kernel/tools/perf/util/machine.c

   1 #include "callchain.h"
   2 #include "debug.h"
   3 #include "event.h"
   4 #include "evsel.h"
   5 #include "hist.h"
   6 #include "machine.h"
   7 #include "map.h"
   8 #include "sort.h"
   9 #include "strlist.h"
  10 #include "thread.h"
  11 #include "vdso.h"
  12 #include <stdbool.h>
  13 #include <symbol/kallsyms.h>
  14 #include "unwind.h"
  15 #include "linux/hash.h"
  16
  17 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock);
  18
  19 static void dsos__init(struct dsos *dsos)
  20 {
  21         INIT_LIST_HEAD(&dsos->head);
  22         dsos->root = RB_ROOT;
  23         pthread_rwlock_init(&dsos->lock, NULL);
  24 }
  25
  26 int machine__init(struct machine *machine, const char *root_dir, pid_t pid)
  27 {
  28         map_groups__init(&machine->kmaps, machine);
  29         RB_CLEAR_NODE(&machine->rb_node);
  30         dsos__init(&machine->dsos);
  31
  32         machine->threads = RB_ROOT;
  33         pthread_rwlock_init(&machine->threads_lock, NULL);
  34         INIT_LIST_HEAD(&machine->dead_threads);
  35         machine->last_match = NULL;
  36
  37         machine->vdso_info = NULL;
  38         machine->env = NULL;
  39
  40         machine->pid = pid;
  41
  42         machine->symbol_filter = NULL;
  43         machine->id_hdr_size = 0;
  44         machine->comm_exec = false;
  45         machine->kernel_start = 0;
  46
  47         machine->root_dir = strdup(root_dir);
  48         if (machine->root_dir == NULL)
  49                 return -ENOMEM;
  50
  51         if (pid != HOST_KERNEL_ID) {
  52                 struct thread *thread = machine__findnew_thread(machine, -1,
  53                                                                 pid);
  54                 char comm[64];
  55
  56                 if (thread == NULL)
  57                         return -ENOMEM;
  58
  59                 snprintf(comm, sizeof(comm), "[guest/%d]", pid);
  60                 thread__set_comm(thread, comm, 0);
  61                 thread__put(thread);
  62         }
  63
  64         machine->current_tid = NULL;
  65
  66         return 0;
  67 }
  68
  69 struct machine *machine__new_host(void)
  70 {
  71         struct machine *machine = malloc(sizeof(*machine));
  72
  73         if (machine != NULL) {
  74                 machine__init(machine, "", HOST_KERNEL_ID);
  75
  76                 if (machine__create_kernel_maps(machine) < 0)
  77                         goto out_delete;
  78         }
  79
  80         return machine;
  81 out_delete:
  82         free(machine);
  83         return NULL;
  84 }
  85
  86 static void dsos__purge(struct dsos *dsos)
  87 {
  88         struct dso *pos, *n;
  89
  90         pthread_rwlock_wrlock(&dsos->lock);
  91
  92         list_for_each_entry_safe(pos, n, &dsos->head, node) {
  93                 RB_CLEAR_NODE(&pos->rb_node);
  94                 pos->root = NULL;
  95                 list_del_init(&pos->node);
  96                 dso__put(pos);
  97         }
  98
  99         pthread_rwlock_unlock(&dsos->lock);
 100 }
 101
 102 static void dsos__exit(struct dsos *dsos)
 103 {
 104         dsos__purge(dsos);
 105         pthread_rwlock_destroy(&dsos->lock);
 106 }
 107
 108 void machine__delete_threads(struct machine *machine)
 109 {
 110         struct rb_node *nd;
 111
 112         pthread_rwlock_wrlock(&machine->threads_lock);
 113         nd = rb_first(&machine->threads);
 114         while (nd) {
 115                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 116
 117                 nd = rb_next(nd);
 118                 __machine__remove_thread(machine, t, false);
 119         }
 120         pthread_rwlock_unlock(&machine->threads_lock);
 121 }
 122
 123 void machine__exit(struct machine *machine)
 124 {
 125         map_groups__exit(&machine->kmaps);
 126         dsos__exit(&machine->dsos);
 127         machine__exit_vdso(machine);
 128         zfree(&machine->root_dir);
 129         zfree(&machine->current_tid);
 130         pthread_rwlock_destroy(&machine->threads_lock);
 131 }
 132
 133 void machine__delete(struct machine *machine)
 134 {
 135         machine__exit(machine);
 136         free(machine);
 137 }
 138
 139 void machines__init(struct machines *machines)
 140 {
 141         machine__init(&machines->host, "", HOST_KERNEL_ID);
 142         machines->guests = RB_ROOT;
 143         machines->symbol_filter = NULL;
 144 }
 145
 146 void machines__exit(struct machines *machines)
 147 {
 148         machine__exit(&machines->host);
 149         /* XXX exit guest */
 150 }
 151
 152 struct machine *machines__add(struct machines *machines, pid_t pid,
 153                               const char *root_dir)
 154 {
 155         struct rb_node **p = &machines->guests.rb_node;
 156         struct rb_node *parent = NULL;
 157         struct machine *pos, *machine = malloc(sizeof(*machine));
 158
 159         if (machine == NULL)
 160                 return NULL;
 161
 162         if (machine__init(machine, root_dir, pid) != 0) {
 163                 free(machine);
 164                 return NULL;
 165         }
 166
 167         machine->symbol_filter = machines->symbol_filter;
 168
 169         while (*p != NULL) {
 170                 parent = *p;
 171                 pos = rb_entry(parent, struct machine, rb_node);
 172                 if (pid < pos->pid)
 173                         p = &(*p)->rb_left;
 174                 else
 175                         p = &(*p)->rb_right;
 176         }
 177
 178         rb_link_node(&machine->rb_node, parent, p);
 179         rb_insert_color(&machine->rb_node, &machines->guests);
 180
 181         return machine;
 182 }
 183
 184 void machines__set_symbol_filter(struct machines *machines,
 185                                  symbol_filter_t symbol_filter)
 186 {
 187         struct rb_node *nd;
 188
 189         machines->symbol_filter = symbol_filter;
 190         machines->host.symbol_filter = symbol_filter;
 191
 192         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 193                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 194
 195                 machine->symbol_filter = symbol_filter;
 196         }
 197 }
 198
 199 void machines__set_comm_exec(struct machines *machines, bool comm_exec)
 200 {
 201         struct rb_node *nd;
 202
 203         machines->host.comm_exec = comm_exec;
 204
 205         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 206                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
 207
 208                 machine->comm_exec = comm_exec;
 209         }
 210 }
 211
 212 struct machine *machines__find(struct machines *machines, pid_t pid)
 213 {
 214         struct rb_node **p = &machines->guests.rb_node;
 215         struct rb_node *parent = NULL;
 216         struct machine *machine;
 217         struct machine *default_machine = NULL;
 218
 219         if (pid == HOST_KERNEL_ID)
 220                 return &machines->host;
 221
 222         while (*p != NULL) {
 223                 parent = *p;
 224                 machine = rb_entry(parent, struct machine, rb_node);
 225                 if (pid < machine->pid)
 226                         p = &(*p)->rb_left;
 227                 else if (pid > machine->pid)
 228                         p = &(*p)->rb_right;
 229                 else
 230                         return machine;
 231                 if (!machine->pid)
 232                         default_machine = machine;
 233         }
 234
 235         return default_machine;
 236 }
 237
 238 struct machine *machines__findnew(struct machines *machines, pid_t pid)
 239 {
 240         char path[PATH_MAX];
 241         const char *root_dir = "";
 242         struct machine *machine = machines__find(machines, pid);
 243
 244         if (machine && (machine->pid == pid))
 245                 goto out;
 246
 247         if ((pid != HOST_KERNEL_ID) &&
 248             (pid != DEFAULT_GUEST_KERNEL_ID) &&
 249             (symbol_conf.guestmount)) {
 250                 sprintf(path, "%s/%d", symbol_conf.guestmount, pid);
 251                 if (access(path, R_OK)) {
 252                         static struct strlist *seen;
 253
 254                         if (!seen)
 255                                 seen = strlist__new(NULL, NULL);
 256
 257                         if (!strlist__has_entry(seen, path)) {
 258                                 pr_err("Can't access file %s\n", path);
 259                                 strlist__add(seen, path);
 260                         }
 261                         machine = NULL;
 262                         goto out;
 263                 }
 264                 root_dir = path;
 265         }
 266
 267         machine = machines__add(machines, pid, root_dir);
 268 out:
 269         return machine;
 270 }
 271
 272 void machines__process_guests(struct machines *machines,
 273                               machine__process_t process, void *data)
 274 {
 275         struct rb_node *nd;
 276
 277         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 278                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 279                 process(pos, data);
 280         }
 281 }
 282
 283 char *machine__mmap_name(struct machine *machine, char *bf, size_t size)
 284 {
 285         if (machine__is_host(machine))
 286                 snprintf(bf, size, "[%s]", "kernel.kallsyms");
 287         else if (machine__is_default_guest(machine))
 288                 snprintf(bf, size, "[%s]", "guest.kernel.kallsyms");
 289         else {
 290                 snprintf(bf, size, "[%s.%d]", "guest.kernel.kallsyms",
 291                          machine->pid);
 292         }
 293
 294         return bf;
 295 }
 296
 297 void machines__set_id_hdr_size(struct machines *machines, u16 id_hdr_size)
 298 {
 299         struct rb_node *node;
 300         struct machine *machine;
 301
 302         machines->host.id_hdr_size = id_hdr_size;
 303
 304         for (node = rb_first(&machines->guests); node; node = rb_next(node)) {
 305                 machine = rb_entry(node, struct machine, rb_node);
 306                 machine->id_hdr_size = id_hdr_size;
 307         }
 308
 309         return;
 310 }
 311
 312 static void machine__update_thread_pid(struct machine *machine,
 313                                        struct thread *th, pid_t pid)
 314 {
 315         struct thread *leader;
 316
 317         if (pid == th->pid_ || pid == -1 || th->pid_ != -1)
 318                 return;
 319
 320         th->pid_ = pid;
 321
 322         if (th->pid_ == th->tid)
 323                 return;
 324
 325         leader = __machine__findnew_thread(machine, th->pid_, th->pid_);
 326         if (!leader)
 327                 goto out_err;
 328
 329         if (!leader->mg)
 330                 leader->mg = map_groups__new(machine);
 331
 332         if (!leader->mg)
 333                 goto out_err;
 334
 335         if (th->mg == leader->mg)
 336                 return;
 337
 338         if (th->mg) {
 339                 /*
 340                  * Maps are created from MMAP events which provide the pid and
 341                  * tid.  Consequently there never should be any maps on a thread
 342                  * with an unknown pid.  Just print an error if there are.
