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