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, u64 *start)
 928 {
 929         char filename[PATH_MAX];
 930         int i, err = -1;
 931         const char *name;
 932         u64 addr = 0;
 933
 934         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 935
 936         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 937                 return 0;
 938
 939         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 940                 err = kallsyms__get_function_start(filename, name, &addr);
 941                 if (!err)
 942                         break;
 943         }
 944
 945         if (err)
 946                 return -1;
 947
 948         if (symbol_name)
 949                 *symbol_name = name;
 950
 951         *start = addr;
 952         return 0;
 953 }
 954
 955 int machine__create_extra_kernel_map(struct machine *machine,
 956                                      struct dso *kernel,
 957                                      struct extra_kernel_map *xm)
 958 {
 959         struct kmap *kmap;
 960         struct map *map;
 961
 962         map = map__new2(xm->start, kernel);
 963         if (!map)
 964                 return -1;
 965
 966         map->end   = xm->end;
 967         map->pgoff = xm->pgoff;
 968
 969         kmap = map__kmap(map);
 970
 971         kmap->kmaps = &machine->kmaps;
 972         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
 973
 974         map_groups__insert(&machine->kmaps, map);
 975
 976         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
 977                   kmap->name, map->start, map->end);
 978
 979         map__put(map);
 980
 981         return 0;
 982 }
 983
 984 static u64 find_entry_trampoline(struct dso *dso)
 985 {
 986         /* Duplicates are removed so lookup all aliases */
 987         const char *syms[] = {
 988                 "_entry_trampoline",
 989                 "__entry_trampoline_start",
 990                 "entry_SYSCALL_64_trampoline",
 991         };
 992         struct symbol *sym = dso__first_symbol(dso);
 993         unsigned int i;
 994
 995         for (; sym; sym = dso__next_symbol(sym)) {
 996                 if (sym->binding != STB_GLOBAL)
 997                         continue;
 998                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
 999                         if (!strcmp(sym->name, syms[i]))
1000                                 return sym->start;
1001                 }
1002         }
1003
1004         return 0;
1005 }
1006
1007 /*
1008  * These values can be used for kernels that do not have symbols for the entry
1009  * trampolines in kallsyms.
1010  */
1011 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1012 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1013 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1014
1015 /* Map x86_64 PTI entry trampolines */
1016 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1017                                           struct dso *kernel)
1018 {
1019         struct map_groups *kmaps = &machine->kmaps;
1020         struct maps *maps = &kmaps->maps;
1021         int nr_cpus_avail, cpu;
1022         bool found = false;
1023         struct map *map;
1024         u64 pgoff;
1025
1026         /*
1027          * In the vmlinux case, pgoff is a virtual address which must now be
1028          * mapped to a vmlinux offset.
1029          */
1030         for (map = maps__first(maps); map; map = map__next(map)) {
1031                 struct kmap *kmap = __map__kmap(map);
1032                 struct map *dest_map;
1033
1034                 if (!kmap || !is_entry_trampoline(kmap->name))
1035                         continue;
1036
1037                 dest_map = map_groups__find(kmaps, map->pgoff);
1038                 if (dest_map != map)
1039                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1040                 found = true;
1041         }
1042         if (found || machine->trampolines_mapped)
1043                 return 0;
1044
1045         pgoff = find_entry_trampoline(kernel);
1046         if (!pgoff)
1047                 return 0;
1048
1049         nr_cpus_avail = machine__nr_cpus_avail(machine);
1050
1051         /* Add a 1 page map for each CPU's entry trampoline */
1052         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1053                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1054                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1055                          X86_64_ENTRY_TRAMPOLINE;
1056                 struct extra_kernel_map xm = {
1057                         .start = va,
1058                         .end   = va + page_size,
1059                         .pgoff = pgoff,
1060                 };
1061
1062                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1063
1064                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1065                         return -1;
1066         }
1067
1068         machine->trampolines_mapped = nr_cpus_avail;
1069
1070         return 0;
1071 }
1072
1073 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1074                                              struct dso *kernel __maybe_unused)
1075 {
1076         return 0;
1077 }
1078
1079 static int
1080 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1081 {
1082         struct kmap *kmap;
1083         struct map *map;
1084
1085         /* In case of renewal the kernel map, destroy previous one */
1086         machine__destroy_kernel_maps(machine);
1087
1088         machine->vmlinux_map = map__new2(0, kernel);
1089         if (machine->vmlinux_map == NULL)
1090                 return -1;
1091
1092         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1093         map = machine__kernel_map(machine);
1094         kmap = map__kmap(map);
1095         if (!kmap)
1096                 return -1;
1097
1098         kmap->kmaps = &machine->kmaps;
1099         map_groups__insert(&machine->kmaps, map);
1100
1101         return 0;
1102 }
1103
1104 void machine__destroy_kernel_maps(struct machine *machine)
1105 {
1106         struct kmap *kmap;
1107         struct map *map = machine__kernel_map(machine);
1108
1109         if (map == NULL)
1110                 return;
1111
1112         kmap = map__kmap(map);
1113         map_groups__remove(&machine->kmaps, map);
1114         if (kmap && kmap->ref_reloc_sym) {
1115                 zfree((char **)&kmap->ref_reloc_sym->name);
1116                 zfree(&kmap->ref_reloc_sym);
1117         }
1118
1119         map__zput(machine->vmlinux_map);
1120 }
1121
1122 int machines__create_guest_kernel_maps(struct machines *machines)
1123 {
1124         int ret = 0;
1125         struct dirent **namelist = NULL;
1126         int i, items = 0;
1127         char path[PATH_MAX];
1128         pid_t pid;
1129         char *endp;
1130
1131         if (symbol_conf.default_guest_vmlinux_name ||
1132             symbol_conf.default_guest_modules ||
1133             symbol_conf.default_guest_kallsyms) {
1134                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1135         }
1136
1137         if (symbol_conf.guestmount) {
1138                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1139                 if (items <= 0)
1140                         return -ENOENT;
1141                 for (i = 0; i < items; i++) {
1142                         if (!isdigit(namelist[i]->d_name[0])) {
1143                                 /* Filter out . and .. */
1144                                 continue;
1145                         }
1146                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1147                         if ((*endp != '\0') ||
1148                             (endp == namelist[i]->d_name) ||
1149                             (errno == ERANGE)) {
1150                                 pr_debug("invalid directory (%s). Skipping.\n",
1151                                          namelist[i]->d_name);
1152                                 continue;
1153                         }
1154                         sprintf(path, "%s/%s/proc/kallsyms",
1155                                 symbol_conf.guestmount,
1156                                 namelist[i]->d_name);
1157                         ret = access(path, R_OK);
1158                         if (ret) {
1159                                 pr_debug("Can't access file %s\n", path);
1160                                 goto failure;
1161                         }
1162                         machines__create_kernel_maps(machines, pid);
1163                 }
1164 failure:
1165                 free(namelist);
1166         }
1167
1168         return ret;
1169 }
1170
1171 void machines__destroy_kernel_maps(struct machines *machines)
1172 {
1173         struct rb_node *next = rb_first_cached(&machines->guests);
1174
1175         machine__destroy_kernel_maps(&machines->host);
1176
1177         while (next) {
1178                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1179
1180                 next = rb_next(&pos->rb_node);
1181                 rb_erase_cached(&pos->rb_node, &machines->guests);
1182                 machine__delete(pos);
1183         }
1184 }
1185
1186 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1187 {
1188         struct machine *machine = machines__findnew(machines, pid);
1189
1190         if (machine == NULL)
1191                 return -1;
1192
1193         return machine__create_kernel_maps(machine);
1194 }
1195
1196 int machine__load_kallsyms(struct machine *machine, const char *filename)
1197 {
1198         struct map *map = machine__kernel_map(machine);
1199         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1200
1201         if (ret > 0) {
1202                 dso__set_loaded(map->dso);
1203                 /*
1204                  * Since /proc/kallsyms will have multiple sessions for the
1205                  * kernel, with modules between them, fixup the end of all
1206                  * sections.
