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(true, 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 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 554                                         const char *filename)
 555 {
 556         struct map *map = NULL;
 557         struct dso *dso;
 558         struct kmod_path m;
 559
 560         if (kmod_path__parse_name(&m, filename))
 561                 return NULL;
 562
 563         map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION,
 564                                        m.name);
 565         if (map)
 566                 goto out;
 567
 568         dso = machine__findnew_module_dso(machine, &m, filename);
 569         if (dso == NULL)
 570                 goto out;
 571
 572         map = map__new2(start, dso, MAP__FUNCTION);
 573         if (map == NULL)
 574                 goto out;
 575
 576         map_groups__insert(&machine->kmaps, map);
 577
 578 out:
 579         free(m.name);
 580         return map;
 581 }
 582
 583 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 584 {
 585         struct rb_node *nd;
 586         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 587
 588         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 589                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 590                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 591         }
 592
 593         return ret;
 594 }
 595
 596 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 597                                      bool (skip)(struct dso *dso, int parm), int parm)
 598 {
 599         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 600 }
 601
 602 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 603                                      bool (skip)(struct dso *dso, int parm), int parm)
 604 {
 605         struct rb_node *nd;
 606         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 607
 608         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 609                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 610                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 611         }
 612         return ret;
 613 }
 614
 615 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 616 {
 617         int i;
 618         size_t printed = 0;
 619         struct dso *kdso = machine->vmlinux_maps[MAP__FUNCTION]->dso;
 620
 621         if (kdso->has_build_id) {
 622                 char filename[PATH_MAX];
 623                 if (dso__build_id_filename(kdso, filename, sizeof(filename)))
 624                         printed += fprintf(fp, "[0] %s\n", filename);
 625         }
 626
 627         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 628                 printed += fprintf(fp, "[%d] %s\n",
 629                                    i + kdso->has_build_id, vmlinux_path[i]);
 630
 631         return printed;
 632 }
 633
 634 size_t machine__fprintf(struct machine *machine, FILE *fp)
 635 {
 636         size_t ret = 0;
 637         struct rb_node *nd;
 638
 639         pthread_rwlock_rdlock(&machine->threads_lock);
 640
 641         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
 642                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 643
 644                 ret += thread__fprintf(pos, fp);
 645         }
 646
 647         pthread_rwlock_unlock(&machine->threads_lock);
 648
 649         return ret;
 650 }
 651
 652 static struct dso *machine__get_kernel(struct machine *machine)
 653 {
 654         const char *vmlinux_name = NULL;
 655         struct dso *kernel;
 656
 657         if (machine__is_host(machine)) {
 658                 vmlinux_name = symbol_conf.vmlinux_name;
 659                 if (!vmlinux_name)
 660                         vmlinux_name = "[kernel.kallsyms]";
 661
 662                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 663                                                  "[kernel]", DSO_TYPE_KERNEL);
 664         } else {
 665                 char bf[PATH_MAX];
 666
 667                 if (machine__is_default_guest(machine))
 668                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 669                 if (!vmlinux_name)
 670                         vmlinux_name = machine__mmap_name(machine, bf,
 671                                                           sizeof(bf));
 672
 673                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 674                                                  "[guest.kernel]",
 675                                                  DSO_TYPE_GUEST_KERNEL);
 676         }
 677
 678         if (kernel != NULL && (!kernel->has_build_id))
 679                 dso__read_running_kernel_build_id(kernel, machine);
 680
 681         return kernel;
 682 }
 683
 684 struct process_args {
 685         u64 start;
 686 };
 687
 688 static void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 689                                            size_t bufsz)
 690 {
 691         if (machine__is_default_guest(machine))
 692                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 693         else
 694                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 695 }
 696
 697 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 698
 699 /* Figure out the start address of kernel map from /proc/kallsyms.
 700  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 701  * symbol_name if it's not that important.
 702  */
 703 static u64 machine__get_running_kernel_start(struct machine *machine,
 704                                              const char **symbol_name)
 705 {
 706         char filename[PATH_MAX];
 707         int i;
 708         const char *name;
 709         u64 addr = 0;
 710
 711         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 712
 713         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 714                 return 0;
 715
 716         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 717                 addr = kallsyms__get_function_start(filename, name);
 718                 if (addr)
 719                         break;
 720         }
 721
 722         if (symbol_name)
 723                 *symbol_name = name;
 724
 725         return addr;
 726 }
 727
 728 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
 729 {
 730         enum map_type type;
 731         u64 start = machine__get_running_kernel_start(machine, NULL);
 732
 733         for (type = 0; type < MAP__NR_TYPES; ++type) {
 734                 struct kmap *kmap;
 735
 736                 machine->vmlinux_maps[type] = map__new2(start, kernel, type);
 737                 if (machine->vmlinux_maps[type] == NULL)
 738                         return -1;
 739
 740                 machine->vmlinux_maps[type]->map_ip =
 741                         machine->vmlinux_maps[type]->unmap_ip =
 742                                 identity__map_ip;
 743                 kmap = map__kmap(machine->vmlinux_maps[type]);
 744                 if (!kmap)
 745                         return -1;
 746
 747                 kmap->kmaps = &machine->kmaps;
 748                 map_groups__insert(&machine->kmaps,
 749                                    machine->vmlinux_maps[type]);
 750         }
 751
 752         return 0;
 753 }
 754
 755 void machine__destroy_kernel_maps(struct machine *machine)
 756 {
 757         enum map_type type;
 758
 759         for (type = 0; type < MAP__NR_TYPES; ++type) {
 760                 struct kmap *kmap;
 761
 762                 if (machine->vmlinux_maps[type] == NULL)
 763                         continue;
 764
 765                 kmap = map__kmap(machine->vmlinux_maps[type]);
 766                 map_groups__remove(&machine->kmaps,
 767                                    machine->vmlinux_maps[type]);
 768                 if (kmap && kmap->ref_reloc_sym) {
 769                         /*
 770                          * ref_reloc_sym is shared among all maps, so free just
 771                          * on one of them.
