1 #include <dirent.h> 2 #include <errno.h> 3 #include <inttypes.h> 4 #include <regex.h> 5 #include "callchain.h" 6 #include "debug.h" 7 #include "event.h" 8 #include "evsel.h" 9 #include "hist.h" 10 #include "machine.h" 11 #include "map.h" 12 #include "sort.h" 13 #include "strlist.h" 14 #include "thread.h" 15 #include "vdso.h" 16 #include <stdbool.h> 17 #include <sys/types.h> 18 #include <sys/stat.h> 19 #include <unistd.h> 20 #include "unwind.h" 21 #include "linux/hash.h" 22 #include "asm/bug.h" 23 24 #include "sane_ctype.h" 25 #include <symbol/kallsyms.h> 26 27 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock); 28 29 static void dsos__init(struct dsos *dsos) 30 { 31 INIT_LIST_HEAD(&dsos->head); 32 dsos->root = RB_ROOT; 33 pthread_rwlock_init(&dsos->lock, NULL); 34 } 35 36 int machine__init(struct machine *machine, const char *root_dir, pid_t pid) 37 { 38 memset(machine, 0, sizeof(*machine)); 39 map_groups__init(&machine->kmaps, machine); 40 RB_CLEAR_NODE(&machine->rb_node); 41 dsos__init(&machine->dsos); 42 43 machine->threads = RB_ROOT; 44 pthread_rwlock_init(&machine->threads_lock, NULL); 45 machine->nr_threads = 0; 46 INIT_LIST_HEAD(&machine->dead_threads); 47 machine->last_match = NULL; 48 49 machine->vdso_info = NULL; 50 machine->env = NULL; 51 52 machine->pid = pid; 53 54 machine->id_hdr_size = 0; 55 machine->kptr_restrict_warned = false; 56 machine->comm_exec = false; 57 machine->kernel_start = 0; 58 59 memset(machine->vmlinux_maps, 0, sizeof(machine->vmlinux_maps)); 60 61 machine->root_dir = strdup(root_dir); 62 if (machine->root_dir == NULL) 63 return -ENOMEM; 64 65 if (pid != HOST_KERNEL_ID) { 66 struct thread *thread = machine__findnew_thread(machine, -1, 67 pid); 68 char comm[64]; 69 70 if (thread == NULL) 71 return -ENOMEM; 72 73 snprintf(comm, sizeof(comm), "[guest/%d]", pid); 74 thread__set_comm(thread, comm, 0); 75 thread__put(thread); 76 } 77 78 machine->current_tid = NULL; 79 80 return 0; 81 } 82 83 struct machine *machine__new_host(void) 84 { 85 struct machine *machine = malloc(sizeof(*machine)); 86 87 if (machine != NULL) { 88 machine__init(machine, "", HOST_KERNEL_ID); 89 90 if (machine__create_kernel_maps(machine) < 0) 91 goto out_delete; 92 } 93 94 return machine; 95 out_delete: 96 free(machine); 97 return NULL; 98 } 99 100 struct machine *machine__new_kallsyms(void) 101 { 102 struct machine *machine = machine__new_host(); 103 /* 104 * FIXME: 105 * 1) MAP__FUNCTION will go away when we stop loading separate maps for 106 * functions and data objects. 107 * 2) We should switch to machine__load_kallsyms(), i.e. not explicitely 108 * ask for not using the kcore parsing code, once this one is fixed 109 * to create a map per module. 110 */ 111 if (machine && __machine__load_kallsyms(machine, "/proc/kallsyms", MAP__FUNCTION, true) <= 0) { 112 machine__delete(machine); 113 machine = NULL; 114 } 115 116 return machine; 117 } 118 119 static void dsos__purge(struct dsos *dsos) 120 { 121 struct dso *pos, *n; 122 123 pthread_rwlock_wrlock(&dsos->lock); 124 125 list_for_each_entry_safe(pos, n, &dsos->head, node) { 126 RB_CLEAR_NODE(&pos->rb_node); 127 pos->root = NULL; 128 list_del_init(&pos->node); 129 dso__put(pos); 130 } 131 132 pthread_rwlock_unlock(&dsos->lock); 133 } 134 135 static void dsos__exit(struct dsos *dsos) 136 { 137 dsos__purge(dsos); 138 pthread_rwlock_destroy(&dsos->lock); 139 } 140 141 void machine__delete_threads(struct machine *machine) 142 { 143 struct rb_node *nd; 144 145 pthread_rwlock_wrlock(&machine->threads_lock); 146 nd = rb_first(&machine->threads); 147 while (nd) { 148 struct thread *t = rb_entry(nd, struct thread, rb_node); 149 150 nd = rb_next(nd); 151 __machine__remove_thread(machine, t, false); 152 } 153 pthread_rwlock_unlock(&machine->threads_lock); 154 } 155 156 void machine__exit(struct machine *machine) 157 { 158 machine__destroy_kernel_maps(machine); 159 map_groups__exit(&machine->kmaps); 160 dsos__exit(&machine->dsos); 161 machine__exit_vdso(machine); 162 zfree(&machine->root_dir); 163 zfree(&machine->current_tid); 164 pthread_rwlock_destroy(&machine->threads_lock); 165 } 166 167 void machine__delete(struct machine *machine) 168 { 169 if (machine) { 170 machine__exit(machine); 171 free(machine); 172 } 173 } 174 175 void machines__init(struct machines *machines) 176 { 177 machine__init(&machines->host, "", HOST_KERNEL_ID); 178 machines->guests = RB_ROOT; 179 } 180 181 void machines__exit(struct machines *machines) 182 { 183 machine__exit(&machines->host); 184 /* XXX exit guest */ 185 } 186 187 struct machine *machines__add(struct machines *machines, pid_t pid, 188 const char *root_dir) 189 { 190 struct rb_node **p = &machines->guests.rb_node; 191 struct rb_node *parent = NULL; 192 struct machine *pos, *machine = malloc(sizeof(*machine)); 193 194 if (machine == NULL) 195 return NULL; 196 197 if (machine__init(machine, root_dir, pid) != 0) { 198 free(machine); 199 return NULL; 200 } 201 202 while (*p != NULL) { 203 parent = *p; 204 pos = rb_entry(parent, struct machine, rb_node); 205 if (pid < pos->pid) 206 p = &(*p)->rb_left; 207 else 208 p = &(*p)->rb_right; 209 } 210 211 rb_link_node(&machine->rb_node, parent, p); 212 rb_insert_color(&machine->rb_node, &machines->guests); 213 214 return machine; 215 } 216 217 void machines__set_comm_exec(struct machines *machines, bool comm_exec) 218 { 219 struct rb_node *nd; 220 221 machines->host.comm_exec = comm_exec; 222 223 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 224 struct machine *machine = rb_entry(nd, struct machine, rb_node); 225 226 machine->comm_exec = comm_exec; 227 } 228 } 229 230 struct machine *machines__find(struct machines *machines, pid_t pid) 231 { 232 struct rb_node **p = &machines->guests.rb_node; 233 struct rb_node *parent = NULL; 234 struct machine *machine; 235 struct machine *default_machine = NULL; 236 237 if (pid == HOST_KERNEL_ID) 238 return &machines->host; 239 240 while (*p != NULL) { 241 parent = *p; 242 machine = rb_entry(parent, struct machine, rb_node); 243 if (pid < machine->pid) 244 p = &(*p)->rb_left; 245 else if (pid > machine->pid) 246 p = &(*p)->rb_right; 247 else 248 return machine; 249 if (!machine->pid) 250 default_machine = machine; 251 } 252 253 return default_machine; 254 } 255 256 struct machine *machines__findnew(struct machines *machines, pid_t pid) 257 { 258 char path[PATH_MAX]; 259 const char *root_dir = ""; 260 struct machine *machine = machines__find(machines, pid); 261 262 if (machine && (machine->pid == pid)) 263 goto out; 264 265 if ((pid != HOST_KERNEL_ID) && 266 (pid != DEFAULT_GUEST_KERNEL_ID) && 267 (symbol_conf.guestmount)) { 268 sprintf(path, "%s/%d", symbol_conf.guestmount, pid); 269 if (access(path, R_OK)) { 270 static struct strlist *seen; 271 272 if (!seen) 273 seen = strlist__new(NULL, NULL); 274 275 if (!strlist__has_entry(seen, path)) { 276 pr_err("Can't access file %s\n", path); 277 strlist__add(seen, path); 278 } 279 machine = NULL; 280 goto out; 281 } 282 root_dir = path; 283 } 284 285 machine = machines__add(machines, pid, root_dir); 286 out: 287 return machine; 288 } 289 290 void machines__process_guests(struct machines *machines, 291 machine__process_t process, void *data) 292 { 293 struct rb_node *nd; 294 295 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 296 struct machine *pos = rb_entry(nd, struct machine, rb_node); 297 process(pos, data); 298 } 299 } 300 301 char *machine__mmap_name(struct machine *machine, char *bf, size_t size) 302 { 303 if (machine__is_host(machine)) 304 snprintf(bf, size, "[%s]", "kernel.kallsyms"); 305 else if (machine__is_default_guest(machine)) 306 snprintf(bf, size, "[%s]", "guest.kernel.kallsyms"); 307 else { 308 snprintf(bf, size, "[%s.