tools/perf/util/machine.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <dirent.h>
   3 #include <errno.h>
   4 #include <inttypes.h>
   5 #include <regex.h>
   6 #include <stdlib.h>
   7 #include "callchain.h"
   8 #include "debug.h"
   9 #include "dso.h"
  10 #include "env.h"
  11 #include "event.h"
  12 #include "evsel.h"
  13 #include "hist.h"
  14 #include "machine.h"
  15 #include "map.h"
  16 #include "map_symbol.h"
  17 #include "branch.h"
  18 #include "mem-events.h"
  19 #include "srcline.h"
  20 #include "symbol.h"
  21 #include "sort.h"
  22 #include "strlist.h"
  23 #include "target.h"
  24 #include "thread.h"
  25 #include "util.h"
  26 #include "vdso.h"
  27 #include <stdbool.h>
  28 #include <sys/types.h>
  29 #include <sys/stat.h>
  30 #include <unistd.h>
  31 #include "unwind.h"
  32 #include "linux/hash.h"
  33 #include "asm/bug.h"
  34 #include "bpf-event.h"
  35 #include <internal/lib.h> // page_size
  36 #include "cgroup.h"
  37
  38 #include <linux/ctype.h>
  39 #include <symbol/kallsyms.h>
  40 #include <linux/mman.h>
  41 #include <linux/string.h>
  42 #include <linux/zalloc.h>
  43
  44 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  45
  46 static struct dso *machine__kernel_dso(struct machine *machine)
  47 {
  48         return machine->vmlinux_map->dso;
  49 }
  50
  51 static void dsos__init(struct dsos *dsos)
  52 {
  53         INIT_LIST_HEAD(&dsos->head);
  54         dsos->root = RB_ROOT;
  55         init_rwsem(&dsos->lock);
  56 }
  57
  58 static void machine__threads_init(struct machine *machine)
  59 {
  60         int i;
  61
  62         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
  63                 struct threads *threads = &machine->threads[i];
  64                 threads->entries = RB_ROOT_CACHED;
  65                 init_rwsem(&threads->lock);
  66                 threads->nr = 0;
  67                 INIT_LIST_HEAD(&threads->dead);
  68                 threads->last_match = NULL;
  69         }
  70 }
  71
  72 static int machine__set_mmap_name(struct machine *machine)
  73 {
  74         if (machine__is_host(machine))
  75                 machine->mmap_name = strdup("[kernel.kallsyms]");
  76         else if (machine__is_default_guest(machine))
  77                 machine->mmap_name = strdup("[guest.kernel.kallsyms]");
  78         else if (asprintf(&machine->mmap_name, "[guest.kernel.kallsyms.%d]",
  79                           machine->pid) < 0)
  80                 machine->mmap_name = NULL;
  81
  82         return machine->mmap_name ? 0 : -ENOMEM;
  83 }
  84
  85 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  86 {
  87         int err = -ENOMEM;
  88
  89         memset(machine, 0, sizeof(*machine));
  90         maps__init(&machine->kmaps, machine);
  91         RB_CLEAR_NODE(&machine->rb_node);
  92         dsos__init(&machine->dsos);
  93
  94         machine__threads_init(machine);
  95
  96         machine->vdso_info = NULL;
  97         machine->env = NULL;
  98
  99         machine->pid = pid;
 100
 101         machine->id_hdr_size = 0;
 102         machine->kptr_restrict_warned = false;
 103         machine->comm_exec = false;
 104         machine->kernel_start = 0;
 105         machine->vmlinux_map = NULL;
 106
 107         machine->root_dir = strdup(root_dir);
 108         if (machine->root_dir == NULL)
 109                 return -ENOMEM;
 110
 111         if (machine__set_mmap_name(machine))
 112                 goto out;
 113
 114         if (pid != HOST_KERNEL_ID) {
 115                 struct thread *thread = machine__findnew_thread(machine, -1,
 116                                                                 pid);
 117                 char comm[64];
 118
 119                 if (thread == NULL)
 120                         goto out;
 121
 122                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
 123                 thread__set_comm(thread, comm, 0);
 124                 thread__put(thread);
 125         }
 126
 127         machine->current_tid = NULL;
 128         err = 0;
 129
 130 out:
 131         if (err) {
 132                 zfree(&machine->root_dir);
 133                 zfree(&machine->mmap_name);
 134         }
 135         return 0;
 136 }
 137
 138 struct machine *machine__new_host(void)
 139 {
 140         struct machine *machine = malloc(sizeof(*machine));
 141
 142         if (machine != NULL) {
 143                 machine__init(machine, "", HOST_KERNEL_ID);
 144
 145                 if (machine__create_kernel_maps(machine) < 0)
 146                         goto out_delete;
 147         }
 148
 149         return machine;
 150 out_delete:
 151         free(machine);
 152         return NULL;
 153 }
 154
 155 struct machine *machine__new_kallsyms(void)
 156 {
 157         struct machine *machine = machine__new_host();
 158         /*
 159          * FIXME:
 160          * 1) We should switch to machine__load_kallsyms(), i.e. not explicitly
 161          *    ask for not using the kcore parsing code, once this one is fixed
 162          *    to create a map per module.
 163          */
 164         if (machine && machine__load_kallsyms(machine, "/proc/kallsyms") <= 0) {
 165                 machine__delete(machine);
 166                 machine = NULL;
 167         }
 168
 169         return machine;
 170 }
 171
 172 static void dsos__purge(struct dsos *dsos)
 173 {
 174         struct dso *pos, *n;
 175
 176         down_write(&dsos->lock);
 177
 178         list_for_each_entry_safe(pos, n, &dsos->head, node) {
 179                 RB_CLEAR_NODE(&pos->rb_node);
 180                 pos->root = NULL;
 181                 list_del_init(&pos->node);
 182                 dso__put(pos);
 183         }
 184
 185         up_write(&dsos->lock);
 186 }
 187
 188 static void dsos__exit(struct dsos *dsos)
 189 {
 190         dsos__purge(dsos);
 191         exit_rwsem(&dsos->lock);
 192 }
 193
 194 void machine__delete_threads(struct machine *machine)
 195 {
 196         struct rb_node *nd;
 197         int i;
 198
 199         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 200                 struct threads *threads = &machine->threads[i];
 201                 down_write(&threads->lock);
 202                 nd = rb_first_cached(&threads->entries);
 203                 while (nd) {
 204                         struct thread *t = rb_entry(nd, struct thread, rb_node);
 205
 206                         nd = rb_next(nd);
 207                         __machine__remove_thread(machine, t, false);
 208                 }
 209                 up_write(&threads->lock);
 210         }
 211 }
 212
 213 void machine__exit(struct machine *machine)
 214 {
 215         int i;
 216
 217         if (machine == NULL)
 218                 return;
 219
 220         machine__destroy_kernel_maps(machine);
 221         maps__exit(&machine->kmaps);
 222         dsos__exit(&machine->dsos);
 223         machine__exit_vdso(machine);
 224         zfree(&machine->root_dir);
 225         zfree(&machine->mmap_name);
 226         zfree(&machine->current_tid);
 227
 228         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 229                 struct threads *threads = &machine->threads[i];
 230                 struct thread *thread, *n;
 231                 /*
 232                  * Forget about the dead, at this point whatever threads were
 233                  * left in the dead lists better have a reference count taken
 234                  * by who is using them, and then, when they drop those references
 235                  * and it finally hits zero, thread__put() will check and see that
 236                  * its not in the dead threads list and will not try to remove it
 237                  * from there, just calling thread__delete() straight away.
 238                  */
 239                 list_for_each_entry_safe(thread, n, &threads->dead, node)
 240                         list_del_init(&thread->node);
 241
 242                 exit_rwsem(&threads->lock);
 243         }
 244 }
 245
 246 void machine__delete(struct machine *machine)
 247 {
 248         if (machine) {
 249                 machine__exit(machine);
 250                 free(machine);
 251         }
 252 }
 253
 254 void machines__init(struct machines *machines)
 255 {
 256         machine__init(&machines->host, "", HOST_KERNEL_ID);
 257         machines->guests = RB_ROOT_CACHED;
 258 }
 259
 260 void machines__exit(struct machines *machines)
 261 {
 262         machine__exit(&machines->host);
 263         /* XXX exit guest */
 264 }
 265
 266 struct machine *machines__add(struct machines *machines, pid_t pid,
 267                               const char *root_dir)
 268 {
 269         struct rb_node **p = &machines->guests.rb_root.rb_node;
 270         struct rb_node *parent = NULL;
 271         struct machine *pos, *machine = malloc(sizeof(*machine));
 272         bool leftmost = true;
 273
 274         if (machine == NULL)
 275                 return NULL;
 276
 277         if (machine__init(machine, root_dir, pid) != 0) {
 278                 free(machine);
 279                 return NULL;
 280         }
 281
 282         while (*p != NULL) {
 283                 parent = *p;
 284                 pos = rb_entry(parent, struct machine, rb_node);
 285                 if (pid < pos->pid)
 286                         p = &(*p)->rb_left;
 287                 else {
 288                         p = &(*p)->rb_right;
 289                         leftmost = false;
 290                 }
 291         }
 292
 293         rb_link_node(&machine->rb_node, parent, p);
 294         rb_insert_color_cached(&machine->rb_node, &machines->guests, leftmost);
 295
 296         return machine;
 297 }
 298
 299 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 300 {
 301         struct rb_node *nd;
 302
 303         machines->host.comm_exec = comm_exec;
 304
 305         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 306                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 307
 308                 machine->comm_exec = comm_exec;
 309         }
 310 }
 311
 312 struct machine *machines__find(struct machines *machines, pid_t pid)
 313 {
 314         struct rb_node **p = &machines->guests.rb_root.rb_node;
 315         struct rb_node *parent = NULL;
 316         struct machine *machine;
 317         struct machine *default_machine = NULL;
 318
 319         if (pid == HOST_KERNEL_ID)
 320                 return &machines->host;
 321
 322         while (*p != NULL) {
 323                 parent = *p;
 324                 machine = rb_entry(parent, struct machine, rb_node);
 325                 if (pid < machine->pid)
 326                         p = &(*p)->rb_left;
 327                 else if (pid > machine->pid)
 328                         p = &(*p)->rb_right;
 329                 else
 330                         return machine;
 331                 if (!machine->pid)
 332                         default_machine = machine;
 333         }
 334
 335         return default_machine;
 336 }
 337
 338 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 339 {
 340         char path[PATH_MAX];
 341         const char *root_dir = "";
 342         struct machine *machine = machines__find(machines, pid);
 343
 344         if (machine && (machine->pid == pid))
 345                 goto out;
 346
 347         if ((pid != HOST_KERNEL_ID) &&
 348             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 349             (symbol_conf.guestmount)) {
 350                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 351                 if (access(path, R_OK)) {
 352                         static struct strlist *seen;
 353
 354                         if (!seen)
 355                                 seen = strlist__new(NULL, NULL);
 356
 357                         if (!strlist__has_entry(seen, path)) {
 358                                 pr_err("Can't access file %s\n", path);
 359                                 strlist__add(seen, path);
 360                         }
 361                         machine = NULL;
 362                         goto out;
 363                 }
 364                 root_dir = path;
 365         }
 366
 367         machine = machines__add(machines, pid, root_dir);
 368 out:
 369         return machine;
 370 }
 371
 372 struct machine *machines__find_guest(struct machines *machines, pid_t pid)
 373 {
 374         struct machine *machine = machines__find(machines, pid);
 375
 376         if (!machine)
 377                 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
 378         return machine;
 379 }
 380
 381 void machines__process_guests(struct machines *machines,
 382                               machine__process_t process, void *data)
 383 {
 384         struct rb_node *nd;
 385
 386         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 387                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 388                 process(pos, data);
 389         }
 390 }
 391
 392 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 393 {
 394         struct rb_node *node;
 395         struct machine *machine;
 396
 397         machines->host.id_hdr_size = id_hdr_size;
 398
 399         for (node = rb_first_cached(&machines->guests); node;
 400              node = rb_next(node)) {
 401                 machine = rb_entry(node, struct machine, rb_node);
 402                 machine->id_hdr_size = id_hdr_size;
 403         }
 404
 405         return;
 406 }
 407
 408 static void machine__update_thread_pid(struct machine *machine,
 409                                        struct thread *th, pid_t pid)
 410 {
 411         struct thread *leader;
 412
 413         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 414                 return;
 415
 416         th->pid_ = pid;
 417
 418         if (th->pid_ == th->tid)
 419                 return;
 420
 421         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 422         if (!leader)
 423                 goto out_err;
 424
 425         if (!leader->maps)
 426                 leader->maps = maps__new(machine);
 427
 428         if (!leader->maps)
 429                 goto out_err;
 430
 431         if (th->maps == leader->maps)
 432                 return;
 433
 434         if (th->maps) {
 435                 /*
 436                  * Maps are created from MMAP events which provide the pid and
 437                  * tid.  Consequently there never should be any maps on a thread
 438                  * with an unknown pid.  Just print an error if there are.