 772                          */
 773                         if (type == MAP__FUNCTION) {
 774                                 zfree((char **)&kmap->ref_reloc_sym->name);
 775                                 zfree(&kmap->ref_reloc_sym);
 776                         } else
 777                                 kmap->ref_reloc_sym = NULL;
 778                 }
 779
 780                 machine->vmlinux_maps[type] = NULL;
 781         }
 782 }
 783
 784 int machines__create_guest_kernel_maps(struct machines *machines)
 785 {
 786         int ret = 0;
 787         struct dirent **namelist = NULL;
 788         int i, items = 0;
 789         char path[PATH_MAX];
 790         pid_t pid;
 791         char *endp;
 792
 793         if (symbol_conf.default_guest_vmlinux_name ||
 794             symbol_conf.default_guest_modules ||
 795             symbol_conf.default_guest_kallsyms) {
 796                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
 797         }
 798
 799         if (symbol_conf.guestmount) {
 800                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
 801                 if (items <= 0)
 802                         return -ENOENT;
 803                 for (i = 0; i < items; i++) {
 804                         if (!isdigit(namelist[i]->d_name[0])) {
 805                                 /* Filter out . and .. */
 806                                 continue;
 807                         }
 808                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
 809                         if ((*endp != '\0') ||
 810                             (endp == namelist[i]->d_name) ||
 811                             (errno == ERANGE)) {
 812                                 pr_debug("invalid directory (%s). Skipping.\n",
 813                                          namelist[i]->d_name);
 814                                 continue;
 815                         }
 816                         sprintf(path, "%s/%s/proc/kallsyms",
 817                                 symbol_conf.guestmount,
 818                                 namelist[i]->d_name);
 819                         ret = access(path, R_OK);
 820                         if (ret) {
 821                                 pr_debug("Can't access file %s\n", path);
 822                                 goto failure;
 823                         }
 824                         machines__create_kernel_maps(machines, pid);
 825                 }
 826 failure:
 827                 free(namelist);
 828         }
 829
 830         return ret;
 831 }
 832
 833 void machines__destroy_kernel_maps(struct machines *machines)
 834 {
 835         struct rb_node *next = rb_first(&machines->guests);
 836
 837         machine__destroy_kernel_maps(&machines->host);
 838
 839         while (next) {
 840                 struct machine *pos = rb_entry(next, struct machine, rb_node);
 841
 842                 next = rb_next(&pos->rb_node);
 843                 rb_erase(&pos->rb_node, &machines->guests);
 844                 machine__delete(pos);
 845         }
 846 }
 847
 848 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
 849 {
 850         struct machine *machine = machines__findnew(machines, pid);
 851
 852         if (machine == NULL)
 853                 return -1;
 854
 855         return machine__create_kernel_maps(machine);
 856 }
 857
 858 int machine__load_kallsyms(struct machine *machine, const char *filename,
 859                            enum map_type type, symbol_filter_t filter)
 860 {
 861         struct map *map = machine->vmlinux_maps[type];
 862         int ret = dso__load_kallsyms(map->dso, filename, map, filter);
 863
 864         if (ret > 0) {
 865                 dso__set_loaded(map->dso, type);
 866                 /*
 867                  * Since /proc/kallsyms will have multiple sessions for the
 868                  * kernel, with modules between them, fixup the end of all
 869                  * sections.
 870                  */
 871                 __map_groups__fixup_end(&machine->kmaps, type);
 872         }
 873
 874         return ret;
 875 }
 876
 877 int machine__load_vmlinux_path(struct machine *machine, enum map_type type,
 878                                symbol_filter_t filter)
 879 {
 880         struct map *map = machine->vmlinux_maps[type];
 881         int ret = dso__load_vmlinux_path(map->dso, map, filter);
 882
 883         if (ret > 0)
 884                 dso__set_loaded(map->dso, type);
 885
 886         return ret;
 887 }
 888
 889 static void map_groups__fixup_end(struct map_groups *mg)
 890 {
 891         int i;
 892         for (i = 0; i < MAP__NR_TYPES; ++i)
 893                 __map_groups__fixup_end(mg, i);
 894 }
 895
 896 static char *get_kernel_version(const char *root_dir)
 897 {
 898         char version[PATH_MAX];
 899         FILE *file;
 900         char *name, *tmp;
 901         const char *prefix = "Linux version ";
 902
 903         sprintf(version, "%s/proc/version", root_dir);
 904         file = fopen(version, "r");
 905         if (!file)
 906                 return NULL;
 907
 908         version[0] = '\0';
 909         tmp = fgets(version, sizeof(version), file);
 910         fclose(file);
 911
 912         name = strstr(version, prefix);
 913         if (!name)
 914                 return NULL;
 915         name += strlen(prefix);
 916         tmp = strchr(name, ' ');
 917         if (tmp)
 918                 *tmp = '\0';
 919
 920         return strdup(name);
 921 }
 922
 923 static bool is_kmod_dso(struct dso *dso)
 924 {
 925         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
 926                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
 927 }
 928
 929 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
 930                                        struct kmod_path *m)
 931 {
 932         struct map *map;
 933         char *long_name;
 934
 935         map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name);
 936         if (map == NULL)
 937                 return 0;
 938
 939         long_name = strdup(path);
 940         if (long_name == NULL)
 941                 return -ENOMEM;
 942
 943         dso__set_long_name(map->dso, long_name, true);
 944         dso__kernel_module_get_build_id(map->dso, "");
 945
 946         /*
 947          * Full name could reveal us kmod compression, so
 948          * we need to update the symtab_type if needed.
