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 <stdlib.h>
   7 #include "callchain.h"
   8 #include "debug.h"
   9 #include "dso.h"
  10 #include "env.h"
  11 #include "event.h"
  12 #include "evsel.h"
  13 #include "hist.h"
  14 #include "machine.h"
  15 #include "map.h"
  16 #include "map_symbol.h"
  17 #include "branch.h"
  18 #include "mem-events.h"
  19 #include "srcline.h"
  20 #include "symbol.h"
  21 #include "sort.h"
  22 #include "strlist.h"
  23 #include "target.h"
  24 #include "thread.h"
  25 #include "util.h"
  26 #include "vdso.h"
  27 #include <stdbool.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <unistd.h>
  31 #include "unwind.h"
  32 #include "linux/hash.h"
  33 #include "asm/bug.h"
  34 #include "bpf-event.h"
  35 #include <internal/lib.h> // page_size
  36
  37 #include <linux/ctype.h>
  38 #include <symbol/kallsyms.h>
  39 #include <linux/mman.h>
  40 #include <linux/string.h>
  41 #include <linux/zalloc.h>
  42
  43 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  44
  45 static struct dso *machine__kernel_dso(struct machine *machine)
  46 {
  47         return machine->vmlinux_map->dso;
  48 }
  49
  50 static void dsos__init(struct dsos *dsos)
  51 {
  52         INIT_LIST_HEAD(&dsos->head);
  53         dsos->root = RB_ROOT;
  54         init_rwsem(&dsos->lock);
  55 }
  56
  57 static void machine__threads_init(struct machine *machine)
  58 {
  59         int i;
  60
  61         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  62                 struct threads *threads = &machine->threads[i];
  63                 threads->entries = RB_ROOT_CACHED;
  64                 init_rwsem(&threads->lock);
  65                 threads->nr = 0;
  66                 INIT_LIST_HEAD(&threads->dead);
  67                 threads->last_match = NULL;
  68         }
  69 }
  70
  71 static int machine__set_mmap_name(struct machine *machine)
  72 {
  73         if (machine__is_host(machine))
  74                 machine->mmap_name = strdup("[kernel.kallsyms]");
  75         else if (machine__is_default_guest(machine))
  76                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  77         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  78                           machine->pid) < 0)
  79                 machine->mmap_name = NULL;
  80
  81         return machine->mmap_name ? 0 : -ENOMEM;
  82 }
  83
  84 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  85 {
  86         int err = -ENOMEM;
  87
  88         memset(machine, 0, sizeof(*machine));
  89         map_groups__init(&machine->kmaps, machine);
  90         RB_CLEAR_NODE(&machine->rb_node);
  91         dsos__init(&machine->dsos);
  92
  93         machine__threads_init(machine);
  94
  95         machine->vdso_info = NULL;
  96         machine->env = NULL;
  97
  98         machine->pid = pid;
  99
 100         machine->id_hdr_size = 0;
 101         machine->kptr_restrict_warned = false;
 102         machine->comm_exec = false;
 103         machine->kernel_start = 0;
 104         machine->vmlinux_map = NULL;
 105
 106         machine->root_dir = strdup(root_dir);
 107         if (machine->root_dir == NULL)
 108                 return -ENOMEM;
 109
 110         if (machine__set_mmap_name(machine))
 111                 goto out;
 112
 113         if (pid != HOST_KERNEL_ID) {
 114                 struct thread *thread = machine__findnew_thread(machine, -1,
 115                                                                 pid);
 116                 char comm[64];
 117
 118                 if (thread == NULL)
 119                         goto out;
 120
 121                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 122                 thread__set_comm(thread, comm, 0);
 123                 thread__put(thread);
 124         }
 125
 126         machine->current_tid = NULL;
 127         err = 0;
 128
 129 out:
 130         if (err) {
 131                 zfree(&machine->root_dir);
 132                 zfree(&machine->mmap_name);
 133         }
 134         return 0;
 135 }
 136
 137 struct machine *machine__new_host(void)
 138 {
 139         struct machine *machine = malloc(sizeof(*machine));
 140
 141         if (machine != NULL) {
 142                 machine__init(machine, "", HOST_KERNEL_ID);
 143
 144                 if (machine__create_kernel_maps(machine) < 0)
 145                         goto out_delete;
 146         }
 147
 148         return machine;
 149 out_delete:
 150         free(machine);
 151         return NULL;
 152 }
 153
 154 struct machine *machine__new_kallsyms(void)
 155 {
 156         struct machine *machine = machine__new_host();
 157         /*
 158          * FIXME:
 159          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 160          *    ask for not using the kcore parsing code, once this one is fixed
 161          *    to create a map per module.
 162          */
 163         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 164                 machine__delete(machine);
 165                 machine = NULL;
 166         }
 167
 168         return machine;
 169 }
 170
 171 static void dsos__purge(struct dsos *dsos)
 172 {
 173         struct dso *pos, *n;
 174
 175         down_write(&dsos->lock);
 176
 177         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 178                 RB_CLEAR_NODE(&pos->rb_node);
 179                 pos->root = NULL;
 180                 list_del_init(&pos->node);
 181                 dso__put(pos);
 182         }
 183
 184         up_write(&dsos->lock);
 185 }
 186
 187 static void dsos__exit(struct dsos *dsos)
 188 {
 189         dsos__purge(dsos);
 190         exit_rwsem(&dsos->lock);
 191 }
 192
 193 void machine__delete_threads(struct machine *machine)
 194 {
 195         struct rb_node *nd;
 196         int i;
 197
 198         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 199                 struct threads *threads = &machine->threads[i];
 200                 down_write(&threads->lock);
 201                 nd = rb_first_cached(&threads->entries);
 202                 while (nd) {
 203                         struct thread *t = rb_entry(nd, struct thread, rb_node);
 204
 205                         nd = rb_next(nd);
 206                         __machine__remove_thread(machine, t, false);
 207                 }
 208                 up_write(&threads->lock);
 209         }
 210 }
 211
 212 void machine__exit(struct machine *machine)
 213 {
 214         int i;
 215
 216         if (machine == NULL)
 217                 return;
 218
 219         machine__destroy_kernel_maps(machine);
 220         map_groups__exit(&machine->kmaps);
 221         dsos__exit(&machine->dsos);
 222         machine__exit_vdso(machine);
 223         zfree(&machine->root_dir);
 224         zfree(&machine->mmap_name);
 225         zfree(&machine->current_tid);
 226
 227         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 228                 struct threads *threads = &machine->threads[i];
 229                 struct thread *thread, *n;
 230                 /*
 231                  * Forget about the dead, at this point whatever threads were
 232                  * left in the dead lists better have a reference count taken
 233                  * by who is using them, and then, when they drop those references
 234                  * and it finally hits zero, thread__put() will check and see that
 235                  * its not in the dead threads list and will not try to remove it
 236                  * from there, just calling thread__delete() straight away.
 237                  */
 238                 list_for_each_entry_safe(thread, n, &threads->dead, node)
 239                         list_del_init(&thread->node);
 240
 241                 exit_rwsem(&threads->lock);
 242         }
 243 }
 244
 245 void machine__delete(struct machine *machine)
 246 {
 247         if (machine) {
 248                 machine__exit(machine);
 249                 free(machine);
 250         }
 251 }
 252
 253 void machines__init(struct machines *machines)
 254 {
 255         machine__init(&machines->host, "", HOST_KERNEL_ID);
 256         machines->guests = RB_ROOT_CACHED;
 257 }
 258
 259 void machines__exit(struct machines *machines)
 260 {
 261         machine__exit(&machines->host);
 262         /* XXX exit guest */
 263 }
 264
 265 struct machine *machines__add(struct machines *machines, pid_t pid,
 266                               const char *root_dir)
 267 {
 268         struct rb_node **p = &machines->guests.rb_root.rb_node;
 269         struct rb_node *parent = NULL;
 270         struct machine *pos, *machine = malloc(sizeof(*machine));
 271         bool leftmost = true;
 272
 273         if (machine == NULL)
 274                 return NULL;
 275
 276         if (machine__init(machine, root_dir, pid) != 0) {
 277                 free(machine);
 278                 return NULL;
 279         }
 280
 281         while (*p != NULL) {
 282                 parent = *p;
 283                 pos = rb_entry(parent, struct machine, rb_node);
 284                 if (pid < pos->pid)
 285                         p = &(*p)->rb_left;
 286                 else {
 287                         p = &(*p)->rb_right;
 288                         leftmost = false;
 289                 }
 290         }
 291
 292         rb_link_node(&machine->rb_node, parent, p);
 293         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 294
 295         return machine;
 296 }
 297
 298 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 299 {
 300         struct rb_node *nd;
 301
 302         machines->host.comm_exec = comm_exec;
 303
 304         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 305                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 306
 307                 machine->comm_exec = comm_exec;
 308         }
 309 }
 310
 311 struct machine *machines__find(struct machines *machines, pid_t pid)
 312 {
 313         struct rb_node **p = &machines->guests.rb_root.rb_node;
 314         struct rb_node *parent = NULL;
 315         struct machine *machine;
 316         struct machine *default_machine = NULL;
 317
 318         if (pid == HOST_KERNEL_ID)
 319                 return &machines->host;
 320
 321         while (*p != NULL) {
 322                 parent = *p;
 323                 machine = rb_entry(parent, struct machine, rb_node);
 324                 if (pid < machine->pid)
 325                         p = &(*p)->rb_left;
 326                 else if (pid > machine->pid)
 327                         p = &(*p)->rb_right;
 328                 else
 329                         return machine;
 330                 if (!machine->pid)
 331                         default_machine = machine;
 332         }
 333
 334         return default_machine;
 335 }
 336
 337 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 338 {
 339         char path[PATH_MAX];
 340         const char *root_dir = "";
 341         struct machine *machine = machines__find(machines, pid);
 342
 343         if (machine && (machine->pid == pid))
 344                 goto out;
 345
 346         if ((pid != HOST_KERNEL_ID) &&
 347             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 348             (symbol_conf.guestmount)) {
 349                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 350                 if (access(path, R_OK)) {
 351                         static struct strlist *seen;
 352
 353                         if (!seen)
 354                                 seen = strlist__new(NULL, NULL);
 355
 356                         if (!strlist__has_entry(seen, path)) {
 357                                 pr_err("Can't access file %s\n", path);
 358                                 strlist__add(seen, path);
 359                         }
 360                         machine = NULL;
 361                         goto out;
 362                 }
 363                 root_dir = path;
 364         }
 365
 366         machine = machines__add(machines, pid, root_dir);
 367 out:
 368         return machine;
 369 }
 370
 371 void machines__process_guests(struct machines *machines,
 372                               machine__process_t process, void *data)
 373 {
 374         struct rb_node *nd;
 375
 376         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 377                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 378                 process(pos, data);
 379         }
 380 }
 381
 382 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 383 {
 384         struct rb_node *node;
 385         struct machine *machine;
 386
 387         machines->host.id_hdr_size = id_hdr_size;
 388
 389         for (node = rb_first_cached(&machines->guests); node;
 390              node = rb_next(node)) {
 391                 machine = rb_entry(node, struct machine, rb_node);
 392                 machine->id_hdr_size = id_hdr_size;
 393         }
 394
 395         return;
 396 }
 397
 398 static void machine__update_thread_pid(struct machine *machine,
 399                                        struct thread *th, pid_t pid)
 400 {
 401         struct thread *leader;
 402
 403         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 404                 return;
 405
 406         th->pid_ = pid;
 407
 408         if (th->pid_ == th->tid)
 409                 return;
 410
 411         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 412         if (!leader)
 413                 goto out_err;
 414
 415         if (!leader->mg)
 416                 leader->mg = map_groups__new(machine);
 417
 418         if (!leader->mg)
 419                 goto out_err;
 420
 421         if (th->mg == leader->mg)
 422                 return;
 423
 424         if (th->mg) {
 425                 /*
 426                  * Maps are created from MMAP events which provide the pid and
 427                  * tid.  Consequently there never should be any maps on a thread
 428                  * with an unknown pid.  Just print an error if there are.
