tools/perf/util/machine.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <dirent.h>
   3 #include <errno.h>
   4 #include <inttypes.h>
   5 #include <regex.h>
   6 #include "callchain.h"
   7 #include "debug.h"
   8 #include "event.h"
   9 #include "evsel.h"
  10 #include "hist.h"
  11 #include "machine.h"
  12 #include "map.h"
  13 #include "symbol.h"
  14 #include "sort.h"
  15 #include "strlist.h"
  16 #include "thread.h"
  17 #include "vdso.h"
  18 #include <stdbool.h>
  19 #include <sys/types.h>
  20 #include <sys/stat.h>
  21 #include <unistd.h>
  22 #include "unwind.h"
  23 #include "linux/hash.h"
  24 #include "asm/bug.h"
  25 #include "bpf-event.h"
  26
  27 #include "sane_ctype.h"
  28 #include <symbol/kallsyms.h>
  29 #include <linux/mman.h>
  30
  31 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  32
  33 static void dsos__init(struct dsos *dsos)
  34 {
  35         INIT_LIST_HEAD(&dsos->head);
  36         dsos->root = RB_ROOT;
  37         init_rwsem(&dsos->lock);
  38 }
  39
  40 static void machine__threads_init(struct machine *machine)
  41 {
  42         int i;
  43
  44         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  45                 struct threads *threads = &machine->threads[i];
  46                 threads->entries = RB_ROOT_CACHED;
  47                 init_rwsem(&threads->lock);
  48                 threads->nr = 0;
  49                 INIT_LIST_HEAD(&threads->dead);
  50                 threads->last_match = NULL;
  51         }
  52 }
  53
  54 static int machine__set_mmap_name(struct machine *machine)
  55 {
  56         if (machine__is_host(machine))
  57                 machine->mmap_name = strdup("[kernel.kallsyms]");
  58         else if (machine__is_default_guest(machine))
  59                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  60         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  61                           machine->pid) < 0)
  62                 machine->mmap_name = NULL;
  63
  64         return machine->mmap_name ? 0 : -ENOMEM;
  65 }
  66
  67 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  68 {
  69         int err = -ENOMEM;
  70
  71         memset(machine, 0, sizeof(*machine));
  72         map_groups__init(&machine->kmaps, machine);
  73         RB_CLEAR_NODE(&machine->rb_node);
  74         dsos__init(&machine->dsos);
  75
  76         machine__threads_init(machine);
  77
  78         machine->vdso_info = NULL;
  79         machine->env = NULL;
  80
  81         machine->pid = pid;
  82
  83         machine->id_hdr_size = 0;
  84         machine->kptr_restrict_warned = false;
  85         machine->comm_exec = false;
  86         machine->kernel_start = 0;
  87         machine->vmlinux_map = NULL;
  88
  89         machine->root_dir = strdup(root_dir);
  90         if (machine->root_dir == NULL)
  91                 return -ENOMEM;
  92
  93         if (machine__set_mmap_name(machine))
  94                 goto out;
  95
  96         if (pid != HOST_KERNEL_ID) {
  97                 struct thread *thread = machine__findnew_thread(machine, -1,
  98                                                                 pid);
  99                 char comm[64];
 100
 101                 if (thread == NULL)
 102                         goto out;
 103
 104                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 105                 thread__set_comm(thread, comm, 0);
 106                 thread__put(thread);
 107         }
 108
 109         machine->current_tid = NULL;
 110         err = 0;
 111
 112 out:
 113         if (err) {
 114                 zfree(&machine->root_dir);
 115                 zfree(&machine->mmap_name);
 116         }
 117         return 0;
 118 }
 119
 120 struct machine *machine__new_host(void)
 121 {
 122         struct machine *machine = malloc(sizeof(*machine));
 123
 124         if (machine != NULL) {
 125                 machine__init(machine, "", HOST_KERNEL_ID);
 126
 127                 if (machine__create_kernel_maps(machine) < 0)
 128                         goto out_delete;
 129         }
 130
 131         return machine;
 132 out_delete:
 133         free(machine);
 134         return NULL;
 135 }
 136
 137 struct machine *machine__new_kallsyms(void)
 138 {
 139         struct machine *machine = machine__new_host();
 140         /*
 141          * FIXME:
 142          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 143          *    ask for not using the kcore parsing code, once this one is fixed
 144          *    to create a map per module.
 145          */
 146         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 147                 machine__delete(machine);
 148                 machine = NULL;
 149         }
 150
 151         return machine;
 152 }
 153
 154 static void dsos__purge(struct dsos *dsos)
 155 {
 156         struct dso *pos, *n;
 157
 158         down_write(&dsos->lock);
 159
 160         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 161                 RB_CLEAR_NODE(&pos->rb_node);
 162                 pos->root = NULL;
 163                 list_del_init(&pos->node);
 164                 dso__put(pos);
 165         }
 166
 167         up_write(&dsos->lock);
 168 }
 169
 170 static void dsos__exit(struct dsos *dsos)
 171 {
 172         dsos__purge(dsos);
 173         exit_rwsem(&dsos->lock);
 174 }
 175
 176 void machine__delete_threads(struct machine *machine)
 177 {
 178         struct rb_node *nd;
 179         int i;
 180
 181         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 182                 struct threads *threads = &machine->threads[i];
 183                 down_write(&threads->lock);
 184                 nd = rb_first_cached(&threads->entries);
 185                 while (nd) {
 186                         struct thread *t = rb_entry(nd, struct thread, rb_node);
 187
 188                         nd = rb_next(nd);
 189                         __machine__remove_thread(machine, t, false);
 190                 }
 191                 up_write(&threads->lock);
 192         }
 193 }
 194
 195 void machine__exit(struct machine *machine)
 196 {
 197         int i;
 198
 199         if (machine == NULL)
 200                 return;
 201
 202         machine__destroy_kernel_maps(machine);
 203         map_groups__exit(&machine->kmaps);
 204         dsos__exit(&machine->dsos);
 205         machine__exit_vdso(machine);
 206         zfree(&machine->root_dir);
 207         zfree(&machine->mmap_name);
 208         zfree(&machine->current_tid);
 209
 210         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 211                 struct threads *threads = &machine->threads[i];
 212                 exit_rwsem(&threads->lock);
 213         }
 214 }
 215
 216 void machine__delete(struct machine *machine)
 217 {
 218         if (machine) {
 219                 machine__exit(machine);
 220                 free(machine);
 221         }
 222 }
 223
 224 void machines__init(struct machines *machines)
 225 {
 226         machine__init(&machines->host, "", HOST_KERNEL_ID);
 227         machines->guests = RB_ROOT_CACHED;
 228 }
 229
 230 void machines__exit(struct machines *machines)
 231 {
 232         machine__exit(&machines->host);
 233         /* XXX exit guest */
 234 }
 235
 236 struct machine *machines__add(struct machines *machines, pid_t pid,
 237                               const char *root_dir)
 238 {
 239         struct rb_node **p = &machines->guests.rb_root.rb_node;
 240         struct rb_node *parent = NULL;
 241         struct machine *pos, *machine = malloc(sizeof(*machine));
 242         bool leftmost = true;
 243
 244         if (machine == NULL)
 245                 return NULL;
 246
 247         if (machine__init(machine, root_dir, pid) != 0) {
 248                 free(machine);
 249                 return NULL;
 250         }
 251
 252         while (*p != NULL) {
 253                 parent = *p;
 254                 pos = rb_entry(parent, struct machine, rb_node);
 255                 if (pid < pos->pid)
 256                         p = &(*p)->rb_left;
 257                 else {
 258                         p = &(*p)->rb_right;
 259                         leftmost = false;
 260                 }
 261         }
 262
 263         rb_link_node(&machine->rb_node, parent, p);
 264         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 265
 266         return machine;
 267 }
 268
 269 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 270 {
 271         struct rb_node *nd;
 272
 273         machines->host.comm_exec = comm_exec;
 274
 275         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 276                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 277
 278                 machine->comm_exec = comm_exec;
 279         }
 280 }
 281
 282 struct machine *machines__find(struct machines *machines, pid_t pid)
 283 {
 284         struct rb_node **p = &machines->guests.rb_root.rb_node;
 285         struct rb_node *parent = NULL;
 286         struct machine *machine;
 287         struct machine *default_machine = NULL;
 288
 289         if (pid == HOST_KERNEL_ID)
 290                 return &machines->host;
 291
 292         while (*p != NULL) {
 293                 parent = *p;
 294                 machine = rb_entry(parent, struct machine, rb_node);
 295                 if (pid < machine->pid)
 296                         p = &(*p)->rb_left;
 297                 else if (pid > machine->pid)
 298                         p = &(*p)->rb_right;
 299                 else
 300                         return machine;
 301                 if (!machine->pid)
 302                         default_machine = machine;
 303         }
 304
 305         return default_machine;
 306 }
 307
 308 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 309 {
 310         char path[PATH_MAX];
 311         const char *root_dir = "";
 312         struct machine *machine = machines__find(machines, pid);
 313
 314         if (machine && (machine->pid == pid))
 315                 goto out;
 316
 317         if ((pid != HOST_KERNEL_ID) &&
 318             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 319             (symbol_conf.guestmount)) {
 320                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 321                 if (access(path, R_OK)) {
 322                         static struct strlist *seen;
 323
 324                         if (!seen)
 325                                 seen = strlist__new(NULL, NULL);
 326
 327                         if (!strlist__has_entry(seen, path)) {
 328                                 pr_err("Can't access file %s\n", path);
 329                                 strlist__add(seen, path);
 330                         }
 331                         machine = NULL;
 332                         goto out;
 333                 }
 334                 root_dir = path;
 335         }
 336
 337         machine = machines__add(machines, pid, root_dir);
 338 out:
 339         return machine;
 340 }
 341
 342 void machines__process_guests(struct machines *machines,
 343                               machine__process_t process, void *data)
 344 {
 345         struct rb_node *nd;
 346
 347         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 348                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 349                 process(pos, data);
 350         }
 351 }
 352
 353 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 354 {
 355         struct rb_node *node;
 356         struct machine *machine;
 357
 358         machines->host.id_hdr_size = id_hdr_size;
 359
 360         for (node = rb_first_cached(&machines->guests); node;
 361              node = rb_next(node)) {
 362                 machine = rb_entry(node, struct machine, rb_node);
 363                 machine->id_hdr_size = id_hdr_size;
 364         }
 365
 366         return;
 367 }
 368
 369 static void machine__update_thread_pid(struct machine *machine,
 370                                        struct thread *th, pid_t pid)
 371 {
 372         struct thread *leader;
 373
 374         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 375                 return;
 376
 377         th->pid_ = pid;
 378
 379         if (th->pid_ == th->tid)
 380                 return;
 381
 382         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 383         if (!leader)
 384                 goto out_err;
 385
 386         if (!leader->mg)
 387                 leader->mg = map_groups__new(machine);
 388
 389         if (!leader->mg)
 390                 goto out_err;
 391
 392         if (th->mg == leader->mg)
 393                 return;
 394
 395         if (th->mg) {
 396                 /*
 397                  * Maps are created from MMAP events which provide the pid and
 398                  * tid.  Consequently there never should be any maps on a thread
 399                  * with an unknown pid.  Just print an error if there are.