 949          */
 950         if (m->comp && is_kmod_dso(map->dso))
 951                 map->dso->symtab_type++;
 952
 953         return 0;
 954 }
 955
 956 static int map_groups__set_modules_path_dir(struct map_groups *mg,
 957                                 const char *dir_name, int depth)
 958 {
 959         struct dirent *dent;
 960         DIR *dir = opendir(dir_name);
 961         int ret = 0;
 962
 963         if (!dir) {
 964                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
 965                 return -1;
 966         }
 967
 968         while ((dent = readdir(dir)) != NULL) {
 969                 char path[PATH_MAX];
 970                 struct stat st;
 971
 972                 /*sshfs might return bad dent->d_type, so we have to stat*/
 973                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
 974                 if (stat(path, &st))
 975                         continue;
 976
 977                 if (S_ISDIR(st.st_mode)) {
 978                         if (!strcmp(dent->d_name, ".") ||
 979                             !strcmp(dent->d_name, ".."))
 980                                 continue;
 981
 982                         /* Do not follow top-level source and build symlinks */
 983                         if (depth == 0) {
 984                                 if (!strcmp(dent->d_name, "source") ||
 985                                     !strcmp(dent->d_name, "build"))
 986                                         continue;
 987                         }
 988
 989                         ret = map_groups__set_modules_path_dir(mg, path,
 990                                                                depth + 1);
 991                         if (ret < 0)
 992                                 goto out;
 993                 } else {
 994                         struct kmod_path m;
 995
 996                         ret = kmod_path__parse_name(&m, dent->d_name);
 997                         if (ret)
 998                                 goto out;
 999
1000                         if (m.kmod)
1001                                 ret = map_groups__set_module_path(mg, path, &m);
1002
1003                         free(m.name);
1004
1005                         if (ret)
1006                                 goto out;
1007                 }
1008         }
1009
1010 out:
1011         closedir(dir);
1012         return ret;
1013 }
1014
1015 static int machine__set_modules_path(struct machine *machine)
1016 {
1017         char *version;
1018         char modules_path[PATH_MAX];
1019
1020         version = get_kernel_version(machine->root_dir);
1021         if (!version)
1022                 return -1;
1023
1024         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1025                  machine->root_dir, version);
1026         free(version);
1027
1028         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1029 }
1030
1031 static int machine__create_module(void *arg, const char *name, u64 start)
1032 {
1033         struct machine *machine = arg;
1034         struct map *map;
1035
1036         map = machine__findnew_module_map(machine, start, name);
1037         if (map == NULL)
1038                 return -1;
1039
1040         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1041
1042         return 0;
1043 }
1044
1045 static int machine__create_modules(struct machine *machine)
1046 {
1047         const char *modules;
1048         char path[PATH_MAX];
1049
1050         if (machine__is_default_guest(machine)) {
1051                 modules = symbol_conf.default_guest_modules;
1052         } else {
1053                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1054                 modules = path;
1055         }
1056
1057         if (symbol__restricted_filename(modules, "/proc/modules"))
1058                 return -1;
1059
1060         if (modules__parse(modules, machine, machine__create_module))
1061                 return -1;
1062
1063         if (!machine__set_modules_path(machine))
1064                 return 0;
1065
1066         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1067
1068         return 0;
1069 }
1070
1071 int machine__create_kernel_maps(struct machine *machine)
1072 {
1073         struct dso *kernel = machine__get_kernel(machine);
1074         const char *name;
1075         u64 addr = machine__get_running_kernel_start(machine, &name);
1076         if (!addr)
1077                 return -1;
1078
1079         if (kernel == NULL ||
1080             __machine__create_kernel_maps(machine, kernel) < 0)
1081                 return -1;
1082
1083         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1084                 if (machine__is_host(machine))
1085                         pr_debug("Problems creating module maps, "
1086                                  "continuing anyway...\n");
1087                 else
1088                         pr_debug("Problems creating module maps for guest %d, "
1089                                  "continuing anyway...\n", machine->pid);
1090         }
1091
1092         /*
1093          * Now that we have all the maps created, just set the ->end of them:
1094          */
1095         map_groups__fixup_end(&machine->kmaps);
1096
1097         if (maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name,
1098                                              addr)) {
1099                 machine__destroy_kernel_maps(machine);
1100                 return -1;
1101         }
1102
1103         return 0;
1104 }
1105
1106 static void machine__set_kernel_mmap_len(struct machine *machine,
1107                                          union perf_event *event)
1108 {
1109         int i;
1110
1111         for (i = 0; i < MAP__NR_TYPES; i++) {
1112                 machine->vmlinux_maps[i]->start = event->mmap.start;
1113                 machine->vmlinux_maps[i]->end   = (event->mmap.start +
1114                                                    event->mmap.len);
1115                 /*
1116                  * Be a bit paranoid here, some perf.data file came with
1117                  * a zero sized synthesized MMAP event for the kernel.