 429                  */
 430                 if (!map_groups__empty(th->mg))
 431                         pr_err("Discarding thread maps for %d:%d\n",
 432                                th->pid_, th->tid);
 433                 map_groups__put(th->mg);
 434         }
 435
 436         th->mg = map_groups__get(leader->mg);
 437 out_put:
 438         thread__put(leader);
 439         return;
 440 out_err:
 441         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 442         goto out_put;
 443 }
 444
 445 /*
 446  * Front-end cache - TID lookups come in blocks,
 447  * so most of the time we dont have to look up
 448  * the full rbtree:
 449  */
 450 static struct thread*
 451 __threads__get_last_match(struct threads *threads, struct machine *machine,
 452                           int pid, int tid)
 453 {
 454         struct thread *th;
 455
 456         th = threads->last_match;
 457         if (th != NULL) {
 458                 if (th->tid == tid) {
 459                         machine__update_thread_pid(machine, th, pid);
 460                         return thread__get(th);
 461                 }
 462
 463                 threads->last_match = NULL;
 464         }
 465
 466         return NULL;
 467 }
 468
 469 static struct thread*
 470 threads__get_last_match(struct threads *threads, struct machine *machine,
 471                         int pid, int tid)
 472 {
 473         struct thread *th = NULL;
 474
 475         if (perf_singlethreaded)
 476                 th = __threads__get_last_match(threads, machine, pid, tid);
 477
 478         return th;
 479 }
 480
 481 static void
 482 __threads__set_last_match(struct threads *threads, struct thread *th)
 483 {
 484         threads->last_match = th;
 485 }
 486
 487 static void
 488 threads__set_last_match(struct threads *threads, struct thread *th)
 489 {
 490         if (perf_singlethreaded)
 491                 __threads__set_last_match(threads, th);
 492 }
 493
 494 /*
 495  * Caller must eventually drop thread->refcnt returned with a successful
 496  * lookup/new thread inserted.
 497  */
 498 static struct thread *____machine__findnew_thread(struct machine *machine,
 499                                                   struct threads *threads,
 500                                                   pid_t pid, pid_t tid,
 501                                                   bool create)
 502 {
 503         struct rb_node **p = &threads->entries.rb_root.rb_node;
 504         struct rb_node *parent = NULL;
 505         struct thread *th;
 506         bool leftmost = true;
 507
 508         th = threads__get_last_match(threads, machine, pid, tid);
 509         if (th)
 510                 return th;
 511
 512         while (*p != NULL) {
 513                 parent = *p;
 514                 th = rb_entry(parent, struct thread, rb_node);
 515
 516                 if (th->tid == tid) {
 517                         threads__set_last_match(threads, th);
 518                         machine__update_thread_pid(machine, th, pid);
 519                         return thread__get(th);
 520                 }
 521
 522                 if (tid < th->tid)
 523                         p = &(*p)->rb_left;
 524                 else {
 525                         p = &(*p)->rb_right;
 526                         leftmost = false;
 527                 }
 528         }
 529
 530         if (!create)
 531                 return NULL;
 532
 533         th = thread__new(pid, tid);
 534         if (th != NULL) {
 535                 rb_link_node(&th->rb_node, parent, p);
 536                 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
 537
 538                 /*
 539                  * We have to initialize map_groups separately
 540                  * after rb tree is updated.
 541                  *
 542                  * The reason is that we call machine__findnew_thread
 543                  * within thread__init_map_groups to find the thread
 544                  * leader and that would screwed the rb tree.
 545                  */
 546                 if (thread__init_map_groups(th, machine)) {
 547                         rb_erase_cached(&th->rb_node, &threads->entries);
 548                         RB_CLEAR_NODE(&th->rb_node);
 549                         thread__put(th);
 550                         return NULL;
 551                 }
 552                 /*
 553                  * It is now in the rbtree, get a ref
 554                  */
 555                 thread__get(th);
 556                 threads__set_last_match(threads, th);
 557                 ++threads->nr;
 558         }
 559
 560         return th;
 561 }
 562
 563 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 564 {
 565         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 566 }
 567
 568 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 569                                        pid_t tid)
 570 {
 571         struct threads *threads = machine__threads(machine, tid);
 572         struct thread *th;
 573
 574         down_write(&threads->lock);
 575         th = __machine__findnew_thread(machine, pid, tid);
 576         up_write(&threads->lock);
 577         return th;
 578 }
 579
 580 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 581                                     pid_t tid)
 582 {
 583         struct threads *threads = machine__threads(machine, tid);
 584         struct thread *th;
 585
 586         down_read(&threads->lock);
 587         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 588         up_read(&threads->lock);
 589         return th;
 590 }
 591
 592 struct comm *machine__thread_exec_comm(struct machine *machine,
 593                                        struct thread *thread)
 594 {
 595         if (machine->comm_exec)
 596                 return thread__exec_comm(thread);
 597         else
 598                 return thread__comm(thread);
 599 }
 600
 601 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 602                                 struct perf_sample *sample)
 603 {
 604         struct thread *thread = machine__findnew_thread(machine,
 605                                                         event->comm.pid,
 606                                                         event->comm.tid);
 607         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 608         int err = 0;
 609
 610         if (exec)
 611                 machine->comm_exec = true;
 612
 613         if (dump_trace)
 614                 perf_event__fprintf_comm(event, stdout);
 615
 616         if (thread == NULL ||
 617             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 618                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 619                 err = -1;
 620         }
 621
 622         thread__put(thread);
 623
 624         return err;
 625 }
 626
 627 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 628                                       union perf_event *event,
 629                                       struct perf_sample *sample __maybe_unused)
 630 {
 631         struct thread *thread = machine__findnew_thread(machine,
 632                                                         event->namespaces.pid,
 633                                                         event->namespaces.tid);
 634         int err = 0;
 635
 636         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 637                   "\nWARNING: kernel seems to support more namespaces than perf"
 638                   " tool.\nTry updating the perf tool..\n\n");
 639
 640         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 641                   "\nWARNING: perf tool seems to support more namespaces than"
 642                   " the kernel.\nTry updating the kernel..\n\n");
 643
 644         if (dump_trace)
 645                 perf_event__fprintf_namespaces(event, stdout);
 646
 647         if (thread == NULL ||
 648             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 649                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 650                 err = -1;
 651         }
 652
 653         thread__put(thread);
 654
 655         return err;
 656 }
 657
 658 int machine__process_lost_event(struct machine *machine __maybe_unused,
 659                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 660 {
 661         dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
 662                     event->lost.id, event->lost.lost);
 663         return 0;
 664 }
 665
 666 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 667                                         union perf_event *event, struct perf_sample *sample)
 668 {
 669         dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
 670                     sample->id, event->lost_samples.lost);
 671         return 0;
 672 }
 673
 674 static struct dso *machine__findnew_module_dso(struct machine *machine,
 675                                                struct kmod_path *m,
 676                                                const char *filename)
 677 {
 678         struct dso *dso;
 679
 680         down_write(&machine->dsos.lock);
 681
 682         dso = __dsos__find(&machine->dsos, m->name, true);
 683         if (!dso) {
 684                 dso = __dsos__addnew(&machine->dsos, m->name);
 685                 if (dso == NULL)
 686                         goto out_unlock;
 687
 688                 dso__set_module_info(dso, m, machine);
 689                 dso__set_long_name(dso, strdup(filename), true);
 690         }
 691
 692         dso__get(dso);
 693 out_unlock:
 694         up_write(&machine->dsos.lock);
 695         return dso;
 696 }
 697
 698 int machine__process_aux_event(struct machine *machine __maybe_unused,
 699                                union perf_event *event)
 700 {
 701         if (dump_trace)
 702                 perf_event__fprintf_aux(event, stdout);
 703         return 0;
 704 }
 705
 706 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 707                                         union perf_event *event)
 708 {
 709         if (dump_trace)
 710                 perf_event__fprintf_itrace_start(event, stdout);
 711         return 0;
 712 }
 713
 714 int machine__process_switch_event(struct machine *machine __maybe_unused,
 715                                   union perf_event *event)
 716 {
 717         if (dump_trace)
 718                 perf_event__fprintf_switch(event, stdout);
 719         return 0;
 720 }
 721
 722 static int machine__process_ksymbol_register(struct machine *machine,
 723                                              union perf_event *event,
 724                                              struct perf_sample *sample __maybe_unused)
 725 {
 726         struct symbol *sym;
 727         struct map *map;
 728
 729         map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
 730         if (!map) {
 731                 map = dso__new_map(event->ksymbol.name);
 732                 if (!map)
 733                         return -ENOMEM;
 734
 735                 map->start = event->ksymbol.addr;
 736                 map->end = map->start + event->ksymbol.len;
 737                 map_groups__insert(&machine->kmaps, map);
 738         }
 739
 740         sym = symbol__new(map->map_ip(map, map->start),
 741                           event->ksymbol.len,
 742                           0, 0, event->ksymbol.name);
 743         if (!sym)
 744                 return -ENOMEM;
 745         dso__insert_symbol(map->dso, sym);
 746         return 0;
 747 }
 748
 749 static int machine__process_ksymbol_unregister(struct machine *machine,
 750                                                union perf_event *event,
 751                                                struct perf_sample *sample __maybe_unused)
 752 {
 753         struct map *map;
 754
 755         map = map_groups__find(&machine->kmaps, event->ksymbol.addr);
 756         if (map)
 757                 map_groups__remove(&machine->kmaps, map);
 758
 759         return 0;
 760 }
 761
 762 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 763                              union perf_event *event,
 764                              struct perf_sample *sample)
 765 {
 766         if (dump_trace)
 767                 perf_event__fprintf_ksymbol(event, stdout);
 768
 769         if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 770                 return machine__process_ksymbol_unregister(machine, event,
 771                                                            sample);
 772         return machine__process_ksymbol_register(machine, event, sample);
 773 }
 774
 775 static void dso__adjust_kmod_long_name(struct dso *dso, const char *filename)
 776 {
 777         const char *dup_filename;
 778
 779         if (!filename || !dso || !dso->long_name)
 780                 return;
 781         if (dso->long_name[0] != '[')
 782                 return;
 783         if (!strchr(filename, '/'))
 784                 return;
 785
 786         dup_filename = strdup(filename);
 787         if (!dup_filename)
 788                 return;
 789
 790         dso__set_long_name(dso, dup_filename, true);
 791 }
 792
 793 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 794                                         const char *filename)
 795 {
 796         struct map *map = NULL;
 797         struct dso *dso = NULL;
 798         struct kmod_path m;
 799
 800         if (kmod_path__parse_name(&m, filename))
 801                 return NULL;
 802
 803         map = map_groups__find_by_name(&machine->kmaps, m.name);
 804         if (map) {
 805                 /*
 806                  * If the map's dso is an offline module, give dso__load()
 807                  * a chance to find the file path of that module by fixing
 808                  * long_name.
