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 static void perf_event__attr_swap(union perf_event *event) 417 { 418 size_t size; 419 420 event->attr.attr.type = bswap_32(event->attr.attr.type); 421 event->attr.attr.size = bswap_32(event->attr.attr.size); 422 event->attr.attr.config = bswap_64(event->attr.attr.config); 423 event->attr.attr.sample_period = bswap_64(event->attr.attr.sample_period); 424 event->attr.attr.sample_type = bswap_64(event->attr.attr.sample_type); 425 event->attr.attr.read_format = bswap_64(event->attr.attr.read_format); 426 event->attr.attr.wakeup_events = bswap_32(event->attr.attr.wakeup_events); 427 event->attr.attr.bp_type = bswap_32(event->attr.attr.bp_type); 428 event->attr.attr.bp_addr = bswap_64(event->attr.attr.bp_addr); 429 event->attr.attr.bp_len = bswap_64(event->attr.attr.bp_len); 430 431 size = event->header.size; 432 size -= (void *)&event->attr.id - (void *)event; 433 mem_bswap_64(event->attr.id, size); 434 } 435 436 static void perf_event__event_type_swap(union perf_event *event) 437 { 438 event->event_type.event_type.event_id = 439 bswap_64(event->event_type.event_type.event_id); 440 } 441 442 static void perf_event__tracing_data_swap(union perf_event *event) 443 { 444 event->tracing_data.size = bswap_32(event->tracing_data.size); 445 } 446 447 typedef void (*perf_event__swap_op)(union perf_event *event); 448 449 static perf_event__swap_op perf_event__swap_ops[] = { 450 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 451 [PERF_RECORD_COMM] = perf_event__comm_swap, 452 [PERF_RECORD_FORK] = perf_event__task_swap, 453 [PERF_RECORD_EXIT] = perf_event__task_swap, 454 [PERF_RECORD_LOST] = perf_event__all64_swap, 455 [PERF_RECORD_READ] = perf_event__read_swap, 456 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 457 [PERF_RECORD_HEADER_ATTR] = perf_event__attr_swap, 458 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 459 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 460 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 461 [PERF_RECORD_HEADER_MAX] = NULL, 462 }; 463 464 struct sample_queue { 465 u64 timestamp; 466 u64 file_offset; 467 union perf_event *event; 468 struct list_head list; 469 }; 470 471 static void perf_session_free_sample_buffers(struct perf_session *session) 472 { 473 struct ordered_samples *os = &session->ordered_samples; 474 475 while (!list_empty(&os->to_free)) { 476 struct sample_queue *sq; 477 478 sq = list_entry(os->to_free.next, struct sample_queue, list); 479 list_del(&sq->list); 480 free(sq); 481 } 482 } 483 484 static int perf_session_deliver_event(struct perf_session *session, 485 union perf_event *event, 486 struct perf_sample *sample, 487 struct perf_event_ops *ops, 488 u64 file_offset); 489 490 static void flush_sample_queue(struct perf_session *s, 491 struct perf_event_ops *ops) 492 { 493 struct ordered_samples *os = &s->ordered_samples; 494 struct list_head *head = &os->samples; 495 struct sample_queue *tmp, *iter; 496 struct perf_sample sample; 497 u64 limit = os->next_flush; 498 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL; 499 int ret; 500 501 if (!ops->ordered_samples || !limit) 502 return; 503 504 list_for_each_entry_safe(iter, tmp, head, list) { 505 if (iter->timestamp > limit) 506 break; 507 508 ret = perf_session__parse_sample(s, iter->event, &sample); 509 if (ret) 510 pr_err("Can't parse sample, err = %d\n", ret); 511 else 512 perf_session_deliver_event(s, iter->event, &sample, ops, 513 iter->file_offset); 514 515 os->last_flush = iter->timestamp; 516 list_del(&iter->list); 517 list_add(&iter->list, &os->sample_cache); 518 } 519 520 if (list_empty(head)) { 521 os->last_sample = NULL; 522 } else if (last_ts <= limit) { 523 os->last_sample = 524 list_entry(head->prev, struct sample_queue, list); 525 } 526 } 527 528 /* 529 * When perf record finishes a pass on every buffers, it records this pseudo 530 * event. 