1118                  */
1119                 if (machine->vmlinux_maps[i]->end == 0)
1120                         machine->vmlinux_maps[i]->end = ~0ULL;
1121         }
1122 }
1123
1124 static bool machine__uses_kcore(struct machine *machine)
1125 {
1126         struct dso *dso;
1127
1128         list_for_each_entry(dso, &machine->dsos.head, node) {
1129                 if (dso__is_kcore(dso))
1130                         return true;
1131         }
1132
1133         return false;
1134 }
1135
1136 static int machine__process_kernel_mmap_event(struct machine *machine,
1137                                               union perf_event *event)
1138 {
1139         struct map *map;
1140         char kmmap_prefix[PATH_MAX];
1141         enum dso_kernel_type kernel_type;
1142         bool is_kernel_mmap;
1143
1144         /* If we have maps from kcore then we do not need or want any others */
1145         if (machine__uses_kcore(machine))
1146                 return 0;
1147
1148         machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix));
1149         if (machine__is_host(machine))
1150                 kernel_type = DSO_TYPE_KERNEL;
1151         else
1152                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1153
1154         is_kernel_mmap = memcmp(event->mmap.filename,
1155                                 kmmap_prefix,
1156                                 strlen(kmmap_prefix) - 1) == 0;
1157         if (event->mmap.filename[0] == '/' ||
1158             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1159                 map = machine__findnew_module_map(machine, event->mmap.start,
1160                                                   event->mmap.filename);
1161                 if (map == NULL)
1162                         goto out_problem;
1163
1164                 map->end = map->start + event->mmap.len;
1165         } else if (is_kernel_mmap) {
1166                 const char *symbol_name = (event->mmap.filename +
1167                                 strlen(kmmap_prefix));
1168                 /*
1169                  * Should be there already, from the build-id table in
1170                  * the header.
1171                  */
1172                 struct dso *kernel = NULL;
1173                 struct dso *dso;
1174
1175                 pthread_rwlock_rdlock(&machine->dsos.lock);
1176
1177                 list_for_each_entry(dso, &machine->dsos.head, node) {
1178
1179                         /*
1180                          * The cpumode passed to is_kernel_module is not the
1181                          * cpumode of *this* event. If we insist on passing
1182                          * correct cpumode to is_kernel_module, we should
1183                          * record the cpumode when we adding this dso to the
1184                          * linked list.
1185                          *
1186                          * However we don't really need passing correct
1187                          * cpumode.  We know the correct cpumode must be kernel
1188                          * mode (if not, we should not link it onto kernel_dsos
1189                          * list).
1190                          *
1191                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1192                          * is_kernel_module() treats it as a kernel cpumode.
1193                          */
1194
1195                         if (!dso->kernel ||
1196                             is_kernel_module(dso->long_name,
1197                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1198                                 continue;
1199
1200
1201                         kernel = dso;
1202                         break;
1203                 }
1204
1205                 pthread_rwlock_unlock(&machine->dsos.lock);
1206
1207                 if (kernel == NULL)
1208                         kernel = machine__findnew_dso(machine, kmmap_prefix);
1209                 if (kernel == NULL)
1210                         goto out_problem;
1211
1212                 kernel->kernel = kernel_type;
1213                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1214                         dso__put(kernel);
1215                         goto out_problem;
1216                 }
1217
1218                 if (strstr(kernel->long_name, "vmlinux"))
1219                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1220
1221                 machine__set_kernel_mmap_len(machine, event);
1222
1223                 /*
1224                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1225                  * symbol. Effectively having zero here means that at record
1226                  * time /proc/sys/kernel/kptr_restrict was non zero.
1227                  */
1228                 if (event->mmap.pgoff != 0) {
1229                         maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps,
1230                                                          symbol_name,
1231                                                          event->mmap.pgoff);
1232                 }
1233
1234                 if (machine__is_default_guest(machine)) {
1235                         /*
1236                          * preload dso of guest kernel and modules
1237                          */
1238                         dso__load(kernel, machine->vmlinux_maps[MAP__FUNCTION],
1239                                   NULL);
1240                 }
1241         }
1242         return 0;
1243 out_problem:
1244         return -1;
1245 }
1246
1247 int machine__process_mmap2_event(struct machine *machine,
1248                                  union perf_event *event,
1249                                  struct perf_sample *sample __maybe_unused)
1250 {
1251         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1252         struct thread *thread;
1253         struct map *map;
1254         enum map_type type;
1255         int ret = 0;
1256
1257         if (dump_trace)
1258                 perf_event__fprintf_mmap2(event, stdout);
1259
1260         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1261             cpumode == PERF_RECORD_MISC_KERNEL) {
1262                 ret = machine__process_kernel_mmap_event(machine, event);
1263                 if (ret < 0)
1264                         goto out_problem;
1265                 return 0;
1266         }
1267
1268         thread = machine__findnew_thread(machine, event->mmap2.pid,
1269                                         event->mmap2.tid);
1270         if (thread == NULL)
1271                 goto out_problem;
1272
1273         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1274                 type = MAP__VARIABLE;
1275         else
1276                 type = MAP__FUNCTION;
1277
1278         map = map__new(machine, event->mmap2.start,
1279                         event->mmap2.len, event->mmap2.pgoff,
1280                         event->mmap2.pid, event->mmap2.maj,
1281                         event->mmap2.min, event->mmap2.ino,
1282                         event->mmap2.ino_generation,
1283                         event->mmap2.prot,
1284                         event->mmap2.flags,
1285                         event->mmap2.filename, type, thread);
1286
1287         if (map == NULL)
1288                 goto out_problem_map;
1289
