tools/perf/util/machine.c

   1 #include "callchain.h"
   2 #include "debug.h"
   3 #include "event.h"
   4 #include "evsel.h"
   5 #include "hist.h"
   6 #include "machine.h"
   7 #include "map.h"
   8 #include "sort.h"
   9 #include "strlist.h"
  10 #include "thread.h"
  11 #include "vdso.h"
  12 #include <stdbool.h>
  13 #include <symbol/kallsyms.h>
  14 #include "unwind.h"
  15 #include "linux/hash.h"
  16
  17 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  18
  19 static void dsos__init(struct dsos *dsos)
  20 {
  21         INIT_LIST_HEAD(&dsos->head);
  22         dsos->root = RB_ROOT;
  23 }
  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         pthread_rwlock_init(&machine->threads_lock, NULL);
  34         INIT_LIST_HEAD(&machine->dead_threads);
  35         machine->last_match = NULL;
  36
  37         machine->vdso_info = NULL;
  38
  39         machine->pid = pid;
  40
  41         machine->symbol_filter = NULL;
  42         machine->id_hdr_size = 0;
  43         machine->comm_exec = false;
  44         machine->kernel_start = 0;
  45
  46         machine->root_dir = strdup(root_dir);
  47         if (machine->root_dir == NULL)
  48                 return -ENOMEM;
  49
  50         if (pid != HOST_KERNEL_ID) {
  51                 struct thread *thread = machine__findnew_thread(machine, -1,
  52                                                                 pid);
  53                 char comm[64];
  54
  55                 if (thread == NULL)
  56                         return -ENOMEM;
  57
  58                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
  59                 thread__set_comm(thread, comm, 0);
  60                 thread__put(thread);
  61         }
  62
  63         machine->current_tid = NULL;
  64
  65         return 0;
  66 }
  67
  68 struct machine *machine__new_host(void)
  69 {
  70         struct machine *machine = malloc(sizeof(*machine));
  71
  72         if (machine != NULL) {
  73                 machine__init(machine, "", HOST_KERNEL_ID);
  74
  75                 if (machine__create_kernel_maps(machine) < 0)
  76                         goto out_delete;
  77         }
  78
  79         return machine;
  80 out_delete:
  81         free(machine);
  82         return NULL;
  83 }
  84
  85 static void dsos__delete(struct dsos *dsos)
  86 {
  87         struct dso *pos, *n;
  88
  89         list_for_each_entry_safe(pos, n, &dsos->head, node) {
  90                 RB_CLEAR_NODE(&pos->rb_node);
  91                 list_del(&pos->node);
  92                 dso__delete(pos);
  93         }
  94 }
  95
  96 void machine__delete_threads(struct machine *machine)
  97 {
  98         struct rb_node *nd;
  99
 100         pthread_rwlock_wrlock(&machine->threads_lock);
 101         nd = rb_first(&machine->threads);
 102         while (nd) {
 103                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 104
 105                 nd = rb_next(nd);
 106                 __machine__remove_thread(machine, t, false);
 107         }
 108         pthread_rwlock_unlock(&machine->threads_lock);
 109 }
 110
 111 void machine__exit(struct machine *machine)
 112 {
 113         map_groups__exit(&machine->kmaps);
 114         dsos__delete(&machine->user_dsos);
 115         dsos__delete(&machine->kernel_dsos);
 116         vdso__exit(machine);
 117         zfree(&machine->root_dir);
 118         zfree(&machine->current_tid);
 119         pthread_rwlock_destroy(&machine->threads_lock);
 120 }
 121
 122 void machine__delete(struct machine *machine)
 123 {
 124         machine__exit(machine);
 125         free(machine);
 126 }
 127
 128 void machines__init(struct machines *machines)
 129 {
 130         machine__init(&machines->host, "", HOST_KERNEL_ID);
 131         machines->guests = RB_ROOT;
 132         machines->symbol_filter = NULL;
 133 }
 134
 135 void machines__exit(struct machines *machines)
 136 {
 137         machine__exit(&machines->host);
 138         /* XXX exit guest */
 139 }
 140
 141 struct machine *machines__add(struct machines *machines, pid_t pid,
 142                               const char *root_dir)
 143 {
 144         struct rb_node **p = &machines->guests.rb_node;
 145         struct rb_node *parent = NULL;
 146         struct machine *pos, *machine = malloc(sizeof(*machine));
 147
 148         if (machine == NULL)
 149                 return NULL;
 150
 151         if (machine__init(machine, root_dir, pid) != 0) {
 152                 free(machine);
 153                 return NULL;
 154         }
 155
 156         machine->symbol_filter = machines->symbol_filter;
 157
 158         while (*p != NULL) {
 159                 parent = *p;
 160                 pos = rb_entry(parent, struct machine, rb_node);
 161                 if (pid < pos->pid)
 162                         p = &(*p)->rb_left;
 163                 else
 164                         p = &(*p)->rb_right;
 165         }
 166
 167         rb_link_node(&machine->rb_node, parent, p);
 168         rb_insert_color(&machine->rb_node, &machines->guests);
 169
 170         return machine;
 171 }
 172
 173 void machines__set_symbol_filter(struct machines *machines,
 174                                  symbol_filter_t symbol_filter)
 175 {
 176         struct rb_node *nd;
 177
 178         machines->symbol_filter = symbol_filter;
 179         machines->host.symbol_filter = symbol_filter;
 180
 181         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 182                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 183
 184                 machine->symbol_filter = symbol_filter;
 185         }
 186 }
 187
 188 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 189 {
 190         struct rb_node *nd;
 191
 192         machines->host.comm_exec = comm_exec;
 193
 194         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 195                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 196
 197                 machine->comm_exec = comm_exec;
 198         }
 199 }
 200
 201 struct machine *machines__find(struct machines *machines, pid_t pid)
 202 {
 203         struct rb_node **p = &machines->guests.rb_node;
 204         struct rb_node *parent = NULL;
 205         struct machine *machine;
 206         struct machine *default_machine = NULL;
 207
 208         if (pid == HOST_KERNEL_ID)
 209                 return &machines->host;
 210
 211         while (*p != NULL) {
 212                 parent = *p;
 213                 machine = rb_entry(parent, struct machine, rb_node);
 214                 if (pid < machine->pid)
 215                         p = &(*p)->rb_left;
 216                 else if (pid > machine->pid)
 217                         p = &(*p)->rb_right;
 218                 else
 219                         return machine;
 220                 if (!machine->pid)
 221                         default_machine = machine;
 222         }
 223
 224         return default_machine;
 225 }
 226
 227 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 228 {
 229         char path[PATH_MAX];
 230         const char *root_dir = "";
 231         struct machine *machine = machines__find(machines, pid);
 232
 233         if (machine && (machine->pid == pid))
 234                 goto out;
 235
 236         if ((pid != HOST_KERNEL_ID) &&
 237             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 238             (symbol_conf.guestmount)) {
 239                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 240                 if (access(path, R_OK)) {
 241                         static struct strlist *seen;
 242
 243                         if (!seen)
 244                                 seen = strlist__new(true, NULL);
 245
 246                         if (!strlist__has_entry(seen, path)) {
 247                                 pr_err("Can't access file %s\n", path);
 248                                 strlist__add(seen, path);
 249                         }
 250                         machine = NULL;
 251                         goto out;
 252                 }
 253                 root_dir = path;
 254         }
 255
 256         machine = machines__add(machines, pid, root_dir);
 257 out:
 258         return machine;
 259 }
 260
 261 void machines__process_guests(struct machines *machines,
 262                               machine__process_t process, void *data)
 263 {
 264         struct rb_node *nd;
 265
 266         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 267                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 268                 process(pos, data);
 269         }
 270 }
 271
 272 char *machine__mmap_name(struct machine *machine, char *bf, size_t size)
 273 {
 274         if (machine__is_host(machine))
 275                 snprintf(bf, size, "[%s]", "kernel.kallsyms");
 276         else if (machine__is_default_guest(machine))
 277                 snprintf(bf, size, "[%s]", "guest.kernel.kallsyms");
 278         else {
 279                 snprintf(bf, size, "[%s.%d]", "guest.kernel.kallsyms",
 280                          machine->pid);
 281         }
 282
 283         return bf;
 284 }
 285
 286 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 287 {
 288         struct rb_node *node;
 289         struct machine *machine;
 290
 291         machines->host.id_hdr_size = id_hdr_size;
 292
 293         for (node = rb_first(&machines->guests); node; node = rb_next(node)) {
 294                 machine = rb_entry(node, struct machine, rb_node);
 295                 machine->id_hdr_size = id_hdr_size;
 296         }
 297
 298         return;
 299 }
 300
 301 static void machine__update_thread_pid(struct machine *machine,
 302                                        struct thread *th, pid_t pid)
 303 {
 304         struct thread *leader;
 305
 306         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 307                 return;
 308
 309         th->pid_ = pid;
 310
 311         if (th->pid_ == th->tid)
 312                 return;
 313
 314         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 315         if (!leader)
 316                 goto out_err;
 317
 318         if (!leader->mg)
 319                 leader->mg = map_groups__new(machine);
 320
 321         if (!leader->mg)
 322                 goto out_err;
 323
 324         if (th->mg == leader->mg)
 325                 return;
 326
 327         if (th->mg) {
 328                 /*
 329                  * Maps are created from MMAP events which provide the pid and
 330                  * tid.  Consequently there never should be any maps on a thread
 331                  * with an unknown pid.  Just print an error if there are.
