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