1207                  */
1208                 map_groups__fixup_end(&machine->kmaps);
1209         }
1210
1211         return ret;
1212 }
1213
1214 int machine__load_vmlinux_path(struct machine *machine)
1215 {
1216         struct map *map = machine__kernel_map(machine);
1217         int ret = dso__load_vmlinux_path(map->dso, map);
1218
1219         if (ret > 0)
1220                 dso__set_loaded(map->dso);
1221
1222         return ret;
1223 }
1224
1225 static char *get_kernel_version(const char *root_dir)
1226 {
1227         char version[PATH_MAX];
1228         FILE *file;
1229         char *name, *tmp;
1230         const char *prefix = "Linux version ";
1231
1232         sprintf(version, "%s/proc/version", root_dir);
1233         file = fopen(version, "r");
1234         if (!file)
1235                 return NULL;
1236
1237         version[0] = '\0';
1238         tmp = fgets(version, sizeof(version), file);
1239         fclose(file);
1240
1241         name = strstr(version, prefix);
1242         if (!name)
1243                 return NULL;
1244         name += strlen(prefix);
1245         tmp = strchr(name, ' ');
1246         if (tmp)
1247                 *tmp = '\0';
1248
1249         return strdup(name);
1250 }
1251
1252 static bool is_kmod_dso(struct dso *dso)
1253 {
1254         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1255                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1256 }
1257
1258 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1259                                        struct kmod_path *m)
1260 {
1261         char *long_name;
1262         struct map *map = map_groups__find_by_name(mg, m->name);
1263
1264         if (map == NULL)
1265                 return 0;
1266
1267         long_name = strdup(path);
1268         if (long_name == NULL)
1269                 return -ENOMEM;
1270
1271         dso__set_long_name(map->dso, long_name, true);
1272         dso__kernel_module_get_build_id(map->dso, "");
1273
1274         /*
1275          * Full name could reveal us kmod compression, so
1276          * we need to update the symtab_type if needed.
1277          */
1278         if (m->comp && is_kmod_dso(map->dso)) {
1279                 map->dso->symtab_type++;
1280                 map->dso->comp = m->comp;
1281         }
1282
1283         return 0;
1284 }
1285
1286 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1287                                 const char *dir_name, int depth)
1288 {
1289         struct dirent *dent;
1290         DIR *dir = opendir(dir_name);
1291         int ret = 0;
1292
1293         if (!dir) {
1294                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1295                 return -1;
1296         }
1297
1298         while ((dent = readdir(dir)) != NULL) {
1299                 char path[PATH_MAX];
1300                 struct stat st;
1301
1302                 /*sshfs might return bad dent->d_type, so we have to stat*/
1303                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1304                 if (stat(path, &st))
1305                         continue;
1306
1307                 if (S_ISDIR(st.st_mode)) {
1308                         if (!strcmp(dent->d_name, ".") ||
1309                             !strcmp(dent->d_name, ".."))
1310                                 continue;
1311
1312                         /* Do not follow top-level source and build symlinks */
1313                         if (depth == 0) {
1314                                 if (!strcmp(dent->d_name, "source") ||
1315                                     !strcmp(dent->d_name, "build"))
1316                                         continue;
1317                         }
1318
1319                         ret = map_groups__set_modules_path_dir(mg, path,
1320                                                                depth + 1);
1321                         if (ret < 0)
1322                                 goto out;
1323                 } else {
1324                         struct kmod_path m;
1325
1326                         ret = kmod_path__parse_name(&m, dent->d_name);
1327                         if (ret)
1328                                 goto out;
1329
1330                         if (m.kmod)
1331                                 ret = map_groups__set_module_path(mg, path, &m);
1332
1333                         free(m.name);
1334
1335                         if (ret)
1336                                 goto out;
1337                 }
1338         }
1339
1340 out:
1341         closedir(dir);
1342         return ret;
1343 }
1344
1345 static int machine__set_modules_path(struct machine *machine)
1346 {
1347         char *version;
1348         char modules_path[PATH_MAX];
1349
1350         version = get_kernel_version(machine->root_dir);
1351         if (!version)
1352                 return -1;
1353
1354         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1355                  machine->root_dir, version);
1356         free(version);
1357
1358         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1359 }
1360 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1361                                 const char *name __maybe_unused)
1362 {
1363         return 0;
1364 }
1365
1366 static int machine__create_module(void *arg, const char *name, u64 start,
1367                                   u64 size)
1368 {
1369         struct machine *machine = arg;
1370         struct map *map;
1371
1372         if (arch__fix_module_text_start(&start, name) < 0)
1373                 return -1;
1374
1375         map = machine__findnew_module_map(machine, start, name);
1376         if (map == NULL)
1377                 return -1;
1378         map->end = start + size;
1379
1380         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1381
1382         return 0;
1383 }
1384
1385 static int machine__create_modules(struct machine *machine)
1386 {
1387         const char *modules;
1388         char path[PATH_MAX];
1389
1390         if (machine__is_default_guest(machine)) {
1391                 modules = symbol_conf.default_guest_modules;
1392         } else {
1393                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1394                 modules = path;
1395         }
1396
1397         if (symbol__restricted_filename(modules, "/proc/modules"))
1398                 return -1;
1399
1400         if (modules__parse(modules, machine, machine__create_module))
1401                 return -1;
1402
1403         if (!machine__set_modules_path(machine))
1404                 return 0;
1405
1406         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1407
1408         return 0;
1409 }
1410
1411 static void machine__set_kernel_mmap(struct machine *machine,
1412                                      u64 start, u64 end)
1413 {
1414         machine->vmlinux_map->start = start;
1415         machine->vmlinux_map->end   = end;
1416         /*
1417          * Be a bit paranoid here, some perf.data file came with
1418          * a zero sized synthesized MMAP event for the kernel.
