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 "mem-info.h"
  20 #include "path.h"
  21 #include "srcline.h"
  22 #include "symbol.h"
  23 #include "sort.h"
  24 #include "strlist.h"
  25 #include "target.h"
  26 #include "thread.h"
  27 #include "util.h"
  28 #include "vdso.h"
  29 #include <stdbool.h>
  30 #include <sys/types.h>
  31 #include <sys/stat.h>
  32 #include <unistd.h>
  33 #include "unwind.h"
  34 #include "linux/hash.h"
  35 #include "asm/bug.h"
  36 #include "bpf-event.h"
  37 #include <internal/lib.h> // page_size
  38 #include "cgroup.h"
  39 #include "arm64-frame-pointer-unwind-support.h"
  40
  41 #include <linux/ctype.h>
  42 #include <symbol/kallsyms.h>
  43 #include <linux/mman.h>
  44 #include <linux/string.h>
  45 #include <linux/zalloc.h>
  46
  47 static struct dso *machine__kernel_dso(struct machine *machine)
  48 {
  49         return map__dso(machine->vmlinux_map);
  50 }
  51
  52 static int machine__set_mmap_name(struct machine *machine)
  53 {
  54         if (machine__is_host(machine))
  55                 machine->mmap_name = strdup("[kernel.kallsyms]");
  56         else if (machine__is_default_guest(machine))
  57                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  58         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  59                           machine->pid) < 0)
  60                 machine->mmap_name = NULL;
  61
  62         return machine->mmap_name ? 0 : -ENOMEM;
  63 }
  64
  65 static void thread__set_guest_comm(struct thread *thread, pid_t pid)
  66 {
  67         char comm[64];
  68
  69         snprintf(comm, sizeof(comm), "[guest/%d]", pid);
  70         thread__set_comm(thread, comm, 0);
  71 }
  72
  73 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  74 {
  75         int err = -ENOMEM;
  76
  77         memset(machine, 0, sizeof(*machine));
  78         machine->kmaps = maps__new(machine);
  79         if (machine->kmaps == NULL)
  80                 return -ENOMEM;
  81
  82         RB_CLEAR_NODE(&machine->rb_node);
  83         dsos__init(&machine->dsos);
  84
  85         threads__init(&machine->threads);
  86
  87         machine->vdso_info = NULL;
  88         machine->env = NULL;
  89
  90         machine->pid = pid;
  91
  92         machine->id_hdr_size = 0;
  93         machine->kptr_restrict_warned = false;
  94         machine->comm_exec = false;
  95         machine->kernel_start = 0;
  96         machine->vmlinux_map = NULL;
  97
  98         machine->root_dir = strdup(root_dir);
  99         if (machine->root_dir == NULL)
 100                 goto out;
 101
 102         if (machine__set_mmap_name(machine))
 103                 goto out;
 104
 105         if (pid != HOST_KERNEL_ID) {
 106                 struct thread *thread = machine__findnew_thread(machine, -1,
 107                                                                 pid);
 108
 109                 if (thread == NULL)
 110                         goto out;
 111
 112                 thread__set_guest_comm(thread, pid);
 113                 thread__put(thread);
 114         }
 115
 116         machine->current_tid = NULL;
 117         err = 0;
 118
 119 out:
 120         if (err) {
 121                 zfree(&machine->kmaps);
 122                 zfree(&machine->root_dir);
 123                 zfree(&machine->mmap_name);
 124         }
 125         return 0;
 126 }
 127
 128 struct machine *machine__new_host(void)
 129 {
 130         struct machine *machine = malloc(sizeof(*machine));
 131
 132         if (machine != NULL) {
 133                 machine__init(machine, "", HOST_KERNEL_ID);
 134
 135                 if (machine__create_kernel_maps(machine) < 0)
 136                         goto out_delete;
 137         }
 138
 139         return machine;
 140 out_delete:
 141         free(machine);
 142         return NULL;
 143 }
 144
 145 struct machine *machine__new_kallsyms(void)
 146 {
 147         struct machine *machine = machine__new_host();
 148         /*
 149          * FIXME:
 150          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 151          *    ask for not using the kcore parsing code, once this one is fixed
 152          *    to create a map per module.
 153          */
 154         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 155                 machine__delete(machine);
 156                 machine = NULL;
 157         }
 158
 159         return machine;
 160 }
 161
 162 void machine__delete_threads(struct machine *machine)
 163 {
 164         threads__remove_all_threads(&machine->threads);
 165 }
 166
 167 void machine__exit(struct machine *machine)
 168 {
 169         if (machine == NULL)
 170                 return;
 171
 172         machine__destroy_kernel_maps(machine);
 173         maps__zput(machine->kmaps);
 174         dsos__exit(&machine->dsos);
 175         machine__exit_vdso(machine);
 176         zfree(&machine->root_dir);
 177         zfree(&machine->mmap_name);
 178         zfree(&machine->current_tid);
 179         zfree(&machine->kallsyms_filename);
 180
 181         threads__exit(&machine->threads);
 182 }
 183
 184 void machine__delete(struct machine *machine)
 185 {
 186         if (machine) {
 187                 machine__exit(machine);
 188                 free(machine);
 189         }
 190 }
 191
 192 void machines__init(struct machines *machines)
 193 {
 194         machine__init(&machines->host, "", HOST_KERNEL_ID);
 195         machines->guests = RB_ROOT_CACHED;
 196 }
 197
 198 void machines__exit(struct machines *machines)
 199 {
 200         machine__exit(&machines->host);
 201         /* XXX exit guest */
 202 }
 203
 204 struct machine *machines__add(struct machines *machines, pid_t pid,
 205                               const char *root_dir)
 206 {
 207         struct rb_node **p = &machines->guests.rb_root.rb_node;
 208         struct rb_node *parent = NULL;
 209         struct machine *pos, *machine = malloc(sizeof(*machine));
 210         bool leftmost = true;
 211
 212         if (machine == NULL)
 213                 return NULL;
 214
 215         if (machine__init(machine, root_dir, pid) != 0) {
 216                 free(machine);
 217                 return NULL;
 218         }
 219
 220         while (*p != NULL) {
 221                 parent = *p;
 222                 pos = rb_entry(parent, struct machine, rb_node);
 223                 if (pid < pos->pid)
 224                         p = &(*p)->rb_left;
 225                 else {
 226                         p = &(*p)->rb_right;
 227                         leftmost = false;
 228                 }
 229         }
 230
 231         rb_link_node(&machine->rb_node, parent, p);
 232         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 233
 234         machine->machines = machines;
 235
 236         return machine;
 237 }
 238
 239 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 240 {
 241         struct rb_node *nd;
 242
 243         machines->host.comm_exec = comm_exec;
 244
 245         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 246                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 247
 248                 machine->comm_exec = comm_exec;
 249         }
 250 }
 251
 252 struct machine *machines__find(struct machines *machines, pid_t pid)
 253 {
 254         struct rb_node **p = &machines->guests.rb_root.rb_node;
 255         struct rb_node *parent = NULL;
 256         struct machine *machine;
 257         struct machine *default_machine = NULL;
 258
 259         if (pid == HOST_KERNEL_ID)
 260                 return &machines->host;
 261
 262         while (*p != NULL) {
 263                 parent = *p;
 264                 machine = rb_entry(parent, struct machine, rb_node);
 265                 if (pid < machine->pid)
 266                         p = &(*p)->rb_left;
 267                 else if (pid > machine->pid)
 268                         p = &(*p)->rb_right;
 269                 else
 270                         return machine;
 271                 if (!machine->pid)
 272                         default_machine = machine;
 273         }
 274
 275         return default_machine;
 276 }
 277
 278 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 279 {
 280         char path[PATH_MAX];
 281         const char *root_dir = "";
 282         struct machine *machine = machines__find(machines, pid);
 283
 284         if (machine && (machine->pid == pid))
 285                 goto out;
 286
 287         if ((pid != HOST_KERNEL_ID) &&
 288             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 289             (symbol_conf.guestmount)) {
 290                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 291                 if (access(path, R_OK)) {
 292                         static struct strlist *seen;
 293
 294                         if (!seen)
 295                                 seen = strlist__new(NULL, NULL);
 296
 297                         if (!strlist__has_entry(seen, path)) {
 298                                 pr_err("Can't access file %s\n", path);
 299                                 strlist__add(seen, path);
 300                         }
 301                         machine = NULL;
 302                         goto out;
 303                 }
 304                 root_dir = path;
 305         }
 306
 307         machine = machines__add(machines, pid, root_dir);
 308 out:
 309         return machine;
 310 }
 311
 312 struct machine *machines__find_guest(struct machines *machines, pid_t pid)
 313 {
 314         struct machine *machine = machines__find(machines, pid);
 315
 316         if (!machine)
 317                 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
 318         return machine;
 319 }
 320
 321 /*
 322  * A common case for KVM test programs is that the test program acts as the
 323  * hypervisor, creating, running and destroying the virtual machine, and
 324  * providing the guest object code from its own object code. In this case,
 325  * the VM is not running an OS, but only the functions loaded into it by the
 326  * hypervisor test program, and conveniently, loaded at the same virtual
 327  * addresses.
 328  *
 329  * Normally to resolve addresses, MMAP events are needed to map addresses
 330  * back to the object code and debug symbols for that object code.
 331  *
 332  * Currently, there is no way to get such mapping information from guests
 333  * but, in the scenario described above, the guest has the same mappings
 334  * as the hypervisor, so support for that scenario can be achieved.
 335  *
 336  * To support that, copy the host thread's maps to the guest thread's maps.
 337  * Note, we do not discover the guest until we encounter a guest event,
 338  * which works well because it is not until then that we know that the host
 339  * thread's maps have been set up.
 340  *
 341  * This function returns the guest thread. Apart from keeping the data
 342  * structures sane, using a thread belonging to the guest machine, instead
 343  * of the host thread, allows it to have its own comm (refer
 344  * thread__set_guest_comm()).
 345  */
 346 static struct thread *findnew_guest_code(struct machine *machine,
 347                                          struct machine *host_machine,
 348                                          pid_t pid)
 349 {
 350         struct thread *host_thread;
 351         struct thread *thread;
 352         int err;
 353
 354         if (!machine)
 355                 return NULL;
 356
 357         thread = machine__findnew_thread(machine, -1, pid);
 358         if (!thread)
 359                 return NULL;
 360
 361         /* Assume maps are set up if there are any */
 362         if (!maps__empty(thread__maps(thread)))
 363                 return thread;
 364
 365         host_thread = machine__find_thread(host_machine, -1, pid);
 366         if (!host_thread)
 367                 goto out_err;
 368
 369         thread__set_guest_comm(thread, pid);
 370
 371         /*
 372          * Guest code can be found in hypervisor process at the same address
 373          * so copy host maps.
 374          */
 375         err = maps__copy_from(thread__maps(thread), thread__maps(host_thread));
 376         thread__put(host_thread);
 377         if (err)
 378                 goto out_err;
 379
 380         return thread;
 381
 382 out_err:
 383         thread__zput(thread);
 384         return NULL;
 385 }
 386
 387 struct thread *machines__findnew_guest_code(struct machines *machines, pid_t pid)
 388 {
 389         struct machine *host_machine = machines__find(machines, HOST_KERNEL_ID);
 390         struct machine *machine = machines__findnew(machines, pid);
 391
 392         return findnew_guest_code(machine, host_machine, pid);
 393 }
 394
 395 struct thread *machine__findnew_guest_code(struct machine *machine, pid_t pid)
 396 {
 397         struct machines *machines = machine->machines;
 398         struct machine *host_machine;
 399
 400         if (!machines)
 401                 return NULL;
 402
 403         host_machine = machines__find(machines, HOST_KERNEL_ID);
 404
 405         return findnew_guest_code(machine, host_machine, pid);
 406 }
 407
 408 void machines__process_guests(struct machines *machines,
 409                               machine__process_t process, void *data)
 410 {
 411         struct rb_node *nd;
 412
 413         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 414                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 415                 process(pos, data);
 416         }
 417 }
 418
 419 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 420 {
 421         struct rb_node *node;
 422         struct machine *machine;
 423
 424         machines->host.id_hdr_size = id_hdr_size;
 425
 426         for (node = rb_first_cached(&machines->guests); node;
 427              node = rb_next(node)) {
 428                 machine = rb_entry(node, struct machine, rb_node);
 429                 machine->id_hdr_size = id_hdr_size;
 430         }
 431
 432         return;
 433 }
 434
 435 static void machine__update_thread_pid(struct machine *machine,
 436                                        struct thread *th, pid_t pid)
 437 {
 438         struct thread *leader;
 439
 440         if (pid == thread__pid(th) || pid == -1 || thread__pid(th) != -1)
 441                 return;
 442
 443         thread__set_pid(th, pid);
 444
 445         if (thread__pid(th) == thread__tid(th))
 446                 return;
 447
 448         leader = machine__findnew_thread(machine, thread__pid(th), thread__pid(th));
 449         if (!leader)
 450                 goto out_err;
 451
 452         if (!thread__maps(leader))
 453                 thread__set_maps(leader, maps__new(machine));
 454
 455         if (!thread__maps(leader))
 456                 goto out_err;
 457
 458         if (thread__maps(th) == thread__maps(leader))
 459                 goto out_put;
 460
 461         if (thread__maps(th)) {
 462                 /*
 463                  * Maps are created from MMAP events which provide the pid and
 464                  * tid.  Consequently there never should be any maps on a thread
 465                  * with an unknown pid.  Just print an error if there are.