 400                  */
 401                 if (!map_groups__empty(th->mg))
 402                         pr_err("Discarding thread maps for %d:%d\n",
 403                                th->pid_, th->tid);
 404                 map_groups__put(th->mg);
 405         }
 406
 407         th->mg = map_groups__get(leader->mg);
 408 out_put:
 409         thread__put(leader);
 410         return;
 411 out_err:
 412         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 413         goto out_put;
 414 }
 415
 416 /*
 417  * Front-end cache - TID lookups come in blocks,
 418  * so most of the time we dont have to look up
 419  * the full rbtree:
 420  */
 421 static struct thread*
 422 __threads__get_last_match(struct threads *threads, struct machine *machine,
 423                           int pid, int tid)
 424 {
 425         struct thread *th;
 426
 427         th = threads->last_match;
 428         if (th != NULL) {
 429                 if (th->tid == tid) {
 430                         machine__update_thread_pid(machine, th, pid);
 431                         return thread__get(th);
 432                 }
 433
 434                 threads->last_match = NULL;
 435         }
 436
 437         return NULL;
 438 }
 439
 440 static struct thread*
 441 threads__get_last_match(struct threads *threads, struct machine *machine,
 442                         int pid, int tid)
 443 {
 444         struct thread *th = NULL;
 445
 446         if (perf_singlethreaded)
 447                 th = __threads__get_last_match(threads, machine, pid, tid);
 448
 449         return th;
 450 }
 451
 452 static void
 453 __threads__set_last_match(struct threads *threads, struct thread *th)
 454 {
 455         threads->last_match = th;
 456 }
 457
 458 static void
 459 threads__set_last_match(struct threads *threads, struct thread *th)
 460 {
 461         if (perf_singlethreaded)
 462                 __threads__set_last_match(threads, th);
 463 }
 464
 465 /*
 466  * Caller must eventually drop thread->refcnt returned with a successful
 467  * lookup/new thread inserted.
 468  */
 469 static struct thread *____machine__findnew_thread(struct machine *machine,
 470                                                   struct threads *threads,
 471                                                   pid_t pid, pid_t tid,
 472                                                   bool create)
 473 {
 474         struct rb_node **p = &threads->entries.rb_root.rb_node;
 475         struct rb_node *parent = NULL;
 476         struct thread *th;
 477         bool leftmost = true;
 478
 479         th = threads__get_last_match(threads, machine, pid, tid);
 480         if (th)
 481                 return th;
 482
 483         while (*p != NULL) {
 484                 parent = *p;
 485                 th = rb_entry(parent, struct thread, rb_node);
 486
 487                 if (th->tid == tid) {
 488                         threads__set_last_match(threads, th);
 489                         machine__update_thread_pid(machine, th, pid);
 490                         return thread__get(th);
 491                 }
 492
 493                 if (tid < th->tid)
 494                         p = &(*p)->rb_left;
 495                 else {
 496                         p = &(*p)->rb_right;
 497                         leftmost = false;
 498                 }
 499         }
 500
 501         if (!create)
 502                 return NULL;
 503
 504         th = thread__new(pid, tid);
 505         if (th != NULL) {
 506                 rb_link_node(&th->rb_node, parent, p);
 507                 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
 508
 509                 /*
 510                  * We have to initialize map_groups separately
 511                  * after rb tree is updated.
 512                  *
 513                  * The reason is that we call machine__findnew_thread
 514                  * within thread__init_map_groups to find the thread
 515                  * leader and that would screwed the rb tree.
 516                  */
 517                 if (thread__init_map_groups(th, machine)) {
 518                         rb_erase_cached(&th->rb_node, &threads->entries);
 519                         RB_CLEAR_NODE(&th->rb_node);
 520                         thread__put(th);
 521                         return NULL;
 522                 }
 523                 /*
 524                  * It is now in the rbtree, get a ref
 525                  */
 526                 thread__get(th);
 527                 threads__set_last_match(threads, th);
 528                 ++threads->nr;
 529         }
 530
 531         return th;
 532 }
 533
 534 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 535 {
 536         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 537 }
 538
 539 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 540                                        pid_t tid)
 541 {
 542         struct threads *threads = machine__threads(machine, tid);
 543         struct thread *th;
 544
 545         down_write(&threads->lock);
 546         th = __machine__findnew_thread(machine, pid, tid);
 547         up_write(&threads->lock);
 548         return th;
 549 }
 550
 551 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 552                                     pid_t tid)
 553 {
 554         struct threads *threads = machine__threads(machine, tid);
 555         struct thread *th;
 556
 557         down_read(&threads->lock);
 558         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 559         up_read(&threads->lock);
 560         return th;
 561 }
 562
 563 struct comm *machine__thread_exec_comm(struct machine *machine,
 564                                        struct thread *thread)
 565 {
 566         if (machine->comm_exec)
 567                 return thread__exec_comm(thread);
 568         else
 569                 return thread__comm(thread);
 570 }
 571
 572 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 573                                 struct perf_sample *sample)
 574 {
 575         struct thread *thread = machine__findnew_thread(machine,
 576                                                         event->comm.pid,
 577                                                         event->comm.tid);
 578         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 579         int err = 0;
 580
 581         if (exec)
 582                 machine->comm_exec = true;
 583
 584         if (dump_trace)
 585                 perf_event__fprintf_comm(event, stdout);
 586
 587         if (thread == NULL ||
 588             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 589                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 590                 err = -1;
 591         }
 592
 593         thread__put(thread);
 594
 595         return err;
 596 }
 597
 598 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 599                                       union perf_event *event,
 600                                       struct perf_sample *sample __maybe_unused)
 601 {
 602         struct thread *thread = machine__findnew_thread(machine,
 603                                                         event->namespaces.pid,
 604                                                         event->namespaces.tid);
 605         int err = 0;
 606
 607         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 608                   "\nWARNING: kernel seems to support more namespaces than perf"
 609                   " tool.\nTry updating the perf tool..\n\n");
 610
 611         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 612                   "\nWARNING: perf tool seems to support more namespaces than"
 613                   " the kernel.\nTry updating the kernel..\n\n");
 614
 615         if (dump_trace)
 616                 perf_event__fprintf_namespaces(event, stdout);
 617
 618         if (thread == NULL ||
 619             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 620                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 621                 err = -1;
 622         }
 623
 624         thread__put(thread);
 625
 626         return err;
 627 }
 628
 629 int machine__process_lost_event(struct machine *machine __maybe_unused,
 630                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 631 {
 632         dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
 633                     event->lost.id, event->lost.lost);
 634         return 0;
 635 }
 636
 637 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 638                                         union perf_event *event, struct perf_sample *sample)
 639 {
 640         dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
 641                     sample->id, event->lost_samples.lost);
 642         return 0;
 643 }
 644
 645 static struct dso *machine__findnew_module_dso(struct machine *machine,
 646                                                struct kmod_path *m,
 647                                                const char *filename)
 648 {
 649         struct dso *dso;
 650
 651         down_write(&machine->dsos.lock);
 652
 653         dso = __dsos__find(&machine->dsos, m->name, true);
 654         if (!dso) {
 655                 dso = __dsos__addnew(&machine->dsos, m->name);
 656                 if (dso == NULL)
 657                         goto out_unlock;
 658
 659                 dso__set_module_info(dso, m, machine);
 660                 dso__set_long_name(dso, strdup(filename), true);
 661         }
 662
 663         dso__get(dso);
 664 out_unlock:
 665         up_write(&machine->dsos.lock);
 666         return dso;
 667 }
 668
 669 int machine__process_aux_event(struct machine *machine __maybe_unused,
 670                                union perf_event *event)
 671 {
 672         if (dump_trace)
 673                 perf_event__fprintf_aux(event, stdout);
 674         return 0;
 675 }
 676
 677 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 678                                         union perf_event *event)
 679 {
 680         if (dump_trace)
 681                 perf_event__fprintf_itrace_start(event, stdout);
 682         return 0;
 683 }
 684
 685 int machine__process_switch_event(struct machine *machine __maybe_unused,
 686                                   union perf_event *event)
 687 {
 688         if (dump_trace)
 689                 perf_event__fprintf_switch(event, stdout);
 690         return 0;
 691 }
 692
 693 static int machine__process_ksymbol_register(struct machine *machine,
 694                                              union perf_event *event,
 695                                              struct perf_sample *sample __maybe_unused)
 696 {
 697         struct symbol *sym;
 698         struct map *map;
 699
 700         map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
 701         if (!map) {
 702                 map = dso__new_map(event->ksymbol_event.name);
 703                 if (!map)
 704                         return -ENOMEM;
 705
 706                 map->start = event->ksymbol_event.addr;
 707                 map->pgoff = map->start;
 708                 map->end = map->start + event->ksymbol_event.len;
 709                 map_groups__insert(&machine->kmaps, map);
 710         }
 711
 712         sym = symbol__new(event->ksymbol_event.addr, event->ksymbol_event.len,
 713                           0, 0, event->ksymbol_event.name);
 714         if (!sym)
 715                 return -ENOMEM;
 716         dso__insert_symbol(map->dso, sym);
 717         return 0;
 718 }
 719
 720 static int machine__process_ksymbol_unregister(struct machine *machine,
 721                                                union perf_event *event,
 722                                                struct perf_sample *sample __maybe_unused)
 723 {
 724         struct map *map;
 725
 726         map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
 727         if (map)
 728                 map_groups__remove(&machine->kmaps, map);
 729
 730         return 0;
 731 }
 732
 733 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 734                              union perf_event *event,
 735                              struct perf_sample *sample)
 736 {
 737         if (dump_trace)
 738                 perf_event__fprintf_ksymbol(event, stdout);
 739
 740         if (event->ksymbol_event.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 741                 return machine__process_ksymbol_unregister(machine, event,
 742                                                            sample);
 743         return machine__process_ksymbol_register(machine, event, sample);
 744 }
 745
 746 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
 747 {
 748         const char *dup_filename;
 749
 750         if (!filename || !dso || !dso->long_name)
 751                 return;
 752         if (dso->long_name[0] != '[')
 753                 return;
 754         if (!strchr(filename, '/'))
 755                 return;
 756
 757         dup_filename = strdup(filename);
 758         if (!dup_filename)
 759                 return;
 760
 761         dso__set_long_name(dso, dup_filename, true);
 762 }
 763
 764 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 765                                         const char *filename)
 766 {
 767         struct map *map = NULL;
 768         struct dso *dso = NULL;
 769         struct kmod_path m;
 770
 771         if (kmod_path__parse_name(&m, filename))
 772                 return NULL;
 773
 774         map = map_groups__find_by_name(&machine->kmaps, m.name);
 775         if (map) {
 776                 /*
 777                  * If the map's dso is an offline module, give dso__load()
 778                  * a chance to find the file path of that module by fixing
 779                  * long_name.
