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 if (machine__is_host(machine)) 576 dso->symtab_type = DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE; 577 else 578 dso->symtab_type = DSO_BINARY_TYPE__GUEST_KMODULE; 579 580 /* _KMODULE_COMP should be next to _KMODULE */ 581 if (m->kmod && m->comp) 582 dso->symtab_type++; 583 584 dso__set_short_name(dso, strdup(m->name), true); 585 dso__set_long_name(dso, strdup(filename), true); 586 } 587 588 dso__get(dso); 589 out_unlock: 590 pthread_rwlock_unlock(&machine->dsos.lock); 591 return dso; 592 } 593 594 int machine__process_aux_event(struct machine *machine __maybe_unused, 595 union perf_event *event) 596 { 597 if (dump_trace) 598 perf_event__fprintf_aux(event, stdout); 599 return 0; 600 } 601 602 int machine__process_itrace_start_event(struct machine *machine __maybe_unused, 603 union perf_event *event) 604 { 605 if (dump_trace) 606 perf_event__fprintf_itrace_start(event, stdout); 607 return 0; 608 } 609 610 int machine__process_switch_event(struct machine *machine __maybe_unused, 611 union perf_event *event) 612 { 613 if (dump_trace) 614 perf_event__fprintf_switch(event, stdout); 615 return 0; 616 } 617 618 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename) 619 { 620 const char *dup_filename; 621 622 if (!filename || !dso || !dso->long_name) 623 return; 624 if (dso->long_name[0] != '[') 625 return; 626 if (!strchr(filename, '/')) 627 return; 628 629 dup_filename = strdup(filename); 630 if (!dup_filename) 631 return; 632 633 dso__set_long_name(dso, dup_filename, true); 634 } 635 636 struct map *machine__findnew_module_map(struct machine *machine, u64 start, 637 const char *filename) 638 { 639 struct map *map = NULL; 640 struct dso *dso = NULL; 641 struct kmod_path m; 642 643 if (kmod_path__parse_name(&m, filename)) 644 return NULL; 645 646 map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION, 647 m.name); 648 if (map) { 649 /* 650 * If the map's dso is an offline module, give dso__load() 651 * a chance to find the file path of that module by fixing 652 * long_name. 653 */ 654 dso__adjust_kmod_long_name(map->dso, filename); 655 goto out; 656 } 657 658 dso = machine__findnew_module_dso(machine, &m, filename); 659 if (dso == NULL) 660 goto out; 661 662 map = map__new2(start, dso, MAP__FUNCTION); 663 if (map == NULL) 664 goto out; 665 666 map_groups__insert(&machine->kmaps, map); 667 668 /* Put the map here because map_groups__insert alread got it */ 669 map__put(map); 670 out: 671 /* put the dso here, corresponding to machine__findnew_module_dso */ 672 dso__put(dso); 673 free(m.name); 674 return map; 675 } 676 677 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp) 678 { 679 struct rb_node *nd; 680 size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp); 681 682 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 683 struct machine *pos = rb_entry(nd, struct machine, rb_node); 684 ret += __dsos__fprintf(&pos->dsos.head, fp); 685 } 686 687 return ret; 688 } 689 690 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp, 691 bool (skip)(struct dso *dso, int parm), int parm) 692 { 693 return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm); 694 } 695 696 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp, 697 bool (skip)(struct dso *dso, int parm), int parm) 698 { 699 struct rb_node *nd; 700 size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm); 701 702 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 703 struct machine *pos = rb_entry(nd, struct machine, rb_node); 704 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm); 705 } 706 return ret; 707 } 708 709 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp) 710 { 711 int i; 712 size_t printed = 0; 713 struct dso *kdso = machine__kernel_map(machine)->dso; 714 715 if (kdso->has_build_id) { 716 char filename[PATH_MAX]; 717 if (dso__build_id_filename(kdso, filename, sizeof(filename))) 718 printed += fprintf(fp, "[0] %s\n", filename); 719 } 720 721 for (i = 0; i < vmlinux_path__nr_entries; ++i) 722 printed += fprintf(fp, "[%d] %s\n", 723 i + kdso->has_build_id, vmlinux_path[i]); 724 725 return printed; 726 } 727 728 size_t machine__fprintf(struct machine *machine, FILE *fp) 729 { 730 size_t ret; 731 struct rb_node *nd; 732 733 pthread_rwlock_rdlock(&machine->threads_lock); 734 735 ret = fprintf(fp, "Threads: %u\n", machine->nr_threads); 736 737 for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) { 738 struct thread *pos = rb_entry(nd, struct thread, rb_node); 739 740 ret += thread__fprintf(pos, fp); 741 } 742 743 pthread_rwlock_unlock(&machine->threads_lock); 744 745 return ret; 746 } 747 748 static struct dso *machine__get_kernel(struct machine *machine) 749 { 750 const char *vmlinux_name = NULL; 751 struct dso *kernel; 752 753 if (machine__is_host(machine)) { 754 vmlinux_name = symbol_conf.vmlinux_name; 755 if (!vmlinux_name) 756 vmlinux_name = DSO__NAME_KALLSYMS; 757 758 kernel = machine__findnew_kernel(machine, vmlinux_name, 759 "[kernel]", DSO_TYPE_KERNEL); 760 } else { 761 char bf[PATH_MAX]; 762 763 if (machine__is_default_guest(machine)) 764 vmlinux_name = symbol_conf.default_guest_vmlinux_name; 765 if (!vmlinux_name) 766 vmlinux_name = machine__mmap_name(machine, bf, 767 sizeof(bf)); 768 769 kernel = machine__findnew_kernel(machine, vmlinux_name, 770 "[guest.kernel]", 771 DSO_TYPE_GUEST_KERNEL); 772 } 773 774 if (kernel != NULL && (!kernel->has_build_id)) 775 dso__read_running_kernel_build_id(kernel, machine); 776 777 return kernel; 778 } 779 780 struct process_args { 781 u64 start; 782 }; 783 784 static void machine__get_kallsyms_filename(struct machine *machine, char *buf, 785 size_t bufsz) 786 { 787 if (machine__is_default_guest(machine)) 788 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms); 789 else 790 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir); 791 } 792 793 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL}; 794 795 /* Figure out the start address of kernel map from /proc/kallsyms. 