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