 439                  */
 440                 if (!maps__empty(th->maps))
 441                         pr_err("Discarding thread maps for %d:%d\n",
 442                                th->pid_, th->tid);
 443                 maps__put(th->maps);
 444         }
 445
 446         th->maps = maps__get(leader->maps);
 447 out_put:
 448         thread__put(leader);
 449         return;
 450 out_err:
 451         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 452         goto out_put;
 453 }
 454
 455 /*
 456  * Front-end cache - TID lookups come in blocks,
 457  * so most of the time we dont have to look up
 458  * the full rbtree:
 459  */
 460 static struct thread*
 461 __threads__get_last_match(struct threads *threads, struct machine *machine,
 462                           int pid, int tid)
 463 {
 464         struct thread *th;
 465
 466         th = threads->last_match;
 467         if (th != NULL) {
 468                 if (th->tid == tid) {
 469                         machine__update_thread_pid(machine, th, pid);
 470                         return thread__get(th);
 471                 }
 472
 473                 threads->last_match = NULL;
 474         }
 475
 476         return NULL;
 477 }
 478
 479 static struct thread*
 480 threads__get_last_match(struct threads *threads, struct machine *machine,
 481                         int pid, int tid)
 482 {
 483         struct thread *th = NULL;
 484
 485         if (perf_singlethreaded)
 486                 th = __threads__get_last_match(threads, machine, pid, tid);
 487
 488         return th;
 489 }
 490
 491 static void
 492 __threads__set_last_match(struct threads *threads, struct thread *th)
 493 {
 494         threads->last_match = th;
 495 }
 496
 497 static void
 498 threads__set_last_match(struct threads *threads, struct thread *th)
 499 {
 500         if (perf_singlethreaded)
 501                 __threads__set_last_match(threads, th);
 502 }
 503
 504 /*
 505  * Caller must eventually drop thread->refcnt returned with a successful
 506  * lookup/new thread inserted.
 507  */
 508 static struct thread *____machine__findnew_thread(struct machine *machine,
 509                                                   struct threads *threads,
 510                                                   pid_t pid, pid_t tid,
 511                                                   bool create)
 512 {
 513         struct rb_node **p = &threads->entries.rb_root.rb_node;
 514         struct rb_node *parent = NULL;
 515         struct thread *th;
 516         bool leftmost = true;
 517
 518         th = threads__get_last_match(threads, machine, pid, tid);
 519         if (th)
 520                 return th;
 521
 522         while (*p != NULL) {
 523                 parent = *p;
 524                 th = rb_entry(parent, struct thread, rb_node);
 525
 526                 if (th->tid == tid) {
 527                         threads__set_last_match(threads, th);
 528                         machine__update_thread_pid(machine, th, pid);
 529                         return thread__get(th);
 530                 }
 531
 532                 if (tid < th->tid)
 533                         p = &(*p)->rb_left;
 534                 else {
 535                         p = &(*p)->rb_right;
 536                         leftmost = false;
 537                 }
 538         }
 539
 540         if (!create)
 541                 return NULL;
 542
 543         th = thread__new(pid, tid);
 544         if (th != NULL) {
 545                 rb_link_node(&th->rb_node, parent, p);
 546                 rb_insert_color_cached(&th->rb_node, &threads->entries, leftmost);
 547
 548                 /*
 549                  * We have to initialize maps separately after rb tree is updated.
 550                  *
 551                  * The reason is that we call machine__findnew_thread
 552                  * within thread__init_maps to find the thread
 553                  * leader and that would screwed the rb tree.
 554                  */
 555                 if (thread__init_maps(th, machine)) {
 556                         rb_erase_cached(&th->rb_node, &threads->entries);
 557                         RB_CLEAR_NODE(&th->rb_node);
 558                         thread__put(th);
 559                         return NULL;
 560                 }
 561                 /*
 562                  * It is now in the rbtree, get a ref
 563                  */
 564                 thread__get(th);
 565                 threads__set_last_match(threads, th);
 566                 ++threads->nr;
 567         }
 568
 569         return th;
 570 }
 571
 572 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 573 {
 574         return ____machine__findnew_thread(machine, machine__threads(machine, tid), pid, tid, true);
 575 }
 576
 577 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 578                                        pid_t tid)
 579 {
 580         struct threads *threads = machine__threads(machine, tid);
 581         struct thread *th;
 582
 583         down_write(&threads->lock);
 584         th = __machine__findnew_thread(machine, pid, tid);
 585         up_write(&threads->lock);
 586         return th;
 587 }
 588
 589 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 590                                     pid_t tid)
 591 {
 592         struct threads *threads = machine__threads(machine, tid);
 593         struct thread *th;
 594
 595         down_read(&threads->lock);
 596         th =  ____machine__findnew_thread(machine, threads, pid, tid, false);
 597         up_read(&threads->lock);
 598         return th;
 599 }
 600
 601 /*
 602  * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
 603  * So here a single thread is created for that, but actually there is a separate
 604  * idle task per cpu, so there should be one 'struct thread' per cpu, but there
 605  * is only 1. That causes problems for some tools, requiring workarounds. For
 606  * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
 607  */
 608 struct thread *machine__idle_thread(struct machine *machine)
 609 {
 610         struct thread *thread = machine__findnew_thread(machine, 0, 0);
 611
 612         if (!thread || thread__set_comm(thread, "swapper", 0) ||
 613             thread__set_namespaces(thread, 0, NULL))
 614                 pr_err("problem inserting idle task for machine pid %d\n", machine->pid);
 615
 616         return thread;
 617 }
 618
 619 struct comm *machine__thread_exec_comm(struct machine *machine,
 620                                        struct thread *thread)
 621 {
 622         if (machine->comm_exec)
 623                 return thread__exec_comm(thread);
 624         else
 625                 return thread__comm(thread);
 626 }
 627
 628 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 629                                 struct perf_sample *sample)
 630 {
 631         struct thread *thread = machine__findnew_thread(machine,
 632                                                         event->comm.pid,
 633                                                         event->comm.tid);
 634         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 635         int err = 0;
 636
 637         if (exec)
 638                 machine->comm_exec = true;
 639
 640         if (dump_trace)
 641                 perf_event__fprintf_comm(event, stdout);
 642
 643         if (thread == NULL ||
 644             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 645                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 646                 err = -1;
 647         }
 648
 649         thread__put(thread);
 650
 651         return err;
 652 }
 653
 654 int machine__process_namespaces_event(struct machine *machine __maybe_unused,
 655                                       union perf_event *event,
 656                                       struct perf_sample *sample __maybe_unused)
 657 {
 658         struct thread *thread = machine__findnew_thread(machine,
 659                                                         event->namespaces.pid,
 660                                                         event->namespaces.tid);
 661         int err = 0;
 662
 663         WARN_ONCE(event->namespaces.nr_namespaces > NR_NAMESPACES,
 664                   "\nWARNING: kernel seems to support more namespaces than perf"
 665                   " tool.\nTry updating the perf tool..\n\n");
 666
 667         WARN_ONCE(event->namespaces.nr_namespaces < NR_NAMESPACES,
 668                   "\nWARNING: perf tool seems to support more namespaces than"
 669                   " the kernel.\nTry updating the kernel..\n\n");
 670
 671         if (dump_trace)
 672                 perf_event__fprintf_namespaces(event, stdout);
 673
 674         if (thread == NULL ||
 675             thread__set_namespaces(thread, sample->time, &event->namespaces)) {
 676                 dump_printf("problem processing PERF_RECORD_NAMESPACES, skipping event.\n");
 677                 err = -1;
 678         }
 679
 680         thread__put(thread);
 681
 682         return err;
 683 }
 684
 685 int machine__process_cgroup_event(struct machine *machine,
 686                                   union perf_event *event,
 687                                   struct perf_sample *sample __maybe_unused)
 688 {
 689         struct cgroup *cgrp;
 690
 691         if (dump_trace)
 692                 perf_event__fprintf_cgroup(event, stdout);
 693
 694         cgrp = cgroup__findnew(machine->env, event->cgroup.id, event->cgroup.path);
 695         if (cgrp == NULL)
 696                 return -ENOMEM;
 697
 698         return 0;
 699 }
 700
 701 int machine__process_lost_event(struct machine *machine __maybe_unused,
 702                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 703 {
 704         dump_printf(": id:%" PRI_lu64 ": lost:%" PRI_lu64 "\n",
 705                     event->lost.id, event->lost.lost);
 706         return 0;
 707 }
 708
 709 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 710                                         union perf_event *event, struct perf_sample *sample)
 711 {
 712         dump_printf(": id:%" PRIu64 ": lost samples :%" PRI_lu64 "\n",
 713                     sample->id, event->lost_samples.lost);
 714         return 0;
 715 }
 716
 717 static struct dso *machine__findnew_module_dso(struct machine *machine,
 718                                                struct kmod_path *m,
 719                                                const char *filename)
 720 {
 721         struct dso *dso;
 722
 723         down_write(&machine->dsos.lock);
 724
 725         dso = __dsos__find(&machine->dsos, m->name, true);
 726         if (!dso) {
 727                 dso = __dsos__addnew(&machine->dsos, m->name);
 728                 if (dso == NULL)
 729                         goto out_unlock;
 730
 731                 dso__set_module_info(dso, m, machine);
 732                 dso__set_long_name(dso, strdup(filename), true);
 733                 dso->kernel = DSO_SPACE__KERNEL;
 734         }
 735
 736         dso__get(dso);
 737 out_unlock:
 738         up_write(&machine->dsos.lock);
 739         return dso;
 740 }
 741
 742 int machine__process_aux_event(struct machine *machine __maybe_unused,
 743                                union perf_event *event)
 744 {
 745         if (dump_trace)
 746                 perf_event__fprintf_aux(event, stdout);
 747         return 0;
 748 }
 749
 750 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 751                                         union perf_event *event)
 752 {
 753         if (dump_trace)
 754                 perf_event__fprintf_itrace_start(event, stdout);
 755         return 0;
 756 }
 757
 758 int machine__process_switch_event(struct machine *machine __maybe_unused,
 759                                   union perf_event *event)
 760 {
 761         if (dump_trace)
 762                 perf_event__fprintf_switch(event, stdout);
 763         return 0;
 764 }
 765
 766 static int machine__process_ksymbol_register(struct machine *machine,
 767                                              union perf_event *event,
 768                                              struct perf_sample *sample __maybe_unused)
 769 {
 770         struct symbol *sym;
 771         struct map *map = maps__find(&machine->kmaps, event->ksymbol.addr);
 772
 773         if (!map) {
 774                 struct dso *dso = dso__new(event->ksymbol.name);
 775
 776                 if (dso) {
 777                         dso->kernel = DSO_SPACE__KERNEL;
 778                         map = map__new2(0, dso);
 779                 }
 780
 781                 if (!dso || !map) {
 782                         dso__put(dso);
 783                         return -ENOMEM;
 784                 }
 785
 786                 if (event->ksymbol.ksym_type == PERF_RECORD_KSYMBOL_TYPE_OOL) {
 787                         map->dso->binary_type = DSO_BINARY_TYPE__OOL;
 788                         map->dso->data.file_size = event->ksymbol.len;
 789                         dso__set_loaded(map->dso);
 790                 }
 791
 792                 map->start = event->ksymbol.addr;
 793                 map->end = map->start + event->ksymbol.len;
 794                 maps__insert(&machine->kmaps, map);
 795                 dso__set_loaded(dso);
 796
 797                 if (is_bpf_image(event->ksymbol.name)) {
 798                         dso->binary_type = DSO_BINARY_TYPE__BPF_IMAGE;
 799                         dso__set_long_name(dso, "", false);
 800                 }
 801         }
 802
 803         sym = symbol__new(map->map_ip(map, map->start),
 804                           event->ksymbol.len,
 805                           0, 0, event->ksymbol.name);
 806         if (!sym)
 807                 return -ENOMEM;
 808         dso__insert_symbol(map->dso, sym);
 809         return 0;
 810 }
 811
 812 static int machine__process_ksymbol_unregister(struct machine *machine,
 813                                                union perf_event *event,
 814                                                struct perf_sample *sample __maybe_unused)
 815 {
 816         struct symbol *sym;
 817         struct map *map;
 818
 819         map = maps__find(&machine->kmaps, event->ksymbol.addr);
 820         if (!map)
 821                 return 0;
 822
 823         if (map != machine->vmlinux_map)
 824                 maps__remove(&machine->kmaps, map);
 825         else {
 826                 sym = dso__find_symbol(map->dso, map->map_ip(map, map->start));
 827                 if (sym)
 828                         dso__delete_symbol(map->dso, sym);
 829         }
 830
 831         return 0;
 832 }
 833
 834 int machine__process_ksymbol(struct machine *machine __maybe_unused,
 835                              union perf_event *event,
 836                              struct perf_sample *sample)
 837 {
 838         if (dump_trace)
 839                 perf_event__fprintf_ksymbol(event, stdout);
 840
 841         if (event->ksymbol.flags & PERF_RECORD_KSYMBOL_FLAGS_UNREGISTER)
 842                 return machine__process_ksymbol_unregister(machine, event,
 843                                                            sample);
 844         return machine__process_ksymbol_register(machine, event, sample);
 845 }
 846
 847 int machine__process_text_poke(struct machine *machine, union perf_event *event,
 848                                struct perf_sample *sample __maybe_unused)
 849 {
 850         struct map *map = maps__find(&machine->kmaps, event->text_poke.addr);
 851         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 852
 853         if (dump_trace)
 854                 perf_event__fprintf_text_poke(event, machine, stdout);
 855
 856         if (!event->text_poke.new_len)
 857                 return 0;
 858
 859         if (cpumode != PERF_RECORD_MISC_KERNEL) {
 860                 pr_debug("%s: unsupported cpumode - ignoring\n", __func__);
 861                 return 0;
 862         }
 863
 864         if (map && map->dso) {
 865                 u8 *new_bytes = event->text_poke.bytes + event->text_poke.old_len;
 866                 int ret;
 867
 868                 /*
 869                  * Kernel maps might be changed when loading symbols so loading
 870                  * must be done prior to using kernel maps.