796 * Returns the name of the start symbol in *symbol_name. Pass in NULL as 797 * symbol_name if it's not that important. 798 */ 799 static u64 machine__get_running_kernel_start(struct machine *machine, 800 const char **symbol_name) 801 { 802 char filename[PATH_MAX]; 803 int i; 804 const char *name; 805 u64 addr = 0; 806 807 machine__get_kallsyms_filename(machine, filename, PATH_MAX); 808 809 if (symbol__restricted_filename(filename, "/proc/kallsyms")) 810 return 0; 811 812 for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) { 813 addr = kallsyms__get_function_start(filename, name); 814 if (addr) 815 break; 816 } 817 818 if (symbol_name) 819 *symbol_name = name; 820 821 return addr; 822 } 823 824 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel) 825 { 826 int type; 827 u64 start = machine__get_running_kernel_start(machine, NULL); 828 829 /* In case of renewal the kernel map, destroy previous one */ 830 machine__destroy_kernel_maps(machine); 831 832 for (type = 0; type < MAP__NR_TYPES; ++type) { 833 struct kmap *kmap; 834 struct map *map; 835 836 machine->vmlinux_maps[type] = map__new2(start, kernel, type); 837 if (machine->vmlinux_maps[type] == NULL) 838 return -1; 839 840 machine->vmlinux_maps[type]->map_ip = 841 machine->vmlinux_maps[type]->unmap_ip = 842 identity__map_ip; 843 map = __machine__kernel_map(machine, type); 844 kmap = map__kmap(map); 845 if (!kmap) 846 return -1; 847 848 kmap->kmaps = &machine->kmaps; 849 map_groups__insert(&machine->kmaps, map); 850 } 851 852 return 0; 853 } 854 855 void machine__destroy_kernel_maps(struct machine *machine) 856 { 857 int type; 858 859 for (type = 0; type < MAP__NR_TYPES; ++type) { 860 struct kmap *kmap; 861 struct map *map = __machine__kernel_map(machine, type); 862 863 if (map == NULL) 864 continue; 865 866 kmap = map__kmap(map); 867 map_groups__remove(&machine->kmaps, map); 868 if (kmap && kmap->ref_reloc_sym) { 869 /* 870 * ref_reloc_sym is shared among all maps, so free just 871 * on one of them. 872 */ 873 if (type == MAP__FUNCTION) { 874 zfree((char **)&kmap->ref_reloc_sym->name); 875 zfree(&kmap->ref_reloc_sym); 876 } else 877 kmap->ref_reloc_sym = NULL; 878 } 879 880 map__put(machine->vmlinux_maps[type]); 881 machine->vmlinux_maps[type] = NULL; 882 } 883 } 884 885 int machines__create_guest_kernel_maps(struct machines *machines) 886 { 887 int ret = 0; 888 struct dirent **namelist = NULL; 889 int i, items = 0; 890 char path[PATH_MAX]; 891 pid_t pid; 892 char *endp; 893 894 if (symbol_conf.default_guest_vmlinux_name || 895 symbol_conf.default_guest_modules || 896 symbol_conf.default_guest_kallsyms) { 897 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID); 898 } 899 900 if (symbol_conf.guestmount) { 901 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL); 902 if (items <= 0) 903 return -ENOENT; 904 for (i = 0; i < items; i++) { 905 if (!isdigit(namelist[i]->d_name[0])) { 906 /* Filter out . and .. */ 907 continue; 908 } 909 pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10); 910 if ((*endp != '\0') || 911 (endp == namelist[i]->d_name) || 912 (errno == ERANGE)) { 913 pr_debug("invalid directory (%s). Skipping.\n", 914 namelist[i]->d_name); 915 continue; 916 } 917 sprintf(path, "%s/%s/proc/kallsyms", 918 symbol_conf.guestmount, 919 namelist[i]->d_name); 920 ret = access(path, R_OK); 921 if (ret) { 922 pr_debug("Can't access file %s\n", path); 923 goto failure; 924 } 925 machines__create_kernel_maps(machines, pid); 926 } 927 failure: 928 free(namelist); 929 } 930 931 return ret; 932 } 933 934 void machines__destroy_kernel_maps(struct machines *machines) 935 { 936 struct rb_node *next = rb_first(&machines->guests); 937 938 machine__destroy_kernel_maps(&machines->host); 939 940 while (next) { 941 struct machine *pos = rb_entry(next, struct machine, rb_node); 942 943 next = rb_next(&pos->rb_node); 944 rb_erase(&pos->rb_node, &machines->guests); 945 machine__delete(pos); 946 } 947 } 948 949 int machines__create_kernel_maps(struct machines *machines, pid_t pid) 950 { 951 struct machine *machine = machines__findnew(machines, pid); 952 953 if (machine == NULL) 954 return -1; 955 956 return machine__create_kernel_maps(machine); 957 } 958 959 int __machine__load_kallsyms(struct machine *machine, const char *filename, 960 enum map_type type, bool no_kcore) 961 { 962 struct map *map = machine__kernel_map(machine); 963 int ret = __dso__load_kallsyms(map->dso, filename, map, no_kcore); 964 965 if (ret > 0) { 966 dso__set_loaded(map->dso, type); 967 /* 968 * Since /proc/kallsyms will have multiple sessions for the 969 * kernel, with modules between them, fixup the end of all 970 * sections. 