 466                  */
 467                 if (!maps__empty(thread__maps(th)))
 468                         pr_err("Discarding thread maps for %d:%d\n",
 469                                 thread__pid(th), thread__tid(th));
 470                 maps__put(thread__maps(th));
 471         }
 472
 473         thread__set_maps(th, maps__get(thread__maps(leader)));
 474 out_put:
 475         thread__put(leader);
 476         return;
 477 out_err:
 478         pr_err("Failed to join map groups for %d:%d\n", thread__pid(th), thread__tid(th));
 479         goto out_put;
 480 }
 481
 482 /*
 483  * Caller must eventually drop thread->refcnt returned with a successful
 484  * lookup/new thread inserted.
 485  */
 486 static struct thread *__machine__findnew_thread(struct machine *machine,
 487                                                 pid_t pid,
 488                                                 pid_t tid,
 489                                                 bool create)
 490 {
 491         struct thread *th = threads__find(&machine->threads, tid);
 492         bool created;
 493
 494         if (th) {
 495                 machine__update_thread_pid(machine, th, pid);
 496                 return th;
 497         }
 498         if (!create)
 499                 return NULL;
 500
 501         th = threads__findnew(&machine->threads, pid, tid, &created);
 502         if (created) {
 503                 /*
 504                  * We have to initialize maps separately after rb tree is
 505                  * updated.
 506                  *
 507                  * The reason is that we call machine__findnew_thread within
 508                  * thread__init_maps to find the thread leader and that would
 509                  * screwed the rb tree.
 510                  */
 511                 if (thread__init_maps(th, machine)) {
 512                         pr_err("Thread init failed thread %d\n", pid);
 513                         threads__remove(&machine->threads, th);
 514                         thread__put(th);
 515                         return NULL;
 516                 }
 517         } else
 518                 machine__update_thread_pid(machine, th, pid);
 519
 520         return th;
 521 }
 522
 523 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 524 {
 525         return __machine__findnew_thread(machine, pid, tid, /*create=*/true);
 526 }
 527
 528 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 529                                     pid_t tid)
 530 {
 531         return __machine__findnew_thread(machine, pid, tid, /*create=*/false);
 532 }
 533
 534 /*
 535  * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
 536  * So here a single thread is created for that, but actually there is a separate
 537  * idle task per cpu, so there should be one 'struct thread' per cpu, but there
 538  * is only 1. That causes problems for some tools, requiring workarounds. For
 539  * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
 540  */
 541 struct thread *machine__idle_thread(struct machine *machine)
 542 {
 543         struct thread *thread = machine__findnew_thread(machine, 0, 0);
 544
 545         if (!thread || thread__set_comm(thread, "swapper", 0) ||
 546             thread__set_namespaces(thread, 0, NULL))
 547                 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
 548
 549         return thread;
 550 }
 551
 552 struct comm *machine__thread_exec_comm(struct machine *machine,
 553                                        struct thread *thread)
 554 {
 555         if (machine->comm_exec)
 556                 return thread__exec_comm(thread);
 557         else
 558                 return thread__comm(thread);
 559 }
 560
 561 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 562                                 struct perf_sample *sample)
 563 {
 564         struct thread *thread = machine__findnew_thread(machine,
 565                                                         event->comm.pid,
 566                                                         event->comm.tid);
 567         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 568         int err = 0;
 569
 570         if (exec)
 571                 machine->comm_exec = true;
 572
 573         if (dump_trace)
 574                 perf_event__fprintf_comm(event, stdout);
 575
 576         if (thread == NULL ||
 577             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 578                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 579                 err = -1;
 580         }
 581
 582         thread__put(thread);
 583
 584         return err;
 585 }
 586
 587 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 588                                       union perf_event *event,
 589                                       struct perf_sample *sample __maybe_unused)
 590 {
 591         struct thread *thread = machine__findnew_thread(machine,
 592                                                         event->namespaces.pid,
 593                                                         event->namespaces.tid);
 594         int err = 0;
 595
 596         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 597                   "\nWARNING: kernel seems to support more namespaces than perf"
 598                   " tool.\nTry updating the perf tool..\n\n");
 599
 600         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 601                   "\nWARNING: perf tool seems to support more namespaces than"
 602                   " the kernel.\nTry updating the kernel..\n\n");
 603
 604         if (dump_trace)
 605                 perf_event__fprintf_namespaces(event, stdout);
 606
 607         if (thread == NULL ||
 608             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 609                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 610                 err = -1;
 611         }
 612
 613         thread__put(thread);
 614
 615         return err;
 616 }
 617
 618 int machine__process_cgroup_event(struct machine *machine,
 619                                   union perf_event *event,
 620                                   struct perf_sample *sample __maybe_unused)
 621 {
 622         struct cgroup *cgrp;
 623
 624         if (dump_trace)
 625                 perf_event__fprintf_cgroup(event, stdout);
 626
 627         cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
 628         if (cgrp == NULL)
 629                 return -ENOMEM;
 630
 631         return 0;
 632 }
 633
 634 int machine__process_lost_event(struct machine *machine __maybe_unused,
 635                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 636 {
 637         dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
 638                     event->lost.id, event->lost.lost);
 639         return 0;
 640 }
 641
 642 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 643                                         union perf_event *event, struct perf_sample *sample)
 644 {
 645         dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
 646                     sample->id, event->lost_samples.lost);
 647         return 0;
 648 }
 649
 650 int machine__process_aux_event(struct machine *machine __maybe_unused,
 651                                union perf_event *event)
 652 {
 653         if (dump_trace)
 654                 perf_event__fprintf_aux(event, stdout);
 655         return 0;
 656 }
 657
 658 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 659                                         union perf_event *event)
 660 {
 661         if (dump_trace)
 662                 perf_event__fprintf_itrace_start(event, stdout);
 663         return 0;
 664 }
 665
 666 int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
 667                                             union perf_event *event)
 668 {
 669         if (dump_trace)
 670                 perf_event__fprintf_aux_output_hw_id(event, stdout);
 671         return 0;
 672 }
 673
 674 int machine__process_switch_event(struct machine *machine __maybe_unused,
 675                                   union perf_event *event)
 676 {
 677         if (dump_trace)
 678                 perf_event__fprintf_switch(event, stdout);
 679         return 0;
 680 }
 681
 682 static int machine__process_ksymbol_register(struct machine *machine,
 683                                              union perf_event *event,
 684                                              struct perf_sample *sample __maybe_unused)
 685 {
 686         struct symbol *sym;
 687         struct dso *dso = NULL;
 688         struct map *map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
 689         int err = 0;
 690
 691         if (!map) {
 692                 dso = dso__new(event->ksymbol.name);
 693
 694                 if (!dso) {
 695                         err = -ENOMEM;
 696                         goto out;
 697                 }
 698                 dso__set_kernel(dso, DSO_SPACE__KERNEL);
 699                 map = map__new2(0, dso);
 700                 if (!map) {
 701                         err = -ENOMEM;
 702                         goto out;
 703                 }
 704                 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
 705                         dso__set_binary_type(dso, DSO_BINARY_TYPE__OOL);
 706                         dso__data(dso)->file_size = event->ksymbol.len;
 707                         dso__set_loaded(dso);
 708                 }
 709
 710                 map__set_start(map, event->ksymbol.addr);
 711                 map__set_end(map, map__start(map) + event->ksymbol.len);
 712                 err = maps__insert(machine__kernel_maps(machine), map);
 713                 if (err) {
 714                         err = -ENOMEM;
 715                         goto out;
 716                 }
 717
 718                 dso__set_loaded(dso);
 719
 720                 if (is_bpf_image(event->ksymbol.name)) {
 721                         dso__set_binary_type(dso, DSO_BINARY_TYPE__BPF_IMAGE);
 722                         dso__set_long_name(dso, "", false);
 723                 }
 724         } else {
 725                 dso = dso__get(map__dso(map));
 726         }
 727
 728         sym = symbol__new(map__map_ip(map, map__start(map)),
 729                           event->ksymbol.len,
 730                           0, 0, event->ksymbol.name);
 731         if (!sym) {
 732                 err = -ENOMEM;
 733                 goto out;
 734         }
 735         dso__insert_symbol(dso, sym);
 736 out:
 737         map__put(map);
 738         dso__put(dso);
 739         return err;
 740 }
 741
 742 static int machine__process_ksymbol_unregister(struct machine *machine,
 743                                                union perf_event *event,
 744                                                struct perf_sample *sample __maybe_unused)
 745 {
 746         struct symbol *sym;
 747         struct map *map;
 748
 749         map = maps__find(machine__kernel_maps(machine), event->ksymbol.addr);
 750         if (!map)
 751                 return 0;
 752
 753         if (!RC_CHK_EQUAL(map, machine->vmlinux_map))
 754                 maps__remove(machine__kernel_maps(machine), map);
 755         else {
 756                 struct dso *dso = map__dso(map);
 757
 758                 sym = dso__find_symbol(dso, map__map_ip(map, map__start(map)));
 759                 if (sym)
 760                         dso__delete_symbol(dso, sym);
 761         }
 762         map__put(map);
 763         return 0;
 764 }
 765
 766 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 767                              union perf_event *event,
 768                              struct perf_sample *sample)
 769 {
 770         if (dump_trace)
 771                 perf_event__fprintf_ksymbol(event, stdout);
 772
 773         if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 774                 return machine__process_ksymbol_unregister(machine, event,
 775                                                            sample);
 776         return machine__process_ksymbol_register(machine, event, sample);
 777 }
 778
 779 int machine__process_text_poke(struct machine *machine, union perf_event *event,
 780                                struct perf_sample *sample __maybe_unused)
 781 {
 782         struct map *map = maps__find(machine__kernel_maps(machine), event->text_poke.addr);
 783         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 784         struct dso *dso = map ? map__dso(map) : NULL;
 785
 786         if (dump_trace)
 787                 perf_event__fprintf_text_poke(event, machine, stdout);
 788
 789         if (!event->text_poke.new_len)
 790                 goto out;
 791
 792         if (cpumode != PERF_RECORD_MISC_KERNEL) {
 793                 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
 794                 goto out;
 795         }
 796
 797         if (dso) {
 798                 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
 799                 int ret;
 800
 801                 /*
 802                  * Kernel maps might be changed when loading symbols so loading
 803                  * must be done prior to using kernel maps.
