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