1419          */
1420         if (start == 0 && end == 0)
1421                 machine->vmlinux_map->end = ~0ULL;
1422 }
1423
1424 static void machine__update_kernel_mmap(struct machine *machine,
1425                                      u64 start, u64 end)
1426 {
1427         struct map *map = machine__kernel_map(machine);
1428
1429         map__get(map);
1430         map_groups__remove(&machine->kmaps, map);
1431
1432         machine__set_kernel_mmap(machine, start, end);
1433
1434         map_groups__insert(&machine->kmaps, map);
1435         map__put(map);
1436 }
1437
1438 int machine__create_kernel_maps(struct machine *machine)
1439 {
1440         struct dso *kernel = machine__get_kernel(machine);
1441         const char *name = NULL;
1442         struct map *map;
1443         u64 addr = 0;
1444         int ret;
1445
1446         if (kernel == NULL)
1447                 return -1;
1448
1449         ret = __machine__create_kernel_maps(machine, kernel);
1450         if (ret < 0)
1451                 goto out_put;
1452
1453         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1454                 if (machine__is_host(machine))
1455                         pr_debug("Problems creating module maps, "
1456                                  "continuing anyway...\n");
1457                 else
1458                         pr_debug("Problems creating module maps for guest %d, "
1459                                  "continuing anyway...\n", machine->pid);
1460         }
1461
1462         if (!machine__get_running_kernel_start(machine, &name, &addr)) {
1463                 if (name &&
1464                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, addr)) {
1465                         machine__destroy_kernel_maps(machine);
1466                         ret = -1;
1467                         goto out_put;
1468                 }
1469
1470                 /*
1471                  * we have a real start address now, so re-order the kmaps
1472                  * assume it's the last in the kmaps
1473                  */
1474                 machine__update_kernel_mmap(machine, addr, ~0ULL);
1475         }
1476
1477         if (machine__create_extra_kernel_maps(machine, kernel))
1478                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1479
1480         /* update end address of the kernel map using adjacent module address */
1481         map = map__next(machine__kernel_map(machine));
1482         if (map)
1483                 machine__set_kernel_mmap(machine, addr, map->start);
1484 out_put:
1485         dso__put(kernel);
1486         return ret;
1487 }
1488
1489 static bool machine__uses_kcore(struct machine *machine)
1490 {
1491         struct dso *dso;
1492
1493         list_for_each_entry(dso, &machine->dsos.head, node) {
1494                 if (dso__is_kcore(dso))
1495                         return true;
1496         }
1497
1498         return false;
1499 }
1500
1501 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1502                                              union perf_event *event)
1503 {
1504         return machine__is(machine, "x86_64") &&
1505                is_entry_trampoline(event->mmap.filename);
1506 }
1507
1508 static int machine__process_extra_kernel_map(struct machine *machine,
1509                                              union perf_event *event)
1510 {
1511         struct map *kernel_map = machine__kernel_map(machine);
1512         struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1513         struct extra_kernel_map xm = {
1514                 .start = event->mmap.start,
1515                 .end   = event->mmap.start + event->mmap.len,
1516                 .pgoff = event->mmap.pgoff,
1517         };
1518
1519         if (kernel == NULL)
1520                 return -1;
1521
1522         strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1523
1524         return machine__create_extra_kernel_map(machine, kernel, &xm);
1525 }
1526
1527 static int machine__process_kernel_mmap_event(struct machine *machine,
1528                                               union perf_event *event)
1529 {
1530         struct map *map;
1531         enum dso_kernel_type kernel_type;
1532         bool is_kernel_mmap;
1533
1534         /* If we have maps from kcore then we do not need or want any others */
1535         if (machine__uses_kcore(machine))
1536                 return 0;
1537
1538         if (machine__is_host(machine))
1539                 kernel_type = DSO_TYPE_KERNEL;
1540         else
1541                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1542
1543         is_kernel_mmap = memcmp(event->mmap.filename,
1544                                 machine->mmap_name,
1545                                 strlen(machine->mmap_name) - 1) == 0;
1546         if (event->mmap.filename[0] == '/' ||
1547             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1548                 map = machine__findnew_module_map(machine, event->mmap.start,
1549                                                   event->mmap.filename);
1550                 if (map == NULL)
1551                         goto out_problem;
1552
1553                 map->end = map->start + event->mmap.len;
1554         } else if (is_kernel_mmap) {
1555                 const char *symbol_name = (event->mmap.filename +
1556                                 strlen(machine->mmap_name));
1557                 /*
1558                  * Should be there already, from the build-id table in
1559                  * the header.
1560                  */
1561                 struct dso *kernel = NULL;
1562                 struct dso *dso;
1563
1564                 down_read(&machine->dsos.lock);
1565
1566                 list_for_each_entry(dso, &machine->dsos.head, node) {
1567
1568                         /*
1569                          * The cpumode passed to is_kernel_module is not the
1570                          * cpumode of *this* event. If we insist on passing
1571                          * correct cpumode to is_kernel_module, we should
1572                          * record the cpumode when we adding this dso to the
1573                          * linked list.
1574                          *
1575                          * However we don't really need passing correct
1576                          * cpumode.  We know the correct cpumode must be kernel
1577                          * mode (if not, we should not link it onto kernel_dsos
1578                          * list).
1579                          *
1580                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1581                          * is_kernel_module() treats it as a kernel cpumode.
1582                          */
1583
1584                         if (!dso->kernel ||
1585                             is_kernel_module(dso->long_name,
1586                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1587                                 continue;
1588
1589
1590                         kernel = dso;
1591                         break;
1592                 }
1593
1594                 up_read(&machine->dsos.lock);
1595
1596                 if (kernel == NULL)
1597                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1598                 if (kernel == NULL)
1599                         goto out_problem;
1600
1601                 kernel->kernel = kernel_type;
1602                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1603                         dso__put(kernel);
1604                         goto out_problem;
1605                 }
1606
1607                 if (strstr(kernel->long_name, "vmlinux"))
1608                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1609
1610                 machine__update_kernel_mmap(machine, event->mmap.start,
1611                                          event->mmap.start + event->mmap.len);
1612
1613                 /*
1614                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1615                  * symbol. Effectively having zero here means that at record
1616                  * time /proc/sys/kernel/kptr_restrict was non zero.
