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 <linux/ctype.h>
  28 #include <symbol/kallsyms.h>
  29 #include <linux/mman.h>
  30 #include <linux/zalloc.h>
  31
  32 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  33
  34 static void dsos__init(struct dsos *dsos)
  35 {
  36         INIT_LIST_HEAD(&dsos->head);
  37         dsos->root = RB_ROOT;
  38         init_rwsem(&dsos->lock);
  39 }
  40
  41 static void machine__threads_init(struct machine *machine)
  42 {
  43         int i;
  44
  45         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  46                 struct threads *threads = &machine->threads[i];
  47                 threads->entries = RB_ROOT_CACHED;
  48                 init_rwsem(&threads->lock);
  49                 threads->nr = 0;
  50                 INIT_LIST_HEAD(&threads->dead);
  51                 threads->last_match = NULL;
  52         }
  53 }
  54
  55 static int machine__set_mmap_name(struct machine *machine)
  56 {
  57         if (machine__is_host(machine))
  58                 machine->mmap_name = strdup("[kernel.kallsyms]");
  59         else if (machine__is_default_guest(machine))
  60                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  61         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  62                           machine->pid) < 0)
  63                 machine->mmap_name = NULL;
  64
  65         return machine->mmap_name ? 0 : -ENOMEM;
  66 }
  67
  68 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  69 {
  70         int err = -ENOMEM;
  71
  72         memset(machine, 0, sizeof(*machine));
  73         map_groups__init(&machine->kmaps, machine);
  74         RB_CLEAR_NODE(&machine->rb_node);
  75         dsos__init(&machine->dsos);
  76
  77         machine__threads_init(machine);
  78
  79         machine->vdso_info = NULL;
  80         machine->env = NULL;
  81
  82         machine->pid = pid;
  83
  84         machine->id_hdr_size = 0;
  85         machine->kptr_restrict_warned = false;
  86         machine->comm_exec = false;
  87         machine->kernel_start = 0;
  88         machine->vmlinux_map = NULL;
  89
  90         machine->root_dir = strdup(root_dir);
  91         if (machine->root_dir == NULL)
  92                 return -ENOMEM;
  93
  94         if (machine__set_mmap_name(machine))
  95                 goto out;
  96
  97         if (pid != HOST_KERNEL_ID) {
  98                 struct thread *thread = machine__findnew_thread(machine, -1,
  99                                                                 pid);
 100                 char comm[64];
 101
 102                 if (thread == NULL)
 103                         goto out;
 104
 105                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 106                 thread__set_comm(thread, comm, 0);
 107                 thread__put(thread);
 108         }
 109
 110         machine->current_tid = NULL;
 111         err = 0;
 112
 113 out:
 114         if (err) {
 115                 zfree(&machine->root_dir);
 116                 zfree(&machine->mmap_name);
 117         }
 118         return 0;
 119 }
 120
 121 struct machine *machine__new_host(void)
 122 {
 123         struct machine *machine = malloc(sizeof(*machine));
 124
 125         if (machine != NULL) {
 126                 machine__init(machine, "", HOST_KERNEL_ID);
 127
 128                 if (machine__create_kernel_maps(machine) < 0)
 129                         goto out_delete;
 130         }
 131
 132         return machine;
 133 out_delete:
 134         free(machine);
 135         return NULL;
 136 }
 137
 138 struct machine *machine__new_kallsyms(void)
 139 {
 140         struct machine *machine = machine__new_host();
 141         /*
 142          * FIXME:
 143          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 144          *    ask for not using the kcore parsing code, once this one is fixed
 145          *    to create a map per module.
 146          */
 147         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 148                 machine__delete(machine);
 149                 machine = NULL;
 150         }
 151
 152         return machine;
 153 }
 154
 155 static void dsos__purge(struct dsos *dsos)
 156 {
 157         struct dso *pos, *n;
 158
 159         down_write(&dsos->lock);
 160
 161         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 162                 RB_CLEAR_NODE(&pos->rb_node);
 163                 pos->root = NULL;
 164                 list_del_init(&pos->node);
 165                 dso__put(pos);
 166         }
 167
 168         up_write(&dsos->lock);
 169 }
 170
 171 static void dsos__exit(struct dsos *dsos)
 172 {
 173         dsos__purge(dsos);
 174         exit_rwsem(&dsos->lock);
 175 }
 176
 177 void machine__delete_threads(struct machine *machine)
 178 {
 179         struct rb_node *nd;
 180         int i;
 181
 182         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 183                 struct threads *threads = &machine->threads[i];
 184                 down_write(&threads->lock);
 185                 nd = rb_first_cached(&threads->entries);
 186                 while (nd) {
 187                         struct thread *t = rb_entry(nd, struct thread, rb_node);
 188
 189                         nd = rb_next(nd);
 190                         __machine__remove_thread(machine, t, false);
 191                 }
 192                 up_write(&threads->lock);
 193         }
 194 }
 195
 196 void machine__exit(struct machine *machine)
 197 {
 198         int i;
 199
 200         if (machine == NULL)
 201                 return;
 202
 203         machine__destroy_kernel_maps(machine);
 204         map_groups__exit(&machine->kmaps);
 205         dsos__exit(&machine->dsos);
 206         machine__exit_vdso(machine);
 207         zfree(&machine->root_dir);
 208         zfree(&machine->mmap_name);
 209         zfree(&machine->current_tid);
 210
 211         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 212                 struct threads *threads = &machine->threads[i];
 213                 struct thread *thread, *n;
 214                 /*
 215                  * Forget about the dead, at this point whatever threads were
 216                  * left in the dead lists better have a reference count taken
 217                  * by who is using them, and then, when they drop those references
 218                  * and it finally hits zero, thread__put() will check and see that
 219                  * its not in the dead threads list and will not try to remove it
 220                  * from there, just calling thread__delete() straight away.
 221                  */
 222                 list_for_each_entry_safe(thread, n, &threads->dead, node)
 223                         list_del_init(&thread->node);
 224
 225                 exit_rwsem(&threads->lock);
 226         }
 227 }
 228
 229 void machine__delete(struct machine *machine)
 230 {
 231         if (machine) {
 232                 machine__exit(machine);
 233                 free(machine);
 234         }
 235 }
 236
 237 void machines__init(struct machines *machines)
 238 {
 239         machine__init(&machines->host, "", HOST_KERNEL_ID);
 240         machines->guests = RB_ROOT_CACHED;
 241 }
 242
 243 void machines__exit(struct machines *machines)
 244 {
 245         machine__exit(&machines->host);
 246         /* XXX exit guest */
 247 }
 248
 249 struct machine *machines__add(struct machines *machines, pid_t pid,
 250                               const char *root_dir)
 251 {
 252         struct rb_node **p = &machines->guests.rb_root.rb_node;
 253         struct rb_node *parent = NULL;
 254         struct machine *pos, *machine = malloc(sizeof(*machine));
 255         bool leftmost = true;
 256
 257         if (machine == NULL)
 258                 return NULL;
 259
 260         if (machine__init(machine, root_dir, pid) != 0) {
 261                 free(machine);
 262                 return NULL;
 263         }
 264
 265         while (*p != NULL) {
 266                 parent = *p;
 267                 pos = rb_entry(parent, struct machine, rb_node);
 268                 if (pid < pos->pid)
 269                         p = &(*p)->rb_left;
 270                 else {
 271                         p = &(*p)->rb_right;
 272                         leftmost = false;
 273                 }
 274         }
 275
 276         rb_link_node(&machine->rb_node, parent, p);
 277         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 278
 279         return machine;
 280 }
 281
 282 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 283 {
 284         struct rb_node *nd;
 285
 286         machines->host.comm_exec = comm_exec;
 287
 288         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 289                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 290
 291                 machine->comm_exec = comm_exec;
 292         }
 293 }
 294
 295 struct machine *machines__find(struct machines *machines, pid_t pid)
 296 {
 297         struct rb_node **p = &machines->guests.rb_root.rb_node;
 298         struct rb_node *parent = NULL;
 299         struct machine *machine;
 300         struct machine *default_machine = NULL;
 301
 302         if (pid == HOST_KERNEL_ID)
 303                 return &machines->host;
 304
 305         while (*p != NULL) {
 306                 parent = *p;
 307                 machine = rb_entry(parent, struct machine, rb_node);
 308                 if (pid < machine->pid)
 309                         p = &(*p)->rb_left;
 310                 else if (pid > machine->pid)
 311                         p = &(*p)->rb_right;
 312                 else
 313                         return machine;
 314                 if (!machine->pid)
 315                         default_machine = machine;
 316         }
 317
 318         return default_machine;
 319 }
 320
 321 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 322 {
 323         char path[PATH_MAX];
 324         const char *root_dir = "";
 325         struct machine *machine = machines__find(machines, pid);
 326
 327         if (machine && (machine->pid == pid))
 328                 goto out;
 329
 330         if ((pid != HOST_KERNEL_ID) &&
 331             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 332             (symbol_conf.guestmount)) {
 333                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 334                 if (access(path, R_OK)) {
 335                         static struct strlist *seen;
 336
 337                         if (!seen)
 338                                 seen = strlist__new(NULL, NULL);
 339
 340                         if (!strlist__has_entry(seen, path)) {
 341                                 pr_err("Can't access file %s\n", path);
 342                                 strlist__add(seen, path);
 343                         }
 344                         machine = NULL;
 345                         goto out;
 346                 }
 347                 root_dir = path;
 348         }
 349
 350         machine = machines__add(machines, pid, root_dir);
 351 out:
 352         return machine;
 353 }
 354
 355 void machines__process_guests(struct machines *machines,
 356                               machine__process_t process, void *data)
 357 {
 358         struct rb_node *nd;
 359
 360         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 361                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 362                 process(pos, data);
 363         }
 364 }
 365
 366 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 367 {
 368         struct rb_node *node;
 369         struct machine *machine;
 370
 371         machines->host.id_hdr_size = id_hdr_size;
 372
 373         for (node = rb_first_cached(&machines->guests); node;
 374              node = rb_next(node)) {
 375                 machine = rb_entry(node, struct machine, rb_node);
 376                 machine->id_hdr_size = id_hdr_size;
 377         }
 378
 379         return;
 380 }
 381
 382 static void machine__update_thread_pid(struct machine *machine,
 383                                        struct thread *th, pid_t pid)
 384 {
 385         struct thread *leader;
 386
 387         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 388                 return;
 389
 390         th->pid_ = pid;
 391
 392         if (th->pid_ == th->tid)
 393                 return;
 394
 395         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 396         if (!leader)
 397                 goto out_err;
 398
 399         if (!leader->mg)
 400                 leader->mg = map_groups__new(machine);
 401
 402         if (!leader->mg)
 403                 goto out_err;
 404
 405         if (th->mg == leader->mg)
 406                 return;
 407
 408         if (th->mg) {
 409                 /*
 410                  * Maps are created from MMAP events which provide the pid and
 411                  * tid.  Consequently there never should be any maps on a thread
 412                  * with an unknown pid.  Just print an error if there are.
