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