1617                  */
1618                 if (event->mmap.pgoff != 0) {
1619                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1620                                                         symbol_name,
1621                                                         event->mmap.pgoff);
1622                 }
1623
1624                 if (machine__is_default_guest(machine)) {
1625                         /*
1626                          * preload dso of guest kernel and modules
1627                          */
1628                         dso__load(kernel, machine__kernel_map(machine));
1629                 }
1630         } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1631                 return machine__process_extra_kernel_map(machine, event);
1632         }
1633         return 0;
1634 out_problem:
1635         return -1;
1636 }
1637
1638 int machine__process_mmap2_event(struct machine *machine,
1639                                  union perf_event *event,
1640                                  struct perf_sample *sample)
1641 {
1642         struct thread *thread;
1643         struct map *map;
1644         int ret = 0;
1645
1646         if (dump_trace)
1647                 perf_event__fprintf_mmap2(event, stdout);
1648
1649         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1650             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1651                 ret = machine__process_kernel_mmap_event(machine, event);
1652                 if (ret < 0)
1653                         goto out_problem;
1654                 return 0;
1655         }
1656
1657         thread = machine__findnew_thread(machine, event->mmap2.pid,
1658                                         event->mmap2.tid);
1659         if (thread == NULL)
1660                 goto out_problem;
1661
1662         map = map__new(machine, event->mmap2.start,
1663                         event->mmap2.len, event->mmap2.pgoff,
1664                         event->mmap2.maj,
1665                         event->mmap2.min, event->mmap2.ino,
1666                         event->mmap2.ino_generation,
1667                         event->mmap2.prot,
1668                         event->mmap2.flags,
1669                         event->mmap2.filename, thread);
1670
1671         if (map == NULL)
1672                 goto out_problem_map;
1673
1674         ret = thread__insert_map(thread, map);
1675         if (ret)
1676                 goto out_problem_insert;
1677
1678         thread__put(thread);
1679         map__put(map);
1680         return 0;
1681
1682 out_problem_insert:
1683         map__put(map);
1684 out_problem_map:
1685         thread__put(thread);
1686 out_problem:
1687         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1688         return 0;
1689 }
1690
1691 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1692                                 struct perf_sample *sample)
1693 {
1694         struct thread *thread;
1695         struct map *map;
1696         u32 prot = 0;
1697         int ret = 0;
1698
1699         if (dump_trace)
1700                 perf_event__fprintf_mmap(event, stdout);
1701
1702         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1703             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1704                 ret = machine__process_kernel_mmap_event(machine, event);
1705                 if (ret < 0)
1706                         goto out_problem;
1707                 return 0;
1708         }
1709
1710         thread = machine__findnew_thread(machine, event->mmap.pid,
1711                                          event->mmap.tid);
1712         if (thread == NULL)
1713                 goto out_problem;
1714
1715         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1716                 prot = PROT_EXEC;
1717
1718         map = map__new(machine, event->mmap.start,
1719                         event->mmap.len, event->mmap.pgoff,
1720                         0, 0, 0, 0, prot, 0,
1721                         event->mmap.filename,
1722                         thread);
1723
1724         if (map == NULL)
1725                 goto out_problem_map;
1726
1727         ret = thread__insert_map(thread, map);
1728         if (ret)
1729                 goto out_problem_insert;
1730
1731         thread__put(thread);
1732         map__put(map);
1733         return 0;
1734
1735 out_problem_insert:
1736         map__put(map);
1737 out_problem_map:
1738         thread__put(thread);
1739 out_problem:
1740         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1741         return 0;
1742 }
1743
1744 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1745 {
1746         struct threads *threads = machine__threads(machine, th->tid);
1747
1748         if (threads->last_match == th)
1749                 threads__set_last_match(threads, NULL);
1750
1751         BUG_ON(refcount_read(&th->refcnt) == 0);
1752         if (lock)
1753                 down_write(&threads->lock);
1754         rb_erase_cached(&th->rb_node, &threads->entries);
1755         RB_CLEAR_NODE(&th->rb_node);
1756         --threads->nr;
1757         /*
1758          * Move it first to the dead_threads list, then drop the reference,
1759          * if this is the last reference, then the thread__delete destructor
1760          * will be called and we will remove it from the dead_threads list.
1761          */
1762         list_add_tail(&th->node, &threads->dead);
1763         if (lock)
1764                 up_write(&threads->lock);
1765         thread__put(th);
1766 }
1767
1768 void machine__remove_thread(struct machine *machine, struct thread *th)
1769 {
1770         return __machine__remove_thread(machine, th, true);
1771 }
1772
1773 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1774                                 struct perf_sample *sample)
1775 {
1776         struct thread *thread = machine__find_thread(machine,
1777                                                      event->fork.pid,
1778                                                      event->fork.tid);
1779         struct thread *parent = machine__findnew_thread(machine,
1780                                                         event->fork.ppid,
1781                                                         event->fork.ptid);
1782         bool do_maps_clone = true;
1783         int err = 0;
1784
1785         if (dump_trace)
1786                 perf_event__fprintf_task(event, stdout);
1787
1788         /*
1789          * There may be an existing thread that is not actually the parent,
1790          * either because we are processing events out of order, or because the
1791          * (fork) event that would have removed the thread was lost. Assume the
1792          * latter case and continue on as best we can.
1793          */
1794         if (parent->pid_ != (pid_t)event->fork.ppid) {
1795                 dump_printf("removing erroneous parent thread %d/%d\n",
1796                             parent->pid_, parent->tid);
1797                 machine__remove_thread(machine, parent);
1798                 thread__put(parent);
1799                 parent = machine__findnew_thread(machine, event->fork.ppid,
1800                                                  event->fork.ptid);
1801         }
1802
1803         /* if a thread currently exists for the thread id remove it */
1804         if (thread != NULL) {
1805                 machine__remove_thread(machine, thread);
1806                 thread__put(thread);
1807         }
1808
1809         thread = machine__findnew_thread(machine, event->fork.pid,
1810                                          event->fork.tid);
1811         /*
1812          * When synthesizing FORK events, we are trying to create thread
1813          * objects for the already running tasks on the machine.
1814          *
1815          * Normally, for a kernel FORK event, we want to clone the parent's
1816          * maps because that is what the kernel just did.
1817          *
1818          * But when synthesizing, this should not be done.  If we do, we end up
1819          * with overlapping maps as we process the sythesized MMAP2 events that
1820          * get delivered shortly thereafter.
1821          *
1822          * Use the FORK event misc flags in an internal way to signal this
1823          * situation, so we can elide the map clone when appropriate.