 413                  */
 414                 if (!map_groups__empty(th->mg))
 415                         pr_err("Discarding thread maps for %d:%d\n",
 416                                th->pid_, th->tid);
 417                 map_groups__put(th->mg);
 418         }
 419
 420         th->mg = map_groups__get(leader->mg);
 421 out_put:
 422         thread__put(leader);
 423         return;
 424 out_err:
 425         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 426         goto out_put;
 427 }
 428
 429 /*
 430  * Front-end cache - TID lookups come in blocks,
 431  * so most of the time we dont have to look up
 432  * the full rbtree:
 433  */
 434 static struct thread*
 435 __threads__get_last_match(struct threads *threads, struct machine *machine,
 436                           int pid, int tid)
 437 {
 438         struct thread *th;
 439
 440         th = threads->last_match;
 441         if (th != NULL) {
 442                 if (th->tid == tid) {
 443                         machine__update_thread_pid(machine, th, pid);
 444                         return thread__get(th);
 445                 }
 446
 447                 threads->last_match = NULL;
 448         }
 449
 450         return NULL;
 451 }
 452
 453 static struct thread*
 454 threads__get_last_match(struct threads *threads, struct machine *machine,
 455                         int pid, int tid)
 456 {
 457         struct thread *th = NULL;
 458
 459         if (perf_singlethreaded)
 460                 th = __threads__get_last_match(threads, machine, pid, tid);
 461
 462         return th;
 463 }
 464
 465 static void
 466 __threads__set_last_match(struct threads *threads, struct thread *th)
 467 {
 468         threads->last_match = th;
 469 }
 470
 471 static void
 472 threads__set_last_match(struct threads *threads, struct thread *th)
 473 {
 474         if (perf_singlethreaded)
 475                 __threads__set_last_match(threads, th);
 476 }
 477
 478 /*
 479  * Caller must eventually drop thread->refcnt returned with a successful
 480  * lookup/new thread inserted.
 481  */
 482 static struct thread *____machine__findnew_thread(struct machine *machine,
 483                                                   struct threads *threads,
 484                                                   pid_t pid, pid_t tid,
 485                                                   bool create)
 486 {
 487         struct rb_node **p = &threads->entries.rb_root.rb_node;
 488         struct rb_node *parent = NULL;
 489         struct thread *th;
 490         bool leftmost = true;
 491
 492         th = threads__get_last_match(threads, machine, pid, tid);
 493         if (th)
 494                 return th;
 495
 496         while (*p != NULL) {
 497                 parent = *p;
 498                 th = rb_entry(parent, struct thread, rb_node);
 499
 500                 if (th->tid == tid) {
 501                         threads__set_last_match(threads, th);
 502                         machine__update_thread_pid(machine, th, pid);
 503                         return thread__get(th);
 504                 }
 505
 506                 if (tid < th->tid)
 507                         p = &(*p)->rb_left;
 508                 else {
 509                         p = &(*p)->rb_right;
 510                         leftmost = false;
 511                 }
 512         }
 513
 514         if (!create)
 515                 return NULL;
 516
 517         th = thread__new(pid, tid);
 518         if (th != NULL) {
 519                 rb_link_node(&th->rb_node, parent, p);
 520                 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
 521
 522                 /*
 523                  * We have to initialize map_groups separately
 524                  * after rb tree is updated.
 525                  *
 526                  * The reason is that we call machine__findnew_thread
 527                  * within thread__init_map_groups to find the thread
 528                  * leader and that would screwed the rb tree.
 529                  */
 530                 if (thread__init_map_groups(th, machine)) {
 531                         rb_erase_cached(&th->rb_node, &threads->entries);
 532                         RB_CLEAR_NODE(&th->rb_node);
 533                         thread__put(th);
 534                         return NULL;
 535                 }
 536                 /*
 537                  * It is now in the rbtree, get a ref
 538                  */
 539                 thread__get(th);
 540                 threads__set_last_match(threads, th);
 541                 ++threads->nr;
 542         }
 543
 544         return th;
 545 }
 546
 547 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 548 {
 549         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 550 }
 551
 552 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 553                                        pid_t tid)
 554 {
 555         struct threads *threads = machine__threads(machine, tid);
 556         struct thread *th;
 557
 558         down_write(&threads->lock);
 559         th = __machine__findnew_thread(machine, pid, tid);
 560         up_write(&threads->lock);
 561         return th;
 562 }
 563
 564 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 565                                     pid_t tid)
 566 {
 567         struct threads *threads = machine__threads(machine, tid);
 568         struct thread *th;
 569
 570         down_read(&threads->lock);
 571         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 572         up_read(&threads->lock);
 573         return th;
 574 }
 575
 576 struct comm *machine__thread_exec_comm(struct machine *machine,
 577                                        struct thread *thread)
 578 {
 579         if (machine->comm_exec)
 580                 return thread__exec_comm(thread);
 581         else
 582                 return thread__comm(thread);
 583 }
 584
 585 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 586                                 struct perf_sample *sample)
 587 {
 588         struct thread *thread = machine__findnew_thread(machine,
 589                                                         event->comm.pid,
 590                                                         event->comm.tid);
 591         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 592         int err = 0;
 593
 594         if (exec)
 595                 machine->comm_exec = true;
 596
 597         if (dump_trace)
 598                 perf_event__fprintf_comm(event, stdout);
 599
 600         if (thread == NULL ||
 601             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 602                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 603                 err = -1;
 604         }
 605
 606         thread__put(thread);
 607
 608         return err;
 609 }
 610
 611 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 612                                       union perf_event *event,
 613                                       struct perf_sample *sample __maybe_unused)
 614 {
 615         struct thread *thread = machine__findnew_thread(machine,
 616                                                         event->namespaces.pid,
 617                                                         event->namespaces.tid);
 618         int err = 0;
 619
 620         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 621                   "\nWARNING: kernel seems to support more namespaces than perf"
 622                   " tool.\nTry updating the perf tool..\n\n");
 623
 624         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 625                   "\nWARNING: perf tool seems to support more namespaces than"
 626                   " the kernel.\nTry updating the kernel..\n\n");
 627
 628         if (dump_trace)
 629                 perf_event__fprintf_namespaces(event, stdout);
 630
 631         if (thread == NULL ||
 632             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 633                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 634                 err = -1;
 635         }
 636
 637         thread__put(thread);
 638
 639         return err;
 640 }
 641
 642 int machine__process_lost_event(struct machine *machine __maybe_unused,
 643                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 644 {
 645         dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
 646                     event->lost.id, event->lost.lost);
 647         return 0;
 648 }
 649
 650 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 651                                         union perf_event *event, struct perf_sample *sample)
 652 {
 653         dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
 654                     sample->id, event->lost_samples.lost);
 655         return 0;
 656 }
 657
 658 static struct dso *machine__findnew_module_dso(struct machine *machine,
 659                                                struct kmod_path *m,
 660                                                const char *filename)
 661 {
 662         struct dso *dso;
 663
 664         down_write(&machine->dsos.lock);
 665
 666         dso = __dsos__find(&machine->dsos, m->name, true);
 667         if (!dso) {
 668                 dso = __dsos__addnew(&machine->dsos, m->name);
 669                 if (dso == NULL)
 670                         goto out_unlock;
 671
 672                 dso__set_module_info(dso, m, machine);
 673                 dso__set_long_name(dso, strdup(filename), true);
 674         }
 675
 676         dso__get(dso);
 677 out_unlock:
 678         up_write(&machine->dsos.lock);
 679         return dso;
 680 }
 681
 682 int machine__process_aux_event(struct machine *machine __maybe_unused,
 683                                union perf_event *event)
 684 {
 685         if (dump_trace)
 686                 perf_event__fprintf_aux(event, stdout);
 687         return 0;
 688 }
 689
 690 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 691                                         union perf_event *event)
 692 {
 693         if (dump_trace)
 694                 perf_event__fprintf_itrace_start(event, stdout);
 695         return 0;
 696 }
 697
 698 int machine__process_switch_event(struct machine *machine __maybe_unused,
 699                                   union perf_event *event)
 700 {
 701         if (dump_trace)
 702                 perf_event__fprintf_switch(event, stdout);
 703         return 0;
 704 }
 705
 706 static int machine__process_ksymbol_register(struct machine *machine,
 707                                              union perf_event *event,
 708                                              struct perf_sample *sample __maybe_unused)
 709 {
 710         struct symbol *sym;
 711         struct map *map;
 712
 713         map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
 714         if (!map) {
 715                 map = dso__new_map(event->ksymbol_event.name);
 716                 if (!map)
 717                         return -ENOMEM;
 718
 719                 map->start = event->ksymbol_event.addr;
 720                 map->end = map->start + event->ksymbol_event.len;
 721                 map_groups__insert(&machine->kmaps, map);
 722         }
 723
 724         sym = symbol__new(map->map_ip(map, map->start),
 725                           event->ksymbol_event.len,
 726                           0, 0, event->ksymbol_event.name);
 727         if (!sym)
 728                 return -ENOMEM;
 729         dso__insert_symbol(map->dso, sym);
 730         return 0;
 731 }
 732
 733 static int machine__process_ksymbol_unregister(struct machine *machine,
 734                                                union perf_event *event,
 735                                                struct perf_sample *sample __maybe_unused)
 736 {
 737         struct map *map;
 738
 739         map = map_groups__find(&machine->kmaps, event->ksymbol_event.addr);
 740         if (map)
 741                 map_groups__remove(&machine->kmaps, map);
 742
 743         return 0;
 744 }
 745
 746 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 747                              union perf_event *event,
 748                              struct perf_sample *sample)
 749 {
 750         if (dump_trace)
 751                 perf_event__fprintf_ksymbol(event, stdout);
 752
 753         if (event->ksymbol_event.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 754                 return machine__process_ksymbol_unregister(machine, event,
 755                                                            sample);
 756         return machine__process_ksymbol_register(machine, event, sample);
 757 }
 758
 759 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
 760 {
 761         const char *dup_filename;
 762
 763         if (!filename || !dso || !dso->long_name)
 764                 return;
 765         if (dso->long_name[0] != '[')
 766                 return;
 767         if (!strchr(filename, '/'))
 768                 return;
 769
 770         dup_filename = strdup(filename);
 771         if (!dup_filename)
 772                 return;
 773
 774         dso__set_long_name(dso, dup_filename, true);
 775 }
 776
 777 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 778                                         const char *filename)
 779 {
 780         struct map *map = NULL;
 781         struct dso *dso = NULL;
 782         struct kmod_path m;
 783
 784         if (kmod_path__parse_name(&m, filename))
 785                 return NULL;
 786
 787         map = map_groups__find_by_name(&machine->kmaps, m.name);
 788         if (map) {
 789                 /*
 790                  * If the map's dso is an offline module, give dso__load()
 791                  * a chance to find the file path of that module by fixing
 792                  * long_name.