 332                  */
 333                 if (!map_groups__empty(th->mg))
 334                         pr_err("Discarding thread maps for %d:%d\n",
 335                                th->pid_, th->tid);
 336                 map_groups__delete(th->mg);
 337         }
 338
 339         th->mg = map_groups__get(leader->mg);
 340
 341         return;
 342
 343 out_err:
 344         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 345 }
 346
 347 static struct thread *____machine__findnew_thread(struct machine *machine,
 348                                                   pid_t pid, pid_t tid,
 349                                                   bool create)
 350 {
 351         struct rb_node **p = &machine->threads.rb_node;
 352         struct rb_node *parent = NULL;
 353         struct thread *th;
 354
 355         /*
 356          * Front-end cache - TID lookups come in blocks,
 357          * so most of the time we dont have to look up
 358          * the full rbtree:
 359          */
 360         th = machine->last_match;
 361         if (th != NULL) {
 362                 if (th->tid == tid) {
 363                         machine__update_thread_pid(machine, th, pid);
 364                         return th;
 365                 }
 366
 367                 machine->last_match = NULL;
 368         }
 369
 370         while (*p != NULL) {
 371                 parent = *p;
 372                 th = rb_entry(parent, struct thread, rb_node);
 373
 374                 if (th->tid == tid) {
 375                         machine->last_match = th;
 376                         machine__update_thread_pid(machine, th, pid);
 377                         return th;
 378                 }
 379
 380                 if (tid < th->tid)
 381                         p = &(*p)->rb_left;
 382                 else
 383                         p = &(*p)->rb_right;
 384         }
 385
 386         if (!create)
 387                 return NULL;
 388
 389         th = thread__new(pid, tid);
 390         if (th != NULL) {
 391                 rb_link_node(&th->rb_node, parent, p);
 392                 rb_insert_color(&th->rb_node, &machine->threads);
 393
 394                 /*
 395                  * We have to initialize map_groups separately
 396                  * after rb tree is updated.
 397                  *
 398                  * The reason is that we call machine__findnew_thread
 399                  * within thread__init_map_groups to find the thread
 400                  * leader and that would screwed the rb tree.
 401                  */
 402                 if (thread__init_map_groups(th, machine)) {
 403                         rb_erase(&th->rb_node, &machine->threads);
 404                         RB_CLEAR_NODE(&th->rb_node);
 405                         thread__delete(th);
 406                         return NULL;
 407                 }
 408                 /*
 409                  * It is now in the rbtree, get a ref
 410                  */
 411                 thread__get(th);
 412                 machine->last_match = th;
 413         }
 414
 415         return th;
 416 }
 417
 418 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 419 {
 420         return ____machine__findnew_thread(machine, pid, tid, true);
 421 }
 422
 423 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 424                                        pid_t tid)
 425 {
 426         struct thread *th;
 427
 428         pthread_rwlock_wrlock(&machine->threads_lock);
 429         th = thread__get(__machine__findnew_thread(machine, pid, tid));
 430         pthread_rwlock_unlock(&machine->threads_lock);
 431         return th;
 432 }
 433
 434 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 435                                     pid_t tid)
 436 {
 437         struct thread *th;
 438         pthread_rwlock_rdlock(&machine->threads_lock);
 439         th =  thread__get(____machine__findnew_thread(machine, pid, tid, false));
 440         pthread_rwlock_unlock(&machine->threads_lock);
 441         return th;
 442 }
 443
 444 struct comm *machine__thread_exec_comm(struct machine *machine,
 445                                        struct thread *thread)
 446 {
 447         if (machine->comm_exec)
 448                 return thread__exec_comm(thread);
 449         else
 450                 return thread__comm(thread);
 451 }
 452
 453 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 454                                 struct perf_sample *sample)
 455 {
 456         struct thread *thread = machine__findnew_thread(machine,
 457                                                         event->comm.pid,
 458                                                         event->comm.tid);
 459         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 460         int err = 0;
 461
 462         if (exec)
 463                 machine->comm_exec = true;
 464
 465         if (dump_trace)
 466                 perf_event__fprintf_comm(event, stdout);
 467
 468         if (thread == NULL ||
 469             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 470                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 471                 err = -1;
 472         }
 473
 474         thread__put(thread);
 475
 476         return err;
 477 }
 478
 479 int machine__process_lost_event(struct machine *machine __maybe_unused,
 480                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 481 {
 482         dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
 483                     event->lost.id, event->lost.lost);
 484         return 0;
 485 }
 486
 487 static struct dso*
 488 machine__module_dso(struct machine *machine, struct kmod_path *m,
 489                     const char *filename)
 490 {
 491         struct dso *dso;
 492
 493         dso = dsos__find(&machine->kernel_dsos, m->name, true);
 494         if (!dso) {
 495                 dso = dsos__addnew(&machine->kernel_dsos, m->name);
 496                 if (dso == NULL)
 497                         return NULL;
 498
 499                 if (machine__is_host(machine))
 500                         dso->symtab_type = DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE;
 501                 else
 502                         dso->symtab_type = DSO_BINARY_TYPE__GUEST_KMODULE;
 503
 504                 /* _KMODULE_COMP should be next to _KMODULE */
 505                 if (m->kmod && m->comp)
 506                         dso->symtab_type++;
 507
 508                 dso__set_short_name(dso, strdup(m->name), true);
 509                 dso__set_long_name(dso, strdup(filename), true);
 510         }
 511
 512         return dso;
 513 }
 514
 515 int machine__process_aux_event(struct machine *machine __maybe_unused,
 516                                union perf_event *event)
 517 {
 518         if (dump_trace)
 519                 perf_event__fprintf_aux(event, stdout);
 520         return 0;
 521 }
 522
 523 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 524                                         union perf_event *event)
 525 {
 526         if (dump_trace)
 527                 perf_event__fprintf_itrace_start(event, stdout);
 528         return 0;
 529 }
 530
 531 struct map *machine__new_module(struct machine *machine, u64 start,
 532                                 const char *filename)
 533 {
 534         struct map *map = NULL;
 535         struct dso *dso;
 536         struct kmod_path m;
 537
 538         if (kmod_path__parse_name(&m, filename))
 539                 return NULL;
 540
 541         map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION,
 542                                        m.name);
 543         if (map)