531 * We record the max timestamp t found in the pass n. 532 * Assuming these timestamps are monotonic across cpus, we know that if 533 * a buffer still has events with timestamps below t, they will be all 534 * available and then read in the pass n + 1. 535 * Hence when we start to read the pass n + 2, we can safely flush every 536 * events with timestamps below t. 537 * 538 * ============ PASS n ================= 539 * CPU 0 | CPU 1 540 * | 541 * cnt1 timestamps | cnt2 timestamps 542 * 1 | 2 543 * 2 | 3 544 * - | 4 <--- max recorded 545 * 546 * ============ PASS n + 1 ============== 547 * CPU 0 | CPU 1 548 * | 549 * cnt1 timestamps | cnt2 timestamps 550 * 3 | 5 551 * 4 | 6 552 * 5 | 7 <---- max recorded 553 * 554 * Flush every events below timestamp 4 555 * 556 * ============ PASS n + 2 ============== 557 * CPU 0 | CPU 1 558 * | 559 * cnt1 timestamps | cnt2 timestamps 560 * 6 | 8 561 * 7 | 9 562 * - | 10 563 * 564 * Flush every events below timestamp 7 565 * etc... 566 */ 567 static int process_finished_round(union perf_event *event __used, 568 struct perf_session *session, 569 struct perf_event_ops *ops) 570 { 571 flush_sample_queue(session, ops); 572 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp; 573 574 return 0; 575 } 576 577 /* The queue is ordered by time */ 578 static void __queue_event(struct sample_queue *new, struct perf_session *s) 579 { 580 struct ordered_samples *os = &s->ordered_samples; 581 struct sample_queue *sample = os->last_sample; 582 u64 timestamp = new->timestamp; 583 struct list_head *p; 584 585 os->last_sample = new; 586 587 if (!sample) { 588 list_add(&new->list, &os->samples); 589 os->max_timestamp = timestamp; 590 return; 591 } 592 593 /* 594 * last_sample might point to some random place in the list as it's 595 * the last queued event. We expect that the new event is close to 596 * this. 597 */ 598 if (sample->timestamp <= timestamp) { 599 while (sample->timestamp <= timestamp) { 600 p = sample->list.next; 601 if (p == &os->samples) { 602 list_add_tail(&new->list, &os->samples); 603 os->max_timestamp = timestamp; 604 return; 605 } 606 sample = list_entry(p, struct sample_queue, list); 607 } 608 list_add_tail(&new->list, &sample->list); 609 } else { 610 while (sample->timestamp > timestamp) { 611 p = sample->list.prev; 612 if (p == &os->samples) { 613 list_add(&new->list, &os->samples); 614 return; 615 } 616 sample = list_entry(p, struct sample_queue, list); 617 } 618 list_add(&new->list, &sample->list); 619 } 620 } 621 622 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue)) 623 624 static int perf_session_queue_event(struct perf_session *s, union perf_event *event, 625 struct perf_sample *sample, u64 file_offset) 626 { 627 struct ordered_samples *os = &s->ordered_samples; 628 struct list_head *sc = &os->sample_cache; 629 u64 timestamp = sample->time; 630 struct sample_queue *new; 631 632 if (!timestamp || timestamp == ~0ULL) 633 return -ETIME; 634 635 if (timestamp < s->ordered_samples.last_flush) { 636 printf("Warning: Timestamp below last timeslice flush\n"); 637 return -EINVAL; 638 } 639 640 if (!list_empty(sc)) { 641 new = list_entry(sc->next, struct sample_queue, list); 642 list_del(&new->list); 643 } else if (os->sample_buffer) { 644 new = os->sample_buffer + os->sample_buffer_idx; 645 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER) 646 os->sample_buffer = NULL; 647 } else { 648 