 793                  */
 794                 dso__adjust_kmod_long_name(map->dso, filename);
 795                 goto out;
 796         }
 797
 798         dso = machine__findnew_module_dso(machine, &m, filename);
 799         if (dso == NULL)
 800                 goto out;
 801
 802         map = map__new2(start, dso);
 803         if (map == NULL)
 804                 goto out;
 805
 806         map_groups__insert(&machine->kmaps, map);
 807
 808         /* Put the map here because map_groups__insert alread got it */
 809         map__put(map);
 810 out:
 811         /* put the dso here, corresponding to  machine__findnew_module_dso */
 812         dso__put(dso);
 813         zfree(&m.name);
 814         return map;
 815 }
 816
 817 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 818 {
 819         struct rb_node *nd;
 820         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 821
 822         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 823                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 824                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 825         }
 826
 827         return ret;
 828 }
 829
 830 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 831                                      bool (skip)(struct dso *dso, int parm), int parm)
 832 {
 833         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 834 }
 835
 836 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 837                                      bool (skip)(struct dso *dso, int parm), int parm)
 838 {
 839         struct rb_node *nd;
 840         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 841
 842         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 843                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 844                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 845         }
 846         return ret;
 847 }
 848
 849 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 850 {
 851         int i;
 852         size_t printed = 0;
 853         struct dso *kdso = machine__kernel_map(machine)->dso;
 854
 855         if (kdso->has_build_id) {
 856                 char filename[PATH_MAX];
 857                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
 858                                            false))
 859                         printed += fprintf(fp, "[0] %s\n", filename);
 860         }
 861
 862         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 863                 printed += fprintf(fp, "[%d] %s\n",
 864                                    i + kdso->has_build_id, vmlinux_path[i]);
 865
 866         return printed;
 867 }
 868
 869 size_t machine__fprintf(struct machine *machine, FILE *fp)
 870 {
 871         struct rb_node *nd;
 872         size_t ret;
 873         int i;
 874
 875         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 876                 struct threads *threads = &machine->threads[i];
 877
 878                 down_read(&threads->lock);
 879
 880                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
 881
 882                 for (nd = rb_first_cached(&threads->entries); nd;
 883                      nd = rb_next(nd)) {
 884                         struct thread *pos = rb_entry(nd, struct thread, rb_node);
 885
 886                         ret += thread__fprintf(pos, fp);
 887                 }
 888
 889                 up_read(&threads->lock);
 890         }
 891         return ret;
 892 }
 893
 894 static struct dso *machine__get_kernel(struct machine *machine)
 895 {
 896         const char *vmlinux_name = machine->mmap_name;
 897         struct dso *kernel;
 898
 899         if (machine__is_host(machine)) {
 900                 if (symbol_conf.vmlinux_name)
 901                         vmlinux_name = symbol_conf.vmlinux_name;
 902
 903                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 904                                                  "[kernel]", DSO_TYPE_KERNEL);
 905         } else {
 906                 if (symbol_conf.default_guest_vmlinux_name)
 907                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 908
 909                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 910                                                  "[guest.kernel]",
 911                                                  DSO_TYPE_GUEST_KERNEL);
 912         }
 913
 914         if (kernel != NULL && (!kernel->has_build_id))
 915                 dso__read_running_kernel_build_id(kernel, machine);
 916
 917         return kernel;
 918 }
 919
 920 struct process_args {
 921         u64 start;
 922 };
 923
 924 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 925                                     size_t bufsz)
 926 {
 927         if (machine__is_default_guest(machine))
 928                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 929         else
 930                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 931 }
 932
 933 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 934
 935 /* Figure out the start address of kernel map from /proc/kallsyms.
 936  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 937  * symbol_name if it's not that important.
 938  */
 939 static int machine__get_running_kernel_start(struct machine *machine,
 940                                              const char **symbol_name,
 941                                              u64 *start, u64 *end)
 942 {
 943         char filename[PATH_MAX];
 944         int i, err = -1;
 945         const char *name;
 946         u64 addr = 0;
 947
 948         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 949
 950         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 951                 return 0;
 952
 953         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 954                 err = kallsyms__get_function_start(filename, name, &addr);
 955                 if (!err)
 956                         break;
 957         }
 958
 959         if (err)
 960                 return -1;
 961
 962         if (symbol_name)
 963                 *symbol_name = name;
 964
 965         *start = addr;
 966
 967         err = kallsyms__get_function_start(filename, "_etext", &addr);
 968         if (!err)
 969                 *end = addr;
 970
 971         return 0;
 972 }
 973
 974 int machine__create_extra_kernel_map(struct machine *machine,
 975                                      struct dso *kernel,
 976                                      struct extra_kernel_map *xm)
 977 {
 978         struct kmap *kmap;
 979         struct map *map;
 980
 981         map = map__new2(xm->start, kernel);
 982         if (!map)
 983                 return -1;
 984
 985         map->end   = xm->end;
 986         map->pgoff = xm->pgoff;
 987
 988         kmap = map__kmap(map);
 989
 990         kmap->kmaps = &machine->kmaps;
 991         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
 992
 993         map_groups__insert(&machine->kmaps, map);
 994
 995         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
 996                   kmap->name, map->start, map->end);
 997
 998         map__put(map);
 999
1000         return 0;
1001 }
1002
1003 static u64 find_entry_trampoline(struct dso *dso)
1004 {
1005         /* Duplicates are removed so lookup all aliases */
1006         const char *syms[] = {
1007                 "_entry_trampoline",
1008                 "__entry_trampoline_start",
1009                 "entry_SYSCALL_64_trampoline",
1010         };
1011         struct symbol *sym = dso__first_symbol(dso);
1012         unsigned int i;
1013
1014         for (; sym; sym = dso__next_symbol(sym)) {
1015                 if (sym->binding != STB_GLOBAL)
1016                         continue;
1017                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1018                         if (!strcmp(sym->name, syms[i]))
1019                                 return sym->start;
1020                 }
1021         }
1022
1023         return 0;
1024 }
1025
1026 /*
1027  * These values can be used for kernels that do not have symbols for the entry
1028  * trampolines in kallsyms.
1029  */
1030 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1031 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1032 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1033
1034 /* Map x86_64 PTI entry trampolines */
1035 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1036                                           struct dso *kernel)
1037 {
1038         struct map_groups *kmaps = &machine->kmaps;
1039         struct maps *maps = &kmaps->maps;
1040         int nr_cpus_avail, cpu;
1041         bool found = false;
1042         struct map *map;
1043         u64 pgoff;
1044
1045         /*
1046          * In the vmlinux case, pgoff is a virtual address which must now be
1047          * mapped to a vmlinux offset.