 544                 goto out;
 545
 546         dso = machine__module_dso(machine, &m, filename);
 547         if (dso == NULL)
 548                 goto out;
 549
 550         map = map__new2(start, dso, MAP__FUNCTION);
 551         if (map == NULL)
 552                 goto out;
 553
 554         map_groups__insert(&machine->kmaps, map);
 555
 556 out:
 557         free(m.name);
 558         return map;
 559 }
 560
 561 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 562 {
 563         struct rb_node *nd;
 564         size_t ret = __dsos__fprintf(&machines->host.kernel_dsos.head, fp) +
 565                      __dsos__fprintf(&machines->host.user_dsos.head, fp);
 566
 567         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 568                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 569                 ret += __dsos__fprintf(&pos->kernel_dsos.head, fp);
 570                 ret += __dsos__fprintf(&pos->user_dsos.head, fp);
 571         }
 572
 573         return ret;
 574 }
 575
 576 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 577                                      bool (skip)(struct dso *dso, int parm), int parm)
 578 {
 579         return __dsos__fprintf_buildid(&m->kernel_dsos.head, fp, skip, parm) +
 580                __dsos__fprintf_buildid(&m->user_dsos.head, fp, skip, parm);
 581 }
 582
 583 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 584                                      bool (skip)(struct dso *dso, int parm), int parm)
 585 {
 586         struct rb_node *nd;
 587         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 588
 589         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 590                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 591                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 592         }
 593         return ret;
 594 }
 595
 596 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 597 {
 598         int i;
 599         size_t printed = 0;
 600         struct dso *kdso = machine->vmlinux_maps[MAP__FUNCTION]->dso;
 601
 602         if (kdso->has_build_id) {
 603                 char filename[PATH_MAX];
 604                 if (dso__build_id_filename(kdso, filename, sizeof(filename)))
 605                         printed += fprintf(fp, "[0] %s\n", filename);
 606         }
 607
 608         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 609                 printed += fprintf(fp, "[%d] %s\n",
 610                                    i + kdso->has_build_id, vmlinux_path[i]);
 611
 612         return printed;
 613 }
 614
 615 size_t machine__fprintf(struct machine *machine, FILE *fp)
 616 {
 617         size_t ret = 0;
 618         struct rb_node *nd;
 619
 620         pthread_rwlock_rdlock(&machine->threads_lock);
 621
 622         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
 623                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 624
 625                 ret += thread__fprintf(pos, fp);
 626         }
 627
 628         pthread_rwlock_unlock(&machine->threads_lock);
 629
 630         return ret;
 631 }
 632
 633 static struct dso *machine__get_kernel(struct machine *machine)
 634 {
 635         const char *vmlinux_name = NULL;
 636         struct dso *kernel;
 637
 638         if (machine__is_host(machine)) {
 639                 vmlinux_name = symbol_conf.vmlinux_name;
 640                 if (!vmlinux_name)
 641                         vmlinux_name = "[kernel.kallsyms]";
 642
 643                 kernel = dso__kernel_findnew(machine, vmlinux_name,
 644                                              "[kernel]",
 645                                              DSO_TYPE_KERNEL);
 646         } else {
 647                 char bf[PATH_MAX];
 648
 649                 if (machine__is_default_guest(machine))
 650                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 651                 if (!vmlinux_name)
 652                         vmlinux_name = machine__mmap_name(machine, bf,
 653                                                           sizeof(bf));
 654
 655                 kernel = dso__kernel_findnew(machine, vmlinux_name,
 656                                              "[guest.kernel]",
 657                                              DSO_TYPE_GUEST_KERNEL);
 658         }
 659
 660         if (kernel != NULL && (!kernel->has_build_id))
 661                 dso__read_running_kernel_build_id(kernel, machine);
 662
 663         return kernel;
 664 }
 665
 666 struct process_args {
 667         u64 start;
 668 };
 669
 670 static void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 671                                            size_t bufsz)
 672 {
 673         if (machine__is_default_guest(machine))
 674                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 675         else
 676                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 677 }
 678
 679 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 680
 681 /* Figure out the start address of kernel map from /proc/kallsyms.
 682  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 683  * symbol_name if it's not that important.
 684  */
 685 static u64 machine__get_running_kernel_start(struct machine *machine,
 686                                              const char **symbol_name)
 687 {
 688         char filename[PATH_MAX];
 689         int i;
 690         const char *name;
 691         u64 addr = 0;
 692
 693         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 694
 695         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 696                 return 0;
 697
 698         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 699                 addr = kallsyms__get_function_start(filename, name);
 700                 if (addr)
 701                         break;
 702         }
 703
 704         if (symbol_name)
 705                 *symbol_name = name;
 706
 707         return addr;
 708 }
 709
 710 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
 711 {
 712         enum map_type type;
 713         u64 start = machine__get_running_kernel_start(machine, NULL);
 714
 715         for (type = 0; type < MAP__NR_TYPES; ++type) {
 716                 struct kmap *kmap;
 717
 718                 machine->vmlinux_maps[type] = map__new2(start, kernel, type);
 719                 if (machine->vmlinux_maps[type] == NULL)
 720                         return -1;
 721
 722                 machine->vmlinux_maps[type]->map_ip =
 723                         machine->vmlinux_maps[type]->unmap_ip =
 724                                 identity__map_ip;
 725                 kmap = map__kmap(machine->vmlinux_maps[type]);
 726                 if (!kmap)
 727                         return -1;
 728
 729                 kmap->kmaps = &machine->kmaps;
 730                 map_groups__insert(&machine->kmaps,
 731                                    machine->vmlinux_maps[type]);
 732         }
 733
 734         return 0;
 735 }
 736
 737 void machine__destroy_kernel_maps(struct machine *machine)
 738 {
 739         enum map_type type;
 740
 741         for (type = 0; type < MAP__NR_TYPES; ++type) {
 742                 struct kmap *kmap;
 743
 744                 if (machine->vmlinux_maps[type] == NULL)
 745                         continue;
 746
 747                 kmap = map__kmap(machine->vmlinux_maps[type]);
 748                 map_groups__remove(&machine->kmaps,
 749                                    machine->vmlinux_maps[type]);
 750                 if (kmap && kmap->ref_reloc_sym) {
 751                         /*
 752                          * ref_reloc_sym is shared among all maps, so free just
 753                          * on one of them.
