine); 119 120 if (ret >= 0) 121 ret = machines__create_guest_kernel_maps(&self->machines); 122 return ret; 123 } 124 125 static void perf_session__destroy_kernel_maps(struct perf_session *self) 126 { 127 machine__destroy_kernel_maps(&self->host_machine); 128 machines__destroy_guest_kernel_maps(&self->machines); 129 } 130 131 struct perf_session *perf_session__new(const char *filename, int mode, 132 bool force, bool repipe, 133 struct perf_event_ops *ops) 134 { 135 size_t len = filename ? strlen(filename) + 1 : 0; 136 struct perf_session *self = zalloc(sizeof(*self) + len); 137 138 if (self == NULL) 139 goto out; 140 141 memcpy(self->filename, filename, len); 142 self->threads = RB_ROOT; 143 INIT_LIST_HEAD(&self->dead_threads); 144 self->last_match = NULL; 145 /* 146 * On 64bit we can mmap the data file in one go. No need for tiny mmap 147 * slices. On 32bit we use 32MB. 148 */ 149 #if BITS_PER_LONG == 64 150 self->mmap_window = ULLONG_MAX; 151 #else 152 self->mmap_window = 32 * 1024 * 1024ULL; 153 #endif 154 self->machines = RB_ROOT; 155 self->repipe = repipe; 156 INIT_LIST_HEAD(&self->ordered_samples.samples); 157 INIT_LIST_HEAD(&self->ordered_samples.sample_cache); 158 INIT_LIST_HEAD(&self->ordered_samples.to_free); 159 machine__init(&self->host_machine, "", HOST_KERNEL_ID); 160 161 if (mode == O_RDONLY) { 162 if (perf_session__open(self, force) < 0) 163 goto out_delete; 164 perf_session__update_sample_type(self); 165 } else if (mode == O_WRONLY) { 166 /* 167 * In O_RDONLY mode this will be performed when reading the 168 * kernel MMAP event, in perf_event__process_mmap(). 169 */ 170 if (perf_session__create_kernel_maps(self) < 0) 171 goto out_delete; 172 } 173 174 if (ops && ops->ordering_requires_timestamps && 175 ops->ordered_samples && !self->sample_id_all) { 176 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 177 ops->ordered_samples = false; 178 } 179 180 out: 181 return self; 182 out_delete: 183 perf_session__delete(self); 184 return NULL; 185 } 186 187 static void perf_session__delete_dead_threads(struct perf_session *self) 188 { 189 struct thread *n, *t; 190 191 list_for_each_entry_safe(t, n, &self->dead_threads, node) { 192 list_del(&t->node); 193 thread__delete(t); 194 } 195 } 196 197 static void perf_session__delete_threads(struct perf_session *self) 198 { 199 struct rb_node *nd = rb_first(&self->threads); 200 201 while (nd) { 202 struct thread *t = rb_entry(nd, struct thread, rb_node); 203 204 rb_erase(&t->rb_node, &self->threads); 205 nd = rb_next(nd); 206 thread__delete(t); 207 } 208 } 209 210 void perf_session__delete(struct perf_session *self) 211 { 212 perf_session__destroy_kernel_maps(self); 213 perf_session__delete_dead_threads(self); 214 perf_session__delete_threads(self); 215 machine__exit(&self->host_machine); 216 close(self->fd); 217 free(self); 218 } 219 220 void perf_session__remove_thread(struct perf_session *self, struct thread *th) 221 { 222 self->last_match = NULL; 223 rb_erase(&th->rb_node, &self->threads); 224 /* 225 * We may have references to this thread, for instance in some hist_entry 226 * instances, so just move them to a separate list. 227 */ 228 list_add_tail(&th->node, &self->dead_threads); 229 } 230 231 static bool symbol__match_parent_regex(struct symbol *sym) 232 { 233 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0)) 234 return 1; 235 236 return 0; 237 } 238 239 int perf_session__resolve_callchain(struct perf_session *self, 240 struct thread *thread, 241 struct ip_callchain *chain, 242 struct symbol **parent) 243 { 244 u8 cpumode = PERF_RECORD_MISC_USER; 245 unsigned int i; 246 int err; 247 248 callchain_cursor_reset(&self->callchain_cursor); 249 250 for (i = 0; i < chain->nr; i++) { 251 u64 ip; 252 struct addr_location al; 253 254 if (callchain_param.order == ORDER_CALLEE) 255 ip = chain->ips[i]; 256 else 257 ip = chain->ips[chain->nr - i - 1]; 258 259 if (ip >= PERF_CONTEXT_MAX) { 260 switch (ip) { 261 case PERF_CONTEXT_HV: 262 cpumode = PERF_RECORD_MISC_HYPERVISOR; break; 263 case PERF_CONTEXT_KERNEL: 264 cpumode = PERF_RECORD_MISC_KERNEL; break; 265 case PERF_CONTEXT_USER: 266 cpumode = PERF_RECORD_MISC_USER; break; 267 default: 268 break; 269 } 270 continue; 271 } 272 273 al.filtered = false; 274 thread__find_addr_location(thread, self, cpumode, 275 MAP__FUNCTION, thread->pid, ip, &al, NULL); 276 if (al.sym != NULL) { 277 if (sort__has_parent && !