1048          */
1049         for (map = maps__first(maps); map; map = map__next(map)) {
1050                 struct kmap *kmap = __map__kmap(map);
1051                 struct map *dest_map;
1052
1053                 if (!kmap || !is_entry_trampoline(kmap->name))
1054                         continue;
1055
1056                 dest_map = map_groups__find(kmaps, map->pgoff);
1057                 if (dest_map != map)
1058                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1059                 found = true;
1060         }
1061         if (found || machine->trampolines_mapped)
1062                 return 0;
1063
1064         pgoff = find_entry_trampoline(kernel);
1065         if (!pgoff)
1066                 return 0;
1067
1068         nr_cpus_avail = machine__nr_cpus_avail(machine);
1069
1070         /* Add a 1 page map for each CPU's entry trampoline */
1071         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1072                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1073                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1074                          X86_64_ENTRY_TRAMPOLINE;
1075                 struct extra_kernel_map xm = {
1076                         .start = va,
1077                         .end   = va + page_size,
1078                         .pgoff = pgoff,
1079                 };
1080
1081                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1082
1083                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1084                         return -1;
1085         }
1086
1087         machine->trampolines_mapped = nr_cpus_avail;
1088
1089         return 0;
1090 }
1091
1092 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1093                                              struct dso *kernel __maybe_unused)
1094 {
1095         return 0;
1096 }
1097
1098 static int
1099 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1100 {
1101         struct kmap *kmap;
1102         struct map *map;
1103
1104         /* In case of renewal the kernel map, destroy previous one */
1105         machine__destroy_kernel_maps(machine);
1106
1107         machine->vmlinux_map = map__new2(0, kernel);
1108         if (machine->vmlinux_map == NULL)
1109                 return -1;
1110
1111         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1112         map = machine__kernel_map(machine);
1113         kmap = map__kmap(map);
1114         if (!kmap)
1115                 return -1;
1116
1117         kmap->kmaps = &machine->kmaps;
1118         map_groups__insert(&machine->kmaps, map);
1119
1120         return 0;
1121 }
1122
1123 void machine__destroy_kernel_maps(struct machine *machine)
1124 {
1125         struct kmap *kmap;
1126         struct map *map = machine__kernel_map(machine);
1127
1128         if (map == NULL)
1129                 return;
1130
1131         kmap = map__kmap(map);
1132         map_groups__remove(&machine->kmaps, map);
1133         if (kmap && kmap->ref_reloc_sym) {
1134                 zfree((char **)&kmap->ref_reloc_sym->name);
1135                 zfree(&kmap->ref_reloc_sym);
1136         }
1137
1138         map__zput(machine->vmlinux_map);
1139 }
1140
1141 int machines__create_guest_kernel_maps(struct machines *machines)
1142 {
1143         int ret = 0;
1144         struct dirent **namelist = NULL;
1145         int i, items = 0;
1146         char path[PATH_MAX];
1147         pid_t pid;
1148         char *endp;
1149
1150         if (symbol_conf.default_guest_vmlinux_name ||
1151             symbol_conf.default_guest_modules ||
1152             symbol_conf.default_guest_kallsyms) {
1153                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1154         }
1155
1156         if (symbol_conf.guestmount) {
1157                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1158                 if (items <= 0)
1159                         return -ENOENT;
1160                 for (i = 0; i < items; i++) {
1161                         if (!isdigit(namelist[i]->d_name[0])) {
1162                                 /* Filter out . and .. */
1163                                 continue;
1164                         }
1165                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1166                         if ((*endp != '\0') ||
1167                             (endp == namelist[i]->d_name) ||
1168                             (errno == ERANGE)) {
1169                                 pr_debug("invalid directory (%s). Skipping.\n",
1170                                          namelist[i]->d_name);
1171                                 continue;
1172                         }
1173                         sprintf(path, "%s/%s/proc/kallsyms",
1174                                 symbol_conf.guestmount,
1175                                 namelist[i]->d_name);
1176                         ret = access(path, R_OK);
1177                         if (ret) {
1178                                 pr_debug("Can't access file %s\n", path);
1179                                 goto failure;
1180                         }
1181                         machines__create_kernel_maps(machines, pid);
1182                 }
1183 failure:
1184                 free(namelist);
1185         }
1186
1187         return ret;
1188 }
1189
1190 void machines__destroy_kernel_maps(struct machines *machines)
1191 {
1192         struct rb_node *next = rb_first_cached(&machines->guests);
1193
1194         machine__destroy_kernel_maps(&machines->host);
1195
1196         while (next) {
1197                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1198
1199                 next = rb_next(&pos->rb_node);
1200                 rb_erase_cached(&pos->rb_node, &machines->guests);
1201                 machine__delete(pos);
1202         }
1203 }
1204
1205 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1206 {
1207         struct machine *machine = machines__findnew(machines, pid);
1208
1209         if (machine == NULL)
1210                 return -1;
1211
1212         return machine__create_kernel_maps(machine);
1213 }
1214
1215 int machine__load_kallsyms(struct machine *machine, const char *filename)
1216 {
1217         struct map *map = machine__kernel_map(machine);
1218         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1219
1220         if (ret > 0) {
1221                 dso__set_loaded(map->dso);
1222                 /*
1223                  * Since /proc/kallsyms will have multiple sessions for the
1224                  * kernel, with modules between them, fixup the end of all
1225                  * sections.
1226                  */
1227                 map_groups__fixup_end(&machine->kmaps);
1228         }
1229
1230         return ret;
1231 }
1232
1233 int machine__load_vmlinux_path(struct machine *machine)
1234 {
1235         struct map *map = machine__kernel_map(machine);
1236         int ret = dso__load_vmlinux_path(map->dso, map);
1237
1238         if (ret > 0)
1239                 dso__set_loaded(map->dso);
1240
1241         return ret;
1242 }
1243
1244 static char *get_kernel_version(const char *root_dir)
1245 {
1246         char version[PATH_MAX];
1247         FILE *file;
1248         char *name, *tmp;
1249         const char *prefix = "Linux version ";
1250
1251         sprintf(version, "%s/proc/version", root_dir);
1252         file = fopen(version, "r");
1253         if (!file)
1254                 return NULL;
1255
1256         tmp = fgets(version, sizeof(version), file);
1257         fclose(file);
1258         if (!tmp)
1259                 return NULL;
1260
1261         name = strstr(version, prefix);
1262         if (!name)
1263                 return NULL;
1264         name += strlen(prefix);
1265         tmp = strchr(name, ' ');
1266         if (tmp)
1267                 *tmp = '\0';
1268
1269         return strdup(name);
1270 }
1271
1272 static bool is_kmod_dso(struct dso *dso)
1273 {
1274         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1275                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1276 }
1277
1278 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1279                                        struct kmod_path *m)
1280 {
1281         char *long_name;
1282         struct map *map = map_groups__find_by_name(mg, m->name);
1283
1284         if (map == NULL)
1285                 return 0;
1286
1287         long_name = strdup(path);
1288         if (long_name == NULL)
1289                 return -ENOMEM;
1290
1291         dso__set_long_name(map->dso, long_name, true);
1292         dso__kernel_module_get_build_id(map->dso, "");
1293
1294         /*
1295          * Full name could reveal us kmod compression, so
1296          * we need to update the symtab_type if needed.
1297          */
1298         if (m->comp && is_kmod_dso(map->dso)) {
1299                 map->dso->symtab_type++;
1300                 map->dso->comp = m->comp;
1301         }
1302
1303         return 0;
1304 }
1305
1306 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1307                                 const char *dir_name, int depth)
1308 {
1309         struct dirent *dent;
1310         DIR *dir = opendir(dir_name);
1311         int ret = 0;
1312
1313         if (!dir) {
1314                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1315                 return -1;
1316         }
1317
1318         while ((dent = readdir(dir)) != NULL) {
1319                 char path[PATH_MAX];
1320                 struct stat st;
1321
1322                 /*sshfs might return bad dent->d_type, so we have to stat*/
1323                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1324                 if (stat(path, &st))
1325                         continue;
1326
1327                 if (S_ISDIR(st.st_mode)) {
1328                         if (!strcmp(dent->d_name, ".") ||
1329                             !strcmp(dent->d_name, ".."))
1330                                 continue;
1331
1332                         /* Do not follow top-level source and build symlinks */
1333                         if (depth == 0) {
1334                                 if (!strcmp(dent->d_name, "source") ||
1335                                     !strcmp(dent->d_name, "build"))
1336                                         continue;
1337                         }
1338
1339                         ret = map_groups__set_modules_path_dir(mg, path,
1340                                                                depth + 1);
1341                         if (ret < 0)
1342                                 goto out;
1343                 } else {
1344                         struct kmod_path m;
1345
1346                         ret = kmod_path__parse_name(&m, dent->d_name);
1347                         if (ret)
1348                                 goto out;
1349
1350                         if (m.kmod)
1351                                 ret = map_groups__set_module_path(mg, path, &m);
1352
1353                         zfree(&m.name);
1354
1355                         if (ret)
1356                                 goto out;
1357                 }
1358         }
1359
1360 out:
1361         closedir(dir);
1362         return ret;
1363 }
1364
1365 static int machine__set_modules_path(struct machine *machine)
1366 {
1367         char *version;
1368         char modules_path[PATH_MAX];
1369
1370         version = get_kernel_version(machine->root_dir);
1371         if (!version)
1372                 return -1;
1373
1374         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1375                  machine->root_dir, version);
1376         free(version);
1377
1378         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1379 }
1380 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1381                                 u64 *size __maybe_unused,
1382                                 const char *name __maybe_unused)
1383 {
1384         return 0;
1385 }
1386
1387 static int machine__create_module(void *arg, const char *name, u64 start,
1388                                   u64 size)
1389 {
1390         struct machine *machine = arg;
1391         struct map *map;
1392
1393         if (arch__fix_module_text_start(&start, &size, name) < 0)
1394                 return -1;
1395
1396         map = machine__findnew_module_map(machine, start, name);
1397         if (map == NULL)
1398                 return -1;
1399         map->end = start + size;
1400
1401         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1402
1403         return 0;
1404 }
1405
1406 static int machine__create_modules(struct machine *machine)
1407 {
1408         const char *modules;
1409         char path[PATH_MAX];
1410
1411         if (machine__is_default_guest(machine)) {
1412                 modules = symbol_conf.default_guest_modules;
1413         } else {
1414                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1415                 modules = path;
1416         }
1417
1418         if (symbol__restricted_filename(modules, "/proc/modules"))
1419                 return -1;
1420
1421         if (modules__parse(modules, machine, machine__create_module))
1422                 return -1;
1423
1424         if (!machine__set_modules_path(machine))
1425                 return 0;
1426
1427         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1428
1429         return 0;
1430 }
1431
1432 static void machine__set_kernel_mmap(struct machine *machine,
1433                                      u64 start, u64 end)
1434 {
1435         machine->vmlinux_map->start = start;
1436         machine->vmlinux_map->end   = end;
1437         /*
1438          * Be a bit paranoid here, some perf.data file came with
1439          * a zero sized synthesized MMAP event for the kernel.