 871                  */
 872                 map__load(map);
 873                 ret = dso__data_write_cache_addr(map->dso, map, machine,
 874                                                  event->text_poke.addr,
 875                                                  new_bytes,
 876                                                  event->text_poke.new_len);
 877                 if (ret != event->text_poke.new_len)
 878                         pr_debug("Failed to write kernel text poke at %#" PRI_lx64 "\n",
 879                                  event->text_poke.addr);
 880         } else {
 881                 pr_debug("Failed to find kernel text poke address map for %#" PRI_lx64 "\n",
 882                          event->text_poke.addr);
 883         }
 884
 885         return 0;
 886 }
 887
 888 static struct map *machine__addnew_module_map(struct machine *machine, u64 start,
 889                                               const char *filename)
 890 {
 891         struct map *map = NULL;
 892         struct kmod_path m;
 893         struct dso *dso;
 894
 895         if (kmod_path__parse_name(&m, filename))
 896                 return NULL;
 897
 898         dso = machine__findnew_module_dso(machine, &m, filename);
 899         if (dso == NULL)
 900                 goto out;
 901
 902         map = map__new2(start, dso);
 903         if (map == NULL)
 904                 goto out;
 905
 906         maps__insert(&machine->kmaps, map);
 907
 908         /* Put the map here because maps__insert already got it */
 909         map__put(map);
 910 out:
 911         /* put the dso here, corresponding to  machine__findnew_module_dso */
 912         dso__put(dso);
 913         zfree(&m.name);
 914         return map;
 915 }
 916
 917 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 918 {
 919         struct rb_node *nd;
 920         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 921
 922         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 923                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 924                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 925         }
 926
 927         return ret;
 928 }
 929
 930 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 931                                      bool (skip)(struct dso *dso, int parm), int parm)
 932 {
 933         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 934 }
 935
 936 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 937                                      bool (skip)(struct dso *dso, int parm), int parm)
 938 {
 939         struct rb_node *nd;
 940         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 941
 942         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
 943                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 944                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 945         }
 946         return ret;
 947 }
 948
 949 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 950 {
 951         int i;
 952         size_t printed = 0;
 953         struct dso *kdso = machine__kernel_dso(machine);
 954
 955         if (kdso->has_build_id) {
 956                 char filename[PATH_MAX];
 957                 if (dso__build_id_filename(kdso, filename, sizeof(filename),
 958                                            false))
 959                         printed += fprintf(fp, "[0] %s\n", filename);
 960         }
 961
 962         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 963                 printed += fprintf(fp, "[%d] %s\n",
 964                                    i + kdso->has_build_id, vmlinux_path[i]);
 965
 966         return printed;
 967 }
 968
 969 size_t machine__fprintf(struct machine *machine, FILE *fp)
 970 {
 971         struct rb_node *nd;
 972         size_t ret;
 973         int i;
 974
 975         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
 976                 struct threads *threads = &machine->threads[i];
 977
 978                 down_read(&threads->lock);
 979
 980                 ret = fprintf(fp, "Threads: %u\n", threads->nr);
 981
 982                 for (nd = rb_first_cached(&threads->entries); nd;
 983                      nd = rb_next(nd)) {
 984                         struct thread *pos = rb_entry(nd, struct thread, rb_node);
 985
 986                         ret += thread__fprintf(pos, fp);
 987                 }
 988
 989                 up_read(&threads->lock);
 990         }
 991         return ret;
 992 }
 993
 994 static struct dso *machine__get_kernel(struct machine *machine)
 995 {
 996         const char *vmlinux_name = machine->mmap_name;
 997         struct dso *kernel;
 998
 999         if (machine__is_host(machine)) {
1000                 if (symbol_conf.vmlinux_name)
1001                         vmlinux_name = symbol_conf.vmlinux_name;
1002
1003                 kernel = machine__findnew_kernel(machine, vmlinux_name,
1004                                                  "[kernel]", DSO_SPACE__KERNEL);
1005         } else {
1006                 if (symbol_conf.default_guest_vmlinux_name)
1007                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
1008
1009                 kernel = machine__findnew_kernel(machine, vmlinux_name,
1010                                                  "[guest.kernel]",
1011                                                  DSO_SPACE__KERNEL_GUEST);
1012         }
1013
1014         if (kernel != NULL && (!kernel->has_build_id))
1015                 dso__read_running_kernel_build_id(kernel, machine);
1016
1017         return kernel;
1018 }
1019
1020 struct process_args {
1021         u64 start;
1022 };
1023
1024 void machine__get_kallsyms_filename(struct machine *machine, char *buf,
1025                                     size_t bufsz)
1026 {
1027         if (machine__is_default_guest(machine))
1028                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
1029         else
1030                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
1031 }
1032
1033 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
1034
1035 /* Figure out the start address of kernel map from /proc/kallsyms.
1036  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
1037  * symbol_name if it's not that important.
1038  */
1039 static int machine__get_running_kernel_start(struct machine *machine,
1040                                              const char **symbol_name,
1041                                              u64 *start, u64 *end)
1042 {
1043         char filename[PATH_MAX];
1044         int i, err = -1;
1045         const char *name;
1046         u64 addr = 0;
1047
1048         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
1049
1050         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
1051                 return 0;
1052
1053         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
1054                 err = kallsyms__get_function_start(filename, name, &addr);
1055                 if (!err)
1056                         break;
1057         }
1058
1059         if (err)
1060                 return -1;
1061
1062         if (symbol_name)
1063                 *symbol_name = name;
1064
1065         *start = addr;
1066
1067         err = kallsyms__get_function_start(filename, "_etext", &addr);
1068         if (!err)
1069                 *end = addr;
1070
1071         return 0;
1072 }
1073
1074 int machine__create_extra_kernel_map(struct machine *machine,
1075                                      struct dso *kernel,
1076                                      struct extra_kernel_map *xm)
1077 {
1078         struct kmap *kmap;
1079         struct map *map;
1080
1081         map = map__new2(xm->start, kernel);
1082         if (!map)
1083                 return -1;
1084
1085         map->end   = xm->end;
1086         map->pgoff = xm->pgoff;
1087
1088         kmap = map__kmap(map);
1089
1090         strlcpy(kmap->name, xm->name, KMAP_NAME_LEN);
1091
1092         maps__insert(&machine->kmaps, map);
1093
1094         pr_debug2("Added extra kernel map %s %" PRIx64 "-%" PRIx64 "\n",
1095                   kmap->name, map->start, map->end);
1096
1097         map__put(map);
1098
1099         return 0;
1100 }
1101
1102 static u64 find_entry_trampoline(struct dso *dso)
1103 {
1104         /* Duplicates are removed so lookup all aliases */
1105         const char *syms[] = {
1106                 "_entry_trampoline",
1107                 "__entry_trampoline_start",
1108                 "entry_SYSCALL_64_trampoline",
1109         };
1110         struct symbol *sym = dso__first_symbol(dso);
1111         unsigned int i;
1112
1113         for (; sym; sym = dso__next_symbol(sym)) {
1114                 if (sym->binding != STB_GLOBAL)
1115                         continue;
1116                 for (i = 0; i < ARRAY_SIZE(syms); i++) {
1117                         if (!strcmp(sym->name, syms[i]))
1118                                 return sym->start;
1119                 }
1120         }
1121
1122         return 0;
1123 }
1124
1125 /*
1126  * These values can be used for kernels that do not have symbols for the entry
1127  * trampolines in kallsyms.
1128  */
1129 #define X86_64_CPU_ENTRY_AREA_PER_CPU   0xfffffe0000000000ULL
1130 #define X86_64_CPU_ENTRY_AREA_SIZE      0x2c000
1131 #define X86_64_ENTRY_TRAMPOLINE         0x6000
1132
1133 /* Map x86_64 PTI entry trampolines */
1134 int machine__map_x86_64_entry_trampolines(struct machine *machine,
1135                                           struct dso *kernel)
1136 {
1137         struct maps *kmaps = &machine->kmaps;
1138         int nr_cpus_avail, cpu;
1139         bool found = false;
1140         struct map *map;
1141         u64 pgoff;
1142
1143         /*
1144          * In the vmlinux case, pgoff is a virtual address which must now be
1145          * mapped to a vmlinux offset.
1146          */
1147         maps__for_each_entry(kmaps, map) {
1148                 struct kmap *kmap = __map__kmap(map);
1149                 struct map *dest_map;
1150
1151                 if (!kmap || !is_entry_trampoline(kmap->name))
1152                         continue;
1153
1154                 dest_map = maps__find(kmaps, map->pgoff);
1155                 if (dest_map != map)
1156                         map->pgoff = dest_map->map_ip(dest_map, map->pgoff);
1157                 found = true;
1158         }
1159         if (found || machine->trampolines_mapped)
1160                 return 0;
1161
1162         pgoff = find_entry_trampoline(kernel);
1163         if (!pgoff)
1164                 return 0;
1165
1166         nr_cpus_avail = machine__nr_cpus_avail(machine);
1167
1168         /* Add a 1 page map for each CPU's entry trampoline */
1169         for (cpu = 0; cpu < nr_cpus_avail; cpu++) {
1170                 u64 va = X86_64_CPU_ENTRY_AREA_PER_CPU +
1171                          cpu * X86_64_CPU_ENTRY_AREA_SIZE +
1172                          X86_64_ENTRY_TRAMPOLINE;
1173                 struct extra_kernel_map xm = {
1174                         .start = va,
1175                         .end   = va + page_size,
1176                         .pgoff = pgoff,
1177                 };
1178
1179                 strlcpy(xm.name, ENTRY_TRAMPOLINE_NAME, KMAP_NAME_LEN);
1180
1181                 if (machine__create_extra_kernel_map(machine, kernel, &xm) < 0)
1182                         return -1;
1183         }
1184
1185         machine->trampolines_mapped = nr_cpus_avail;
1186
1187         return 0;
1188 }
1189
1190 int __weak machine__create_extra_kernel_maps(struct machine *machine __maybe_unused,
1191                                              struct dso *kernel __maybe_unused)
1192 {
1193         return 0;
1194 }
1195
1196 static int
1197 __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
1198 {
1199         /* In case of renewal the kernel map, destroy previous one */
1200         machine__destroy_kernel_maps(machine);
1201
1202         machine->vmlinux_map = map__new2(0, kernel);
1203         if (machine->vmlinux_map == NULL)
1204                 return -1;
1205
1206         machine->vmlinux_map->map_ip = machine->vmlinux_map->unmap_ip = identity__map_ip;
1207         maps__insert(&machine->kmaps, machine->vmlinux_map);
1208         return 0;
1209 }
1210
1211 void machine__destroy_kernel_maps(struct machine *machine)
1212 {
1213         struct kmap *kmap;
1214         struct map *map = machine__kernel_map(machine);
1215
1216         if (map == NULL)
1217                 return;
1218
1219         kmap = map__kmap(map);
1220         maps__remove(&machine->kmaps, map);
1221         if (kmap && kmap->ref_reloc_sym) {
1222                 zfree((char **)&kmap->ref_reloc_sym->name);
1223                 zfree(&kmap->ref_reloc_sym);
1224         }
1225
1226         map__zput(machine->vmlinux_map);
1227 }
1228
1229 int machines__create_guest_kernel_maps(struct machines *machines)
1230 {
1231         int ret = 0;
1232         struct dirent **namelist = NULL;
1233         int i, items = 0;
1234         char path[PATH_MAX];
1235         pid_t pid;
1236         char *endp;
1237
1238         if (symbol_conf.default_guest_vmlinux_name ||
1239             symbol_conf.default_guest_modules ||
1240             symbol_conf.default_guest_kallsyms) {
1241                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
1242         }
1243
1244         if (symbol_conf.guestmount) {
1245                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
1246                 if (items <= 0)
1247                         return -ENOENT;
1248                 for (i = 0; i < items; i++) {
1249                         if (!isdigit(namelist[i]->d_name[0])) {
1250                                 /* Filter out . and .. */
1251                                 continue;
1252                         }
1253                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
1254                         if ((*endp != '\0') ||
1255                             (endp == namelist[i]->d_name) ||
1256                             (errno == ERANGE)) {
1257                                 pr_debug("invalid directory (%s). Skipping.\n",
1258                                          namelist[i]->d_name);
1259                                 continue;
1260                         }
1261                         sprintf(path, "%s/%s/proc/kallsyms",
1262                                 symbol_conf.guestmount,
1263                                 namelist[i]->d_name);
1264                         ret = access(path, R_OK);
1265                         if (ret) {
1266                                 pr_debug("Can't access file %s\n", path);
1267                                 goto failure;
1268                         }
1269                         machines__create_kernel_maps(machines, pid);
1270                 }
1271 failure:
1272                 free(namelist);
1273         }
1274
1275         return ret;
1276 }
1277
1278 void machines__destroy_kernel_maps(struct machines *machines)
1279 {
1280         struct rb_node *next = rb_first_cached(&machines->guests);
1281
1282         machine__destroy_kernel_maps(&machines->host);
1283
1284         while (next) {
1285                 struct machine *pos = rb_entry(next, struct machine, rb_node);
1286
1287                 next = rb_next(&pos->rb_node);
1288                 rb_erase_cached(&pos->rb_node, &machines->guests);
1289                 machine__delete(pos);
1290         }
1291 }
1292
1293 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
1294 {
1295         struct machine *machine = machines__findnew(machines, pid);
1296
1297         if (machine == NULL)
1298                 return -1;
1299
1300         return machine__create_kernel_maps(machine);
1301 }
1302
1303 int machine__load_kallsyms(struct machine *machine, const char *filename)
1304 {
1305         struct map *map = machine__kernel_map(machine);
1306         int ret = __dso__load_kallsyms(map->dso, filename, map, true);
1307
1308         if (ret > 0) {
1309                 dso__set_loaded(map->dso);
1310                 /*
1311                  * Since /proc/kallsyms will have multiple sessions for the
1312                  * kernel, with modules between them, fixup the end of all
1313                  * sections.