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new)); 649 if (!os->sample_buffer) 650 return -ENOMEM; 651 list_add(&os->sample_buffer->list, &os->to_free); 652 os->sample_buffer_idx = 2; 653 new = os->sample_buffer + 1; 654 } 655 656 new->timestamp = timestamp; 657 new->file_offset = file_offset; 658 new->event = event; 659 660 __queue_event(new, s); 661 662 return 0; 663 } 664 665 static void callchain__printf(struct perf_sample *sample) 666 { 667 unsigned int i; 668 669 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr); 670 671 for (i = 0; i < sample->callchain->nr; i++) 672 printf("..... %2d: %016" PRIx64 "\n", 673 i, sample->callchain->ips[i]); 674 } 675 676 static void perf_session__print_tstamp(struct perf_session *session, 677 union perf_event *event, 678 struct perf_sample *sample) 679 { 680 if (event->header.type != PERF_RECORD_SAMPLE && 681 !session->sample_id_all) { 682 fputs("-1 -1 ", stdout); 683 return; 684 } 685 686 if ((session->sample_type & PERF_SAMPLE_CPU)) 687 printf("%u ", sample->cpu); 688 689 if (session->sample_type & PERF_SAMPLE_TIME) 690 printf("%" PRIu64 " ", sample->time); 691 } 692 693 static void dump_event(struct perf_session *session, union perf_event *event, 694 u64 file_offset, struct perf_sample *sample) 695 { 696 if (!dump_trace) 697 return; 698 699 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 700 file_offset, event->header.size, event->header.type); 701 702 trace_event(event); 703 704 if (sample) 705 perf_session__print_tstamp(session, event, sample); 706 707 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 708 event->header.size, perf_event__name(event->header.type)); 709 } 710 711 static void dump_sample(struct perf_session *session, union perf_event *event, 712 struct perf_sample *sample) 713 { 714 if (!dump_trace) 715 return; 716 717 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 718 event->header.misc, sample->pid, sample->tid, sample->ip, 719 sample->period, sample->addr); 720 721 if (session->sample_type & PERF_SAMPLE_CALLCHAIN) 722 callchain__printf(sample); 723 } 724 725 static int perf_session_deliver_event(struct perf_session *session, 726 union perf_event *event, 727 struct perf_sample *sample, 728 struct perf_event_ops *ops, 729 u64 file_offset) 730 { 731 struct perf_evsel *evsel; 732 733 dump_event(session, event, file_offset, sample); 734 735 switch (event->header.type) { 736 case PERF_RECORD_SAMPLE: 737 dump_sample(session, event, sample); 738 evsel = perf_evlist__id2evsel(session->evlist, sample->id); 739 if (evsel == NULL) { 740 ++session->hists.stats.nr_unknown_id; 741 return -1; 742 } 743 return ops->sample(event, sample, evsel, session); 744 case PERF_RECORD_MMAP: 745 return ops->mmap(event, sample, session); 746 case PERF_RECORD_COMM: 747 return ops->comm(event, sample, session); 748 case PERF_RECORD_FORK: 749 return ops->fork(event, sample, session); 750 case PERF_RECORD_EXIT: 751 return ops->exit(event, sample, session); 752 case PERF_RECORD_LOST: 753 return ops->lost(event, sample, session); 754 case PERF_RECORD_READ: 755 return ops->read(event, sample, session); 756 case PERF_RECORD_THROTTLE: 757 return ops->throttle(event, sample, session); 758 case PERF_RECORD_UNTHROTTLE: 759 return ops->unthrottle(event, sample, session); 760 default: 761 ++session->hists.stats.nr_unknown_events; 762 return -1; 763 } 764 } 765 766 static int perf_session__preprocess_sample(struct perf_session *session, 767 union perf_event *event, struct perf_sample *sample) 768 { 769 if (event->header.type != PERF_RECORD_SAMPLE || 770 !(session->sample_type & PERF_SAMPLE_CALLCHAIN)) 771 return 0; 772 773 if (!ip_callchain__valid(sample->callchain, event)) { 774 pr_debug("call-chain problem with event, skipping it.