*parent && 278 symbol__match_parent_regex(al.sym)) 279 *parent = al.sym; 280 if (!symbol_conf.use_callchain) 281 break; 282 } 283 284 err = callchain_cursor_append(&self->callchain_cursor, 285 ip, al.map, al.sym); 286 if (err) 287 return err; 288 } 289 290 return 0; 291 } 292 293 static int process_event_synth_stub(union perf_event *event __used, 294 struct perf_session *session __used) 295 { 296 dump_printf(": unhandled!\n"); 297 return 0; 298 } 299 300 static int process_event_sample_stub(union perf_event *event __used, 301 struct perf_sample *sample __used, 302 struct perf_evsel *evsel __used, 303 struct perf_session *session __used) 304 { 305 dump_printf(": unhandled!\n"); 306 return 0; 307 } 308 309 static int process_event_stub(union perf_event *event __used, 310 struct perf_sample *sample __used, 311 struct perf_session *session __used) 312 { 313 dump_printf(": unhandled!\n"); 314 return 0; 315 } 316 317 static int process_finished_round_stub(union perf_event *event __used, 318 struct perf_session *session __used, 319 struct perf_event_ops *ops __used) 320 { 321 dump_printf(": unhandled!\n"); 322 return 0; 323 } 324 325 static int process_finished_round(union perf_event *event, 326 struct perf_session *session, 327 struct perf_event_ops *ops); 328 329 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler) 330 { 331 if (handler->sample == NULL) 332 handler->sample = process_event_sample_stub; 333 if (handler->mmap == NULL) 334 handler->mmap = process_event_stub; 335 if (handler->comm == NULL) 336 handler->comm = process_event_stub; 337 if (handler->fork == NULL) 338 handler->fork = process_event_stub; 339 if (handler->exit == NULL) 340 handler->exit = process_event_stub; 341 if (handler->lost == NULL) 342 handler->lost = perf_event__process_lost; 343 if (handler->read == NULL) 344 handler->read = process_event_stub; 345 if (handler->throttle == NULL) 346 handler->throttle = process_event_stub; 347 if (handler->unthrottle == NULL) 348 handler->unthrottle = process_event_stub; 349 if (handler->attr == NULL) 350 handler->attr = process_event_synth_stub; 351 if (handler->event_type == NULL) 352 handler->event_type = process_event_synth_stub; 353 if (handler->tracing_data == NULL) 354 handler->tracing_data = process_event_synth_stub; 355 if (handler->build_id == NULL) 356 handler->build_id = process_event_synth_stub; 357 if (handler->finished_round == NULL) { 358 if (handler->ordered_samples) 359 handler->finished_round = process_finished_round; 360 else 361 handler->finished_round = process_finished_round_stub; 362 } 363 } 364 365 void mem_bswap_64(void *src, int byte_size) 366 { 367 u64 *m = src; 368 369 while (byte_size > 0) { 370 *m = bswap_64(*m); 371 byte_size -= sizeof(u64); 372 ++m; 373 } 374 } 375 376 static void perf_event__all64_swap(union perf_event *event) 377 { 378 struct perf_event_header *hdr = &event->header; 379 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 380 } 381 382 static void perf_event__comm_swap(union perf_event *event) 383 { 384 event->comm.pid = bswap_32(event->comm.pid); 385 event->comm.tid = bswap_32(event->comm.tid); 386 } 387 388 static void perf_event__mmap_swap(union perf_event *event) 389 { 390 event->mmap.pid = bswap_32(event->mmap.pid); 391 event->mmap.tid = bswap_32(event->mmap.tid); 392 event->mmap.start = bswap_64(event->mmap.start); 393 event->mmap.len = bswap_64(event->mmap.len); 394 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 395 } 396 397 static void perf_event__task_swap(union perf_event *event) 398 { 399 event->fork.pid = bswap_32(event->fork.pid); 400 event->fork.tid = bswap_32(event->fork.tid); 401 event->fork.ppid = bswap_32(event->fork.ppid); 402 event->fork.ptid = bswap_32(event->fork.ptid); 403 event->fork.time = bswap_64(event->fork.time); 404 } 405 406 static void perf_event__read_swap(union perf_event *event) 407 { 408 event->read.pid = bswap_32(event->read.pid); 409 event->read.tid = bswap_32(event->read.tid); 410 event->read.value = bswap_64(event->read.value); 411 event->read.time_enabled = bswap_64(event->read.time_enabled); 412 event->read.time_running = bswap_64(event->read.time_running); 413 event->read.id = bswap_64(event->read.id); 414 } 415 416 /* exported for swapping attributes in file header */ 417 void perf_event__attr_swap(struct perf_event_attr *attr) 418 { 419 attr->type = bswap_32(attr->type); 420 attr->size = bswap_32(attr->size); 421 attr->config = bswap_64(attr->config); 422 attr->sample_period = bswap_64(attr->sample_period); 423 attr->sample_type = bswap_64(attr->sample_type); 424 attr->read_format = bswap_64(attr->read_format); 425 attr->wakeup_events = bswap_32(attr->wakeup_events); 426 attr->bp_type = bswap_32(attr->bp_type); 427 attr->bp_addr = bswap_64(attr->bp_addr); 428 attr->bp_len = bswap_64(attr->bp_len); 429 } 430 431 static void perf_event__hdr_attr_swap(union perf_event *event) 432 { 433 size_t size; 434 435 perf_event__attr_swap(&event->attr.attr); 436 437 size = event->header.size; 438 size -= (void *)&event->attr.id - (void *)event; 439 mem_bswap_64(event->attr.id, size); 440 } 441 442 static void perf_event__event_type_swap(union perf_event *event) 443 { 444 event->event_type.event_type.event_id = 445 bswap_64(event->event_type.event_type.event_id); 446 } 447 448 static void perf_event__tracing_data_swap(union perf_event *event) 449 { 450 event->tracing_data.size = bswap_32(event->tracing_data.size); 451 } 452 453 typedef void (*perf_event__swap_op)(union perf_event *event); 454 455 static perf_event__swap_op perf_event__swap_ops[] = { 456 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 457 [PERF_RECORD_COMM] = perf_event__comm_swap, 458 [PERF_RECORD_FORK] = perf_event__task_swap, 459 [PERF_RECORD_EXIT] = perf_event__task_swap, 460 [PERF_RECORD_LOST] = perf_event__all64_swap, 461 [PERF_RECORD_READ] = perf_event__read_swap, 462 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 463 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 464 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 465 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 466 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 467 [PERF_RECORD_HEADER_MAX] = NULL, 468 }; 469 470 struct sample_queue { 471 u64 timestamp; 472 u64 file_offset; 473 union perf_event *event; 474 struct list_head list; 475 }; 476 477 static void perf_session_free_sample_buffers(struct perf_session *session) 478 { 479 struct ordered_samples *os = &session->ordered_samples; 480 481 while (!list_empty(&os->to_free)) { 482 struct sample_queue *sq; 483 484 sq = list_entry(os->to_free.next, struct sample_queue, list); 485 list_del(&sq->list); 486 free(sq); 487 } 488 } 489 490 static int perf_session_deliver_event(struct perf_session *session, 491 union perf_event *event, 492 struct perf_sample *sample, 493 struct perf_event_ops *ops, 494 u64 file_offset); 495 496 static void flush_sample_queue(struct perf_session *s, 497 struct perf_event_ops *ops) 498 { 499 struct ordered_samples *os = &s->ordered_samples; 500 struct list_head *head = &os->samples; 501 struct sample_queue *tmp, *iter; 502 struct perf_sample sample; 503 u64 limit = os->next_flush; 504 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL; 505 unsigned idx = 0, progress_next = os->nr_samples / 16; 506 int ret; 507 508 if (!ops->ordered_samples || !limit) 509 return; 510 511 list_for_each_entry_safe(iter, tmp, head, list) { 512 if (iter->timestamp > limit) 513 break; 514 515 ret = perf_session__parse_sample(s, iter->event, &sample); 516 if (ret) 517 pr_err("Can't parse sample, err = %d\n", ret); 518 else 519 perf_session_deliver_event(s, iter->event, &sample, ops, 520 iter->file_offset); 521 522 os->last_flush = iter->timestamp; 523 list_del(&iter->list); 524 list_add(&iter->list, &os->sample_cache); 525 if (++idx >= progress_next) { 526 progress_next += os->nr_samples / 16; 527 ui_progress__update(idx, os->nr_samples, 528 "Processing time ordered events..."); 529 } 530 } 531 532 if (list_empty(head)) { 533 os->last_sample = NULL; 534 } else if (last_ts <= limit) { 535 os->last_sample = 536 list_entry(head->prev, struct sample_queue, list); 537 } 538 539 os->nr_samples = 0; 540 } 541 542 /* 543 * When perf record finishes a pass on every buffers, it records this pseudo 544 * event. 545 * We record the max timestamp t found in the pass n. 546 * Assuming these timestamps are monotonic across cpus, we know that if 547 * a buffer still has events with timestamps below t, they will be all 548 * available and then read in the pass n + 1. 