%d]", "guest.kernel.kallsyms", 309 machine->pid); 310 } 311 312 return bf; 313 } 314 315 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size) 316 { 317 struct rb_node *node; 318 struct machine *machine; 319 320 machines->host.id_hdr_size = id_hdr_size; 321 322 for (node = rb_first(&machines->guests); node; node = rb_next(node)) { 323 machine = rb_entry(node, struct machine, rb_node); 324 machine->id_hdr_size = id_hdr_size; 325 } 326 327 return; 328 } 329 330 static void machine__update_thread_pid(struct machine *machine, 331 struct thread *th, pid_t pid) 332 { 333 struct thread *leader; 334 335 if (pid == th->pid_ || pid == -1 || th->pid_ != -1) 336 return; 337 338 th->pid_ = pid; 339 340 if (th->pid_ == th->tid) 341 return; 342 343 leader = __machine__findnew_thread(machine, th->pid_, th->pid_); 344 if (!leader) 345 goto out_err; 346 347 if (!leader->mg) 348 leader->mg = map_groups__new(machine); 349 350 if (!leader->mg) 351 goto out_err; 352 353 if (th->mg == leader->mg) 354 return; 355 356 if (th->mg) { 357 /* 358 * Maps are created from MMAP events which provide the pid and 359 * tid. Consequently there never should be any maps on a thread 360 * with an unknown pid. Just print an error if there are. 361 */ 362 if (!map_groups__empty(th->mg)) 363 pr_err("Discarding thread maps for %d:%d\n", 364 th->pid_, th->tid); 365 map_groups__put(th->mg); 366 } 367 368 th->mg = map_groups__get(leader->mg); 369 out_put: 370 thread__put(leader); 371 return; 372 out_err: 373 pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid); 374 goto out_put; 375 } 376 377 /* 378 * Caller must eventually drop thread->refcnt returned with a successful 379 * lookup/new thread inserted. 380 */ 381 static struct thread *____machine__findnew_thread(struct machine *machine, 382 pid_t pid, pid_t tid, 383 bool create) 384 { 385 struct rb_node **p = &machine->threads.rb_node; 386 struct rb_node *parent = NULL; 387 struct thread *th; 388 389 /* 390 * Front-end cache - TID lookups come in blocks, 391 * so most of the time we dont have to look up 392 * the full rbtree: 393 */ 394 th = machine->last_match; 395 if (th != NULL) { 396 if (th->tid == tid) { 397 machine__update_thread_pid(machine, th, pid); 398 return thread__get(th); 399 } 400 401 machine->last_match = NULL; 402 } 403 404 while (*p != NULL) { 405 parent = *p; 406 th = rb_entry(parent, struct thread, rb_node); 407 408 if (th->tid == tid) { 409 machine->last_match = th; 410 machine__update_thread_pid(machine, th, pid); 411 return thread__get(th); 412 } 413 414 if (tid < th->tid) 415 p = &(*p)->rb_left; 416 else 417 p = &(*p)->rb_right; 418 } 419 420 if (!create) 421 return NULL; 422 423 th = thread__new(pid, tid); 424 if (th != NULL) { 425 rb_link_node(&th->rb_node, parent, p); 426 rb_insert_color(&th->rb_node, &machine->threads); 427 428 /* 429 * We have to initialize map_groups separately 430 * after rb tree is updated. 431 * 432 * The reason is that we call machine__findnew_thread 433 * within thread__init_map_groups to find the thread 434 * leader and that would screwed the rb tree. 435 */ 436 if (thread__init_map_groups(th, machine)) { 437 rb_erase_init(&th->rb_node, &machine->threads); 438 RB_CLEAR_NODE(&th->rb_node); 439 thread__put(th); 440 return NULL; 441 } 442 /* 443 * It is now in the rbtree, get a ref 444 */ 445 thread__get(th); 446 machine->last_match = th; 447 ++machine->nr_threads; 448 } 449 450 return th; 451 } 452 453 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid) 454 { 455 return ____machine__findnew_thread(machine, pid, tid, true); 456 } 457 458 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, 459 pid_t tid) 460 { 461 struct thread *th; 462 463 pthread_rwlock_wrlock(&machine->threads_lock); 464 th = __machine__findnew_thread(machine, pid, tid); 465 pthread_rwlock_unlock(&machine->threads_lock); 466 return th; 467 } 468 469 struct thread *machine__find_thread(struct machine *machine, pid_t pid, 470 pid_t tid) 471 { 472 struct thread *th; 473 pthread_rwlock_rdlock(&machine->threads_lock); 474 th = ____machine__findnew_thread(machine, pid, tid, false); 475 pthread_rwlock_unlock(&machine->threads_lock); 476 return th; 477 } 478 479 struct comm *machine__thread_exec_comm(struct machine *machine, 480 struct thread *thread) 481 { 482 if (machine->comm_exec) 483 return thread__exec_comm(thread); 484 else 485 return thread__comm(thread); 486 } 487 488 int machine__process_comm_event(struct machine *machine, union perf_event *event, 489 struct perf_sample *sample) 490 { 491 struct thread *thread = machine__findnew_thread(machine, 492 event->comm.pid, 493 event->comm.tid); 494 bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC; 495 int err = 0; 496 497 if (exec) 498 machine->comm_exec = true; 499 500 if (dump_trace) 501 perf_event__fprintf_comm(event, stdout); 502 503 if (thread == NULL || 504 __thread__set_comm(thread, event->comm.comm, sample->time, exec)) { 505 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n"); 506 err = -1; 507 } 508 509 thread__put(thread); 510 511 return err; 512 } 513 514 int machine__process_namespaces_event(struct machine *machine __maybe_unused, 515 union perf_event *event, 516 struct perf_sample *sample __maybe_unused) 517 { 518 struct thread *thread = machine__findnew_thread(machine, 519 event->namespaces.pid, 520 event->namespaces.tid); 521 int err = 0; 522 523 WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES, 524 "\nWARNING: kernel seems to support more namespaces than perf" 525 " tool.\nTry updating the perf tool..\n\n"); 526 527 WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES, 528 "\nWARNING: perf tool seems to support more namespaces than" 529 " the kernel.\nTry updating the kernel..\n\n"); 530 531 if (dump_trace) 532 perf_event__fprintf_namespaces(event, stdout); 533 534 if (thread == NULL || 535 thread__set_namespaces(thread, sample->time, &event->namespaces)) { 536 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n"); 537 err = -1; 538 } 539 540 thread__put(thread); 541 542 return err; 543 } 544 545 int machine__process_lost_event(struct machine *machine __maybe_unused, 546 union perf_event *event, struct perf_sample *sample __maybe_unused) 547 { 548 dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n", 549 event->lost.id, event->lost.lost); 550 return 0; 551 } 552 553 int machine__process_lost_samples_event(struct machine *machine __maybe_unused, 554 union perf_event *event, struct perf_sample *sample) 555 { 556 dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n", 557 sample->id, event->lost_samples.lost); 558 return 0; 559 } 560 561 static struct dso *machine__findnew_module_dso(struct machine *machine, 562 struct kmod_path *m, 563 const char *filename) 564 { 565 struct dso *dso; 566 567 pthread_rwlock_wrlock(&machine->dsos.lock); 568 569 dso = __dsos__find(&machine->dsos, m->name, true); 570 if (!dso) { 571 dso = __dsos__addnew(&machine->dsos, m->name); 572 if (dso == NULL) 573 goto out_unlock; 574 575 dso__set_module_info(dso, m, machine); 576 dso__set_long_name(dso, strdup(filename), true); 577 } 578 579 dso__get(dso); 580 out_unlock: 581 pthread_rwlock_unlock(&machine->dsos.lock); 582 return dso; 583 } 584 585 int machine__process_aux_event(struct machine *machine __maybe_unused, 586 union perf_event *event) 587 { 588 if (dump_trace) 589 perf_event__fprintf_aux(event, stdout); 590 return 0; 591 } 592 593 int machine__process_itrace_start_event(struct machine *machine __maybe_unused, 594 union perf_event *event) 595 { 596 if (dump_trace) 597 perf_event__fprintf_itrace_start(event, stdout); 598 return 0; 599 } 600 601 int machine__process_switch_event(struct machine *machine __maybe_unused, 602 union perf_event *event) 603 { 604 if (dump_trace) 605 perf_event__fprintf_switch(event, stdout); 606 return 0; 607 } 608 609 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename) 610 { 611 const char *dup_filename; 612 613 if (!filename || !dso || !dso->long_name) 614 return; 615 if (dso->long_name[0] != '[') 616 return; 617 if (!strchr(filename, '/')) 618 return; 619 620 dup_filename = strdup(filename); 621 if (!dup_filename) 622 return; 623 624 dso__set_long_name(dso, dup_filename, true); 625 } 626 627 struct map *machine__findnew_module_map(struct machine *machine, u64 start, 628 const char *filename) 629 { 630 struct map *map = NULL; 631 struct dso *dso = NULL; 632 struct kmod_path m; 633 634 if (kmod_path__parse_name(&m, filename)) 635 return NULL; 636 637 map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION, 638 m.name); 639 if (map) { 640 /* 641 * If the map's dso is an offline module, give dso__load() 642 * a chance to find the file path of that module by fixing 643 * long_name. 644 */ 645 dso__adjust_kmod_long_name(map->dso, filename); 646 goto out; 647 } 648 649 dso = machine__findnew_module_dso(machine, &m, filename); 650 if (dso == NULL) 651 goto out; 652 653 map = map__new2(start, dso, MAP__FUNCTION); 654 if (map == NULL) 655 goto out; 656 657 map_groups__insert(&machine->kmaps, map); 658 659 /* Put the map here because map_groups__insert alread got it */ 660 map__put(map); 661 out: 662 /* put the dso here, corresponding to machine__findnew_module_dso */ 663 dso__put(dso); 664 free(m.name); 665 return map; 666 } 667 668 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp) 669 { 670 struct rb_node *nd; 671 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp); 672 673 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 674 struct machine *pos = rb_entry(nd, struct machine, rb_node); 675 ret += __dsos__fprintf(&pos->dsos.head, fp); 676 } 677 678 return ret; 679 } 680 681 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp, 682 bool (skip)(struct dso *dso, int parm), int parm) 683 { 684 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm); 685 } 686 687 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp, 688 bool (skip)(struct dso *dso, int parm), int parm) 689 { 690 struct rb_node *nd; 691 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm); 692 693 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 694 struct machine *pos = rb_entry(nd, struct machine, rb_node); 695 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm); 696 } 697 return ret; 698 } 699 700 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp) 701 { 702 int i; 703 size_t printed = 0; 704 struct dso *kdso = machine__kernel_map(machine)->dso; 705 706 if (kdso->has_build_id) { 707 char filename[PATH_MAX]; 708 if (dso__build_id_filename(kdso, filename, sizeof(filename))) 709 printed += fprintf(fp, "[0] %s\n", filename); 710 } 711 712 for (i = 0; i < vmlinux_path__nr_entries; ++i) 713 printed += fprintf(fp, "[%d] %s\n", 714 i + kdso->has_build_id, vmlinux_path[i]); 715 716 return printed; 717 } 718 719 size_t machine__fprintf(struct machine *machine, FILE *fp) 720 { 721 size_t ret; 722 struct rb_node *nd; 723 724 pthread_rwlock_rdlock(&machine->threads_lock); 725 726 ret = fprintf(fp, "Threads: %u\n", machine->nr_threads); 727 728 for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) { 729 struct thread *pos = rb_entry(nd, struct thread, rb_node); 730 731 ret += thread__fprintf(pos, fp); 732 } 733 734 pthread_rwlock_unlock(&machine->threads_lock); 735 736 return ret; 737 } 738 739 static struct dso *machine__get_kernel(struct machine *machine) 740 { 741 const char *vmlinux_name = NULL; 742 struct dso *kernel; 743 744 if (machine__is_host(machine)) { 745 vmlinux_name = symbol_conf.vmlinux_name; 746 if (!vmlinux_name) 747 vmlinux_name = DSO__NAME_KALLSYMS; 748 749 kernel = machine__findnew_kernel(machine, vmlinux_name, 750 "[kernel]", DSO_TYPE_KERNEL); 751 } else { 752 char bf[PATH_MAX]; 753 754 if (machine__is_default_guest(machine)) 755 vmlinux_name = symbol_conf.default_guest_vmlinux_name; 756 if (!vmlinux_name) 757 vmlinux_name = machine__mmap_name(machine, bf, 758 sizeof(bf)); 759 760 kernel = machine__findnew_kernel(machine, vmlinux_name, 761 "[guest.kernel]", 762 DSO_TYPE_GUEST_KERNEL); 763 } 764 765 if (kernel != NULL && (!kernel->has_build_id)) 766 dso__read_running_kernel_build_id(kernel, machine); 767 768 return kernel; 769 } 770 771 struct process_args { 772 u64 start; 773 }; 774 775 static void machine__get_kallsyms_filename(struct machine *machine, char *buf, 776 size_t bufsz) 777 { 778 if (machine__is_default_guest(machine)) 779 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms); 780 else 781 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir); 782 } 783 784 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL}; 785 786 /* Figure out the start address of kernel map from /proc/kallsyms. 787 * Returns the name of the start symbol in *symbol_name. Pass in NULL as 788 * symbol_name if it's not that important. 789 */ 790 static int machine__get_running_kernel_start(struct machine *machine, 791 const char **symbol_name, u64 *start) 792 { 793 char filename[PATH_MAX]; 794 int i, err = -1; 795 const char *name; 796 u64 addr = 0; 797 798 machine__get_kallsyms_filename(machine, filename, PATH_MAX); 799 800 if (symbol__restricted_filename(filename, "/proc/kallsyms")) 801 return 0; 802 803 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) { 804 err = kallsyms__get_function_start(filename, name, &addr); 805 if (!err) 806 break; 807 } 808 809 if (err) 810 return -1; 811 812 if (symbol_name) 813 *symbol_name = name; 814 815 *start = addr; 816 return 0; 817 } 818 819 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel) 820 { 821 int type; 822 u64 start = 0; 823 824 if (machine__get_running_kernel_start(machine, NULL, &start)) 825 return -1; 826 827 /* In case of renewal the kernel map, destroy previous one */ 828 machine__destroy_kernel_maps(machine); 829 830 for (type = 0; type < MAP__NR_TYPES; ++type) { 831 struct kmap *kmap; 832 struct map *map; 833 834 machine->vmlinux_maps[type] = map__new2(start, kernel, type); 835 if (machine->vmlinux_maps[type] == NULL) 836 return -1; 837 838 machine->vmlinux_maps[type]->map_ip = 839 machine->vmlinux_maps[type]->unmap_ip = 840 identity__map_ip; 841 map = __machine__kernel_map(machine, type); 842 kmap = map__kmap(map); 843 if (!kmap) 844 return -1; 845 846 kmap->kmaps = &machine->kmaps; 847 map_groups__insert(&machine->kmaps, map); 848 } 849 850 return 0; 851 } 852 853 void machine__destroy_kernel_maps(struct machine *machine) 854 { 855 int type; 856 857 for (type = 0; type < MAP__NR_TYPES; ++type) { 858 struct kmap *kmap; 859 struct map *map = __machine__kernel_map(machine, type); 860 861 if (map == NULL) 862 continue; 863 864 kmap = map__kmap(map); 865 map_groups__remove(&machine->kmaps, map); 866 if (kmap && kmap->ref_reloc_sym) { 867 /* 868 * ref_reloc_sym is shared among all maps, so free just 869 * on one of them. 870 */ 