1824          */
1825         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1826                 do_maps_clone = false;
1827
1828         if (thread == NULL || parent == NULL ||
1829             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1830                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1831                 err = -1;
1832         }
1833         thread__put(thread);
1834         thread__put(parent);
1835
1836         return err;
1837 }
1838
1839 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1840                                 struct perf_sample *sample __maybe_unused)
1841 {
1842         struct thread *thread = machine__find_thread(machine,
1843                                                      event->fork.pid,
1844                                                      event->fork.tid);
1845
1846         if (dump_trace)
1847                 perf_event__fprintf_task(event, stdout);
1848
1849         if (thread != NULL) {
1850                 thread__exited(thread);
1851                 thread__put(thread);
1852         }
1853
1854         return 0;
1855 }
1856
1857 int machine__process_event(struct machine *machine, union perf_event *event,
1858                            struct perf_sample *sample)
1859 {
1860         int ret;
1861
1862         switch (event->header.type) {
1863         case PERF_RECORD_COMM:
1864                 ret = machine__process_comm_event(machine, event, sample); break;
1865         case PERF_RECORD_MMAP:
1866                 ret = machine__process_mmap_event(machine, event, sample); break;
1867         case PERF_RECORD_NAMESPACES:
1868                 ret = machine__process_namespaces_event(machine, event, sample); break;
1869         case PERF_RECORD_MMAP2:
1870                 ret = machine__process_mmap2_event(machine, event, sample); break;
1871         case PERF_RECORD_FORK:
1872                 ret = machine__process_fork_event(machine, event, sample); break;
1873         case PERF_RECORD_EXIT:
1874                 ret = machine__process_exit_event(machine, event, sample); break;
1875         case PERF_RECORD_LOST:
1876                 ret = machine__process_lost_event(machine, event, sample); break;
1877         case PERF_RECORD_AUX:
1878                 ret = machine__process_aux_event(machine, event); break;
1879         case PERF_RECORD_ITRACE_START:
1880                 ret = machine__process_itrace_start_event(machine, event); break;
1881         case PERF_RECORD_LOST_SAMPLES:
1882                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1883         case PERF_RECORD_SWITCH:
1884         case PERF_RECORD_SWITCH_CPU_WIDE:
1885                 ret = machine__process_switch_event(machine, event); break;
1886         case PERF_RECORD_KSYMBOL:
1887                 ret = machine__process_ksymbol(machine, event, sample); break;
1888         case PERF_RECORD_BPF_EVENT:
1889                 ret = machine__process_bpf_event(machine, event, sample); break;
1890         default:
1891                 ret = -1;
1892                 break;
1893         }
1894
1895         return ret;
1896 }
1897
1898 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1899 {
1900         if (!regexec(regex, sym->name, 0, NULL, 0))
1901                 return 1;
1902         return 0;
1903 }
1904
1905 static void ip__resolve_ams(struct thread *thread,
1906                             struct addr_map_symbol *ams,
1907                             u64 ip)
1908 {
1909         struct addr_location al;
1910
1911         memset(&al, 0, sizeof(al));
1912         /*
1913          * We cannot use the header.misc hint to determine whether a
1914          * branch stack address is user, kernel, guest, hypervisor.
1915          * Branches may straddle the kernel/user/hypervisor boundaries.
1916          * Thus, we have to try consecutively until we find a match
1917          * or else, the symbol is unknown
1918          */
1919         thread__find_cpumode_addr_location(thread, ip, &al);
1920
1921         ams->addr = ip;
1922         ams->al_addr = al.addr;
1923         ams->sym = al.sym;
1924         ams->map = al.map;
1925         ams->phys_addr = 0;
1926 }
1927
1928 static void ip__resolve_data(struct thread *thread,
1929                              u8 m, struct addr_map_symbol *ams,
1930                              u64 addr, u64 phys_addr)
1931 {
1932         struct addr_location al;
1933
1934         memset(&al, 0, sizeof(al));
1935
1936         thread__find_symbol(thread, m, addr, &al);
1937
1938         ams->addr = addr;
1939         ams->al_addr = al.addr;
1940         ams->sym = al.sym;
1941         ams->map = al.map;
1942         ams->phys_addr = phys_addr;
1943 }
1944
1945 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1946                                      struct addr_location *al)
1947 {
1948         struct mem_info *mi = mem_info__new();
1949
1950         if (!mi)
1951                 return NULL;
1952
1953         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1954         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1955                          sample->addr, sample->phys_addr);
1956         mi->data_src.val = sample->data_src;
1957
1958         return mi;
1959 }
1960
1961 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
1962 {
1963         char *srcline = NULL;
1964
1965         if (!map || callchain_param.key == CCKEY_FUNCTION)
1966                 return srcline;
1967
1968         srcline = srcline__tree_find(&map->dso->srclines, ip);
1969         if (!srcline) {
1970                 bool show_sym = false;
1971                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
1972
1973                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
1974                                       sym, show_sym, show_addr, ip);
1975                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
1976         }
1977
1978         return srcline;
1979 }
1980
1981 struct iterations {
1982         int nr_loop_iter;
1983         u64 cycles;
1984 };
1985
1986 static int add_callchain_ip(struct thread *thread,
1987                             struct callchain_cursor *cursor,
1988                             struct symbol **parent,
1989                             struct addr_location *root_al,
1990                             u8 *cpumode,
1991                             u64 ip,
1992                             bool branch,
1993                             struct branch_flags *flags,
1994                             struct iterations *iter,
1995                             u64 branch_from)
1996 {
1997         struct addr_location al;
1998         int nr_loop_iter = 0;
1999         u64 iter_cycles = 0;
2000         const char *srcline = NULL;
2001
2002         al.filtered = 0;
2003         al.sym = NULL;
2004         if (!cpumode) {
2005                 thread__find_cpumode_addr_location(thread, ip, &al);
2006         } else {
2007                 if (ip >= PERF_CONTEXT_MAX) {
2008                         switch (ip) {
2009                         case PERF_CONTEXT_HV:
2010                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2011                                 break;
2012                         case PERF_CONTEXT_KERNEL:
2013                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2014                                 break;
2015                         case PERF_CONTEXT_USER:
2016                                 *cpumode = PERF_RECORD_MISC_USER;
2017                                 break;
2018                         default:
2019                                 pr_debug("invalid callchain context: "
2020                                          "%"PRId64"\n", (s64) ip);
2021                                 /*
2022                                  * It seems the callchain is corrupted.
2023                                  * Discard all.