 754                          */
 755                         if (type == MAP__FUNCTION) {
 756                                 zfree((char **)&kmap->ref_reloc_sym->name);
 757                                 zfree(&kmap->ref_reloc_sym);
 758                         } else
 759                                 kmap->ref_reloc_sym = NULL;
 760                 }
 761
 762                 map__delete(machine->vmlinux_maps[type]);
 763                 machine->vmlinux_maps[type] = NULL;
 764         }
 765 }
 766
 767 int machines__create_guest_kernel_maps(struct machines *machines)
 768 {
 769         int ret = 0;
 770         struct dirent **namelist = NULL;
 771         int i, items = 0;
 772         char path[PATH_MAX];
 773         pid_t pid;
 774         char *endp;
 775
 776         if (symbol_conf.default_guest_vmlinux_name ||
 777             symbol_conf.default_guest_modules ||
 778             symbol_conf.default_guest_kallsyms) {
 779                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
 780         }
 781
 782         if (symbol_conf.guestmount) {
 783                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
 784                 if (items <= 0)
 785                         return -ENOENT;
 786                 for (i = 0; i < items; i++) {
 787                         if (!isdigit(namelist[i]->d_name[0])) {
 788                                 /* Filter out . and .. */
 789                                 continue;
 790                         }
 791                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
 792                         if ((*endp != '\0') ||
 793                             (endp == namelist[i]->d_name) ||
 794                             (errno == ERANGE)) {
 795                                 pr_debug("invalid directory (%s). Skipping.\n",
 796                                          namelist[i]->d_name);
 797                                 continue;
 798                         }
 799                         sprintf(path, "%s/%s/proc/kallsyms",
 800                                 symbol_conf.guestmount,
 801                                 namelist[i]->d_name);
 802                         ret = access(path, R_OK);
 803                         if (ret) {
 804                                 pr_debug("Can't access file %s\n", path);
 805                                 goto failure;
 806                         }
 807                         machines__create_kernel_maps(machines, pid);
 808                 }
 809 failure:
 810                 free(namelist);
 811         }
 812
 813         return ret;
 814 }
 815
 816 void machines__destroy_kernel_maps(struct machines *machines)
 817 {
 818         struct rb_node *next = rb_first(&machines->guests);
 819
 820         machine__destroy_kernel_maps(&machines->host);
 821
 822         while (next) {
 823                 struct machine *pos = rb_entry(next, struct machine, rb_node);
 824
 825                 next = rb_next(&pos->rb_node);
 826                 rb_erase(&pos->rb_node, &machines->guests);
 827                 machine__delete(pos);
 828         }
 829 }
 830
 831 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
 832 {
 833         struct machine *machine = machines__findnew(machines, pid);
 834
 835         if (machine == NULL)
 836                 return -1;
 837
 838         return machine__create_kernel_maps(machine);
 839 }
 840
 841 int machine__load_kallsyms(struct machine *machine, const char *filename,
 842                            enum map_type type, symbol_filter_t filter)
 843 {
 844         struct map *map = machine->vmlinux_maps[type];
 845         int ret = dso__load_kallsyms(map->dso, filename, map, filter);
 846
 847         if (ret > 0) {
 848                 dso__set_loaded(map->dso, type);
 849                 /*
 850                  * Since /proc/kallsyms will have multiple sessions for the
 851                  * kernel, with modules between them, fixup the end of all
 852                  * sections.
 853                  */
 854                 __map_groups__fixup_end(&machine->kmaps, type);
 855         }
 856
 857         return ret;
 858 }
 859
 860 int machine__load_vmlinux_path(struct machine *machine, enum map_type type,
 861                                symbol_filter_t filter)
 862 {
 863         struct map *map = machine->vmlinux_maps[type];
 864         int ret = dso__load_vmlinux_path(map->dso, map, filter);
 865
 866         if (ret > 0)
 867                 dso__set_loaded(map->dso, type);
 868
 869         return ret;
 870 }
 871
 872 static void map_groups__fixup_end(struct map_groups *mg)
 873 {
 874         int i;
 875         for (i = 0; i < MAP__NR_TYPES; ++i)
 876                 __map_groups__fixup_end(mg, i);
 877 }
 878
 879 static char *get_kernel_version(const char *root_dir)
 880 {
 881         char version[PATH_MAX];
 882         FILE *file;
 883         char *name, *tmp;
 884         const char *prefix = "Linux version ";
 885
 886         sprintf(version, "%s/proc/version", root_dir);
 887         file = fopen(version, "r");
 888         if (!file)
 889                 return NULL;
 890
 891         version[0] = '\0';
 892         tmp = fgets(version, sizeof(version), file);
 893         fclose(file);
 894
 895         name = strstr(version, prefix);
 896         if (!name)
 897                 return NULL;
 898         name += strlen(prefix);
 899         tmp = strchr(name, ' ');
 900         if (tmp)
 901                 *tmp = '\0';
 902
 903         return strdup(name);
 904 }
 905
 906 static bool is_kmod_dso(struct dso *dso)
 907 {
 908         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
 909                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
 910 }
 911
 912 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
 913                                        struct kmod_path *m)
 914 {
 915         struct map *map;
 916         char *long_name;
 917
 918         map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name);
 919         if (map == NULL)
 920                 return 0;
 921
 922         long_name = strdup(path);
 923         if (long_name == NULL)
 924                 return -ENOMEM;
 925
 926         dso__set_long_name(map->dso, long_name, true);
 927         dso__kernel_module_get_build_id(map->dso, "");
 928
 929         /*
 930          * Full name could reveal us kmod compression, so
 931          * we need to update the symtab_type if needed.
 932          */
 933         if (m->comp && is_kmod_dso(map->dso))
 934                 map->dso->symtab_type++;
 935
 936         return 0;
 937 }
 938
 939 static int map_groups__set_modules_path_dir(struct map_groups *mg,
 940                                 const char *dir_name, int depth)
 941 {
 942         struct dirent *dent;
 943         DIR *dir = opendir(dir_name);
 944         int ret = 0;
 945
 946         if (!dir) {
 947                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
 948                 return -1;
 949         }
 950
 951         while ((dent = readdir(dir)) != NULL) {
 952                 char path[PATH_MAX];
 953                 struct stat st;
 954
 955                 /*sshfs might return bad dent->d_type, so we have to stat*/
 956                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
 957                 if (stat(path, &st))
 958                         continue;
 959
 960                 if (S_ISDIR(st.st_mode)) {
 961                         if (!strcmp(dent->d_name, ".") ||
 962                             !strcmp(dent->d_name, ".."))