871 if (type == MAP__FUNCTION) { 872 zfree((char **)&kmap->ref_reloc_sym->name); 873 zfree(&kmap->ref_reloc_sym); 874 } else 875 kmap->ref_reloc_sym = NULL; 876 } 877 878 map__put(machine->vmlinux_maps[type]); 879 machine->vmlinux_maps[type] = NULL; 880 } 881 } 882 883 int machines__create_guest_kernel_maps(struct machines *machines) 884 { 885 int ret = 0; 886 struct dirent **namelist = NULL; 887 int i, items = 0; 888 char path[PATH_MAX]; 889 pid_t pid; 890 char *endp; 891 892 if (symbol_conf.default_guest_vmlinux_name || 893 symbol_conf.default_guest_modules || 894 symbol_conf.default_guest_kallsyms) { 895 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID); 896 } 897 898 if (symbol_conf.guestmount) { 899 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL); 900 if (items <= 0) 901 return -ENOENT; 902 for (i = 0; i < items; i++) { 903 if (!isdigit(namelist[i]->d_name[0])) { 904 /* Filter out . and .. */ 905 continue; 906 } 907 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10); 908 if ((*endp != '\0') || 909 (endp == namelist[i]->d_name) || 910 (errno == ERANGE)) { 911 pr_debug("invalid directory (%s). Skipping.\n", 912 namelist[i]->d_name); 913 continue; 914 } 915 sprintf(path, "%s/%s/proc/kallsyms", 916 symbol_conf.guestmount, 917 namelist[i]->d_name); 918 ret = access(path, R_OK); 919 if (ret) { 920 pr_debug("Can't access file %s\n", path); 921 goto failure; 922 } 923 machines__create_kernel_maps(machines, pid); 924 } 925 failure: 926 free(namelist); 927 } 928 929 return ret; 930 } 931 932 void machines__destroy_kernel_maps(struct machines *machines) 933 { 934 struct rb_node *next = rb_first(&machines->guests); 935 936 machine__destroy_kernel_maps(&machines->host); 937 938 while (next) { 939 struct machine *pos = rb_entry(next, struct machine, rb_node); 940 941 next = rb_next(&pos->rb_node); 942 rb_erase(&pos->rb_node, &machines->guests); 943 machine__delete(pos); 944 } 945 } 946 947 int machines__create_kernel_maps(struct machines *machines, pid_t pid) 948 { 949 struct machine *machine = machines__findnew(machines, pid); 950 951 if (machine == NULL) 952 return -1; 953 954 return machine__create_kernel_maps(machine); 955 } 956 957 int __machine__load_kallsyms(struct machine *machine, const char *filename, 958 enum map_type type, bool no_kcore) 959 { 960 struct map *map = machine__kernel_map(machine); 961 int ret = __dso__load_kallsyms(map->dso, filename, map, no_kcore); 962 963 if (ret > 0) { 964 dso__set_loaded(map->dso, type); 965 /* 966 * Since /proc/kallsyms will have multiple sessions for the 967 * kernel, with modules between them, fixup the end of all 968 * sections. 969 */ 970 __map_groups__fixup_end(&machine->kmaps, type); 971 } 972 973 return ret; 974 } 975 976 int machine__load_kallsyms(struct machine *machine, const char *filename, 977 enum map_type type) 978 { 979 return __machine__load_kallsyms(machine, filename, type, false); 980 } 981 982 int machine__load_vmlinux_path(struct machine *machine, enum map_type type) 983 { 984 struct map *map = machine__kernel_map(machine); 985 int ret = dso__load_vmlinux_path(map->dso, map); 986 987 if (ret > 0) 988 dso__set_loaded(map->dso, type); 989 990 return ret; 991 } 992 993 static void map_groups__fixup_end(struct map_groups *mg) 994 { 995 int i; 996 for (i = 0; i < MAP__NR_TYPES; ++i) 997 __map_groups__fixup_end(mg, i); 998 } 999 1000 static char *get_kernel_version(const char *root_dir) 1001 { 1002 char version[PATH_MAX]; 1003 FILE *file; 1004 char *name, *tmp; 1005 const char *prefix = "Linux version "; 1006 1007 sprintf(version, "%s/proc/version", root_dir); 1008 file = fopen(version, "r"); 1009 if (!file) 1010 return NULL; 1011 1012 version[0] = '\0'; 1013 tmp = fgets(version, sizeof(version), file); 1014 fclose(file); 1015 1016 name = strstr(version, prefix); 1017 if (!name) 1018 return NULL; 1019 name += strlen(prefix); 1020 tmp = strchr(name, ' '); 1021 if (tmp) 1022 *tmp = '\0'; 1023 1024 return strdup(name); 1025 } 1026 1027 static bool is_kmod_dso(struct dso *dso) 1028 { 1029 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE || 1030 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE; 1031 } 1032 1033 static int map_groups__set_module_path(struct map_groups *mg, const char *path, 1034 struct kmod_path *m) 1035 { 1036 struct map *map; 1037 char *long_name; 1038 1039 map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name); 1040 if (map == NULL) 1041 return 0; 1042 1043 long_name = strdup(path); 1044 if (long_name == NULL) 1045 return -ENOMEM; 1046 1047 dso__set_long_name(map->dso, long_name, true); 1048 dso__kernel_module_get_build_id(map->dso, ""); 1049 1050 /* 1051 * Full name could reveal us kmod compression, so 1052 * we need to update the symtab_type if needed. 1053 */ 1054 if (m->comp && is_kmod_dso(map->dso)) 1055 map->dso->symtab_type++; 1056 1057 return 0; 1058 } 1059 1060 static int map_groups__set_modules_path_dir(struct map_groups *mg, 1061 const char *dir_name, int depth) 1062 { 1063 struct dirent *dent; 1064 DIR *dir = opendir(dir_name); 1065 int ret = 0; 1066 1067 if (!dir) { 1068 pr_debug("%s: cannot open %s dir\n", __func__, dir_name); 1069 return -1; 1070 } 1071 1072 while ((dent = readdir(dir)) != NULL) { 1073 char path[PATH_MAX]; 1074 struct stat st; 1075 1076 /*sshfs might return bad dent->d_type, so we have to stat*/ 1077 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name); 1078 if (stat(path, &st)) 1079 continue; 1080 1081 if (S_ISDIR(st.st_mode)) { 1082 if (!strcmp(dent->d_name, ".") || 1083 !strcmp(dent->d_name, "..")) 1084 continue; 1085 1086 /* Do not follow top-level source and build symlinks */ 1087 if (depth == 0) { 1088 if (!strcmp(dent->d_name, "source") || 1089 !strcmp(dent->d_name, "build")) 1090 continue; 1091 } 1092 1093 ret = map_groups__set_modules_path_dir(mg, path, 1094 depth + 1); 1095 if (ret < 0) 1096 goto out; 1097 } else { 1098 struct kmod_path m; 1099 1100 ret = kmod_path__parse_name(&m, dent->d_name); 1101 if (ret) 1102 goto out; 1103 1104 if (m.kmod) 1105 ret = map_groups__set_module_path(mg, path, &m); 1106 1107 free(m.name); 1108 1109 if (ret) 1110 goto out; 1111 } 1112 } 1113 1114 out: 1115 closedir(dir); 1116 return ret; 1117 } 1118 1119 static int machine__set_modules_path(struct machine *machine) 1120 { 1121 char *version; 1122 char modules_path[PATH_MAX]; 1123 1124 version = get_kernel_version(machine->root_dir); 1125 if (!version) 1126 return -1; 1127 1128 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s", 1129 machine->root_dir, version); 1130 free(version); 1131 1132 return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0); 1133 } 1134 int __weak arch__fix_module_text_start(u64 *start __maybe_unused, 1135 const char *name __maybe_unused) 1136 { 1137 return 0; 1138 } 1139 1140 static int machine__create_module(void *arg, const char *name, u64 start) 1141 { 1142 struct machine *machine = arg; 1143 struct map *map; 1144 1145 if (arch__fix_module_text_start(&start, name) < 0) 1146 return -1; 1147 1148 map = machine__findnew_module_map(machine, start, name); 1149 if (map == NULL) 1150 return -1; 1151 1152 dso__kernel_module_get_build_id(map->dso, machine->root_dir); 1153 1154 return 0; 1155 } 1156 1157 static int machine__create_modules(struct machine *machine) 1158 { 1159 const char *modules; 1160 char path[PATH_MAX]; 1161 1162 if (machine__is_default_guest(machine)) { 1163 modules = symbol_conf.default_guest_modules; 1164 } else { 1165 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir); 1166 modules = path; 1167 } 1168 1169 if (symbol__restricted_filename(modules, "/proc/modules")) 1170 return -1; 1171 1172 if (modules__parse(modules, machine, machine__create_module)) 1173 return -1; 1174 1175 if (!machine__set_modules_path(machine)) 1176 return 0; 1177 1178 pr_debug("Problems setting modules path maps, continuing anyway...