 809                  */
 810                 dso__adjust_kmod_long_name(map->dso, filename);
 811                 goto out;
 812         }
 813
 814         dso = machine__findnew_module_dso(machine, &m, filename);
 815         if (dso == NULL)
 816                 goto out;
 817
 818         map = map__new2(start, dso);
 819         if (map == NULL)
 820                 goto out;
 821
 822         map_groups__insert(&machine->kmaps, map);
 823
 824         /* Put the map here because map_groups__insert alread got it */
 825         map__put(map);
 826 out:
 827         /* put the dso here, corresponding to  machine__findnew_module_dso */
 828         dso__put(dso);
 829         zfree(&m.name);
 830         return map;
 831 }
 832
 833 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 834 {
 835         struct rb_node *nd;
 836         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 837
 838         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 839                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 840                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 841         }
 842
 843         return ret;
 844 }
 845
 846 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 847                                      bool (skip)(struct dso *dso, int parm), int parm)
 848 {
 849         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 850 }
 851
 852 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 853                                      bool (skip)(struct dso *dso, int parm), int parm)
 854 {
 855         struct rb_node *nd;
 856         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 857
 858         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 859                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 860                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 861         }
 862         return ret;
 863 }
 864
 865 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 866 {
 867         int i;
 868         size_t printed = 0;
 869         struct dso *kdso = machine__kernel_dso(machine);
 870
 871         if (kdso->has_build_id) {
 872                 char filename[PATH_MAX];
 873                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
 874                                            false))
 875                         printed += fprintf(fp, "[0] %s\n", filename);
 876         }
 877
 878         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 879                 printed += fprintf(fp, "[%d] %s\n",
 880                                    i + kdso->has_build_id, vmlinux_path[i]);
 881
 882         return printed;
 883 }
 884
 885 size_t machine__fprintf(struct machine *machine, FILE *fp)
 886 {
 887         struct rb_node *nd;
 888         size_t ret;
 889         int i;
 890
 891         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 892                 struct threads *threads = &machine->threads[i];
 893
 894                 down_read(&threads->lock);
 895
 896                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
 897
 898                 for (nd = rb_first_cached(&threads->entries); nd;
 899                      nd = rb_next(nd)) {
 900                         struct thread *pos = rb_entry(nd, struct thread, rb_node);
 901
 902                         ret += thread__fprintf(pos, fp);
 903                 }
 904
 905                 up_read(&threads->lock);
 906         }
 907         return ret;
 908 }
 909
 910 static struct dso *machine__get_kernel(struct machine *machine)
 911 {
 912         const char *vmlinux_name = machine->mmap_name;
 913         struct dso *kernel;
 914
 915         if (machine__is_host(machine)) {
 916                 if (symbol_conf.vmlinux_name)
 917                         vmlinux_name = symbol_conf.vmlinux_name;
 918
 919                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 920                                                  "[kernel]", DSO_TYPE_KERNEL);
 921         } else {
 922                 if (symbol_conf.default_guest_vmlinux_name)
 923                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 924
 925                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 926                                                  "[guest.kernel]",
 927                                                  DSO_TYPE_GUEST_KERNEL);
 928         }
 929
 930         if (kernel != NULL && (!kernel->has_build_id))
 931                 dso__read_running_kernel_build_id(kernel, machine);
 932
 933         return kernel;
 934 }
 935
 936 struct process_args {
 937         u64 start;
 938 };
 939
 940 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 941                                     size_t bufsz)
 942 {
 943         if (machine__is_default_guest(machine))
 944                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 945         else
 946                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 947 }
 948
 949 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 950
 951 /* Figure out the start address of kernel map from /proc/kallsyms.
 952  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 953  * symbol_name if it's not that important.
 954  */
 955 static int machine__get_running_kernel_start(struct machine *machine,
 956                                              const char **symbol_name,
 957                                              u64 *start, u64 *end)
 958 {
 959         char filename[PATH_MAX];
 960         int i, err = -1;
 961         const char *name;
 962         u64 addr = 0;
 963
 964         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 965
 966         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 967                 return 0;
 968
 969         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 970                 err = kallsyms__get_function_start(filename, name, &addr);
 971                 if (!err)
 972                         break;
 973         }
 974
 975         if (err)
 976                 return -1;
 977
 978         if (symbol_name)
 979                 *symbol_name = name;
 980
 981         *start = addr;
 982
 983         err = kallsyms__get_function_start(filename, "_etext", &addr);
 984         if (!err)
 985                 *end = addr;
 986
 987         return 0;
 988 }
 989
 990 int machine__create_extra_kernel_map(struct machine *machine,
 991                                      struct dso *kernel,
 992                                      struct extra_kernel_map *xm)
 993 {
 994         struct kmap *kmap;
 995         struct map *map;
 996
 997         map = map__new2(xm->start, kernel);
 998         if (!map)
 999                 return -1;
1000
1001         map->end   = xm->end;
1002         map->pgoff = xm->pgoff;
1003
1004         kmap = map__kmap(map);
1005
1006         kmap->kmaps = &machine->kmaps;
1007         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1008
1009         map_groups__insert(&machine->kmaps, map);
1010
1011         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1012                   kmap->name, map->start, map->end);
1013
1014         map__put(map);
1015
1016         return 0;
1017 }
1018
1019 static u64 find_entry_trampoline(struct dso *dso)
1020 {
1021         /* Duplicates are removed so lookup all aliases */
1022         const char *syms[] = {
1023                 "_entry_trampoline",
1024                 "__entry_trampoline_start",
1025                 "entry_SYSCALL_64_trampoline",
1026         };
1027         struct symbol *sym = dso__first_symbol(dso);
1028         unsigned int i;
1029
1030         for (; sym; sym = dso__next_symbol(sym)) {
1031                 if (sym->binding != STB_GLOBAL)
1032                         continue;
1033                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1034                         if (!strcmp(sym->name, syms[i]))
1035                                 return sym->start;
1036                 }
1037         }
1038
1039         return 0;
1040 }
1041
1042 /*
1043  * These values can be used for kernels that do not have symbols for the entry
1044  * trampolines in kallsyms.
1045  */
1046 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1047 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1048 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1049
1050 /* Map x86_64 PTI entry trampolines */
1051 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1052                                           struct dso *kernel)
1053 {
1054         struct map_groups *kmaps = &machine->kmaps;
1055         struct maps *maps = &kmaps->maps;
1056         int nr_cpus_avail, cpu;
1057         bool found = false;
1058         struct map *map;
1059         u64 pgoff;
1060
1061         /*
1062          * In the vmlinux case, pgoff is a virtual address which must now be
1063          * mapped to a vmlinux offset.
