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