tools/perf/util/machine.c

   1 #include "callchain.h"
   2 #include "debug.h"
   3 #include "event.h"
   4 #include "evsel.h"
   5 #include "hist.h"
   6 #include "machine.h"
   7 #include "map.h"
   8 #include "sort.h"
   9 #include "strlist.h"
  10 #include "thread.h"
  11 #include "vdso.h"
  12 #include <stdbool.h>
  13 #include <symbol/kallsyms.h>
  14 #include "unwind.h"
  15 #include "linux/hash.h"
  16
  17 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  18
  19 static void dsos__init(struct dsos *dsos)
  20 {
  21         INIT_LIST_HEAD(&dsos->head);
  22         dsos->root = RB_ROOT;
  23         pthread_rwlock_init(&dsos->lock, NULL);
  24 }
  25
  26 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  27 {
  28         map_groups__init(&machine->kmaps, machine);
  29         RB_CLEAR_NODE(&machine->rb_node);
  30         dsos__init(&machine->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__purge(struct dsos *dsos)
  86 {
  87         struct dso *pos, *n;
  88
  89         pthread_rwlock_wrlock(&dsos->lock);
  90
  91         list_for_each_entry_safe(pos, n, &dsos->head, node) {
  92                 RB_CLEAR_NODE(&pos->rb_node);
  93                 list_del_init(&pos->node);
  94                 dso__put(pos);
  95         }
  96
  97         pthread_rwlock_unlock(&dsos->lock);
  98 }
  99
 100 static void dsos__exit(struct dsos *dsos)
 101 {
 102         dsos__purge(dsos);
 103         pthread_rwlock_destroy(&dsos->lock);
 104 }
 105
 106 void machine__delete_threads(struct machine *machine)
 107 {
 108         struct rb_node *nd;
 109
 110         pthread_rwlock_wrlock(&machine->threads_lock);
 111         nd = rb_first(&machine->threads);
 112         while (nd) {
 113                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 114
 115                 nd = rb_next(nd);
 116                 __machine__remove_thread(machine, t, false);
 117         }
 118         pthread_rwlock_unlock(&machine->threads_lock);
 119 }
 120
 121 void machine__exit(struct machine *machine)
 122 {
 123         map_groups__exit(&machine->kmaps);
 124         dsos__exit(&machine->dsos);
 125         machine__exit_vdso(machine);
 126         zfree(&machine->root_dir);
 127         zfree(&machine->current_tid);
 128         pthread_rwlock_destroy(&machine->threads_lock);
 129 }
 130
 131 void machine__delete(struct machine *machine)
 132 {
 133         machine__exit(machine);
 134         free(machine);
 135 }
 136
 137 void machines__init(struct machines *machines)
 138 {
 139         machine__init(&machines->host, "", HOST_KERNEL_ID);
 140         machines->guests = RB_ROOT;
 141         machines->symbol_filter = NULL;
 142 }
 143
 144 void machines__exit(struct machines *machines)
 145 {
 146         machine__exit(&machines->host);
 147         /* XXX exit guest */
 148 }
 149
 150 struct machine *machines__add(struct machines *machines, pid_t pid,
 151                               const char *root_dir)
 152 {
 153         struct rb_node **p = &machines->guests.rb_node;
 154         struct rb_node *parent = NULL;
 155         struct machine *pos, *machine = malloc(sizeof(*machine));
 156
 157         if (machine == NULL)
 158                 return NULL;
 159
 160         if (machine__init(machine, root_dir, pid) != 0) {
 161                 free(machine);
 162                 return NULL;
 163         }
 164
 165         machine->symbol_filter = machines->symbol_filter;
 166
 167         while (*p != NULL) {
 168                 parent = *p;
 169                 pos = rb_entry(parent, struct machine, rb_node);
 170                 if (pid < pos->pid)
 171                         p = &(*p)->rb_left;
 172                 else
 173                         p = &(*p)->rb_right;
 174         }
 175
 176         rb_link_node(&machine->rb_node, parent, p);
 177         rb_insert_color(&machine->rb_node, &machines->guests);
 178
 179         return machine;
 180 }
 181
 182 void machines__set_symbol_filter(struct machines *machines,
 183                                  symbol_filter_t symbol_filter)
 184 {
 185         struct rb_node *nd;
 186
 187         machines->symbol_filter = symbol_filter;
 188         machines->host.symbol_filter = symbol_filter;
 189
 190         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 191                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 192
 193                 machine->symbol_filter = symbol_filter;
 194         }
 195 }
 196
 197 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 198 {
 199         struct rb_node *nd;
 200
 201         machines->host.comm_exec = comm_exec;
 202
 203         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 204                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 205
 206                 machine->comm_exec = comm_exec;
 207         }
 208 }
 209
 210 struct machine *machines__find(struct machines *machines, pid_t pid)
 211 {
 212         struct rb_node **p = &machines->guests.rb_node;
 213         struct rb_node *parent = NULL;
 214         struct machine *machine;
 215         struct machine *default_machine = NULL;
 216
 217         if (pid == HOST_KERNEL_ID)
 218                 return &machines->host;
 219
 220         while (*p != NULL) {
 221                 parent = *p;
 222                 machine = rb_entry(parent, struct machine, rb_node);
 223                 if (pid < machine->pid)
 224                         p = &(*p)->rb_left;
 225                 else if (pid > machine->pid)
 226                         p = &(*p)->rb_right;
 227                 else
 228                         return machine;
 229                 if (!machine->pid)
 230                         default_machine = machine;
 231         }
 232
 233         return default_machine;
 234 }
 235
 236 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 237 {
 238         char path[PATH_MAX];
 239         const char *root_dir = "";
 240         struct machine *machine = machines__find(machines, pid);
 241
 242         if (machine && (machine->pid == pid))
 243                 goto out;
 244
 245         if ((pid != HOST_KERNEL_ID) &&
 246             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 247             (symbol_conf.guestmount)) {
 248                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 249                 if (access(path, R_OK)) {
 250                         static struct strlist *seen;
 251
 252                         if (!seen)
 253                                 seen = strlist__new(NULL, NULL);
 254
 255                         if (!strlist__has_entry(seen, path)) {
 256                                 pr_err("Can't access file %s\n", path);
 257                                 strlist__add(seen, path);
 258                         }
 259                         machine = NULL;
 260                         goto out;
 261                 }
 262                 root_dir = path;
 263         }
 264
 265         machine = machines__add(machines, pid, root_dir);
 266 out:
 267         return machine;
 268 }
 269
 270 void machines__process_guests(struct machines *machines,
 271                               machine__process_t process, void *data)
 272 {
 273         struct rb_node *nd;
 274
 275         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 276                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 277                 process(pos, data);
 278         }
 279 }
 280
 281 char *machine__mmap_name(struct machine *machine, char *bf, size_t size)
 282 {
 283         if (machine__is_host(machine))
 284                 snprintf(bf, size, "[%s]", "kernel.kallsyms");
 285         else if (machine__is_default_guest(machine))
 286                 snprintf(bf, size, "[%s]", "guest.kernel.kallsyms");
 287         else {
 288                 snprintf(bf, size, "[%s.%d]", "guest.kernel.kallsyms",
 289                          machine->pid);
 290         }
 291
 292         return bf;
 293 }
 294
 295 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 296 {
 297         struct rb_node *node;
 298         struct machine *machine;
 299
 300         machines->host.id_hdr_size = id_hdr_size;
 301
 302         for (node = rb_first(&machines->guests); node; node = rb_next(node)) {
 303                 machine = rb_entry(node, struct machine, rb_node);
 304                 machine->id_hdr_size = id_hdr_size;
 305         }
 306
 307         return;
 308 }
 309
 310 static void machine__update_thread_pid(struct machine *machine,
 311                                        struct thread *th, pid_t pid)
 312 {
 313         struct thread *leader;
 314
 315         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 316                 return;
 317
 318         th->pid_ = pid;
 319
 320         if (th->pid_ == th->tid)
 321                 return;
 322
 323         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 324         if (!leader)
 325                 goto out_err;
 326
 327         if (!leader->mg)
 328                 leader->mg = map_groups__new(machine);
 329
 330         if (!leader->mg)
 331                 goto out_err;
 332
 333         if (th->mg == leader->mg)
 334                 return;
 335
 336         if (th->mg) {
 337                 /*
 338                  * Maps are created from MMAP events which provide the pid and
 339                  * tid.  Consequently there never should be any maps on a thread
 340                  * with an unknown pid.  Just print an error if there are.