1064          */
1065         maps__for_each_entry(maps, map) {
1066                 struct kmap *kmap = __map__kmap(map);
1067                 struct map *dest_map;
1068
1069                 if (!kmap || !is_entry_trampoline(kmap->name))
1070                         continue;
1071
1072                 dest_map = map_groups__find(kmaps, map->pgoff);
1073                 if (dest_map != map)
1074                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1075                 found = true;
1076         }
1077         if (found || machine->trampolines_mapped)
1078                 return 0;
1079
1080         pgoff = find_entry_trampoline(kernel);
1081         if (!pgoff)
1082                 return 0;
1083
1084         nr_cpus_avail = machine__nr_cpus_avail(machine);
1085
1086         /* Add a 1 page map for each CPU's entry trampoline */
1087         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1088                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1089                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1090                          X86_64_ENTRY_TRAMPOLINE;
1091                 struct extra_kernel_map xm = {
1092                         .start = va,
1093                         .end   = va + page_size,
1094                         .pgoff = pgoff,
1095                 };
1096
1097                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1098
1099                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1100                         return -1;
1101         }
1102
1103         machine->trampolines_mapped = nr_cpus_avail;
1104
1105         return 0;
1106 }
1107
1108 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1109                                              struct dso *kernel __maybe_unused)
1110 {
1111         return 0;
1112 }
1113
1114 static int
1115 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1116 {
1117         struct kmap *kmap;
1118         struct map *map;
1119
1120         /* In case of renewal the kernel map, destroy previous one */
1121         machine__destroy_kernel_maps(machine);
1122
1123         machine->vmlinux_map = map__new2(0, kernel);
1124         if (machine->vmlinux_map == NULL)
1125                 return -1;
1126
1127         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1128         map = machine__kernel_map(machine);
1129         kmap = map__kmap(map);
1130         if (!kmap)
1131                 return -1;
1132
1133         kmap->kmaps = &machine->kmaps;
1134         map_groups__insert(&machine->kmaps, map);
1135
1136         return 0;
1137 }
1138
1139 void machine__destroy_kernel_maps(struct machine *machine)
1140 {
1141         struct kmap *kmap;
1142         struct map *map = machine__kernel_map(machine);
1143
1144         if (map == NULL)
1145                 return;
1146
1147         kmap = map__kmap(map);
1148         map_groups__remove(&machine->kmaps, map);
1149         if (kmap && kmap->ref_reloc_sym) {
1150                 zfree((char **)&kmap->ref_reloc_sym->name);
1151                 zfree(&kmap->ref_reloc_sym);
1152         }
1153
1154         map__zput(machine->vmlinux_map);
1155 }
1156
1157 int machines__create_guest_kernel_maps(struct machines *machines)
1158 {
1159         int ret = 0;
1160         struct dirent **namelist = NULL;
1161         int i, items = 0;
1162         char path[PATH_MAX];
1163         pid_t pid;
1164         char *endp;
1165
1166         if (symbol_conf.default_guest_vmlinux_name ||
1167             symbol_conf.default_guest_modules ||
1168             symbol_conf.default_guest_kallsyms) {
1169                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1170         }
1171
1172         if (symbol_conf.guestmount) {
1173                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1174                 if (items <= 0)
1175                         return -ENOENT;
1176                 for (i = 0; i < items; i++) {
1177                         if (!isdigit(namelist[i]->d_name[0])) {
1178                                 /* Filter out . and .. */
1179                                 continue;
1180                         }
1181                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1182                         if ((*endp != '\0') ||
1183                             (endp == namelist[i]->d_name) ||
1184                             (errno == ERANGE)) {
1185                                 pr_debug("invalid directory (%s). Skipping.\n",
1186                                          namelist[i]->d_name);
1187                                 continue;
1188                         }
1189                         sprintf(path, "%s/%s/proc/kallsyms",
1190                                 symbol_conf.guestmount,
1191                                 namelist[i]->d_name);
1192                         ret = access(path, R_OK);
1193                         if (ret) {
1194                                 pr_debug("Can't access file %s\n", path);
1195                                 goto failure;
1196                         }
1197                         machines__create_kernel_maps(machines, pid);
1198                 }
1199 failure:
1200                 free(namelist);
1201         }
1202
1203         return ret;
1204 }
1205
1206 void machines__destroy_kernel_maps(struct machines *machines)
1207 {
1208         struct rb_node *next = rb_first_cached(&machines->guests);
1209
1210         machine__destroy_kernel_maps(&machines->host);
1211
1212         while (next) {
1213                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1214
1215                 next = rb_next(&pos->rb_node);
1216                 rb_erase_cached(&pos->rb_node, &machines->guests);
1217                 machine__delete(pos);
1218         }
1219 }
1220
1221 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1222 {
1223         struct machine *machine = machines__findnew(machines, pid);
1224
1225         if (machine == NULL)
1226                 return -1;
1227
1228         return machine__create_kernel_maps(machine);
1229 }
1230
1231 int machine__load_kallsyms(struct machine *machine, const char *filename)
1232 {
1233         struct map *map = machine__kernel_map(machine);
1234         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1235
1236         if (ret > 0) {
1237                 dso__set_loaded(map->dso);
1238                 /*
1239                  * Since /proc/kallsyms will have multiple sessions for the
1240                  * kernel, with modules between them, fixup the end of all
1241                  * sections.
1242                  */
1243                 map_groups__fixup_end(&machine->kmaps);
1244         }
1245
1246         return ret;
1247 }
1248
1249 int machine__load_vmlinux_path(struct machine *machine)
1250 {
1251         struct map *map = machine__kernel_map(machine);
1252         int ret = dso__load_vmlinux_path(map->dso, map);
1253
1254         if (ret > 0)
1255                 dso__set_loaded(map->dso);
1256
1257         return ret;
1258 }
1259
1260 static char *get_kernel_version(const char *root_dir)
1261 {
1262         char version[PATH_MAX];
1263         FILE *file;
1264         char *name, *tmp;
1265         const char *prefix = "Linux version ";
1266
1267         sprintf(version, "%s/proc/version", root_dir);
1268         file = fopen(version, "r");
1269         if (!file)
1270                 return NULL;
1271
1272         tmp = fgets(version, sizeof(version), file);
1273         fclose(file);
1274         if (!tmp)
1275                 return NULL;
1276
1277         name = strstr(version, prefix);
1278         if (!name)
1279                 return NULL;
1280         name += strlen(prefix);
1281         tmp = strchr(name, ' ');
1282         if (tmp)
1283                 *tmp = '\0';
1284
1285         return strdup(name);
1286 }
1287
1288 static bool is_kmod_dso(struct dso *dso)
1289 {
1290         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1291                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1292 }
1293
1294 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
1295                                        struct kmod_path *m)
1296 {
1297         char *long_name;
1298         struct map *map = map_groups__find_by_name(mg, m->name);
1299
1300         if (map == NULL)
1301                 return 0;
1302
1303         long_name = strdup(path);
1304         if (long_name == NULL)
1305                 return -ENOMEM;
1306
1307         dso__set_long_name(map->dso, long_name, true);
1308         dso__kernel_module_get_build_id(map->dso, "");
1309
1310         /*
1311          * Full name could reveal us kmod compression, so
1312          * we need to update the symtab_type if needed.
1313          */
1314         if (m->comp && is_kmod_dso(map->dso)) {
1315                 map->dso->symtab_type++;
1316                 map->dso->comp = m->comp;
1317         }
1318
1319         return 0;
1320 }
1321
1322 static int map_groups__set_modules_path_dir(struct map_groups *mg,
1323                                 const char *dir_name, int depth)
1324 {
1325         struct dirent *dent;
1326         DIR *dir = opendir(dir_name);
1327         int ret = 0;
1328
1329         if (!dir) {
1330                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1331                 return -1;
1332         }
1333
1334         while ((dent = readdir(dir)) != NULL) {
1335                 char path[PATH_MAX];
1336                 struct stat st;
1337
1338                 /*sshfs might return bad dent->d_type, so we have to stat*/
1339                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1340                 if (stat(path, &st))
1341                         continue;
1342
1343                 if (S_ISDIR(st.st_mode)) {
1344                         if (!strcmp(dent->d_name, ".") ||
1345                             !strcmp(dent->d_name, ".."))
1346                                 continue;
1347
1348                         /* Do not follow top-level source and build symlinks */
1349                         if (depth == 0) {
1350                                 if (!strcmp(dent->d_name, "source") ||
1351                                     !strcmp(dent->d_name, "build"))
1352                                         continue;
1353                         }
1354
1355                         ret = map_groups__set_modules_path_dir(mg, path,
1356                                                                depth + 1);
1357                         if (ret < 0)
1358                                 goto out;
1359                 } else {
1360                         struct kmod_path m;
1361
1362                         ret = kmod_path__parse_name(&m, dent->d_name);
1363                         if (ret)
1364                                 goto out;
1365
1366                         if (m.kmod)
1367                                 ret = map_groups__set_module_path(mg, path, &m);
1368
1369                         zfree(&m.name);
1370
1371                         if (ret)
1372                                 goto out;
1373                 }
1374         }
1375
1376 out:
1377         closedir(dir);
1378         return ret;
1379 }
1380
1381 static int machine__set_modules_path(struct machine *machine)
1382 {
1383         char *version;
1384         char modules_path[PATH_MAX];
1385
1386         version = get_kernel_version(machine->root_dir);
1387         if (!version)
1388                 return -1;
1389
1390         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1391                  machine->root_dir, version);
1392         free(version);
1393
1394         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1395 }
1396 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1397                                 u64 *size __maybe_unused,
1398                                 const char *name __maybe_unused)
1399 {
1400         return 0;
1401 }
1402
1403 static int machine__create_module(void *arg, const char *name, u64 start,
1404                                   u64 size)
1405 {
1406         struct machine *machine = arg;
1407         struct map *map;
1408
1409         if (arch__fix_module_text_start(&start, &size, name) < 0)
1410                 return -1;
1411
1412         map = machine__findnew_module_map(machine, start, name);
1413         if (map == NULL)
1414                 return -1;
1415         map->end = start + size;
1416
1417         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1418
1419         return 0;
1420 }
1421
1422 static int machine__create_modules(struct machine *machine)
1423 {
1424         const char *modules;
1425         char path[PATH_MAX];
1426
1427         if (machine__is_default_guest(machine)) {
1428                 modules = symbol_conf.default_guest_modules;
1429         } else {
1430                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1431                 modules = path;
1432         }
1433
1434         if (symbol__restricted_filename(modules, "/proc/modules"))
1435                 return -1;
1436
1437         if (modules__parse(modules, machine, machine__create_module))
1438                 return -1;
1439
1440         if (!machine__set_modules_path(machine))
1441                 return 0;
1442
1443         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1444
1445         return 0;
1446 }
1447
1448 static void machine__set_kernel_mmap(struct machine *machine,
1449                                      u64 start, u64 end)
1450 {
1451         machine->vmlinux_map->start = start;
1452         machine->vmlinux_map->end   = end;
1453         /*
1454          * Be a bit paranoid here, some perf.data file came with
1455          * a zero sized synthesized MMAP event for the kernel.