1440          */
1441         if (start == 0 && end == 0)
1442                 machine->vmlinux_map->end = ~0ULL;
1443 }
1444
1445 static void machine__update_kernel_mmap(struct machine *machine,
1446                                      u64 start, u64 end)
1447 {
1448         struct map *map = machine__kernel_map(machine);
1449
1450         map__get(map);
1451         map_groups__remove(&machine->kmaps, map);
1452
1453         machine__set_kernel_mmap(machine, start, end);
1454
1455         map_groups__insert(&machine->kmaps, map);
1456         map__put(map);
1457 }
1458
1459 int machine__create_kernel_maps(struct machine *machine)
1460 {
1461         struct dso *kernel = machine__get_kernel(machine);
1462         const char *name = NULL;
1463         struct map *map;
1464         u64 start = 0, end = ~0ULL;
1465         int ret;
1466
1467         if (kernel == NULL)
1468                 return -1;
1469
1470         ret = __machine__create_kernel_maps(machine, kernel);
1471         if (ret < 0)
1472                 goto out_put;
1473
1474         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1475                 if (machine__is_host(machine))
1476                         pr_debug("Problems creating module maps, "
1477                                  "continuing anyway...\n");
1478                 else
1479                         pr_debug("Problems creating module maps for guest %d, "
1480                                  "continuing anyway...\n", machine->pid);
1481         }
1482
1483         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1484                 if (name &&
1485                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1486                         machine__destroy_kernel_maps(machine);
1487                         ret = -1;
1488                         goto out_put;
1489                 }
1490
1491                 /*
1492                  * we have a real start address now, so re-order the kmaps
1493                  * assume it's the last in the kmaps
1494                  */
1495                 machine__update_kernel_mmap(machine, start, end);
1496         }
1497
1498         if (machine__create_extra_kernel_maps(machine, kernel))
1499                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1500
1501         if (end == ~0ULL) {
1502                 /* update end address of the kernel map using adjacent module address */
1503                 map = map__next(machine__kernel_map(machine));
1504                 if (map)
1505                         machine__set_kernel_mmap(machine, start, map->start);
1506         }
1507
1508 out_put:
1509         dso__put(kernel);
1510         return ret;
1511 }
1512
1513 static bool machine__uses_kcore(struct machine *machine)
1514 {
1515         struct dso *dso;
1516
1517         list_for_each_entry(dso, &machine->dsos.head, node) {
1518                 if (dso__is_kcore(dso))
1519                         return true;
1520         }
1521
1522         return false;
1523 }
1524
1525 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1526                                              union perf_event *event)
1527 {
1528         return machine__is(machine, "x86_64") &&
1529                is_entry_trampoline(event->mmap.filename);
1530 }
1531
1532 static int machine__process_extra_kernel_map(struct machine *machine,
1533                                              union perf_event *event)
1534 {
1535         struct map *kernel_map = machine__kernel_map(machine);
1536         struct dso *kernel = kernel_map ? kernel_map->dso : NULL;
1537         struct extra_kernel_map xm = {
1538                 .start = event->mmap.start,
1539                 .end   = event->mmap.start + event->mmap.len,
1540                 .pgoff = event->mmap.pgoff,
1541         };
1542
1543         if (kernel == NULL)
1544                 return -1;
1545
1546         strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1547
1548         return machine__create_extra_kernel_map(machine, kernel, &xm);
1549 }
1550
1551 static int machine__process_kernel_mmap_event(struct machine *machine,
1552                                               union perf_event *event)
1553 {
1554         struct map *map;
1555         enum dso_kernel_type kernel_type;
1556         bool is_kernel_mmap;
1557
1558         /* If we have maps from kcore then we do not need or want any others */
1559         if (machine__uses_kcore(machine))
1560                 return 0;
1561
1562         if (machine__is_host(machine))
1563                 kernel_type = DSO_TYPE_KERNEL;
1564         else
1565                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1566
1567         is_kernel_mmap = memcmp(event->mmap.filename,
1568                                 machine->mmap_name,
1569                                 strlen(machine->mmap_name) - 1) == 0;
1570         if (event->mmap.filename[0] == '/' ||
1571             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1572                 map = machine__findnew_module_map(machine, event->mmap.start,
1573                                                   event->mmap.filename);
1574                 if (map == NULL)
1575                         goto out_problem;
1576
1577                 map->end = map->start + event->mmap.len;
1578         } else if (is_kernel_mmap) {
1579                 const char *symbol_name = (event->mmap.filename +
1580                                 strlen(machine->mmap_name));
1581                 /*
1582                  * Should be there already, from the build-id table in
1583                  * the header.
1584                  */
1585                 struct dso *kernel = NULL;
1586                 struct dso *dso;
1587
1588                 down_read(&machine->dsos.lock);
1589
1590                 list_for_each_entry(dso, &machine->dsos.head, node) {
1591
1592                         /*
1593                          * The cpumode passed to is_kernel_module is not the
1594                          * cpumode of *this* event. If we insist on passing
1595                          * correct cpumode to is_kernel_module, we should
1596                          * record the cpumode when we adding this dso to the
1597                          * linked list.
1598                          *
1599                          * However we don't really need passing correct
1600                          * cpumode.  We know the correct cpumode must be kernel
1601                          * mode (if not, we should not link it onto kernel_dsos
1602                          * list).
1603                          *
1604                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1605                          * is_kernel_module() treats it as a kernel cpumode.
1606                          */
1607
1608                         if (!dso->kernel ||
1609                             is_kernel_module(dso->long_name,
1610                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1611                                 continue;
1612
1613
1614                         kernel = dso;
1615                         break;
1616                 }
1617
1618                 up_read(&machine->dsos.lock);
1619
1620                 if (kernel == NULL)
1621                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1622                 if (kernel == NULL)
1623                         goto out_problem;
1624
1625                 kernel->kernel = kernel_type;
1626                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1627                         dso__put(kernel);
1628                         goto out_problem;
1629                 }
1630
1631                 if (strstr(kernel->long_name, "vmlinux"))
1632                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1633
1634                 machine__update_kernel_mmap(machine, event->mmap.start,
1635                                          event->mmap.start + event->mmap.len);
1636
1637                 /*
1638                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1639                  * symbol. Effectively having zero here means that at record
1640                  * time /proc/sys/kernel/kptr_restrict was non zero.
1641                  */
1642                 if (event->mmap.pgoff != 0) {
1643                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1644                                                         symbol_name,
1645                                                         event->mmap.pgoff);
1646                 }
1647
1648                 if (machine__is_default_guest(machine)) {
1649                         /*
1650                          * preload dso of guest kernel and modules
1651                          */
1652                         dso__load(kernel, machine__kernel_map(machine));
1653                 }
1654         } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1655                 return machine__process_extra_kernel_map(machine, event);
1656         }
1657         return 0;
1658 out_problem:
1659         return -1;
1660 }
1661
1662 int machine__process_mmap2_event(struct machine *machine,
1663                                  union perf_event *event,
1664                                  struct perf_sample *sample)
1665 {
1666         struct thread *thread;
1667         struct map *map;
1668         int ret = 0;
1669
1670         if (dump_trace)
1671                 perf_event__fprintf_mmap2(event, stdout);
1672
1673         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1674             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1675                 ret = machine__process_kernel_mmap_event(machine, event);
1676                 if (ret < 0)
1677                         goto out_problem;
1678                 return 0;
1679         }
1680
1681         thread = machine__findnew_thread(machine, event->mmap2.pid,
1682                                         event->mmap2.tid);
1683         if (thread == NULL)
1684                 goto out_problem;
1685
1686         map = map__new(machine, event->mmap2.start,
1687                         event->mmap2.len, event->mmap2.pgoff,
1688                         event->mmap2.maj,
1689                         event->mmap2.min, event->mmap2.ino,
1690                         event->mmap2.ino_generation,
1691                         event->mmap2.prot,
1692                         event->mmap2.flags,
1693                         event->mmap2.filename, thread);
1694
1695         if (map == NULL)
1696                 goto out_problem_map;
1697
1698         ret = thread__insert_map(thread, map);
1699         if (ret)
1700                 goto out_problem_insert;
1701
1702         thread__put(thread);
1703         map__put(map);
1704         return 0;
1705
1706 out_problem_insert:
1707         map__put(map);
1708 out_problem_map:
1709         thread__put(thread);
1710 out_problem:
1711         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1712         return 0;
1713 }
1714
1715 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1716                                 struct perf_sample *sample)
1717 {
1718         struct thread *thread;
1719         struct map *map;
1720         u32 prot = 0;
1721         int ret = 0;
1722
1723         if (dump_trace)
1724                 perf_event__fprintf_mmap(event, stdout);
1725
1726         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1727             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1728                 ret = machine__process_kernel_mmap_event(machine, event);
1729                 if (ret < 0)
1730                         goto out_problem;
1731                 return 0;
1732         }
1733
1734         thread = machine__findnew_thread(machine, event->mmap.pid,
1735                                          event->mmap.tid);
1736         if (thread == NULL)
1737                 goto out_problem;
1738
1739         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1740                 prot = PROT_EXEC;
1741
1742         map = map__new(machine, event->mmap.start,
1743                         event->mmap.len, event->mmap.pgoff,
1744                         0, 0, 0, 0, prot, 0,
1745                         event->mmap.filename,
1746                         thread);
1747
1748         if (map == NULL)
1749                 goto out_problem_map;
1750
1751         ret = thread__insert_map(thread, map);
1752         if (ret)
1753                 goto out_problem_insert;
1754
1755         thread__put(thread);
1756         map__put(map);
1757         return 0;
1758
1759 out_problem_insert:
1760         map__put(map);
1761 out_problem_map:
1762         thread__put(thread);
1763 out_problem:
1764         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1765         return 0;
1766 }
1767
1768 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1769 {
1770         struct threads *threads = machine__threads(machine, th->tid);
1771
1772         if (threads->last_match == th)
1773                 threads__set_last_match(threads, NULL);
1774
1775         if (lock)
1776                 down_write(&threads->lock);
1777
1778         BUG_ON(refcount_read(&th->refcnt) == 0);
1779
1780         rb_erase_cached(&th->rb_node, &threads->entries);
1781         RB_CLEAR_NODE(&th->rb_node);
1782         --threads->nr;
1783         /*
1784          * Move it first to the dead_threads list, then drop the reference,
1785          * if this is the last reference, then the thread__delete destructor
1786          * will be called and we will remove it from the dead_threads list.
1787          */
1788         list_add_tail(&th->node, &threads->dead);
1789
1790         /*
1791          * We need to do the put here because if this is the last refcount,
1792          * then we will be touching the threads->dead head when removing the
1793          * thread.
