1 #define _FILE_OFFSET_BITS 64 2 3 #include <linux/kernel.h> 4 5 #include <byteswap.h> 6 #include <unistd.h> 7 #include <sys/types.h> 8 #include <sys/mman.h> 9 10 #include "evlist.h" 11 #include "evsel.h" 12 #include "session.h" 13 #include "tool.h" 14 #include "sort.h" 15 #include "util.h" 16 #include "cpumap.h" 17 #include "event-parse.h" 18 #include "perf_regs.h" 19 #include "unwind.h" 20 21 static int perf_session__open(struct perf_session *self, bool force) 22 { 23 struct stat input_stat; 24 25 if (!strcmp(self->filename, "-")) { 26 self->fd_pipe = true; 27 self->fd = STDIN_FILENO; 28 29 if (perf_session__read_header(self, self->fd) < 0) 30 pr_err("incompatible file format (rerun with -v to learn more)"); 31 32 return 0; 33 } 34 35 self->fd = open(self->filename, O_RDONLY); 36 if (self->fd < 0) { 37 int err = errno; 38 39 pr_err("failed to open %s: %s", self->filename, strerror(err)); 40 if (err == ENOENT && !strcmp(self->filename, "perf.data")) 41 pr_err(" (try 'perf record' first)"); 42 pr_err("\n"); 43 return -errno; 44 } 45 46 if (fstat(self->fd, &input_stat) < 0) 47 goto out_close; 48 49 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) { 50 pr_err("file %s not owned by current user or root\n", 51 self->filename); 52 goto out_close; 53 } 54 55 if (!input_stat.st_size) { 56 pr_info("zero-sized file (%s), nothing to do!\n", 57 self->filename); 58 goto out_close; 59 } 60 61 if (perf_session__read_header(self, self->fd) < 0) { 62 pr_err("incompatible file format (rerun with -v to learn more)"); 63 goto out_close; 64 } 65 66 if (!perf_evlist__valid_sample_type(self->evlist)) { 67 pr_err("non matching sample_type"); 68 goto out_close; 69 } 70 71 if (!perf_evlist__valid_sample_id_all(self->evlist)) { 72 pr_err("non matching sample_id_all"); 73 goto out_close; 74 } 75 76 self->size = input_stat.st_size; 77 return 0; 78 79 out_close: 80 close(self->fd); 81 self->fd = -1; 82 return -1; 83 } 84 85 void perf_session__set_id_hdr_size(struct perf_session *session) 86 { 87 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist); 88 89 session->host_machine.id_hdr_size = id_hdr_size; 90 machines__set_id_hdr_size(&session->machines, id_hdr_size); 91 } 92 93 int perf_session__create_kernel_maps(struct perf_session *self) 94 { 95 int ret = machine__create_kernel_maps(&self->host_machine); 96 97 if (ret >= 0) 98 ret = machines__create_guest_kernel_maps(&self->machines); 99 return ret; 100 } 101 102 static void perf_session__destroy_kernel_maps(struct perf_session *self) 103 { 104 machine__destroy_kernel_maps(&self->host_machine); 105 machines__destroy_guest_kernel_maps(&self->machines); 106 } 107 108 struct perf_session *perf_session__new(const char *filename, int mode, 109 bool force, bool repipe, 110 struct perf_tool *tool) 111 { 112 struct perf_session *self; 113 struct stat st; 114 size_t len; 115 116 if (!filename || !strlen(filename)) { 117 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 118 filename = "-"; 119 else 120 filename = "perf.data"; 121 } 122 123 len = strlen(filename); 124 self = zalloc(sizeof(*self) + len); 125 126 if (self == NULL) 127 goto out; 128 129 memcpy(self->filename, filename, len); 130 /* 131 * On 64bit we can mmap the data file in one go. No need for tiny mmap 132 * slices. On 32bit we use 32MB. 133 */ 134 #if BITS_PER_LONG == 64 135 self->mmap_window = ULLONG_MAX; 136 #else 137 self->mmap_window = 32 * 1024 * 1024ULL; 138 #endif 139 self->machines = RB_ROOT; 140 self->repipe = repipe; 141 INIT_LIST_HEAD(&self->ordered_samples.samples); 142 INIT_LIST_HEAD(&self->ordered_samples.sample_cache); 143 INIT_LIST_HEAD(&self->ordered_samples.to_free); 144 machine__init(&self->host_machine, "", HOST_KERNEL_ID); 145 hists__init(&self->hists); 146 147 if (mode == O_RDONLY) { 148 if (perf_session__open(self, force) < 0) 149 goto out_delete; 150 perf_session__set_id_hdr_size(self); 151 } else if (mode == O_WRONLY) { 152 /* 153 * In O_RDONLY mode this will be performed when reading the 154 * kernel MMAP event, in perf_event__process_mmap(). 155 */ 156 if (perf_session__create_kernel_maps(self) < 0) 157 goto out_delete; 158 } 159 160 if (tool && tool->ordering_requires_timestamps && 161 tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) { 162 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 163 tool->ordered_samples = false; 164 } 165 166 out: 167 return self; 168 out_delete: 169 perf_session__delete(self); 170 return NULL; 171 } 172 173 static void machine__delete_dead_threads(struct machine *machine) 174 { 175 struct thread *n, *t; 176 177 list_for_each_entry_safe(t, n, &machine->dead_threads, node) { 178 list_del(&t->node); 179 thread__delete(t); 180 } 181 } 182 183 static void perf_session__delete_dead_threads(struct perf_session *session) 184 { 185 machine__delete_dead_threads(&session->host_machine); 186 } 187 188 static void machine__delete_threads(struct machine *self) 189 { 190 struct rb_node *nd = rb_first(&self->threads); 191 192 while (nd) { 193 struct thread *t = rb_entry(nd, struct thread, rb_node); 194 195 rb_erase(&t->rb_node, &self->threads); 196 nd = rb_next(nd); 197 thread__delete(t); 198 } 199 } 200 201 static void perf_session__delete_threads(struct perf_session *session) 202 { 203 machine__delete_threads(&session->host_machine); 204 } 205 206 void perf_session__delete(struct perf_session *self) 207 { 208 perf_session__destroy_kernel_maps(self); 209 perf_session__delete_dead_threads(self); 210 perf_session__delete_threads(self); 211 machine__exit(&self->host_machine); 212 close(self->fd); 213 free(self); 214 } 215 216 void machine__remove_thread(struct machine *self, struct thread *th) 217 { 218 self->last_match = NULL; 219 rb_erase(&th->rb_node, &self->threads); 220 /* 221 * We may have references to this thread, for instance in some hist_entry 222 * instances, so just move them to a separate list. 223 */ 224 list_add_tail(&th->node, &self->dead_threads); 225 } 226 227 static bool symbol__match_parent_regex(struct symbol *sym) 228 { 229 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0)) 230 return 1; 231 232 return 0; 233 } 234 235 static const u8 cpumodes[] = { 236 PERF_RECORD_MISC_USER, 237 PERF_RECORD_MISC_KERNEL, 238 PERF_RECORD_MISC_GUEST_USER, 239 PERF_RECORD_MISC_GUEST_KERNEL 240 }; 241 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8)) 242 243 static void ip__resolve_ams(struct machine *self, struct thread *thread, 244 struct addr_map_symbol *ams, 245 u64 ip) 246 { 247 struct addr_location al; 248 size_t i; 249 u8 m; 250 251 memset(&al, 0, sizeof(al)); 252 253 for (i = 0; i < NCPUMODES; i++) { 254 m = cpumodes[i]; 255 /* 256 * We cannot use the header.misc hint to determine whether a 257 * branch stack address is user, kernel, guest, hypervisor. 