 963                                 continue;
 964
 965                         /* Do not follow top-level source and build symlinks */
 966                         if (depth == 0) {
 967                                 if (!strcmp(dent->d_name, "source") ||
 968                                     !strcmp(dent->d_name, "build"))
 969                                         continue;
 970                         }
 971
 972                         ret = map_groups__set_modules_path_dir(mg, path,
 973                                                                depth + 1);
 974                         if (ret < 0)
 975                                 goto out;
 976                 } else {
 977                         struct kmod_path m;
 978
 979                         ret = kmod_path__parse_name(&m, dent->d_name);
 980                         if (ret)
 981                                 goto out;
 982
 983                         if (m.kmod)
 984                                 ret = map_groups__set_module_path(mg, path, &m);
 985
 986                         free(m.name);
 987
 988                         if (ret)
 989                                 goto out;
 990                 }
 991         }
 992
 993 out:
 994         closedir(dir);
 995         return ret;
 996 }
 997
 998 static int machine__set_modules_path(struct machine *machine)
 999 {
1000         char *version;
1001         char modules_path[PATH_MAX];
1002
1003         version = get_kernel_version(machine->root_dir);
1004         if (!version)
1005                 return -1;
1006
1007         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1008                  machine->root_dir, version);
1009         free(version);
1010
1011         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1012 }
1013
1014 static int machine__create_module(void *arg, const char *name, u64 start)
1015 {
1016         struct machine *machine = arg;
1017         struct map *map;
1018
1019         map = machine__new_module(machine, start, name);
1020         if (map == NULL)
1021                 return -1;
1022
1023         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1024
1025         return 0;
1026 }
1027
1028 static int machine__create_modules(struct machine *machine)
1029 {
1030         const char *modules;
1031         char path[PATH_MAX];
1032
1033         if (machine__is_default_guest(machine)) {
1034                 modules = symbol_conf.default_guest_modules;
1035         } else {
1036                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1037                 modules = path;
1038         }
1039
1040         if (symbol__restricted_filename(modules, "/proc/modules"))
1041                 return -1;
1042
1043         if (modules__parse(modules, machine, machine__create_module))
1044                 return -1;
1045
1046         if (!machine__set_modules_path(machine))
1047                 return 0;
1048
1049         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1050
1051         return 0;
1052 }
1053
1054 int machine__create_kernel_maps(struct machine *machine)
1055 {
1056         struct dso *kernel = machine__get_kernel(machine);
1057         const char *name;
1058         u64 addr = machine__get_running_kernel_start(machine, &name);
1059         if (!addr)
1060                 return -1;
1061
1062         if (kernel == NULL ||
1063             __machine__create_kernel_maps(machine, kernel) < 0)
1064                 return -1;
1065
1066         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1067                 if (machine__is_host(machine))
1068                         pr_debug("Problems creating module maps, "
1069                                  "continuing anyway...\n");
1070                 else
1071                         pr_debug("Problems creating module maps for guest %d, "
1072                                  "continuing anyway...\n", machine->pid);
1073         }
1074
1075         /*
1076          * Now that we have all the maps created, just set the ->end of them:
1077          */
1078         map_groups__fixup_end(&machine->kmaps);
1079
1080         if (maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name,
1081                                              addr)) {
1082                 machine__destroy_kernel_maps(machine);
1083                 return -1;
1084         }
1085
1086         return 0;
1087 }
1088
1089 static void machine__set_kernel_mmap_len(struct machine *machine,
1090                                          union perf_event *event)
1091 {
1092         int i;
1093
1094         for (i = 0; i < MAP__NR_TYPES; i++) {
1095                 machine->vmlinux_maps[i]->start = event->mmap.start;
1096                 machine->vmlinux_maps[i]->end   = (event->mmap.start +
1097                                                    event->mmap.len);
1098                 /*
1099                  * Be a bit paranoid here, some perf.data file came with
1100                  * a zero sized synthesized MMAP event for the kernel.
1101                  */
1102                 if (machine->vmlinux_maps[i]->end == 0)
1103                         machine->vmlinux_maps[i]->end = ~0ULL;
1104         }
1105 }
1106
1107 static bool machine__uses_kcore(struct machine *machine)
1108 {
1109         struct dso *dso;
1110
1111         list_for_each_entry(dso, &machine->kernel_dsos.head, node) {
1112                 if (dso__is_kcore(dso))
1113                         return true;
1114         }
1115
1116         return false;
1117 }
1118
1119 static int machine__process_kernel_mmap_event(struct machine *machine,
1120                                               union perf_event *event)
1121 {
1122         struct map *map;
1123         char kmmap_prefix[PATH_MAX];
1124         enum dso_kernel_type kernel_type;
1125         bool is_kernel_mmap;
1126
1127         /* If we have maps from kcore then we do not need or want any others */
1128         if (machine__uses_kcore(machine))
1129                 return 0;
1130
1131         machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix));
1132         if (machine__is_host(machine))
1133                 kernel_type = DSO_TYPE_KERNEL;
1134         else
1135                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1136
1137         is_kernel_mmap = memcmp(event->mmap.filename,
1138                                 kmmap_prefix,
1139                                 strlen(kmmap_prefix) - 1) == 0;
1140         if (event->mmap.filename[0] == '/' ||
1141             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1142                 map = machine__new_module(machine, event->mmap.start,
1143                                           event->mmap.filename);
1144                 if (map == NULL)
1145                         goto out_problem;
1146
1147                 map->end = map->start + event->mmap.len;
1148         } else if (is_kernel_mmap) {
1149                 const char *symbol_name = (event->mmap.filename +
1150                                 strlen(kmmap_prefix));
1151                 /*
1152                  * Should be there already, from the build-id table in
1153                  * the header.
1154                  */
1155                 struct dso *kernel = NULL;
1156                 struct dso *dso;
1157
1158                 list_for_each_entry(dso, &machine->kernel_dsos.head, node) {
1159                         if (is_kernel_module(dso->long_name))
1160                                 continue;
1161
1162                         kernel = dso;
1163                         break;
1164                 }
1165
1166                 if (kernel == NULL)
1167                         kernel = __dsos__findnew(&machine->kernel_dsos,
1168                                                  kmmap_prefix);
1169                 if (kernel == NULL)
1170                         goto out_problem;
1171
1172                 kernel->kernel = kernel_type;
1173                 if (__machine__create_kernel_maps(machine, kernel) < 0)
1174                         goto out_problem;
1175
1176                 if (strstr(kernel->long_name, "vmlinux"))
1177                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1178
1179                 machine__set_kernel_mmap_len(machine, event);
1180
1181                 /*
1182                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1183                  * symbol. Effectively having zero here means that at record
1184                  * time /proc/sys/kernel/kptr_restrict was non zero.