 804                  */
 805                 map__load(map);
 806                 ret = dso__data_write_cache_addr(dso, map, machine,
 807                                                  event->text_poke.addr,
 808                                                  new_bytes,
 809                                                  event->text_poke.new_len);
 810                 if (ret != event->text_poke.new_len)
 811                         pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
 812                                  event->text_poke.addr);
 813         } else {
 814                 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
 815                          event->text_poke.addr);
 816         }
 817 out:
 818         map__put(map);
 819         return 0;
 820 }
 821
 822 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
 823                                               const char *filename)
 824 {
 825         struct map *map = NULL;
 826         struct kmod_path m;
 827         struct dso *dso;
 828         int err;
 829
 830         if (kmod_path__parse_name(&m, filename))
 831                 return NULL;
 832
 833         dso = dsos__findnew_module_dso(&machine->dsos, machine, &m, filename);
 834         if (dso == NULL)
 835                 goto out;
 836
 837         map = map__new2(start, dso);
 838         if (map == NULL)
 839                 goto out;
 840
 841         err = maps__insert(machine__kernel_maps(machine), map);
 842         /* If maps__insert failed, return NULL. */
 843         if (err) {
 844                 map__put(map);
 845                 map = NULL;
 846         }
 847 out:
 848         /* put the dso here, corresponding to  machine__findnew_module_dso */
 849         dso__put(dso);
 850         zfree(&m.name);
 851         return map;
 852 }
 853
 854 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 855 {
 856         struct rb_node *nd;
 857         size_t ret = dsos__fprintf(&machines->host.dsos, fp);
 858
 859         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 860                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 861                 ret += dsos__fprintf(&pos->dsos, fp);
 862         }
 863
 864         return ret;
 865 }
 866
 867 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 868                                      bool (skip)(struct dso *dso, int parm), int parm)
 869 {
 870         return dsos__fprintf_buildid(&m->dsos, fp, skip, parm);
 871 }
 872
 873 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 874                                      bool (skip)(struct dso *dso, int parm), int parm)
 875 {
 876         struct rb_node *nd;
 877         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 878
 879         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 880                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 881                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 882         }
 883         return ret;
 884 }
 885
 886 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 887 {
 888         int i;
 889         size_t printed = 0;
 890         struct dso *kdso = machine__kernel_dso(machine);
 891
 892         if (dso__has_build_id(kdso)) {
 893                 char filename[PATH_MAX];
 894
 895                 if (dso__build_id_filename(kdso, filename, sizeof(filename), false))
 896                         printed += fprintf(fp, "[0] %s\n", filename);
 897         }
 898
 899         for (i = 0; i < vmlinux_path__nr_entries; ++i) {
 900                 printed += fprintf(fp, "[%d] %s\n", i + dso__has_build_id(kdso),
 901                                    vmlinux_path[i]);
 902         }
 903         return printed;
 904 }
 905
 906 struct machine_fprintf_cb_args {
 907         FILE *fp;
 908         size_t printed;
 909 };
 910
 911 static int machine_fprintf_cb(struct thread *thread, void *data)
 912 {
 913         struct machine_fprintf_cb_args *args = data;
 914
 915         /* TODO: handle fprintf errors. */
 916         args->printed += thread__fprintf(thread, args->fp);
 917         return 0;
 918 }
 919
 920 size_t machine__fprintf(struct machine *machine, FILE *fp)
 921 {
 922         struct machine_fprintf_cb_args args = {
 923                 .fp = fp,
 924                 .printed = 0,
 925         };
 926         size_t ret = fprintf(fp, "Threads: %zu\n", threads__nr(&machine->threads));
 927
 928         machine__for_each_thread(machine, machine_fprintf_cb, &args);
 929         return ret + args.printed;
 930 }
 931
 932 static struct dso *machine__get_kernel(struct machine *machine)
 933 {
 934         const char *vmlinux_name = machine->mmap_name;
 935         struct dso *kernel;
 936
 937         if (machine__is_host(machine)) {
 938                 if (symbol_conf.vmlinux_name)
 939                         vmlinux_name = symbol_conf.vmlinux_name;
 940
 941                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 942                                                  "[kernel]", DSO_SPACE__KERNEL);
 943         } else {
 944                 if (symbol_conf.default_guest_vmlinux_name)
 945                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 946
 947                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 948                                                  "[guest.kernel]",
 949                                                  DSO_SPACE__KERNEL_GUEST);
 950         }
 951
 952         if (kernel != NULL && (!dso__has_build_id(kernel)))
 953                 dso__read_running_kernel_build_id(kernel, machine);
 954
 955         return kernel;
 956 }
 957
 958 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 959                                     size_t bufsz)
 960 {
 961         if (machine__is_default_guest(machine))
 962                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 963         else
 964                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 965 }
 966
 967 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 968
 969 /* Figure out the start address of kernel map from /proc/kallsyms.
 970  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 971  * symbol_name if it's not that important.
 972  */
 973 static int machine__get_running_kernel_start(struct machine *machine,
 974                                              const char **symbol_name,
 975                                              u64 *start, u64 *end)
 976 {
 977         char filename[PATH_MAX];
 978         int i, err = -1;
 979         const char *name;
 980         u64 addr = 0;
 981
 982         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 983
 984         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 985                 return 0;
 986
 987         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 988                 err = kallsyms__get_function_start(filename, name, &addr);
 989                 if (!err)
 990                         break;
 991         }
 992
 993         if (err)
 994                 return -1;
 995
 996         if (symbol_name)
 997                 *symbol_name = name;
 998
 999         *start = addr;
1000
1001         err = kallsyms__get_symbol_start(filename, "_edata", &addr);
1002         if (err)
1003                 err = kallsyms__get_function_start(filename, "_etext", &addr);
1004         if (!err)
1005                 *end = addr;
1006
1007         return 0;
1008 }
1009
1010 int machine__create_extra_kernel_map(struct machine *machine,
1011                                      struct dso *kernel,
1012                                      struct extra_kernel_map *xm)
1013 {
1014         struct kmap *kmap;
1015         struct map *map;
1016         int err;
1017
1018         map = map__new2(xm->start, kernel);
1019         if (!map)
1020                 return -ENOMEM;
1021
1022         map__set_end(map, xm->end);
1023         map__set_pgoff(map, xm->pgoff);
1024
1025         kmap = map__kmap(map);
1026
1027         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1028
1029         err = maps__insert(machine__kernel_maps(machine), map);
1030
1031         if (!err) {
1032                 pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1033                         kmap->name, map__start(map), map__end(map));
1034         }
1035
1036         map__put(map);
1037
1038         return err;
1039 }
1040
1041 static u64 find_entry_trampoline(struct dso *dso)
1042 {
1043         /* Duplicates are removed so lookup all aliases */
1044         const char *syms[] = {
1045                 "_entry_trampoline",
1046                 "__entry_trampoline_start",
1047                 "entry_SYSCALL_64_trampoline",
1048         };
1049         struct symbol *sym = dso__first_symbol(dso);
1050         unsigned int i;
1051
1052         for (; sym; sym = dso__next_symbol(sym)) {
1053                 if (sym->binding != STB_GLOBAL)
1054                         continue;
1055                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1056                         if (!strcmp(sym->name, syms[i]))
1057                                 return sym->start;
1058                 }
1059         }
1060
1061         return 0;
1062 }
1063
1064 /*
1065  * These values can be used for kernels that do not have symbols for the entry
1066  * trampolines in kallsyms.
1067  */
1068 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1069 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1070 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1071
1072 struct machine__map_x86_64_entry_trampolines_args {
1073         struct maps *kmaps;
1074         bool found;
1075 };
1076
1077 static int machine__map_x86_64_entry_trampolines_cb(struct map *map, void *data)
1078 {
1079         struct machine__map_x86_64_entry_trampolines_args *args = data;
1080         struct map *dest_map;
1081         struct kmap *kmap = __map__kmap(map);
1082
1083         if (!kmap || !is_entry_trampoline(kmap->name))
1084                 return 0;
1085
1086         dest_map = maps__find(args->kmaps, map__pgoff(map));
1087         if (RC_CHK_ACCESS(dest_map) != RC_CHK_ACCESS(map))
1088                 map__set_pgoff(map, map__map_ip(dest_map, map__pgoff(map)));
1089
1090         map__put(dest_map);
1091         args->found = true;
1092         return 0;
1093 }
1094
1095 /* Map x86_64 PTI entry trampolines */
1096 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1097                                           struct dso *kernel)
1098 {
1099         struct machine__map_x86_64_entry_trampolines_args args = {
1100                 .kmaps = machine__kernel_maps(machine),
1101                 .found = false,
1102         };
1103         int nr_cpus_avail, cpu;
1104         u64 pgoff;
1105
1106         /*
1107          * In the vmlinux case, pgoff is a virtual address which must now be
1108          * mapped to a vmlinux offset.
1109          */
1110         maps__for_each_map(args.kmaps, machine__map_x86_64_entry_trampolines_cb, &args);
1111
1112         if (args.found || machine->trampolines_mapped)
1113                 return 0;
1114
1115         pgoff = find_entry_trampoline(kernel);
1116         if (!pgoff)
1117                 return 0;
1118
1119         nr_cpus_avail = machine__nr_cpus_avail(machine);
1120
1121         /* Add a 1 page map for each CPU's entry trampoline */
1122         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1123                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1124                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1125                          X86_64_ENTRY_TRAMPOLINE;
1126                 struct extra_kernel_map xm = {
1127                         .start = va,
1128                         .end   = va + page_size,
1129                         .pgoff = pgoff,
1130                 };
1131
1132                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1133
1134                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1135                         return -1;
1136         }
1137
1138         machine->trampolines_mapped = nr_cpus_avail;
1139
1140         return 0;
1141 }
1142
1143 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1144                                              struct dso *kernel __maybe_unused)
1145 {
1146         return 0;
1147 }
1148
1149 static int
1150 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1151 {
1152         /* In case of renewal the kernel map, destroy previous one */
1153         machine__destroy_kernel_maps(machine);
1154
1155         map__put(machine->vmlinux_map);
1156         machine->vmlinux_map = map__new2(0, kernel);
1157         if (machine->vmlinux_map == NULL)
1158                 return -ENOMEM;
1159
1160         map__set_mapping_type(machine->vmlinux_map, MAPPING_TYPE__IDENTITY);
1161         return maps__insert(machine__kernel_maps(machine), machine->vmlinux_map);
1162 }
1163
1164 void machine__destroy_kernel_maps(struct machine *machine)
1165 {
1166         struct kmap *kmap;
1167         struct map *map = machine__kernel_map(machine);
1168
1169         if (map == NULL)
1170                 return;
1171
1172         kmap = map__kmap(map);
1173         maps__remove(machine__kernel_maps(machine), map);
1174         if (kmap && kmap->ref_reloc_sym) {
1175                 zfree((char **)&kmap->ref_reloc_sym->name);
1176                 zfree(&kmap->ref_reloc_sym);
1177         }
1178
1179         map__zput(machine->vmlinux_map);
1180 }
1181
1182 int machines__create_guest_kernel_maps(struct machines *machines)
1183 {
1184         int ret = 0;
1185         struct dirent **namelist = NULL;
1186         int i, items = 0;
1187         char path[PATH_MAX];
1188         pid_t pid;
1189         char *endp;
1190
1191         if (symbol_conf.default_guest_vmlinux_name ||
1192             symbol_conf.default_guest_modules ||
1193             symbol_conf.default_guest_kallsyms) {
1194                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1195         }
1196
1197         if (symbol_conf.guestmount) {
1198                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1199                 if (items <= 0)
1200                         return -ENOENT;
1201                 for (i = 0; i < items; i++) {
1202                         if (!isdigit(namelist[i]->d_name[0])) {
1203                                 /* Filter out . and .. */
1204                                 continue;
1205                         }
1206                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1207                         if ((*endp != '\0') ||
1208                             (endp == namelist[i]->d_name) ||
1209                             (errno == ERANGE)) {
1210                                 pr_debug("invalid directory (%s). Skipping.\n",
1211                                          namelist[i]->d_name);
1212                                 continue;
1213                         }
1214                         sprintf(path, "%s/%s/proc/kallsyms",
1215                                 symbol_conf.guestmount,
1216                                 namelist[i]->d_name);
1217                         ret = access(path, R_OK);
1218                         if (ret) {
1219                                 pr_debug("Can't access file %s\n", path);
1220                                 goto failure;
1221                         }
1222                         machines__create_kernel_maps(machines, pid);
1223                 }
1224 failure:
1225                 free(namelist);
1226         }
1227
1228         return ret;
1229 }
1230
1231 void machines__destroy_kernel_maps(struct machines *machines)
1232 {
1233         struct rb_node *next = rb_first_cached(&machines->guests);
1234
1235         machine__destroy_kernel_maps(&machines->host);
1236
1237         while (next) {
1238                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1239
1240                 next = rb_next(&pos->rb_node);
1241                 rb_erase_cached(&pos->rb_node, &machines->guests);
1242                 machine__delete(pos);
1243         }
1244 }
1245
1246 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1247 {
1248         struct machine *machine = machines__findnew(machines, pid);
1249
1250         if (machine == NULL)
1251                 return -1;
1252
1253         return machine__create_kernel_maps(machine);
1254 }
1255
1256 int machine__load_kallsyms(struct machine *machine, const char *filename)
1257 {
1258         struct map *map = machine__kernel_map(machine);
1259         struct dso *dso = map__dso(map);
1260         int ret = __dso__load_kallsyms(dso, filename, map, true);
1261
1262         if (ret > 0) {
1263                 dso__set_loaded(dso);
1264                 /*
1265                  * Since /proc/kallsyms will have multiple sessions for the
1266                  * kernel, with modules between them, fixup the end of all
1267                  * sections.