971 */ 972 __map_groups__fixup_end(&machine->kmaps, type); 973 } 974 975 return ret; 976 } 977 978 int machine__load_kallsyms(struct machine *machine, const char *filename, 979 enum map_type type) 980 { 981 return __machine__load_kallsyms(machine, filename, type, false); 982 } 983 984 int machine__load_vmlinux_path(struct machine *machine, enum map_type type) 985 { 986 struct map *map = machine__kernel_map(machine); 987 int ret = dso__load_vmlinux_path(map->dso, map); 988 989 if (ret > 0) 990 dso__set_loaded(map->dso, type); 991 992 return ret; 993 } 994 995 static void map_groups__fixup_end(struct map_groups *mg) 996 { 997 int i; 998 for (i = 0; i < MAP__NR_TYPES; ++i) 999 __map_groups__fixup_end(mg, i); 1000 } 1001 1002 static char *get_kernel_version(const char *root_dir) 1003 { 1004 char version[PATH_MAX]; 1005 FILE *file; 1006 char *name, *tmp; 1007 const char *prefix = "Linux version "; 1008 1009 sprintf(version, "%s/proc/version", root_dir); 1010 file = fopen(version, "r"); 1011 if (!file) 1012 return NULL; 1013 1014 version[0] = '\0'; 1015 tmp = fgets(version, sizeof(version), file); 1016 fclose(file); 1017 1018 name = strstr(version, prefix); 1019 if (!name) 1020 return NULL; 1021 name += strlen(prefix); 1022 tmp = strchr(name, ' '); 1023 if (tmp) 1024 *tmp = '\0'; 1025 1026 return strdup(name); 1027 } 1028 1029 static bool is_kmod_dso(struct dso *dso) 1030 { 1031 return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE || 1032 dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE; 1033 } 1034 1035 static int map_groups__set_module_path(struct map_groups *mg, const char *path, 1036 struct kmod_path *m) 1037 { 1038 struct map *map; 1039 char *long_name; 1040 1041 map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name); 1042 if (map == NULL) 1043 return 0; 1044 1045 long_name = strdup(path); 1046 if (long_name == NULL) 1047 return -ENOMEM; 1048 1049 dso__set_long_name(map->dso, long_name, true); 1050 dso__kernel_module_get_build_id(map->dso, ""); 1051 1052 /* 1053 * Full name could reveal us kmod compression, so 1054 * we need to update the symtab_type if needed. 1055 */ 1056 if (m->comp && is_kmod_dso(map->dso)) 1057 map->dso->symtab_type++; 1058 1059 return 0; 1060 } 1061 1062 static int map_groups__set_modules_path_dir(struct map_groups *mg, 1063 const char *dir_name, int depth) 1064 { 1065 struct dirent *dent; 1066 DIR *dir = opendir(dir_name); 1067 int ret = 0; 1068 1069 if (!dir) { 1070 pr_debug("%s: cannot open %s dir\n", __func__, dir_name); 1071 return -1; 1072 } 1073 1074 while ((dent = readdir(dir)) != NULL) { 1075 char path[PATH_MAX]; 1076 struct stat st; 1077 1078 /*sshfs might return bad dent->d_type, so we have to stat*/ 1079 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name); 1080 if (stat(path, &st)) 1081 continue; 1082 1083 if (S_ISDIR(st.st_mode)) { 1084 if (!strcmp(dent->d_name, ".") || 1085 !strcmp(dent->d_name, "..")) 1086 continue; 1087 1088 /* Do not follow top-level source and build symlinks */ 1089 if (depth == 0) { 1090 if (!strcmp(dent->d_name, "source") || 1091 !strcmp(dent->d_name, "build")) 1092 continue; 1093 } 1094 1095 ret = map_groups__set_modules_path_dir(mg, path, 1096 depth + 1); 1097 if (ret < 0) 1098 goto out; 1099 } else { 1100 struct kmod_path m; 1101 1102 ret = kmod_path__parse_name(&m, dent->d_name); 1103 if (ret) 1104 goto out; 1105 1106 if (m.kmod) 1107 ret = map_groups__set_module_path(mg, path, &m); 1108 1109 free(m.name); 1110 1111 if (ret) 1112 goto out; 1113 } 1114 } 1115 1116 out: 1117 closedir(dir); 1118 return ret; 1119 } 1120 1121 static int machine__set_modules_path(struct machine *machine) 1122 { 1123 char *version; 1124 char modules_path[PATH_MAX]; 1125 1126 version = get_kernel_version(machine->root_dir); 1127 if (!version) 1128 return -1; 1129 1130 snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s", 1131 machine->root_dir, version); 1132 free(version); 1133 1134 return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0); 1135 } 1136 int __weak arch__fix_module_text_start(u64 *start __maybe_unused, 1137 const char *name __maybe_unused) 1138 { 1139 return 0; 1140 } 1141 1142 static int machine__create_module(void *arg, const char *name, u64 start) 1143 { 1144 struct machine *machine = arg; 1145 struct map *map; 1146 1147 if (arch__fix_module_text_start(&start, name) < 0) 1148 return -1; 1149 1150 map = machine__findnew_module_map(machine, start, name); 1151 if (map == NULL) 1152 return -1; 1153 1154 dso__kernel_module_get_build_id(map->dso, machine->root_dir); 1155 1156 return 0; 1157 } 1158 1159 static int machine__create_modules(struct machine *machine) 1160 { 1161 const char *modules; 1162 char path[PATH_MAX]; 1163 1164 if (machine__is_default_guest(machine)) { 1165 modules = symbol_conf.default_guest_modules; 1166 } else { 1167 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir); 1168 modules = path; 1169 } 1170 1171 if (symbol__restricted_filename(modules, "/proc/modules")) 1172 return -1; 1173 1174 if (modules__parse(modules, machine, machine__create_module)) 1175 return -1; 1176 1177 if (!machine__set_modules_path(machine)) 1178 return 0; 1179 1180 pr_debug("Problems setting modules path maps, continuing anyway...\n"); 1181 1182 return 0; 1183 } 1184 1185 int machine__create_kernel_maps(struct machine *machine) 1186 { 1187 struct dso *kernel = machine__get_kernel(machine); 1188 const char *name; 1189 u64 addr; 1190 int ret; 1191 1192 if (kernel == NULL) 1193 return -1; 1194 1195 ret = __machine__create_kernel_maps(machine, kernel); 1196 dso__put(kernel); 1197 if (ret < 0) 1198 return -1; 1199 1200 if (symbol_conf.use_modules && machine__create_modules(machine) < 0) { 1201 if (machine__is_host(machine)) 1202 pr_debug("Problems creating module maps, " 1203 "continuing anyway...\n"); 1204 else 1205 pr_debug("Problems creating module maps for guest %d, " 1206 "continuing anyway...\n", machine->pid); 1207 } 1208 1209 /* 1210 * Now that we have all the maps created, just set the ->end of them: 1211 */ 1212 map_groups__fixup_end(&machine->kmaps); 1213 1214 addr = machine__get_running_kernel_start(machine, &name); 1215 if (!addr) { 1216 } else if (maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name, addr)) { 1217 machine__destroy_kernel_maps(machine); 1218 return -1; 1219 } 1220 1221 return 0; 1222 } 1223 1224 static void machine__set_kernel_mmap_len(struct machine *machine, 1225 union perf_event *event) 1226 { 1227 int i; 1228 1229 for (i = 0; i < MAP__NR_TYPES; i++) { 1230 machine->vmlinux_maps[i]->start = event->mmap.start; 1231 machine->vmlinux_maps[i]->end = (event->mmap.start + 1232 event->mmap.len); 1233 /* 1234 * Be a bit paranoid here, some perf.data file came with 1235 * a zero sized synthesized MMAP event for the kernel. 