 343                  */
 344                 if (!map_groups__empty(th->mg))
 345                         pr_err("Discarding thread maps for %d:%d\n",
 346                                th->pid_, th->tid);
 347                 map_groups__put(th->mg);
 348         }
 349
 350         th->mg = map_groups__get(leader->mg);
 351
 352         return;
 353
 354 out_err:
 355         pr_err("Failed to join map groups for %d:%d\n", th->pid_, th->tid);
 356 }
 357
 358 static struct thread *____machine__findnew_thread(struct machine *machine,
 359                                                   pid_t pid, pid_t tid,
 360                                                   bool create)
 361 {
 362         struct rb_node **p = &machine->threads.rb_node;
 363         struct rb_node *parent = NULL;
 364         struct thread *th;
 365
 366         /*
 367          * Front-end cache - TID lookups come in blocks,
 368          * so most of the time we dont have to look up
 369          * the full rbtree:
 370          */
 371         th = machine->last_match;
 372         if (th != NULL) {
 373                 if (th->tid == tid) {
 374                         machine__update_thread_pid(machine, th, pid);
 375                         return th;
 376                 }
 377
 378                 machine->last_match = NULL;
 379         }
 380
 381         while (*p != NULL) {
 382                 parent = *p;
 383                 th = rb_entry(parent, struct thread, rb_node);
 384
 385                 if (th->tid == tid) {
 386                         machine->last_match = th;
 387                         machine__update_thread_pid(machine, th, pid);
 388                         return th;
 389                 }
 390
 391                 if (tid < th->tid)
 392                         p = &(*p)->rb_left;
 393                 else
 394                         p = &(*p)->rb_right;
 395         }
 396
 397         if (!create)
 398                 return NULL;
 399
 400         th = thread__new(pid, tid);
 401         if (th != NULL) {
 402                 rb_link_node(&th->rb_node, parent, p);
 403                 rb_insert_color(&th->rb_node, &machine->threads);
 404
 405                 /*
 406                  * We have to initialize map_groups separately
 407                  * after rb tree is updated.
 408                  *
 409                  * The reason is that we call machine__findnew_thread
 410                  * within thread__init_map_groups to find the thread
 411                  * leader and that would screwed the rb tree.
 412                  */
 413                 if (thread__init_map_groups(th, machine)) {
 414                         rb_erase_init(&th->rb_node, &machine->threads);
 415                         RB_CLEAR_NODE(&th->rb_node);
 416                         thread__delete(th);
 417                         return NULL;
 418                 }
 419                 /*
 420                  * It is now in the rbtree, get a ref
 421                  */
 422                 thread__get(th);
 423                 machine->last_match = th;
 424         }
 425
 426         return th;
 427 }
 428
 429 struct thread *__machine__findnew_thread(struct machine *machine, pid_t pid, pid_t tid)
 430 {
 431         return ____machine__findnew_thread(machine, pid, tid, true);
 432 }
 433
 434 struct thread *machine__findnew_thread(struct machine *machine, pid_t pid,
 435                                        pid_t tid)
 436 {
 437         struct thread *th;
 438
 439         pthread_rwlock_wrlock(&machine->threads_lock);
 440         th = thread__get(__machine__findnew_thread(machine, pid, tid));
 441         pthread_rwlock_unlock(&machine->threads_lock);
 442         return th;
 443 }
 444
 445 struct thread *machine__find_thread(struct machine *machine, pid_t pid,
 446                                     pid_t tid)
 447 {
 448         struct thread *th;
 449         pthread_rwlock_rdlock(&machine->threads_lock);
 450         th =  thread__get(____machine__findnew_thread(machine, pid, tid, false));
 451         pthread_rwlock_unlock(&machine->threads_lock);
 452         return th;
 453 }
 454
 455 struct comm *machine__thread_exec_comm(struct machine *machine,
 456                                        struct thread *thread)
 457 {
 458         if (machine->comm_exec)
 459                 return thread__exec_comm(thread);
 460         else
 461                 return thread__comm(thread);
 462 }
 463
 464 int machine__process_comm_event(struct machine *machine, union perf_event *event,
 465                                 struct perf_sample *sample)
 466 {
 467         struct thread *thread = machine__findnew_thread(machine,
 468                                                         event->comm.pid,
 469                                                         event->comm.tid);
 470         bool exec = event->header.misc & PERF_RECORD_MISC_COMM_EXEC;
 471         int err = 0;
 472
 473         if (exec)
 474                 machine->comm_exec = true;
 475
 476         if (dump_trace)
 477                 perf_event__fprintf_comm(event, stdout);
 478
 479         if (thread == NULL ||
 480             __thread__set_comm(thread, event->comm.comm, sample->time, exec)) {
 481                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
 482                 err = -1;
 483         }
 484
 485         thread__put(thread);
 486
 487         return err;
 488 }
 489
 490 int machine__process_lost_event(struct machine *machine __maybe_unused,
 491                                 union perf_event *event, struct perf_sample *sample __maybe_unused)
 492 {
 493         dump_printf(": id:%" PRIu64 ": lost:%" PRIu64 "\n",
 494                     event->lost.id, event->lost.lost);
 495         return 0;
 496 }
 497
 498 int machine__process_lost_samples_event(struct machine *machine __maybe_unused,
 499                                         union perf_event *event, struct perf_sample *sample)
 500 {
 501         dump_printf(": id:%" PRIu64 ": lost samples :%" PRIu64 "\n",
 502                     sample->id, event->lost_samples.lost);
 503         return 0;
 504 }
 505
 506 static struct dso *machine__findnew_module_dso(struct machine *machine,
 507                                                struct kmod_path *m,
 508                                                const char *filename)
 509 {
 510         struct dso *dso;
 511
 512         pthread_rwlock_wrlock(&machine->dsos.lock);
 513
 514         dso = __dsos__find(&machine->dsos, m->name, true);
 515         if (!dso) {
 516                 dso = __dsos__addnew(&machine->dsos, m->name);
 517                 if (dso == NULL)
 518                         goto out_unlock;
 519
 520                 if (machine__is_host(machine))
 521                         dso->symtab_type = DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE;
 522                 else
 523                         dso->symtab_type = DSO_BINARY_TYPE__GUEST_KMODULE;
 524
 525                 /* _KMODULE_COMP should be next to _KMODULE */
 526                 if (m->kmod && m->comp)
 527                         dso->symtab_type++;
 528
 529                 dso__set_short_name(dso, strdup(m->name), true);
 530                 dso__set_long_name(dso, strdup(filename), true);
 531         }
 532
 533         dso__get(dso);
 534 out_unlock:
 535         pthread_rwlock_unlock(&machine->dsos.lock);
 536         return dso;
 537 }
 538
 539 int machine__process_aux_event(struct machine *machine __maybe_unused,
 540                                union perf_event *event)
 541 {
 542         if (dump_trace)
 543                 perf_event__fprintf_aux(event, stdout);
 544         return 0;
 545 }
 546
 547 int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 548                                         union perf_event *event)
 549 {
 550         if (dump_trace)
 551                 perf_event__fprintf_itrace_start(event, stdout);
 552         return 0;
 553 }
 554
 555 int machine__process_switch_event(struct machine *machine __maybe_unused,
 556                                   union perf_event *event)
 557 {
 558         if (dump_trace)
 559                 perf_event__fprintf_switch(event, stdout);