\n"); 1179 1180 return 0; 1181 } 1182 1183 int machine__create_kernel_maps(struct machine *machine) 1184 { 1185 struct dso *kernel = machine__get_kernel(machine); 1186 const char *name = NULL; 1187 u64 addr = 0; 1188 int ret; 1189 1190 if (kernel == NULL) 1191 return -1; 1192 1193 ret = __machine__create_kernel_maps(machine, kernel); 1194 dso__put(kernel); 1195 if (ret < 0) 1196 return -1; 1197 1198 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) { 1199 if (machine__is_host(machine)) 1200 pr_debug("Problems creating module maps, " 1201 "continuing anyway...\n"); 1202 else 1203 pr_debug("Problems creating module maps for guest %d, " 1204 "continuing anyway...\n", machine->pid); 1205 } 1206 1207 /* 1208 * Now that we have all the maps created, just set the ->end of them: 1209 */ 1210 map_groups__fixup_end(&machine->kmaps); 1211 1212 if (!machine__get_running_kernel_start(machine, &name, &addr)) { 1213 if (name && 1214 maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name, addr)) { 1215 machine__destroy_kernel_maps(machine); 1216 return -1; 1217 } 1218 } 1219 1220 return 0; 1221 } 1222 1223 static void machine__set_kernel_mmap_len(struct machine *machine, 1224 union perf_event *event) 1225 { 1226 int i; 1227 1228 for (i = 0; i < MAP__NR_TYPES; i++) { 1229 machine->vmlinux_maps[i]->start = event->mmap.start; 1230 machine->vmlinux_maps[i]->end = (event->mmap.start + 1231 event->mmap.len); 1232 /* 1233 * Be a bit paranoid here, some perf.data file came with 1234 * a zero sized synthesized MMAP event for the kernel. 1235 */ 1236 if (machine->vmlinux_maps[i]->end == 0) 1237 machine->vmlinux_maps[i]->end = ~0ULL; 1238 } 1239 } 1240 1241 static bool machine__uses_kcore(struct machine *machine) 1242 { 1243 struct dso *dso; 1244 1245 list_for_each_entry(dso, &machine->dsos.head, node) { 1246 if (dso__is_kcore(dso)) 1247 return true; 1248 } 1249 1250 return false; 1251 } 1252 1253 static int machine__process_kernel_mmap_event(struct machine *machine, 1254 union perf_event *event) 1255 { 1256 struct map *map; 1257 char kmmap_prefix[PATH_MAX]; 1258 enum dso_kernel_type kernel_type; 1259 bool is_kernel_mmap; 1260 1261 /* If we have maps from kcore then we do not need or want any others */ 1262 if (machine__uses_kcore(machine)) 1263 return 0; 1264 1265 machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix)); 1266 if (machine__is_host(machine)) 1267 kernel_type = DSO_TYPE_KERNEL; 1268 else 1269 kernel_type = DSO_TYPE_GUEST_KERNEL; 1270 1271 is_kernel_mmap = memcmp(event->mmap.filename, 1272 kmmap_prefix, 1273 strlen(kmmap_prefix) - 1) == 0; 1274 if (event->mmap.filename[0] == '/' || 1275 (!is_kernel_mmap && event->mmap.filename[0] == '[')) { 1276 map = machine__findnew_module_map(machine, event->mmap.start, 1277 event->mmap.filename); 1278 if (map == NULL) 1279 goto out_problem; 1280 1281 map->end = map->start + event->mmap.len; 1282 } else if (is_kernel_mmap) { 1283 const char *symbol_name = (event->mmap.filename + 1284 strlen(kmmap_prefix)); 1285 /* 1286 * Should be there already, from the build-id table in 1287 * the header. 1288 */ 1289 struct dso *kernel = NULL; 1290 struct dso *dso; 1291 1292 pthread_rwlock_rdlock(&machine->dsos.lock); 1293 1294 list_for_each_entry(dso, &machine->dsos.head, node) { 1295 1296 /* 1297 * The cpumode passed to is_kernel_module is not the 1298 * cpumode of *this* event. If we insist on passing 1299 * correct cpumode to is_kernel_module, we should 1300 * record the cpumode when we adding this dso to the 1301 * linked list. 1302 * 1303 * However we don't really need passing correct 1304 * cpumode. We know the correct cpumode must be kernel 1305 * mode (if not, we should not link it onto kernel_dsos 1306 * list). 1307 * 1308 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN. 1309 * is_kernel_module() treats it as a kernel cpumode. 1310 */ 1311 1312 if (!dso->kernel || 1313 is_kernel_module(dso->long_name, 1314 PERF_RECORD_MISC_CPUMODE_UNKNOWN)) 1315 continue; 1316 1317 1318 kernel = dso; 1319 break; 1320 } 1321 1322 pthread_rwlock_unlock(&machine->dsos.lock); 1323 1324 if (kernel == NULL) 1325 kernel = machine__findnew_dso(machine, kmmap_prefix); 1326 if (kernel == NULL) 1327 goto out_problem; 1328 1329 kernel->kernel = kernel_type; 1330 if (__machine__create_kernel_maps(machine, kernel) < 0) { 1331 dso__put(kernel); 1332 goto out_problem; 1333 } 1334 1335 if (strstr(kernel->long_name, "vmlinux")) 1336 dso__set_short_name(kernel, "[kernel.vmlinux]", false); 1337 1338 machine__set_kernel_mmap_len(machine, event); 1339 1340 /* 1341 * Avoid using a zero address (kptr_restrict) for the ref reloc 1342 * symbol. Effectively having zero here means that at record 1343 * time /proc/sys/kernel/kptr_restrict was non zero. 1344 */ 1345 if (event->mmap.pgoff != 0) { 1346 maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, 1347 symbol_name, 1348 event->mmap.pgoff); 1349 } 1350 1351 if (machine__is_default_guest(machine)) { 1352 /* 1353 * preload dso of guest kernel and modules 1354 */ 1355 dso__load(kernel, machine__kernel_map(machine)); 1356 } 1357 } 1358 return 0; 1359 out_problem: 1360 return -1; 1361 } 1362 1363 int machine__process_mmap2_event(struct machine *machine, 1364 union perf_event *event, 1365 struct perf_sample *sample) 1366 { 1367 struct thread *thread; 1368 struct map *map; 1369 enum map_type type; 1370 int ret = 0; 1371 1372 if (dump_trace) 1373 perf_event__fprintf_mmap2(event, stdout); 1374 1375 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL || 1376 sample->cpumode == PERF_RECORD_MISC_KERNEL) { 1377 ret = machine__process_kernel_mmap_event(machine, event); 1378 if (ret < 0) 1379 goto out_problem; 1380 return 0; 1381 } 1382 1383 thread = machine__findnew_thread(machine, event->mmap2.pid, 1384 event->mmap2.tid); 1385 if (thread == NULL) 1386 goto out_problem; 1387 1388 if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) 1389 type = MAP__VARIABLE; 1390 else 1391 type = MAP__FUNCTION; 1392 1393 map = map__new(machine, event->mmap2.start, 1394 event->mmap2.len, event->mmap2.pgoff, 1395 event->mmap2.pid, event->mmap2.maj, 1396 event->mmap2.min, event->mmap2.ino, 1397 event->mmap2.ino_generation, 1398 event->mmap2.prot, 1399 event->mmap2.flags, 1400 event->mmap2.filename, type, thread); 1401 1402 if (map == NULL) 1403 goto out_problem_map; 1404 1405 ret = thread__insert_map(thread, map); 1406 if (ret) 1407 goto out_problem_insert; 1408 1409 thread__put(thread); 1410 map__put(map); 1411 return 0; 1412 1413 out_problem_insert: 1414 map__put(map); 1415 out_problem_map: 1416 thread__put(thread); 1417 out_problem: 1418 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n"); 1419 return 0; 1420 } 1421 1422 int machine__process_mmap_event(struct machine *machine, union perf_event *event, 1423 struct perf_sample *sample) 1424 { 1425 struct thread *thread; 1426 struct map *map; 1427 enum map_type type; 1428 int ret = 0; 1429 1430 if (dump_trace) 1431 perf_event__fprintf_mmap(event, stdout); 1432 1433 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL || 1434 sample->cpumode == PERF_RECORD_MISC_KERNEL) { 1435 ret = machine__process_kernel_mmap_event(machine, event); 1436 if (ret < 0) 1437 goto out_problem; 1438 return 0; 1439 } 1440 1441 thread = machine__findnew_thread(machine, event->mmap.pid, 1442 event->mmap.tid); 1443 if (thread == NULL) 1444 goto out_problem; 1445 1446 if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) 1447 type = MAP__VARIABLE; 1448 else 1449 type = MAP__FUNCTION; 1450 1451 map = map__new(machine, event->mmap.start, 1452 event->mmap.len, event->mmap.pgoff, 1453 event->mmap.pid, 0, 0, 0, 0, 0, 0, 1454 event->mmap.filename, 1455 type, thread); 1456 1457 if (map == NULL) 1458 goto out_problem_map; 1459 1460 ret = thread__insert_map(thread, map); 1461 if (ret) 1462 goto out_problem_insert; 1463 1464 thread__put(thread); 1465 map__put(map); 1466 return 0; 1467 1468 out_problem_insert: 1469 map__put(map); 1470 out_problem_map: 1471 thread__put(thread); 1472 out_problem: 1473 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n"); 1474 return 0; 1475 } 1476 1477 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock) 1478 { 1479 if (machine->last_match == th) 1480 machine->last_match = NULL; 1481 1482 BUG_ON(refcount_read(&th->refcnt) == 0); 1483 if (lock) 1484 pthread_rwlock_wrlock(&machine->threads_lock); 1485 rb_erase_init(&th->rb_node, &machine->threads); 1486 RB_CLEAR_NODE(&th->rb_node); 1487 --machine->nr_threads; 1488 /* 1489 * Move it first to the dead_threads list, then drop the reference, 1490 * if this is the last reference, then the thread__delete destructor 1491 * will be called and we will remove it from the dead_threads list. 1492 */ 1493 list_add_tail(&th->node, &machine->dead_threads); 1494 if (lock) 1495 pthread_rwlock_unlock(&machine->threads_lock); 1496 thread__put(th); 1497 } 1498 1499 void machine__remove_thread(struct machine *machine, struct thread *th) 1500 { 1501 return __machine__remove_thread(machine, th, true); 1502 } 1503 1504 int machine__process_fork_event(struct machine *machine, union perf_event *event, 1505 struct perf_sample *sample) 1506 { 1507 struct thread *thread = machine__find_thread(machine, 1508 event->fork.pid, 1509 event->fork.tid); 1510 struct thread *parent = machine__findnew_thread(machine, 1511 event->fork.ppid, 1512 event->fork.ptid); 1513 int err = 0; 1514 1515 if (dump_trace) 1516 perf_event__fprintf_task(event, stdout); 1517 1518 /* 1519 * There may be an existing thread that is not actually the parent, 1520 * either because we are processing events out of order, or because the 1521 * (fork) event that would have removed the thread was lost. Assume the 1522 * latter case and continue on as best we can. 1523 */ 1524 if (parent->pid_ != (pid_t)event->fork.ppid) { 1525 dump_printf("removing erroneous parent thread %d/%d\n", 1526 parent->pid_, parent->tid); 1527 machine__remove_thread(machine, parent); 1528 thread__put(parent); 1529 parent = machine__findnew_thread(machine, event->fork.ppid, 1530 event->fork.ptid); 1531 } 1532 1533 /* if a thread currently exists for the thread id remove it */ 1534 if (thread != NULL) { 1535 machine__remove_thread(machine, thread); 1536 thread__put(thread); 1537 } 1538 1539 thread = machine__findnew_thread(machine, event->fork.pid, 1540 event->fork.tid); 1541 1542 if (thread == NULL || parent == NULL || 1543 thread__fork(thread, parent, sample->time) < 0) { 1544 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n"); 1545 err = -1; 1546 } 1547 thread__put(thread); 1548 thread__put(parent); 1549 1550 return err; 1551 } 1552 1553 int machine__process_exit_event(struct machine *machine, union perf_event *event, 1554 struct perf_sample *sample __maybe_unused) 1555 { 1556 struct thread *thread = machine__find_thread(machine, 1557 event->fork.pid, 1558 event->fork.tid); 1559 1560 if (dump_trace) 1561 perf_event__fprintf_task(event, stdout); 1562 1563 if (thread != NULL) { 1564 thread__exited(thread); 1565 thread__put(thread); 1566 } 1567 1568 return 0; 1569 } 1570 1571 int machine__process_event(struct machine *machine, union perf_event *event, 1572 struct perf_sample *sample) 1573 { 1574 int ret; 1575 1576 switch (event->header.type) { 1577 case PERF_RECORD_COMM: 1578 ret = machine__process_comm_event(machine, event, sample); break; 1579 case PERF_RECORD_MMAP: 1580 ret = machine__process_mmap_event(machine, event, sample); break; 1581 case PERF_RECORD_NAMESPACES: 1582 ret = machine__process_namespaces_event(machine, event, sample); break; 1583 case PERF_RECORD_MMAP2: 1584 ret = machine__process_mmap2_event(machine, event, sample); break; 1585 case PERF_RECORD_FORK: 1586 ret = machine__process_fork_event(machine, event, sample); break; 1587 case PERF_RECORD_EXIT: 1588 ret = machine__process_exit_event(machine, event, sample); break; 1589 case PERF_RECORD_LOST: 1590 ret = machine__process_lost_event(machine, event, sample); break; 1591 case PERF_RECORD_AUX: 1592 ret = machine__process_aux_event(machine, event); break; 1593 case PERF_RECORD_ITRACE_START: 1594 ret = machine__process_itrace_start_event(machine, event); break; 1595 case PERF_RECORD_LOST_SAMPLES: 1596 ret = machine__process_lost_samples_event(machine, event, sample); break; 1597 case PERF_RECORD_SWITCH: 1598 case PERF_RECORD_SWITCH_CPU_WIDE: 1599 ret = machine__process_switch_event(machine, event); break; 1600 default: 1601 ret = -1; 1602 break; 1603 } 1604 1605 return ret; 1606 } 1607 1608 static bool symbol__match_regex(struct symbol *sym, regex_t *regex) 1609 { 1610 if (!regexec(regex, sym->name, 0, NULL, 0)) 1611 return 1; 1612 return 0; 1613 } 1614 1615 static void ip__resolve_ams(struct thread *thread, 1616 struct addr_map_symbol *ams, 1617 u64 ip) 1618 { 1619 struct addr_location al; 1620 1621 memset(&al, 0, sizeof(al)); 1622 /* 1623 * We cannot use the header.misc hint to determine whether a 1624 * branch stack address is user, kernel, guest, hypervisor. 1625 * Branches may straddle the kernel/user/hypervisor boundaries. 1626 * Thus, we have to try consecutively until we find a match 1627 * or else, the symbol is unknown 1628 */ 1629 thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al); 1630 1631 ams->addr = ip; 1632 ams->al_addr = al.addr; 1633 ams->sym = al.sym; 1634 ams->map = al.map; 1635 } 1636 1637 static void ip__resolve_data(struct thread *thread, 1638 u8 m, struct addr_map_symbol *ams, u64 addr) 1639 { 1640 struct addr_location al; 1641 1642 memset(&al, 0, sizeof(al)); 1643 1644 thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al); 1645 if (al.map == NULL) { 1646 /* 1647 * some shared data regions have execute bit set which puts 1648 * their mapping in the MAP__FUNCTION type array. 1649 * Check there as a fallback option before dropping the sample. 1650 */ 1651 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al); 1652 } 1653 1654 ams->addr = addr; 1655 ams->al_addr = al.addr; 1656 ams->sym = al.sym; 1657 ams->map = al.map; 1658 } 1659 1660 struct mem_info *sample__resolve_mem(struct perf_sample *sample, 1661 struct addr_location *al) 1662 { 1663 struct mem_info *mi = zalloc(sizeof(*mi)); 1664 1665 if (!mi) 1666 return NULL; 1667 1668 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip); 1669 ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr); 1670 mi->data_src.val = sample->data_src; 1671 1672 return mi; 1673 } 1674 1675 static int add_callchain_ip(struct thread *thread, 1676 struct callchain_cursor *cursor, 1677 struct symbol **parent, 1678 struct addr_location *root_al, 1679 u8 *cpumode, 1680 u64 ip, 1681 bool branch, 1682 struct branch_flags *flags, 1683 int nr_loop_iter, 1684 int samples) 1685 { 1686 struct addr_location al; 1687 1688 al.filtered = 0; 1689 al.sym = NULL; 1690 if (!cpumode) { 1691 thread__find_cpumode_addr_location(thread, MAP__FUNCTION, 1692 ip, &al); 1693 } else { 1694 if (ip >= PERF_CONTEXT_MAX) { 1695 switch (ip) { 1696 case PERF_CONTEXT_HV: 1697 *cpumode = PERF_RECORD_MISC_HYPERVISOR; 1698 break; 1699 case PERF_CONTEXT_KERNEL: 1700 *cpumode = PERF_RECORD_MISC_KERNEL; 1701 break; 1702 case PERF_CONTEXT_USER: 1703 *cpumode = PERF_RECORD_MISC_USER; 1704 break; 1705 default: 1706 pr_debug("invalid callchain context: " 1707 "%"PRId64"\n", (s64) ip); 1708 /* 1709 * It seems the callchain is corrupted. 