 341                  */
 342                 if (!map_groups__empty(th->mg))
 343                         pr_err("Discarding thread maps for %d:%d\n",
 344                                th->pid_, th->tid);
 345                 map_groups__put(th->mg);
 346         }
 347
 348         th->mg = map_groups__get(leader->mg);
 349
 350         return;
 351
 352 out_err:
 353         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 354 }
 355
 356 static struct thread *____machine__findnew_thread(struct machine *machine,
 357                                                   pid_t pid, pid_t tid,
 358                                                   bool create)
 359 {
 360         struct rb_node **p = &machine->threads.rb_node;
 361         struct rb_node *parent = NULL;
 362         struct thread *th;
 363
 364         /*
 365          * Front-end cache - TID lookups come in blocks,
 366          * so most of the time we dont have to look up
 367          * the full rbtree:
 368          */
 369         th = machine->last_match;
 370         if (th != NULL) {
 371                 if (th->tid == tid) {
 372                         machine__update_thread_pid(machine, th, pid);
 373                         return th;
 374                 }
 375
 376                 machine->last_match = NULL;
 377         }
 378
 379         while (*p != NULL) {
 380                 parent = *p;
 381                 th = rb_entry(parent, struct thread, rb_node);
 382
 383                 if (th->tid == tid) {
 384                         machine->last_match = th;
 385                         machine__update_thread_pid(machine, th, pid);
 386                         return th;
 387                 }
 388
 389                 if (tid < th->tid)
 390                         p = &(*p)->rb_left;
 391                 else
 392                         p = &(*p)->rb_right;
 393         }
 394
 395         if (!create)
 396                 return NULL;
 397
 398         th = thread__new(pid, tid);
 399         if (th != NULL) {
 400                 rb_link_node(&th->rb_node, parent, p);
 401                 rb_insert_color(&th->rb_node, &machine->threads);
 402
 403                 /*
 404                  * We have to initialize map_groups separately
 405                  * after rb tree is updated.
 406                  *
 407                  * The reason is that we call machine__findnew_thread
 408                  * within thread__init_map_groups to find the thread
 409                  * leader and that would screwed the rb tree.
 410                  */
 411                 if (thread__init_map_groups(th, machine)) {
 412                         rb_erase_init(&th->rb_node, &machine->threads);
 413                         RB_CLEAR_NODE(&th->rb_node);
 414                         thread__delete(th);
 415                         return NULL;
 416                 }
 417                 /*
 418                  * It is now in the rbtree, get a ref
 419                  */
 420                 thread__get(th);
 421                 machine->last_match = th;
 422         }
 423
 424         return th;
 425 }
 426
 427 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 428 {
 429         return ____machine__findnew_thread(machine, pid, tid, true);
 430 }
 431
 432 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 433                                        pid_t tid)
 434 {
 435         struct thread *th;
 436
 437         pthread_rwlock_wrlock(&machine->threads_lock);
 438         th = thread__get(__machine__findnew_thread(machine, pid, tid));
 439         pthread_rwlock_unlock(&machine->threads_lock);
 440         return th;
 441 }
 442
 443 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 444                                     pid_t tid)
 445 {
 446         struct thread *th;
 447         pthread_rwlock_rdlock(&machine->threads_lock);
 448         th =  thread__get(____machine__findnew_thread(machine, pid, tid, false));
 449         pthread_rwlock_unlock(&machine->threads_lock);
 450         return th;
 451 }
 452
 453 struct comm *machine__thread_exec_comm(struct machine *machine,
 454                                        struct thread *thread)
 455 {
 456         if (machine->comm_exec)
 457                 return thread__exec_comm(thread);
 458         else
 459                 return thread__comm(thread);
 460 }
 461
 462 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 463                                 struct perf_sample *sample)
 464 {
 465         struct thread *thread = machine__findnew_thread(machine,
 466                                                         event->comm.pid,
 467                                                         event->comm.tid);
 468         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 469         int err = 0;
 470
 471         if (exec)
 472                 machine->comm_exec = true;
 473
 474         if (dump_trace)
 475                 perf_event__fprintf_comm(event, stdout);
 476
 477         if (thread == NULL ||
 478             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 479                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 480                 err = -1;
 481         }
 482
 483         thread__put(thread);
 484
 485         return err;
 486 }
 487
 488 int machine__process_lost_event(struct machine *machine __maybe_unused,
 489                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 490 {
 491         dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
 492                     event->lost.id, event->lost.lost);
 493         return 0;
 494 }
 495
 496 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 497                                         union perf_event *event, struct perf_sample *sample)
 498 {
 499         dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
 500                     sample->id, event->lost_samples.lost);
 501         return 0;
 502 }
 503
 504 static struct dso *machine__findnew_module_dso(struct machine *machine,
 505                                                struct kmod_path *m,
 506                                                const char *filename)
 507 {
 508         struct dso *dso;
 509
 510         pthread_rwlock_wrlock(&machine->dsos.lock);
 511
 512         dso = __dsos__find(&machine->dsos, m->name, true);
 513         if (!dso) {
 514                 dso = __dsos__addnew(&machine->dsos, m->name);
 515                 if (dso == NULL)
 516                         goto out_unlock;
 517
 518                 if (machine__is_host(machine))
 519                         dso->symtab_type = DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE;
 520                 else
 521                         dso->symtab_type = DSO_BINARY_TYPE__GUEST_KMODULE;
 522
 523                 /* _KMODULE_COMP should be next to _KMODULE */
 524                 if (m->kmod && m->comp)
 525                         dso->symtab_type++;
 526
 527                 dso__set_short_name(dso, strdup(m->name), true);
 528                 dso__set_long_name(dso, strdup(filename), true);
 529         }
 530
 531         dso__get(dso);
 532 out_unlock:
 533         pthread_rwlock_unlock(&machine->dsos.lock);
 534         return dso;
 535 }
 536
 537 int machine__process_aux_event(struct machine *machine __maybe_unused,
 538                                union perf_event *event)
 539 {
 540         if (dump_trace)
 541                 perf_event__fprintf_aux(event, stdout);
 542         return 0;
 543 }
 544
 545 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 546                                         union perf_event *event)
 547 {
 548         if (dump_trace)
 549                 perf_event__fprintf_itrace_start(event, stdout);
 550         return 0;
 551 }
 552
 553 int machine__process_switch_event(struct machine *machine __maybe_unused,
 554                                   union perf_event *event)
 555 {
 556         if (dump_trace)
 557                 perf_event__fprintf_switch(event, stdout);
 558         return 0;
 559 }
 560
 561 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 562                                         const char *filename)
 563 {
 564         struct map *map = NULL;
 565         struct dso *dso;
 566         struct kmod_path m;
 567
 568         if (kmod_path__parse_name(&m, filename))
 569                 return NULL;
 570
 571         map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION,
 572                                        m.name);
 573         if (map)
 574                 goto out;
 575
 576         dso = machine__findnew_module_dso(machine, &m, filename);
 577         if (dso == NULL)
 578                 goto out;
 579
 580         map = map__new2(start, dso, MAP__FUNCTION);
 581         if (map == NULL)
 582                 goto out;
 583
 584         map_groups__insert(&machine->kmaps, map);
 585
 586 out:
 587         free(m.name);
 588         return map;
 589 }
 590
 591 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 592 {
 593         struct rb_node *nd;
 594         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 595
 596         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 597                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 598                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 599         }
 600
 601         return ret;
 602 }
 603
 604 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 605                                      bool (skip)(struct dso *dso, int parm), int parm)
 606 {
 607         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 608 }
 609
 610 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 611                                      bool (skip)(struct dso *dso, int parm), int parm)
 612 {
 613         struct rb_node *nd;
 614         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 615
