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