1314                  */
1315                 maps__fixup_end(&machine->kmaps);
1316         }
1317
1318         return ret;
1319 }
1320
1321 int machine__load_vmlinux_path(struct machine *machine)
1322 {
1323         struct map *map = machine__kernel_map(machine);
1324         int ret = dso__load_vmlinux_path(map->dso, map);
1325
1326         if (ret > 0)
1327                 dso__set_loaded(map->dso);
1328
1329         return ret;
1330 }
1331
1332 static char *get_kernel_version(const char *root_dir)
1333 {
1334         char version[PATH_MAX];
1335         FILE *file;
1336         char *name, *tmp;
1337         const char *prefix = "Linux version ";
1338
1339         sprintf(version, "%s/proc/version", root_dir);
1340         file = fopen(version, "r");
1341         if (!file)
1342                 return NULL;
1343
1344         tmp = fgets(version, sizeof(version), file);
1345         fclose(file);
1346         if (!tmp)
1347                 return NULL;
1348
1349         name = strstr(version, prefix);
1350         if (!name)
1351                 return NULL;
1352         name += strlen(prefix);
1353         tmp = strchr(name, ' ');
1354         if (tmp)
1355                 *tmp = '\0';
1356
1357         return strdup(name);
1358 }
1359
1360 static bool is_kmod_dso(struct dso *dso)
1361 {
1362         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
1363                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
1364 }
1365
1366 static int maps__set_module_path(struct maps *maps, const char *path, struct kmod_path *m)
1367 {
1368         char *long_name;
1369         struct map *map = maps__find_by_name(maps, m->name);
1370
1371         if (map == NULL)
1372                 return 0;
1373
1374         long_name = strdup(path);
1375         if (long_name == NULL)
1376                 return -ENOMEM;
1377
1378         dso__set_long_name(map->dso, long_name, true);
1379         dso__kernel_module_get_build_id(map->dso, "");
1380
1381         /*
1382          * Full name could reveal us kmod compression, so
1383          * we need to update the symtab_type if needed.
1384          */
1385         if (m->comp && is_kmod_dso(map->dso)) {
1386                 map->dso->symtab_type++;
1387                 map->dso->comp = m->comp;
1388         }
1389
1390         return 0;
1391 }
1392
1393 static int maps__set_modules_path_dir(struct maps *maps, const char *dir_name, int depth)
1394 {
1395         struct dirent *dent;
1396         DIR *dir = opendir(dir_name);
1397         int ret = 0;
1398
1399         if (!dir) {
1400                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
1401                 return -1;
1402         }
1403
1404         while ((dent = readdir(dir)) != NULL) {
1405                 char path[PATH_MAX];
1406                 struct stat st;
1407
1408                 /*sshfs might return bad dent->d_type, so we have to stat*/
1409                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
1410                 if (stat(path, &st))
1411                         continue;
1412
1413                 if (S_ISDIR(st.st_mode)) {
1414                         if (!strcmp(dent->d_name, ".") ||
1415                             !strcmp(dent->d_name, ".."))
1416                                 continue;
1417
1418                         /* Do not follow top-level source and build symlinks */
1419                         if (depth == 0) {
1420                                 if (!strcmp(dent->d_name, "source") ||
1421                                     !strcmp(dent->d_name, "build"))
1422                                         continue;
1423                         }
1424
1425                         ret = maps__set_modules_path_dir(maps, path, depth + 1);
1426                         if (ret < 0)
1427                                 goto out;
1428                 } else {
1429                         struct kmod_path m;
1430
1431                         ret = kmod_path__parse_name(&m, dent->d_name);
1432                         if (ret)
1433                                 goto out;
1434
1435                         if (m.kmod)
1436                                 ret = maps__set_module_path(maps, path, &m);
1437
1438                         zfree(&m.name);
1439
1440                         if (ret)
1441                                 goto out;
1442                 }
1443         }
1444
1445 out:
1446         closedir(dir);
1447         return ret;
1448 }
1449
1450 static int machine__set_modules_path(struct machine *machine)
1451 {
1452         char *version;
1453         char modules_path[PATH_MAX];
1454
1455         version = get_kernel_version(machine->root_dir);
1456         if (!version)
1457                 return -1;
1458
1459         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1460                  machine->root_dir, version);
1461         free(version);
1462
1463         return maps__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1464 }
1465 int __weak arch__fix_module_text_start(u64 *start __maybe_unused,
1466                                 u64 *size __maybe_unused,
1467                                 const char *name __maybe_unused)
1468 {
1469         return 0;
1470 }
1471
1472 static int machine__create_module(void *arg, const char *name, u64 start,
1473                                   u64 size)
1474 {
1475         struct machine *machine = arg;
1476         struct map *map;
1477
1478         if (arch__fix_module_text_start(&start, &size, name) < 0)
1479                 return -1;
1480
1481         map = machine__addnew_module_map(machine, start, name);
1482         if (map == NULL)
1483                 return -1;
1484         map->end = start + size;
1485
1486         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1487
1488         return 0;
1489 }
1490
1491 static int machine__create_modules(struct machine *machine)
1492 {
1493         const char *modules;
1494         char path[PATH_MAX];
1495
1496         if (machine__is_default_guest(machine)) {
1497                 modules = symbol_conf.default_guest_modules;
1498         } else {
1499                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1500                 modules = path;
1501         }
1502
1503         if (symbol__restricted_filename(modules, "/proc/modules"))
1504                 return -1;
1505
1506         if (modules__parse(modules, machine, machine__create_module))
1507                 return -1;
1508
1509         if (!machine__set_modules_path(machine))
1510                 return 0;
1511
1512         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1513
1514         return 0;
1515 }
1516
1517 static void machine__set_kernel_mmap(struct machine *machine,
1518                                      u64 start, u64 end)
1519 {
1520         machine->vmlinux_map->start = start;
1521         machine->vmlinux_map->end   = end;
1522         /*
1523          * Be a bit paranoid here, some perf.data file came with
1524          * a zero sized synthesized MMAP event for the kernel.
1525          */
1526         if (start == 0 && end == 0)
1527                 machine->vmlinux_map->end = ~0ULL;
1528 }
1529
1530 static void machine__update_kernel_mmap(struct machine *machine,
1531                                      u64 start, u64 end)
1532 {
1533         struct map *map = machine__kernel_map(machine);
1534
1535         map__get(map);
1536         maps__remove(&machine->kmaps, map);
1537
1538         machine__set_kernel_mmap(machine, start, end);
1539
1540         maps__insert(&machine->kmaps, map);
1541         map__put(map);
1542 }
1543
1544 int machine__create_kernel_maps(struct machine *machine)
1545 {
1546         struct dso *kernel = machine__get_kernel(machine);
1547         const char *name = NULL;
1548         struct map *map;
1549         u64 start = 0, end = ~0ULL;
1550         int ret;
1551
1552         if (kernel == NULL)
1553                 return -1;
1554
1555         ret = __machine__create_kernel_maps(machine, kernel);
1556         if (ret < 0)
1557                 goto out_put;
1558
1559         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1560                 if (machine__is_host(machine))
1561                         pr_debug("Problems creating module maps, "
1562                                  "continuing anyway...\n");
1563                 else
1564                         pr_debug("Problems creating module maps for guest %d, "
1565                                  "continuing anyway...\n", machine->pid);
1566         }
1567
1568         if (!machine__get_running_kernel_start(machine, &name, &start, &end)) {
1569                 if (name &&
1570                     map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map, name, start)) {
1571                         machine__destroy_kernel_maps(machine);
1572                         ret = -1;
1573                         goto out_put;
1574                 }
1575
1576                 /*
1577                  * we have a real start address now, so re-order the kmaps
1578                  * assume it's the last in the kmaps
1579                  */
1580                 machine__update_kernel_mmap(machine, start, end);
1581         }
1582
1583         if (machine__create_extra_kernel_maps(machine, kernel))
1584                 pr_debug("Problems creating extra kernel maps, continuing anyway...\n");
1585
1586         if (end == ~0ULL) {
1587                 /* update end address of the kernel map using adjacent module address */
1588                 map = map__next(machine__kernel_map(machine));
1589                 if (map)
1590                         machine__set_kernel_mmap(machine, start, map->start);
1591         }
1592
1593 out_put:
1594         dso__put(kernel);
1595         return ret;
1596 }
1597
1598 static bool machine__uses_kcore(struct machine *machine)
1599 {
1600         struct dso *dso;
1601
1602         list_for_each_entry(dso, &machine->dsos.head, node) {
1603                 if (dso__is_kcore(dso))
1604                         return true;
1605         }
1606
1607         return false;
1608 }
1609
1610 static bool perf_event__is_extra_kernel_mmap(struct machine *machine,
1611                                              struct extra_kernel_map *xm)
1612 {
1613         return machine__is(machine, "x86_64") &&
1614                is_entry_trampoline(xm->name);
1615 }
1616
1617 static int machine__process_extra_kernel_map(struct machine *machine,
1618                                              struct extra_kernel_map *xm)
1619 {
1620         struct dso *kernel = machine__kernel_dso(machine);
1621
1622         if (kernel == NULL)
1623                 return -1;
1624
1625         return machine__create_extra_kernel_map(machine, kernel, xm);
1626 }
1627
1628 static int machine__process_kernel_mmap_event(struct machine *machine,
1629                                               struct extra_kernel_map *xm,
1630                                               struct build_id *bid)
1631 {
1632         struct map *map;
1633         enum dso_space_type dso_space;
1634         bool is_kernel_mmap;
1635
1636         /* If we have maps from kcore then we do not need or want any others */
1637         if (machine__uses_kcore(machine))
1638                 return 0;
1639
1640         if (machine__is_host(machine))
1641                 dso_space = DSO_SPACE__KERNEL;
1642         else
1643                 dso_space = DSO_SPACE__KERNEL_GUEST;
1644
1645         is_kernel_mmap = memcmp(xm->name, machine->mmap_name,
1646                                 strlen(machine->mmap_name) - 1) == 0;
1647         if (xm->name[0] == '/' ||
1648             (!is_kernel_mmap && xm->name[0] == '[')) {
1649                 map = machine__addnew_module_map(machine, xm->start,
1650                                                  xm->name);
1651                 if (map == NULL)
1652                         goto out_problem;
1653
1654                 map->end = map->start + xm->end - xm->start;
1655
1656                 if (build_id__is_defined(bid))
1657                         dso__set_build_id(map->dso, bid);
1658
1659         } else if (is_kernel_mmap) {
1660                 const char *symbol_name = (xm->name + strlen(machine->mmap_name));
1661                 /*
1662                  * Should be there already, from the build-id table in
1663                  * the header.
1664                  */
1665                 struct dso *kernel = NULL;
1666                 struct dso *dso;
1667
1668                 down_read(&machine->dsos.lock);
1669
1670                 list_for_each_entry(dso, &machine->dsos.head, node) {
1671
1672                         /*
1673                          * The cpumode passed to is_kernel_module is not the
1674                          * cpumode of *this* event. If we insist on passing
1675                          * correct cpumode to is_kernel_module, we should
1676                          * record the cpumode when we adding this dso to the
1677                          * linked list.