 616         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 617                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 618                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 619         }
 620         return ret;
 621 }
 622
 623 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 624 {
 625         int i;
 626         size_t printed = 0;
 627         struct dso *kdso = machine->vmlinux_maps[MAP__FUNCTION]->dso;
 628
 629         if (kdso->has_build_id) {
 630                 char filename[PATH_MAX];
 631                 if (dso__build_id_filename(kdso, filename, sizeof(filename)))
 632                         printed += fprintf(fp, "[0] %s\n", filename);
 633         }
 634
 635         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 636                 printed += fprintf(fp, "[%d] %s\n",
 637                                    i + kdso->has_build_id, vmlinux_path[i]);
 638
 639         return printed;
 640 }
 641
 642 size_t machine__fprintf(struct machine *machine, FILE *fp)
 643 {
 644         size_t ret = 0;
 645         struct rb_node *nd;
 646
 647         pthread_rwlock_rdlock(&machine->threads_lock);
 648
 649         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
 650                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 651
 652                 ret += thread__fprintf(pos, fp);
 653         }
 654
 655         pthread_rwlock_unlock(&machine->threads_lock);
 656
 657         return ret;
 658 }
 659
 660 static struct dso *machine__get_kernel(struct machine *machine)
 661 {
 662         const char *vmlinux_name = NULL;
 663         struct dso *kernel;
 664
 665         if (machine__is_host(machine)) {
 666                 vmlinux_name = symbol_conf.vmlinux_name;
 667                 if (!vmlinux_name)
 668                         vmlinux_name = "[kernel.kallsyms]";
 669
 670                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 671                                                  "[kernel]", DSO_TYPE_KERNEL);
 672         } else {
 673                 char bf[PATH_MAX];
 674
 675                 if (machine__is_default_guest(machine))
 676                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 677                 if (!vmlinux_name)
 678                         vmlinux_name = machine__mmap_name(machine, bf,
 679                                                           sizeof(bf));
 680
 681                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 682                                                  "[guest.kernel]",
 683                                                  DSO_TYPE_GUEST_KERNEL);
 684         }
 685
 686         if (kernel != NULL && (!kernel->has_build_id))
 687                 dso__read_running_kernel_build_id(kernel, machine);
 688
 689         return kernel;
 690 }
 691
 692 struct process_args {
 693         u64 start;
 694 };
 695
 696 static void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 697                                            size_t bufsz)
 698 {
 699         if (machine__is_default_guest(machine))
 700                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 701         else
 702                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 703 }
 704
 705 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 706
 707 /* Figure out the start address of kernel map from /proc/kallsyms.
 708  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 709  * symbol_name if it's not that important.
 710  */
 711 static u64 machine__get_running_kernel_start(struct machine *machine,
 712                                              const char **symbol_name)
 713 {
 714         char filename[PATH_MAX];
 715         int i;
 716         const char *name;
 717         u64 addr = 0;
 718
 719         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 720
 721         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 722                 return 0;
 723
 724         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 725                 addr = kallsyms__get_function_start(filename, name);
 726                 if (addr)
 727                         break;
 728         }
 729
 730         if (symbol_name)
 731                 *symbol_name = name;
 732
 733         return addr;
 734 }
 735
 736 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
 737 {
 738         enum map_type type;
 739         u64 start = machine__get_running_kernel_start(machine, NULL);
 740
 741         for (type = 0; type < MAP__NR_TYPES; ++type) {
 742                 struct kmap *kmap;
 743
 744                 machine->vmlinux_maps[type] = map__new2(start, kernel, type);
 745                 if (machine->vmlinux_maps[type] == NULL)
 746                         return -1;
 747
 748                 machine->vmlinux_maps[type]->map_ip =
 749                         machine->vmlinux_maps[type]->unmap_ip =
 750                                 identity__map_ip;
 751                 kmap = map__kmap(machine->vmlinux_maps[type]);
 752                 if (!kmap)
 753                         return -1;
 754
 755                 kmap->kmaps = &machine->kmaps;
 756                 map_groups__insert(&machine->kmaps,
 757                                    machine->vmlinux_maps[type]);
 758         }
 759
 760         return 0;
 761 }
 762
 763 void machine__destroy_kernel_maps(struct machine *machine)
 764 {
 765         enum map_type type;
 766
 767         for (type = 0; type < MAP__NR_TYPES; ++type) {
 768                 struct kmap *kmap;
 769
 770                 if (machine->vmlinux_maps[type] == NULL)
 771                         continue;
 772
 773                 kmap = map__kmap(machine->vmlinux_maps[type]);
 774                 map_groups__remove(&machine->kmaps,
 775                                    machine->vmlinux_maps[type]);
 776                 if (kmap && kmap->ref_reloc_sym) {
 777                         /*
 778                          * ref_reloc_sym is shared among all maps, so free just
 779                          * on one of them.
 780                          */
 781                         if (type == MAP__FUNCTION) {
 782                                 zfree((char **)&kmap->ref_reloc_sym->name);
 783                                 zfree(&kmap->ref_reloc_sym);
 784                         } else
 785                                 kmap->ref_reloc_sym = NULL;
 786                 }
 787
 788                 machine->vmlinux_maps[type] = NULL;
 789         }
 790 }
 791
 792 int machines__create_guest_kernel_maps(struct machines *machines)
 793 {
 794         int ret = 0;
 795         struct dirent **namelist = NULL;
 796         int i, items = 0;
 797         char path[PATH_MAX];
 798         pid_t pid;
 799         char *endp;
 800
 801         if (symbol_conf.default_guest_vmlinux_name ||
 802             symbol_conf.default_guest_modules ||
 803             symbol_conf.default_guest_kallsyms) {
 804                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
 805         }
 806
 807         if (symbol_conf.guestmount) {
 808                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
 809                 if (items <= 0)
 810                         return -ENOENT;
 811                 for (i = 0; i < items; i++) {
 812                         if (!isdigit(namelist[i]->d_name[0])) {
 813                                 /* Filter out . and .. */
 814                                 continue;
 815                         }
 816                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
 817                         if ((*endp != '\0') ||
 818                             (endp == namelist[i]->d_name) ||
 819                             (errno == ERANGE)) {
 820                                 pr_debug("invalid directory (%s). Skipping.\n",
 821                                          namelist[i]->d_name);
 822                                 continue;
 823                         }
 824                         sprintf(path, "%s/%s/proc/kallsyms",
 825                                 symbol_conf.guestmount,
 826                                 namelist[i]->d_name);
 827                         ret = access(path, R_OK);
 828                         if (ret) {
 829                                 pr_debug("Can't access file %s\n", path);
 830                                 goto failure;
 831                         }
 832                         machines__create_kernel_maps(machines, pid);
 833                 }
 834 failure:
 835                 free(namelist);
 836         }
 837
 838         return ret;
 839 }
 840
 841 void machines__destroy_kernel_maps(struct machines *machines)
 842 {
 843         struct rb_node *next = rb_first(&machines->guests);
 844
 845         machine__destroy_kernel_maps(&machines->host);
 846
 847         while (next) {
 848                 struct machine *pos = rb_entry(next, struct machine, rb_node);
 849
 850                 next = rb_next(&pos->rb_node);
 851                 rb_erase(&pos->rb_node, &machines->guests);
 852                 machine__delete(pos);
 853         }
 854 }
 855
 856 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
 857 {
 858         struct machine *machine = machines__findnew(machines, pid);
 859
 860         if (machine == NULL)
 861                 return -1;
 862
 863         return machine__create_kernel_maps(machine);
 864 }
 865
 866 int machine__load_kallsyms(struct machine *machine, const char *filename,
 867                            enum map_type type, symbol_filter_t filter)
 868 {
 869         struct map *map = machine->vmlinux_maps[type];
 870         int ret = dso__load_kallsyms(map->dso, filename, map, filter);
 871
 872         if (ret > 0) {
 873                 dso__set_loaded(map->dso, type);
 874                 /*
 875                  * Since /proc/kallsyms will have multiple sessions for the
 876                  * kernel, with modules between them, fixup the end of all
 877                  * sections.