1185                  */
1186                 if (event->mmap.pgoff != 0) {
1187                         maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps,
1188                                                          symbol_name,
1189                                                          event->mmap.pgoff);
1190                 }
1191
1192                 if (machine__is_default_guest(machine)) {
1193                         /*
1194                          * preload dso of guest kernel and modules
1195                          */
1196                         dso__load(kernel, machine->vmlinux_maps[MAP__FUNCTION],
1197                                   NULL);
1198                 }
1199         }
1200         return 0;
1201 out_problem:
1202         return -1;
1203 }
1204
1205 int machine__process_mmap2_event(struct machine *machine,
1206                                  union perf_event *event,
1207                                  struct perf_sample *sample __maybe_unused)
1208 {
1209         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1210         struct thread *thread;
1211         struct map *map;
1212         enum map_type type;
1213         int ret = 0;
1214
1215         if (dump_trace)
1216                 perf_event__fprintf_mmap2(event, stdout);
1217
1218         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1219             cpumode == PERF_RECORD_MISC_KERNEL) {
1220                 ret = machine__process_kernel_mmap_event(machine, event);
1221                 if (ret < 0)
1222                         goto out_problem;
1223                 return 0;
1224         }
1225
1226         thread = machine__findnew_thread(machine, event->mmap2.pid,
1227                                         event->mmap2.tid);
1228         if (thread == NULL)
1229                 goto out_problem;
1230
1231         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1232                 type = MAP__VARIABLE;
1233         else
1234                 type = MAP__FUNCTION;
1235
1236         map = map__new(machine, event->mmap2.start,
1237                         event->mmap2.len, event->mmap2.pgoff,
1238                         event->mmap2.pid, event->mmap2.maj,
1239                         event->mmap2.min, event->mmap2.ino,
1240                         event->mmap2.ino_generation,
1241                         event->mmap2.prot,
1242                         event->mmap2.flags,
1243                         event->mmap2.filename, type, thread);
1244
1245         if (map == NULL)
1246                 goto out_problem_map;
1247
1248         thread__insert_map(thread, map);
1249         thread__put(thread);
1250         return 0;
1251
1252 out_problem_map:
1253         thread__put(thread);
1254 out_problem:
1255         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1256         return 0;
1257 }
1258
1259 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1260                                 struct perf_sample *sample __maybe_unused)
1261 {
1262         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1263         struct thread *thread;
1264         struct map *map;
1265         enum map_type type;
1266         int ret = 0;
1267
1268         if (dump_trace)
1269                 perf_event__fprintf_mmap(event, stdout);
1270
1271         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1272             cpumode == PERF_RECORD_MISC_KERNEL) {
1273                 ret = machine__process_kernel_mmap_event(machine, event);
1274                 if (ret < 0)
1275                         goto out_problem;
1276                 return 0;
1277         }
1278
1279         thread = machine__findnew_thread(machine, event->mmap.pid,
1280                                          event->mmap.tid);
1281         if (thread == NULL)
1282                 goto out_problem;
1283
1284         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1285                 type = MAP__VARIABLE;
1286         else
1287                 type = MAP__FUNCTION;
1288
1289         map = map__new(machine, event->mmap.start,
1290                         event->mmap.len, event->mmap.pgoff,
1291                         event->mmap.pid, 0, 0, 0, 0, 0, 0,
1292                         event->mmap.filename,
1293                         type, thread);
1294
1295         if (map == NULL)
1296                 goto out_problem_map;
1297
1298         thread__insert_map(thread, map);
1299         thread__put(thread);
1300         return 0;
1301
1302 out_problem_map:
1303         thread__put(thread);
1304 out_problem:
1305         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1306         return 0;
1307 }
1308
1309 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1310 {
1311         if (machine->last_match == th)
1312                 machine->last_match = NULL;
1313
1314         BUG_ON(th->refcnt.counter == 0);
1315         if (lock)
1316                 pthread_rwlock_wrlock(&machine->threads_lock);
1317         rb_erase(&th->rb_node, &machine->threads);
1318         RB_CLEAR_NODE(&th->rb_node);
1319         /*
1320          * Move it first to the dead_threads list, then drop the reference,
1321          * if this is the last reference, then the thread__delete destructor
1322          * will be called and we will remove it from the dead_threads list.
1323          */
1324         list_add_tail(&th->node, &machine->dead_threads);
1325         if (lock)
1326                 pthread_rwlock_unlock(&machine->threads_lock);
1327         thread__put(th);
1328 }
1329
1330 void machine__remove_thread(struct machine *machine, struct thread *th)
1331 {
1332         return __machine__remove_thread(machine, th, true);
1333 }
1334
1335 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1336                                 struct perf_sample *sample)
1337 {
1338         struct thread *thread = machine__find_thread(machine,
1339                                                      event->fork.pid,
1340                                                      event->fork.tid);
1341         struct thread *parent = machine__findnew_thread(machine,
1342                                                         event->fork.ppid,
1343                                                         event->fork.ptid);
1344         int err = 0;
1345
1346         /* if a thread currently exists for the thread id remove it */
1347         if (thread != NULL) {
1348                 machine__remove_thread(machine, thread);
1349                 thread__put(thread);
1350         }
1351
1352         thread = machine__findnew_thread(machine, event->fork.pid,
1353                                          event->fork.tid);
1354         if (dump_trace)
1355                 perf_event__fprintf_task(event, stdout);
1356
1357         if (thread == NULL || parent == NULL ||
1358             thread__fork(thread, parent, sample->time) < 0) {
1359                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1360                 err = -1;
1361         }
1362         thread__put(thread);
1363         thread__put(parent);
1364
1365         return err;
1366 }
1367
1368 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1369                                 struct perf_sample *sample __maybe_unused)
1370 {
1371         struct thread *thread = machine__find_thread(machine,
1372                                                      event->fork.pid,
1373                                                      event->fork.tid);
1374
1375         if (dump_trace)
1376                 perf_event__fprintf_task(event, stdout);
1377
1378         if (thread != NULL) {
1379                 thread__exited(thread);
1380                 thread__put(thread);
1381         }
1382
1383         return 0;
1384 }
1385
1386 int machine__process_event(struct machine *machine, union perf_event *event,
1387                            struct perf_sample *sample)
1388 {
1389         int ret;
1390
1391         switch (event->header.type) {
1392         case PERF_RECORD_COMM:
1393                 ret = machine__process_comm_event(machine, event, sample); break;
1394         case PERF_RECORD_MMAP:
1395                 ret = machine__process_mmap_event(machine, event, sample); break;
1396         case PERF_RECORD_MMAP2:
1397                 ret = machine__process_mmap2_event(machine, event, sample); break;
1398         case PERF_RECORD_FORK:
1399                 ret = machine__process_fork_event(machine, event, sample); break;
1400         case PERF_RECORD_EXIT:
1401                 ret = machine__process_exit_event(machine, event, sample); break;
1402         case PERF_RECORD_LOST:
1403                 ret = machine__process_lost_event(machine, event, sample); break;
1404         case PERF_RECORD_AUX:
1405                 ret = machine__process_aux_event(machine, event); break;
1406         case PERF_RECORD_ITRACE_START:
1407                 ret = machine__process_itrace_start_event(machine, event);
1408                 break;
1409         default:
1410                 ret = -1;
1411                 break;
1412         }
1413
1414         return ret;
1415 }
1416
1417 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1418 {
1419         if (sym->name && !regexec(regex, sym->name, 0, NULL, 0))
1420                 return 1;
1421         return 0;
1422 }
1423
1424 static void ip__resolve_ams(struct thread *thread,
1425                             struct addr_map_symbol *ams,
1426                             u64 ip)
1427 {
1428         struct addr_location al;
1429
1430         memset(&al, 0, sizeof(al));
1431         /*
1432          * We cannot use the header.misc hint to determine whether a
1433          * branch stack address is user, kernel, guest, hypervisor.
1434          * Branches may straddle the kernel/user/hypervisor boundaries.
1435          * Thus, we have to try consecutively until we find a match
1436          * or else, the symbol is unknown
1437          */
1438         thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al);
1439
1440         ams->addr = ip;
1441         ams->al_addr = al.addr;
1442         ams->sym = al.sym;
1443         ams->map = al.map;
1444 }
1445
1446 static void ip__resolve_data(struct thread *thread,
1447                              u8 m, struct addr_map_symbol *ams, u64 addr)
1448 {
1449         struct addr_location al;
1450
1451         memset(&al, 0, sizeof(al));
1452
1453         thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al);
1454         if (al.map == NULL) {
1455                 /*
1456                  * some shared data regions have execute bit set which puts
1457                  * their mapping in the MAP__FUNCTION type array.
1458                  * Check there as a fallback option before dropping the sample.