258 * Branches may straddle the kernel/user/hypervisor boundaries. 259 * Thus, we have to try consecutively until we find a match 260 * or else, the symbol is unknown 261 */ 262 thread__find_addr_location(thread, self, m, MAP__FUNCTION, 263 ip, &al, NULL); 264 if (al.sym) 265 goto found; 266 } 267 found: 268 ams->addr = ip; 269 ams->al_addr = al.addr; 270 ams->sym = al.sym; 271 ams->map = al.map; 272 } 273 274 struct branch_info *machine__resolve_bstack(struct machine *self, 275 struct thread *thr, 276 struct branch_stack *bs) 277 { 278 struct branch_info *bi; 279 unsigned int i; 280 281 bi = calloc(bs->nr, sizeof(struct branch_info)); 282 if (!bi) 283 return NULL; 284 285 for (i = 0; i < bs->nr; i++) { 286 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to); 287 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from); 288 bi[i].flags = bs->entries[i].flags; 289 } 290 return bi; 291 } 292 293 static int machine__resolve_callchain_sample(struct machine *machine, 294 struct thread *thread, 295 struct ip_callchain *chain, 296 struct symbol **parent) 297 298 { 299 u8 cpumode = PERF_RECORD_MISC_USER; 300 unsigned int i; 301 int err; 302 303 callchain_cursor_reset(&callchain_cursor); 304 305 if (chain->nr > PERF_MAX_STACK_DEPTH) { 306 pr_warning("corrupted callchain. skipping...\n"); 307 return 0; 308 } 309 310 for (i = 0; i < chain->nr; i++) { 311 u64 ip; 312 struct addr_location al; 313 314 if (callchain_param.order == ORDER_CALLEE) 315 ip = chain->ips[i]; 316 else 317 ip = chain->ips[chain->nr - i - 1]; 318 319 if (ip >= PERF_CONTEXT_MAX) { 320 switch (ip) { 321 case PERF_CONTEXT_HV: 322 cpumode = PERF_RECORD_MISC_HYPERVISOR; 323 break; 324 case PERF_CONTEXT_KERNEL: 325 cpumode = PERF_RECORD_MISC_KERNEL; 326 break; 327 case PERF_CONTEXT_USER: 328 cpumode = PERF_RECORD_MISC_USER; 329 break; 330 default: 331 pr_debug("invalid callchain context: " 332 "%"PRId64"\n", (s64) ip); 333 /* 334 * It seems the callchain is corrupted. 335 * Discard all. 336 */ 337 callchain_cursor_reset(&callchain_cursor); 338 return 0; 339 } 340 continue; 341 } 342 343 al.filtered = false; 344 thread__find_addr_location(thread, machine, cpumode, 345 MAP__FUNCTION, ip, &al, NULL); 346 if (al.sym != NULL) { 347 if (sort__has_parent && !*parent && 348 symbol__match_parent_regex(al.sym)) 349 *parent = al.sym; 350 if (!symbol_conf.use_callchain) 351 break; 352 } 353 354 err = callchain_cursor_append(&callchain_cursor, 355 ip, al.map, al.sym); 356 if (err) 357 return err; 358 } 359 360 return 0; 361 } 362 363 static int unwind_entry(struct unwind_entry *entry, void *arg) 364 { 365 struct callchain_cursor *cursor = arg; 366 return callchain_cursor_append(cursor, entry->ip, 367 entry->map, entry->sym); 368 } 369 370 int machine__resolve_callchain(struct machine *machine, 371 struct perf_evsel *evsel, 372 struct thread *thread, 373 struct perf_sample *sample, 374 struct symbol **parent) 375 376 { 377 int ret; 378 379 callchain_cursor_reset(&callchain_cursor); 380 381 ret = machine__resolve_callchain_sample(machine, thread, 382 sample->callchain, parent); 383 if (ret) 384 return ret; 385 386 /* Can we do dwarf post unwind? */ 387 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) && 388 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER))) 389 return 0; 390 391 return unwind__get_entries(unwind_entry, &callchain_cursor, machine, 392 thread, evsel->attr.sample_regs_user, 393 sample); 394 395 } 396 397 static int process_event_synth_tracing_data_stub(union perf_event *event __used, 398 struct perf_session *session __used) 399 { 400 dump_printf(": unhandled!\n"); 401 return 0; 402 } 403 404 static int process_event_synth_attr_stub(union perf_event *event __used, 405 struct perf_evlist **pevlist __used) 406 { 407 dump_printf(": unhandled!\n"); 408 return 0; 409 } 410 411 static int process_event_sample_stub(struct perf_tool *tool __used, 412 union perf_event *event __used, 413 struct perf_sample *sample __used, 414 struct perf_evsel *evsel __used, 415 struct machine *machine __used) 416 { 417 dump_printf(": unhandled!\n"); 418 return 0; 419 } 420 421 static int process_event_stub(struct perf_tool *tool __used, 422 union perf_event *event __used, 423 struct perf_sample *sample __used, 424 struct machine *machine __used) 425 { 426 dump_printf(": unhandled!\n"); 427 return 0; 428 } 429 430 static int process_finished_round_stub(struct perf_tool *tool __used, 431 union perf_event *event __used, 432 struct perf_session *perf_session __used) 433 { 434 dump_printf(": unhandled!\n"); 435 return 0; 436 } 437 438 static int process_event_type_stub(struct perf_tool *tool __used, 439 union perf_event *event __used) 440 { 441 dump_printf(": unhandled!