1268                  */
1269                 maps__fixup_end(machine__kernel_maps(machine));
1270         }
1271
1272         return ret;
1273 }
1274
1275 int machine__load_vmlinux_path(struct machine *machine)
1276 {
1277         struct map *map = machine__kernel_map(machine);
1278         struct dso *dso = map__dso(map);
1279         int ret = dso__load_vmlinux_path(dso, map);
1280
1281         if (ret > 0)
1282                 dso__set_loaded(dso);
1283
1284         return ret;
1285 }
1286
1287 static char *get_kernel_version(const char *root_dir)
1288 {
1289         char version[PATH_MAX];
1290         FILE *file;
1291         char *name, *tmp;
1292         const char *prefix = "Linux version ";
1293
1294         sprintf(version, "%s/proc/version", root_dir);
1295         file = fopen(version, "r");
1296         if (!file)
1297                 return NULL;
1298
1299         tmp = fgets(version, sizeof(version), file);
1300         fclose(file);
1301         if (!tmp)
1302                 return NULL;
1303
1304         name = strstr(version, prefix);
1305         if (!name)
1306                 return NULL;
1307         name += strlen(prefix);
1308         tmp = strchr(name, ' ');
1309         if (tmp)
1310                 *tmp = '\0';
1311
1312         return strdup(name);
1313 }
1314
1315 static bool is_kmod_dso(struct dso *dso)
1316 {
1317         return dso__symtab_type(dso) == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1318                dso__symtab_type(dso) == DSO_BINARY_TYPE__GUEST_KMODULE;
1319 }
1320
1321 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1322 {
1323         char *long_name;
1324         struct dso *dso;
1325         struct map *map = maps__find_by_name(maps, m->name);
1326
1327         if (map == NULL)
1328                 return 0;
1329
1330         long_name = strdup(path);
1331         if (long_name == NULL) {
1332                 map__put(map);
1333                 return -ENOMEM;
1334         }
1335
1336         dso = map__dso(map);
1337         dso__set_long_name(dso, long_name, true);
1338         dso__kernel_module_get_build_id(dso, "");
1339
1340         /*
1341          * Full name could reveal us kmod compression, so
1342          * we need to update the symtab_type if needed.
1343          */
1344         if (m->comp && is_kmod_dso(dso)) {
1345                 dso__set_symtab_type(dso, dso__symtab_type(dso));
1346                 dso__set_comp(dso, m->comp);
1347         }
1348         map__put(map);
1349         return 0;
1350 }
1351
1352 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1353 {
1354         struct dirent *dent;
1355         DIR *dir = opendir(dir_name);
1356         int ret = 0;
1357
1358         if (!dir) {
1359                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1360                 return -1;
1361         }
1362
1363         while ((dent = readdir(dir)) != NULL) {
1364                 char path[PATH_MAX];
1365                 struct stat st;
1366
1367                 /*sshfs might return bad dent->d_type, so we have to stat*/
1368                 path__join(path, sizeof(path), dir_name, dent->d_name);
1369                 if (stat(path, &st))
1370                         continue;
1371
1372                 if (S_ISDIR(st.st_mode)) {
1373                         if (!strcmp(dent->d_name, ".") ||
1374                             !strcmp(dent->d_name, ".."))
1375                                 continue;
1376
1377                         /* Do not follow top-level source and build symlinks */
1378                         if (depth == 0) {
1379                                 if (!strcmp(dent->d_name, "source") ||
1380                                     !strcmp(dent->d_name, "build"))
1381                                         continue;
1382                         }
1383
1384                         ret = maps__set_modules_path_dir(maps, path, depth + 1);
1385                         if (ret < 0)
1386                                 goto out;
1387                 } else {
1388                         struct kmod_path m;
1389
1390                         ret = kmod_path__parse_name(&m, dent->d_name);
1391                         if (ret)
1392                                 goto out;
1393
1394                         if (m.kmod)
1395                                 ret = maps__set_module_path(maps, path, &m);
1396
1397                         zfree(&m.name);
1398
1399                         if (ret)
1400                                 goto out;
1401                 }
1402         }
1403
1404 out:
1405         closedir(dir);
1406         return ret;
1407 }
1408
1409 static int machine__set_modules_path(struct machine *machine)
1410 {
1411         char *version;
1412         char modules_path[PATH_MAX];
1413
1414         version = get_kernel_version(machine->root_dir);
1415         if (!version)
1416                 return -1;
1417
1418         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1419                  machine->root_dir, version);
1420         free(version);
1421
1422         return maps__set_modules_path_dir(machine__kernel_maps(machine), modules_path, 0);
1423 }
1424 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1425                                 u64 *size __maybe_unused,
1426                                 const char *name __maybe_unused)
1427 {
1428         return 0;
1429 }
1430
1431 static int machine__create_module(void *arg, const char *name, u64 start,
1432                                   u64 size)
1433 {
1434         struct machine *machine = arg;
1435         struct map *map;
1436
1437         if (arch__fix_module_text_start(&start, &size, name) < 0)
1438                 return -1;
1439
1440         map = machine__addnew_module_map(machine, start, name);
1441         if (map == NULL)
1442                 return -1;
1443         map__set_end(map, start + size);
1444
1445         dso__kernel_module_get_build_id(map__dso(map), machine->root_dir);
1446         map__put(map);
1447         return 0;
1448 }
1449
1450 static int machine__create_modules(struct machine *machine)
1451 {
1452         const char *modules;
1453         char path[PATH_MAX];
1454
1455         if (machine__is_default_guest(machine)) {
1456                 modules = symbol_conf.default_guest_modules;
1457         } else {
1458                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1459                 modules = path;
1460         }
1461
1462         if (symbol__restricted_filename(modules, "/proc/modules"))
1463                 return -1;
1464
1465         if (modules__parse(modules, machine, machine__create_module))
1466                 return -1;
1467
1468         if (!machine__set_modules_path(machine))
1469                 return 0;
1470
1471         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1472
1473         return 0;
1474 }
1475
1476 static void machine__set_kernel_mmap(struct machine *machine,
1477                                      u64 start, u64 end)
1478 {
1479         map__set_start(machine->vmlinux_map, start);
1480         map__set_end(machine->vmlinux_map, end);
1481         /*
1482          * Be a bit paranoid here, some perf.data file came with
1483          * a zero sized synthesized MMAP event for the kernel.
1484          */
1485         if (start == 0 && end == 0)
1486                 map__set_end(machine->vmlinux_map, ~0ULL);
1487 }
1488
1489 static int machine__update_kernel_mmap(struct machine *machine,
1490                                      u64 start, u64 end)
1491 {
1492         struct map *orig, *updated;
1493         int err;
1494
1495         orig = machine->vmlinux_map;
1496         updated = map__get(orig);
1497
1498         machine->vmlinux_map = updated;
1499         maps__remove(machine__kernel_maps(machine), orig);
1500         machine__set_kernel_mmap(machine, start, end);
1501         err = maps__insert(machine__kernel_maps(machine), updated);
1502         map__put(orig);
1503
1504         return err;
1505 }
1506
1507 int machine__create_kernel_maps(struct machine *machine)
1508 {
1509         struct dso *kernel = machine__get_kernel(machine);
1510         const char *name = NULL;
1511         u64 start = 0, end = ~0ULL;
1512         int ret;
1513
1514         if (kernel == NULL)
1515                 return -1;
1516
1517         ret = __machine__create_kernel_maps(machine, kernel);
1518         if (ret < 0)
1519                 goto out_put;
1520
1521         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1522                 if (machine__is_host(machine))
1523                         pr_debug("Problems creating module maps, "
1524                                  "continuing anyway...\n");
1525                 else
1526                         pr_debug("Problems creating module maps for guest %d, "
1527                                  "continuing anyway...\n", machine->pid);
1528         }
1529
1530         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1531                 if (name &&
1532                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1533                         machine__destroy_kernel_maps(machine);
1534                         ret = -1;
1535                         goto out_put;
1536                 }
1537
1538                 /*
1539                  * we have a real start address now, so re-order the kmaps
1540                  * assume it's the last in the kmaps
1541                  */
1542                 ret = machine__update_kernel_mmap(machine, start, end);
1543                 if (ret < 0)
1544                         goto out_put;
1545         }
1546
1547         if (machine__create_extra_kernel_maps(machine, kernel))
1548                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1549
1550         if (end == ~0ULL) {
1551                 /* update end address of the kernel map using adjacent module address */
1552                 struct map *next = maps__find_next_entry(machine__kernel_maps(machine),
1553                                                          machine__kernel_map(machine));
1554
1555                 if (next) {
1556                         machine__set_kernel_mmap(machine, start, map__start(next));
1557                         map__put(next);
1558                 }
1559         }
1560
1561 out_put:
1562         dso__put(kernel);
1563         return ret;
1564 }
1565
1566 static int machine__uses_kcore_cb(struct dso *dso, void *data __maybe_unused)
1567 {
1568         return dso__is_kcore(dso) ? 1 : 0;
1569 }
1570
1571 static bool machine__uses_kcore(struct machine *machine)
1572 {
1573         return dsos__for_each_dso(&machine->dsos, machine__uses_kcore_cb, NULL) != 0 ? true : false;
1574 }
1575
1576 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1577                                              struct extra_kernel_map *xm)
1578 {
1579         return machine__is(machine, "x86_64") &&
1580                is_entry_trampoline(xm->name);
1581 }
1582
1583 static int machine__process_extra_kernel_map(struct machine *machine,
1584                                              struct extra_kernel_map *xm)
1585 {
1586         struct dso *kernel = machine__kernel_dso(machine);
1587
1588         if (kernel == NULL)
1589                 return -1;
1590
1591         return machine__create_extra_kernel_map(machine, kernel, xm);
1592 }
1593
1594 static int machine__process_kernel_mmap_event(struct machine *machine,
1595                                               struct extra_kernel_map *xm,
1596                                               struct build_id *bid)
1597 {
1598         enum dso_space_type dso_space;
1599         bool is_kernel_mmap;
1600         const char *mmap_name = machine->mmap_name;
1601
1602         /* If we have maps from kcore then we do not need or want any others */
1603         if (machine__uses_kcore(machine))
1604                 return 0;
1605
1606         if (machine__is_host(machine))
1607                 dso_space = DSO_SPACE__KERNEL;
1608         else
1609                 dso_space = DSO_SPACE__KERNEL_GUEST;
1610
1611         is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1612         if (!is_kernel_mmap && !machine__is_host(machine)) {
1613                 /*
1614                  * If the event was recorded inside the guest and injected into
1615                  * the host perf.data file, then it will match a host mmap_name,
1616                  * so try that - see machine__set_mmap_name().
1617                  */
1618                 mmap_name = "[kernel.kallsyms]";
1619                 is_kernel_mmap = memcmp(xm->name, mmap_name, strlen(mmap_name) - 1) == 0;
1620         }
1621         if (xm->name[0] == '/' ||
1622             (!is_kernel_mmap && xm->name[0] == '[')) {
1623                 struct map *map = machine__addnew_module_map(machine, xm->start, xm->name);
1624
1625                 if (map == NULL)
1626                         goto out_problem;
1627
1628                 map__set_end(map, map__start(map) + xm->end - xm->start);
1629
1630                 if (build_id__is_defined(bid))
1631                         dso__set_build_id(map__dso(map), bid);
1632
1633                 map__put(map);
1634         } else if (is_kernel_mmap) {
1635                 const char *symbol_name = xm->name + strlen(mmap_name);
1636                 /*
1637                  * Should be there already, from the build-id table in
1638                  * the header.
1639                  */
1640                 struct dso *kernel = dsos__find_kernel_dso(&machine->dsos);
1641
1642                 if (kernel == NULL)
1643                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1644                 if (kernel == NULL)
1645                         goto out_problem;
1646
1647                 dso__set_kernel(kernel, dso_space);
1648                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1649                         dso__put(kernel);
1650                         goto out_problem;
1651                 }
1652
1653                 if (strstr(dso__long_name(kernel), "vmlinux"))
1654                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1655
1656                 if (machine__update_kernel_mmap(machine, xm->start, xm->end) < 0) {
1657                         dso__put(kernel);
1658                         goto out_problem;
1659                 }
1660
1661                 if (build_id__is_defined(bid))
1662                         dso__set_build_id(kernel, bid);
1663
1664                 /*
1665                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1666                  * symbol. Effectively having zero here means that at record
1667                  * time /proc/sys/kernel/kptr_restrict was non zero.