1290         thread__insert_map(thread, map);
1291         thread__put(thread);
1292         map__put(map);
1293         return 0;
1294
1295 out_problem_map:
1296         thread__put(thread);
1297 out_problem:
1298         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1299         return 0;
1300 }
1301
1302 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1303                                 struct perf_sample *sample __maybe_unused)
1304 {
1305         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1306         struct thread *thread;
1307         struct map *map;
1308         enum map_type type;
1309         int ret = 0;
1310
1311         if (dump_trace)
1312                 perf_event__fprintf_mmap(event, stdout);
1313
1314         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1315             cpumode == PERF_RECORD_MISC_KERNEL) {
1316                 ret = machine__process_kernel_mmap_event(machine, event);
1317                 if (ret < 0)
1318                         goto out_problem;
1319                 return 0;
1320         }
1321
1322         thread = machine__findnew_thread(machine, event->mmap.pid,
1323                                          event->mmap.tid);
1324         if (thread == NULL)
1325                 goto out_problem;
1326
1327         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1328                 type = MAP__VARIABLE;
1329         else
1330                 type = MAP__FUNCTION;
1331
1332         map = map__new(machine, event->mmap.start,
1333                         event->mmap.len, event->mmap.pgoff,
1334                         event->mmap.pid, 0, 0, 0, 0, 0, 0,
1335                         event->mmap.filename,
1336                         type, thread);
1337
1338         if (map == NULL)
1339                 goto out_problem_map;
1340
1341         thread__insert_map(thread, map);
1342         thread__put(thread);
1343         map__put(map);
1344         return 0;
1345
1346 out_problem_map:
1347         thread__put(thread);
1348 out_problem:
1349         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1350         return 0;
1351 }
1352
1353 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1354 {
1355         if (machine->last_match == th)
1356                 machine->last_match = NULL;
1357
1358         BUG_ON(atomic_read(&th->refcnt) == 0);
1359         if (lock)
1360                 pthread_rwlock_wrlock(&machine->threads_lock);
1361         rb_erase_init(&th->rb_node, &machine->threads);
1362         RB_CLEAR_NODE(&th->rb_node);
1363         /*
1364          * Move it first to the dead_threads list, then drop the reference,
1365          * if this is the last reference, then the thread__delete destructor
1366          * will be called and we will remove it from the dead_threads list.
1367          */
1368         list_add_tail(&th->node, &machine->dead_threads);
1369         if (lock)
1370                 pthread_rwlock_unlock(&machine->threads_lock);
1371         thread__put(th);
1372 }
1373
1374 void machine__remove_thread(struct machine *machine, struct thread *th)
1375 {
1376         return __machine__remove_thread(machine, th, true);
1377 }
1378
1379 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1380                                 struct perf_sample *sample)
1381 {
1382         struct thread *thread = machine__find_thread(machine,
1383                                                      event->fork.pid,
1384                                                      event->fork.tid);
1385         struct thread *parent = machine__findnew_thread(machine,
1386                                                         event->fork.ppid,
1387                                                         event->fork.ptid);
1388         int err = 0;
1389
1390         /* if a thread currently exists for the thread id remove it */
1391         if (thread != NULL) {
1392                 machine__remove_thread(machine, thread);
1393                 thread__put(thread);
1394         }
1395
1396         thread = machine__findnew_thread(machine, event->fork.pid,
1397                                          event->fork.tid);
1398         if (dump_trace)
1399                 perf_event__fprintf_task(event, stdout);
1400
1401         if (thread == NULL || parent == NULL ||
1402             thread__fork(thread, parent, sample->time) < 0) {
1403                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1404                 err = -1;
1405         }
1406         thread__put(thread);
1407         thread__put(parent);
1408
1409         return err;
1410 }
1411
1412 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1413                                 struct perf_sample *sample __maybe_unused)
1414 {
1415         struct thread *thread = machine__find_thread(machine,
1416                                                      event->fork.pid,
1417                                                      event->fork.tid);
1418
1419         if (dump_trace)
1420                 perf_event__fprintf_task(event, stdout);
1421
1422         if (thread != NULL) {
1423                 thread__exited(thread);
1424                 thread__put(thread);
1425         }
1426
1427         return 0;
1428 }
1429
1430 int machine__process_event(struct machine *machine, union perf_event *event,
1431                            struct perf_sample *sample)
1432 {
1433         int ret;
1434
1435         switch (event->header.type) {
1436         case PERF_RECORD_COMM:
1437                 ret = machine__process_comm_event(machine, event, sample); break;
1438         case PERF_RECORD_MMAP:
1439                 ret = machine__process_mmap_event(machine, event, sample); break;
1440         case PERF_RECORD_MMAP2:
1441                 ret = machine__process_mmap2_event(machine, event, sample); break;
1442         case PERF_RECORD_FORK:
1443                 ret = machine__process_fork_event(machine, event, sample); break;
1444         case PERF_RECORD_EXIT:
1445                 ret = machine__process_exit_event(machine, event, sample); break;
1446         case PERF_RECORD_LOST:
1447                 ret = machine__process_lost_event(machine, event, sample); break;
1448         case PERF_RECORD_AUX:
1449                 ret = machine__process_aux_event(machine, event); break;
1450         case PERF_RECORD_ITRACE_START:
1451                 ret = machine__process_itrace_start_event(machine, event); break;
1452         case PERF_RECORD_LOST_SAMPLES:
1453                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1454         default:
1455                 ret = -1;
1456                 break;
1457         }
1458
1459         return ret;
1460 }
1461
1462 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1463 {
1464         if (sym->name && !regexec(regex, sym->name, 0, NULL, 0))
1465                 return 1;
1466         return 0;
1467 }
1468
1469 static void ip__resolve_ams(struct thread *thread,
1470                             struct addr_map_symbol *ams,
1471                             u64 ip)
1472 {
1473         struct addr_location al;
1474
1475         memset(&al, 0, sizeof(al));
1476         /*
1477          * We cannot use the header.misc hint to determine whether a
1478          * branch stack address is user, kernel, guest, hypervisor.