2024                                  */
2025                                 callchain_cursor_reset(cursor);
2026                                 return 1;
2027                         }
2028                         return 0;
2029                 }
2030                 thread__find_symbol(thread, *cpumode, ip, &al);
2031         }
2032
2033         if (al.sym != NULL) {
2034                 if (perf_hpp_list.parent && !*parent &&
2035                     symbol__match_regex(al.sym, &parent_regex))
2036                         *parent = al.sym;
2037                 else if (have_ignore_callees && root_al &&
2038                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2039                         /* Treat this symbol as the root,
2040                            forgetting its callees. */
2041                         *root_al = al;
2042                         callchain_cursor_reset(cursor);
2043                 }
2044         }
2045
2046         if (symbol_conf.hide_unresolved && al.sym == NULL)
2047                 return 0;
2048
2049         if (iter) {
2050                 nr_loop_iter = iter->nr_loop_iter;
2051                 iter_cycles = iter->cycles;
2052         }
2053
2054         srcline = callchain_srcline(al.map, al.sym, al.addr);
2055         return callchain_cursor_append(cursor, ip, al.map, al.sym,
2056                                        branch, flags, nr_loop_iter,
2057                                        iter_cycles, branch_from, srcline);
2058 }
2059
2060 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2061                                            struct addr_location *al)
2062 {
2063         unsigned int i;
2064         const struct branch_stack *bs = sample->branch_stack;
2065         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2066
2067         if (!bi)
2068                 return NULL;
2069
2070         for (i = 0; i < bs->nr; i++) {
2071                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2072                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2073                 bi[i].flags = bs->entries[i].flags;
2074         }
2075         return bi;
2076 }
2077
2078 static void save_iterations(struct iterations *iter,
2079                             struct branch_entry *be, int nr)
2080 {
2081         int i;
2082
2083         iter->nr_loop_iter++;
2084         iter->cycles = 0;
2085
2086         for (i = 0; i < nr; i++)
2087                 iter->cycles += be[i].flags.cycles;
2088 }
2089
2090 #define CHASHSZ 127
2091 #define CHASHBITS 7
2092 #define NO_ENTRY 0xff
2093
2094 #define PERF_MAX_BRANCH_DEPTH 127
2095
2096 /* Remove loops. */
2097 static int remove_loops(struct branch_entry *l, int nr,
2098                         struct iterations *iter)
2099 {
2100         int i, j, off;
2101         unsigned char chash[CHASHSZ];
2102
2103         memset(chash, NO_ENTRY, sizeof(chash));
2104
2105         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2106
2107         for (i = 0; i < nr; i++) {
2108                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2109
2110                 /* no collision handling for now */
2111                 if (chash[h] == NO_ENTRY) {
2112                         chash[h] = i;
2113                 } else if (l[chash[h]].from == l[i].from) {
2114                         bool is_loop = true;
2115                         /* check if it is a real loop */
2116                         off = 0;
2117                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2118                                 if (l[j].from != l[i + off].from) {
2119                                         is_loop = false;
2120                                         break;
2121                                 }
2122                         if (is_loop) {
2123                                 j = nr - (i + off);
2124                                 if (j > 0) {
2125                                         save_iterations(iter + i + off,
2126                                                 l + i, off);
2127
2128                                         memmove(iter + i, iter + i + off,
2129                                                 j * sizeof(*iter));
2130
2131                                         memmove(l + i, l + i + off,
2132                                                 j * sizeof(*l));
2133                                 }
2134
2135                                 nr -= off;
2136                         }
2137                 }
2138         }
2139         return nr;
2140 }
2141
2142 /*
2143  * Recolve LBR callstack chain sample
2144  * Return:
2145  * 1 on success get LBR callchain information
2146  * 0 no available LBR callchain information, should try fp
2147  * negative error code on other errors.
2148  */
2149 static int resolve_lbr_callchain_sample(struct thread *thread,
2150                                         struct callchain_cursor *cursor,
2151                                         struct perf_sample *sample,
2152                                         struct symbol **parent,
2153                                         struct addr_location *root_al,
2154                                         int max_stack)
2155 {
2156         struct ip_callchain *chain = sample->callchain;
2157         int chain_nr = min(max_stack, (int)chain->nr), i;
2158         u8 cpumode = PERF_RECORD_MISC_USER;
2159         u64 ip, branch_from = 0;
2160
2161         for (i = 0; i < chain_nr; i++) {
2162                 if (chain->ips[i] == PERF_CONTEXT_USER)
2163                         break;
2164         }
2165
2166         /* LBR only affects the user callchain */
2167         if (i != chain_nr) {
2168                 struct branch_stack *lbr_stack = sample->branch_stack;
2169                 int lbr_nr = lbr_stack->nr, j, k;
2170                 bool branch;
2171                 struct branch_flags *flags;
2172                 /*
2173                  * LBR callstack can only get user call chain.
2174                  * The mix_chain_nr is kernel call chain
2175                  * number plus LBR user call chain number.
2176                  * i is kernel call chain number,
2177                  * 1 is PERF_CONTEXT_USER,
2178                  * lbr_nr + 1 is the user call chain number.
2179                  * For details, please refer to the comments
2180                  * in callchain__printf
2181                  */
2182                 int mix_chain_nr = i + 1 + lbr_nr + 1;
2183
2184                 for (j = 0; j < mix_chain_nr; j++) {
2185                         int err;
2186                         branch = false;
2187                         flags = NULL;
2188
2189                         if (callchain_param.order == ORDER_CALLEE) {
2190                                 if (j < i + 1)
2191                                         ip = chain->ips[j];
2192                                 else if (j > i + 1) {
2193                                         k = j - i - 2;
2194                                         ip = lbr_stack->entries[k].from;
2195                                         branch = true;
2196                                         flags = &lbr_stack->entries[k].flags;
2197                                 } else {
2198                                         ip = lbr_stack->entries[0].to;
2199                                         branch = true;
2200                                         flags = &lbr_stack->entries[0].flags;
2201                                         branch_from =
2202                                                 lbr_stack->entries[0].from;
2203                                 }
2204                         } else {
2205                                 if (j < lbr_nr) {
2206                                         k = lbr_nr - j - 1;
2207                                         ip = lbr_stack->entries[k].from;
2208                                         branch = true;
2209                                         flags = &lbr_stack->entries[k].flags;
2210                                 }
2211                                 else if (j > lbr_nr)
2212                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
2213                                 else {
2214                                         ip = lbr_stack->entries[0].to;
2215                                         branch = true;
2216                                         flags = &lbr_stack->entries[0].flags;
2217                                         branch_from =
2218                                                 lbr_stack->entries[0].from;
2219                                 }
2220                         }
2221
2222                         err = add_callchain_ip(thread, cursor, parent,
2223                                                root_al, &cpumode, ip,
2224                                                branch, flags, NULL,
2225                                                branch_from);
2226                         if (err)
2227                                 return (err < 0) ? err : 0;
2228                 }
2229                 return 1;
2230         }
2231
2232         return 0;
2233 }
2234
2235 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2236                              struct callchain_cursor *cursor,
2237                              struct symbol **parent,
2238                              struct addr_location *root_al,
2239                              u8 *cpumode, int ent)
2240 {
2241         int err = 0;
2242
2243         while (--ent >= 0) {
2244                 u64 ip = chain->ips[ent];
2245
2246                 if (ip >= PERF_CONTEXT_MAX) {
2247                         err = add_callchain_ip(thread, cursor, parent,
2248                                                root_al, cpumode, ip,
2249                                                false, NULL, NULL, 0);
2250                         break;
2251                 }
2252         }
2253         return err;
2254 }
2255
2256 static int thread__resolve_callchain_sample(struct thread *thread,
2257                                             struct callchain_cursor *cursor,
2258                                             struct perf_evsel *evsel,
2259                                             struct perf_sample *sample,
2260                                             struct symbol **parent,
2261                                             struct addr_location *root_al,
2262                                             int max_stack)
2263 {
2264         struct branch_stack *branch = sample->branch_stack;
2265         struct ip_callchain *chain = sample->callchain;
2266         int chain_nr = 0;
2267         u8 cpumode = PERF_RECORD_MISC_USER;
2268         int i, j, err, nr_entries;
2269         int skip_idx = -1;
2270         int first_call = 0;
2271
2272         if (chain)
2273                 chain_nr = chain->nr;
2274
2275         if (perf_evsel__has_branch_callstack(evsel)) {
2276                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2277                                                    root_al, max_stack);
2278                 if (err)
2279                         return (err < 0) ? err : 0;
2280         }
2281
2282         /*
2283          * Based on DWARF debug information, some architectures skip
2284          * a callchain entry saved by the kernel.