1710 * Discard all. 1711 */ 1712 callchain_cursor_reset(cursor); 1713 return 1; 1714 } 1715 return 0; 1716 } 1717 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION, 1718 ip, &al); 1719 } 1720 1721 if (al.sym != NULL) { 1722 if (perf_hpp_list.parent && !*parent && 1723 symbol__match_regex(al.sym, &parent_regex)) 1724 *parent = al.sym; 1725 else if (have_ignore_callees && root_al && 1726 symbol__match_regex(al.sym, &ignore_callees_regex)) { 1727 /* Treat this symbol as the root, 1728 forgetting its callees. */ 1729 *root_al = al; 1730 callchain_cursor_reset(cursor); 1731 } 1732 } 1733 1734 if (symbol_conf.hide_unresolved && al.sym == NULL) 1735 return 0; 1736 return callchain_cursor_append(cursor, al.addr, al.map, al.sym, 1737 branch, flags, nr_loop_iter, samples); 1738 } 1739 1740 struct branch_info *sample__resolve_bstack(struct perf_sample *sample, 1741 struct addr_location *al) 1742 { 1743 unsigned int i; 1744 const struct branch_stack *bs = sample->branch_stack; 1745 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info)); 1746 1747 if (!bi) 1748 return NULL; 1749 1750 for (i = 0; i < bs->nr; i++) { 1751 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to); 1752 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from); 1753 bi[i].flags = bs->entries[i].flags; 1754 } 1755 return bi; 1756 } 1757 1758 #define CHASHSZ 127 1759 #define CHASHBITS 7 1760 #define NO_ENTRY 0xff 1761 1762 #define PERF_MAX_BRANCH_DEPTH 127 1763 1764 /* Remove loops. */ 1765 static int remove_loops(struct branch_entry *l, int nr) 1766 { 1767 int i, j, off; 1768 unsigned char chash[CHASHSZ]; 1769 1770 memset(chash, NO_ENTRY, sizeof(chash)); 1771 1772 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255); 1773 1774 for (i = 0; i < nr; i++) { 1775 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ; 1776 1777 /* no collision handling for now */ 1778 if (chash[h] == NO_ENTRY) { 1779 chash[h] = i; 1780 } else if (l[chash[h]].from == l[i].from) { 1781 bool is_loop = true; 1782 /* check if it is a real loop */ 1783 off = 0; 1784 for (j = chash[h]; j < i && i + off < nr; j++, off++) 1785 if (l[j].from != l[i + off].from) { 1786 is_loop = false; 1787 break; 1788 } 1789 if (is_loop) { 1790 memmove(l + i, l + i + off, 1791 (nr - (i + off)) * sizeof(*l)); 1792 nr -= off; 1793 } 1794 } 1795 } 1796 return nr; 1797 } 1798 1799 /* 1800 * Recolve LBR callstack chain sample 1801 * Return: 1802 * 1 on success get LBR callchain information 1803 * 0 no available LBR callchain information, should try fp 1804 * negative error code on other errors. 1805 */ 1806 static int resolve_lbr_callchain_sample(struct thread *thread, 1807 struct callchain_cursor *cursor, 1808 struct perf_sample *sample, 1809 struct symbol **parent, 1810 struct addr_location *root_al, 1811 int max_stack) 1812 { 1813 struct ip_callchain *chain = sample->callchain; 1814 int chain_nr = min(max_stack, (int)chain->nr), i; 1815 u8 cpumode = PERF_RECORD_MISC_USER; 1816 u64 ip; 1817 1818 for (i = 0; i < chain_nr; i++) { 1819 if (chain->ips[i] == PERF_CONTEXT_USER) 1820 break; 1821 } 1822 1823 /* LBR only affects the user callchain */ 1824 if (i != chain_nr) { 1825 struct branch_stack *lbr_stack = sample->branch_stack; 1826 int lbr_nr = lbr_stack->nr, j, k; 1827 bool branch; 1828 struct branch_flags *flags; 1829 /* 1830 * LBR callstack can only get user call chain. 1831 * The mix_chain_nr is kernel call chain 1832 * number plus LBR user call chain number. 1833 * i is kernel call chain number, 1834 * 1 is PERF_CONTEXT_USER, 1835 * lbr_nr + 1 is the user call chain number. 1836 * For details, please refer to the comments 1837 * in callchain__printf 1838 */ 1839 int mix_chain_nr = i + 1 + lbr_nr + 1; 1840 1841 for (j = 0; j < mix_chain_nr; j++) { 1842 int err; 1843 branch = false; 1844 flags = NULL; 1845 1846 if (callchain_param.order == ORDER_CALLEE) { 1847 if (j < i + 1) 1848 ip = chain->ips[j]; 1849 else if (j > i + 1) { 1850 k = j - i - 2; 1851 ip = lbr_stack->entries[k].from; 1852 branch = true; 1853 flags = &lbr_stack->entries[k].flags; 1854 } else { 1855 ip = lbr_stack->entries[0].to; 1856 branch = true; 1857 flags = &lbr_stack->entries[0].flags; 1858 } 1859 } else { 1860 if (j < lbr_nr) { 1861 k = lbr_nr - j - 1; 1862 ip = lbr_stack->entries[k].from; 1863 branch = true; 1864 flags = &lbr_stack->entries[k].flags; 1865 } 1866 else if (j > lbr_nr) 1867 ip = chain->ips[i + 1 - (j - lbr_nr)]; 1868 else { 1869 ip = lbr_stack->entries[0].to; 1870 branch = true; 1871 flags = &lbr_stack->entries[0].flags; 1872 } 1873 } 1874 1875 err = add_callchain_ip(thread, cursor, parent, 1876 root_al, &cpumode, ip, 1877 branch, flags, 0, 0); 1878 if (err) 1879 return (err < 0) ? err : 0; 1880 } 1881 return 1; 1882 } 1883 1884 return 0; 1885 } 1886 1887 static int thread__resolve_callchain_sample(struct thread *thread, 1888 struct callchain_cursor *cursor, 1889 struct perf_evsel *evsel, 1890 struct perf_sample *sample, 1891 struct symbol **parent, 1892 struct addr_location *root_al, 1893 int max_stack) 1894 { 1895 struct branch_stack *branch = sample->branch_stack; 1896 struct ip_callchain *chain = sample->callchain; 1897 int chain_nr = chain->nr; 1898 u8 cpumode = PERF_RECORD_MISC_USER; 1899 int i, j, err, nr_entries; 1900 int skip_idx = -1; 1901 int first_call = 0; 1902 int nr_loop_iter; 1903 1904 if (perf_evsel__has_branch_callstack(evsel)) { 1905 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent, 1906 root_al, max_stack); 1907 if (err) 1908 return (err < 0) ? err : 0; 1909 } 1910 1911 /* 1912 * Based on DWARF debug information, some architectures skip 1913 * a callchain entry saved by the kernel. 1914 */ 1915 skip_idx = arch_skip_callchain_idx(thread, chain); 1916 1917 /* 1918 * Add branches to call stack for easier browsing. This gives 1919 * more context for a sample than just the callers. 1920 * 1921 * This uses individual histograms of paths compared to the 1922 * aggregated histograms the normal LBR mode uses. 1923 * 1924 * Limitations for now: 1925 * - No extra filters 1926 * - No annotations (should annotate somehow) 1927 */ 1928 1929 if (branch && callchain_param.branch_callstack) { 1930 int nr = min(max_stack, (int)branch->nr); 1931 struct branch_entry be[nr]; 1932 1933 if (branch->nr > PERF_MAX_BRANCH_DEPTH) { 1934 pr_warning("corrupted branch chain. skipping...\n"); 1935 goto check_calls; 1936 } 1937 1938 for (i = 0; i < nr; i++) { 1939 if (callchain_param.order == ORDER_CALLEE) { 1940 be[i] = branch->entries[i]; 1941 /* 1942 * Check for overlap into the callchain. 1943 * The return address is one off compared to 1944 * the branch entry. To adjust for this 1945 * assume the calling instruction is not longer 1946 * than 8 bytes. 