 780                  */
 781                 dso__adjust_kmod_long_name(map->dso, filename);
 782                 goto out;
 783         }
 784
 785         dso = machine__findnew_module_dso(machine, &m, filename);
 786         if (dso == NULL)
 787                 goto out;
 788
 789         map = map__new2(start, dso);
 790         if (map == NULL)
 791                 goto out;
 792
 793         map_groups__insert(&machine->kmaps, map);
 794
 795         /* Put the map here because map_groups__insert alread got it */
 796         map__put(map);
 797 out:
 798         /* put the dso here, corresponding to  machine__findnew_module_dso */
 799         dso__put(dso);
 800         free(m.name);
 801         return map;
 802 }
 803
 804 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 805 {
 806         struct rb_node *nd;
 807         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 808
 809         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 810                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 811                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 812         }
 813
 814         return ret;
 815 }
 816
 817 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 818                                      bool (skip)(struct dso *dso, int parm), int parm)
 819 {
 820         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 821 }
 822
 823 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 824                                      bool (skip)(struct dso *dso, int parm), int parm)
 825 {
 826         struct rb_node *nd;
 827         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 828
 829         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 830                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 831                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 832         }
 833         return ret;
 834 }
 835
 836 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 837 {
 838         int i;
 839         size_t printed = 0;
 840         struct dso *kdso = machine__kernel_map(machine)->dso;
 841
 842         if (kdso->has_build_id) {
 843                 char filename[PATH_MAX];
 844                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
 845                                            false))
 846                         printed += fprintf(fp, "[0] %s\n", filename);
 847         }
 848
 849         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 850                 printed += fprintf(fp, "[%d] %s\n",
 851                                    i + kdso->has_build_id, vmlinux_path[i]);
 852
 853         return printed;
 854 }
 855
 856 size_t machine__fprintf(struct machine *machine, FILE *fp)
 857 {
 858         struct rb_node *nd;
 859         size_t ret;
 860         int i;
 861
 862         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 863                 struct threads *threads = &machine->threads[i];
 864
 865                 down_read(&threads->lock);
 866
 867                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
 868
 869                 for (nd = rb_first_cached(&threads->entries); nd;
 870                      nd = rb_next(nd)) {
 871                         struct thread *pos = rb_entry(nd, struct thread, rb_node);
 872
 873                         ret += thread__fprintf(pos, fp);
 874                 }
 875
 876                 up_read(&threads->lock);
 877         }
 878         return ret;
 879 }
 880
 881 static struct dso *machine__get_kernel(struct machine *machine)
 882 {
 883         const char *vmlinux_name = machine->mmap_name;
 884         struct dso *kernel;
 885
 886         if (machine__is_host(machine)) {
 887                 if (symbol_conf.vmlinux_name)
 888                         vmlinux_name = symbol_conf.vmlinux_name;
 889
 890                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 891                                                  "[kernel]", DSO_TYPE_KERNEL);
 892         } else {
 893                 if (symbol_conf.default_guest_vmlinux_name)
 894                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 895
 896                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 897                                                  "[guest.kernel]",
 898                                                  DSO_TYPE_GUEST_KERNEL);
 899         }
 900
 901         if (kernel != NULL && (!kernel->has_build_id))
 902                 dso__read_running_kernel_build_id(kernel, machine);
 903
 904         return kernel;
 905 }
 906
 907 struct process_args {
 908         u64 start;
 909 };
 910
 911 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 912                                     size_t bufsz)
 913 {
 914         if (machine__is_default_guest(machine))
 915                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 916         else
 917                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 918 }
 919
 920 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 921
 922 /* Figure out the start address of kernel map from /proc/kallsyms.
 923  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 924  * symbol_name if it's not that important.
 925  */
 926 static int machine__get_running_kernel_start(struct machine *machine,
 927                                              const char **symbol_name,
 928                                              u64 *start, u64 *end)
 929 {
 930         char filename[PATH_MAX];
 931         int i, err = -1;
 932         const char *name;
 933         u64 addr = 0;
 934
 935         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 936
 937         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 938                 return 0;
 939
 940         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 941                 err = kallsyms__get_function_start(filename, name, &addr);
 942                 if (!err)
 943                         break;
 944         }
 945
 946         if (err)
 947                 return -1;
 948
 949         if (symbol_name)
 950                 *symbol_name = name;
 951
 952         *start = addr;
 953
 954         err = kallsyms__get_function_start(filename, "_etext", &addr);
 955         if (!err)
 956                 *end = addr;
 957
 958         return 0;
 959 }
 960
 961 int machine__create_extra_kernel_map(struct machine *machine,
 962                                      struct dso *kernel,
 963                                      struct extra_kernel_map *xm)
 964 {
 965         struct kmap *kmap;
 966         struct map *map;
 967
 968         map = map__new2(xm->start, kernel);
 969         if (!map)
 970                 return -1;
 971
 972         map->end   = xm->end;
 973         map->pgoff = xm->pgoff;
 974
 975         kmap = map__kmap(map);
 976
 977         kmap->kmaps = &machine->kmaps;
 978         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
 979
 980         map_groups__insert(&machine->kmaps, map);
 981
 982         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
 983                   kmap->name, map->start, map->end);
 984
 985         map__put(map);
 986
 987         return 0;
 988 }
 989
 990 static u64 find_entry_trampoline(struct dso *dso)
 991 {
 992         /* Duplicates are removed so lookup all aliases */
 993         const char *syms[] = {
 994                 "_entry_trampoline",
 995                 "__entry_trampoline_start",
 996                 "entry_SYSCALL_64_trampoline",
 997         };
 998         struct symbol *sym = dso__first_symbol(dso);
 999         unsigned int i;
1000
1001         for (; sym; sym = dso__next_symbol(sym)) {
1002                 if (sym->binding != STB_GLOBAL)
1003                         continue;
1004                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1005                         if (!strcmp(sym->name, syms[i]))
1006                                 return sym->start;
1007                 }
1008         }
1009
1010         return 0;
1011 }
1012
1013 /*
1014  * These values can be used for kernels that do not have symbols for the entry
1015  * trampolines in kallsyms.
1016  */
1017 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1018 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1019 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1020
1021 /* Map x86_64 PTI entry trampolines */
1022 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1023                                           struct dso *kernel)
1024 {
1025         struct map_groups *kmaps = &machine->kmaps;
1026         struct maps *maps = &kmaps->maps;
1027         int nr_cpus_avail, cpu;
1028         bool found = false;
1029         struct map *map;
1030         u64 pgoff;
1031
1032         /*
1033          * In the vmlinux case, pgoff is a virtual address which must now be
1034          * mapped to a vmlinux offset.