1794          */
1795         thread__put(th);
1796
1797         if (lock)
1798                 up_write(&threads->lock);
1799 }
1800
1801 void machine__remove_thread(struct machine *machine, struct thread *th)
1802 {
1803         return __machine__remove_thread(machine, th, true);
1804 }
1805
1806 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1807                                 struct perf_sample *sample)
1808 {
1809         struct thread *thread = machine__find_thread(machine,
1810                                                      event->fork.pid,
1811                                                      event->fork.tid);
1812         struct thread *parent = machine__findnew_thread(machine,
1813                                                         event->fork.ppid,
1814                                                         event->fork.ptid);
1815         bool do_maps_clone = true;
1816         int err = 0;
1817
1818         if (dump_trace)
1819                 perf_event__fprintf_task(event, stdout);
1820
1821         /*
1822          * There may be an existing thread that is not actually the parent,
1823          * either because we are processing events out of order, or because the
1824          * (fork) event that would have removed the thread was lost. Assume the
1825          * latter case and continue on as best we can.
1826          */
1827         if (parent->pid_ != (pid_t)event->fork.ppid) {
1828                 dump_printf("removing erroneous parent thread %d/%d\n",
1829                             parent->pid_, parent->tid);
1830                 machine__remove_thread(machine, parent);
1831                 thread__put(parent);
1832                 parent = machine__findnew_thread(machine, event->fork.ppid,
1833                                                  event->fork.ptid);
1834         }
1835
1836         /* if a thread currently exists for the thread id remove it */
1837         if (thread != NULL) {
1838                 machine__remove_thread(machine, thread);
1839                 thread__put(thread);
1840         }
1841
1842         thread = machine__findnew_thread(machine, event->fork.pid,
1843                                          event->fork.tid);
1844         /*
1845          * When synthesizing FORK events, we are trying to create thread
1846          * objects for the already running tasks on the machine.
1847          *
1848          * Normally, for a kernel FORK event, we want to clone the parent's
1849          * maps because that is what the kernel just did.
1850          *
1851          * But when synthesizing, this should not be done.  If we do, we end up
1852          * with overlapping maps as we process the sythesized MMAP2 events that
1853          * get delivered shortly thereafter.
1854          *
1855          * Use the FORK event misc flags in an internal way to signal this
1856          * situation, so we can elide the map clone when appropriate.
1857          */
1858         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1859                 do_maps_clone = false;
1860
1861         if (thread == NULL || parent == NULL ||
1862             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1863                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1864                 err = -1;
1865         }
1866         thread__put(thread);
1867         thread__put(parent);
1868
1869         return err;
1870 }
1871
1872 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1873                                 struct perf_sample *sample __maybe_unused)
1874 {
1875         struct thread *thread = machine__find_thread(machine,
1876                                                      event->fork.pid,
1877                                                      event->fork.tid);
1878
1879         if (dump_trace)
1880                 perf_event__fprintf_task(event, stdout);
1881
1882         if (thread != NULL) {
1883                 thread__exited(thread);
1884                 thread__put(thread);
1885         }
1886
1887         return 0;
1888 }
1889
1890 int machine__process_event(struct machine *machine, union perf_event *event,
1891                            struct perf_sample *sample)
1892 {
1893         int ret;
1894
1895         switch (event->header.type) {
1896         case PERF_RECORD_COMM:
1897                 ret = machine__process_comm_event(machine, event, sample); break;
1898         case PERF_RECORD_MMAP:
1899                 ret = machine__process_mmap_event(machine, event, sample); break;
1900         case PERF_RECORD_NAMESPACES:
1901                 ret = machine__process_namespaces_event(machine, event, sample); break;
1902         case PERF_RECORD_MMAP2:
1903                 ret = machine__process_mmap2_event(machine, event, sample); break;
1904         case PERF_RECORD_FORK:
1905                 ret = machine__process_fork_event(machine, event, sample); break;
1906         case PERF_RECORD_EXIT:
1907                 ret = machine__process_exit_event(machine, event, sample); break;
1908         case PERF_RECORD_LOST:
1909                 ret = machine__process_lost_event(machine, event, sample); break;
1910         case PERF_RECORD_AUX:
1911                 ret = machine__process_aux_event(machine, event); break;
1912         case PERF_RECORD_ITRACE_START:
1913                 ret = machine__process_itrace_start_event(machine, event); break;
1914         case PERF_RECORD_LOST_SAMPLES:
1915                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1916         case PERF_RECORD_SWITCH:
1917         case PERF_RECORD_SWITCH_CPU_WIDE:
1918                 ret = machine__process_switch_event(machine, event); break;
1919         case PERF_RECORD_KSYMBOL:
1920                 ret = machine__process_ksymbol(machine, event, sample); break;
1921         case PERF_RECORD_BPF_EVENT:
1922                 ret = machine__process_bpf_event(machine, event, sample); break;
1923         default:
1924                 ret = -1;
1925                 break;
1926         }
1927
1928         return ret;
1929 }
1930
1931 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1932 {
1933         if (!regexec(regex, sym->name, 0, NULL, 0))
1934                 return 1;
1935         return 0;
1936 }
1937
1938 static void ip__resolve_ams(struct thread *thread,
1939                             struct addr_map_symbol *ams,
1940                             u64 ip)
1941 {
1942         struct addr_location al;
1943
1944         memset(&al, 0, sizeof(al));
1945         /*
1946          * We cannot use the header.misc hint to determine whether a
1947          * branch stack address is user, kernel, guest, hypervisor.
1948          * Branches may straddle the kernel/user/hypervisor boundaries.
1949          * Thus, we have to try consecutively until we find a match
1950          * or else, the symbol is unknown
1951          */
1952         thread__find_cpumode_addr_location(thread, ip, &al);
1953
1954         ams->addr = ip;
1955         ams->al_addr = al.addr;
1956         ams->sym = al.sym;
1957         ams->map = al.map;
1958         ams->phys_addr = 0;
1959 }
1960
1961 static void ip__resolve_data(struct thread *thread,
1962                              u8 m, struct addr_map_symbol *ams,
1963                              u64 addr, u64 phys_addr)
1964 {
1965         struct addr_location al;
1966
1967         memset(&al, 0, sizeof(al));
1968
1969         thread__find_symbol(thread, m, addr, &al);
1970
1971         ams->addr = addr;
1972         ams->al_addr = al.addr;
1973         ams->sym = al.sym;
1974         ams->map = al.map;
1975         ams->phys_addr = phys_addr;
1976 }
1977
1978 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1979                                      struct addr_location *al)
1980 {
1981         struct mem_info *mi = mem_info__new();
1982
1983         if (!mi)
1984                 return NULL;
1985
1986         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1987         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
1988                          sample->addr, sample->phys_addr);
1989         mi->data_src.val = sample->data_src;
1990
1991         return mi;
1992 }
1993
1994 static char *callchain_srcline(struct map *map, struct symbol *sym, u64 ip)
1995 {
1996         char *srcline = NULL;
1997
1998         if (!map || callchain_param.key == CCKEY_FUNCTION)
1999                 return srcline;
2000
2001         srcline = srcline__tree_find(&map->dso->srclines, ip);
2002         if (!srcline) {
2003                 bool show_sym = false;
2004                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2005
2006                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2007                                       sym, show_sym, show_addr, ip);
2008                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2009         }
2010
2011         return srcline;
2012 }
2013
2014 struct iterations {
2015         int nr_loop_iter;
2016         u64 cycles;
2017 };
2018
2019 static int add_callchain_ip(struct thread *thread,
2020                             struct callchain_cursor *cursor,
2021                             struct symbol **parent,
2022                             struct addr_location *root_al,
2023                             u8 *cpumode,
2024                             u64 ip,
2025                             bool branch,
2026                             struct branch_flags *flags,
2027                             struct iterations *iter,
2028                             u64 branch_from)
2029 {
2030         struct addr_location al;
2031         int nr_loop_iter = 0;
2032         u64 iter_cycles = 0;
2033         const char *srcline = NULL;
2034
2035         al.filtered = 0;
2036         al.sym = NULL;
2037         if (!cpumode) {
2038                 thread__find_cpumode_addr_location(thread, ip, &al);
2039         } else {
2040                 if (ip >= PERF_CONTEXT_MAX) {
2041                         switch (ip) {
2042                         case PERF_CONTEXT_HV:
2043                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2044                                 break;
2045                         case PERF_CONTEXT_KERNEL:
2046                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2047                                 break;
2048                         case PERF_CONTEXT_USER:
2049                                 *cpumode = PERF_RECORD_MISC_USER;
2050                                 break;
2051                         default:
2052                                 pr_debug("invalid callchain context: "
2053                                          "%"PRId64"\n", (s64) ip);
2054                                 /*
2055                                  * It seems the callchain is corrupted.
2056                                  * Discard all.