549 * Hence when we start to read the pass n + 2, we can safely flush every 550 * events with timestamps below t. 551 * 552 * ============ PASS n ================= 553 * CPU 0 | CPU 1 554 * | 555 * cnt1 timestamps | cnt2 timestamps 556 * 1 | 2 557 * 2 | 3 558 * - | 4 <--- max recorded 559 * 560 * ============ PASS n + 1 ============== 561 * CPU 0 | CPU 1 562 * | 563 * cnt1 timestamps | cnt2 timestamps 564 * 3 | 5 565 * 4 | 6 566 * 5 | 7 <---- max recorded 567 * 568 * Flush every events below timestamp 4 569 * 570 * ============ PASS n + 2 ============== 571 * CPU 0 | CPU 1 572 * | 573 * cnt1 timestamps | cnt2 timestamps 574 * 6 | 8 575 * 7 | 9 576 * - | 10 577 * 578 * Flush every events below timestamp 7 579 * etc... 580 */ 581 static int process_finished_round(union perf_event *event __used, 582 struct perf_session *session, 583 struct perf_event_ops *ops) 584 { 585 flush_sample_queue(session, ops); 586 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp; 587 588 return 0; 589 } 590 591 /* The queue is ordered by time */ 592 static void __queue_event(struct sample_queue *new, struct perf_session *s) 593 { 594 struct ordered_samples *os = &s->ordered_samples; 595 struct sample_queue *sample = os->last_sample; 596 u64 timestamp = new->timestamp; 597 struct list_head *p; 598 599 ++os->nr_samples; 600 os->last_sample = new; 601 602 if (!sample) { 603 list_add(&new->list, &os->samples); 604 os->max_timestamp = timestamp; 605 return; 606 } 607 608 /* 609 * last_sample might point to some random place in the list as it's 610 * the last queued event. We expect that the new event is close to 611 * this. 612 */ 613 if (sample->timestamp <= timestamp) { 614 while (sample->timestamp <= timestamp) { 615 p = sample->list.next; 616 if (p == &os->samples) { 617 list_add_tail(&new->list, &os->samples); 618 os->max_timestamp = timestamp; 619 return; 620 } 621 sample = list_entry(p, struct sample_queue, list); 622 } 623 list_add_tail(&new->list, &sample->list); 624 } else { 625 while (sample->timestamp > timestamp) { 626 p = sample->list.prev; 627 if (p == &os->samples) { 628 list_add(&new->list, &os->samples); 629 return; 630 } 631 sample = list_entry(p, struct sample_queue, list); 632 } 633 list_add(&new->list, &sample->list); 634 } 635 } 636 637 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue)) 638 639 static int perf_session_queue_event(struct perf_session *s, union perf_event *event, 640 struct perf_sample *sample, u64 file_offset) 641 { 642 struct ordered_samples *os = &s->ordered_samples; 643 struct list_head *sc = &os->sample_cache; 644 u64 timestamp = sample->time; 645 struct sample_queue *new; 646 647 if (!timestamp || timestamp == ~0ULL) 648 return -ETIME; 649 650 if (timestamp < s->ordered_samples.last_flush) { 651 printf("Warning: Timestamp below last timeslice flush\n"); 652 return -EINVAL; 653 } 654 655 if (!list_empty(sc)) { 656 new = list_entry(sc->next, struct sample_queue, list); 657 list_del(&new->list); 658 } else if (os->sample_buffer) { 659 new = os->sample_buffer + os->sample_buffer_idx; 660 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER) 661 os->sample_buffer = NULL; 662 } else { 663 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new)); 664 if (!os->sample_buffer) 665 return -ENOMEM; 666 list_add(&os->sample_buffer->list, &os->to_free); 667 os->sample_buffer_idx = 2; 668 new = os->sample_buffer + 1; 669 } 670 671 new->timestamp = timestamp; 672 new->file_offset = file_offset; 673 new->event = event; 674 675 __queue_event(new, s); 676 677 return 0; 678 } 679 680 static void callchain__printf(struct perf_sample *sample) 681 { 682 unsigned int i; 683 684 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr); 685 686 for (i = 0; i < sample->callchain->nr; i++) 687 printf("..... %2d: %016" PRIx64 "\n", 688 i, sample->callchain->ips[i]); 689 } 690 691 static void perf_session__print_tstamp(struct perf_session *session, 692 union perf_event *event, 693 struct perf_sample *sample) 694 { 695 if (event->header.type != PERF_RECORD_SAMPLE && 696 !session->sample_id_all) { 697 fputs("-1 -1 ", stdout); 698 return; 699 } 700 701 if ((session->sample_type & PERF_SAMPLE_CPU)) 702 printf("%u ", sample->cpu); 703 704 if (session->sample_type & PERF_SAMPLE_TIME) 705 printf("%" PRIu64 " ", sample->time); 706 } 707 708 static void dump_event(struct perf_session *session, union perf_event *event, 709 u64 file_offset, struct perf_sample *sample) 710 { 711 if (!dump_trace) 712 return; 713 714 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 715 file_offset, event->header.size, event->header.type); 716 717 trace_event(event); 718 719 if (sample) 720 perf_session__print_tstamp(session, event, sample); 721 722 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 723 event->header.size, perf_event__name(event->header.type)); 