1035          */
1036         for (map = maps__first(maps); map; map = map__next(map)) {
1037                 struct kmap *kmap = __map__kmap(map);
1038                 struct map *dest_map;
1039
1040                 if (!kmap || !is_entry_trampoline(kmap->name))
1041                         continue;
1042
1043                 dest_map = map_groups__find(kmaps, map->pgoff);
1044                 if (dest_map != map)
1045                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1046                 found = true;
1047         }
1048         if (found || machine->trampolines_mapped)
1049                 return 0;
1050
1051         pgoff = find_entry_trampoline(kernel);
1052         if (!pgoff)
1053                 return 0;
1054
1055         nr_cpus_avail = machine__nr_cpus_avail(machine);
1056
1057         /* Add a 1 page map for each CPU's entry trampoline */
1058         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1059                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1060                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1061                          X86_64_ENTRY_TRAMPOLINE;
1062                 struct extra_kernel_map xm = {
1063                         .start = va,
1064                         .end   = va + page_size,
1065                         .pgoff = pgoff,
1066                 };
1067
1068                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1069
1070                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1071                         return -1;
1072         }
1073
1074         machine->trampolines_mapped = nr_cpus_avail;
1075
1076         return 0;
1077 }
1078
1079 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1080                                              struct dso *kernel __maybe_unused)
1081 {
1082         return 0;
1083 }
1084
1085 static int
1086 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1087 {
1088         struct kmap *kmap;
1089         struct map *map;
1090
1091         /* In case of renewal the kernel map, destroy previous one */
1092         machine__destroy_kernel_maps(machine);
1093
1094         machine->vmlinux_map = map__new2(0, kernel);
1095         if (machine->vmlinux_map == NULL)
1096                 return -1;
1097
1098         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1099         map = machine__kernel_map(machine);
1100         kmap = map__kmap(map);
1101         if (!kmap)
1102                 return -1;
1103
1104         kmap->kmaps = &machine->kmaps;
1105         map_groups__insert(&machine->kmaps, map);
1106
1107         return 0;
1108 }
1109
1110 void machine__destroy_kernel_maps(struct machine *machine)
1111 {
1112         struct kmap *kmap;
1113         struct map *map = machine__kernel_map(machine);
1114
1115         if (map == NULL)
1116                 return;
1117
1118         kmap = map__kmap(map);
1119         map_groups__remove(&machine->kmaps, map);
1120         if (kmap && kmap->ref_reloc_sym) {
1121                 zfree((char **)&kmap->ref_reloc_sym->name);
1122                 zfree(&kmap->ref_reloc_sym);
1123         }
1124
1125         map__zput(machine->vmlinux_map);
1126 }
1127
1128 int machines__create_guest_kernel_maps(struct machines *machines)
1129 {
1130         int ret = 0;
1131         struct dirent **namelist = NULL;
1132         int i, items = 0;
1133         char path[PATH_MAX];
1134         pid_t pid;
1135         char *endp;
1136
1137         if (symbol_conf.default_guest_vmlinux_name ||
1138             symbol_conf.default_guest_modules ||
1139             symbol_conf.default_guest_kallsyms) {
1140                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1141         }
1142
1143         if (symbol_conf.guestmount) {
1144                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1145                 if (items <= 0)
1146                         return -ENOENT;
1147                 for (i = 0; i < items; i++) {
1148                         if (!isdigit(namelist[i]->d_name[0])) {
1149                                 /* Filter out . and .. */
1150                                 continue;
1151                         }
1152                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1153                         if ((*endp != '\0') ||
1154                             (endp == namelist[i]->d_name) ||
1155                             (errno == ERANGE)) {
1156                                 pr_debug("invalid directory (%s). Skipping.\n",
1157                                          namelist[i]->d_name);
1158                                 continue;
1159                         }
1160                         sprintf(path, "%s/%s/proc/kallsyms",
1161                                 symbol_conf.guestmount,
1162                                 namelist[i]->d_name);
1163                         ret = access(path, R_OK);
1164                         if (ret) {
1165                                 pr_debug("Can't access file %s\n", path);
1166                                 goto failure;
1167                         }
1168                         machines__create_kernel_maps(machines, pid);
1169                 }
1170 failure:
1171                 free(namelist);
1172         }
1173
1174         return ret;
1175 }
1176
1177 void machines__destroy_kernel_maps(struct machines *machines)
1178 {
1179         struct rb_node *next = rb_first_cached(&machines->guests);
1180
1181         machine__destroy_kernel_maps(&machines->host);
1182
1183         while (next) {
1184                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1185
1186                 next = rb_next(&pos->rb_node);
1187                 rb_erase_cached(&pos->rb_node, &machines->guests);
1188                 machine__delete(pos);
1189         }
1190 }
1191
1192 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1193 {
1194         struct machine *machine = machines__findnew(machines, pid);
1195
1196         if (machine == NULL)
1197                 return -1;
1198
1199         return machine__create_kernel_maps(machine);
1200 }
1201
1202 int machine__load_kallsyms(struct machine *machine, const char *filename)
1203 {
1204         struct map *map = machine__kernel_map(machine);
1205         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1206
1207         if (ret > 0) {
1208                 dso__set_loaded(map->dso);
1209                 /*
1210                  * Since /proc/kallsyms will have multiple sessions for the
1211                  * kernel, with modules between them, fixup the end of all
1212                  * sections.
1213                  */
1214                 map_groups__fixup_end(&machine->kmaps);
1215         }
1216
1217         return ret;
1218 }
1219
1220 int machine__load_vmlinux_path(struct machine *machine)
1221 {
1222         struct map *map = machine__kernel_map(machine);
1223         int ret = dso__load_vmlinux_path(map->dso, map);
1224
1225         if (ret > 0)
1226                 dso__set_loaded(map->dso);
1227
1228         return ret;
1229 }
1230
1231 static char *get_kernel_version(const char *root_dir)
1232 {
1233         char version[PATH_MAX];
1234         FILE *file;
1235         char *name, *tmp;
1236         const char *prefix = "Linux version ";
1237
1238         sprintf(version, "%s/proc/version", root_dir);
1239         file = fopen(version, "r");
1240         if (!file)
1241                 return NULL;
1242
1243         tmp = fgets(version, sizeof(version), file);
1244         if (!tmp)
1245                 *version = '\0';
1246         fclose(file);
1247
1248         name = strstr(version, prefix);
1249         if (!name)
1250                 return NULL;
1251         name += strlen(prefix);
1252         tmp = strchr(name, ' ');
1253         if (tmp)
1254                 *tmp = '\0';
1255
1256         return strdup(name);
1257 }
1258
1259 static bool is_kmod_dso(struct dso *dso)
1260 {
1261         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1262                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1263 }
1264
1265 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1266                                        struct kmod_path *m)
1267 {
1268         char *long_name;
1269         struct map *map = map_groups__find_by_name(mg, m->name);
1270
1271         if (map == NULL)
1272                 return 0;
1273
1274         long_name = strdup(path);
1275         if (long_name == NULL)
1276                 return -ENOMEM;
1277
1278         dso__set_long_name(map->dso, long_name, true);
1279         dso__kernel_module_get_build_id(map->dso, "");
1280
1281         /*
1282          * Full name could reveal us kmod compression, so
1283          * we need to update the symtab_type if needed.
1284          */
1285         if (m->comp && is_kmod_dso(map->dso)) {
1286                 map->dso->symtab_type++;
1287                 map->dso->comp = m->comp;
1288         }
1289
1290         return 0;
1291 }
1292
1293 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1294                                 const char *dir_name, int depth)
1295 {
1296         struct dirent *dent;
1297         DIR *dir = opendir(dir_name);
1298         int ret = 0;
1299
1300         if (!dir) {
1301                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1302                 return -1;
1303         }
1304
1305         while ((dent = readdir(dir)) != NULL) {
1306                 char path[PATH_MAX];
1307                 struct stat st;
1308
1309                 /*sshfs might return bad dent->d_type, so we have to stat*/
1310                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1311                 if (stat(path, &st))
1312                         continue;
1313
1314                 if (S_ISDIR(st.st_mode)) {
1315                         if (!strcmp(dent->d_name, ".") ||
1316                             !strcmp(dent->d_name, ".."))
1317                                 continue;
1318
1319                         /* Do not follow top-level source and build symlinks */
1320                         if (depth == 0) {
1321                                 if (!strcmp(dent->d_name, "source") ||
1322                                     !strcmp(dent->d_name, "build"))
1323                                         continue;
1324                         }
1325
1326                         ret = map_groups__set_modules_path_dir(mg, path,
1327                                                                depth + 1);
1328                         if (ret < 0)
1329                                 goto out;
1330                 } else {
1331                         struct kmod_path m;
1332
1333                         ret = kmod_path__parse_name(&m, dent->d_name);
1334                         if (ret)
1335                                 goto out;
1336
1337                         if (m.kmod)
1338                                 ret = map_groups__set_module_path(mg, path, &m);
1339
1340                         free(m.name);
1341
1342                         if (ret)
1343                                 goto out;
1344                 }
1345         }
1346
1347 out:
1348         closedir(dir);
1349         return ret;
1350 }
1351
1352 static int machine__set_modules_path(struct machine *machine)
1353 {
1354         char *version;
1355         char modules_path[PATH_MAX];
1356
1357         version = get_kernel_version(machine->root_dir);
1358         if (!version)
1359                 return -1;
1360
1361         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1362                  machine->root_dir, version);
1363         free(version);
1364
1365         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1366 }
1367 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1368                                 const char *name __maybe_unused)
1369 {
1370         return 0;
1371 }
1372
1373 static int machine__create_module(void *arg, const char *name, u64 start,
1374                                   u64 size)
1375 {
1376         struct machine *machine = arg;
1377         struct map *map;
1378
1379         if (arch__fix_module_text_start(&start, name) < 0)
1380                 return -1;
1381
1382         map = machine__findnew_module_map(machine, start, name);
1383         if (map == NULL)
1384                 return -1;
1385         map->end = start + size;
1386
1387         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1388
1389         return 0;
1390 }
1391
1392 static int machine__create_modules(struct machine *machine)
1393 {
1394         const char *modules;
1395         char path[PATH_MAX];
1396
1397         if (machine__is_default_guest(machine)) {
1398                 modules = symbol_conf.default_guest_modules;
1399         } else {
1400                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1401                 modules = path;
1402         }
1403
1404         if (symbol__restricted_filename(modules, "/proc/modules"))
1405                 return -1;
1406
1407         if (modules__parse(modules, machine, machine__create_module))
1408                 return -1;
1409
1410         if (!machine__set_modules_path(machine))
1411                 return 0;
1412
1413         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1414
1415         return 0;
1416 }
1417
1418 static void machine__set_kernel_mmap(struct machine *machine,
1419                                      u64 start, u64 end)
1420 {
1421         machine->vmlinux_map->start = start;
1422         machine->vmlinux_map->end   = end;
1423         /*
1424          * Be a bit paranoid here, some perf.data file came with
1425          * a zero sized synthesized MMAP event for the kernel.