1456          */
1457         if (start == 0 && end == 0)
1458                 machine->vmlinux_map->end = ~0ULL;
1459 }
1460
1461 static void machine__update_kernel_mmap(struct machine *machine,
1462                                      u64 start, u64 end)
1463 {
1464         struct map *map = machine__kernel_map(machine);
1465
1466         map__get(map);
1467         map_groups__remove(&machine->kmaps, map);
1468
1469         machine__set_kernel_mmap(machine, start, end);
1470
1471         map_groups__insert(&machine->kmaps, map);
1472         map__put(map);
1473 }
1474
1475 int machine__create_kernel_maps(struct machine *machine)
1476 {
1477         struct dso *kernel = machine__get_kernel(machine);
1478         const char *name = NULL;
1479         struct map *map;
1480         u64 start = 0, end = ~0ULL;
1481         int ret;
1482
1483         if (kernel == NULL)
1484                 return -1;
1485
1486         ret = __machine__create_kernel_maps(machine, kernel);
1487         if (ret < 0)
1488                 goto out_put;
1489
1490         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1491                 if (machine__is_host(machine))
1492                         pr_debug("Problems creating module maps, "
1493                                  "continuing anyway...\n");
1494                 else
1495                         pr_debug("Problems creating module maps for guest %d, "
1496                                  "continuing anyway...\n", machine->pid);
1497         }
1498
1499         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1500                 if (name &&
1501                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1502                         machine__destroy_kernel_maps(machine);
1503                         ret = -1;
1504                         goto out_put;
1505                 }
1506
1507                 /*
1508                  * we have a real start address now, so re-order the kmaps
1509                  * assume it's the last in the kmaps
1510                  */
1511                 machine__update_kernel_mmap(machine, start, end);
1512         }
1513
1514         if (machine__create_extra_kernel_maps(machine, kernel))
1515                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1516
1517         if (end == ~0ULL) {
1518                 /* update end address of the kernel map using adjacent module address */
1519                 map = map__next(machine__kernel_map(machine));
1520                 if (map)
1521                         machine__set_kernel_mmap(machine, start, map->start);
1522         }
1523
1524 out_put:
1525         dso__put(kernel);
1526         return ret;
1527 }
1528
1529 static bool machine__uses_kcore(struct machine *machine)
1530 {
1531         struct dso *dso;
1532
1533         list_for_each_entry(dso, &machine->dsos.head, node) {
1534                 if (dso__is_kcore(dso))
1535                         return true;
1536         }
1537
1538         return false;
1539 }
1540
1541 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1542                                              union perf_event *event)
1543 {
1544         return machine__is(machine, "x86_64") &&
1545                is_entry_trampoline(event->mmap.filename);
1546 }
1547
1548 static int machine__process_extra_kernel_map(struct machine *machine,
1549                                              union perf_event *event)
1550 {
1551         struct dso *kernel = machine__kernel_dso(machine);
1552         struct extra_kernel_map xm = {
1553                 .start = event->mmap.start,
1554                 .end   = event->mmap.start + event->mmap.len,
1555                 .pgoff = event->mmap.pgoff,
1556         };
1557
1558         if (kernel == NULL)
1559                 return -1;
1560
1561         strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1562
1563         return machine__create_extra_kernel_map(machine, kernel, &xm);
1564 }
1565
1566 static int machine__process_kernel_mmap_event(struct machine *machine,
1567                                               union perf_event *event)
1568 {
1569         struct map *map;
1570         enum dso_kernel_type kernel_type;
1571         bool is_kernel_mmap;
1572
1573         /* If we have maps from kcore then we do not need or want any others */
1574         if (machine__uses_kcore(machine))
1575                 return 0;
1576
1577         if (machine__is_host(machine))
1578                 kernel_type = DSO_TYPE_KERNEL;
1579         else
1580                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1581
1582         is_kernel_mmap = memcmp(event->mmap.filename,
1583                                 machine->mmap_name,
1584                                 strlen(machine->mmap_name) - 1) == 0;
1585         if (event->mmap.filename[0] == '/' ||
1586             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1587                 map = machine__findnew_module_map(machine, event->mmap.start,
1588                                                   event->mmap.filename);
1589                 if (map == NULL)
1590                         goto out_problem;
1591
1592                 map->end = map->start + event->mmap.len;
1593         } else if (is_kernel_mmap) {
1594                 const char *symbol_name = (event->mmap.filename +
1595                                 strlen(machine->mmap_name));
1596                 /*
1597                  * Should be there already, from the build-id table in
1598                  * the header.
1599                  */
1600                 struct dso *kernel = NULL;
1601                 struct dso *dso;
1602
1603                 down_read(&machine->dsos.lock);
1604
1605                 list_for_each_entry(dso, &machine->dsos.head, node) {
1606
1607                         /*
1608                          * The cpumode passed to is_kernel_module is not the
1609                          * cpumode of *this* event. If we insist on passing
1610                          * correct cpumode to is_kernel_module, we should
1611                          * record the cpumode when we adding this dso to the
1612                          * linked list.
1613                          *
1614                          * However we don't really need passing correct
1615                          * cpumode.  We know the correct cpumode must be kernel
1616                          * mode (if not, we should not link it onto kernel_dsos
1617                          * list).
1618                          *
1619                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1620                          * is_kernel_module() treats it as a kernel cpumode.
1621                          */
1622
1623                         if (!dso->kernel ||
1624                             is_kernel_module(dso->long_name,
1625                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1626                                 continue;
1627
1628
1629                         kernel = dso;
1630                         break;
1631                 }
1632
1633                 up_read(&machine->dsos.lock);
1634
1635                 if (kernel == NULL)
1636                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1637                 if (kernel == NULL)
1638                         goto out_problem;
1639
1640                 kernel->kernel = kernel_type;
1641                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1642                         dso__put(kernel);
1643                         goto out_problem;
1644                 }
1645
1646                 if (strstr(kernel->long_name, "vmlinux"))
1647                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1648
1649                 machine__update_kernel_mmap(machine, event->mmap.start,
1650                                          event->mmap.start + event->mmap.len);
1651
1652                 /*
1653                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1654                  * symbol. Effectively having zero here means that at record
1655                  * time /proc/sys/kernel/kptr_restrict was non zero.
1656                  */
1657                 if (event->mmap.pgoff != 0) {
1658                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1659                                                         symbol_name,
1660                                                         event->mmap.pgoff);
1661                 }
1662
1663                 if (machine__is_default_guest(machine)) {
1664                         /*
1665                          * preload dso of guest kernel and modules
1666                          */
1667                         dso__load(kernel, machine__kernel_map(machine));
1668                 }
1669         } else if (perf_event__is_extra_kernel_mmap(machine, event)) {
1670                 return machine__process_extra_kernel_map(machine, event);
1671         }
1672         return 0;
1673 out_problem:
1674         return -1;
1675 }
1676
1677 int machine__process_mmap2_event(struct machine *machine,
1678                                  union perf_event *event,
1679                                  struct perf_sample *sample)
1680 {
1681         struct thread *thread;
1682         struct map *map;
1683         int ret = 0;
1684
1685         if (dump_trace)
1686                 perf_event__fprintf_mmap2(event, stdout);
1687
1688         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1689             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1690                 ret = machine__process_kernel_mmap_event(machine, event);
1691                 if (ret < 0)
1692                         goto out_problem;
1693                 return 0;
1694         }
1695
1696         thread = machine__findnew_thread(machine, event->mmap2.pid,
1697                                         event->mmap2.tid);
1698         if (thread == NULL)
1699                 goto out_problem;
1700
1701         map = map__new(machine, event->mmap2.start,
1702                         event->mmap2.len, event->mmap2.pgoff,
1703                         event->mmap2.maj,
1704                         event->mmap2.min, event->mmap2.ino,
1705                         event->mmap2.ino_generation,
1706                         event->mmap2.prot,
1707                         event->mmap2.flags,
1708                         event->mmap2.filename, thread);
1709
1710         if (map == NULL)
1711                 goto out_problem_map;
1712
1713         ret = thread__insert_map(thread, map);
1714         if (ret)
1715                 goto out_problem_insert;
1716
1717         thread__put(thread);
1718         map__put(map);
1719         return 0;
1720
1721 out_problem_insert:
1722         map__put(map);
1723 out_problem_map:
1724         thread__put(thread);
1725 out_problem:
1726         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1727         return 0;
1728 }
1729
1730 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1731                                 struct perf_sample *sample)
1732 {
1733         struct thread *thread;
1734         struct map *map;
1735         u32 prot = 0;
1736         int ret = 0;
1737
1738         if (dump_trace)
1739                 perf_event__fprintf_mmap(event, stdout);
1740
1741         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1742             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1743                 ret = machine__process_kernel_mmap_event(machine, event);
1744                 if (ret < 0)
1745                         goto out_problem;
1746                 return 0;
1747         }
1748
1749         thread = machine__findnew_thread(machine, event->mmap.pid,
1750                                          event->mmap.tid);
1751         if (thread == NULL)
1752                 goto out_problem;
1753
1754         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1755                 prot = PROT_EXEC;
1756
1757         map = map__new(machine, event->mmap.start,
1758                         event->mmap.len, event->mmap.pgoff,
1759                         0, 0, 0, 0, prot, 0,
1760                         event->mmap.filename,
1761                         thread);
1762
1763         if (map == NULL)
1764                 goto out_problem_map;
1765
1766         ret = thread__insert_map(thread, map);
1767         if (ret)
1768                 goto out_problem_insert;
1769
1770         thread__put(thread);
1771         map__put(map);
1772         return 0;
1773
1774 out_problem_insert:
1775         map__put(map);
1776 out_problem_map:
1777         thread__put(thread);
1778 out_problem:
1779         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1780         return 0;
1781 }
1782
1783 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1784 {
1785         struct threads *threads = machine__threads(machine, th->tid);
1786
1787         if (threads->last_match == th)
1788                 threads__set_last_match(threads, NULL);
1789
1790         if (lock)
1791                 down_write(&threads->lock);
1792
1793         BUG_ON(refcount_read(&th->refcnt) == 0);
1794
1795         rb_erase_cached(&th->rb_node, &threads->entries);
1796         RB_CLEAR_NODE(&th->rb_node);
1797         --threads->nr;
1798         /*
1799          * Move it first to the dead_threads list, then drop the reference,
1800          * if this is the last reference, then the thread__delete destructor
1801          * will be called and we will remove it from the dead_threads list.