2285          */
2286         skip_idx = arch_skip_callchain_idx(thread, chain);
2287
2288         /*
2289          * Add branches to call stack for easier browsing. This gives
2290          * more context for a sample than just the callers.
2291          *
2292          * This uses individual histograms of paths compared to the
2293          * aggregated histograms the normal LBR mode uses.
2294          *
2295          * Limitations for now:
2296          * - No extra filters
2297          * - No annotations (should annotate somehow)
2298          */
2299
2300         if (branch && callchain_param.branch_callstack) {
2301                 int nr = min(max_stack, (int)branch->nr);
2302                 struct branch_entry be[nr];
2303                 struct iterations iter[nr];
2304
2305                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2306                         pr_warning("corrupted branch chain. skipping...\n");
2307                         goto check_calls;
2308                 }
2309
2310                 for (i = 0; i < nr; i++) {
2311                         if (callchain_param.order == ORDER_CALLEE) {
2312                                 be[i] = branch->entries[i];
2313
2314                                 if (chain == NULL)
2315                                         continue;
2316
2317                                 /*
2318                                  * Check for overlap into the callchain.
2319                                  * The return address is one off compared to
2320                                  * the branch entry. To adjust for this
2321                                  * assume the calling instruction is not longer
2322                                  * than 8 bytes.
2323                                  */
2324                                 if (i == skip_idx ||
2325                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2326                                         first_call++;
2327                                 else if (be[i].from < chain->ips[first_call] &&
2328                                     be[i].from >= chain->ips[first_call] - 8)
2329                                         first_call++;
2330                         } else
2331                                 be[i] = branch->entries[branch->nr - i - 1];
2332                 }
2333
2334                 memset(iter, 0, sizeof(struct iterations) * nr);
2335                 nr = remove_loops(be, nr, iter);
2336
2337                 for (i = 0; i < nr; i++) {
2338                         err = add_callchain_ip(thread, cursor, parent,
2339                                                root_al,
2340                                                NULL, be[i].to,
2341                                                true, &be[i].flags,
2342                                                NULL, be[i].from);
2343
2344                         if (!err)
2345                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2346                                                        NULL, be[i].from,
2347                                                        true, &be[i].flags,
2348                                                        &iter[i], 0);
2349                         if (err == -EINVAL)
2350                                 break;
2351                         if (err)
2352                                 return err;
2353                 }
2354
2355                 if (chain_nr == 0)
2356                         return 0;
2357
2358                 chain_nr -= nr;
2359         }
2360
2361 check_calls:
2362         if (callchain_param.order != ORDER_CALLEE) {
2363                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2364                                         &cpumode, chain->nr - first_call);
2365                 if (err)
2366                         return (err < 0) ? err : 0;
2367         }
2368         for (i = first_call, nr_entries = 0;
2369              i < chain_nr && nr_entries < max_stack; i++) {
2370                 u64 ip;
2371
2372                 if (callchain_param.order == ORDER_CALLEE)
2373                         j = i;
2374                 else
2375                         j = chain->nr - i - 1;
2376
2377 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2378                 if (j == skip_idx)
2379                         continue;
2380 #endif
2381                 ip = chain->ips[j];
2382                 if (ip < PERF_CONTEXT_MAX)
2383                        ++nr_entries;
2384                 else if (callchain_param.order != ORDER_CALLEE) {
2385                         err = find_prev_cpumode(chain, thread, cursor, parent,
2386                                                 root_al, &cpumode, j);
2387                         if (err)
2388                                 return (err < 0) ? err : 0;
2389                         continue;
2390                 }
2391
2392                 err = add_callchain_ip(thread, cursor, parent,
2393                                        root_al, &cpumode, ip,
2394                                        false, NULL, NULL, 0);
2395
2396                 if (err)
2397                         return (err < 0) ? err : 0;
2398         }
2399
2400         return 0;
2401 }
2402
2403 static int append_inlines(struct callchain_cursor *cursor,
2404                           struct map *map, struct symbol *sym, u64 ip)
2405 {
2406         struct inline_node *inline_node;
2407         struct inline_list *ilist;
2408         u64 addr;
2409         int ret = 1;
2410
2411         if (!symbol_conf.inline_name || !map || !sym)
2412                 return ret;
2413
2414         addr = map__map_ip(map, ip);
2415         addr = map__rip_2objdump(map, addr);
2416
2417         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2418         if (!inline_node) {
2419                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2420                 if (!inline_node)
2421                         return ret;
2422                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2423         }
2424
2425         list_for_each_entry(ilist, &inline_node->val, list) {
2426                 ret = callchain_cursor_append(cursor, ip, map,
2427                                               ilist->symbol, false,
2428                                               NULL, 0, 0, 0, ilist->srcline);
2429
2430                 if (ret != 0)
2431                         return ret;
2432         }
2433
2434         return ret;
2435 }
2436
2437 static int unwind_entry(struct unwind_entry *entry, void *arg)
2438 {
2439         struct callchain_cursor *cursor = arg;
2440         const char *srcline = NULL;
2441         u64 addr = entry->ip;
2442
2443         if (symbol_conf.hide_unresolved && entry->sym == NULL)
2444                 return 0;
2445
2446         if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2447                 return 0;
2448
2449         /*
2450          * Convert entry->ip from a virtual address to an offset in
2451          * its corresponding binary.