1668                  */
1669                 if (xm->pgoff != 0) {
1670                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1671                                                         symbol_name,
1672                                                         xm->pgoff);
1673                 }
1674
1675                 if (machine__is_default_guest(machine)) {
1676                         /*
1677                          * preload dso of guest kernel and modules
1678                          */
1679                         dso__load(kernel, machine__kernel_map(machine));
1680                 }
1681                 dso__put(kernel);
1682         } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1683                 return machine__process_extra_kernel_map(machine, xm);
1684         }
1685         return 0;
1686 out_problem:
1687         return -1;
1688 }
1689
1690 int machine__process_mmap2_event(struct machine *machine,
1691                                  union perf_event *event,
1692                                  struct perf_sample *sample)
1693 {
1694         struct thread *thread;
1695         struct map *map;
1696         struct dso_id dso_id = {
1697                 .maj = event->mmap2.maj,
1698                 .min = event->mmap2.min,
1699                 .ino = event->mmap2.ino,
1700                 .ino_generation = event->mmap2.ino_generation,
1701         };
1702         struct build_id __bid, *bid = NULL;
1703         int ret = 0;
1704
1705         if (dump_trace)
1706                 perf_event__fprintf_mmap2(event, stdout);
1707
1708         if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1709                 bid = &__bid;
1710                 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1711         }
1712
1713         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1714             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1715                 struct extra_kernel_map xm = {
1716                         .start = event->mmap2.start,
1717                         .end   = event->mmap2.start + event->mmap2.len,
1718                         .pgoff = event->mmap2.pgoff,
1719                 };
1720
1721                 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1722                 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1723                 if (ret < 0)
1724                         goto out_problem;
1725                 return 0;
1726         }
1727
1728         thread = machine__findnew_thread(machine, event->mmap2.pid,
1729                                         event->mmap2.tid);
1730         if (thread == NULL)
1731                 goto out_problem;
1732
1733         map = map__new(machine, event->mmap2.start,
1734                         event->mmap2.len, event->mmap2.pgoff,
1735                         &dso_id, event->mmap2.prot,
1736                         event->mmap2.flags, bid,
1737                         event->mmap2.filename, thread);
1738
1739         if (map == NULL)
1740                 goto out_problem_map;
1741
1742         ret = thread__insert_map(thread, map);
1743         if (ret)
1744                 goto out_problem_insert;
1745
1746         thread__put(thread);
1747         map__put(map);
1748         return 0;
1749
1750 out_problem_insert:
1751         map__put(map);
1752 out_problem_map:
1753         thread__put(thread);
1754 out_problem:
1755         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1756         return 0;
1757 }
1758
1759 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1760                                 struct perf_sample *sample)
1761 {
1762         struct thread *thread;
1763         struct map *map;
1764         u32 prot = 0;
1765         int ret = 0;
1766
1767         if (dump_trace)
1768                 perf_event__fprintf_mmap(event, stdout);
1769
1770         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1771             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1772                 struct extra_kernel_map xm = {
1773                         .start = event->mmap.start,
1774                         .end   = event->mmap.start + event->mmap.len,
1775                         .pgoff = event->mmap.pgoff,
1776                 };
1777
1778                 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1779                 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
1780                 if (ret < 0)
1781                         goto out_problem;
1782                 return 0;
1783         }
1784
1785         thread = machine__findnew_thread(machine, event->mmap.pid,
1786                                          event->mmap.tid);
1787         if (thread == NULL)
1788                 goto out_problem;
1789
1790         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1791                 prot = PROT_EXEC;
1792
1793         map = map__new(machine, event->mmap.start,
1794                         event->mmap.len, event->mmap.pgoff,
1795                         NULL, prot, 0, NULL, event->mmap.filename, thread);
1796
1797         if (map == NULL)
1798                 goto out_problem_map;
1799
1800         ret = thread__insert_map(thread, map);
1801         if (ret)
1802                 goto out_problem_insert;
1803
1804         thread__put(thread);
1805         map__put(map);
1806         return 0;
1807
1808 out_problem_insert:
1809         map__put(map);
1810 out_problem_map:
1811         thread__put(thread);
1812 out_problem:
1813         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1814         return 0;
1815 }
1816
1817 void machine__remove_thread(struct machine *machine, struct thread *th)
1818 {
1819         return threads__remove(&machine->threads, th);
1820 }
1821
1822 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1823                                 struct perf_sample *sample)
1824 {
1825         struct thread *thread = machine__find_thread(machine,
1826                                                      event->fork.pid,
1827                                                      event->fork.tid);
1828         struct thread *parent = machine__findnew_thread(machine,
1829                                                         event->fork.ppid,
1830                                                         event->fork.ptid);
1831         bool do_maps_clone = true;
1832         int err = 0;
1833
1834         if (dump_trace)
1835                 perf_event__fprintf_task(event, stdout);
1836
1837         /*
1838          * There may be an existing thread that is not actually the parent,
1839          * either because we are processing events out of order, or because the
1840          * (fork) event that would have removed the thread was lost. Assume the
1841          * latter case and continue on as best we can.
1842          */
1843         if (thread__pid(parent) != (pid_t)event->fork.ppid) {
1844                 dump_printf("removing erroneous parent thread %d/%d\n",
1845                             thread__pid(parent), thread__tid(parent));
1846                 machine__remove_thread(machine, parent);
1847                 thread__put(parent);
1848                 parent = machine__findnew_thread(machine, event->fork.ppid,
1849                                                  event->fork.ptid);
1850         }
1851
1852         /* if a thread currently exists for the thread id remove it */
1853         if (thread != NULL) {
1854                 machine__remove_thread(machine, thread);
1855                 thread__put(thread);
1856         }
1857
1858         thread = machine__findnew_thread(machine, event->fork.pid,
1859                                          event->fork.tid);
1860         /*
1861          * When synthesizing FORK events, we are trying to create thread
1862          * objects for the already running tasks on the machine.
1863          *
1864          * Normally, for a kernel FORK event, we want to clone the parent's
1865          * maps because that is what the kernel just did.
1866          *
1867          * But when synthesizing, this should not be done.  If we do, we end up
1868          * with overlapping maps as we process the synthesized MMAP2 events that
1869          * get delivered shortly thereafter.
1870          *
1871          * Use the FORK event misc flags in an internal way to signal this
1872          * situation, so we can elide the map clone when appropriate.
1873          */
1874         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1875                 do_maps_clone = false;
1876
1877         if (thread == NULL || parent == NULL ||
1878             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1879                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1880                 err = -1;
1881         }
1882         thread__put(thread);
1883         thread__put(parent);
1884
1885         return err;
1886 }
1887
1888 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1889                                 struct perf_sample *sample __maybe_unused)
1890 {
1891         struct thread *thread = machine__find_thread(machine,
1892                                                      event->fork.pid,
1893                                                      event->fork.tid);
1894
1895         if (dump_trace)
1896                 perf_event__fprintf_task(event, stdout);
1897
1898         if (thread != NULL) {
1899                 if (symbol_conf.keep_exited_threads)
1900                         thread__set_exited(thread, /*exited=*/true);
1901                 else
1902                         machine__remove_thread(machine, thread);
1903         }
1904         thread__put(thread);
1905         return 0;
1906 }
1907
1908 int machine__process_event(struct machine *machine, union perf_event *event,
1909                            struct perf_sample *sample)
1910 {
1911         int ret;
1912
1913         switch (event->header.type) {
1914         case PERF_RECORD_COMM:
1915                 ret = machine__process_comm_event(machine, event, sample); break;
1916         case PERF_RECORD_MMAP:
1917                 ret = machine__process_mmap_event(machine, event, sample); break;
1918         case PERF_RECORD_NAMESPACES:
1919                 ret = machine__process_namespaces_event(machine, event, sample); break;
1920         case PERF_RECORD_CGROUP:
1921                 ret = machine__process_cgroup_event(machine, event, sample); break;
1922         case PERF_RECORD_MMAP2:
1923                 ret = machine__process_mmap2_event(machine, event, sample); break;
1924         case PERF_RECORD_FORK:
1925                 ret = machine__process_fork_event(machine, event, sample); break;
1926         case PERF_RECORD_EXIT:
1927                 ret = machine__process_exit_event(machine, event, sample); break;
1928         case PERF_RECORD_LOST:
1929                 ret = machine__process_lost_event(machine, event, sample); break;
1930         case PERF_RECORD_AUX:
1931                 ret = machine__process_aux_event(machine, event); break;
1932         case PERF_RECORD_ITRACE_START:
1933                 ret = machine__process_itrace_start_event(machine, event); break;
1934         case PERF_RECORD_LOST_SAMPLES:
1935                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1936         case PERF_RECORD_SWITCH:
1937         case PERF_RECORD_SWITCH_CPU_WIDE:
1938                 ret = machine__process_switch_event(machine, event); break;
1939         case PERF_RECORD_KSYMBOL:
1940                 ret = machine__process_ksymbol(machine, event, sample); break;
1941         case PERF_RECORD_BPF_EVENT:
1942                 ret = machine__process_bpf(machine, event, sample); break;
1943         case PERF_RECORD_TEXT_POKE:
1944                 ret = machine__process_text_poke(machine, event, sample); break;
1945         case PERF_RECORD_AUX_OUTPUT_HW_ID:
1946                 ret = machine__process_aux_output_hw_id_event(machine, event); break;
1947         default:
1948                 ret = -1;
1949                 break;
1950         }
1951
1952         return ret;
1953 }
1954
1955 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1956 {
1957         return regexec(regex, sym->name, 0, NULL, 0) == 0;
1958 }
1959
1960 static void ip__resolve_ams(struct thread *thread,
1961                             struct addr_map_symbol *ams,
1962                             u64 ip)
1963 {
1964         struct addr_location al;
1965
1966         addr_location__init(&al);
1967         /*
1968          * We cannot use the header.misc hint to determine whether a
1969          * branch stack address is user, kernel, guest, hypervisor.
1970          * Branches may straddle the kernel/user/hypervisor boundaries.
1971          * Thus, we have to try consecutively until we find a match
1972          * or else, the symbol is unknown
1973          */
1974         thread__find_cpumode_addr_location(thread, ip, &al);
1975
1976         ams->addr = ip;
1977         ams->al_addr = al.addr;
1978         ams->al_level = al.level;
1979         ams->ms.maps = maps__get(al.maps);
1980         ams->ms.sym = al.sym;
1981         ams->ms.map = map__get(al.map);
1982         ams->phys_addr = 0;
1983         ams->data_page_size = 0;
1984         addr_location__exit(&al);
1985 }
1986
1987 static void ip__resolve_data(struct thread *thread,
1988                              u8 m, struct addr_map_symbol *ams,
1989                              u64 addr, u64 phys_addr, u64 daddr_page_size)
1990 {
1991         struct addr_location al;
1992
1993         addr_location__init(&al);
1994
1995         thread__find_symbol(thread, m, addr, &al);
1996
1997         ams->addr = addr;
1998         ams->al_addr = al.addr;
1999         ams->al_level = al.level;
2000         ams->ms.maps = maps__get(al.maps);
2001         ams->ms.sym = al.sym;
2002         ams->ms.map = map__get(al.map);
2003         ams->phys_addr = phys_addr;
2004         ams->data_page_size = daddr_page_size;
2005         addr_location__exit(&al);
2006 }
2007
2008 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2009                                      struct addr_location *al)
2010 {
2011         struct mem_info *mi = mem_info__new();
2012
2013         if (!mi)
2014                 return NULL;
2015
2016         ip__resolve_ams(al->thread, mem_info__iaddr(mi), sample->ip);
2017         ip__resolve_data(al->thread, al->cpumode, mem_info__daddr(mi),
2018                          sample->addr, sample->phys_addr,
2019                          sample->data_page_size);
2020         mem_info__data_src(mi)->val = sample->data_src;
2021
2022         return mi;
2023 }
2024
2025 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2026 {
2027         struct map *map = ms->map;
2028         char *srcline = NULL;
2029         struct dso *dso;
2030
2031         if (!map || callchain_param.key == CCKEY_FUNCTION)
2032                 return srcline;
2033
2034         dso = map__dso(map);
2035         srcline = srcline__tree_find(dso__srclines(dso), ip);
2036         if (!srcline) {
2037                 bool show_sym = false;
2038                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2039
2040                 srcline = get_srcline(dso, map__rip_2objdump(map, ip),
2041                                       ms->sym, show_sym, show_addr, ip);
2042                 srcline__tree_insert(dso__srclines(dso), ip, srcline);
2043         }
2044
2045         return srcline;
2046 }
2047
2048 struct iterations {
2049         int nr_loop_iter;
2050         u64 cycles;
2051 };
2052
2053 static int add_callchain_ip(struct thread *thread,
2054                             struct callchain_cursor *cursor,
2055                             struct symbol **parent,
2056                             struct addr_location *root_al,
2057                             u8 *cpumode,
2058                             u64 ip,
2059                             bool branch,
2060                             struct branch_flags *flags,
2061                             struct iterations *iter,
2062                             u64 branch_from)
2063 {
2064         struct map_symbol ms = {};
2065         struct addr_location al;
2066         int nr_loop_iter = 0, err = 0;
2067         u64 iter_cycles = 0;
2068         const char *srcline = NULL;
2069
2070         addr_location__init(&al);
2071         al.filtered = 0;
2072         al.sym = NULL;
2073         al.srcline = NULL;
2074         if (!cpumode) {
2075                 thread__find_cpumode_addr_location(thread, ip, &al);
2076         } else {
2077                 if (ip >= PERF_CONTEXT_MAX) {
2078                         switch (ip) {
2079                         case PERF_CONTEXT_HV:
2080                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2081                                 break;
2082                         case PERF_CONTEXT_KERNEL:
2083                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2084                                 break;
2085                         case PERF_CONTEXT_USER:
2086                                 *cpumode = PERF_RECORD_MISC_USER;
2087                                 break;
2088                         default:
2089                                 pr_debug("invalid callchain context: "
2090                                          "%"PRId64"\n", (s64) ip);
2091                                 /*
2092                                  * It seems the callchain is corrupted.