 560         return 0;
 561 }
 562
 563 struct map *machine__findnew_module_map(struct machine *machine, u64 start,
 564                                         const char *filename)
 565 {
 566         struct map *map = NULL;
 567         struct dso *dso;
 568         struct kmod_path m;
 569
 570         if (kmod_path__parse_name(&m, filename))
 571                 return NULL;
 572
 573         map = map_groups__find_by_name(&machine->kmaps, MAP__FUNCTION,
 574                                        m.name);
 575         if (map)
 576                 goto out;
 577
 578         dso = machine__findnew_module_dso(machine, &m, filename);
 579         if (dso == NULL)
 580                 goto out;
 581
 582         map = map__new2(start, dso, MAP__FUNCTION);
 583         if (map == NULL)
 584                 goto out;
 585
 586         map_groups__insert(&machine->kmaps, map);
 587
 588 out:
 589         free(m.name);
 590         return map;
 591 }
 592
 593 size_t machines__fprintf_dsos(struct machines *machines, FILE *fp)
 594 {
 595         struct rb_node *nd;
 596         size_t ret = __dsos__fprintf(&machines->host.dsos.head, fp);
 597
 598         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 599                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 600                 ret += __dsos__fprintf(&pos->dsos.head, fp);
 601         }
 602
 603         return ret;
 604 }
 605
 606 size_t machine__fprintf_dsos_buildid(struct machine *m, FILE *fp,
 607                                      bool (skip)(struct dso *dso, int parm), int parm)
 608 {
 609         return __dsos__fprintf_buildid(&m->dsos.head, fp, skip, parm);
 610 }
 611
 612 size_t machines__fprintf_dsos_buildid(struct machines *machines, FILE *fp,
 613                                      bool (skip)(struct dso *dso, int parm), int parm)
 614 {
 615         struct rb_node *nd;
 616         size_t ret = machine__fprintf_dsos_buildid(&machines->host, fp, skip, parm);
 617
 618         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
 619                 struct machine *pos = rb_entry(nd, struct machine, rb_node);
 620                 ret += machine__fprintf_dsos_buildid(pos, fp, skip, parm);
 621         }
 622         return ret;
 623 }
 624
 625 size_t machine__fprintf_vmlinux_path(struct machine *machine, FILE *fp)
 626 {
 627         int i;
 628         size_t printed = 0;
 629         struct dso *kdso = machine__kernel_map(machine)->dso;
 630
 631         if (kdso->has_build_id) {
 632                 char filename[PATH_MAX];
 633                 if (dso__build_id_filename(kdso, filename, sizeof(filename)))
 634                         printed += fprintf(fp, "[0] %s\n", filename);
 635         }
 636
 637         for (i = 0; i < vmlinux_path__nr_entries; ++i)
 638                 printed += fprintf(fp, "[%d] %s\n",
 639                                    i + kdso->has_build_id, vmlinux_path[i]);
 640
 641         return printed;
 642 }
 643
 644 size_t machine__fprintf(struct machine *machine, FILE *fp)
 645 {
 646         size_t ret = 0;
 647         struct rb_node *nd;
 648
 649         pthread_rwlock_rdlock(&machine->threads_lock);
 650
 651         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
 652                 struct thread *pos = rb_entry(nd, struct thread, rb_node);
 653
 654                 ret += thread__fprintf(pos, fp);
 655         }
 656
 657         pthread_rwlock_unlock(&machine->threads_lock);
 658
 659         return ret;
 660 }
 661
 662 static struct dso *machine__get_kernel(struct machine *machine)
 663 {
 664         const char *vmlinux_name = NULL;
 665         struct dso *kernel;
 666
 667         if (machine__is_host(machine)) {
 668                 vmlinux_name = symbol_conf.vmlinux_name;
 669                 if (!vmlinux_name)
 670                         vmlinux_name = "[kernel.kallsyms]";
 671
 672                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 673                                                  "[kernel]", DSO_TYPE_KERNEL);
 674         } else {
 675                 char bf[PATH_MAX];
 676
 677                 if (machine__is_default_guest(machine))
 678                         vmlinux_name = symbol_conf.default_guest_vmlinux_name;
 679                 if (!vmlinux_name)
 680                         vmlinux_name = machine__mmap_name(machine, bf,
 681                                                           sizeof(bf));
 682
 683                 kernel = machine__findnew_kernel(machine, vmlinux_name,
 684                                                  "[guest.kernel]",
 685                                                  DSO_TYPE_GUEST_KERNEL);
 686         }
 687
 688         if (kernel != NULL && (!kernel->has_build_id))
 689                 dso__read_running_kernel_build_id(kernel, machine);
 690
 691         return kernel;
 692 }
 693
 694 struct process_args {
 695         u64 start;
 696 };
 697
 698 static void machine__get_kallsyms_filename(struct machine *machine, char *buf,
 699                                            size_t bufsz)
 700 {
 701         if (machine__is_default_guest(machine))
 702                 scnprintf(buf, bufsz, "%s", symbol_conf.default_guest_kallsyms);
 703         else
 704                 scnprintf(buf, bufsz, "%s/proc/kallsyms", machine->root_dir);
 705 }
 706
 707 const char *ref_reloc_sym_names[] = {"_text", "_stext", NULL};
 708
 709 /* Figure out the start address of kernel map from /proc/kallsyms.
 710  * Returns the name of the start symbol in *symbol_name. Pass in NULL as
 711  * symbol_name if it's not that important.
 712  */
 713 static u64 machine__get_running_kernel_start(struct machine *machine,
 714                                              const char **symbol_name)
 715 {
 716         char filename[PATH_MAX];
 717         int i;
 718         const char *name;
 719         u64 addr = 0;
 720
 721         machine__get_kallsyms_filename(machine, filename, PATH_MAX);
 722
 723         if (symbol__restricted_filename(filename, "/proc/kallsyms"))
 724                 return 0;
 725
 726         for (i = 0; (name = ref_reloc_sym_names[i]) != NULL; i++) {
 727                 addr = kallsyms__get_function_start(filename, name);
 728                 if (addr)
 729                         break;
 730         }
 731
 732         if (symbol_name)
 733                 *symbol_name = name;
 734
 735         return addr;
 736 }
 737
 738 int __machine__create_kernel_maps(struct machine *machine, struct dso *kernel)
 739 {
 740         enum map_type type;
 741         u64 start = machine__get_running_kernel_start(machine, NULL);
 742
 743         for (type = 0; type < MAP__NR_TYPES; ++type) {
 744                 struct kmap *kmap;
 745                 struct map *map;
 746
 747                 machine->vmlinux_maps[type] = map__new2(start, kernel, type);
 748                 if (machine->vmlinux_maps[type] == NULL)
 749                         return -1;
 750
 751                 machine->vmlinux_maps[type]->map_ip =
 752                         machine->vmlinux_maps[type]->unmap_ip =
 753                                 identity__map_ip;
 754                 map = __machine__kernel_map(machine, type);
 755                 kmap = map__kmap(map);
 756                 if (!kmap)
 757                         return -1;
 758
 759                 kmap->kmaps = &machine->kmaps;
 760                 map_groups__insert(&machine->kmaps, map);
 761         }
 762
 763         return 0;
 764 }
 765
 766 void machine__destroy_kernel_maps(struct machine *machine)
 767 {
 768         enum map_type type;
 769
 770         for (type = 0; type < MAP__NR_TYPES; ++type) {
 771                 struct kmap *kmap;
 772                 struct map *map = __machine__kernel_map(machine, type);
 773
 774                 if (map == NULL)
 775                         continue;
 776
 777                 kmap = map__kmap(map);
 778                 map_groups__remove(&machine->kmaps, map);
 779                 if (kmap && kmap->ref_reloc_sym) {
 780                         /*
 781                          * ref_reloc_sym is shared among all maps, so free just
 782                          * on one of them.