1678                          *
1679                          * However we don't really need passing correct
1680                          * cpumode.  We know the correct cpumode must be kernel
1681                          * mode (if not, we should not link it onto kernel_dsos
1682                          * list).
1683                          *
1684                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1685                          * is_kernel_module() treats it as a kernel cpumode.
1686                          */
1687
1688                         if (!dso->kernel ||
1689                             is_kernel_module(dso->long_name,
1690                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1691                                 continue;
1692
1693
1694                         kernel = dso;
1695                         break;
1696                 }
1697
1698                 up_read(&machine->dsos.lock);
1699
1700                 if (kernel == NULL)
1701                         kernel = machine__findnew_dso(machine, machine->mmap_name);
1702                 if (kernel == NULL)
1703                         goto out_problem;
1704
1705                 kernel->kernel = dso_space;
1706                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1707                         dso__put(kernel);
1708                         goto out_problem;
1709                 }
1710
1711                 if (strstr(kernel->long_name, "vmlinux"))
1712                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1713
1714                 machine__update_kernel_mmap(machine, xm->start, xm->end);
1715
1716                 if (build_id__is_defined(bid))
1717                         dso__set_build_id(kernel, bid);
1718
1719                 /*
1720                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1721                  * symbol. Effectively having zero here means that at record
1722                  * time /proc/sys/kernel/kptr_restrict was non zero.
1723                  */
1724                 if (xm->pgoff != 0) {
1725                         map__set_kallsyms_ref_reloc_sym(machine->vmlinux_map,
1726                                                         symbol_name,
1727                                                         xm->pgoff);
1728                 }
1729
1730                 if (machine__is_default_guest(machine)) {
1731                         /*
1732                          * preload dso of guest kernel and modules
1733                          */
1734                         dso__load(kernel, machine__kernel_map(machine));
1735                 }
1736         } else if (perf_event__is_extra_kernel_mmap(machine, xm)) {
1737                 return machine__process_extra_kernel_map(machine, xm);
1738         }
1739         return 0;
1740 out_problem:
1741         return -1;
1742 }
1743
1744 int machine__process_mmap2_event(struct machine *machine,
1745                                  union perf_event *event,
1746                                  struct perf_sample *sample)
1747 {
1748         struct thread *thread;
1749         struct map *map;
1750         struct dso_id dso_id = {
1751                 .maj = event->mmap2.maj,
1752                 .min = event->mmap2.min,
1753                 .ino = event->mmap2.ino,
1754                 .ino_generation = event->mmap2.ino_generation,
1755         };
1756         struct build_id __bid, *bid = NULL;
1757         int ret = 0;
1758
1759         if (dump_trace)
1760                 perf_event__fprintf_mmap2(event, stdout);
1761
1762         if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) {
1763                 bid = &__bid;
1764                 build_id__init(bid, event->mmap2.build_id, event->mmap2.build_id_size);
1765         }
1766
1767         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1768             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1769                 struct extra_kernel_map xm = {
1770                         .start = event->mmap2.start,
1771                         .end   = event->mmap2.start + event->mmap2.len,
1772                         .pgoff = event->mmap2.pgoff,
1773                 };
1774
1775                 strlcpy(xm.name, event->mmap2.filename, KMAP_NAME_LEN);
1776                 ret = machine__process_kernel_mmap_event(machine, &xm, bid);
1777                 if (ret < 0)
1778                         goto out_problem;
1779                 return 0;
1780         }
1781
1782         thread = machine__findnew_thread(machine, event->mmap2.pid,
1783                                         event->mmap2.tid);
1784         if (thread == NULL)
1785                 goto out_problem;
1786
1787         map = map__new(machine, event->mmap2.start,
1788                         event->mmap2.len, event->mmap2.pgoff,
1789                         &dso_id, event->mmap2.prot,
1790                         event->mmap2.flags, bid,
1791                         event->mmap2.filename, thread);
1792
1793         if (map == NULL)
1794                 goto out_problem_map;
1795
1796         ret = thread__insert_map(thread, map);
1797         if (ret)
1798                 goto out_problem_insert;
1799
1800         thread__put(thread);
1801         map__put(map);
1802         return 0;
1803
1804 out_problem_insert:
1805         map__put(map);
1806 out_problem_map:
1807         thread__put(thread);
1808 out_problem:
1809         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1810         return 0;
1811 }
1812
1813 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1814                                 struct perf_sample *sample)
1815 {
1816         struct thread *thread;
1817         struct map *map;
1818         u32 prot = 0;
1819         int ret = 0;
1820
1821         if (dump_trace)
1822                 perf_event__fprintf_mmap(event, stdout);
1823
1824         if (sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1825             sample->cpumode == PERF_RECORD_MISC_KERNEL) {
1826                 struct extra_kernel_map xm = {
1827                         .start = event->mmap.start,
1828                         .end   = event->mmap.start + event->mmap.len,
1829                         .pgoff = event->mmap.pgoff,
1830                 };
1831
1832                 strlcpy(xm.name, event->mmap.filename, KMAP_NAME_LEN);
1833                 ret = machine__process_kernel_mmap_event(machine, &xm, NULL);
1834                 if (ret < 0)
1835                         goto out_problem;
1836                 return 0;
1837         }
1838
1839         thread = machine__findnew_thread(machine, event->mmap.pid,
1840                                          event->mmap.tid);
1841         if (thread == NULL)
1842                 goto out_problem;
1843
1844         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_DATA))
1845                 prot = PROT_EXEC;
1846
1847         map = map__new(machine, event->mmap.start,
1848                         event->mmap.len, event->mmap.pgoff,
1849                         NULL, prot, 0, NULL, event->mmap.filename, thread);
1850
1851         if (map == NULL)
1852                 goto out_problem_map;
1853
1854         ret = thread__insert_map(thread, map);
1855         if (ret)
1856                 goto out_problem_insert;
1857
1858         thread__put(thread);
1859         map__put(map);
1860         return 0;
1861
1862 out_problem_insert:
1863         map__put(map);
1864 out_problem_map:
1865         thread__put(thread);
1866 out_problem:
1867         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1868         return 0;
1869 }
1870
1871 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1872 {
1873         struct threads *threads = machine__threads(machine, th->tid);
1874
1875         if (threads->last_match == th)
1876                 threads__set_last_match(threads, NULL);
1877
1878         if (lock)
1879                 down_write(&threads->lock);
1880
1881         BUG_ON(refcount_read(&th->refcnt) == 0);
1882
1883         rb_erase_cached(&th->rb_node, &threads->entries);
1884         RB_CLEAR_NODE(&th->rb_node);
1885         --threads->nr;
1886         /*
1887          * Move it first to the dead_threads list, then drop the reference,
1888          * if this is the last reference, then the thread__delete destructor
1889          * will be called and we will remove it from the dead_threads list.
1890          */
1891         list_add_tail(&th->node, &threads->dead);
1892
1893         /*
1894          * We need to do the put here because if this is the last refcount,
1895          * then we will be touching the threads->dead head when removing the
1896          * thread.
1897          */
1898         thread__put(th);
1899
1900         if (lock)
1901                 up_write(&threads->lock);
1902 }
1903
1904 void machine__remove_thread(struct machine *machine, struct thread *th)
1905 {
1906         return __machine__remove_thread(machine, th, true);
1907 }
1908
1909 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1910                                 struct perf_sample *sample)
1911 {
1912         struct thread *thread = machine__find_thread(machine,
1913                                                      event->fork.pid,
1914                                                      event->fork.tid);
1915         struct thread *parent = machine__findnew_thread(machine,
1916                                                         event->fork.ppid,
1917                                                         event->fork.ptid);
1918         bool do_maps_clone = true;
1919         int err = 0;
1920
1921         if (dump_trace)
1922                 perf_event__fprintf_task(event, stdout);
1923
1924         /*
1925          * There may be an existing thread that is not actually the parent,
1926          * either because we are processing events out of order, or because the
1927          * (fork) event that would have removed the thread was lost. Assume the
1928          * latter case and continue on as best we can.
1929          */
1930         if (parent->pid_ != (pid_t)event->fork.ppid) {
1931                 dump_printf("removing erroneous parent thread %d/%d\n",
1932                             parent->pid_, parent->tid);
1933                 machine__remove_thread(machine, parent);
1934                 thread__put(parent);
1935                 parent = machine__findnew_thread(machine, event->fork.ppid,
1936                                                  event->fork.ptid);
1937         }
1938
1939         /* if a thread currently exists for the thread id remove it */
1940         if (thread != NULL) {
1941                 machine__remove_thread(machine, thread);
1942                 thread__put(thread);
1943         }
1944
1945         thread = machine__findnew_thread(machine, event->fork.pid,
1946                                          event->fork.tid);
1947         /*
1948          * When synthesizing FORK events, we are trying to create thread
1949          * objects for the already running tasks on the machine.
1950          *
1951          * Normally, for a kernel FORK event, we want to clone the parent's
1952          * maps because that is what the kernel just did.
1953          *
1954          * But when synthesizing, this should not be done.  If we do, we end up
1955          * with overlapping maps as we process the synthesized MMAP2 events that
1956          * get delivered shortly thereafter.
1957          *
1958          * Use the FORK event misc flags in an internal way to signal this
1959          * situation, so we can elide the map clone when appropriate.
1960          */
1961         if (event->fork.header.misc & PERF_RECORD_MISC_FORK_EXEC)
1962                 do_maps_clone = false;
1963
1964         if (thread == NULL || parent == NULL ||
1965             thread__fork(thread, parent, sample->time, do_maps_clone) < 0) {
1966                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1967                 err = -1;
1968         }
1969         thread__put(thread);
1970         thread__put(parent);
1971
1972         return err;
1973 }
1974
1975 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1976                                 struct perf_sample *sample __maybe_unused)
1977 {
1978         struct thread *thread = machine__find_thread(machine,
1979                                                      event->fork.pid,
1980                                                      event->fork.tid);
1981
1982         if (dump_trace)
1983                 perf_event__fprintf_task(event, stdout);
1984
1985         if (thread != NULL) {
1986                 thread__exited(thread);
1987                 thread__put(thread);
1988         }
1989
1990         return 0;
1991 }
1992
1993 int machine__process_event(struct machine *machine, union perf_event *event,
1994                            struct perf_sample *sample)
1995 {
1996         int ret;
1997
1998         switch (event->header.type) {
1999         case PERF_RECORD_COMM:
2000                 ret = machine__process_comm_event(machine, event, sample); break;
2001         case PERF_RECORD_MMAP:
2002                 ret = machine__process_mmap_event(machine, event, sample); break;
2003         case PERF_RECORD_NAMESPACES:
2004                 ret = machine__process_namespaces_event(machine, event, sample); break;
2005         case PERF_RECORD_CGROUP:
2006                 ret = machine__process_cgroup_event(machine, event, sample); break;
2007         case PERF_RECORD_MMAP2:
2008                 ret = machine__process_mmap2_event(machine, event, sample); break;
2009         case PERF_RECORD_FORK:
2010                 ret = machine__process_fork_event(machine, event, sample); break;
2011         case PERF_RECORD_EXIT:
2012                 ret = machine__process_exit_event(machine, event, sample); break;
2013         case PERF_RECORD_LOST:
2014                 ret = machine__process_lost_event(machine, event, sample); break;
2015         case PERF_RECORD_AUX:
2016                 ret = machine__process_aux_event(machine, event); break;
2017         case PERF_RECORD_ITRACE_START:
2018                 ret = machine__process_itrace_start_event(machine, event); break;
2019         case PERF_RECORD_LOST_SAMPLES:
2020                 ret = machine__process_lost_samples_event(machine, event, sample); break;
2021         case PERF_RECORD_SWITCH:
2022         case PERF_RECORD_SWITCH_CPU_WIDE:
2023                 ret = machine__process_switch_event(machine, event); break;
2024         case PERF_RECORD_KSYMBOL:
2025                 ret = machine__process_ksymbol(machine, event, sample); break;
2026         case PERF_RECORD_BPF_EVENT:
2027                 ret = machine__process_bpf(machine, event, sample); break;
2028         case PERF_RECORD_TEXT_POKE:
2029                 ret = machine__process_text_poke(machine, event, sample); break;
2030         default:
2031                 ret = -1;
2032                 break;
2033         }
2034
2035         return ret;
2036 }
2037
2038 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
2039 {
2040         if (!regexec(regex, sym->name, 0, NULL, 0))
2041                 return true;
2042         return false;
2043 }
2044
2045 static void ip__resolve_ams(struct thread *thread,
2046                             struct addr_map_symbol *ams,
2047                             u64 ip)
2048 {
2049         struct addr_location al;
2050
2051         memset(&al, 0, sizeof(al));
2052         /*
2053          * We cannot use the header.misc hint to determine whether a
2054          * branch stack address is user, kernel, guest, hypervisor.