1426          */
1427         if (start == 0 && end == 0)
1428                 machine->vmlinux_map->end = ~0ULL;
1429 }
1430
1431 static void machine__update_kernel_mmap(struct machine *machine,
1432                                      u64 start, u64 end)
1433 {
1434         struct map *map = machine__kernel_map(machine);
1435
1436         map__get(map);
1437         map_groups__remove(&machine->kmaps, map);
1438
1439         machine__set_kernel_mmap(machine, start, end);
1440
1441         map_groups__insert(&machine->kmaps, map);
1442         map__put(map);
1443 }
1444
1445 int machine__create_kernel_maps(struct machine *machine)
1446 {
1447         struct dso *kernel = machine__get_kernel(machine);
1448         const char *name = NULL;
1449         struct map *map;
1450         u64 start = 0, end = ~0ULL;
1451         int ret;
1452
1453         if (kernel == NULL)
1454                 return -1;
1455
1456         ret = __machine__create_kernel_maps(machine, kernel);
1457         if (ret < 0)
1458                 goto out_put;
1459
1460         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1461                 if (machine__is_host(machine))
1462                         pr_debug("Problems creating module maps, "
1463                                  "continuing anyway...\n");
1464                 else
1465                         pr_debug("Problems creating module maps for guest %d, "
1466                                  "continuing anyway...\n", machine->pid);
1467         }
1468
1469         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1470                 if (name &&
1471                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1472                         machine__destroy_kernel_maps(machine);
1473                         ret = -1;
1474                         goto out_put;
1475                 }
1476
1477                 /*
1478                  * we have a real start address now, so re-order the kmaps
1479                  * assume it's the last in the kmaps
1480                  */
1481                 machine__update_kernel_mmap(machine, start, end);
1482         }
1483
1484         if (machine__create_extra_kernel_maps(machine, kernel))
1485                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1486
1487         if (end == ~0ULL) {
1488                 /* update end address of the kernel map using adjacent module address */
1489                 map = map__next(machine__kernel_map(machine));
1490                 if (map)
1491                         machine__set_kernel_mmap(machine, start, map->start);
1492         }
1493
1494 out_put:
1495         dso__put(kernel);
1496         return ret;
1497 }
1498
1499 static bool machine__uses_kcore(struct machine *machine)
1500 {
1501         struct dso *dso;
1502
1503         list_for_each_entry(dso, &machine->dsos.head, node) {
1504                 if (dso__is_kcore(dso))
1505                         return true;
1506         }
1507
1508         return false;
1509 }
1510
1511 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1512                                              union perf_event *event)
1513 {
1514         return machine__is(machine, "x86_64") &&
1515                is_entry_trampoline(event->mmap.filename);
1516 }
1517
1518 static int machine__process_extra_kernel_map(struct machine *machine,
1519                                              union perf_event *event)
1520 {
1521         struct map *kernel_map = machine__kernel_map(machine);
1522         struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1523         struct extra_kernel_map xm = {
1524                 .start = event->mmap.start,
1525                 .end   = event->mmap.start + event->mmap.len,
1526                 .pgoff = event->mmap.pgoff,
1527         };
1528
1529         if (kernel == NULL)
1530                 return -1;
1531
1532         strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1533
1534         return machine__create_extra_kernel_map(machine, kernel, &xm);
1535 }
1536
1537 static int machine__process_kernel_mmap_event(struct machine *machine,
1538                                               union perf_event *event)
1539 {
1540         struct map *map;
1541         enum dso_kernel_type kernel_type;
1542         bool is_kernel_mmap;
1543
1544         /* If we have maps from kcore then we do not need or want any others */
1545         if (machine__uses_kcore(machine))
1546                 return 0;
1547
1548         if (machine__is_host(machine))
1549                 kernel_type = DSO_TYPE_KERNEL;
1550         else
1551                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1552
1553         is_kernel_mmap = memcmp(event->mmap.filename,
1554                                 machine->mmap_name,
1555                                 strlen(machine->mmap_name) - 1) == 0;
1556         if (event->mmap.filename[0] == '/' ||
1557             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1558                 map = machine__findnew_module_map(machine, event->mmap.start,
1559                                                   event->mmap.filename);
1560                 if (map == NULL)
1561                         goto out_problem;
1562
1563                 map->end = map->start + event->mmap.len;
1564         } else if (is_kernel_mmap) {
1565                 const char *symbol_name = (event->mmap.filename +
1566                                 strlen(machine->mmap_name));
1567                 /*
1568                  * Should be there already, from the build-id table in
1569                  * the header.
1570                  */
1571                 struct dso *kernel = NULL;
1572                 struct dso *dso;
1573
1574                 down_read(&machine->dsos.lock);
1575
1576                 list_for_each_entry(dso, &machine->dsos.head, node) {
1577
1578                         /*
1579                          * The cpumode passed to is_kernel_module is not the
1580                          * cpumode of *this* event. If we insist on passing
1581                          * correct cpumode to is_kernel_module, we should
1582                          * record the cpumode when we adding this dso to the
1583                          * linked list.
1584                          *
1585                          * However we don't really need passing correct
1586                          * cpumode.  We know the correct cpumode must be kernel
1587                          * mode (if not, we should not link it onto kernel_dsos
1588                          * list).
1589                          *
1590                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1591                          * is_kernel_module() treats it as a kernel cpumode.
1592                          */
1593
1594                         if (!dso->kernel ||
1595                             is_kernel_module(dso->long_name,
1596                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1597                                 continue;
1598
1599
1600                         kernel = dso;
1601                         break;
1602                 }
1603
1604                 up_read(&machine->dsos.lock);
1605
1606                 if (kernel == NULL)
1607                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1608                 if (kernel == NULL)
1609                         goto out_problem;
1610
1611                 kernel->kernel = kernel_type;
1612                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1613                         dso__put(kernel);
1614                         goto out_problem;
1615                 }
1616
1617                 if (strstr(kernel->long_name, "vmlinux"))
1618                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1619
1620                 machine__update_kernel_mmap(machine, event->mmap.start,
1621                                          event->mmap.start + event->mmap.len);
1622
1623                 /*
1624                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1625                  * symbol. Effectively having zero here means that at record
1626                  * time /proc/sys/kernel/kptr_restrict was non zero.
1627                  */
1628                 if (event->mmap.pgoff != 0) {
1629                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1630                                                         symbol_name,
1631                                                         event->mmap.pgoff);
1632                 }
1633
1634                 if (machine__is_default_guest(machine)) {
1635                         /*
1636                          * preload dso of guest kernel and modules
1637                          */
1638                         dso__load(kernel, machine__kernel_map(machine));
1639                 }
1640         } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1641                 return machine__process_extra_kernel_map(machine, event);
1642         }
1643         return 0;
1644 out_problem:
1645         return -1;
1646 }
1647
1648 int machine__process_mmap2_event(struct machine *machine,
1649                                  union perf_event *event,
1650                                  struct perf_sample *sample)
1651 {
1652         struct thread *thread;
1653         struct map *map;
1654         int ret = 0;
1655
1656         if (dump_trace)
1657                 perf_event__fprintf_mmap2(event, stdout);
1658
1659         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1660             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1661                 ret = machine__process_kernel_mmap_event(machine, event);
1662                 if (ret < 0)
1663                         goto out_problem;
1664                 return 0;
1665         }
1666
1667         thread = machine__findnew_thread(machine, event->mmap2.pid,
1668                                         event->mmap2.tid);
1669         if (thread == NULL)
1670                 goto out_problem;
1671
1672         map = map__new(machine, event->mmap2.start,
1673                         event->mmap2.len, event->mmap2.pgoff,
1674                         event->mmap2.maj,
1675                         event->mmap2.min, event->mmap2.ino,
1676                         event->mmap2.ino_generation,
1677                         event->mmap2.prot,
1678                         event->mmap2.flags,
1679                         event->mmap2.filename, thread);
1680
1681         if (map == NULL)
1682                 goto out_problem_map;
1683
1684         ret = thread__insert_map(thread, map);
1685         if (ret)
1686                 goto out_problem_insert;
1687
1688         thread__put(thread);
1689         map__put(map);
1690         return 0;
1691
1692 out_problem_insert:
1693         map__put(map);
1694 out_problem_map:
1695         thread__put(thread);
1696 out_problem:
1697         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1698         return 0;
1699 }
1700
1701 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1702                                 struct perf_sample *sample)
1703 {
1704         struct thread *thread;
1705         struct map *map;
1706         u32 prot = 0;
1707         int ret = 0;
1708
1709         if (dump_trace)
1710                 perf_event__fprintf_mmap(event, stdout);
1711
1712         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1713             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1714                 ret = machine__process_kernel_mmap_event(machine, event);
1715                 if (ret < 0)
1716                         goto out_problem;
1717                 return 0;
1718         }
1719
1720         thread = machine__findnew_thread(machine, event->mmap.pid,
1721                                          event->mmap.tid);
1722         if (thread == NULL)
1723                 goto out_problem;
1724
1725         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1726                 prot = PROT_EXEC;
1727
1728         map = map__new(machine, event->mmap.start,
1729                         event->mmap.len, event->mmap.pgoff,
1730                         0, 0, 0, 0, prot, 0,
1731                         event->mmap.filename,
1732                         thread);
1733
1734         if (map == NULL)
1735                 goto out_problem_map;
1736
1737         ret = thread__insert_map(thread, map);
1738         if (ret)
1739                 goto out_problem_insert;
1740
1741         thread__put(thread);
1742         map__put(map);
1743         return 0;
1744
1745 out_problem_insert:
1746         map__put(map);
1747 out_problem_map:
1748         thread__put(thread);
1749 out_problem:
1750         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1751         return 0;
1752 }
1753
1754 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1755 {
1756         struct threads *threads = machine__threads(machine, th->tid);
1757
1758         if (threads->last_match == th)
1759                 threads__set_last_match(threads, NULL);
1760
1761         BUG_ON(refcount_read(&th->refcnt) == 0);
1762         if (lock)
1763                 down_write(&threads->lock);
1764         rb_erase_cached(&th->rb_node, &threads->entries);
1765         RB_CLEAR_NODE(&th->rb_node);
1766         --threads->nr;
1767         /*
1768          * Move it first to the dead_threads list, then drop the reference,
1769          * if this is the last reference, then the thread__delete destructor
1770          * will be called and we will remove it from the dead_threads list.
