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