 878                  */
 879                 __map_groups__fixup_end(&machine->kmaps, type);
 880         }
 881
 882         return ret;
 883 }
 884
 885 int machine__load_vmlinux_path(struct machine *machine, enum map_type type,
 886                                symbol_filter_t filter)
 887 {
 888         struct map *map = machine->vmlinux_maps[type];
 889         int ret = dso__load_vmlinux_path(map->dso, map, filter);
 890
 891         if (ret > 0)
 892                 dso__set_loaded(map->dso, type);
 893
 894         return ret;
 895 }
 896
 897 static void map_groups__fixup_end(struct map_groups *mg)
 898 {
 899         int i;
 900         for (i = 0; i < MAP__NR_TYPES; ++i)
 901                 __map_groups__fixup_end(mg, i);
 902 }
 903
 904 static char *get_kernel_version(const char *root_dir)
 905 {
 906         char version[PATH_MAX];
 907         FILE *file;
 908         char *name, *tmp;
 909         const char *prefix = "Linux version ";
 910
 911         sprintf(version, "%s/proc/version", root_dir);
 912         file = fopen(version, "r");
 913         if (!file)
 914                 return NULL;
 915
 916         version[0] = '\0';
 917         tmp = fgets(version, sizeof(version), file);
 918         fclose(file);
 919
 920         name = strstr(version, prefix);
 921         if (!name)
 922                 return NULL;
 923         name += strlen(prefix);
 924         tmp = strchr(name, ' ');
 925         if (tmp)
 926                 *tmp = '\0';
 927
 928         return strdup(name);
 929 }
 930
 931 static bool is_kmod_dso(struct dso *dso)
 932 {
 933         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
 934                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
 935 }
 936
 937 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
 938                                        struct kmod_path *m)
 939 {
 940         struct map *map;
 941         char *long_name;
 942
 943         map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name);
 944         if (map == NULL)
 945                 return 0;
 946
 947         long_name = strdup(path);
 948         if (long_name == NULL)
 949                 return -ENOMEM;
 950
 951         dso__set_long_name(map->dso, long_name, true);
 952         dso__kernel_module_get_build_id(map->dso, "");
 953
 954         /*
 955          * Full name could reveal us kmod compression, so
 956          * we need to update the symtab_type if needed.
 957          */
 958         if (m->comp && is_kmod_dso(map->dso))
 959                 map->dso->symtab_type++;
 960
 961         return 0;
 962 }
 963
 964 static int map_groups__set_modules_path_dir(struct map_groups *mg,
 965                                 const char *dir_name, int depth)
 966 {
 967         struct dirent *dent;
 968         DIR *dir = opendir(dir_name);
 969         int ret = 0;
 970
 971         if (!dir) {
 972                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
 973                 return -1;
 974         }
 975
 976         while ((dent = readdir(dir)) != NULL) {
 977                 char path[PATH_MAX];
 978                 struct stat st;
 979
 980                 /*sshfs might return bad dent->d_type, so we have to stat*/
 981                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
 982                 if (stat(path, &st))
 983                         continue;
 984
 985                 if (S_ISDIR(st.st_mode)) {
 986                         if (!strcmp(dent->d_name, ".") ||
 987                             !strcmp(dent->d_name, ".."))
 988                                 continue;
 989
 990                         /* Do not follow top-level source and build symlinks */
 991                         if (depth == 0) {
 992                                 if (!strcmp(dent->d_name, "source") ||
 993                                     !strcmp(dent->d_name, "build"))
 994                                         continue;
 995                         }
 996
 997                         ret = map_groups__set_modules_path_dir(mg, path,
 998                                                                depth + 1);
 999                         if (ret < 0)
1000                                 goto out;
1001                 } else {
1002                         struct kmod_path m;
1003
1004                         ret = kmod_path__parse_name(&m, dent->d_name);
1005                         if (ret)
1006                                 goto out;
1007
1008                         if (m.kmod)
1009                                 ret = map_groups__set_module_path(mg, path, &m);
1010
1011                         free(m.name);
1012
1013                         if (ret)
1014                                 goto out;
1015                 }
1016         }
1017
1018 out:
1019         closedir(dir);
1020         return ret;
1021 }
1022
1023 static int machine__set_modules_path(struct machine *machine)
1024 {
1025         char *version;
1026         char modules_path[PATH_MAX];
1027
1028         version = get_kernel_version(machine->root_dir);
1029         if (!version)
1030                 return -1;
1031
1032         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1033                  machine->root_dir, version);
1034         free(version);
1035
1036         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1037 }
1038
1039 static int machine__create_module(void *arg, const char *name, u64 start)
1040 {
1041         struct machine *machine = arg;
1042         struct map *map;
1043
1044         map = machine__findnew_module_map(machine, start, name);
1045         if (map == NULL)
1046                 return -1;
1047
1048         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1049
1050         return 0;
1051 }
1052
1053 static int machine__create_modules(struct machine *machine)
1054 {
1055         const char *modules;
1056         char path[PATH_MAX];
1057
1058         if (machine__is_default_guest(machine)) {
1059                 modules = symbol_conf.default_guest_modules;
1060         } else {
1061                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1062                 modules = path;
1063         }
1064
1065         if (symbol__restricted_filename(modules, "/proc/modules"))
1066                 return -1;
1067
1068         if (modules__parse(modules, machine, machine__create_module))
1069                 return -1;
1070
1071         if (!machine__set_modules_path(machine))
1072                 return 0;
1073
1074         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1075
1076         return 0;
1077 }
1078
1079 int machine__create_kernel_maps(struct machine *machine)
1080 {
1081         struct dso *kernel = machine__get_kernel(machine);
1082         const char *name;
1083         u64 addr = machine__get_running_kernel_start(machine, &name);
1084         if (!addr)
1085                 return -1;
1086
1087         if (kernel == NULL ||
1088             __machine__create_kernel_maps(machine, kernel) < 0)
1089                 return -1;
1090
1091         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1092                 if (machine__is_host(machine))
1093                         pr_debug("Problems creating module maps, "
1094                                  "continuing anyway...\n");
1095                 else
1096                         pr_debug("Problems creating module maps for guest %d, "
1097                                  "continuing anyway...\n", machine->pid);
1098         }
1099
1100         /*
1101          * Now that we have all the maps created, just set the ->end of them:
1102          */
1103         map_groups__fixup_end(&machine->kmaps);
1104
1105         if (maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name,
1106                                              addr)) {
1107                 machine__destroy_kernel_maps(machine);
1108                 return -1;
1109         }
1110
1111         return 0;
1112 }
1113
1114 static void machine__set_kernel_mmap_len(struct machine *machine,
1115                                          union perf_event *event)
1116 {
1117         int i;
1118
1119         for (i = 0; i < MAP__NR_TYPES; i++) {
1120                 machine->vmlinux_maps[i]->start = event->mmap.start;
1121                 machine->vmlinux_maps[i]->end   = (event->mmap.start +
1122                                                    event->mmap.len);
1123                 /*
1124                  * Be a bit paranoid here, some perf.data file came with
1125                  * a zero sized synthesized MMAP event for the kernel.