\n"); 442 return 0; 443 } 444 445 static int process_finished_round(struct perf_tool *tool, 446 union perf_event *event, 447 struct perf_session *session); 448 449 static void perf_tool__fill_defaults(struct perf_tool *tool) 450 { 451 if (tool->sample == NULL) 452 tool->sample = process_event_sample_stub; 453 if (tool->mmap == NULL) 454 tool->mmap = process_event_stub; 455 if (tool->comm == NULL) 456 tool->comm = process_event_stub; 457 if (tool->fork == NULL) 458 tool->fork = process_event_stub; 459 if (tool->exit == NULL) 460 tool->exit = process_event_stub; 461 if (tool->lost == NULL) 462 tool->lost = perf_event__process_lost; 463 if (tool->read == NULL) 464 tool->read = process_event_sample_stub; 465 if (tool->throttle == NULL) 466 tool->throttle = process_event_stub; 467 if (tool->unthrottle == NULL) 468 tool->unthrottle = process_event_stub; 469 if (tool->attr == NULL) 470 tool->attr = process_event_synth_attr_stub; 471 if (tool->event_type == NULL) 472 tool->event_type = process_event_type_stub; 473 if (tool->tracing_data == NULL) 474 tool->tracing_data = process_event_synth_tracing_data_stub; 475 if (tool->build_id == NULL) 476 tool->build_id = process_finished_round_stub; 477 if (tool->finished_round == NULL) { 478 if (tool->ordered_samples) 479 tool->finished_round = process_finished_round; 480 else 481 tool->finished_round = process_finished_round_stub; 482 } 483 } 484 485 void mem_bswap_32(void *src, int byte_size) 486 { 487 u32 *m = src; 488 while (byte_size > 0) { 489 *m = bswap_32(*m); 490 byte_size -= sizeof(u32); 491 ++m; 492 } 493 } 494 495 void mem_bswap_64(void *src, int byte_size) 496 { 497 u64 *m = src; 498 499 while (byte_size > 0) { 500 *m = bswap_64(*m); 501 byte_size -= sizeof(u64); 502 ++m; 503 } 504 } 505 506 static void swap_sample_id_all(union perf_event *event, void *data) 507 { 508 void *end = (void *) event + event->header.size; 509 int size = end - data; 510 511 BUG_ON(size % sizeof(u64)); 512 mem_bswap_64(data, size); 513 } 514 515 static void perf_event__all64_swap(union perf_event *event, 516 bool sample_id_all __used) 517 { 518 struct perf_event_header *hdr = &event->header; 519 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 520 } 521 522 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 523 { 524 event->comm.pid = bswap_32(event->comm.pid); 525 event->comm.tid = bswap_32(event->comm.tid); 526 527 if (sample_id_all) { 528 void *data = &event->comm.comm; 529 530 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 531 swap_sample_id_all(event, data); 532 } 533 } 534 535 static void perf_event__mmap_swap(union perf_event *event, 536 bool sample_id_all) 537 { 538 event->mmap.pid = bswap_32(event->mmap.pid); 539 event->mmap.tid = bswap_32(event->mmap.tid); 540 event->mmap.start = bswap_64(event->mmap.start); 541 event->mmap.len = bswap_64(event->mmap.len); 542 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 543 544 if (sample_id_all) { 545 void *data = &event->mmap.filename; 546 547 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 548 swap_sample_id_all(event, data); 549 } 550 } 551 552 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 553 { 554 event->fork.pid = bswap_32(event->fork.pid); 555 event->fork.tid = bswap_32(event->fork.tid); 556 event->fork.ppid = bswap_32(event->fork.ppid); 557 event->fork.ptid = bswap_32(event->fork.ptid); 558 event->fork.time = bswap_64(event->fork.time); 559 560 if (sample_id_all) 561 swap_sample_id_all(event, &event->fork + 1); 562 } 563 564 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 565 { 566 event->read.pid = bswap_32(event->read.pid); 567 event->read.tid = bswap_32(event->read.tid); 568 event->read.value = bswap_64(event->read.value); 569 event->read.time_enabled = bswap_64(event->read.time_enabled); 570 event->read.time_running = bswap_64(event->read.time_running); 571 event->read.id = bswap_64(event->read.id); 572 573 if (sample_id_all) 574 swap_sample_id_all(event, &event->read + 1); 575 } 576 577 static u8 revbyte(u8 b) 578 { 579 int rev = (b >> 4) | ((b & 0xf) << 4); 580 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 581 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 582 return (u8) rev; 583 } 584 585 /* 586 * XXX this is hack in attempt to carry flags bitfield 587 * throught endian village. ABI says: 588 * 589 * Bit-fields are allocated from right to left (least to most significant) 590 * on little-endian implementations and from left to right (most to least 591 * significant) on big-endian implementations. 592 * 593 * The above seems to be byte specific, so we need to reverse each 594 * byte of the bitfield. 'Internet' also says this might be implementation 595 * specific and we probably need proper fix and carry perf_event_attr 596 * bitfield flags in separate data file FEAT_ section. Thought this seems 597 * to work for now. 598 */ 599 static void swap_bitfield(u8 *p, unsigned len) 600 { 601 unsigned i; 602 603 for (i = 0; i < len; i++) { 604 *p = revbyte(*p); 605 p++; 606 } 607 } 608 609 /* exported for swapping attributes in file header */ 610 void perf_event__attr_swap(struct perf_event_attr *attr) 611 { 612 attr->type = bswap_32(attr->type); 613 attr->size = bswap_32(attr->size); 614 attr->config = bswap_64(attr->config); 615 attr->sample_period = bswap_64(attr->sample_period); 616 attr->sample_type = bswap_64(attr->sample_type); 617 attr->read_format = bswap_64(attr->read_format); 618 attr->wakeup_events = bswap_32(attr->wakeup_events); 619 attr->bp_type = bswap_32(attr->bp_type); 620 attr->bp_addr = bswap_64(attr->bp_addr); 621 attr->bp_len = bswap_64(attr->bp_len); 622 623 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64)); 624 } 625 626 static void perf_event__hdr_attr_swap(union perf_event *event, 627 bool sample_id_all __used) 628 { 629 size_t size; 630 631 perf_event__attr_swap(&event->attr.attr); 632 633 size = event->header.size; 634 size -= (void *)&event->attr.id - (void *)event; 635 mem_bswap_64(event->attr.id, size); 636 } 637 638 static void perf_event__event_type_swap(union perf_event *event, 639 bool sample_id_all __used) 640 { 641 event->event_type.event_type.event_id = 642 bswap_64(event->event_type.event_type.event_id); 643 } 644 645 static void perf_event__tracing_data_swap(union perf_event *event, 646 bool sample_id_all __used) 647 { 648 event->tracing_data.size = bswap_32(event->tracing_data.size); 649 } 650 651 typedef void (*perf_event__swap_op)(union perf_event *event, 652 bool sample_id_all); 653 654 static perf_event__swap_op perf_event__swap_ops[] = { 655 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 656 [PERF_RECORD_COMM] = perf_event__comm_swap, 657 [PERF_RECORD_FORK] = perf_event__task_swap, 658 [PERF_RECORD_EXIT] = perf_event__task_swap, 659 [PERF_RECORD_LOST] = perf_event__all64_swap, 660 [PERF_RECORD_READ] = perf_event__read_swap, 661 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 662 