\n"); 775 ++session->hists.stats.nr_invalid_chains; 776 session->hists.stats.total_invalid_chains += sample->period; 777 return -EINVAL; 778 } 779 return 0; 780 } 781 782 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event, 783 struct perf_event_ops *ops, u64 file_offset) 784 { 785 dump_event(session, event, file_offset, NULL); 786 787 /* These events are processed right away */ 788 switch (event->header.type) { 789 case PERF_RECORD_HEADER_ATTR: 790 return ops->attr(event, session); 791 case PERF_RECORD_HEADER_EVENT_TYPE: 792 return ops->event_type(event, session); 793 case PERF_RECORD_HEADER_TRACING_DATA: 794 /* setup for reading amidst mmap */ 795 lseek(session->fd, file_offset, SEEK_SET); 796 return ops->tracing_data(event, session); 797 case PERF_RECORD_HEADER_BUILD_ID: 798 return ops->build_id(event, session); 799 case PERF_RECORD_FINISHED_ROUND: 800 return ops->finished_round(event, session, ops); 801 default: 802 return -EINVAL; 803 } 804 } 805 806 static int perf_session__process_event(struct perf_session *session, 807 union perf_event *event, 808 struct perf_event_ops *ops, 809 u64 file_offset) 810 { 811 struct perf_sample sample; 812 int ret; 813 814 if (session->header.needs_swap && 815 perf_event__swap_ops[event->header.type]) 816 perf_event__swap_ops[event->header.type](event); 817 818 if (event->header.type >= PERF_RECORD_HEADER_MAX) 819 return -EINVAL; 820 821 hists__inc_nr_events(&session->hists, event->header.type); 822 823 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 824 return perf_session__process_user_event(session, event, ops, file_offset); 825 826 /* 827 * For all kernel events we get the sample data 828 */ 829 ret = perf_session__parse_sample(session, event, &sample); 830 if (ret) 831 return ret; 832 833 /* Preprocess sample records - precheck callchains */ 834 if (perf_session__preprocess_sample(session, event, &sample)) 835 return 0; 836 837 if (ops->ordered_samples) { 838 ret = perf_session_queue_event(session, event, &sample, 839 file_offset); 840 if (ret != -ETIME) 841 return ret; 842 } 843 844 return perf_session_deliver_event(session, event, &sample, ops, 845 file_offset); 846 } 847 848 void perf_event_header__bswap(struct perf_event_header *self) 849 { 850 self->type = bswap_32(self->type); 851 self->misc = bswap_16(self->misc); 852 self->size = bswap_16(self->size); 853 } 854 855 static struct thread *perf_session__register_idle_thread(struct perf_session *self) 856 { 857 struct thread *thread = perf_session__findnew(self, 0); 858 859 if (thread == NULL || thread__set_comm(thread, "swapper")) { 860 pr_err("problem inserting idle task.\n"); 861 thread = NULL; 862 } 863 864 return thread; 865 } 866 867 static void perf_session__warn_about_errors(const struct perf_session *session, 868 const struct perf_event_ops *ops) 869 { 870 if (ops->lost == perf_event__process_lost && 871 session->hists.stats.total_lost != 0) { 872 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64 873 "!\n\nCheck IO/CPU overload!\n\n", 874 session->hists.stats.total_period, 875 session->hists.stats.total_lost); 876 } 877 878 if (session->hists.stats.nr_unknown_events != 0) { 879 ui__warning("Found %u unknown events!\n\n" 880 "Is this an older tool processing a perf.data " 881 "file generated by a more recent tool?\n\n" 882 "If that is not the case, consider " 883 "reporting to linux-kernel@vger.kernel.org.\n\n", 884 session->hists.stats.nr_unknown_events); 885 } 886 887 if (session->hists.stats.nr_unknown_id != 0) { 888 ui__warning("%u samples with id not present in the header\n", 889 session->hists.stats.nr_unknown_id); 890 } 891 892 if (session->hists.stats.nr_invalid_chains != 0) { 893 ui__warning("Found invalid callchains!