1126                  */
1127                 if (machine->vmlinux_maps[i]->end == 0)
1128                         machine->vmlinux_maps[i]->end = ~0ULL;
1129         }
1130 }
1131
1132 static bool machine__uses_kcore(struct machine *machine)
1133 {
1134         struct dso *dso;
1135
1136         list_for_each_entry(dso, &machine->dsos.head, node) {
1137                 if (dso__is_kcore(dso))
1138                         return true;
1139         }
1140
1141         return false;
1142 }
1143
1144 static int machine__process_kernel_mmap_event(struct machine *machine,
1145                                               union perf_event *event)
1146 {
1147         struct map *map;
1148         char kmmap_prefix[PATH_MAX];
1149         enum dso_kernel_type kernel_type;
1150         bool is_kernel_mmap;
1151
1152         /* If we have maps from kcore then we do not need or want any others */
1153         if (machine__uses_kcore(machine))
1154                 return 0;
1155
1156         machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix));
1157         if (machine__is_host(machine))
1158                 kernel_type = DSO_TYPE_KERNEL;
1159         else
1160                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1161
1162         is_kernel_mmap = memcmp(event->mmap.filename,
1163                                 kmmap_prefix,
1164                                 strlen(kmmap_prefix) - 1) == 0;
1165         if (event->mmap.filename[0] == '/' ||
1166             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1167                 map = machine__findnew_module_map(machine, event->mmap.start,
1168                                                   event->mmap.filename);
1169                 if (map == NULL)
1170                         goto out_problem;
1171
1172                 map->end = map->start + event->mmap.len;
1173         } else if (is_kernel_mmap) {
1174                 const char *symbol_name = (event->mmap.filename +
1175                                 strlen(kmmap_prefix));
1176                 /*
1177                  * Should be there already, from the build-id table in
1178                  * the header.
1179                  */
1180                 struct dso *kernel = NULL;
1181                 struct dso *dso;
1182
1183                 pthread_rwlock_rdlock(&machine->dsos.lock);
1184
1185                 list_for_each_entry(dso, &machine->dsos.head, node) {
1186
1187                         /*
1188                          * The cpumode passed to is_kernel_module is not the
1189                          * cpumode of *this* event. If we insist on passing
1190                          * correct cpumode to is_kernel_module, we should
1191                          * record the cpumode when we adding this dso to the
1192                          * linked list.
1193                          *
1194                          * However we don't really need passing correct
1195                          * cpumode.  We know the correct cpumode must be kernel
1196                          * mode (if not, we should not link it onto kernel_dsos
1197                          * list).
1198                          *
1199                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1200                          * is_kernel_module() treats it as a kernel cpumode.
1201                          */
1202
1203                         if (!dso->kernel ||
1204                             is_kernel_module(dso->long_name,
1205                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1206                                 continue;
1207
1208
1209                         kernel = dso;
1210                         break;
1211                 }
1212
1213                 pthread_rwlock_unlock(&machine->dsos.lock);
1214
1215                 if (kernel == NULL)
1216                         kernel = machine__findnew_dso(machine, kmmap_prefix);
1217                 if (kernel == NULL)
1218                         goto out_problem;
1219
1220                 kernel->kernel = kernel_type;
1221                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1222                         dso__put(kernel);
1223                         goto out_problem;
1224                 }
1225
1226                 if (strstr(kernel->long_name, "vmlinux"))
1227                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1228
1229                 machine__set_kernel_mmap_len(machine, event);
1230
1231                 /*
1232                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1233                  * symbol. Effectively having zero here means that at record
1234                  * time /proc/sys/kernel/kptr_restrict was non zero.
1235                  */
1236                 if (event->mmap.pgoff != 0) {
1237                         maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps,
1238                                                          symbol_name,
1239                                                          event->mmap.pgoff);
1240                 }
1241
1242                 if (machine__is_default_guest(machine)) {
1243                         /*
1244                          * preload dso of guest kernel and modules
1245                          */
1246                         dso__load(kernel, machine->vmlinux_maps[MAP__FUNCTION],
1247                                   NULL);
1248                 }
1249         }
1250         return 0;
1251 out_problem:
1252         return -1;
1253 }
1254
1255 int machine__process_mmap2_event(struct machine *machine,
1256                                  union perf_event *event,
1257                                  struct perf_sample *sample __maybe_unused)
1258 {
1259         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1260         struct thread *thread;
1261         struct map *map;
1262         enum map_type type;
1263         int ret = 0;
1264
1265         if (dump_trace)
1266                 perf_event__fprintf_mmap2(event, stdout);
1267
1268         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1269             cpumode == PERF_RECORD_MISC_KERNEL) {
1270                 ret = machine__process_kernel_mmap_event(machine, event);
1271                 if (ret < 0)
1272                         goto out_problem;
1273                 return 0;
1274         }
1275
1276         thread = machine__findnew_thread(machine, event->mmap2.pid,
1277                                         event->mmap2.tid);
1278         if (thread == NULL)
1279                 goto out_problem;
1280
1281         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1282                 type = MAP__VARIABLE;
1283         else
1284                 type = MAP__FUNCTION;
1285
1286         map = map__new(machine, event->mmap2.start,
1287                         event->mmap2.len, event->mmap2.pgoff,
1288                         event->mmap2.pid, event->mmap2.maj,
1289                         event->mmap2.min, event->mmap2.ino,
1290                         event->mmap2.ino_generation,
1291                         event->mmap2.prot,
1292                         event->mmap2.flags,
1293                         event->mmap2.filename, type, thread);
1294
1295         if (map == NULL)
1296                 goto out_problem_map;
1297
1298         thread__insert_map(thread, map);
1299         thread__put(thread);
1300         map__put(map);
1301         return 0;
1302
1303 out_problem_map:
1304         thread__put(thread);
1305 out_problem:
1306         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1307         return 0;
1308 }
1309
1310 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1311                                 struct perf_sample *sample __maybe_unused)
1312 {
1313         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1314         struct thread *thread;
1315         struct map *map;
1316         enum map_type type;
1317         int ret = 0;
1318
1319         if (dump_trace)
1320                 perf_event__fprintf_mmap(event, stdout);
1321
1322         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1323             cpumode == PERF_RECORD_MISC_KERNEL) {
1324                 ret = machine__process_kernel_mmap_event(machine, event);
1325                 if (ret < 0)
1326                         goto out_problem;
1327                 return 0;
1328         }
1329
1330         thread = machine__findnew_thread(machine, event->mmap.pid,
1331                                          event->mmap.tid);
1332         if (thread == NULL)
1333                 goto out_problem;
1334
1335         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1336                 type = MAP__VARIABLE;
1337         else
1338                 type = MAP__FUNCTION;
1339
1340         map = map__new(machine, event->mmap.start,
1341                         event->mmap.len, event->mmap.pgoff,
1342                         event->mmap.pid, 0, 0, 0, 0, 0, 0,
1343                         event->mmap.filename,
1344                         type, thread);
1345
1346         if (map == NULL)
1347                 goto out_problem_map;
1348
1349         thread__insert_map(thread, map);
1350         thread__put(thread);
1351         map__put(map);
1352         return 0;
1353
1354 out_problem_map:
1355         thread__put(thread);
1356 out_problem:
1357         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1358         return 0;
1359 }
1360
1361 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1362 {
1363         if (machine->last_match == th)
1364                 machine->last_match = NULL;
1365
1366         BUG_ON(atomic_read(&th->refcnt) == 0);
1367         if (lock)
1368                 pthread_rwlock_wrlock(&machine->threads_lock);
1369         rb_erase_init(&th->rb_node, &machine->threads);
1370         RB_CLEAR_NODE(&th->rb_node);
1371         /*
1372          * Move it first to the dead_threads list, then drop the reference,
1373          * if this is the last reference, then the thread__delete destructor
1374          * will be called and we will remove it from the dead_threads list.