2057                                  */
2058                                 callchain_cursor_reset(cursor);
2059                                 return 1;
2060                         }
2061                         return 0;
2062                 }
2063                 thread__find_symbol(thread, *cpumode, ip, &al);
2064         }
2065
2066         if (al.sym != NULL) {
2067                 if (perf_hpp_list.parent && !*parent &&
2068                     symbol__match_regex(al.sym, &parent_regex))
2069                         *parent = al.sym;
2070                 else if (have_ignore_callees && root_al &&
2071                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2072                         /* Treat this symbol as the root,
2073                            forgetting its callees. */
2074                         *root_al = al;
2075                         callchain_cursor_reset(cursor);
2076                 }
2077         }
2078
2079         if (symbol_conf.hide_unresolved && al.sym == NULL)
2080                 return 0;
2081
2082         if (iter) {
2083                 nr_loop_iter = iter->nr_loop_iter;
2084                 iter_cycles = iter->cycles;
2085         }
2086
2087         srcline = callchain_srcline(al.map, al.sym, al.addr);
2088         return callchain_cursor_append(cursor, ip, al.map, al.sym,
2089                                        branch, flags, nr_loop_iter,
2090                                        iter_cycles, branch_from, srcline);
2091 }
2092
2093 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2094                                            struct addr_location *al)
2095 {
2096         unsigned int i;
2097         const struct branch_stack *bs = sample->branch_stack;
2098         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2099
2100         if (!bi)
2101                 return NULL;
2102
2103         for (i = 0; i < bs->nr; i++) {
2104                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2105                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2106                 bi[i].flags = bs->entries[i].flags;
2107         }
2108         return bi;
2109 }
2110
2111 static void save_iterations(struct iterations *iter,
2112                             struct branch_entry *be, int nr)
2113 {
2114         int i;
2115
2116         iter->nr_loop_iter++;
2117         iter->cycles = 0;
2118
2119         for (i = 0; i < nr; i++)
2120                 iter->cycles += be[i].flags.cycles;
2121 }
2122
2123 #define CHASHSZ 127
2124 #define CHASHBITS 7
2125 #define NO_ENTRY 0xff
2126
2127 #define PERF_MAX_BRANCH_DEPTH 127
2128
2129 /* Remove loops. */
2130 static int remove_loops(struct branch_entry *l, int nr,
2131                         struct iterations *iter)
2132 {
2133         int i, j, off;
2134         unsigned char chash[CHASHSZ];
2135
2136         memset(chash, NO_ENTRY, sizeof(chash));
2137
2138         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2139
2140         for (i = 0; i < nr; i++) {
2141                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2142
2143                 /* no collision handling for now */
2144                 if (chash[h] == NO_ENTRY) {
2145                         chash[h] = i;
2146                 } else if (l[chash[h]].from == l[i].from) {
2147                         bool is_loop = true;
2148                         /* check if it is a real loop */
2149                         off = 0;
2150                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2151                                 if (l[j].from != l[i + off].from) {
2152                                         is_loop = false;
2153                                         break;
2154                                 }
2155                         if (is_loop) {
2156                                 j = nr - (i + off);
2157                                 if (j > 0) {
2158                                         save_iterations(iter + i + off,
2159                                                 l + i, off);
2160
2161                                         memmove(iter + i, iter + i + off,
2162                                                 j * sizeof(*iter));
2163
2164                                         memmove(l + i, l + i + off,
2165                                                 j * sizeof(*l));
2166                                 }
2167
2168                                 nr -= off;
2169                         }
2170                 }
2171         }
2172         return nr;
2173 }
2174
2175 /*
2176  * Recolve LBR callstack chain sample
2177  * Return:
2178  * 1 on success get LBR callchain information
2179  * 0 no available LBR callchain information, should try fp
2180  * negative error code on other errors.
2181  */
2182 static int resolve_lbr_callchain_sample(struct thread *thread,
2183                                         struct callchain_cursor *cursor,
2184                                         struct perf_sample *sample,
2185                                         struct symbol **parent,
2186                                         struct addr_location *root_al,
2187                                         int max_stack)
2188 {
2189         struct ip_callchain *chain = sample->callchain;
2190         int chain_nr = min(max_stack, (int)chain->nr), i;
2191         u8 cpumode = PERF_RECORD_MISC_USER;
2192         u64 ip, branch_from = 0;
2193
2194         for (i = 0; i < chain_nr; i++) {
2195                 if (chain->ips[i] == PERF_CONTEXT_USER)
2196                         break;
2197         }
2198
2199         /* LBR only affects the user callchain */
2200         if (i != chain_nr) {
2201                 struct branch_stack *lbr_stack = sample->branch_stack;
2202                 int lbr_nr = lbr_stack->nr, j, k;
2203                 bool branch;
2204                 struct branch_flags *flags;
2205                 /*
2206                  * LBR callstack can only get user call chain.
2207                  * The mix_chain_nr is kernel call chain
2208                  * number plus LBR user call chain number.
2209                  * i is kernel call chain number,
2210                  * 1 is PERF_CONTEXT_USER,
2211                  * lbr_nr + 1 is the user call chain number.
2212                  * For details, please refer to the comments
2213                  * in callchain__printf
2214                  */
2215                 int mix_chain_nr = i + 1 + lbr_nr + 1;
2216
2217                 for (j = 0; j < mix_chain_nr; j++) {
2218                         int err;
2219                         branch = false;
2220                         flags = NULL;
2221
2222                         if (callchain_param.order == ORDER_CALLEE) {
2223                                 if (j < i + 1)
2224                                         ip = chain->ips[j];
2225                                 else if (j > i + 1) {
2226                                         k = j - i - 2;
2227                                         ip = lbr_stack->entries[k].from;
2228                                         branch = true;
2229                                         flags = &lbr_stack->entries[k].flags;
2230                                 } else {
2231                                         ip = lbr_stack->entries[0].to;
2232                                         branch = true;
2233                                         flags = &lbr_stack->entries[0].flags;
2234                                         branch_from =
2235                                                 lbr_stack->entries[0].from;
2236                                 }
2237                         } else {
2238                                 if (j < lbr_nr) {
2239                                         k = lbr_nr - j - 1;
2240                                         ip = lbr_stack->entries[k].from;
2241                                         branch = true;
2242                                         flags = &lbr_stack->entries[k].flags;
2243                                 }
2244                                 else if (j > lbr_nr)
2245                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
2246                                 else {
2247                                         ip = lbr_stack->entries[0].to;
2248                                         branch = true;
2249                                         flags = &lbr_stack->entries[0].flags;
2250                                         branch_from =
2251                                                 lbr_stack->entries[0].from;
2252                                 }
2253                         }
2254
2255                         err = add_callchain_ip(thread, cursor, parent,
2256                                                root_al, &cpumode, ip,
2257                                                branch, flags, NULL,
2258                                                branch_from);
2259                         if (err)
2260                                 return (err < 0) ? err : 0;
2261                 }
2262                 return 1;
2263         }
2264
2265         return 0;
2266 }
2267
2268 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2269                              struct callchain_cursor *cursor,
2270                              struct symbol **parent,
2271                              struct addr_location *root_al,
2272                              u8 *cpumode, int ent)
2273 {
2274         int err = 0;
2275
2276         while (--ent >= 0) {
2277                 u64 ip = chain->ips[ent];
2278
2279                 if (ip >= PERF_CONTEXT_MAX) {
2280                         err = add_callchain_ip(thread, cursor, parent,
2281                                                root_al, cpumode, ip,
2282                                                false, NULL, NULL, 0);
2283                         break;
2284                 }
2285         }
2286         return err;
2287 }
2288
2289 static int thread__resolve_callchain_sample(struct thread *thread,
2290                                             struct callchain_cursor *cursor,
2291                                             struct perf_evsel *evsel,
2292                                             struct perf_sample *sample,
2293                                             struct symbol **parent,
2294                                             struct addr_location *root_al,
2295                                             int max_stack)
2296 {
2297         struct branch_stack *branch = sample->branch_stack;
2298         struct ip_callchain *chain = sample->callchain;
2299         int chain_nr = 0;
2300         u8 cpumode = PERF_RECORD_MISC_USER;
2301         int i, j, err, nr_entries;
2302         int skip_idx = -1;
2303         int first_call = 0;
2304
2305         if (chain)
2306                 chain_nr = chain->nr;
2307
2308         if (perf_evsel__has_branch_callstack(evsel)) {
2309                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2310                                                    root_al, max_stack);
2311                 if (err)
2312                         return (err < 0) ? err : 0;
2313         }
2314
2315         /*
2316          * Based on DWARF debug information, some architectures skip
2317          * a callchain entry saved by the kernel.
2318          */
2319         skip_idx = arch_skip_callchain_idx(thread, chain);
2320
2321         /*
2322          * Add branches to call stack for easier browsing. This gives
2323          * more context for a sample than just the callers.
2324          *
2325          * This uses individual histograms of paths compared to the
2326          * aggregated histograms the normal LBR mode uses.
2327          *
2328          * Limitations for now:
2329          * - No extra filters
2330          * - No annotations (should annotate somehow)
2331          */
2332
2333         if (branch && callchain_param.branch_callstack) {
2334                 int nr = min(max_stack, (int)branch->nr);
2335                 struct branch_entry be[nr];
2336                 struct iterations iter[nr];
2337
2338                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2339                         pr_warning("corrupted branch chain. skipping...\n");
2340                         goto check_calls;
2341                 }
2342
2343                 for (i = 0; i < nr; i++) {
2344                         if (callchain_param.order == ORDER_CALLEE) {
2345                                 be[i] = branch->entries[i];
2346
2347                                 if (chain == NULL)
2348                                         continue;
2349
2350                                 /*
2351                                  * Check for overlap into the callchain.
2352                                  * The return address is one off compared to
2353                                  * the branch entry. To adjust for this
2354                                  * assume the calling instruction is not longer
2355                                  * than 8 bytes.