1459                  */
1460                 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al);
1461         }
1462
1463         ams->addr = addr;
1464         ams->al_addr = al.addr;
1465         ams->sym = al.sym;
1466         ams->map = al.map;
1467 }
1468
1469 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1470                                      struct addr_location *al)
1471 {
1472         struct mem_info *mi = zalloc(sizeof(*mi));
1473
1474         if (!mi)
1475                 return NULL;
1476
1477         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1478         ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr);
1479         mi->data_src.val = sample->data_src;
1480
1481         return mi;
1482 }
1483
1484 static int add_callchain_ip(struct thread *thread,
1485                             struct symbol **parent,
1486                             struct addr_location *root_al,
1487                             u8 *cpumode,
1488                             u64 ip)
1489 {
1490         struct addr_location al;
1491
1492         al.filtered = 0;
1493         al.sym = NULL;
1494         if (!cpumode) {
1495                 thread__find_cpumode_addr_location(thread, MAP__FUNCTION,
1496                                                    ip, &al);
1497         } else {
1498                 if (ip >= PERF_CONTEXT_MAX) {
1499                         switch (ip) {
1500                         case PERF_CONTEXT_HV:
1501                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
1502                                 break;
1503                         case PERF_CONTEXT_KERNEL:
1504                                 *cpumode = PERF_RECORD_MISC_KERNEL;
1505                                 break;
1506                         case PERF_CONTEXT_USER:
1507                                 *cpumode = PERF_RECORD_MISC_USER;
1508                                 break;
1509                         default:
1510                                 pr_debug("invalid callchain context: "
1511                                          "%"PRId64"\n", (s64) ip);
1512                                 /*
1513                                  * It seems the callchain is corrupted.
1514                                  * Discard all.
1515                                  */
1516                                 callchain_cursor_reset(&callchain_cursor);
1517                                 return 1;
1518                         }
1519                         return 0;
1520                 }
1521                 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION,
1522                                            ip, &al);
1523         }
1524
1525         if (al.sym != NULL) {
1526                 if (sort__has_parent && !*parent &&
1527                     symbol__match_regex(al.sym, &parent_regex))
1528                         *parent = al.sym;
1529                 else if (have_ignore_callees && root_al &&
1530                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
1531                         /* Treat this symbol as the root,
1532                            forgetting its callees. */
1533                         *root_al = al;
1534                         callchain_cursor_reset(&callchain_cursor);
1535                 }
1536         }
1537
1538         return callchain_cursor_append(&callchain_cursor, al.addr, al.map, al.sym);
1539 }
1540
1541 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
1542                                            struct addr_location *al)
1543 {
1544         unsigned int i;
1545         const struct branch_stack *bs = sample->branch_stack;
1546         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
1547
1548         if (!bi)
1549                 return NULL;
1550
1551         for (i = 0; i < bs->nr; i++) {
1552                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
1553                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
1554                 bi[i].flags = bs->entries[i].flags;
1555         }
1556         return bi;
1557 }
1558
1559 #define CHASHSZ 127
1560 #define CHASHBITS 7
1561 #define NO_ENTRY 0xff
1562
1563 #define PERF_MAX_BRANCH_DEPTH 127
1564
1565 /* Remove loops. */
1566 static int remove_loops(struct branch_entry *l, int nr)
1567 {
1568         int i, j, off;
1569         unsigned char chash[CHASHSZ];
1570
1571         memset(chash, NO_ENTRY, sizeof(chash));
1572
1573         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
1574
1575         for (i = 0; i < nr; i++) {
1576                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
1577
1578                 /* no collision handling for now */
1579                 if (chash[h] == NO_ENTRY) {
1580                         chash[h] = i;
1581                 } else if (l[chash[h]].from == l[i].from) {
1582                         bool is_loop = true;
1583                         /* check if it is a real loop */
1584                         off = 0;
1585                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
1586                                 if (l[j].from != l[i + off].from) {
1587                                         is_loop = false;
1588                                         break;
1589                                 }
1590                         if (is_loop) {
1591                                 memmove(l + i, l + i + off,
1592                                         (nr - (i + off)) * sizeof(*l));
1593                                 nr -= off;
1594                         }
1595                 }
1596         }
1597         return nr;
1598 }
1599
1600 /*
1601  * Recolve LBR callstack chain sample
1602  * Return:
1603  * 1 on success get LBR callchain information
1604  * 0 no available LBR callchain information, should try fp
1605  * negative error code on other errors.
1606  */
1607 static int resolve_lbr_callchain_sample(struct thread *thread,
1608                                         struct perf_sample *sample,
1609                                         struct symbol **parent,
1610                                         struct addr_location *root_al,
1611                                         int max_stack)
1612 {
1613         struct ip_callchain *chain = sample->callchain;
1614         int chain_nr = min(max_stack, (int)chain->nr);
1615         u8 cpumode = PERF_RECORD_MISC_USER;
1616         int i, j, err;
1617         u64 ip;
1618
1619         for (i = 0; i < chain_nr; i++) {
1620                 if (chain->ips[i] == PERF_CONTEXT_USER)
1621                         break;
1622         }
1623
1624         /* LBR only affects the user callchain */
1625         if (i != chain_nr) {
1626                 struct branch_stack *lbr_stack = sample->branch_stack;
1627                 int lbr_nr = lbr_stack->nr;
1628                 /*
1629                  * LBR callstack can only get user call chain.
1630                  * The mix_chain_nr is kernel call chain
1631                  * number plus LBR user call chain number.
1632                  * i is kernel call chain number,
1633                  * 1 is PERF_CONTEXT_USER,
1634                  * lbr_nr + 1 is the user call chain number.
1635                  * For details, please refer to the comments
1636                  * in callchain__printf
1637                  */
1638                 int mix_chain_nr = i + 1 + lbr_nr + 1;
1639
1640                 if (mix_chain_nr > PERF_MAX_STACK_DEPTH + PERF_MAX_BRANCH_DEPTH) {
1641                         pr_warning("corrupted callchain. skipping...\n");
1642                         return 0;
1643                 }
1644
1645                 for (j = 0; j < mix_chain_nr; j++) {
1646                         if (callchain_param.order == ORDER_CALLEE) {
1647                                 if (j < i + 1)
1648                                         ip = chain->ips[j];
1649                                 else if (j > i + 1)
1650                                         ip = lbr_stack->entries[j - i - 2].from;
1651                                 else
1652                                         ip = lbr_stack->entries[0].to;
1653                         } else {
1654                                 if (j < lbr_nr)
1655                                         ip = lbr_stack->entries[lbr_nr - j - 1].from;
1656                                 else if (j > lbr_nr)
1657                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
1658                                 else
1659                                         ip = lbr_stack->entries[0].to;
1660                         }
1661
1662                         err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1663                         if (err)
1664                                 return (err < 0) ? err : 0;
1665                 }
1666                 return 1;
1667         }
1668
1669         return 0;
1670 }
1671
1672 static int thread__resolve_callchain_sample(struct thread *thread,
1673                                             struct perf_evsel *evsel,
1674                                             struct perf_sample *sample,
1675                                             struct symbol **parent,
1676                                             struct addr_location *root_al,
1677                                             int max_stack)
1678 {
1679         struct branch_stack *branch = sample->branch_stack;
1680         struct ip_callchain *chain = sample->callchain;
1681         int chain_nr = min(max_stack, (int)chain->nr);
1682         u8 cpumode = PERF_RECORD_MISC_USER;
1683         int i, j, err;
1684         int skip_idx = -1;
1685         int first_call = 0;
1686
1687         callchain_cursor_reset(&callchain_cursor);
1688
1689         if (has_branch_callstack(evsel)) {
1690                 err = resolve_lbr_callchain_sample(thread, sample, parent,
1691                                                    root_al, max_stack);
1692                 if (err)
1693                         return (err < 0) ? err : 0;
1694         }
1695
1696         /*
1697          * Based on DWARF debug information, some architectures skip
1698          * a callchain entry saved by the kernel.