1771          */
1772         list_add_tail(&th->node, &threads->dead);
1773         if (lock)
1774                 up_write(&threads->lock);
1775         thread__put(th);
1776 }
1777
1778 void machine__remove_thread(struct machine *machine, struct thread *th)
1779 {
1780         return __machine__remove_thread(machine, th, true);
1781 }
1782
1783 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1784                                 struct perf_sample *sample)
1785 {
1786         struct thread *thread = machine__find_thread(machine,
1787                                                      event->fork.pid,
1788                                                      event->fork.tid);
1789         struct thread *parent = machine__findnew_thread(machine,
1790                                                         event->fork.ppid,
1791                                                         event->fork.ptid);
1792         bool do_maps_clone = true;
1793         int err = 0;
1794
1795         if (dump_trace)
1796                 perf_event__fprintf_task(event, stdout);
1797
1798         /*
1799          * There may be an existing thread that is not actually the parent,
1800          * either because we are processing events out of order, or because the
1801          * (fork) event that would have removed the thread was lost. Assume the
1802          * latter case and continue on as best we can.
1803          */
1804         if (parent->pid_ != (pid_t)event->fork.ppid) {
1805                 dump_printf("removing erroneous parent thread %d/%d\n",
1806                             parent->pid_, parent->tid);
1807                 machine__remove_thread(machine, parent);
1808                 thread__put(parent);
1809                 parent = machine__findnew_thread(machine, event->fork.ppid,
1810                                                  event->fork.ptid);
1811         }
1812
1813         /* if a thread currently exists for the thread id remove it */
1814         if (thread != NULL) {
1815                 machine__remove_thread(machine, thread);
1816                 thread__put(thread);
1817         }
1818
1819         thread = machine__findnew_thread(machine, event->fork.pid,
1820                                          event->fork.tid);
1821         /*
1822          * When synthesizing FORK events, we are trying to create thread
1823          * objects for the already running tasks on the machine.
1824          *
1825          * Normally, for a kernel FORK event, we want to clone the parent's
1826          * maps because that is what the kernel just did.
1827          *
1828          * But when synthesizing, this should not be done.  If we do, we end up
1829          * with overlapping maps as we process the sythesized MMAP2 events that
1830          * get delivered shortly thereafter.
1831          *
1832          * Use the FORK event misc flags in an internal way to signal this
1833          * situation, so we can elide the map clone when appropriate.
1834          */
1835         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1836                 do_maps_clone = false;
1837
1838         if (thread == NULL || parent == NULL ||
1839             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1840                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1841                 err = -1;
1842         }
1843         thread__put(thread);
1844         thread__put(parent);
1845
1846         return err;
1847 }
1848
1849 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1850                                 struct perf_sample *sample __maybe_unused)
1851 {
1852         struct thread *thread = machine__find_thread(machine,
1853                                                      event->fork.pid,
1854                                                      event->fork.tid);
1855
1856         if (dump_trace)
1857                 perf_event__fprintf_task(event, stdout);
1858
1859         if (thread != NULL) {
1860                 thread__exited(thread);
1861                 thread__put(thread);
1862         }
1863
1864         return 0;
1865 }
1866
1867 int machine__process_event(struct machine *machine, union perf_event *event,
1868                            struct perf_sample *sample)
1869 {
1870         int ret;
1871
1872         switch (event->header.type) {
1873         case PERF_RECORD_COMM:
1874                 ret = machine__process_comm_event(machine, event, sample); break;
1875         case PERF_RECORD_MMAP:
1876                 ret = machine__process_mmap_event(machine, event, sample); break;
1877         case PERF_RECORD_NAMESPACES:
1878                 ret = machine__process_namespaces_event(machine, event, sample); break;
1879         case PERF_RECORD_MMAP2:
1880                 ret = machine__process_mmap2_event(machine, event, sample); break;
1881         case PERF_RECORD_FORK:
1882                 ret = machine__process_fork_event(machine, event, sample); break;
1883         case PERF_RECORD_EXIT:
1884                 ret = machine__process_exit_event(machine, event, sample); break;
1885         case PERF_RECORD_LOST:
1886                 ret = machine__process_lost_event(machine, event, sample); break;
1887         case PERF_RECORD_AUX:
1888                 ret = machine__process_aux_event(machine, event); break;
1889         case PERF_RECORD_ITRACE_START:
1890                 ret = machine__process_itrace_start_event(machine, event); break;
1891         case PERF_RECORD_LOST_SAMPLES:
1892                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1893         case PERF_RECORD_SWITCH:
1894         case PERF_RECORD_SWITCH_CPU_WIDE:
1895                 ret = machine__process_switch_event(machine, event); break;
1896         case PERF_RECORD_KSYMBOL:
1897                 ret = machine__process_ksymbol(machine, event, sample); break;
1898         case PERF_RECORD_BPF_EVENT:
1899                 ret = machine__process_bpf_event(machine, event, sample); break;
1900         default:
1901                 ret = -1;
1902                 break;
1903         }
1904
1905         return ret;
1906 }
1907
1908 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1909 {
1910         if (!regexec(regex, sym->name, 0, NULL, 0))
1911                 return 1;
1912         return 0;
1913 }
1914
1915 static void ip__resolve_ams(struct thread *thread,
1916                             struct addr_map_symbol *ams,
1917                             u64 ip)
1918 {
1919         struct addr_location al;
1920
1921         memset(&al, 0, sizeof(al));
1922         /*
1923          * We cannot use the header.misc hint to determine whether a
1924          * branch stack address is user, kernel, guest, hypervisor.
1925          * Branches may straddle the kernel/user/hypervisor boundaries.
1926          * Thus, we have to try consecutively until we find a match
1927          * or else, the symbol is unknown
1928          */
1929         thread__find_cpumode_addr_location(thread, ip, &al);
1930
1931         ams->addr = ip;
1932         ams->al_addr = al.addr;
1933         ams->sym = al.sym;
1934         ams->map = al.map;
1935         ams->phys_addr = 0;
1936 }
1937
1938 static void ip__resolve_data(struct thread *thread,
1939                              u8 m, struct addr_map_symbol *ams,
1940                              u64 addr, u64 phys_addr)
1941 {
1942         struct addr_location al;
1943
1944         memset(&al, 0, sizeof(al));
1945
1946         thread__find_symbol(thread, m, addr, &al);
1947
1948         ams->addr = addr;
1949         ams->al_addr = al.addr;
1950         ams->sym = al.sym;
1951         ams->map = al.map;
1952         ams->phys_addr = phys_addr;
1953 }
1954
1955 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1956                                      struct addr_location *al)
1957 {
1958         struct mem_info *mi = mem_info__new();
1959
1960         if (!mi)
1961                 return NULL;
1962
1963         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1964         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1965                          sample->addr, sample->phys_addr);
1966         mi->data_src.val = sample->data_src;
1967
1968         return mi;
1969 }
1970
1971 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
1972 {
1973         char *srcline = NULL;
1974
1975         if (!map || callchain_param.key == CCKEY_FUNCTION)
1976                 return srcline;
1977
1978         srcline = srcline__tree_find(&map->dso->srclines, ip);
1979         if (!srcline) {
1980                 bool show_sym = false;
1981                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
1982
1983                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
1984                                       sym, show_sym, show_addr, ip);
1985                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
1986         }
1987
1988         return srcline;
1989 }
1990
1991 struct iterations {
1992         int nr_loop_iter;
1993         u64 cycles;
1994 };
1995
1996 static int add_callchain_ip(struct thread *thread,
1997                             struct callchain_cursor *cursor,
1998                             struct symbol **parent,
1999                             struct addr_location *root_al,
2000                             u8 *cpumode,
2001                             u64 ip,
2002                             bool branch,
2003                             struct branch_flags *flags,
2004                             struct iterations *iter,
2005                             u64 branch_from)
2006 {
2007         struct addr_location al;
2008         int nr_loop_iter = 0;
2009         u64 iter_cycles = 0;
2010         const char *srcline = NULL;
2011
2012         al.filtered = 0;
2013         al.sym = NULL;
2014         if (!cpumode) {
2015                 thread__find_cpumode_addr_location(thread, ip, &al);
2016         } else {
2017                 if (ip >= PERF_CONTEXT_MAX) {
2018                         switch (ip) {
2019                         case PERF_CONTEXT_HV:
2020                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2021                                 break;
2022                         case PERF_CONTEXT_KERNEL:
2023                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2024                                 break;
2025                         case PERF_CONTEXT_USER:
2026                                 *cpumode = PERF_RECORD_MISC_USER;
2027                                 break;
2028                         default:
2029                                 pr_debug("invalid callchain context: "
2030                                          "%"PRId64"\n", (s64) ip);
2031                                 /*
2032                                  * It seems the callchain is corrupted.
2033                                  * Discard all.