1479          * Branches may straddle the kernel/user/hypervisor boundaries.
1480          * Thus, we have to try consecutively until we find a match
1481          * or else, the symbol is unknown
1482          */
1483         thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al);
1484
1485         ams->addr = ip;
1486         ams->al_addr = al.addr;
1487         ams->sym = al.sym;
1488         ams->map = al.map;
1489 }
1490
1491 static void ip__resolve_data(struct thread *thread,
1492                              u8 m, struct addr_map_symbol *ams, u64 addr)
1493 {
1494         struct addr_location al;
1495
1496         memset(&al, 0, sizeof(al));
1497
1498         thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al);
1499         if (al.map == NULL) {
1500                 /*
1501                  * some shared data regions have execute bit set which puts
1502                  * their mapping in the MAP__FUNCTION type array.
1503                  * Check there as a fallback option before dropping the sample.
1504                  */
1505                 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al);
1506         }
1507
1508         ams->addr = addr;
1509         ams->al_addr = al.addr;
1510         ams->sym = al.sym;
1511         ams->map = al.map;
1512 }
1513
1514 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1515                                      struct addr_location *al)
1516 {
1517         struct mem_info *mi = zalloc(sizeof(*mi));
1518
1519         if (!mi)
1520                 return NULL;
1521
1522         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1523         ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr);
1524         mi->data_src.val = sample->data_src;
1525
1526         return mi;
1527 }
1528
1529 static int add_callchain_ip(struct thread *thread,
1530                             struct symbol **parent,
1531                             struct addr_location *root_al,
1532                             u8 *cpumode,
1533                             u64 ip)
1534 {
1535         struct addr_location al;
1536
1537         al.filtered = 0;
1538         al.sym = NULL;
1539         if (!cpumode) {
1540                 thread__find_cpumode_addr_location(thread, MAP__FUNCTION,
1541                                                    ip, &al);
1542         } else {
1543                 if (ip >= PERF_CONTEXT_MAX) {
1544                         switch (ip) {
1545                         case PERF_CONTEXT_HV:
1546                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
1547                                 break;
1548                         case PERF_CONTEXT_KERNEL:
1549                                 *cpumode = PERF_RECORD_MISC_KERNEL;
1550                                 break;
1551                         case PERF_CONTEXT_USER:
1552                                 *cpumode = PERF_RECORD_MISC_USER;
1553                                 break;
1554                         default:
1555                                 pr_debug("invalid callchain context: "
1556                                          "%"PRId64"\n", (s64) ip);
1557                                 /*
1558                                  * It seems the callchain is corrupted.
1559                                  * Discard all.
1560                                  */
1561                                 callchain_cursor_reset(&callchain_cursor);
1562                                 return 1;
1563                         }
1564                         return 0;
1565                 }
1566                 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION,
1567                                            ip, &al);
1568         }
1569
1570         if (al.sym != NULL) {
1571                 if (sort__has_parent && !*parent &&
1572                     symbol__match_regex(al.sym, &parent_regex))
1573                         *parent = al.sym;
1574                 else if (have_ignore_callees && root_al &&
1575                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
1576                         /* Treat this symbol as the root,
1577                            forgetting its callees. */
1578                         *root_al = al;
1579                         callchain_cursor_reset(&callchain_cursor);
1580                 }
1581         }
1582
1583         return callchain_cursor_append(&callchain_cursor, al.addr, al.map, al.sym);
1584 }
1585
1586 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
1587                                            struct addr_location *al)
1588 {
1589         unsigned int i;
1590         const struct branch_stack *bs = sample->branch_stack;
1591         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
1592
1593         if (!bi)
1594                 return NULL;
1595
1596         for (i = 0; i < bs->nr; i++) {
1597                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
1598                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
1599                 bi[i].flags = bs->entries[i].flags;
1600         }
1601         return bi;
1602 }
1603
1604 #define CHASHSZ 127
1605 #define CHASHBITS 7
1606 #define NO_ENTRY 0xff
1607
1608 #define PERF_MAX_BRANCH_DEPTH 127
1609
1610 /* Remove loops. */
1611 static int remove_loops(struct branch_entry *l, int nr)
1612 {
1613         int i, j, off;
1614         unsigned char chash[CHASHSZ];
1615
1616         memset(chash, NO_ENTRY, sizeof(chash));
1617
1618         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
1619
1620         for (i = 0; i < nr; i++) {
1621                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
1622
1623                 /* no collision handling for now */
1624                 if (chash[h] == NO_ENTRY) {
1625                         chash[h] = i;
1626                 } else if (l[chash[h]].from == l[i].from) {
1627                         bool is_loop = true;
1628                         /* check if it is a real loop */
1629                         off = 0;
1630                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
1631                                 if (l[j].from != l[i + off].from) {
1632                                         is_loop = false;
1633                                         break;
1634                                 }
1635                         if (is_loop) {
1636                                 memmove(l + i, l + i + off,
1637                                         (nr - (i + off)) * sizeof(*l));
1638                                 nr -= off;
1639                         }
1640                 }
1641         }
1642         return nr;
1643 }
1644
1645 /*
1646  * Recolve LBR callstack chain sample
1647  * Return:
1648  * 1 on success get LBR callchain information
1649  * 0 no available LBR callchain information, should try fp
1650  * negative error code on other errors.