724 } 725 726 static void dump_sample(struct perf_session *session, union perf_event *event, 727 struct perf_sample *sample) 728 { 729 if (!dump_trace) 730 return; 731 732 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 733 event->header.misc, sample->pid, sample->tid, sample->ip, 734 sample->period, sample->addr); 735 736 if (session->sample_type & PERF_SAMPLE_CALLCHAIN) 737 callchain__printf(sample); 738 } 739 740 static int perf_session_deliver_event(struct perf_session *session, 741 union perf_event *event, 742 struct perf_sample *sample, 743 struct perf_event_ops *ops, 744 u64 file_offset) 745 { 746 struct perf_evsel *evsel; 747 748 dump_event(session, event, file_offset, sample); 749 750 evsel = perf_evlist__id2evsel(session->evlist, sample->id); 751 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) { 752 /* 753 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here 754 * because the tools right now may apply filters, discarding 755 * some of the samples. For consistency, in the future we 756 * should have something like nr_filtered_samples and remove 757 * the sample->period from total_sample_period, etc, KISS for 758 * now tho. 759 * 760 * Also testing against NULL allows us to handle files without 761 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the 762 * future probably it'll be a good idea to restrict event 763 * processing via perf_session to files with both set. 764 */ 765 hists__inc_nr_events(&evsel->hists, event->header.type); 766 } 767 768 switch (event->header.type) { 769 case PERF_RECORD_SAMPLE: 770 dump_sample(session, event, sample); 771 if (evsel == NULL) { 772 ++session->hists.stats.nr_unknown_id; 773 return -1; 774 } 775 return ops->sample(event, sample, evsel, session); 776 case PERF_RECORD_MMAP: 777 return ops->mmap(event, sample, session); 778 case PERF_RECORD_COMM: 779 return ops->comm(event, sample, session); 780 case PERF_RECORD_FORK: 781 return ops->fork(event, sample, session); 782 case PERF_RECORD_EXIT: 783 return ops->exit(event, sample, session); 784 case PERF_RECORD_LOST: 785 return ops->lost(event, sample, session); 786 case PERF_RECORD_READ: 787 return ops->read(event, sample, session); 788 case PERF_RECORD_THROTTLE: 789 return ops->throttle(event, sample, session); 790 case PERF_RECORD_UNTHROTTLE: 791 return ops->unthrottle(event, sample, session); 792 default: 793 ++session->hists.stats.nr_unknown_events; 794 return -1; 795 } 796 } 797 798 static int perf_session__preprocess_sample(struct perf_session *session, 799 union perf_event *event, struct perf_sample *sample) 800 { 801 if (event->header.type != PERF_RECORD_SAMPLE || 802 !(session->sample_type & PERF_SAMPLE_CALLCHAIN)) 803 return 0; 804 805 if (!ip_callchain__valid(sample->callchain, event)) { 806 pr_debug("call-chain problem with event, skipping it.\n"); 807 ++session->hists.stats.nr_invalid_chains; 808 session->hists.stats.total_invalid_chains += sample->period; 809 return -EINVAL; 810 } 811 return 0; 812 } 813 814 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event, 815 struct perf_event_ops *ops, u64 file_offset) 816 { 817 dump_event(session, event, file_offset, NULL); 818 819 /* These events are processed right away */ 820 switch (event->header.type) { 821 case PERF_RECORD_HEADER_ATTR: 822 return ops->attr(event, session); 823 case PERF_RECORD_HEADER_EVENT_TYPE: 824 return ops->event_type(event, session); 825 case PERF_RECORD_HEADER_TRACING_DATA: 826 /* setup for reading amidst mmap */ 827 lseek(session->fd, file_offset, SEEK_SET); 828 return ops->tracing_data(event, session); 829 case PERF_RECORD_HEADER_BUILD_ID: 830 return ops->build_id(event, session); 831 case PERF_RECORD_FINISHED_ROUND: 832 return ops->finished_round(event, session, ops); 833 default: 834 return -EINVAL; 835 } 836 } 837 838 static int perf_session__process_event(struct perf_session *session, 839 union perf_event *event, 840 struct perf_event_ops *ops, 841 u64 file_offset) 842 { 843 struct perf_sample sample; 844 int ret; 845 846 if (session->header.needs_swap && 847 perf_event__swap_ops[event->header.type]) 848 perf_event__swap_ops[event->header.type](event); 849 850 if (event->header.type >= PERF_RECORD_HEADER_MAX) 851 return -EINVAL; 852 853 hists__inc_nr_events(&session->hists, event->header.type); 854 855 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 856 return perf_session__process_user_event(session, event, ops, file_offset); 857 858 /* 859 * For all kernel events we get the sample data 860 */ 861 ret = perf_session__parse_sample(session, event, &sample); 862 if (ret) 863 return ret; 864 865 /* Preprocess sample records - precheck callchains */ 866 if (perf_session__preprocess_sample(session, event, &sample)) 867 return 0; 868 869 if (ops->ordered_samples) { 870 ret = perf_session_queue_event(session, event, &sample, 871 file_offset); 872 if (ret != -ETIME) 873 return ret; 874 } 875 876 return perf_session_deliver_event(session, event, &sample, ops, 877 file_offset); 878 } 879 880 void perf_event_header__bswap(struct perf_event_header *self) 881 { 882 self->type = bswap_32(self->type); 883 self->misc = bswap_16(self->misc); 884 self->size = bswap_16(self->size); 885 } 886 887 static struct thread *perf_session__register_idle_thread(struct perf_session *self) 888 { 889 struct thread *thread = perf_session__findnew(self, 0); 890 891 if (thread == NULL || thread__set_comm(thread, "swapper")) { 892 pr_err("problem inserting idle task.\n"); 893 thread = NULL; 894 } 895 896 return thread; 897 } 898 899 static void perf_session__warn_about_errors(const struct perf_session *session, 900 const struct perf_event_ops *ops) 901 { 902 if (ops->lost == perf_event__process_lost && 903 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) { 904 ui__warning("Processed %d events and lost %d chunks!\n\n" 905 "Check IO/CPU overload!\n\n", 906 session->hists.stats.nr_events[0], 907 session->hists.stats.nr_events[PERF_RECORD_LOST]); 908 } 909 910 if (session->hists.stats.nr_unknown_events != 0) { 911 ui__warning("Found %u unknown events!\n\n" 912 "Is this an older tool processing a perf.data " 913 "file generated by a more recent tool?\n\n" 914 "If that is not the case, consider " 915 "reporting to linux-kernel@vger.kernel.org.\n\n", 916 session->hists.stats.nr_unknown_events); 917 } 918 919 if (session->hists.stats.nr_unknown_id != 0) { 920 ui__warning("%u samples with id not present in the header\n", 921 session->hists.stats.nr_unknown_id); 922 } 923 924 if (session->hists.stats.nr_invalid_chains != 0) { 925 ui__warning("Found invalid callchains!\n\n" 926 "%u out of %u events were discarded for this reason.\n\n" 927 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 928 session->hists.stats.nr_invalid_chains, 929 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]); 930 } 931 } 932 933 #define session_done() (*(volatile int *)(&session_done)) 934 volatile int session_done; 935 936 static int __perf_session__process_pipe_events(struct perf_session *self, 937 struct perf_event_ops *ops) 938 { 939 union perf_event event; 940 uint32_t size; 941 int skip = 0; 942 u64 head; 943 int err; 944 void *p; 945 946 perf_event_ops__fill_defaults(ops); 947 948 head = 0; 949 more: 950 err = readn(self->fd, &event, sizeof(struct perf_event_header)); 951 if (err <= 0) { 952 if (err == 0) 953 goto done; 954 955 pr_err("failed to read event header\n"); 956 goto out_err; 957 } 958 959 if (self->header.needs_swap) 960 perf_event_header__bswap(&event.header); 961 962 size = event.header.size; 963 if (size == 0) 964 size = 8; 965 966 p = &event; 967 p += sizeof(struct perf_event_header); 968 969 if (size - sizeof(struct perf_event_header)) { 970 err = readn(self->fd, p, size - sizeof(struct perf_event_header)); 971 if (err <= 0) { 972 if (err == 0) { 973 pr_err("unexpected end of event stream\n"); 974 goto done; 975 } 976 977 pr_err("failed to read event data\n"); 978 goto out_err; 979 } 980 } 981 982 if (size == 0 || 983 (skip = perf_session__process_event(self, &event, ops, head)) < 0) { 984 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 985 head, event.header.size, event.header.type); 986 /* 987 * assume we lost track of the stream, check alignment, and 988 * increment a single u64 in the hope to catch on again 'soon'. 989 */ 990 if (unlikely(head & 7)) 991 head &= ~7ULL; 992 993 size = 8; 994 } 995 996 head += size; 997 998 if (skip > 0) 999 head += skip; 1000 1001 if (!session_done()) 1002 goto more; 1003 done: 1004 err = 0; 1005 out_err: 1006 perf_session__warn_about_errors(self, ops); 1007 perf_session_free_sample_buffers(self); 1008 return err; 1009 } 1010 1011 static union perf_event * 1012 fetch_mmaped_event(struct perf_session *session, 1013 u64 head, size_t mmap_size, char *buf) 1014 { 1015 union perf_event *event; 1016 1017 /* 1018 * Ensure we have enough space remaining to read 1019 * the size of the event in the headers. 