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 663 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 664 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 665 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 666 [PERF_RECORD_HEADER_MAX] = NULL, 667 }; 668 669 struct sample_queue { 670 u64 timestamp; 671 u64 file_offset; 672 union perf_event *event; 673 struct list_head list; 674 }; 675 676 static void perf_session_free_sample_buffers(struct perf_session *session) 677 { 678 struct ordered_samples *os = &session->ordered_samples; 679 680 while (!list_empty(&os->to_free)) { 681 struct sample_queue *sq; 682 683 sq = list_entry(os->to_free.next, struct sample_queue, list); 684 list_del(&sq->list); 685 free(sq); 686 } 687 } 688 689 static int perf_session_deliver_event(struct perf_session *session, 690 union perf_event *event, 691 struct perf_sample *sample, 692 struct perf_tool *tool, 693 u64 file_offset); 694 695 static int flush_sample_queue(struct perf_session *s, 696 struct perf_tool *tool) 697 { 698 struct ordered_samples *os = &s->ordered_samples; 699 struct list_head *head = &os->samples; 700 struct sample_queue *tmp, *iter; 701 struct perf_sample sample; 702 u64 limit = os->next_flush; 703 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL; 704 unsigned idx = 0, progress_next = os->nr_samples / 16; 705 int ret; 706 707 if (!tool->ordered_samples || !limit) 708 return 0; 709 710 list_for_each_entry_safe(iter, tmp, head, list) { 711 if (iter->timestamp > limit) 712 break; 713 714 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample, 715 s->header.needs_swap); 716 if (ret) 717 pr_err("Can't parse sample, err = %d\n", ret); 718 else { 719 ret = perf_session_deliver_event(s, iter->event, &sample, tool, 720 iter->file_offset); 721 if (ret) 722 return ret; 723 } 724 725 os->last_flush = iter->timestamp; 726 list_del(&iter->list); 727 list_add(&iter->list, &os->sample_cache); 728 if (++idx >= progress_next) { 729 progress_next += os->nr_samples / 16; 730 ui_progress__update(idx, os->nr_samples, 731 "Processing time ordered events..."); 732 } 733 } 734 735 if (list_empty(head)) { 736 os->last_sample = NULL; 737 } else if (last_ts <= limit) { 738 os->last_sample = 739 list_entry(head->prev, struct sample_queue, list); 740 } 741 742 os->nr_samples = 0; 743 744 return 0; 745 } 746 747 /* 748 * When perf record finishes a pass on every buffers, it records this pseudo 749 * event. 750 * We record the max timestamp t found in the pass n. 751 * Assuming these timestamps are monotonic across cpus, we know that if 752 * a buffer still has events with timestamps below t, they will be all 753 * available and then read in the pass n + 1. 754 * Hence when we start to read the pass n + 2, we can safely flush every 755 * events with timestamps below t. 756 * 757 * ============ PASS n ================= 758 * CPU 0 | CPU 1 759 * | 760 * cnt1 timestamps | cnt2 timestamps 761 * 1 | 2 762 * 2 | 3 763 * - | 4 <--- max recorded 764 * 765 * ============ PASS n + 1 ============== 766 * CPU 0 | CPU 1 767 * | 768 * cnt1 timestamps | cnt2 timestamps 769 * 3 | 5 770 * 4 | 6 771 * 5 | 7 <---- max recorded 772 * 773 * Flush every events below timestamp 4 774 * 775 * ============ PASS n + 2 ============== 776 * CPU 0 | CPU 1 777 * | 778 * cnt1 timestamps | cnt2 timestamps 779 * 6 | 8 780 * 7 | 9 781 * - | 10 782 * 783 * Flush every events below timestamp 7 784 * etc... 785 */ 786 static int process_finished_round(struct perf_tool *tool, 787 union perf_event *event __used, 788 struct perf_session *session) 789 { 790 int ret = flush_sample_queue(session, tool); 791 if (!ret) 792 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp; 793 794 return ret; 795 } 796 797 /* The queue is ordered by time */ 798 static void __queue_event(struct sample_queue *new, struct perf_session *s) 799 { 800 struct ordered_samples *os = &s->ordered_samples; 801 struct sample_queue *sample = os->last_sample; 802 u64 timestamp = new->timestamp; 803 struct list_head *p; 804 805 ++os->nr_samples; 806 os->last_sample = new; 807 808 if (!sample) { 809 list_add(&new->list, &os->samples); 810 os->max_timestamp = timestamp; 811 return; 812 } 813 814 /* 815 * last_sample might point to some random place in the list as it's 816 * the last queued event. We expect that the new event is close to 817 * this. 818 */ 819 if (sample->timestamp <= timestamp) { 820 while (sample->timestamp <= timestamp) { 821 p = sample->list.next; 822 if (p == &os->samples) { 823 list_add_tail(&new->list, &os->samples); 824 os->max_timestamp = timestamp; 825 return; 826 } 827 sample = list_entry(p, struct sample_queue, list); 828 } 829 list_add_tail(&new->list, &sample->list); 830 } else { 831 while (sample->timestamp > timestamp) { 832 p = sample->list.prev; 833 if (p == &os->samples) { 834 list_add(&new->list, &os->samples); 835 return; 836 } 837 sample = list_entry(p, struct sample_queue, list); 838 } 839 list_add(&new->list, &sample->list); 840 } 841 } 842 843 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue)) 844 845 static int perf_session_queue_event(struct perf_session *s, union perf_event *event, 846 struct perf_sample *sample, u64 file_offset) 847 { 848 struct ordered_samples *os = &s->ordered_samples; 849 struct list_head *sc = &os->sample_cache; 850 u64 timestamp = sample->time; 851 struct sample_queue *new; 852 853 if (!timestamp || timestamp == ~0ULL) 854 return -ETIME; 855 856 if (timestamp < s->ordered_samples.last_flush) { 857 printf("Warning: Timestamp below last timeslice flush\n"); 858 return -EINVAL; 859 } 860 861 if (!list_empty(sc)) { 862 new = list_entry(sc->next, struct sample_queue, list); 863 list_del(&new->list); 864 } else if (os->sample_buffer) { 865 new = os->sample_buffer + os->sample_buffer_idx; 866 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER) 867 os->sample_buffer = NULL; 868 } else { 869 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new)); 870 if (!os->sample_buffer) 871 return -ENOMEM; 872 list_add(&os->sample_buffer->list, &os->to_free); 873 os->sample_buffer_idx = 2; 874 new = os->sample_buffer + 1; 875 } 