2055          * Branches may straddle the kernel/user/hypervisor boundaries.
2056          * Thus, we have to try consecutively until we find a match
2057          * or else, the symbol is unknown
2058          */
2059         thread__find_cpumode_addr_location(thread, ip, &al);
2060
2061         ams->addr = ip;
2062         ams->al_addr = al.addr;
2063         ams->ms.maps = al.maps;
2064         ams->ms.sym = al.sym;
2065         ams->ms.map = al.map;
2066         ams->phys_addr = 0;
2067         ams->data_page_size = 0;
2068 }
2069
2070 static void ip__resolve_data(struct thread *thread,
2071                              u8 m, struct addr_map_symbol *ams,
2072                              u64 addr, u64 phys_addr, u64 daddr_page_size)
2073 {
2074         struct addr_location al;
2075
2076         memset(&al, 0, sizeof(al));
2077
2078         thread__find_symbol(thread, m, addr, &al);
2079
2080         ams->addr = addr;
2081         ams->al_addr = al.addr;
2082         ams->ms.maps = al.maps;
2083         ams->ms.sym = al.sym;
2084         ams->ms.map = al.map;
2085         ams->phys_addr = phys_addr;
2086         ams->data_page_size = daddr_page_size;
2087 }
2088
2089 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
2090                                      struct addr_location *al)
2091 {
2092         struct mem_info *mi = mem_info__new();
2093
2094         if (!mi)
2095                 return NULL;
2096
2097         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
2098         ip__resolve_data(al->thread, al->cpumode, &mi->daddr,
2099                          sample->addr, sample->phys_addr,
2100                          sample->data_page_size);
2101         mi->data_src.val = sample->data_src;
2102
2103         return mi;
2104 }
2105
2106 static char *callchain_srcline(struct map_symbol *ms, u64 ip)
2107 {
2108         struct map *map = ms->map;
2109         char *srcline = NULL;
2110
2111         if (!map || callchain_param.key == CCKEY_FUNCTION)
2112                 return srcline;
2113
2114         srcline = srcline__tree_find(&map->dso->srclines, ip);
2115         if (!srcline) {
2116                 bool show_sym = false;
2117                 bool show_addr = callchain_param.key == CCKEY_ADDRESS;
2118
2119                 srcline = get_srcline(map->dso, map__rip_2objdump(map, ip),
2120                                       ms->sym, show_sym, show_addr, ip);
2121                 srcline__tree_insert(&map->dso->srclines, ip, srcline);
2122         }
2123
2124         return srcline;
2125 }
2126
2127 struct iterations {
2128         int nr_loop_iter;
2129         u64 cycles;
2130 };
2131
2132 static int add_callchain_ip(struct thread *thread,
2133                             struct callchain_cursor *cursor,
2134                             struct symbol **parent,
2135                             struct addr_location *root_al,
2136                             u8 *cpumode,
2137                             u64 ip,
2138                             bool branch,
2139                             struct branch_flags *flags,
2140                             struct iterations *iter,
2141                             u64 branch_from)
2142 {
2143         struct map_symbol ms;
2144         struct addr_location al;
2145         int nr_loop_iter = 0;
2146         u64 iter_cycles = 0;
2147         const char *srcline = NULL;
2148
2149         al.filtered = 0;
2150         al.sym = NULL;
2151         if (!cpumode) {
2152                 thread__find_cpumode_addr_location(thread, ip, &al);
2153         } else {
2154                 if (ip >= PERF_CONTEXT_MAX) {
2155                         switch (ip) {
2156                         case PERF_CONTEXT_HV:
2157                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
2158                                 break;
2159                         case PERF_CONTEXT_KERNEL:
2160                                 *cpumode = PERF_RECORD_MISC_KERNEL;
2161                                 break;
2162                         case PERF_CONTEXT_USER:
2163                                 *cpumode = PERF_RECORD_MISC_USER;
2164                                 break;
2165                         default:
2166                                 pr_debug("invalid callchain context: "
2167                                          "%"PRId64"\n", (s64) ip);
2168                                 /*
2169                                  * It seems the callchain is corrupted.
2170                                  * Discard all.
2171                                  */
2172                                 callchain_cursor_reset(cursor);
2173                                 return 1;
2174                         }
2175                         return 0;
2176                 }
2177                 thread__find_symbol(thread, *cpumode, ip, &al);
2178         }
2179
2180         if (al.sym != NULL) {
2181                 if (perf_hpp_list.parent && !*parent &&
2182                     symbol__match_regex(al.sym, &parent_regex))
2183                         *parent = al.sym;
2184                 else if (have_ignore_callees && root_al &&
2185                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
2186                         /* Treat this symbol as the root,
2187                            forgetting its callees. */
2188                         *root_al = al;
2189                         callchain_cursor_reset(cursor);
2190                 }
2191         }
2192
2193         if (symbol_conf.hide_unresolved && al.sym == NULL)
2194                 return 0;
2195
2196         if (iter) {
2197                 nr_loop_iter = iter->nr_loop_iter;
2198                 iter_cycles = iter->cycles;
2199         }
2200
2201         ms.maps = al.maps;
2202         ms.map = al.map;
2203         ms.sym = al.sym;
2204         srcline = callchain_srcline(&ms, al.addr);
2205         return callchain_cursor_append(cursor, ip, &ms,
2206                                        branch, flags, nr_loop_iter,
2207                                        iter_cycles, branch_from, srcline);
2208 }
2209
2210 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
2211                                            struct addr_location *al)
2212 {
2213         unsigned int i;
2214         const struct branch_stack *bs = sample->branch_stack;
2215         struct branch_entry *entries = perf_sample__branch_entries(sample);
2216         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
2217
2218         if (!bi)
2219                 return NULL;
2220
2221         for (i = 0; i < bs->nr; i++) {
2222                 ip__resolve_ams(al->thread, &bi[i].to, entries[i].to);
2223                 ip__resolve_ams(al->thread, &bi[i].from, entries[i].from);
2224                 bi[i].flags = entries[i].flags;
2225         }
2226         return bi;
2227 }
2228
2229 static void save_iterations(struct iterations *iter,
2230                             struct branch_entry *be, int nr)
2231 {
2232         int i;
2233
2234         iter->nr_loop_iter++;
2235         iter->cycles = 0;
2236
2237         for (i = 0; i < nr; i++)
2238                 iter->cycles += be[i].flags.cycles;
2239 }
2240
2241 #define CHASHSZ 127
2242 #define CHASHBITS 7
2243 #define NO_ENTRY 0xff
2244
2245 #define PERF_MAX_BRANCH_DEPTH 127
2246
2247 /* Remove loops. */
2248 static int remove_loops(struct branch_entry *l, int nr,
2249                         struct iterations *iter)
2250 {
2251         int i, j, off;
2252         unsigned char chash[CHASHSZ];
2253
2254         memset(chash, NO_ENTRY, sizeof(chash));
2255
2256         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
2257
2258         for (i = 0; i < nr; i++) {
2259                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
2260
2261                 /* no collision handling for now */
2262                 if (chash[h] == NO_ENTRY) {
2263                         chash[h] = i;
2264                 } else if (l[chash[h]].from == l[i].from) {
2265                         bool is_loop = true;
2266                         /* check if it is a real loop */
2267                         off = 0;
2268                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
2269                                 if (l[j].from != l[i + off].from) {
2270                                         is_loop = false;
2271                                         break;
2272                                 }
2273                         if (is_loop) {
2274                                 j = nr - (i + off);
2275                                 if (j > 0) {
2276                                         save_iterations(iter + i + off,
2277                                                 l + i, off);
2278
2279                                         memmove(iter + i, iter + i + off,
2280                                                 j * sizeof(*iter));
2281
2282                                         memmove(l + i, l + i + off,
2283                                                 j * sizeof(*l));
2284                                 }
2285
2286                                 nr -= off;
2287                         }
2288                 }
2289         }
2290         return nr;
2291 }
2292
2293 static int lbr_callchain_add_kernel_ip(struct thread *thread,
2294                                        struct callchain_cursor *cursor,
2295                                        struct perf_sample *sample,
2296                                        struct symbol **parent,
2297                                        struct addr_location *root_al,
2298                                        u64 branch_from,
2299                                        bool callee, int end)
2300 {
2301         struct ip_callchain *chain = sample->callchain;
2302         u8 cpumode = PERF_RECORD_MISC_USER;
2303         int err, i;
2304
2305         if (callee) {
2306                 for (i = 0; i < end + 1; i++) {
2307                         err = add_callchain_ip(thread, cursor, parent,
2308                                                root_al, &cpumode, chain->ips[i],
2309                                                false, NULL, NULL, branch_from);
2310                         if (err)
2311                                 return err;
2312                 }
2313                 return 0;
2314         }
2315
2316         for (i = end; i >= 0; i--) {
2317                 err = add_callchain_ip(thread, cursor, parent,
2318                                        root_al, &cpumode, chain->ips[i],
2319                                        false, NULL, NULL, branch_from);
2320                 if (err)
2321                         return err;
2322         }
2323
2324         return 0;
2325 }
2326
2327 static void save_lbr_cursor_node(struct thread *thread,
2328                                  struct callchain_cursor *cursor,
2329                                  int idx)
2330 {
2331         struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2332
2333         if (!lbr_stitch)
2334                 return;
2335
2336         if (cursor->pos == cursor->nr) {
2337                 lbr_stitch->prev_lbr_cursor[idx].valid = false;
2338                 return;
2339         }
2340
2341         if (!cursor->curr)
2342                 cursor->curr = cursor->first;
2343         else
2344                 cursor->curr = cursor->curr->next;
2345         memcpy(&lbr_stitch->prev_lbr_cursor[idx], cursor->curr,
2346                sizeof(struct callchain_cursor_node));
2347
2348         lbr_stitch->prev_lbr_cursor[idx].valid = true;
2349         cursor->pos++;
2350 }
2351
2352 static int lbr_callchain_add_lbr_ip(struct thread *thread,
2353                                     struct callchain_cursor *cursor,
2354                                     struct perf_sample *sample,
2355                                     struct symbol **parent,
2356                                     struct addr_location *root_al,
2357                                     u64 *branch_from,
2358                                     bool callee)
2359 {
2360         struct branch_stack *lbr_stack = sample->branch_stack;
2361         struct branch_entry *entries = perf_sample__branch_entries(sample);
2362         u8 cpumode = PERF_RECORD_MISC_USER;
2363         int lbr_nr = lbr_stack->nr;
2364         struct branch_flags *flags;
2365         int err, i;
2366         u64 ip;
2367
2368         /*
2369          * The curr and pos are not used in writing session. They are cleared
2370          * in callchain_cursor_commit() when the writing session is closed.
2371          * Using curr and pos to track the current cursor node.
2372          */
2373         if (thread->lbr_stitch) {
2374                 cursor->curr = NULL;
2375                 cursor->pos = cursor->nr;
2376                 if (cursor->nr) {
2377                         cursor->curr = cursor->first;
2378                         for (i = 0; i < (int)(cursor->nr - 1); i++)
2379                                 cursor->curr = cursor->curr->next;
2380                 }
2381         }
2382
2383         if (callee) {
2384                 /* Add LBR ip from first entries.to */
2385                 ip = entries[0].to;
2386                 flags = &entries[0].flags;
2387                 *branch_from = entries[0].from;
2388                 err = add_callchain_ip(thread, cursor, parent,
2389                                        root_al, &cpumode, ip,
2390                                        true, flags, NULL,
2391                                        *branch_from);
2392                 if (err)
2393                         return err;
2394
2395                 /*
2396                  * The number of cursor node increases.
2397                  * Move the current cursor node.
2398                  * But does not need to save current cursor node for entry 0.
2399                  * It's impossible to stitch the whole LBRs of previous sample.