1020 */ 1021 if (head + sizeof(event->header) > mmap_size) 1022 return NULL; 1023 1024 event = (union perf_event *)(buf + head); 1025 1026 if (session->header.needs_swap) 1027 perf_event_header__bswap(&event->header); 1028 1029 if (head + event->header.size > mmap_size) 1030 return NULL; 1031 1032 return event; 1033 } 1034 1035 int __perf_session__process_events(struct perf_session *session, 1036 u64 data_offset, u64 data_size, 1037 u64 file_size, struct perf_event_ops *ops) 1038 { 1039 u64 head, page_offset, file_offset, file_pos, progress_next; 1040 int err, mmap_prot, mmap_flags, map_idx = 0; 1041 size_t page_size, mmap_size; 1042 char *buf, *mmaps[8]; 1043 union perf_event *event; 1044 uint32_t size; 1045 1046 perf_event_ops__fill_defaults(ops); 1047 1048 page_size = sysconf(_SC_PAGESIZE); 1049 1050 page_offset = page_size * (data_offset / page_size); 1051 file_offset = page_offset; 1052 head = data_offset - page_offset; 1053 1054 if (data_offset + data_size < file_size) 1055 file_size = data_offset + data_size; 1056 1057 progress_next = file_size / 16; 1058 1059 mmap_size = session->mmap_window; 1060 if (mmap_size > file_size) 1061 mmap_size = file_size; 1062 1063 memset(mmaps, 0, sizeof(mmaps)); 1064 1065 mmap_prot = PROT_READ; 1066 mmap_flags = MAP_SHARED; 1067 1068 if (session->header.needs_swap) { 1069 mmap_prot |= PROT_WRITE; 1070 mmap_flags = MAP_PRIVATE; 1071 } 1072 remap: 1073 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd, 1074 file_offset); 1075 if (buf == MAP_FAILED) { 1076 pr_err("failed to mmap file\n"); 1077 err = -errno; 1078 goto out_err; 1079 } 1080 mmaps[map_idx] = buf; 1081 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 1082 file_pos = file_offset + head; 1083 1084 more: 1085 event = fetch_mmaped_event(session, head, mmap_size, buf); 1086 if (!event) { 1087 if (mmaps[map_idx]) { 1088 munmap(mmaps[map_idx], mmap_size); 1089 mmaps[map_idx] = NULL; 1090 } 1091 1092 page_offset = page_size * (head / page_size); 1093 file_offset += page_offset; 1094 head -= page_offset; 1095 goto remap; 1096 } 1097 1098 size = event->header.size; 1099 1100 if (size == 0 || 1101 perf_session__process_event(session, event, ops, file_pos) < 0) { 1102 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 1103 file_offset + head, event->header.size, 1104 event->header.type); 1105 /* 1106 * assume we lost track of the stream, check alignment, and 1107 * increment a single u64 in the hope to catch on again 'soon'. 1108 */ 1109 if (unlikely(head & 7)) 1110 head &= ~7ULL; 1111 1112 size = 8; 1113 } 1114 1115 head += size; 1116 file_pos += size; 1117 1118 if (file_pos >= progress_next) { 1119 progress_next += file_size / 16; 1120 ui_progress__update(file_pos, file_size, 1121 "Processing events..."); 1122 } 1123 1124 if (file_pos < file_size) 1125 goto more; 1126 1127 err = 0; 1128 /* do the final flush for ordered samples */ 1129 session->ordered_samples.next_flush = ULLONG_MAX; 1130 flush_sample_queue(session, ops); 1131 out_err: 1132 perf_session__warn_about_errors(session, ops); 1133 perf_session_free_sample_buffers(session); 1134 return err; 1135 } 1136 1137 int perf_session__process_events(struct perf_session *self, 1138 struct perf_event_ops *ops) 1139 { 1140 int err; 1141 1142 if (perf_session__register_idle_thread(self) == NULL) 1143 return -ENOMEM; 1144 1145 if (!self->fd_pipe) 1146 err = __perf_session__process_events(self, 1147 self->header.data_offset, 1148 self->header.data_size, 1149 self->size, ops); 1150 else 1151 err = __perf_session__process_pipe_events(self, ops); 1152 1153 return err; 1154 } 1155 1156 bool perf_session__has_traces(struct perf_session *self, const char *msg) 1157 { 1158 if (!(self->sample_type & PERF_SAMPLE_RAW)) { 1159 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 1160 return false; 1161 } 1162 1163 return true; 1164 } 1165 1166 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps, 1167 const char *symbol_name, 1168 u64 addr) 1169 { 1170 char *bracket; 1171 enum map_type i; 1172 struct ref_reloc_sym *ref; 1173 1174 ref = zalloc(sizeof(struct ref_reloc_sym)); 1175 if (ref == NULL) 1176 return -ENOMEM; 1177 1178 ref->name = strdup(symbol_name); 1179 if (ref->name == NULL) { 1180 free(ref); 1181 return -ENOMEM; 1182 } 1183 1184 bracket = strchr(ref->name, ']'); 1185 if (bracket) 1186 *bracket = '\0'; 1187 1188 ref->addr = addr; 1189 1190 for (i = 0; i < MAP__NR_TYPES; ++i) { 1191 