1236 */ 1237 if (machine->vmlinux_maps[i]->end == 0) 1238 machine->vmlinux_maps[i]->end = ~0ULL; 1239 } 1240 } 1241 1242 static bool machine__uses_kcore(struct machine *machine) 1243 { 1244 struct dso *dso; 1245 1246 list_for_each_entry(dso, &machine->dsos.head, node) { 1247 if (dso__is_kcore(dso)) 1248 return true; 1249 } 1250 1251 return false; 1252 } 1253 1254 static int machine__process_kernel_mmap_event(struct machine *machine, 1255 union perf_event *event) 1256 { 1257 struct map *map; 1258 char kmmap_prefix[PATH_MAX]; 1259 enum dso_kernel_type kernel_type; 1260 bool is_kernel_mmap; 1261 1262 /* If we have maps from kcore then we do not need or want any others */ 1263 if (machine__uses_kcore(machine)) 1264 return 0; 1265 1266 machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix)); 1267 if (machine__is_host(machine)) 1268 kernel_type = DSO_TYPE_KERNEL; 1269 else 1270 kernel_type = DSO_TYPE_GUEST_KERNEL; 1271 1272 is_kernel_mmap = memcmp(event->mmap.filename, 1273 kmmap_prefix, 1274 strlen(kmmap_prefix) - 1) == 0; 1275 if (event->mmap.filename[0] == '/' || 1276 (!is_kernel_mmap && event->mmap.filename[0] == '[')) { 1277 map = machine__findnew_module_map(machine, event->mmap.start, 1278 event->mmap.filename); 1279 if (map == NULL) 1280 goto out_problem; 1281 1282 map->end = map->start + event->mmap.len; 1283 } else if (is_kernel_mmap) { 1284 const char *symbol_name = (event->mmap.filename + 1285 strlen(kmmap_prefix)); 1286 /* 1287 * Should be there already, from the build-id table in 1288 * the header. 1289 */ 1290 struct dso *kernel = NULL; 1291 struct dso *dso; 1292 1293 pthread_rwlock_rdlock(&machine->dsos.lock); 1294 1295 list_for_each_entry(dso, &machine->dsos.head, node) { 1296 1297 /* 1298 * The cpumode passed to is_kernel_module is not the 1299 * cpumode of *this* event. If we insist on passing 1300 * correct cpumode to is_kernel_module, we should 1301 * record the cpumode when we adding this dso to the 1302 * linked list. 1303 * 1304 * However we don't really need passing correct 1305 * cpumode. We know the correct cpumode must be kernel 1306 * mode (if not, we should not link it onto kernel_dsos 1307 * list). 1308 * 1309 * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN. 1310 * is_kernel_module() treats it as a kernel cpumode. 1311 */ 1312 1313 if (!dso->kernel || 1314 is_kernel_module(dso->long_name, 1315 PERF_RECORD_MISC_CPUMODE_UNKNOWN)) 1316 continue; 1317 1318 1319 kernel = dso; 1320 break; 1321 } 1322 1323 pthread_rwlock_unlock(&machine->dsos.lock); 1324 1325 if (kernel == NULL) 1326 kernel = machine__findnew_dso(machine, kmmap_prefix); 1327 if (kernel == NULL) 1328 goto out_problem; 1329 1330 kernel->kernel = kernel_type; 1331 if (__machine__create_kernel_maps(machine, kernel) < 0) { 1332 dso__put(kernel); 1333 goto out_problem; 1334 } 1335 1336 if (strstr(kernel->long_name, "vmlinux")) 1337 dso__set_short_name(kernel, "[kernel.vmlinux]", false); 1338 1339 machine__set_kernel_mmap_len(machine, event); 1340 1341 /* 1342 * Avoid using a zero address (kptr_restrict) for the ref reloc 1343 * symbol. Effectively having zero here means that at record 1344 * time /proc/sys/kernel/kptr_restrict was non zero. 1345 */ 1346 if (event->mmap.pgoff != 0) { 1347 maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, 1348 symbol_name, 1349 event->mmap.pgoff); 1350 } 1351 1352 if (machine__is_default_guest(machine)) { 1353 /* 1354 * preload dso of guest kernel and modules 1355 */ 1356 dso__load(kernel, machine__kernel_map(machine)); 1357 } 1358 } 1359 return 0; 1360 out_problem: 1361 return -1; 1362 } 1363 1364 int machine__process_mmap2_event(struct machine *machine, 1365 union perf_event *event, 1366 struct perf_sample *sample) 1367 { 1368 struct thread *thread; 1369 struct map *map; 1370 enum map_type type; 1371 int ret = 0; 1372 1373 if (dump_trace) 1374 perf_event__fprintf_mmap2(event, stdout); 1375 1376 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL || 1377 sample->cpumode == PERF_RECORD_MISC_KERNEL) { 1378 ret = machine__process_kernel_mmap_event(machine, event); 1379 if (ret < 0) 1380 goto out_problem; 1381 return 0; 1382 } 1383 1384 thread = machine__findnew_thread(machine, event->mmap2.pid, 1385 event->mmap2.tid); 1386 if (thread == NULL) 1387 goto out_problem; 1388 1389 if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) 1390 type = MAP__VARIABLE; 1391 else 1392 type = MAP__FUNCTION; 1393 1394 map = map__new(machine, event->mmap2.start, 1395 event->mmap2.len, event->mmap2.pgoff, 1396 event->mmap2.pid, event->mmap2.maj, 1397 event->mmap2.min, event->mmap2.ino, 1398 event->mmap2.ino_generation, 1399 event->mmap2.prot, 1400 event->mmap2.flags, 1401 event->mmap2.filename, type, thread); 1402 1403 if (map == NULL) 1404 goto out_problem_map; 1405 1406 ret = thread__insert_map(thread, map); 1407 if (ret) 1408 goto out_problem_insert; 1409 1410 thread__put(thread); 1411 map__put(map); 1412 return 0; 1413 1414 out_problem_insert: 1415 map__put(map); 1416 out_problem_map: 1417 thread__put(thread); 1418 out_problem: 1419 dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n"); 1420 return 0; 1421 } 1422 1423 int machine__process_mmap_event(struct machine *machine, union perf_event *event, 1424 struct perf_sample *sample) 1425 { 1426 struct thread *thread; 1427 struct map *map; 1428 enum map_type type; 1429 int ret = 0; 1430 1431 if (dump_trace) 1432 perf_event__fprintf_mmap(event, stdout); 1433 1434 if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL || 1435 sample->cpumode == PERF_RECORD_MISC_KERNEL) { 1436 ret = machine__process_kernel_mmap_event(machine, event); 1437 if (ret < 0) 1438 goto out_problem; 1439 return 0; 1440 } 1441 1442 thread = machine__findnew_thread(machine, event->mmap.pid, 1443 event->mmap.tid); 1444 if (thread == NULL) 1445 goto out_problem; 1446 1447 if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) 1448 type = MAP__VARIABLE; 1449 else 1450 type = MAP__FUNCTION; 1451 1452 map = map__new(machine, event->mmap.start, 1453 event->mmap.len, event->mmap.pgoff, 1454 event->mmap.pid, 0, 0, 0, 0, 0, 0, 1455 event->mmap.filename, 1456 type, thread); 1457 1458 if (map == NULL) 1459 goto out_problem_map; 1460 1461 ret = thread__insert_map(thread, map); 1462 if (ret) 1463 goto out_problem_insert; 1464 1465 thread__put(thread); 1466 map__put(map); 1467 return 0; 1468 1469 out_problem_insert: 1470 map__put(map); 1471 out_problem_map: 1472 thread__put(thread); 1473 out_problem: 1474 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n"); 1475 return 0; 1476 } 1477 1478 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock) 1479 { 1480 if (machine->last_match == th) 1481 machine->last_match = NULL; 1482 1483 BUG_ON(refcount_read(&th->refcnt) == 0); 1484 if (lock) 1485 pthread_rwlock_wrlock(&machine->threads_lock); 1486 rb_erase_init(&th->rb_node, &machine->threads); 1487 RB_CLEAR_NODE(&th->rb_node); 1488 --machine->nr_threads; 1489 /* 1490 * Move it first to the dead_threads list, then drop the reference, 1491 * if this is the last reference, then the thread__delete destructor 1492 * will be called and we will remove it from the dead_threads list. 1493 */ 1494 list_add_tail(&th->node, &machine->dead_threads); 1495 if (lock) 1496 pthread_rwlock_unlock(&machine->threads_lock); 1497 thread__put(th); 1498 } 1499 1500 void machine__remove_thread(struct machine *machine, struct thread *th) 1501 { 1502 return __machine__remove_thread(machine, th, true); 1503 } 1504 1505 int machine__process_fork_event(struct machine *machine, union perf_event *event, 1506 struct perf_sample *sample) 1507 { 1508 struct thread *thread = machine__find_thread(machine, 1509 event->fork.pid, 1510 event->fork.tid); 1511 struct thread *parent = machine__findnew_thread(machine, 1512 event->fork.ppid, 1513 event->fork.ptid); 1514 int err = 0; 1515 1516 if (dump_trace) 1517 perf_event__fprintf_task(event, stdout); 1518 1519 /* 1520 * There may be an existing thread that is not actually the parent, 1521 * either because we are processing events out of order, or because the 1522 * (fork) event that would have removed the thread was lost. Assume the 1523 * latter case and continue on as best we can. 1524 */ 1525 if (parent->pid_ != (pid_t)event->fork.ppid) { 1526 dump_printf("removing erroneous parent thread %d/%d\n", 1527 parent->pid_, parent->tid); 1528 machine__remove_thread(machine, parent); 1529 thread__put(parent); 1530 parent = machine__findnew_thread(machine, event->fork.ppid, 1531 event->fork.ptid); 1532 } 1533 1534 /* if a thread currently exists for the thread id remove it */ 1535 if (thread != NULL) { 1536 machine__remove_thread(machine, thread); 1537 thread__put(thread); 1538 } 1539 1540 thread = machine__findnew_thread(machine, event->fork.pid, 1541 event->fork.tid); 1542 1543 if (thread == NULL || parent == NULL || 1544 thread__fork(thread, parent, sample->time) < 0) { 1545 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n"); 1546 err = -1; 1547 } 1548 thread__put(thread); 1549 thread__put(parent); 1550 1551 return err; 1552 } 1553 1554 int machine__process_exit_event(struct machine *machine, union perf_event *event, 1555 struct perf_sample *sample __maybe_unused) 1556 { 1557 struct thread *thread = machine__find_thread(machine, 1558 event->fork.pid, 1559 event->fork.tid); 1560 1561 if (dump_trace) 1562 perf_event__fprintf_task(event, stdout); 1563 1564 if (thread != NULL) { 1565 thread__exited(thread); 1566 thread__put(thread); 1567 } 1568 1569 return 0; 1570 } 1571 1572 int machine__process_event(struct machine *machine, union perf_event *event, 1573 struct perf_sample *sample) 1574 { 1575 int ret; 1576 1577 switch (event->header.type) { 1578 case PERF_RECORD_COMM: 1579 ret = machine__process_comm_event(machine, event, sample); break; 1580 case PERF_RECORD_MMAP: 1581 ret = machine__process_mmap_event(machine, event, sample); break; 1582 case PERF_RECORD_NAMESPACES: 1583 ret = machine__process_namespaces_event(machine, event, sample); break; 1584 case PERF_RECORD_MMAP2: 1585 ret = machine__process_mmap2_event(machine, event, sample); break; 1586 case PERF_RECORD_FORK: 1587 ret = machine__process_fork_event(machine, event, sample); break; 1588 case PERF_RECORD_EXIT: 1589 ret = machine__process_exit_event(machine, event, sample); break; 1590 case PERF_RECORD_LOST: 1591 ret = machine__process_lost_event(machine, event, sample); break; 1592 case PERF_RECORD_AUX: 1593 ret = machine__process_aux_event(machine, event); break; 1594 case PERF_RECORD_ITRACE_START: 1595 ret = machine__process_itrace_start_event(machine, event); break; 1596 case PERF_RECORD_LOST_SAMPLES: 1597 ret = machine__process_lost_samples_event(machine, event, sample); break; 1598 case PERF_RECORD_SWITCH: 1599 case PERF_RECORD_SWITCH_CPU_WIDE: 1600 ret = machine__process_switch_event(machine, event); break; 1601 default: 1602 ret = -1; 1603 break; 1604 } 1605 1606 return ret; 1607 } 1608 1609 static bool symbol__match_regex(struct symbol *sym, regex_t *regex) 1610 { 1611 if (!regexec(regex, sym->name, 0, NULL, 0)) 1612 return 1; 1613 return 0; 1614 } 1615 1616 static void ip__resolve_ams(struct thread *thread, 1617 struct addr_map_symbol *ams, 1618 u64 ip) 1619 { 1620 struct addr_location al; 1621 1622 memset(&al, 0, sizeof(al)); 1623 /* 1624 * We cannot use the header.misc hint to determine whether a 1625 * branch stack address is user, kernel, guest, hypervisor. 