876 877 new->timestamp = timestamp; 878 new->file_offset = file_offset; 879 new->event = event; 880 881 __queue_event(new, s); 882 883 return 0; 884 } 885 886 static void callchain__printf(struct perf_sample *sample) 887 { 888 unsigned int i; 889 890 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr); 891 892 for (i = 0; i < sample->callchain->nr; i++) 893 printf("..... %2d: %016" PRIx64 "\n", 894 i, sample->callchain->ips[i]); 895 } 896 897 static void branch_stack__printf(struct perf_sample *sample) 898 { 899 uint64_t i; 900 901 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr); 902 903 for (i = 0; i < sample->branch_stack->nr; i++) 904 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n", 905 i, sample->branch_stack->entries[i].from, 906 sample->branch_stack->entries[i].to); 907 } 908 909 static void regs_dump__printf(u64 mask, u64 *regs) 910 { 911 unsigned rid, i = 0; 912 913 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 914 u64 val = regs[i++]; 915 916 printf(".... %-5s 0x%" PRIx64 "\n", 917 perf_reg_name(rid), val); 918 } 919 } 920 921 static void regs_user__printf(struct perf_sample *sample, u64 mask) 922 { 923 struct regs_dump *user_regs = &sample->user_regs; 924 925 if (user_regs->regs) { 926 printf("... user regs: mask 0x%" PRIx64 "\n", mask); 927 regs_dump__printf(mask, user_regs->regs); 928 } 929 } 930 931 static void stack_user__printf(struct stack_dump *dump) 932 { 933 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 934 dump->size, dump->offset); 935 } 936 937 static void perf_session__print_tstamp(struct perf_session *session, 938 union perf_event *event, 939 struct perf_sample *sample) 940 { 941 u64 sample_type = perf_evlist__sample_type(session->evlist); 942 943 if (event->header.type != PERF_RECORD_SAMPLE && 944 !perf_evlist__sample_id_all(session->evlist)) { 945 fputs("-1 -1 ", stdout); 946 return; 947 } 948 949 if ((sample_type & PERF_SAMPLE_CPU)) 950 printf("%u ", sample->cpu); 951 952 if (sample_type & PERF_SAMPLE_TIME) 953 printf("%" PRIu64 " ", sample->time); 954 } 955 956 static void dump_event(struct perf_session *session, union perf_event *event, 957 u64 file_offset, struct perf_sample *sample) 958 { 959 if (!dump_trace) 960 return; 961 962 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 963 file_offset, event->header.size, event->header.type); 964 965 trace_event(event); 966 967 if (sample) 968 perf_session__print_tstamp(session, event, sample); 969 970 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 971 event->header.size, perf_event__name(event->header.type)); 972 } 973 974 static void dump_sample(struct perf_evsel *evsel, union perf_event *event, 975 struct perf_sample *sample) 976 { 977 u64 sample_type; 978 979 if (!dump_trace) 980 return; 981 982 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 983 event->header.misc, sample->pid, sample->tid, sample->ip, 984 sample->period, sample->addr); 985 986 sample_type = evsel->attr.sample_type; 987 988 if (sample_type & PERF_SAMPLE_CALLCHAIN) 989 callchain__printf(sample); 990 991 if (sample_type & PERF_SAMPLE_BRANCH_STACK) 992 branch_stack__printf(sample); 993 994 if (sample_type & PERF_SAMPLE_REGS_USER) 995 regs_user__printf(sample, evsel->attr.sample_regs_user); 996 997 if (sample_type & PERF_SAMPLE_STACK_USER) 998 stack_user__printf(&sample->user_stack); 999 } 1000 1001 static struct machine * 1002 perf_session__find_machine_for_cpumode(struct perf_session *session, 1003 union perf_event *event) 1004 { 1005 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 1006 1007 if (perf_guest && 1008 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1009 (cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1010 u32 pid; 1011 1012 if (event->header.type == PERF_RECORD_MMAP) 1013 pid = event->mmap.pid; 1014 else 1015 pid = event->ip.pid; 1016 1017 return perf_session__findnew_machine(session, pid); 1018 } 1019 1020 return perf_session__find_host_machine(session); 1021 } 1022 1023 static int perf_session_deliver_event(struct perf_session *session, 1024 union perf_event *event, 1025 struct perf_sample *sample, 1026 struct perf_tool *tool, 1027 u64 file_offset) 1028 { 1029 struct perf_evsel *evsel; 1030 struct machine *machine; 1031 1032 dump_event(session, event, file_offset, sample); 1033 1034 evsel = perf_evlist__id2evsel(session->evlist, sample->id); 1035 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) { 1036 /* 1037 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here 1038 * because the tools right now may apply filters, discarding 1039 * some of the samples. For consistency, in the future we 1040 * should have something like nr_filtered_samples and remove 1041 * the sample->period from total_sample_period, etc, KISS for 1042 * now tho. 1043 * 1044 * Also testing against NULL allows us to handle files without 1045 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the 1046 * future probably it'll be a good idea to restrict event 1047 * processing via perf_session to files with both set. 1048 */ 1049 hists__inc_nr_events(&evsel->hists, event->header.type); 1050 } 1051 1052 machine = perf_session__find_machine_for_cpumode(session, event); 1053 1054 switch (event->header.type) { 1055 case PERF_RECORD_SAMPLE: 1056 dump_sample(evsel, event, sample); 1057 if (evsel == NULL) { 1058 ++session->hists.stats.nr_unknown_id; 1059 return 0; 1060 } 1061 if (machine == NULL) { 1062 ++session->hists.stats.nr_unprocessable_samples; 1063 return 0; 1064 } 1065 return tool->sample(tool, event, sample, evsel, machine); 1066 case PERF_RECORD_MMAP: 1067 return tool->mmap(tool, event, sample, machine); 1068 case PERF_RECORD_COMM: 1069 return tool->comm(tool, event, sample, machine); 1070 case PERF_RECORD_FORK: 1071 return tool->fork(tool, event, sample, machine); 1072 case PERF_RECORD_EXIT: 1073 return tool->exit(tool, event, sample, machine); 1074 case PERF_RECORD_LOST: 1075 if (tool->lost == perf_event__process_lost) 1076 session->hists.stats.total_lost += event->lost.lost; 1077 return tool->lost(tool, event, sample, machine); 1078 case PERF_RECORD_READ: 1079 return tool->read(tool, event, sample, evsel, machine); 1080 case PERF_RECORD_THROTTLE: 1081 return tool->throttle(tool, event, sample, machine); 1082 case PERF_RECORD_UNTHROTTLE: 1083 return tool->unthrottle(tool, event, sample, machine); 1084 default: 1085 ++session->hists.stats.nr_unknown_events; 1086 return -1; 1087 } 1088 } 1089 1090 static int perf_session__preprocess_sample(struct perf_session *session, 1091 union perf_event *event, struct perf_sample *sample) 1092 { 1093 if (event->header.type != PERF_RECORD_SAMPLE || 1094 !