1699          */
1700         if (chain->nr < PERF_MAX_STACK_DEPTH)
1701                 skip_idx = arch_skip_callchain_idx(thread, chain);
1702
1703         /*
1704          * Add branches to call stack for easier browsing. This gives
1705          * more context for a sample than just the callers.
1706          *
1707          * This uses individual histograms of paths compared to the
1708          * aggregated histograms the normal LBR mode uses.
1709          *
1710          * Limitations for now:
1711          * - No extra filters
1712          * - No annotations (should annotate somehow)
1713          */
1714
1715         if (branch && callchain_param.branch_callstack) {
1716                 int nr = min(max_stack, (int)branch->nr);
1717                 struct branch_entry be[nr];
1718
1719                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
1720                         pr_warning("corrupted branch chain. skipping...\n");
1721                         goto check_calls;
1722                 }
1723
1724                 for (i = 0; i < nr; i++) {
1725                         if (callchain_param.order == ORDER_CALLEE) {
1726                                 be[i] = branch->entries[i];
1727                                 /*
1728                                  * Check for overlap into the callchain.
1729                                  * The return address is one off compared to
1730                                  * the branch entry. To adjust for this
1731                                  * assume the calling instruction is not longer
1732                                  * than 8 bytes.
1733                                  */
1734                                 if (i == skip_idx ||
1735                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
1736                                         first_call++;
1737                                 else if (be[i].from < chain->ips[first_call] &&
1738                                     be[i].from >= chain->ips[first_call] - 8)
1739                                         first_call++;
1740                         } else
1741                                 be[i] = branch->entries[branch->nr - i - 1];
1742                 }
1743
1744                 nr = remove_loops(be, nr);
1745
1746                 for (i = 0; i < nr; i++) {
1747                         err = add_callchain_ip(thread, parent, root_al,
1748                                                NULL, be[i].to);
1749                         if (!err)
1750                                 err = add_callchain_ip(thread, parent, root_al,
1751                                                        NULL, be[i].from);
1752                         if (err == -EINVAL)
1753                                 break;
1754                         if (err)
1755                                 return err;
1756                 }
1757                 chain_nr -= nr;
1758         }
1759
1760 check_calls:
1761         if (chain->nr > PERF_MAX_STACK_DEPTH) {
1762                 pr_warning("corrupted callchain. skipping...\n");
1763                 return 0;
1764         }
1765
1766         for (i = first_call; i < chain_nr; i++) {
1767                 u64 ip;
1768
1769                 if (callchain_param.order == ORDER_CALLEE)
1770                         j = i;
1771                 else
1772                         j = chain->nr - i - 1;
1773
1774 #ifdef HAVE_SKIP_CALLCHAIN_IDX
1775                 if (j == skip_idx)
1776                         continue;
1777 #endif
1778                 ip = chain->ips[j];
1779
1780                 err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1781
1782                 if (err)
1783                         return (err < 0) ? err : 0;
1784         }
1785
1786         return 0;
1787 }
1788
1789 static int unwind_entry(struct unwind_entry *entry, void *arg)
1790 {
1791         struct callchain_cursor *cursor = arg;
1792         return callchain_cursor_append(cursor, entry->ip,
1793                                        entry->map, entry->sym);
1794 }
1795
1796 int thread__resolve_callchain(struct thread *thread,
1797                               struct perf_evsel *evsel,
1798                               struct perf_sample *sample,
1799                               struct symbol **parent,
1800                               struct addr_location *root_al,
1801                               int max_stack)
1802 {
1803         int ret = thread__resolve_callchain_sample(thread, evsel,
1804                                                    sample, parent,
1805                                                    root_al, max_stack);
1806         if (ret)
1807                 return ret;
1808
1809         /* Can we do dwarf post unwind? */
1810         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
1811               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
1812                 return 0;
1813
1814         /* Bail out if nothing was captured. */
1815         if ((!sample->user_regs.regs) ||
1816             (!sample->user_stack.size))
1817                 return 0;
1818
1819         return unwind__get_entries(unwind_entry, &callchain_cursor,
1820                                    thread, sample, max_stack);
1821
1822 }
1823
1824 int machine__for_each_thread(struct machine *machine,
1825                              int (*fn)(struct thread *thread, void *p),
1826                              void *priv)
1827 {
1828         struct rb_node *nd;
1829         struct thread *thread;
1830         int rc = 0;
1831
1832         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
1833                 thread = rb_entry(nd, struct thread, rb_node);
1834                 rc = fn(thread, priv);
1835                 if (rc != 0)
1836                         return rc;
1837         }
1838
1839         list_for_each_entry(thread, &machine->dead_threads, node) {
1840                 rc = fn(thread, priv);
1841                 if (rc != 0)
1842                         return rc;
1843         }
1844         return rc;
1845 }
1846
1847 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1848                                   struct target *target, struct thread_map *threads,
1849                                   perf_event__handler_t process, bool data_mmap)
1850 {
1851         if (target__has_task(target))
1852                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
1853         else if (target__has_cpu(target))
1854                 return perf_event__synthesize_threads(tool, process, machine, data_mmap);
1855         /* command specified */
1856         return 0;
1857 }
1858
1859 pid_t machine__get_current_tid(struct machine *machine, int cpu)
1860 {
1861         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
1862                 return -1;
1863
1864         return machine->current_tid[cpu];
1865 }
1866
1867 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
1868                              pid_t tid)
1869 {
1870         struct thread *thread;
1871
1872         if (cpu < 0)
1873                 return -EINVAL;
1874
1875         if (!machine->current_tid) {
1876                 int i;
1877
1878                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
1879                 if (!machine->current_tid)
1880                         return -ENOMEM;
1881                 for (i = 0; i < MAX_NR_CPUS; i++)
1882                         machine->current_tid[i] = -1;
1883         }
1884
1885         if (cpu >= MAX_NR_CPUS) {
1886                 pr_err("Requested CPU %d too large. ", cpu);
1887                 pr_err("Consider raising MAX_NR_CPUS\n");
1888                 return -EINVAL;
1889         }
1890
1891         machine->current_tid[cpu] = tid;
1892
1893         thread = machine__findnew_thread(machine, pid, tid);
1894         if (!thread)
1895                 return -ENOMEM;
1896
1897         thread->cpu = cpu;
1898         thread__put(thread);
1899
1900         return 0;
1901 }
1902
1903 int machine__get_kernel_start(struct machine *machine)
1904 {
1905         struct map *map = machine__kernel_map(machine, MAP__FUNCTION);
1906         int err = 0;
1907
1908         /*
1909          * The only addresses above 2^63 are kernel addresses of a 64-bit
1910          * kernel.  Note that addresses are unsigned so that on a 32-bit system
1911          * all addresses including kernel addresses are less than 2^32.  In
1912          * that case (32-bit system), if the kernel mapping is unknown, all
1913          * addresses will be assumed to be in user space - see
1914          * machine__kernel_ip().
1915          */
1916         machine->kernel_start = 1ULL << 63;
1917         if (map) {
1918                 err = map__load(map, machine->symbol_filter);
1919                 if (map->start)
1920                         machine->kernel_start = map->start;
1921         }
1922         return err;
1923 }