1626 * Branches may straddle the kernel/user/hypervisor boundaries. 1627 * Thus, we have to try consecutively until we find a match 1628 * or else, the symbol is unknown 1629 */ 1630 thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al); 1631 1632 ams->addr = ip; 1633 ams->al_addr = al.addr; 1634 ams->sym = al.sym; 1635 ams->map = al.map; 1636 } 1637 1638 static void ip__resolve_data(struct thread *thread, 1639 u8 m, struct addr_map_symbol *ams, u64 addr) 1640 { 1641 struct addr_location al; 1642 1643 memset(&al, 0, sizeof(al)); 1644 1645 thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al); 1646 if (al.map == NULL) { 1647 /* 1648 * some shared data regions have execute bit set which puts 1649 * their mapping in the MAP__FUNCTION type array. 1650 * Check there as a fallback option before dropping the sample. 1651 */ 1652 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al); 1653 } 1654 1655 ams->addr = addr; 1656 ams->al_addr = al.addr; 1657 ams->sym = al.sym; 1658 ams->map = al.map; 1659 } 1660 1661 struct mem_info *sample__resolve_mem(struct perf_sample *sample, 1662 struct addr_location *al) 1663 { 1664 struct mem_info *mi = zalloc(sizeof(*mi)); 1665 1666 if (!mi) 1667 return NULL; 1668 1669 ip__resolve_ams(al->thread, &mi->iaddr, sample->ip); 1670 ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr); 1671 mi->data_src.val = sample->data_src; 1672 1673 return mi; 1674 } 1675 1676 static int add_callchain_ip(struct thread *thread, 1677 struct callchain_cursor *cursor, 1678 struct symbol **parent, 1679 struct addr_location *root_al, 1680 u8 *cpumode, 1681 u64 ip, 1682 bool branch, 1683 struct branch_flags *flags, 1684 int nr_loop_iter, 1685 int samples) 1686 { 1687 struct addr_location al; 1688 1689 al.filtered = 0; 1690 al.sym = NULL; 1691 if (!cpumode) { 1692 thread__find_cpumode_addr_location(thread, MAP__FUNCTION, 1693 ip, &al); 1694 } else { 1695 if (ip >= PERF_CONTEXT_MAX) { 1696 switch (ip) { 1697 case PERF_CONTEXT_HV: 1698 *cpumode = PERF_RECORD_MISC_HYPERVISOR; 1699 break; 1700 case PERF_CONTEXT_KERNEL: 1701 *cpumode = PERF_RECORD_MISC_KERNEL; 1702 break; 1703 case PERF_CONTEXT_USER: 1704 *cpumode = PERF_RECORD_MISC_USER; 1705 break; 1706 default: 1707 pr_debug("invalid callchain context: " 1708 "%"PRId64"\n", (s64) ip); 1709 /* 1710 * It seems the callchain is corrupted. 1711 * Discard all. 1712 */ 1713 callchain_cursor_reset(cursor); 1714 return 1; 1715 } 1716 return 0; 1717 } 1718 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION, 1719 ip, &al); 1720 } 1721 1722 if (al.sym != NULL) { 1723 if (perf_hpp_list.parent && !*parent && 1724 symbol__match_regex(al.sym, &parent_regex)) 1725 *parent = al.sym; 1726 else if (have_ignore_callees && root_al && 1727 symbol__match_regex(al.sym, &ignore_callees_regex)) { 1728 /* Treat this symbol as the root, 1729 forgetting its callees. */ 1730 *root_al = al; 1731 callchain_cursor_reset(cursor); 1732 } 1733 } 1734 1735 if (symbol_conf.hide_unresolved && al.sym == NULL) 1736 return 0; 1737 return callchain_cursor_append(cursor, al.addr, al.map, al.sym, 1738 branch, flags, nr_loop_iter, samples); 1739 } 1740 1741 struct branch_info *sample__resolve_bstack(struct perf_sample *sample, 1742 struct addr_location *al) 1743 { 1744 unsigned int i; 1745 const struct branch_stack *bs = sample->branch_stack; 1746 struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info)); 1747 1748 if (!bi) 1749 return NULL; 1750 1751 for (i = 0; i < bs->nr; i++) { 1752 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to); 1753 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from); 1754 bi[i].flags = bs->entries[i].flags; 1755 } 1756 return bi; 1757 } 1758 1759 #define CHASHSZ 127 1760 #define CHASHBITS 7 1761 #define NO_ENTRY 0xff 1762 1763 #define PERF_MAX_BRANCH_DEPTH 127 1764 1765 /* Remove loops. */ 1766 static int remove_loops(struct branch_entry *l, int nr) 1767 { 1768 int i, j, off; 1769 unsigned char chash[CHASHSZ]; 1770 1771 memset(chash, NO_ENTRY, sizeof(chash)); 1772 1773 BUG_ON(PERF_MAX_BRANCH_DEPTH > 255); 1774 1775 for (i = 0; i < nr; i++) { 1776 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ; 1777 1778 /* no collision handling for now */ 1779 if (chash[h] == NO_ENTRY) { 1780 chash[h] = i; 1781 } else if (l[chash[h]].from == l[i].from) { 1782 bool is_loop = true; 1783 /* check if it is a real loop */ 1784 off = 0; 1785 for (j = chash[h]; j < i && i + off < nr; j++, off++) 1786 if (l[j].from != l[i + off].from) { 1787 is_loop = false; 1788 break; 1789 } 1790 if (is_loop) { 1791 memmove(l + i, l + i + off, 1792 (nr - (i + off)) * sizeof(*l)); 1793 nr -= off; 1794 } 1795 } 1796 } 1797 return nr; 1798 } 1799 1800 /* 1801 * Recolve LBR callstack chain sample 1802 * Return: 1803 * 1 on success get LBR callchain information 1804 * 0 no available LBR callchain information, should try fp 1805 * negative error code on other errors. 