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