 783                          */
 784                         if (type == MAP__FUNCTION) {
 785                                 zfree((char **)&kmap->ref_reloc_sym->name);
 786                                 zfree(&kmap->ref_reloc_sym);
 787                         } else
 788                                 kmap->ref_reloc_sym = NULL;
 789                 }
 790
 791                 machine->vmlinux_maps[type] = NULL;
 792         }
 793 }
 794
 795 int machines__create_guest_kernel_maps(struct machines *machines)
 796 {
 797         int ret = 0;
 798         struct dirent **namelist = NULL;
 799         int i, items = 0;
 800         char path[PATH_MAX];
 801         pid_t pid;
 802         char *endp;
 803
 804         if (symbol_conf.default_guest_vmlinux_name ||
 805             symbol_conf.default_guest_modules ||
 806             symbol_conf.default_guest_kallsyms) {
 807                 machines__create_kernel_maps(machines, DEFAULT_GUEST_KERNEL_ID);
 808         }
 809
 810         if (symbol_conf.guestmount) {
 811                 items = scandir(symbol_conf.guestmount, &namelist, NULL, NULL);
 812                 if (items <= 0)
 813                         return -ENOENT;
 814                 for (i = 0; i < items; i++) {
 815                         if (!isdigit(namelist[i]->d_name[0])) {
 816                                 /* Filter out . and .. */
 817                                 continue;
 818                         }
 819                         pid = (pid_t)strtol(namelist[i]->d_name, &endp, 10);
 820                         if ((*endp != '\0') ||
 821                             (endp == namelist[i]->d_name) ||
 822                             (errno == ERANGE)) {
 823                                 pr_debug("invalid directory (%s). Skipping.\n",
 824                                          namelist[i]->d_name);
 825                                 continue;
 826                         }
 827                         sprintf(path, "%s/%s/proc/kallsyms",
 828                                 symbol_conf.guestmount,
 829                                 namelist[i]->d_name);
 830                         ret = access(path, R_OK);
 831                         if (ret) {
 832                                 pr_debug("Can't access file %s\n", path);
 833                                 goto failure;
 834                         }
 835                         machines__create_kernel_maps(machines, pid);
 836                 }
 837 failure:
 838                 free(namelist);
 839         }
 840
 841         return ret;
 842 }
 843
 844 void machines__destroy_kernel_maps(struct machines *machines)
 845 {
 846         struct rb_node *next = rb_first(&machines->guests);
 847
 848         machine__destroy_kernel_maps(&machines->host);
 849
 850         while (next) {
 851                 struct machine *pos = rb_entry(next, struct machine, rb_node);
 852
 853                 next = rb_next(&pos->rb_node);
 854                 rb_erase(&pos->rb_node, &machines->guests);
 855                 machine__delete(pos);
 856         }
 857 }
 858
 859 int machines__create_kernel_maps(struct machines *machines, pid_t pid)
 860 {
 861         struct machine *machine = machines__findnew(machines, pid);
 862
 863         if (machine == NULL)
 864                 return -1;
 865
 866         return machine__create_kernel_maps(machine);
 867 }
 868
 869 int machine__load_kallsyms(struct machine *machine, const char *filename,
 870                            enum map_type type, symbol_filter_t filter)
 871 {
 872         struct map *map = machine__kernel_map(machine);
 873         int ret = dso__load_kallsyms(map->dso, filename, map, filter);
 874
 875         if (ret > 0) {
 876                 dso__set_loaded(map->dso, type);
 877                 /*
 878                  * Since /proc/kallsyms will have multiple sessions for the
 879                  * kernel, with modules between them, fixup the end of all
 880                  * sections.
 881                  */
 882                 __map_groups__fixup_end(&machine->kmaps, type);
 883         }
 884
 885         return ret;
 886 }
 887
 888 int machine__load_vmlinux_path(struct machine *machine, enum map_type type,
 889                                symbol_filter_t filter)
 890 {
 891         struct map *map = machine__kernel_map(machine);
 892         int ret = dso__load_vmlinux_path(map->dso, map, filter);
 893
 894         if (ret > 0)
 895                 dso__set_loaded(map->dso, type);
 896
 897         return ret;
 898 }
 899
 900 static void map_groups__fixup_end(struct map_groups *mg)
 901 {
 902         int i;
 903         for (i = 0; i < MAP__NR_TYPES; ++i)
 904                 __map_groups__fixup_end(mg, i);
 905 }
 906
 907 static char *get_kernel_version(const char *root_dir)
 908 {
 909         char version[PATH_MAX];
 910         FILE *file;
 911         char *name, *tmp;
 912         const char *prefix = "Linux version ";
 913
 914         sprintf(version, "%s/proc/version", root_dir);
 915         file = fopen(version, "r");
 916         if (!file)
 917                 return NULL;
 918
 919         version[0] = '\0';
 920         tmp = fgets(version, sizeof(version), file);
 921         fclose(file);
 922
 923         name = strstr(version, prefix);
 924         if (!name)
 925                 return NULL;
 926         name += strlen(prefix);
 927         tmp = strchr(name, ' ');
 928         if (tmp)
 929                 *tmp = '\0';
 930
 931         return strdup(name);
 932 }
 933
 934 static bool is_kmod_dso(struct dso *dso)
 935 {
 936         return dso->symtab_type == DSO_BINARY_TYPE__SYSTEM_PATH_KMODULE ||
 937                dso->symtab_type == DSO_BINARY_TYPE__GUEST_KMODULE;
 938 }
 939
 940 static int map_groups__set_module_path(struct map_groups *mg, const char *path,
 941                                        struct kmod_path *m)
 942 {
 943         struct map *map;
 944         char *long_name;
 945
 946         map = map_groups__find_by_name(mg, MAP__FUNCTION, m->name);
 947         if (map == NULL)
 948                 return 0;
 949
 950         long_name = strdup(path);
 951         if (long_name == NULL)
 952                 return -ENOMEM;
 953
 954         dso__set_long_name(map->dso, long_name, true);
 955         dso__kernel_module_get_build_id(map->dso, "");
 956
 957         /*
 958          * Full name could reveal us kmod compression, so
 959          * we need to update the symtab_type if needed.
 960          */
 961         if (m->comp && is_kmod_dso(map->dso))
 962                 map->dso->symtab_type++;
 963
 964         return 0;
 965 }
 966
 967 static int map_groups__set_modules_path_dir(struct map_groups *mg,
 968                                 const char *dir_name, int depth)
 969 {
 970         struct dirent *dent;
 971         DIR *dir = opendir(dir_name);
 972         int ret = 0;
 973
 974         if (!dir) {
 975                 pr_debug("%s: cannot open %s dir\n", __func__, dir_name);
 976                 return -1;
 977         }
 978
 979         while ((dent = readdir(dir)) != NULL) {
 980                 char path[PATH_MAX];
 981                 struct stat st;
 982
 983                 /*sshfs might return bad dent->d_type, so we have to stat*/
 984                 snprintf(path, sizeof(path), "%s/%s", dir_name, dent->d_name);
 985                 if (stat(path, &st))
 986                         continue;
 987
 988                 if (S_ISDIR(st.st_mode)) {
 989                         if (!strcmp(dent->d_name, ".") ||
 990                             !strcmp(dent->d_name, ".."))