1802          */
1803         list_add_tail(&th->node, &threads->dead);
1804
1805         /*
1806          * We need to do the put here because if this is the last refcount,
1807          * then we will be touching the threads->dead head when removing the
1808          * thread.
1809          */
1810         thread__put(th);
1811
1812         if (lock)
1813                 up_write(&threads->lock);
1814 }
1815
1816 void machine__remove_thread(struct machine *machine, struct thread *th)
1817 {
1818         return __machine__remove_thread(machine, th, true);
1819 }
1820
1821 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1822                                 struct perf_sample *sample)
1823 {
1824         struct thread *thread = machine__find_thread(machine,
1825                                                      event->fork.pid,
1826                                                      event->fork.tid);
1827         struct thread *parent = machine__findnew_thread(machine,
1828                                                         event->fork.ppid,
1829                                                         event->fork.ptid);
1830         bool do_maps_clone = true;
1831         int err = 0;
1832
1833         if (dump_trace)
1834                 perf_event__fprintf_task(event, stdout);
1835
1836         /*
1837          * There may be an existing thread that is not actually the parent,
1838          * either because we are processing events out of order, or because the
1839          * (fork) event that would have removed the thread was lost. Assume the
1840          * latter case and continue on as best we can.
1841          */
1842         if (parent->pid_ != (pid_t)event->fork.ppid) {
1843                 dump_printf("removing erroneous parent thread %d/%d\n",
1844                             parent->pid_, parent->tid);
1845                 machine__remove_thread(machine, parent);
1846                 thread__put(parent);
1847                 parent = machine__findnew_thread(machine, event->fork.ppid,
1848                                                  event->fork.ptid);
1849         }
1850
1851         /* if a thread currently exists for the thread id remove it */
1852         if (thread != NULL) {
1853                 machine__remove_thread(machine, thread);
1854                 thread__put(thread);
1855         }
1856
1857         thread = machine__findnew_thread(machine, event->fork.pid,
1858                                          event->fork.tid);
1859         /*
1860          * When synthesizing FORK events, we are trying to create thread
1861          * objects for the already running tasks on the machine.
1862          *
1863          * Normally, for a kernel FORK event, we want to clone the parent's
1864          * maps because that is what the kernel just did.
1865          *
1866          * But when synthesizing, this should not be done.  If we do, we end up
1867          * with overlapping maps as we process the sythesized MMAP2 events that
1868          * get delivered shortly thereafter.
1869          *
1870          * Use the FORK event misc flags in an internal way to signal this
1871          * situation, so we can elide the map clone when appropriate.
1872          */
1873         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1874                 do_maps_clone = false;
1875
1876         if (thread == NULL || parent == NULL ||
1877             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1878                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1879                 err = -1;
1880         }
1881         thread__put(thread);
1882         thread__put(parent);
1883
1884         return err;
1885 }
1886
1887 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1888                                 struct perf_sample *sample __maybe_unused)
1889 {
1890         struct thread *thread = machine__find_thread(machine,
1891                                                      event->fork.pid,
1892                                                      event->fork.tid);
1893
1894         if (dump_trace)
1895                 perf_event__fprintf_task(event, stdout);
1896
1897         if (thread != NULL) {
1898                 thread__exited(thread);
1899                 thread__put(thread);
1900         }
1901
1902         return 0;
1903 }
1904
1905 int machine__process_event(struct machine *machine, union perf_event *event,
1906                            struct perf_sample *sample)
1907 {
1908         int ret;
1909
1910         switch (event->header.type) {
1911         case PERF_RECORD_COMM:
1912                 ret = machine__process_comm_event(machine, event, sample); break;
1913         case PERF_RECORD_MMAP:
1914                 ret = machine__process_mmap_event(machine, event, sample); break;
1915         case PERF_RECORD_NAMESPACES:
1916                 ret = machine__process_namespaces_event(machine, event, sample); break;
1917         case PERF_RECORD_MMAP2:
1918                 ret = machine__process_mmap2_event(machine, event, sample); break;
1919         case PERF_RECORD_FORK:
1920                 ret = machine__process_fork_event(machine, event, sample); break;
1921         case PERF_RECORD_EXIT:
1922                 ret = machine__process_exit_event(machine, event, sample); break;
1923         case PERF_RECORD_LOST:
1924                 ret = machine__process_lost_event(machine, event, sample); break;
1925         case PERF_RECORD_AUX:
1926                 ret = machine__process_aux_event(machine, event); break;
1927         case PERF_RECORD_ITRACE_START:
1928                 ret = machine__process_itrace_start_event(machine, event); break;
1929         case PERF_RECORD_LOST_SAMPLES:
1930                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1931         case PERF_RECORD_SWITCH:
1932         case PERF_RECORD_SWITCH_CPU_WIDE:
1933                 ret = machine__process_switch_event(machine, event); break;
1934         case PERF_RECORD_KSYMBOL:
1935                 ret = machine__process_ksymbol(machine, event, sample); break;
1936         case PERF_RECORD_BPF_EVENT:
1937                 ret = machine__process_bpf(machine, event, sample); break;
1938         default:
1939                 ret = -1;
1940                 break;
1941         }
1942
1943         return ret;
1944 }
1945
1946 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1947 {
1948         if (!regexec(regex, sym->name, 0, NULL, 0))
1949                 return 1;
1950         return 0;
1951 }
1952
1953 static void ip__resolve_ams(struct thread *thread,
1954                             struct addr_map_symbol *ams,
1955                             u64 ip)
1956 {
1957         struct addr_location al;
1958
1959         memset(&al, 0, sizeof(al));
1960         /*
1961          * We cannot use the header.misc hint to determine whether a
1962          * branch stack address is user, kernel, guest, hypervisor.
1963          * Branches may straddle the kernel/user/hypervisor boundaries.
1964          * Thus, we have to try consecutively until we find a match
1965          * or else, the symbol is unknown
1966          */
1967         thread__find_cpumode_addr_location(thread, ip, &al);
1968
1969         ams->addr = ip;
1970         ams->al_addr = al.addr;
1971         ams->ms.sym = al.sym;
1972         ams->ms.map = al.map;
1973         ams->phys_addr = 0;
1974 }
1975
1976 static void ip__resolve_data(struct thread *thread,
1977                              u8 m, struct addr_map_symbol *ams,
1978                              u64 addr, u64 phys_addr)
1979 {
1980         struct addr_location al;
1981
1982         memset(&al, 0, sizeof(al));
1983
1984         thread__find_symbol(thread, m, addr, &al);
1985
1986         ams->addr = addr;
1987         ams->al_addr = al.addr;
1988         ams->ms.sym = al.sym;
1989         ams->ms.map = al.map;
1990         ams->phys_addr = phys_addr;
1991 }
1992
1993 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1994                                      struct addr_location *al)
1995 {
1996         struct mem_info *mi = mem_info__new();
1997
1998         if (!mi)
1999                 return NULL;
2000
2001         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2002         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2003                          sample->addr, sample->phys_addr);
2004         mi->data_src.val = sample->data_src;
2005
2006         return mi;
2007 }
2008
2009 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2010 {
2011         struct map *map = ms->map;
2012         char *srcline = NULL;
2013
2014         if (!map || callchain_param.key == CCKEY_FUNCTION)
2015                 return srcline;
2016
2017         srcline = srcline__tree_find(&map->dso->srclines, ip);
2018         if (!srcline) {
2019                 bool show_sym = false;
2020                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2021
2022                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2023                                       ms->sym, show_sym, show_addr, ip);
2024                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2025         }
2026
2027         return srcline;
2028 }
2029
2030 struct iterations {
2031         int nr_loop_iter;
2032         u64 cycles;
2033 };
2034
2035 static int add_callchain_ip(struct thread *thread,
2036                             struct callchain_cursor *cursor,
2037                             struct symbol **parent,
2038                             struct addr_location *root_al,
2039                             u8 *cpumode,
2040                             u64 ip,
2041                             bool branch,
2042                             struct branch_flags *flags,
2043                             struct iterations *iter,
2044                             u64 branch_from)
2045 {
2046         struct map_symbol ms;
2047         struct addr_location al;
2048         int nr_loop_iter = 0;
2049         u64 iter_cycles = 0;
2050         const char *srcline = NULL;
2051
2052         al.filtered = 0;
2053         al.sym = NULL;
2054         if (!cpumode) {
2055                 thread__find_cpumode_addr_location(thread, ip, &al);
2056         } else {
2057                 if (ip >= PERF_CONTEXT_MAX) {
2058                         switch (ip) {
2059                         case PERF_CONTEXT_HV:
2060                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2061                                 break;
2062                         case PERF_CONTEXT_KERNEL:
2063                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2064                                 break;
2065                         case PERF_CONTEXT_USER:
2066                                 *cpumode = PERF_RECORD_MISC_USER;
2067                                 break;
2068                         default:
2069                                 pr_debug("invalid callchain context: "
2070                                          "%"PRId64"\n", (s64) ip);
2071                                 /*
2072                                  * It seems the callchain is corrupted.
2073                                  * Discard all.