1375          */
1376         list_add_tail(&th->node, &machine->dead_threads);
1377         if (lock)
1378                 pthread_rwlock_unlock(&machine->threads_lock);
1379         thread__put(th);
1380 }
1381
1382 void machine__remove_thread(struct machine *machine, struct thread *th)
1383 {
1384         return __machine__remove_thread(machine, th, true);
1385 }
1386
1387 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1388                                 struct perf_sample *sample)
1389 {
1390         struct thread *thread = machine__find_thread(machine,
1391                                                      event->fork.pid,
1392                                                      event->fork.tid);
1393         struct thread *parent = machine__findnew_thread(machine,
1394                                                         event->fork.ppid,
1395                                                         event->fork.ptid);
1396         int err = 0;
1397
1398         if (dump_trace)
1399                 perf_event__fprintf_task(event, stdout);
1400
1401         /*
1402          * There may be an existing thread that is not actually the parent,
1403          * either because we are processing events out of order, or because the
1404          * (fork) event that would have removed the thread was lost. Assume the
1405          * latter case and continue on as best we can.
1406          */
1407         if (parent->pid_ != (pid_t)event->fork.ppid) {
1408                 dump_printf("removing erroneous parent thread %d/%d\n",
1409                             parent->pid_, parent->tid);
1410                 machine__remove_thread(machine, parent);
1411                 thread__put(parent);
1412                 parent = machine__findnew_thread(machine, event->fork.ppid,
1413                                                  event->fork.ptid);
1414         }
1415
1416         /* if a thread currently exists for the thread id remove it */
1417         if (thread != NULL) {
1418                 machine__remove_thread(machine, thread);
1419                 thread__put(thread);
1420         }
1421
1422         thread = machine__findnew_thread(machine, event->fork.pid,
1423                                          event->fork.tid);
1424
1425         if (thread == NULL || parent == NULL ||
1426             thread__fork(thread, parent, sample->time) < 0) {
1427                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1428                 err = -1;
1429         }
1430         thread__put(thread);
1431         thread__put(parent);
1432
1433         return err;
1434 }
1435
1436 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1437                                 struct perf_sample *sample __maybe_unused)
1438 {
1439         struct thread *thread = machine__find_thread(machine,
1440                                                      event->fork.pid,
1441                                                      event->fork.tid);
1442
1443         if (dump_trace)
1444                 perf_event__fprintf_task(event, stdout);
1445
1446         if (thread != NULL) {
1447                 thread__exited(thread);
1448                 thread__put(thread);
1449         }
1450
1451         return 0;
1452 }
1453
1454 int machine__process_event(struct machine *machine, union perf_event *event,
1455                            struct perf_sample *sample)
1456 {
1457         int ret;
1458
1459         switch (event->header.type) {
1460         case PERF_RECORD_COMM:
1461                 ret = machine__process_comm_event(machine, event, sample); break;
1462         case PERF_RECORD_MMAP:
1463                 ret = machine__process_mmap_event(machine, event, sample); break;
1464         case PERF_RECORD_MMAP2:
1465                 ret = machine__process_mmap2_event(machine, event, sample); break;
1466         case PERF_RECORD_FORK:
1467                 ret = machine__process_fork_event(machine, event, sample); break;
1468         case PERF_RECORD_EXIT:
1469                 ret = machine__process_exit_event(machine, event, sample); break;
1470         case PERF_RECORD_LOST:
1471                 ret = machine__process_lost_event(machine, event, sample); break;
1472         case PERF_RECORD_AUX:
1473                 ret = machine__process_aux_event(machine, event); break;
1474         case PERF_RECORD_ITRACE_START:
1475                 ret = machine__process_itrace_start_event(machine, event); break;
1476         case PERF_RECORD_LOST_SAMPLES:
1477                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1478         case PERF_RECORD_SWITCH:
1479         case PERF_RECORD_SWITCH_CPU_WIDE:
1480                 ret = machine__process_switch_event(machine, event); break;
1481         default:
1482                 ret = -1;
1483                 break;
1484         }
1485
1486         return ret;
1487 }
1488
1489 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1490 {
1491         if (sym->name && !regexec(regex, sym->name, 0, NULL, 0))
1492                 return 1;
1493         return 0;
1494 }
1495
1496 static void ip__resolve_ams(struct thread *thread,
1497                             struct addr_map_symbol *ams,
1498                             u64 ip)
1499 {
1500         struct addr_location al;
1501
1502         memset(&al, 0, sizeof(al));
1503         /*
1504          * We cannot use the header.misc hint to determine whether a
1505          * branch stack address is user, kernel, guest, hypervisor.
1506          * Branches may straddle the kernel/user/hypervisor boundaries.
1507          * Thus, we have to try consecutively until we find a match
1508          * or else, the symbol is unknown
1509          */
1510         thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al);
1511
1512         ams->addr = ip;
1513         ams->al_addr = al.addr;
1514         ams->sym = al.sym;
1515         ams->map = al.map;
1516 }
1517
1518 static void ip__resolve_data(struct thread *thread,
1519                              u8 m, struct addr_map_symbol *ams, u64 addr)
1520 {
1521         struct addr_location al;
1522
1523         memset(&al, 0, sizeof(al));
1524
1525         thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al);
1526         if (al.map == NULL) {
1527                 /*
1528                  * some shared data regions have execute bit set which puts
1529                  * their mapping in the MAP__FUNCTION type array.
1530                  * Check there as a fallback option before dropping the sample.
1531                  */
1532                 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al);
1533         }
1534
1535         ams->addr = addr;
1536         ams->al_addr = al.addr;
1537         ams->sym = al.sym;
1538         ams->map = al.map;
1539 }
1540
1541 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1542                                      struct addr_location *al)
1543 {
1544         struct mem_info *mi = zalloc(sizeof(*mi));
1545
1546         if (!mi)
1547                 return NULL;
1548
1549         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1550         ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr);
1551         mi->data_src.val = sample->data_src;
1552
1553         return mi;
1554 }
1555
1556 static int add_callchain_ip(struct thread *thread,
1557                             struct symbol **parent,
1558                             struct addr_location *root_al,
1559                             u8 *cpumode,
1560                             u64 ip)
1561 {
1562         struct addr_location al;
1563
1564         al.filtered = 0;
1565         al.sym = NULL;
1566         if (!cpumode) {
1567                 thread__find_cpumode_addr_location(thread, MAP__FUNCTION,
1568                                                    ip, &al);
1569         } else {
1570                 if (ip >= PERF_CONTEXT_MAX) {
1571                         switch (ip) {
1572                         case PERF_CONTEXT_HV:
1573                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
1574                                 break;
1575                         case PERF_CONTEXT_KERNEL:
1576                                 *cpumode = PERF_RECORD_MISC_KERNEL;
1577                                 break;
1578                         case PERF_CONTEXT_USER:
1579                                 *cpumode = PERF_RECORD_MISC_USER;
1580                                 break;
1581                         default:
1582                                 pr_debug("invalid callchain context: "
1583                                          "%"PRId64"\n", (s64) ip);
1584                                 /*
1585                                  * It seems the callchain is corrupted.