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN)) 1095 return 0; 1096 1097 if (!ip_callchain__valid(sample->callchain, event)) { 1098 pr_debug("call-chain problem with event, skipping it.\n"); 1099 ++session->hists.stats.nr_invalid_chains; 1100 session->hists.stats.total_invalid_chains += sample->period; 1101 return -EINVAL; 1102 } 1103 return 0; 1104 } 1105 1106 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event, 1107 struct perf_tool *tool, u64 file_offset) 1108 { 1109 int err; 1110 1111 dump_event(session, event, file_offset, NULL); 1112 1113 /* These events are processed right away */ 1114 switch (event->header.type) { 1115 case PERF_RECORD_HEADER_ATTR: 1116 err = tool->attr(event, &session->evlist); 1117 if (err == 0) 1118 perf_session__set_id_hdr_size(session); 1119 return err; 1120 case PERF_RECORD_HEADER_EVENT_TYPE: 1121 return tool->event_type(tool, event); 1122 case PERF_RECORD_HEADER_TRACING_DATA: 1123 /* setup for reading amidst mmap */ 1124 lseek(session->fd, file_offset, SEEK_SET); 1125 return tool->tracing_data(event, session); 1126 case PERF_RECORD_HEADER_BUILD_ID: 1127 return tool->build_id(tool, event, session); 1128 case PERF_RECORD_FINISHED_ROUND: 1129 return tool->finished_round(tool, event, session); 1130 default: 1131 return -EINVAL; 1132 } 1133 } 1134 1135 static void event_swap(union perf_event *event, bool sample_id_all) 1136 { 1137 perf_event__swap_op swap; 1138 1139 swap = perf_event__swap_ops[event->header.type]; 1140 if (swap) 1141 swap(event, sample_id_all); 1142 } 1143 1144 static int perf_session__process_event(struct perf_session *session, 1145 union perf_event *event, 1146 struct perf_tool *tool, 1147 u64 file_offset) 1148 { 1149 struct perf_sample sample; 1150 int ret; 1151 1152 if (session->header.needs_swap) 1153 event_swap(event, perf_evlist__sample_id_all(session->evlist)); 1154 1155 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1156 return -EINVAL; 1157 1158 hists__inc_nr_events(&session->hists, event->header.type); 1159 1160 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1161 return perf_session__process_user_event(session, event, tool, file_offset); 1162 1163 /* 1164 * For all kernel events we get the sample data 1165 */ 1166 ret = perf_evlist__parse_sample(session->evlist, event, &sample, 1167 session->header.needs_swap); 1168 if (ret) 1169 return ret; 1170 1171 /* Preprocess sample records - precheck callchains */ 1172 if (perf_session__preprocess_sample(session, event, &sample)) 1173 return 0; 1174 1175 if (tool->ordered_samples) { 1176 ret = perf_session_queue_event(session, event, &sample, 1177 file_offset); 1178 if (ret != -ETIME) 1179 return ret; 1180 } 1181 1182 return perf_session_deliver_event(session, event, &sample, tool, 1183 file_offset); 1184 } 1185 1186 void perf_event_header__bswap(struct perf_event_header *self) 1187 { 1188 self->type = bswap_32(self->type); 1189 self->misc = bswap_16(self->misc); 1190 self->size = bswap_16(self->size); 1191 } 1192 1193 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1194 { 1195 return machine__findnew_thread(&session->host_machine, pid); 1196 } 1197 1198 static struct thread *perf_session__register_idle_thread(struct perf_session *self) 1199 { 1200 struct thread *thread = perf_session__findnew(self, 0); 1201 1202 if (thread == NULL || thread__set_comm(thread, "swapper")) { 1203 pr_err("problem inserting idle task.\n"); 1204 thread = NULL; 1205 } 1206 1207 return thread; 1208 } 1209 1210 static void perf_session__warn_about_errors(const struct perf_session *session, 1211 const struct perf_tool *tool) 1212 { 1213 if (tool->lost == perf_event__process_lost && 1214 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) { 1215 ui__warning("Processed %d events and lost %d chunks!\n\n" 1216 "Check IO/CPU overload!\n\n", 1217 session->hists.stats.nr_events[0], 1218 session->hists.stats.nr_events[PERF_RECORD_LOST]); 1219 } 1220 1221 if (session->hists.stats.nr_unknown_events != 0) { 1222 ui__warning("Found %u unknown events!\n\n" 1223 "Is this an older tool processing a perf.data " 1224 "file generated by a more recent tool?\n\n" 1225 "If that is not the case, consider " 1226 "reporting to linux-kernel@vger.kernel.org.\n\n", 1227 session->hists.stats.nr_unknown_events); 1228 } 1229 1230 if (session->hists.stats.nr_unknown_id != 0) { 1231 ui__warning("%u samples with id not present in the header\n", 1232 session->hists.stats.nr_unknown_id); 1233 } 1234 1235 if (session->hists.stats.nr_invalid_chains != 0) { 1236 ui__warning("Found invalid callchains!\n\n" 1237 "%u out of %u events were discarded for this reason.\n\n" 1238 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1239 session->hists.stats.nr_invalid_chains, 1240 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]); 1241 } 1242 1243 if (session->hists.stats.nr_unprocessable_samples != 0) { 1244 ui__warning("%u unprocessable samples recorded.