2093                                  * Discard all.
2094                                  */
2095                                 callchain_cursor_reset(cursor);
2096                                 err = 1;
2097                                 goto out;
2098                         }
2099                         goto out;
2100                 }
2101                 thread__find_symbol(thread, *cpumode, ip, &al);
2102         }
2103
2104         if (al.sym != NULL) {
2105                 if (perf_hpp_list.parent && !*parent &&
2106                     symbol__match_regex(al.sym, &parent_regex))
2107                         *parent = al.sym;
2108                 else if (have_ignore_callees && root_al &&
2109                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2110                         /* Treat this symbol as the root,
2111                            forgetting its callees. */
2112                         addr_location__copy(root_al, &al);
2113                         callchain_cursor_reset(cursor);
2114                 }
2115         }
2116
2117         if (symbol_conf.hide_unresolved && al.sym == NULL)
2118                 goto out;
2119
2120         if (iter) {
2121                 nr_loop_iter = iter->nr_loop_iter;
2122                 iter_cycles = iter->cycles;
2123         }
2124
2125         ms.maps = maps__get(al.maps);
2126         ms.map = map__get(al.map);
2127         ms.sym = al.sym;
2128         srcline = callchain_srcline(&ms, al.addr);
2129         err = callchain_cursor_append(cursor, ip, &ms,
2130                                       branch, flags, nr_loop_iter,
2131                                       iter_cycles, branch_from, srcline);
2132 out:
2133         addr_location__exit(&al);
2134         map_symbol__exit(&ms);
2135         return err;
2136 }
2137
2138 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2139                                            struct addr_location *al)
2140 {
2141         unsigned int i;
2142         const struct branch_stack *bs = sample->branch_stack;
2143         struct branch_entry *entries = perf_sample__branch_entries(sample);
2144         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2145
2146         if (!bi)
2147                 return NULL;
2148
2149         for (i = 0; i < bs->nr; i++) {
2150                 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2151                 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2152                 bi[i].flags = entries[i].flags;
2153         }
2154         return bi;
2155 }
2156
2157 static void save_iterations(struct iterations *iter,
2158                             struct branch_entry *be, int nr)
2159 {
2160         int i;
2161
2162         iter->nr_loop_iter++;
2163         iter->cycles = 0;
2164
2165         for (i = 0; i < nr; i++)
2166                 iter->cycles += be[i].flags.cycles;
2167 }
2168
2169 #define CHASHSZ 127
2170 #define CHASHBITS 7
2171 #define NO_ENTRY 0xff
2172
2173 #define PERF_MAX_BRANCH_DEPTH 127
2174
2175 /* Remove loops. */
2176 static int remove_loops(struct branch_entry *l, int nr,
2177                         struct iterations *iter)
2178 {
2179         int i, j, off;
2180         unsigned char chash[CHASHSZ];
2181
2182         memset(chash, NO_ENTRY, sizeof(chash));
2183
2184         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2185
2186         for (i = 0; i < nr; i++) {
2187                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2188
2189                 /* no collision handling for now */
2190                 if (chash[h] == NO_ENTRY) {
2191                         chash[h] = i;
2192                 } else if (l[chash[h]].from == l[i].from) {
2193                         bool is_loop = true;
2194                         /* check if it is a real loop */
2195                         off = 0;
2196                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2197                                 if (l[j].from != l[i + off].from) {
2198                                         is_loop = false;
2199                                         break;
2200                                 }
2201                         if (is_loop) {
2202                                 j = nr - (i + off);
2203                                 if (j > 0) {
2204                                         save_iterations(iter + i + off,
2205                                                 l + i, off);
2206
2207                                         memmove(iter + i, iter + i + off,
2208                                                 j * sizeof(*iter));
2209
2210                                         memmove(l + i, l + i + off,
2211                                                 j * sizeof(*l));
2212                                 }
2213
2214                                 nr -= off;
2215                         }
2216                 }
2217         }
2218         return nr;
2219 }
2220
2221 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2222                                        struct callchain_cursor *cursor,
2223                                        struct perf_sample *sample,
2224                                        struct symbol **parent,
2225                                        struct addr_location *root_al,
2226                                        u64 branch_from,
2227                                        bool callee, int end)
2228 {
2229         struct ip_callchain *chain = sample->callchain;
2230         u8 cpumode = PERF_RECORD_MISC_USER;
2231         int err, i;
2232
2233         if (callee) {
2234                 for (i = 0; i < end + 1; i++) {
2235                         err = add_callchain_ip(thread, cursor, parent,
2236                                                root_al, &cpumode, chain->ips[i],
2237                                                false, NULL, NULL, branch_from);
2238                         if (err)
2239                                 return err;
2240                 }
2241                 return 0;
2242         }
2243
2244         for (i = end; i >= 0; i--) {
2245                 err = add_callchain_ip(thread, cursor, parent,
2246                                        root_al, &cpumode, chain->ips[i],
2247                                        false, NULL, NULL, branch_from);
2248                 if (err)
2249                         return err;
2250         }
2251
2252         return 0;
2253 }
2254
2255 static void save_lbr_cursor_node(struct thread *thread,
2256                                  struct callchain_cursor *cursor,
2257                                  int idx)
2258 {
2259         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2260
2261         if (!lbr_stitch)
2262                 return;
2263
2264         if (cursor->pos == cursor->nr) {
2265                 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2266                 return;
2267         }
2268
2269         if (!cursor->curr)
2270                 cursor->curr = cursor->first;
2271         else
2272                 cursor->curr = cursor->curr->next;
2273         memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2274                sizeof(struct callchain_cursor_node));
2275
2276         lbr_stitch->prev_lbr_cursor[idx].valid = true;
2277         cursor->pos++;
2278 }
2279
2280 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2281                                     struct callchain_cursor *cursor,
2282                                     struct perf_sample *sample,
2283                                     struct symbol **parent,
2284                                     struct addr_location *root_al,
2285                                     u64 *branch_from,
2286                                     bool callee)
2287 {
2288         struct branch_stack *lbr_stack = sample->branch_stack;
2289         struct branch_entry *entries = perf_sample__branch_entries(sample);
2290         u8 cpumode = PERF_RECORD_MISC_USER;
2291         int lbr_nr = lbr_stack->nr;
2292         struct branch_flags *flags;
2293         int err, i;
2294         u64 ip;
2295
2296         /*
2297          * The curr and pos are not used in writing session. They are cleared
2298          * in callchain_cursor_commit() when the writing session is closed.
2299          * Using curr and pos to track the current cursor node.
2300          */
2301         if (thread__lbr_stitch(thread)) {
2302                 cursor->curr = NULL;
2303                 cursor->pos = cursor->nr;
2304                 if (cursor->nr) {
2305                         cursor->curr = cursor->first;
2306                         for (i = 0; i < (int)(cursor->nr - 1); i++)
2307                                 cursor->curr = cursor->curr->next;
2308                 }
2309         }
2310
2311         if (callee) {
2312                 /* Add LBR ip from first entries.to */
2313                 ip = entries[0].to;
2314                 flags = &entries[0].flags;
2315                 *branch_from = entries[0].from;
2316                 err = add_callchain_ip(thread, cursor, parent,
2317                                        root_al, &cpumode, ip,
2318                                        true, flags, NULL,
2319                                        *branch_from);
2320                 if (err)
2321                         return err;
2322
2323                 /*
2324                  * The number of cursor node increases.
2325                  * Move the current cursor node.
2326                  * But does not need to save current cursor node for entry 0.
2327                  * It's impossible to stitch the whole LBRs of previous sample.
2328                  */
2329                 if (thread__lbr_stitch(thread) && (cursor->pos != cursor->nr)) {
2330                         if (!cursor->curr)
2331                                 cursor->curr = cursor->first;
2332                         else
2333                                 cursor->curr = cursor->curr->next;
2334                         cursor->pos++;
2335                 }
2336
2337                 /* Add LBR ip from entries.from one by one. */
2338                 for (i = 0; i < lbr_nr; i++) {
2339                         ip = entries[i].from;
2340                         flags = &entries[i].flags;
2341                         err = add_callchain_ip(thread, cursor, parent,
2342                                                root_al, &cpumode, ip,
2343                                                true, flags, NULL,
2344                                                *branch_from);
2345                         if (err)
2346                                 return err;
2347                         save_lbr_cursor_node(thread, cursor, i);
2348                 }
2349                 return 0;
2350         }
2351
2352         /* Add LBR ip from entries.from one by one. */
2353         for (i = lbr_nr - 1; i >= 0; i--) {
2354                 ip = entries[i].from;
2355                 flags = &entries[i].flags;
2356                 err = add_callchain_ip(thread, cursor, parent,
2357                                        root_al, &cpumode, ip,
2358                                        true, flags, NULL,
2359                                        *branch_from);
2360                 if (err)
2361                         return err;
2362                 save_lbr_cursor_node(thread, cursor, i);
2363         }
2364
2365         if (lbr_nr > 0) {
2366                 /* Add LBR ip from first entries.to */
2367                 ip = entries[0].to;
2368                 flags = &entries[0].flags;
2369                 *branch_from = entries[0].from;
2370                 err = add_callchain_ip(thread, cursor, parent,
2371                                 root_al, &cpumode, ip,
2372                                 true, flags, NULL,
2373                                 *branch_from);
2374                 if (err)
2375                         return err;
2376         }
2377
2378         return 0;
2379 }
2380
2381 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2382                                              struct callchain_cursor *cursor)
2383 {
2384         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2385         struct callchain_cursor_node *cnode;
2386         struct stitch_list *stitch_node;
2387         int err;
2388
2389         list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2390                 cnode = &stitch_node->cursor;
2391
2392                 err = callchain_cursor_append(cursor, cnode->ip,
2393                                               &cnode->ms,
2394                                               cnode->branch,
2395                                               &cnode->branch_flags,
2396                                               cnode->nr_loop_iter,
2397                                               cnode->iter_cycles,
2398                                               cnode->branch_from,
2399                                               cnode->srcline);
2400                 if (err)
2401                         return err;
2402         }
2403         return 0;
2404 }
2405
2406 static struct stitch_list *get_stitch_node(struct thread *thread)
2407 {
2408         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2409         struct stitch_list *stitch_node;
2410
2411         if (!list_empty(&lbr_stitch->free_lists)) {
2412                 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2413                                                struct stitch_list, node);
2414                 list_del(&stitch_node->node);
2415
2416                 return stitch_node;
2417         }
2418
2419         return malloc(sizeof(struct stitch_list));
2420 }
2421
2422 static bool has_stitched_lbr(struct thread *thread,
2423                              struct perf_sample *cur,
2424                              struct perf_sample *prev,
2425                              unsigned int max_lbr,
2426                              bool callee)
2427 {
2428         struct branch_stack *cur_stack = cur->branch_stack;
2429         struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2430         struct branch_stack *prev_stack = prev->branch_stack;
2431         struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2432         struct lbr_stitch *lbr_stitch = thread__lbr_stitch(thread);
2433         int i, j, nr_identical_branches = 0;
2434         struct stitch_list *stitch_node;
2435         u64 cur_base, distance;
2436
2437         if (!cur_stack || !prev_stack)
2438                 return false;
2439
2440         /* Find the physical index of the base-of-stack for current sample. */
2441         cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2442
2443         distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2444                                                      (max_lbr + prev_stack->hw_idx - cur_base);
2445         /* Previous sample has shorter stack. Nothing can be stitched. */
2446         if (distance + 1 > prev_stack->nr)
2447                 return false;
2448
2449         /*
2450          * Check if there are identical LBRs between two samples.
2451          * Identical LBRs must have same from, to and flags values. Also,
2452          * they have to be saved in the same LBR registers (same physical
2453          * index).