 991                                 continue;
 992
 993                         /* Do not follow top-level source and build symlinks */
 994                         if (depth == 0) {
 995                                 if (!strcmp(dent->d_name, "source") ||
 996                                     !strcmp(dent->d_name, "build"))
 997                                         continue;
 998                         }
 999
1000                         ret = map_groups__set_modules_path_dir(mg, path,
1001                                                                depth + 1);
1002                         if (ret < 0)
1003                                 goto out;
1004                 } else {
1005                         struct kmod_path m;
1006
1007                         ret = kmod_path__parse_name(&m, dent->d_name);
1008                         if (ret)
1009                                 goto out;
1010
1011                         if (m.kmod)
1012                                 ret = map_groups__set_module_path(mg, path, &m);
1013
1014                         free(m.name);
1015
1016                         if (ret)
1017                                 goto out;
1018                 }
1019         }
1020
1021 out:
1022         closedir(dir);
1023         return ret;
1024 }
1025
1026 static int machine__set_modules_path(struct machine *machine)
1027 {
1028         char *version;
1029         char modules_path[PATH_MAX];
1030
1031         version = get_kernel_version(machine->root_dir);
1032         if (!version)
1033                 return -1;
1034
1035         snprintf(modules_path, sizeof(modules_path), "%s/lib/modules/%s",
1036                  machine->root_dir, version);
1037         free(version);
1038
1039         return map_groups__set_modules_path_dir(&machine->kmaps, modules_path, 0);
1040 }
1041
1042 static int machine__create_module(void *arg, const char *name, u64 start)
1043 {
1044         struct machine *machine = arg;
1045         struct map *map;
1046
1047         map = machine__findnew_module_map(machine, start, name);
1048         if (map == NULL)
1049                 return -1;
1050
1051         dso__kernel_module_get_build_id(map->dso, machine->root_dir);
1052
1053         return 0;
1054 }
1055
1056 static int machine__create_modules(struct machine *machine)
1057 {
1058         const char *modules;
1059         char path[PATH_MAX];
1060
1061         if (machine__is_default_guest(machine)) {
1062                 modules = symbol_conf.default_guest_modules;
1063         } else {
1064                 snprintf(path, PATH_MAX, "%s/proc/modules", machine->root_dir);
1065                 modules = path;
1066         }
1067
1068         if (symbol__restricted_filename(modules, "/proc/modules"))
1069                 return -1;
1070
1071         if (modules__parse(modules, machine, machine__create_module))
1072                 return -1;
1073
1074         if (!machine__set_modules_path(machine))
1075                 return 0;
1076
1077         pr_debug("Problems setting modules path maps, continuing anyway...\n");
1078
1079         return 0;
1080 }
1081
1082 int machine__create_kernel_maps(struct machine *machine)
1083 {
1084         struct dso *kernel = machine__get_kernel(machine);
1085         const char *name;
1086         u64 addr = machine__get_running_kernel_start(machine, &name);
1087         if (!addr)
1088                 return -1;
1089
1090         if (kernel == NULL ||
1091             __machine__create_kernel_maps(machine, kernel) < 0)
1092                 return -1;
1093
1094         if (symbol_conf.use_modules && machine__create_modules(machine) < 0) {
1095                 if (machine__is_host(machine))
1096                         pr_debug("Problems creating module maps, "
1097                                  "continuing anyway...\n");
1098                 else
1099                         pr_debug("Problems creating module maps for guest %d, "
1100                                  "continuing anyway...\n", machine->pid);
1101         }
1102
1103         /*
1104          * Now that we have all the maps created, just set the ->end of them:
1105          */
1106         map_groups__fixup_end(&machine->kmaps);
1107
1108         if (maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps, name,
1109                                              addr)) {
1110                 machine__destroy_kernel_maps(machine);
1111                 return -1;
1112         }
1113
1114         return 0;
1115 }
1116
1117 static void machine__set_kernel_mmap_len(struct machine *machine,
1118                                          union perf_event *event)
1119 {
1120         int i;
1121
1122         for (i = 0; i < MAP__NR_TYPES; i++) {
1123                 machine->vmlinux_maps[i]->start = event->mmap.start;
1124                 machine->vmlinux_maps[i]->end   = (event->mmap.start +
1125                                                    event->mmap.len);
1126                 /*
1127                  * Be a bit paranoid here, some perf.data file came with
1128                  * a zero sized synthesized MMAP event for the kernel.
1129                  */
1130                 if (machine->vmlinux_maps[i]->end == 0)
1131                         machine->vmlinux_maps[i]->end = ~0ULL;
1132         }
1133 }
1134
1135 static bool machine__uses_kcore(struct machine *machine)
1136 {
1137         struct dso *dso;
1138
1139         list_for_each_entry(dso, &machine->dsos.head, node) {
1140                 if (dso__is_kcore(dso))
1141                         return true;
1142         }
1143
1144         return false;
1145 }
1146
1147 static int machine__process_kernel_mmap_event(struct machine *machine,
1148                                               union perf_event *event)
1149 {
1150         struct map *map;
1151         char kmmap_prefix[PATH_MAX];
1152         enum dso_kernel_type kernel_type;
1153         bool is_kernel_mmap;
1154
1155         /* If we have maps from kcore then we do not need or want any others */
1156         if (machine__uses_kcore(machine))
1157                 return 0;
1158
1159         machine__mmap_name(machine, kmmap_prefix, sizeof(kmmap_prefix));
1160         if (machine__is_host(machine))
1161                 kernel_type = DSO_TYPE_KERNEL;
1162         else
1163                 kernel_type = DSO_TYPE_GUEST_KERNEL;
1164
1165         is_kernel_mmap = memcmp(event->mmap.filename,
1166                                 kmmap_prefix,
1167                                 strlen(kmmap_prefix) - 1) == 0;
1168         if (event->mmap.filename[0] == '/' ||
1169             (!is_kernel_mmap && event->mmap.filename[0] == '[')) {
1170                 map = machine__findnew_module_map(machine, event->mmap.start,
1171                                                   event->mmap.filename);
1172                 if (map == NULL)
1173                         goto out_problem;
1174
1175                 map->end = map->start + event->mmap.len;
1176         } else if (is_kernel_mmap) {
1177                 const char *symbol_name = (event->mmap.filename +
1178                                 strlen(kmmap_prefix));
1179                 /*
1180                  * Should be there already, from the build-id table in
1181                  * the header.
1182                  */
1183                 struct dso *kernel = NULL;
1184                 struct dso *dso;
1185
1186                 pthread_rwlock_rdlock(&machine->dsos.lock);
1187
1188                 list_for_each_entry(dso, &machine->dsos.head, node) {
1189
1190                         /*
1191                          * The cpumode passed to is_kernel_module is not the
1192                          * cpumode of *this* event. If we insist on passing
1193                          * correct cpumode to is_kernel_module, we should
1194                          * record the cpumode when we adding this dso to the
1195                          * linked list.
1196                          *
1197                          * However we don't really need passing correct
1198                          * cpumode.  We know the correct cpumode must be kernel
1199                          * mode (if not, we should not link it onto kernel_dsos
1200                          * list).
1201                          *
1202                          * Therefore, we pass PERF_RECORD_MISC_CPUMODE_UNKNOWN.
1203                          * is_kernel_module() treats it as a kernel cpumode.
1204                          */
1205
1206                         if (!dso->kernel ||
1207                             is_kernel_module(dso->long_name,
1208                                              PERF_RECORD_MISC_CPUMODE_UNKNOWN))
1209                                 continue;
1210
1211
1212                         kernel = dso;
1213                         break;
1214                 }
1215
1216                 pthread_rwlock_unlock(&machine->dsos.lock);
1217
1218                 if (kernel == NULL)
1219                         kernel = machine__findnew_dso(machine, kmmap_prefix);
1220                 if (kernel == NULL)
1221                         goto out_problem;
1222
1223                 kernel->kernel = kernel_type;
1224                 if (__machine__create_kernel_maps(machine, kernel) < 0) {
1225                         dso__put(kernel);
1226                         goto out_problem;
1227                 }
1228
1229                 if (strstr(kernel->long_name, "vmlinux"))
1230                         dso__set_short_name(kernel, "[kernel.vmlinux]", false);
1231
1232                 machine__set_kernel_mmap_len(machine, event);
1233
1234                 /*
1235                  * Avoid using a zero address (kptr_restrict) for the ref reloc
1236                  * symbol. Effectively having zero here means that at record
1237                  * time /proc/sys/kernel/kptr_restrict was non zero.