1651  */
1652 static int resolve_lbr_callchain_sample(struct thread *thread,
1653                                         struct perf_sample *sample,
1654                                         struct symbol **parent,
1655                                         struct addr_location *root_al,
1656                                         int max_stack)
1657 {
1658         struct ip_callchain *chain = sample->callchain;
1659         int chain_nr = min(max_stack, (int)chain->nr);
1660         u8 cpumode = PERF_RECORD_MISC_USER;
1661         int i, j, err;
1662         u64 ip;
1663
1664         for (i = 0; i < chain_nr; i++) {
1665                 if (chain->ips[i] == PERF_CONTEXT_USER)
1666                         break;
1667         }
1668
1669         /* LBR only affects the user callchain */
1670         if (i != chain_nr) {
1671                 struct branch_stack *lbr_stack = sample->branch_stack;
1672                 int lbr_nr = lbr_stack->nr;
1673                 /*
1674                  * LBR callstack can only get user call chain.
1675                  * The mix_chain_nr is kernel call chain
1676                  * number plus LBR user call chain number.
1677                  * i is kernel call chain number,
1678                  * 1 is PERF_CONTEXT_USER,
1679                  * lbr_nr + 1 is the user call chain number.
1680                  * For details, please refer to the comments
1681                  * in callchain__printf
1682                  */
1683                 int mix_chain_nr = i + 1 + lbr_nr + 1;
1684
1685                 if (mix_chain_nr > PERF_MAX_STACK_DEPTH + PERF_MAX_BRANCH_DEPTH) {
1686                         pr_warning("corrupted callchain. skipping...\n");
1687                         return 0;
1688                 }
1689
1690                 for (j = 0; j < mix_chain_nr; j++) {
1691                         if (callchain_param.order == ORDER_CALLEE) {
1692                                 if (j < i + 1)
1693                                         ip = chain->ips[j];
1694                                 else if (j > i + 1)
1695                                         ip = lbr_stack->entries[j - i - 2].from;
1696                                 else
1697                                         ip = lbr_stack->entries[0].to;
1698                         } else {
1699                                 if (j < lbr_nr)
1700                                         ip = lbr_stack->entries[lbr_nr - j - 1].from;
1701                                 else if (j > lbr_nr)
1702                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
1703                                 else
1704                                         ip = lbr_stack->entries[0].to;
1705                         }
1706
1707                         err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1708                         if (err)
1709                                 return (err < 0) ? err : 0;
1710                 }
1711                 return 1;
1712         }
1713
1714         return 0;
1715 }
1716
1717 static int thread__resolve_callchain_sample(struct thread *thread,
1718                                             struct perf_evsel *evsel,
1719                                             struct perf_sample *sample,
1720                                             struct symbol **parent,
1721                                             struct addr_location *root_al,
1722                                             int max_stack)
1723 {
1724         struct branch_stack *branch = sample->branch_stack;
1725         struct ip_callchain *chain = sample->callchain;
1726         int chain_nr = min(max_stack, (int)chain->nr);
1727         u8 cpumode = PERF_RECORD_MISC_USER;
1728         int i, j, err;
1729         int skip_idx = -1;
1730         int first_call = 0;
1731
1732         callchain_cursor_reset(&callchain_cursor);
1733
1734         if (has_branch_callstack(evsel)) {
1735                 err = resolve_lbr_callchain_sample(thread, sample, parent,
1736                                                    root_al, max_stack);
1737                 if (err)
1738                         return (err < 0) ? err : 0;
1739         }
1740
1741         /*
1742          * Based on DWARF debug information, some architectures skip
1743          * a callchain entry saved by the kernel.
1744          */
1745         if (chain->nr < PERF_MAX_STACK_DEPTH)
1746                 skip_idx = arch_skip_callchain_idx(thread, chain);
1747
1748         /*
1749          * Add branches to call stack for easier browsing. This gives
1750          * more context for a sample than just the callers.
1751          *
1752          * This uses individual histograms of paths compared to the
1753          * aggregated histograms the normal LBR mode uses.
1754          *
1755          * Limitations for now:
1756          * - No extra filters
1757          * - No annotations (should annotate somehow)
1758          */
1759
1760         if (branch && callchain_param.branch_callstack) {
1761                 int nr = min(max_stack, (int)branch->nr);
1762                 struct branch_entry be[nr];
1763
1764                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
1765                         pr_warning("corrupted branch chain. skipping...\n");
1766                         goto check_calls;
1767                 }
1768
1769                 for (i = 0; i < nr; i++) {
1770                         if (callchain_param.order == ORDER_CALLEE) {
1771                                 be[i] = branch->entries[i];
1772                                 /*
1773                                  * Check for overlap into the callchain.
1774                                  * The return address is one off compared to
1775                                  * the branch entry. To adjust for this
1776                                  * assume the calling instruction is not longer
1777                                  * than 8 bytes.