1586                                  * Discard all.
1587                                  */
1588                                 callchain_cursor_reset(&callchain_cursor);
1589                                 return 1;
1590                         }
1591                         return 0;
1592                 }
1593                 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION,
1594                                            ip, &al);
1595         }
1596
1597         if (al.sym != NULL) {
1598                 if (sort__has_parent && !*parent &&
1599                     symbol__match_regex(al.sym, &parent_regex))
1600                         *parent = al.sym;
1601                 else if (have_ignore_callees && root_al &&
1602                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
1603                         /* Treat this symbol as the root,
1604                            forgetting its callees. */
1605                         *root_al = al;
1606                         callchain_cursor_reset(&callchain_cursor);
1607                 }
1608         }
1609
1610         return callchain_cursor_append(&callchain_cursor, al.addr, al.map, al.sym);
1611 }
1612
1613 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
1614                                            struct addr_location *al)
1615 {
1616         unsigned int i;
1617         const struct branch_stack *bs = sample->branch_stack;
1618         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
1619
1620         if (!bi)
1621                 return NULL;
1622
1623         for (i = 0; i < bs->nr; i++) {
1624                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
1625                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
1626                 bi[i].flags = bs->entries[i].flags;
1627         }
1628         return bi;
1629 }
1630
1631 #define CHASHSZ 127
1632 #define CHASHBITS 7
1633 #define NO_ENTRY 0xff
1634
1635 #define PERF_MAX_BRANCH_DEPTH 127
1636
1637 /* Remove loops. */
1638 static int remove_loops(struct branch_entry *l, int nr)
1639 {
1640         int i, j, off;
1641         unsigned char chash[CHASHSZ];
1642
1643         memset(chash, NO_ENTRY, sizeof(chash));
1644
1645         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
1646
1647         for (i = 0; i < nr; i++) {
1648                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
1649
1650                 /* no collision handling for now */
1651                 if (chash[h] == NO_ENTRY) {
1652                         chash[h] = i;
1653                 } else if (l[chash[h]].from == l[i].from) {
1654                         bool is_loop = true;
1655                         /* check if it is a real loop */
1656                         off = 0;
1657                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
1658                                 if (l[j].from != l[i + off].from) {
1659                                         is_loop = false;
1660                                         break;
1661                                 }
1662                         if (is_loop) {
1663                                 memmove(l + i, l + i + off,
1664                                         (nr - (i + off)) * sizeof(*l));
1665                                 nr -= off;
1666                         }
1667                 }
1668         }
1669         return nr;
1670 }
1671
1672 /*
1673  * Recolve LBR callstack chain sample
1674  * Return:
1675  * 1 on success get LBR callchain information
1676  * 0 no available LBR callchain information, should try fp
1677  * negative error code on other errors.
1678  */
1679 static int resolve_lbr_callchain_sample(struct thread *thread,
1680                                         struct perf_sample *sample,
1681                                         struct symbol **parent,
1682                                         struct addr_location *root_al,
1683                                         int max_stack)
1684 {
1685         struct ip_callchain *chain = sample->callchain;
1686         int chain_nr = min(max_stack, (int)chain->nr);
1687         u8 cpumode = PERF_RECORD_MISC_USER;
1688         int i, j, err;
1689         u64 ip;
1690
1691         for (i = 0; i < chain_nr; i++) {
1692                 if (chain->ips[i] == PERF_CONTEXT_USER)
1693                         break;
1694         }
1695
1696         /* LBR only affects the user callchain */
1697         if (i != chain_nr) {
1698                 struct branch_stack *lbr_stack = sample->branch_stack;
1699                 int lbr_nr = lbr_stack->nr;
1700                 /*
1701                  * LBR callstack can only get user call chain.
1702                  * The mix_chain_nr is kernel call chain
1703                  * number plus LBR user call chain number.
1704                  * i is kernel call chain number,
1705                  * 1 is PERF_CONTEXT_USER,
1706                  * lbr_nr + 1 is the user call chain number.
1707                  * For details, please refer to the comments
1708                  * in callchain__printf
1709                  */
1710                 int mix_chain_nr = i + 1 + lbr_nr + 1;
1711
1712                 if (mix_chain_nr > PERF_MAX_STACK_DEPTH + PERF_MAX_BRANCH_DEPTH) {
1713                         pr_warning("corrupted callchain. skipping...\n");
1714                         return 0;
1715                 }
1716
1717                 for (j = 0; j < mix_chain_nr; j++) {
1718                         if (callchain_param.order == ORDER_CALLEE) {
1719                                 if (j < i + 1)
1720                                         ip = chain->ips[j];
1721                                 else if (j > i + 1)
1722                                         ip = lbr_stack->entries[j - i - 2].from;
1723                                 else
1724                                         ip = lbr_stack->entries[0].to;
1725                         } else {
1726                                 if (j < lbr_nr)
1727                                         ip = lbr_stack->entries[lbr_nr - j - 1].from;
1728                                 else if (j > lbr_nr)
1729                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
1730                                 else
1731                                         ip = lbr_stack->entries[0].to;
1732                         }
1733
1734                         err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1735                         if (err)
1736                                 return (err < 0) ? err : 0;
1737                 }
1738                 return 1;
1739         }
1740
1741         return 0;
1742 }
1743
1744 static int thread__resolve_callchain_sample(struct thread *thread,
1745                                             struct perf_evsel *evsel,
1746                                             struct perf_sample *sample,
1747                                             struct symbol **parent,
1748                                             struct addr_location *root_al,
1749                                             int max_stack)
1750 {
1751         struct branch_stack *branch = sample->branch_stack;
1752         struct ip_callchain *chain = sample->callchain;
1753         int chain_nr = min(max_stack, (int)chain->nr);
1754         u8 cpumode = PERF_RECORD_MISC_USER;
1755         int i, j, err;
1756         int skip_idx = -1;
1757         int first_call = 0;
1758
1759         callchain_cursor_reset(&callchain_cursor);
1760
1761         if (has_branch_callstack(evsel)) {
1762                 err = resolve_lbr_callchain_sample(thread, sample, parent,
1763                                                    root_al, max_stack);
1764                 if (err)
1765                         return (err < 0) ? err : 0;
1766         }
1767
1768         /*
1769          * Based on DWARF debug information, some architectures skip
1770          * a callchain entry saved by the kernel.
1771          */
1772         if (chain->nr < PERF_MAX_STACK_DEPTH)
1773                 skip_idx = arch_skip_callchain_idx(thread, chain);
1774
1775         /*
1776          * Add branches to call stack for easier browsing. This gives
1777          * more context for a sample than just the callers.
1778          *
1779          * This uses individual histograms of paths compared to the
1780          * aggregated histograms the normal LBR mode uses.
