tools/perf/util/machine.c

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