\n\n" 894 "%u out of %u events were discarded for this reason.\n\n" 895 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 896 session->hists.stats.nr_invalid_chains, 897 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]); 898 } 899 } 900 901 #define session_done() (*(volatile int *)(&session_done)) 902 volatile int session_done; 903 904 static int __perf_session__process_pipe_events(struct perf_session *self, 905 struct perf_event_ops *ops) 906 { 907 union perf_event event; 908 uint32_t size; 909 int skip = 0; 910 u64 head; 911 int err; 912 void *p; 913 914 perf_event_ops__fill_defaults(ops); 915 916 head = 0; 917 more: 918 err = readn(self->fd, &event, sizeof(struct perf_event_header)); 919 if (err <= 0) { 920 if (err == 0) 921 goto done; 922 923 pr_err("failed to read event header\n"); 924 goto out_err; 925 } 926 927 if (self->header.needs_swap) 928 perf_event_header__bswap(&event.header); 929 930 size = event.header.size; 931 if (size == 0) 932 size = 8; 933 934 p = &event; 935 p += sizeof(struct perf_event_header); 936 937 if (size - sizeof(struct perf_event_header)) { 938 err = readn(self->fd, p, size - sizeof(struct perf_event_header)); 939 if (err <= 0) { 940 if (err == 0) { 941 pr_err("unexpected end of event stream\n"); 942 goto done; 943 } 944 945 pr_err("failed to read event data\n"); 946 goto out_err; 947 } 948 } 949 950 if (size == 0 || 951 (skip = perf_session__process_event(self, &event, ops, head)) < 0) { 952 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 953 head, event.header.size, event.header.type); 954 /* 955 * assume we lost track of the stream, check alignment, and 956 * increment a single u64 in the hope to catch on again 'soon'. 957 */ 958 if (unlikely(head & 7)) 959 head &= ~7ULL; 960 961 size = 8; 962 } 963 964 head += size; 965 966 if (skip > 0) 967 head += skip; 968 969 if (!session_done()) 970 goto more; 971 done: 972 err = 0; 973 out_err: 974 perf_session__warn_about_errors(self, ops); 975 perf_session_free_sample_buffers(self); 976 return err; 977 } 978 979 static union perf_event * 980 fetch_mmaped_event(struct perf_session *session, 981 u64 head, size_t mmap_size, char *buf) 982 { 983 union perf_event *event; 984 985 /* 986 * Ensure we have enough space remaining to read 987 * the size of the event in the headers. 988 */ 989 if (head + sizeof(event->header) > mmap_size) 990 return NULL; 991 992 event = (union perf_event *)(buf + head); 993 994 if (session->header.needs_swap) 995 perf_event_header__bswap(&event->header); 996 997 if (head + event->header.size > mmap_size) 998 return NULL; 999 1000 return event; 1001 } 1002 1003 int __perf_session__process_events(struct perf_session *session, 1004 u64 data_offset, u64 data_size, 1005 u64 file_size, struct perf_event_ops *ops) 1006 { 1007 u64 head, page_offset, file_offset, file_pos, progress_next; 1008 int err, mmap_prot, mmap_flags, map_idx = 0; 1009 struct ui_progress *progress; 1010 size_t page_size, mmap_size; 1011 char *buf, *mmaps[8]; 1012 union perf_event *event; 1013 uint32_t size; 1014 1015 perf_event_ops__fill_defaults(ops); 1016 1017 page_size = sysconf(_SC_PAGESIZE); 1018 1019 page_offset = page_size * (data_offset / page_size); 1020 file_offset = page_offset; 1021 head = data_offset - page_offset; 1022 1023 if (data_offset + data_size < file_size) 1024 file_size = data_offset + data_size; 1025 1026 progress_next = file_size / 16; 1027 progress = ui_progress__new("Processing events...", file_size); 1028 if (progress == NULL) 1029 return -1; 1030 1031 