2454          *
2455          * Starts from the base-of-stack of current sample.
2456          */
2457         for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2458                 if ((prev_entries[i].from != cur_entries[j].from) ||
2459                     (prev_entries[i].to != cur_entries[j].to) ||
2460                     (prev_entries[i].flags.value != cur_entries[j].flags.value))
2461                         break;
2462                 nr_identical_branches++;
2463         }
2464
2465         if (!nr_identical_branches)
2466                 return false;
2467
2468         /*
2469          * Save the LBRs between the base-of-stack of previous sample
2470          * and the base-of-stack of current sample into lbr_stitch->lists.
2471          * These LBRs will be stitched later.
2472          */
2473         for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2474
2475                 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2476                         continue;
2477
2478                 stitch_node = get_stitch_node(thread);
2479                 if (!stitch_node)
2480                         return false;
2481
2482                 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2483                        sizeof(struct callchain_cursor_node));
2484
2485                 if (callee)
2486                         list_add(&stitch_node->node, &lbr_stitch->lists);
2487                 else
2488                         list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2489         }
2490
2491         return true;
2492 }
2493
2494 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2495 {
2496         if (thread__lbr_stitch(thread))
2497                 return true;
2498
2499         thread__set_lbr_stitch(thread, zalloc(sizeof(struct lbr_stitch)));
2500         if (!thread__lbr_stitch(thread))
2501                 goto err;
2502
2503         thread__lbr_stitch(thread)->prev_lbr_cursor =
2504                 calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2505         if (!thread__lbr_stitch(thread)->prev_lbr_cursor)
2506                 goto free_lbr_stitch;
2507
2508         INIT_LIST_HEAD(&thread__lbr_stitch(thread)->lists);
2509         INIT_LIST_HEAD(&thread__lbr_stitch(thread)->free_lists);
2510
2511         return true;
2512
2513 free_lbr_stitch:
2514         free(thread__lbr_stitch(thread));
2515         thread__set_lbr_stitch(thread, NULL);
2516 err:
2517         pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2518         thread__set_lbr_stitch_enable(thread, false);
2519         return false;
2520 }
2521
2522 /*
2523  * Resolve LBR callstack chain sample
2524  * Return:
2525  * 1 on success get LBR callchain information
2526  * 0 no available LBR callchain information, should try fp
2527  * negative error code on other errors.
2528  */
2529 static int resolve_lbr_callchain_sample(struct thread *thread,
2530                                         struct callchain_cursor *cursor,
2531                                         struct perf_sample *sample,
2532                                         struct symbol **parent,
2533                                         struct addr_location *root_al,
2534                                         int max_stack,
2535                                         unsigned int max_lbr)
2536 {
2537         bool callee = (callchain_param.order == ORDER_CALLEE);
2538         struct ip_callchain *chain = sample->callchain;
2539         int chain_nr = min(max_stack, (int)chain->nr), i;
2540         struct lbr_stitch *lbr_stitch;
2541         bool stitched_lbr = false;
2542         u64 branch_from = 0;
2543         int err;
2544
2545         for (i = 0; i < chain_nr; i++) {
2546                 if (chain->ips[i] == PERF_CONTEXT_USER)
2547                         break;
2548         }
2549
2550         /* LBR only affects the user callchain */
2551         if (i == chain_nr)
2552                 return 0;
2553
2554         if (thread__lbr_stitch_enable(thread) && !sample->no_hw_idx &&
2555             (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2556                 lbr_stitch = thread__lbr_stitch(thread);
2557
2558                 stitched_lbr = has_stitched_lbr(thread, sample,
2559                                                 &lbr_stitch->prev_sample,
2560                                                 max_lbr, callee);
2561
2562                 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2563                         list_replace_init(&lbr_stitch->lists,
2564                                           &lbr_stitch->free_lists);
2565                 }
2566                 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2567         }
2568
2569         if (callee) {
2570                 /* Add kernel ip */
2571                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2572                                                   parent, root_al, branch_from,
2573                                                   true, i);
2574                 if (err)
2575                         goto error;
2576
2577                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2578                                                root_al, &branch_from, true);
2579                 if (err)
2580                         goto error;
2581
2582                 if (stitched_lbr) {
2583                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2584                         if (err)
2585                                 goto error;
2586                 }
2587
2588         } else {
2589                 if (stitched_lbr) {
2590                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2591                         if (err)
2592                                 goto error;
2593                 }
2594                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2595                                                root_al, &branch_from, false);
2596                 if (err)
2597                         goto error;
2598
2599                 /* Add kernel ip */
2600                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2601                                                   parent, root_al, branch_from,
2602                                                   false, i);
2603                 if (err)
2604                         goto error;
2605         }
2606         return 1;
2607
2608 error:
2609         return (err < 0) ? err : 0;
2610 }
2611
2612 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2613                              struct callchain_cursor *cursor,
2614                              struct symbol **parent,
2615                              struct addr_location *root_al,
2616                              u8 *cpumode, int ent)
2617 {
2618         int err = 0;
2619
2620         while (--ent >= 0) {
2621                 u64 ip = chain->ips[ent];
2622
2623                 if (ip >= PERF_CONTEXT_MAX) {
2624                         err = add_callchain_ip(thread, cursor, parent,
2625                                                root_al, cpumode, ip,
2626                                                false, NULL, NULL, 0);
2627                         break;
2628                 }
2629         }
2630         return err;
2631 }
2632
2633 static u64 get_leaf_frame_caller(struct perf_sample *sample,
2634                 struct thread *thread, int usr_idx)
2635 {
2636         if (machine__normalized_is(maps__machine(thread__maps(thread)), "arm64"))
2637                 return get_leaf_frame_caller_aarch64(sample, thread, usr_idx);
2638         else
2639                 return 0;
2640 }
2641
2642 static int thread__resolve_callchain_sample(struct thread *thread,
2643                                             struct callchain_cursor *cursor,
2644                                             struct evsel *evsel,
2645                                             struct perf_sample *sample,
2646                                             struct symbol **parent,
2647                                             struct addr_location *root_al,
2648                                             int max_stack)
2649 {
2650         struct branch_stack *branch = sample->branch_stack;
2651         struct branch_entry *entries = perf_sample__branch_entries(sample);
2652         struct ip_callchain *chain = sample->callchain;
2653         int chain_nr = 0;
2654         u8 cpumode = PERF_RECORD_MISC_USER;
2655         int i, j, err, nr_entries, usr_idx;
2656         int skip_idx = -1;
2657         int first_call = 0;
2658         u64 leaf_frame_caller;
2659
2660         if (chain)
2661                 chain_nr = chain->nr;
2662
2663         if (evsel__has_branch_callstack(evsel)) {
2664                 struct perf_env *env = evsel__env(evsel);
2665
2666                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2667                                                    root_al, max_stack,
2668                                                    !env ? 0 : env->max_branches);
2669                 if (err)
2670                         return (err < 0) ? err : 0;
2671         }
2672
2673         /*
2674          * Based on DWARF debug information, some architectures skip
2675          * a callchain entry saved by the kernel.
2676          */
2677         skip_idx = arch_skip_callchain_idx(thread, chain);
2678
2679         /*
2680          * Add branches to call stack for easier browsing. This gives
2681          * more context for a sample than just the callers.
2682          *
2683          * This uses individual histograms of paths compared to the
2684          * aggregated histograms the normal LBR mode uses.
2685          *
2686          * Limitations for now:
2687          * - No extra filters
2688          * - No annotations (should annotate somehow)
2689          */
2690
2691         if (branch && callchain_param.branch_callstack) {
2692                 int nr = min(max_stack, (int)branch->nr);
2693                 struct branch_entry be[nr];
2694                 struct iterations iter[nr];
2695
2696                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2697                         pr_warning("corrupted branch chain. skipping...\n");
2698                         goto check_calls;
2699                 }
2700
2701                 for (i = 0; i < nr; i++) {
2702                         if (callchain_param.order == ORDER_CALLEE) {
2703                                 be[i] = entries[i];
2704
2705                                 if (chain == NULL)
2706                                         continue;
2707
2708                                 /*
2709                                  * Check for overlap into the callchain.
2710                                  * The return address is one off compared to
2711                                  * the branch entry. To adjust for this
2712                                  * assume the calling instruction is not longer
2713                                  * than 8 bytes.
2714                                  */
2715                                 if (i == skip_idx ||
2716                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2717                                         first_call++;
2718                                 else if (be[i].from < chain->ips[first_call] &&
2719                                     be[i].from >= chain->ips[first_call] - 8)
2720                                         first_call++;
2721                         } else
2722                                 be[i] = entries[branch->nr - i - 1];
2723                 }
2724
2725                 memset(iter, 0, sizeof(struct iterations) * nr);
2726                 nr = remove_loops(be, nr, iter);
2727
2728                 for (i = 0; i < nr; i++) {
2729                         err = add_callchain_ip(thread, cursor, parent,
2730                                                root_al,
2731                                                NULL, be[i].to,
2732                                                true, &be[i].flags,
2733                                                NULL, be[i].from);
2734
2735                         if (!err)
2736                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2737                                                        NULL, be[i].from,
2738                                                        true, &be[i].flags,
2739                                                        &iter[i], 0);
2740                         if (err == -EINVAL)
2741                                 break;
2742                         if (err)
2743                                 return err;
2744                 }
2745
2746                 if (chain_nr == 0)
2747                         return 0;
2748
2749                 chain_nr -= nr;
2750         }
2751
2752 check_calls:
2753         if (chain && callchain_param.order != ORDER_CALLEE) {
2754                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2755                                         &cpumode, chain->nr - first_call);
2756                 if (err)
2757                         return (err < 0) ? err : 0;
2758         }
2759         for (i = first_call, nr_entries = 0;
2760              i < chain_nr && nr_entries < max_stack; i++) {
2761                 u64 ip;
2762
2763                 if (callchain_param.order == ORDER_CALLEE)
2764                         j = i;
2765                 else
2766                         j = chain->nr - i - 1;
2767
2768 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2769                 if (j == skip_idx)
2770                         continue;
2771 #endif
2772                 ip = chain->ips[j];
2773                 if (ip < PERF_CONTEXT_MAX)
2774                        ++nr_entries;
2775                 else if (callchain_param.order != ORDER_CALLEE) {
2776                         err = find_prev_cpumode(chain, thread, cursor, parent,
2777                                                 root_al, &cpumode, j);
2778                         if (err)
2779                                 return (err < 0) ? err : 0;
2780                         continue;
2781                 }
2782
2783                 /*
2784                  * PERF_CONTEXT_USER allows us to locate where the user stack ends.
2785                  * Depending on callchain_param.order and the position of PERF_CONTEXT_USER,
2786                  * the index will be different in order to add the missing frame
2787                  * at the right place.
2788                  */
2789
2790                 usr_idx = callchain_param.order == ORDER_CALLEE ? j-2 : j-1;
2791
2792                 if (usr_idx >= 0 && chain->ips[usr_idx] == PERF_CONTEXT_USER) {
2793
2794                         leaf_frame_caller = get_leaf_frame_caller(sample, thread, usr_idx);
2795
2796                         /*
2797                          * check if leaf_frame_Caller != ip to not add the same
2798                          * value twice.
2799                          */
2800
2801                         if (leaf_frame_caller && leaf_frame_caller != ip) {
2802
2803                                 err = add_callchain_ip(thread, cursor, parent,
2804                                                root_al, &cpumode, leaf_frame_caller,
2805                                                false, NULL, NULL, 0);
2806                                 if (err)
2807                                         return (err < 0) ? err : 0;
2808                         }
2809                 }
2810
2811                 err = add_callchain_ip(thread, cursor, parent,
2812                                        root_al, &cpumode, ip,
2813                                        false, NULL, NULL, 0);
2814
2815                 if (err)
2816                         return (err < 0) ? err : 0;
2817         }
2818
2819         return 0;
2820 }
2821
2822 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2823 {
2824         struct symbol *sym = ms->sym;
2825         struct map *map = ms->map;
2826         struct inline_node *inline_node;
2827         struct inline_list *ilist;
2828         struct dso *dso;
2829         u64 addr;
2830         int ret = 1;
2831         struct map_symbol ilist_ms;
2832
2833         if (!symbol_conf.inline_name || !map || !sym)
2834                 return ret;
2835
2836         addr = map__dso_map_ip(map, ip);
2837         addr = map__rip_2objdump(map, addr);
2838         dso = map__dso(map);
2839
2840         inline_node = inlines__tree_find(dso__inlined_nodes(dso), addr);
2841         if (!inline_node) {
2842                 inline_node = dso__parse_addr_inlines(dso, addr, sym);
2843                 if (!inline_node)
2844                         return ret;
2845                 inlines__tree_insert(dso__inlined_nodes(dso), inline_node);
2846         }
2847
2848         ilist_ms = (struct map_symbol) {
2849                 .maps = maps__get(ms->maps),
2850                 .map = map__get(map),
2851         };
2852         list_for_each_entry(ilist, &inline_node->val, list) {
2853                 ilist_ms.sym = ilist->symbol;
2854                 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2855                                               NULL, 0, 0, 0, ilist->srcline);
2856
2857                 if (ret != 0)
2858                         return ret;
2859         }
2860         map_symbol__exit(&ilist_ms);
2861
2862         return ret;
2863 }
2864
2865 static int unwind_entry(struct unwind_entry *entry, void *arg)
2866 {
2867         struct callchain_cursor *cursor = arg;
2868         const char *srcline = NULL;
2869         u64 addr = entry->ip;
2870
2871         if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2872                 return 0;
2873
2874         if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2875                 return 0;
2876
2877         /*
2878          * Convert entry->ip from a virtual address to an offset in
2879          * its corresponding binary.