\n" 1245 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1246 session->hists.stats.nr_unprocessable_samples); 1247 } 1248 } 1249 1250 #define session_done() (*(volatile int *)(&session_done)) 1251 volatile int session_done; 1252 1253 static int __perf_session__process_pipe_events(struct perf_session *self, 1254 struct perf_tool *tool) 1255 { 1256 union perf_event *event; 1257 uint32_t size, cur_size = 0; 1258 void *buf = NULL; 1259 int skip = 0; 1260 u64 head; 1261 int err; 1262 void *p; 1263 1264 perf_tool__fill_defaults(tool); 1265 1266 head = 0; 1267 cur_size = sizeof(union perf_event); 1268 1269 buf = malloc(cur_size); 1270 if (!buf) 1271 return -errno; 1272 more: 1273 event = buf; 1274 err = readn(self->fd, event, sizeof(struct perf_event_header)); 1275 if (err <= 0) { 1276 if (err == 0) 1277 goto done; 1278 1279 pr_err("failed to read event header\n"); 1280 goto out_err; 1281 } 1282 1283 if (self->header.needs_swap) 1284 perf_event_header__bswap(&event->header); 1285 1286 size = event->header.size; 1287 if (size == 0) 1288 size = 8; 1289 1290 if (size > cur_size) { 1291 void *new = realloc(buf, size); 1292 if (!new) { 1293 pr_err("failed to allocate memory to read event\n"); 1294 goto out_err; 1295 } 1296 buf = new; 1297 cur_size = size; 1298 event = buf; 1299 } 1300 p = event; 1301 p += sizeof(struct perf_event_header); 1302 1303 if (size - sizeof(struct perf_event_header)) { 1304 err = readn(self->fd, p, size - sizeof(struct perf_event_header)); 1305 if (err <= 0) { 1306 if (err == 0) { 1307 pr_err("unexpected end of event stream\n"); 1308 goto done; 1309 } 1310 1311 pr_err("failed to read event data\n"); 1312 goto out_err; 1313 } 1314 } 1315 1316 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) { 1317 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1318 head, event->header.size, event->header.type); 1319 err = -EINVAL; 1320 goto out_err; 1321 } 1322 1323 head += size; 1324 1325 if (skip > 0) 1326 head += skip; 1327 1328 if (!session_done()) 1329 goto more; 1330 done: 1331 err = 0; 1332 out_err: 1333 free(buf); 1334 perf_session__warn_about_errors(self, tool); 1335 perf_session_free_sample_buffers(self); 1336 return err; 1337 } 1338 1339 static union perf_event * 1340 fetch_mmaped_event(struct perf_session *session, 1341 u64 head, size_t mmap_size, char *buf) 1342 { 1343 union perf_event *event; 1344 1345 /* 1346 * Ensure we have enough space remaining to read 1347 * the size of the event in the headers. 1348 */ 1349 if (head + sizeof(event->header) > mmap_size) 1350 return NULL; 1351 1352 event = (union perf_event *)(buf + head); 1353 1354 if (session->header.needs_swap) 1355 perf_event_header__bswap(&event->header); 1356 1357 if (head + event->header.size > mmap_size) 1358 return NULL; 1359 1360 return event; 1361 } 1362 1363 int __perf_session__process_events(struct perf_session *session, 1364 u64 data_offset, u64 data_size, 1365 u64 file_size, struct perf_tool *tool) 1366 { 1367 u64 head, page_offset, file_offset, file_pos, progress_next; 1368 int err, mmap_prot, mmap_flags, map_idx = 0; 1369 size_t page_size, mmap_size; 1370 char *buf, *mmaps[8]; 1371 union perf_event *event; 1372 uint32_t size; 1373 1374 perf_tool__fill_defaults(tool); 1375 1376 page_size = sysconf(_SC_PAGESIZE); 1377 1378 page_offset = page_size * (data_offset / page_size); 1379 file_offset = page_offset; 1380 head = data_offset - page_offset; 1381 1382 if (data_offset + data_size < file_size) 1383 file_size = data_offset + data_size; 1384 1385 progress_next = file_size / 16; 1386 1387 mmap_size = session->mmap_window; 1388 if (mmap_size > file_size) 1389 mmap_size = file_size; 1390 1391 memset(mmaps, 0, sizeof(mmaps)); 1392 1393 mmap_prot = PROT_READ; 1394 mmap_flags = MAP_SHARED; 1395 1396 if (session->header.needs_swap) { 1397 mmap_prot |= PROT_WRITE; 1398 mmap_flags = MAP_PRIVATE; 1399 } 1400 remap: 1401 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd, 1402 file_offset); 1403 if (buf == MAP_FAILED) { 1404 pr_err("failed to mmap file\n"); 1405 err = -errno; 1406 goto out_err; 1407 } 1408 mmaps[map_idx] = buf; 1409 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 1410 file_pos = file_offset + head; 1411 1412 more: 1413 event = fetch_mmaped_event(session, head, mmap_size, buf); 1414 if (!event) { 1415 if (mmaps[map_idx]) { 1416 munmap(mmaps[map_idx], mmap_size); 1417 mmaps[map_idx] = NULL; 1418 } 1419 1420 page_offset = page_size * (head / page_size); 1421 file_offset += page_offset; 1422 head -= page_offset; 1423 goto remap; 1424 } 1425 1426 size = event->header.size; 1427 1428 if (size == 0 || 1429 perf_session__process_event(session, event, tool, file_pos) < 0) { 1430 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1431 file_offset + head, event->header.size, 1432 event->header.type); 1433 err = -EINVAL; 1434 goto out_err; 1435 } 1436 1437 head += size; 1438 file_pos += size; 1439 1440 if (file_pos >= progress_next) { 1441 progress_next += file_size / 16; 1442 ui_progress__update(file_pos, file_size, 1443 "Processing events..."); 1444 } 1445 1446 if (file_pos < file_size) 1447 goto more; 1448 1449 err = 0; 1450 /* do the final flush for ordered samples */ 1451 session->ordered_samples.next_flush = ULLONG_MAX; 1452 err = flush_sample_queue(session, tool); 1453 out_err: 1454 perf_session__warn_about_errors(session, tool); 1455 perf_session_free_sample_buffers(session); 1456 return err; 1457 } 1458 1459 int perf_session__process_events(struct perf_session *self, 1460 struct perf_tool *tool) 1461 { 1462 int err; 1463 1464 if (perf_session__register_idle_thread(self) == NULL) 1465 return -ENOMEM; 1466 1467 if (!self->fd_pipe) 1468 err = __perf_session__process_events(self, 1469 self->header.data_offset, 1470 self->header.data_size, 1471 self->size, tool); 1472 else 1473 err = __perf_session__process_pipe_events(self, tool); 1474 1475 return err; 1476 } 1477 1478 bool perf_session__has_traces(struct perf_session *session, const char *msg) 1479 { 1480 if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) { 1481 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 1482 return false; 1483 } 1484 1485 return true; 1486 } 1487 1488 int maps__set_kallsyms_ref_reloc_sym(struct map **maps, 1489 const char *symbol_name, u64 addr) 1490 { 1491 char *bracket; 1492 enum map_type i; 1493 struct ref_reloc_sym *ref; 1494 1495 ref = zalloc(sizeof(struct ref_reloc_sym)); 1496 if (ref == NULL) 1497 return -ENOMEM; 1498 1499 ref->name = strdup(symbol_name); 1500 if (ref->name == NULL) { 1501 free(ref); 1502 return -ENOMEM; 1503 } 1504 1505 bracket = strchr(ref->name, ']'); 1506 if (bracket) 1507 *bracket = '\0'; 1508 1509 ref->addr = addr; 1510 1511 for (i = 0; i < MAP__NR_TYPES; ++i) { 1512 struct kmap *kmap = map__kmap(maps[i]); 1513 kmap->ref_reloc_sym = ref; 1514 } 1515 1516 return 0; 