2074                                  */
2075                                 callchain_cursor_reset(cursor);
2076                                 return 1;
2077                         }
2078                         return 0;
2079                 }
2080                 thread__find_symbol(thread, *cpumode, ip, &al);
2081         }
2082
2083         if (al.sym != NULL) {
2084                 if (perf_hpp_list.parent && !*parent &&
2085                     symbol__match_regex(al.sym, &parent_regex))
2086                         *parent = al.sym;
2087                 else if (have_ignore_callees && root_al &&
2088                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2089                         /* Treat this symbol as the root,
2090                            forgetting its callees. */
2091                         *root_al = al;
2092                         callchain_cursor_reset(cursor);
2093                 }
2094         }
2095
2096         if (symbol_conf.hide_unresolved && al.sym == NULL)
2097                 return 0;
2098
2099         if (iter) {
2100                 nr_loop_iter = iter->nr_loop_iter;
2101                 iter_cycles = iter->cycles;
2102         }
2103
2104         ms.map = al.map;
2105         ms.sym = al.sym;
2106         srcline = callchain_srcline(&ms, al.addr);
2107         return callchain_cursor_append(cursor, ip, &ms,
2108                                        branch, flags, nr_loop_iter,
2109                                        iter_cycles, branch_from, srcline);
2110 }
2111
2112 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2113                                            struct addr_location *al)
2114 {
2115         unsigned int i;
2116         const struct branch_stack *bs = sample->branch_stack;
2117         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2118
2119         if (!bi)
2120                 return NULL;
2121
2122         for (i = 0; i < bs->nr; i++) {
2123                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
2124                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
2125                 bi[i].flags = bs->entries[i].flags;
2126         }
2127         return bi;
2128 }
2129
2130 static void save_iterations(struct iterations *iter,
2131                             struct branch_entry *be, int nr)
2132 {
2133         int i;
2134
2135         iter->nr_loop_iter++;
2136         iter->cycles = 0;
2137
2138         for (i = 0; i < nr; i++)
2139                 iter->cycles += be[i].flags.cycles;
2140 }
2141
2142 #define CHASHSZ 127
2143 #define CHASHBITS 7
2144 #define NO_ENTRY 0xff
2145
2146 #define PERF_MAX_BRANCH_DEPTH 127
2147
2148 /* Remove loops. */
2149 static int remove_loops(struct branch_entry *l, int nr,
2150                         struct iterations *iter)
2151 {
2152         int i, j, off;
2153         unsigned char chash[CHASHSZ];
2154
2155         memset(chash, NO_ENTRY, sizeof(chash));
2156
2157         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2158
2159         for (i = 0; i < nr; i++) {
2160                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2161
2162                 /* no collision handling for now */
2163                 if (chash[h] == NO_ENTRY) {
2164                         chash[h] = i;
2165                 } else if (l[chash[h]].from == l[i].from) {
2166                         bool is_loop = true;
2167                         /* check if it is a real loop */
2168                         off = 0;
2169                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2170                                 if (l[j].from != l[i + off].from) {
2171                                         is_loop = false;
2172                                         break;
2173                                 }
2174                         if (is_loop) {
2175                                 j = nr - (i + off);
2176                                 if (j > 0) {
2177                                         save_iterations(iter + i + off,
2178                                                 l + i, off);
2179
2180                                         memmove(iter + i, iter + i + off,
2181                                                 j * sizeof(*iter));
2182
2183                                         memmove(l + i, l + i + off,
2184                                                 j * sizeof(*l));
2185                                 }
2186
2187                                 nr -= off;
2188                         }
2189                 }
2190         }
2191         return nr;
2192 }
2193
2194 /*
2195  * Recolve LBR callstack chain sample
2196  * Return:
2197  * 1 on success get LBR callchain information
2198  * 0 no available LBR callchain information, should try fp
2199  * negative error code on other errors.
2200  */
2201 static int resolve_lbr_callchain_sample(struct thread *thread,
2202                                         struct callchain_cursor *cursor,
2203                                         struct perf_sample *sample,
2204                                         struct symbol **parent,
2205                                         struct addr_location *root_al,
2206                                         int max_stack)
2207 {
2208         struct ip_callchain *chain = sample->callchain;
2209         int chain_nr = min(max_stack, (int)chain->nr), i;
2210         u8 cpumode = PERF_RECORD_MISC_USER;
2211         u64 ip, branch_from = 0;
2212
2213         for (i = 0; i < chain_nr; i++) {
2214                 if (chain->ips[i] == PERF_CONTEXT_USER)
2215                         break;
2216         }
2217
2218         /* LBR only affects the user callchain */
2219         if (i != chain_nr) {
2220                 struct branch_stack *lbr_stack = sample->branch_stack;
2221                 int lbr_nr = lbr_stack->nr, j, k;
2222                 bool branch;
2223                 struct branch_flags *flags;
2224                 /*
2225                  * LBR callstack can only get user call chain.
2226                  * The mix_chain_nr is kernel call chain
2227                  * number plus LBR user call chain number.
2228                  * i is kernel call chain number,
2229                  * 1 is PERF_CONTEXT_USER,
2230                  * lbr_nr + 1 is the user call chain number.
2231                  * For details, please refer to the comments
2232                  * in callchain__printf
2233                  */
2234                 int mix_chain_nr = i + 1 + lbr_nr + 1;
2235
2236                 for (j = 0; j < mix_chain_nr; j++) {
2237                         int err;
2238                         branch = false;
2239                         flags = NULL;
2240
2241                         if (callchain_param.order == ORDER_CALLEE) {
2242                                 if (j < i + 1)
2243                                         ip = chain->ips[j];
2244                                 else if (j > i + 1) {
2245                                         k = j - i - 2;
2246                                         ip = lbr_stack->entries[k].from;
2247                                         branch = true;
2248                                         flags = &lbr_stack->entries[k].flags;
2249                                 } else {
2250                                         ip = lbr_stack->entries[0].to;
2251                                         branch = true;
2252                                         flags = &lbr_stack->entries[0].flags;
2253                                         branch_from =
2254                                                 lbr_stack->entries[0].from;
2255                                 }
2256                         } else {
2257                                 if (j < lbr_nr) {
2258                                         k = lbr_nr - j - 1;
2259                                         ip = lbr_stack->entries[k].from;
2260                                         branch = true;
2261                                         flags = &lbr_stack->entries[k].flags;
2262                                 }
2263                                 else if (j > lbr_nr)
2264                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
2265                                 else {
2266                                         ip = lbr_stack->entries[0].to;
2267                                         branch = true;
2268                                         flags = &lbr_stack->entries[0].flags;
2269                                         branch_from =
2270                                                 lbr_stack->entries[0].from;
2271                                 }
2272                         }
2273
2274                         err = add_callchain_ip(thread, cursor, parent,
2275                                                root_al, &cpumode, ip,
2276                                                branch, flags, NULL,
2277                                                branch_from);
2278                         if (err)
2279                                 return (err < 0) ? err : 0;
2280                 }
2281                 return 1;
2282         }
2283
2284         return 0;
2285 }
2286
2287 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2288                              struct callchain_cursor *cursor,
2289                              struct symbol **parent,
2290                              struct addr_location *root_al,
2291                              u8 *cpumode, int ent)
2292 {
2293         int err = 0;
2294
2295         while (--ent >= 0) {
2296                 u64 ip = chain->ips[ent];
2297
2298                 if (ip >= PERF_CONTEXT_MAX) {
2299                         err = add_callchain_ip(thread, cursor, parent,
2300                                                root_al, cpumode, ip,
2301                                                false, NULL, NULL, 0);
2302                         break;
2303                 }
2304         }
2305         return err;
2306 }
2307
2308 static int thread__resolve_callchain_sample(struct thread *thread,
2309                                             struct callchain_cursor *cursor,
2310                                             struct evsel *evsel,
2311                                             struct perf_sample *sample,
2312                                             struct symbol **parent,
2313                                             struct addr_location *root_al,
2314                                             int max_stack)
2315 {
2316         struct branch_stack *branch = sample->branch_stack;
2317         struct ip_callchain *chain = sample->callchain;
2318         int chain_nr = 0;
2319         u8 cpumode = PERF_RECORD_MISC_USER;
2320         int i, j, err, nr_entries;
2321         int skip_idx = -1;
2322         int first_call = 0;
2323
2324         if (chain)
2325                 chain_nr = chain->nr;
2326
2327         if (perf_evsel__has_branch_callstack(evsel)) {
2328                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2329                                                    root_al, max_stack);
2330                 if (err)
2331                         return (err < 0) ? err : 0;
2332         }
2333
2334         /*
2335          * Based on DWARF debug information, some architectures skip
2336          * a callchain entry saved by the kernel.
2337          */
2338         skip_idx = arch_skip_callchain_idx(thread, chain);
2339
2340         /*
2341          * Add branches to call stack for easier browsing. This gives
2342          * more context for a sample than just the callers.
2343          *
2344          * This uses individual histograms of paths compared to the
2345          * aggregated histograms the normal LBR mode uses.
2346          *
2347          * Limitations for now:
2348          * - No extra filters
2349          * - No annotations (should annotate somehow)
2350          */
2351
2352         if (branch && callchain_param.branch_callstack) {
2353                 int nr = min(max_stack, (int)branch->nr);
2354                 struct branch_entry be[nr];
2355                 struct iterations iter[nr];
2356
2357                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2358                         pr_warning("corrupted branch chain. skipping...\n");
2359                         goto check_calls;
2360                 }
2361
2362                 for (i = 0; i < nr; i++) {
2363                         if (callchain_param.order == ORDER_CALLEE) {
2364                                 be[i] = branch->entries[i];
2365
2366                                 if (chain == NULL)
2367                                         continue;
2368
2369                                 /*
2370                                  * Check for overlap into the callchain.
2371                                  * The return address is one off compared to
2372                                  * the branch entry. To adjust for this
2373                                  * assume the calling instruction is not longer
2374                                  * than 8 bytes.
