tools/perf/util/machine.c

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