1781          *
1782          * Limitations for now:
1783          * - No extra filters
1784          * - No annotations (should annotate somehow)
1785          */
1786
1787         if (branch && callchain_param.branch_callstack) {
1788                 int nr = min(max_stack, (int)branch->nr);
1789                 struct branch_entry be[nr];
1790
1791                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
1792                         pr_warning("corrupted branch chain. skipping...\n");
1793                         goto check_calls;
1794                 }
1795
1796                 for (i = 0; i < nr; i++) {
1797                         if (callchain_param.order == ORDER_CALLEE) {
1798                                 be[i] = branch->entries[i];
1799                                 /*
1800                                  * Check for overlap into the callchain.
1801                                  * The return address is one off compared to
1802                                  * the branch entry. To adjust for this
1803                                  * assume the calling instruction is not longer
1804                                  * than 8 bytes.
1805                                  */
1806                                 if (i == skip_idx ||
1807                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
1808                                         first_call++;
1809                                 else if (be[i].from < chain->ips[first_call] &&
1810                                     be[i].from >= chain->ips[first_call] - 8)
1811                                         first_call++;
1812                         } else
1813                                 be[i] = branch->entries[branch->nr - i - 1];
1814                 }
1815
1816                 nr = remove_loops(be, nr);
1817
1818                 for (i = 0; i < nr; i++) {
1819                         err = add_callchain_ip(thread, parent, root_al,
1820                                                NULL, be[i].to);
1821                         if (!err)
1822                                 err = add_callchain_ip(thread, parent, root_al,
1823                                                        NULL, be[i].from);
1824                         if (err == -EINVAL)
1825                                 break;
1826                         if (err)
1827                                 return err;
1828                 }
1829                 chain_nr -= nr;
1830         }
1831
1832 check_calls:
1833         if (chain->nr > PERF_MAX_STACK_DEPTH) {
1834                 pr_warning("corrupted callchain. skipping...\n");
1835                 return 0;
1836         }
1837
1838         for (i = first_call; i < chain_nr; i++) {
1839                 u64 ip;
1840
1841                 if (callchain_param.order == ORDER_CALLEE)
1842                         j = i;
1843                 else
1844                         j = chain->nr - i - 1;
1845
1846 #ifdef HAVE_SKIP_CALLCHAIN_IDX
1847                 if (j == skip_idx)
1848                         continue;
1849 #endif
1850                 ip = chain->ips[j];
1851
1852                 err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1853
1854                 if (err)
1855                         return (err < 0) ? err : 0;
1856         }
1857
1858         return 0;
1859 }
1860
1861 static int unwind_entry(struct unwind_entry *entry, void *arg)
1862 {
1863         struct callchain_cursor *cursor = arg;
1864         return callchain_cursor_append(cursor, entry->ip,
1865                                        entry->map, entry->sym);
1866 }
1867
1868 int thread__resolve_callchain(struct thread *thread,
1869                               struct perf_evsel *evsel,
1870                               struct perf_sample *sample,
1871                               struct symbol **parent,
1872                               struct addr_location *root_al,
1873                               int max_stack)
1874 {
1875         int ret = thread__resolve_callchain_sample(thread, evsel,
1876                                                    sample, parent,
1877                                                    root_al, max_stack);
1878         if (ret)
1879                 return ret;
1880
1881         /* Can we do dwarf post unwind? */
1882         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
1883               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
1884                 return 0;
1885
1886         /* Bail out if nothing was captured. */
1887         if ((!sample->user_regs.regs) ||
1888             (!sample->user_stack.size))
1889                 return 0;
1890
1891         return unwind__get_entries(unwind_entry, &callchain_cursor,
1892                                    thread, sample, max_stack);
1893
1894 }
1895
1896 int machine__for_each_thread(struct machine *machine,
1897                              int (*fn)(struct thread *thread, void *p),
1898                              void *priv)
1899 {
1900         struct rb_node *nd;
1901         struct thread *thread;
1902         int rc = 0;
1903
1904         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
1905                 thread = rb_entry(nd, struct thread, rb_node);
1906                 rc = fn(thread, priv);
1907                 if (rc != 0)
1908                         return rc;
1909         }
1910
1911         list_for_each_entry(thread, &machine->dead_threads, node) {
1912                 rc = fn(thread, priv);
1913                 if (rc != 0)
1914                         return rc;
1915         }
1916         return rc;
1917 }
1918
1919 int machines__for_each_thread(struct machines *machines,
1920                               int (*fn)(struct thread *thread, void *p),
1921                               void *priv)
1922 {
1923         struct rb_node *nd;
1924         int rc = 0;
1925
1926         rc = machine__for_each_thread(&machines->host, fn, priv);
1927         if (rc != 0)
1928                 return rc;
1929
1930         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
1931                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
1932
1933                 rc = machine__for_each_thread(machine, fn, priv);
1934                 if (rc != 0)
1935                         return rc;
1936         }
1937         return rc;
1938 }
1939
1940 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1941                                   struct target *target, struct thread_map *threads,
1942                                   perf_event__handler_t process, bool data_mmap,
1943                                   unsigned int proc_map_timeout)
1944 {
1945         if (target__has_task(target))
1946                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap, proc_map_timeout);
1947         else if (target__has_cpu(target))
1948                 return perf_event__synthesize_threads(tool, process, machine, data_mmap, proc_map_timeout);
1949         /* command specified */
1950         return 0;
1951 }
1952
1953 pid_t machine__get_current_tid(struct machine *machine, int cpu)
1954 {
1955         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
1956                 return -1;
1957
1958         return machine->current_tid[cpu];
1959 }
1960
1961 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
1962                              pid_t tid)
1963 {
1964         struct thread *thread;
1965
1966         if (cpu < 0)
1967                 return -EINVAL;
1968
1969         if (!machine->current_tid) {
1970                 int i;
1971
1972                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
1973                 if (!machine->current_tid)
1974                         return -ENOMEM;
1975                 for (i = 0; i < MAX_NR_CPUS; i++)
1976                         machine->current_tid[i] = -1;
1977         }
1978
1979         if (cpu >= MAX_NR_CPUS) {
1980                 pr_err("Requested CPU %d too large. ", cpu);
1981                 pr_err("Consider raising MAX_NR_CPUS\n");
1982                 return -EINVAL;
1983         }
1984
1985         machine->current_tid[cpu] = tid;
1986
1987         thread = machine__findnew_thread(machine, pid, tid);
1988         if (!thread)
1989                 return -ENOMEM;
1990
1991         thread->cpu = cpu;
1992         thread__put(thread);
1993
1994         return 0;
1995 }
1996
1997 int machine__get_kernel_start(struct machine *machine)
1998 {
1999         struct map *map = machine__kernel_map(machine, MAP__FUNCTION);
2000         int err = 0;
2001
2002         /*
2003          * The only addresses above 2^63 are kernel addresses of a 64-bit
2004          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2005          * all addresses including kernel addresses are less than 2^32.  In
2006          * that case (32-bit system), if the kernel mapping is unknown, all
2007          * addresses will be assumed to be in user space - see
2008          * machine__kernel_ip().
2009          */
2010         machine->kernel_start = 1ULL << 63;
2011         if (map) {
2012                 err = map__load(map, machine->symbol_filter);
2013                 if (map->start)
2014                         machine->kernel_start = map->start;
2015         }
2016         return err;
2017 }
2018
2019 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2020 {
2021         return dsos__findnew(&machine->dsos, filename);
2022 }
2023
2024 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2025 {
2026         struct machine *machine = vmachine;
2027         struct map *map;
2028         struct symbol *sym = map_groups__find_symbol(&machine->kmaps, MAP__FUNCTION, *addrp, &map,  NULL);
2029
2030         if (sym == NULL)
2031                 return NULL;
2032
2033         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2034         *addrp = map->unmap_ip(map, sym->start);
2035         return sym->name;
2036 }