mmap_size = session->mmap_window; 1032 if (mmap_size > file_size) 1033 mmap_size = file_size; 1034 1035 memset(mmaps, 0, sizeof(mmaps)); 1036 1037 mmap_prot = PROT_READ; 1038 mmap_flags = MAP_SHARED; 1039 1040 if (session->header.needs_swap) { 1041 mmap_prot |= PROT_WRITE; 1042 mmap_flags = MAP_PRIVATE; 1043 } 1044 remap: 1045 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd, 1046 file_offset); 1047 if (buf == MAP_FAILED) { 1048 pr_err("failed to mmap file\n"); 1049 err = -errno; 1050 goto out_err; 1051 } 1052 mmaps[map_idx] = buf; 1053 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 1054 file_pos = file_offset + head; 1055 1056 more: 1057 event = fetch_mmaped_event(session, head, mmap_size, buf); 1058 if (!event) { 1059 if (mmaps[map_idx]) { 1060 munmap(mmaps[map_idx], mmap_size); 1061 mmaps[map_idx] = NULL; 1062 } 1063 1064 page_offset = page_size * (head / page_size); 1065 file_offset += page_offset; 1066 head -= page_offset; 1067 goto remap; 1068 } 1069 1070 size = event->header.size; 1071 1072 if (size == 0 || 1073 perf_session__process_event(session, event, ops, file_pos) < 0) { 1074 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n", 1075 file_offset + head, event->header.size, 1076 event->header.type); 1077 /* 1078 * assume we lost track of the stream, check alignment, and 1079 * increment a single u64 in the hope to catch on again 'soon'. 1080 */ 1081 if (unlikely(head & 7)) 1082 head &= ~7ULL; 1083 1084 size = 8; 1085 } 1086 1087 head += size; 1088 file_pos += size; 1089 1090 if (file_pos >= progress_next) { 1091 progress_next += file_size / 16; 1092 ui_progress__update(progress, file_pos); 1093 } 1094 1095 if (file_pos < file_size) 1096 goto more; 1097 1098 err = 0; 1099 /* do the final flush for ordered samples */ 1100 session->ordered_samples.next_flush = ULLONG_MAX; 1101 flush_sample_queue(session, ops); 1102 out_err: 1103 ui_progress__delete(progress); 1104 perf_session__warn_about_errors(session, ops); 1105 perf_session_free_sample_buffers(session); 1106 return err; 1107 } 1108 1109 int perf_session__process_events(struct perf_session *self, 1110 struct perf_event_ops *ops) 1111 { 1112 int err; 1113 1114 if (perf_session__register_idle_thread(self) == NULL) 1115 return -ENOMEM; 1116 1117 if (!self->fd_pipe) 1118 err = __perf_session__process_events(self, 1119 self->header.data_offset, 1120 self->header.data_size, 1121 self->size, ops); 1122 else 1123 err = __perf_session__process_pipe_events(self, ops); 1124 1125 return err; 1126 } 1127 1128 bool perf_session__has_traces(struct perf_session *self, const char *msg) 1129 { 1130 if (!(self->sample_type & PERF_SAMPLE_RAW)) { 1131 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 1132 return false; 1133 } 1134 1135 return true; 1136 } 1137 1138 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps, 1139 const char *symbol_name, 1140 u64 addr) 1141 { 1142 char *bracket; 1143 enum map_type i; 1144 struct ref_reloc_sym *ref; 1145 1146 ref = zalloc(sizeof(struct ref_reloc_sym)); 1147 if (ref == NULL) 1148 return -ENOMEM; 1149 1150 ref->name = strdup(symbol_name); 1151 if (ref->name == NULL) { 1152 free(ref); 1153 return -ENOMEM; 1154 } 1155 1156 bracket = strchr(ref->name, ']'); 1157 if (bracket) 1158 *bracket = '\0'; 1159 1160 ref->addr = addr; 1161 1162 for (i = 0; i < MAP__NR_TYPES; ++i) { 1163 struct kmap *kmap = map__kmap(maps[i]); 1164 kmap->ref_reloc_sym = ref; 1165 } 1166 1167 return 0; 1168 } 1169 1170 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp) 1171 { 1172 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) + 1173 __dsos__fprintf(&self->host_machine.user_dsos, fp) + 1174 