1238                  */
1239                 if (event->mmap.pgoff != 0) {
1240                         maps__set_kallsyms_ref_reloc_sym(machine->vmlinux_maps,
1241                                                          symbol_name,
1242                                                          event->mmap.pgoff);
1243                 }
1244
1245                 if (machine__is_default_guest(machine)) {
1246                         /*
1247                          * preload dso of guest kernel and modules
1248                          */
1249                         dso__load(kernel, machine__kernel_map(machine), NULL);
1250                 }
1251         }
1252         return 0;
1253 out_problem:
1254         return -1;
1255 }
1256
1257 int machine__process_mmap2_event(struct machine *machine,
1258                                  union perf_event *event,
1259                                  struct perf_sample *sample __maybe_unused)
1260 {
1261         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1262         struct thread *thread;
1263         struct map *map;
1264         enum map_type type;
1265         int ret = 0;
1266
1267         if (dump_trace)
1268                 perf_event__fprintf_mmap2(event, stdout);
1269
1270         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1271             cpumode == PERF_RECORD_MISC_KERNEL) {
1272                 ret = machine__process_kernel_mmap_event(machine, event);
1273                 if (ret < 0)
1274                         goto out_problem;
1275                 return 0;
1276         }
1277
1278         thread = machine__findnew_thread(machine, event->mmap2.pid,
1279                                         event->mmap2.tid);
1280         if (thread == NULL)
1281                 goto out_problem;
1282
1283         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1284                 type = MAP__VARIABLE;
1285         else
1286                 type = MAP__FUNCTION;
1287
1288         map = map__new(machine, event->mmap2.start,
1289                         event->mmap2.len, event->mmap2.pgoff,
1290                         event->mmap2.pid, event->mmap2.maj,
1291                         event->mmap2.min, event->mmap2.ino,
1292                         event->mmap2.ino_generation,
1293                         event->mmap2.prot,
1294                         event->mmap2.flags,
1295                         event->mmap2.filename, type, thread);
1296
1297         if (map == NULL)
1298                 goto out_problem_map;
1299
1300         thread__insert_map(thread, map);
1301         thread__put(thread);
1302         map__put(map);
1303         return 0;
1304
1305 out_problem_map:
1306         thread__put(thread);
1307 out_problem:
1308         dump_printf("problem processing PERF_RECORD_MMAP2, skipping event.\n");
1309         return 0;
1310 }
1311
1312 int machine__process_mmap_event(struct machine *machine, union perf_event *event,
1313                                 struct perf_sample *sample __maybe_unused)
1314 {
1315         u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1316         struct thread *thread;
1317         struct map *map;
1318         enum map_type type;
1319         int ret = 0;
1320
1321         if (dump_trace)
1322                 perf_event__fprintf_mmap(event, stdout);
1323
1324         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL ||
1325             cpumode == PERF_RECORD_MISC_KERNEL) {
1326                 ret = machine__process_kernel_mmap_event(machine, event);
1327                 if (ret < 0)
1328                         goto out_problem;
1329                 return 0;
1330         }
1331
1332         thread = machine__findnew_thread(machine, event->mmap.pid,
1333                                          event->mmap.tid);
1334         if (thread == NULL)
1335                 goto out_problem;
1336
1337         if (event->header.misc & PERF_RECORD_MISC_MMAP_DATA)
1338                 type = MAP__VARIABLE;
1339         else
1340                 type = MAP__FUNCTION;
1341
1342         map = map__new(machine, event->mmap.start,
1343                         event->mmap.len, event->mmap.pgoff,
1344                         event->mmap.pid, 0, 0, 0, 0, 0, 0,
1345                         event->mmap.filename,
1346                         type, thread);
1347
1348         if (map == NULL)
1349                 goto out_problem_map;
1350
1351         thread__insert_map(thread, map);
1352         thread__put(thread);
1353         map__put(map);
1354         return 0;
1355
1356 out_problem_map:
1357         thread__put(thread);
1358 out_problem:
1359         dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
1360         return 0;
1361 }
1362
1363 static void __machine__remove_thread(struct machine *machine, struct thread *th, bool lock)
1364 {
1365         if (machine->last_match == th)
1366                 machine->last_match = NULL;
1367
1368         BUG_ON(atomic_read(&th->refcnt) == 0);
1369         if (lock)
1370                 pthread_rwlock_wrlock(&machine->threads_lock);
1371         rb_erase_init(&th->rb_node, &machine->threads);
1372         RB_CLEAR_NODE(&th->rb_node);
1373         /*
1374          * Move it first to the dead_threads list, then drop the reference,
1375          * if this is the last reference, then the thread__delete destructor
1376          * will be called and we will remove it from the dead_threads list.
1377          */
1378         list_add_tail(&th->node, &machine->dead_threads);
1379         if (lock)
1380                 pthread_rwlock_unlock(&machine->threads_lock);
1381         thread__put(th);
1382 }
1383
1384 void machine__remove_thread(struct machine *machine, struct thread *th)
1385 {
1386         return __machine__remove_thread(machine, th, true);
1387 }
1388
1389 int machine__process_fork_event(struct machine *machine, union perf_event *event,
1390                                 struct perf_sample *sample)
1391 {
1392         struct thread *thread = machine__find_thread(machine,
1393                                                      event->fork.pid,
1394                                                      event->fork.tid);
1395         struct thread *parent = machine__findnew_thread(machine,
1396                                                         event->fork.ppid,
1397                                                         event->fork.ptid);
1398         int err = 0;
1399
1400         if (dump_trace)
1401                 perf_event__fprintf_task(event, stdout);
1402
1403         /*
1404          * There may be an existing thread that is not actually the parent,
1405          * either because we are processing events out of order, or because the
1406          * (fork) event that would have removed the thread was lost. Assume the
1407          * latter case and continue on as best we can.
1408          */
1409         if (parent->pid_ != (pid_t)event->fork.ppid) {
1410                 dump_printf("removing erroneous parent thread %d/%d\n",
1411                             parent->pid_, parent->tid);
1412                 machine__remove_thread(machine, parent);
1413                 thread__put(parent);
1414                 parent = machine__findnew_thread(machine, event->fork.ppid,
1415                                                  event->fork.ptid);
1416         }
1417
1418         /* if a thread currently exists for the thread id remove it */
1419         if (thread != NULL) {
1420                 machine__remove_thread(machine, thread);
1421                 thread__put(thread);
1422         }
1423
1424         thread = machine__findnew_thread(machine, event->fork.pid,
1425                                          event->fork.tid);
1426
1427         if (thread == NULL || parent == NULL ||
1428             thread__fork(thread, parent, sample->time) < 0) {
1429                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
1430                 err = -1;
1431         }
1432         thread__put(thread);
1433         thread__put(parent);
1434
1435         return err;
1436 }
1437
1438 int machine__process_exit_event(struct machine *machine, union perf_event *event,
1439                                 struct perf_sample *sample __maybe_unused)
1440 {
1441         struct thread *thread = machine__find_thread(machine,
1442                                                      event->fork.pid,
1443                                                      event->fork.tid);
1444
1445         if (dump_trace)
1446                 perf_event__fprintf_task(event, stdout);
1447
1448         if (thread != NULL) {
1449                 thread__exited(thread);
1450                 thread__put(thread);
1451         }
1452
1453         return 0;
1454 }
1455
1456 int machine__process_event(struct machine *machine, union perf_event *event,
1457                            struct perf_sample *sample)
1458 {
1459         int ret;
1460
1461         switch (event->header.type) {
1462         case PERF_RECORD_COMM:
1463                 ret = machine__process_comm_event(machine, event, sample); break;
1464         case PERF_RECORD_MMAP:
1465                 ret = machine__process_mmap_event(machine, event, sample); break;
1466         case PERF_RECORD_MMAP2:
1467                 ret = machine__process_mmap2_event(machine, event, sample); break;
1468         case PERF_RECORD_FORK:
1469                 ret = machine__process_fork_event(machine, event, sample); break;
1470         case PERF_RECORD_EXIT:
1471                 ret = machine__process_exit_event(machine, event, sample); break;
1472         case PERF_RECORD_LOST:
1473                 ret = machine__process_lost_event(machine, event, sample); break;
1474         case PERF_RECORD_AUX:
1475                 ret = machine__process_aux_event(machine, event); break;
1476         case PERF_RECORD_ITRACE_START:
1477                 ret = machine__process_itrace_start_event(machine, event); break;
1478         case PERF_RECORD_LOST_SAMPLES:
1479                 ret = machine__process_lost_samples_event(machine, event, sample); break;
1480         case PERF_RECORD_SWITCH:
1481         case PERF_RECORD_SWITCH_CPU_WIDE:
1482                 ret = machine__process_switch_event(machine, event); break;
1483         default:
1484                 ret = -1;
1485                 break;
1486         }
1487
1488         return ret;
1489 }
1490
1491 static bool symbol__match_regex(struct symbol *sym, regex_t *regex)
1492 {
1493         if (sym->name && !regexec(regex, sym->name, 0, NULL, 0))
1494                 return 1;
1495         return 0;
1496 }
1497
1498 static void ip__resolve_ams(struct thread *thread,
1499                             struct addr_map_symbol *ams,
1500                             u64 ip)
1501 {
1502         struct addr_location al;
1503
1504         memset(&al, 0, sizeof(al));
1505         /*
1506          * We cannot use the header.misc hint to determine whether a
1507          * branch stack address is user, kernel, guest, hypervisor.
1508          * Branches may straddle the kernel/user/hypervisor boundaries.
1509          * Thus, we have to try consecutively until we find a match
1510          * or else, the symbol is unknown
1511          */
1512         thread__find_cpumode_addr_location(thread, MAP__FUNCTION, ip, &al);
1513
1514         ams->addr = ip;
1515         ams->al_addr = al.addr;
1516         ams->sym = al.sym;
1517         ams->map = al.map;
1518 }
1519
1520 static void ip__resolve_data(struct thread *thread,
1521                              u8 m, struct addr_map_symbol *ams, u64 addr)
1522 {
1523         struct addr_location al;
1524
1525         memset(&al, 0, sizeof(al));
1526
1527         thread__find_addr_location(thread, m, MAP__VARIABLE, addr, &al);
1528         if (al.map == NULL) {
1529                 /*
1530                  * some shared data regions have execute bit set which puts
1531                  * their mapping in the MAP__FUNCTION type array.
1532                  * Check there as a fallback option before dropping the sample.