1947 */ 1948 if (i == skip_idx || 1949 chain->ips[first_call] >= PERF_CONTEXT_MAX) 1950 first_call++; 1951 else if (be[i].from < chain->ips[first_call] && 1952 be[i].from >= chain->ips[first_call] - 8) 1953 first_call++; 1954 } else 1955 be[i] = branch->entries[branch->nr - i - 1]; 1956 } 1957 1958 nr_loop_iter = nr; 1959 nr = remove_loops(be, nr); 1960 1961 /* 1962 * Get the number of iterations. 1963 * It's only approximation, but good enough in practice. 1964 */ 1965 if (nr_loop_iter > nr) 1966 nr_loop_iter = nr_loop_iter - nr + 1; 1967 else 1968 nr_loop_iter = 0; 1969 1970 for (i = 0; i < nr; i++) { 1971 if (i == nr - 1) 1972 err = add_callchain_ip(thread, cursor, parent, 1973 root_al, 1974 NULL, be[i].to, 1975 true, &be[i].flags, 1976 nr_loop_iter, 1); 1977 else 1978 err = add_callchain_ip(thread, cursor, parent, 1979 root_al, 1980 NULL, be[i].to, 1981 true, &be[i].flags, 1982 0, 0); 1983 1984 if (!err) 1985 err = add_callchain_ip(thread, cursor, parent, root_al, 1986 NULL, be[i].from, 1987 true, &be[i].flags, 1988 0, 0); 1989 if (err == -EINVAL) 1990 break; 1991 if (err) 1992 return err; 1993 } 1994 chain_nr -= nr; 1995 } 1996 1997 check_calls: 1998 for (i = first_call, nr_entries = 0; 1999 i < chain_nr && nr_entries < max_stack; i++) { 2000 u64 ip; 2001 2002 if (callchain_param.order == ORDER_CALLEE) 2003 j = i; 2004 else 2005 j = chain->nr - i - 1; 2006 2007 #ifdef HAVE_SKIP_CALLCHAIN_IDX 2008 if (j == skip_idx) 2009 continue; 2010 #endif 2011 ip = chain->ips[j]; 2012 2013 if (ip < PERF_CONTEXT_MAX) 2014 ++nr_entries; 2015 2016 err = add_callchain_ip(thread, cursor, parent, 2017 root_al, &cpumode, ip, 2018 false, NULL, 0, 0); 2019 2020 if (err) 2021 return (err < 0) ? err : 0; 2022 } 2023 2024 return 0; 2025 } 2026 2027 static int unwind_entry(struct unwind_entry *entry, void *arg) 2028 { 2029 struct callchain_cursor *cursor = arg; 2030 2031 if (symbol_conf.hide_unresolved && entry->sym == NULL) 2032 return 0; 2033 return callchain_cursor_append(cursor, entry->ip, 2034 entry->map, entry->sym, 2035 false, NULL, 0, 0); 2036 } 2037 2038 static int thread__resolve_callchain_unwind(struct thread *thread, 2039 struct callchain_cursor *cursor, 2040 struct perf_evsel *evsel, 2041 struct perf_sample *sample, 2042 int max_stack) 2043 { 2044 /* Can we do dwarf post unwind? */ 2045 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) && 2046 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER))) 2047 return 0; 2048 2049 /* Bail out if nothing was captured. */ 2050 if ((!sample->user_regs.regs) || 2051 (!sample->user_stack.size)) 2052 return 0; 2053 2054 return unwind__get_entries(unwind_entry, cursor, 2055 thread, sample, max_stack); 2056 } 2057 2058 int thread__resolve_callchain(struct thread *thread, 2059 struct callchain_cursor *cursor, 2060 struct perf_evsel *evsel, 2061 struct perf_sample *sample, 2062 struct symbol **parent, 2063 struct addr_location *root_al, 2064 int max_stack) 2065 { 2066 int ret = 0; 2067 2068 callchain_cursor_reset(&callchain_cursor); 2069 2070 if (callchain_param.order == ORDER_CALLEE) { 2071 ret = thread__resolve_callchain_sample(thread, cursor, 2072 evsel, sample, 2073 parent, root_al, 2074 max_stack); 2075 if (ret) 2076 return ret; 2077 ret = thread__resolve_callchain_unwind(thread, cursor, 2078 evsel, sample, 2079 max_stack); 2080 } else { 2081 ret = thread__resolve_callchain_unwind(thread, cursor, 2082 evsel, sample, 2083 max_stack); 2084 if (ret) 2085 return ret; 2086 ret = thread__resolve_callchain_sample(thread, cursor, 2087 evsel, sample, 2088 parent, root_al, 2089 max_stack); 2090 } 2091 2092 return ret; 2093 } 2094 2095 int machine__for_each_thread(struct machine *machine, 2096 int (*fn)(struct thread *thread, void *p), 2097 void *priv) 2098 { 2099 struct rb_node *nd; 2100 struct thread *thread; 2101 int rc = 0; 2102 2103 for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) { 2104 thread = rb_entry(nd, struct thread, rb_node); 2105 rc = fn(thread, priv); 2106 if (rc != 0) 2107 return rc; 2108 } 2109 2110 list_for_each_entry(thread, &machine->dead_threads, node) { 2111 rc = fn(thread, priv); 2112 if (rc != 0) 2113 return rc; 2114 } 2115 return rc; 2116 } 2117 2118 int machines__for_each_thread(struct machines *machines, 2119 int (*fn)(struct thread *thread, void *p), 2120 void *priv) 2121 { 2122 struct rb_node *nd; 2123 int rc = 0; 2124 2125 rc = machine__for_each_thread(&machines->host, fn, priv); 2126 if (rc != 0) 2127 return rc; 2128 2129 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 2130 struct machine *machine = rb_entry(nd, struct machine, rb_node); 2131 2132 rc = machine__for_each_thread(machine, fn, priv); 2133 if (rc != 0) 2134 return rc; 2135 } 2136 return rc; 2137 } 2138 2139 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 2140 struct target *target, struct thread_map *threads, 2141 perf_event__handler_t process, bool data_mmap, 2142 unsigned int proc_map_timeout) 2143 { 2144 if (target__has_task(target)) 2145 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap, proc_map_timeout); 2146 else if (target__has_cpu(target)) 2147 return perf_event__synthesize_threads(tool, process, machine, data_mmap, proc_map_timeout); 2148 /* command specified */ 2149 return 0; 2150 } 2151 2152 pid_t machine__get_current_tid(struct machine *machine, int cpu) 2153 { 2154 if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid) 2155 return -1; 2156 2157 return machine->current_tid[cpu]; 2158 } 2159 2160 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid, 2161 pid_t tid) 2162 { 2163 struct thread *thread; 2164 2165 if (cpu < 0) 2166 return -EINVAL; 2167 2168 if (!machine->current_tid) { 2169 int i; 2170 2171 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t)); 2172 if (!machine->current_tid) 2173 return -ENOMEM; 2174 for (i = 0; i < MAX_NR_CPUS; i++) 2175 machine->current_tid[i] = -1; 2176 } 2177 2178 if (cpu >= MAX_NR_CPUS) { 2179 pr_err("Requested CPU %d too large. ", cpu); 2180 pr_err("Consider raising MAX_NR_CPUS\n"); 2181 return -EINVAL; 2182 } 2183 2184 machine->current_tid[cpu] = tid; 2185 2186 thread = machine__findnew_thread(machine, pid, tid); 2187 if (!thread) 2188 return -ENOMEM; 2189 2190 thread->cpu = cpu; 2191 thread__put(thread); 2192 2193 return 0; 2194 } 2195 2196 int machine__get_kernel_start(struct machine *machine) 2197 { 2198 struct map *map = machine__kernel_map(machine); 2199 int err = 0; 2200 2201 /* 2202 * The only addresses above 2^63 are kernel addresses of a 64-bit 2203 * kernel. Note that addresses are unsigned so that on a 32-bit system 2204 * all addresses including kernel addresses are less than 2^32. In 2205 * that case (32-bit system), if the kernel mapping is unknown, all 2206 * addresses will be assumed to be in user space - see 2207 * machine__kernel_ip(). 2208 */ 2209 machine->kernel_start = 1ULL << 63; 2210 if (map) { 2211 err = map__load(map); 2212 if (!err) 2213 machine->kernel_start = map->start; 2214 } 2215 return err; 2216 } 2217 2218 struct dso *machine__findnew_dso(struct machine *machine, const char *filename) 2219 { 2220 return dsos__findnew(&machine->dsos, filename); 2221 } 2222 2223 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp) 2224 { 2225 struct machine *machine = vmachine; 2226 struct map *map; 2227 struct symbol *sym = map_groups__find_symbol(&machine->kmaps, MAP__FUNCTION, *addrp, &map); 2228 2229 if (sym == NULL) 2230 return NULL; 2231 2232 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL; 2233 *addrp = map->unmap_ip(map, sym->start); 2234 return sym->name; 2235 } 2236