2034                                  */
2035                                 callchain_cursor_reset(cursor);
2036                                 return 1;
2037                         }
2038                         return 0;
2039                 }
2040                 thread__find_symbol(thread, *cpumode, ip, &al);
2041         }
2042
2043         if (al.sym != NULL) {
2044                 if (perf_hpp_list.parent && !*parent &&
2045                     symbol__match_regex(al.sym, &parent_regex))
2046                         *parent = al.sym;
2047                 else if (have_ignore_callees && root_al &&
2048                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2049                         /* Treat this symbol as the root,
2050                            forgetting its callees. */
2051                         *root_al = al;
2052                         callchain_cursor_reset(cursor);
2053                 }
2054         }
2055
2056         if (symbol_conf.hide_unresolved && al.sym == NULL)
2057                 return 0;
2058
2059         if (iter) {
2060                 nr_loop_iter = iter->nr_loop_iter;
2061                 iter_cycles = iter->cycles;
2062         }
2063
2064         srcline = callchain_srcline(al.map, al.sym, al.addr);
2065         return callchain_cursor_append(cursor, ip, al.map, al.sym,
2066                                        branch, flags, nr_loop_iter,
2067                                        iter_cycles, branch_from, srcline);
2068 }
2069
2070 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2071                                            struct addr_location *al)
2072 {
2073         unsigned int i;
2074         const struct branch_stack *bs = sample->branch_stack;
2075         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2076
2077         if (!bi)
2078                 return NULL;
2079
2080         for (i = 0; i < bs->nr; i++) {
2081                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2082                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2083                 bi[i].flags = bs->entries[i].flags;
2084         }
2085         return bi;
2086 }
2087
2088 static void save_iterations(struct iterations *iter,
2089                             struct branch_entry *be, int nr)
2090 {
2091         int i;
2092
2093         iter->nr_loop_iter++;
2094         iter->cycles = 0;
2095
2096         for (i = 0; i < nr; i++)
2097                 iter->cycles += be[i].flags.cycles;
2098 }
2099
2100 #define CHASHSZ 127
2101 #define CHASHBITS 7
2102 #define NO_ENTRY 0xff
2103
2104 #define PERF_MAX_BRANCH_DEPTH 127
2105
2106 /* Remove loops. */
2107 static int remove_loops(struct branch_entry *l, int nr,
2108                         struct iterations *iter)
2109 {
2110         int i, j, off;
2111         unsigned char chash[CHASHSZ];
2112
2113         memset(chash, NO_ENTRY, sizeof(chash));
2114
2115         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2116
2117         for (i = 0; i < nr; i++) {
2118                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2119
2120                 /* no collision handling for now */
2121                 if (chash[h] == NO_ENTRY) {
2122                         chash[h] = i;
2123                 } else if (l[chash[h]].from == l[i].from) {
2124                         bool is_loop = true;
2125                         /* check if it is a real loop */
2126                         off = 0;
2127                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2128                                 if (l[j].from != l[i + off].from) {
2129                                         is_loop = false;
2130                                         break;
2131                                 }
2132                         if (is_loop) {
2133                                 j = nr - (i + off);
2134                                 if (j > 0) {
2135                                         save_iterations(iter + i + off,
2136                                                 l + i, off);
2137
2138                                         memmove(iter + i, iter + i + off,
2139                                                 j * sizeof(*iter));
2140
2141                                         memmove(l + i, l + i + off,
2142                                                 j * sizeof(*l));
2143                                 }
2144
2145                                 nr -= off;
2146                         }
2147                 }
2148         }
2149         return nr;
2150 }
2151
2152 /*
2153  * Recolve LBR callstack chain sample
2154  * Return:
2155  * 1 on success get LBR callchain information
2156  * 0 no available LBR callchain information, should try fp
2157  * negative error code on other errors.
2158  */
2159 static int resolve_lbr_callchain_sample(struct thread *thread,
2160                                         struct callchain_cursor *cursor,
2161                                         struct perf_sample *sample,
2162                                         struct symbol **parent,
2163                                         struct addr_location *root_al,
2164                                         int max_stack)
2165 {
2166         struct ip_callchain *chain = sample->callchain;
2167         int chain_nr = min(max_stack, (int)chain->nr), i;
2168         u8 cpumode = PERF_RECORD_MISC_USER;
2169         u64 ip, branch_from = 0;
2170
2171         for (i = 0; i < chain_nr; i++) {
2172                 if (chain->ips[i] == PERF_CONTEXT_USER)
2173                         break;
2174         }
2175
2176         /* LBR only affects the user callchain */
2177         if (i != chain_nr) {
2178                 struct branch_stack *lbr_stack = sample->branch_stack;
2179                 int lbr_nr = lbr_stack->nr, j, k;
2180                 bool branch;
2181                 struct branch_flags *flags;
2182                 /*
2183                  * LBR callstack can only get user call chain.
2184                  * The mix_chain_nr is kernel call chain
2185                  * number plus LBR user call chain number.
2186                  * i is kernel call chain number,
2187                  * 1 is PERF_CONTEXT_USER,
2188                  * lbr_nr + 1 is the user call chain number.
2189                  * For details, please refer to the comments
2190                  * in callchain__printf
2191                  */
2192                 int mix_chain_nr = i + 1 + lbr_nr + 1;
2193
2194                 for (j = 0; j < mix_chain_nr; j++) {
2195                         int err;
2196                         branch = false;
2197                         flags = NULL;
2198
2199                         if (callchain_param.order == ORDER_CALLEE) {
2200                                 if (j < i + 1)
2201                                         ip = chain->ips[j];
2202                                 else if (j > i + 1) {
2203                                         k = j - i - 2;
2204                                         ip = lbr_stack->entries[k].from;
2205                                         branch = true;
2206                                         flags = &lbr_stack->entries[k].flags;
2207                                 } else {
2208                                         ip = lbr_stack->entries[0].to;
2209                                         branch = true;
2210                                         flags = &lbr_stack->entries[0].flags;
2211                                         branch_from =
2212                                                 lbr_stack->entries[0].from;
2213                                 }
2214                         } else {
2215                                 if (j < lbr_nr) {
2216                                         k = lbr_nr - j - 1;
2217                                         ip = lbr_stack->entries[k].from;
2218                                         branch = true;
2219                                         flags = &lbr_stack->entries[k].flags;
2220                                 }
2221                                 else if (j > lbr_nr)
2222                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
2223                                 else {
2224                                         ip = lbr_stack->entries[0].to;
2225                                         branch = true;
2226                                         flags = &lbr_stack->entries[0].flags;
2227                                         branch_from =
2228                                                 lbr_stack->entries[0].from;
2229                                 }
2230                         }
2231
2232                         err = add_callchain_ip(thread, cursor, parent,
2233                                                root_al, &cpumode, ip,
2234                                                branch, flags, NULL,
2235                                                branch_from);
2236                         if (err)
2237                                 return (err < 0) ? err : 0;
2238                 }
2239                 return 1;
2240         }
2241
2242         return 0;
2243 }
2244
2245 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2246                              struct callchain_cursor *cursor,
2247                              struct symbol **parent,
2248                              struct addr_location *root_al,
2249                              u8 *cpumode, int ent)
2250 {
2251         int err = 0;
2252
2253         while (--ent >= 0) {
2254                 u64 ip = chain->ips[ent];
2255
2256                 if (ip >= PERF_CONTEXT_MAX) {
2257                         err = add_callchain_ip(thread, cursor, parent,
2258                                                root_al, cpumode, ip,
2259                                                false, NULL, NULL, 0);
2260                         break;
2261                 }
2262         }
2263         return err;
2264 }
2265
2266 static int thread__resolve_callchain_sample(struct thread *thread,
2267                                             struct callchain_cursor *cursor,
2268                                             struct perf_evsel *evsel,
2269                                             struct perf_sample *sample,
2270                                             struct symbol **parent,
2271                                             struct addr_location *root_al,
2272                                             int max_stack)
2273 {
2274         struct branch_stack *branch = sample->branch_stack;
2275         struct ip_callchain *chain = sample->callchain;
2276         int chain_nr = 0;
2277         u8 cpumode = PERF_RECORD_MISC_USER;
2278         int i, j, err, nr_entries;
2279         int skip_idx = -1;
2280         int first_call = 0;
2281
2282         if (chain)
2283                 chain_nr = chain->nr;
2284
2285         if (perf_evsel__has_branch_callstack(evsel)) {
2286                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2287                                                    root_al, max_stack);
2288                 if (err)
2289                         return (err < 0) ? err : 0;
2290         }
2291
2292         /*
2293          * Based on DWARF debug information, some architectures skip
2294          * a callchain entry saved by the kernel.
2295          */
2296         skip_idx = arch_skip_callchain_idx(thread, chain);
2297
2298         /*
2299          * Add branches to call stack for easier browsing. This gives
2300          * more context for a sample than just the callers.
2301          *
2302          * This uses individual histograms of paths compared to the
2303          * aggregated histograms the normal LBR mode uses.
2304          *
2305          * Limitations for now:
2306          * - No extra filters
2307          * - No annotations (should annotate somehow)
2308          */
2309
2310         if (branch && callchain_param.branch_callstack) {
2311                 int nr = min(max_stack, (int)branch->nr);
2312                 struct branch_entry be[nr];
2313                 struct iterations iter[nr];
2314
2315                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2316                         pr_warning("corrupted branch chain. skipping...\n");
2317                         goto check_calls;
2318                 }
2319
2320                 for (i = 0; i < nr; i++) {
2321                         if (callchain_param.order == ORDER_CALLEE) {
2322                                 be[i] = branch->entries[i];
2323
2324                                 if (chain == NULL)
2325                                         continue;
2326
2327                                 /*
2328                                  * Check for overlap into the callchain.
2329                                  * The return address is one off compared to
2330                                  * the branch entry. To adjust for this
2331                                  * assume the calling instruction is not longer
2332                                  * than 8 bytes.