1778                                  */
1779                                 if (i == skip_idx ||
1780                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
1781                                         first_call++;
1782                                 else if (be[i].from < chain->ips[first_call] &&
1783                                     be[i].from >= chain->ips[first_call] - 8)
1784                                         first_call++;
1785                         } else
1786                                 be[i] = branch->entries[branch->nr - i - 1];
1787                 }
1788
1789                 nr = remove_loops(be, nr);
1790
1791                 for (i = 0; i < nr; i++) {
1792                         err = add_callchain_ip(thread, parent, root_al,
1793                                                NULL, be[i].to);
1794                         if (!err)
1795                                 err = add_callchain_ip(thread, parent, root_al,
1796                                                        NULL, be[i].from);
1797                         if (err == -EINVAL)
1798                                 break;
1799                         if (err)
1800                                 return err;
1801                 }
1802                 chain_nr -= nr;
1803         }
1804
1805 check_calls:
1806         if (chain->nr > PERF_MAX_STACK_DEPTH) {
1807                 pr_warning("corrupted callchain. skipping...\n");
1808                 return 0;
1809         }
1810
1811         for (i = first_call; i < chain_nr; i++) {
1812                 u64 ip;
1813
1814                 if (callchain_param.order == ORDER_CALLEE)
1815                         j = i;
1816                 else
1817                         j = chain->nr - i - 1;
1818
1819 #ifdef HAVE_SKIP_CALLCHAIN_IDX
1820                 if (j == skip_idx)
1821                         continue;
1822 #endif
1823                 ip = chain->ips[j];
1824
1825                 err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1826
1827                 if (err)
1828                         return (err < 0) ? err : 0;
1829         }
1830
1831         return 0;
1832 }
1833
1834 static int unwind_entry(struct unwind_entry *entry, void *arg)
1835 {
1836         struct callchain_cursor *cursor = arg;
1837         return callchain_cursor_append(cursor, entry->ip,
1838                                        entry->map, entry->sym);
1839 }
1840
1841 int thread__resolve_callchain(struct thread *thread,
1842                               struct perf_evsel *evsel,
1843                               struct perf_sample *sample,
1844                               struct symbol **parent,
1845                               struct addr_location *root_al,
1846                               int max_stack)
1847 {
1848         int ret = thread__resolve_callchain_sample(thread, evsel,
1849                                                    sample, parent,
1850                                                    root_al, max_stack);
1851         if (ret)
1852                 return ret;
1853
1854         /* Can we do dwarf post unwind? */
1855         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
1856               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
1857                 return 0;
1858
1859         /* Bail out if nothing was captured. */
1860         if ((!sample->user_regs.regs) ||
1861             (!sample->user_stack.size))
1862                 return 0;
1863
1864         return unwind__get_entries(unwind_entry, &callchain_cursor,
1865                                    thread, sample, max_stack);
1866
1867 }
1868
1869 int machine__for_each_thread(struct machine *machine,
1870                              int (*fn)(struct thread *thread, void *p),
1871                              void *priv)
1872 {
1873         struct rb_node *nd;
1874         struct thread *thread;
1875         int rc = 0;
1876
1877         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
1878                 thread = rb_entry(nd, struct thread, rb_node);
1879                 rc = fn(thread, priv);
1880                 if (rc != 0)
1881                         return rc;
1882         }
1883
1884         list_for_each_entry(thread, &machine->dead_threads, node) {
1885                 rc = fn(thread, priv);
1886                 if (rc != 0)
1887                         return rc;
1888         }
1889         return rc;
1890 }
1891
1892 int machines__for_each_thread(struct machines *machines,
1893                               int (*fn)(struct thread *thread, void *p),
1894                               void *priv)
1895 {
1896         struct rb_node *nd;
1897         int rc = 0;
1898
1899         rc = machine__for_each_thread(&machines->host, fn, priv);
1900         if (rc != 0)
1901                 return rc;
1902
1903         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
1904                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
1905
1906                 rc = machine__for_each_thread(machine, fn, priv);
1907                 if (rc != 0)
1908                         return rc;
1909         }
1910         return rc;
1911 }
1912
1913 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1914                                   struct target *target, struct thread_map *threads,
1915                                   perf_event__handler_t process, bool data_mmap,
1916                                   unsigned int proc_map_timeout)
1917 {
1918         if (target__has_task(target))
1919                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap, proc_map_timeout);
1920         else if (target__has_cpu(target))
1921                 return perf_event__synthesize_threads(tool, process, machine, data_mmap, proc_map_timeout);
1922         /* command specified */
1923         return 0;
1924 }
1925
1926 pid_t machine__get_current_tid(struct machine *machine, int cpu)
1927 {
1928         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
1929                 return -1;
1930
1931         return machine->current_tid[cpu];
1932 }
1933
1934 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
1935                              pid_t tid)
1936 {
1937         struct thread *thread;
1938
1939         if (cpu < 0)
1940                 return -EINVAL;
1941
1942         if (!machine->current_tid) {
1943                 int i;
1944
1945                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
1946                 if (!machine->current_tid)
1947                         return -ENOMEM;
1948                 for (i = 0; i < MAX_NR_CPUS; i++)
1949                         machine->current_tid[i] = -1;
1950         }
1951
1952         if (cpu >= MAX_NR_CPUS) {
1953                 pr_err("Requested CPU %d too large. ", cpu);
1954                 pr_err("Consider raising MAX_NR_CPUS\n");
1955                 return -EINVAL;
1956         }
1957
1958         machine->current_tid[cpu] = tid;
1959
1960         thread = machine__findnew_thread(machine, pid, tid);
1961         if (!thread)
1962                 return -ENOMEM;
1963
1964         thread->cpu = cpu;
1965         thread__put(thread);
1966
1967         return 0;
1968 }
1969
1970 int machine__get_kernel_start(struct machine *machine)
1971 {
1972         struct map *map = machine__kernel_map(machine, MAP__FUNCTION);
1973         int err = 0;
1974
1975         /*
1976          * The only addresses above 2^63 are kernel addresses of a 64-bit
1977          * kernel.  Note that addresses are unsigned so that on a 32-bit system
1978          * all addresses including kernel addresses are less than 2^32.  In
1979          * that case (32-bit system), if the kernel mapping is unknown, all
1980          * addresses will be assumed to be in user space - see
1981          * machine__kernel_ip().
1982          */
1983         machine->kernel_start = 1ULL << 63;
1984         if (map) {
1985                 err = map__load(map, machine->symbol_filter);
1986                 if (map->start)
1987                         machine->kernel_start = map->start;
1988         }
1989         return err;
1990 }
1991
1992 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
1993 {
1994         return dsos__findnew(&machine->dsos, filename);
1995 }