1533                  */
1534                 thread__find_addr_location(thread, m, MAP__FUNCTION, addr, &al);
1535         }
1536
1537         ams->addr = addr;
1538         ams->al_addr = al.addr;
1539         ams->sym = al.sym;
1540         ams->map = al.map;
1541 }
1542
1543 struct mem_info *sample__resolve_mem(struct perf_sample *sample,
1544                                      struct addr_location *al)
1545 {
1546         struct mem_info *mi = zalloc(sizeof(*mi));
1547
1548         if (!mi)
1549                 return NULL;
1550
1551         ip__resolve_ams(al->thread, &mi->iaddr, sample->ip);
1552         ip__resolve_data(al->thread, al->cpumode, &mi->daddr, sample->addr);
1553         mi->data_src.val = sample->data_src;
1554
1555         return mi;
1556 }
1557
1558 static int add_callchain_ip(struct thread *thread,
1559                             struct symbol **parent,
1560                             struct addr_location *root_al,
1561                             u8 *cpumode,
1562                             u64 ip)
1563 {
1564         struct addr_location al;
1565
1566         al.filtered = 0;
1567         al.sym = NULL;
1568         if (!cpumode) {
1569                 thread__find_cpumode_addr_location(thread, MAP__FUNCTION,
1570                                                    ip, &al);
1571         } else {
1572                 if (ip >= PERF_CONTEXT_MAX) {
1573                         switch (ip) {
1574                         case PERF_CONTEXT_HV:
1575                                 *cpumode = PERF_RECORD_MISC_HYPERVISOR;
1576                                 break;
1577                         case PERF_CONTEXT_KERNEL:
1578                                 *cpumode = PERF_RECORD_MISC_KERNEL;
1579                                 break;
1580                         case PERF_CONTEXT_USER:
1581                                 *cpumode = PERF_RECORD_MISC_USER;
1582                                 break;
1583                         default:
1584                                 pr_debug("invalid callchain context: "
1585                                          "%"PRId64"\n", (s64) ip);
1586                                 /*
1587                                  * It seems the callchain is corrupted.
1588                                  * Discard all.
1589                                  */
1590                                 callchain_cursor_reset(&callchain_cursor);
1591                                 return 1;
1592                         }
1593                         return 0;
1594                 }
1595                 thread__find_addr_location(thread, *cpumode, MAP__FUNCTION,
1596                                            ip, &al);
1597         }
1598
1599         if (al.sym != NULL) {
1600                 if (sort__has_parent && !*parent &&
1601                     symbol__match_regex(al.sym, &parent_regex))
1602                         *parent = al.sym;
1603                 else if (have_ignore_callees && root_al &&
1604                   symbol__match_regex(al.sym, &ignore_callees_regex)) {
1605                         /* Treat this symbol as the root,
1606                            forgetting its callees. */
1607                         *root_al = al;
1608                         callchain_cursor_reset(&callchain_cursor);
1609                 }
1610         }
1611
1612         return callchain_cursor_append(&callchain_cursor, al.addr, al.map, al.sym);
1613 }
1614
1615 struct branch_info *sample__resolve_bstack(struct perf_sample *sample,
1616                                            struct addr_location *al)
1617 {
1618         unsigned int i;
1619         const struct branch_stack *bs = sample->branch_stack;
1620         struct branch_info *bi = calloc(bs->nr, sizeof(struct branch_info));
1621
1622         if (!bi)
1623                 return NULL;
1624
1625         for (i = 0; i < bs->nr; i++) {
1626                 ip__resolve_ams(al->thread, &bi[i].to, bs->entries[i].to);
1627                 ip__resolve_ams(al->thread, &bi[i].from, bs->entries[i].from);
1628                 bi[i].flags = bs->entries[i].flags;
1629         }
1630         return bi;
1631 }
1632
1633 #define CHASHSZ 127
1634 #define CHASHBITS 7
1635 #define NO_ENTRY 0xff
1636
1637 #define PERF_MAX_BRANCH_DEPTH 127
1638
1639 /* Remove loops. */
1640 static int remove_loops(struct branch_entry *l, int nr)
1641 {
1642         int i, j, off;
1643         unsigned char chash[CHASHSZ];
1644
1645         memset(chash, NO_ENTRY, sizeof(chash));
1646
1647         BUG_ON(PERF_MAX_BRANCH_DEPTH > 255);
1648
1649         for (i = 0; i < nr; i++) {
1650                 int h = hash_64(l[i].from, CHASHBITS) % CHASHSZ;
1651
1652                 /* no collision handling for now */
1653                 if (chash[h] == NO_ENTRY) {
1654                         chash[h] = i;
1655                 } else if (l[chash[h]].from == l[i].from) {
1656                         bool is_loop = true;
1657                         /* check if it is a real loop */
1658                         off = 0;
1659                         for (j = chash[h]; j < i && i + off < nr; j++, off++)
1660                                 if (l[j].from != l[i + off].from) {
1661                                         is_loop = false;
1662                                         break;
1663                                 }
1664                         if (is_loop) {
1665                                 memmove(l + i, l + i + off,
1666                                         (nr - (i + off)) * sizeof(*l));
1667                                 nr -= off;
1668                         }
1669                 }
1670         }
1671         return nr;
1672 }
1673
1674 /*
1675  * Recolve LBR callstack chain sample
1676  * Return:
1677  * 1 on success get LBR callchain information
1678  * 0 no available LBR callchain information, should try fp
1679  * negative error code on other errors.
1680  */
1681 static int resolve_lbr_callchain_sample(struct thread *thread,
1682                                         struct perf_sample *sample,
1683                                         struct symbol **parent,
1684                                         struct addr_location *root_al,
1685                                         int max_stack)
1686 {
1687         struct ip_callchain *chain = sample->callchain;
1688         int chain_nr = min(max_stack, (int)chain->nr);
1689         u8 cpumode = PERF_RECORD_MISC_USER;
1690         int i, j, err;
1691         u64 ip;
1692
1693         for (i = 0; i < chain_nr; i++) {
1694                 if (chain->ips[i] == PERF_CONTEXT_USER)
1695                         break;
1696         }
1697
1698         /* LBR only affects the user callchain */
1699         if (i != chain_nr) {
1700                 struct branch_stack *lbr_stack = sample->branch_stack;
1701                 int lbr_nr = lbr_stack->nr;
1702                 /*
1703                  * LBR callstack can only get user call chain.
1704                  * The mix_chain_nr is kernel call chain
1705                  * number plus LBR user call chain number.
1706                  * i is kernel call chain number,
1707                  * 1 is PERF_CONTEXT_USER,
1708                  * lbr_nr + 1 is the user call chain number.
1709                  * For details, please refer to the comments
1710                  * in callchain__printf
1711                  */
1712                 int mix_chain_nr = i + 1 + lbr_nr + 1;
1713
1714                 if (mix_chain_nr > PERF_MAX_STACK_DEPTH + PERF_MAX_BRANCH_DEPTH) {
1715                         pr_warning("corrupted callchain. skipping...\n");
1716                         return 0;
1717                 }
1718
1719                 for (j = 0; j < mix_chain_nr; j++) {
1720                         if (callchain_param.order == ORDER_CALLEE) {
1721                                 if (j < i + 1)
1722                                         ip = chain->ips[j];
1723                                 else if (j > i + 1)
1724                                         ip = lbr_stack->entries[j - i - 2].from;
1725                                 else
1726                                         ip = lbr_stack->entries[0].to;
1727                         } else {
1728                                 if (j < lbr_nr)
1729                                         ip = lbr_stack->entries[lbr_nr - j - 1].from;
1730                                 else if (j > lbr_nr)
1731                                         ip = chain->ips[i + 1 - (j - lbr_nr)];
1732                                 else
1733                                         ip = lbr_stack->entries[0].to;
1734                         }
1735
1736                         err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1737                         if (err)
1738                                 return (err < 0) ? err : 0;
1739                 }
1740                 return 1;
1741         }
1742
1743         return 0;
1744 }
1745
1746 static int thread__resolve_callchain_sample(struct thread *thread,
1747                                             struct perf_evsel *evsel,
1748                                             struct perf_sample *sample,
1749                                             struct symbol **parent,
1750                                             struct addr_location *root_al,
1751                                             int max_stack)
1752 {
1753         struct branch_stack *branch = sample->branch_stack;
1754         struct ip_callchain *chain = sample->callchain;
1755         int chain_nr = min(max_stack, (int)chain->nr);
1756         u8 cpumode = PERF_RECORD_MISC_USER;
1757         int i, j, err;
1758         int skip_idx = -1;
1759         int first_call = 0;
1760
1761         callchain_cursor_reset(&callchain_cursor);
1762
1763         if (has_branch_callstack(evsel)) {
1764                 err = resolve_lbr_callchain_sample(thread, sample, parent,
1765                                                    root_al, max_stack);
1766                 if (err)
1767                         return (err < 0) ? err : 0;
1768         }
1769
1770         /*
1771          * Based on DWARF debug information, some architectures skip
1772          * a callchain entry saved by the kernel.
1773          */
1774         if (chain->nr < PERF_MAX_STACK_DEPTH)
1775                 skip_idx = arch_skip_callchain_idx(thread, chain);
1776
1777         /*
1778          * Add branches to call stack for easier browsing. This gives
1779          * more context for a sample than just the callers.
1780          *
1781          * This uses individual histograms of paths compared to the
1782          * aggregated histograms the normal LBR mode uses.