2452          */
2453         if (entry->map)
2454                 addr = map__map_ip(entry->map, entry->ip);
2455
2456         srcline = callchain_srcline(entry->map, entry->sym, addr);
2457         return callchain_cursor_append(cursor, entry->ip,
2458                                        entry->map, entry->sym,
2459                                        false, NULL, 0, 0, 0, srcline);
2460 }
2461
2462 static int thread__resolve_callchain_unwind(struct thread *thread,
2463                                             struct callchain_cursor *cursor,
2464                                             struct perf_evsel *evsel,
2465                                             struct perf_sample *sample,
2466                                             int max_stack)
2467 {
2468         /* Can we do dwarf post unwind? */
2469         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2470               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
2471                 return 0;
2472
2473         /* Bail out if nothing was captured. */
2474         if ((!sample->user_regs.regs) ||
2475             (!sample->user_stack.size))
2476                 return 0;
2477
2478         return unwind__get_entries(unwind_entry, cursor,
2479                                    thread, sample, max_stack);
2480 }
2481
2482 int thread__resolve_callchain(struct thread *thread,
2483                               struct callchain_cursor *cursor,
2484                               struct perf_evsel *evsel,
2485                               struct perf_sample *sample,
2486                               struct symbol **parent,
2487                               struct addr_location *root_al,
2488                               int max_stack)
2489 {
2490         int ret = 0;
2491
2492         callchain_cursor_reset(cursor);
2493
2494         if (callchain_param.order == ORDER_CALLEE) {
2495                 ret = thread__resolve_callchain_sample(thread, cursor,
2496                                                        evsel, sample,
2497                                                        parent, root_al,
2498                                                        max_stack);
2499                 if (ret)
2500                         return ret;
2501                 ret = thread__resolve_callchain_unwind(thread, cursor,
2502                                                        evsel, sample,
2503                                                        max_stack);
2504         } else {
2505                 ret = thread__resolve_callchain_unwind(thread, cursor,
2506                                                        evsel, sample,
2507                                                        max_stack);
2508                 if (ret)
2509                         return ret;
2510                 ret = thread__resolve_callchain_sample(thread, cursor,
2511                                                        evsel, sample,
2512                                                        parent, root_al,
2513                                                        max_stack);
2514         }
2515
2516         return ret;
2517 }
2518
2519 int machine__for_each_thread(struct machine *machine,
2520                              int (*fn)(struct thread *thread, void *p),
2521                              void *priv)
2522 {
2523         struct threads *threads;
2524         struct rb_node *nd;
2525         struct thread *thread;
2526         int rc = 0;
2527         int i;
2528
2529         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2530                 threads = &machine->threads[i];
2531                 for (nd = rb_first_cached(&threads->entries); nd;
2532                      nd = rb_next(nd)) {
2533                         thread = rb_entry(nd, struct thread, rb_node);
2534                         rc = fn(thread, priv);
2535                         if (rc != 0)
2536                                 return rc;
2537                 }
2538
2539                 list_for_each_entry(thread, &threads->dead, node) {
2540                         rc = fn(thread, priv);
2541                         if (rc != 0)
2542                                 return rc;
2543                 }
2544         }
2545         return rc;
2546 }
2547
2548 int machines__for_each_thread(struct machines *machines,
2549                               int (*fn)(struct thread *thread, void *p),
2550                               void *priv)
2551 {
2552         struct rb_node *nd;
2553         int rc = 0;
2554
2555         rc = machine__for_each_thread(&machines->host, fn, priv);
2556         if (rc != 0)
2557                 return rc;
2558
2559         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2560                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2561
2562                 rc = machine__for_each_thread(machine, fn, priv);
2563                 if (rc != 0)
2564                         return rc;
2565         }
2566         return rc;
2567 }
2568
2569 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
2570                                   struct target *target, struct thread_map *threads,
2571                                   perf_event__handler_t process, bool data_mmap,
2572                                   unsigned int nr_threads_synthesize)
2573 {
2574         if (target__has_task(target))
2575                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
2576         else if (target__has_cpu(target))
2577                 return perf_event__synthesize_threads(tool, process,
2578                                                       machine, data_mmap,
2579                                                       nr_threads_synthesize);
2580         /* command specified */
2581         return 0;
2582 }
2583
2584 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2585 {
2586         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
2587                 return -1;
2588
2589         return machine->current_tid[cpu];
2590 }
2591
2592 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2593                              pid_t tid)
2594 {
2595         struct thread *thread;
2596
2597         if (cpu < 0)
2598                 return -EINVAL;
2599
2600         if (!machine->current_tid) {
2601                 int i;
2602
2603                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
2604                 if (!machine->current_tid)
2605                         return -ENOMEM;
2606                 for (i = 0; i < MAX_NR_CPUS; i++)
2607                         machine->current_tid[i] = -1;
2608         }
2609
2610         if (cpu >= MAX_NR_CPUS) {
2611                 pr_err("Requested CPU %d too large. ", cpu);
2612                 pr_err("Consider raising MAX_NR_CPUS\n");
2613                 return -EINVAL;
2614         }
2615
2616         machine->current_tid[cpu] = tid;
2617
2618         thread = machine__findnew_thread(machine, pid, tid);
2619         if (!thread)
2620                 return -ENOMEM;
2621
2622         thread->cpu = cpu;
2623         thread__put(thread);
2624
2625         return 0;
2626 }
2627
2628 /*
2629  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2630  * normalized arch is needed.
2631  */
2632 bool machine__is(struct machine *machine, const char *arch)
2633 {
2634         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2635 }
2636
2637 int machine__nr_cpus_avail(struct machine *machine)
2638 {
2639         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2640 }
2641
2642 int machine__get_kernel_start(struct machine *machine)
2643 {
2644         struct map *map = machine__kernel_map(machine);
2645         int err = 0;
2646
2647         /*
2648          * The only addresses above 2^63 are kernel addresses of a 64-bit
2649          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2650          * all addresses including kernel addresses are less than 2^32.  In
2651          * that case (32-bit system), if the kernel mapping is unknown, all
2652          * addresses will be assumed to be in user space - see
2653          * machine__kernel_ip().
2654          */
2655         machine->kernel_start = 1ULL << 63;
2656         if (map) {
2657                 err = map__load(map);
2658                 /*
2659                  * On x86_64, PTI entry trampolines are less than the
2660                  * start of kernel text, but still above 2^63. So leave
2661                  * kernel_start = 1ULL << 63 for x86_64.
2662                  */
2663                 if (!err && !machine__is(machine, "x86_64"))
2664                         machine->kernel_start = map->start;
2665         }
2666         return err;
2667 }
2668
2669 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2670 {
2671         u8 addr_cpumode = cpumode;
2672         bool kernel_ip;
2673
2674         if (!machine->single_address_space)
2675                 goto out;
2676
2677         kernel_ip = machine__kernel_ip(machine, addr);
2678         switch (cpumode) {
2679         case PERF_RECORD_MISC_KERNEL:
2680         case PERF_RECORD_MISC_USER:
2681                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2682                                            PERF_RECORD_MISC_USER;
2683                 break;
2684         case PERF_RECORD_MISC_GUEST_KERNEL:
2685         case PERF_RECORD_MISC_GUEST_USER:
2686                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2687                                            PERF_RECORD_MISC_GUEST_USER;
2688                 break;
2689         default:
2690                 break;
2691         }
2692 out:
2693         return addr_cpumode;
2694 }
2695
2696 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2697 {
2698         return dsos__findnew(&machine->dsos, filename);
2699 }
2700
2701 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2702 {
2703         struct machine *machine = vmachine;
2704         struct map *map;
2705         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2706
2707         if (sym == NULL)
2708                 return NULL;
2709
2710         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2711         *addrp = map->unmap_ip(map, sym->start);
2712         return sym->name;
2713 }