2400                  */
2401                 if (thread->lbr_stitch && (cursor->pos != cursor->nr)) {
2402                         if (!cursor->curr)
2403                                 cursor->curr = cursor->first;
2404                         else
2405                                 cursor->curr = cursor->curr->next;
2406                         cursor->pos++;
2407                 }
2408
2409                 /* Add LBR ip from entries.from one by one. */
2410                 for (i = 0; i < lbr_nr; i++) {
2411                         ip = entries[i].from;
2412                         flags = &entries[i].flags;
2413                         err = add_callchain_ip(thread, cursor, parent,
2414                                                root_al, &cpumode, ip,
2415                                                true, flags, NULL,
2416                                                *branch_from);
2417                         if (err)
2418                                 return err;
2419                         save_lbr_cursor_node(thread, cursor, i);
2420                 }
2421                 return 0;
2422         }
2423
2424         /* Add LBR ip from entries.from one by one. */
2425         for (i = lbr_nr - 1; i >= 0; i--) {
2426                 ip = entries[i].from;
2427                 flags = &entries[i].flags;
2428                 err = add_callchain_ip(thread, cursor, parent,
2429                                        root_al, &cpumode, ip,
2430                                        true, flags, NULL,
2431                                        *branch_from);
2432                 if (err)
2433                         return err;
2434                 save_lbr_cursor_node(thread, cursor, i);
2435         }
2436
2437         /* Add LBR ip from first entries.to */
2438         ip = entries[0].to;
2439         flags = &entries[0].flags;
2440         *branch_from = entries[0].from;
2441         err = add_callchain_ip(thread, cursor, parent,
2442                                root_al, &cpumode, ip,
2443                                true, flags, NULL,
2444                                *branch_from);
2445         if (err)
2446                 return err;
2447
2448         return 0;
2449 }
2450
2451 static int lbr_callchain_add_stitched_lbr_ip(struct thread *thread,
2452                                              struct callchain_cursor *cursor)
2453 {
2454         struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2455         struct callchain_cursor_node *cnode;
2456         struct stitch_list *stitch_node;
2457         int err;
2458
2459         list_for_each_entry(stitch_node, &lbr_stitch->lists, node) {
2460                 cnode = &stitch_node->cursor;
2461
2462                 err = callchain_cursor_append(cursor, cnode->ip,
2463                                               &cnode->ms,
2464                                               cnode->branch,
2465                                               &cnode->branch_flags,
2466                                               cnode->nr_loop_iter,
2467                                               cnode->iter_cycles,
2468                                               cnode->branch_from,
2469                                               cnode->srcline);
2470                 if (err)
2471                         return err;
2472         }
2473         return 0;
2474 }
2475
2476 static struct stitch_list *get_stitch_node(struct thread *thread)
2477 {
2478         struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2479         struct stitch_list *stitch_node;
2480
2481         if (!list_empty(&lbr_stitch->free_lists)) {
2482                 stitch_node = list_first_entry(&lbr_stitch->free_lists,
2483                                                struct stitch_list, node);
2484                 list_del(&stitch_node->node);
2485
2486                 return stitch_node;
2487         }
2488
2489         return malloc(sizeof(struct stitch_list));
2490 }
2491
2492 static bool has_stitched_lbr(struct thread *thread,
2493                              struct perf_sample *cur,
2494                              struct perf_sample *prev,
2495                              unsigned int max_lbr,
2496                              bool callee)
2497 {
2498         struct branch_stack *cur_stack = cur->branch_stack;
2499         struct branch_entry *cur_entries = perf_sample__branch_entries(cur);
2500         struct branch_stack *prev_stack = prev->branch_stack;
2501         struct branch_entry *prev_entries = perf_sample__branch_entries(prev);
2502         struct lbr_stitch *lbr_stitch = thread->lbr_stitch;
2503         int i, j, nr_identical_branches = 0;
2504         struct stitch_list *stitch_node;
2505         u64 cur_base, distance;
2506
2507         if (!cur_stack || !prev_stack)
2508                 return false;
2509
2510         /* Find the physical index of the base-of-stack for current sample. */
2511         cur_base = max_lbr - cur_stack->nr + cur_stack->hw_idx + 1;
2512
2513         distance = (prev_stack->hw_idx > cur_base) ? (prev_stack->hw_idx - cur_base) :
2514                                                      (max_lbr + prev_stack->hw_idx - cur_base);
2515         /* Previous sample has shorter stack. Nothing can be stitched. */
2516         if (distance + 1 > prev_stack->nr)
2517                 return false;
2518
2519         /*
2520          * Check if there are identical LBRs between two samples.
2521          * Identical LBRs must have same from, to and flags values. Also,
2522          * they have to be saved in the same LBR registers (same physical
2523          * index).
2524          *
2525          * Starts from the base-of-stack of current sample.
2526          */
2527         for (i = distance, j = cur_stack->nr - 1; (i >= 0) && (j >= 0); i--, j--) {
2528                 if ((prev_entries[i].from != cur_entries[j].from) ||
2529                     (prev_entries[i].to != cur_entries[j].to) ||
2530                     (prev_entries[i].flags.value != cur_entries[j].flags.value))
2531                         break;
2532                 nr_identical_branches++;
2533         }
2534
2535         if (!nr_identical_branches)
2536                 return false;
2537
2538         /*
2539          * Save the LBRs between the base-of-stack of previous sample
2540          * and the base-of-stack of current sample into lbr_stitch->lists.
2541          * These LBRs will be stitched later.
2542          */
2543         for (i = prev_stack->nr - 1; i > (int)distance; i--) {
2544
2545                 if (!lbr_stitch->prev_lbr_cursor[i].valid)
2546                         continue;
2547
2548                 stitch_node = get_stitch_node(thread);
2549                 if (!stitch_node)
2550                         return false;
2551
2552                 memcpy(&stitch_node->cursor, &lbr_stitch->prev_lbr_cursor[i],
2553                        sizeof(struct callchain_cursor_node));
2554
2555                 if (callee)
2556                         list_add(&stitch_node->node, &lbr_stitch->lists);
2557                 else
2558                         list_add_tail(&stitch_node->node, &lbr_stitch->lists);
2559         }
2560
2561         return true;
2562 }
2563
2564 static bool alloc_lbr_stitch(struct thread *thread, unsigned int max_lbr)
2565 {
2566         if (thread->lbr_stitch)
2567                 return true;
2568
2569         thread->lbr_stitch = zalloc(sizeof(*thread->lbr_stitch));
2570         if (!thread->lbr_stitch)
2571                 goto err;
2572
2573         thread->lbr_stitch->prev_lbr_cursor = calloc(max_lbr + 1, sizeof(struct callchain_cursor_node));
2574         if (!thread->lbr_stitch->prev_lbr_cursor)
2575                 goto free_lbr_stitch;
2576
2577         INIT_LIST_HEAD(&thread->lbr_stitch->lists);
2578         INIT_LIST_HEAD(&thread->lbr_stitch->free_lists);
2579
2580         return true;
2581
2582 free_lbr_stitch:
2583         zfree(&thread->lbr_stitch);
2584 err:
2585         pr_warning("Failed to allocate space for stitched LBRs. Disable LBR stitch\n");
2586         thread->lbr_stitch_enable = false;
2587         return false;
2588 }
2589
2590 /*
2591  * Resolve LBR callstack chain sample
2592  * Return:
2593  * 1 on success get LBR callchain information
2594  * 0 no available LBR callchain information, should try fp
2595  * negative error code on other errors.
2596  */
2597 static int resolve_lbr_callchain_sample(struct thread *thread,
2598                                         struct callchain_cursor *cursor,
2599                                         struct perf_sample *sample,
2600                                         struct symbol **parent,
2601                                         struct addr_location *root_al,
2602                                         int max_stack,
2603                                         unsigned int max_lbr)
2604 {
2605         bool callee = (callchain_param.order == ORDER_CALLEE);
2606         struct ip_callchain *chain = sample->callchain;
2607         int chain_nr = min(max_stack, (int)chain->nr), i;
2608         struct lbr_stitch *lbr_stitch;
2609         bool stitched_lbr = false;
2610         u64 branch_from = 0;
2611         int err;
2612
2613         for (i = 0; i < chain_nr; i++) {
2614                 if (chain->ips[i] == PERF_CONTEXT_USER)
2615                         break;
2616         }
2617
2618         /* LBR only affects the user callchain */
2619         if (i == chain_nr)
2620                 return 0;
2621
2622         if (thread->lbr_stitch_enable && !sample->no_hw_idx &&
2623             (max_lbr > 0) && alloc_lbr_stitch(thread, max_lbr)) {
2624                 lbr_stitch = thread->lbr_stitch;
2625
2626                 stitched_lbr = has_stitched_lbr(thread, sample,
2627                                                 &lbr_stitch->prev_sample,
2628                                                 max_lbr, callee);
2629
2630                 if (!stitched_lbr && !list_empty(&lbr_stitch->lists)) {
2631                         list_replace_init(&lbr_stitch->lists,
2632                                           &lbr_stitch->free_lists);
2633                 }
2634                 memcpy(&lbr_stitch->prev_sample, sample, sizeof(*sample));
2635         }
2636
2637         if (callee) {
2638                 /* Add kernel ip */
2639                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2640                                                   parent, root_al, branch_from,
2641                                                   true, i);
2642                 if (err)
2643                         goto error;
2644
2645                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2646                                                root_al, &branch_from, true);
2647                 if (err)
2648                         goto error;
2649
2650                 if (stitched_lbr) {
2651                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2652                         if (err)
2653                                 goto error;
2654                 }
2655
2656         } else {
2657                 if (stitched_lbr) {
2658                         err = lbr_callchain_add_stitched_lbr_ip(thread, cursor);
2659                         if (err)
2660                                 goto error;
2661                 }
2662                 err = lbr_callchain_add_lbr_ip(thread, cursor, sample, parent,
2663                                                root_al, &branch_from, false);
2664                 if (err)
2665                         goto error;
2666
2667                 /* Add kernel ip */
2668                 err = lbr_callchain_add_kernel_ip(thread, cursor, sample,
2669                                                   parent, root_al, branch_from,
2670                                                   false, i);
2671                 if (err)
2672                         goto error;
2673         }
2674         return 1;
2675
2676 error:
2677         return (err < 0) ? err : 0;
2678 }
2679
2680 static int find_prev_cpumode(struct ip_callchain *chain, struct thread *thread,
2681                              struct callchain_cursor *cursor,
2682                              struct symbol **parent,
2683                              struct addr_location *root_al,
2684                              u8 *cpumode, int ent)
2685 {
2686         int err = 0;
2687
2688         while (--ent >= 0) {
2689                 u64 ip = chain->ips[ent];
2690
2691                 if (ip >= PERF_CONTEXT_MAX) {
2692                         err = add_callchain_ip(thread, cursor, parent,
2693                                                root_al, cpumode, ip,
2694                                                false, NULL, NULL, 0);
2695                         break;
2696                 }
2697         }
2698         return err;
2699 }
2700
2701 static int thread__resolve_callchain_sample(struct thread *thread,
2702                                             struct callchain_cursor *cursor,
2703                                             struct evsel *evsel,
2704                                             struct perf_sample *sample,
2705                                             struct symbol **parent,
2706                                             struct addr_location *root_al,
2707                                             int max_stack)
2708 {
2709         struct branch_stack *branch = sample->branch_stack;
2710         struct branch_entry *entries = perf_sample__branch_entries(sample);
2711         struct ip_callchain *chain = sample->callchain;
2712         int chain_nr = 0;
2713         u8 cpumode = PERF_RECORD_MISC_USER;
2714         int i, j, err, nr_entries;
2715         int skip_idx = -1;
2716         int first_call = 0;
2717
2718         if (chain)
2719                 chain_nr = chain->nr;
2720
2721         if (evsel__has_branch_callstack(evsel)) {
2722                 struct perf_env *env = evsel__env(evsel);
2723
2724                 err = resolve_lbr_callchain_sample(thread, cursor, sample, parent,
2725                                                    root_al, max_stack,
2726                                                    !env ? 0 : env->max_branches);
2727                 if (err)
2728                         return (err < 0) ? err : 0;
2729         }
2730
2731         /*
2732          * Based on DWARF debug information, some architectures skip
2733          * a callchain entry saved by the kernel.
2734          */
2735         skip_idx = arch_skip_callchain_idx(thread, chain);
2736
2737         /*
2738          * Add branches to call stack for easier browsing. This gives
2739          * more context for a sample than just the callers.
2740          *
2741          * This uses individual histograms of paths compared to the
2742          * aggregated histograms the normal LBR mode uses.
2743          *
2744          * Limitations for now:
2745          * - No extra filters
2746          * - No annotations (should annotate somehow)
2747          */
2748
2749         if (branch && callchain_param.branch_callstack) {
2750                 int nr = min(max_stack, (int)branch->nr);
2751                 struct branch_entry be[nr];
2752                 struct iterations iter[nr];
2753
2754                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
2755                         pr_warning("corrupted branch chain. skipping...\n");
2756                         goto check_calls;
2757                 }
2758
2759                 for (i = 0; i < nr; i++) {
2760                         if (callchain_param.order == ORDER_CALLEE) {
2761                                 be[i] = entries[i];
2762
2763                                 if (chain == NULL)
2764                                         continue;
2765
2766                                 /*
2767                                  * Check for overlap into the callchain.
2768                                  * The return address is one off compared to
2769                                  * the branch entry. To adjust for this
2770                                  * assume the calling instruction is not longer
2771                                  * than 8 bytes.