2356                                  */
2357                                 if (i == skip_idx ||
2358                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2359                                         first_call++;
2360                                 else if (be[i].from < chain->ips[first_call] &&
2361                                     be[i].from >= chain->ips[first_call] - 8)
2362                                         first_call++;
2363                         } else
2364                                 be[i] = branch->entries[branch->nr - i - 1];
2365                 }
2366
2367                 memset(iter, 0, sizeof(struct iterations) * nr);
2368                 nr = remove_loops(be, nr, iter);
2369
2370                 for (i = 0; i < nr; i++) {
2371                         err = add_callchain_ip(thread, cursor, parent,
2372                                                root_al,
2373                                                NULL, be[i].to,
2374                                                true, &be[i].flags,
2375                                                NULL, be[i].from);
2376
2377                         if (!err)
2378                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2379                                                        NULL, be[i].from,
2380                                                        true, &be[i].flags,
2381                                                        &iter[i], 0);
2382                         if (err == -EINVAL)
2383                                 break;
2384                         if (err)
2385                                 return err;
2386                 }
2387
2388                 if (chain_nr == 0)
2389                         return 0;
2390
2391                 chain_nr -= nr;
2392         }
2393
2394 check_calls:
2395         if (callchain_param.order != ORDER_CALLEE) {
2396                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2397                                         &cpumode, chain->nr - first_call);
2398                 if (err)
2399                         return (err < 0) ? err : 0;
2400         }
2401         for (i = first_call, nr_entries = 0;
2402              i < chain_nr && nr_entries < max_stack; i++) {
2403                 u64 ip;
2404
2405                 if (callchain_param.order == ORDER_CALLEE)
2406                         j = i;
2407                 else
2408                         j = chain->nr - i - 1;
2409
2410 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2411                 if (j == skip_idx)
2412                         continue;
2413 #endif
2414                 ip = chain->ips[j];
2415                 if (ip < PERF_CONTEXT_MAX)
2416                        ++nr_entries;
2417                 else if (callchain_param.order != ORDER_CALLEE) {
2418                         err = find_prev_cpumode(chain, thread, cursor, parent,
2419                                                 root_al, &cpumode, j);
2420                         if (err)
2421                                 return (err < 0) ? err : 0;
2422                         continue;
2423                 }
2424
2425                 err = add_callchain_ip(thread, cursor, parent,
2426                                        root_al, &cpumode, ip,
2427                                        false, NULL, NULL, 0);
2428
2429                 if (err)
2430                         return (err < 0) ? err : 0;
2431         }
2432
2433         return 0;
2434 }
2435
2436 static int append_inlines(struct callchain_cursor *cursor,
2437                           struct map *map, struct symbol *sym, u64 ip)
2438 {
2439         struct inline_node *inline_node;
2440         struct inline_list *ilist;
2441         u64 addr;
2442         int ret = 1;
2443
2444         if (!symbol_conf.inline_name || !map || !sym)
2445                 return ret;
2446
2447         addr = map__map_ip(map, ip);
2448         addr = map__rip_2objdump(map, addr);
2449
2450         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2451         if (!inline_node) {
2452                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2453                 if (!inline_node)
2454                         return ret;
2455                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2456         }
2457
2458         list_for_each_entry(ilist, &inline_node->val, list) {
2459                 ret = callchain_cursor_append(cursor, ip, map,
2460                                               ilist->symbol, false,
2461                                               NULL, 0, 0, 0, ilist->srcline);
2462
2463                 if (ret != 0)
2464                         return ret;
2465         }
2466
2467         return ret;
2468 }
2469
2470 static int unwind_entry(struct unwind_entry *entry, void *arg)
2471 {
2472         struct callchain_cursor *cursor = arg;
2473         const char *srcline = NULL;
2474         u64 addr = entry->ip;
2475
2476         if (symbol_conf.hide_unresolved && entry->sym == NULL)
2477                 return 0;
2478
2479         if (append_inlines(cursor, entry->map, entry->sym, entry->ip) == 0)
2480                 return 0;
2481
2482         /*
2483          * Convert entry->ip from a virtual address to an offset in
2484          * its corresponding binary.
2485          */
2486         if (entry->map)
2487                 addr = map__map_ip(entry->map, entry->ip);
2488
2489         srcline = callchain_srcline(entry->map, entry->sym, addr);
2490         return callchain_cursor_append(cursor, entry->ip,
2491                                        entry->map, entry->sym,
2492                                        false, NULL, 0, 0, 0, srcline);
2493 }
2494
2495 static int thread__resolve_callchain_unwind(struct thread *thread,
2496                                             struct callchain_cursor *cursor,
2497                                             struct perf_evsel *evsel,
2498                                             struct perf_sample *sample,
2499                                             int max_stack)
2500 {
2501         /* Can we do dwarf post unwind? */
2502         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2503               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
2504                 return 0;
2505
2506         /* Bail out if nothing was captured. */
2507         if ((!sample->user_regs.regs) ||
2508             (!sample->user_stack.size))
2509                 return 0;
2510
2511         return unwind__get_entries(unwind_entry, cursor,
2512                                    thread, sample, max_stack);
2513 }
2514
2515 int thread__resolve_callchain(struct thread *thread,
2516                               struct callchain_cursor *cursor,
2517                               struct perf_evsel *evsel,
2518                               struct perf_sample *sample,
2519                               struct symbol **parent,
2520                               struct addr_location *root_al,
2521                               int max_stack)
2522 {
2523         int ret = 0;
2524
2525         callchain_cursor_reset(cursor);
2526
2527         if (callchain_param.order == ORDER_CALLEE) {
2528                 ret = thread__resolve_callchain_sample(thread, cursor,
2529                                                        evsel, sample,
2530                                                        parent, root_al,
2531                                                        max_stack);
2532                 if (ret)
2533                         return ret;
2534                 ret = thread__resolve_callchain_unwind(thread, cursor,
2535                                                        evsel, sample,
2536                                                        max_stack);
2537         } else {
2538                 ret = thread__resolve_callchain_unwind(thread, cursor,
2539                                                        evsel, sample,
2540                                                        max_stack);
2541                 if (ret)
2542                         return ret;
2543                 ret = thread__resolve_callchain_sample(thread, cursor,
2544                                                        evsel, sample,
2545                                                        parent, root_al,
2546                                                        max_stack);
2547         }
2548
2549         return ret;
2550 }
2551
2552 int machine__for_each_thread(struct machine *machine,
2553                              int (*fn)(struct thread *thread, void *p),
2554                              void *priv)
2555 {
2556         struct threads *threads;
2557         struct rb_node *nd;
2558         struct thread *thread;
2559         int rc = 0;
2560         int i;
2561
2562         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2563                 threads = &machine->threads[i];
2564                 for (nd = rb_first_cached(&threads->entries); nd;
2565                      nd = rb_next(nd)) {
2566                         thread = rb_entry(nd, struct thread, rb_node);
2567                         rc = fn(thread, priv);
2568                         if (rc != 0)
2569                                 return rc;
2570                 }
2571
2572                 list_for_each_entry(thread, &threads->dead, node) {
2573                         rc = fn(thread, priv);
2574                         if (rc != 0)
2575                                 return rc;
2576                 }
2577         }
2578         return rc;
2579 }
2580
2581 int machines__for_each_thread(struct machines *machines,
2582                               int (*fn)(struct thread *thread, void *p),
2583                               void *priv)
2584 {
2585         struct rb_node *nd;
2586         int rc = 0;
2587
2588         rc = machine__for_each_thread(&machines->host, fn, priv);
2589         if (rc != 0)
2590                 return rc;
2591
2592         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2593                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2594
2595                 rc = machine__for_each_thread(machine, fn, priv);
2596                 if (rc != 0)
2597                         return rc;
2598         }
2599         return rc;
2600 }
2601
2602 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
2603                                   struct target *target, struct thread_map *threads,
2604                                   perf_event__handler_t process, bool data_mmap,
2605                                   unsigned int nr_threads_synthesize)
2606 {
2607         if (target__has_task(target))
2608                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap);
2609         else if (target__has_cpu(target))
2610                 return perf_event__synthesize_threads(tool, process,
2611                                                       machine, data_mmap,
2612                                                       nr_threads_synthesize);
2613         /* command specified */
2614         return 0;
2615 }
2616
2617 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2618 {
2619         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
2620                 return -1;
2621
2622         return machine->current_tid[cpu];
2623 }
2624
2625 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2626                              pid_t tid)
2627 {
2628         struct thread *thread;
2629
2630         if (cpu < 0)
2631                 return -EINVAL;
2632
2633         if (!machine->current_tid) {
2634                 int i;
2635
2636                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
2637                 if (!machine->current_tid)
2638                         return -ENOMEM;
2639                 for (i = 0; i < MAX_NR_CPUS; i++)
2640                         machine->current_tid[i] = -1;
2641         }
2642
2643         if (cpu >= MAX_NR_CPUS) {
2644                 pr_err("Requested CPU %d too large. ", cpu);
2645                 pr_err("Consider raising MAX_NR_CPUS\n");
2646                 return -EINVAL;
2647         }
2648
2649         machine->current_tid[cpu] = tid;
2650
2651         thread = machine__findnew_thread(machine, pid, tid);
2652         if (!thread)
2653                 return -ENOMEM;
2654
2655         thread->cpu = cpu;
2656         thread__put(thread);
2657
2658         return 0;
2659 }
2660
2661 /*
2662  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2663  * normalized arch is needed.
2664  */
2665 bool machine__is(struct machine *machine, const char *arch)
2666 {
2667         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2668 }
2669
2670 int machine__nr_cpus_avail(struct machine *machine)
2671 {
2672         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2673 }
2674
2675 int machine__get_kernel_start(struct machine *machine)
2676 {
2677         struct map *map = machine__kernel_map(machine);
2678         int err = 0;
2679
2680         /*
2681          * The only addresses above 2^63 are kernel addresses of a 64-bit
2682          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2683          * all addresses including kernel addresses are less than 2^32.  In
2684          * that case (32-bit system), if the kernel mapping is unknown, all
2685          * addresses will be assumed to be in user space - see
2686          * machine__kernel_ip().
2687          */
2688         machine->kernel_start = 1ULL << 63;
2689         if (map) {
2690                 err = map__load(map);
2691                 /*
2692                  * On x86_64, PTI entry trampolines are less than the
2693                  * start of kernel text, but still above 2^63. So leave
2694                  * kernel_start = 1ULL << 63 for x86_64.
2695                  */
2696                 if (!err && !machine__is(machine, "x86_64"))
2697                         machine->kernel_start = map->start;
2698         }
2699         return err;
2700 }
2701
2702 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2703 {
2704         u8 addr_cpumode = cpumode;
2705         bool kernel_ip;
2706
2707         if (!machine->single_address_space)
2708                 goto out;
2709
2710         kernel_ip = machine__kernel_ip(machine, addr);
2711         switch (cpumode) {
2712         case PERF_RECORD_MISC_KERNEL:
2713         case PERF_RECORD_MISC_USER:
2714                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2715                                            PERF_RECORD_MISC_USER;
2716                 break;
2717         case PERF_RECORD_MISC_GUEST_KERNEL:
2718         case PERF_RECORD_MISC_GUEST_USER:
2719                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2720                                            PERF_RECORD_MISC_GUEST_USER;
2721                 break;
2722         default:
2723                 break;
2724         }
2725 out:
2726         return addr_cpumode;
2727 }
2728
2729 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2730 {
2731         return dsos__findnew(&machine->dsos, filename);
2732 }
2733
2734 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2735 {
2736         struct machine *machine = vmachine;
2737         struct map *map;
2738         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2739
2740         if (sym == NULL)
2741                 return NULL;
2742
2743         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2744         *addrp = map->unmap_ip(map, sym->start);
2745         return sym->name;
2746 }