1783          *
1784          * Limitations for now:
1785          * - No extra filters
1786          * - No annotations (should annotate somehow)
1787          */
1788
1789         if (branch && callchain_param.branch_callstack) {
1790                 int nr = min(max_stack, (int)branch->nr);
1791                 struct branch_entry be[nr];
1792
1793                 if (branch->nr > PERF_MAX_BRANCH_DEPTH) {
1794                         pr_warning("corrupted branch chain. skipping...\n");
1795                         goto check_calls;
1796                 }
1797
1798                 for (i = 0; i < nr; i++) {
1799                         if (callchain_param.order == ORDER_CALLEE) {
1800                                 be[i] = branch->entries[i];
1801                                 /*
1802                                  * Check for overlap into the callchain.
1803                                  * The return address is one off compared to
1804                                  * the branch entry. To adjust for this
1805                                  * assume the calling instruction is not longer
1806                                  * than 8 bytes.
1807                                  */
1808                                 if (i == skip_idx ||
1809                                     chain->ips[first_call] >= PERF_CONTEXT_MAX)
1810                                         first_call++;
1811                                 else if (be[i].from < chain->ips[first_call] &&
1812                                     be[i].from >= chain->ips[first_call] - 8)
1813                                         first_call++;
1814                         } else
1815                                 be[i] = branch->entries[branch->nr - i - 1];
1816                 }
1817
1818                 nr = remove_loops(be, nr);
1819
1820                 for (i = 0; i < nr; i++) {
1821                         err = add_callchain_ip(thread, parent, root_al,
1822                                                NULL, be[i].to);
1823                         if (!err)
1824                                 err = add_callchain_ip(thread, parent, root_al,
1825                                                        NULL, be[i].from);
1826                         if (err == -EINVAL)
1827                                 break;
1828                         if (err)
1829                                 return err;
1830                 }
1831                 chain_nr -= nr;
1832         }
1833
1834 check_calls:
1835         if (chain->nr > PERF_MAX_STACK_DEPTH && (int)chain->nr > max_stack) {
1836                 pr_warning("corrupted callchain. skipping...\n");
1837                 return 0;
1838         }
1839
1840         for (i = first_call; i < chain_nr; i++) {
1841                 u64 ip;
1842
1843                 if (callchain_param.order == ORDER_CALLEE)
1844                         j = i;
1845                 else
1846                         j = chain->nr - i - 1;
1847
1848 #ifdef HAVE_SKIP_CALLCHAIN_IDX
1849                 if (j == skip_idx)
1850                         continue;
1851 #endif
1852                 ip = chain->ips[j];
1853
1854                 err = add_callchain_ip(thread, parent, root_al, &cpumode, ip);
1855
1856                 if (err)
1857                         return (err < 0) ? err : 0;
1858         }
1859
1860         return 0;
1861 }
1862
1863 static int unwind_entry(struct unwind_entry *entry, void *arg)
1864 {
1865         struct callchain_cursor *cursor = arg;
1866         return callchain_cursor_append(cursor, entry->ip,
1867                                        entry->map, entry->sym);
1868 }
1869
1870 int thread__resolve_callchain(struct thread *thread,
1871                               struct perf_evsel *evsel,
1872                               struct perf_sample *sample,
1873                               struct symbol **parent,
1874                               struct addr_location *root_al,
1875                               int max_stack)
1876 {
1877         int ret = thread__resolve_callchain_sample(thread, evsel,
1878                                                    sample, parent,
1879                                                    root_al, max_stack);
1880         if (ret)
1881                 return ret;
1882
1883         /* Can we do dwarf post unwind? */
1884         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
1885               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
1886                 return 0;
1887
1888         /* Bail out if nothing was captured. */
1889         if ((!sample->user_regs.regs) ||
1890             (!sample->user_stack.size))
1891                 return 0;
1892
1893         return unwind__get_entries(unwind_entry, &callchain_cursor,
1894                                    thread, sample, max_stack);
1895
1896 }
1897
1898 int machine__for_each_thread(struct machine *machine,
1899                              int (*fn)(struct thread *thread, void *p),
1900                              void *priv)
1901 {
1902         struct rb_node *nd;
1903         struct thread *thread;
1904         int rc = 0;
1905
1906         for (nd = rb_first(&machine->threads); nd; nd = rb_next(nd)) {
1907                 thread = rb_entry(nd, struct thread, rb_node);
1908                 rc = fn(thread, priv);
1909                 if (rc != 0)
1910                         return rc;
1911         }
1912
1913         list_for_each_entry(thread, &machine->dead_threads, node) {
1914                 rc = fn(thread, priv);
1915                 if (rc != 0)
1916                         return rc;
1917         }
1918         return rc;
1919 }
1920
1921 int machines__for_each_thread(struct machines *machines,
1922                               int (*fn)(struct thread *thread, void *p),
1923                               void *priv)
1924 {
1925         struct rb_node *nd;
1926         int rc = 0;
1927
1928         rc = machine__for_each_thread(&machines->host, fn, priv);
1929         if (rc != 0)
1930                 return rc;
1931
1932         for (nd = rb_first(&machines->guests); nd; nd = rb_next(nd)) {
1933                 struct machine *machine = rb_entry(nd, struct machine, rb_node);
1934
1935                 rc = machine__for_each_thread(machine, fn, priv);
1936                 if (rc != 0)
1937                         return rc;
1938         }
1939         return rc;
1940 }
1941
1942 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool,
1943                                   struct target *target, struct thread_map *threads,
1944                                   perf_event__handler_t process, bool data_mmap,
1945                                   unsigned int proc_map_timeout)
1946 {
1947         if (target__has_task(target))
1948                 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap, proc_map_timeout);
1949         else if (target__has_cpu(target))
1950                 return perf_event__synthesize_threads(tool, process, machine, data_mmap, proc_map_timeout);
1951         /* command specified */
1952         return 0;
1953 }
1954
1955 pid_t machine__get_current_tid(struct machine *machine, int cpu)
1956 {
1957         if (cpu < 0 || cpu >= MAX_NR_CPUS || !machine->current_tid)
1958                 return -1;
1959
1960         return machine->current_tid[cpu];
1961 }
1962
1963 int machine__set_current_tid(struct machine *machine, int cpu, pid_t pid,
1964                              pid_t tid)
1965 {
1966         struct thread *thread;
1967
1968         if (cpu < 0)
1969                 return -EINVAL;
1970
1971         if (!machine->current_tid) {
1972                 int i;
1973
1974                 machine->current_tid = calloc(MAX_NR_CPUS, sizeof(pid_t));
1975                 if (!machine->current_tid)
1976                         return -ENOMEM;
1977                 for (i = 0; i < MAX_NR_CPUS; i++)
1978                         machine->current_tid[i] = -1;
1979         }
1980
1981         if (cpu >= MAX_NR_CPUS) {
1982                 pr_err("Requested CPU %d too large. ", cpu);
1983                 pr_err("Consider raising MAX_NR_CPUS\n");
1984                 return -EINVAL;
1985         }
1986
1987         machine->current_tid[cpu] = tid;
1988
1989         thread = machine__findnew_thread(machine, pid, tid);
1990         if (!thread)
1991                 return -ENOMEM;
1992
1993         thread->cpu = cpu;
1994         thread__put(thread);
1995
1996         return 0;
1997 }
1998
1999 int machine__get_kernel_start(struct machine *machine)
2000 {
2001         struct map *map = machine__kernel_map(machine);
2002         int err = 0;
2003
2004         /*
2005          * The only addresses above 2^63 are kernel addresses of a 64-bit
2006          * kernel.  Note that addresses are unsigned so that on a 32-bit system
2007          * all addresses including kernel addresses are less than 2^32.  In
2008          * that case (32-bit system), if the kernel mapping is unknown, all
2009          * addresses will be assumed to be in user space - see
2010          * machine__kernel_ip().
2011          */
2012         machine->kernel_start = 1ULL << 63;
2013         if (map) {
2014                 err = map__load(map, machine->symbol_filter);
2015                 if (map->start)
2016                         machine->kernel_start = map->start;
2017         }
2018         return err;
2019 }
2020
2021 struct dso *machine__findnew_dso(struct machine *machine, const char *filename)
2022 {
2023         return dsos__findnew(&machine->dsos, filename);
2024 }
2025
2026 char *machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
2027 {
2028         struct machine *machine = vmachine;
2029         struct map *map;
2030         struct symbol *sym = map_groups__find_symbol(&machine->kmaps, MAP__FUNCTION, *addrp, &map,  NULL);
2031
2032         if (sym == NULL)
2033                 return NULL;
2034
2035         *modp = __map__is_kmodule(map) ? (char *)map->dso->short_name : NULL;
2036         *addrp = map->unmap_ip(map, sym->start);
2037         return sym->name;
2038 }