2772                                  */
2773                                 if (i == skip_idx ||
2774                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
2775                                         first_call++;
2776                                 else if (be[i].from < chain->ips[first_call] &&
2777                                     be[i].from >= chain->ips[first_call] - 8)
2778                                         first_call++;
2779                         } else
2780                                 be[i] = entries[branch->nr - i - 1];
2781                 }
2782
2783                 memset(iter, 0, sizeof(struct iterations) * nr);
2784                 nr = remove_loops(be, nr, iter);
2785
2786                 for (i = 0; i < nr; i++) {
2787                         err = add_callchain_ip(thread, cursor, parent,
2788                                                root_al,
2789                                                NULL, be[i].to,
2790                                                true, &be[i].flags,
2791                                                NULL, be[i].from);
2792
2793                         if (!err)
2794                                 err = add_callchain_ip(thread, cursor, parent, root_al,
2795                                                        NULL, be[i].from,
2796                                                        true, &be[i].flags,
2797                                                        &iter[i], 0);
2798                         if (err == -EINVAL)
2799                                 break;
2800                         if (err)
2801                                 return err;
2802                 }
2803
2804                 if (chain_nr == 0)
2805                         return 0;
2806
2807                 chain_nr -= nr;
2808         }
2809
2810 check_calls:
2811         if (chain && callchain_param.order != ORDER_CALLEE) {
2812                 err = find_prev_cpumode(chain, thread, cursor, parent, root_al,
2813                                         &cpumode, chain->nr - first_call);
2814                 if (err)
2815                         return (err < 0) ? err : 0;
2816         }
2817         for (i = first_call, nr_entries = 0;
2818              i < chain_nr && nr_entries < max_stack; i++) {
2819                 u64 ip;
2820
2821                 if (callchain_param.order == ORDER_CALLEE)
2822                         j = i;
2823                 else
2824                         j = chain->nr - i - 1;
2825
2826 #ifdef HAVE_SKIP_CALLCHAIN_IDX
2827                 if (j == skip_idx)
2828                         continue;
2829 #endif
2830                 ip = chain->ips[j];
2831                 if (ip < PERF_CONTEXT_MAX)
2832                        ++nr_entries;
2833                 else if (callchain_param.order != ORDER_CALLEE) {
2834                         err = find_prev_cpumode(chain, thread, cursor, parent,
2835                                                 root_al, &cpumode, j);
2836                         if (err)
2837                                 return (err < 0) ? err : 0;
2838                         continue;
2839                 }
2840
2841                 err = add_callchain_ip(thread, cursor, parent,
2842                                        root_al, &cpumode, ip,
2843                                        false, NULL, NULL, 0);
2844
2845                 if (err)
2846                         return (err < 0) ? err : 0;
2847         }
2848
2849         return 0;
2850 }
2851
2852 static int append_inlines(struct callchain_cursor *cursor, struct map_symbol *ms, u64 ip)
2853 {
2854         struct symbol *sym = ms->sym;
2855         struct map *map = ms->map;
2856         struct inline_node *inline_node;
2857         struct inline_list *ilist;
2858         u64 addr;
2859         int ret = 1;
2860
2861         if (!symbol_conf.inline_name || !map || !sym)
2862                 return ret;
2863
2864         addr = map__map_ip(map, ip);
2865         addr = map__rip_2objdump(map, addr);
2866
2867         inline_node = inlines__tree_find(&map->dso->inlined_nodes, addr);
2868         if (!inline_node) {
2869                 inline_node = dso__parse_addr_inlines(map->dso, addr, sym);
2870                 if (!inline_node)
2871                         return ret;
2872                 inlines__tree_insert(&map->dso->inlined_nodes, inline_node);
2873         }
2874
2875         list_for_each_entry(ilist, &inline_node->val, list) {
2876                 struct map_symbol ilist_ms = {
2877                         .maps = ms->maps,
2878                         .map = map,
2879                         .sym = ilist->symbol,
2880                 };
2881                 ret = callchain_cursor_append(cursor, ip, &ilist_ms, false,
2882                                               NULL, 0, 0, 0, ilist->srcline);
2883
2884                 if (ret != 0)
2885                         return ret;
2886         }
2887
2888         return ret;
2889 }
2890
2891 static int unwind_entry(struct unwind_entry *entry, void *arg)
2892 {
2893         struct callchain_cursor *cursor = arg;
2894         const char *srcline = NULL;
2895         u64 addr = entry->ip;
2896
2897         if (symbol_conf.hide_unresolved && entry->ms.sym == NULL)
2898                 return 0;
2899
2900         if (append_inlines(cursor, &entry->ms, entry->ip) == 0)
2901                 return 0;
2902
2903         /*
2904          * Convert entry->ip from a virtual address to an offset in
2905          * its corresponding binary.
2906          */
2907         if (entry->ms.map)
2908                 addr = map__map_ip(entry->ms.map, entry->ip);
2909
2910         srcline = callchain_srcline(&entry->ms, addr);
2911         return callchain_cursor_append(cursor, entry->ip, &entry->ms,
2912                                        false, NULL, 0, 0, 0, srcline);
2913 }
2914
2915 static int thread__resolve_callchain_unwind(struct thread *thread,
2916                                             struct callchain_cursor *cursor,
2917                                             struct evsel *evsel,
2918                                             struct perf_sample *sample,
2919                                             int max_stack)
2920 {
2921         /* Can we do dwarf post unwind? */
2922         if (!((evsel->core.attr.sample_type & PERF_SAMPLE_REGS_USER) &&
2923               (evsel->core.attr.sample_type & PERF_SAMPLE_STACK_USER)))
2924                 return 0;
2925
2926         /* Bail out if nothing was captured. */
2927         if ((!sample->user_regs.regs) ||
2928             (!sample->user_stack.size))
2929                 return 0;
2930
2931         return unwind__get_entries(unwind_entry, cursor,
2932                                    thread, sample, max_stack);
2933 }
2934
2935 int thread__resolve_callchain(struct thread *thread,
2936                               struct callchain_cursor *cursor,
2937                               struct evsel *evsel,
2938                               struct perf_sample *sample,
2939                               struct symbol **parent,
2940                               struct addr_location *root_al,
2941                               int max_stack)
2942 {
2943         int ret = 0;
2944
2945         callchain_cursor_reset(cursor);
2946
2947         if (callchain_param.order == ORDER_CALLEE) {
2948                 ret = thread__resolve_callchain_sample(thread, cursor,
2949                                                        evsel, sample,
2950                                                        parent, root_al,
2951                                                        max_stack);
2952                 if (ret)
2953                         return ret;
2954                 ret = thread__resolve_callchain_unwind(thread, cursor,
2955                                                        evsel, sample,
2956                                                        max_stack);
2957         } else {
2958                 ret = thread__resolve_callchain_unwind(thread, cursor,
2959                                                        evsel, sample,
2960                                                        max_stack);
2961                 if (ret)
2962                         return ret;
2963                 ret = thread__resolve_callchain_sample(thread, cursor,
2964                                                        evsel, sample,
2965                                                        parent, root_al,
2966                                                        max_stack);
2967         }
2968
2969         return ret;
2970 }
2971
2972 int machine__for_each_thread(struct machine *machine,
2973                              int (*fn)(struct thread *thread, void *p),
2974                              void *priv)
2975 {
2976         struct threads *threads;
2977         struct rb_node *nd;
2978         struct thread *thread;
2979         int rc = 0;
2980         int i;
2981
2982         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
2983                 threads = &machine->threads[i];
2984                 for (nd = rb_first_cached(&threads->entries); nd;
2985                      nd = rb_next(nd)) {
2986                         thread = rb_entry(nd, struct thread, rb_node);
2987                         rc = fn(thread, priv);
2988                         if (rc != 0)
2989                                 return rc;
2990                 }
2991
2992                 list_for_each_entry(thread, &threads->dead, node) {
2993                         rc = fn(thread, priv);
2994                         if (rc != 0)
2995                                 return rc;
2996                 }
2997         }
2998         return rc;
2999 }
3000
3001 int machines__for_each_thread(struct machines *machines,
3002                               int (*fn)(struct thread *thread, void *p),
3003                               void *priv)
3004 {
3005         struct rb_node *nd;
3006         int rc = 0;
3007
3008         rc = machine__for_each_thread(&machines->host, fn, priv);
3009         if (rc != 0)
3010                 return rc;
3011
3012         for (nd = rb_first_cached(&machines->guests); nd; nd = rb_next(nd)) {
3013                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
3014
3015                 rc = machine__for_each_thread(machine, fn, priv);
3016                 if (rc != 0)
3017                         return rc;
3018         }
3019         return rc;
3020 }
3021
3022 pid_t machine__get_current_tid(struct machine *machine, int cpu)
3023 {
3024         int nr_cpus = min(machine->env->nr_cpus_avail, MAX_NR_CPUS);
3025
3026         if (cpu < 0 || cpu >= nr_cpus || !machine->current_tid)
3027                 return -1;
3028
3029         return machine->current_tid[cpu];
3030 }
3031
3032 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
3033                              pid_t tid)
3034 {
3035         struct thread *thread;
3036         int nr_cpus = min(machine->env->nr_cpus_avail, MAX_NR_CPUS);
3037
3038         if (cpu < 0)
3039                 return -EINVAL;
3040
3041         if (!machine->current_tid) {
3042                 int i;
3043
3044                 machine->current_tid = calloc(nr_cpus, sizeof(pid_t));
3045                 if (!machine->current_tid)
3046                         return -ENOMEM;
3047                 for (i = 0; i < nr_cpus; i++)
3048                         machine->current_tid[i] = -1;
3049         }
3050
3051         if (cpu >= nr_cpus) {
3052                 pr_err("Requested CPU %d too large. ", cpu);
3053                 pr_err("Consider raising MAX_NR_CPUS\n");
3054                 return -EINVAL;
3055         }
3056
3057         machine->current_tid[cpu] = tid;
3058
3059         thread = machine__findnew_thread(machine, pid, tid);
3060         if (!thread)
3061                 return -ENOMEM;
3062
3063         thread->cpu = cpu;
3064         thread__put(thread);
3065
3066         return 0;
3067 }
3068
3069 /*
3070  * Compares the raw arch string. N.B. see instead perf_env__arch() if a
3071  * normalized arch is needed.
3072  */
3073 bool machine__is(struct machine *machine, const char *arch)
3074 {
3075         return machine && !strcmp(perf_env__raw_arch(machine->env), arch);
3076 }
3077
3078 int machine__nr_cpus_avail(struct machine *machine)
3079 {
3080         return machine ? perf_env__nr_cpus_avail(machine->env) : 0;
3081 }
3082
3083 int machine__get_kernel_start(struct machine *machine)
3084 {
3085         struct map *map = machine__kernel_map(machine);
3086         int err = 0;
3087
3088         /*
3089          * The only addresses above 2^63 are kernel addresses of a 64-bit
3090          * kernel.  Note that addresses are unsigned so that on a 32-bit system
3091          * all addresses including kernel addresses are less than 2^32.  In
3092          * that case (32-bit system), if the kernel mapping is unknown, all
3093          * addresses will be assumed to be in user space - see
3094          * machine__kernel_ip().
3095          */
3096         machine->kernel_start = 1ULL << 63;
3097         if (map) {
3098                 err = map__load(map);
3099                 /*
3100                  * On x86_64, PTI entry trampolines are less than the
3101                  * start of kernel text, but still above 2^63. So leave
3102                  * kernel_start = 1ULL << 63 for x86_64.
3103                  */
3104                 if (!err && !machine__is(machine, "x86_64"))
3105                         machine->kernel_start = map->start;
3106         }
3107         return err;
3108 }
3109
3110 u8 machine__addr_cpumode(struct machine *machine, u8 cpumode, u64 addr)
3111 {
3112         u8 addr_cpumode = cpumode;
3113         bool kernel_ip;
3114
3115         if (!machine->single_address_space)
3116                 goto out;
3117
3118         kernel_ip = machine__kernel_ip(machine, addr);
3119         switch (cpumode) {
3120         case PERF_RECORD_MISC_KERNEL:
3121         case PERF_RECORD_MISC_USER:
3122                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_KERNEL :
3123                                            PERF_RECORD_MISC_USER;
3124                 break;
3125         case PERF_RECORD_MISC_GUEST_KERNEL:
3126         case PERF_RECORD_MISC_GUEST_USER:
3127                 addr_cpumode = kernel_ip ? PERF_RECORD_MISC_GUEST_KERNEL :
3128                                            PERF_RECORD_MISC_GUEST_USER;
3129                 break;
3130         default:
3131                 break;
3132         }
3133 out:
3134         return addr_cpumode;
3135 }
3136
3137 struct dso *machine__findnew_dso_id(struct machine *machine, const char *filename, struct dso_id *id)
3138 {
3139         return dsos__findnew_id(&machine->dsos, filename, id);
3140 }
3141
3142 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
3143 {
3144         return machine__findnew_dso_id(machine, filename, NULL);
3145 }
3146
3147 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
3148 {
3149         struct machine *machine = vmachine;
3150         struct map *map;
3151         struct symbol *sym = machine__find_kernel_symbol(machine, *addrp, &map);
3152
3153         if (sym == NULL)
3154                 return NULL;
3155
3156         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
3157         *addrp = map->unmap_ip(map, sym->start);
3158         return sym->name;
3159 }
3160
3161 int machine__for_each_dso(struct machine *machine, machine__dso_t fn, void *priv)
3162 {
3163         struct dso *pos;
3164         int err = 0;
3165
3166         list_for_each_entry(pos, &machine->dsos.head, node) {
3167                 if (fn(pos, machine, priv))
3168                         err = -1;
3169         }
3170         return err;
3171 }