1806 */ 1807 static int resolve_lbr_callchain_sample(struct thread *thread, 1808 struct callchain_cursor *cursor, 1809 struct perf_sample *sample, 1810 struct symbol **parent, 1811 struct addr_location *root_al, 1812 int max_stack) 1813 { 1814 struct ip_callchain *chain = sample->callchain; 1815 int chain_nr = min(max_stack, (int)chain->nr), i; 1816 u8 cpumode = PERF_RECORD_MISC_USER; 1817 u64 ip; 1818 1819 for (i = 0; i < chain_nr; i++) { 1820 if (chain->ips[i] == PERF_CONTEXT_USER) 1821 break; 1822 } 1823 1824 /* LBR only affects the user callchain */ 1825 if (i != chain_nr) { 1826 struct branch_stack *lbr_stack = sample->branch_stack; 1827 int lbr_nr = lbr_stack->nr, j, k; 1828 bool branch; 1829 struct branch_flags *flags; 1830 /* 1831 * LBR callstack can only get user call chain. 1832 * The mix_chain_nr is kernel call chain 1833 * number plus LBR user call chain number. 1834 * i is kernel call chain number, 1835 * 1 is PERF_CONTEXT_USER, 1836 * lbr_nr + 1 is the user call chain number. 1837 * For details, please refer to the comments 1838 * in callchain__printf 1839 */ 1840 int mix_chain_nr = i + 1 + lbr_nr + 1; 1841 1842 for (j = 0; j < mix_chain_nr; j++) { 1843 int err; 1844 branch = false; 1845 flags = NULL; 1846 1847 if (callchain_param.order == ORDER_CALLEE) { 1848 if (j < i + 1) 1849 ip = chain->ips[j]; 1850 else if (j > i + 1) { 1851 k = j - i - 2; 1852 ip = lbr_stack->entries[k].from; 1853 branch = true; 1854 flags = &lbr_stack->entries[k].flags; 1855 } else { 1856 ip = lbr_stack->entries[0].to; 1857 branch = true; 1858 flags = &lbr_stack->entries[0].flags; 1859 } 1860 } else { 1861 if (j < lbr_nr) { 1862 k = lbr_nr - j - 1; 1863 ip = lbr_stack->entries[k].from; 1864 branch = true; 1865 flags = &lbr_stack->entries[k].flags; 1866 } 1867 else if (j > lbr_nr) 1868 ip = chain->ips[i + 1 - (j - lbr_nr)]; 1869 else { 1870 ip = lbr_stack->entries[0].to; 1871 branch = true; 1872 flags = &lbr_stack->entries[0].flags; 1873 } 1874 } 1875 1876 err = add_callchain_ip(thread, cursor, parent, 1877 root_al, &cpumode, ip, 1878 branch, flags, 0, 0); 1879 if (err) 1880 return (err < 0) ? err : 0; 1881 } 1882 return 1; 1883 } 1884 1885 return 0; 1886 } 1887 1888 static int thread__resolve_callchain_sample(struct thread *thread, 1889 struct callchain_cursor *cursor, 1890 struct perf_evsel *evsel, 1891 struct perf_sample *sample, 1892 struct symbol **parent, 1893 struct addr_location *root_al, 1894 int max_stack) 1895 { 1896 struct branch_stack *branch = sample->branch_stack; 1897 struct ip_callchain *chain = sample->callchain; 1898 int chain_nr = chain->nr; 1899 u8 cpumode = PERF_RECORD_MISC_USER; 1900 int i, j, err, nr_entries; 1901 int skip_idx = -1; 1902 int first_call = 0; 1903 int nr_loop_iter; 1904 1905 if (perf_evsel__has_branch_callstack(evsel)) { 1906 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent, 1907 root_al, max_stack); 1908 if (err) 1909 return (err < 0) ? err : 0; 1910 } 1911 1912 /* 1913 * Based on DWARF debug information, some architectures skip 1914 * a callchain entry saved by the kernel. 1915 */ 1916 skip_idx = arch_skip_callchain_idx(thread, chain); 1917 1918 /* 1919 * Add branches to call stack for easier browsing. This gives 1920 * more context for a sample than just the callers. 1921 * 1922 * This uses individual histograms of paths compared to the 1923 * aggregated histograms the normal LBR mode uses. 1924 * 1925 * Limitations for now: 1926 * - No extra filters 1927 * - No annotations (should annotate somehow) 1928 */ 1929 1930 if (branch && callchain_param.branch_callstack) { 1931 int nr = min(max_stack, (int)branch->nr); 1932 struct branch_entry be[nr]; 1933 1934 if (branch->nr > PERF_MAX_BRANCH_DEPTH) { 1935 pr_warning("corrupted branch chain. skipping...\n"); 1936 goto check_calls; 1937 } 1938 1939 for (i = 0; i < nr; i++) { 1940 if (callchain_param.order == ORDER_CALLEE) { 1941 be[i] = branch->entries[i]; 1942 /* 1943 * Check for overlap into the callchain. 1944 * The return address is one off compared to 1945 * the branch entry. To adjust for this 1946 * assume the calling instruction is not longer 1947 * than 8 bytes. 1948 */ 1949 if (i == skip_idx || 1950 chain->ips[first_call] >= PERF_CONTEXT_MAX) 1951 first_call++; 1952 else if (be[i].from < chain->ips[first_call] && 1953 be[i].from >= chain->ips[first_call] - 8) 1954 first_call++; 1955 } else 1956 be[i] = branch->entries[branch->nr - i - 1]; 1957 } 1958 1959 nr_loop_iter = nr; 1960 nr = remove_loops(be, nr); 1961 1962 /* 1963 * Get the number of iterations. 1964 * It's only approximation, but good enough in practice. 1965 */ 1966 if (nr_loop_iter > nr) 1967 nr_loop_iter = nr_loop_iter - nr + 1; 1968 else 1969 nr_loop_iter = 0; 1970 1971 for (i = 0; i < nr; i++) { 1972 if (i == nr - 1) 1973 err = add_callchain_ip(thread, cursor, parent, 1974 root_al, 1975 NULL, be[i].to, 1976 true, &be[i].flags, 1977 nr_loop_iter, 1); 1978 else 1979 err = add_callchain_ip(thread, cursor, parent, 1980 root_al, 1981 NULL, be[i].to, 1982 true, &be[i].flags, 1983 0, 0); 1984 1985 if (!err) 1986 err = add_callchain_ip(thread, cursor, parent, root_al, 1987 NULL, be[i].from, 1988 true, &be[i].flags, 1989 0, 0); 1990 if (err == -EINVAL) 1991 break; 1992 if (err) 1993 return err; 1994 } 1995 chain_nr -= nr; 1996 } 1997 1998 check_calls: 1999 for (i = first_call, nr_entries = 0; 2000 i < chain_nr && nr_entries < max_stack; i++) { 2001 u64 ip; 2002 2003 if (callchain_param.order == ORDER_CALLEE) 2004 j = i; 2005 else 2006 j = chain->nr - i - 1; 2007 2008 #ifdef HAVE_SKIP_CALLCHAIN_IDX 2009 if (j == skip_idx) 2010 continue; 2011 #endif 2012 ip = chain->ips[j]; 2013 2014 if (ip < PERF_CONTEXT_MAX) 2015 ++nr_entries; 2016 2017 err = add_callchain_ip(thread, cursor, parent, 2018 root_al, &cpumode, ip, 2019 false, NULL, 0, 0); 2020 2021 if (err) 2022 return (err < 0) ? err : 0; 2023 } 2024 2025 return 0; 2026 } 2027 2028 static int unwind_entry(struct unwind_entry *entry, void *arg) 2029 { 2030 struct callchain_cursor *cursor = arg; 2031 2032 if (symbol_conf.hide_unresolved && entry->sym == NULL) 2033 return 0; 2034 return callchain_cursor_append(cursor, entry->ip, 2035 entry->map, entry->sym, 2036 false, NULL, 0, 0); 2037 } 2038 2039 static int thread__resolve_callchain_unwind(struct thread *thread, 2040 struct callchain_cursor *cursor, 2041 struct perf_evsel *evsel, 2042 struct perf_sample *sample, 2043 int max_stack) 2044 { 2045 /* Can we do dwarf post unwind? */ 2046 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) && 2047 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER))) 2048 return 0; 2049 2050 /* Bail out if nothing was captured. */ 2051 if ((!sample->user_regs.regs) || 2052 (!sample->user_stack.size)) 2053 return 0; 2054 2055 return unwind__get_entries(unwind_entry, cursor, 2056 thread, sample, max_stack); 2057 } 2058 2059 int thread__resolve_callchain(struct thread *thread, 2060 struct callchain_cursor *cursor, 2061 struct perf_evsel *evsel, 2062 struct perf_sample *sample, 2063 struct symbol **parent, 2064 struct addr_location *root_al, 2065 int max_stack) 2066 { 2067 int ret = 0; 2068 2069 callchain_cursor_reset(&callchain_cursor); 2070 2071 if (callchain_param.order == ORDER_CALLEE) { 2072 ret = thread__resolve_callchain_sample(thread, cursor, 2073 evsel, sample, 2074 parent, root_al, 2075 max_stack); 2076 if (ret) 2077 return ret; 2078 ret = thread__resolve_callchain_unwind(thread, cursor, 2079 evsel, sample, 2080 max_stack); 2081 } else { 2082 ret = thread__resolve_callchain_unwind(thread, cursor, 2083 evsel, sample, 2084 max_stack); 2085 if (ret) 2086 return ret; 2087 ret = thread__resolve_callchain_sample(thread, cursor, 2088 evsel, sample, 2089 parent, root_al, 2090 max_stack); 2091 } 2092 2093 return ret; 2094 } 2095 2096 int machine__for_each_thread(struct machine *machine, 2097 int (*fn)(struct thread *thread, void *p), 2098 void *priv) 2099 { 2100 struct rb_node *nd; 2101 struct thread *thread; 2102 int rc = 0; 2103 2104 for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) { 2105 thread = rb_entry(nd, struct thread, rb_node); 2106 rc = fn(thread, priv); 2107 if (rc != 0) 2108 return rc; 2109 } 2110 2111 list_for_each_entry(thread, &machine->dead_threads, node) { 2112 rc = fn(thread, priv); 2113 if (rc != 0) 2114 return rc; 2115 } 2116 return rc; 2117 } 2118 2119 int machines__for_each_thread(struct machines *machines, 2120 int (*fn)(struct thread *thread, void *p), 2121 void *priv) 2122 { 2123 struct rb_node *nd; 2124 int rc = 0; 2125 2126 rc = machine__for_each_thread(&machines->host, fn, priv); 2127 if (rc != 0) 2128 return rc; 2129 2130 for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) { 2131 struct machine *machine = rb_entry(nd, struct machine, rb_node); 2132 2133 rc = machine__for_each_thread(machine, fn, priv); 2134 if (rc != 0) 2135 return rc; 2136 } 2137 return rc; 2138 } 2139 2140 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 2141 struct target *target, struct thread_map *threads, 2142 perf_event__handler_t process, bool data_mmap, 2143 unsigned int proc_map_timeout) 2144 { 2145 if (target__has_task(target)) 2146 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap, proc_map_timeout); 2147 else if (target__has_cpu(target)) 2148 return perf_event__synthesize_threads(tool, process, machine, data_mmap, proc_map_timeout); 2149 /* command specified */ 2150 return 0; 2151 } 2152 2153 pid_t machine__get_current_tid(struct machine *machine, int cpu) 2154 { 2155 if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid) 2156 return -1; 2157 2158 return machine->current_tid[cpu]; 2159 } 2160 2161 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid, 2162 pid_t tid) 2163 { 2164 struct thread *thread; 2165 2166 if (cpu < 0) 2167 return -EINVAL; 2168 2169 if (!machine->current_tid) { 2170 int i; 2171 2172 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t)); 2173 if (!machine->current_tid) 2174 return -ENOMEM; 2175 for (i = 0; i < MAX_NR_CPUS; i++) 2176 machine->current_tid[i] = -1; 2177 } 2178 2179 if (cpu >= MAX_NR_CPUS) { 2180 pr_err("Requested CPU %d too large. ", cpu); 2181 pr_err("Consider raising MAX_NR_CPUS\n"); 2182 return -EINVAL; 2183 } 2184 2185 machine->current_tid[cpu] = tid; 2186 2187 thread = machine__findnew_thread(machine, pid, tid); 2188 if (!thread) 2189 return -ENOMEM; 2190 2191 thread->cpu = cpu; 2192 thread__put(thread); 2193 2194 return 0; 2195 } 2196 2197 int machine__get_kernel_start(struct machine *machine) 2198 { 2199 struct map *map = machine__kernel_map(machine); 2200 int err = 0; 2201 2202 /* 2203 * The only addresses above 2^63 are kernel addresses of a 64-bit 2204 * kernel. Note that addresses are unsigned so that on a 32-bit system 2205 * all addresses including kernel addresses are less than 2^32. In 2206 * that case (32-bit system), if the kernel mapping is unknown, all 2207 * addresses will be assumed to be in user space - see 2208 * machine__kernel_ip(). 2209 */ 2210 machine->kernel_start = 1ULL << 63; 2211 if (map) { 2212 err = map__load(map); 2213 if (map->start) 2214 machine->kernel_start = map->start; 2215 } 2216 return err; 2217 } 2218 2219 struct dso *machine__findnew_dso(struct machine *machine, const char *filename) 2220 { 2221 return dsos__findnew(&machine->dsos, filename); 2222 } 2223 2224 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp) 2225 { 2226 struct machine *machine = vmachine; 2227 struct map *map; 2228 struct symbol *sym = map_groups__find_symbol(&machine->kmaps, MAP__FUNCTION, *addrp, &map); 2229 2230 if (sym == NULL) 2231 return NULL; 2232 2233 *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL; 2234 *addrp = map->unmap_ip(map, sym->start); 2235 return sym->name; 2236 } 2237