2375                                  */
2376                                 if (i == skip_idx ||
2377                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2378                                         first_call++;
2379                                 else if (be[i].from < chain->ips[first_call] &&
2380                                     be[i].from >= chain->ips[first_call] - 8)
2381                                         first_call++;
2382                         } else
2383                                 be[i] = branch->entries[branch->nr - i - 1];
2384                 }
2385
2386                 memset(iter, 0, sizeof(struct iterations) * nr);
2387                 nr = remove_loops(be, nr, iter);
2388
2389                 for (i = 0; i < nr; i++) {
2390                         err = add_callchain_ip(thread, cursor, parent,
2391                                                root_al,
2392                                                NULL, be[i].to,
2393                                                true, &be[i].flags,
2394                                                NULL, be[i].from);
2395
2396                         if (!err)
2397                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2398                                                        NULL, be[i].from,
2399                                                        true, &be[i].flags,
2400                                                        &iter[i], 0);
2401                         if (err == -EINVAL)
2402                                 break;
2403                         if (err)
2404                                 return err;
2405                 }
2406
2407                 if (chain_nr == 0)
2408                         return 0;
2409
2410                 chain_nr -= nr;
2411         }
2412
2413 check_calls:
2414         if (callchain_param.order != ORDER_CALLEE) {
2415                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2416                                         &cpumode, chain->nr - first_call);
2417                 if (err)
2418                         return (err < 0) ? err : 0;
2419         }
2420         for (i = first_call, nr_entries = 0;
2421              i < chain_nr && nr_entries < max_stack; i++) {
2422                 u64 ip;
2423
2424                 if (callchain_param.order == ORDER_CALLEE)
2425                         j = i;
2426                 else
2427                         j = chain->nr - i - 1;
2428
2429 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2430                 if (j == skip_idx)
2431                         continue;
2432 #endif
2433                 ip = chain->ips[j];
2434                 if (ip < PERF_CONTEXT_MAX)
2435                        ++nr_entries;
2436                 else if (callchain_param.order != ORDER_CALLEE) {
2437                         err = find_prev_cpumode(chain, thread, cursor, parent,
2438                                                 root_al, &cpumode, j);
2439                         if (err)
2440                                 return (err < 0) ? err : 0;
2441                         continue;
2442                 }
2443
2444                 err = add_callchain_ip(thread, cursor, parent,
2445                                        root_al, &cpumode, ip,
2446                                        false, NULL, NULL, 0);
2447
2448                 if (err)
2449                         return (err < 0) ? err : 0;
2450         }
2451
2452         return 0;
2453 }
2454
2455 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2456 {
2457         struct symbol *sym = ms->sym;
2458         struct map *map = ms->map;
2459         struct inline_node *inline_node;
2460         struct inline_list *ilist;
2461         u64 addr;
2462         int ret = 1;
2463
2464         if (!symbol_conf.inline_name || !map || !sym)
2465                 return ret;
2466
2467         addr = map__map_ip(map, ip);
2468         addr = map__rip_2objdump(map, addr);
2469
2470         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2471         if (!inline_node) {
2472                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2473                 if (!inline_node)
2474                         return ret;
2475                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2476         }
2477
2478         list_for_each_entry(ilist, &inline_node->val, list) {
2479                 struct map_symbol ilist_ms = {
2480                         .map = map,
2481                         .sym = ilist->symbol,
2482                 };
2483                 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2484                                               NULL, 0, 0, 0, ilist->srcline);
2485
2486                 if (ret != 0)
2487                         return ret;
2488         }
2489
2490         return ret;
2491 }
2492
2493 static int unwind_entry(struct unwind_entry *entry, void *arg)
2494 {
2495         struct callchain_cursor *cursor = arg;
2496         const char *srcline = NULL;
2497         u64 addr = entry->ip;
2498
2499         if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2500                 return 0;
2501
2502         if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2503                 return 0;
2504
2505         /*
2506          * Convert entry->ip from a virtual address to an offset in
2507          * its corresponding binary.
2508          */
2509         if (entry->ms.map)
2510                 addr = map__map_ip(entry->ms.map, entry->ip);
2511
2512         srcline = callchain_srcline(&entry->ms, addr);
2513         return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2514                                        false, NULL, 0, 0, 0, srcline);
2515 }
2516
2517 static int thread__resolve_callchain_unwind(struct thread *thread,
2518                                             struct callchain_cursor *cursor,
2519                                             struct evsel *evsel,
2520                                             struct perf_sample *sample,
2521                                             int max_stack)
2522 {
2523         /* Can we do dwarf post unwind? */
2524         if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2525               (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2526                 return 0;
2527
2528         /* Bail out if nothing was captured. */
2529         if ((!sample->user_regs.regs) ||
2530             (!sample->user_stack.size))
2531                 return 0;
2532
2533         return unwind__get_entries(unwind_entry, cursor,
2534                                    thread, sample, max_stack);
2535 }
2536
2537 int thread__resolve_callchain(struct thread *thread,
2538                               struct callchain_cursor *cursor,
2539                               struct evsel *evsel,
2540                               struct perf_sample *sample,
2541                               struct symbol **parent,
2542                               struct addr_location *root_al,
2543                               int max_stack)
2544 {
2545         int ret = 0;
2546
2547         callchain_cursor_reset(cursor);
2548
2549         if (callchain_param.order == ORDER_CALLEE) {
2550                 ret = thread__resolve_callchain_sample(thread, cursor,
2551                                                        evsel, sample,
2552                                                        parent, root_al,
2553                                                        max_stack);
2554                 if (ret)
2555                         return ret;
2556                 ret = thread__resolve_callchain_unwind(thread, cursor,
2557                                                        evsel, sample,
2558                                                        max_stack);
2559         } else {
2560                 ret = thread__resolve_callchain_unwind(thread, cursor,
2561                                                        evsel, sample,
2562                                                        max_stack);
2563                 if (ret)
2564                         return ret;
2565                 ret = thread__resolve_callchain_sample(thread, cursor,
2566                                                        evsel, sample,
2567                                                        parent, root_al,
2568                                                        max_stack);
2569         }
2570
2571         return ret;
2572 }
2573
2574 int machine__for_each_thread(struct machine *machine,
2575                              int (*fn)(struct thread *thread, void *p),
2576                              void *priv)
2577 {
2578         struct threads *threads;
2579         struct rb_node *nd;
2580         struct thread *thread;
2581         int rc = 0;
2582         int i;
2583
2584         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2585                 threads = &machine->threads[i];
2586                 for (nd = rb_first_cached(&threads->entries); nd;
2587                      nd = rb_next(nd)) {
2588                         thread = rb_entry(nd, struct thread, rb_node);
2589                         rc = fn(thread, priv);
2590                         if (rc != 0)
2591                                 return rc;
2592                 }
2593
2594                 list_for_each_entry(thread, &threads->dead, node) {
2595                         rc = fn(thread, priv);
2596                         if (rc != 0)
2597                                 return rc;
2598                 }
2599         }
2600         return rc;
2601 }
2602
2603 int machines__for_each_thread(struct machines *machines,
2604                               int (*fn)(struct thread *thread, void *p),
2605                               void *priv)
2606 {
2607         struct rb_node *nd;
2608         int rc = 0;
2609
2610         rc = machine__for_each_thread(&machines->host, fn, priv);
2611         if (rc != 0)
2612                 return rc;
2613
2614         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2615                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2616
2617                 rc = machine__for_each_thread(machine, fn, priv);
2618                 if (rc != 0)
2619                         return rc;
2620         }
2621         return rc;
2622 }
2623
2624 pid_t machine__get_current_tid(struct machine *machine, int cpu)
2625 {
2626         int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2627
2628         if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
2629                 return -1;
2630
2631         return machine->current_tid[cpu];
2632 }
2633
2634 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
2635                              pid_t tid)
2636 {
2637         struct thread *thread;
2638         int nr_cpus = min(machine->env->nr_cpus_online, MAX_NR_CPUS);
2639
2640         if (cpu < 0)
2641                 return -EINVAL;
2642
2643         if (!machine->current_tid) {
2644                 int i;
2645
2646                 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
2647                 if (!machine->current_tid)
2648                         return -ENOMEM;
2649                 for (i = 0; i < nr_cpus; i++)
2650                         machine->current_tid[i] = -1;
2651         }
2652
2653         if (cpu >= nr_cpus) {
2654                 pr_err("Requested CPU %d too large. ", cpu);
2655                 pr_err("Consider raising MAX_NR_CPUS\n");
2656                 return -EINVAL;
2657         }
2658
2659         machine->current_tid[cpu] = tid;
2660
2661         thread = machine__findnew_thread(machine, pid, tid);
2662         if (!thread)
2663                 return -ENOMEM;
2664
2665         thread->cpu = cpu;
2666         thread__put(thread);
2667
2668         return 0;
2669 }
2670
2671 /*
2672  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
2673  * normalized arch is needed.
2674  */
2675 bool machine__is(struct machine *machine, const char *arch)
2676 {
2677         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
2678 }
2679
2680 int machine__nr_cpus_avail(struct machine *machine)
2681 {
2682         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
2683 }
2684
2685 int machine__get_kernel_start(struct machine *machine)
2686 {
2687         struct map *map = machine__kernel_map(machine);
2688         int err = 0;
2689
2690         /*
2691          * The only addresses above 2^63 are kernel addresses of a 64-bit
2692          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2693          * all addresses including kernel addresses are less than 2^32.  In
2694          * that case (32-bit system), if the kernel mapping is unknown, all
2695          * addresses will be assumed to be in user space - see
2696          * machine__kernel_ip().
2697          */
2698         machine->kernel_start = 1ULL << 63;
2699         if (map) {
2700                 err = map__load(map);
2701                 /*
2702                  * On x86_64, PTI entry trampolines are less than the
2703                  * start of kernel text, but still above 2^63. So leave
2704                  * kernel_start = 1ULL << 63 for x86_64.
2705                  */
2706                 if (!err && !machine__is(machine, "x86_64"))
2707                         machine->kernel_start = map->start;
2708         }
2709         return err;
2710 }
2711
2712 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
2713 {
2714         u8 addr_cpumode = cpumode;
2715         bool kernel_ip;
2716
2717         if (!machine->single_address_space)
2718                 goto out;
2719
2720         kernel_ip = machine__kernel_ip(machine, addr);
2721         switch (cpumode) {
2722         case PERF_RECORD_MISC_KERNEL:
2723         case PERF_RECORD_MISC_USER:
2724                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
2725                                            PERF_RECORD_MISC_USER;
2726                 break;
2727         case PERF_RECORD_MISC_GUEST_KERNEL:
2728         case PERF_RECORD_MISC_GUEST_USER:
2729                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
2730                                            PERF_RECORD_MISC_GUEST_USER;
2731                 break;
2732         default:
2733                 break;
2734         }
2735 out:
2736         return addr_cpumode;
2737 }
2738
2739 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2740 {
2741         return dsos__findnew(&machine->dsos, filename);
2742 }
2743
2744 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2745 {
2746         struct machine *machine = vmachine;
2747         struct map *map;
2748         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
2749
2750         if (sym == NULL)
2751                 return NULL;
2752
2753         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2754         *addrp = map->unmap_ip(map, sym->start);
2755         return sym->name;
2756 }