struct kmap *kmap = map__kmap(maps[i]); 1192 kmap->ref_reloc_sym = ref; 1193 } 1194 1195 return 0; 1196 } 1197 1198 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp) 1199 { 1200 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) + 1201 __dsos__fprintf(&self->host_machine.user_dsos, fp) + 1202 machines__fprintf_dsos(&self->machines, fp); 1203 } 1204 1205 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp, 1206 bool with_hits) 1207 { 1208 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits); 1209 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits); 1210 } 1211 1212 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 1213 { 1214 struct perf_evsel *pos; 1215 size_t ret = fprintf(fp, "Aggregated stats:\n"); 1216 1217 ret += hists__fprintf_nr_events(&session->hists, fp); 1218 1219 list_for_each_entry(pos, &session->evlist->entries, node) { 1220 ret += fprintf(fp, "%s stats:\n", event_name(pos)); 1221 ret += hists__fprintf_nr_events(&pos->hists, fp); 1222 } 1223 1224 return ret; 1225 } 1226 1227 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session, 1228 unsigned int type) 1229 { 1230 struct perf_evsel *pos; 1231 1232 list_for_each_entry(pos, &session->evlist->entries, node) { 1233 if (pos->attr.type == type) 1234 return pos; 1235 } 1236 return NULL; 1237 } 1238 1239 void perf_session__print_ip(union perf_event *event, 1240 struct perf_sample *sample, 1241 struct perf_session *session, 1242 int print_sym, int print_dso) 1243 { 1244 struct addr_location al; 1245 const char *symname, *dsoname; 1246 struct callchain_cursor *cursor = &session->callchain_cursor; 1247 struct callchain_cursor_node *node; 1248 1249 if (perf_event__preprocess_sample(event, session, &al, sample, 1250 NULL) < 0) { 1251 error("problem processing %d event, skipping it.\n", 1252 event->header.type); 1253 return; 1254 } 1255 1256 if (symbol_conf.use_callchain && sample->callchain) { 1257 1258 if (perf_session__resolve_callchain(session, al.thread, 1259 sample->callchain, NULL) != 0) { 1260 if (verbose) 1261 error("Failed to resolve callchain. Skipping\n"); 1262 return; 1263 } 1264 callchain_cursor_commit(cursor); 1265 1266 while (1) { 1267 node = callchain_cursor_current(cursor); 1268 if (!node) 1269 break; 1270 1271 printf("\t%16" PRIx64, node->ip); 1272 if (print_sym) { 1273 if (node->sym && node->sym->name) 1274 symname = node->sym->name; 1275 else 1276 symname = ""; 1277 1278 printf(" %s", symname); 1279 } 1280 if (print_dso) { 1281 if (node->map && node->map->dso && node->map->dso->name) 1282 dsoname = node->map->dso->name; 1283 else 1284 dsoname = ""; 1285 1286 printf(" (%s)", dsoname); 1287 } 1288 printf("\n"); 1289 1290 callchain_cursor_advance(cursor); 1291 } 1292 1293 } else { 1294 printf("%16" PRIx64, sample->ip); 1295 if (print_sym) { 1296 if (al.sym && al.sym->name) 1297 symname = al.sym->name; 1298 else 1299 symname = ""; 1300 1301 printf(" %s", symname); 1302 } 1303 1304 if (print_dso) { 1305 if (al.map && al.map->dso && al.map->dso->name) 1306 dsoname = al.map->dso->name; 1307 else 1308 dsoname = ""; 1309 1310 printf(" (%s)", dsoname); 1311 } 1312 } 1313 } 1314 1315 int perf_session__cpu_bitmap(struct perf_session *session, 1316 const char *cpu_list, unsigned long *cpu_bitmap) 1317 { 1318 int i; 1319 struct cpu_map *map; 1320 1321 for (i = 0; i < PERF_TYPE_MAX; ++i) { 1322 struct perf_evsel *evsel; 1323 1324 evsel = perf_session__find_first_evtype(session, i); 1325 if (!evsel) 1326 continue; 1327 1328 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) { 1329 pr_err("File does not contain CPU events. " 1330 "Remove -c option to proceed.\n"); 1331 return -1; 1332 } 1333 } 1334 1335 map = cpu_map__new(cpu_list); 1336 1337 for (i = 0; i < map->nr; i++) { 1338 int cpu = map->map[i]; 1339 1340 if (cpu >= MAX_NR_CPUS) { 1341 pr_err("Requested CPU %d too large. " 1342 "Consider raising MAX_NR_CPUS\n", cpu); 1343 return -1; 1344 } 1345 1346 set_bit(cpu, cpu_bitmap); 1347 } 1348 1349 return 0; 1350 } 1351 1352 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 1353 bool full) 1354 { 1355 struct stat st; 1356 int ret; 1357 1358 if (session == NULL || fp == NULL) 1359 return; 1360 1361 ret = fstat(session->fd, &st); 1362 if (ret == -1) 1363 return; 1364 1365 fprintf(fp, "# ========\n"); 1366 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime)); 1367 perf_header__fprintf_info(session, fp, full); 1368 fprintf(fp, "# ========\n#\n"); 1369 } 1370