2333                                  */
2334                                 if (i == skip_idx ||
2335                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2336                                         first_call++;
2337                                 else if (be[i].from < chain->ips[first_call] &&
2338                                     be[i].from >= chain->ips[first_call] - 8)
2339                                         first_call++;
2340                         } else
2341                                 be[i] = branch->entries[branch->nr - i - 1];
2342                 }
2343
2344                 memset(iter, 0, sizeof(struct iterations) * nr);
2345                 nr = remove_loops(be, nr, iter);
2346
2347                 for (i = 0; i < nr; i++) {
2348                         err = add_callchain_ip(thread, cursor, parent,
2349                                                root_al,
2350                                                NULL, be[i].to,
2351                                                true, &be[i].flags,
2352                                                NULL, be[i].from);
2353
2354                         if (!err)
2355                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2356                                                        NULL, be[i].from,
2357                                                        true, &be[i].flags,
2358                                                        &iter[i], 0);
2359                         if (err == -EINVAL)
2360                                 break;
2361                         if (err)
2362                                 return err;
2363                 }
2364
2365                 if (chain_nr == 0)
2366                         return 0;
2367
2368                 chain_nr -= nr;
2369         }
2370
2371 check_calls:
2372         if (callchain_param.order != ORDER_CALLEE) {
2373                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2374                                         &cpumode, chain->nr - first_call);
2375                 if (err)
2376                         return (err < 0) ? err : 0;
2377         }
2378         for (i = first_call, nr_entries = 0;
2379              i < chain_nr && nr_entries < max_stack; i++) {
2380                 u64 ip;
2381
2382                 if (callchain_param.order == ORDER_CALLEE)
2383                         j = i;
2384                 else
2385                         j = chain->nr - i - 1;
2386
2387 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2388                 if (j == skip_idx)
2389                         continue;
2390 #endif
2391                 ip = chain->ips[j];
2392                 if (ip < PERF_CONTEXT_MAX)
2393                        ++nr_entries;
2394                 else if (callchain_param.order != ORDER_CALLEE) {
2395                         err = find_prev_cpumode(chain, thread, cursor, parent,
2396                                                 root_al, &cpumode, j);
2397                         if (err)
2398                                 return (err < 0) ? err : 0;
2399                         continue;
2400                 }
2401
2402                 err = add_callchain_ip(thread, cursor, parent,
2403                                        root_al, &cpumode, ip,
2404                                        false, NULL, NULL, 0);
2405
2406                 if (err)
2407                         return (err < 0) ? err : 0;
2408         }
2409
2410         return 0;
2411 }
2412
2413 static int append_inlines(struct callchain_cursor *cursor,
2414                           struct map *map, struct symbol *sym, u64 ip)
2415 {
2416         struct inline_node *inline_node;
2417         struct inline_list *ilist;
2418         u64 addr;
2419         int ret = 1;
2420
2421         if (!symbol_conf.inline_name || !map || !sym)
2422                 return ret;
2423
2424         addr = map__map_ip(map, ip);
2425         addr = map__rip_2objdump(map, addr);
2426
2427         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2428         if (!inline_node) {
2429                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2430                 if (!inline_node)
2431                         return ret;
2432                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2433         }
2434
2435         list_for_each_entry(ilist, &inline_node->val, list) {
2436                 ret = callchain_cursor_append(cursor, ip, map,
2437                                               ilist->symbol, false,
2438                                               NULL, 0, 0, 0, ilist->srcline);
2439
2440                 if (ret != 0)
2441                         return ret;
2442         }
2443
2444         return ret;
2445 }
2446
2447 static int unwind_entry(struct unwind_entry *entry, void *arg)
2448 {
2449         struct callchain_cursor *cursor = arg;
2450         const char *srcline = NULL;
2451         u64 addr = entry->ip;
2452
2453         if (symbol_conf.hide_unresolved && entry->sym == NULL)
2454                 return 0;
2455
2456         if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2457                 return 0;
2458
2459         /*
2460          * Convert entry->ip from a virtual address to an offset in
2461          * its corresponding binary.
2462          */
2463         if (entry->map)
2464                 addr = map__map_ip(entry->map, entry->ip);
2465
2466         srcline = callchain_srcline(entry->map, entry->sym, addr);
2467         return callchain_cursor_append(cursor, entry->ip,
2468                                        entry->map, entry->sym,
2469                                        false, NULL, 0, 0, 0, srcline);
2470 }
2471
2472 static int thread__resolve_callchain_unwind(struct thread *thread,
2473                                             struct callchain_cursor *cursor,
2474                                             struct perf_evsel *evsel,
2475                                             struct perf_sample *sample,
2476                                             int max_stack)
2477 {
2478         /* Can we do dwarf post unwind? */
2479         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2480               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
2481                 return 0;
2482
2483         /* Bail out if nothing was captured. */
2484         if ((!sample->user_regs.regs) ||
2485             (!sample->user_stack.size))
2486                 return 0;
2487
2488         return unwind__get_entries(unwind_entry, cursor,
2489                                    thread, sample, max_stack);
2490 }
2491
2492 int thread__resolve_callchain(struct thread *thread,
2493                               struct callchain_cursor *cursor,
2494                               struct perf_evsel *evsel,
2495                               struct perf_sample *sample,
2496                               struct symbol **parent,
2497                               struct addr_location *root_al,
2498                               int max_stack)
2499 {
2500         int ret = 0;
2501
2502         callchain_cursor_reset(cursor);
2503
2504         if (callchain_param.order == ORDER_CALLEE) {
2505                 ret = thread__resolve_callchain_sample(thread, cursor,
2506                                                        evsel, sample,
2507                                                        parent, root_al,
2508                                                        max_stack);
2509                 if (ret)
2510                         return ret;
2511                 ret = thread__resolve_callchain_unwind(thread, cursor,
2512                                                        evsel, sample,
2513                                                        max_stack);
2514         } else {
2515                 ret = thread__resolve_callchain_unwind(thread, cursor,
2516                                                        evsel, sample,
2517                                                        max_stack);
2518                 if (ret)
2519                         return ret;
2520                 ret = thread__resolve_callchain_sample(thread, cursor,
2521                                                        evsel, sample,
2522                                                        parent, root_al,
2523                                                        max_stack);
2524         }
2525
2526         return ret;
2527 }
2528
2529 int machine__for_each_thread(struct machine *machine,
2530                              int (*fn)(struct thread *thread, void *p),
2531                              void *priv)
2532 {
2533         struct threads *threads;
2534         struct rb_node *nd;
2535         struct thread *thread;
2536         int rc = 0;
2537         int i;
2538
2539         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2540                 threads = &machine->threads[i];
2541                 for (nd = rb_first_cached(&threads->entries); nd;
2542                      nd = rb_next(nd)) {
2543                         thread = rb_entry(nd, struct thread, rb_node);
2544                         rc = fn(thread, priv);
2545                         if (rc != 0)
2546                                 return rc;
2547                 }
2548
2549                 list_for_each_entry(thread, &threads->dead, node) {
2550                         rc = fn(thread, priv);
2551                         if (rc != 0)
2552                                 return rc;
2553                 }
2554         }
2555         return rc;
2556 }
2557
2558 int machines__for_each_thread(struct machines *machines,
2559                               int (*fn)(struct thread *thread, void *p),
2560                               void *priv)
2561 {
2562         struct rb_node *nd;
2563         int rc = 0;
2564
2565         rc = machine__for_each_thread(&machines->host, fn, priv);
2566         if (rc != 0)
2567                 return rc;
2568
2569         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2570                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2571
2572                 rc = machine__for_each_thread(machine, fn, priv);
2573                 if (rc != 0)
2574                         return rc;
2575         }
2576         return rc;
2577 }
2578
2579 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
2580                                   struct target *target, struct thread_map *threads,
2581                                   perf_event__handler_t process, bool data_mmap,
2582                                   unsigned int nr_threads_synthesize)
2583 {
2584         if (target__has_task(target))
2585                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
2586         else if (target__has_cpu(target))
2587                 return perf_event__synthesize_threads(tool, process,
2588                                                       machine, data_mmap,
2589                                                       nr_threads_synthesize);
2590         /* command specified */
2591         return 0;
2592 }
2593
2594 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2595 {
2596         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
2597                 return -1;
2598
2599         return machine->current_tid[cpu];
2600 }
2601
2602 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2603                              pid_t tid)
2604 {
2605         struct thread *thread;
2606
2607         if (cpu < 0)
2608                 return -EINVAL;
2609
2610         if (!machine->current_tid) {
2611                 int i;
2612
2613                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
2614                 if (!machine->current_tid)
2615                         return -ENOMEM;
2616                 for (i = 0; i < MAX_NR_CPUS; i++)
2617                         machine->current_tid[i] = -1;
2618         }
2619
2620         if (cpu >= MAX_NR_CPUS) {
2621                 pr_err("Requested CPU %d too large. ", cpu);
2622                 pr_err("Consider raising MAX_NR_CPUS\n");
2623                 return -EINVAL;
2624         }
2625
2626         machine->current_tid[cpu] = tid;
2627
2628         thread = machine__findnew_thread(machine, pid, tid);
2629         if (!thread)
2630                 return -ENOMEM;
2631
2632         thread->cpu = cpu;
2633         thread__put(thread);
2634
2635         return 0;
2636 }
2637
2638 /*
2639  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2640  * normalized arch is needed.
2641  */
2642 bool machine__is(struct machine *machine, const char *arch)
2643 {
2644         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2645 }
2646
2647 int machine__nr_cpus_avail(struct machine *machine)
2648 {
2649         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2650 }
2651
2652 int machine__get_kernel_start(struct machine *machine)
2653 {
2654         struct map *map = machine__kernel_map(machine);
2655         int err = 0;
2656
2657         /*
2658          * The only addresses above 2^63 are kernel addresses of a 64-bit
2659          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2660          * all addresses including kernel addresses are less than 2^32.  In
2661          * that case (32-bit system), if the kernel mapping is unknown, all
2662          * addresses will be assumed to be in user space - see
2663          * machine__kernel_ip().
2664          */
2665         machine->kernel_start = 1ULL << 63;
2666         if (map) {
2667                 err = map__load(map);
2668                 /*
2669                  * On x86_64, PTI entry trampolines are less than the
2670                  * start of kernel text, but still above 2^63. So leave
2671                  * kernel_start = 1ULL << 63 for x86_64.
2672                  */
2673                 if (!err && !machine__is(machine, "x86_64"))
2674                         machine->kernel_start = map->start;
2675         }
2676         return err;
2677 }
2678
2679 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2680 {
2681         u8 addr_cpumode = cpumode;
2682         bool kernel_ip;
2683
2684         if (!machine->single_address_space)
2685                 goto out;
2686
2687         kernel_ip = machine__kernel_ip(machine, addr);
2688         switch (cpumode) {
2689         case PERF_RECORD_MISC_KERNEL:
2690         case PERF_RECORD_MISC_USER:
2691                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2692                                            PERF_RECORD_MISC_USER;
2693                 break;
2694         case PERF_RECORD_MISC_GUEST_KERNEL:
2695         case PERF_RECORD_MISC_GUEST_USER:
2696                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2697                                            PERF_RECORD_MISC_GUEST_USER;
2698                 break;
2699         default:
2700                 break;
2701         }
2702 out:
2703         return addr_cpumode;
2704 }
2705
2706 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2707 {
2708         return dsos__findnew(&machine->dsos, filename);
2709 }
2710
2711 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2712 {
2713         struct machine *machine = vmachine;
2714         struct map *map;
2715         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2716
2717         if (sym == NULL)
2718                 return NULL;
2719
2720         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2721         *addrp = map->unmap_ip(map, sym->start);
2722         return sym->name;
2723 }