1517 } 1518 1519 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp) 1520 { 1521 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) + 1522 __dsos__fprintf(&self->host_machine.user_dsos, fp) + 1523 machines__fprintf_dsos(&self->machines, fp); 1524 } 1525 1526 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp, 1527 bool with_hits) 1528 { 1529 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits); 1530 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits); 1531 } 1532 1533 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 1534 { 1535 struct perf_evsel *pos; 1536 size_t ret = fprintf(fp, "Aggregated stats:\n"); 1537 1538 ret += hists__fprintf_nr_events(&session->hists, fp); 1539 1540 list_for_each_entry(pos, &session->evlist->entries, node) { 1541 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos)); 1542 ret += hists__fprintf_nr_events(&pos->hists, fp); 1543 } 1544 1545 return ret; 1546 } 1547 1548 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 1549 { 1550 /* 1551 * FIXME: Here we have to actually print all the machines in this 1552 * session, not just the host... 1553 */ 1554 return machine__fprintf(&session->host_machine, fp); 1555 } 1556 1557 void perf_session__remove_thread(struct perf_session *session, 1558 struct thread *th) 1559 { 1560 /* 1561 * FIXME: This one makes no sense, we need to remove the thread from 1562 * the machine it belongs to, perf_session can have many machines, so 1563 * doing it always on ->host_machine is wrong. Fix when auditing all 1564 * the 'perf kvm' code. 1565 */ 1566 machine__remove_thread(&session->host_machine, th); 1567 } 1568 1569 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session, 1570 unsigned int type) 1571 { 1572 struct perf_evsel *pos; 1573 1574 list_for_each_entry(pos, &session->evlist->entries, node) { 1575 if (pos->attr.type == type) 1576 return pos; 1577 } 1578 return NULL; 1579 } 1580 1581 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event, 1582 struct perf_sample *sample, struct machine *machine, 1583 int print_sym, int print_dso, int print_symoffset) 1584 { 1585 struct addr_location al; 1586 struct callchain_cursor_node *node; 1587 1588 if (perf_event__preprocess_sample(event, machine, &al, sample, 1589 NULL) < 0) { 1590 error("problem processing %d event, skipping it.\n", 1591 event->header.type); 1592 return; 1593 } 1594 1595 if (symbol_conf.use_callchain && sample->callchain) { 1596 1597 1598 if (machine__resolve_callchain(machine, evsel, al.thread, 1599 sample, NULL) != 0) { 1600 if (verbose) 1601 error("Failed to resolve callchain. Skipping\n"); 1602 return; 1603 } 1604 callchain_cursor_commit(&callchain_cursor); 1605 1606 while (1) { 1607 node = callchain_cursor_current(&callchain_cursor); 1608 if (!node) 1609 break; 1610 1611 printf("\t%16" PRIx64, node->ip); 1612 if (print_sym) { 1613 printf(" "); 1614 symbol__fprintf_symname(node->sym, stdout); 1615 } 1616 if (print_dso) { 1617 printf(" ("); 1618 map__fprintf_dsoname(node->map, stdout); 1619 printf(")"); 1620 } 1621 printf("\n"); 1622 1623 callchain_cursor_advance(&callchain_cursor); 1624 } 1625 1626 } else { 1627 printf("%16" PRIx64, sample->ip); 1628 if (print_sym) { 1629 printf(" "); 1630 if (print_symoffset) 1631 symbol__fprintf_symname_offs(al.sym, &al, 1632 stdout); 1633 else 1634 symbol__fprintf_symname(al.sym, stdout); 1635 } 1636 1637 if (print_dso) { 1638 printf(" ("); 1639 map__fprintf_dsoname(al.map, stdout); 1640 printf(")"); 1641 } 1642 } 1643 } 1644 1645 int perf_session__cpu_bitmap(struct perf_session *session, 1646 const char *cpu_list, unsigned long *cpu_bitmap) 1647 { 1648 int i; 1649 struct cpu_map *map; 1650 1651 for (i = 0; i < PERF_TYPE_MAX; ++i) { 1652 struct perf_evsel *evsel; 1653 1654 evsel = perf_session__find_first_evtype(session, i); 1655 if (!evsel) 1656 continue; 1657 1658 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) { 1659 pr_err("File does not contain CPU events. " 1660 "Remove -c option to proceed.\n"); 1661 return -1; 1662 } 1663 } 1664 1665 map = cpu_map__new(cpu_list); 1666 if (map == NULL) { 1667 pr_err("Invalid cpu_list\n"); 1668 return -1; 1669 } 1670 1671 for (i = 0; i < map->nr; i++) { 1672 int cpu = map->map[i]; 1673 1674 if (cpu >= MAX_NR_CPUS) { 1675 pr_err("Requested CPU %d too large. " 1676 "Consider raising MAX_NR_CPUS\n", cpu); 1677 return -1; 1678 } 1679 1680 set_bit(cpu, cpu_bitmap); 1681 } 1682 1683 return 0; 1684 } 1685 1686 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 1687 bool full) 1688 { 1689 struct stat st; 1690 int ret; 1691 1692 if (session == NULL || fp == NULL) 1693 return; 1694 1695 ret = fstat(session->fd, &st); 1696 if (ret == -1) 1697 return; 1698 1699 fprintf(fp, "# ========\n"); 1700 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime)); 1701 perf_header__fprintf_info(session, fp, full); 1702 fprintf(fp, "# ========\n#\n"); 1703 } 1704 1705 1706 int __perf_session__set_tracepoints_handlers(struct perf_session *session, 1707 const struct perf_evsel_str_handler *assocs, 1708 size_t nr_assocs) 1709 { 1710 struct perf_evlist *evlist = session->evlist; 1711 struct event_format *format; 1712 struct perf_evsel *evsel; 1713 char *tracepoint, *name; 1714 size_t i; 1715 int err; 1716 1717 for (i = 0; i < nr_assocs; i++) { 1718 err = -ENOMEM; 1719 tracepoint = strdup(assocs[i].name); 1720 if (tracepoint == NULL) 1721 goto out; 1722 1723 err = -ENOENT; 1724 name = strchr(tracepoint, ':'); 1725 if (name == NULL) 1726 goto out_free; 1727 1728 *name++ = '\0'; 1729 format = pevent_find_event_by_name(session->pevent, 1730 tracepoint, name); 1731 if (format == NULL) { 1732 /* 1733 * Adding a handler for an event not in the session, 1734 * just ignore it. 1735 */ 1736 goto next; 1737 } 1738 1739 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id); 1740 if (evsel == NULL) 1741 goto next; 1742 1743 err = -EEXIST; 1744 if (evsel->handler.func != NULL) 1745 goto out_free; 1746 evsel->handler.func = assocs[i].handler; 1747 next: 1748 free(tracepoint); 1749 } 1750 1751 err = 0; 1752 out: 1753 return err; 1754 1755 out_free: 1756 free(tracepoint); 1757 goto out; 1758 } 1759