machines__fprintf_dsos(&self->machines, fp); 1175 } 1176 1177 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp, 1178 bool with_hits) 1179 { 1180 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits); 1181 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits); 1182 } 1183 1184 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 1185 { 1186 struct perf_evsel *pos; 1187 size_t ret = fprintf(fp, "Aggregated stats:\n"); 1188 1189 ret += hists__fprintf_nr_events(&session->hists, fp); 1190 1191 list_for_each_entry(pos, &session->evlist->entries, node) { 1192 ret += fprintf(fp, "%s stats:\n", event_name(pos)); 1193 ret += hists__fprintf_nr_events(&pos->hists, fp); 1194 } 1195 1196 return ret; 1197 } 1198 1199 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session, 1200 unsigned int type) 1201 { 1202 struct perf_evsel *pos; 1203 1204 list_for_each_entry(pos, &session->evlist->entries, node) { 1205 if (pos->attr.type == type) 1206 return pos; 1207 } 1208 return NULL; 1209 } 1210 1211 void perf_session__print_ip(union perf_event *event, 1212 struct perf_sample *sample, 1213 struct perf_session *session, 1214 int print_sym, int print_dso) 1215 { 1216 struct addr_location al; 1217 const char *symname, *dsoname; 1218 struct callchain_cursor *cursor = &session->callchain_cursor; 1219 struct callchain_cursor_node *node; 1220 1221 if (perf_event__preprocess_sample(event, session, &al, sample, 1222 NULL) < 0) { 1223 error("problem processing %d event, skipping it.\n", 1224 event->header.type); 1225 return; 1226 } 1227 1228 if (symbol_conf.use_callchain && sample->callchain) { 1229 1230 if (perf_session__resolve_callchain(session, al.thread, 1231 sample->callchain, NULL) != 0) { 1232 if (verbose) 1233 error("Failed to resolve callchain. Skipping\n"); 1234 return; 1235 } 1236 callchain_cursor_commit(cursor); 1237 1238 while (1) { 1239 node = callchain_cursor_current(cursor); 1240 if (!node) 1241 break; 1242 1243 printf("\t%16" PRIx64, node->ip); 1244 if (print_sym) { 1245 if (node->sym && node->sym->name) 1246 symname = node->sym->name; 1247 else 1248 symname = ""; 1249 1250 printf(" %s", symname); 1251 } 1252 if (print_dso) { 1253 if (node->map && node->map->dso && node->map->dso->name) 1254 dsoname = node->map->dso->name; 1255 else 1256 dsoname = ""; 1257 1258 printf(" (%s)", dsoname); 1259 } 1260 printf("\n"); 1261 1262 callchain_cursor_advance(cursor); 1263 } 1264 1265 } else { 1266 printf("%16" PRIx64, al.addr); 1267 if (print_sym) { 1268 if (al.sym && al.sym->name) 1269 symname = al.sym->name; 1270 else 1271 symname = ""; 1272 1273 printf(" %s", symname); 1274 } 1275 1276 if (print_dso) { 1277 if (al.map && al.map->dso && al.map->dso->name) 1278 dsoname = al.map->dso->name; 1279 else 1280 dsoname = ""; 1281 1282 printf(" (%s)", dsoname); 1283 } 1284 } 1285 } 1286 1287 int perf_session__cpu_bitmap(struct perf_session *session, 1288 const char *cpu_list, unsigned long *cpu_bitmap) 1289 { 1290 int i; 1291 struct cpu_map *map; 1292 1293 for (i = 0; i < PERF_TYPE_MAX; ++i) { 1294 struct perf_evsel *evsel; 1295 1296 evsel = perf_session__find_first_evtype(session, i); 1297 if (!evsel) 1298 continue; 1299 1300 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) { 1301 pr_err("File does not contain CPU events. " 1302 "Remove -c option to proceed.\n"); 1303 return -1; 1304 } 1305 } 1306 1307 map = cpu_map__new(cpu_list); 1308 1309 for (i = 0; i < map->nr; i++) { 1310 int cpu = map->map[i]; 1311 1312 if (cpu >= MAX_NR_CPUS) { 1313 pr_err("Requested CPU %d too large. " 1314 "Consider raising MAX_NR_CPUS\n", cpu); 1315 return -1; 1316 } 1317 1318 set_bit(cpu, cpu_bitmap); 1319 } 1320 1321 return 0; 1322 } 1323