2880          */
2881         if (entry->ms.map)
2882                 addr = map__dso_map_ip(entry->ms.map, entry->ip);
2883
2884         srcline = callchain_srcline(&entry->ms, addr);
2885         return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2886                                        false, NULL, 0, 0, 0, srcline);
2887 }
2888
2889 static int thread__resolve_callchain_unwind(struct thread *thread,
2890                                             struct callchain_cursor *cursor,
2891                                             struct evsel *evsel,
2892                                             struct perf_sample *sample,
2893                                             int max_stack)
2894 {
2895         /* Can we do dwarf post unwind? */
2896         if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2897               (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2898                 return 0;
2899
2900         /* Bail out if nothing was captured. */
2901         if ((!sample->user_regs.regs) ||
2902             (!sample->user_stack.size))
2903                 return 0;
2904
2905         return unwind__get_entries(unwind_entry, cursor,
2906                                    thread, sample, max_stack, false);
2907 }
2908
2909 int thread__resolve_callchain(struct thread *thread,
2910                               struct callchain_cursor *cursor,
2911                               struct evsel *evsel,
2912                               struct perf_sample *sample,
2913                               struct symbol **parent,
2914                               struct addr_location *root_al,
2915                               int max_stack)
2916 {
2917         int ret = 0;
2918
2919         if (cursor == NULL)
2920                 return -ENOMEM;
2921
2922         callchain_cursor_reset(cursor);
2923
2924         if (callchain_param.order == ORDER_CALLEE) {
2925                 ret = thread__resolve_callchain_sample(thread, cursor,
2926                                                        evsel, sample,
2927                                                        parent, root_al,
2928                                                        max_stack);
2929                 if (ret)
2930                         return ret;
2931                 ret = thread__resolve_callchain_unwind(thread, cursor,
2932                                                        evsel, sample,
2933                                                        max_stack);
2934         } else {
2935                 ret = thread__resolve_callchain_unwind(thread, cursor,
2936                                                        evsel, sample,
2937                                                        max_stack);
2938                 if (ret)
2939                         return ret;
2940                 ret = thread__resolve_callchain_sample(thread, cursor,
2941                                                        evsel, sample,
2942                                                        parent, root_al,
2943                                                        max_stack);
2944         }
2945
2946         return ret;
2947 }
2948
2949 int machine__for_each_thread(struct machine *machine,
2950                              int (*fn)(struct thread *thread, void *p),
2951                              void *priv)
2952 {
2953         return threads__for_each_thread(&machine->threads, fn, priv);
2954 }
2955
2956 int machines__for_each_thread(struct machines *machines,
2957                               int (*fn)(struct thread *thread, void *p),
2958                               void *priv)
2959 {
2960         struct rb_node *nd;
2961         int rc = 0;
2962
2963         rc = machine__for_each_thread(&machines->host, fn, priv);
2964         if (rc != 0)
2965                 return rc;
2966
2967         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
2968                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
2969
2970                 rc = machine__for_each_thread(machine, fn, priv);
2971                 if (rc != 0)
2972                         return rc;
2973         }
2974         return rc;
2975 }
2976
2977
2978 static int thread_list_cb(struct thread *thread, void *data)
2979 {
2980         struct list_head *list = data;
2981         struct thread_list *entry = malloc(sizeof(*entry));
2982
2983         if (!entry)
2984                 return -ENOMEM;
2985
2986         entry->thread = thread__get(thread);
2987         list_add_tail(&entry->list, list);
2988         return 0;
2989 }
2990
2991 int machine__thread_list(struct machine *machine, struct list_head *list)
2992 {
2993         return machine__for_each_thread(machine, thread_list_cb, list);
2994 }
2995
2996 void thread_list__delete(struct list_head *list)
2997 {
2998         struct thread_list *pos, *next;
2999
3000         list_for_each_entry_safe(pos, next, list, list) {
3001                 thread__zput(pos->thread);
3002                 list_del(&pos->list);
3003                 free(pos);
3004         }
3005 }
3006
3007 pid_t machine__get_current_tid(struct machine *machine, int cpu)
3008 {
3009         if (cpu < 0 || (size_t)cpu >= machine->current_tid_sz)
3010                 return -1;
3011
3012         return machine->current_tid[cpu];
3013 }
3014
3015 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3016                              pid_t tid)
3017 {
3018         struct thread *thread;
3019         const pid_t init_val = -1;
3020
3021         if (cpu < 0)
3022                 return -EINVAL;
3023
3024         if (realloc_array_as_needed(machine->current_tid,
3025                                     machine->current_tid_sz,
3026                                     (unsigned int)cpu,
3027                                     &init_val))
3028                 return -ENOMEM;
3029
3030         machine->current_tid[cpu] = tid;
3031
3032         thread = machine__findnew_thread(machine, pid, tid);
3033         if (!thread)
3034                 return -ENOMEM;
3035
3036         thread__set_cpu(thread, cpu);
3037         thread__put(thread);
3038
3039         return 0;
3040 }
3041
3042 /*
3043  * Compares the raw arch string. N.B. see instead perf_env__arch() or
3044  * machine__normalized_is() if a normalized arch is needed.
3045  */
3046 bool machine__is(struct machine *machine, const char *arch)
3047 {
3048         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3049 }
3050
3051 bool machine__normalized_is(struct machine *machine, const char *arch)
3052 {
3053         return machine && !strcmp(perf_env__arch(machine->env), arch);
3054 }
3055
3056 int machine__nr_cpus_avail(struct machine *machine)
3057 {
3058         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3059 }
3060
3061 int machine__get_kernel_start(struct machine *machine)
3062 {
3063         struct map *map = machine__kernel_map(machine);
3064         int err = 0;
3065
3066         /*
3067          * The only addresses above 2^63 are kernel addresses of a 64-bit
3068          * kernel.  Note that addresses are unsigned so that on a 32-bit system
3069          * all addresses including kernel addresses are less than 2^32.  In
3070          * that case (32-bit system), if the kernel mapping is unknown, all
3071          * addresses will be assumed to be in user space - see
3072          * machine__kernel_ip().
3073          */
3074         machine->kernel_start = 1ULL << 63;
3075         if (map) {
3076                 err = map__load(map);
3077                 /*
3078                  * On x86_64, PTI entry trampolines are less than the
3079                  * start of kernel text, but still above 2^63. So leave
3080                  * kernel_start = 1ULL << 63 for x86_64.
3081                  */
3082                 if (!err && !machine__is(machine, "x86_64"))
3083                         machine->kernel_start = map__start(map);
3084         }
3085         return err;
3086 }
3087
3088 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3089 {
3090         u8 addr_cpumode = cpumode;
3091         bool kernel_ip;
3092
3093         if (!machine->single_address_space)
3094                 goto out;
3095
3096         kernel_ip = machine__kernel_ip(machine, addr);
3097         switch (cpumode) {
3098         case PERF_RECORD_MISC_KERNEL:
3099         case PERF_RECORD_MISC_USER:
3100                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3101                                            PERF_RECORD_MISC_USER;
3102                 break;
3103         case PERF_RECORD_MISC_GUEST_KERNEL:
3104         case PERF_RECORD_MISC_GUEST_USER:
3105                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3106                                            PERF_RECORD_MISC_GUEST_USER;
3107                 break;
3108         default:
3109                 break;
3110         }
3111 out:
3112         return addr_cpumode;
3113 }
3114
3115 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3116 {
3117         return dsos__findnew_id(&machine->dsos, filename, id);
3118 }
3119
3120 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3121 {
3122         return machine__findnew_dso_id(machine, filename, NULL);
3123 }
3124
3125 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3126 {
3127         struct machine *machine = vmachine;
3128         struct map *map;
3129         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3130
3131         if (sym == NULL)
3132                 return NULL;
3133
3134         *modp = __map__is_kmodule(map) ? (char *)dso__short_name(map__dso(map)) : NULL;
3135         *addrp = map__unmap_ip(map, sym->start);
3136         return sym->name;
3137 }
3138
3139 struct machine__for_each_dso_cb_args {
3140         struct machine *machine;
3141         machine__dso_t fn;
3142         void *priv;
3143 };
3144
3145 static int machine__for_each_dso_cb(struct dso *dso, void *data)
3146 {
3147         struct machine__for_each_dso_cb_args *args = data;
3148
3149         return args->fn(dso, args->machine, args->priv);
3150 }
3151
3152 int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3153 {
3154         struct machine__for_each_dso_cb_args args = {
3155                 .machine = machine,
3156                 .fn = fn,
3157                 .priv = priv,
3158         };
3159
3160         return dsos__for_each_dso(&machine->dsos, machine__for_each_dso_cb, &args);
3161 }
3162
3163 int machine__for_each_kernel_map(struct machine *machine, machine__map_t fn, void *priv)
3164 {
3165         struct maps *maps = machine__kernel_maps(machine);
3166
3167         return maps__for_each_map(maps, fn, priv);
3168 }
3169
3170 bool machine__is_lock_function(struct machine *machine, u64 addr)
3171 {
3172         if (!machine->sched.text_start) {
3173                 struct map *kmap;
3174                 struct symbol *sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_start", &kmap);
3175
3176                 if (!sym) {
3177                         /* to avoid retry */
3178                         machine->sched.text_start = 1;
3179                         return false;
3180                 }
3181
3182                 machine->sched.text_start = map__unmap_ip(kmap, sym->start);
3183
3184                 /* should not fail from here */
3185                 sym = machine__find_kernel_symbol_by_name(machine, "__sched_text_end", &kmap);
3186                 machine->sched.text_end = map__unmap_ip(kmap, sym->start);
3187
3188                 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_start", &kmap);
3189                 machine->lock.text_start = map__unmap_ip(kmap, sym->start);
3190
3191                 sym = machine__find_kernel_symbol_by_name(machine, "__lock_text_end", &kmap);
3192                 machine->lock.text_end = map__unmap_ip(kmap, sym->start);
3193
3194                 sym = machine__find_kernel_symbol_by_name(machine, "__traceiter_contention_begin", &kmap);
3195                 if (sym) {
3196                         machine->traceiter.text_start = map__unmap_ip(kmap, sym->start);
3197                         machine->traceiter.text_end = map__unmap_ip(kmap, sym->end);
3198                 }
3199                 sym = machine__find_kernel_symbol_by_name(machine, "trace_contention_begin", &kmap);
3200                 if (sym) {
3201                         machine->trace.text_start = map__unmap_ip(kmap, sym->start);
3202                         machine->trace.text_end = map__unmap_ip(kmap, sym->end);
3203                 }
3204         }
3205
3206         /* failed to get kernel symbols */
3207         if (machine->sched.text_start == 1)
3208                 return false;
3209
3210         /* mutex and rwsem functions are in sched text section */
3211         if (machine->sched.text_start <= addr && addr < machine->sched.text_end)
3212                 return true;
3213
3214         /* spinlock functions are in lock text section */
3215         if (machine->lock.text_start <= addr && addr < machine->lock.text_end)
3216                 return true;
3217
3218         /* traceiter functions currently don't have their own section
3219          * but we consider them lock functions
3220          */
3221         if (machine->traceiter.text_start != 0) {
3222                 if (machine->traceiter.text_start <= addr && addr < machine->traceiter.text_end)
3223                         return true;
3224         }
3225
3226         if (machine->trace.text_start != 0) {
3227                 if (machine->trace.text_start <= addr && addr < machine->trace.text_end)
3228                         return true;
3229         }
3230
3231         return false;
3232 }
3233
3234 int machine__hit_all_dsos(struct machine *machine)
3235 {
3236         return dsos__hit_all(&machine->dsos);
3237 }